OPTIMAL CONTROL APPLICATIONS AND METHODS

Optim. Control Appl. Meth. 2010; 31:289–301

Published online 10 July 2009 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/oca.905

Optimal time switching control for multi-reaction batch process

D. Mazouni1, 2 , J. Harmand1, 2, ∗, † , A. Rapaport2, 3 and H. Hammouri4

1 INRA, Laboratoire de Biotechnologie de l’Environnement, avenue des Etangs, Narbonne, France
2 INRA-INRIA Research Project, UMR Analyse des Systèmes et Biométrie, 2 place Viala, Montpellier, France
3 UMR Analyse des Systèmes et Biométrie, 2 place Viala, Montpellier, France
4 Laboratoire d’Automatique et de Génie des Procédés, Université Claude Bernard, Lyon, France

SUMMARY
This paper presents an approach for solving minimal time problems for biological batch models with switching controls.
Multi-reaction processes operated in a sequence of operating modes are considered, where the control variable is the switch
between different operating modes. In order to determine the optimal switching instants that minimize the final time to
reach the target, a reachability analysis and a study of dynamical properties are first investigated. This leads to a reduced
model and the characterization of reachable switching sequences. Optimization tools are then used to determine the optimal
switching instants. Copyright q 2009 John Wiley & Sons, Ltd.

Received 2 November 2006; Revised 12 September 2008; Accepted 30 May 2009

KEY WORDS: sequencing batch reactors (SBR); biological treatment; optimal control; switching system; controllability

1. INTRODUCTION operation can be preferred since the hydrodynamic

conditions are then completely controlled. Such an
Biological processes are often used to treat wastewater. operation mode is called a batch operation. The opti-
The principle of such processes is first to transform mization of batch and fed-batch biological processes
soluble organic compounds (the pollutants also called has been studied with success for single reaction
substrates) into a solid form (biomass) under adequate described by a single triplet of variables: biomass,
environmental conditions, and then to separate the solid substrate and product. These processes are modeled by
phase (the biomass) from the cleaned water (the liquid four mass-balance equations and the control variable
phase with substrate concentrations less than normative is most of the time the volumetric substrate feed-rate
constraints) before rejecting it into the environment. (see, e.g. [1–3]). Here, we are concerned with the
One of the most important limiting factors of such optimization of a multi-reaction sequencing batch
processes is the separation of the solids from the liquid bioprocess. By ‘multi-reaction bioprocess’, we refer to
phase. In order to optimize this separation step, batch biological processes requiring different aeration condi-
tions in order to remove from wastewater both organic
∗ Correspondence to: J. Harmand, INRA, Laboratoire de Biotech-
carbon and nitrogen compounds. Indeed, in such a
nologie de l’Environnement, avenue des Etangs, Narbonne, case, it is necessary to perform both carbon removal
†
France. (either in aerobic or anoxic conditions), nitrification
E-mail: [email protected], [email protected] (only in aerobic conditions) and de-nitrification (only
Contract/grant sponsor: INCO programme of the European in anoxic conditions). More particularly, we propose a
Community; contract/grant number: ICA4-CT-2002-10012 solution to a problem of minimal time control for the

Copyright q 2009 John Wiley & Sons, Ltd.

 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
290 D. MAZOUNI ET AL.

optimization of a biological sequencing batch reactor each autonomous system (aerobic or anoxic mode),
(SBR). Since our manipulated variable is the aeration which allows us to perform the controllability anal-
mode of operation (presence or absence of oxygen), ysis. Then, we enlarge the discrete control set to a
such a process can be viewed as a switching system. continuous one in order to make convex the set of
Switching systems are a particular class of hybrid admissible trajectories. Next, we analyze the optimal
systems that consist of a family of continuous-time solution using Pontryagin’s maximum principle and
dynamical systems and a control (here the aeration) characterize the set of bang–bang trajectories when they
that determines the active system at each time. Many exist. Finally, for the other trajectories (trajectories with
results have appeared in the literature regarding optimal a singular arc) a parameterization of switching instants
time control problems for non-linear hybrid systems is introduced and the cost function is optimized using
and switching systems. The major difficulty relies numerical algorithms. Additional assumptions about
on the existence of an optimal control. Most of the the number of switches can be added to solve practi-
solutions consist in reformulating the problem and cally the problem.
augmenting it into a larger family of systems for
which the necessary and sufficient conditions of opti-
mality are fulfilled. In [4] a parameterization of the 2. DESCRIPTION OF THE SBR
switching instants to transform the switching system
into an equivalent system is proposed. The optimal Let us consider a bioprocess operated in sequencing
solution is then determined using a numerical approach mode used for carbon and nitrogen removal from
based on the derivative of the cost function. In the wastewater. This biological reactor requires different
case of switching between two systems only, Bengea aerating conditions to be applied sequentially.
and DeCarlo [5] propose to transform the discrete
switching control {0, 1} into a continuous control • Anoxic mode: In the absence of oxygen and the
in the interval [0, 1]. Hence, this approach avoids presence of both nitrates and carbon, de-nitrifi-
restrictions in the number of switchings [4, 6, 7]. The cation takes place. A set of microorganisms (also
augmented system is constructed such that neces- called the biomass b1 ) consumes simultaneously
sary conditions of optimality are fulfilled [5, 8]. In the carbon (x) and the nitrates (y) with respect to
[9, 10], it is shown that the optimal solution for such a the following synthetic reaction scheme:
system, when it exists, is a concatenation of a limited r3
number of ‘bang-singular-bang’ controls. To adapt x + y −→ b1 (1)
this solution to our initial problem, for which singular Notice that this reaction only takes place if both
control is not allowed, it is possible to approximate x and y are available.
the singular control by successive switches (for more • Aerobic mode: In the presence of oxygen, two set
details see [8]). However, there are many systems of microorganisms (b1 and b2 ) consume, respec-
for which this solution cannot be applied because of tively, the carbon and the nitrogen (z) through two
some physical limitation or model robustness limi- independent reactions: carbon removal (reaction 2)
tation at high switching frequency. Moreover, the and nitrogen removal (reaction 3) where nitrogen
difficulty of the problem increases with the number is converted into nitrates:
of variables and the complexity of the problem
gets beyond control when there are more than three r1
x + O2 −→ b1 (2)
variables. r2
In this paper, we focus on minimal time control z + O2 −→ b2 + y (3)
problems of a biological SBR where the systems asso-
ciated with the presence or the absence of the oxygen This process is operated by alternating two operating
have autonomous dynamics. To solve this problem, modes in order to eliminate the substrates, i.e. to reach
we first propose to study the dynamical behavior of at final time desired concentrations of substrates (that

