Chapter 5

Model Predictive Control

EDWARD S. MEADOWS
Departamento de Eletronica
Pontificia Universidade Catolica do Rio Grande do Sul
Porto Alegre, RS BRAZIL

JAMES B. RAWLINGS
Department of Chemical Engineering
The University of Wisconsin
Madison, WI USA

5.1 Introduction
Model predictive control (MPC) is an optimal-control based method to select
control inputs by minimizing an objective function. The objective function
is defined in terms of both present and predicted system variables and is
evaluated using an explicit model to predict future process outputs. This
chapter discusses the current practice in MPC, computational and design
issues, performance and stability and future research directions. Model pre-
dictive control incorporates ideas from systems theory, system identification
and optimization. The application of these ideas is illustrated using simula-
tion results from two chemical reactor models.
Because MPC is implemented with digital computers, much of this chap-
ter deals with discrete-time formulations. In discrete time, a nonlinear, de-
terministic process model can be represented by
x
k+l = .f ( x k,'u k) (5.1)

233
234 CHAPTER 5. MODEL PREDICTIVE CONTROL

Yk = 9(xk) (5.2)

The states Xk, controls Uk and outputs Yk are vectors.

The MPC control problem is as follows: With knowledge of the current
output y k , we seek a control that minimizes the objective function

N-1
J
= c( y k+N^k) + 1, L(yk +jl k,uk +j l k, Uk k
+jl ) (5.3)
j=0

Of the N-move control sequence that minimizes the above objective, only
the first is implemented. When another measurement becomes available,
the parameters of the problem are updated and a new optimization prob-
lem is formulated whose solution provides the next control. This repeated
optimization using an objective function that is modified through process
feedback is one of the principal defining features of MPC.
The objective function that appears in Equation 5.3 illustrates some of
the commonly used terminology in the literature of MPC. States, outputs
and controls are usually doubly indexed as x k+j I k, Yk+jlk and u k+jlk to indi-
cate values at time k +j given information up to and including time k. Since
MPC requires prediction, the double subscript also carries the connotation
of prediction when the first index is larger than the second. Variables that
are singly indexed may be used to represent controls, states or outputs that
have already been implemented, computed or measured, up to the current
time k. An optimal sequence of controls is often indicated using an asterisk
(uk+jIk). The corresponding state and output values then become xk+j^k
and yk +jlk , respectively.
Another feature of MPC illustrated by Equation 5.3 is the presence of
Au
Auk+j^k in the objective, in which k+iJk = u k+jIk — uk+j—llk• In chem-
ical engineering applications, the speed of control action may be limited.
For example, a large valve may require several seconds to change position.
Rather than model the dynamics of control actuators, it is common practice
to include limitations on actuator speed through penalties or constraints on
Ank+jlk. Aside from the issue of actuator limitations, inclusion of Auk+jlk
offers performance advantages in that control action can be limited through
restrictions or penalties on Duk+jjk without introducing permanent offset in
the output. This issue is discussed further in Section 5.7.
The objective function stage-cost L in Equation 5.3 may be chosen to
meet a variety of process objectives including maximization of profits, mini-
mization of operating costs or energy usage, or may be a measure of deviation
5.2. SOME EXAMPLES 235

of the process output from a reference trajectory. In subsequent discussion,

we will emphasize the case of regulation to a setpoint using a quadratic error
criterion. In this case, the objective function of Equation 5.3 may be written
N-1
'J = I IYk+N^k — Yref11 + ( H yk +jJ k — YreflIQ
j=0

+ I I n k+j k — u
reflln + I1Auk+jlklls) (5.4)

In Equation 5.4, Q, R and S indicate weighting matrices in vector norms.

When penalties or constraints on Auk+jlk are used, we admit the special
notation uk_ ij k to indicate the most recently implemented control value,
which is not a decision variable.
If the process model is linear, MPC with the objective of Equation 5.4
reduces to the linear-quadratic optimal control problem. LQ optimal control
theory can be viewed as a (very important) special case of MPC. Many of the
key theoretical results from LQ optimal control have analogs or extensions to
the nonlinear case. For example, it is possible to guarantee nominal stability
of an LQ optimal control in the closed loop based on properties of solutions
of the Riccati difference equation [5]. In nonlinear MPC, this concept may be
extended to provide stability results for nonlinear MPC based on iterations
using the theory of dynamic programming [48].
Another key feature of MPC is the ability to handle constraints directly,
including constraints on the control inputs, outputs or internal states. The
constraint handling property has been one of the most significant contribu-
tors to its success in industrial applications. Some typical constraint formu-
lations are discussed in Section 5.4.

