Process Measurement and Control,!

On-line Estimation
and Adaptive
Control of Bioreactors

G. Bastin and D. Dochain

Laboratoire d'Automatique, Dynamique et Analyse des Systomes, Universite
Catholique de Louvain, 1348 Louvain-la-Neuve, Belgium

ELSEVIER
Amsterdam — Oxford — New York — Tokyo 1990
ELSEVIER SCIENCE PUBLISHERS B.V.
Sara Burgerhartstraat 25
P.O. Box 2 1 1 , 1000 AE Amsterdam, The Netherlands

Distributors for the United States and Canada:

ELSEVIER SCIENCE PUBLISHING COMPANY INC.

655, Avenue of the Americas
New York, NY 10010, U.S.A.

ISBN 0­444­88430­0

© Elsevier Science Publishers B. V., 1990

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system or
transmitted in any form or by any means, electronic, mechanical, photocopying, recording or
otherwise, without the prior written permission of the publisher, Elsevier Science Publishers B.V./
Physical Sciences & Engineering Division, P.O. Box 330, 1000 AH Amsterdam, The Netherlands.

Special regulations for readers in the USA ­ This publication has been registered with the Copyright
Clearance Center Inc. (CCC), Salem, Massachusetts. Information can be obtained from the CCC
about conditions under which photocopies of parts of this publication may be made in the USA. All
other copyright questions, including photocopying outside of the USA, should be referred to the
publisher.

No responsibility is assumed by the Publisher for any injury and/or damage to persons or property
as a matter of products liability, negligence or otherwise, or from any use or operation of any meth­
ods, products, instructions or ideas contained in the material herein.

Printed in The Netherlands

To Chantal, Kevin, Simon and Arnaud
To Eveline, Frederic and Antoine
 PREFACE

The development of modern control science in biotechnology has been

hampered by two important obstacles.

First, since bioprocesses involve living organisms, their dynamics are often
poorly understood, strongly nonlinear and non­stationary. The reproducibility
of experiments is uncertain. The model parameters do not remain constant
over long periods, due to metabolic variations and physiological modifications.

Another essential difficulty lies in the absence, in most cases, of cheap and
reliable instrumentation suited to real time monitoring. To date, the market
offers very few sensors capable of providing reliable, direct, on­line
measurements of the biological variables (such as biomass or metabolite
concentration) required to implement high performance control strategies.

The aim of this book is to present an integrated theoretical framework which

overcomes both of these difficulties. The notion of minimal modelling will be
central throughout the book. We shall show how efficient monitoring and
control algorithms can be designed, in a systematic and rigourous way, from a
minimal knowledge of the process kinetics. In particular, a great part of the text
will be devoted to the design of software sensors for the on­line monitoring of
biological variables and reaction kinetics, which are capable of coping with the
lack of instrumental sensors just mentioned.

One of our objectives is also to acquaint the reader with the application of
mathematical modelling techniques and dynamical systems analysis in the
solution of engineering problems in bioreactors. The methodology is
abundantly illuminated and illustrated by a variety of practical examples drawn
via PPE^

from the experience we have gained in collaboration with our colleagues of

the bioengineering laboratories within the Biotechnology Action Programme of
the Commission of the European Communities, and also of the Research
Institute of Biotechnolgy in Montreal, Canada.

The book is the result of an intensive joint research effort by the authors during
the last decade. (The authors' names therefore appear in a purely alphabetical
order.) It is intended as a graduate level text for students of bioengineering as
well as a reference text for scientists and engineers involved in the design and
optimisation of bioprocesses.

In the first chapter the mathematical framework necessary for the analysis of
bioreactor dynamics is established. In particular, it is shown how a general
dynamical model of a biotechnological process may be derived from the
reaction network. The main mathematical properties of this general dynamical
model are also analysed.

A review of the estimation and control problems that we address in the

remaining chapters is given in Chapter 2.

In Chapter 3 the design of state observers and on­line kinetic estimators is

examined in great detail under the assumption that the yield coefficients of the
process are known. The same issues are discussed in Chapter 4, but on the
basis of the yield coefficients being unknown.

Finally, Chapter 5 deals with adaptive control of bioreactors by combining

feedback linearization techniques with the observers and estimators
presented in the previous chapters.

The writing of this book would not have been possible without the support of
the Biotechnology Action Programme of the European Communities which is
gratefully aknowledged. In particular, we would like to thank Dr Economidis
for his constant interest in our work. We are also indebted to G. Corrieu, C.
Beal, P. Louvet and E. Spinnler from the Institut National de Recherche
Agronomique (France), C. Sola and M. Poch from the Universitat Autonoma de
Barcelona (Spain), A. Cheruy, M.P. Bernier, J.F. Beteau, R. Montellano and C.
Vialas from the Institut National Polytechnique de Grenoble (France), J.C.
 PREFACE

Duarte and E. Ferreira from the Laboratorio Nacional de Engenheria et

Technologia Industrial (Portugal), A. Rozzi from the Politecnico di Milano
(Italy), M.C. Tomei of the Istituto de Ricerce sulle Acque ­ CNR (Italy).

Our interest in the application of systems analysis to biological processes goes

back to the late seventies. We are particularly indebted to M. Installe, who
introduced us to the subject, and to our colleagues of the Unit of
Bioengineering in Louvain­la­Neuve : H. Naveau, E.J. Nyns, D. Poncelet and
P. Renard.

During the writing of this book we also had the pleasure of interacting with
many people, who helped us to penetrate the mysteries of biology and of
adaptive systems and gave us useful hints, insights, advice and criticism. We
would particularly like to thank M.Y. Andersen, G. Andre, J.P. Axelsson, J.P.
Babary, B. Bitmead, G. Campion, G. Chamilothoris, C. Chavarie, L. Chen, B.
Coupal, B. Dahou, S. Dasgupta, P. De Larminat, M. Dewan, L. Dugard, Y.
Goma, R. Gorez, A.M. Guillaume, M. Haest, L Joassin, S.B. Jorgensen, J.
Levine, I. Mareels, A.J. Morris, P. Peringer, Y. Pomerleau, L. Praly, Y. Prigent,
Y. Sevely, E. Sinvitu, V. Wertz.

Part of the book has been written during the stays of D. Dochain at the Ecole
Polytechnique de Montreal (Canada) in 1987­88 and at the LAAS (Laboratoire
d' Automatique et d'Analyse des Systemes) of the CNRS in Toulouse (France)
during the spring of 1989 : these institutions are gratefully acknowledged.

Several people have read the manuscript in its various versions and have
given us many valuable comments. Many thanks are due to A. Cheruy, M.
Gevers, D. Johnson, A. Pauss, M. Perrier and V. Van Breusegem.

The typing of the many successive versions of the manuscript was patiently
and impeccably carried out by M. Termolle. Without her constant good
humour, even when our whims were excessive, the book would never have
been completed. We also want to thank V. Vermeulen and G. Donders for
χ

their invaluable help in preparing the manuscript and in helping us where we

were baffled by one or other feature of the Macintosh. We finally would like to
thank M.D. Dochain for the design of the figure on the front cover.

Louvain-la-Neuve, 1 January 1990

Georges Bastin and Denis Dochain
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS

CHAPTER 1

DYNAMICAL MODELS OF BIOREACTORS

1.0­ Introduction

Basically, a bioreactor is a tank in which several biological reactions occur

simultaneously in a liquid medium. A standard schematic diagram of a
completely mixed continuous stirred tank (CST) bioreactor is shown in Fig.1.1.
The biological reactions which are involved in the process may be roughly
classified into two categories : microbial growth reactions (often referred to as
microbiological reactions) and enzyme catalysed reactions (also termed
biochemical reactions or biotransformations).

The growth of the microorganisms (bacteria, yeasts, etc.) proceeds by

consumption of appropriate nutrients or substrates (involving carbon, nitrogen,
oxygen, etc.) provided the environmental conditions (temperature, pH, etc.) are
favourable. The mass of living microorganisms or living cells is called the
biomass.

Associated with cell growth, but often proceeding at a different rate, are the
enzyme catalysed reactions in which some reactants are transformed into
products (sometimes called metabolites) through the catalytic action of
intracellular or extracellular enzymes.
 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Once a bioreactor has been designed, one of the main challenges of the
bioengineer is the implementation of efficient strategies for on­line monitoring
and control of the process. Our main concern in this book will be to present
(and to illustrate with practical applications) a general methodology, based on
mathematical modelling and dynamical systems analysis, for the solution of
such engineering problems.

In this first chapter we introduce mathematical models of the dynamics of

bioreactors. The objective is not to provide a comprehensive overview of the
modelling problem of biotechnological processes. The goal is rather to
establish a firm background for the design and the analysis of estimation and
control algorithms which will be discussed in subsequent chapters.

Outline of the chapter

In the first three sections we follow the conventional route for the description of
dynamical models of stirred tank bioreactors. The basic model of the growth of
a single microorganism population on a single substrate is presented in
Section 1.1. This model is shown to be valid for different operating modes :
batch, fed­batch, continuous operations. The way in which the basic model
may be extended for complex cultures is outlined in Section 1.2, using the
example of product synthesis and that of oxygen consumption in aerobic
fermentations. The most commonly used kinetic models of the specific growth
rate, which is a key parameter for bioreactor description, are reviewed in
Section 1.3.

From Section 1.4 we move to a more general viewpoint. Our aim is to establish
a global and rigorous theoretical framework for the analysis of bioreactor
dynamics. The concept of "reaction scheme" of a biotechnological process is
introduced and illustrated with four practical examples : yeast fermentation,
anaerobic digestion, production of PHB acid, and lactic fermentation. Then, in
Section 1.5, it is shown that, once the reaction scheme of a biotechnological
process is given, the derivation of a general state space model for the process
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS

is made fully automatic. Several practical examples are described in Section

1.6.

Sections 1.7 to 1.9 are devoted to the mathematical properties of the general
state space model introduced in Section 1.5.

A basic algebraic structural property which will be critical for both estimation
and control design is presented in Section 1.7.

The general state­space model can be fairly complex and may involve a large
number of differential equations. But there are many practical applications
where a simplified model is sufficient from an engineering viewpoint. Model
reduction can be achieved via the technique of singular perturbations : this
leads to a general rule for the reduction of bioreactor models presented in
Section 1.8.

The stability of the general state­space model of bioreactors is analyzed in

Section 1.9. Two issues are discussed : global bounded­input bounded­state
stability and local stability of equilibrium points.

Finally, the extension of the state space model to more general situations than
the stirred tank bioreactors (such as fixed bed reactors and recycle reactors) is
considered in Section 1.10.

1.1. The Basic Dynamics of Microbial Growth in Stirred Tank

Reactors

In stirred tank reactors, the process is assumed to be in a completely mixed

condition : this implies that the composition of the medium is homogeneous in
the reactor.
 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

The dynamical behaviour of the growth of one population of microorganisms

on a single limiting substrate in a stirred tank reactor (Fig.1.1) is most often
expressed by equations (1.1), which are obtained from straightforward mass
balances.

Fig.1.1. Stirred tank bioreactor

The net accumulation of biomass in the reactor

d(VX)
­μ ν Χ ­FoutX (1.1.a)
dt

The net accumulation of substrate in the reactor

d(VS)
= ­k1μ VX + Fi„Sin ­ F o ^ S (1.1.b)
dt

The variation of volume

— 'in 'out (1.1.c)

with X being the concentration of the microorganism population (the biomass)

in the reactor and in the effluent, S the substrate concentration in the reactor
and in the effluent, Sjn the substrate concentration in the influent, Fmthe
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS

influent flow rate, Fout the effluent flow rate, μ the specific microbial growth
rate, ki the yield coefficient of the substrate consumption by the biomass, and
V the volume of the culture medium.

In these equations the only modelling assumption is that the biomass growth
term (μ Χ ) and the substrate consumption term (Κ ι μ Χ ) are proportional to the
biomass concentration X. This assumption has been validated many times
and has become commonly accepted since Monod introduced it in 1942.

Two additional terms are sometimes included in equations (1.1.a­b).

­ A decay term (­kdVX) in the biomass growth equation (1.1.a) to account for
the natural death of microorganisms.
­ A maintenance term (­kmVX) in the substrate consumption equation (1.1 .b) to
account for that part of the substrate used for biomass survival.

With these terms, equations (1.1 .a­b) may be rewritten as :

d(VX)
= (μ ­kd)VX­FoutX (1.2.a)
dt

d(VS)
= ­ ( k 1 μ + k J V X + Fi,Sin­FoutS (1.2.b)
dt

However the coefficients kd and km are considered to be negligible in many

industrial fermentations. Therefore, except where otherwise specified (see for
instance Section 1.4.3), they will frequently be omitted in this book.

On the other hand, by defining the dilution rate :

a useful alternative formulation of equations (1.1.a­c) is obtained as follows :

 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

dX
:(μ ­0)Χ (1.3.a)
dt

dS
:­Κ ι μ Χ + 0(5|η ­8) (1.3.b)
dt
dV
DV­Fout (1.3.C)
dt

We now describe three particular cases of this model corresponding to the

most frequently encountered operating conditions : batch, fed­batch and
continuous operations.

a) The batch reactor

A batch reactor is a reactor with neither inflow nor outflow :

Fin=Fo,t = 0 (1.4)

The tank is initially filled with a large amount of substrate and a small amount
of biomass (the inoculum). No substrate is introduced during the fermentation,
which is stopped when enough substrate has been consumed. The total
amount of biomass produced (and possibly of by­products) is then harvested.
Clearly the culture volume is constant and the dynamical model is given by
(1.3.a­b) with D = 0.

b) The fed-batch reactor

A fed­batch reactor is a reactor with no outflow :

Fout=0 (1.5)
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS

The tank initially contains a small amount of both substrate and biomass and is
progressively filled with the influent substrate. The model is then given by
(1.3.a,b,c) with Fout = 0

c) The continuous stirred tank reactor (CSTR)

In the continuous cultivation of microorganisms, the reactor is continuously fed

with the substrate influent. The rate of outflow is equal to the rate of inflow and
the volume of culture remains constant:

Fin = Fout = F (1­6)

The model is therefore given by (1.3.a­b) with :

D = ­^ (1.8)

1.2. Extensions to the Basic Dynamics

The situation encountered in many fermentation applications is much more

complex than the one described by the basic dynamical model of Section 1 . 1 . :
several biochemical and microbial growth reactions can coexist in the
bioreactor, each of them possibly involving several limiting substrates and
several reaction products or metabolites. In these cases, other dynamical
equations need to be introduced in order to make the dynamical description of
the process more complete. Later in this chapter we shall present a general
theoretical framework designed to describe the complexity of the entire
process for a wide range of potential applications. In this section we limit
ourselves to the introduction of the issue with two specific situations :
 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

­ the formation of an extracellular synthesis product

­ the dynamics of dissolved oxygen in aerobic fermentations.

Formation of a synthesis product

The growth of microorganisms in bioreactors is often accompanied by the

formation of products which are either soluble in the culture or which are given
off in gaseous form. The mass balance relative to the product in the bioreactor
is given by :

^ = O X ­ D P ­ Q (1.9)
dt

with Ρ being the synthesis product concentration (in the liquid phase), Q the
rate of mass outflow of the product from the reactor in gaseous form, and υ the
specific production rate.

The term υ Χ in (1.9) represents the rate of product formation : it expresses the
fact that the production is, in some way, "catalysed" by the biomass X.

In some instances (e.g. methane CH4), the liquid concentration is negligible

(P^O). The gaseous outflow rate is then usually considered as being equal to
the production rate and is written as follows :

0 = υ Χ (1.10)

An important special case arises when the product formation is "growth­

associated", i.e. the specific production rate is assumed to be proportional to
the specific growth rate :

υ = Κ 2μ (1.11)

with k2 a yield coefficient.

 Chap.1. DYNAMICAL MODELS OF BIOREACTORS

However, the specific production rate may also be completely or partially

independent of the specific growth rate. A classical example is the lactic
fermentation for which Luedeking and Piret (1959) have proposed the
following expression :

υ = k2μ + p (1.12)

where ρ is the non­growth associated specific production rate.

Dissolved

Aerobic fermentations are processes in which the microorganisms need

oxygen in order to develop properly. Typical examples are yeast growth
processes and activated sludge processes. In such cases dissolved oxygen
in the culture medium can be considered as an additional substrate.

The dissolved oxygen (DO) mass balance in the bioreactor is described as

follows :

^ = OTR­OUR­DC (1.13)
dt

where C is the DO concentration in the reactor, OTR is the oxygen transfer rate
and OUR is the oxygen uptake rate.

The oxygen uptake rate OUR obviously depends on the growth of the biomass.
This is often expressed as follows :

0UR = k3μ X (1.14)

with ka being a yield coefficient. A term proportional to the biomass

concentration (krX) is sometimes included in equation (1.14) to account for the
endogenous respiration.
 1ο ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

By using a line of reasoning based on Henry's law to model the liquid­gas

transfer dynamics, the oxygen transfer rate, OTR, is expressed as follows :

OTR = k L a ( C s ­ C ) (1.15)

where kia is the mass transfer coefficient and Cs is the oxygen saturation
concentration.

However expression (1.15) is often useless because the coefficients Cs and

kLa may be unknown and may vary greatly with time : it is well known that the
oxygen saturation concentration Cs depends on variables such as the oxygen
partial pressure in the surrounding atmosphere, temperature, salinity and
concentration of surfactants in the liquid, and that factors such as the type and
geometry of the aerator or the air flow rate determine the value of kLa.

Fortunately, in most industrial applications, the input and output gaseous

oxygen flow rates are measured on­line. Hence, if the liquid­gas transfer
dynamics are negligible (as is often assumed), the OTR can simply be
expressed from the gaseous oxygen balance :

0TR = Q i ­ Q 2 (1.16)

where Qi and Q2 are respectively the input and output oxygen flow rates (per
volume unit).

1.3. Models of the Specific Growth Rate

It can clearly be seen from equations (1.3.a­b), (1.11) and (1.14) that the
specific growth rate μ is a key parameter for the description of biomass growth,
substrate consumption and product formation. Biochemical experiments
carried out over more than half a century on pure cultures as well as on open
cultures (with non­sterile substrates) have clearly indicated that the parameter
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 11

μ varies with time and is influenced by many physico­chemical and biological

environmental factors among which the most important ones are : substrate
concentration, biomass concentration, product concentration, pH, temperature,
dissolved oxygen concentration, light intensity and various inhibitors of
microbial growth.

The specific growth rate is then commonly expressed by the multiplication of

individual terms, each of them referring to one of the influencing factors :

μ ω = μ (8) μ (Χ ) μ (Ρ ) μ (ρ Η ) μ (Τ ) μ ( 0 μ ί ϋ μ (Ι )... (1.17)

where Χ , S, Ρ , C have been defined above while Τ , L, and I refer to

temperature, light intensity and inhibitor concentration respectively.

We shall present, in the following paragraphs, some of the most commonly

used kinetic models for the different terms of equation (1.17). A list including
more than fifty different growth rate structures may be found in Appendix 1.

Influence of the S

The most widespread analytical specific growth rate model is certainly the
"Michaelis­Menten law", also often called the "Monod law", which expresses
the dependence of μ on the substrate concentration S as follows (Fig. 1.2) :

where μ * is the maximum growth rate and KM is the "Michaelis­Menten"

constant.

In fact, this expression was initially proposed by Michaelis and Menten in 1913
and physically justified by Briggs and Haldane later, in 1925, to express the
reaction rate of enzyme­catalysed reactions with a single substrate. It was
 12 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

extended by Monod in 1942 to the case of microorganism growth, but without

any specific physical justification.

!^
O

CO

Substrate Concentration S
Fig. 1.2. The Monod law

Δ * ·· »· »» iZT
• · >» >· »· / y

Mgrn cf^ G/ucose parhtre\

25 50 75" 100 25 1^0

Fig.1.3. Fitting of the Monod law to experimental data (reprinted by

permission from J. Monod (1942). Recherches suria
Croissance des Cultures Bacterierines. Hermann, Paris).
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 13

Expression (1.19) was adopted by Monod because it fits experimental data

well (see, for instance, Fig.1.3). But, as Monod himself recognized, "toute
courbe d'allure sigmo'ide pourrait etre ajustee aux donnees experimentales"
(any sigmoidal curve could be fitted to the experimental data). Besides,
another expression was suggested by Tessier in the same year of 1942 :

μ (8) = μ * 1 ­ exp (1.19)

KM

Clearly this equation could fit the Monod data equally well. Many different
(and more or less esoteric) formulas have been proposed since then and
these are partially listed in Appendix 1.

A drawback of the Monod or Teissier laws is that they do not allow any account
to be taken of possible substrate inhibitory effects at high concentrations
(overloading).

0.3 Η
= 6.3 h­^
Φ

0.2 Η Κ , = 0.1 g/l

I
Ο
ο
»^
ο
&
(Ο

1 2 3 4 5
Substrate Concentration S
Fig. 1.4. The Haldane law
 14 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Andrews suggested that substrate inhibition be treated by the "Haldane law"

which was initially derived by Haldane to describe inhibition in enzyme­
substrate reactions (Fig.1.4) :

μ (8) = ^ — — (1.20.a)
KM + S­f S^/K,

where K| is the "inhibition parameter and :

μ ο =μ (1.20.b)
(•·^/f)

If the substrate inhibition is neglected, the Haldane law reduces to the Monod
law.

Influence of the biomass concentration X

The biomass growth is often presumed to be slowed down at high biomass

concentration (and this has been experimentally observed in particular
instances). A simple model that accomodates for this situation assumes that
the specific growth rate decreases linearly with the biomass concentration :

μ (Χ ) = μ *(1 ­ a X ) (1.21)

where μ * is the maximum growth rate and a the inhibition constant. It is often
called the "logistic model" and was proposed by Verhulst in 1838. Another
model which is a function of both S and X is the following :

with Kc constant.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 15

μ * = 0.3 h"^

2 4
Biomass Concentration X
Fig. 1.5. The Contois model

This model was proposed by Contois in 1959, and is illustrated in Fig. 1.5,
which shows the inhibition dependence of μ with respect to X (for constant S).

Influence of the synthesis product concentration Ρ

It is a well known fact that, in particular fermentations, the synthesis product

can also inhibit the biomass growth. Typical examples are alcoholic or
ethanolic fermentations on glucose for which the following models have been
suggested (Fig. 1.6) :

Kp
μ (Ρ ) = (1.23)
Kp+P

μ (P) = exp(­K.P) (1.24)

with Kp constant.
 16 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Kp= 1 g/l

1 2 3 4
Product Concentration Ρ
Fig. 1.6. μ (Ρ ) = exp(­Kp/P)

Influence of pH and temperature

As we have indicated in the introduction, the biomass growth can actually take
place only if pH and temperature lie within (usually small) ranges of
admissible values.

In anaerobic digestion, for instance, the process is known to operate correctly

only for nearly neutral pH (= 7). For this process, Rozzi proposes treating the
influence of pH by a parabolic law derived from experimental data (Fig.1.7) :

μ (pH) = a p H ^ + b p H + c (1.26)

with a, b, c constants.

In a similar case, Jackson and Edwards suggest using the Haldane function
(1.21) in terms of hydrogen ion concentration :
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 17

Η "
μ (Η +) = (1.27)
K M + H ^ + K|(H*)^

with H+ representing the hydrogen ion concentration.

1.0 i

0.8 ­ \ ® = ­1
(y \ ^ = 14
H 0,6 ­ \ ^ = ­48
I 0.4 ­
δ
ο
Ξ 0.2­
i 0,0 ­
5.5 6.0 6.5 7,0 7,5 8.0 8.5
pH
Fig.1.7. μ (ρ Η )

On the other hand, the influence of temperature is most often modelled by an

Arrhenius­type law, as has been done, for instance, by Topiwala and Sinclair :

μ (Τ ) = a^expC­Ei/RT) ­ a2exp(­E2/RT) ­ b if <Τ <

=0 ifT<TiorT>T2 (1.28)

with E i , E2 : activation energies

R : gas constant
a i , a2, b : constants

This expression shows that the specific growth rate increases continuously
with temperature up to a maximum value T2 (at which the cells die) (see
Fig. 1.8).
 18 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

0.8 h = 273 κ , T2 = 320 Κ

=L
Ε = 16000 J/g.mole
Β
CD
0,6 h Eg = 34500 J/g.mole
a: b = 0.037
? 0.4 h a , = 2.4x10
2 1 24
ο a^ = 1.02x10
ο
!^ 0.2 h
Ο

0.0
270 280 290 300 310 320
Temperature Τ (Κ )

Fig.1.8. μ (Τ )

1.4. The Reaction Scheme of a Biotechnological Process

So far in this chapter the standard intuitive way of presenting the modelling of
biotechnological processes has been followed. Our aim, however, is to
propose a more global and rigorous theoretical framework for the analysis of
bioreactor dynamics. In the next section we shall present a general class of
state space models which can describe a wide class of complex
biotechnological processes in stirred tank bioreactors. It will be shown that the
elaboration of these models can be made fully systematic provided a process
description in terms of reaction sctiemes is available. These reaction schemes,
which are presented in this section, are analogous (but not equivalent) to
those commonly used in classical chemical engineering.

The section is organised as follows. An abstract definition of the notion of

reaction scheme is given in Section 1.4.1. This allows a precise definition of
the concept of biotectinological process in the sense used in this book
 Chap.1. DYTJAMICAL MODELS OF BIOREACTORS 19

(Section 1.4.2). The definition is then illustrated and explained with a number
of concrete situations and applications in Section 1.4.3.

1.4.1. The notion of reaction scheme

Scheme of

A simple irreversible reaction involving two reactants and yielding a reaction

product is commonly represented by a scheme of the following form :

ξ ι + ξ 2 ^ ξ 3 (1­29)

where ξ ι and ξ 2 are the two which are irreversibly combined to

give the product ξ 3. φ is the reaction rate, i.e. the rate of consumption
of the reactants, which is equal to the rate of formation of the product.

Generally, the number of reaction components (reactants and/or products) is

arbitrary and the scheme is written :

Σ ξ ι ^ Σ ξ ; (1.30)

A reaction is a reaction in which a component (termed a

appears on both sides of the scheme : this means that the catalyst is assumed
to be consumed and produced simultaneously at the same rate by the reaction
(in such a way that the balance of this component is continuously in
equilibrium).
 20 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Autocatalysed reactions

In an autocatalysed reaction, one product is a catalyst of its own production. In

that case, the scheme is a follows :

ξ ι ­ ^ ξ 2 +ξ 3 (1.31)

The feedback arrow indicates the presence of an autocatalyst %2·'t is a

pseudoreactant which is not consumed by the reaction, but which can be
accumulated in the reactor.

1.4.2. Definition of a biotechnological process

A biotechnological process is a set of Μ reactions involving Ν components (i.e.

reactants and reaction products). The various applications addressed later in
this section will show that this definition is efficient and allows a very wide
class of practical situations to be covered. However, the following comments
are important for a correct understanding of this definition and to avoid any
confusion regarding the notion of reaction scheme and the definition of
biotechnological processes as they are used in this book.

1) The reaction scheme does nof represent a stoichiometric relationship

between the components, in contrast to the common practice in chemical
kinetics. It simply represents a qualitative relation. This allows us to
include chemical, biochemical and microbial growth processes in a unified
approach.

2) The components ξ ι of the reaction scheme are basically of four kinds :

­ populations of microorganisms
­ enzymes
­ external substrates (i.e. substrates which are introduced into the
reactor from outside)
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 21

­ products/internal substrates (i.e. components which are produced by

one reaction and can possibly be the substrate of another reaction).
However, other chemical components can be included in the reaction
scheme if they are useful from an engineering viewpoint (for instance,
components used for controlling the pH of a reaction).

3) The reaction scheme of a biotechnological process is a tool for deriving an

operational dynamical model of the process and for solving engineering
problems. It is never an exhaustive description of the process. For
instance, substrates which are not limiting are most often omitted from the
scheme. Similarly, by­products of a reaction which are not substrates in
other reactions and are of no interest to the user may also be omitted.
Consequently, the reaction scheme may be inconsistent with the law of
conservation of mass without causing any damage from an engineering
viewpoint.

We shall now illustrate the concept of a reaction scheme of a biotechnological

process using several typical examples.

1.4.3. Illustration

Microbial growth, death and maintenance

We consider a biotechnological process involving simultaneous growth, death

and maintenance of microorganisms on a single limiting substrate (as
described by the model (1.2)). The process is represented by the following set
of Μ = 3 reactions involving Ν = 3 components :

— (1.32.a)

Xd (1.32.b)

S+X >X (1.32.C)

 22 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

The three components are the living biomass X, the dead biomass Xd and the
limiting substrate S.

Notice that the three kinds of reactions mentioned above are present in the
scheme : simple reaction, catalysis, and autocatalysis.

The first reaction (1.32.a) represents the growth of the microorganism

population (growth rate φ g) : it is obvious that this reaction is autocatalysed
since the biomass X is a catalyst of its own production (there is no biomass
growth without initial biomass) but is not consumed by the growth reaction.

The second reaction (1.32.b) represents the death of microorganisms (decay

rate 9d ) : it is a simple irreversible reaction.

The third reaction (1.32.c) represents the maintenance of the microorganisms

(maintenance rate 9m) : it is a catalysed reaction (catalyst X) since the biomass
is neither consumed nor produced by the reaction, but simply maintained.

It is however important to note that in many practical applications, the death

and maintenance phenomena may be neglected in the analysis, so that the
process reduces to the simple autocatalytic reaction :

φ ^
S—L>X (1.33)

This reaction will very often serve as a basic example throughout the book to
illustrate the theory, under the name of simple microbial growth process or
single biomass/single substrate process.

Production by enzymatic catalysis is certainly one of the most common

biotechnological processes. The reaction scheme is as follows :
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 23

S+E >P + E (1.34)

with S representing the substrate, Ρ the product and Ε the enzyme. It may
occur that the enzyme cannot be isolated from the microorganism with which it
is associated and that the enzymatic catalysis is possible only in the presence
of biomass growing on the same substrate. In that case, the reaction scheme is
as follows:

S—i^X + E (1.35.a)

S+E >P + E (1.35.b)

It is then realistic to assume that the enzyme concentration is proportional to

the biomass concentration in such a way that the biomass itself can be
considered as the catalyst:

S ­ ^ X (1.36.a)

S+X >P + X (1.36.b)

In the sequel we shall usually characterise the enzyme catalysed (or non­
growth associated) production of a synthesis product Ρ by a reaction scheme
of the type(1.36.b).

Yeast growth

Yeast (Saccharomyces cerevisiae) growth is usually characterised by three

metabolic reactions :

­ respiratory growth on glucose

­ fermentative growth on glucose
­ respiratory growth on ethanol
 24 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

It can be represented by the following set of Μ = 3 reactions involving Ν = 5

components :

S + C —U X+P (1.37.a)

S—4 X­I­E­HP (1.37.b)

E­hC — L > X + P (1.37.C)

with X : yeast concentration

S : glucose concentration
C : dissolved oxygen concentration
Ε : ethanol concentration
Ρ : dissolved carbon dioxide concentration
φ ι : glucose respiration reaction rate
9 2 : glucose fermentation reaction rate
9 3 : ethanol respiration reaction rate

Anaerobic digestion

Anaerobic digestion is a process for the biological treatment of organic wastes

with production of methane gas. Four metabolic paths can be identified in this
process : two for acidogenesis and two for methanization (Fig.1.9). In the first
acidogenic path (Path 1), glucose is decomposed into fatty volatile acids
(acetate, propionate), hydrogen and inorganic carbon by acidogenic bacteria.
In the second acidogenic path (Path 2), OHPA (Obligate Hydrogen Producing
Acetogens) decompose propionate into acetate, hydrogen and inorganic
carbon. In a first methanization path (Path 3), acetate is transformed into
methane and inorganic carbon by acetoclastic methanogenic bacteria. While
in the second methanization path (Path 4), hydrogen combines with inorganic
carbon to produce methane under the action of hydrogenophilic
methanogenic bacteria.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 25

glucose
acidogenic
bacteria

propionate
­OHPA

£ I hydrogen
1
acetate
inorganic
carbon
acetoclastic
I methanogenic
hydrogenophi
hydrogenophilic r bacteria
methanogenic
bacteria
CO. CH

Fig.1.9. Reaction scheme of anaerobic digestion

The following reaction scheme follows from Fig.1.9 (M = 4 reactions, Ν = 10

components) :

Si — ^ Xi+S2 + S3+S4+S5 (1.38.a)

<P2
S2 — C x g + S g + P i (1.38.b)

<P3

S3 U X3 + S2 + S4 + Sg (1.38.C)

Φ 4
S 4 + S 5 — ­ C X 4 + P1 (I.SB.d)

with X i : acidogenic bacteria concentration

 26 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

X2: acetoclastic methanogenic bacteria concentration

X3: OHPA concentration
X4: hydrogenophilic methanogenic bacteria concentration
S i : glucose concentration
S 2 : acetate concentration
S 3 : propionate concentration
S 4 : hydrogen concentration
S5 : inorganic carbon concentration
Pi : methane concentration
φ ι : 1 St acidogenesis reaction rate
92 : 1^^ methanization reaction rate
93 : 2"^ acidogenesis reaction rate
94 : 2"^ methanization reaction rate

Intracellular production of PHB acid

The process under consideration is an aerobic culture of Alcaligenes

eutrophus with intracellular production of poly­p­hydroxybutyric (PHB) acid.

The main features of this process are as follows :

1) Two limiting substrates are needed for the microbial growth : fructose as
carbon substrate and ammonia as nitrogen substrate.

2) The intracellular production of the PHB acid, by fructose degradation, can

take two different paths :
­ the first being associated with the growth but with a very small yield;
­ the second being enzyme catalysed and completely inhibited by nitrogen.

3) Both microbial growth and product formation yield gaseous carbon dioxide.
The respiratory quotient is close to unity.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 27

According to these features 1) to 3), the process is described by the following

reaction scheme (M = 2, Ν = 5):

S1 + S 2 + C C"x + Pi + P2 (1.39.a)

51 + C + X > X + Pi + P2 (1.39.b)

with X : biomass concentration

S i : fructose concentration
52 : ammonia concentration
Pi : PHB acid concentration
P2 : dissolved carbon dioxide concentration
C : dissolved oxygen concentration
φ ι : growth associated product formation rate
92 : enzyme­catalysed product formation rate

Lactic fermentation

We consider the coculture of two thermophilic bacteria : Lactobacillus

and used for producing starters for the
manufacture of yoghurt.

In this anaerobic mixed culture, the influent substrate, in this case lactose, is
hydrolyzed into galactose and glucose which is in turn the unique limiting
substrate of four different biochemical reactions :

­ growth of Lactobacillus with production of CO2

­ production of D­lactate and CO2 catalysed by Lactobacillus
­ growth of

­ production of L­lactate catalysed by

Hence, the process is represented by the following reaction scheme (M = 5,

 28 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Ν = 8):

Φ ι

S >Pi + P2 (1.40.a)

Pi—4 X1 + P3 (1.40.b)

Φ 3

Pi + Xi > Xi + P3 + P4 (1.40.C)

P i — ^ X2 (1.40.d)

Φ 5

P1+X2 > X2 + P5 (1.40.e)

with S : lactose concentration

Xi : Streptococcus thermophilus concentration
X2 : Lactobacillus bulgaricus concentration
Pi : glucose concentration
P2 : galactose concentration
P3 : dissolved carbon dioxide concentration
P4 : L­lactate concentration
P5 : D­ lactate concentration
φ ι : lactose hydrolysis reaction rate
92 : St. thermophilus growth reaction rate
93 : L­lactate formation reaction rate
94 : L bulgaricus growth reaction rate
95 : D­lactate formation reaction rate
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 29

1.5 General Dynamical Model of Bioreactors

in this section we shall present a general state space model for the description
of biotechnological processes in stirred tank bioreactors. To simplify the
presentation, the same notation ξ \ will be used to denote a component or its
concentration (unit mass/unit volume) in the liquid phase of the reactor.

1.5.1. Dynamics of the process components

Once the reaction scheme of a biotechnological process is given, the

derivation of a dynamical model is made fully systematic by applying the
following rules :

R1) The reaction scheme of the process involves Ν components ξ , (i = 1 ,...,N)

and Μ reactions ( j = 1 M ) . The reaction rates are denoted φ \ 0=1 ,...M)

R2) The dynamics of the concentration of each component ξ \ are written as

follows :

§^=Σ (=^)^ϋ φ )­Ο ξ |­0| + Ρ | (1.41)

R3) the notation j'­i means that the summation is taken on the reactions with
index j which involve the component with index i.

R4) kjj are strictly positive constant yield coefficients, without dimension (i.e.
units of mass / mass). They have a s i g n when ξ ι is a reactant (i.e. when
it appears on the left hand side of the reaction scheme) and a sign
when ξ ί is a product of the reaction (i.e. it appears on the right hand side).

R5) Qi is the rate of mass outflow of the component ξ ί from the reactor in
gaseous form.
 30 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

R6) Fj is the mass feed rate in the reactor of the component ξ \ if it is an external
substrate. OthenA/ise R = 0.

1.5.2. State space model

We introduce the following matrix notations :

ξ ^=[ξ ι ,ξ 2....ξ Ν ] (1­42.a)

φ ^=[φ ι .φ 2.···.Φ Μ ΐ (i.42.b)

Q^=[QI,....QN] (1.42.C)

ρ Τ =[ρ ^,...,Ρ ^] (1.42.d)

Κ =[Kij] : Ν XΜ matrix with Ky = (±) ky if JH

K|j = 0 othenvise (1.42.e)

Then, from (1.41), the dynamics of biotechnological processes can thus be

represented by the following general nonlinear state space model, written in
matrix form :

General Dynamical Model

^ = Κ φ (ξ .0­Ο ξ ­Ο (ξ ) + Ρ (1.43)

In this expression we have introduced the notations φ (ξ ,t) and 0 ( ξ ) to

emphasize that φ and Q (and also F in some instances) may be time varying
and may depend on the process state ξ . The issue of modelling φ , Q and F as
functions of ξ will be addressed later.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 31

The State space model (1.43) is in fact the backbone of this book because it
will serve as the basic ingredient for the derivation of a//the estimation and
control algorithms proposed throughout the subsequent chapters. It should be
stressed for the reader that, in our view, the model (1.43) is not a mathematical
oddity but has a definite physical meaning :

­ the first term Kφ (ξ ,t) describes the kinetics of the biochemical and
microbiological reactions which are involved in the process;

­ the remaining terms ­Ο ξ ­ 0(ξ ) + F describe the transport dynamics of

the components through the bioreactor.

The state space model (1.43) is thus mainly a way to express, in a single
compact mathematical form, the two physical phenomena (namely kinetics
and transport dynamics) which are in intimate interaction in a bioreactor.

Example: Microbial growth, death and maintenance (continued)

The process involves Ν = 3 components and Μ = 3 reactions and is described

by the reaction scheme (1.32). The following change of notation is
introduced:

ξ ι =Χ ξ 2=δ ξ 3 = Xd (1.44.a)

9i = 9g 92 = 9d 93 = 9m (1.44.b)

Then, applying the rules R1 to R6 to the scheme (1.32), the dynamical model is
easily written as follows :
 32 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

­ = Ι <ι ι φ ι ­Κ ΐ 2φ 2­0ξ ι (1.45.a)

dt

dξ 2
= ­Κ 2ΐ φ ι ­Κ 23φ 3­Ο ξ 2 + Ρ 2 (1.45.b)

dt
= k3292 ­ Ο ξ 3 (1.45.C)

dt
or, in matrix forrn

= 0 • "Φ Γ ­D + Ό ' (1.46)

dt "ξ ι "
­k2i 0 ­^23 92 ξ 2 F2
Μ . 0 k32 0 .93. Μ . .0.

It is easy to check that the first two equations of the state space model (1.46)
exactly coincide with the "classical" model (1.2), provided the following change
of notation is considered (in addition to (1.44)) :

ki 1 = 1 ki2 =1 k2i = k^ k23 = 1 k32 = 1

φ ΐ =φ 9 = μ Χ 92 = 9d=kdX 93 = 9m = kmX

It should also be noted that this example clearly shows that the reaction rates
φ ι , φ 2 and 93 depend on the process state.

1.5.3. Modelling the reaction rates

The reaction rate φ (ξ ,t) is most often a very complex function of the operating
conditions and of the state of the process. The analytical modelling of this
function is often cumbersome and is the subject of continuing, intensive
investigation (and sometimes of controversy). For instance, in the case where
φ is proportional to the specific growth rate μ , there are several dozens of
possible models, as shown in Section 1.3 and in Appendix 1.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 33

In this book, we adopt a unifying (but not simplifying) stance for modelling φ
based on the following fact: the reaction can take place only if all the reactants
are present in the reactor. In other words, the reaction rate is necessarily zero
whenever the concentration of one of the reactants is zero. This is
represented as follows :

φ j(ξ ,t) = α j(ξ ,t) (1.47.a)

0<α j(ξ ,t)<α ^ax (1.47.b)

The notation n­j means that the multiplication Π is taken over the components
with index η which are reactants in the reaction j {including the autocatalysts)
which are considered as reactants when writing (1.47.a)).

α j(ξ ,t) is called the specific reaction rate since it is the reaction rate per unit of
each reactant. It must be a bounded function for evident reasons of
mathematical consistency.

Define the vector α and the matrix 0(ξ )

α '^=[α ι ,...,α Μ ] (1.48.a)

0(ξ )=_(1ί 3||­ (1.48.b)

n­j

The notation "diag" is used to denote a diagonal matrix, i.e. :

 34 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Η
diac
Λ π 4=
n~]
η ~1

ο Π ξπ ο
η ­2
Ο ο .. ο

η ~Μ J

The state­space model (1.43) is then rewritten :

• ^ = Κ Ο (ξ )α ­Ο ξ ­Η Ρ ­0(ξ ) (1.49)
dt

An important special case occurs when the specific reaction rates α are
independent of the state ξ and depend only on the temperature (e.g. according
to the Arrhenius law) :

α (ξ ,t) = α (T(t)) (1.50)

In particular, when the temperature is regulated at a constant value, a

multilinear state space model with constant parameters is obtained. In terms of
chemical kinetic theory such a model corresponds to the assumption that all
the reactions are governed by the law of mass action with a unit partial order
with respect to each reactant.

The specific growth rate

In the particular case where the reaction rate φ (ξ ) is actually a microbial growtti
rate, an alternative representation, which has already been evoked several
times, is commonly used in bioengineering studies.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 35

Let US suppose that φ ](ξ ) denotes the growth rate of a particular biomass
population Xj. It is clear, from our preceding discussion, that the concentration
Xj necessarily appears as a factor in the mutiplication term of (1.47.a). Hence
the reaction rate φ ](ξ ) can equivalents be represented as follows :

φ )(ξ ) = μ ί (ξ )Χ | (1.51)

where μ )(ξ ) is termed the specific growth rate, since it is the growth rate per unit
of biomass.

Remark that the notation has been implicitly used many times in Sections 1.1
to 1.3, and explicitly in the example of Section 1.5.2 (namely 9g = μ Χ ).

The definitions of specific reaction rate and specific growth rate are now further
illustrated by an example.

Example : Basic dynamics of microbial growth

Consider the simple microbial growth process (1.33)

s^Cx

for which the following dynamical model is derived from the rules R1­R6 :

­ki φ ­DrS (1.52)

X X

The growth rate φ can be represented either as

φ (8.Χ ) = a(S.X) SX (1.53)

or a s :
 36 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

φ (8,Χ ) = μ (8,Χ )Χ (1.54)

Notice that, with the latter expression, the model (1.52) coincides with the
"classical" representation (1.3.a­b), for k2 = 1 and F = DSin.

Suppose that the specific growth rate μ (8,Χ ) is described by the "Contois law"
(1.22):

μ (8,Χ ) = ­ ^

This implies that the reaction rate φ (Χ ,8) is written :

„*qy

φ (Χ ,8) = μ (Χ .8)Χ = ι ^ 5 ^ (1.55)

which in turn implies that the specific reaction rate a(X,8) is defined as :

The example shows the consistency between the two alternative

representations (1.53) and (1.54) when the Contois law is used.

1.5.4. Modelling the gaseous outflow rates

In equation (1.41), Qi represents the rate of mass removal in gaseous form for
those components ξ , which are soluble in the liquid phase and gasifiable at
atmospheric pressure. According to common industrial practice, we suppose
that these compounds are freely given off from the reactor. As long as the
concentration is lower than the saturation level, if we neglect the liquid­gas
transfer dynamics, it is natural to assume that the outflow rate in the gaseous
phase is simply proportional to the concentration in the liquid phase :
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 37

α ϊ = β ίξ ΐ 0<β | 0<ξ |<ξ |3 (1.57)

where β ­, is the specific liquid-gas transfer rate and ξ ί β is the saturation

concentration.

Obviously, β ί = 0 if the component ξ ι is not gasifiable (as for biomass or

enzymes, for instance).

Defining the matrix Β :

B = diag{β |} i = 1,....N (1.58)

the general dynamical model (1.43) is rewritten :

^ = Kφ (ξ )­Dξ + F ­ B ξ (1.59)

If ξ ί > ξ ί 8, it is clear that the mass balance expressed by equation (1.59) is no

longer valid. This point will be revisited in Section 1.8.2.

1.5.5. Modelling the feed rates

In equation (1.41), Fj represents the feed rate (per unit of volume) of those
components ξ ί which are external substrates introduced in the reactor from the
outside. The way the reactor is fed and the nature of the substrate (liquid or
gaseous) lead to various methods of modelling the feed rates.

Liquid

There are essentially two ways of introducing a liquid substrate into the
reactor: either diluted in the influent stream (Fig.l.lO.b) or independently of the
influent stream (Fig.l.lO.a).
 38 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

concentrated concentrated
substrate substrate

water water
dilution
tank
r
I do
a) b)

Fig.1.10. Influent liquid substrates

In the latter case (Fig.l.lO.a), there is no need for specific further modelling of
the substrate feed rate Fj. However, in the former case, when the substrate is
diluted in the water stream, the feed rate is proportional to the substrate
concentration in the influent. The proportionality coefficient is precisely the
dilution rate D (see for instance model (1.3)). In that case, each nonzero feed
rate Fj is written:

Fi=DS|,.| (1.60)

where Sjn,! denotes the influent substrate concentration.

When all the external substrates are in liquid form, we can define the vector:

Sin = [Sinj , Sin,2 ^\n,U^^ -^^ )

where, obviously, Sin.i = 0 when ξ ί is not an external substrate; then :

F = DSiin (1.62)

and the general dynamical model (1.43) is rewritten :

= Kφ (ξ )~Dξ + DSin­Q(ξ ) (1.63)

dt
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 39

Gaseous substrates

It may occur that substrates are introduced into the bioreactor in gaseous form.
The archetype is the class of aerobic processes where the reactor is fed with
gaseous oxygen by aeration of the culture medium. Let us suppose that ξ η
denotes such a gaseous substrate. Then, as we have seen in Section 1.2
(e.g. expression (1.15)), the corresponding feed rate is written :

Fn=kLa(Fg) ( ξ η ε ­ ξ η )

where kLa(Fg) is a time­varying transfer coefficient depending on the

"aeration" feed rate Fg (and also on other factors such as, for instance, the
geometry of the aerator) and ξ η 8 is the saturation concentration.

It is worth noting that in this particular case the feed rate Fn is itself a function of
the process state ξ η ­

1.5.6. Scaling of the yield coefficients

The rates of consumption of reactants and the rates of production of products

are expressed in the model (1.41) by terms of the form (±)kij9j whose common
dimension is unit of mass per unit of time.

In each reaction, let us choose, one component ξ η which is called the nominal
component of the reaction. It can easily be seen that the fraction kjj/kjn
represents the yield of consumption or production of the component ξ ί per unit
of consumption or production of the nominal component ξ η . It is then obvious
that the coefficient knj can be arbitrarily fixed to 1 without any loss of generality
in the general dynamical model (1.43). This method of scaling the yield
coefficients will be used repeteadly throughout the book and, in particular, in
the examples of models presented in the next section.
 40 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

1.6. Examples of State Space Models

In this section the state space models (1.43) and (1.49) are established for the
examples of Section 1.4.3.

1.6.1. Yeast growth

The external influent substrates are glucose and oxygen. The application of
rules R1­R6 to the scheme (1.37) leads to the following definitions :

ξ '^=[Χ ,8,Ε .0,Ρ ] (1.64.a)

F^=[0 Fi=DSin 0 QinO] (1.64.b)

Q'''=[0 0 0 0 Q I ] (1.64.C)

K' = 1 ­ki 0 ­ k g k7 (1.64.d)

1 ­kg k4 0 kj

1 0 ­ k s ­ k g kg

(1.64.e)

(1.64. f)

α (ξ )= sex 0 0 (1.64.g)
0 SX 0
0 0 ECX

with Qin: oxygen feed rate

Q i : output carbon dioxide flow rate
Sin : influent glucose concentration
kj (i = 1 to 9 ) : yield coefficients
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 41

ttj (i = 1 to 3 ) : specific reaction rates

Note the scaling of the yield coefficients in the matrix Κ (1.64.d).

1.6.2. Anaerobic digestion

The following model is associated with the reaction scheme (1.38) :

ξ ^= [X­j Si X2 S2 X3 S3 X4 S4 S5 P^] (1.65.a)

QT=[0000000 Q3Q2Q1] (1.65.b)

FT=[OFi=DSinOOOOOOOO] (1.65.C)

1 ­kgi 0 0 kei 0 kg^ kg^ 0 ' (1.65.d)

0 0 1 ­k42 0 0 0 0 kgg ko2
0 0 0 k43 1 ­kea 0 kgg kg^ 0
0 0 0 0 0 0 1 ­k84 ­k94 ko4

(1.65.e)

a ^ = [ a i ttg Og a4] (1.65.f)

α (ξ ) = ["8ι Χ ι 0 0 0 (1.65.g)
0 S2X2 0 0
0 0 S3X3 0
0 0 0 S4S5X5

with Q i : methane gas flow rate

Q2 : gaseous CO2 flow rate
Q3 : gaseous hydrogen flow rate
Sin: influent glucose concentration
kij (i=0 to 9, j=1 to 4 ) : yield coefficients
 42 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

ttj 0=1 to 4 ) : specific growth rate of Xj

1.6.3. Intracellular production of PHB acid

The following model is associated with the reaction scheme (1.39) :

ξ "^ = [Χ Si S2 Pi C P2] (1.66.a)

Q"^=[0 0 0 0 0 Q I ] (1.66.b)

F^=[0 DSi,in DS2.|nOQin 0] (1.66.C)

—Γ 1 —k­| —kg k3 —k4 k j (1.66.d)

0 ­kg 0 1 ­kg kg

(1.66.e)

= [α ϊ og] (1.66.0

0(ξ ) = S i S j C X Ο (1.66.g)
Ο SiCX

Qi : output gaseous carbon dioxide flow rate

Qin: oxygen feed rate
Si,in: influent fructose concentration
S2,in: influent ammonia concentration
kj (j=1 to 6 ) : yield coefficients
a, (1=1 to 2 ) : specific reaction rates

1.6.4. Lactic fermentation

The following model is associated with the reaction scheme (1.40)

 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 43

%^=[S X 1 X 2 P 1 P 2 P 3 P 4 P 5 ] (1.67.a)

Q^ = [OOOOOQi 0 0] (1.67.b)

F'^ = [Fi=DSin0 0 0 0 0 0 0] (1.67.C)

­ki 0 0 1 kg 0 0 0 (1.67.d)
0 1 0 ­ k g 0 ky 0 0
0 0 0 ­ k 3 0 kg kg 0
0 0 1 ­k4 0 0 0 0
0 0 0 ­kg 0 0 0 kioJ

(1.67.e)

a ^ = [tti Ogttaa4 05] (1.67.f)

0(ξ ) = S 0 0 0 0 (1.67.g)
0 P1X1 0 0 0
0 0 P1X1 0 0
0 0 0 PgXg 0
0 0 0 0 P2X2J

with Qi : gaseous CO2 flow rate

Sin : influent lactose concentration
kj 0=1 to 10): yield coefficients
tti (i=1 to 5 ) : specific reaction rates

1.7. A Basic Structural Property of the General Dynamical Model

We consider the general dynamical model (1.43):

= Kφ (ξ .t)­Dξ ­Q(ξ ) + F (1.43)

dt

with dim(ξ ) = dim(F) = dim(Q) = N, dim(9) = Μ and dim(K) = NxM.