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OPTIMAL TIME SWITCHING CONTROL 291

define the target). The optimal control problem to be 3. MODELING AND SIMPLIFICATION
solved is the minimization of the total duration required
to reach the target. We aim to find the optimal switching Biological processes are modeled using the standard
between aerobic and anoxic modes. mass-balance conservation law (see for instance [12]).
From a practical point of view, nitrogen removal We also assume, according to [13], that the yield coef-
in batch reactors is usually carried out using two ficient of biomass b1 is the same for both aerobic and
successive steps only: de-nitrification and nitrifica- anoxic conditions.
tion. The process takes place in this order to take With respect to carbon and nitrogen removal
advantage of the presence of organic carbon in the described by reactions (1)–(3), two models are derived:
input (recall that it is necessary to have simultaneous
• Aerobic model:
x and y for the de-nitrification to take place and
to avoid an external addition of carbon during the x
ḃ1 = r1 (x, b1 ) = 1 max b1 (4)
de-nitrification phase [11]). On the other hand, the Kx + x
processes are theoretically designed for a given range z
of input concentrations. The tank volume and the ḃ2 = r2 (z, b2 ) = 2 max b2 (5)
Kz + z
duration of the phases are pre-established. However,
when operated with significant variations in influent x
ẋ = −k x1r1 (x, b1 ) = −k x1 1 max b1 (6)
concentrations, no guarantee is given for the effluent to Kx + x
reach the desired concentrations. For such processes, z
notice that some reactions are mutually interdependent ẏ = k y2r2 (z, b2 ) = k y2 2 max b2 (7)
Kz + z
and the final concentrations have to reach terminal
levels defined by a norm that constitutes the target. z
ż = −k z2r2 (z, b2 ) = −k z2 2 max b2 (8)
Prior to any control design, it is then important to Kz + z
conduct a controllability analysis in order to obtain
the set of initial conditions for which there exists • Anoxic model:
a control sequence driving the process toward the
ḃ1 = r3 (x, y, b1 )
target set. To handle this problem we proceed as
follows: x y
= 1 max  b1 (9)
Kx + x K y + y
• For each operating mode (presence or absence
of oxygen) the batch process was modeled ḃ2 = 0 (10)
using standard mass-balance equations including
biological kinetics. The dynamical behaviors of ẋ = −k x1r3 (x, y, b1 )
the two systems are studied and the determi- x y
= −k x1 1 max  b1 (11)
nation of an invariant set in the state space is Kx + x K y + y
used to reduce the model.
• The target being described as a set of terminal ẏ = −k y3r3 (x, y, b1 )
state constraints, the reachable set is determined. x y
Admissible solutions are analyzed to estimate the = −k y3 1 max  b1 (12)
Kx + x K y + y
number of switchings necessary for the system to
reach the target. ż = 0 (13)
• The minimal time problem is formulated. An
initial estimate of optimal trajectory was deter- It is assumed that during the aerobic phase, the
mined using Pontryagin’s maximum principle oxygen concentration is controlled and fixed at a
and numerical algorithms are used to obtain non-limiting (constant) value. In addition, we assume
suboptimal control law. that the switches between aerobic and anoxic phase

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
292 D. MAZOUNI ET AL.