5.2 Some Examples

Before proceeding with more mathematical developments, we wish to in-
troduce two example problems that will be used to illustrate some of the
principles of nonlinear MPC. Our examples are both taken from the chemi-
cal engineering literature and represent chemical reactors commonly used in
industry.

5.2.1 Isothermal CSTR

Here we consider control of a continuous stirred-tank reactor (CSTR) using
dilution rate as the manipulated variable. The volume of the reactor is
236 CHAPTER 5. MODEL PREDICTIVE CONTROL

constant. The dynamics and control of this reactor have been studied by
several researchers, including Van de Vusse [75], Kantor [31], Kravaris and
Daoutidis [38], and Sistu and Bequette [69]. The reactor model is given by

xl = -k1x1 - k3xi + (XF - xl)u (5.5)

X2 = k 1 x 1 — k 2 x 2 — X 2 U(5.6)

which models the reactions

AFB B3C 2A D

The state variables x l and x 2 represent the concentrations of species A and

B, respectively, XF represents the concentration of A in the feed and u rep-
resents the dilution (feed) rate. The parameters {k 1 , k2 , k 3 , x F } have nomi-
nal values {50 hr', 100 hr -1 ,10 liter/(mol hr), 10 mol/liter}. The primary
control objective is to maintain the output variable y = x 2 at a setpoint of
1.0.
Assuming a perfect model, we can solve Equations 5.5 and 5.6 at steady
state to find values of fx 1 , x 2 , u} that give the desired output. Because of
the quadratic term in Equation 5.5, there are two solutions: {2.5, 1.0, 25}
and {6.6667, 1.0, 233.33}. According to Sistu and Bequette [69], lower feed
rates are preferable to achieve higher conversion; therefore as a secondary
control objective, we seek to maintain a low dilution rate.
Assuming that the steady-state model solution provides setpoints for
{x l , x 2 , u}, we can form an objective function
t+T
=
ry l [x ' (r)— 2.5] 2 +-12 [x 2 (r) —
1.0 2 2
] +73 [u (T) — 25] dr (5.7)
A
in which { ly l, 72,73 } represent weight factors that may be used for scaling
or weighting the relative importance of each contribution.
Equations 5.5, 5.6 and the objective function of Equation 5.7 are defined
for continuous time systems. In order to solve with a computer, the differ-
ential equations must be discretized and the controller must be expressed
in a finite parameterization. We discretized the dynamic equations and the
objective function using orthogonal collocation on finite elements. Following
the example of Sistu and Bequette [69], each finite element was 0.002 hours
(7.2 seconds) long. To solve the model equations, we used three interior col-
location points and a point on the element boundaries. (More details on the
5.2. SOME EXAMPLES 237

orthogonal collocation technique are provided in Section 5.4.) In the non-

linear program that solved the MPC problem, we implemented constraints
on control and state variables to ensure that they remained non-negative.
Using y l = -y2 = 1000 and 'y3 = 1, MPC results for the nominal system
are shown in Figure 5.1 for horizon length 5 (five finite elements of 7.2
seconds each). The initial condition was {2.9997,1.11697}, which was the
same initial condition used by Sistu and Bequette [69]. The points in the
diagrams show the sampling times. Between sampling times the curves were
computed by integrating the nominal model. Both states and the control
converge quickly to the desired target values. Similar results were obtained
for simulations with horizon lengths from 1 to 7.
In most processes, all of the internal model states are not measured.
If we use only the output variable x 2 in the objective (7 1 = rya = 0), we
find that the MPC controller is able to bring the output to the setpoint,
but the controller drives the system to the steady-state values at the higher
(undesirable) dilution rate. These results are shown in Figure 5.2. Similar
results were also obtained for horizon lengths from 1 to 7.
Since using only the output variable failed to satisfy the secondary con-
trol objective of maintaining low dilution rate, a logical attempt to satisfy
the secondary objective might be to place an upper limit on the control in-
put. Although not provided in the problem description supplied by Sistu
and Bequette [69], we can reasonably assume that the nominal steady-state
feed rate (25) lies near the middle of the operating range and we therefore
apply a maximum feed rate of 50. The results are provided in Figure 5.3,
which shows that the steady-state value of the output x 2 has offset. The
same effect is seen with controllers with horizon lengths 1 through 7. The
length of the horizon has an effect on the steady-state values observed in
the simulations. The steady-state value of the output versus horizon length
is shown in Figure 5.4. We can see that longer horizons produce less offset.
With "well-behaved" systems, steady-state offset will disappear in the limit
as the horizon increases to infinity.
Since none of the closed-loop control values in Figure 5.3 are at the
upper limit, it is not obvious why the input constraint causes steady-state
offset. To see why this occurs, we must examine the open-loop control pro-
file computed at each time step. For each of the horizon lengths tested in
simulation, we found that the input constraint is active for some portion of
the prediction horizon. For N = 1 to 4, the input constraint is active for the
initial open-loop control; therefore, the steady-state value of x 2 corresponds
to u = 50. For N = 5, the initial open-loop control is not constrained, but
238 CHAPTER 5. MODEL PREDICTIVE CONTROL