We define :
 44 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

ρ = rank (K);
Ka a (pxM) full rank arbitrary submatrix of K;
Kb the remaining submatrix of K;
(ξ β , ξ ό ). (Qa, Qb) and (Fa, Fb) the partitions of ξ , Q and F induced by
(Ka.Kb).

The general dynamical model (1.43) is rewritten :

dt = + (1­68)

dt = Κ 6 φ ( ξ 3 . ξ 6 ) ­ 0 ξ 6 ­ α 6 + Ρ 6 (1­69)

We then have the following property :

There exists a state transformation :

Ζ =Α ο ξ 3+ξ 6 (1­70)

where Aq, of dimension (N­p)xp, is the unique solution of the matrix

equation :

AoKa + Kb = 0 (1.71)

such that the state­space model is equivalent to :

dξ _a
= Κ 3φ (ξ 3­ξ 6)­0ξ ,­α β +Ρ 3 (1.72)
dt

^ = ­ D Z + Ao(F3­Q3) + ( F b ­ Q b ) (1.73)
 Chap.1. DYNAMICAL MODB.S OF BIOREACTORS 45

When (Fa ­ Qa) is identically zero, the partition (ξ a.ξ b) is said to be nice
(because the dynamics of Ζ are independent of both Κ and φ ).

Example: Intracellular production of PHB acid

X ξ 6 = S2
Si Pl
c
Pa.

Consider the state space model (1.66) for the intracellular production of PHB.
If we choose the partition :

X ξ 6 = S2
LSiJ Pi
c
LP2.

we have from (1.56.d):

Ka = 1 0 Kk = ­ k 2 0•
L­ki ­ksj kg 1
­k4 ­ke
Lk7 kg.

Clearly the matrix Ka is full rank and the partition is admissible. Hence, from
(1.71):

1 .
kgkg 0 •
­kgks+ki 1
+k4k5—k^kg —kg
—ksky+k^kg kg
 46 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

i.e. the auxiliary state transformation (1.70) is as follows

z = "Ζ Γ + S2

Z2 (k3+ki/k5)X +Si/kg + Pi
Z3 ­(k4 + ki kg/kg) X ­ keSi/kg + C
.Z4. ­(ky ­ kikg/kg) X + k8Si/k5+ P2.

This is not, however, a nice partition since Fa = [0 DSjn]"^ is not identically

zero. We shall obtain a nice partition, in this example, by considering the
following partition (it is easy to check that it is also the only nice partition) :

ξ 3= ξ 6 = Si"
u

c
LP2.

1.8. Reduction of the General Dynamical Model

The examples of biotechnological processes in Section 1.6 have shown that a

bioreactor dynamical model may be fairly complex in some instances and
involve a large number of differential equations. But there are many practical
applications where a simplified reduced order model is sufficient from an
engineering viewpoint. Model simplification can be achieved by using the
singular perturbation technique, which is a technique for transforming a set of
n+m differential equations into a set of η differential equations and a set of m
algebraic equations. This technique is suitable when neglecting the dynamics
of substrates and of products with low solubility in the liquid phase. The
method will be illustrated with two specific examples before stating the general
rule for order reduction.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 47

1.8.1. Singular perturbation technique for substrates

We return to the simple microbial growth process (1.33) :

The dynamics corresponding to this scheme are represented by the model

(1.1.a­b) rewritten as follows :

dXj
dt ^ = μ Χ τ ­ ϋ Χ γ (1.74.a)

dS
= ­Κ ι μ Χ τ ­ε 05 + Ρ , (1.74.b)
^ dt

where X j is the total amount of biomass in the reactor (i.e. not the
concentration ), S is the substrate concentration, ε = V is the reactor volume,
and Fr is the substrate feed rate.

Notice that equations (1.74.a­b) are true mass balance equations, expressed
in terms of total quantities and not in terms of concentrations.

Clearly, for any ε > 0, the system (1.74) consists of two differential equations.
However, if ε = 0, the system consists of one differential equation and one
algebraic equation, because (1.74.b) reduces to :

Κ ιμ Χ τ = Ρ Γ (1.75)

Consequently, the first differential equation (1.74.a) may be rewritten :

dXj 1
^ = ­DX,+ ^ F , (1.76)
 48 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Setting ε = 0 in (1.74.b) is called a singular perturbation. The objective is to

examine the simplified model (1.76) in order to draw conclusions about the
original system with ε 0. It must be well understood that we are not assuming
that the volume of the reactor is zero. We actually assume that the volume V is
small enough to neglect ε (dS/dt) and ε DS in (1.74.b). This means that it is
legitimate to consider the simplified model with the actual nonzero value of V
and hence rewrite it in the standard form (see Section 1.5.3) :

^ = a(X,S)XS ­ DX (1.77.a)
dt

kia(X.S)XS = DSin (1.77.b)

with X = XT/V and φ = a(X,S)XS = μ (X,S)X

1.8.2. Singular perturbation technique for products

Let us consider a biochemical reaction described by the following reaction

scheme :

S >P

where Ρ is a volatile product which can be given off in gaseous form and has
low solubility in the liquid phase. According to the rules of Section 1.5.1, the
dynamical model is as follows :

4^ = ­ 9 ­ D S + P (1.78.a)
dt

^ =k9­DP­Q (1.78.b)
dt

The consistency of this model requires that the product concentration Ρ be

lower than a saturation concentration representative of the product solubility,
which is expressed as :
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 49

P = nPsat 0<n(t)<1

where Psat is the saturation concentration which is constant in a stable

physico­chemical environment. The model (1.78) is rewritten in the standard
singular perturbation form, with ε = Psat ·

^ =­ 9 ­ D S + F (1.79.a)
dt

e ^ =k9­eDn­Q (1.79.b)
dt

If the solubility is very low, we obtain a reduced order model by setting ε = 0

and replacing the differential equation (1.79.b) by the algebraic one :

Q = k9 (1.80)

1.8.3. A general rule for order reduction

The above examples show that the rule for model simplification is actually very
simple and that an explicit singular perturbation analysis is not really needed.
Consider that, for some i, the dynamics of the component ξ ί are to be
neglected. The dynamics of ξ \ are described by equation (1.41) :

^ = Χ (±)Κ ί φ ί ­Ο ξ ί ­α ί + Ρ ί (1.41)
j~l

The simplification is then achieved by setting ξ ι and dξ j/dt to zero i.e. by

replacing the differential equation (1.39) by the following algebraic equation :

5^(±)k|j9j= Q i ­ F | (1.81)
 50 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

1.8.4. Example : The anaerobic digestion process

We shall now show how these rules for order reduction can be applied to
derive a simplified model of the anaerobic digestion process. The third
reaction (decomposition of propionate by the OHPA) and the fourth reaction
(decomposition of hydrogen) are supposed to be characterised by fast
dynamics. This is expressed by setting S3, dSa/dt, and S4, dS4/dt to zero in
the dynamical model (1.65). We derive the following algebraic equations :

93 = F ^ 9 i (1.82)
^63

kei ^83 k63k8i +k6ik83 oon

Ψ λ = ­Γ —Φ ι + ι Γ "Ψ 3 = 9i (1 ­83)

Moreover, it is well known that the solubility of methane is extremely low.

Therefore we write :

Q, = ko*­k04^:^i^^^, (1.84)
K63K84

By using the above approximations (1.82), (1.83), (1.84), we finally obtain the
classical two­step (acidogenesis­methanisation) dynamical representation of
anaerobic digestion processes :

dXi
dt = Φ ι ­ DXi (1.85.a)

dS
1 = ­ | < , φ ^ ­ D S i + DSin (1.85.b)
dt

dXg
= 92­DX2 (1.85.C)
"df
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 51

dSg
= Ι <3 9i ­ k292 ­ DS2 (1.85.d)
"df

dPg
= K49I+ K592­DP2­Q2 (1.85.e)
"dT

Q i = k 6 9 2 + k79i (1.85.f)

with :
^61^43
KI = K^ ­I, K2 ­ K42, K3 — K4­| +
^63

. . ^61^93 . (^63^81 ^61 ^33 ^

^63 ^63^84

^63^81 +K6IK83
= K92, KG = KO2I KY = KO4
^63^84

P2=S5

This model corresponds to the following simplified reaction scheme (compare

with scheme (1.38)) :

Si — 4 ­ S 2 + Pi + P2 (1.86.a)

S2 ^ X 2 + P i +P2 (1.86.b)

Comment

In some instances (e.g. when the influent substrate is not composed mainly of
glucose) the presence, formation and decomposition of hydrogen are
negligible. The formation of methane is then assumed to result almost
exclusively from the decomposition of acetate (i.e. K4 = KY = 0), and the
reaction scheme simplifies as follows :
 52 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Φ ι

S i — C x i + S 2 + P2 (1.87.a)

S 2 — ^ X 2 + Pi+P2 (1.87.b)

while equation (1.85.f) reduces to :

Qi=k692 (1.88)

1.9. Stability Analysis

In this section, we analyse the stability of the state space dynamical model. In
Section 1.9.1, we shall prove the global Bounded Input Bounded State (BIBS)
stability of the system (in accordance with the physical reality). In Section
1.9.2, we introduce the concept of equilibrium state and we show that
bioprocesses generically possess several equilibrium states whose stability is
analysed in Section 1.9.3.

1.9.1. Bounded Input Bounded State stability

The BIBS stability will be analysed under the following assumptions :

A I . The dilution rate is bounded below as follows :

0<D^in^D(t) Vt (1.89)

A2. The feed rates are bounded as follows :

0<F|(t)<F^ax Vi,Vt (1.90)

 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 53

A3. Each reaction involves at least one reactant which is neither a catalyst nor
an autocatalyst.

We have the following stability theorem.

Theorem 1.1 : Under Assumptions AI to A3, the state variables of the general
dynamical model (1.59) are positive and bounded for all t.

Proof : The proof of the theorem will be divided into two parts. We shall first
show that the state variables are positive, and then that they have upper
bounds.

1) the state variables are positive : ξ ι ^) >0 W, Vt

Suppose that ξ i(t) = 0 for some i. Then, from (1.47) and (1.59), the dynamical
equation reduces to :

­ ^ = I ( + ) k j j 9 j + F|>0 (1.91)

where the summation is taken only over those reactions with index j which
involve ξ ι as a product but neither as a substrate nor as an autocatalyst.
Hence, the right hand side of (1.91) is necessarily nonnegative and, since ξ ΐ (0)
>0,ξ i(t)>Oforallt.

2) The state variables are upper bounded.

Let us select one reactant of the process which is, at least in one reaction,
neither a catalyst nor an autocatalyst. This reactant (which necessarily exists
by Assumption A3) is denoted ξ η (η e [1, Ν ]).

We define two sets of indices :

J = (mi. m2,..., mp) I = (m, n2,..., nq) (1.92)

 54 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

mj ( i = 1 p ) are the indices of the reactions which involve ξ η as a reactant

(i.e. not as a product).

nj (i=1,..., q) are the indices of the components (excluding ξ η ) involved in

the reactions with an index mj e J.

We define the auxiliary variable :

z, = a , ξ , + Σ ξ i (1.93)

with

Σ kij
an = m a x — >0 (1.94)
­knj

where ky denotes entry (i,j) of the matrix K.

The dynamics of Zn can then easily shown to be :

^ =Σ [a, R,j 4­ Σ R|jl 9j ­ Dz, ­ a, β , ξ , ­ Σ ξ ^ ­f a , ­f Σ F| (1.95)

From Assumption A2 and the definition of an it follows that :

dzn
< ­ D z , + (an+q)F^3, (1.96)
dt

At this point, there are two possibilities depending on the initial value of zn.

(an + q)Fmax
a) Zn(0) >

It is clear from Assumption AI and (1.96) that the time derivative of Zn(0) is
negative and hence that Zn(t) < Zn(0) for all t.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 55

'^min

It then follows from (1.96) that Zn(t) is bounded as follows for all t :

, ^ ( t ) , i ! n ; ^ (1.97)

The boundedness of Zn(t) obviously implies that of ξ j(t), i e I, and of ξ n(t).

The theorem follows by repeating the same argumentation for all the reactants.
Q.E.D.

An explicit expression for the upper bound on the state variables can be
formulated by introducing the following assumption on the initial conditions :

A4. The initial values of the state variables ξ j(t) have upper bounds as follows :

8„ξ η (0) + . Σ ξ , ( 0 ) < ^ ^ " ΐ ^ ^ ' ' ­ (1.98)

'^' *^min

for all η and i e I as defined in Theorem 1.1.

Corollary 1.2 : Under Assumptions AI to A4, the state variables ξ i(t) of the
general dynamical model (1.59) are non­negative and bounded as follows for
all t :

0 < ξ i(t) < maxj 1 (a, + q ) ^ (1.99)

Proof: Straightforward from (1.97).

 56 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Comment

Assumption A1 is rather restrictive, since it requires a strictly positive lower

bound Dmin on the dilution rate D(t). It can easily be relaxed in cases where all
the external substrates are in the liquid phase (with the exception of any
gaseous external reactant such as oxygen, see Section 1.5.5). In that case, the
vector F of feed rates is written (see (1.61) (1.62)):

F = DSin = D [Sin.l. Sin.2, .·., Sin,Nr (1 ­100)

Assumptions A I , A2 and A4 are replaced by the following ones :

AV. D(t)>0 Vt (1.101)

A2'.0<Sin.i(t)<Smax Vi, Vt (1.102)

A4'. a,ξ ,(0) + Σ ξ ί (0) < (a, + q) S^^x (1­103)

and Corollary 1.2 is reformulated as follows :

Corollary 1.3 : Under Assumptions AV, A2', A3 and A4', the state variables ξ i(t)
of the general dynamical model (1.59) are positive and bounded as follows for
all t :

0<ξ i(t)<max^ 1 , — (an + q)Sa,ax (1­104)

Proof: Straightforward from Theorem 1.1 and Corollary 1.2 by noting that
equations (1.96) and (1.97) become :

^ < D ( ­ 2 n + (an + q)S^ax) (1­105)

2n<(an + q)S,ax (1­106)

Q.E.D.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 57

Consider an autocatalytic reaction with one substrate and one product:

s^-Qx + p (1.107)

The dynamics are as follows :

­ki φ ­DrS· 0 (1.108)

dt
1 X 0 0
0 lQI.

If we directly apply the result of the above Corollary 1.2 by setting :

ξ η =8 (1.109)

1 +k2
(1.110)

and if the initial values of S, X and Ρ are bounded according to Assumption

A4' :

S(0) + X(0) + P(0) < i — 1 (1.111)

v1 + k 2 y J +k2.

it follows that the concentrations of the substrate S, the biomass X and the
product Ρ are bounded as follows, for all t :

0 < S(t) < S Vt (1.112.a)

max
1 +k.
0<X(t)<­j_iS„,a, Vt (1.112.b)

1 +k.
0 < P(t) < Vt (1.112.C)
 58 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

where Smax is the maximum substrate concentration in the influent (0 < Sin(t) <
Smax, Vt).

Expressions (1.112.a­c) strengthen the plausibility of the dynamical model

since they guarantee that the different concentrations remain positive, that the
substrate concentration S remains lower than the higher influent substrate
concentration Smax and that the concentrations of the two "products" X and Ρ
are upper bounded proportionally to the (maximum) influent substrate
concentration, the proportionality factor being determined by the yield
coefficients.

However, the general formulation of Theorem 1.1 and Corollaries 1.2 and 1.3
often leads in practice to rather conservative upper bounds. By taking
advantage of the structure of the reaction scheme in specific applications, less
conservative bounds may be written down. In fact, this is easily achieved in the
above example by defining two auxiliary variables (instead of one in Theorem
1.1):

Zi = T ^ S + X (1.113.a)

Z2 = 7 ^ S + P (1.113.b)

These are immediately derived from the basic structural property (Section 1.7)
by considering :

The dynamics of Z^ and Z2 are clearly governed by the following equations :

^ = ­DZ, + ^ D S i , (1.115.a)

dZg
= ­DZ2 + T ^ D S i n ­ Q (1.115.b)
"dT ' ki
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 59

From Assumption A2', tlie time derivatives of Z i and Z2 are bounded as

follows :

dZi 1

_ i < _ D Z i + ­DS„,ax (1­116.a)

dZp kp
_ 2 < _ D Z 2 + ^DS„,3x (1­116.b)
It follows immediately that if, at the initial time, S. X and Ρ satisfy the
inequalities :

S(0) + kiX(0) < S„,ax (1.117.a)

S(0) + ^ P ( 0 ) < S ^ a x (1.117.b)

k2

then they have the following upper bounds :

S(t)<S^ax Vt (1.118.a)

X(t)< —Smax Vt (1.118.b)

P(t)<­R^S^ax vt (1.118.C)

These upper bounds are clearly less conservative than those given directly by
(1.112). They are also more realistic in the sense that they are closer to the
mass balances of the biochemical reaction : it is obvious, for instance, that,
according to (1.118.b), the maximum quantity of biomass that can be
synthesised is given by the maximum available amount of substrate divided by
the yield coefficient ki from S to X.
 60 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

1.9.2. Equilibrium states

An equilibhum state is, by definition, a constant state, denoted ξ , which

satisfies the equation of the general dynamical model (1.43), i.e. the following
nonlinear multivariable algebraic equation :

^ =0 => Κ φ (ξ )­ϋ ξ + Ρ ­α (ξ ) = 0 (1.119)

for given constant values D and F of the dilution rate D and the vector of feed
rates F.

The problem of calculating the equilibrium state ξ of a bioreactor model is that

of solving equation (1.119). The latter has no general analytical solution and
can be solved only in specific applications.

Furthermore, it is a characteristic of biotechnological processes that they

exhibit multiple equilibrium states (i.e. several different solutions to (1.119)), as
a consequence of the autocatalytic action of microorganisms. This issue,
which will be extensively discussed in the following paragraphs, is introduced
via a simple example.

Example

We begin with the example of a simple irreversible reaction of the form :

φ
S >Pi + P2 (1.120)

with one reactant S and two products Pi and P2. This is, for instance, a
plausible model of the hydrolysis of lactose into glucose and galactose, which
is the first step (1.40.a) of lactic fermentation.

We suppose that the reaction rate φ is modelled by equation (1.47), i.e. :

 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 61

φ =α 8 a = constant (1.121)

which corresponds to first order kinetics. The dynamical model (1.63) may
therefore be written :

= ­ a S + DSin­DS (1.122.a)
dt

dP
i = kiaS­DPi (1.122.b)
dt

dPg
= k2aS­DP2 (1.122.C)
"dT

The equilibrium state (Si,Pi,P2) is the solution of the following algebraic

equations :

­ a S + DSin­DS=0 (1.123.a)

kiaS­DPi=0 (1.123.b)

k2aS­DP2 = 0 (1.123.C)

This equilibrium state is unique and given by :

DSin ­ kiaSin ­ kpttSin

a+5 a+D a+D

Multiple equilibrium states

We shall now explain why biotechnological processes, as a class, possess

multiple equilibrium states. The issue is first discussed on the basis of an
 62 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

example. Consider the reaction scheme (1.35.a­b) of a microbial growth

process with associated enzymatic production :

S—UX +E
92

S + E >P­HE

Applying the rules of Section 1.5.1 and the feed rate modelling (1.62), the
general state space model is written :

1 0• ­ D [ X" + 0 (1.125)
dt
­ k i ­kg ,<P2j s
ka 0 Ε 0
0 1 P. 0

According to Section 1.5.3, we assume in addition that the reaction rates φ ι

and 92 are modelled as follows :

φ 1=μ (X,S)X 0<μ (X,S)<μ * (1.126.a)

92 = a(E,S)ES (1.126.b)

where μ (X,S) is the specific growth rate and a(E,S) is the specific enzymatic
production rate.

An equilibrium state is defined as a set of constant states (χ , s, E, P) which

satisfy d(X,S,E,P)/dt = 0 for given constant values D and Sjn. From the model
(1.125)­(1.126), it follows readily that there are necessarily two different sets of
equilibrium states :

X=0 E=0 P=0 S = Sin (1.127)

 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 63

b)

k3X = E (1.128.a)

μ (Χ .8) = 0 (1.128.b)

kiX + k2P = D(Sin­S) (1.128.C)

a(S,E)ES = DP (1.128.d)

The states (1.127) are so called because they

correspond to a wash­out of the biomass from the reactor. They can occur for
any value of D and S|n. They are also the only possible equilibrium states
when D > μ * It is obvious that wash­out equilibrium states are undesirable
and must be avoided as far as possible in industrial applications.

On the other hand, the operational equilibrium states implicitly defined by

(1.128) are those of practical interest. Their characterization obviously requires
the specification of analytical models for the specific rates μ (Χ ,8) and a(E,S).
In general, however, the explicit calculation of the operational equilibrium
states is impossible because the set of nonlinear algebraic equations (1.128)
has no explicit analytical solution and can only be solved numerically. An
analytical solution of (1.128) can nevertheless be calculated using simple
models for μ (Χ ,5) and a(E,S). This is shown in the next paragraph.

Example of calculation of the operational equilibrium states

We assume that the specific growth rate μ (Χ ,5) obeys the Contois law (1.22) :

μ (Χ ,8) = ­ 4 ^ (1.129)
KcX + S

while the specific production rate a(E,S) is constant.

 64 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

The system (1.128.a­d) is rewritten as follows :

k3X = E (1.130.a)

5κ ^Χ = (μ *­5)8 (1.130.b)

kiX + k2P = D S i n ­ D S (1.130.c)

aES = DP (1.130.d)

Equation (1.130.b) is obtained by substituting (1.129) into (1.128.b).

We notice that (1.130.b) and (1.130.c) make sense only if

5<μ * S<Sin (1.131)

This means that, as pointed out above, the operational equilibrium states exist
only if these inequalities are satisfied. Otherwise, only the wash­out
equilibrium states are possible.

Substituting (1.130.b) into (1.130.a) gives :

­ k3(μ *­ D) ­
Ε = ^3Χ = — S (1.132)

which implies that the equilibrium values E,X,S are proportional to one
another. Substituting (1.132) into (1.130.d) gives :

P = fMil£Is^ (1.,33)

which means that the product equilibrium concentration Ρ is proportional to

the square of the substrate concentration S.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 65

Finally, substituting (1.133) into (1.30.c) leads to the following second degree
equation :

ak2k3AS^+ D(D + K^A) S ­ D^Sjn = 0 (1.134)

with A =

the solutions of which are written :

. ~D(D 4­ kiA) ± J(D^+ DkiA)^ + 4 akgkaAD^Sjn

S = · ο I. Λ — ' (1 · "135)
2 ak2k3A

It appears immediately that only one of these solutions is positive, and hence
physically plausible (the one which corresponds to the sign). Then the
equilibrium state is therefore fully analytically characterised by equations
(1.132), (1.133) and (1.135).

This simple example clearly shows that the explicit calculation of equilibrium
states may rapidly become very involved when the process is a combination of
several biological reactions and when plausible analytical models are
adopted for the reaction rates.

With (1.127)­(1.128), we have emphasised the existence of multiple

equilibrium states in a particular example. It is, however, obvious from that
example that it is a generic situation which occurs in every case where the
biotechnological process is a combination of autocatalysed microbial growth
reactions and enzyme catalysed productions.

In particular, if the process dynamics are described by the model (1.59) :

 66 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

^ = Κ φ (ξ )­0ξ 4­Ρ ­Β ξ

it can easily be shown that the wash­out equilibrium states are characterised
as follows :

­ concentrations of biomass and products/internal substrates : ξ ί = 0

­ concentrations of external substrates : ξ ι = τ

D+β ι

1.9.3. Stability of equilibrium states

It is a basic feature of nonlinear systems in general, and consequently of

bioreactor models in particular, that the equilibrium states can be stable or
unstable depending on the operating point; that is, on the values of D and F. It
is beyond the scope of this book to give a mathematical presentation of the
concept of stability, the intuitive meaning being sufficient for our purpose. We
shall limit ourselves to exposing the technique of stability analysis by
linearization (Lyapunov's first method, see Appendix 2) and to giving an
example of its application.

Stability analysis

Consider a biotechnological process described by the model (1.59) :

= Κ φ (ξ ) ­ Ο ξ ­h F ­ Β ξ ΐ g(ξ ,F,D) (1.136)

dt

where the function g is introduced for convenience. Suppose that we are

concerned with checking whether some equilibrium state (ξ , D,F) is stable or
not. Then, the linearized approximation of the model (1.136) around the
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 67

equilibrium state (which is also called the "linearized tangent model") is

defined as follows :

|(?­C) = ( | ) < i ­ D . ( § ) ( 0 ­ b ) . ( | ) ( F ­ F , (1.,37)

where the partial derivatives are evaluated at the equilibrium state. An explicit
calculation of the "linearized tangent model" of (1.136) readily yields :

^ ( ξ ­ ξ ) = Α (ξ .Ο .Ρ )(ξ ­ ξ ) ­ (D ­ 6)ξ + (F ­ F) (1.138.a)

with :

3φ
Mib,F) = (1.138.b)

Lyapunov's first method utilizes the eigenvalues of the matrix Α (ξ ,5,Ρ ) to check
on the stability of the equilibrium state. If the real parts of all the eigenvalues
are negative, the equilibrium state is stable. If any of the real parts of the
eigenvalues are positive, the equilibrium state is unstable. No conclusion may
be drawn in the case of all eigenvalues having zero real parts.

Example : substrate Inhibition

We consider the case of a simple microbial growth process (1.33)

S Qx

for which the dynamical model is as follows :

dX
= μ Χ ~ DX (1.139.a)
dt
 68 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

(1.139.b)

i i

il μ ^= 6.3 h"^

Φ
K M = 10 g/l
1
K| = 0.1 g/l
% 1

k
Ο
ο

(f> 0.0 τ — , ^ I—n >

0 1 2 3 4 5 6
Substrate Concentration S
Fig.1.11. Stable and unstable equilibrium states with the Haldane law

We assume that substrate inhibition occurs at high substrate concentrations

and is expressed by the Haldane law (1.20) :

μ (8) = (1.140)
KM + S + S^/K,

which is graphically represented in Fig 1.11.

From this figure and the model (1.139), we conclude that, apart from the wash­
out equilibrium state, two operational equilibrium states may exist, which are
denoted :

Sin­Si Sin­S2^
S i . X i = ^ 82, X2 — (1.141)
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 69

and which fulfil the following conditions

0 < §1 < V K M K | < §2 (1.142.a)

μ (8ι ) = μ (82) = 5 (1.142.b)

( S i , X i ) exists for all 5 < μ * while (Sg,Xg) ex'sts only when μ (8|η )<5<μ *
(since Sg cannot be physically greater than δ |η , see (1.118)).

In this particular example, the coefficient matrix of the linearized tangent model
is written :

A(X,8,D) = 0 Ω (1.143)
­Κ ι μ (8) ­ k i Q ­ D

with :

>2Λ
μ ο Χ KM­
Κ I )
Ω = (1.144)

KM + 8 + ­ΐ ^­

The eigenvalues of this matrix are (­Κ Ω ) and (­D). It is obvious that the
equilibrium state is stable if and only if:

Ω >0 or 8 < ^ K ^ K , (1.145)

Thus, it is clear that the equilibrium state ( 8 i , X i ) is stable while (82,X2) i s

unstable.

In fact, it is a general property of biological reactions which involve substrate

inhibition that they exhibit unstable equilibrium states.
 70 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

, Wash­out equilibrium state

D = 0.2 h'^
Sin = 5 g/l
k, = 2
μ ο = 6.3ί ι ^
KM = 10 g/l
K. = 0 . 1 g/l

0.40
Stable equilibrium state

Fig.1.12. Phase plane representation

An important point in the above example is obviously the determination of the

set of initial conditions (X(0), 8(0)) for which the steady­states X^, 8^ are stable
over a broad range. 8uch an investigation can be carried out by using phase
plane analysis. A typical phase­plane representation is shown in Fig.1.12. The
steady­state X i , 8^ is asymptotically stable with respect to the initial conditions
below the separatrix, while the initial conditions above the separatrix
irreversibly lead to a wash­out steady­state.

Therefore, when substrate inhibition takes place in a bioreactor, we discover

not only that the process can exhibit unstable behaviour but also that it can
lead to wash­out steady­states, in which the microbial life disappears
completely and the reactor comes to a definite stop. Therefore, in such a
situation, it is clear that industrial fermentations in the continuous operation
mode absolutely require feedback control in order to stabilise the process : this
issue will be addressed in Chapter 5.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 71

1.10. Extending the General Dynamical Model

So far, in this chapter, we have focused on the derivation and the analysis of
the general dynamical model (1.43) for biotechnological processes which take
place in a single, perfectly mixed bioreactor:

^ = Κ φ ( ξ ) ­ Ο ξ + Ρ ­α (ξ ) (1.146)

This model may easily be extended to deal with complex processes consisting
of several interconnected tanks, possibly with special constructional features
which allow, for instance, for biomass recycling or biomass accumulation. The
extension is simply obtained by considering an appropriate matrix dilution
rate:

D = [Dij] (1.147)

in equation (1.146) instead of the single scalar dilution rate adopted so far.
This issue is illustrated with two examples (Sections 1.10.1 and 1.10.2).
Furthermore, distributed parameter extensions of the general dynamical model
can also be derived, if necessary, as is shown in Section 1.10.3 for fixed bed
reactors.

1.10.1 Recycle bioreactor

In some fermentation processes (e.g. activated sludge processes), part of the

effluent biomass is recycled to the bioreactor. The bioreactor effluent is fed to
a clarifier (settler) which is used to separate substrate and biomass (Fig. 1.13).

Part of the settled biomass is then fed back to the bioreactor, while the surplus
biomass is removed from the process. We suppose that the process is a
"single substrate ­ single biomass" fermentation (1.33) and that the clarifier is
 72 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

such that the whole biomass is settled (none is left in the supernatent of the
settler). The dynamics of the process are written :

4dt x " 1 μ Χ ­ 'Db 0 ­Di" "Χ ' + 0 (1.148)

S ­ki 0 Db 0 S DinSin
XRJ . 0. ­D2 0 Ds 0

where X is the biomass in the bioreactor, S is the substrate in the bioreactor

and XR is the recycled biomass.

F
IN
bioreactor settler 1 Fe

V Vs

Fig. 1.13. Schematic view of a recycle bioreactor

It appears clearly that this model exactly fits the form of equation (1.146) but
with a matrix dilution rate. The entries of this matrix are defined as follows :

F|N+FR
Dh = bioreactor dilution rate (1.149.a)
V

FR + FW
settler dilution rate (1.149.b)

(1.149.C)
V
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 73

F|N+FR
D2= .. (1.149.d)
's

D,N = ^ (1.149.e)

with the notations of Fig. 1.13, i.e. F|N is the influent flow rate, FR the recycle
flow rate, Fw the waste flow rate, V the bioreactor volume, and Vs the settler
volume.

It should be noted also that, in this extended model, the matrix Κ may contain
rows of zeros corresponding to tanks in which no reaction occurs (as in the
clarifier of the example).

1.10.2. Two­Stage anaerobic digestion process with biomass

accumulation

Several times in this chapter (Sections 1.4.3, 1.6.2, 1.8.4) we have described
and discussed in detail the general dynamical model of the anaerobic
digestion process in a single bioreactor in which all the reactions involved take
place simultaneously. It is, however, common practice to implement the
process in a "two­stage" bioreactor as depicted in Fig.1.14. Assuming that the
reduced order scheme (1.87) holds, the idea is to devote the first tank to the
first reaction (acidification) and the second tank to the second reaction
(methanization). This means that the reaction scheme (1.87) has to be
modified as follows. In the first tank, we have :

Si i­^ Xi + S21 + P21 (1.150)

where Si is the organic substrate, Xi the acidogenic biomass, S21 the acetate
in the first tank and P21 the inorganic carbon in the first tank.

Moreover, in the second tank, we have :

 74 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Φ 2 ^

S22 ^ X2 + Pi + P22 (1.151)

where S22 is acetate in the second tank, X2 is the methanogenic biomass, Pi

is the methane and P22 the inorganic carbon in the second tank.

'in

F 21
m
D
S:
in
Λ , }

Fig.1.14. Two­stage anaerobic digestion reactor

The reaction scheme (1.150)­(1.151) excludes transfer of acidogenic biomass

between the two tanks, which is supposed to be achieved by an appropriate
filtration device. We also note that, as shown in Fig.1.14, the volume V2 of the
second reactor is larger than that of the first tank V i , because methanization is
the limiting step of the process. Then, applying the rules of Section 1.5.1, we
obtain the following dynamical model (compare with (1.85)):

dXi
(1.152.a)
"dT = 9i­DiXi

dS
l = ­ k i 9 i ­ D i S i + DiSi, (1.152.b)
"dt
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 75

dS 21
dt
=Μ ι ­ D1S21 (1.152.C)

dXg
= 92 ­ D2X2 (1.152.d)
dt

dS22
(1.152.e)
" d T ~ ~^2Φ 2 ~ D2S22 "t" ^2^21

(The dynamics of P i , P21, P22 are omitted for simplicity).

The model is written in matrix form :

d_
dt 'xr = ' 1 0 •Γ φ ι" - 0 0 0 0' • χ Γ + 0 (1.153)
Si ­ki 0 0 Di 0 0 0 Si DiSin
S21 k3 0 0 0 Di 0 0 S21 0
X2 0 1 0 0 0 D2 0 X2 0
S22. 0 ­k2. .0 0 ­D2 0 D2. S22. 0

In this model, Di = Fjn/Vi, D2 = Fjn/V2, and the yield coefficients k i , k2, ka have
exactly the same meaning as in (1.85).

Again, we see that equation (1.153) has exactly the same format as (1.146) but
with a gfeneraZ/seof dilution rate matrix.

1.10.3­ A distributed paranfieter model for a fixed bed bioreactor

A fixed bed bioreactor is a reactor where the biomass is immobilised on fixed

carriers such as polymers, porous glass or ceramics.

We consider a horizontal bioreactor operating under plug flow conditions

(rather than complete mixing conditions) as shown in Fig.1.15.
 76 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

influent effluent
F(S^0S/92)dz^

Fig. 1.15. plug­flow reactor

We suppose that a "single biomass/single substrate" reaction (1.33) takes

place in the reactor:

with a reaction rate φ = μ Χ .

The length of the bioreactor is equal to L. Consider a section of the reactor

with length "dz" located at a distance ζ from the bioreactor input (shaded area
in Fig 1.15). The constant reactor cross­section is denoted A. The volume of
the section is thus Adz.

The mass balance for the substrate concentration S around this section is
written as follows :

|­[SAdz] FS ­F^ S + | | d z ] ­k1μ XAdz (1.154)

dt

time variation influent effluent amount of

of the amount substrate substrate substrate
of substrate into the out of the consumed
in the section dz section dz section dz by the biomass
in the section dz

Equation (1.154) is rewritten :

 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 77

f =­ I f ­ M X a.,55,

while the mass balance of the biomass leads to :

^ =μ Χ (1.156)

where the hydraulic term has disappeared since the biomass is immobilised.

Equations (1.155) ­ (1.156) constitute the distributed parameter equivalent of

the dynamical model (1.52).

For completeness, we must define the limit conditions :

S(t,z=0) = S^(t) X(0,z) = Xo(z) S(0, 0 < ζ < L) = So(z) (1.157)

where Sin(t) is the influent substrate concentration, and Xo(z) and So(z) the
initial immobilised biomass and substrate concentrations.

1.11. References and Bibliography

The :

Bailey J.E. and D.F. OIlis (1986).

Mac­Graw Hill Kogakusha, Tokyo, 2"^ edition.
Moser A. (1988). Springer
Verlag, New York.
 78 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Section 1.3

Many papers are The

gives a survey specific :

Frederickson A.G. and H.M. Tsuchiya (1977), Microbial kinetics and dynamics,
in L. Lapidus and N.R. Amudson (Eds.), Chemical
Prentice­Hall, Englewood Cliffs, NJ.

A large number specific are in Appendix

1. The references in 1.3 are :

Andrews J.F. (1968). A mathematical model for the continuous culture of

microorganisms utilizing inhibiting substrates.
707­723.
Briggs G.E. and J.B.S. Haldane (1925). J., 242, 3973.
Contois D. (1959). Kinetics of bacterial growth relationship between
population density and specific growth rate of continuous cultures. J.
Gen. 21, 40­50.
Haldane J.B.S (1930). Enzymes, Longmans, London.
Jackson J.V. and V.H. Edwards (1975). Kinetics of substrate inhibition of
exponential yeast growth. 17, 943­964.
Michaelis L. and M.L. Menten (1913). Die kinetic der Invertinwirkung.
49, 334­369.
Monod J. (1942). Recherches sur la des
Hermann, Paris.
Rozzi A. (1984). Modelling and control of anaerobic digestion processes.
Trans. Meas. 6, n°3, 153­159.
Tessier G. (1942). Croissance des populations bacteriennes et quantites
d'aliments disponibles. Rev. Sci., Paris, 80­209.
Topiwala H. and C.G. Sinclair (1971). Temperature relationship in continuous
culture. 13, 795­813.
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 79

Verhulst R. (1838). Notice sur la loi que la population suit dans son
accroissement. Corn Math. etPhys., A. Quetelet (Ed.), t. X, 113.

The specific growth rate models including product inhibition are drawn from :

Aiba S., M. Shoda and M. Nagatani (1968). Kinetics of product inhibition in

alcohol fermentation. Biotechnol. Bioeng., 10, 845­864.

Section 1.4

The process examples presented in this section are adapted from the
following references :

Yeast growtti:
Sonnleitner B. and O. Kappeli (1986). Growth of Saccharomyces cerevisiae is
controlled by its limited respiratory capacity : formulation and verification
of a hypothesis. Biotechnol. Bioeng., 28, 927­937.

Anaerobic digestion :
Mosey F.E. (1983). Mathematical modelling of the anaerobic digestion
process: regulatory mechanisms for the formation of short­chain volatile
acids from glucose. Water Sci. Technol., 15, 209­232.
Sinechal X., M. Installe and E.J. Nyns (1979). Differentiation between acetate
and higher volatile acids in the modeling of the anaerobic
biomethanization process. Biotechnol. Lett., 1, 309­314.

lactic fermentation:
Spinnler H.E., C. Bouillanne, M.S. Desmazeaud and G. Corrieu (1987).
Measurement of the partial pressure of dissolved C O 2 for estimating the
concentration of St. thermophilus in coculture with Lb. bulgaricus. Appl.
Microbiol. Biotechnol., 25, 464­470.
 80 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Section 1.5

Reaction networl<s and some of ttieir algebraic properties tiave been studied
by:

Feinberg M. and F.J.M. Horn (1974). Dynamics of open chemical systems and
the algebraic structure of the underlying reaction network. Chem. Eng.
Sci., 29, 775­787.

Ttie modelling of tfie gaseous outflow rates is studied in particular in :

Bellgart K.H., W. Kuhlman and H.D. Meyer (1983). Deterministic growth model
of Saccharomyces cerevisiae. Parameter identification and simulation; In
A. Halme (Ed.), Modelling and Control of Biotechnical Processes,
Pergamon, Oxford, 67­74.

Section 1.7

In this section and throughout the book we refer, when needed, to properties of
matrices and linear algebra. The following book constitutes a good
introduction and summary on the subject:

Strang G. (1980). Linear Agebra and its Applications. Academic Press,

Orlando.

Section 1.8

Singular perturbation techniques have been used for reducing the general
dynamical model. The theory of singular perturbation is treated in :
 Chap.1. DYNAMICAL MODELS OF BIOREACTORS 81

Kokotovic P., H.K. Khalil, J. O'Reilly (1986). Singular Perturbation Methods in

Control: Analysis and Design. Academic Press, London.

Section 1.9

The concepts and notions of stability theory of linear and nonlinear systems
are described e.g. in :

Vidyasagar M. (1978). Nonlinear Systems Analysis. Prentice­Hall, Englewood

Cliffs, NJ.
Willems J.L (1970). Stability Theory of Dynamical Systems. Nelson, London.

Equilibrium states of bioreactors have been widely studied, especially the

appearance of possible unstable equilibrium states, e.g. in :

Agrawal P., C. Lee, H.C. Lim and D. Ramkrishna (1982). Theoretical

investigations of dynamic behavior of isothermal continuous stirred tank
biological reactors. Chem. Eng. Sci., 37, n°3, 453­462.
Agrawal P. and H.C. Lim (1984). Analysis of various control schemes for
continuous bioreactors. Adv. Biochem. Eng., 30, 61­90.
Antunes S. and M. Installe (1981). The use of phase­plane analysis in the
modelling and the control of a biomethanization process. Proc. &^ IFAC
World Congress, 22, 165­170.
DiBiasio D., H.C. Lim, W.A. Weigand and G.T. Tsao (1978). Phase­plane
analysis of feedback control of unstable steady­states in a biological
reactor. AlChEJ., 24, n°4, 686­693.
Van den Heuvel J.C. and R.J. Zoetmeyer (1982). Stability of the methane
reactor : a simple model including substrate inhibition and cell recycle.
Process Biochem., May­June, 14­19.
 82 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Section 1.10

Biotechnological processes in fixed bed and fluidised bed reactors are

presented e.g. in:

Poncelet D., R. Binot, H. Naveau and E.J. Nyns (1985). Biotechnologie des lits
fluidises en reacteurs cylindriques et tronconiques. Trib. Cebedeau, 38
(494). 3­12; 38 (497), 33­48.
Schugerl K. (1989). Biofluidization : application of the fluidization technique in
biotechnology. Can. J. Chem. Eng., 67, 178­184.
Wang D.I.C., C.L Cooney, A.L Demain, P. Dunhill, A.E. Humphrey and M.D.
Lilly (1979). Fermentation and Enzyme Technology. John Wiley & Sons,
New York.
Wiseman A. (1978). Topics in Enzyme and Fermentation Biotechnology. John
Wiley & Sons, New York.
 Chap.2. KINETIC MODELLING, ESTIMATION AND CONTROL 83

CHAPTER 2

KINETIC MODELLING, ESTIMATION

AND CONTROL IN BIOREACTORS :
AN OVERVIEW

2.0. Introduction

When a bioengineer is faced with the problem of developing a mathematical

model for a bioreactor (with a view to performance assessment or control
design), his first task is to establish the reaction scheme for the process.
Numerous examples of this have been given in Chapter 1. He can then easily
write down the General Dynamical Model (1.43) associated with the scheme
by using the rules of Section 1.5.2. This is simply a set of ordinary differential
equations of the following form :

^ = Κ φ (ξ )­Ο ξ + Ρ ­ α (1.43) = (2.1)

where ξ is the state vector (i.e. the set of component concentrations), Κ a yield
coefficient matrix, φ (ξ ) the vector of reaction kinetics (also called reaction
rates), D the dilution rate, F a set of feed rates and Q a set of gaseous outflow
rates.

Once equation (2.1) is written, however, the model construction is far from
being completed. Actually, the most difficult task remains to be performed,
namely the task of modelling the reaction kinetics φ (ξ ) which is discussed in
 84 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Section 2.1. On the grounds of that discussion, the concept of minimal

modelling of the kinetics is introduced and illustrated with simple examples in
Section 2.2.

A review of the monitoring and control problems which are addressed in the
remaining chapters of the book, is then presented in Sections 2.3 and 2.4.

2.1. Difficulties in Modelling the Reaction Kinetics

In Section 1.4.4, the reaction rates φ j(ξ ) (j = 1 M ) were represented by :

φ )(ξ ) = α )(ξ )ί Π ξ η (1.47) = (2.2)

V n~j

The significance of this equation is as follows : the reaction rate φ \ is

proportional to the concentrations of the reactants ξ η (including the auto­
catalysts) involved in the reaction and to another function α ](ξ ), termed the
specific reaction rate. The usual approach in bioreactor modelling consists in
adopting a particular analytical structure for each specific reaction rate α ](ξ )
and in calibrating its internal coefficients on the basis of experimental data.

However, this modelling exercise is often very hazardous because it comes up

against three major difficulties :

1) First of all, one has to select, for each particular fermentation, the biological
and physico­chemical factors which are supposed to influence the kinetics
and must therefore be incorporated in the model.

2) Second, once this selection has been made, one has to choose an
appropriate analytical description for each α ](ξ ), for instance among the
(long) list of Appendix 1; this critical choice is the subject of continuing
controversy in the literature (which is probably why the list of Appendix 1 is
so long); most often these analytical expressions take the form of rational
 Chap.2. KINETIC MODELLING, ESTIMATION AND CONTROL 85

functions (i.e. polynomial ratios) of both the state variables and a set of
constant kinetic coefficients.

3) Finally, once a specific structure has been chosen one faces intricate
parameter identifiability difficulties when trying to calibrate the kinetic
coefficients from experimental data; the difficulties arise mainly from a lack
of experimental reproducibility and a lack of statistical significance of the
data; this issue will be discussed further on in this section.

A typical example of kinetic modelling has been given in Section 1.4 for a
simple microbial growth process governed by the Contois law. The reaction
rate was written :

9(X,S) = a(X,S) XS = μ (X,S) X (2.3)

with a(X,S) = κ X+S ^^'^^

In that example, there are two kinetic coefficients : μ * and Kc. It can also be
seen that a(X,S) effectively has the form of a rational function. Another
example follows.

Example of kinetic modelling

Consider a culture of L Bulgaricus in a lactic fermentation process, described

by the reaction scheme (1.40.d)­(1.40.e), which is written as follows :

S—Ux (2.5.a)

S+X >P + X (2.5.b)

 86 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

with glucose S, biomass X and lactic acid P. Note that it is the standard
scheme (1.36) of microbial growth with associated enzyme catalysed
production. It can easily be shown that the dynamical model can be written :

­1 ­1 ­Df DSi, (2.6)

dt
ki 0 0
0 ko 0

We first suppose that the enzymatic catalysis is governed by the conventional

Michaelis­Menten law. The reaction rate φ ο is therefore as follows :

μ ;Χ 3
9c(X,S) = (2.7)
K i + S

On the other hand, it is well known that the bacterial growth is inhibited by
lactic acid. Hence, according to Section 1.3, it is a reasonable assumption to
represent the specific growth rate by the combination of a Monod law (1.18)
and a product inhibition function of the form (1.23), as follows :

Kp
μ (8,Ρ ) = (2.8)
Kg+S K p + P

which implies that the growth kinetics are as follows :

HaKpSX
(2.9)
^g^^'^'^^=(K2+S)(Kp+P)

The full model contains five kinetic coefficients :

(2.10)

while the specific reaction rates are the following rational functions :
 Chap.2. KINETIC MODELLING. ESTIMATION AND CONTROL 87

Thus, with the definitions (2.11), we have adopted one plausible kinetic model
for this fermentation process. The reader should, however, easily realize that
many other models, just as plausible as (2.11), could have been chosen from
the list of Appendix 1. In fact, the choice is a priori virtually unlimited. It should,
however, be obvious that identification techniques could help to discriminate
between the models.

On the Idnetic

Clearly, kinetic modelling (with a view to performance assessment in particular

applications) makes sense only if the kinetic coefficients have a well specified
numerical value. Since these values are generally not known a priori, they
need to be estimated (identified) from process measurements. Unfortunately,
numerous studies (see references) devoted to parameter estimation in
biological models have shown that, in practice, the identifiability of the kinetic
coefficients is far from being guaranteed and may even be an insuperable
difficulty in most applications. A detailed discussion of this question is beyond
the scope of this book.

The reader will, however, be convinced of the relevance of the issue simply by
considering Table 2.1 and Fig.2.1 taken from Holmberg (1983). Table 2.1
summarizes the identification results of the kinetic coefficients μ * and Km of a
Monod law from batch cultivation data of B. In each experiment,
the environmental conditions were identical except for the initial substrate
concentration (So in the Table). It is clear that the different identification runs
produce a wide dispersion of the numerical values of the two parameters.
Furthermore, the confidence intervals are so large that, in some instances, the
coefficients could just as well be negative. Fig 2.1 illustrates the point in a
complementary way by showing that very different sets of kinetic coefficients
can actually provide quite similar simulation results which fit the data equally
well.
88 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Cultivation Initial substrate Estimates Kinetic Coeff.

So (g/l) μ * (h­1) KM (g/l)
1 11.6 1.0 ± 2 . 0 6.8 ± 22.3
2 7.0 1.1 ± 0 . 8 1.8 ± 2 . 2
3 18.2 0.7 ± 0 . 2 12.9 ±
4 25.0 0.3 ± 0.2 7.0 ± 11.9

Table 2.1.Identification results of the kinetic coefficients μ and KM

(Reprinted by permission from A. Holmberg (1983), On thie
accuracy of estimating the parameters of models containing
Michaelis-Menten type nonlinearities. In G.C, Vansteenkiste
and P,C, Young (Eds,), Modelling and Data Analysis in
Biotechnology and Medical Engineering, North-Holland,
Amsterdam, p.202)

Ε
κ CONCENTRATION OF BACTERIA
• CONCENTRATION OF TOTAL SUGAR
SIMULATED ( = 1 4 . K, = 12 )
SIMULATED m m = 3 5 , K , = CO)

10 15

TIME (hoors)

Fig.2.1. Simulation with two different sets of kinetic coefficients

(Reprinted by permission from A, Holmberg (1983), On
the accuracy of estimating the parameters of models
containing Michaelis-Menten type nonlinearities. In G.C,
Vansteenkiste and P,C, Young (Eds,), Modelling and
Data Analysis in Biotechnology and Medical
Engineering, North-Holland, Amsterdam, p.202)
 Chap.2. KINETIC MODELLING. ESTIMATION AND CONTROL 89

2.2. Minimal Modelling of Reaction Kinetics

On the grounds of the previous discussion, there is a clear incentive, at least

from an engineering viewpoint, to try to avoid kinetic modelling. This will be
the main challenge of this book :

To
of

This notion of is formalised as follows. We shall represent

the reaction rates φ j(ξ ) 0 = 1 ,..·, M) by :

φ j(ξ ) = hj(ξ )pj(ξ ) (2.12)

where hj(ξ ) is a known function of the state ξ while ρ ](ξ ) is an unknown

function. More generally, the vector φ (ξ ) of reaction rates will be written as :

φ (ξ ) = Η (ξ )ρ (ξ ) (2.13)

with Η (ξ ) an Mxr known matrix and ρ (ξ ) a vector of unknown functions of ξ , with

dim ρ ( ξ ) = r.