(and vice versa) are instantaneous (Comparing with the We assume that b1 (t0 ) and b2 (t0 ) are strictly positive.
biological time scale, time to switch when the number Otherwise, x(t) = x(t0 ) and x(t) = x(t0 ). This means
of switches is limited, is expected to be negligible). that without biomass the reactions do not take place in
the reactor.
Definition 1 Now, one can replace expressions of b1 and b2
U = {0, 1} is the set of the control variable that corre- given by Equations (19) and (20) to obtain the reduced
sponds to the switching signal: u = 1 for aerobic mode dynamics
and u = 0 for anoxic mode.
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
The two dynamics can be gathered into the following ẋ f (x) f (x)g(y)
matrix form: ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ ẏ ⎦ = ⎣ −h(z)⎦ u + ⎣  f (x)g(y)⎦ (1−u) (21)
⎡ ⎤ ⎡⎡ ⎤ ⎡ ⎤⎤
ḃ1 0 0 1 1 0 −1 ż h(z) 0
⎢ ⎥ ⎢⎢ ⎥ ⎢ ⎥⎥
⎢ ẋ ⎥ ⎢⎢ 0 0 −k x1⎥ ⎢ −k x1 0 k x1 ⎥⎥
⎢ ⎥ ⎢⎢ ⎥ ⎢ ⎥⎥ where
⎢ ⎥ ⎢⎢ ⎥ ⎢ ⎥⎥
⎢ ẏ ⎥ = ⎢⎢ 0 0 −k y3⎥ +u(t) ⎢ 0 k y2 k y3 ⎥⎥
⎢ ⎥ ⎢⎢ ⎥ ⎢ ⎥⎥ k y2 k y3 x(x −m 1 )
⎢ ⎥ ⎢⎢ 0 0 0 ⎥ ⎢ 0 0 ⎥ ⎥ = , = , f (x) = 1 max
⎣ ḃ2 ⎦ ⎣⎣ ⎦ ⎣ 1 ⎦⎦ k z2 k x1 Kx + x
ż 0 0 0 0 −k z2 0
y z(z −m 2 )
⎡ ⎤ g(y) =  and h(z) = 2 max
r1 (·) Ky + y Kz + z
⎢ ⎥
× ⎣ r2 (·)⎦ (14)
A compact form of the model is given by
r3 (·)
Ẋ = F(X )+ G(X )u
from which one can easily derive the following rela- (22)
tionships: X (t0 ) = X 0 = [x0 z 0 y0 ]t

ẋ +k x1 ḃ1 = 0 (15) where X t = [x, y, z] is the state vector and F and G are
the analytic vector fields in R3 .
ż +k z2 ḃ2 = 0 (16)
and by the integration, we get
4. ACCESSIBILITY AND REACHABILITY
m 1 = x +k x1 b1 = x(t0 )+k x1 b1 (t0 ) (17) ANALYSIS
m 2 = z +k z2 b2 = z(t0 )+k z2 b2 (t0 ) (18) 4.1. Dynamical behavior
where m 1 and m 2 depend only on the initial conditions In this section, the dynamical behavior of system (22)
and remain constant throughout the batch treatment, is studied in order the characterize the behavior of the
whatever is the control. This is a consequence of the trajectories.
mass conservation principle which states that all quan-
tities of x and z that disappear are transformed into Proposition 1
biomass. [3, 14]. Thus, one obtains a linear relation Consider X (t, X 0 , u) the solution of the differential
between x and b1 (resp. z and b2 ): equation (22), where X 0 is a positive vector, b1 (t0 ) > 0
and b2 (t0 ) > 0. Let x(·), y(·) and z(·) be the coordinates
x −m 1
b1 = − (19) of the state vector X (·).
k x1
• When u = 1: t  → x(t, X 0 , 1) and t  → z(t, X 0 , 1)
z −m 2
b2 = − (20) are two positive decreasing maps that tend toward
k z2 zero as t tends to infinity.

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OPTIMAL TIME SWITCHING CONTROL 293

• When u = 0: t  → x(t, X 0 , 0) and t  → y(t, X 0 , 0)

are two positive decreasing maps.

Proof of Proposition 1
Notice first that since initial condition of dynamics (14)
is positive, its solution remains positive at any t > t0 .
• When u = 1, one has m 1 = x +k x1 b1 > x (resp.
m 2 = z +k z2 b2 > z) at any time, so as long as x
(resp. z) stays positive, we have f (x) < 0 (resp.
h(z) < 0). One can deduce that ẋ < 0 (resp. ż < 0)
when u = 1 and as long as x (resp. z) stays Figure 1. Subsets (X 0 ), (X 0 ) and (X 0 ).
positive. Moreover, one has ẋ = 0 (resp. ż = 0)
when x = 0 (resp. z = 0). Consequently, x(·) and
z(·) are decreasing with respect to time and tend 4.2. Accessibility and reachability
asymptotically toward zero. Definition 2 (Accessibility)
• In the same way, one has ẋ < 0 (resp. ẏ < 0) when Consider the system (22) and let X (t, X 0 , u) be its
u = 0. Then, t  → x(t, X 0 , 0) and t  → y(t, X 0 , 0) maximal solution. The set of accessible points at time
are decreasing maps.  T > 0 is
Let us now introduce the following notations: 
A+ (X 0 , T ) = X (T, X 0 , u) (23)
u∈{0,1}
(X 0 ) = {X ∈ R3+ /y − y0 = (x − x0 )−(z − z 0 )}
and the accessibility set is given as
(X 0 ) = {X ∈ R3+ /x < x0 , z < z 0 , y − y0 = −(z − z 0 )}  +
A+ (X 0 ) = A (X 0 , T ) (24)
×∪{X 0 } T >0