2.8

2.4
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04

Time

N
1.16

1.12

1.08
a 1.04
u
1
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04

Time

30

20

0
4D 10
0
U
0L
0 U.UU5 0.01 0.015 0.02 0.025 0.03 0.035 0.04

Time

Figure 5.1: Isothermal CSTR: Nominal system with horizon length 5.

5.2. SOME EXAMPLES 239

11

`^ 5

cr 3

0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04

Time

1.12

1.08

1.04

0.96
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04

Time

300

200

100

01
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04

Time

Figure 5.2: Isothermal CSTR: Nominal system with horizon length 5 using
only the output.
240 CHAPTER 5. MODEL PREDICTIVE CONTROL

2.8

2.6
UO

0 0.01 0.02 0.03 0.04 0.05 0.06

Time

1.15

^ 1.1

1.05

0 0.01 0.02 0.03 0.04 0.05 0.06

Time

30
41

28
0
26

241 I I I I I I

0 0.01 0.02 0.03 0.04 0.05 0.06

Time

Figure 5.3: Isothermal CSTR: Nominal system with horizon length 5 using
only the output with input constraints.
5.2. SOME EXAMPLES 241

D 0---El--U
1.2

1.1
Ct
a)

1 ----

1 2 3 4 5 6 7 8 9

Horizon Length
Figure 5.4: Isothermal CSTR: Steady-state output versus horizon length.

subsequent values in the prediction horizon are constrained. With the initial
state used in the examples, the initial open-loop control sequence for N = 5
is {24.7, 50, 50, 50, 50}. For horizon lengths greater than 5, the open-loop
control profiles start at the lower constraint U min = 0 and reach the up-
per constraint subsequently within the control horizon. For example, with
N = 6, the open-loop control profile is {0, 46.4, 50, 50, 50, 50}. To minimize
the objective function while satisfying the input constraints, the optimal
open-loop control sequence begins at the lower bound and increases during
the prediction horizon until it reaches the upper bound.
It should be emphasized that the observed offset is not due to model error
or unmeasured disturbances. The controller and the simulation both use the
nominal model and no noise or disturbances were included. The observed
offset is the result of using a controller formulation that is asymptotically
stable only in the limit as N —* oo. The control values and corresponding
states reflect the best finite horizon solution to each constrained optimization
problem, but do not necessarily result in zero offset. In the limit as the
horizon length approaches infinity, we would expect the closed-loop output
to approach its setpoint with zero offset, since steady-state offset would result
in an unbounded objective function. This effect can be seen in Figure 5.4.
Since imposing control bounds failed to provide a satisfactory controller,
we look for other strategies to improve the performance of the nominal sys-
tem in closed loop. Figure 5.2 shows that, without the control penalty, the
control initially jumps to almost 148 on its way to the final undesirable
steady-state value at the undesirable high dilution rate. To prevent exces-
sive jumps in the control, we can include a penalty on Au k + j ^ k . Figure 5.5
shows the effect of including a 1 percent penalty on Duk+^^k.
 CHAPTER 5. MODEL PREDICTIVE CONTROL
242

3
N=3 -
F3
N=6 ---
N=8 ----
2.8

PKy

2.4
0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
0
Time

1.02
N
\ N = 3 -

1.01 ,
\ N = 8

O I I
0.99 '
1Q
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 n

Time

25.4
N-3
41
N= 6 ---
Cdr N=8 ----
H-i
25
41

24.6
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18

Time

Figure 5.5: Isothermal CSTR: System response with Du penalty.