The idea is to insert into Η (ξ ) only the (possibly very limited) prior knowledge
which is available regarding the kinetics and then to consider ρ as a
completely unknown time varying parameter.

The form of equation (2.13) is very flexible and allows us to account for various
kinds of uncertainty and to cover a wide spectrum of practical situations.

A first situation occurs when there is no prior knowledge at all concerning the
kinetics. Minimal modelling then reduces to no i.e. the reaction
kinetics are the unknowns of the model:

ρ (ξ ) = φ (ξ ) Η (ξ ) = ΐ Μ (2.14)
 90 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Another possibility is to consider that the unknown parameters are the specific
reaction rates α (ξ ) introduced in Section 1.5.3. In that case, we have :

ρ (ξ ) = α (ξ ) Η (ξ ) = 0(ξ ) (2.15)

But many other (more or less complicated) variants are obviously possible.
This is illustrated by some examples.

Examples of minimal modelling

Consider again the reaction scheme (2.5) of an enzyme catalysed production

process :

s—Ux

+X >X + P

If there is no prior knowledge of the kinetics, we simply define the unknown

parameters as follows :

Ρ ι (ξ ) = 9g(S,X,P) Ρ 2(ξ ) = 9c(S,X,P) (2.16)

If the specific reaction rates are the unknown parameters, we have :

Ρ ι (ξ ) = ag(S,P) Ρ 2(ξ ) = ae(S,X) (2.17)

while the corresponding matrix Η (ξ ) is written :

Η (ξ ) = SX 0 (2.19)
0 SX

so that we have (compare with (2.7), (2.9)) :

 Chap.2. KINETIC MODELLING. ESTIMATION AND CONTROL 91

>g(S.X,P)' = SX 0 " "Ρ Γ (2.20)

. 9c(S.X) . . 0 S X . .P2.

If we wish the specific growth rate μ ( 8 , Ρ ) (see (2.8)) to be treated as an

unknown parameter of the model, we may define :

Ρ ι (ξ ) = μ (8,Ρ ) ρ 2(ξ ) = α ^(8,Χ ) (2.21)

The corresponding Η (ξ ) is then :

Η (ξ ) = Γ Χ 0 (2.22)
0 SX

We may also suppose that, in the specific growth rate model (2.8), the kinetic
coefficients and K2 are known, while the product inhibition function is
unknown, and that the specific production of the enzymatic reaction has to be
treated as an unknown parameter. Hence we have (compare again with (2.7),
(2.9)):

9g(S,X,P) Pi(P) (2.23)

9c(S,X) . LP2(S)
0 X

2.3. Software Sensors for Bioreactors

Besides the problems inherent to kinetic modelling, another essential difficulty

lies in the absence, in most applications, of cheap and reliable sensors
capable of providing direct real­time measurements of the state variables
required to implement advanced monitoring and control methods on
bioreactors. The main variables (such as biomass, substrate or metabolite
concentrations) generally need to be determined through laboratory analyses.
The cost and duration of these analyses limit the frequency of the sampling. In
 92 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

addition, even when on­line sensors exist for biomass or metabolites, they are
often not robust enough for routine industrial applications.

The design of Software Sensors to cope with the lack of instrumental sensors
in bioreactors is one of the main concerns of this book. A software sensor is an
algorithm for the on­line estimation of the state variables and the parameters
which are not measurable in real time, on the basis of related measurements
which are more easily accessible. The General Dynamical Model of the
process (2.1) is the main ingredient for the design of a software sensor.

In control science terminology, a software sensor for the estimation of state

variables is called a state observer (or simply an observer) while for the
estimation of internal model parameters, it is called a parameter estimator. We
shall adopt this terminology in the sequel. The design of software sensors, i.e.
of state and parameter estimation algorithms, will be extensively studied in
Chapters 3 and 4.

PRIOR KNOWLEDGE

Reaction Kinetics Yield Coefficients

(vector φ (ξ )) (matrix K)

Chapter 3 ­ Section 3.2 Known Known

Chapter 3 ­ Section 3.3 Unknown Known

Chapter 4 ­ Section 4.3 Unknown Unknown

Table 2.2. Assumptions for the design of state observers

 Chap.2. KINETIC MODELLING. ESTIMATION AND CONTROL 93

Different state observers will be presented, depending on the prior knowledge

available concerning the reaction kinetics and the yield coefficients. The
presentation is organised according to Table 2.2.

With the definition (2.13), the general dynamical model (1.43) is rewritten :

= Κ Η (ξ )ρ (ξ )­Ο ξ +F­Q (2.24)

dt

As we have mentioned previously, the parameters ρ (ξ ) will be considered

throughout the book (except in Sections 3.1 and 3.2), as completely unknown
time varying parameters. They are therefore of primary concern in the design
of parameter estimators. But, in some instances, the yield coefficients (matrix
K) are also badly known and/or time varying, and therefore also concerned.
An overview of the presentation of the parameter estimators is given in Table
2.3.

Chapter 3 ­ Section 3.4 Estimation of ρ with known Κ

Chapter 4 ­ Section 4.1 Estimation of ρ = α independently of Κ

Chapter 4 ­ Section 4.2 Joint estimation of ρ = α and Κ

Chapter 4 ­ Section 4.4 Estimation of Κ independently of φ = Hp

Table 2.3. Overview of the presentation of parameter estimators

The advantages of considering ρ (ξ ) as unknown time varying parameters to be

estimated on­line are threefold :
 94 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

1) First of all, it allows avoidance of the difficult choice of a particular analytical

expression for the process kinetics of the application under study.

2) Secondly, it allows avoidance of the identifiability difficulties (mentioned and

illustrated in Section 2.1) arising when trying to calibrate the coefficients of
the chosen analytical expression from input­output data.

3) Finally, it allows a search for possible correlations between the estimated

reaction rates and the different physico­chemical and biochemical factors
which are assumed to influence them. This latter point will be illustrated in
Chapter 4 (Section 4.1.3).

2.4. Adaptive Control of Bioreactors

In industry, the control of bioreactors is most often limited to the regulation of

pH and temperature. There is, however, no doubt that computer control of the
biochemical state variables (such as substrate or product concentrations) can
help to increase the process performance significantly. This issue will be
treated in Chapter 5, where we shall present several methods for the design of
adaptive linearizing controllers.

The purpose of computer control is to keep some process state variable close
to a prespecified value (called a set point), in spite of disturbances and
variations in the process kinetics, by acting on the feed rate of an external
substrate. The task of the controller is to determine, at each instant, the best
control action (i.e. the best substrate feed rate), on the basis of the data
collected on line by the sensors and of the estimates provided by the state
observers (which then more particularly deserve the name of software
sensors).
 Chap.2. KINETIC MODELLING, ESTIMATION AND CONTROL 95

Adaptive Control

Chapter 5 ­ Section 5.2 independently of the kinetics, with known or

unknown yield coefficients

Chapter 5 ­ Sections 5.3, 5.4 with unknown kinetics and known yield
coefficients

Chapter 5­ Section 5.5 with unknown kinetics and unknown yield

coefficients.

Table 2.4. Assumptions for adaptive control design

The control algorithms will be obtained by appropriate algebraic manipulations

of the General Dynamical Model (2.1). But, since the kinetics are not fully
known (recall that our viewpoint is a minimal modelling of the kinetics), they
are replaced in the control algorithm by on­line estimates provided by one of
the various parameter estimators listed in Table 2.3. A control algorithm which
incorporates a parameter estimator is called an adaptive controller because it
has the potential to adapt itself to variations in the kinetics. Different adaptive
controllers will be presented depending on the prior knowledge available
concerning the reaction kinetics and the yield coefficients. The presentation is
organised according to Table 2.4.
 96 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

2.5. Conclusions and Perspectives

An overview of the monitoring and control problems, which are addressed in

the subsequent chapters of this book, has been presented. The software
sensors and the adaptive control algorithms will be designed on the
assumption of a poor knowledge and a minimal modelling of the kinetics. This
is justified by the difficulties that usually appear when trying to identify
analytical models of the kinetics of bioprocesses, as has been discussed in
Section 2.1.

We would also like to stress the point that, in the rest of the book, we shall be
concerned only with stirred tank bioreactors (either in batch, fed­batch or
continuous operating modes). The extension of the estimation and control
algorithms to more general biotechnological systems (such as those
mentioned in Section 1.10 : biomass recycling, biomass accumulation, fixed
and fluidised beds, distributed models) is far from straightforward and has
received little attention in the literature. Some relevant references are listed in
the next section.

2.6. References and Bibliography

Ttie following paper is a survey on various aspects (instrumentation,

estimation, control) of computer applications to fermentation processes :

Wang N.S. and G.N. Stephanopoulos (1984). Computer applications to

fermentation processes. CRC Critical Reviews in Biotectinology, 2, 1­103.

Section 2.1

Tt)e problem of ctioosing one particular analytical structure for ttie specific
growtti rate is discussed in e.g. :
 Chap.2. KINETIC MODELLING. ESTIMATION AND CONTROL 97

Spriet J. (1982). Modelling the growth of microorganisms : a critical appraisal.

In S. Rinaldi (Ed.), Environmental Systems Analysis and Management,
North­Holland, Amsterdam, 451­465.

The practical identifiability of the specific growth rate models is discussed e.g.
in:

Holmberg A.(1982). On the practical identifiability of microbial growth models

incorporating Michaelis­Menten nonlinearities. I^ath. Biosc. 62, 23­43.
Holmberg A. (1983). On the accuracy of estimating the parameters of models
containing Michaelis­Menten type nonlinearities. In G.C. Vansteenkiste
and P.C. Young (Eds.), Modelling and Data Analysis in Biotechnology
and Medical Engineering, North­Holland, Amsterdam, 199­208.
Munack A. (1989). Optimal feeding strategy for identification of Monod­type
models by fed­batch experiments. In N.M. Fish, R.I. Fox and N.F. Thornhill
(Eds.), Computer Applications in Fermentation Technology : Modelling
and Control of Biotechnological Processes, Elsevier, Amsterdam, 195­
204.
Munack A. and C. Posten (1989). Design of optimal dynamical experiments for
parameter estimation. Proc. ACC, Pittsburgh.
Nihtila M. and J. Virkkunen (1977). Practical identifiability of growth and
substrate consumption models. Biotechnol. Bioeng., 19, 1831­1850.
Vialas C , A. Cheruy and S. Gentil (1986). An experimental approach to
improve the Monod model identification. In A. Johnson (Ed.), Modelling
and Control of Biotechnological Processes, Pergamon, Oxford, 155­160.

Section 2.3

A tutorial on identification and state estimation with a special emphasis on

bioprocesses can be found in :

Beck M.C. and P.C. Young (1988). An Introduction to System Identification,

Parameter and State Estimation. In N.M. Fish, R.I. Fox & N.F. Thornhill
 98 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

(Eds), Computer Applications in Fermentation Technology : Modelling

and Control of Biotechnological Processes, Elsevier, 129­158.

Section 2.4

Surveys of the control problems which arise in particular classes of biological

systems can be found in :

Beck M.C. (1986). Identification, estimation and control of biological waste­

water treatment processes. Proc. lEE, Part D, 133, 254­264.
Johnson A. (1987). The control of fed­batch fermentation processes. A survey.
Automatica, 23, 691­705.

Section 2.5

Some typical problems of modelling and control of distributed parameter

bioreactors are treated e.g. in :

Dochain D. and J.P. Babary (1989). Control of distributed parameter

bioreactors via orthogonal collocation. Proc. 5^^ IFAC Symposium on
Control of Distributed Parameter Systems. Pergamon, Oxford.
Isaacs S., A. Munack and M. Thoma (1986). Use of orthogonal collocation
approximation for parameter identification and control optimization of a
distributed parameter biological reactor. AlChE Meeting, Miami Beach.
Jorgensen S.B. (1986). Fixed bed reactor dynamics and control ­ A review.
Proc. IFAC Symposium on Control of Distillation Columns and Chemical
Reactors, Pergamon, Oxford, 11­24.
Luttman R., A. Munack and M. Thoma (1985). Mathematical modelling,
parameter identification and adaptive control of single cell protein
processes in tower loop bioreactors. Advances in Biochemical Eng.,
Springer Verlag, 32, 95­205.
Ray W.H. (1981). Advanced Process Control. Mc­Graw Hill, New York.
 Chap.2. KINETIC MODELLING. ESTIMATION AND CONTROL 99

The
:

Hamalainen R.P., A. Halme and A. Gyllenberg (1975). A control model for

activated sludge wastewater treatment process. 6^^ IFAC
Boston, the Instrument Society of America, Paper 61:6.
Holmberg A. (1983). A microprocessor­based estimation and control system
for the activated sludge process. In A. Halme (Ed.),
Pergamon, Oxford, 111­119.
Marsili­Libelli S. (1980). Reduced­order modelling of the activated sludge
process. 9,15­32.
Marsili­Libelli S. (1984). Aggregated model of the activated sludge processes.
24, 171­190.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 101

CHAPTER 3

STATE AND PARAMETER ESTIMATION

WITH KNOWN YIELD COEFFICIENTS

3.0. Introduction

This chapter deals with state estimation and parameter estimation in

bioreactors when the yield coefficients are known.

In Sections 3.1, 3.2 and 3.3 we address the state estimation problem, that is
the problem of reconstructing the time evolution of the non measured state
variables from the measured ones. An algorithm designed for that purpose is
called an observer.

In Section 3.1 the concept of observability is discussed in relation to the

structure of the general dynamical model of biotechnological processes. We
introduce a distinction between exponential observers (for which it is possible
arbitrarily to fix the speed of convergence of the estimated variables towards
their true values) and asymptotic observers (for which the speed of
convergence is exclusively determined by the experimental conditions).
Examples of lack of exponential observability are also given.

Section 3.2 is dedicated to the design of exponential observers for

biotechnological processes, under the (hypothetical) assumption of full model
 102 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

knowledge : we assume in particular a knowledge of the mathematical

structure of the kinetic functions φ (ξ ) and of their parameters. The design
procedure for extended Luenberger and Kalman observers is presented and
illustrated with practical examples.

In Section 3.3 we show how specific dedicated asymptotic observers can be

designed in the (realistic) situation of partial model knowledge, namely when
the mathematical structure of the kinetics φ (ξ ) is unknown.

Two real life applications of asymptotic observers are presented : the first one
concerns biomass estimation in an anaerobic digestion process from acetic
acid and methane measurements; the second one refers to biomass and
product estimation in a PHB acid production process from oxygen
measurements. Some practical implementation aspects are also briefly
discussed.

Section 3.4 deals with the on­line estimation of the reaction kinetics φ (ξ ) in
bioreactors. Two different parameter estimators are proposed : the first one is
called an observer­based estimator because it is based on a variant of the
observers described before; the second one is based on the reformulation of
the process model in a linear regression form. An example of convergence
analysis is given. A case study devoted to the on­line estimation of microbial
specific growth rates is then carried out in depth. The use of extended Kalman
filtering for on­line estimation of the kinetics is also presented and compared
with the previous algorithms. Three experimental applications are described :
estimation of the specific growth rate in a fed­batch ethanolic fermentation
process, estimation of the specific reaction rates in a continuous anaerobic
digestion process, estimation of the reaction rates in a lactic fermentation
process.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 103

3.1. On State Observation in Bioreactors

We consider a biotechnological process described by the general state space

model (1.43) introduced in Chapter 1, that is :

^ = Κ φ (ξ ) ­ Ο ξ ­ Q + F (1.43) = (3.1)

We suppose that we have a full knowledge of this model (see Chapter 2) : the
structure of the reaction kinetics φ (ξ ) is completely known; also the numerical
values of all the coefficients involved in the model (i.e. yield coefficients and
kinetic coefficients) are given.

In addition, we assume that the dilution rate D, the feed rates F and the
gaseous outflow rates Q are known on­line, and that a subset of the state
variables is measured on­line. The vector of these measurements is denoted
ξ ι and is related to the state of the system as follows :

ξ ι = ί ξ

where the qxN matrix L is an elementary matrix which selects the measured
components of ξ . On the other hand, the vector of unmeasured states is
denoted ξ 2, so that ( ξ ι , ξ ζ ) constitutes a partition of ξ .

A state observer is an algorithm designed to reconstruct the non measured

state variables from the measured ones.

A general class of state observers for nonlinear systems of the form (3.1) is as
follows :
 104 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

General state observer

^ = Κ φ (ξ ) ­ Ο ξ ­ Q + F + Ω (ξ )[ξ ι ­ ξ ,] (3.2)

where ξ denotes the on­line estimate of ξ , Ω (ξ ) is an Nxq gain matrix

depending on ξ , and ξ ι = L|.

The observer equation (3.2) can be interpreted as a copy of the model (3.1)
with an additional driving term which is proportional to the observation error of
the measured part of the state ( ξ ι ­ ξ ι ) , and which disappears in the case of
perfect estimation.

In practice, the on­line reconstruction (observation) of the missing states is

obtained simply by integrating equation (3.2) on the supervising computer.
We note that the estimated state vector ξ includes all the state variables, even
those which are measured.

The state observer design problem reduces to that of a reasonable choice of

the gain matrix Ω (ξ ).

To solve this problem, we introduce the observation error:

e=ξ ­ξ (3.3)

the dynamics of which are easily shown to be governed by the following

differential equation (obtained by subtracting (3.2) from (3.1)) :

^ = Κ [φ (ξ + e) ­ φ (ξ )] ­ De ­ Q(^)Le (3.4)

A zero observation error e = 0 is clearly an equilibrium point of the error model

(3.4). Hence, it makes sense to consider the linearized tangent approximation
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 105

of (3.4) around e = 0 (see Section 1.10.3 for a definition of the linearized

tangent approximation of a nonlinear model). It is written as follows :

^ = [A(^)­Q(|)L]e (3.5.a)

with :

Α (ξ ) = Κ (ξ ) ^ ^ ­ DIN (3.5.b)

Thus, the design problem can be stated as the problem of choosing Ω (ξ ) in

such a way that the linear time­varying model (3.5) has desirable properties.

Exponential observability

The particular form (3.5.a) of the observation error dynamics indicates that we
can hope to assign an arbitrarily fast (exponential) rate of convergence of the
observation |(t) to its true value ξ (t) if we can freely fix the eigenvalues of the
matrix [Α (ξ ) ­ Ω (ξ )1] by an appropriate choice of the matrix Ω (ξ ).

When such a possibility exists, the system (3.1) is said to be exponentially

observable and the observation scheme (3.2) is called an exponential
observer.

A simple test to check whether a given system is exponentially observable or

not is expressed by the following property.

Property 3.1. A necessary condition for the process (3.1) to be exponentially

observable is that the following rank condition holds :
 106 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Exponential observability condition

Δ
rank(0) = rank L = N Α (ξ ) = Κ ­DIN (3.6)
Ι Α (ξ )

.ί Α (ξ )'­\

along the process trajectories. The matrix Ο is called the observability matrix.

It is important to note that condition (3.6) is only a necessary (but not a

sufficient) exponential observability condition. Its main interest is to provide an
easy way to detect those processes which are generically not exponentially
observable. This is illustrated by two simple examples.

Example 1

Consider the simple microbial growth process (1.33)

S—Cx

the dynamics of which are written :

d_,—, r
­ki φ ­DrS
dt
1 X 0

It is easy to see that the matrix Α ( ξ ) can be written as follows

Α ( ξ ) = Γ ­Κ ι %­0 ­ k i 9 x '
. 9s Φ χ ­D.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 107

Assume now that the substrate concentration S is measured on­line. Thus,

using the notations introduced before, we have :

ξ = [ξ ι .ξ 2Γ ξ ι =8 ξ 2=Χ L = [1 0]

The observability matrix Ο of property 3.1 is then equal to:

' L" = 1 0
.LA ­ki9s­D ­ki9x

The observability condition (3.6) is fulfilled if rank(O) = 2, i.e. if φ ^^ 0. If φ = μ Χ

(see 1.51), we must have :

φ χ = μ +a^x^o

In the particular (but widely used) case when the specific growth rate μ is only
a function of the substrate concentration S, 3μ /9Χ = 0, we can expect the
biomass concentration X to be exponentially observed from S by using an
observation scheme such as (3.2) whenever μ (8) ^ 0.

Consider now a process similar to the previous one but with an additional
synthesis product Ρ :

­Cx + i (3.7)

for which the dynamics are written :

 108 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

­ki φ ­Df DSin (3.8)

dt
1 0
Lk2j 0

The matrix Α (ξ ) is as follows :

Α (ξ ) = ­ki9s­D ­Κ ι φ χ ­Κ ι φ ρ •
Φ 3 Φ χ ­D φ ρ
k29s kgcpx k29p­D

Suppose that the product concentration Ρ is the only component available for
on­line measurement. The matrix L is equal to :

L = [0 0 1]

The observability matrix Ο is obtained after some calculation

0 0 0 1
k29s k2φ χ k29p­D

k29s^ k2φ χ φ k29p<iH­D^.

with φ = <Px ­ ki9s + k29p ­ 2D

It is straightforward to check that the determinant of Ο is equal to zero

whatever the values of 95, ψ χ and φ ρ . Therefore, we know from the
observability test (3.6) that the process components S and X cannot be
reconstructed at an arbitrary exponential rate from measurements of Ρ by
using the observation scheme (3.2).

As a matter of fact, the above calculations can be repeated with S or X as the

only measured variable : they lead to the same result, i.e. the determinant of Ο
is equal to zero.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 109

We can then draw the following important conclusion : the above process (3.7)
is not exponentially observable if is its state
cannot be exponentially reconstructed, at an arbitrarily fast rate, with an
observation scheme of the form (3.2). This result is since it is
independent of the values of 9 5 , φ χ and φ ρ , and therefore of the structure of
the reaction rate φ .

Asymptotic observers

When the system (3.1) is not exponentially observable (i.e. when the
eigenvalues of Α (ξ ) ­ Ω (ξ )1 cannot be freely assigned) but when the error
system (3.5.a) nevertheless has an asymptotically stable equilibrium point at e
= 0, the process can still be observed, but its dynamics will be partially
determined by the experimental conditions through Α (ξ ). Observers of this kind
are called (since one cannot freely assign their
dynamics).

Other forms of asymptotic observers can be derived from equivalent state

space models, obtained through suitable state transformations such as that
introduced in Section 1.7. This will be discussed at length in Section 3.3.
Before we move to this, we shall present in the next section the standard
design methods for exponential observers.

3.2. Extended Luenberger and Kaiman Observers

Assume that the system (3.1) is exponentially observable and let us return to
the observation scheme (3.2) :

^ = Κ φ (ξ ) ­ Ο ξ ­ 0 + F + Ω (ξ )[ξ ι ­ ξ ^] (3.2)
 11 ο ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

As we have already mentioned, the observer design problem is the problem of

selecting the gain matrix Ω (ξ ). Two standard solutions are the extended
Luenberger observer and the extended Kalman observer, respectively, which
are now presented.

3.2.1. The extended Luenberger observer

The design rule for Luenberger observers is to choose the matrix Ω (ξ ) in such
a way that e = 0 is an asymptotically stable equilibrium point (see Definition
A2.1, Appendix 2) of the linear tangent error model (3.5). This is achieved by
choosing Ω (ξ ) such that :

a) the matrix Α (ξ ) ­ Ω (ξ )1 and its time derivative are bounded :

||Α (ξ )­Ω (ξ )ί || <Ci ν ξ

||||:[Α (ξ )­Ω (ξ )υ || < C 2 ν ξ

b) the eigenvalues of Α (ξ ) ­ Ω (ξ )ί have strictly negative real parts :

Re{ λ ί [Α (ξ ) ­ Ω (ξ )υ } < C3 < Ο ν ξ i = 1 to Ν

Under these conditions, it can be shown that e = 0 is a stable equilibrium point

of (3.5) provided C2 is sufficiently small (see Theorem A.2.4, Appendix 2). A
practical example of calculation of Ω (ξ ) will be given in Section 3.2.3.

3.2.2. The extended Kalman observer

In a Kalman observer, the design is based on a "quadratic optimisation"

approach. The problem is to find the gain matrix Ω (ξ ) which minimises the
mean square observation error:
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 111

E= Ι Ι ξ ­ξ Ρ ο Ι τ = f ||e(T)PdT (3.9.a)
Jo

under the constraint of the linear tangent error model (3.5).

The solution is as follows :

Ω (ξ ) = Π (ξ ) J (3.9.b)

where the NxN square symmetric matrix R(|) is generated by the Riccati
equation :

^ = ­ RL\r + RA'^(I) + A(|)R (3.10)

Note that this yields a time­varying gain Ω , even in those cases where A is
independent of ξ .

3.2.3. Example : Basic dynamics of microbial growth

Consider again the simple microbial growth process (1.33) :

Suppose that the reaction rate φ is governed by the "Contois law" (see (1.22)) :

where μ * and Kc are constant kinetic coefficients. Then the state space model
is written :
 112 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Assume that the substrate concentration S is measured on­line.

Hence, we are in the situation where :

N=2 q=1 ξ ^ = [ 8 X] ξ ι =8 L = [ 1 0] (3.13)

The matrix Α (ξ ) = A(S, X) is as follows :

Α (ξ ) = ­ k i ­D 1 0 (3.14.a)
as ax ­ΐ ξ =ξ . 0 1
1

(3.14.b)
Ψ 5 Φ χ ­Dj

Δ 3φ ι μ *Κ ς Χ '
with 9s = I ξ =ξ = (3.15.a)
(KcX+S)^
. Δ 3φ . μ *
φ χ =9χ ξ =ξ = (3.15.b)
(KcX+S)^

The observer (3.2) is written :

^ = ­ k i ­ t i l ^ ­ DS + DSin + ω ι (Χ .8) (S ­ S) (3.16.a)

KcX+S

^ = ­ DX ω 2(Χ ,5) (S ­ S) (3.16.b)

°* KcX+S

The Luenberger and Kaiman observers differ from one another by the way in
which the gain matrix Ω Τ (Χ ,8) = [ω ι (Χ ,8) co2(X,S)] is computed.

In the case of the Luenberger observer one desires to assign to the

eigenvalues of A(X,S) ­ Ω ( χ , 8 ) ί some prespecified negative real values, say
λ ι . λ 2 ( < 0).
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFRCIENTS 113

The gains ω ι (Χ ,8) and co2(X,S) can then be shown to be :

ω ι (Χ ,8) = ­ λ ι ­ λ 2 ­ ki9s + φ χ ­ 2D (3.17.a)

CU2(X,S) = ι ­ ί ­ { ­ λ ι λ 2 ­ (λ ι + λ 2)(0 ­ φ χ ) ­ (D ­ φ χ )^ + Ι ^ι φ χ φ δ ) (3.17.b)

Note that the estimated value S must not be allowed to become zero in order
to avoid division by zero in (3.17.b).

In the case of the Kaiman observer, the gain Ω (χ ,8) is calculated as follows :

R = Ri R3 (3.18.a)
LR3 R2J

= ­ R i ­ 2Ρ ι (Κ ι φ 5 + D) ­ 2Ρ 3ΐ <ι φ χ (3.18.b)
dt

dRo 0
= ­ R | + 2Ρ 3φ 5 + 2Ρ 2(φ χ ­ D) (3.18.c)

dRa
(3.18.d)
= ­ R i R3 ­ Ρ 3(Κ ι φ 5 ­ φ χ + 2D) + Ρ ι φ δ ­ Ρ 2ΐ <ι φ χ
(3.18.e)
ω ι (Χ ,8) = Ρ ι co2(X,S) = R3
Figs.3.1 and 3.2 illustrate the performances of both observers in a simulation
performed under the following conditions :

1)μ * = 0.3 h­1,Kc = 0.2, k i = 2 0 .

2) a square wave influent substrate concentration 8in(t) (from 50 to 40
g/l)
3) a constant value for the dilution rate : D(t) = 0.1 h"""
4) X(0) = 2.49 g/l, 8(0) = 0.25 g/l
114 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Sin (g/i)

50

40 ^
time (hours)

υ 12 24 36 48
0.4
λ ι = λ 2 = ­1
S (g/i) λ ι = λ 2 ­ ­0.5

0.3­f
vs λ ι = λ 2 ­ ­0.1

0.2 4

time (hours)

time (hours)

Fig. 3.1. State observation with an extended Luenberger observer

 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 115

Fig.3.1 Shows the simulation results of the extended Luenberger observer for
three sets of eigenvalues λ ι and λ ζ

a) λ ι = λ 2 = ­1
b) λ ι = λ 2 = ­0.5
c) λ ι = λ 2 = ­0.1

which correspond to decreasing response times.

0.41 S (g/l) : R1(0) = R2(0) = 500

:R1(0).R2(0)»100
:R1(0) = R2(0) = 20

1 ^
0.2
time(hours)
I I •

time (hours)
r

Fig.3.2. State observation with an extended Kalman filter

 116 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

In Fig.3.2, the simulation results of the extended Kaiman filter are presented
under three sets of initial conditions :

a) Ri(0) = R2(0) = 500. R3(0) = 0

b) Ri(0) = R2(0) = 100, R3(0) = 0
c) Ri(0) = R2(0) = 20, R3(0) = 0

which also result in decreasing response times.

The figures clearly illustrate the ability of the observers to guarantee an

arbitrarily fast convergence.

3.3. Asymptotic Observers for State Estimation when the

Reaction Rates are Unknov\^n

In Section 3.2 we presented the design of state observers under two very
restrictive conditions :

­ the model (3.1) is exponentially observable

­ the structure of the model is fully known.

However there are many simple processes of practical interest which are
known not to be exponentially observable (see Example 2 in Section 3.1).
Moreover, the requirement of a full knowledge of the process kinetics is too
severe in most engineering applications, as was shown in Chapter 2. There is
therefore a clear incentive to search for a category of observers which allows
one to asymptotically reconstnjct the missing states even when the process is
not exponentially observable and the kinetics are unknown. In this section we
present a design method of such asymptotic observers under the following
conditions :

­ the reaction rate function φ (ξ ) is unknown

 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 117

­ the yield coefficients (matrix K) are known

­ the number q of measured state variables is equal to or greater than
the rank of the matrix Κ : q = dim (ξ ι ) > ρ = rank (Κ )

3.3­1. Statement of the asymptotic observer

From the basic structural property of Section 1.7, we know that there exists (at
least) one partition (ξ 3, ξ b) of the state, one (N­p)xp matrix AQ, and one vector Ζ
of dimension N­p defined as the following linear combination of the state
variables :

Ζ = Α ο ξ 3+ ξ 6 (1.70) = (3.19)

and whose dynamics, given by ( 1 . 7 3 ) , are independent of the reaction rate

φ (ξ ):

^ = ­ DZ + Ao (F^ ­ Q^) 4­ (Fb ­ Qb) (1.73) = (3.20)

Now, it is clear that we can rewrite this vector Ζ explicitly as a linear

combination of the vectors ξ ι and ξ 2 of measured and non­measured state
variables :

Ζ = Α ι ξ ι +Α 2ξ 2 (3.21)

with appropriate definitions of the (N­p)xq and (N­p)x(N­q) matrices Ai and A2.

Example : The anaerobic digestion process

We consider the simplified reaction scheme ( 1 . 8 6 ) obtained in Section 1.8.4

 118 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Si ­I­ S2 + Pi + P2 (3.22.a)

S2 —0<2 + Pi + P2 (3.22.b)

Suppose, furthermore, that we are not interested in the products Pi and P2 , so

that we can omit them in the analysis (see Section 1.4.2). The reaction
scheme reduces to :

Si — C X i ­I­ S2 (3.23.a)

S2—CX2 (3.23.b)

According to the rules of Section 1.5, the state space model is as follows :

d r
Si ­ki 0 ­DTSi (3.24)
dt
Xi 1 0 ι φ 2] Xi 0
S2 ka ­k2 0
LX2J L 0 1 0
LX2J

The matrix Κ is clearly of rank 2. An admissible state partition is

(3.25)
ξ 3 = "Χ ι ' ξ 6= "sr
.X2. .S2.

The vector Ζ is as follows :

z= " ki 0 ' •Χ ι ' + ' S i ' (3.26)

7-2. ­ka k2 .Xa. .S2.

If we assume that Xi and S2 are measured on­line, then

(3.27)
xr ξ 2 = •sr
.Sa. .X2.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 119

and we can rewrite Ζ :

Ζ = Α ι ξ ι +Α 2ξ 2 (3.28)

with :

Ai = ' k i 0' A2 = Ί 0' (3.29)

_­k3 1_ 0 k2_

Similarly, if we assume that S i , X2, S2 are measured on­line, then :

ξ ι = "sr ξ 2= (3.30.a)
S2
.X2.

Ai = Ί 0 0" A2 = ' k i ' (3.30. b)

0 1 k2 .­k3.

The asymptotic observer

If the (N­p)x(N­q) matrix A2 has a left inverse, the following asymptotic

observer follows quite naturally from (3.20) and (3.21):

Asymptotic observer

^ = ­ DZ + Ao(Fa ­ Qa) + (Fb ­ Qb) (3.31 .a)

ξ 2 = Α |(Ζ ­Α ι ξ ι ) (3.31.b)

where A j is a left inverse of A j . Ζ and ^2 denote on­line estimates of Ζ and ξ 2

respectively.
 120 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

This algorithm is completely independent of φ (ξ ) and can be used without any

knowledge of the kinetics (i.e. of the reaction rates).

The convergence of this algorithm is proved by the following theorem :

Theorem 3.Ϊ : If the dilution rate D(t) is persistently exciting (see Appendix 3),
i.e. if there exist positive constants δ and β ι such that:

0<β ι < D(T)dT Vt (3.32)

then :

lim Ι 1ξ2(1)­ξ2(1)11 = 0 (3.33)

t­>oo

The proof of this theorem is immediate when one observes that, from (3.20),
(3.21) and (3.31), the dynamics of the estimation error are as follows :

^(ξ 2­ξ 2) = ­0(ξ 2­ξ 2) (3.34)

The assumption (3.32) simply implies that the dilution rate D(t) does not
remain equal to zero for excessively long periods of time. Hence, the
convergence of the observer is valid only for fed­batch and continuous
operating conditions. In batch reactors, any initial estimation error will persist
throughout the estimation procedure.

Unlike the Luenberger and Kalman observers of Section 3.1, the speed of
convergence of the estimation is completely determined by the experimental
conditions (through the value of the dilution rate). This is why this algorithm is
called an asymptotic observer. One consequence is that the accuracy of the
initial estimate may be critical when the observer is implemented for a period
of time which is not very long with respect to the residence time (i.e. the
inverse of the dilution rate).
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 121

Example : The anaerobic digestion process (continued)

We return to the process described by the reaction scheme (3.23). We

consider three different situations.

1) Suppose that Xi and S2 are measured on­line and that Si and X2 need to
be estimated.

The vectors ξ β , ξ 6. Ζ , ξ ι , ξ 2 and the matrices A i , A2 have been defined

above ((3.25), (3.26), (3.27), (3.29)).

Clearly the matrix A2 is square and nonsingular. Hence its left inverse is :

­1
Ai = Ai' =
_r
1 0 (3.35)
0 ki^j

The asymptotic observer (3.31) is as follows

d_r^n
= ­D + •FI" (3.36.a)
dt
.0.

'Si' = Ί 0 • 'ki 0' Xi" (3.36.b)

- ­k3 1 .S2J /
.X2. 0 \q\ .Z2.

2) We now assume that S i , X2, S2 are measured on­line.

The on­line estimation of Z i , Z2 is obviously given by the same expression

(3.36.a). The matrix A2 is necessarily full rank, and its (left) pseudo­inverse
is:

Ao = (3.37)
L­k3.
 122 ON­LINE ESTM
I ATO
I N AND ADAPTV
I E CONTROL OF BO
I REACTORS

The on­line estimate of Xi is :

1
[ k i ( Z i ­ S i ) ­ k 3 (Zg­kgXg­Sg)] (3.38)
Xi =

In this case, it is interesting to note that the observer provides an estimate

which is in fact a weighted average between two estimates, which could be
computed separately from Si and Z^ on the one hand, and from S2, X2 and
Z2 on the other.

3) A necessary condition for applying the asymptotic observer is that the

number of measurements be equal to or greater than the rank of Κ : q > p.
But this condition is not always sufficient. Suppose for instance that Si and
X i are measured on­line. Then the matrices Ai and A2 are defined as
follows :

Ai = "1 ki" A2 = "0 0' (3.39)

0 ­kg 1 k2

It is clear that, although the number of measurements is sufficient (q=p=2),

the matrix A2 is singular and cannot be inverted : it is not possible to
estimate ξ 2 = [S2, Χ 2 Γ from ξ ι = [SI,XI]T with an asymptotic observer of the
form (3.31).

3.3.2. A simulation experiment : estimation of substrate and biomass

concentration from synthesis product data

We consider the example 2 of Section 3.1, i.e. a process characterised by the

following reaction scheme :

s—Cx + p (3.40)
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 123

In Section 3.1 we showed that this process is not exponentially observable.

But, as we shall see, it is asymptotically observable.

The dynamics of the process are described by the state space model (3.8).
We choose the following state partition :

(3.41)
.1

ζ and its time derivative are then equal to

z = "Zi" = " kiX+S' (3.42.a)

.Z2. ­kgX+P,

dZi
­ ^ = ­ D Z i + DSjn (3.42.b)

dZg
= ­DZ, (3.42.C)
"dT

Since rank(K) is obviously equal to 1, the measurement of only one

component is needed to estimate the other two. Assume that Ρ is measured
on­line. ξ ι , ξ 2, A i , A2 are then defined as follows :

(3.43.a)

Α ι =Γ θ 1 A,= 1 ki" (3.43.b)

1 0 ­k2

The asymptotic observer (3.31) for the estimation of the substrate

concentration S and the biomass concentration X is written :
 124 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

= ­ D Z i + DSi, (3.44.a)
dt

dZg
"dT = ­ D Z , (3.44.b)

(3.44.C)

X = ^(Z2­P) (3.44.d)

The behaviour of the observer is illustrated by a simulation experiment

{Fig.3.3).

The kinetics of the "tme" process are assumed to be governed by the following
growth rate model (see Appendix 1)

Kp
μ (1) = μ * (3.45)
KM + S ­ I ­ S ^2/1^.
/KI Kp­HP V PLJ

with the following set of parameter values :

μ * = 0.23 h­1, KM = 0.26 g/l, K| = 297 g/l, Kp = 7.77 g/l, PL = 85.81 g/l.

The simulation of Fig.3.3 was performed under the following conditions :

a) X(0) = 0.365 g / l , S(0) = 90 g / l , P(0) = 5.1 g/l

D(0) = 0.1 h­l,Sin(0) = 100g/l

b) two different sets of initial conditions for Z^ and Zg:

Zi(0) = ­ 1 and 1, Z2(0) = 80 and 110

c) a trapezoidal signal for D(t) (from 0.1 to 0.05 h­^)

 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 125

0.06
time (hours)
>• Γ ' Γ • 1
0 40 80 120 160 200
•S(g/I)
:Zi(0) = ­1,Z2(0) = 80
- \ :Zi(0) = 1, Z2(0) = 110
\
90 : true value of the variables
/ ""
/
/
/
/
80 ' /
/
/
" /
/ time (hours)
70
0 40 80 120 160 200
1.4 • X (g/i)
\

0.7
^

• /
./
1 time (hours)
­0.7.
0 40 80 120 160 200

Fig.3.3.a. State observation with the asymptotic observer: simulation result

126 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

13^ Ρ (g/l)

9 4

' 1 "
r 1 r ' 1
0 40 80 120 160 200
1 ­ Zi(g/i)
\
\
\
\
\
\

/
/ time (hours)
­1 /'
Γ ·
40 80 120 160 200
110 2.Z (g/l)

100­ ^

/
/
1
/
90­ /
/
/
/
/
time(hours)
80 —Γ ­ 1 —
r— r • 1 · 1 200
(3 40 80 120 160

Fig.3.3.b. State observation witli the asymptotic observer: simulation result

 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 127

Comments

1 ) It is worth noting that the asymptotic observer (3.44) is extremely simple

compared with the Luenberger and Kaiman observers. Compare for
instance (3.44) with (3.16) to (3.18).

2) Besides its simplicity, the asymptotic observer does not require any
knowledge of the reaction rate structure (here the particular and quite
complex structure (3.45) has been used to simulate the process, but clearly
this structure is ignored by the observer) and, nevertheless, has a most
acceptable behaviour. The drawback, however, is that the speed of
convergence is fixed by the experimental conditions and cannot be
manipulated by the user. In the simulation, convergence is achieved after
about 40 hours.

3) Last but not least we know (see example 2, Section 3.1) that, in this
example, the process is nof exponentially observable, and therefore cannot
be observed at an arbitrary exponential rate by using e.g. extended
Luenberger or Kaiman observers. However, as illustrated by Fig.3.3, the
state is efficiently reconstructed with the simple asymptotic observer (3.44).

3.3.3. Real life experiment 1 : An anaerobic digestion process

The asymptotic observer (3.31) has been applied to a pilot­scale anaerobic

digestion process for the on­line estimation of organic substrate concentration
and of acidogenic and methanigenic bacterial concentrations. The results of
this experiment (performed by the Unit of Bioengineering, Universite
Catholique de Louvain, Belgium) are now reported.

The process is assumed to be described by the simplified reaction scheme

(1.86):
 128 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

•Pi
S i — ­ f ­ S 2 + P 2 (3.45.a)

<P2
S 2 — C X 2 + P 1 + P2 (3.45.b)

with X i being the acidogenic biomass, S i the organic substrate, X2 the

methanogenic biomass, S2 the acetate, Pi the methane and P2 the inorganic
carbon.

The dynamics of the process are described by the model (1.85) which is
rewritten as follows :

4rx^
dt 1
^1 0• ­•Γ χ Ο ο ' (3.46)
Si ­ki0 Si Fi ο
X2 0 1 Χ 2 Ο ο
S2 S2
Ο ο
0 kg ο Qi
Pi Ρ ι
>4 k5.
ο LQ2.
LPs J P2J

We choose the partition ξ 3 = [ Χ ι , X2]''", ξ b = [ S i , S2, P i . P2]''' so that the

auxiliary state Ζ is as follows :

= "sr ­ "­ki 0• (3.47)

Z2 S2 kg ­k2

Z3 Pi 0 ke
.Z4. .P2. k4 ks _

We are interested in the problem of estimating the biomass concentrations X i

and X2 and the substrate concentration Si from on­line measurements of S2
(acetate concentration), Fi = DSijn (substrate feedrate) and Q i (methane gas
flow rate). We know also that the dissolved methane concentration and its time
derivative may be neglected (see Section 1.8.4) : Pi = dPi/dt = 0.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 129

We immediately note that, for our purpose, the fourth auxiliary state Z4 is
superfluous and hence that the yield coefficients k4 and ks are not required.

The asymptotic observer (3.31) is as follows :

dZ
= ­DZi+Fi (3.48.a)
dt

d i g
= ­ D Z o (3.48.b)
"dT

dZg
= ­DZ3­Qi (3.48.C)
dt

Z3
(3.48.d)
<6

Λ k2
S2 ­ Z2 ­ ­ j ^ Z3 (3.48.e)

Si=Zi­^ S2­Z2­^Z (3.48.f)

Experimental results

This asymptotic observer has been implemented over 90 days on a 60 liter

pilot bioreactor with the yield coefficients :

ki = 3.2 k2 = 28.3 k3 = 5.7 k6 = 27.3 (3.49)

and the following three sets of initial conditions :

130 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

4 $2 (gCOD/i)

0 .
2­1
QcH4

0.08 Η

0.04 J

time (days)

Fig.3.4. State observation in a continuous anaerobic digestion process

on­line data
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 131

Si (gCOD/l)
164

:Z,(0) = 11,Z2(0) = 3.9, ZgiO) = 0

8 Λ : Z,(0) = 1 1 , Z2(0) = 4, ZjiO) = ­6

: Ζ ,ί Ο ) = 11, 2.2(0) = 6.65, ZgiO) = ­13.65

40 80 120

Xi (gCOD/l)

• X2(gC0D/l)

0.5

time (days)

0 40 80 120

Fig.3.5. State observation in a continuous anaerobic digestion process

estimation results
 132 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

1)Zi(0) = 11,Z2(0) = 3.9,Z3(0) = 0 (3.50.a)

2)Zi(0) = 11,Z2(0) = 4.Z3(0) = ­ 6 (3.50.b)
3) Zi (0) = 11, Z2(0) = 6.65, Z^(0) = ­13.65 (3.50.C)

The influent substrate concentration S i j n . the dilution rate D and the

measurements of S2 (acetate) and Qi (methane gas) are shown in Fig.3.4.

The on­line estimates of X i , X2, Si are shown in Fig.3.5.

An of the observer can be performed as follows. A

measurement of the biomass (i.e Xi + X2) in the reactor is also available :
it is obtained from the difference between total COD and solubilised COD
measurements. We can thus compare the estimated total biomass Xi + X2
with its actual value : this is done in Fig.3.6 where a very good agreement is
found (after the initial transient). We see also that convergence is achieved
within 70 days.

Xi + X2 (gCOD/l)
8J
^ : off­line analyses

4J

time (days)
—I—
40 80 120

Fig. 3.6. State observation in a continuous anaerobic digestion process

experimental validation
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 133

3.3.4. Real life experiment 2 : A PHB production process

The dynamical model (1.66) of the intracellular production of PHB was

described in Section 1.6. In this section an experimental validation (performed
by the SOLVAY Company of Belgium) of the asymptotic observer (3.31) on a
fed­batch PHB production pilot plant is reported.

The choice of usefui measurements

In this pilot plant, both oxygen and carbon dioxide are available for on­line
measurement. However, as we shall show, only one of these two components
can be used for the implementation of the observer because the respiratory
quotient (RQ) is equal to one (RQ = 1). The RQ is defined as follows :

moles COp formed

RQ= ^
moles O2 consumed

This can be written in terms of our dynamical model as follows :

RQ = (3.51)
'k49i+k692

where γ is a conversion factor from grams to moles.

However, since RQ is equal to 1 in both reactions, as can be checked

experimentally, we necessarily have :

k7 = 7k4 (3.52.a)

k8 = 7k6 (3.52.b)

Suppose now that we try to estimate the state variables of the process from the
measurements of oxygen (C) and carbon dioxide (P2).
 134 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Since the matrix Κ is presently of rank 2, we can define ξ β and ξ b as follows :

ξ 6 = (3.53)
LPIJ S2
c
LP2.

Then, using (3.52), Ζ is equal to :

Z= (ki­k3k5)X­l­k5Pi­)­Si (3.54)
"I" S2
(k4­k3k6)X­hk6Pi+C
.­/[(k4­k3k6)X­Hk6Pi]­HP2.

A priori, we could believe that the state of the process could be reconstructed
using the asymptotic observer from the measurement of any components.
However, this is not true if the estimation is based on oxygen and carbon
dioxide data. Indeed ξ ι , %z, ^^ and A2 would be defined as follows :

C ξ 2 = X (3.55.a)
P2. Si
S2
Pi,

Ai = Ό 0" A2 = ki ­ k3k5 1 0 ks • (3.55.b)

0 0 k2 0 1 0
0 1 k4­k3k6 0 0 ke
.0 1.
­7(k4­k3k6) 0 0 ­Tkg

A2 is a square matrix but not full rank (lines 3 and 4 are proportional) and
therefore not invertible : we cannot estimate the state of the PHB process from
the measurements of O2 and C O 2 . This is simply the mathematical
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 135

confirmation of the fact that, in view of the fact that RQ = 1, the measurements
of O2 and CO2 do not give independent information on the state of the
process. This highlights the fact that the measurements used for the state
observation need to be carefully chosen.

The observer equations

In order to reconstruct the state of this PHB process, it is necessary either to

have another component available for on­line measurement, or to look for
additional prior knowledge on the process dynamics. As the first possibility
was excluded from an operational point of view, we opted for the second one.

From the description of the process given in Section 1.4.3, it can be seen that
the two biological paths of the PHB production are mutually exclusive. The
first path (microbial growth (with a small growth­associated production) with
rate φ ι ) requires fructose and nitrogen as substrates. The second path
(enzyme catalysed production with rate 92) requires fructose but is completely
inhibited by nitrogen. The results, clearly, is that only one path is activated at
each time.

In practice, the process was conducted in two successive steps :

1) During the first step the process is fed with the two substrates (fnjctose and
ammonia) : this is a growth step without enzyme­catalysed production.

2) During the second step the process is fed with fructose only : this is a
production step without growth.

This allows us to consider two different dynamical models to describe each

step. Both are special cases of the general dynamical model (1.49).
 136 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

MODEL A : Growth without production (in presence of nitrogen)

ξ ^= [X Si S2 Pi 0 P2]

Q"^ = [0 0 0 0 0 Q I ]

F^ = [0 DSijn DSjjn 0 Gin 0] (3.56)

k'^ = [1 ­ki ­k2 kg ­ k 4 kj]

G(ξ ) = S1S2CX

MODEL Β : Production without growth (absence of nitrogen)

ξ •^=[XS1 P i C P2]

Q''' = [0 0 0 0 Q I ]

F^=[ODSijnOQjn 0] (3.57)

K ' ^ = [ 0 ­ k 5 l ­kg kg]

G(ξ ) = S1CX

In both models, the rank of matrix Κ is equal to 1, and the measurement of one
component can be used for the state observation. We decided to use the on­
line measurements of oxygen in preference to those of CO2 because the
oxygen sensors were in practice more reliable and less sensitive to pH
variations.

Hence ξ ι and ξ ζ are defined as follows :

ξ ι =0

m o d e l A . · ξ | = [ X S i S2P1 P2]

modelB : ξ ^ = [ Χ Si Pi P2]
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 137

For both models, we choose the partition = ξ ι and = ξ 2· Then :

1 ­k­( k3
­iT
Aq ­ ­ (model A ) (3.58.a)
k4 k4 k4 k4

(model B) (3.58.b)
ke kg J

A2 = l (3.58.0)

In this study, we were interested in reconstructing the time evolution of the

biomass concentration X, the PHB concentration Pi and the nitrogen
concentration Safrom the on­line measurement of C (dissolved oxygen), D
(dilution rate), S2,in (influent nitrogen concentration), and Qjn (oxygen feed
rate).