Similarly, we denote by A− (X f , T ) the set of initial

(X 0 ) = {X ∈ R3+ /x < x0 , z = z 0 , y − y0 = (x − x0 )} conditions from which it is possible to reach X f at time
T . A− (X f ) is the set of all reachable points.
×∪{X 0 }
Definition 3 (Reachability set)
Corollary 1 Consider the target set
For u(·) piecewise constant, the solution of system (22)
C N = {X = [x, y, z]t /xx N , yy N −z, zz N }
fulfills the following properties:
• ∀X 0 ∈ R3+ , X (t, X 0 , 1) ∈ (X 0 ), ∀t > 0 where x N , y N and z N are strictly positive numbers
(that represent the threshold allowed to be rejected
• ∀X 0 ∈ R3+ , X (t, X 0 , 0) ∈ (X 0 ), ∀t > 0.
from wastewater treatments). Denote also that X N =
Proof of Corollary 1 [0, y N , 0]t .
When u = 1, one can write ẏ = −ż and so one has The reachability set (C N ) is the set of points from
y − y0 = −(z − z 0 ) at any time. Moreover, from Propo- which the target can be reached, using the dynamics (22):
sition 1, one has xt < x0 and z t < z 0 at any time t > 0,
(C N ) = {X ∈ A− (X f )/ X f ∈ C N }
so we deduce that X (t, X 0 , 1) ∈ (X 0 ) (Figure 1).
If u = 0, one has ż = 0 and ẏ = ẋ and then properties
4.3. Study of the reachability set
z = z 0 and y − y0 = (x − x0 ) are fulfilled at any time.
Moreover, from Proposition 1, one has x t < x0 at any The study of the dynamical behavior shows that the set
time, so we deduce that X (t, X 0 , 0) ∈ (z 0 ).  of trajectories belonging to a family of planes (X )

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
294 D. MAZOUNI ET AL.

and (X ) have there backward solutions X (−t, X f , u)

belonging to the same planes. Consequently, we char-
acterize the reachability set, considering the target C N
as a final set of points X f , and by a backward combina-
tion of planes  and , we construct sets of reachable
points A− (X f ).

Definition 4
It is convenient to introduce the sets that are represented
in Figure 3.

 A (C N ) = {X ∈ R3+ , z−1 y N , y < x −z + y N }

 B (C N ) = {X ∈ R3+ , z > −1 y N , y < x −z

+(y − y N )}\ D (C N )
Figure 2. Sets  A (C N ) and C (C N ).
C (C N ) = {X ∈ R3+ , z > −1 y N , y < x −z +(y − y N )}
 D (C N ) = {X 0 ∈ C (C N )/∃t > 0,
∀u ∈ {0, 1}, z(t, X 0 , u) > y N and
y(t, X 0 , u) < x(t, X 0 , u)−z(t, X 0 , u)}

Notice that  D (C N ) is the set of initial conditions

X 0 from which the trajectory X (t, X 0 , u) leaves the set
C (C N ) before reaching the set  A (C N ).  D (C N ) can
be determined numerically. We shall also consider the
following sets (Figure 2):

1 (C N ) = {X ∈ R3+ /z−1 y N , y < −z + y N }

2 (C N ) = {X ∈ R3+ /z−1 y N ,
−z + y N y < x −z + y N }
Figure 3. Sets 1 (C N ), 2 (C N ),  B (C N ) and  D (C N ).

Proposition 2
The reachability set is given by the following union of
sets: one has clearly (X 0 ) ⊂ C N ∨(X 0 ) ⊂ C N . Since
X (t, X 0 , 1) ⊂ (X 0 ) and X (t, X 0 , 0) ⊂ (X 0 ) for any
(C N ) =  A (C N )∪ B (C N ) t > 0, one has also X (t, X 0 , u) ⊂ C N , for any t and u
we have X (t, X 0 , u) ⊂ C N .
Proof of Proposition 2
Let us divide  A (C N )∪ B (C N ) into three sets:
We first show that the target set is invariant under the
1 (C N ), 2 (C N ) and  B (C N ), as depicted in Figure 3.
dynamics, whatever is the control (Figure 3).
Now, we shall show that
Let X 0 be an initial condition belonging to C N .
One has x0 < x N , z 0 < z N and y0 < y N −z 0 . Then, (C N ) = 1 (C N )∪2 (C N )∪ B (C N )

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OPTIMAL TIME SWITCHING CONTROL 295

Figure 4. Illustration of X (t, X 0 , 0) when X 0 ∈ 2 (C N ).