5.2. SOME EXAMPLES 243

For this case, qualitative differences appear in the responses for different
horizon lengths. Figure 5.5 shows the results for horizon lengths of 3, 6
and 8. Notice that the speed of response has been greatly slowed due to
the penalty on Au, compared to Figures 5.1 or 5.2. The time scale on the
figures has increased by a factor of three from the previous figure. (The use
of symbols to identify the sampling instants obscures the results and has
been dropped.) We find that for horizons less than 6, the desired steady-
state operating point y = 1.0 is not asymptotically stable. The case N = 3,
illustrated in Figure 5.5, shows a continuous increase in the control following
the initial transient. The simulation ended before a steady-state value was
reached. All horizons shorter that 6 showed similar unstable behavior.

In the other cases illustrated, N = 6 and N = 8, the controller brings

the system to the desired steady state. With N = 6, there is no overshoot,
while N = 8 shows overshoot, both in the control and in the output. The
vertical axis of the graph of the output variable, x 2 , has been expanded to
more clearly show the different behaviors near the setpoint.

When the output variable is x 2 only and the flow rate is low, the lin-
earized reactor model contains a right half-plane zero and therefore exhibits
nonminimum phase characteristics. For linear systems, it is known that
model predictive control can induce unstable closed-loop behavior for non-
minimum phase systems, even if the open-loop system is stable [23]. Some
of the linear systems results can be used as a guide to explain the observed
behavior of the nominal system under MPC that was examined in this sec-
tion.

In Figure 5.1, two output were assumed, i.e., y = [x l , x 2 ] T . In this

case, the system cannot be described as non-minimum phase. Figure 5.2
shows the results of using y = x 2 , illustrating the non-minimum phase effect.
Although the desired setpoint was reached, the control continued to increase
until it reached the second steady state at the higher setpoint. This is
unstable behavior. A similar non-minimum phase linear system would not
have reached an upper steady state, but would have demonstrated control
action that continued to increase without limit.

To prevent MPC from returning an unstable controller for non-minimum

phase linear systems, it is necessary to penalize control action or decrease
the control horizon relative to the prediction horizon [23, 25]. We chose the
first course and the results are provided in Figure 5.5.
244 CHAPTER 5. MODEL PREDICTIVE CONTROL

5.2.2 Fluidized Bed Reactor

In this section, we demonstrate model predictive control through simulation

of the fluidized-bed reactor described by Elnashaie, et al. [22]. The model
describes the catalytic exothermic consecutive reactions

A B C

in a freely bubbling fluidized-bed reactor. After a series of simplifying as-

sumptions, the dimensionless reactor model is given by the following:
1 dXA
Q(
XA , F — XA) — rl (5.8)
Leg dt
1 dx B _
= ( X B,F — xB) + rl — r2 (5.9)
L dt
dT _
r
— Q (TF — T + u) + i l +2r2 (5.10)
dt

ri = ai exp xA (5.11)
(—T)

r2 = a 2 exp I — T I XB (5.12)

In the model, Le A and Le B represent Lewis numbers, 3 the residence time,

ctr l , a 2 pre-exponential factors for rate constants, 'i i , rye activation energies,
and ,6 1 , / 32 heats of reaction. Concentrations of A and B and the temperature
are given by X A , XB and T, respectively; XA,F, XB,F and TF represent their
values in the inlet stream. For details concerning derivation of the reactor
model and the numerical values of the constants, the reader is referred to
the original publications [21, 22].
In our version of the model, the control enters the equations in a slightly
different way than in the original work of Elnashaie et al. [22]. Their pa-
per was concerned with chaotic behavior of this reactor using proportional
control. Their original model can be recovered by setting 'u = K ( Tset — T).
The measured output of the system is the temperature of the reactor
effluent T. The control objective is to maintain the effluent temperature
at the setpoint Tset that corresponds to the maximum yield of species B.
Using the nominal model, we computed this optimal (dimensionless) effluent
temperature of 0.9213, which will be used as the setpoint for our simulation
studies. At this temperature, the nominal steady state for the reactor is XA =
5.3. MODELS FOR MPC 245