The following auxiliary variables are defined :

Model A : Zi = X + ^ (3.59.a)
K4

kgC
Z2=S2­ (3.59.b)

kaC
Z3 = P 1 ­ Κ (3.59.C)

Model Β : Z4 = X

(3.59.d)

The asymptotic observer (3.31) is as follows :

 138 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

A. Whenever δ ο > 0 :

dt

dip kgQin
(3.60.b)

^ = _ D Z 3 ­ H ^ (3.60.C)
dt ^ k4

X= ~ (3.60.d)
k4

§2 = Ζ 2 ­ Η ^ (3.60.e)
K4

P l = Z 3 ­ ^ (3.60.f)

B. When§p = 0

^ = ­DZ4 (3.61.8)
dt

= _DZ, + ­ ^ (3.61 .b)

X = Z4 (3.61.c)

Ρ ι =Ζ 5­·^ (3.61.d)

Note that, since the nitrogen concentration is not accessible from on­line
measurement, we use the estimate S2 to switch from the first algorithm (3.60)
to the second one (3.61).
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 139

Fig.3.7. State observation in a PHB producing process : on­line data

140 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

time (hours)
. G3 tPl

Fig.3.8. State observation in a PHB producing process : estimation results

 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 141

Implementation on a fed-batch bioreactor

We present the estimation results on one experiment carried out by the

SOLVAY Company (Belgium) on a 20 liter pilot bioreactor. The numerical
values of the yield coefficients are as follows :

k2 = 0.83, kg =0.657, k4 = k6 = 0.769

The on line data are shown in Fig.3.7. (Note that Qin = Qo2.in ­ Qo2,out)­ The
estimates of X, S2 and Ρ are shown in Fig.3.8. The agreement between these
estimates and off­line data from chemical analysis is evident. In addition, it
can be observed that the switching from algorithm A to algorithm Β is very
efficiently driven by the estimate of the nitrogen substrate S2.

3.3.5. Practical aspects of implementation

Numerical Implementation

Practical computer implementation of the asymptotic observer (3.31) requires

that it be rewritten in a discrete­time form. This can be done simply by
replacing the time derivative of Ζ by a finite difference (using a first order
forward Euler approximation) :

where Τ is the sampling period and t and t+1 are time indices.

The asymptotic observer (3.31) is then written as follows :

Discrete-time asymptotic observer

Zt^i = Zt ­ TDtZt + TFbt ­ TQbt + AoT(Fat ­ Qat) (3.63)

 142 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Choice of the sampiing period Τ

The switch from the continuous­time equation (3.31) to the discrete­time

version (3.61) induces the presence of the sampling period T. The continuous­
time equation (3.31) is unconditionally stable. For the discrete­time equation
(3.63), the value of Τ plays an important role in the stability. In fact, if the
dilution rate is bounded as follows :

0<D(t)<Dmax Vt

then equation (3.63) will remain stable as long as the value of T/2 is smaller
than •"''ax:

I < ­ J ­ (3.64)

This is easily shown by considering the following positive definite, decrescent

candidate Lyapunov function :

Wt = (Zt ­ z/(Z^ ­ i t ) > 0 (3.65)

the time difference of which is equal to :

­ Wt = {TDt(TDt ­ 2)}(Zt ­ z / ( Z t ­ Z^) (3.66)

which is nonpositive definite, decrescent as long as inequality (3.64) holds.

iVIatrix inversion and conditioning

The implementation of the asymptotic observer (3.31) requires the inversion of

yield coefficient matrices in the computation of AQ and A2 . It is well known
that, if the matrices to be inverted are ill­conditioned, the algorithm is extremely
sensitive to any numerical error and its performance will be dramatically
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 143

degraded. The conditioning of a matrix is evaluated by the condition number.

Ill­conditioning is characteristic of nearly singular matrices, i.e. matrices with
one (or more) row (or columns) which are an almost exact linear combination
of some of the other rows (or columns).

This is typical of reaction networks in which one reaction is almost

indistingishable from some of the others, or in which some components
appear almost proportionally. The extreme situation is exact proportionality,
which has been illustrated in the PHB example (Section 3.3.4).

As an illustration of matrix conditioning and condition number, let us consider

the following reaction network :

S + C—Cx
S+C+X >X + P

The matrix Κ and the component vector ξ are defined as follows :

K = ­ki ­k2 ξ = S"

­ka ­k4 c
1 0 X
. 0 1 .p.

Let us choose a state partition (ξ a,ξ b)^ As has already been pointed out, this
choice is arbitrary. It is then usually possible, in the absence of any constraint,
to select a state partition for which the matrix Ka is (almost) diagonal with Ί " on
the diagonal and for which inversion is easy. Here, if we choose = [X P]^
and ξ b = [S C]T, then

Ka = l

and it follows that AQ is simply equal to ­Kb

Ao = ~KbK^^=­Kb
 144 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Assume now that the objective is to estimate X and Ρ from the measurements
of S and C. The implementation of the asymptotic observer (3.31) requires the
inversion of A2 :

A2 =
L­k3 ­k4.

The computation of A2 is possible only if A2 is of full rank, i.e. if its columns (or
its rows) are linearly independent. Assume that A2 is symmetric (k2 = ka) and
that the yield coefficients k i , k2, ka and k4 are such that they can be expressed
as follows :

(k^, k3) = a(k2, k4 + b)

with a GiR and b a small positive real number. A2 will be invertible as long as
b is different from zero. However its condition number is highly dependent on
b. In fact, due to the symmetry of A2, the condition number is simply the ratio of
the maximum and minimum norm of the eigenvalues of A 2 :

max λ (Α ο )|
CN(A2) =
ί η |λ (Α 2)Ι
mm

The condition number is an expression of the propagation of numerical errors

(arising from roundoff or experimental data for instance) from ^­Α ­ι ξ ^ to ^2­
can be shown that, in our example, a relative error on ^­Α ^ξ ^,
||δ (Ζ ­Α ι ξ ι )||/||2­Α ι ξ ι ||, will give a relative error on ξ 2, Ι |δ ξ 2ΐ |/||ξ 2ΐ 1, which is
related to the preceding one by the condition number CN(A2) as follows :

Ι Ι Ζ ­Α ι ξ ι

In our example, the condition number of A2 is approximately equal to

 Chap.3. ESTIMATION WITO KNOWN YIELD COEFFICIENTS 145

The performances of the estimation algorithm will fall off as b becomes

smaller.

If, for instance, ki = k2 = ka = 1 and k4 = 1.01 (i.e. a = 1 and b = 0.01), the

condition number CN(A2) is equal to :

CN(A2) = 200

and any numerical error in ^­Α ^ξ ^ might propagate in with a factor equal to
200!

Further illustration of degraded performances of the estimation algorithm

(3.31) in the presence of ill­conditioned matrices can be found in the
references.

In order to avoid the degradation of the estimation performances in the

presence of ill­conditioned matrices, it is clear that the estimation algorithm
must be modified. In general, one possible solution consists in rearranging
the reaction network. This may eventually lead to the suppression of some
reactions which are almost redundant and can hardly be distinguished from
the others. Or, as in the PHB process, the rearrangement can take advantage
of additional a priori information and the practical operation of the process. As
shown in the different examples of Section 3.3, there usually exists a state
partition for which Ka is (almost) diagonal with Ί " on the diagonal. We may
therefore expect that the eigenvalues will be equal, or close, to 1, and the
condition number for the inversion of Ka will then be small. However, if the ill­
conditioning affects the matrix A2, a solution may be to look for other on­line
measurements if available, so as to improve the conditioning of the inversion
of A2.
 146 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

3.4. On­line Estimation of Reaction Rates

So far in this chapter, we have dealt with the state estimation problem in
bioreactors. In Section 3.3, this problem was considered in the case where
the reaction rates φ (ξ ) are unknown. In this section, we address the
complementary question, namely the problem of estimating the reaction rates
from on­line knowledge of the state variables (either from measurements or by
means of asymptotic observers).

3.4.1. Statement of the estimation problem

We consider a biotechnological process described by the general state space

model (1.43) :

^ = Κ φ (ξ ) ­ Ο ξ ­ Q + F (1.43) = (3.67)

We assume that:

a) The yield coefficients (matrix K) are known.

b) The dilution rate D, the feed rates F and the gaseous outflow rates Q are
measured on­line.
c) The vector of state variables ξ is known on­line either by measurement or by
estimation via the asymptotic observer of Section 3.3.

We assume that the vector φ (ξ ) of reaction rates is partially unknown and

written as follows :

φ (ξ ) = Η (ξ )ρ (ξ ) (2.13) = (3.68)

where Η (ξ ) is an Mxr matrix of l<nown functions of the state and ρ (ξ ) a vector of

unknown functions of ξ , with dim ρ (ξ ) = r.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 147

Several examples of this representation of φ (ξ ) have been given in Chapter 2.

With this definition, the process dynamics (3.67) are rewritten :

^ = Κ Η (ξ )ρ (ξ ) ­ Ο ξ ­ Q + F (2.24) = (3.69)

We are concerned with the problem of estimating ρ (ξ ) from the on­line

knowledge of D, F, Q and ξ .

Two different solutions to this problem are presented. The first one (Section
3.4.2) is called an observer-based estimator because it is based on the use of
a variant of the state observers described in Section 3.1. The second one
(Section 3.4.3) is based on a reformulation of the process model in a linear
regression form.

3.4.2. An observer­based estimator

The basic idea is to use a state observer, not to estimate the state (since it is
known) but to provide on­line information for the updating of the estimate of
Ρ (ξ ).

The estimation algorithm is as follows :

Observer-based estimator

^ = Κ Η (ξ )ρ (ΐ ) ­ Ο ξ ­ Q + F ­ Ω (ξ ­ξ ) (3.70.a)

^ = [Κ Η (ξ )]ν (ξ ­ξ ) (3.70.b)

where ρ denotes the on­line estimate of ρ (ξ ).

 148 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

An intuitive justification for this algorithm is as follows. Equation (3.70.a) is

clearly similar to the general state observer (3.2) but with three decisive
modifications:

a) The square matrix Ω may be dependent on ξ but must be stable for all ξ (t).

b) The actual value of the state ξ is used in the right hand side of the equation
(terms Η (ξ ) and Ο ξ ).

c) The actual value of ρ (ξ ), which is unknown, is replaced by an estimate ρ

which is updated by the second equation (3.70.b).

Γ is the gain matrix of the updating law (3.70.b); it must chosen such that the
matrix Ω ^Γ + Γ Ω is negative definite.

This updating law (3.70.b) is in turn driven by the deviation (ξ ­ ξ ) which is

supposed to reflect the mismatch between p(t) and ρ (ξ ).

Tuning

The matrices Ω and Γ are design parameters at the disposal of the user for the
control of the stability and the tracking properties of the algorithm. A common
choice is to take :

Ω = diag {-ω ^} Γ = diag {η } ω ί ,η e iR^ (3.71)

Μ .....Ν j=1 r

With this choice, the condition on Ω "^Γ + Γ Ω is automatically verified. In such a

case the tuning of the estimator reduces to the calibration (by trial and error) of
the Ν + r scalar constants ω ί and η . This issue will be further discussed and
illustrated with the examples of Sections 3.4.5. and 3.4.7.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 149

Analysis

Define the observation error e and the tracking error ρ as follows

β = ξ ­ξ ρ = ρ (ξ )­ρ (3.72)

Their dynamics are easily shown to be governed by the following linear time­
varying system (obtained by substracting (3.70.a) from (3.67)):

= A(ξ )Γ e + V (3.73.a)
LpJ ­PJ

with

dp
Α (ξ )=Γ Ω Κ Η (ξ )' ν = (3.73.b)
dt
.­Η '^(ξ )Κ '^Γ Ο

Suppose that the assumptions of Section 1.9.1 hold, that Ω is constant with all
its eigenvalues having strictly real parts, and that ρ (ξ ) is a differentiable
function of ξ . Then it is a standard result of adaptive system theory that the
error system (3.73) is stable if the matrix Κ Η (ξ ) is persistently exciting, which is
easily established from Theorem A2.6 (Appendix 2) and Theorem A3.2
(Appendix 3).

Numerical implementation

The numerical implementation of the estimator requires a discrete­time

formulation. A forward Euler discretization of the continuous­time algorithm
(3.70) gives :
 150 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Discrete-time observer-based estimator

= It + Τ [KH(ypt ­ ϋ ξ ,­ Qt + Ft ­ Ω (ξ ^4)] (3.74.a)

= Pt + Τ [ K H ( y ] ^ Γ (ξ t­ξ t) (3.74.b)

where Τ denotes the sampling period.

As for the numerical implementation of asymptotic observers (Section 3.3), the

switch from the continuous­time observer­based estimator (3.70) to the
discrete­time version sets specific stability questions in which the sampling
period T, the design parameters Ω and Γ and the regressor Κ Η {ξ χ ) play an
important role. This will be illustrated at the end of Section 3.4.5 with a simple
example.

The observer­based estimator (3.70) and its discrete­time counterpart (3.74)

are based on the full dynamical model of the process. In practice, however,
this is not always necessary. It is often sufficient to select a subset of the state
equations, provided they involve all the parameters ρ (ξ ) which need to be
estimated.

In such a case, the entire theory presented in this chapter holds "mutatis
mutandis". For instance, if we denote by ξ 8 the selected part of the state and
by (Ks, Qs. Fs) the corresponding parts of (K,Q,F), the observer­based
estimator (3.70) is rewritten :
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 151

^ = Κ 5Η (ξ )ρ α ) ­ Ο ξ δ ­ Qs + Fs ­ Ω ( ξ 5 " I s ) (3.75.a)

^ = [Κ δ Η (ξ Γ Π ξ ς ­ Ι δ ) (3.75.b)

This will be illustrated many times throughout the book, namely in the
experimental applications of Sections 3.4.8 and 3.4.9.

3.4.3. A linear regression estimator

We present an alternative estimator of ρ (ξ ) based on a model reformulation in

terms of linear regression.

Model reformulation

The solution of the differential equation (3.69) which describes the process
dynamics can be shown to be written as follows :

ξ = ψ '^ρ + ψ ο + ε

ψ , and ε are the outputs of the following linear filters :

^ = Ω ψ ^ + Κ Η (ξ ) (3.76.a)
dt

dvo
= Ω ψ ο ­ (Ω + •ΐ Ν )ξ ­ Q + F (3.76.b)
dt

^-^Wf (3.77,

with Ω an arbitrary stable symmetric matrix.

 152 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Introduce the notation :

Υ = ξ ­ψ ο (3.78)

It is then obvious that the process dynamics (3.69) are described by the
following linear regression model :

y = \\f^p + (3.79)

where y is the output, ψ is the regressor, ρ is the unknown parameter to be

estimated and ε is interpreted as an unknown additive disturbance.

Note that the output y and the regressor ψ can be calculated on­line from the
available data (ξ , D, Q, F) with the aid of the filtering equations (3.76). The
important point is that the linear regression model (3.79) allows us to use a
standard least squares algorithm for the estimation of ρ (ξ ).

The principle of least squares estimation is the computation of the on­line

estimate p(t) which minimises the following quadratic criterion :

J(p) = eA(t­) ||(y­YTp)||2dx (3.80)

where the parameter λ is called the forgetting factor.

The algorithm for the on­line computation of p, involving the computation of ψ

and ψ ο via (3.76), is as follows :
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 153

Least squares estimator

^'=Ω ψ ^+Κ Η (ξ ) (3.81 .a)

dt

^ = Ω ψ ο ­(Ω + Ο ΐ Ν )ξ ­0 + Ρ (3.81 .b)

^ = Γ ψ (ξ ­ψ ο ­ψ ^ρ ) (3.81 .c)

^ = ­Γ ψ ψ '^Γ + λ Γ Γ (0)>0 (3.81 .d)

Tuning

The matrix Ω and the forgetting factor λ are design parameters at the disposal
of the user for the control of the stability and the tracking properties of the
estimator.

The matrix Ω is commonly chosen as follows :

Ω = diag {-ω ^ ω | e iR^ (3.82)

i=1 Ν

Hence, as for the observer­based estimator, the tuning of the estimator

reduces to the calibration (by trial and error) of the η + 1 scalar constants ω ί
and λ .

We define the tracking error ρ

 154 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Ρ = ρ (ξ ) ­ Ρ (3.83)

Then, from (3.77) and (3.81 .c), the dynamics of the estimator are easily shown
to be governed by the following linear time­varying system :

d_
ε = Α (ξ )Γ ε +v (3.84.a)
dt
LPJ ­PJ

with :

Α (ξ ) = Ω Ο V= ( V lf ^ (3.84. b)
.­Γ ψ ­Γ ψ ψ ^.

Suppose that the assumptions of Section 1.9.1 hold, that Ω is constant and
stable and that ρ (ξ ) is a differentiable function of ξ . It is a standard result of
adaptive systems theory that the system (3.84) is stable if the matrix ψ is
persistently exciting (from Theorem A2.6 (Appendix 2) and Theorem A3.1
(Appendix 3)).

Numerical implementation

The numerical implementation of the estimator (3.81) requires a discrete­time

formulation. Euler discretization of the filters (3.81 .a­b) which provide ψ and
Ψ ο gives :

Ψ ^=(ΐ Ν + Ω Τ )ψ [ + Τ Κ Η (ξ ί ) (3.85.a)

Ψ ο .Η ΐ = (IN + Ω Τ ) ψ ο ,ί + Τ [ ( Ω ­ DIN) ξ ι ­ Qt + F^] (3.85.b)

Moreover, the discrete­time least squares algorithm for the linear regression
model (3.79) is written :
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 155

Pt+1 = Pt + TTt \\fl^ ­ ^OM ­ ψ [+1 Ptl (3.85.C)

Ft Γ

Π ^ι = Γ [ I ­ TVt^i ί λ Ι + TV^i^Vt^, rVliFt, (3.85.d)

An interesting special case occurs when the matrix Ω is chosen as :

Ω = ­(Τ )­^Ι Ν (3.86)

Indeed, in that case, the discrete­time algorithm reduces to :

PM = Pt + τ η K H ( y {ξ ^ ­ ξ ι ­ τ [ K H ( y pt ­ ­ Qt ­Η (3.87.a)

I ­ Τ ^Η %)κ \λ \ + KH(ξ t)Γ t] (3.87.b)

This algorithm (3.87) is simply the recursive least squares algorithm obtained
by applying a linear regression technique to the first order Euler discretized
counterpart of the model equation (3.67).

As for the observer­based estimator, the influence of the sampling period T, of

the design parameters Π and λ , and of the regressor Κ Η (ξ ι ) on the stability of
the discrete­time least square estimator (3.87) will be illustrated in Section
3.4.5 with a simple example.

3.4.4. An example of convergence analysis

The stability and convergence properties of the estimators presented in the

previous sections are closely related to so­called "persistence of excitation"
 156 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

conditions (see Appendix 3). In this section we illustrate, with a particular

example, how the full analysis can be carried out.

We consider the situation in which the reaction rates φ ](ξ ) are described by the
multilinear model motivated in Section 1.5.2:

(1.47.a) = (3.88)

Here ξ η denotes the reactants involved in the reaction with index j and aj is the
"specific reaction rate".

With this definition, the state space model of the process is written (equation
(1.49)):

(1.49) = (3.89.a)
dt = Κ Ο ( ξ ) α ­ Ο ξ + Ρ ­α

with 0(ξ ) being a diagonal matrix (equation (1.48.b)):

«?>>f..'^a{n5n (1.48.b) = (3.89.b)

Several examples of this way of modelling were given in Section 1.6.

The reactor is supposed to be operating under assumptions A I to A4 (Section

1.9.1) which guarantee the global input­output stability of the process, and in
particular that the state variables have upper bounds (see Corollary 1.2) :

max
ξ i(t)<r. Vi,Vt (3.90)

In addition, it is assumed that:

 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 157

A5. The experimental conditions are such that the state variables are bounded
away from zero :

0<ξ ,|η <ξ ,ω Vi.Vt (3.91)

A6. The specific reaction rates aj 0=1 M) are constant parameters.

A7. The matrix Κ of yield coefficients is full rank.

We address the problem of the on­line estimation of parameters aj from the on­
line [knowledge of D, F, Q, ξ . We note that this estimation problem is stated in
the format of Section 3.4.1 with the definitions :

Η (ξ ) = 3(ξ ) ρ (ξ ) = α (3.92)

We suppose that the linear regression estimator (3.81) is used for the
estimation of a. In this particular case, the estimator is written as follows :

dv^ τ
^ ­ ω ψ ^ + Κ Ο (ξ ) (3.93.a)
dt

dvo
dt = ­ ω ψ ο +(ω ­0)ξ ­α +Ρ (3.93.b)

^ = Γ ψ (ξ ­ψ ο ­ψ '^ά ) (3.93.C)

— = ­Γ ψ ψ '^Γ Γ (0)>0 (3.93.d)

Note that we have chosen a scalar parametrization of the filters (parameter ω )

and a zero forgetting factor (λ = 0) because the aj are constant.

The convergence of this algorithm is established in the following theorem.

 158 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Theorem 3.2: Under Assumptions A1 to A7,

lim a(t) = a (3.94)

t­>oo

Proof: The proof proceeds in 2 steps.

Step 1 : We establish that ψ is persistently exciting (Appendix 3).

It follows from Theorem 1.1 and Assumption A5 that the diagonal matrix 0(ξ ) is
positive definite for all t. This, in turn, implies from Assumption A7 that the
matrix 0 ( ξ ) Κ Τ Κ 0 ( ξ ) is also positive definite, and hence that Κ 6(ξ ) is
persistently exciting (see Definition A3.1, Appendix 3). Since the filter (3.93.a)
is state reachable, then ψ is persistently exciting (Theorem A.3.4, Appendix 3).

Step2\\Ne define :

ε = ξ ­ψ ^ά and ά =α ­ά (3.95)

It is easily shown that these quantities are governed by the following

dynamics:

' ε " = • ­ω Ι 0 • "ε " (3.96)

dt
_­Γ ψ ­Γ ψ ψ ^. .0C.

The theorem follows immediately from step 1 (Theorem A3.1, Appendix 3).
Q.E.D

3.4.5. Case study : Estimation of microbial specific growth rates

In Sections 3.3.1 to 3.3.3, we presented a general theory for the estimation of

reaction rates. In this section, the theory is extensively studied through a
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 159

simple application : the on­line estimation of the microbial growth rate in a

simple biological culture which involves a single biomass growing on a single
substrate and yielding a single product. This case study is carried out for two
main reasons. First, the issue of tracking microbial growth rates, even in
simple processes, is clearly a relevant problem by itself in many
biotechnological applications. Secondly, the case study will serve to show the
power of the theory and to illustrate its main features and variants, while
keeping the mathematical derivations as simple as possible.

Process definition

We consider a process characterised by the following reaction scheme :

(3.97)

where S is the substrate, X the biomass and Ρ a gasifiable synthesis product.

The process dynamics are as follows :

1 0 (3.98)
dt
­ki
Lk2j 0

where Fi is the substrate feed rate and Q is the gaseous outflow rate.

According to the usual approach in fermentation modelling, the reaction rate is

assumed to be described as follows (see Section 1.5.3):

φ = μ (X,S,P)X (3.99)

where μ (X,S,P) is termed "specific growth rate"

We note that φ is in the format of Section 3.4.1 with :

 160 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Η (ξ )^Χ and ρ (ξ ) = μ (Χ .8,Ρ ) (3.100)

The problem is the on­line estimation of the time­varying specific growth rate μ .

Using

In this particular case, the observer­based parameter estimator (3.70)

specialises as follows :

^ = μ Χ ­ DX + ω ι (Χ ­Χ ) (3.101 .a)

^ = ­Κ ^μ Χ ­ DS + F + ω 2(8­8) (3.101 .b)

^ kzliX ­ DP ­ Q + (03(P­P) (3.101 .c)

^ = Yi X (X­X) ­ Y2 ki X (S­S) + 73 ksX (P­P) (3.101 .d)

We remark that the design matrices have been chosen diagonal :

Ω = diagi­coj} Γ = diagiyj} (3.102)

The implementation of the estimator (3.101) requires full state measurement

(X,S,P) as well as measurement of F and Q. Various alternatives, however,
are possible in the case of incomplete measurements.

We know from Section 3.2 that only one state measurement is actually needed
to design an asymptotic observer able to reconstruct the process state.
 Chap.3. ESTIMATION Wim KNOWN YIELD COEFFICIENTS 161

independently of the specific growth rate μ . In the case of incomplete state

measurement, the solution is obviously to use this asymptotic observer to
compute on­line estimates to replace the missing states in the algorithm
(3.101). This is illustrated by an example.

Suppose the product concentration Ρ is the only measurable state variable.

Then, applying the theory of Section 3.2, we define the auxiliary variables :

1 k
Zi=X­i^P Z2 = S + ­r^P (3.103)

for which we have the following asymptotic observer:

dZi 1
­ ­ ^ = ­ D Z i + 7 ^ Q (3.104.a)
dt k
<2o

dZo
—^ = ­DZ2+ F (3.104.b)
dt ^

Xo = Z^ + ^P (3.104.C)
2

. ki
So = Z 2 ­ 7 ^ P (3.104.d)
KO

The "observed" states XQ and So provided by this observer are then used, in
the algorithm (3.101), in place of the "true but unknown" states X and S.

If the substrate feed rate F or the gaseous outflow rate Q (or both) are not
accessible on­line, it is nevertheless possible to derive an efficient estimator of
 162 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

μ provided the biomass X is measured on­line. The solution is simply to drop,

from the estimator, equation (3.101.b) which contains F or (and) equation
(3.101.c) which contains Q, and to set 72 or (and) 73 to zero in the updating
equation (3.102.d). In such a case, the observer­based estimator reduces to :

^ = μ Χ ­ DX + ω (Χ ­Χ ) (3.105.a)

^ = 7Χ (Χ ­Χ ) (3.105.b)

Stability analysis

The structural simplicity of the reduced order estimator (3.105) allows one to
perform a full stability analysis with a moderate mathematical effort. The
analysis will be carried out under the following assumptions :

HI. The biomass data, denoted Xm(t), are corrupted by an additive

measurement noise e(t):

Xm(t) = X(t) + e(t) (3.106)

H2. Xm(t) is strictly positive for all t :

XJt)>η >0 (3.107)

H3. The specific growth rate μ (t) is time­varying and bounded :

0 < μ (t) = μ (Χ ,8,Ρ ) < μ * Vt (3.108)

H4. The time derivative of the specific growth rate is bounded :

|­^|<Mi Vt (3.109)
dt
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 163

H5. The measurement noise is bounded :

|e(t)|<M2 Vt (3.110)

H6. The dilution rate is bounded :

0<D(t)<D^3, Vt (3.111)

H7. The substrate feed rate is expressed as (see Section 1.5.5):

Fi = DS|n (3.112)

where Sin is the influent substrate concentration, assumed to be

bounded for all t :

0<SJt)<S,3x (3.113)

H8. The design parameter ω is chosen to be proportional to the biomass

measurement :

ω =σ Χ ^ (3.114)

Under these assumptions the parameter estimator (3.105) is rewritten as

follows :

^ = [μ ­ D + oiX^ ­ X)] X^ (3.115.a)

^=γ Χ ,(Χ ,­Χ ) (3.115.b)

If we define the errors X = X ­ X and μ = μ ­ μ , the following error system

derives from (3.115) :
 164 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

(3.116)

with

x= "x A = "­σ 1"

.μ . .­γ 0.

v = (D­μ ­σ XJε

dt

Let υ ι and V2 be the eigenvalues of A related (by definition) to σ and γ as

follows :

σ = ­ ( υ ι ­^υ 2) γ =υ ι υ 2 (3.117)

Assume that:

H9. The design parameters σ and γ are chosen such that A has real distinct
eigenvalues : υ 2 < υ ι <0 (i.e γ < σ ^Μ ).

We have the following stability result.

Theorem 3.3. : Under assumptions HI to H9, the estimation errors X and μ

are bounded for all t and asymptotically bounded as follows :

Ml ,,2 SmaxM2 2β ι δ Μ 2
lim sup |x(t)| < ­ — + Μ 2 + —Γ + (3.118)
ν σ ^­4γ

,2 Smax M2 ^1
lim suplii(t)| < : J ­ | ­ j M i ­ H p i M 2 + a (3.119)
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 165

with :

Δ
••max (3.120.a)
β ι = max Dmax. μ * + σ

8 i
{^1 (3.120.b)
Ul

Proof : From Assumptions H5 and H7, the measured biomass concentration

Xm has an upper bound, as follows from Corollary 1.3 :

X , ( t ) < ^ + M2 Vt (3.121)

Moreover, since Xm(s) > η > 0 for all s (Assumption H2), the following change
of time scale :

dt = X J s ) d s (3.122)

leads to rewriting the system (3.116) as follows :

dx
= Ax + (3.123)
dt Xm

From Assumptions H3, H4, H5, H6, H7 and (3.121), the input v/Xm of this
system is bounded. Hence, since A is a stable matrix, the state χ is also
bounded (see Theorem A2.6, Appendix 2) and the first part of the theorem
(boundedness of X and μ ) follows.

The second part of the theorem follows from the technical Lemma A2.2
(Appendix 2) with the following definitions of Bi and B2 :

Δ Mp ,
Bi=—'η 3χ |0^3Χ ,μ * + σ { (3.124.a)
k
 166 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

iS
max (3.124.b)

Q.E.D.

In the case of noise­free measurement of the biomass (i.e. M2 = 0), the

asymptotic upper bounds of X and μ are simplified as follows :

lim|x(t)l<­^ (3.125)

ΙΐΓ π Ιμ ω Ι (3.126)
t­»oo Π Τ

An important consequence of these expressions is that the estimation errors

X(t) and ii(t) can be made asymptotically arbitrarily small, by choosing a
sufficiently large design parameter γ . This means that, in the noise free case, a
perfect tracking accuracy can be expected although the parameter μ (t) is time­
varying. In practice, however, large values of γ will make the estimation
sensitive to measurement noise. The choice of γ is thus a compromise
between asymptotic accuracy and sensitivity to noise. The issue is discussed
in the next paragraph.

The characteristic polynomial of the matrix A (see (3.116)) is written :

s^ + 2Cws + w^ (3.127)

with the natural frequency w = and the damping coefficient ζ = σ /2ν γ .

The choice of the design parameters γ and σ is thus equivalent to the choice of
ζ and w. A common engineering rule of thumb for the design of second order
systems is to fix the damping coefficient ζ close to 1. Once ζ is fixed, optimal
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 167

values of σ and γ can be defined, both from the point of view of asymptotic
accuracy as well as transient behaviour.

Asymptotic accuracy

The optimal value of σ is defined as the value which minimises the asymptotic
upper bound (3.119) on the tracking error μ :

^opt = arg mjn lim s u p l j i l (3.128)

t­>oo

This is easily shown to be :

ν = 2ζ ^/. (3.129)
'Vl2^max

The optimal value of γ follows from the definition of ζ :

Yopt = ­ ^ (3.130)

Transient behaviour

The ingredient for the selection of optimal design parameters is the mean
square tracking error over a given period of time :

J(a) = ­ V lii(T)l^dT (3.131)

The period of time (t2­ti) may, for instance, be the duration of a standard fed­
batch operation.
 168 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

The optimal value of σ is defined as the value which minimises the criterion
J(a). The optimal value of γ follows as in the previous case (3.130). The
computation of these optimal values is illustrated in the next paragraph.

The reduced order estimator (3.115) is simulated under the following

conditions.

1 °) Figure 3.9 :

a) The specific growth rate to be estimated is assumed to be a Monod

growth rate (1.18) :

^ KM + S

with μ * = 0.33 h""" and KM = 5 g/l.

b) The influent substrate concentration is constant (Sin = 5 g/l) but the

dilution rate D(t) is a square wave signal (Fig.3.9).

c) The yield coefficient ki= 2.

d) The process state and the estimation algorithm are initialised as

follows :

X(0) = 2.054 g/l S(0) = 0.893 g/l μ (0) = 0.06 h"^

 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 169

­ D U )

ο
d

in
ο
6
t ( h )
1 I Γ >
• 1 • 1

Γ "»• •••••• • ­ 1 1
6 4 8 12 16 2 0

ii .
iii :
i

\ \
­i\

. 1\

in
.\\
ο
d
t (h)

—1 1 1 ^
12 16 2 0

Fig.3.9. On­line estimation of μ from noise­free data of X

 170 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

e) Three different sets of design parameters are considered :

(i) σ = 5, γ = 6
(ii) σ = 1, γ =0.24

(iii) σ = 0.5, γ = 0.06

which correspond to decreasing response times.

2°) Figure 3.10:

a) The specific growth rate to be estimated is assumed to be a Haldane

growth rate (1.20.a)

_ μ * 8 _ _
KM + S + S^/K,

with μ * = 0.33 h"^. KM = 5 g/l, K, = 25 g/l

b) The dilution rate is constant, D = 0.05 h­i, but the influent substrate
concentration, Sin(t), is a square wave signal (Fig.3.10.a).

c) The yield coefficient ki = 2.

d) The biomass data are corrupted by a measurement noise with a

standard deviation of 0.04 (i.e. about 5% of the biomass mean
value).

e) The process state and the estimation algorithm are initialised as

follows :

X(0) = 2.05 g/l S(0) = 0.9 g/l μ (0) = 0.06 h"^

f) The design parameters are set to the following values :

σ = 1 , γ =0.24
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 171

SIN<T)

ID

T (H)

Fig.3.10. On­line estimation of μ from noisy data of X

 172 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

The behaviour of the process and of the alogrithm are shown in Figs.3.9 and
3.10. We note the excellent performance of the estimator, which is indeed
able to track the specific growth rate, even in the presence of noisy
measurements (Fig.3.10), without any knowledge of the Monod or Haldane
structures.

The optimal choice of design parameters is illustrated in Fig.3.11. The

criterion (3.131) is drawn with respect to σ for different values of ζ . The figure
clearly shows the existence of an optimal value of σ for each ζ . In particular,
for ζ = 1, the optimal parameters are :

σ = 2.25 γ =1.27

Design parameter σ

Fig. 3.11. Optimal values of σ for different values of ζ

 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 173

Numerical implementatiori

Let us now briefly analyse the stability of the discrete­time versions of the two
parameter estimators (observer­based (3.74) and least squares (3.87)
algorithms) proposed in this section. For the continuous­time version, the
choice of the design parameters is quite free : for instance, with diagonal gain
matrices (see (3.71) and (3.82)), the estimators will be stable for any positive
value of ω ί and η . However, as for the asymptotic observer (see Section 3.3.5),
the stability of the discrete­time version of the estimators depends on
additional constraints, in which the sampling period Τ also plays an important
role.

It should be pointed out that, basically, the values of the design parameters
must be bounded proportionally to Τ ""· and |Κ Η (ξ )|­'' in order to keep the
estimators stable.

We illustrate this with the on­line estimation of the specific growth rate from
biomass measurements.

1) Let us begin with the observer­based estimator. The discrete­time version of

(3.115) is written as follows :

Xt^i=Xt + TA,Xt + TaXt(XrXt)

M,^i=iit + TYXt(Xt­Xt)

Let us write the parameter γ as in Section 3.4.5 :

a'
4ζ ^

To simplify the analysis, we shall consider that ζ = 1. The estimator equations

are then equal to :
 174 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Xt+1 = AJ Xt
At+iJ LAiJ

with :

At = 1 ­ oTXt TXt b, = TcXr

σ
4

The i/A7/brceoi system

+1 = AJX,

L^t+iJ L^tJ

is uniformly asymptotically stable if

0<η <Χ ί <Χ ^3χ

2
0 <σ <
Τ Χ ,max

Indeed, At is bounded and the time difference of the following positive definite,
decrescent candidate Lyapunov function W t :

Wt = G^xf ­ 2aXtiit + 2{if = (cXt ­ iixf+ {if > 0

for the unforced system is equal to :

Wt,i ­ Wt = ­ oTXt (2 ­ aTXt(iit ­ ψ ­ ^ ( 6 ­ σ Τ Χ ,) X,^

It is clearly decrescent and negative definite under the above assumptions.

Hence, since bt is bounded via boundedness of Xt and σ , the estimator is
stable.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 175

2) Let US now consider the linear regression estimator. The discrete­time least
squares algorithm is written from (3.87) as follows :

μ ι ^ι = μ ι + TYtXtiX^i ­ Xt ­ T^^Xt + TD^Xtl

Yt
Yt+1 =
λ +Τ ^Χ ^γ ^

It is easy to verify that and y\ will remain bounded under Assumptions H2

and H6 (Section 3.4.5), if λ and γ ο are chosen as follows :

max 0 , 1 ­ 2 ­ ^ | < λ < 1 Yo^^^4

Xmax J λ Τ η

Therefore, the values of the design parameters σ and γ ο must remain even
smaller as the sampling period Τ and the bounds on the regressor X, i.e. η and
Xmax, are large.

3.4.6. Extended Kaiman filtering for on­line estimation of reaction rates

In Section 3.1 we showed how extended Kaiman observers can be designed

for on­line state estimation when the model structure (and hence the kinetic
structure) is fully known. The same technique can be used to solve the dual
problem of estimating the reaction rates from full state measurement. This has
given rise to numerous studies and applications (see the bibliography).

The basic idea is to consider the unknown parameter ρ as an additional

unknown state of the process. The state space model (3.69) is therefore
augmented as follows :

= "Κ Η (ξ )ρ ­Ο ξ + Ρ ­ Ο " (3.132)

dt "ξ "
­p. . w(t)

where w(t) is a completely unknown time­varying function.

 176 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

A STATE observer, analogous to (3.2), is written for that system :

dt - "KH(|)P­D|­I­F­Q' ­1­ Ω ι (ξ ,ρ ) (3.133)

0
J L"2(I.P)J

We note that w(t) is replaced by zero in the observer equation since it is an a

priori unknown quantity.

The extended Kalman filtering (EKF) technique is then used to compute the
gain matrices Ω ι and Ω 2 on­line.

The linearised tangent approximation of (3.133) is written (compare with

(3.5.a)):

= "Α (ξ ,ρ )­Ω ι (ξ ,ρ ) Κ Η (ξ )' (3.134)

dt " ξ ­ ξ "
­Ω 2(ξ ,ρ ) 0

9[Η (ξ )ρ ]
with Α (ξ ,ρ ) = Κ ­DIK
9ξ ­ΐξ =ξ

The gain matrices Ω ι and Ω 2 are computed in order to minimise the following
quadratic criterion :

^dpj
J(t) = (τ ) (f(x)).IU(x)­l(x)P}dt (3.135)
dt

under the constraint of the linearised tangent model (3.134)

This criterion is a weighted sum of two terms (weighting matrix Σ ""·) : the first
term penalises rapid variations of the parameter estimate ρ while the second
term penalises excessively large deviations between the actual state value ξ
and its estimate ξ .
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 177

The solution of this optimisation is as follows. The time­varying symmetric

matrix R(t):

Δ ^
R(t) = Rii(t) Ri2(t) (3.136)
_Rl2(t) R22(t).

is updated via the following Riccati equation :

^ = ­ REqR + RAjd. p) + Aod. p)R + Zq (3.137.a)

where

Δ ^ Δ ^
En = IN 0 0 0 (3.137.b)
0 0 0 Σ

and Α ο (ξ , ρ ) is the coefficient matrix of the system (3.134).

The gains Ω ι and Ω 2 are given by :

Ω 1(ξ ,p) = R11(t) (3.138.a)

" 2 ( i p ) = Ri2(t) (3.138. b)

The EKF estimator thus consists of equations (3.133), (3.137) and (3.138). It
appears to have a similar structure to that of the observer­based estimator
(3.70). However the computational complexity is increased by the need to
solve the Riccati equation (3.137) on­line.

An additional drawback of the EKF approach is that its stability and

convergence properties are extremely difficult to analyse and, as far as we
know, are still an open problem in the case (which is our concern) of
parameter estimation in nonlinear systems. It is well known, however, that an
EKF estimator may give biased estimates or may even diverge if it is not
 178 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

carefully initialised (Ljung, 1979). Therefore, although the EKF estimator may
often yields interesting results, there is no a priori guarantee as to its stability.
This issue is illustrated with a simple example.

Example : EKF rates

Our aim is to compare the EKF estimator with the reduced order observer­
based estimator (3.105) for the on­line estimation of microbial specific growth
rates in simple cultures. The process is assumed to be described by the
equation :

^ = μ Χ ­Ο Χ (3.139)

Applying the EKF technique to this model yields the following algorithm :

^ = μ Χ ­ DX + r, 1 (X­X) (3.140.a)

^=ri2(X­X) (3.140.b)

^ = 2(μ ­0)Γ ι ι +2Χ Γ ΐ 2­Γ ^ι (3.140.C)

dri2
" = ( i l ­ D ) Γ Ι 2 + Χ Γ 22­Γ Ι Ι Γ Ι 2 (3.140.d)
dt

­^=­ή ^ +α (3.140.e)

where σ denotes the weighting factor of the associated quadratic criterion

(3.135) and r n , r^2· ^22 are the entries of the matrix R :
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 179

R = •"n ''12 (3.141)

r•^2 r22j

It immediately appears that this EKF algorithm is far more complex than the
reduced order estimator (3.105).

μ (h­1)

15 — · — : Extended Kalman filtering

: Observer­based estimator

or
time (hours)
­15
8
X (g/l)

J
time (hours)
8

Fig.3.12. Comparison between the EKF and the observer­based

estimator

A simulation experiment using this algorithm has been performed with the
same process model as in Fig.3.10 and is shown in Fig.3.12. We examine the
case in which the process is in a steady state. This implies that μ is constant
 180 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

and hence dμ /dt = 0. In addition, we assume noise­free measurements of X.

Therefore Mi = M2 = 0 in expression (3.118) and (3.119), and we know from
Theorem 3.2 that the estimate μ {\) will converge to its true value with the
reduced order estimator (3.105).

Both the reduced order estimator and the EKF estimator are initialised with :

il(0)=Oh­^ X(0) = Og/l

The design parameters of the reduced order estimator are σ = 1, γ = 0.24.

The design parameters of the EKF estimator are :

σ = 0.01 rii(0) = 1 ri2(0) = ­ 99 r22(0) = 10000

(with these values, the initial matrix R(0) is positive definite).

The estimation results are shown in Fig. 3.12. With the EKF estimator, the
estimate μ (Χ ) converges to a biased value equal to 22 h­i (i.e. a relative error of
4400 %), while with the simple reduced order estimator ji(t) converges, as
expected, to its true constant value (μ = 0.05 h­"").

3.4.7. Experimental application 1 : Estimation of the specific growth rate

in a fed­batch ethanolic fermentation process

In this section, we present an experimental application, on a laboratory pilot

scale process, of the parameter estimator (3.101) which has been proposed
for the on line estimation of the specific growth rate μ . The process under
consideration is a yeast growth process assumed to be characterised by the
reaction scheme (1.37). However the reactor is not aerated, so that only the
second metabolic path (1.37.b) of the reaction is activated. Hence, the
reaction scheme reduces to the following single reaction :
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 181

S ­ ^ X +P

where S, X and Ρ denote, respectively, the glucose, yeast and ethanol

concentrations (we ignore, here, the CO2 production).

We notice that this scheme is identical to (3.97). The dynamics of the process
are then given by equations (3.98). We suppose that ethanol is the only
component available for on­line measurement. Moreover, the feed rate F and
the dilution rate D are known while the gaseous ethanol outflow rate Q is
negligible. Therefore we can directly apply a discrete­time version of the
estimation scheme (3.101) (3.103) (3.104) for the specific growth rate μ , i.e.:

Zit^,=Zit­TDtZit (3.142.a)

Z2t+i = Z2t ­ TDt Z2t + TFt (3.142.b)

Xot + i = Z i t + i + T^Pt+i (3.142.C)

k2

Sot^i=Z2t^i­^Pt^i (3.142.d)

Xt+1 = Xt + Tilt Xot ­ TDtXot + ω ι Τ (XQI ­ Xt) (3.142.e)

SM = St ­ Tkiilt Xot ­ TDtSot + TFt + WgT (Sot ­ St) (3.142.f)

Pt^i = Pt + Tkgiit Xot ­ TDtPt + ω ^Τ (Ρ χ ­ Pt) (3.142.g)

At^i = At + TYiXot(Xot ­ Xt) ­ TY2kiXot(Sot ­ St) TY3k2Xot(Pt ­ Pt)

(3.142.h)

The estimation algorithm (3.142) has been implemented on a fed­batch

ethanolic bioreactor over a period of 44 hours (Fig.3.13) at the Unit of
Bioengineering (Universite Catholique de Louvain, Belgium). The reactor
operated under the following conditions : the initial volume of culture was
equal to 35 liters; the reactor was fed twice with a 5­minute pulse of glucose
(F = 600 mMol/h) and the volume was increased by 8 liters at each step to
 182 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

finally reach 51 liters. Fig.3.13.a shows the evolution of the ethanol

concentration Ρ with time during the experiment.

The values of the yield coefficients are equal to :

I rN mMol , ^ ^ mMol ,
•^^^^­^^W ^2 = 0 . 3 — (3.143)

(NTU : Normal Turbidimetry Units)

The result of the estimation of the specific growth rate μ with the algorithm
(3.142) is shown in Fig.3.13.b. It was obtained under the following conditions :

μ ο = 0.15 h""* Zio = 13 NTU Z20 = 54.7 mMol (3.144.a)

of

Qi=aiXot ω 2 = σ 2Κ ι Χ ο ί 003 = 031^2X01 Υ ι = — ( i =1,2,3) (3.144.b)

σ ι =0.018 02 = 0.086 0 3 = 0.086 (3.144.C)

Note that the form of ω ι , ω 2, ω β is in line with the argument presented in
Section 3.4.5 and Assumption H8. An experimental validation of this result
has been carried out as follows. Off­line data on biomass X (via a
turbidimeter) and on glucose S were available. The validation procedure is
based on a comparison between estimates of X and S obtained via the
estimated specific growth rate μ and the true measured values of X and S.

Recall first that the dynamical equations governing X and S are :

dS
= ­k1μ X + F ­ D S (3.145.a)
dt

^ = μ Χ ­ϋ Χ (3.145.b)
dt
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 183

'Ρ (g/l)

X (NTU)

400 Η
• : off­line analyses
200 φ

time (hours)

10 "20" "30" 40

Fig.3.13. Estimation of μ in a fed­batch ethanolic fermentation

process
 184 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

We can compute on­line estimates of X and S by introducing the estimated

specific growth rate μ , given by (3.142), into these equations, i.e. :

=­kiiiXo+F­DSv (3.146.a)
dt

dXy
"dT = i i X v ­ D X v (3.146.b)

where Xv and Sv denote the validation estimates of X and S. Discretization of

(3.146) gives :

Sv.t+1 = Sv,t ­ TkiiitXot + TFt ­ TDtSv.t (3.147.a)

Xv.t+1 = Xv,t + TiitXv,t ­ TFt ­ TDtXv,t (3.147.b)

Note that the use of XQ in (3.146.a) makes both estimations (3.146.a) and
(3.146.b) independent of one another.

The validation results are shown in Fig.3.13.c­d. They exhibit a very good
agreement between the off­line data of S and X and their estimates provided
through the estimated specific growth rate μ .

Calibration of the design parameters

In Section 3.3 we have used, when available, off­line data to validate the state
observation performed by the asymptotic observer. Here again, as explained
above and as will be true in the following applications, a similar experimental
validation of the parameter estimation has been carried out. However, the
validation results shown in Fig.3.13 are somewhat unfair. Actually, we used
the experimental validation as a tool for choosing appropriate values for the
design parameters. These are calibrated so as to obtain the reaction rate
estimates which give rise to the best estimates of the unmeasured
 Chap.3. ESTIMATION WITO KNOWN YIELD COEFFICIENTS 185

— — : asinFig.3.13

ω ι =ω 2=7ι =Υ 2=0·

cu^.OOSekgXot.

^3 =0.000018

S(g/1)
50 •·
\
25

\
\
­1

X (NTU)

600

400H

200

time (hours)
0 ^ ^ ^ ^ —T­
10 lo" 30 40

Fig 3.14. calibration of the design parameters :

comparison between two sets of design parameters
 186 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

components for which there exist some off­line data. Fig.3.13 shows the
estimation of the specific growth rate with a good set of design parameters, i.e.
those which correctly validate the estimation of the substrate and biomass
concentrations. For comparison purposes. Fig.3.14 shows what happens
when the design parameters are not well calibrated. The estimation of the
specific growth rate is performed under the same initial conditions as in
Fig.3.13 but with a different set of design parameters : ω ι , CU2, γ ι and γ 2 are set
to zero and ω 3 and γ β are divided by a factor of 10 and 100, respectively.

ω ι = ω 2 = γ ι =γ 2 = 0, ω 3 = σ 3Κ 2Χ ο (. Y3 = ­4­, σ 3 = 0.0086

The values of ω β and ys are clearly too low to make the estimator capable of
tracking the rapid variations of the specific growth rate, μ does not vary quickly
enough from its initial value of 0.15 h­i, and the experimental validation gives
negative values for the substrate concentration S and excessively large values
for the biomass concentration.

3.4.8. Experimental application 2 : Estimation of the reaction rates in a

continuous anaerobic digestion process

In this section, we present an experimental application of the linear regression

estimator proposed in Section 3.4.3, on the anaerobic digestion application
already described in Section 3.3.3. The reaction scheme and the dynamics
are as follows :

5^ L^Xl+S2 + l

92

S2 — C X 2 + Pl + P2
 Ch^.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 187

d
dt 1 0 φ ι ­ D f Xi 0' ­ " 0'
Si ­ki 0 Φ 2ϋ Si Fi 0
0 1 X2 0 0
S2 ks ~^2 S2 0 0
Pi 0 kfi Pi
0 Qi
.P2. >4 ksj P2
0. Q2.

with Xi being the acidogenic biomass, Si the organic substrate, X2 the

methanogenic biomass, S2 the acetate, Pi the methane and P2 the inorganic
carbon.

We recall that:

1) the only components which are available from on­line measurements are
the acetate S2 and the methane P i ;
2) the biomass concentrations X i and X2 and the organic substrate
concentrations Si are given on­line by the asymptotic observer (3.48);
3) there are two unknown reaction rates : the acidogenesis reaction rate φ ι ,
and the methanization reaction rate 92, which can be expressed (see
Section 1.5.3) by one of the following expressions :

9i = μ ι Χ ι = a i X i S i
Δ Δ
Φ 2 ~ V^2^2 ~ ^2^2^2

We decide to design an algorithm for the estimation o f :

ρ (ξ ) = Γ μ Γ
.«2.

This choice of Η (ξ ) and ρ (ξ ) illustrates the flexibility of our approach to the

design of parameter estimators, since it will allow us jointly to estimate two
 188 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

parameters of different dimensions : a specific growth rate μ ι (time­i) and a

specific reaction rate a2 (concentration­^ χ time­'').

As we have mentioned earlier (see the remark at the end of Section 3.4.2), the
algorithm does not necessarily need to be based on the full dynamical model
of the process. It may be sufficient to select a subset of the state equations,
provided they involve all the parameters we want to estimate. Here we have
selected the following two equations :

d_r^ ­I Γ
S2 ­ D T S
dt
Pi. [0 kej 0 SaXaJL^a. LPIJ LQI

but, obviously, many other choices are possible.