Let X 0 be an initial condition belonging to  A (C N )∪

 B (C N )\C N , and distinguish three cases: Figure 5. Reachable trajectories.
1. If X 0 ∈ 1 (C N ) then X (t, X 0 , 1) ⊂ (X 0 ) when
t → ∞, and X (t, X 0 , 1) → [0, y < y N , 0] ∈ C N Hence, X (t, X 0out , u)∩C (C N ) = ∅ for any u
according to Proposition 1. and since C N ⊂ C (C N ) one has X (t, X 0out , u)∩
2. If X 0 ∈ 2 (C N ), then X (t, X 0 , 0) ⊂ (X 0 ) when C N =∅.
t → ∞, according to Proposition 1. One get : (cf. 2. If X 0out ∈  D (C N ), then by definition we have
Figure 4) X (t, X 0out , u) ∈  D (C N )∪C (C N ) ∀u ∈U, ∀t > 0
y(t) → 0 and X (t, X 0 , 0) → [x̃, 0, z 0 ]t ∈ 1 (C N ) so X (t, X 0out , u) cannot reach the target C N .
with x̃ > 0.
or x(t) → 0 and X (t, X 0 , 0) → [0, ỹ, z 0 ]t ∈ Finally, we have shown (C N ) ⊂  A (C N )∪
1 (C N ) with ỹ > 0.  B (C N ). 
Let t1 > 0 be such that X 1 = X (t1 , X 0 , 0) ∈
Corollary 2
1 (C N ) and t2 > 0 such that X 2 = X (t2 , X 1 , 1) ∈
From any initial condition X 0 ∈ (C N )\1 (C N ), the
C N . Then one has X (t f , X 0 , u) ∈ C N for t f =
target can be reached using at most one anoxic phase.
t1 +t2 .
3. If X 0 ∈  B (C N ), then by definition there exists Proof of Corollary 2
t1 > 0 such that X (t1 , X 0 , 1) ∈ 2 (C N ). Then, In order to reach the target, the trajectories have to
there exists t f > t1 such that X (t f , X 0 , u) reach first the set 1 (C N ). Notice that each trajectory
∈ CN . X (t, X 0 , 1) stays parallel to (X N ), and consequently
Thus, we have shown  A (C N )∪ B (C N ) ⊂ (C N ). the set 1 (C N ) is reached using one anoxic phase only.
Let X 0out ∈ (C N ) and consider two cases: X 0out ∈ Then, accordingly to Proposition 1, an anaerobic phase
C (C N ) and X 0out ∈  D (C N ). is necessary to reach the target if z(0) > z N (Figure 5).
Notice that (C N )∪ D (C N ) = C (C N ) and (X N )
split R3+ into C (C N ) and its complement.
5. OPTIMAL CONTROL
1. If X 0out ∈
/ C (C N ), then one has X (t, X 0out , 1) ⊂
(X 0 ), but (X 0out )∩C (C N ) = ∅ (because
out
5.1. About optimal control of switching autonomous
plans (X 0out ) and (X N ) are parallel). Thus,
systems
the trajectory X (t, X 0out , 1) does not cross (X N )
and does not reach C N . Switching systems belong to a particular class of hybrid
In the same way, X (t, X 0out , 0) ⊂ (X 0out ), but systems. They are composed of several systems and
(X 0out )∩C (C N ) = ∅ (because (X 0out ) and switching laws that determine the active system at each
(X N ) are parallel), and X (t, X 0out , 0) does not instant in time. Many real chemical and biological
cross the attainable set. processes can be modeled as such systems.

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
296 D. MAZOUNI ET AL.

We focus here on the particular case of a set of two with the boundary condition
autonomous systems where the system 1 is active on
intervals i=0...n [t
2i , t2i+1 ), whereas the system 2 is
∗ (t ∗ ) ∈ (T(X ∗ ) (C))⊥ (29)
active on intervals i=0...n [t2i+1 , t2i+2 )
where the Hamiltonian H being defined by

Ẋ = F1 (t, X (t)) when t ∈ [t2i , t2i+1 )
i=0...n H (X, , u) = t (F(X )+ G(X )u)+0 (30)
 (25)
Ẋ = F2 (t, X (t)) when t ∈ [t2i+1 , t2i+2 ) fulfilled
i=0...n

where 2n +1 is the number of switches for t ∈ [t0 , t f ] H (X ∗ (t), ∗ (t), u ∗ (t))

with t f = t2n+2 , t1 –tn are the switching instants to be
determined in order to minimize the functional cost = min H (X ∗ (t), ∗ (t), u) = 0 (31)
tf u∈U
t0 dt.
Such switching systems can be embedded into The optimal control u ∗ (t) is defined for almost any
control affine systems [5] t ∈ [0, t ∗ ] as follows:
Ẋ (t) = F(X (t))+ G(X (t))u(t) (26) u min if t (t)G(X (t)) > 0
by introducing a control taking continuous value in u ∗ (t) = u s if t (t)G(X (t)) = 0 (32)
the interval u ∈ [u min , u max ] such that F1 (t, X (t)) = t
u max if  (t)G(X (t)) < 0
F(X (t))+ G(X (t))u min and F2 (t, X (t)) = F(X (t))+
G(X (t))u max . Optimal control theory can then be where u s is a singular control derived by solving the
applied to determine optimal control laws by mini- following set of equations:
mizing the cost function. Afterwards, one can look for
optimal laws that take values of u min and u max only. d t
( (t)G(X (t))) = 0
Consider now the control affine system in Rn of dt
the form (26) where F and G are two analytical ..
vector fields in Rn , u is a measurable map defined . (33)
on R+ that takes values on the compact convex set dk t
U = [u min , u max ]. For X 0 ∈ Rn we write X (t, X 0 , u) for ( (t)G(X (t))) = 0
the solution of the differential equation with the initial dt k
condition X 0 at t = t0 . k being chosen such that the control variable u appears
Let C be a regular sub manifold in Rn and T(X ) C explicitly in the kth derivative of the switching function
the tangent space of C at the point X . t  → t (t)G(X (t)).
We denote X ∗ (t, X 0 , u ∗ ) a minimum time trajec- In the case where the optimal solution is the concate-
tory driving the initial state X 0 to the target C in nation of bang–bang arcs only (i.e. it takes values u min
minimal time t ∗ , (i.e X ∗ (t ∗ , X 0 , u ∗ ) ∈ C) . From or u max only), the optimal solution can be applied to
Pontryaguin’s maximum principle, there exists a triplet the original switched system. If the optimal trajectory
(X ∗ (t), ∗ (t), u ∗ (t)), where ∗ (t) = 0, and a non- contains a singular arc (i.e. it takes values different from
negative number 0 verifying the following system of u min or u max during a non-null interval of time), the
equations (PMP) (see for instance [15]): optimal solution has to be approximated for the original
switching problem.
*H ∗
Ẋ ∗ (t) = (X (t), ∗ (t), u ∗ (t)) (27) In the following, we define the switching function
*
(t) = t (t)G(X (t))
˙∗ *H ∗
 (t) = − (X (t), ∗ (t), u ∗ (t)) (28)
*X along any trajectory of Equations (27) and (28).