0.2355, XB = 0.6448, which is an unstable steady state. The eigenvalues of

the linearized system matrix at this point are { - 0.2706, —0.0507, 0.6323}.
No sampling interval was specified by Elnashaie et al., since their study
was not concerned with digital control. We selected a sampling interval of
1.09 dimensionless time units, slightly shorter than the doubling time of the
unstable mode. Using the objective function
N-1 (k+1) OT
J = 100 [T (T) — 0.9213] 2 dT + ,C3 ( nk)2
Ik OT

with prediction and control horizon equal to 4, simulation results for the
nominal system are provided in Figure 5.6. As in the previous example,
we employed the orthogonal collocation technique in a simultaneous model
solution/optimization approach. Both cases 13 = 0 and 13 = 0.1 are shown.
Without any penalty on control action, Figure 5.6 shows a control range of
approximately —2 to 3.5 with a maximum of approximately 5.5. Based
on the values of the numerical results reported by Elnashaie et al. [22], we
can infer that this control is probably more aggressive than the system can
support. Using a 0.1 percent penalty on Du (Q = 0.1) moderates the control
action and yields a slower response by the internal states and the output
variable.
Another common approach that dampens excessive control action is to
select a control horizon M that is shorter than the prediction horizon N.
After the M-th interval, the control is held constant through the end of the
prediction horizon. Figure 5.7 shows the results of this approach for N = 4
without control penalty (3 = 0), using control horizons M = 1 and M = 2.
Using M = 1 greatly dampens the control input and slows the response of
the system. Note the increase in scale in Figure 5.7 necessary to illustrate
the transient of XA and X B . Increasing the prediction horizon from M = 1
to M = 2 increases both the aggressiveness of the control action and the
speed of the system response. With M = 2 we observed system behavior
very similar to that shown in Figure 5.6 using the 0.1 percent penalty on
Ou.

5.3 Models for MPC

The heart of model predictive control is the process model itself. Models
can be classified by various features. Since MPC requires the solution of a
model to predict future process outputs, the form of the model selected has
246 CHAPTER 5. MODEL PREDICTIVE CONTROL

0 .5 '^^^` Without Du penalty

With Du penalty ---

^' 0.35 ^'^--_

0.2
0 1 2 3 4 5 6 7 8

0 .6 ___ ---------------

^--^ Without Au penalty

0.5 _ With Du penalty ---

0 1 2 3 4 5 6 7 8

`' 1

0.8 ' Without Au penalty

With Du penalty
E^ 0.6
0 1 2 3 4 5 6 7 8

_ Without Du penalty
With Ou penalty - - -

-3 E I I I I I I I -1
0 1 2 3 4 5 6 7 8

Dimensionless Time

Figure 5.6: Fluidized-bed reactor: Effect of penalty on Au.

5.3. MODELS FOR MPG 247

0.7
M=1 -
M=2 -

0.3

0 5 10 15 20 25 30

---- ------ -------------

M=1
0.5 % M = 2 ---

0.4
0 5 10 15 20 25 30

I
C, 1
-----

/ M = 1
// M = 2 - --

0.6
0 1 2 3 4 5 6 7 8

3 ---- -i M =1

2 M= 2 ---

1
0 -- ------ IL----- ------ - - - - _

-1
0 1 2 3 4 5 6 7 8
Dimensionless Time

Figure 5.7: Fluidized-bed reactor: Effect of control horizon.

248 CHAPTER 5. MODEL PREDICTIVE CONTROL

a large impact on our ability to implement MPC. Some specific categories

of models are discussed below:

• Linear or nonlinear : The response of linear dynamic systems obeys

the principle of superposition, i.e., the response of the system to a
linear combination of inputs is a linear combination of their responses
to each input separately. Many (perhaps most) systems of engineering
interest approximate this behavior for small inputs, which accounts for
the universal study and application of linear control theory. In almost
any control application, linear design techniques are usually the first
to be attempted and are completely satisfactory for many engineering
applications, especially those involving regulation about a steady-state
operating point. Linear models are used extensively in the industrial
practice of MPC.
Nonlinear models have no specific characteristics except that they
don't include the linear case. This makes it difficult to generalize, since
nonlinear models can be devised to have almost any characteristic or
demonstrate any strange behavior. For many engineering applications,
it is usually appropriate to assume that the models are continuous with
respect to model parameters and that differential or difference equation
models satisfy growth conditions that permit solutions to be computed.