The least squares discrete­time algorithm (3.87) is then written as follows for
this model (in which the methane P i is assumed, as before, to be a low
solubility product) :

"Aif+r = "Alt' kaXit ­ ^2821X21 Yt

.«2t. ^6821X21

with :

S2t+1 ­ ^21 ­ T kaXit ­ k2S2t^2t ­ D f Sat 0

0 k3S2t^2t «2tJ 0 LQit.

and an appropriate equation (3.87.b) for the updating of the matrix Γ χ .

The estimates of ^ i , and ^2t are provided by the asymptotic observer (3.48).

The algorithm (3.152) has been implemented on the same set of data as in
Section 3.3.3 (Fig. 3.4) under the following conditions :
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 189

Γ ι ο = 0.01 λ ι =0.9 μ ι ο = 0.06 d"^ (3.153.a)

Γ 2ο = 0.01 λ 2 = 0.9 020 = 0.01 l/g.d (3.153.b)

Xio = 1.1g/l X2o = 0.22g/l (3.153.0)

The results are presented in Fig.3.15.a­b.

An experimental validation of these results has been carried out as follows by

using, as in Section 3.3, the off­line data of total biomass concentration Xj
(=Xi + X 2 ) . The time derivative of Xj is equal to :

dXj dXi DXO

= μ ι Xi + a2S2X2­ D(Xi + X2) (3.154.b)

μ ^X^+ DXj (3.154.c)

It is then natural to compute an on­line estimate of X T from the estimates of μ ι

and a2, as follows :

^ = IIIXI + a2S2X2 ­ DXt (3.155)

or, in discrete time :

XT.T+1 = XT.T + TiiitXit+ Ta2tS2tX2t ­ TDX­r^t (3.156)

The successful comparison between the actual measured total biomass X T

and its estimate X j calculated with the aid of the parameter estimates

μ ^ and 0.2 is shown in Fig.3.15.c.

190 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

0.08H

0.04J

0.04H

O.O2J

Xi + X2 ( g c o D / i )

Si
Φ : off­line analyses

Λ 4

time (days)

40 80 ­I 1—
120

Fig.3.15. Estimation of the reaction rates

in a continuous anaerobic digestion process
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 191

3.4.9. Experimental application 3 : Estimation of reaction rates in a batch

process of lactic acid production

In this example, we illustrate the on­line estimation of the reaction rates φ \ on a

lactic acid fermentation process (Laboratoire de Genie des Precedes
Biotechnologiques et Agro­alimentaires, INRA, Grignon, France)(see Section
1.4 (1.40)). The process under study in this example is a pure culture of
in which the yield of glucose consumption for
biomass growth appears to be negligible. The process can then be described
by the reaction schemes (1.40.a) and (1.40.e) :

Φ Ι
S > Pi + Pg (1.40.a) = (3.157.a)

Φ 2

Pi 4­Xp >X2 + P5 (1.40.e) = (3.157.b)

where S, P i , P2, X2 and P5 are, respectively, the lactose, glucose, galactose,

biomass and lactic acid.

The reactions take place in a 15 liter batch bioreactor maintained at a

temperature of 44°C. pH is regulated at a value of 5.8 by addition of NaOH.
From stoichiometric arguments, we know that the yield coefficients k i 2 , k22
and k i 3 are equal to :

k i 2 = 0.5, k 2 2 = 1, k i 3 = 0.5

Taking into account that, here, F = Q = D = 0, the process dynamics are

described by the following model :

_d
'S " = '­1 0 " >r
dt
­ k 2 2 .Φ 2.

P2. ki3 0
.P5. . 0 1
 192 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Assume that we wish to estimate the reaction rates φ ι and 92 from the glucose
Pi and lactic acid P5 concentrations. Let us apply for instance the observer­
based estimator. We consider a reduced order estimator based on the
dynamical equations in Pi and P5 (i.e. with ^ = [ P i , P5] in (3.75)). If we
choose diagonal gain matrices Ω and Γ , the estimator is written in discrete time
(see (3.74)):

Pi.t+i = Pi.t + Tki29i.t ­ Tk2292.t + ω ι Τ (Ρ ι _, ­ Ρ ,,,) (3.138.a)

P5.t+1 = P5.t + Τ φ 2,, + ω 2Τ (Ρ 5.ι ­ Ps.t) (3.138.b)

9i.t+i = 9i,t + YiTki2(Pi,t­ Pi.t) (3.138.C)

φ 2.1+1 = Φ 2,1 ­ YlTk22(Pl.t ­ Pl.t) + Y2T(P5.t " Ps.t) (3.138.d)

Ω = "­ω ι 0' Γ = "Υ ι 0 '

0 ­0)2 0 Υ 2.

In this application, glucose is measured on­line with a biosensor. Besides,

NaOH essentially neutralises the lactic acid produced : as shown in Fig.3.16,
NaOH addition and lactic acid concentration are proportional (with a
proportionality coefficient of 6.5). Since NaOH is representative of lactic acid
and easier to measure on­line, we used NaOH data (instead of lactic acid
data) in the implementation of the estimation algorithm (3.158).

NaOH (g/l)

130.

65.

time (hours)

'^"^ 2 '4 6

Fig.3.16. Lactic acid fermentation : NaOH and lactic acid data

 Ct\ap3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 193

1$I{G/I.H)
7i

time (hours)
6
—1 1 ·
2
6 φ 2 (g/i.h)

4^

time (hours)

Fig.3.17. Estimation of the reaction rates in a batch lactic fermentation

process : on­line data and estimation results
 194 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Estimation results of φ ι and 92 are shown in Fig.3.17. They were obtained

under the following conditions :

ω ι =2 C02=1 Yi=2 Y2 = 0.25 T = 1/3h

9io = 6l/g.h 92o = 6l/g.h Pio = Pio P5o=P5o

Fig.3.18. Estimation of the reaction rates in a batch lactic

fermentation process: experimental validation

As in Sections 3.4.7 and 3.4.8, the experimental validation of the reaction rate
estimation was carried out on the remaining components, i.e. lactose S and
galactose P2, i.e. :
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 195

^2,t+i = ^2,χ -^^^3Ψ ^,x^

where §o and Pg.o were set to :

So = So ^2,0=^2.0

The experimental validation results are shown in Fig.3.18. Note here again
the good fit between experimental data and estimated values of S and P2.

3.5. References and Bibliography

Sections 3.1 and 3.2

Kwakernaak H. & R. Sivan (1972). John

Wiley, New York.

Williamson D. (1977). Observation of bilinear systems with application to

biological control. 13, 243­254.

:
 196 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Aborhey S. and D. Williamson (1978). State and parameter estimation of

microbial growth processes. Automatica, 14, 493­498.

Section 3.3

A simplified presentation of asymptotic observers for biotechinological

processes is given in :

Bastin G. (1988). State estimation and adaptive control of multilinear

compartmental systems : theoretical framework and application to
biotechnological processes, in New Trends in Nonlinear Systems Theory,
Lecture Notes on Control and Information Science, n°122. Springer Veriag,
341­352.

Ttie experimental application of anaerobic digestion of Section 3.3.3 is

developed in :

Renard P., V. VanBreusegem, G. Bastin, H. Naveau and E.J. Nyns (1990).

Active biomass estimation in a biomethanation process. Submitted for
publication.
Renard P. (1990). Verification Experimentale de TApplication de la Theorie de
I'Estimation et du Controle Adaptatif a des Systemes Biologiques Non
Lineaires. Doctoral Dissertation, Univ. Cath. Louvain, Belgium.

Ttie experimental application of Section 3.3.4 is developed in :

Dochain D., E. De Buy I and G. Bastin (1988). Experimental validation of a

methodology for on­line state estimation in bioreactors. In N.IVI. Fish, R.I.
Fox & N.F. Thornhill (Eds.), Computer Applications in Fermentation
Technology : Modelling and Control of Biotechnological Processes,
Elsevier, pp187­194.
 Chap.3. ESTIMATION WITH KNOWN YIELD COEFFICIENTS 197

Effects of ill'Conditioned matrices on tfie performances of asymptotic observers

are presented in :

Pomerleau Y. and M. Perrier (1989). Estimation of multiple specific growth

rates in bioprocesses. AlCfiE J., in press.
Pomerleau Y. (1990). Modelisation et controle d'un prodece fed­batch de
culture des levures a pain., Ph. D. Thesis, Ecole Polytechnique de
Montreal, Canada.

Section 3.4

On­line estimation of reaction rates is treated in :

Bastin G. and D. Dochain (1986). On­line estimation of microbial specific

growth rates. Automatica, 22, n°6, 705­709.

Applications can be found in ttie above paper and in the following ones :

Flaus J.M., A. Cheruy , M.N. Pons & J.M. Engasser (1989). Estimation de
I'etat physiologique des microorganismes dans un bioprocede. Rairo, AP
11,23, 211­220.
Flaus J.M., M.N. Pons, A. Cheruy and J.M. Engasser (1989). Adaptive
algorithm for estimation and control of fed­batch bioprocesses. In N.M. Fish,
R.I. Fox and N.F. Thornhill (Eds.) Computer Applications in Fermentation
Technology : hnodelling and Control of Biotechnological Processes, 349­
353.
Pomerleau Y. and M. Perrier (1989). Nonlinear estimation and adaptive
control of a fed­batch fermentor. In N.M. Fish, R.I. Fox and N.F. Thornhill
(Eds)., Computer Applications in Fermentation Technology. Modelling and
Control of Biotechnological Processes, 361­365.
D. Atroune, A. Cheruy, J.P. Flandrois, G. Carret (1988). Choice of the specific
growth rate formulation in biotechnological processes. IMACS 12^^ World
Congress on scientific Computation.
 198 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Flaus J.M., A. Cheruy, M.N. Pons and J.M. Engasser (1987). Multivariable
adaptative estimation applied to Baker's yeast fermentation. Proc. 4^^
1987, Elsevier, Amsterdam, 3, 139­
142.

of of is

Pomerleau Y. and M. Perrier (1988). Non­linear estimation of specific growth

rates in baker's yeast fed­batch fermentation.
Washington, DC, paper 132B.
Pomerleau Y. (1988). Modelisation et controle d'un precede fed­batch de
culture des levures a pain Ecole
Polytechnique de Montreal, Canada, internal report.
Pomerleau Y. and M. Perrier (1989). Estimation of multiple specific growth
rates in bioprocesses. J., in press.
Pomerleau Y. (1990). Modelisation et controle d'un prodece fed­batch de
culture des levures a pain.. Ph. D. Thesis, Ecole Polytechnique de
Montreal, Canada.

(EKF) for of growth

is :

Stephanopoulos G. and K.Y. San (1984). Studies on on­line bioreactor

identification. 26, 1176­1188.

of EKF for is :

Ljung L. (1979). Asymptotic behavior of the extended Kaiman filter as a

parameter estimator for linear systems. IEEE Aut. 24, 36­50.

of EKF to :
 Chap.3. ESTIMATION WITO KNOWN YIELD COEFFICIENTS 199

Bellgardt K.H., H.D. Meyer, W. Kuhlmann, K. Schurgerl and M. Thoma (1984).

On­line estimation of biomass and fermentation parameters by a Kalman
filter during a cultivation of Saccharomyces cerevisiae. 3rd European Cong.
Biotechnol, Verlag Chemie, Weinheim, 2, 607­616.
Caminal G., J. Lafuente, J. Lopez­Santin, M. Pooh and C. Sola (1987).
Application of Extended Kalman Filter to Identification of Enzymatic
Deactivation. Biotechnol. Bioeng., 29, 366­369.
Dekkers R.M. (1983). State estimation of a fed­batch baker's yeast
fermentation. In A. Halme (Ed.), Modelling and Control of Biotechnical
Processes, Pergamon, Oxford, 201­212.
Katoh M., H. Kanoh and M. Masubuchi (1981). Analysis and state estimation
of enzyme kinetic model for microbial growth processes. In Preprints &^
Triennal IFAC World Congress, Kyoto, Japan, 22,159­164.
Kotob S., 0. Albrecht, M. Jaridan and H. Kafeety (1987). Adaptive bioreactor
identification using sensitivity of process stoichiometry. Prod0^^ IFAC
World Congress, Munich, July 27­31, 2, 344­349.
Leigh J.R. and M.H. Ng (1984). Estimation of biomass and secondary product
in batch fermentation. In A. Bensoussan and J.L. Lions (Eds.), Analysis and
Optimization of Systems, Springer Verlag, Berlin, part 2, 456­467.
Moilanen U.R. (1981). State and parameter estimation experiments by a
fermentation process. In Preprints 8^^ Triennal IFAC World Congress,
Kyoto, Japan, 22, 185­191.
Nihtila M., P. Harmo and M. Perttula (1984). Real­time growth estimation in
batch fermentation. Proc. 9^^ IFAC World Congress, Budapest, July 2­6,
225­230.
Shimizu H., T. Takamatsu, S. Shioya and K.I. Suga (1989). An algorithmic
approach to constructing the on­line estimation system for the specific
growth rate. Biootechnol. Bioeng., 33, 354­364.
Staniskis J., and R. Simutis (1986). A measuring System for Biotechnical
Processes Based on Discrete Methods of Estimation. Biotechnol. Bioeng.,
28, 362­371.
Takamatsu T., S. Shioya, K. Yokoyama, Y. Kurome and K. Morisaki (1981).
On­line monitoring and control of biochemical reaction processes. Proc. &^
Triennal IFAC World Congress, Kyoto, Japan, Vol.22, 146­151.
 200 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Yoo Y.J., J. Hong and R.T. Hatch (1985). Sequential estimation of states and
kinetic parameters and optimization of fermentation processes. Proc, ACC,
Boston, USA, 2, 866­871.

The experiments on lactic acid production used in Section 3.4.9 are described
in detail in:

Beal C , P. Louvet, G. Corrieu (1989). Influence of controlled pH and

temperature on the growth and acidification of pure culture of S.
thermophilus 404 and L Bulgaricus 398, Appl. Microbiol. Biotechnol., in
press.
Pique D., G. Corrieu (1989). On line monitoring of fermentation processes
using a new sterilizable sampling device, 5^^ Int. Congress on Engineering
and Food, Cologne.
 ChapA ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 201

CHAPTER 4

STATE AND PARAMETER ESTIMATION WITH

UNKNOWN YIELD COEFFICIENTS

4.0. Introduction

Chapter 4 is a natural extension of Chapter 3. In the preceding chapter we

have considered different estimation problems when the yield coefficients are
known. Here we shall concentrate on the same estimation problems but when
the yield coefficients are (partially) unknown.

In Section 4.1 we address the on­line estimation problem of the specific

reaction rates. We first examine the conditions under which one can derive an
alternative dynamical model of the process from which the unknown yield
coefficients have disappeared. When this is possible, it is straightforward to
extend the observer­based and the linear regression algorithms of Chapter 3
to perform on­line estimation of the kinetic related parameter vector ρ (ξ ),
independently of the unknown yield coefficients. The issue is discussed in
Sections 4.1.1 and 4.1.2 and illustrated with practical applications (ethanolic
fermentation and anaerobic digestion) in Sections 4.1.3 and 4.1.4.

The approach proposed in Section 4.1 can be followed only if some of the
yield coefficients are unknown. Therefore, in Section 4.2, we examine the
situation where all the yield coefficients are unknown and we explore the
possibility of simultaneous on­line estimation of both specific reaction rates
 202 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

and yield coefficients from full state measurement (Section 4.2.1) and partial
state measurement (Section 4.2.2).

It is a natural extension to investigate the possibility of jointly estimating the

yield coefficients, the specific reaction rates and the missing states. Algorithms
designed for that purpose are briefly presented in Section 4.3. They are called
adaptive observers because they are state observers, as defined in Chapter 3,
which are made adaptive by introducing on­line parameter estimation.

When the yield coefficients are unknown, the simplest method may be to try to
estimate them from experimental data before going on to implement state
observers or on­line parameter estimators. The problem of estimating the yield
coefficients independently of the reaction rates, from full state measurement, is
therefore considered in Section 4.4.

Unknown parameters can also be hidden in the transport dynamics part of the
General Dynamical Model (1.43) (see Section 1.2), for which the estimation
methods presented in this book also apply. Two such issues are briefly
treated in Section 4.5 : estimation of specific liquid­gas transfer rates (Section
4.5.1) and estimation of oxygen related parameters (Section 4.5.2).

4.1. On­line Estimation of the Specific Reaction Rates

We consider a biotechnological process described by the general state space

model (2.24) :

^ = ­Ο ξ Κ Η (ξ )ρ (ξ ) + F ­ Q (2.24) = (4.1)

i.e. where the reaction rate vector φ (ξ ) is expressed as the product of an Mxr
matrix Η (ξ ) of known functions of ξ and a vector ρ (ξ ) of unknown functions of ξ
(dim p(ξ ) = r ) :
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 203

φ (ξ ) = Η (ξ )ρ (ξ )

In addition, we suppose now that (some of) the yield coefficients (matrix K) are
unknown.

In this section we examine under what conditions the estimation of the

parameter vector ρ (ξ ) could be performed independently of the unknown yield
coefficients. It is obvious that this will be possible only if an alternative state
space representation exists which is independent of these unknown yield
coefficients.

Let us begin with an example.

4.1.1. Example 1 : Simple microbial growth

Consider the simple microbial growth process with a single limiting substrate
(1.33):

S—Ax (1.33) = (4.2)

the dynamical model of which is as follows :

^ = ­k^aSX ­ DS + DSin (4.3.a)

^ = aSX­DX (4.3. b)
dt

Assume that the substrate concentration S is measured on­line (as well as the
dilution rate D and the influent substrate concentration Sjn). Define the
parameter ρ as follows :

p = a
 204 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

It is immediately apparent that, in this simple example, (ξ β = X, = S) is a nice

partition, so that the auxiliary state :

Z = S + kiX (4.4)

has dynamics :

dZ
^ = ­ D Z + DSin (4.5)
dt

independent of the yield coefficient ki and the specific reaction rate a.

We know from the structural property of Section 1.7 that the dynamics (4.3.a­b)
of the process are equivalently represented by equations (4.3.b) and (4.5). By
substituting S in equation (4.3.b) by its expression in (4.4) (i.e. S = Ζ ­ kiX),
the equivalent state space model (4.3.b)(4.5) is written as follows :

^ = a(Z­kiX)X­DX (4.6.a)
dt
dZ
^ = ­DZ + DS|n (4.6.b)
dt

We note, however, that this system representation explicitly depends on the

yield coefficient k i , and is useless for our purpose since it involves the state X
which is not measured.

But it is clear that equations (4.3.a) and (4.5) constitute an alternative

equivalent dynamical model. This one is explicitly written as follows by
replacing kiX by its expression in (4.4) (i.e. kiX = Ζ ­ S) :

^ = ­aS(Z ­ S) ­ DS + DS|n (4.7.a)

^ = ­ D Z + DS|n (4.7.b)
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 205

This model is now independent of the yield coefficient ki and involves the
measured state S. It is clear that, as a consequence, it can be used for on­line
estimation of the parameter α independently of ki. By way of example, if we
choose for instance an observer­based estimator (Section 3.4.2), the
estimation of α will be carried out as follows :

dZ
^ = ­ D Z + DS|n + ω (S ­ S) (4.8.a)
dt

^ = ­aS(Z ­ S) ­ DS + DSjn + w^iS - S) (4.8.b)

^ = YiS(S­Z)(S­S) (4.8.C)

where ω , ω ι and y^ are positive design parameters.

This example suggests an argument for the estimation of ρ independently of Κ

which is presented in the next paragraph.

4.1.2. Conditions for estimating ρ independently of the unknown yield

coefficients

Recall that the dynamics of the vector Ζ :

Ζ = Α ο ξ 3+ ξ 6 (1.70) = (4.9)

defined in the structural property of Section 1.7 are equal to :

dZ
= ­DZ + Ao(Fa ­ Qa) + (Fb ­ Qb) (1.73) = (4.10)
dt

Assume that, in line with Section 3.3.1, the (N­p) vector Ζ can be expressed as
a linear combination of the vectors ξ ι and ξ ζ of measured and non measured
variables, i.e.:
 206 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Ζ = Α ι ξ ι +Α 2ξ 2 (4.11)

The dynamics of the measured variables ξ ι are equal to :

^ = Κ ι Η ( ξ „ ξ 2)ρ (ξ ι , ξ 2) ­ Ο ξ ι + F, ­ Q, (4.12)

The kinetics term Κ ι Η ( ξ ι , ξ 2)ρ (ξ ι . ξ 2) is in general a function of some of the

unknown components of ξ 2 (and not necessarily of all of them, see e.g.
example 2 below). Assume that these components can be expressed from
(4.11) as functions of ξ ι and Ζ only (in the particular situation when a//the
components of ξ 2 appear explicitly in Κ ι Η (ξ ι , ξ 2)ρ (ξ ι , ξ 2), this means that then
A2 must be left invertible, as been shown in Section 3.3.1). Then the kinetic
term Κ ι Η ( ξ ι , ξ 2)ρ can be rewritten in terms of the measured states ξ ι and of
the auxiliary states Ζ :

Κ ι Η (ξ ι .ξ 2)ρ = Φ (ξ ι ,Ζ )ρ (4.13)

where Φ ( ξ ι , Ζ ) is a qxM matrix which is a function of ξ ι and Z. (Recall, for

instance, that in example 1, Φ (ξ ι , Ζ ) was equal to S(S ­ Z)).

Assume also that F and Q are known.

In conclusion, the parameter vector ρ can be estimated independently of the

unknown yield coefficients under the two following conditions :

C I . There exists a state transformation Ζ (4.9) whose dynamics (4.10) are

independent of the unknown yield coefficients.

C2. The reformulation (4.13) of the kinetic term Κ ι Η ( ξ ι , Ζ )ρ (ξ ι , Ζ ) is such that

the matrix Φ (ξ ι , Ζ ) is independent of the unknown yield coefficients.
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFRCIENTS 207

Remark

Note that the nice partition (i.e. with Fa ­ Qa = 0) is a particular case of

condition C 1 , since then the dynamics of the auxiliary state vector Ζ are
independent of a//the yield coefficients.

Let us now introduce two more examples by way of illustration.

4.1.3. Example 2 : Estimation of the specific growth rate independently of

the yield coefficients

We shall here extend the first example to a single aerobic biochemical

reaction with two reactants (substrate S and oxygen C), one biomass X and
one (possibly gasifiable) product P, i.e. :

S+C ­ ^ X +Ρ (4.14)

with φ = μ Χ .

The dynamical model of the system is then written as follows :

dX
= μ Χ ­Ο Χ (4.15.a)
dt

dS
= ­k1μ X­DS + F (4.15.b)
dt

dC
= ­ Ι < 2 μ Χ ­ 0 0 + 0|η (4.15.C)
dt

dP
= Κ 3μ Χ ­0Ρ ­α (4.15.d)
dt
 208 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

We shall now study the on­line estimation of the specific growth rate μ (i.e. ρ =
μ ) from the measurements of one of the process components X, S, C or P,
independently of the yield coefficients k i , k2 and ka.

If the biomass X is measured on­line, equation (4.15.a) is the relevant special

case of (4.12) with ρ = μ and Φ (ξ ι , Ζ ) = Χ . This equation is clearly independent
of the kjs. The on­line estimation of μ with this equation has been quite
exhaustively treated in Chapter 3 (Section 3.4.5).

Consider now the estimation of μ from one of the other three components S, C
or P.

(ξ β = X, ξ ό = [S, C ,P]T) is a nice partition with the following associated state

transformation :

"Zi" = X + S' (4.16)

Z2 c
.p.
.Z3. .­k3.

The dynamics are equal to :

d_r,­,
= ­ D "Zi + ' F" (4.17)
dt
Qin
- "0"0
Z2 Z2
. 0 . .Q.
.Z3. .Z3.

They are independent of the yield coefficients k i , ka and ka. Therefore,

condition CI is fulfilled.

Now let us see how to replace the unknown component ξ 2 by a function of ξ ι

and Ζ in the kinetic term Κ ι Η (ξ ι ,ξ 2)ρ of equation (4.12). Note that here the
kinetic term does not explicitly depend on all the unknown components of ξ ζ ·
Suppose, for instance, that S is the only measured component. The kinetic
term Κ ι Η (ξ ι , ξ 2)ρ is equal to :
 ChapA ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 209

Κ ι Η (ξ ι .ξ 2)ρ = ­Κ ι μ Χ (4.18)

which can be rewritten as follows by using the first component Z i of the

auxiliary vector (4.16):

Κ ι Η (ξ ι ,ξ 2)ρ = μ ( 8 ­ Ζ ι ) (4.19)

i.e. withO = S ­ Z i .

We see that we do not have to invert the matrix A2 in order to write (4.19).

The same reasoning can be used if the measured component ξ ι is C or P. As

a matter of fact, it is straightforward to see that, in each situation, the dynamical
equation of the measured variable ξ ι (which is alternatively S, C or P) may be
written as follows:

dξ
1 = μ (ξ ι ­Ζ )­0ξ ι +Ρ ι ­α ι (4.20)
dt

but with a different auxiliary variable Ζ and a different meaning for Fi and Q i
depending on which component is accessible for on­line measurement (see
Table 4.1).

ζ Fi Qi
S S + kiX F 0
C C + k2X Qin 0
Ρ P­ksX 0 Q

Table 4.1. Estimation of μ from S, C or Ρ

The dynamics of the scalar variable Ζ appearing in (4.20) are given by

the following equation :
 210 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

^^=­DZ + Fi­Qi (4.21)

dt

Note that equations (4.20) and (4.21) are independent of the yield coefficients
k i , k2 and ka. Condition C2 is then fulfilled and these equations can be used
to estimate the specific growth rate independently of k i , k2 and ka from any of
the component S, C or P. The observer­based estimator is, for instance,
written here as follows :

= ­ DZ Fi ­ Qi 4­ ω (ξ ι ­ ξ ^) (4.22.a)
dt

(1ξ
1 = μ (ξ ι ­ Ζ ) ­ ϋ ξ ι + Fi ­ Qi + ω ι (ξ ι ­ ξ ^) (4.22.b)
dt

=γ ι (ξ ι ­Ζ )(ξ ι ­ξ ι ) (4.22.C)
dt

with ω , ω ι and γ ι positive design parameters.

Experimental application : estimation of the specific growth rate μ in a

batch ethanolic bioreactor

The estimation algorithm has been implemented on an ethanolic fermentation

bioreactor described (as in Section 3.4.7) by the following reaction scheme :

S­^­Cx + P (4.23)

where S, X and Ρ are glucose, yeast and ethanol, respectively.

The only difference from the scheme (4.14) is that there is no dissolved
oxygen C because the reaction is anaerobic. It is easy to see that, in this case,
the algorithm (4.22) remains valid for estimating the specific growth rate μ
from measurements of the ethanol concentration P.
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 211

30 40 50 70
Time (hours)

0.25 Η

0.15 h

0.05 h

­0,05
10 20 30 40 50 60 70
Time (hours)

Fig.4.1. On­line estimation of the specific growth

rate in an ethanolic batch bioreactor

The process was operated in a batch reactor (Unit of Bioengineering, Catholic

University of Louvain, Belgium). Considering that the process operates in
batch mode (D = 0) and that the presence of ethanol Ρ in the gas phase is
negligible, a discrete­time version of (4.22) is written as follows :
 212 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Ζ „ ι = Ζ ι + ω Τ (Ρ ,­Ρ ,) (4.24.a)

Pt+1 = Pt + iitT(P, ­ Z,) + ω ι Τ (Ρ , ­ Ρ ,) (4.24.b)

iM+i=iit + YiT(Pt­Z,)(P,­P,) (4.24.C)

The data of the ethanol concentration Ρ and the estimates of the specific
growth rate μ , obtained with (4.24), are shown in Figs.4.1.a and 4.1.b,
respectively. The estimation was carried out under the following conditions :

­1
ZQ = 0 mmol/L, PQ = 0 mmol/L, μ ο = 0 h (4.25.a)

ω = 0, ω ι = 3.8(P, ­ Ζ ,) h"^, γ ι = 3.9 L^/mmol^.h^ (4.25.b)

The calibration of the design parameters was performed (Fig.4.2) by

comparing some off­line yeast concentration data X (measured by cell
counting) and an estimate calculated from the estimate of μ with a discrete­
time version of the dynamical equation of X in (4.15) with D = 0, as follows :

Χ ^ , ι = Χ , + Τ μ ,Χ , (4.26)

2> 80 h
to
ε
Ο

40 h • : Off­line analyses

10 20 30 40 50 60 70
Time (hours)

Fig.4.2. Experimental validation of the on­line estimation of μ

 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 213

Note the good fit between the few off­line data of X and its estimate in Fig.4.2.

Comment

A similar application of the estimation of the specific growth rate has already
been treated in Section 3.4.5. The advantage of the estimation scheme (4.24)
compared to (3.101)(3.104) is that it does not require any knowledge of the
yield coefficients. The drawback, however, is that it does not take advantage of
data on substrate feed rate F, which are usually available on­line.

Correlation between μ and the glucose concentration

In Chapter 2 we mentioned that there are three advantages to the

consideration of the parameters ρ (ξ ) as unknown time­varying parameters to
be estimated on­line :

1) it avoids the difficult choice of an analytical expression for ρ (ξ );

2) it avoids the identifiability difficulties inherent in the expression chosen ;
3) it allows, once the estimation has been processed, a search for possible
correlations between ρ (ξ ) and the factors which are capable of influencing
it. For instance, how does the substrate concentration influence ρ (ξ )? Is
there any inhibition effect which can be emphasised?

We now illustrate this idea. In the experimental application, some off­line

glucose concentration data were also available. In Fig.4.3 the glucose
concentration values have been drawn with respect to the corresponding
specific growth rate estimates μ (calculated with the algorithm (4.24)). It is
clear from this figure that the dependence between the specific growth rate
and the glucose concentration appears to be "Monod­like" , i.e. μ is an
increasing function of the glucose concentration until a certain saturation level
is reached. It is worth noting that, in order to emphasise this dependence, we
 214 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

do not need to formulate any a priori assumption about the analytical structure
of μ . Moreover, Fig.4.3 does not mean that this dependence is a Monod law. In
fact, any equivalent mathematical expression (such as the models of
Blackman or Tessier (see Appendix 1), for instance) could as well be used to
represent it.

The purpose when carrying out such an analysis does not consist of finding
the "best" analytical relationship between the specific growth rate and the
biochemical and physico­chemical variables. It is more specifically aimed at
emphasising possible correlations. In practice, this means that the above
procedure does not necessarily imply the mathematical formulation of the
correlation (with, as a consequence, the calibration of the kinetic coefficients).
In a process "modelling" context (or rather a process description context), this
has the advantage of avoiding the hazardous interpretation of uncertain
coefficients of the model (see Section 2.1).

Sugar (mmol/L)

Fig.4.3. Correlation between the estimated specific growth rate and the
glucose concentration
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 215

4.1.4. Example 3 : Estimation of the specific reaction rates in an

anaerobic digestion process

In this example, we shall show that the existence of a nice partition is not
enough to estimate the specific reaction rates independently of the (unknown)
yield coefficients.

Consider the same example as in Section 3.3.1, i.e. the simplified scheme of
the anaerobic digestion without the two products Pi and P2 (see (3.23)) :

Φ Ι
Si — + S2 (4.27.a)
φ

So ^ X o (4.27.b)

The dynamical model is defined as follows :

d_
"sr = ­ki 0 " 'SiXi 0 • "«Γ ­ D "SI­ "DSin­ (4.28)
dt
1 0 0 S2X2. .«2. XI 0
S2 ka —k2 S2 0
.X2. . 0 1 . . 0
.X2.

Assume that the substrate concentrations Si and S2 are measured on­line.

We can define the following state transformation which corresponds to a nice

partition :

z= " z ; = • ki 0" •χ Γ + 's/ (4.29.a)

.Z2. —k3 k2 .X2. .S2.

(4.29.b)

with Fi =
0
 216 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

But the dynamical behaviour of ξ ι (= [Si 82]^) does not become independent
of the yield coefficients with the introduction of the above state transformation.
In fact, Κ ι Η (ξ ι , Ζ ) is equal to :

Κ ι Η (ξ ι ,Ζ ) = Si(Si­Zi) 0 (4.30)

^Si(Zi­Si) S2is2­Z2­K^(Si­Zi)
Lki

Therefore, a knowledge of the yield coefficients ki and ks (but not k2) is

necessary to perform the on­line estimation of a i and a 2 with one of the
algorithms of Section 3.4. More precisely, the specific reaction rate a i can be
estimated independently of the yield coefficients, while the estimation of a 2
with the above formulation requires the knowledge of ks/ki (which, as a matter
of fact, is the yield coefficient from Si into S2, i.e., in the two­step anaerobic
digestion example, from organic matter into acetate).

4.2. Joint Estimation of Yield Coefficients and Specific Reaction

Rates

The issue of estimating the reaction rates φ (ξ ) (or related quantities like ρ (ξ ),
α (ξ ), μ (ξ )) has been already addressed in this book. A general method was
described in Chapter 3 (Section 3.4) in the case in which the yield coefficents
(i.e. the matrix K) are known. Later on, in Section 4.1, we discussed the case
in which some of the yield coefficients are unknown but can be eliminated from
the estimation procedure by appropriate model transformations. In this
section, we shall examine the situation in which all the yield coefficients are
unknown. We explore the possibility of simultaneous on­line estimation of
both specific reaction rates α (ξ ) and yield coefficients (matrix K), from full state
measurement (Section 4.2.1) and partial state measurement (Section 4.2.2).
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 217

4.2.1. Full State measurement

We consider biotechnological processes described by the state space model

(1.49):

^ = Κ Ο (ξ )α ­ Ο ξ + F ~ Q (1.49) = (4.31)

We assume that:

1) the full state ξ is measured on­line;

2) the dilution rate D, the feed rates F and the gaseous outflow rates are
known on­line;
3) the yield coefficients (matrix K) and the specific reaction rates α are
unknown.

The problem is to design an estimator of Κ and a.

Recall that Ν denotes the number of components and Μ the number of

reactions involved in the process. The first point to notice, then, is that the
matrix Κ necessarily contains Μ nonzero entries which may be freely assigned
by the user and are usually set to Ί " as was discussed in Section 1.5.6.

If we denote by Nj the number of components involved in the reaction with

index j , it is then easy to see that the total number m of unknown yield
coefficients in Κ (denoted k i , kz km) is given by :

Μ
m=^(Nj­1) (4.32)
i=i

Moreover, there are obviously Μ parameters aj Q = 1, M) in the model

(4.31). Hence the total number of unknown parameters is :
 218 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

m Μ = 2^ Nj (4.33)
i=i

Now, due to its particular structure, the vector Κ 6(ξ )α may be rewritten in a
linear regression form as follows :

Κ Ο (ξ )α = Φ (ξ )θ (4.34)

In this expression, Φ (ξ ) is a Nx(m+M) matrix of known mty/f/V/nearcombinations

of the state variables. 0"'' = [θ ι , θ 2 , . . . Θ γ π +Μ ] is a vector of d/V/nearcombinations
of the unknown parameters kj (i = 1 , . . . m) and aj 0 = 1, M). It is formally
written as follows :

e = f(a,k) (4.35)

The definitions of Φ and θ will be soon illuminated by several examples. An

important point is that the function f in (4.35) is invertible, that is α and k can be
computed uniquely from θ :

r a l = rVe) (4.36)
k

With the definition (4.34), the process model (4.31) is rewritten as follows :

$ = Φ (ξ )θ ­Ο ξ ­Η Ρ ­0 (4.37)
dt

Since the unknown parameter θ enters linearly in (4.37), a natural solution is

to estimate it with algorithms similar to those of Chapter 3 (Section 3.4), and
then to recover α and k with the inverse function (4.36). The algorithms are
written as follows :
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 219

Observer-based

Φ (ξ )θ ­ο ξ + ρ ­ α ­ Ω (ξ ­ξ ) (4.38.a)
dt

d§
= Φ ^(ξ )Γ (ξ ­ξ ) (4.38.b)
dt
ά = Γ \θ ) (4.38.C)

^ = Ω ψ Τ + Φ (ξ ) (4.39.8)
dt

= Ω ψ ο + (Ω ­Ο ΐ Ν )ξ ­0+Ρ (4.39.b)
dt

^ = Γ ψ (ξ ­ψ ο ­ψ ^θ ) (4.39.C)

dr
= ­Γ ψ ψ '^Γ + λ Γ (4.39.d)
dt

ά = r\Q) (4.39.e)
R

Consider the basic process (1.33):

s—JCX (4.40)

whose dynamics are written :

 220 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

d_rw­l Γ

dt
SX 0 α ­Dfx (4.41)
0 ­SX aki .SJ DSi„
We notice that we have here :

M =1 m=1 dim(e) = m ­h Μ = 2 (4.42)

θ = "Θ Γ — α ' = f(a,ki) Φ (ξ ) = SX 0 (4.43)

aki . 0 ­SX.

The function f is inverted as follows

= rVe) = θ ι (4.44)
θ 2
θ ι

If we choose diagonal matrices for Ω and Γ ι , the observer­based estimator

(4.38) resolves t o :

^ = SXθ 1­DX­^ω 1(X­X) (4.45.a)

dt

^ = ­ SXGa ­ DS ­I­ DSin + ω 2(8 ­ δ ) (4.45.b)

^ = YiSX(X­X) (4.45.C)

dSj (4.45.d)
­ ^ = ­Y2SX(S­S)

ά =θ ι (4.45.e)

R (4.45.f)
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 221

Example : Yeast fermentation process

Consider a yeast fermentation described by the reaction scheme (1.37).

Suppose, however, that the reaction of respiratory growth on ethanol is not
activated. Hence the reaction scheme reduces to :

Φ Ι
s + c—Cx + i
92

S—Cx + E + P

with X being biomass, S glucose, C oxygen, Ε ethanol and Ρ carbon dioxide.

The associated dynamical model is as follows :

= ki k2 "xsc 0 • 'α ι " ­Dfx 0 0

dt
­k3 ­k4 0 XS. . « 2 . S DSin 0
0 kg Ε 0 0
C Qin 0
­ke 0
Ρ 0 LQi.
. 1 1 .

Expression (4.34) is written here :

Φ (ξ )θ = XSC XS 0 0 0 0 0 0 • 'Θ Γ
0 0 xsc XS 0 0 0 0 02
0 0 0 0 XS 0 0 0 θ 3
0 0 0 0 0 xsc 0 0 θ 4
0 0 0 0 0 0 xsc XS.
θ 5
θ 6
θ 7
.θ 8.

with the reparametrization θ = f(a,k) :

 222 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

θ ^ι = KI (Xi Θ 2 = K2(X2 Θ 3 — KAOCI Θ 4 = — K4(X2

Θ 5 = ^5Α 2 e6 = ­ K 6 A I 67 = Α Ϊ 08 = α 2

which is clearly invertible.

4.2.2. Partial state measurement

We shall now reconsider the problem of joint estimation of yield coefficients

and specific reaction rates, but when only a part of the state is measured on­
line.

We assume that:

1) the subset of measured state variables is denoted ξ ι ; the set of remaining

non measured states is denoted ξ 2, so that ( ξ ι , ξ 2) is a state partition which
in turn induces the similar partitions (Ki, K2), (Fi, F2), (Qi. Q2) as before;

2) there exists a nice state partition (ξ β , ξ b), with an associated auxiliary state Ζ
(see sections 1.7 and 3.3.1):

Ζ = Α ο ξ 3 + ξ 6 = Α ι ξ ι ­Η Α 2ξ 2 (4.46)

which obeys the following dynamics :

^ = ­DZ­HFb­Qb (4.47)

3) the matrix A2 is left invertible.

Under these assumptions, the state space model (4.31) is equivalent to :

 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 223

^ = Κ ι Ο (ξ ι . A ^ ( Z ­ Α ι ξ ι ) ) α ­ Ο ξ ι + Fi ­ Qi (4.48.a)

dZ
= ­ D Z + Fb­Qb (4.48.b)
dt

Now, as in the previous section, it is easily seen that the term

Κ ι Ο (ξ ι , (Z ­ Α ι ξ ι ))α may be rewritten as :

Κ ι Ο (ξ ι , A ^ ( Z ­ Α ι ξ ι ) ) α = Φ (ξ ι , Ζ )θ (4.49)

The structures of Φ and Θ , however, are slightly different from the previous
ones. Φ (ξ ι ,Ζ ) is now a matrix of known polynomial combinations of the
measured states ξ ι and the auxiliary state Z. θ is a vector of polynomial
combinations of some of the unknown parameters kj and aj, denoted as before
in (4.35):

e = f(a,k) (4.50)

It may arise, however, that this function f is not completely invertible . Then, in
contrast with the previous case, only a few kj and aj can be recovered from Θ .
This is illustrated in the examples below.

With the (4.49), the model (4.48) is rewritten :

^ = Φ (ξ ι ,Ζ )θ ­Ο ξ ι +Ρ ι ­α ι (4.51 .a)
dt
dZ
= ­ D Z + Fb­Qb (4.51 .b)
dt

On the grounds of our previous discussions, the solution to the estimation

problem readily appears : it suffices to combine a parameter estimator for θ
(Section 4.2.1) with an asymptotic observer for Ζ (Section 3.3.1). For instance,
an observer­based estimator is written as follows :
 224 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Observer-based

^ = ­D2­fFb­Qb (4.52.a)

^ = Φ (ξ ι , 2)θ ­ Ο ξ ι + Fi ­ Qi ­ Ω ι (ξ ι ­ | i ) (4.52.b)

^ = Φ Τ (ξ ι .2)Γ (ξ ι ­ξ ι ) (4.52.C)

Example 1. Simple microbial growth process

Consider the basic microbial growth process (1.33)

S—Cx (4.53)

the dynamics of which are written :

SX 0 α ­Dfx 0 (4.54)
0 ­SX akij S DSin

Assume that only the biomass concentration X is measured on­line. Hence,

we have :

ξ ι =Χ ξ 2=δ (4.55.a)

K«j = 1 K2 = — k­j (4.55.b)

Z = S + kiX (Ao = A i = k i , A2 = 1) (4.55.C)

The equivalent model (4.51) is written here :

 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 225

dX Γ
zx ­χ 2 α ­DX (4.56.a)
dt
ak U

dZ
^ = ­ D Z + DSiIN (4.56.b)
dt

That i s :

Φ ( ξ ι . Ζ ) = Lzx ­X^J (4.57.a)

θ = =• α ' (4.57.b)
β 2.

We notice that, in this example, the parameter θ is identical to that which was
identifiable when both S and X were measured (example 1 of Section 4.2.1).

2.

We consider the simplified reaction scheme (1.86) obtained in Section 1.8.4 :

Si — + S2 + Pi + P2

S 2 — C X 2 + P 1 + P2

We suppose, furthermore, that we are not interested in the products P i and P2,
so that we can omit them in the analysis. Then the state space model is as
follows :

dt
= ­ki 0 ' SiXi 0 •«Γ ­ D f ' S I ­ + Ρ ί ­ (4.58)
Xi 1 0 0 S2X2 . « 2 . XI ο
S2 k3 ­k2 S2 0
. 0 1 . .0.
.X2. .X2.

We assume that the substrates S i , S2 are measured on­line :

 226 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

ξ ΐ =[8ι S2]^ ξ 2=[Χ ΐ X2]^ (4.59)

We notice that (ξ β = ξ 2 . = ξ ι ) constitutes a nice partition of the state.

The auxiliary state Ζ is as follows :

z= = ξ 6 + Α ο ξ 3 = Α ι ξ ι + Α 2 ξ 2 = ' s r + ' ki 0' (4.60)

.Z2. .S2. .­ka k2

Clearly the matrix A2 is full rank. Straightfonward calculations then lead to the
following expressions :

Φ (ξ ι .Ζ ) = S i ( S i ­ Z i ) 0 0 0 (4.61.a)
Ο S i ( S i ­ Z i ) S2(S2­Z2) S 2 ( S i ­ Z i )

θ = θ ι «1 (4.61.b)
θ 2

θ 3
θ 4 «2

«2k7.
We notice that the new model is now "underparametrized" in the sense that it
contains only four parameters 6 1 , 0 2 , 6 3 , 6 4 , although the initial system
contained five unknown parameters, α ϊ , α 2, k i , k2, ks. Therefore, as appears
clearly from (4.61), there is no hope to recover all the aj and kj from Θ .

On the other hand, if we consider the same example, but assume that the
measured state is ξ ι = [Xi X2], then the calculations lead to the following
expressions for Φ and 6:

Φ (ξ ι .Ζ ) = XiZi ­ X i (4.62. a)

0 0 X2Z2 X1X2 ­xi.

θ ^ = [ θ ι Θ 2θ 3 64θ 5]^=[α ι a i k i α 2 a2k3 a2k2] (4.62.b)

 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 227

In this case, the reparametrization θ is identical to that which would be used

when the full state is measured (Section 4.2.1). Therefore, it is clearly
invertible:

Θ 2 Θ 4 Θ 5
α ι =θ ι ki = ­ ^ α 2= θ 3 ^3 = . ^ ^^"θ ^

1) The equivalent model (4.51) has been introduced under the condition that
the matrix A2 is left invertible. There are however particular situations in
which a suitable model of the form (4.51) can be derived even if the matrix
A2 is not left invertible. This issue has been already discussed in Section
4.1.

2) As in Chapter 3, the convergence of the parameter estimators presented

above is guaranteed only if the regressor matrix Φ (ξ ) o r Φ ( ξ 1 , Ζ ) is
persistently exciting (see Appendix 3). This condition is easily fulfilled
when the yield coefficients are known (see for example Theorems 3.2 and
3.3). But the problem is dramatically more difficult when the yield
coefficients are unknown. Actually the problem of deriving sufficient
experimental conditions so as to assure the persistency of excitation of
regressors in nonlinear systems is at the time of writing still an open
question in the field of control science. Only very limited results are
available. As a matter of illustration, it can be shown for example, that the
following conditions are sufficient to guarantee that the regressor (4.57.a) of
example 1 (simple microbial growth) is persistently exciting :

C I . The dilution rate is positive and constant:

0 < D = constant

C2. The influent sustrate concentration Sjn is :

 228 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

­ persistently exciting of order 2 (see Appendix 3)

­ piecewise continuous
­ bounded as follows :

^1 min

where Oimin and e2max, respectively, are known lower and upper bounds
on θ ι and Θ 2 such that 2e2max ^ θ ΐ π ΐ π ­

Simulation result

Fig.4.4 shows a simulation result of the estimation problem of example 1

(simple microbial growth). The simulation has been performed under the
following conditions :

α = 0.4 l/g.d ki = 0.25

X(0) = 0.7 g/l S(0) = 0.25 g/l
D(0) = 0.1 d­1 S|n(0) = 2g/I

This means that a Blackman model (see Appendix 1) in limiting conditions has
been chosen :

μ = α 8

with α = constant

Random amplitude step functions of the inputs D and Sjn have been applied to
the process (Fig. 4.4.a&b).

A discrete­time least squares estimation algorithm has been implemented to

estimate θ ι and Θ 2. It specialises here as follows :
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 229

Z , + i = Z , ­ T D , Z , + TD,Sin., (4.63.a)

^IT+I =Γ θ ΐ ί 1+ Τ Γ ,φ /ί Χ ,+ι ­ χ , + TD,Xt ­ &i,T2tX, + GatTXf) (4.63.b)

.^21+1. .^2t.

I­ (4.63.C)

with r , = Yl1t Yl2t (4.63.d)

LYl2t Y22tJ

φ , = [Ζ ,Χ , ­Xf] (4.63.e)

0<λ <1 (4.63. f)

A sampling period Τ of 1 hour has been chosen. The following initial

conditions and forgetting factor have been used in the estimator:

§10 = 620 = 0 Zo = 2g/l Γ ο = 10^*Ι λ = 1

Fig.4.4.c&d show the convergence of the estimates of θ ι and Θ 2 to their true

constant values. From (4.45.e), we know that the estimation of the specific
reaction rate α is directly given by :

at = ei,

Fig.4.4.e shows the reconstruction (4.45.f) of the yield coefficient ki from the
estimates of θ ι and Θ 2 :

Kit —
θ 1t

Note that, in the simulation of Fig.4.4, both the influent substrate Sjn, as well as
the dilution rate D, fulfill the first requirement of condition C2 (persistence of
excitation of order 2). It has been observed in practical implementation that a
time­varying dilution rate D(t) induces more excitation to the estimator and
therefore improves the convergence rate of the algorithm.
230 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

D(d­1)
0.11

0.1
time (hours)

480 960
Sin (g/l)
2.1

2.

time (hours)
1.9
480 960

time (hours)

480 960

time (hours)

480 960

time(hours)

480 960

Fig.4.4. On­line estimation of the specific reaction rate

and the yield coefficient in a basic microbial growth process
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 231

4.3. Adaptive Observers

In Section 4.2.2 we examined the problem of joint estimation of the yield

coefficients k,­ and the specific reaction rates ajfrom partial state measurement
ξ ι . It is a natural extension to investigate the possibility of jointly estimating
the yield coefficients, the specific reaction rates and the missing states.
Algorithms designed for that purpose are called adaptive observers because
these are state observers as defined in Chapter 3, which are made adaptive
by introducing on­line parameter estimation. The design of adaptive
observers for nonlinear systems in general is still, nowadays, a matter of
intense research. It is beyond the scope of this book to report on the state­of­
the­art of that topic. We limit ourselves to a review of three possible solutions,
without further comment.

Extended Luenberger adaptive observer

A first solution is the combination of a Luenberger observer (3.2) with the

observer­based parameter estimator (4.52). The adaptive observer is as
follows :

Extended Luenberger adaptive observer

$ = Φ (ξ )θ ­Ο ξ + Ρ ­ α + Ω (ξ ­ξ ) (4.64.a)
dt

^=φ \ξ )Γ (ξ ι ­ξ ι ) (4.64.b)

Extended Kaiman fiitering

As we have discussed in Chapter 3, extended Kaiman filtering can be used

either for state estimation (Section 3.2.2) or for parameter estimation (Section
 232 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

3.4.6). Obviously, it can also be used for both together. In that case, the
adaptive observer is written as follows (compare with (3.133)):

Extended Kalman

^ = Φ (Ξ )θ ­ Ο Ξ + F ­ Q + Ω ι ( Ξ , Θ ) ( ξ ι ­ | i ) (4.65.a)

^ = Ω 2(ξ .θ )(ξ ι ­ξ ι ) (4.65.b)

where Ω ι and Ω 2 are updated by an appropriate Riccati equation.

An

A third solution, which is more closely related to the results of Section 4.2.2, is
the following one. We suppose that ail the yield coefficients involved in A i and
A2 (equation (4.46)) can be calculated from Θ . An adaptive observer is then
readily obtained by augmenting the observer based parameter estimator
(4.52) as follows :

dZ
= ­ D Z + Fb­Qb (4.66.a)
dt

1 = Φ (ξ ι ,Ζ )Θ ­ Ο ξ ι + Fi ­ Qi ­ Ω ι (ξ ι ­ ξ ι ) (4.66.b)
dt

^ = Φ Τ (ξ ΐ ,Ζ )Γ (ξ ι ­ξ ι ) (4.66.C)

ξ 2 = Α ^(Ζ ­Α ι ξ ι ) (4.66.d)
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 233

where A2 and A^ respectively denote the matrices A2 and A i whose nonzero

entries are substitued by the corresponding estimates.