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OPTIMAL TIME SWITCHING CONTROL 297

5.2. Optimal control problem formulation Proposition 3

From any initial condition belonging to 1 (C N )\C N ,
Consider again the dynamical system defined in (21).
the optimal solution is u ∗ (t) = 1 for any t ∈ [t0 , t f ].
In this following, we consider only trajectories with
initial condition belonging to the reachable set (C N ): Proof of Proposition 3
Notice first that for such initial conditions, it is not
Ẋ = F(X )+ G(X )u possible to reach the target using a control u(t) = 0 for
(34)
X (t0 ) = X 0 ∈ (C N ) any t0. Thus, either the control u(t) = 1 for any t0
is optimal, or the optimal solution consists in at least
The problem to be solved here is to find an optimal one switch.
control u ∗ (·) minimizing the total reaction time t reac = If the optimal solution reaches Px ∪ Pz , then the
t0 −t f , where X (t f ) ∈ C N , with C N being the target following transversality conditions are fulfilled, accord-
given in Definition 3. Accordingly to Proposition 1 and ingly to Equation (29): 2 (t f ) = 0, 1 (t f )0 and
Corollary 2, we consider only trajectories with at most 3 (t f )0. The switching function becomes:
one anoxic phase, and distinguish two cases:
(t f ) = 1 (t f )( f (x(t f ))(1− g(y(t f ))))
• the target is reached without any switch of the
control, +3 (t f )h(z(t f )) (35)
• the target is reached with only one anoxic phase
(i.e. with one or two switches). Since (t) = 0 for any t, f (x) < 0, h(z) < 0 and
0 < g(y) < 1 we have (t f ) < 0 so one has u ∗ (t f ) = 1
Denote Px , Py and Pz the three face of the target according to Equation (32).
C N (cf. Figure 6). Notice the following properties of With u(·) = 1 on an interval [t, t f ] with t < t f , the
the trajectories: adjoint vectors are given by
• If a trajectory reaches the first face Py , then neces-
* f (x(t))
sarily the control u(t) ends with an anoxic phase ˙ 1 (t) = −1 (t) (36)
u = 0. *x
• when the control ends with an aerobic phase u = 1,
the trajectory reaches first one of the two faces Px ˙ 2 (t) = 0 (37)
or Pz . *h(z(t))
˙ 3 (t) = −(2 (t)+3 (t)) (38)
*z
5.2.1. Solution without switching.
with boundary conditions 2 (t f ) = 0, 1 (t f )0 and
3 (t f )0. The solutions of the differential equations
(36)–(38) verify 2 (t) = 0, 1 (t)0 and 3 (t)0. Thus,
the switching function is negative and its sign does not
change in the interval [t, t f ] whatever are (t, t f ) such
that 0t < t f . This means that X (., X 0 , 1) verifies the
(PMP) conditions and is a candidate to be an optimal
solution.
Let us now consider the case where the optimal
control at a time t1 is u ∗ (t1 ) = 0, which means that the
switching function verifies

(t1 ) = (1 (t1 )+2 (t1 )) f (x(t1 ))

Figure 6. Trajectories without switching. +(g(y(t1 ))) > 0 (39)

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
298 D. MAZOUNI ET AL.

where f (·) and g(·) are positive referring to the following properties:
Proposition 1. We deduce that 1 (t1 )+2 (t1 ) is
• If X 0 ∈ 1 (C N ) then t0 = t1 = t2 , t f being defined
strictly positive. From the adjoint equations
as the time at which the trajectory X (t f , X (t2 ), 1)
reaches the target.
*
˙ 1 (t) = −(1 (t)+2 (t)) f (x)g(y) (40) • If X 0 ∈ 1 (C N ), t2 is given by the time at which
*x the trajectory X (t2 , X (t1 ), 0) reaches 1 (C N ). The
* plane (X N ) is playing the role of a switching
˙ 2 (t) = −(1 (t)+2 (t)) g(y) f (x) (41) surface because it is the upper boundary of the set
*y 1 (C N ).
˙ 3 (t) = 0 (42) • If X (t2 , X (t1 ), 0) reaches Py , then t f = t2 .
Finally, the only unknown parameter is t1 . In the
we deduce the following property: following, we study first the duration of the anoxic
phase only.
˙ 1 (t)+˙ 2 (t)
5.2.3. Duration of the anoxic phase. Between t1 and t2 ,
= −(1 (t)+2 (t)) u = 0 and z is constant, while x(·) and y(·) are linearly
* * dependent, according to the first integral:
× f (x)g(y)+ g(y) f (x) (43)
*x *y ẏ = ẋ (45)