• Continuous-time or discrete-time : Most of the physical laws that are

used by engineers to develop models are presented as differential equa-
tions with time as the independent variable. A typical representation
is of the form ± = f (x, u). Before the widespread availability of dig-
ital computers, differential equation models were the central tool of
researchers and control engineers for the study of dynamic systems. If
simulations were needed, analog computers could be used. Prediction
was rarely used as an element of control engineering, since complex
models could not be solved using paper-and-pencil methods.
With the advent of digital computers, the study of difference equa-
tions, which had been relegated to a minor role previously, assumed
a new significance. We can write a nonlinear difference equation as
x
k+1 = f ( x k, U k ). Since MPC is universally implemented via digital
computer, difference equations are the model of choice. Differential
equation models (as in Section 5.2) must be discretized for computer
solution. In this chapter, our examples are based on continuous-time
models that are discretized using orthogonal collocation.
5.3. MODELS FOR MPC 249

• Distributed parameter or lumped parameter : A distributed parame-

ter model involves partial differential equations, instead of ordinary
differential equations. An example of a distributed parameter model
would be that for a plug flow reactor, in which the changes in chemi-
cal concentrations in the reactor are subject to both spatial variation
and variations in time. Although we do not specifically consider dis-
tributed parameter models in this chapter, the basic MPC concept is
completely applicable to systems described by distributed parameter
models, with a corresponding increase in computations necessary to
solve the model.

• Deterministic or stochastic : All physical processes are subject to un-

predictable disturbances. Disturbances can affect MPC control design
and operation in at least two distinct ways:

— In the process of model identification, a model is often selected

from some class based on experimental results. The selection
process implicitly or explicitly uses assumptions about the distur-
bances to select and evaluate the best model of the class. These
assumptions have a direct impact on the model selected.
— After the model identification phase is complete, the assumptions
about the disturbances are sometimes discarded, and control de-
sign may be based on this "nominal" model. If the model allows us
to predict the statistics of process variables based on assumptions
about random effects on the models, we say that it is a stochastic
model. We can represent discrete-time stochastic models using w
and v to represent state and output disturbances as follows:
X
k+1 = f ( x k, u k, w k) (5.13)

Yk = 9( x k, v k) (5.14)

In general, better control can be achieved by using explicit stochas-

tic models [16, 17, 51] in the prediction phase of MPC, but with
a much higher computational burden. The traditional approach
to stochastic MPC is through dynamic programming [2, 3], which
suffers from combinatorial growth in the number of optimization
variables with increasing horizon length (Bellman's "curse of di-
mensionality"). Better ways to incorporate stochastic models in
MPC are active areas of research.
250 CHAPTER 5. MODEL PREDICTIVE CONTROL

• Input-output or state-space : As indicated by the name, input-output

models provide a relation between the process input and the output,
without reference to variables internal to the process. An input-output
model of a distillation column, for example, might relate the tempera-
ture of a side-product stream to reboiler duty, without consideration of
individual tray temperatures or compositions; whereas, a state-space
model might include equations relating all internal compositions and
flow rates to the reboiler duty, with the side product temperature be-
ing provided as a function of the composition. Because most nonlinear
state-space models are based on heat, mass and momentum balances,
each state has a physical interpretation. States may also be gener-
ated as mathematically convenient intermediate variables of an input-
output process model. In the examples discussed in this chapter, the
states have definite physical significance; however, the principles of
MPC are applicable in either case.
As a special class of input-output model, we should mention that artifi-
cial neural network models are becoming important in many engineer-
ing applications, including model predictive control, with a number of
researchers providing important new results [39, 40, 45, 59, 61, 20].
This chapter does not specifically discuss neural networks in MPC, al-
though the above cited works make clear that artificial neural networks
may be successfully used in MPC. Chapter 7 by McAvoy in this book
contains further discussion of the use of artificial neural networks in
control.

• Frequency domain or time domain : Frequency domain models in con-

tinuous time are based on the Laplace transform of continuous linear
systems and are not used for model predictive control, except for the
linear unconstrained problem. For frequency domain descriptions in
discrete time, it is not possible to be so definitive, since the shift op-
eration (q or q — ') is often interchanged with the z transform oper-
ator. Depending on the context, these models may be described as
input-output or discrete-time frequency domain models. Since few of
the concepts of frequency domain analysis are applicable for nonlinear
systems, frequency domain concepts play little or no role in nonlinear
MPC.