Example : simple microbial growth process

Let us return to example 1 of Section 4.2.2. The Luenberger adaptive

observer is written :

^ = 8Χ θ ι ­ DX + ω ι (Χ ­ X) (4.67.a)

^ = ­ 8 Χ θ 2 ­ 0 8 + 08|η + ω 2 ( Χ ­ Χ ) (4.67.b)

­ ^ = YiSX(X­X) (4.67.C)

­ ^ = ­Y2SX(X­X) (4.67.d)

The asymptotic adaptive observer is written :

^=­DZ+DSin (4.68.a)

dX
= Ζ Χ Θ ι ­ Χ ^θ 2 ­ DX + ω (Χ ­ Χ ) (4.68.b)
dt
dO
I = YiZX(X­X) (4.68.C)
dt

= ­Y2X^(X­X) (4.68.d)
dt

S = Z ­ t 4 X (4.68.e)
θ ι
 234 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Simulation result

Fig.4.5 Shows the complementary simulation result of Fig.4.4, i.e. the adaptive
observation of the substrate concentration S with equation (4.68.e) where
θ ι and §2 are given by (4.63.b) as shown in Fig.4.4.c&d.

[S(g/I)

^ time(hours) ^60

Fig.4.5. Adaptive observation of S

4.4. Estimation of Yield Coefficients

4.4.1. The estimation problem

In this section we shall address the problem of estimating the yield coefficients
independently of the reaction rates. We consider biotechnological processes
described by the general dynamical model (1.43) :

= Kφ (ξ )­Dξ ­HF­Q (1.43) = (4.69)

dt

under the following conditions :

1) the yield coefficients (matrix K) are unknown;

2) the reaction rates φ (ξ ) are unknown;
3) the full state ξ is measured;
 ChapA ESTIMATION WITO UNKNOWN YIELD COEFFICIENTS 235

4) the substrate feed rates F, the gaseous outflow rates Q and the culture
volume V are measured.

On this basis, we investigate the possibility of estimating the yield coefficients

independently of the reaction rates. In a first step we derive a linear
regression form equivalent to the model (4.69) but independent of the reaction
rates φ (ξ ).

4.4.2. A linear regression form

We consider a partition of the model (4.69), as defined in Section 1.7 :

dt =Κ 3φ (ξ) ­ + Fa ~ Qa (1.68) = (4.70)

dξ b
dt
= Kbφ (ξ ) ­ Dξ b + Fb ­ Qb (1.69) = (4.71)

with dim(ξ a) = ρ = rank(K) and dim(ξ b) = N­p

We define the auxiliary time­varying quantities :

η β ί ί ) = V(t)ξ 3(t) ­ ν (0)ξ 3(0) + J [Foutξ a + y(Qa ­ Fa)]CLT (4.72)

0

Ά ^(Χ ) = V(t)ξ b(t) ­ V(0)ξ b(0) + [Foutξ b + V(Qb ­ Fb)]dT (4.73)

0

where V denotes the culture volume and Fout the effluent flow rate (see
Section 1.1). Note in the integral of expressions (4.72) and (4.73) that the first
term (Foutξ a and Foutξ b. respectively) is equal to zero in batch and fed­batch
reactors and that the second term (V(Qa ­ Fa) and V(Qb ­ Fb), respectively)
represents the cumulated mass of feed introduced and of gas produced.
 236 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Then, according to the structural property of Section 1.7, straightforward

calculations and integration lead to the following relationship between η a(t)
and η b(t):

η 6ω = ­Α ο η 3ω (4.74)

where AQ is the solution of the matrix equation ( 1 . 7 1 ) :

AoKa + Kb = 0 (4.75)

Most often, Ka is a square MxM nonsingular matrix, such that AQ = ­KbK^^ and
( 4 . 7 4 ) is rewritten :

nb(t) = (KbK;;^)na(t) (4.76)

Let us now denote by θ = [θ ι ,...,θ |]"^ (dim(e) = I) the vector of the entries of the
matrix AQ which are not identically zero. Expression (4.76) is rewritten in the
following linear regression form :

η 6(0 = π \η 3ω )θ (4.77)

where Π "^(η 3(1)) is a (N­p)xl matrix, function of η a(t).

It is worth noting that this form is independent of the kinetics.

4.4.3 Example 1 : PHB acid production process

Consider the state space model (1.66) for intracellular production of PHB acid,
with the following state partition :
 ChapA ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 237

X ξ 6 = S2 (4.78)
LSi
c
LP2.

with X being biomass, Si fructose, S2 ammonia, P i PHB acid, P2 carbon

dioxide, and C oxygen. The corresponding partition of the matrix Κ is written :

Ka = 1 0 ­k2 0• (4.79)
kg 1
­ k 4 ­ke
1^7 ke.

The matrix ­ A Q = KbK^^ is calculated as follows

­k2 (4.80)
­1
k3­r^
"kg

"'^^Π ^ k^
kik
1^8 ke
ky
ks "ks

We see that this matrix contains I = 7 nonzero entries, so that we define the
vector θ = [θ ι θ γ ]^:

k­] k­] kg kik,

^1 = ­k2 Θ 2 = ­ks ­ ­j^ Θ 3 = k4 ­ ­j^ 04 = k 7 ­ .i^e
ks
1 ke ke
(4.81)
ks ks

Finally the matrix H is defined as follows :

 238 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Π ^(η 3) = η β ΐ 0 0 0 0 0 0 (4.82)
Ο Ο Ο Ο Ο
Ο Ο Ο Ο Ο
Ο Ο Ο Ο Ο η 32.

where :

V(t)X(t)­V(0)X(0) + FoutXdx (4.83)

,^a2. Λ

V(t)Si(t) ­V(0)Si(0) + jjFoutSi ­ FinSuJdT

4.4.4. Estimators of the parameter vector θ

Suppose that measurements of ξ , F and Q have been collected during a time

interval [0, tf]. Obviously, expression (4.77) naturally suggests the use of a
linear regression technique to estimate θ independently of the reaction rates
φ (ξ ). The standard least squares algorithm is written as follows in this case :

[R(na)nb]dx (4.84)
I Jo J Jo

The matrix inversion can be avoided by using a recursive form of this algorithm
(compare, for instance, with (4.39)):

^ = rR(na)[nb­RV)Q] (4.85.a)

^=­Γ R(η a)R^(η a)Γ (4.85.b)

 ChapA ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 239

4.4.5. Example 2 : Basic microbial growth process witli a synthesis

product

Consider a basic microbial growth process with a synthesis product Ρ :

S—Cx + P (4.86)

Its dynamics are described by the following model:

d_
dt
­ki φ ­iDfS DSiin (4.87)
1 X 0
L . J L^gJ 0

Consider for instance the following state partition {ζ ^, ξ b)

ξ 3=Ρ ξ ί =[3,Χ ] (4.88)

Then Ao is clearly equal to :

Ao = (4.89)
k2
­1
k2j

Equation (4.76) specialises here as follows

(4.90)

with :
 240 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

V(t)S(t) ­ V(0)S(0) + f [Fou,S ­ FinSinldx (4.91 .a)

V(t)X(t)­V(0)X(0)+ f FoutXdx
Jo

« ( Η Β ) = ­ N A ( t ) = ­ v ( t ) P ( t ) + v(o)P(o) ­ ^ [ F o u , P + vQidt (4.91.b)

θ ι (4.91.0)
k2
L92J
­1
L

The above relation is obviously invertible :

(4.92.a)
02

k . = ­ ^ (4.92.b)
02

And the yield coefficient k i and k 2 can be estimated from S, X and P.

If X is not available for measurement, but only S and P, we can still estimate 0 i
= ki/k2, which is the expression of the yield from S to P. This is now illustrated
with a real­life experiment.

Real-life implementation on an anaerobic digestion pilot reactor

Consider a two­step anaerobic digestion process in which the methanization

phase is assumed to be rate­limiting. The process can then be described by
the reaction scheme (4.86) where S is the organic matter and Ρ the methane.

The estimation of ki/k2 has been implemented on a 60 liter CSTR pilot reactor
(Unit of Biogineering, Catholic University of Louvain, Belgium) over a period of
14 days by using a discrete­time least square estimation algorithm.
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 241

By taking account of the low solubility of the methane (see Section 1.8.4), a
discrete­time version of the estimator (4.85) specialises here as follows (with
an additional forgetting factor):

Θ ι ,ί +1 = θ ι ,ί + γ ί Ρ ( η 3 ί ) [ η 6 ΐ ­ θ 1tR(η at)] (4.93.a)

Yt= — r­ 0<λ <1 Yo>0 (4.93.b)

Ρ (η 3ΐ ) =­ VQdx (4.93.C)
Jo

η bt = V(t)S(t)­V(0)S(0) + [FS­FSJdx (4.93.d)

with F = Fout = Fin = DV (since the bioreactor is a CSTR).

The reactor was operated under the following conditions. The dilution rate D
was set to 0.1 d­i and a step of influent substrate concentration Sjn (from 10 to
20 gCOD/l) was applied during the second day of the experiment (Fig.4.6.a).
The on­line measurements of the methane gas flow rate Q and of the substrate
concentration S are shown in Fig.4.6.b­c. The sampling period Τ is equal to 1
hour. The estimation of θ ι shown in Fig.4.6.d was performed under the
following conditions :

= 1.15, 70 = 0.3, λ = 0.99 (4.94)

Note that converges to a value close to 1.1, which corresponds to the

following physical value.

Simple

An approximate value of the yield coefficient θ ι = k i / k 2 can be deduced by

using the following line of reasoning.
242 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

"Sint(9C0D/l)

­ a

240

tCh)
240

Fig.4.6. Estimation of θ ι
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 243

It is usually considered that in an anaerobic digestion process, when 1 g COD

is decomposed, in the mean 0.1 g is used for the production of acidogenic
bacteria and 0.9 g for the production of volatile acids (VA). Of this 0.9 g, 0.02 g
is used for the production of methanogenic bacteria and 0.88 g under the form
of methane gas CH4.

1 g COD 0.1 g COD : Xacldogenlc

^ 0.9 g COD : VA ­> 0.02 g COD Xmethanogenic

^ 0.88 g COD CH4

At a temperature Τ = 35°C and a pressure ρ = 1 atm (which corresponds to the

physical operating conditions), 1 mole of CH4 correspond to 64 g COD and
25.3 I CH4.

1 mole CH4 64 g COD 25.3 I CH4

This gives a conversion ratio CH4/COD equal to 2.53 g COD/I CH4. A mean
theoretical value of θ ι can then be deduced :

Q _ 2.53x24 _^ gCODxh
^ 0.88 X 60 I CH4 X day X I reactor

The term 24/60 is introduced in order to take account of the fact that the
methane gas flow rate is expressed in h­"" (instead of day''x I reactor) (the
volume of the reactor is here equal t°60 I).

4 . 4 . 6 . Structural identifiability of the yield coefficients

With the algorithm (4.84), we have a tool for estimating the parameter Θ . But
our goal is actually to estimate the yield coefficents, i.e. the vector k = k i , k2,...
km). Clearly, as illustrated with (4.81), θ is a nonlinear rational function of k,
which we write formally as follows :
 244 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

θ = f (k)

Therefore the question arises whether k can be calculated from a knowledge

of θ or not. This is called the structural identifiability issue. It is evidently of
interest to be able to detect beforehand those models which are not
identifiable whatever the experimental conditions. A first trivial condition is
obviously that the number of unknown yield coefficients cannot be larger than
the number of parameters θ ί . For example the yield coefficients of the PHB
adic production process are certainly not identifiable (independently of the
kinetics) since the number of unknowns is eight while there are only seven
parameters θ | (see (4.81)). (For further details on the structural identifiability
issue of the yield coefficients kj, see the bibliography.)

4.5. Other Parameter Estimation Issues in Bioreactors

At this stage, we would like to remind the reader of the important remark which
we made in Section 1.5.2, regarding the general dynamical model (1.43) :

^ = Κ φ (ξ ) ­ Ο ξ + F ­ Q (1.43) = (4.95)

This model is a combination of two terms : Κ φ (ξ ) describes the process kinetics

while ­Ο ξ + F ­ Q represents the mass transport (or dilution) dynamics.

In Chapters 3 and 4, we have focused either on the real time estimation of the
state ξ or on the estimation of the parameters related to the kinetics Κ φ (ξ ). But
unknown parameters can also be hidden in the transport dynamics, for which
the methods followed so far may easily be extended. Two such issues are
briefly examined in this section.
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 245

4.5.1. Estimation of specific liquid gas transfer rates

In Section 1.5.4 we argued that it is legitimate to consider that the gaseous

outflow rates Qj are proportional to the concentrations ξ ί in the liquid phase :

α ί = β ΐξ |

as long as ξ ί is lower than the saturation level. The parameters β ι are termed
specific liquid-gas transfer rates. Define the vector β containing those
parameters β ί . It is then clear that the model (4.95) can be rewritten as follows:

^ = Φ (ξ )θ ­ Dξ + F ­ Ψ (ξ )β (4.96)

where Ψ (ξ ) is an Nxdim(β ) matrix which depends linearly on ξ .

If we define :

Φ *(ξ ) = [Φ (ξ ).ψ (ξ )] θ *= θ (4.97)

ίβ .

this model (4.96) can also be written

^ = Φ *(ξ )θ *­Dξ + F (4.98)

It is then evident that the augmented parameter Θ * can be estimated by any of

the algorithms which have been built for Θ , only replacing Φ by Φ * and θ by Θ *.
It is, however, necessary that the augmented regressor Φ * is persistently
exciting. A simple example illustrates the definitions.
 246 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Example

Consider the process

S­^­Cx­i­P

where S is the substrate, X the biomass and Ρ a gasifiable product (e.g. CO2).
The dynamical model is as follows :

4rx^
dt
SX 0 0 α l­DrSl­i­r 0 β
0 ­SX 0 DSi,
0 0 SX

"φ (ξ )"

Hence the augmented regressor and parameter are defined :

Φ *(ξ ) = SX 0 0 0 α
0 ­SX 0 0 ­kitt
0 0 SX ρ
L β J

4.5.2. Estimation of oxygen related parameters

Another typical application, which has given rise to numerous studies (see the
bibliography), concerns the parameters related to the dissolved oxygen
concentration, which is often the most easily measurable state. Consider an
aerobic process described by the following scheme :

S­i­C­ ­Cx
with S the substrate, X the biomass and C the dissolved oxygen. The biomass
and dissolved oxygen dynamics are written (see Section 1.5.5) :
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 247

dX
= μ Χ ­ DX (4.99.a)
dt

dC
= ­ Κ 2 μ Χ ­ D C + KLSCCS ­ C ) (4.99.b)
dt

where kLa and Cs denote the mass transfer coefficient and the saturation
concentration, respectively. These equations may be reformulated as follows :

—Γ χ ­· X 0 0 0 Γ μ ­ D "X"
dt'^
0 ­X 1 ­c .c.
kLaCs
ki^a

= Φ *(ξ )θ *­Dξ

with

Φ *(ξ ) = X 0 0 0 and Θ * = θ ι μ
0 ­X 1 C k2μ
Θ ;
kLaCs
θ 3

The model is thus clearly in the correct format for parameter estimation. We
notice, in addition, that the reparametrization Θ * is invertible (see the
discussion in Section 4.2.1), that is estimates of μ , k2, kLa and Cs can be
recovered as follows :

μ =θ ί k2 = ^ Μ =θ 4 Cs ^
θ ι θ 4

In addition, an on­line estimate of the OUR (see Section

1.2) is given by :
 248 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

0υ Ρ = θ 2Χ

Numerous variants for the estimation of oxygen related parameters in aerobic

processes have been proposed in the literature (see the bibliography). For
instance, one can assume that the mass transfer coefficient KLA is proportional
to the aeration flow rate (see Section 1.5.5) :

KLA = koFg

AND HENCE ONE CAN REWRITE (4.99) as FOLLOWS :

dC r
OUR ­ D C
koCs
ko J

i.e., one can proceed directly to the estimation of the oxygen uptake rate OUR,
together with the parameters ko and Cs­

4.6. References and Bibliography

Section 4.1

Several examples of the on-line estimation of the specific growth rates

independently of the yield coefficients are presented in the following paper :

Bastin G. and D. Dochain (1986). On­line estimation of microbial specific

growth rates. Automatica, vol.22, n°6, 705­711.

Example 3 is presented in full detail in :

Dochain D. and A. Pauss (1988). On­line estimation of specific growth rates :

an illustrative case study. Can. J. Chem. Eng., vol.66, n°4, 626­631.
 Chap.4. ESTIMATION WITH UNKNOWN YIELD COEFFICIENTS 249

Section 4.2

The C1 C2
1 :

Bastin G., LM.Y. Mareels, D. Dochain and M. Gevers (1988). Sufficient

experimental conditions for the convergence of an adaptive observer for
nonlinear systems. IFAC
Bejing, China, 1762­1766.

Section 4.3

be :

Bastin G. and M. Gevers (1988). Adaptive observers for nonlinear time varying
systems. IEEE AC­33, n°7, 650­658.
Kreisselmeier G. (1977). Adaptive observers with exponential rate of
convergence. IEEE AC­22, n°1, 2­8.
Mareels I.M.Y., M. Gevers, G. Bastin and D. Dochain (1987). Exponential
convergence of a new error system arising from adaptive observers.
26^^ IEEE Los Angeles, 16­19.
Narendra K.S. and A.M. Annaswany (1989).
Prentice­Hall, Englewood Cliffs.

Section 4.4

The
:

Chen L , G. Bastin and D. Dochain (1990). Parameter identifiability of a class

of nonlinear compartmental models for bioprocesses. IFAC
Tallinn, USSR.
 250 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Chen L (1990). Identification and Control of Biological Systems. Doctoral

Dissertation, Univ. Cath. Louvain, Belgique.

Section 4.5

The e.g.
:

Goodwin G.C, B.C. Mclnnis and R.S. Long (1980). Adaptive control algorithms
for waste water treatment and pH neutralization. 3,
443­459.
Holmberg A. and J. Ranta (1982). Procedures for parameter and state
estimation of microbial growth process models. 18, 181­193.

Holmberg A. (1983b). A microprocessor­based estimation and control system

for the activated sludge process. In A. Halme (Ed.),
Pergamon, Oxford, 111 ­120.
Holmberg U. and G. Olsson (1986). Simultaneous on­line estimation of
oxygen transfer rate and respiration rate. In A. Johnson (Ed.),
Pergamon, Oxford, 205­209.
Ko K.Y., B.C. Mc Innis and G.C. Goodwin (1982). Adaptive control and
identification of the dissolved oxygen process. 18, 727­730.
Marsili­Libelli S. (1983). On­line estimation of bioactivities in activated sludge
processes. In A. Halme (ed.).
Pergamon, Oxford, 121­126.
Marsili­Libelli S., R. Giardi and M. Lasagni (1985). Self­tuning control of the
activated sludge process. 6 (12),
576­583.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 251

CHAPTER 5

ADAPTIVE CONTROL OF BIOREACTORS

5.0. Introduction

We are concerned in this chapter with the control of state variables in

bioprocesses. Roughly speaking, the purpose of control is to keep some
process state variable (or some function of the state variables) close to a
prespecified reference value, in spite of disturbances and variations in process
kinetics. In the special case where the reference value is constant, it is termed
a set point and control is called regulation. A computer controlled bioreactor is
schematically represented in Fig.5.1. The feedback loop is clearly apparent : it
involves sucessively the reactor, a set of sensors and measuring devices, the
computer and an actuator (here a valve). The closed loop system is defined
as the combination of the process and the controller within the feedback loop.
The process is thus controlled by the computer through the manipulation of the
feed rate of one external substrate. The task of the control algorithm,
implemented on the computer, is to determine, at each instant, the control
action, on the basis of information collected in real time by the sensors or
provided by the user via the keyboard. As we see in Fig.5.1, the regulator may
also incorporate information from state observers and parameter estimators
such as those which were described in Chapters 3 and 4. Our main concern
in this chapter will be the design of control algorithms for bioreactors.
Depending on the context, a control algorithm will also be called a control law,
controller or regulator. Before embarking on this, we would like to discuss
briefly the significance of control in bioreactors.
 252 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Monitoring screen

set J
point Keyboard Computer

Control Software
algorithm sensors
Control
action

exit
Concentrated gas ο
substrate

liquid
d o ο
water stream sampling
I J

bioreactor Sensors : analysers

and measurement
devices

Fig.5.1 Computer control system for bioreactors

Significance of controi in bioreactors

The control of bioreactors, in industrial applications, is usually restricted to the

regulation of pH and temperature at constant values which are supposed to be
favourable to the microbial growth. However, it should be clear (as we shall
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 253

demonstrate in this chapter) that significant performance improvements (in

terms of yield and/or productivity) may be expected from the control of the
biological variables themselves (such as biomass, substrate or synthesis
product). The following examples are typical.

1) Regulation of substrate concentration in the case where biomass growth is

inhibited by high substrate concentrations (see Section 1.3). The challenge
is twofold : first, feedback regulation may be needed to stabilise the process,
which is intrinsically unstable in open loop (see Section 1.9.3); secondly, a
good choice of the set point may contribute to the optimisation of the
biomass production as will be discussed in Section 5.7. Moreover, in
biological waste water treatment processes, regulation of the substrate
concentration is equivalent to controlling the pollution level in the effluent of
the plant.

2) Regulation of dissolved oxygen concentration. In aerobic bioprocesses this

concentration must be maintained sufficiently high to guarantee microbial
activity. But excessive aeration is undesirable from an economic viewpoint
(oxygen in excess is just lost to the atmosphere) and, in addition, can
induce biomass settling problems in activated sludge processes.

3) Regulation of product concentration (e.g. ethanol in yeast fermentation

processes). Fed­batch biotechnological processes are often characterised
by a conflict between yield and productivity (see Section 5.7). In many
instances regulation of the product concentration is an attractive method to
maintain the process at an operating point which corresponds to a good
trade­off between yield and productivity (see also Section 5.2.4).

4) Regulation of gaseous outflow rates : for example regulation of the methane

flow rate in a biomethanization plant in order to achieve a match between
gas production and energy demand (see Section 5.8).
 254 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Why controi ?

As we shall see throughout the chapter, the control design problem is solved
by appropriate algebraic manipulations of the general dynamical model (1.43)
of the process under consideration. The design, however, is carried out as if
the process kinetics were known exactly . But, as we have argued many times
in this book, since the kinetics are most often poorly known, they are replaced
in the control law by suitable on­line estimates provided by one of the various
parameter estimators which have been described in Chapters 3 and 4, or
others which will be specifically designed in this chapter. Such a controller,
equipped with parameter estimation (see Fig.5.1), is called an adaptive
controller because it has the potential to adapt itself to variations in the
kinetics.

Why not controi ?

Various conventional control techniques such as PID control or minimum

variance control have been used many times (even by the authors of this
book) to regulate biological variables in bioreactors. Several relevant
references are listed at the end of this chapter. However, these techniques will
not be presented here because, in our view, they suffer from the drawback of
being based on linear tangent model approximations (see Section 1.9.3) or
even on linear "black box" models.

Yet, as we have discussed at length in Chapter 1, we have a crucial prior

knowledge of the nonlinear dynamics of the processes, formalised in the
general dynamical model (1.43) and its variants. Therefore, improved control
performance can be expected from an exploitation of the nonlinear structure of
the model in solving the control design problem. In this chapter we shall follow
such a nonlinear control approach by repeatedly using a design technique
which is called Exact Linearizing Control. The difference between this
technique and conventional control lies in the way that linearization is
introduced in the problem. In a standard approach, one first calculates a
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 255

linearized approximation of the model, and then one designs a linear

controller iorXh\s approximate model. But, the closed loop remains non/Znear

Conventional Control

set point Linear Nonlinear output

Controller Process
i —

Nonlinear Closed Loop

Linearizing Control

set point Nonlinear Nonlinear output

Controller" Process

— r e

Linear Closed Loop

Fig.5.2. Conventional and linearizing control schemes

and is guaranteed to be stabilised only locally but not over a wide range of
operating points or around the transient trajectory in fed­batch operation. In
the exact linearizing control approach we obtain a nonlinear controller which
is precisely designed to achieve a linear closed loop \Nh\ch is unconditionally
stable whatever the operating point or the transient fed­batch trajectory. This is
summarized in Fig.5.2.
 256 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Outline of the chapter

The control objective and the principle of linearizing control are presented,
formalised and illustrated with simple examples in Section 5.1. Some basic
notations and definitions used in the sequel and an essential remark on the
closed loop stability are also given.

In Section 5.2 we address the linearizing control problem of bioreactors when

the kinetics φ (ξ ) are unknown but can be alleviated by a systematic application
of the singular perturbation technique. The theory is presented first. Then its
motivation is discussed and illustrated with three applications : depollution
control, propionate regulation in anaerobic digestion, and ethanol regulation
in yeast fermentation. Adaptive implementation is described and real life
experiment is reported.

In Section 5.3 we treat the adaptive linearizing control problem when the lack
of knowledge of the kinetics φ (ξ ) is explicitly taken into account (by
incorporating on­line parameter estimation in the control law) but under the
assumption that the yield coefficients are known. Two particular cases are
considered : substrate control and product control.

A general solution to the problem is also presented in Section 5.4 for a class of
CST bioreactors. The concept of relative degree is introduced.

In Section 5.5, we treat the adaptive linearizing control problem in the case
where both reaction rates φ (ξ ) and yield coefficients are unknown. The price
to pay, however, is that more restrictive structural assumptions are to be
considered.

In Section 5.6 we discuss some practical implementation aspects of the

proposed control laws.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 257

In Section 5.7 a first case study is carried out in detail : it concerns the adaptive
linearizing control of fed­batch bioreactors. The goal is to give a concrete
illustration of the design procedure and to emphasize the benefit to be
expected from using adaptive control in practical engineering applications, in
particular when a yield­productivity conflict occurs. A comparison with optimal
control is also presented.

In Section 5.8 a second case study is considered : the adaptive linearizing

control of gaseous outflow rates. It is an extension of the preceding adaptive
linearizing laws to the regulation of a nonlinear combination of the process
variables (not of one process component). We also introduce a modification of
the linearizing control algorithm in order to avoid possible division by zero.

5.1. Principle of Linearizing Control and Remarks on Closed

Loop Stability

In this chapter, as we have mentioned in the introduction, we shall concentrate

on a particular control strategy which is called state feedback linearizing
control and which is formalised in the present section.

We consider biotechnological processes described by the general dynamical

model (1.43) :

^ = Κ φ (ξ ) ­ Ο ξ + F ­ Q (1.43) = (5.1)

with dim(ξ ) = Ν and dim(9) = M.

The objective is to control a scalar output variable which is a measured linear
combination of the state variables :
 258 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Ν
V
Σ C\ (5.2)
i=1

where [ci^ C2 , C N ] is a vector of known constants.

The control input (denoted by "u", as is usual in control science terminology) is

the feed rate of one external substrate of the process (i.e. a substrate which is
introduced to the reactor from the outside). This is denoted as follows :

u = Fj for some i

F = bu + f (5.3)
b"^ = [bi,b2,...,bN] b| = 1, bj = 0 Vj^i

f^=[fl,f2.­..fN] fi = 0, fj = Fj Vj^^i

With this definition the model (5.1) is rewritten :

^ = Κ φ ( ξ ) ­ Ο ξ + bu + f ­ Q (5.4)

Throughout the chapter it will be assumed that f and Q are measured on­line
and that ξ is known on­line either by measurement or from an asymptotic
observer (Section 3.3).

The various notations and definitions are illustrated by an example.

Consider an anaerobic digestion process used for the depollution control.

Suppose that the dynamics of the process are suitably described by the model
(1.85.a­e), (1.88) :
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 259

dt "Χ Γ
=' 1 0 • + "ο ' ­ Ό " (5.5)
Si ­ki 0 .Φ 2. Si Fi 0
0 1 X2 0 0
X2
k3 ­k2 0 0
S2 S2
0 ke 0 Qi
.0. Q2
.P2. .k4 ks. .P2.

where X i is the acidogenic biomass, Si the organic substrate, X2 the

methanogenic biomass, S2 the acetate, Pi the methane and P2 the carbon
dioxide.

Clearly, the output pollution level (expressed in COD (Chemical Oxygen

Demand) units for instance) is defined as :

y = CiSi + C 2 S 2 (5.6)

where ci and C2 are the unit conversion constants from (g/l) to (g COD/I).

Suppose, in addition, that the influent feed rate Fi is the control input. It is
easy to check that, in this example :

CT = [0 ci 0 C2 0 0]
bT = [0 1 0 0 0 0]
f=0

5.1.1. Principle of linearizing control

The control objective is to track a reference output signal denoted y*(t). In the
special case of a constant reference, y*(= constant) is called the set point, and
control is called regulation.
 260 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

The principle of linearizing control is to find a control law υ (ξ , Q, f, y*) which is

a multivariable nonlinear function of ξ , Q, f, y* such that the tracking error (y*­y)
is governed by a prespecified stable linear differential equation called a
reference model.

Linearizing control design is a three­step procedure.

Step 1. The first task is to derive an input/output model (abr.: I/O model) by
appropriate manipulations (e.g. successive differentiations) of the
general dynamical model (5.4). This model takes the form of a δ ^^^
order differential equation:

d^
^ = f o ( t ) + u(t)fi(t) (5.7.a)
dt^

where δ is called the relative degree. The way to derive input/output

models will be illustrated many times in this chapter. At this stage, let
us just emphasise that:

­ depending on the application, fo(t) and may be highly complex

functions of ξ , Q, F and their successive derivatives;

­ as a consequence of the specific structure of the general dynamical

model (5.4), the input/output model (5.7.a) is //near with respect to
the control input u(t).

Step 2. A stable linear reference model of the tracking error (y*(t) ­ y(t)) is
selected as follows:

.i λ Η · ^ [ γ * ( ί ) ­ y(t)] = 0 λ ο =1 (5.7.b)
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 261

It is a model of how we want the tracking error to decrease (Fig.5.3).

The coefficients λ δ _] are arbitrary except that they have obviously to be
chosen so that the differential equation (5.7.b) is stable. The reference
model is independent of the particular process operating point.

y*(t)­y(t)

Fig. 5.3. Reference model of the tracking error

Step 3. Finally the control design consists of calculating the control action u(t)
such that the input/output model (5.7.a) exactly matches the reference
model (5.7.b). This is achieved by eliminating {d^y/d\^) between both
equations. The solution, as can readily be seen, is written:

Linearizing Control Law

δ ­1 d' d^v*
u(t) = ­fo(t) + Σ X5_~[y*(t) ­ y(t)] + ^ (5.7.C)
fi(t) \=o dt^
 262 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

In this chapter we shall examine three ways of solving the linearizing control
problem for bioprocesses described by the model (5.4). They are arranged
according to the complexity of the assumptions regarding the prior knowledge
available . We first give some simple examples.

Examples of linearizing control:

We consider the basic microbial growth process (1.33)

s—C\
for which the dynamical model (1.52)­(1.55) holds :

dX
= μ (Χ ,8)Χ ­ϋ Χ (5.8.a)
dt
dS
= ­k1μ (X,S)X­DS + F (5.8.b)
dt

Case 7. Suppose that we want to regulate the substrate concentration 8 = y at

a given constant value S*. Notice that (5.8.b) is an input/output model with
5=1. We select a first order reference model for the regulation error:

^ ( 8 * ­ 8 ) + λ ι (8*­8) = 0

which also implies (since d8*/dt = 0 because 8* = constant) :

X,(S­­S)=f

Now, by substituting (5.8.b) in tiiis equation, we obtain :

λ ι (8* ­ S) = ­ ί <ι μ (Χ ,8)Χ ­ DS + F

 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 263

The linearizing control law is the expression of the feed rate F which satisfies
this equation, that is :

F = λ ι (5* ­ S) + Κ ι μ (Χ ,8)Χ + DS (5.8.c)

Case 2. Suppose now that the feed rate F is expressed as follows (see
Section 1.5.5) :

F = DSin

and that we want to use the dilution rate D as the control action. It is
straightfonvard to see that the linearizing control law is written :

^ _ λ ι ( 8 * ­ 8 ) + Κ ι μ (Χ ,8)Χ
Sin"S

Case 3. Suppose, finally, that we want to regulate the biomass concentration

X = y at a constant value X* and the dilution rate is constant. Clearly the first
model equation (5.8.a) is not suitable for deriving the linearizing control law
since it does not explicitly express the connection between the feed rate F and
the biomass X (which is to be regulated). This apparent difficulty is solved as
follows. We differentiate (5.8.a) :

d^X _ 3φ dX ^ 3φ dS _ ^ d X
dtvdt; ax dt as dt dt

where φ (Χ ,8) = μ (Χ ,8)Χ as usual.

Then, substituting (5.8.a) and (5.8.b) for (dX/dt) and (dS/dt) in this equation, we
obtain :
 264 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

This equation is an input/output model, that is an explicit differential

relationship between the control F and the controlled output X. Because it is a
second order differential equation, we have also to select a second order
reference model for the regulation error :

d^ . d
— ( X * ­ X) + λ ι ^ ( X * ­ X) + λ 2(Χ *­ X) = 0
dt clt

or equivalently, since X* is assumed to be constant:

^ + λ , ^ = λ 2(Χ *­Χ )
dt^

Straightforward calculations then lead to the linearizing control law

5.1.2. An essential remark regarding closed loop stability

It is generally claimed in control textbooks that the main objective of feedback

control is plant stabilisation. It is very important to draw the reader's attention
on the fact that, in fed-batch biotechnological processes, the objective is
actually to destabilise the plant. Let us explain this apparent paradox.

A fed­batch bioreactor is, by definition, operated during a finite time. The goal
is generally to exponentially accumulate some reaction product, which is
harvested at the end of the operation. In mathematical terms, this means that
some state variable follows an exponentially growing trajectory, which is
characteristic of an unstable behaviour. As a matter of fact, in many instances,
this exponential trajectory can be made more or less optimal by regulating
some other state variables at appropriate reference values. Hence the goal is
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 265

clearly not to stabilise the process globally but rather to keep an unstable
trajectory under control. This issue, which has also been pointed out by other
authors (see the bibliography), will be illustrated in depth in Section 5.7.

In contrast, it is obviously clear that, for CST bioreactors, the global stability of
the closed loop is a primary requirement which has to be checked in advance
by the control designer. Unfortunately, depending on the particular form of the
kinetics, the same method of designing a linearizing controller may produce a
stable as well as an unstable closed loop. Furthermore, checking the stability
beforehand with the aid of a process model is hazardous in our context of
minimal modelling of the kinetics, since our aim is precisely to propose
adaptive feedback control strategies when the kinetics are poorly known and
hence when a full process model is not available. Our best recommendation is
therefore to experiment very carefully with new controllers on bioreactors.

Let us illustrate this point with an example.

Example

Consider the basic microbial growth process (5.8.a­b) with the linearizing
control law (5.8.c) to regulate the substrate concentration S at the constant set
point S*.

Suppose first that the specific growth rate μ (3) depends only on the substrate.
Then it can be easily checked that, if μ (8*) > D, there is no equilibrium state in
closed loop. The dynamics of the biomass growth tend asymptotically to :

^ = WS*)­D]X > 0

Hence the biomass concentration is exponentially increasing without limit. If

the process operates in the fed­batch mode, with a view to biomass
production, this unstable behaviour is exactly the desired objective. But if the
 266 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

process operates in the continuous mode this is obviously a completely

unacceptable behaviour.

On the other hand, if the specific growth rate μ (Χ ,8) depends on both biomass
X and substrate S, then the closed loop has an equilibrium state defined by :

S = S* and X such that μ (Χ ,5*) = D

which can even be globally attractive, depending on the shape of the function
μ (Χ ,8).

Hence we see that the same control law applied to the same process may give
either a globally stable or a completely unstable behaviour: it is just a matter of
how the kinetics depend on the biomass concentration, an issue which is very
poorly known in most practical applications.

5.2. Singular Perturbation Design of Linearizing Controllers

In this section we shall establish a general method for the design of linearizing
controllers based on so­called "fully reduced" models which are obtained by a
systematic application of the singular perturbation technique (Section 1.8).
The theory is presented first. Its biological motivation is then discussed and
illustrated with three typical examples.

5.2.1. Theory

A "fully reduced model" is defined as follows :

a) there exists a state partition ξ "^= [ξ ^, ξ [], κ "^= [KJ, Κ [] with dim(ξ f) = Μ and Kf
full rank. The components of are referred to as the "slow" components
and those of ξ i are the "fast" components.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 267

b) the dynamics of ξ ι are supposed to be sufficiently fast to allow singular

perturbation reduction to a set of algebraic equations, so that the model is
written :

= Ks9 ­ + Fs ­ Qs (5.9.a)
dt

Κ ,φ + F, ­ Q, = 0 (5.9.b)

c) the controlled output y is a combination of the slow states only :

y= cIξ s (5.9.C)

Since Kf is supposed to be full rank, it follows from (5.9.b) that the vector φ (ξ ) of
reaction rates can be written :

φ (ξ ) = ­Κ Γ ν Ρ ,­α ,) (5.10)

Substituting (5.10) into (5.9.a) gives :

^ = ­0ξ 3 + [ΐ Ν _Μ ­KsKr^(F­Q) (5.11)

Denoting J = [IN­M ­ K s K T ^ , the dynamics of y are written :

dy
= ­ D y + C^J(F ­ Q) (5.12)
dt

We now turn to the control design problem. Substituting (5.3) into (5.12), we
get the following input/output fully reduced model:

dy
^ = ­Dy + C^Jbu + C^J(f­Q) (5.13)
dt

We suppose that:
 268 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Cs J b ?t 0

We select a first order reference model for the tracking error (see the definition
(5.7.b)):

^ ( y * ­ y ) + Xi(y*­y) = 0 (5.14)

According to the principle of linearizing control, we have to find a control law

u(y, y*, f, Q) such that the linear reference equation (5.14) is satisfied (with
(dy/dt) given by (5.13)). This control law is readily obtained by substituting
(5.13) into (5.14), as follows:

Linearizing control law by singular perturbation design

u(y, y*, f, Q) = (Cl J b)"^ + λ 1(y*­y) + Dy + C^J(Q­f) | (5.15)

We notice that this control law involves a feedforward compensation of the

feed rates f, which are not used to control the process.

The theory will now be illustrated by three examples.

5.2.2. Depollution control in an anaerobic digestion process

The process is described by the model (5.5)­(5.6) which is clearly not "fully
reduced" although it already results from a reduction by singular perturbation
(see Section 1.8.4). In order to perform the full reduction, we select the
following state partition :
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 269

ξ 3 = Xil 4f=rP
Pii l (5.16)

Si LP2J
X2

S2.

We notice that the "slow" state involves the two substrate concentrations Si
and S2, a combination of which is to be controlled, since (5.6):

y = C1S1 + C2S2 (5.6) = (5.17)

The induced partition of Κ is as follows :

Ks = 1 0 0 kg (5.18)
­ki 0 Lk4 ksJ
0 1
Lka ­K2J

We observe that Kf is invertible :

­ks 1
(5.19)
k4k6 k4

In this particular situation, equation (5.10) is then written :

­ks 1
"Qi" Qi (5.20)
K4K6 K4
.92j _Q2_ LQ2J

Substituting (5.20) in the first four equations of the model (5.5) leads to the
"fully reduced model" (compare with (5.11)) :
 270 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

d_
= ­ D •χ Γ +Ό " +" 1 0 "
dt "Χ Γ
Q;
k4ke k4

Si Si Fi ­ki 0 LQ2.

X2 X2 0 0 1
.S2. .S2. .0. ­k3 ­k2.

J(F­Q)

We verify that CsJb = ci ^ 0.

The equation expressing output dynamics is then written as follows (compare

with (5.13)) :

dv
Qi (5.21 .a)
• ^ = ­ D y + CiU + [ei Θ 2]
LQ2J
i
C|jb CsJ(f­Q)

with :

Δ Cikik5­C2(k3k5 + k2k4)
(5.21 .b)
k^k^

Δ Ciki +C2k3
θ 2= , (5.21 .C)

Hence, the linearizing control law (5.15) is specialised as follows :

u = Fi = C i + λ 1(y*­y) + Dy ­ θ ι Ο , ­ 0202 (5.22)

It is worth noting that this control law not only makes use of the measured
output y (which is the pollution level we want to control) but also of the other
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 271

available on­line measurements, namely methane gas Qi and carbon dioxide

Q2 flow rates.

5.2.3. Propionate control in an anaerobic digestion process

Motivation

Anaerobic digestion is well known to exhibit unstable equilibrium states (see

Section 1.9.3) and feedback control is therefore required to stabilise the
process. In Section 5.2.2, a control algorithm has been designed with a view to
depollution control. Our purpose, in this section, is to show how linearizing
control can be designed with a specific objective of stabilization. The issue of
analysing the sources of instability of anaerobic digesters has received much
attention in the scientific literature (see the references). Propionate
accumulation is often described as one of the main symptoms. This implies a
twofold hypothesis regarding the process model :

a) propionate is an inhibiting factor for the growth of methanogenic bacteria;

b) propionate degradation is considered to be a critical limiting reaction for the

process (otherwise propionate accumulation would not be possible).

It is therefore relevant to try to stabilise the process by regulation of the

propionate concentration in the reactor. And the fact that propionate
degradation is presumed to be limiting makes the fully reduced model
approach adequate for control design.

Γ Λ β fuiiy reduced modei

We consider that the process is described by the model (1.65) and we select
the following set of fesf state variables :

ξ ί = [ 8 ι . 8 2 . S4, P i ] T
 272 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

where we recall that Si is the influent organic substrate, S2 is the acetic acid,
S4 is the hydrogen and Pi is the methane.

The important point is that the propionate S3 is not involved in ξ ι : propionate

is considered as a "slow" state according to the assumption that propionate
degradation is limiting.

The corresponding matrix Kf is as follows :

K,=| ­^21 0 0 0 • (5.23)

­k42 k43 0
kei 0 kaa ­k84
0 ko2 0 ko4 .

The controlled output is :

(5.24)

while the control input is :

u = Fi (5.25)

Then the (fairly tedious) calculations of the fully reduced model lead to the
following input/output dynamics if Q3 is assumed to be negligible :

dy
= ­Dy­i­eiQi +e2U (5.26.a)
dt

with :

^ ko2k84k43^ ^ 'k84^
Θ ι =­Κ 63 ksa + (5.26.b)
^04^42 1^04/
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 273

^ . (. *^02^84^43 ^ ^ ( ^ 8 1 ^02^QΛ ^4^

^2 = 1^— + K53 + — 1 I Iτ — + (5.26.C)
2^ Ι <04^42 J ν ^2ΐ k2iko4k42J

From (5.26), the control law is derived as follows :

(dyVdt) + X,(y*­y) + D y ­ 9 i Q i
u = Fi = 5 (5.27)
«2

5.2.4. Ethanol regulation in a yeast fermentation process

Fed­batch yeast fermentation processes are known to be characterised by a

conflict between yield and productivity. This can be roughly explained as
follows. A high glucose feed rate will generally induce high biomass
productivity but with a low yield due to the additional production of ethanol
which inhibits glucose respiration. The converse is obviously true for low
glucose feeding. Simulation studies have shown that ethanol regulation in the
process (with glucose feed rate as control action) allows to maintain the
process to be maintained at operating points which correspond to a good
trade­off (from an economic viewpoint) between yield and productivity. This
section is therefore devoted to the design of a linearizing regulator of the
ethanol concentration.

We suppose that the process is described by the model (1.64) and we select
the following set of "fast" state variables :

ξ ί =[8,0,ρ Γ (5.28)
 274 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

with S being the glucose concentration, C the dissolved oxygen concentration

and Ρ the dissolved CO2 concentration.

We notice that ethanol concentration E, which is the state variable to be

regulated, does not belongs to ξ ι . The corresponding matrix Kf is as follows :

Kf = ­ k i ­k2 0" (5.29)

­ks 0 ­ke

L kg kg J

The controlled output is y = Ε while the control input is u = F i . The calculations

lead the following fully reduced input/output model :

dy_
= ­Dy + θ ι Ο ι + θ 2 α ΐ η + θ 3Ρ ι (5.30.a)
dt

with Qi the CO2 gaseous outflow rate, Qjn the oxygen transfer rate and :

rkik4 kgkg
θ ι = (5.30.b)
ν ^2 kg '"

fkik4k9 kskskg^
θρ = (5.30.C)
k2k6

fk­|k4k8 k3k5k8 k4
03 = (5.30.d)
k2k6
4
kskg k^ke^
Θ = k7­ (5.30.e)
~ϊ <6 kT";

Control law

From (5.30), the linearizing control law (5.15) is written

 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 275

d/
+ λ ι (χ *­ y) + Dy ­ Q^Q^ - 02Qin} (5.31)
I dt

5.2.5. Direct adaptive linearizing control

The practical implementation of the control law (5.15) requires the knowledge
of the yield coefficients which are present in the matrix J = [IN_M ­KsKf^]. It
may arise, however, that some of these coefficients are either badly known or
time­varying and, hence, that the linearizing controller is not well adapted to
the process. The remedy offered by control science in that case is known as
Adaptive Control. It is a design methodology which provides controllers or
regulators capable of adapting themselves to modelling uncertainties.

We consider the nonzero unknown components of the vector CsJ as a set of

unknown parameters, denoted Θ . This means that CsJ can be factorised as
follows :

CgJ = e^Jo (5.32)

with Jo being a known matrix.

The model (5.13) is then rewritten :

dy
= ­ Dy + e"^Jo(bu + f ­ Q) = ­ Dy + θ "^φ (Ρ ,α ) (5.33)
dt

Δ
where φ (Ρ ,Ο ) = Jo(bu + f ­ Q) is a so­called linear regressor.

Examples of such parametrization have been given above for anaerobic

digestion and yeast fermentation processes (equations (5.21), (5.26), (5.30)).

This implies that the linearizing control law (5.15) is reformulated as :

 276 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

u = (e'^Job) ^ + λ ι (y*­ y) + Dy + e^Jo(Q ­ o l (5.34)

V at y

An adaptive control law is then obtained by replacing the true unknown

parameter θ in (5.34) by a time­varying parameter value e(t) calculated on­line
with a so­called adaptation law. The adaptive control law is written :

Adaptive linearizing control law

^ = h(y.F,Q) (5.35)

u = (e"^Job)"^ + λ ι (y*­ y) + Dy + e'^Jo(Q ­ o l (5.36)

\ Qt

Equation (5.35) is the adaptation law. The design problem consists of

selecting a suitable function h(y,F,Q) for the adaptation law, which guarantees
the closed loop stability of the control system despite the lack of knowledge of
the true parameter Θ . So called direct and indirect adaptive algorithms can be
formulated. In a direct scheme the parameter adaptation is driven by the
tracking error (y*­ y). In an indirect scheme it is driven by an auxiliary
observation error. The direct scheme obtained from a Lyapunov design is now
presented. The indirect one is the subject of the next section.

Lyapunov design

The direct adaptive control technique is based on Lyapunov stability theory

(see Appendix 2).

We thus suppose that an adaptive control law (5.35)­(5.36) is applied to the

process (5.33). We select the following candidate Lyapunov function :

W(t) = ­^[(y*­ yf + (Θ ­ θ )"^ (Θ ­ Θ )] (5.37)

 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 277

where Γ is an arbitrary positive definite matrix.

Taking the time derivative of W(t) and using (5.33) and (5.35), we obtain :

^ = ­ λ 1 ( y * ­ y)^­ (Θ ­ Θ )V'h(y,F,Q) ­f φ (Ρ ,Ο ) (y*­ y)] (5.38)

If the adaptation function h(y,F,Q) is chosen as follows :

h(y,F,Q) = ­Γ φ (F,Q) (y*~y) (5.39)

and if we substitute (5.39) into (5.38), we find that dW/dt is semi­negative

definite along the system trajectory since it reduces to :

(5.40)

Thus the direct adaptive control law (5.35)­(5.36) is written as follows :

Direct adaptive linearizing control law

^­­Γ φ (Ρ ,0)(ν *­ν ) (5.41 .a)

u = (eJobf^ + Xi(y*­y) + Dy + §"^Jo(Q­f)' (5.41.b)

Example : Depollution control

If we apply the above theory to example 5.2.2 with a diagonal matrix

r=diag{y.|, Yg}, straightforward calculations lead to the following adaptive
counterpart of the control law (5.22) :
 278 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

­ ^ = YiQi(y*­y) (5.42.a)

Ο θ 2
= Y2Q2(y*­y) (5.42.b)
dt

u = c j ^ f ^ + Xi(y*­ y) + Dy ~ Q^Q^ - (5.42.c)

Although the theory may seem rather involved to those who are not familiar
with adaptive systems, it is worth noting that it leads to a very simple control
structure in practice, as is illustrated by this example.

5.2.6. Indirect adaptive linearizing control

Indirect adaptive control is based on the idea of using a parameter estimator

as the adaptation law. The two different estimators (observer­based and linear
regression), which have been presented and discussed several times in this
book, can be used. In this section we analyse in detail the design of an
adaptive controller with an observer­based estimator as the adaptation law.
The use of linear regression (least squares algorithm) will be illustrated on the
basis of the experimental application of Section 5.2.8.

An observer­based estimator, similar to (3.70), is written as follows from the

fully reduced model (5.33) :

^ = ­Dy ­f θ "^φ (Ρ ,α ) ­ ω (y ­ y) y(0) = y(0) (5.43.a)

^=^(F,Q)(y­y) (5.43.b)

with ω and γ being scalar design parameters (which play the same role as the
matrices Ω and Γ in (3.70)).
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 279

It is interesting to notice that the adaptation law (5.43.b) is quite parallel to the
preceding one (5.41 .a), the only difference being that the tracking error (y*­ y)
is replaced by the observation error (y ­ y).

The corresponding adaptive control law is then just obtained by using (5.36)
but with θ provided by (5.43.b).

Indirect adaptive linearizing control law

^ = ­Dy + eVF,Q)­ro(y­y) y(0) = y(0) (5.43.a)

(5.43.b)

u = (e^Job)"^ + Xi(y*­ y) + Dy + q\(Q - f)' (5.36)

The stability analysis of indirect adaptive control is carried out under the
following realistic assumptions.

A.5.1. The reference signal y*, its time derivative dyVdt and the measured
signals F and Q are continuous bounded functions of time t.

A.5.2. The initial parameter estimate θ (0) belongs to a closed ball Be c: iRdime^
centered on Θ , such that each GgBq has all nonzero components.

This latter assumption, which is fairly technical, is illustrated in Fig.5.4.