If 1 (t)+2 (t) = 0 at a certain time t then this Define

quantity remains equal to zero for any future time. 1
l(t) = x(t1 )− x(t) = (y(t1 )− y(t)) and  =l(t2 )
Consequently, the function t  → 1 (t)+2 (t) cannot 
vanish or change its sign for any t, and consequently
Then, one has
no switch is possible. However, at least one switch
from 0 to 1 is necessary to reach the target, thus a dl=− f (x(t1 )−l)g(y(t1 )−l) dt with l(t1 )=0 (46)
contradiction. 
Recall from Proposition 1 that during the anoxic phase,
t  → x(t) is a monotone decreasing map. Then, one has
5.2.2. Solution with one anoxic phase. Consider now

the set of initial conditions (C N )\1 (C N ). The reach-
t anox (x(t1 )) = t2 −t1 = L(x(t1 ), y(t1 ),l) dl (47)
ability analysis shows that from any point in this set, 0
the use of one anoxic phase is necessary and sufficient
where
to reach the target. In the following, we characterize the
optimal solution of the minimal time control problem 1
L(x, y,l) =
for which only one of the phases is anoxic, i.e. with at (x −l)(y −l)
most two switches ‘aerobic–anoxic–aerobic’. To do so, Remark
we consider control sequences The optimal trajectory crosses the switching surface
1 for t ∈ [t0 , t1 ] (X N ) at time t2 . Then one has y(t2 ) = y N −z(t2 ) =
y N −z(t1 ) and  = (1/)(y(t1 )+z(t1 )− y N ). More-
u(t) = 0 for t ∈ [t1 , t2 ] (44) over, t1 is the final instant of the first aerobic phase,
1 for t ∈ [t2 , t f ] which fulfills the properties y(t1 )+z(t1 ) = y(t0 )+
z(t0 ) (accordingly to the first integral ẏ = ż). Conse-
with t1 t2 t f to be defined. Corresponding trajectories quently, one has  = (1/)(y(0)+z(0)− y N ), which
can be written as the concatenation: X (tt1 , X (t0 ), 1)∪ depends only on the initial condition and the terminal
X (tt2 , X (t1 ), 0)∪ X (tt f , X (t2 ), 1). One can notice point.

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OPTIMAL TIME SWITCHING CONTROL 299

z
5.2.4. The total reaction time. When there exists a last Since the integral z(tN0 ) dz/h(z) does not depend on the
anaerobic phase (i.e. t f > t2 ), we distinguish two cases, switching time, minimizing the total time is equivalent
depending on the plane Px or Pz reached by the trajec- to minimizing the duration of the anoxic phase, defined
tory at the final time. in (47) only.
When X (·, X (t2 ), 1) reaches Px (i.e. x(t f ) = x N ), the
total reaction time t reac is given by: 5.2.5. Computation of the optimal solution. To opti-
t1 t2 tf mize the total cycle time, one needs to compute t1∗1 and
t reac = dt + dt + dt t1∗2 only, which minimize, respectively, t1  → t anox (t1 )
t0 t1 t2 and t1  → tmanox (t1 ), and then to compare their corre-
x(t1 ) dx t2 x(t f ) dx sponding total reaction time t reac (t1 ). t1∗1 and t1∗2 can
= + dt + (48) be computed numerically, criteria (47) and (50) being
x(t0 ) f (x) t1 x(t2 ) f (x)
of the form
x(t2 )
Adding and subtracting the term x(t1 ) dx/ f (x), we 
get: J ((t1 )) = L((t1 ),l) dl
0
x(t f ) dx t2 x(t2 ) dx
t reac = + dt − where (·) is the solution of an autonomous system
x(t0 ) f (x) t1 x(t1 ) f (x) ˙ = F(), (t0 ) = 0 . Assuming   → L(,l) to be
xN t2 x(t2 ) differentiable, one has
dx dx
= + dt − (49)  
x(t0 ) f (x) t1 x(t1 ) f (x) dJ ((t1 ))  *L
= ((t1 ),l) dl F((t1 ))
0 *
x dt1
Notice in this last expression that integral x(tN0 ) dx/ f (x)
does not depend on the switching time. To minimize the
t x(t ) which can be determined numerically along the
total time, the expression t12 dt − x(t12) dx/ f (x) only
trajectory. The first-order optimality condition gives
needs to be minimized. Define
dJ ((t1 ))/dt1 = 0. Similarly, d2 J ((t1 ))/dt12 can
1 be computed to write the second-order optimality
L m (x, y,l) = L(x, y,l)−
f (x −l) condition.
and this amounts to minimize the quantity

6. NUMERICAL RESULTS AND DISCUSSION
tmanox (x(t1 ), y(t1 ), ) = L m (x(t1 ), y(t1 ),l) dl (50)
0
In this section we illustrate the improvement of the
When X (·, X (t2 ), 1) reaches Pz (i.e. z(t f ) = z N ), the control law proposed in this paper by comparing the
total reaction time t reac is given by: optimal solution with the standard solution used in
t1 t2 tf the major industrial applications. The standard solution
t reac = dt + dt + dt consists in starting systematically with u = 0 and then
t0 t1 t2
switching to u = 1. This means that t1 = 0. On the oppo-
z(t1 ) dz t2 z(t f ) dz site, our control scheme starts with u = 1, then possibly
= + dt + (51) switches once to 0 and finally ended with u = 1. The
z(t0 ) h(z) t1 z(t2 ) h(z)
parameters of the model we have used for the simula-
Since z(·) is constant when u = 0, then one has z(t1 ) = tions have been identified on a real batch process.
z(t2 ) and the former expression becomes Figure 7 shows the evaluation of t anox (t1 ) and
zN t2 tm (t1 ) with respect to the first switching time t1 , for
anox
dz
t reac = + dt (52) a given initial condition X 0 . The total reaction time has
z(t0 ) h(z) t1 been evaluated for both scenario (one anoxic phase

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
300 D. MAZOUNI ET AL.