For the parameter estimator (5.43), under Assumptions A.5.1,

A.5.2,
 280 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

a) the observation error e = y ­ y a n d the parameter error θ = θ ­ θ are

bounded as follows :

e(t)|< Ι θ (θ )| Ι θ ω 11<||θ (0)|| vt>o (5.44)

b) the observation error converges asymptotically to zero

lim |e(t)| = 0 (5.45)

Proof (Lyapunov analysis). The following candidate Lyapunov function is

considered :

W(e,e) = + θ ^θ ) (5.46)

The time derivative of W along the trajectories of (5.43) is easily shown to be :

dW 2
(5.47)
dt = ­(ΰ γ β

which is clearly a semi­negative definite function. Hence W(e,e) is a positive

decreasing function. Since e(0) = y(0) ­ y(0) = 0 (see 5.43.a), we have :

w(e(t),e(t))<w(e(o),e(o))­ I e(o) (5.48)

Part a) follows.

On the other hand, (5.47) and (5.48) also imply :

| e ( T ) r d x < W ( 0 ) ­ W ( t ) < |θ (0) (5.49)

•0

Prom Assumption A.5.1 and (5.44), φ (Ρ ,Ω ) and hence (de/dt) are bounded.
Therefore e(t) is continuous and (5.49) implies part b).
QED.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 281

Fig.5.4. Be

It can now be shown by straightforward calculation that the adaptive control

law (5.36) with the parameter estimator (5.43) applied to the system (5.33)
leads to the following expression of the closed loop dynamics :

^ ( y * ­ y ) + λ 1(y*­y) = • ^ + ω e (5.50)

That is the tracking error (y* ­ y) is the output of a stable linear filter driven by
the observation error e. Then the convergence of e to zero (Lemma 5.1)
necessarily implies the convergence of (y*­ y) to zero as stated in the following
theorem.

Theorem 5.1. For the adaptive control law (5.36) with the parameter estimator
(5.43) applied to the system (5.33), under Assumptions A.5.1 and A.5.2, the
tracking error (y*­ y) converges asymptotically to zero :

lim|y*(t)­y(t)| = 0 (5.51)
 282 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Example: Depollution control

If we apply the foregoing theory to Example 5.2.2, we obtain the following

adaptive control law :

dy
= ­Dy ­I­ CiU ­I­ e^Qi ­i­ &2Q2 + ω (y ­ y) (5.52.a)
dt

de
(5.52.b)
dt­ = 7 Q i ( y ­ y )

d&s
= 7Q2(y­y) (5.52.C)
"dT

dy*
U = Ci^ + Xi(y*­ y) ­I­ Dy ­ θ ι Ο ι ­ &2Q2 (5.52.d)
V dt

This indirect control law has to be compared to that obtained by the direct
method (5.42).

5.2.7. Using the dilution rate D as control action

So far, we have supposed that the process control input u = Fj (see (5.3)) is the
feed rate of some external substrate which is introduced into the reactor
independently of the dilution water stream. An alternative (see Fig.1.10),
which is common in industrial practice, is to suppose that the feeding substrate
is diluted in the water stream and to use the dilution rate as the manipulated
variable to control the process (Fig.5.5). in that case, the derivation of the
control algorithm is a direct consequence of our previous theory. Indeed, the
I/O fully reduced model, (5.12), is simply rewritten :

^ = D(C|jSj,­y)­C^JQ (5.53)
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 283

where F = DSjn according to (1.62).

computer

contro
concentrated action
substrate u=D

Tlutionl^
tank Sensors

Fig.5.5. Computer control system for bioreactors with the

dilution rate as the control action

The principle of linearizing control then produces the following control law
(compare with 5.15) :

Linearizing control law witti D as control action

D = (C^JSin ­ y)"' + Xi(y*­ y) + C^JQ (5.54)

or, using the parametrization θ (5.32):

 284 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

D = (G^JoSin ­ y)"^ + λ ι (y*­ y) + e'^JoQ (5.55)

Adaptive versions of (5.55) are simply obtained by replacing θ by the

parameter estimates θ provided either by the direct adaptation law (5.41.a) or
the observer­based adaptation law (5.43).

5.2.8. Experimental application. Adaptive regulation of propionate

concentration in an anaerobic digestion pilot plant

Adaptive regulation of propionate concentration, as described in Section 5.2.3,

has been implemented on a pilot­scale anaerobic digestion process by the
unit of Bioengineering (Universite Catholique de Louvain, Belgium). The
dilution rate D was used as control action so that (as explained in the previous
paragraph), the control law (5.27) was rewritten :

^^x,(S3­S3)­e,Q, ^^^^

where S3 is the propionate regulation which is to be regulated at the set point

S3, Sin is the influent substrate concentration and Q i is the methane gas
outflow rate.

The process has been operated during more than 150 days under an indirect
adaptive version of the control law (5.56). However the parameter adaptation
was applied only to the parameter Θ 2 because the value of θ ι was known
from a preliminary identification study :

θ ι =­0.94gr.COD/lit (5.57)

The adaptation of Θ 2 was performed by a discrete­time least squares

estimator, analogous to (3.87), written as follows :
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 285

θ 2.ί ­1 = θ 2 , ί + TYtSin,AiS3.t+1­S3,t­T[eiQi,t­ DtS3,t+ Os.tDtSin.tl>

yM = — (5.58) (5.59)
λ ­f (TDtSin.t)Yt

The forgetting factor λ was set to 0.9 while a sampling period Τ = 2 days was
used. The adaptive controller was thus constituted by the control law (5.56)
rewritten in discrete time as follows :

combined with the parameter adaptation law (5.58)­(5.59).

An excerpt of the experimental results is shown in Fig.5.6 from which the

following conclusions are drawn :

1) During the first part of the experiment (until the 65^·^ day), we observe an
excellent behaviour of the control system which keeps the actual propionate
concentration S3 very close to the set point. The regulation is robust against
a big step of influent organic matter on day 38, from 40 g COD/I to 50 g
COD/I (Fig.5.6.a).

2) On day 66, the regulator has been switched off and the process operates in
open loop with a slightly increasing dilution rate. We see that the process is
rapidly destabilised and that propionate is exponentially increasing.

3) On day 82, the regulator is switched on again and brings the propionate
concentration S3 back to the set point.

4) The control law (5.60) takes advantage of the on­line measurement of the
methane gas flow rate Q i (Fig.5.6.d) which gives valuable on­line
information on the biomass activity.
286 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

0.04

0.02H

time (days)

Fig.5.6. Adaptive regulation of proprionate concentration

in an anaerobic pilot plant.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 287

5.3. Adaptive Linearizing Control (Known Yield Coefficients)

in Section 5.2 we have discussed the design of linearizing controllers for

bioreactors when a (fairly drastic) model reduction is authorised (on the basis
of available prior knowledge).

In the present section we are concerned with adaptive control in bioreactors

when such a model reduction is not desirable but, however, the yield
coefficients are known (either from stoichiometry or preliminary parameter
identification, see Chapter 4). The case where the yield coefficients are
unknown will be treated in the next section.

5.3.1. Statement of the adaptive control problem

We consider biotechnological processes described by the state space model

(1.49):

^ = Κ Ο (ξ )α " Ο ξ + F ­ Q (1.49) = (5.61)

where it is recalled that 6(ξ ) is a diagonal matrix depending only on the

reactants in the bioreactor, while α (ξ ,t) is a set of time­varying, state dependent
specific

As in Section 5.2, we suppose that the objective is to control a scalar output

variable y = Ο ^ξ by using some external substrate feedrate as control action.
Using (5.3) it is then convenient to rewrite the model (5.61) as follows :

^ = Κ Ο (ξ )α ­ Ο ξ + bu + f ­ Q (5.62)

: throughout this section, we assume that :

­ f and Q are measured on­line;

 288 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

­ ξ is known on­line either by measurement or from an asymptotic

observer (Section 3.2);
­ the specific raction rates α are unknown.

Crucial comment:

The reader is advised to notice that, in this section (unlike in the previous one)
we suppose from the beginning that the process model is parametrized by a
set of unknown parameters a. This motivates an immediate search for
adaptive controllers, i.e. controllers which are able to cope with parameter
uncertainty and nonstationarity.

The control design problem

The problem is to find an adaptive control law u(ξ ,f,Q,ά ), which is a function of
the measurements ξ , f, Q and of on­line estimates ά , such that the dynamics of
the tracking error (y*­ y) are governed, at least asymptotically and in the
neighbourhood of the reference trajectory y*, by stable linear dynamics. This
is called

The derivation of a general solution to this problem is extremely involved and

has not yet received a complete answer for nonlinear systems in general.
Therefore, we shall be content to examine three particular situations, which
nevertheless cover a wide range of potential applications.

The first situation, presented in Section 5.3.2, is termed "substrate control"

because it relies on the assumption that one of the components involved in the
output y is precisely the external substrate which is used as control input.

The second situation, presented in Section 5.3.3, is termed "product control"

because it relies on the assumption that the output y involves only liquid
products (including internal substrates).
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 289

Finally, the third situation concerns a particular class of continuous stirred tank
bioreactors. It will be presented in Section 5.4.

5.3.2. Substrate control

Here we suppose that one of the components involved in the controlled output
y is precisely the external substrate which is used as control input. That is the
vectors b and C are such that: C^b Φ 0.

From the state space model (5.62), we readily derive the following I/O
dynamics :

^ = θ "^Κ Ο (ξ )α ­ Dy + c"^bu + f ­ 0 (5.63)

A straightforward application of the linearizing control principle leads to the

following control law :

U= (CV^ ^ + Xi(y*­ y) ­ Ο ^Κ Ο (ξ )α + Dy ­ C^(f ­ Q) (5.64)

which necessarily exists, since C"'"b ^ 0 by assumption, and which guarantees

the stable linear dynamics of the tracking error:

^ ( y * ­ y ) + λ 1(y*­y) = 0 (5.65)

if the parameter α is known. But our hypothesis is precisely that α is unknown.

Then (5.64) is transformed into an adaptive control law by using one of the
parameter estimators which have been proposed in Chapter 3. For instance
the linear regression estimator (3.93), with a forgetting factor λ :
 290 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

= ­<0ψ '^­Ι ­Κ Ο (ξ ) (3.93.a) = (5.66.8)

dt

= ­ω ψ ο + ( ω ­ 0 ) ξ ­ α ­ ι ­ Ρ (3.93.b) = (5.66.b)
dt

da _
(3.93.C) = (5.66.C)
— = Γ ψ (ξ ­ψ ο ­ψ α )
Γ (0) >ο (3.93.d) = (5.66.d)
dr
— = ­Γ ψ ψ '''Γ ­Ι ­λ ψ
The full adaptive controller is thus the combination of the control law (5.64)
and the parameter estimator (5.66) with α substituted by ά .

Example: Depollution control

Again, we come back to Example 5.2.2. The process model is as follows :

= " 1 0 1'X1S1 0 ' ­OrXi 0 0 (5.67.a)

dt •Χ ι "
Si ­ki 0 0 X2S2 .«2. Si Fi
- 0
X2 0 1 X2 0 0
k3 ­k2 0 0
S2 S2
0 ke 0 Qi
Pi Pi
k4 k5_ 0. .Q2.
P2.

•CiS 1 + C2S2 (5.67.b)

We notice that in this example :

C'^b = Ci ^ 0 (5.68)

The I/O dynamics are written (compare to (5.63)):

 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 291

dy
= [(C2K3 ­ CIKI) XiSi ­ C2K2X2S2] «1 ­ D y + CiU (5.69)
dt
«2.

SO THAT THE LINEARIZING CONTROL LAW IS WRITTEN :

U= ^ + Λ Ι (y*­ y) + Dy ­ (C2K3 ­ C2KI) X I S I Α Ϊ + C2K2X2S2A2 (5.70)

Practical implementation

Let us suppose that the output pollution level y, the acetic acid concentration
S2 and the methane gas flow rate Q i are the only measurements which are
available on­line. We shall see how to implement an adaptive version of (5.70)
by using the algorithmic tools which have been proposed in the preceding
chapters. First of all we have, from (5.67.b):

y ­ C2S2
SI = (5.71)

Then we use the asymptotic observer (3.48) to compute on­line estimates of

Xi and X2 :

DZ2
= ­D2, (5.72.a)
dt
d^a
= ­DZ3­QI (5.72.b)
dt

Xp = ­ (5.72.C)
KE

^1­17" i>2 ­ ^^2 + I T ^3 (5.72.d)

 292 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

We then use the parameter estimator (5.66) to compute on­line estimates of

ά ι and 0.2.

As we mentioned in Chapter 3, we can apply this parameter estimator to a

subset of the state equations, provided they involve the two unknown
parameters a i and 02. We select the dynamics of Si and S2 :

dS
= ­ k i X i S i a i ­ DSi ­1­F1
dt
dSg
= k3X.|S­)(X­| — k2X2S2CC2
"dT

Then the algorithm (5.66) is specialised as follows :

dvi
= ­ω ψ ι + X i S i (5.73.a)
dt
d¥2
= ­ω ψ 2 + X2S2 (5.73.b)
dt

= ­ω ψ ο ι ­Η (ω ­ D) Si ­ι ­ Fi (5.73.C)
dt
d¥o2
= ­ω ψ ο 2 + (ω ­ D) $2 (5.73.d)
dt
d . . ­,
= rT­kiyi kgYi Si ­ Ψ 0 1 +Ι <ι Ψ ι ά ι (5.73.e)
dt α ^Ί = 1 ­κ
^2} [ Ο ­k2y2lS2­Yo2­k3Viai+Ι <2Ψ 2ά 2]
dr (5.73.f)
dt = ­Γ ψ ψ '''Γ ­ι ­λ Γ

with :
­ki¥i 0 (5.73.g)
[kgyi ­k2V2j

Finally, the full adaptive controller is made up of the combination of (5.71),

(5.72), (5.73) with the control law (5.70) rewritten :
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 293

u= ^ + y) + Dy ­ (C2k3 ­ C2ki) X^Siai + C2k2X2S2a2 (5.74)

5.3.3. Product control

We suppose now that the output y = Ο ^ξ involves only liquid products

(including internal substrates as we shall see in the example hereafter). We
notice immediately that this implies :

­ Q ) = 0 and, in particluar, c"^b = 0 (5.75)

We assume in addition that at least one of these products is produced by a

reaction to which the control substrate also belongs. This is formalised by
assuming that the following quantity is not identically zero :

c•^K^[G(ξ )α ]b;^0 (5.76)

Under these assumptions, it can be observed that :

^ = C'^KG(ξ )α ­Dy (5.77)

This equation does not involve the control input u so that it is useless for the
design of a linearizing control law. Now, if we differentiate ( 5 . 7 7 ) , we obtain,
thanks to (5.75) :

y ­ D [θ '^Κ Ο (ξ )α ­ Dy] + θ '^Κ Ο (ξ )­

+ C ^ K ^ [0(ξ )α ] {Κ Ο (ξ )α ­ Dξ + bu + (f ­ Q)} (5.78)

Clearly, this equation now contains an explicit relation between u and y. To

simplify the notation, (5.78) is rewritten :
 294 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

2
^ = 9 ο ( ξ . « . ^ ) + 9ι (ξ .α ) (f ­ Q) + ς ι (ξ .α )6υ (5.79.a)

with

9ο (ξ .α .^) = ­ ­ 0[Ο ^Κ Ο (ξ )α ­ Dy] + θ ' ^ Κ ^ [ 0 ( ξ ) α ] { Κ α ( ξ ) α ­ Ο ξ }

(5.79.b)

9ι (ξ ,α ) = θ '^Κ ^[Θ (ξ )α ] (5.79.C)

We select a second order linear reference model

• ^ ( y * ­ y) + λ 2 • ^ ( y * ­ y) + λ 1(y*­ y) = 0 (5.80)
dt^ dt

The principle of linearizing control requires the calculation of the control law
u ί ξ , α , ^ l in such a way that (5.80) is satisfied with dy/dt and d 2 y / d t 2 .
V dt y
respectively, given by (5.77) and (5.78). The solution is as follows :

ξ ,α ,^]=[g1(ξ ,α )bΓ '

­ λ 2(θ '^Κ Ο (ξ )α ­ Dy) ­ 9 ( ( ξ . α . · ^ ) " 9ι (ξ .α )(ί ­ Q)] (5.81)

We see that this control law not only depends, as before, on the unknown
value of the specific reaction rates a, but also on their derivatives da/dt. This
apparent difficulty, however, is easily resolved : we replace α by an on­line
estimate ά , and da/dt by da/dt (see 5.66.c) in order to make the control law
adaptive.
 Chap.5. ADAPTIVE CONTRCDL OF BIOREACTORS 295

Example : Ethanol control In a yeast production process

In Section 5.2.4, we derived a linearizing controller based on a fully reduced

model of the yeast fermentation process. We shall now derive an alternative
controller using the theory presented above. The process dynamics (1.54) are
as follows :

d_rv.n

dt
Γ
1 1 1 " XSC0 0 Τ α ι ' ­D X' + • 0 ­• " 0 •
­ki -k2 0 0 XS 0 «2 S
0
0 k4 ­k3 0 0 XEC­L«3j Ε
0 0
­ks 0 ­ke C Qin 0
L kg kg _
.P.
0 .Qi.
(5.82)

The goal is to regulate the ethanol concentration :

γ = 0Τ ξ = Ε withCT=[001 00] (5.83)

The control action is the glucose feed rate Fg:

u = Fg i.e. bT=[01 0 00] (5.84)

According to (5.75), we verify that C'''(F ­ Q) = 0 and C^b = 0, and also that,
according to (5.76) : c ' ^ K ^ [ G ( ξ ) α ] b = k4Xa2 ^ 0.
Equations (5.77) and (5.78) are specialised as follows :

dE
= k4XSa2 ­ kaXECos ­ D E (5.85)
dt

j2.
^ = g o ί ξ . α . 7 ? V g 1 ( ξ . α ) [ f ­ Q + bu)] (5.86.a)
dt^ V dt y

with :
 296 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

with
( dα ^
go ξ , α , — = (a2k4S ­ aakaEC) (XSCa^ + XSa2 + XECag ­ DX)
V at y
a2k4X ( ­ kiXSCai ­ k2XSa2 ­ DS)

­ a3k3XE ( ­ ksXSCtti ­ keXECoa ­ DC)

­ ( D + a3k3XC)(a2k4XS­ a3k3XEC­DE)

dap dap dD
+ k4XS — ^ ­ k 3 X E C — ^ ­ Ε ^ (5.86.b)
dt dt dt

g1(ξ ,α )[f ­ Q + bu] = a2k4XFg ­ a3k3XEQ|n (5.86.C)

The computation of the control law (5.81) follows.

Comment

In contrast to Section 5.2, we have designed a linearizing adaptive controller

without model reduction, i.e. a controller which is able to compensate for the
full process nonlinearities. Expressions (5.81) and (5.86) show that the price
to pay is, however, a much greater complexity of the controller equations as
compared to (5.27).

5.4. A General Solution to the Linearizing Control Problem for a

Class of CST Bioreactors

Several particular ways of deriving llinearizing controllers for bioreactors have

been introduced, either through model simplification by singular perturbations
(Section 5.2) or under specific structural assumptions regarding the link
between the control action u and the component which is to be regulated y
(Section 5.3). In this section, we would like to give an idea of the way in which
the theory of nonlinear control systems can be applied to solve the problem in
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 297

general. However, in order to limit the mathematical complexity, we introduce

some simplifying hypotheses.

Model description

We consider a class of bioreactors described by the state space model (1.49),

with the particular model of the gaseous outflow rates given by (1.57), that is :

^ = Κ Ο (ξ )α ­ Ο ξ ­ Β ξ + F (5.87)

We suppose that:

1) the dilution rate D is constant;

2) the specific reaction rates α are constant;
3) the specific liquid­gas transfer rates β ί (see Section 1.5.4) are constant.
Hence the matrix Β = diag {β ί } is also constant;
4) the process involves only one external substrate such that the vector F is
written (compare to (5.3)) :

F = bu (5.88)

Under these assumptions, the model is rewritten in the following convenient,

compact form :

^ = g(ξ ) + bu (5.89)

Δ
WITH 9(ξ ) = Κ 0 ( ξ ) α ­ 0 ξ ­ Β ξ
 298 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Technical definitions

Let Μ (ξ ) : [RN ^ (RN denote a real valued vector function. The gradient of h,
denoted Vh, is the following matrix :

9hi 9hi
9ξ 7

(5.90)

9h, ahN
3ξ ι

The gradient of h, along an N­vector g, is itself the vector obtained by

multiplying Vh and g. This is denoted :

Vgh = (Vh)g (5.91)

The components of this vector are thus defined as follows :

Ν
9hj
j= 1 Ν (5.92)
i=1

It is remarkable that the notation (5.91) can be used recursively :

V^h = Vg(v5­^h) (5.93)

Moreover, the vector g can itself be a real valued vector function

g(ξ ): IR"^ ^ [R'^ (5.94)

We then notice that, with these definitions, the model (5.89) can be rewritten :
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 299

$ =ν α ξ +ό υ (5.95.a)
dt y

γ = θ "^ξ (5.95.b)

The relative degree δ of the model is the smaller integer δ such that:

C'^Vb(V^ξ ) = 0 γ = 1.2 δ ­2 (5.96.a)

θ Χ (ν ^^ξ )^0 (5.96.b)

Derivation of an input/output model

The next step is to derive an input/output model in order to be able to apply the
linearizing control principle. This is done as follows. We first differentiate the
output equation (5.95.b) :

Substituting (5.95.a) gives :

^ =θ \ ξ + C"^bu = + (C"^Vbξ )u (5.98)

If C^b = Ο ^ ^ ξ is not identically equal to zero :

c'^b = c X ξ ; ^ 0

(which corresponds to a relative degree δ = 1, see (5.96.b)), (5.98) is a

convenient input/output model and the derivation procedure has come to an
end. But, if the relative degree δ > 1, C^b = 0 and (5.98) reduces to :

^ =θ \ ξ (5.99)
 300 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

which does not involve an explicit connection between u and y. Then we

proceed by differentiating (5.99) further to obtain :

^ = Ο ^ν ^ξ + c X ( V g ξ ) u (5.100)

As before, if δ = 2, (5.100) is convenient for linearizing control design.

Otherwise, we continue the differentiation up to the relative degree δ of the
model, to obtain :

^ =Ο ^ ξ + cX(Vg^^ξ )u (5.101)

We see that the meaning of the concept of relative degree δ is that we have to
differentiate the output y δ times before terms involving the input u appear on
the right hand side of the input/output model (5.101).

Linearizing control design

With the input­output model (5.101) at hand, we are in a position to design the
control law. We adopt a reference model (5.7) of order δ , that is :

^δ ­1
• ^ ( y * ­ y) 4­ λ 1 ­ ^ ( y * ­ y) + + λ δ (y*­ y) = 0 (5.102)
dt^ dt^^

Then substituting the expressions (5.99) to (5.101) of the successive

derivatives of y into (5.102) readily yields the linearizing control law :

^­1
u(ξ ,α ) = [ c X (ν ^'ξ )_
dt V dt ^

(5.103)

As will by now be familiar to the reader, an adaptive version of this control law
is obtained by replacing the non­measured states and the parameter α by on­
line estimates provided by the algorithms described in Chapter 3.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 301

5.5. Adaptive Linearizing Control (Unknown Yield Coefficients)

5.5.1. Introduction and basic assumptions

In Section 5.2 we addressed the problem of the (adaptive) linearizing control

of bioreactors when the reaction rates φ (ξ ) are unknown, but can be alleviated
by using singular perturbations. In Sections 5.3 and 5.4 we addressed the
problem in which the lack of knowledge of the reaction rates explicitly taken
into account (by incorporating on­line parameter estimation in the control law)
but under the assumption that the yield coefficients (matrix K) are known. In
this section, we shall deal with the case where both reaction rates φ (ξ ) and
yield coefficients (matrix K) are unknown. The price to pay, however, is that
more restrictive structural assumptions have to be considered.

We suppose that the bioreactor dynamics are described by the general state
space model (1.49) :

^ = Κ Ο (ξ )α ­ Ο ξ ­ Q + F (1.49) = (5.104)

under the following assumptions :

1) There exists a state partition ξ = (ξ ι ,ξ 2) where ξ ι is a set of measured states

while ξ 2 is the complementary set of non measured states. The partition
(ξ ι .ξ 2) induces similar partitions (Ki, K2), (Qi, Q2), (Fi, F2) as usual.

2) The output to be controlled is a linear combination of the measured states :

y = C|ξ 1 (5.105)

3) On the other hand, there exists a nice partition (ξ 3, ξ b) and an auxiliary state
(see Sections 1.7 and 3.3.1) :

Ζ = Α ο ξ 3 + ξ 6 = Α ι ξ ι +Α 2ξ 2 (5.106)
 302 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

4) The matrix A2 is left invertible. The left inverse is denoted A2.

Under these assumptions, it follows that the state space model (5.104) is
equivalent to :

(1ξ
' = Κ ι Θ (ξ ι , A^(Z ­ Α ι ξ ι ))α ­ Ο ξ ^ + ­ Qi (5.108.a)
dt

dZ
= ­DZ + F b ­ Q b (5.108.b)
dt

The problem is to control the system (5.108) with K i , α and Ζ being unknown.
As we have argued in Chapter 4 (Section 4.2.2), the term K­jGa in (5.108.a)
can be reparametrized as follows :

Κ ι Ο (ξ ι , Α ^(Ζ ­Α ι ξ ι ))α = Φ (ξ ι ,Ζ )θ (5.109)

where Φ (ξ ι ,Ζ ) is a matrix of known functions of ξ ι and Ζ while θ is an unknown

parameter vector which is itself a nonlinear combination of the yield
coefficients kj and the specific rates aj. Several examples have been given in
Chapter 4, see for instance (4.19).

With the reparametrization (5.109), the model (5.108) is rewritten :

^ = Φ (ξ 1,Z)θ ­Dξ 1+F1­Q1 (5.110.a)

^ = ­DZ + F b ­ Q b (5.110.b)

The technique for constructing this model (5.110) will be illustrated in Section
5.7 with a simple practical example.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 303

5.5.2. Adaptive control design : the relative degree one case

We now suppose that the control input is the feed rate of one external
substrate chosen so that:

Fi = biu + fi with c | b i ^ O (5.111)

This means that the system has relative degree one and, therefore, that the
input/output model is obtained from (5.105) and (5.110) as follows :

^ = θ |φ (ξ ι ,Ζ )Θ ­ Dy + C|biu + Cj(fi ­ Q^) (5.112)

According to the linearizing control principle, we select first order linear

dynamics for the tracking error:

| ^ ( y * ­ y ) + λ 1(y*­y) = 0 (5.113)

The linearizing control law is then calculated by substituting (5.112) into

(5.113) :

u(ξ ,θ ) = ^ + Ci(y*­ y) + Dy ­ θ '^φ (ξ ι ,Ζ )θ + C^(U ­ α ι ) ] (TU

Λ ι )

(5.114)

In practice, the auxiliary state Ζ and the parameter vector θ are unknown. As
before, they will be replaced by on­line estimates so as to make the control law
(5.114) adaptive. On the grounds of Chapter 3, a natural asymptotic observer
for the on­line estimation of Ζ is obtained as follows from (5.110.b) :
 304 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

^ = ­D2 + Fb­Qb (5.115)

On the other hand, both direct and indirect parameter adaptation laws can be
derived, which are similar to the ones of Sections 5.2.5 and 5.2.6.

Direct adaptive control

By analogy with (5.41 .a), the parameter adaptation law is written

de
= Γ Φ (ξ 1,Z)(y*­y) (5.116)
dt

It can be justified by using exactly the same Lyapunov function (5.37) as

before.

Indirect adaptive control

In this case, the solution is clearly to use one of the parameter estimators
(4.38) (4.39) which were introduced in Section 4.2 for the design of adaptive
observers. Here also, the closed loop stability may be demonstrated by an
argumentation similar to that of Section 5.2.6.

5.5.3. Adaptive control design : relative degrees higher than one

In the case where the control feed rate is chosen in such a way that condition
(5.109) is not satisfied, it is clear that the system has a relative degree higher
than one. The solution is then to extend the theory which has been developed
in Sections 5.3 and 5.4 when the yield coefficents are known. It will not be
pursued here.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 305

5.6. Practical Aspects of Implementation

We would like in this section to draw the reader's attention to some practical
problems which may arise in the implementation of adaptive linearizing control
laws in bioreactors.

5.6.1. Saturation of the control input

A practical control limitation, which is not specific to adaptive linearizing

control schemes but is usually present in any controller, is the saturation of the
control input. Throughout the chapter, we have considered a flow rate (a feed
rate Fj or the dilution rate D) as the command input. It is obvious that, in
practice, the flow rate which is physically applied to the bioreactor will be
bounded; it must be positive and upper bounded :

0<u(t)<Umax (5.117)

In presence of bounds on the control input, the performances of the closed

loop system may be degrading with respect to the ideal case of unbounded
inputs. The effect may even be disastrous in some instances, as will be
illustrated in Sections 5.6.3 and 5.8. Ideally, an exhaustive theoretical stability
analysis should take their influence into account. An example of such an
analysis is mentioned in the bibliography. However, in general, an analysis
taking account of the input saturation is fairly involved and is beyond the scope
of this book. Usually, careful simulation studies constitute a useful tool for the
analysis of the influence of the input bounds on the control performances. In
the case study of Section 5.7 we also emphasise theoretical conditions on the
choice of the design parameters which avoid setting the control input to zero.
 306 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

5.6.2. Introduction of an integral action in the control law

The introduction of integral action in control laws is a very important and

commonly used technique for eliminating steady­state errors in the closed
loop system. It is worth noting that, in our case, integral action is introduced
naturally in the linearizing control laws proposed in the preceding sections via
the parameter adaptation algorithms. Indeed, as appears in (5.41 .a) and
(5.43.b), these equations are in integral form. Our preference is for the
introduction of the integral action via the adaptive law (and not for the simple
addition of an integrator into the controller) for the following reasons. First of
all, in the adaptive control schemes, the integrator is incorporated within a
control framework for which the theoretical stability and convergence has been
carefully studied. Second, the adaptation algorithm will give extra information
about the uncertain process parameters, which can be very valuable to the
process operator.

Let us also mention one classical problem of integral control in the presence of
bounded inputs : the reset windup. When the input reaches the bounds, the
integrator still integrates the regulation error : this may lead to undesirable
oscillations and degradation of the control performance. This problem can be
easily avoided by using an antirest windup mechanism. Let us give an
example.

Example : basic microbial growth process with a synthesis product

Consider the following bioprocess :

S—Cx + P (5.118)

characterised by the following dynamics :

 Ch^.5. ADAPTIVE CONTROL OF BIOREACTORS 307

S' = • ­ k i " φ ­ D S" π ­ ρ ­ (5.119)

dt
X 1 X ο
.p. .k2.
. P . .0.

with φ = aSX

Let us follow the procedure of Section 5.2.7 to derive a direct adaptive

linearizing controller of the substrate concentration S with the dilution rate D
as the control action. If S* denotes the set point of S and the parameter θ is
defined as follows :

θ = kiaX (5.120)

the discrete­time version of the direct adaptive linearizing controller, with input
saturation and antireset windup mechanism, is written as follows :

Xi(S*­St) + etSt
u» — · (5.121 .a)

D.= r D f i f O s D f SD^ax (5.121.b)

0 ifD,°<0 (5.121.c)
(5.121.d)

θ ,,ι =h, + Y T S , ( S ­ ­ S.) + (D, ­ Df) ^ ' " ­ 1 " ^ (5.121 .e)
5>t

In the above equations, Df is the value of the dilution rate which is calculated
by the control law (5.120) and Dt is the value which is effectively applied Xo the
bioreactor. The last terrn of the adaptation equation (5.121 .e) is the antireset
windup mechanism : when the input Dt saturates, the integral adaptation law is
reset to a value that gives 0 or Dmax. respectively.
 308 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

5.6.3. Control of a product or biomass with the dilution rate D

It is straightfonvard to see from the general dynamical model (5.1) that, for any
process component, the relative degree is always equal to one when the
dilution rate D is chosen as the control input. Therefore it appears a priori very
attractive to directly use, for the synthesis of the linearizing control law, the
dynamical equation of the component to be controlled. Due to the presence of
the input bounds, this might easily give disastrous results when the controlled
variable is a product or a biomass, as we now illustrate.

Let us consider the example of basic microbial growth with a synthesis product
(5.118) already considered above. From (5.119) we can directly synthesise the
following linearizing control law for the regulation of Ρ with D as the control
input:

λ ι (Ρ *­Ρ )­Κ 2φ
D^ = - — — (5.122)

where P* is the reference value of P. In practice, we must take account of the

bounds (5.121.b,c,d).

The problem is that this controller may be diverging. As a matter of fact, if at

some instant, the reference P* becomes larger than Ρ + k29/Xi, then D° is
negative and D is set to zero. The process is set to a batch operation : if there
is not enough substrate available to make Ρ + k 2 9 A i larger than P*, it will
remain in batch mode and Ρ will never be able to reach the desired value P*.
This is illustrated in Fig.5.7 which shows a simulation performed under the
following conditions :

φ = (Monod model) with μ * = 0.4 h""" and KM = 1 g/l

k i = 2 , k2 = 0.5
Sin(O) = 5 g/l, D(0) = 0.2 h­1, S(0) = 1 g/l, X(0) = 2 g/l, P(0) = 1 g/l
λ ι =1
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 309

0.25

τ
2 4 6 8
time (hours)
Fig.5.7. Effect of input saturation on the llinearizing control of Ρ with D

The process is initially in open loop. At time t = 1 hour the influent substrate
concentration Sjn is set to 7 g/l. The loop is closed with the control algorithm
(5.122)(5.121.b,c,d). P* has been set to 1.5 g/l. Note that this set point is a
priori accessible, since it corresponds, with the new value of Sjn equal to 7 g/l,
to a stable equilibrium point, defined by the following input and state variable
values :
 310 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

D = 0.2 h­1. Sin = 7 g/l, Ρ = 1.5 g/l, X = 3 g/l, S = 1 g/l

However, the controller appears to be unable to drive the process to it. Indeed,
the control input D is instantaneously set to zero (Fig.5.7.b); there is an
increase of the product concentration which stops when all the substrate which
remained in the bioreactor has been consumed. By that time, φ has become
equal to zero since S is equal to zero. Because the increase of Ρ is not large
enough, D° remains negative, the batch operation (D = 0) goes on and Ρ will
never reach the desired value P*.

Another example will be given in Section 5.8, showing the control of the
production rate of a gaseous product.

One solution to this problem has been suggested in the literature (see the
references): it consists of trying to reach the final set point P* step­by­step by
introducing intermediate reference values within a "setpoint definition"
mechanism. Another possible approach is to change the control law either by
choosing a feed rate Fj (and not the dilution rate D) as the control input, or to
follow the procedure proposed in Section 5.2.7, i.e. to derive a controller with
an external substrate Fj as the input and then rewrite it by choosing the dilution
rate D as the manipulated variable. By comparing (5.121 .a) with (5.122), we
see that (5.122) contains an explicit feedforward term Sjn and that its
denominator is usually positive.

5.6.4. Division by zero

It is obvious that, in practice, it is undesirable in the computation of the control

law to divide by zero (or by values close to zero). Rapid inspection of equation
(5.54) shows that the problem of division by zero will be present if the
controlled output y is equal to CsJSjn (note that this will not happen with the
controller (5.15), i.e. when the input is a feed rate Fj).
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 311

It is therefore important to check (theoretically and/or by simulation) if the

operation of the bioreactor might possibly lead the controlled output y to a
value close to CsJSjp. In many practical situations division by zero is most
unlikely. For instance, in the example of the regulation of S (5.121.a), the
bioreactor is usually operated in regions where S is strictly smaller than Sjn.
But there may exist control situations where division by zero is more likely to
appear. We shall give an example in Section 5.8 and we shall then show how
to modify the controller so as to deal with that problem.

5.7. Case Study : Adaptive Linearizing Control of Fed­Batch

Reactors

Our purpose in this section is to highlight the issue of the adaptive control of
bioreactors by discussing in detail its application to a specific class of simple
fed­batch processes. The aim is not only to describe the design procedure but
also to emphasize the benefit which is to be expected from using adaptive
control in practical engineering applications.

5.7.1. Process description

We consider a microbial growth process with an associated enzyme­catalysed

product described by the reaction scheme (1.36) :

S—i^X (1.36.a) = (5.123.a)

S+X >Ρ +X (1.36.b) = (5.123.b)

According to the modelling assumptions of Section 1.5.3, we suppose that the

reaction rates φ g and φ ο are expressed as follows :
 312 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

φ 9 = μ (8)Χ = α (8)8Χ (5.124)

9c = v(S)X = p(S)SX (5.125)

The corresponding dynamical model is written

dX
= μ (8)Χ ­ DX (5.126.a)
dt

dS
= ­ Ι <ι μ (8)Χ ­ k2v(8)X ­ D8 + DSi^ (5.126.b)
dt

^=v(S)X­DP (5.126.C)
dt

We suppose, in addition, that the yield of substrate consumption for product

formation is negligible. Therefore, ka = 0 in the model (5.126), which is
rewritten in matrix form, with the definitions (5.124)­(5.125):

χ ­ = ' 1 0' •8X 0 • ' a ' ­ D f (5.127)

dt
8 ­ k i 0 . 0 8X. ­β . D8iin
P. . 0 1.

We now explicitly adopt assumptions 1) to 4) of 8ection 5.5 in this particular

case.

1) We suppose that only the substrate 8 is measured on­line. That is :

ξ ι =8 ξ 2 = [Χ Ρ Γ (5.128.a)

K i = [ ­ k i 0] K2 = 1 0 (5.128.b)
0 1

Fi = D8in F2 = [0 of (5.128.C)

2) Consequently, the substrate 8 is also the controlled output:

 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 313

γ = ξ ΐ =8 Ci=1 (5.129)

3) Clearly, the model (5.127) exhibits a nice partition :

^a = [X P]^ ^b = S (5.130)

Z = S + kiX (5.131)

^ = ­ D Z + DSin (5.132)

4) The unmeasured state X can be recovered from the auxiliary state Ζ as

follows :

X= ­ ^ ^ (5.133)

5.7.2. Adaptive control design

From (5.127) and (5.133), the input/output dynamics are written (recall that y =
S):
dS
^ = ­ S(Z­S)a ­ DS + DSin (5.134)
dt

Notice that this equation is in the format of the model (5.112) with :

Φ (ξ ι ,Ζ ) = ­ 8 ( Ζ ­ 8 ) θ =α (5.135)

We suppose that the dilution rate D(t) is imposed by the experimental

conditions, that the control input is the influent substrate concentration Sin:

u = Sin (5.136)
 314 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

and that the control objective is to regulate the substrate concentration at a

constant desired value S* (the "set point"). Then, the linearizing control
principle (see (5.114)) gives the following linearizing control law :

Sin(Z.a) = ­1 [ λ ι ( 8 * ­ 8 ) + DS + S(Z ­ S ) a ] (5.137)

where λ ι is a design parameter.

Since the reaction rate α is presumed to be unknown and Ζ is not measured

on­line, an adaptive version of the above control law (5.137) is implemented :

Sin(2.a) = 1 [λ ι ( S * ­ S ) ­h DS + S(2 ­S)a] (5.138)

where Ζ is updated as follows :

^ = ­•2+DSjn (5.139)

and ά is computed according to the following direct adaptation scheme

(Lyapunov design, see (5.116)) :

^ = YiS(2­S)(S*­S) (5.140)

doc

except if ά = 0 and S*< S, then =0

(in order to avoid negative values of ά )

Comment

The full adaptive controller is thus constituted by equations (5.138)­(5.140). It

is obtained by a straightforward implementation of the Lyapunov design
technique presented in Section 5.2.5, but with a slight additional modification
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 315

to force the positivity of ά , which is critical to guaranteeing the control

convergence established in the next subsection.

5.7.3. Convergence analysis

The convergence analysis of the above adaptive controller is based on

Lyapunov stability theory (Appendix 3).

We define the following candidate Lyapunov function :

W(S.a7) = + γ Τ ^ά ^ + (5.141)

with: S = S*­S. ά =α ­ά . Z= Z­2

and the domain of attraction D :

D = {S.a^Z I W(S,a7) < Co, Co > 0} (5.142)

We introduce the following assumptions :

H I . The specific reaction rate α is constant.

H2. The dilution rate D has a lower bound :

0<D^i,<D(t) Vt

(recall that the process is operated in the fed­batch mode).

H3. The minimum dilution rate Dmin. the design parameter λ ι and the constant
Co are chosen such that:
 316 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

(It is easy to check that such a choice is necessarily possible).

H4. The initial estimate of Ζ is chosen such that:

Z(0) > S(0) (5.144)

We have the following convergence result.

Theorem 5. Under assumptions HI to H4, if

{S(0).a(0).2(0)}6D

then :{§(t).a(t),2(t)}€D Vt (5.145)

lim(Z­Z) = 0 limS(t) = S* (5.146)

t­»o, t­»~

Proof. First, it can easily be shown that the upper bound in H3 and H4
(together with the positivity of ά ) implies that the influent substrate
concentration Sin(t) is positive for all t. Straightforward calculations then show
that the time derivative of the Lyapunov function W(S,a,Z) can be bounded as
follows :

dW
<­λ ι δ 2­Ο ζ 2­Η α 8δ Ζ
dt

a(S*­i­Co)
<­[|s , Iz Isl (5.147)

a(S* + Co) Izl

'mm

which is a negative definite quadratic form by assumption H3. The theorem

follows (see Appendix 2).
QED.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 317

5.7.4. Yield­productivity conflict in fed­batch reactors

So far in this section we have been concerned with the design and analysis of
an adaptive regulator of substrate concentration for fed­batch reactors
characterized by the reaction scheme (5.123). Now the question arises : why
choose substrate regulation in fed­batch processes? There are several
answers to that question. One of them relies on the fact that substrate
regulation can constitute an efficient tool to manage the yield­productivity
conflict which occurs in many practical applications. We shall illustrate this
point with an example. Bacterial production of amino acids (e.g. lysine) is
known to be inhibited by high substrate concentration. The dynamics of such
processes are described in the literature (see the references) by the state
space model (5.127), with a Monod model for the specific growth rate :

μ <8) = ^ (5.148)

and a parabolic specific production rate :

v(S) = S(vo­ViS) 0<S<^ (5.149.a)

1

0 S>— (5.149.b)
V1

According to (5.124)­(5.125), the corresponding specific reaction rates are as

follows, for 0 < S < VQ/VI :

«(S) = ­ r r ^ P(S) = vo ­ ViS (5.150)

Suppose now that the fed­batch operation has a fixed duration T. We define
the productivity as the final product quantity per unit of time :

PRi«Ii=M (5.151)
 318 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

and the yield as the ratio of the final product quantity to the amount of substrate
which has been consumed :

P(T)V(T)
(5.152)
DSindx

These parameters PR and Yl are plotted in Fig.5.8 with respect to the set point
S*, for a series of simulations, performed with the model (5.127), (5.150),
under the adaptive control (5.138)­(5.140). From this figure it clearly appears
that the set point S* (and obviously the associated adaptive controller) can be
viewed as a means to modulate the process between productivity
maximization (but with a low yield) obtained with S*=1, and yield maximization
(with a low productivity) obtained with S*= 2.

It must be clear that this conclusion holds even if the analytical models of μ (8)
and v(S) are unknown to the user, provided that they have the same shape as
in Fig.5.9. In that case, the optimization procedure should consist of looking,
over successive fed­batch experiments, for the best set point S*; that is, the set
point which corresponds, from the user's viewpoint, to the best trade­off
between yield and productivity.

1 14 2
Fig.5.8. Final product quantity V(T)P(T) (x) and yield Yl (0) vs set point S*
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 319

1 2
Substrate Cone. S

Fig.5.9. Specific production rate v(S)

5.7.5. Comparison with optimal control

We now consider the case where the biomass growth itself is inhibited by high
substrate concentration. We assume that the specific growth rate μ (S) is
suitably described by a Haldane law (1.20) :

μ (S) = with μ ο = μ *Γ ΐ + V^WK| (5.153)

KM + S + S^/K,

A simulation experiment is shown in Fig.5.10 for the following set of

parameters:

k^ = 1 KM = 10 K, = 0.1 μ *=5

and the following initial and operating conditions :

S(0) = 1 X(0) = 0.1 V(0) = 10 Sin(O) = 3.5

Fo = 0.1 Vmax =12

 320 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

= 10 Yi = 10 ά (0) = 0.2 Z(0) = S(0)

Fig.5.10 compares closed loop and open loop operations of the fed­batch
process. In both cases, the reactor is fed with the same amount of substrate.
In open loop, a constant input concentration Sjn(t) is used while, in closed
loop, Sin(t) is computed by the control algorithm (5.138)­(5.140). One hour
after the end of the feeding period, X(t) has already reached its maximum
value. In open loop the substrate concentration S(t) increases rapidly to high
values which inhibits the growth and the final biomass concentration remains
below 10% of the value reached in closed loop.

It is interesting to compare this result with that which would be obtained by

using optimal control theory. In this application, the optimal control problem is
the problem maximizing the final biomass production (X(T)V(T) under the
constraint of the dynamical model (5.127).

It can be shown that, if:

S(0) (5.154)

the optimal control is written :

, V M^W '^M'^I

^ ^ ' ^ • ^ V i i w ^ ' ''''''

with V and X computed from :

,y ^ . J i V ^ X . ^ X (5.156)

In contrast with the adaptive regulator (5.138)­(5.140) which just requires the
choice of the set point S*, it appears that the main drawback of this optimal
control is the need for a knowledge of the specific growth rate structure and, in
particular, of the constants μ *. KM and K|.
 Ch^.5. ADAPTIVE CONTROL OF BIOREACTORS 321

: adaptive control
^—: optimal control
: open loop
υ
§
Υ

«ο

ο
φ
S*=i

TIME (H)

5 g
c

2 Ο
TIME (H)
' Γ · 1 ' R '

>

TIME (H)

TIME (H)

10 20

Fig.5.10. Adaptive control of fed­batch reactors : maximization of X(t)

322 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

X o adaptive control
optimal control
ο
§
υ
Μ
Μ
Ι
ο
ω _1 time^ (h)
Τ 1 1 ' Γ
'Ο 10

^ 2
Ο
C
/
ο /
ο S* = 0.9
φ /

η
Ώ
time (h)
1 1 Γ ­τ ^
1 ο 10 20
/

ο
/
2 ^ /
/ c
/ /
g o
^
,
time (h)
1 ­­I 1 1 τ
10 20

Fig.5.11. Comparison adaptive control ­ optimal control: robustness with

respect to model inaccuraccies
 Ch£p.5. ADAPTIVE CONTROL OF BIOREACTORS 323

Besides, as shown in Fig.5.10, the behaviour of the adaptive controller is so

close to the optimal one that no significant discrimination between both
trajectories can be made, despite the initial errors in the estimates
Z(0) and ά (0).

Obviously, the quasi­optimality of the adaptive controller relies critically on a

precise knowledge of the set point which maximizes the specific growth
rate. It is therefore relevant to study the sensitivity of the control performance
with respect to the value of S*. In Fig.5.11 we have considered that the "best"
value of S * was known with a 10% error. This means that, in our simulations,
the adaptive controller (5.138)­(5.140) has been implemented with a desired
set point S * = 0.9 (while the true S * = 1). Accordingly, the value of KM has
been set to 8.1 (instead of 10) in the optimal control law (5.150) because :

Fig.5.11 shows the better robustness of the adaptive controller towards that
inaccuracy. Although the biomass production is no more optimal, it remains
very close to the optimum, while the (open loop) optimal controller diverges
completely form the optimal trajectory, which induces a loss of more than 50%
of the production.

5.8. Case Study : Adaptive Control of the Gaseous Production

Rate of a Synthesis Product

Throughout the chapter we have discussed the control of an output which is a

linear combination of the process variables. We shall now illustrate how to
extend the methodology to the control of variables which are complex
nonlinear combinations of the process variables. We shall also see, in Section
5.8.4 how to modify the linearizing control law in order to avoid division by
zero.
 324 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

5.8.1. Description of the process

Consider a single substrate, single biomass process with a low solubility

product Ρ :

(5.157)

By considering the singular pertrubation approximation for Ρ (Section 1.8), the

dynamical equations of the process are given by :

x" = " 1 • φ + 0 "­ D X" (5.158.a)

dt
.s. .­ki. .S.

Q= k29 (5.158.b)

where Q is the gaseous outflow rate of the synthesis product Ρ and where the
reaction rate φ can be expressed as follows according to (1.53) (1.54) :

φ =μ Χ (5.159.a)
= aSX (5.159.b)

Assume that the objective here is to regulate the gaseous production rate Q at
a desired level Q* by acting on the dilution rate D.

5.8.2. Exampie : anaerobic digestion process

A typical example is anaerobic digestion. In the preceding sections we have

discussed the application of anaerobic digestion processes in wastewater
treatment. Another possible application, potentially combined with the first
one, is methane production. Methane may then be used as an energy supply
to feed e.g. gas ovens in industry or as a fuel for light or heating purposes on
farms. If the process is used for methane production, there is a clear incentive
to optimize its production rate Q and to have a methane production that fits the
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 325

energy demand. As a matter of fact, we can expect that the optimal desired
value Q* will be an increasing function of the available influent substrate
concentration Sin(t). as illustrated in Fig.5.12. The control objective consists
here of regulating the methane production rate Q at a desired value Q* by
acting on the dilution rate D. Sjn may be chosen so as to define an "optimal"
value of Q*, which itself corresponds to the energy demand.

Sin(g/'1
Fig.5.12. Q* = f(Sin)

Let us now describe two different control algorithms.

5.8.3. Control algorithm #1

It is clear from equation (5.159.b) that the production rate Q is a nonlinear

function of the state (X, S). Moreover, its time evolution is described explicitly
by a differential equation in the model (5.158). In order to synthesise a
linearizing controller, let us first calculate its time derivative.
 326 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

A bad solution consists of using equation (5.159.a) (which does exhibit the
explicit dependence with respect to S) as a basis for the time derivation.
Indeed , we then obtain :

(5.160)

with :
1 ^Ι μ
(5.161)

This equation is similar to the dynamical equation of the biomass in a basic

microbial growth process (with θ instead of μ ). As has been pointed out in
Section 5.6.2, the linearizing control directly based on this equation is not a
good solution since, as a result of input saturation, the control may diverge.
This is illustrated in the simulation of Fig.5.13.

0 100 200 (h)

Fig.5.13. Divergence of the controller due to the input saturation

 α ι ^.5. ADAPTIVE Ο Ο Ν Τ Π α OF BIOREACTORS 327

The simulation has been performed under the following conditions. A Monod
model (1.18) has been chosen for the specific growth rate μ :

μ *5
μ = (5.162)
KM + S

with μ * = 0.4 d­1 and KM = 0.4 g/l

The initial conditions have been set to :

X(0) = 0.095 g/l. S(0) = 0.4 g/l. Sin(O) = 3 g/l, D(0) = 0.2 d­i

and the yield coefficients are :

ki = 27.3 k2 = 75 |2d/g.h

The upper bound on the input D is :

Dmax = 0.39 d ^

The process is initially operated in open loop (without control) with a white
noise input signal D(t). At time t = 72. Sjn has been set to 3.4 g/l, and according
to Fig. 5.12, Q* has been set to 1.58 l/h. The gain of the controller λ ι is equal
to 1 h­1. The dilution rate is instantaneously set to zero. It remains at this
bound and Q tends to zero (since S, and therefore μ (8), tends to zero).

Therefore we propose to use equation (5.159.b) to compute the time derivative

of Q. This gives :

dQ
= Θ ι 80 ­ + OaQ + D Q [ ^ ­ 2 (5.163)
dt

with :

[θ ι , θ 2. Θ 3] = ilL I d a (5.164)
" · k2' α dt
 328 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

The adaptive linearizing adaptive control law then specialises here as follows :

λ ι ( α * ­ Q) ­ §iSQ + GaQ^/S ­ OgQ

(5.165)
Q[Sin/S­2]

In practice, we have to take account of the physical bounds on the flow rate
and the control law is then implemented as follows :

= D' if 0 < D" < D, (5.166.a)

max
=0 if D ° < 0 (5.166.b)
= D,max if D°>D, (5.166.C)
max

In (5.165) and (5.166) 0 ° is the value of the control input calculated by the
computer, and D is the input value which is effectively applied at the control
value.

1.6

Q­
1.5

1.4
0.4

0.3

0.2 η /*Λ /^^^­ι Γ Ί Λ ί Κ t

0 100 200 (h\

Fig.5.14. Adaptive control of the production rate Q

 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 329

This control law (5.165) appears to be more effective. This is mainly due to the
explicit dependence of Q on the limiting substrate S. This is illustrated in
Fig.5.14, which shows the performance of the controller (5.165) in the same
experiment as in Fig.5.13.

0.4

0.3

0.2

0.1

0.0

2.0 s.
\
\
Sin/2
1.0

0.0
0 100 200 300 400

Fig.5.15. Problems of division by zero

However, the drawback of the controller (5.165) is the possibility of the

occurrence of a division by zero. Indeed it is obvious that values of the
 330 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

substrate concentration S close to Sjn 12 are possible in practice. This is

illustrated in Fig.5.15. At time t = 72, the desired value Q* has been set to 1.4
l/h (Sin is equal to 3 g/l). The control input D is oscillating between its lower
and upper bounds, and the controller is able to drive the output Q to its desired
value Q*.

The simulation has been performed in almost similar conditions as before but
with the following initial conditions :

X(0) = 0.04 g / l , S(0) = 1.9 g/l, Sin(O) = 3 g/l, D(0) = 0.33 d­l

5.8.4. Control algorithm #2

In order to avoid division by zero wich arises when S is equal to Sjn/2, we

suggest the use of the following dynamic control law :

λ ι ( 0 * ­ Q) ­ GiSQ + 0 2 · ^ ­ OgQ (5.167)

dt ω +ς ^

where ω is an arbitrary positive constant (ω > 0) and g is the denominator of

(5.165), i.e.:

>in
g=Q ­2 (5.168)

It is clear that for values of ω different from zero, division by zero is avoided.

Note that, in the steady­state (dOO/dt = 0), the control law (5.168) reduces to
the preceding one (5.165).

It is also worth noting that (5.168) with λ ι = 1 is a continuous time equivalent of

the generalized minimum variance (GMV) controller.
 Ch^.5. ADAPTIVE CONTROL OF BIOREACTORS 331

By way of comparison, Fig.5.16 reproduces the same simulation experiment

as in Fig.5.15 but with the application of the controller (5.168) (with ω = 0.16)
and shows its ability to avoid division by zero and reach the desired set point
Q*.

04

Dt
0.2

0.0

2.0
' ^'
1.0

t
0.0
0 100 200 300 400 <h)

Fig.5.16. Dynamic control of the production rate Q

 332 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

5.9. References and Bibliography

Section 5.0

Adaptive linear control of biotechnological processes is investigated e.g, in :

Axelsson J.P. (1989). Modelling and Control of Fermentation Processes.

Doctoral Dissertation, Dept. Aut. Cent., Lund Inst. Technol., Sweden,
LUTFD2/(TFRT­1030).
Bastin G. D., D. Dochain, M. Haest, M. Installe and Ph. Opdenacker (1983).
Modelling and adaptive control of a continuous anaerobic fermentation
process. In A. Halme (Ed.), Modelling and Control of Biotechnical
Processes, Pergamon, Oxford, 299­306.*
Dekkers R.M.and M.H. Voetter (1986). Adaptive control of a fed­batch baker's
yeast fermentation. In A. Johnson (Ed.), Modelling and Control of
Biotechnological Processes, Pergamon, Oxford. 73­80.
Frueh K., Th. Lorenz, J. Niehoff, J. Diekmann, R. Hiddessen and K. Schurgerl
(1986). On­line measurement and control of penicillin V production. In A.
Johnson (Ed.), Modelling and Control of Biotechnological Processes,
Pergamon, Oxford, 45­48.
Montague G.A., A.J. Morris, A.R. Wright, M. Aynsley and A.C. Ward (1986). On­
line estimation and adaptive control of penicillin fermentation, lEE Proc,
133, Pt. D, 240­246.
Rolf M.J. and H.C. Lim (1984). Adaptive on­line optimization of continuous
bioreactors, Chem. Eng. Com., 29, 229­255.
Verbruggen H.B., G.H.B. Eelderinck and P.M. Van den Broecke (1986)
Multiloop controlled fed­batch fermentation process using a self­tuning
controller. In A. Johnson (Ed.), Modelling and Control of Biotechnological
Processes, Pergamon, Oxford, 91­97.
Williams D., P. Yousefpour and E.M.H. Wellington (1986). On­line adaptive
control of a fed­batch fermentation of Saccharomyces cerevisiae.
Biotechnol. Bioeng., 28, 631­645.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 333

Section 5.1

The be
:

Isidori A. (1989). : An 2"d edition

Springer Verlag.

Hoc K.H. and J.C. Kantor (1986). Linear feedback equivalence and control of
an unstable biological reactor. Chem. Eng. 46, 385­389.
Hoo K.H. and J.C. Kantor (1986). Global linearization and control of a mixed­
culture bioreactor with competition and external inhibition. 82,
43­62.

The

Axelsson J.P. (1989). Modelling and Control of Fermentation Processes.

Doctoral Dissertation, Dept. Aut. Cont., Lund Inst. Technol., Sweden,
LUTFD2/(TFRT­1030).
Pomerleau Y. (1990). Modelisation et controle d'un prodece fed­batch de
culture des levures a pain.. Ph. D. Thesis, Ecole Polytechnique de
Montreal, Canada.

Sections 5.2, 5.3 and 5.4

Bastin G. (1988). Lyapunov design of convergent adaptive controllers for

multilinear systems arising from biotechnological applications. In N.M. Fish,
 334 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

R.L Fox & N.F. Thornhill (Eds.), Computer Applications in Fermentation

Tectmology : Modelling and Control of Biotectinological Processes,
Elsevier, 331­339.

Γ Λ β ttieory of Lyapunov design of adaptive controllers is further invertigated in:

Taylor D. P.V. Kokotovic, R. Marino and I. Kanellakopoulos (1989). Adaptive

Regulation of Nonlinear Systems with Unmodeled Dynamics. IEEE Trans.
Aut. Control, 34, 405­412.

Indirect adaptive linearizing control of bioreactors is treated in :

Bastin G. and D. Dochain (1988). Nonlinear adaptive control of

biotechnological processes. Proc. ACC, Invited Session on Advances in
Control of Biotechnical Processes, 1124­1128.

The stability analysis of indirect adaptive controllers (Lemma 5.1 and Theorem
5.1) follows the argumentation in:

Bastin G. and G. Campion (1989). Adaptive Control for Linearly Parametrized

Nonlinear Systems, Proc. IFAC Symposium ACASP, Glasgow.

Related results are presented in :

Campion G. and G. Bastin (1989). Indirect Adaptive State Feedback Control of

Linearly Parametrized Nonlinear Systems. Internal report AP89­09, Lab.
Aut. Dyn. Anal. Syst., UCL, Belgium.

The application to propionate regulation (Section 5.2.8) is adapted from :

Renard P., V. Van Breusegem, N. Nguyen, H. Naveau and E.J. Nyns (1990).
Implementation of an adaptive controller for the start­up and steady­state
working of a biomethanation process operated in the CSTR mode.
Biotechnol. Bioeng., to appear.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 335

A :

Renard P., D. Dochain, G. Bastin, H. Naveau and E.J. Nyns (1988). Adaptive
control of anaerobic digestion processes ­ a pilot­scale application.
3 1 , 287­294.

5.3
:

Bastin G. (1988). State estimation and adaptive control of multilinear

compartimental systems : theoretical framework and application to
biotechnological processes, in
Lecture Notes on Control and Information Science, n°142. Springer Verlag,
341­352.

of :

Chen L., (1987). Contribution a la commando adaptative de precedes de

fermentation aerobies. Master's degree thesis, Universite Catholique de
Louvain, Belgium.

Section 5 . 6

of of
:

Dochain D. (1986). On­line Parameter Estimation, Adaptive State Estimation

and Adaptive Control of Fermentation Processes. Doctoral Dissertation,
Univ. Cath. Louvain, Belgium.

A
:
 336 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Pomerleau Y., M. Perrier and D. Dochain (1989). Adaptive nonlinear control of

the baker's yeast fed­batch fermentation. Proc, ACC, Invited session on
Intelligent Systems and Advanced Control Strategies in Biotechnology.

of is
in following of as
in 5,6.3 is

Chamilothoris G. (1987). Techniques adaptatives pour le suivi et la conduite

des processus de fermentation. Doctoral Dissertation, LAAS, CNRS,
Toulouse, France.
Chamilothoris G. and Y. Sovely (1988). Adaptive control of biomass and
substrate concentration in a continuous­flow fermentation process. Rairo
Λ Ρ //, 22, 159­175.

of of
also in :

Chamilothoris G., P.Y. Renaud, Y. Sovely and P. Vigie (1988). Adaptive

predictive control of a multistage fermentation process. J. 48,
1089­1105.
Ko K.Y., B.C. Mc Innis and G.C. Goodwin (1982). Adaptive control and
identification of the dissolved oxygen process. 18, 727­730.
Rundqwist L. (1986). Self­tuning control of the dissolved oxygen concentration
in activated sludge processes. Master's degree thesis, Lund Inst. Tech.,
Sweden.

Section 5 . 7

on of was in :

Dochain D. and G. Bastin (1989). Adaptive control of fedbatch reactors.

Eng. Com., to appear.
 Chap.5. ADAPTIVE CONTROL OF BIOREACTORS 337

The model of lysine production in Section 5.7.4 is adapted from :

Ohno H., E. Nakanishi and T. Takamatsu (1976). Optimal control of a

semibatch fermentation. Biotechnol. Bioeng., 28, 847­864.

Further examples on yield-productivity conflict in fed-batch processes can be

found in:

Chen L. (1990). Identification and Control of Biological Systems. Doctoral

Dissertation, Univ. Cath. Louvain, Belgium.

Section 5.8

The case study on adaptive control of the gaseous production rate is taken
from:

Dochain D. and G. Bastin (1984). Adaptive identification and control

algorithms for non linear bacterial growth systems. Automatica, 20, 621­634.
 Appendix 1. MODELS OF THE SPECIFIC GROWHH RATE 339

APPENDIX 1

MODELS OF THE SPECIFIC GROWTH RATE

1 . Dependence on the Substrate Concentration : μ ( 8 )

Monod (1942)

μ *8(1)
μ (8) =
KM + S(t)

(1905)

μ (8)=Γ ^8(0 if 8(0 < KM

Κ ,Μ

μ if8(t)>KM

(1942)

8(t)
μ (8) = μ * 1 ­ exp
ν K,

(Andrews, 1968)

μ ο 8(ί )
μ (8) =
KM + 8(t) + 8^(t)/K,
 340 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Moser (1958)

μ *8\ΐ )
μ (8) =
Κ Μ + 8\0
with λ > Ο

Powell (1967)

μ (8) = μ KM + 8(t) ­ V K M + S ( t ) ^ ­ 4 K M 8
2KML

with KM > Ο

Konak (1974)

with λ > 0

Aborhey and Williamson {X 977)

a ) ^ = ί <[μ *8(t)­μ [8(t) + K M Ϊ

μ *8(0
b ) ^ =k ­μ
dt LKM + S(t)

with k: constant

Takamatsu et ai. (1983)

μ Β (8) = Κ Β (8(0­8B)

μ ρ (8) = Κ ρ ( 8 ( 0 ­ 8 F )

with μ Β (8): bulking sludge specific growth rate

 Appendix 1. MODELS OF THE SPECIFIC GROVVTH RATE 341

μ ρ ( S ) : floc­forming sludge specific growth rate

KB, S B , KF,SF : constants

Axelsson et al (1984)

with Y : yield coefficient

7i­(S(t)) : negative part of π = p(S­Sc)
Sc : critical substrate concentration.

Hoppe and Hansford {^ 982)

^^"^^ KM + S(\) KM + Y[Sin­S(t)]

with KM, Y : constants

Cfien and Hasliimoto (1978)

^^""^ K [ S i n ­ S ( t ) ] + S(t)

with Κ : constant

y'ue(Roques et al., 1982)

μ (8) = μ *S(t)
K[Sin­S(t)] + KM+S(t)

Jost et a/(1973)

(Knii + S(t))(K^2 + S(t))

with Kmi, Km2 : saturation constants
 342 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Shehata and Man (1971)

Km1+S(t) K^2 + S(t)

with + μ 2=μ *

Sokol and Howell (1981)

KiS
1)μ (8) =
K2+ 8^

Κ ι 8
2) μ (8) =
Kp + 8*^3

Ming et ai (1988)

μ (8) =μ

2. Dependence on the Biomass Concentration : μ (Χ )

Verhulst (1838)

μ (Χ ) = μ * 1 ­

with Χ Μ : constant
 Appendix 1. MODELS OF THE SPECIFIC GROWTH RATE 343

Kono (1969)

μ (Χ ) = Κ χ Φ
with Κ χ = growth rate constant
Φ = apparent coefficient of growth activity
Φ = 0, induction phase
Φ = φ , 0 < φ < 1, transient phase
φ = at
Φ = 1 exponential phase
Cxc ­ Cx
Φ =
CxM ­ Cxc Cx
with Cxc: critical cell concentration
CxM : maximum cell concentration
Cx : cell concentration

3. Dependence on the Substrate Concentration S(t)

and on the Biomass Concentration X(t) : μ ( 8 , Χ )

(1959)

μ *S(t)
KcX(t) + S(t)
with Kc: saturation constant

(1977)

C(t)
μ (8,Χ ) = Ki
X(t)
dC
= K2S(t)[X(t) ­ C(t)] ­ K3C(t)
with C(t): cell­substrate complex
K i , K2, K3 : constants
 344 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Kishimoto et al (1983)

μ (8,Χ ) = μ + Qi(X(t) ­ X) + QsiSCt) ­ S)

with μ , X, S : mean arithmetic values

ofμ (t),X(t). S(t)
Q i , Q2: partial regression coefficients

Staniskis and Levisauskas {^ 984)

μ (S,X) = K1S(t)­K2X(t)
with K i , K2 : constants

4. Dependence on the Product Concentration P(t) : μ (Ρ )

Hinslieiwood (1946)

μ (Ρ ) = μ * ­ Κ ι [ Ρ ( 0 ­Kg]
with K i , K2 : positive constants

Aiba et al (1968)

μ (P) = μ *e­^''(«
with Ki : constant

Jerusaliwski and Engambervediev {^ 969)

with Kp : saturation constant

μ *Κ ρ
 Appendx 1. MODELS OF THE SPECIFIC GROWTH RATE 345

Z­eve/isp/e/(1980)

P(t)
μ (Ρ ) = μ * 1 ­

with P L : limiting product concentration

η : toxic power

HSgglund (1983)

μ (Ρ ) = Κ ι ­ Κ 2
K3 + P

5. Dependence on the Substrate Concentration S(t)

and the Product Concentration P(t) : μ ( 8 , Ρ )

and (1979)

μ ο 5α )
μ (8,Ρ ) =
K M + S(t) + S^(t)/Ki

Jin a/(1981)

S(t)
μ (8.Ρ ) = μ ^ ^ ^ 3 ( ^ ^ exp {­Κ ^Ρ ί Ο ­ KgSd)}

with K i , K2 : constants

e/ a/ (1981)

S(t) Kp ρ ω ^
μ (8,Ρ ) = μ * 1­
KM + S(t) Kp + P(t) PL J
 346 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Aborhey and Williamson {^ 977)

Six) P(t)
a) μ (8,Ρ ) = μ
KM + Six) Kp + P(t)

. Six) Six)
b) μ (8,Ρ ) = μ
KM + Six) 8(t)­HKpP(t)

with Kp : constant

Dourado and Caivet (1983)

8(t) Kp
μ (8,Ρ ) = μ *
KM+Six) + SHX)/K; Kp + P ( t ) l , PL j

Moulin et al (1980)

μ (8,Ρ ) = μ * exp{­Ki P(t) + Kg (8(t) ­ K3) ­ K4 P(t) (8(t) ­K3)}

with K i , K2. K3, K4 : constants

Aiba et a/(1968)

Bazua and Wilke (1977)

μ (3.Ρ ) = μ · ­ '^'^'
K2­P(t) KM + 8(t)
 Appendix 1. MODELS OF THE SPECIFIC GROWTH RATE 347

6. Dependence on the Substrate Concentration S(t)

and the Dissolved Oxygen Concentration C(t) :

O/sson (1976)

S(t) C(t)
μ (8,0 μ K^^s(t) Kc + C(t)
with Kc : saturation constant

Peringer et a/. (1972)

• S(t) C(t) ^ 1
^ KM + S(t) Kc + C(t) + M2
i+KiC(t)
with Ki : constant
^2: specific endogenous respiration rate
2μ ΐ +μ 2=μ *

7. Dependence on the Substrate Concentration S(t),

the Dissolved Oxygen Concentration C(t) and the
Product Concentration P(t) : μ ( 8 , 0 , Ρ )

Williams et ai. (1984)

μ (8,0.Ρ ) J KiSd) ^ K2P(t)_

K^Pft) U
· ' ' l K M + s ( t ) ^ K p + P(t)p(t)n
with K i , K2, K3, K4 : constants
 348 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

8. Dependence on the Substrate Concentration S(t)

and the Inhibitory Metabolite Concentration l(t) :
μ (3,Ι )

Kishimoto et al (1983)

3)μ (3.Ι ) = Ki­Η K2S(t)­Η KaKt)

with K i , K2, K3 : constants
l(t): inhibitory metabolite concentration

ω μ (8.Ι ) = /^—L­
KM + S(t) n . | 2 ( t )

9. Dependence on the pH : μ (pH)

Andreyeva and Biryuliov (1973)

μ (ρ Η ) = α ρ Η 2­ι ­β ρ Η ­ι ­γ
with a, β , γ : constants

10. Dependence on the pH and the Temperature : μ ( ρ Η , Τ )

Cheruy and Durand (^ 979)

μ (ρ Η ) = α ρ Η ^ ­ι ­ β pH (δ ­ι ­ Τ ) ­ι ­ γ
with α , β , γ , δ : constants
 Appendix 1. MODELS OF THE SPECIFIC GROWTH RATE 349

11. D e p e n d e n c e on the p H a n d the Substrate

Concentration S(t) : μ ( ρ Η , 8 )

Andreyevs and Biryukov (1973)

3)μ (ρ Η ,8) = —
KM + 8H^ + ( 8 H Y / K i
with H+ : hydrogen ion concentration
pH : ­ log H+

μ *8
b) μ (ρ Η ,8) =

KM 1 1+ + 77­ + 8

with Ka, Kb: constants

l KbJ

μ *8
c) μ (ρ Η .8) =
^ K3
KM + 8
Η * Kb
I KJ
with Kg, K^: constants

Ο μ (ρ Η .8) = ^ ^
KM + S KH + H +

with KH : constant

KQH
e) μ (ρ Η .8) = ­ ^
KM + S KOH+1/H^

with KoH : constant

μ *8 KH KQH
f) μ (ρ Η ,8) =
KM + S KH + H"" KQH + I / H *
 350 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

μ *8 KH
g) μ (ρ Η ,8) =
^ 5 KH + H+ KOH + 1/H^
K,Μ

μ ·8 KoH
h) μ (ρ Η ,8) =
1+ ­i­S KH + H"^ KOH + I/H"^
KM

Jackson and Edwards (1975)

μ *8
μ =
(Κ Μ + 8­Η 8^/Κ |[1­Η Κ 3/Η *])

with K i , K2, K3 : constants

12. Dependence on the Temperature Τ : μ (Τ )

Topiwala and Sinclair (1971)

μ (Τ ) = Ai e­^i'""^ + A2 e­^^'"'^ ­ A3
with A l , A2, A3 : constants
E i , E2 : activation energies
R : gas constant
Hashimoto (1982)

μ σ )= α Τ ­β
with a, β : constants
 Appendix 1. MODELS OF Τ Ή Ε SPECIFIC GROWTH RATE 351

13. Dependence on the Temperature Τ and the

Biomass Concentration X(t) : μ (Τ , X)

Constantlnides (1970)

X(t) ^
μ (Τ .Χ ) = Ο ι σ ) 1 ­
b2(T)

bi(T) = Ki­K2a(t)­K3)^

b2a) = K 4 ­ K 5 ( T ( t ) ­ K 6 ) ^

with K i , K2, K3, K4, K5, Ke : constants

14. Dependence on the Light Intensity l(t) and the

Biomass Concentration X(t) : μ (Ι , X).

Tamiya et a/(1953)

ZeX(t) μ *+θ Ι ο Θ χ ρ [­ε Ζ Χ (ΐ )]

with ε : extinction coefficient of the micro­organisms

Ζ : culture's depth
lo : incident intensity.
 352 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

References

Aborhey S. and D. Williamson (1977). Modelling of lactic acid production by

Streptococcus cremoris HP. J. Gen. Appl. Microbiol., 23, 7­21.
Aiba S., M. Shoda and M. Nagatani (1968). Kinetics of product inhibition in
alcohol fermentation. Biotechnol. Bioeng., 10, 845­864.
Andrews J.F. (1968). A mathematical model for the continuous culture of
microorganisms utilizing inhibiting substrates. Biotechnol. Bioeng., 10,
707­723.
Andreyeva L.N. and V.V. Biryukov (1973). Analysis of mathematical models of
the effect of pH on fermentation processes and their use for calculating
optimal fermentation conditions. Biotechnol. Bioeng., 4, 61­76.
Axelsson J.P. and P. Hagander (1984). Control of baker's yeast production
based on ethanol measurements. Proc. 3^^ European Congress on
Biotechnology, Verlag Chemie, Weiheim, 2, 629­636.
Bazua C D . and C.R. Wilke (1977). Ethanol effects on the kinetics of a
continuous fermentation with Saccharomyces cerevisiae. Biotechnol.
Bioeng.,7, 105­118.
Blackman F. (1905). Optima and limiting factors. Ann. Bot, 19. 281­295.
Briggs G.E. and J.B.S. Haldane (1925). Biochem. J., 242, 3973.
Chen Y.R. and A.G. Hashimoto (1978). Kinetics of methane fermentation.
Symp. on Biotech, in Energy Prod, and Cons., May.
Cheruy A. and A. Durand (1979). Optimization of erythromycin biosynthesis by
controlling pH and temperature : theoretical aspects and practical
applications. Biotechnol. Bioeng. Symp., 9, 303­320.
Constantinides Α ., J.L Spencer and E.L Gaden (1970). Optimization of batch
fermentation processes. I. Development of mathematical models for batch
penicillin fermentations. Biotechnol. Bioeng., 12, 803­830.
Contois D. (1959). Kinetics of bacterial growth relationship between
population density and specific growth rate of continuous cultures. J.
Gen. Microb., 21, 40­50.
Dourado A. and J.L. Caivet (1983). Static optimization of the ethanol
production in a cascade reactor. In A. Halme (Ed.), Modelling and Control
of Biotechnical Processes, Pergamon, Oxford, 177­184.
 Appendix 1. MODELS OF THE SPECIFIC GROWTH RATE 353

Ghose T.K. and R.D. Tyagi (1979). Rapid ethanol fermentation of cellulose
hydrolysate : II. Product and substrate inhibition and optimization of
fermentor design.Biotechnol. Bioeng., 2Λ , 1387­1400.
Hagglund T. (1983). Incubation time prediction in yoghurt manufacturing. In A.
Halme (Ed.), Modelling and Control of Biotechnical Processes,
Pergamon, Oxford, 307­311.
Haldane J.B.S (1930). Enzymes, Longmans, London.
Hashimoto A.G. (1982).Methane from cattle waste : effects of temperature,
hydraulic retention time and influent substrate concentrationon kinetic
parameter (K). Biotechnol. Bioeng., 24, 2039­2052.
Hinshelwood C M . (1946). Chemical Kinetics of the Bacterial Cell. Oxford
Univ. Press, London.
Hoppe G.K. and G.S. Hansford (1983). Ethanol inhibition of continuous
anaerobic yeast growth. Biotechnol. Lett., 4 , 1 , 39­44.
Jackson J.V. and V.H. Edwards (1975). Kinetics of substrate inhibition of
exponential yeast growth. Biotechnol. Bioeng., 17, 943­964.
Jerusalimski N.D. and N.B. Engamberdiev (1969). Continuous Cultivation of
Microorganisms. Academic Press, New York, 517.
Jin C.K., H.L Chiang and S.S. Wang (1981). Steady­state analysis of the
enhancement in ethanol productivity of a continuous fermentation
process employing a protein­phosphotifind complex as a protecting
agent. Enzyme Microbiol. Technol., 3, 249­257
Jost J.L., J.F. Drake, A.G. Frederickson and H.M. Tsuchiya (1973). Interactions
of Tetrahymena pyroformis, Escherichia coli, Azobacter vinelandii and
glucose in a minimal medium. J. Bacterid., 113, 834­840.
Kishimoto M., T. Sawano, T. Yoshida and T. Taguchi (1983). Optimization of a
fed­batch culture by statistical data analysis. In A. Halme (Ed.), Modelling
and Control of Biotechnical Processes, Pergamon, Oxford,161 ­168.
Konak A.R. (1974). Derivation of a generalised Monod equation and its
applications. J. Applied Chem. Biotechnol., 24, 453.
Kono T. and T. Asai (1969). Kinetics of fermentation processes. Biotechnol.
Bioeng., U, 293­321.
Levenspiel O. (1980). The Monod equation : a revisit and a generalization to
product inhibition situations. Biotechnol. Bioeng., 22, 1671­1687.
 354 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Michaelis L. and M.L. Menten (1913). Die kinetic der Invertinwirkung.

Biochemische Zeitschrift, 4 9 , 334­369.
Ming F., J.A. Howell and M. Canovas­Diaz (1988). Mathematical simulation of
anaerobic stratified biofilm processes, in N.M. Fish, R.I. Fox and N.F.
Thornhill (Eds.),Computer Applications in Fermentation Technology :
Modelling and Control of Biotechnological Processes, Elsevier.
Monod J. (1942). Recherches sur la Croissance des Cultures Bacteriennes.
Hermann, Paris.
Moser H. (1958). The dynamics of bacterial populations in the chemostat.
Carnegie Inst Publication, n°614, Washington.
Moulin G., H. Boze and P. Galzy (1980). Inhibition of alcoholic fermentation by
substrate and ethanol. Biotechnol. Bioeng., 22, 2375­2381.
Nihtila M. and J. Virkkunen (1977). Practical identifiability of growth in batch
fermentation. Proc. 9^^ IFAC World Congress, Budapest, July 2­6, 225­
230.
Olsson G. (1976). State of the art in sewage treatment plant control. AlChE
Symposium Series, n°159, vol.72, 52­76.
Peringer P., H. Blachere, G. Corrieu and A.G. Lane (1972). Mathematical
model of the kinetics of growth of Saccharomyces cerevisia. 4^^ Int. Perm.
Symp., Kyoto.
Powell E. (1967). Growth rate of microorganisms as a function of substrate
consumption. Microbial Physiology and Continuous Culture, 3'^^ Symp.,
34­56, HMSO, London.
Roques H., S. Yue, S. Saipanich and B. Gapdeville (1982). Is Monod's
approach adequate for the modelisation of purification processes using
biological treatment? Water Resources, 16, 839­847.
Rozzi A. (1984). Modelling and control of anaerobic digestion processes.
Trans. Inst. Meas. Control., 6, n°3, 153­159.
Sevely Y., J.P. Pourciel, G. Rauzy and J.P. Bovee (1981). Modelling,
identification and control of the alcohol fermentation in a cascade reactor.
Proc. &^ IFAC World Congress, Kyoto, 22, 177­184.
Shehata E. Talaat and A.G. Marr (1971). Effect of nutrient concentration on the
growth of Escherichia coli. J. Bacterid., 107, 210­216.
 Appendix 1. MODELS OF THE SPECIFIC GROWTH RATE 355

Sokol W. and J.A. Howell (1981). Kinetics of phenol oxidation by washed

cells. Biotechnol. Bioeng., 23, 2039­2049.
Staniskis J. and D. Levisauskas (1984). An adaptive control algorithm for fed­
batch culture. Biotechnol. Bioeng., 26, 419­425.
Takamatsu T., S. Shioya and H. Kurome (1983). Dynamics and control of a
mixed culture in activated sludge process. In A. Halme (Ed.), Modelling
and Control of Biotechnical Processes, Pergamon, Oxford, 103­110.
Tamiya H., E. Hase, K. Shibata, A. Mituya, Nihei and T. Sasa (1953). Kinetics
of growth of Chorella, with special reference to its dependence on
quantity of available light and on temperature. In J.S. Burlew (Ed.), Algal
Culture from Laboratory to Pilot Plant, Carnegie Inst. Washington, D.C.,
Chap.16.
Tessier G. (1942). Croissance des populations bacteriennes et quantites
d'aliments disponibles. Rev. Sci, Paris, 80­209.
Topiwala H. and C.G. Sinclair (1971). Temperature relationship in continuous
culture. Biotechnol. Bioeng., 13, 795­813.
Verhulst R. (1838). Notice sur la loi que la population suit dans son
accroissement. Corr. Math. etPhys., A. Quetelet (Ed.), t­ X, 113.
Williams D., P. Yousefpour and B.H. Swanick (1984). On­line adaptive control
of a fermentation process. lEE Proc, 131, 4, 117­124.
 Appendix 2. ELEMENTS OF STABILITY THEORY 357

APPENDIX 2

ELEMENTS OF STABILITY THEORY

In this appendix we present an elementary and intuitive presentation of the

stability theory of nonlinear and linear time­varying systems. The reader is
referred to the books listed in the references for a comprehensive treatment of
the subject.

A2.1. Nonlinear Systems

Consider a nonlinear system described by the differential equation :

^ = f(x) x i ( x i , X 2 , . . . , X n ) e [R" Vt > 0 (A2.1)

Suppose that χ ξ 0 is an equilibrium state of the system :

f(0) Ξ 0 (A2.2)

Definition A2.1
The equilibrium state χ = 0 is asymptotically stable (a.s.) if there exists a
positive constant ε ι > 0 such that, if ||x(0)|| < ε ι , then :

lim||x(t)|| = 0 (A2.3)
t­>oo
 358 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

The Lyapunov methods are techniques for testing the stability of an

equilibrium state of a nonlinear system without having to calculate the
trajectories x(t).

Definition A2.2
A Lyapunov function W(x) for the system (A2.1) is defined as follows :
a) W(x) is a continuously differentiable scalar function of χ
b) W(x) is positive definite, i.e. W(x) > 0 Vx, W(x) = 0 if χ = 0
c) ^ W ( x ) = ­ ^ ^ ^ f i x ) is negative­definite along the solutions of (A2.1).
dt σ Χ

Tlieorem A2.1 (Direct Lyapunov's method)

If there exists a positive constant ε ι > 0 such that, if ||x(0)|| < ε ι , there exists a
Lyapunov function for the system (A2.1), then χ ξ 0 is an asymptotically stable
equilibrium state.

Definition A2.3
The equilibrium state χ ξ Ο is exponentially stable if there exist three positive
constants ε ι , C i , C2 such that, for all ||x(0)|| < ε ι , the solution x(t) of (A2.1) is
bounded as follows :

IIx(t)II<Ci exp (­Cgt)IIx(0) II Vt>0 (A2.4)

(Notice that exponential stability implies asymptotic stability).

 Appendix 2. ELEMENTS OF STABILITY THEORY 359

Theorem A2.2 (First Lyapunov method)

Consider the matrix : Δ = ί —1 (A2.5)

ν 3x 7χ =ο

­ If all the eigenvalues of the matrix Δ have negative real parts, then the
equilibrium state χ ξ ο is exponentially stable.
­ If any of the real parts of the eigenvalues of Δ are positive, then the
equilibrium state χ ξ 0 is unstable.
­ No conclusion may be drawn in case of eigenvalues having zero real parts.

Total stability

Suppose that the system (A2.1) is perturiDed as follows :

^ = f ( x ) + g(x,t) (A2.6)

where g(x,t) is a continuously differentiable perturbation function.

Defim'tion A2.4
The equilibrium state of (A2.1) is totally stable if there exist two positive
constants ε ι and E 2 such that, if ||x(0)|| < e i and ||g(x,t)|| < E 2 , then the solution
x(t) of (A2.6) is bounded for all t.

Theorem A2.3
If the equilibrium state x = 0 of (A2.1) is exponentially stable, then it is totally
stable.

Remark A2,1. A special case of (A2.6) occurs when the perturbation g(x,t) t
g(x)u(t) with u(t) an external input. In such a case, total stability is also called
BIBS (for Bounded Input Bounded State) stability.
 360 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

A2.2. Linear Time­Varying Systems

Consider a linear time­varying system described by the following non

stationary differential equation :

^ = A(t)x(t) t>to (A2.7)

with x = (xi, X2,..., Xn) e iR".

and A(t) a square nxn real valued matrix of continuously differentiable

bounded time functions.

We notice that χ Ξ 0 is necessarily an equilibrium state of the system (A2.7).

Furthermore, it is an isolated equilibrium state if A(t) is nonsingular for some t.

Definition A2.5
The equilibrium state χ Ξ Q of system (A2.7) is exponentially stable
(synonymous : uniformly asymptotically stable), if there exist two positive
constants Ci and C2, independent of to and x(to), such that x(t) is bounded as
follows for all to and x(to):

x(t) II < [Ci exp(­C2(t ­ to))] II x(to) II Vt > to (A2.8)

Tlieorem A2.4
If there exists a positive constant ε ι , such that for each t > 0, the eigenvalues of
A(t) all have real parts less than or equal to ­EI , then there exists another
constant £2 such that the equilibrium state χ Ξ 0 of the system (A2.7) is
exponentially stable whenever:

II ^ II < 82 Vt>0 (A2.9)

where |[.[| is the matrix norm induced by the vector norm used in (A2.8)
 Appendix 2. ELEMENTS OF STABILITY THEORY 361

Definition A2.6
A scalar function W(x,t) is decrescent in a closed regions ofthex­space
containing the origin χ Ξ 0, if a positive definite function Wi(x) exists such that,
for all X in S and all t,

|W(x,t)| <Wi(x)

Ttieorem A2.5
If the matrix A(t) is bounded, then the equilibrium state χ Ξ ο of system (A2.7) is
uniformly asymptotically stable if and only if there exists a time­varying,
quadratic, positive definite, decrescent Lyapunov function whose derivative
along the solutions of (A2.7) has a time­varying negative definite decrescent
quadratic form.

BIBS (Bounded Input Bounded State) Stability

Consider a linear time­varying system with an external input u(t) described by

the following differential equation :

^ = A ( t ) x ( t ) + B(t)u(t) (A2.10)
at

Ttieorem A2.6
If the equilibrium state χ Ξ 0 of (A2.7) is exponentially stable, if the matrix B(t) is
bounded (||B(t)|| < Mi), if the input u(t) is bounded (||u(t)|| < M), then the state of
the system (A2.10) is bounded as follows :

||x(t)||<cJ|x(0)IU^^ Vt>0
 362 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Second Order Systems

Technical Lemma A2.1

If λ 2 < λ ι < 0, then

1 λ 2 expiX^s) ­ λ 2 ο χ ρ (λ 28) I ds
Jo
= expa^t) ­ ο χ ρ ( λ ι ) + 2[ expa^o*) ­ expago*)] (A2.11)

Ι η ί λ ρ /λ ι )
with σ *= . \ forallt>a* (A2.12)
^1 "^2

Proof: the value σ * defined by (A2.12) is clearly the unique value such that:

λ ι expa^o*) = λ 2 ο χ ρ (λ 2σ *) (A2.13)

Then it is easy to see that:

I λ ι expiX^s) ­ λ 2 exp(X2s) ds
Ό
σ
[λ ι exp(Xis) ­ λ 2 exp(X2s)] ds

JXi expa^ s) ­ λ 2 expags)] ds (A2.14)

and expression (A2.11) follows readily.

QED

Consider the second order system :

^ = Ax(t)4­b(t) (A2.15)
dt
 Appendx 2. ELEMENTS OF STABILrTY THEORY 363

with X(t) = xi(ty (A2.16)

Lx2(t).

A = ­Ci 1 (A2.17)
­C2 0
b(t) = bi(ty (A2.18)
Lb2(t)

Technical Lemma A2.2

Assume that bi(t) and b2(t) are bounded :

|bi(t)| < B i Ib2(t)| <B2 (A2.19)

and that A has real distinct eigenvalues λ 2 < λ ι < 0 (i.e. C2 < Cy4).

Then :

,imsup|xi(t)|<j^||^5^

(A2.20)

λ ι +λ 2
lim sup I X2(t) I < Bi ­ ­ ^ ^ ^ B2 (A2.21)

Proof: the solution x(t) of A2.15 may be written as follows

x(t) = PJ(t)p­^x(0) = PJ(t­x)P"^b(x)dT (A2.22)

Jo

with ρ = 1 1 : the eigenvector matrix of A

"^2 ­ λ ι
 364 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

λ ι Ο
λ ι ­λ 2
1­λ 2 ­1

J(t) = Θ Χ Ρ (λ ι Ο Ο
Ο Θ Χ Ρ (λ 2ΐ )

For sufficiently large t (i.e. after the effect of initial conditions has disappeard),
the solution of (A2.15) is :

1
XL(t) = j ^ — j j ­ j^{[Xiexp(Xi[t­x]) ­ λ 2exp(λ 2[t­τ ])]b1(τ )

+ [exp(Xi[t­T]) ­ exp(X2[t­T])]b2(T)}dx (A2.23)

1
X2(t) = {λ 1λ 2[exp(λ 2[­t­τ ]) ­ exp(X2[t­T])]bi(t)
^1~^2 Jo
­I­ [λ 1exp(λ 1[t­τ ]) ­ X2exp(Xi[t­x])]b2(x)}dt (A2.24)

It is then easy to see that:

exp(Xi[t­x]) ­ expiXgtt­x]) > 0 (A2.25)

and λ 1exp(λ 1[t­τ ]) ­ λ 2exp(λ 2[t­τ ]) > 0 (A2.26)

We now examine xi(t) and X2(t) separately. Using (A2.12) (A2.26) and the
Cauchy­Schwartz inequality, we have from (A2.23) :

I xi(t) I + γ Λ ­ Γ { Β ι I λ ι [exp(Xi [t­τ ]) ­λ 2exp(λ 2[t­τ ])] I

Λ ι ­Λ ,2 Jo
­I­ B2 [exp(λ 1 [t­c]) ­ exp(λ 2[t­τ ])}dτ (A2.27)

and from Technical Lemma A2.1 :

 Appendix 2. ELEMENTS OF STABILITY THEORY 365

Ixi(t) l< T ­ V {Β ι [θ χ ρ (λ 2ί ) ­ expa^t) + 2{Θ χ ρ (λ ι σ *)

­ θ χ ρ (λ 2σ *)}] + B2 ^ θ χ ρ (λ ι Ο ­ 1 ­ ^(Θ χ ρ (λ 2ΐ ) ­ 1) } (A2.28)

with σ * given by (A2.12).

Hence taking the limit of the RHS of (A2.28) for t ^ o o , we obtain inequality
(A2.20) after some simple manipulations.

In an analogous way, using (A2.25) and (A2.26), we have from (A2.11) :

I X2(t) l< ­ j ^ ^ ^ { Β ι λ ι λ 2 ^(^expait) ­ 1 ­ (expa2t) ­ 1)

λ λ
+ B2 (expa2t) ­ 1) ­ ^ (expa^t) ­ 1) } (A2.29)

Hence taking the limit of (A2.29) for t ^ o o , we obtain inequality (A2.21).

Then expression (A2.20) and (A2.21) hold and the lemma is proved.
QED

Corollary A2,1
If C2 is chosen such that:

C2 = aC?/4, 0<a<1

then expression (A2.20)(A2.21) can be rewritten as follows

 366 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

­([1 ­^lΰ aV2^Γ Γ 5)

limsuplx,(t)U­^[|f:^|
t­^ C i V l ­ a LI 1 + V l ­ a J
­([i+/roi)/2VT^L
1 ­>Γ Ρ α 1 4B2
(A2.30)
1 ­ι ­Λ Γ Π α

lim sup I X2(t) I < Bi + B2 (A2.31)

straightforward by noting that:

—c
λ 2 = ­2^[1 +VT^]

References and Bibliography

Rouche N., P. Habets and M. Laloy (1977). Stability by

Springer Verlag, Applied Mathematical Sciences, New York.
Vidyasagar M. (1978). Prentice­Hall,Englewood
Cliffs.
Willems J.L (1970). Stability Nelson.
 Appendix 3. PERSISTENCE OF EXCITATION 367

APPENDIX 3

PERSISTENCE OF EXCITATION.
CONVERGENCE OF ADAPTIVE ESTIMATORS

In this appendix we introduce several definitions and theoretical results

concerning the persistence of excitation of regressors which is a key concept
for the analysis of adaptive estimators such as those which are described in
Chapters 3 and 4 of this book.

Definition A3.1
An mxn matrix 0(t), t > 0, of continuously differentiable bounded time functions
is persistently exciting if there exist positive constants α and ε such that:

al < I Φ (τ )φ \τ )ά τ Vt > 0 (A3.1)

Convergence of adaptive estimators

Definition A3.2
If the equilibrium state χ Ξ 0 of the time­varying linear system (A2.7) is
exponentially stable, then by extension the system itself is said to be
exponentially stable.
 368 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Theorem A3.1
If Φ (\) is persistently exciting, if Γ is a constant mxm positive definite matrix,
then the following linear time­varying system is exponentially stable :

^ = ­ro(t)o\t)e (A3.2)
dt

Theorem A3.2
If 0(t) is persistently exciting, if A(t) is a stable nxn matrix of bounded piecewise
continuous­time functions, if P(t) is a symmetric positive definite matrix of
bounded continuous­time functions such that dP/dt + PA + A^P is negative
definite, then the following linear time­varying system is exponentially stable :

"x" = A(t) (A3.3)

dt
θ L

Transfer of excitation

The above theorems are useful for demonstrating the convergence of adaptive
estimators if the "regressor" 0(t) is the state of stable linear filter. In such a
case, the following results are relevant.

Let 9(t) denote one arbitrary column of the matrix 0(t). Let u(t) be a scalar, real
valued, time­varying, continuously differentiable signal.

Definition A3.3
The signal u(t)is said to be sufficiently rich of order m if the following vector:

Π Τ
du dju d­^­^u
W(t) = u. (A3.4)
dt' dt^dt"^"^

is persistently exciting.
 Appendx 3. PERSISTENCE OF EXCITATION 369

Theorem A3.4
If each column of 0(t) is generated by a linear stable time invariant filter

^ = A9(t) + Bu(t) cpGlR^ (A3.5)

If the following matrix is full rank :

AB A^B A^'­^B] (43.6)

If u(t) is sufficiently rich of order m, then 0(t) is persistently exciting.

Remark : when the matrix (A3.6) is full rank, the system (A3.5) is said to be
"state reachable".

References and Bibliography

Mareels I.M.Y. and M. Gevers (1988). Persistency of excitation criteria for

linear, multivariable, time­varying systems. Math.
203­226.
Narendra K.S. and Annaswamy A.M. (1989) Stable adaptive systems.
Prentice­Hall, Englewood Cliffs.
 NOMENCLATURE 371

NOMENCLATURE

A : bioreactor cross­section (Section 1.10)

Α (ξ ) : state matrix of the linear tangent model
a,b,c,ai, a2 : constants
Ao, A i , A2 : (yield coefficient nonlinear combination) matrices
Β : diagonal matrix of the specific liquid­gas transfer rates
b,f : feed rate vectors
C : dissolved oxygen (concentration); output matrix (Chapter 4)
Ci (i = 1,2,3) : constants (Chapter 3)
Cs : oxygen saturation constant
D : dilution rate
Ε : enzyme
E i , E2 : activation energies
F : flow rate (Sections 1.1 and 5.7) or feed rate vector
Fg : aeration feed rate
Fr : substrate feed rate
Η : matrix of known functions of the state
h : adaptation law
I : inhibitor concentration
J : cost function
Κ : yield coefficient matrix
Kc Kp : specific growth rate constants of the models
kd : death coefficient
kj, kjj (i,JG N) : yield coefficient
kLa : mass transfer coefficient
KM : Michaelis­Menten constant
km : maintenance coefficient
 372 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

L : bioreactor length (Section 1.10); light intensity (Section 1.3);

output matrix (Chapter 3)
Mi,M2 : constant
Μ : number of reactions
Ν : number of components
0 : observability matrix
OTR : oxygen transfer rate
OUR : oxygen uptake rate
ρ : dimension of the matrix Κ
Ρ : product (concentration)
pH :pH
Psat ' product saturation concentration
q : number of measured components
Q : gaseous flow rate
R : ideal gas constant; symmetric matrix (Kaiman filter)
Π ι , Μ 2, ^22 · elements of R
Rj (i = 1 to 3 ) : elements of R
RQ : respiratory quotient
S : substrate (concentration)
Τ : sampling period; temperature (Chapter 1)
t : time
V : volume
V : vector
w : natural frequency ( = \/γ )
W : Lyapunov function
X : biomass (concentration)
y : output

Greek letters

α : specific reaction rate (vector)

β : specific liquid­gas transfer rate
 NOMENCLATURE 373

ε : singular perturbation variable (Section 1.8) or linear filter output

(Section 3.4.2)
Φ : regressor matrix
Γ : gain matrix
Υ ΐ 'Ύ 2'Ύ 3 ­ elements of Γ
η ,δ : constants
φ : reaction rate
λ : forgetting factor
μ : specific growth rate
μ * : maximum specific growth rate
μ ο : Haldane growth rate parameter
ν : specific production rate
v i ,V2 : eigenvalues of the 2x2 matrix A
Π : P/Psat (Section 1.8)
θ : unknown parameter vector
ρ : non­growth associated specific production rate (Section 1.3) or
vector of unknown functions
σ : design parameter
Ω : gain matrix
ω ,γ : tuning parameters
ω ι ,co2 : elements of Ω
ξ : biochemical reaction component (vector)
ψ ,ψ ο : linear filter output
ζ : damping coefficient
Σ ­1 : weighting matrix

Subscripts

t : time index
in : influent
out : effluent
i,j : number indices
a,b,s,f: partition indices
 374 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

min, max : minimum, maximum

m : measured
R : recycle
Τ : total
W : waste

Superscripts

τ : transposed matrix or vector

: estimate
­ : error
: desired

Mathematical notations

Σ : sum

Π : product

V : for all

lim : limit for time tending to infinity

t­>oo

min : minimum

max : maximum

sup : supremum

diag{.} : diagonal matrix

^ : summation on the reactions with index j which involve the component

with index i

e : belongs to
 NOMENCLATURE 375

4­ time derivative
dt

•^, ^ : partial derivative with respect to time t and length ζ

dt σ Ζ
arg min F(x): value of the argument χ for which F(x) is minimum with
respect to χ

elementary matrix : matrix which contains only terms equal to 1 or 0

A­i : inverse of the (square) matrix A, i.e. such that Α Α ­"· = Α ­"Ά = I (where I is
the identity matrix)
 INDEX 377

INDEX

Associated
growth 9; 27
non­growth 9; 23; 27
Autocatalyst 20; 33; 84
Catalyst 19
Control
adaptive 94; 254; 277; 279; 287; 301; 305; 311; 328
linearizing 94; 254; 261; 266; 276; 277; 279; 283; 287; 296; 300; 323
Dynamics 3; 7; 30
transport 31
Enzymes 20; 37
Equilibrium 60; 104
stable 68; 109; 110; 309
unstable 68
Estimation
linear regression 151; 175; 186; 278; 289
observer­based 147; 160; 173; 177; 181; 192; 205; 219; 224; 278
parameter 87; 92; 202; 216; 231; 244; 278; 279; 281
state 103; 116; 231; 300
General dynamical model 30; 93; 103; 146; 202; 218; 257
Identifiability 87; 243
Inhibition
substrate 14; 67; 319
product 15; 86; 91
Kinetics 11; 31; 34; 84; 94; 95
 378 ON­LINE ESTIMATION AND ADAPTIVE CONTROL OF BIOREACTORS

Maintenance 5; 21; 31
Mass transfer coefficient 10
Metabolites 7; 91
Microbial death 5; 2 1 ; 31
Microorganisms 4; 20
Observer
adaptive 231; 304
asymptotic 109; 116; 146; 258; 288; 291; 303
exponential 105; 109
Kalman 109; 175; 231
Luenberger 109; 231
Observability 105
Order reduction 46; 49
0TR9
OUR 9; 247
pH 16; 94; 136
Process
anaerobic digestion 24; 4 1 ; 50; 73; 117; 121; 127; 186; 215; 240; 258; 268;
271 ;284; 324
lactic fermentation 27; 42; 60; 85; 194
simple microbial growth 4; 21; 35; 47; 62; 67; 85; 106; 111; 203; 219; 224;
233; 239; 306;311
PHB 26; 42; 45; 133; 236; 244
yeast growth 9; 23; 40; 180; 210; 273; 295
Reaction rate 19; 29; 32; 84; 116; 146; 202; 216; 231
Reactors
batch 6; 120; 191; 211; 235; 311; 320
CSTR 7; 113; 120; 124; 127; 168; 170; 186; 240; 284; 327
fed­batch 6; 120; 133; 141; 180; 235; 264; 273; 311
fixed bed 71
Respiratory quotient (RQ) 26; 133
Sampling 141; 150; 155; 173; 285
Singular perturbation 46; 266; 296; 301
Specific growth rate 5; 8; 10; 34; 68; 86; 91; 160; 180; 181; 188; 207; 327
Blackman law 228
 INDEX 379

Contois law 14; 36; 63; 85; 111

Haldane law 14; 68; 124; 170; 319
Monod law 11; 86; 87; 168; 317; 327
Specific production rate 8; 63; 317
Specific reaction rate 33; 84
Stability 52; 120; 142; 150; 155; 162; 173; 177; 264; 276; 279; 304; 305; 306;
315; 357
State partition 44; 117; 145; 206; 222; 235; 239; 267; 269; 301
nice 46; 207; 222; 301
State transformation 44; 46; 109; 206; 208; 215; 216
Temperature 16; 94
Wash­out 63