Table I. Optimization results.

z0 % total gain % anoxic gain
10 5.2 7.82
20 13.6 16.01
30 24.6 23.91
40 39.0 31.67
50 53.7 39.68

and
anoxic −t anoxic
tstandar optimal
anoxic gain = anoxic
∗100
tstandar

Figure 7. t anox and tmanox evolutions w.r.t. t1 .

7. CONCLUSION

In this paper, we have tackled a problem of the

synthesis of a minimal time feedback law for a class
of bioprocesses controlled by switches. The studied
model possesses a non-linear dynamics with five state
variables. Our approach consists in studying first the
reachability set and characterizing control strategies
leading to the target. First, integrals of the dynamics
have been exhibited in order to reduce the dynamics
and face a truly controllable system. With the help of
the maximum principle, we have characterized a set of
initial conditions for which trajectories with no switch
are extremal solutions. For other initial conditions,
we have proposed to optimize a simple sequence of
Figure 8. Standard and optimal trajectories. switches that reach the target, leading to a suboptimal
feedback law. The overall scheme has been compared
or anoxic–aerobic–anoxic switches), and the optimal numerically with the control strategies currently used
solution is compared with the standard one in Figure 8. in the industrial processes.
For the following parameters

x0 y0 b1 b2 Kx K y Kz   1 max 2 max x N yN zN
700 30 4000 500 145 130 30 0.35 0.058 0.003 0.024 35 25 10

gains in time are summarized in Table I, for different

A relevant extension of this work would be to
z 0 , where
consider the additional control of the dissolved oxygen,
reaction −t reaction
tstandar as an input taken continuous values. The objective to
optimal
total gain = reaction
∗100 be minimized would be the total reaction time and the
tstandar energy consumption.

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca
 10991514, 2010, 4, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/oca.905 by Inrae - Dipso-Paris, Wiley Online Library on [08/04/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
OPTIMAL TIME SWITCHING CONTROL 301

ACKNOWLEDGEMENTS 7. Schattler H. The local structure of time optimal

trajectories in dimension three under generic conditions.
This paper includes results from the EOLI project that SIAM Journal on Control and Optimization 1988; 26(4):
was supported by the INCO programme of the European 897–918.
Community (Contract number ICA4-CT-2002-10012). The 8. Lee E, Markus L. Foundation of Optimal Control Theory.
model parameters used for the simulation were identified The SIAM Series in Applied Mathematics. Wiley: London,
within the EOLI project. We thank very warmly C. Lobry 1967.
for his unvaluable help. 9. Sussmann HJ. The structure of time optimal trajectories for
single input systems in the plane: the cinf nonsingular case.
SIAM Journal on Control and Optimization 1987; 25(2):
REFERENCES 433–465.
10. Bonnard B, Pelletier M. Time minimal synthesis for
1. Impe JV, Bastin G. Optimal adaptive control of fed-batch planar systems in the neighborhood of a terminal manifold
fermentation processes. Control Engineering Practice 1995; of codimension one. Journal of Mathematical Systems,
3(7):939–954. Estimation and Control 1995; 5(3):1–22.
2. Smets I, Claes E, November E, Eva J, Bastin G, Impe 11. Henze M, Grady J, Gujer Wea. Activated Sludge Model
JV. Optimal adaptive control of (bio)chemical reactors: No. 1 (ASM1), IAWPRC Scientific and Technical Report No. 1.
past, present and future. Journal of Process Control 2004; IWAPRC Publishing: London, 1987.
14(7):795–805. 12. Dochain D, Vanrolleghem PA. Dynamical Modelling
3. Moreno J. Optimal time control of bioreactors for the and Estimation in Wastewater Treatment Processes. IWA
wastewater treatment. Optimal Control Application and Publishing: London, 2001.
Methods 1999; 20:145–164. 13. Copp J, Dold P. Comparing sludge production under aerobic
4. Xu X. Optimal control of switched systems based on and anoxic conditions. Water Science and Technology 1998;
parameterization of the switching instants. IEEE Transaction 38(1):285–294.
on Automatic Control 2004; 49(1):2–16. 14. Szederkenyi G, Kovcs M, Hangos KM. Reachability of
5. Bengea S, DeCarlo R. Optimal control of switching systems. nonlinear fed-batch fermentation processes. International
Automatica 2005; 41(1):11–27. Journal of Robust and Nonlinear Control 2002; 12(12):
6. Sussmann HJ. Regular synthesis for time optimal control of 1109–1124.
single input real analytic systems in the plane. SIAM Journal 15. Bryson AE, Ho YC. Applied Optimal Control. Wiley: London,
of Control and Optimization 1987; 25(5):1145–1162. 1975.

Copyright q 2009 John Wiley & Sons, Ltd. Optim. Control Appl. Meth. 2010; 31:289–301
DOI: 10.1002/oca