• First-principles or "black box" : Models that are derived from heat,

mass and momentum balances are frequently called "first-principles"
5.3. MODELS FOR MPC 251

or "fundamental" models, in contrast to other modeling schemes that

fit a data set to an arbitrary form. Both approaches have been used
successfully in MPC applications. An example of the "black box"
approach is that used in model identification using step or impulse re-
sponse models, in which model coefficients are fitted to process data
using statistical methods without regard to underlying physical prin-
ciples. The first principles approach is reflected in the models used in
Section 5.2. Basic principles of chemistry were applied to arrive at a
model of the chemical reactions involved and mass balances were used
to create an overall reactor model.
The clearest tradeoff between the first-principles and black-box mod-
els lies in our confidence in predictions made by the model. A first-
principles model presumably can be used to predict over a wide range
of conditions, even without prior operating experience, provided that
the basic assumptions of the model remain valid. On the other hand, a
model based on (for example) artificial neural networks has almost no
predictive value outside the range of operating conditions where data
has been collected.
The best features of both approaches can perhaps be incorporated
into a hybrid approach in which fundamental modeling is used for
those portions of a process where the physical phenomena are well
understood, and an input-output model used for the remainder of the
process model.
Each of these model characteristics has an impact on the implementation of
MPC. Since MPC is an on-line control method, the speed of the computer
algorithm used is essential. Linear models are very well suited to MPC
because they may be solved quickly and the optimization problem may be
posed as linear or quadratic programming problems, for which robust and
reliable software is available. Using continuous models in MPC also impacts
the speed of solution because it usually requires significantly more variables
to represent a discretized model. Distributed parameter models require more
time to solve because of the additional difficulty in solving partial differential
equation models.
From the class of input-output models, step or impulse response mod-
els have been favored for industrial application. In view of recent results
concerning the use of linear state-space models [55], there appears to be
little advantage in continuing with this practice. An equivalent state-space
model is no less easy to identify, contains fewer unknown coefficients, can
252 CHAPTER 5. MODEL PREDICTIVE CONTROL

be evaluated more quickly by computer and can represent a wider range of

processes.
For the nonlinear case, the decision to use first-principles or input-output
models is less clear. For small systems with well-understood physical phe-
nomena, fundamental modeling is preferable, because of the ability of the
model to predict beyond the range of existing operating data. On the other
hand, model identification is easier for larger systems using black-box mod-
els.
In conventional industrial MPC, installation of an MPC controller usu-
ally begins with identification of a linear input-output model in discrete
time. Little or no effort is made to determine a first-principles model based
on mass, energy or momentum balances. The objective of the modeling pro-
cess is to determine a model that can be numerically evaluated quickly and
that adequately describes the process dynamics in a neighborhood of some
desired steady-state operating point. MPC based on these models has been
extraordinarily successful in industry, but it is only applicable for stable
processes that operate continuously near the nominal steady-state values.
Although recent research results have extended the use of linear models for
unstable processes [55], the range of operating conditions remains limited to
those near the nominal steady state.
MPC using nonlinear models is especially suited for batch or semi-batch
operations in which process conditions can vary significantly over the course
of a batch, or for continuous processes that are expected to experience widely
varying operating conditions.
In the past, the expense of computation often prevented even the consid-
eration of the use of nonlinear models. As the power of computers available
for on-line computations continues to increase, it has become feasible to
consider using more complex models of various kinds. Today, the princi-
ple limitations in applying nonlinear models for on-line process control are
model identification and robustness of control software and algorithms. A
discussion of current research issues is contained in Section 5.9.

5.4 Computational Issues

The MPC control depends on finding solutions to a nonlinear programming

problem at each sample time. This section discusses solutions techniques
that are readily implementable using current optimization software. Since
the complexity of the nonlinear program (NLP) is directly related to the
5.4. COMPUTATIONAL ISSUES 253

model form, we consider several of the most common forms.

To solve the model predictive control problem, it is necessary to both
solve an optimization problem and solve the system model. These two pro-
cedures may be implemented either sequentially or simultaneously. For se-
quential solutions, the model is solved at each iteration of the optimization
routine. The controls are the decision variables, which are supplied to a rou-
tine that computes the model solution. The model solution is then used to
compute the objective function, and the computed value is returned to the
optimization code. Since gradient information is used by most modern opti-
mization routines, it can be obtained through numerical differentiation (not
recommended, see below), involving repeated calls to the routine that solves
the model, or the sensitivity equations can be evaluated along with the model
equations using, for example, the software routine DDASAC [7]. Our expe-
rience using numerical derivatives obtained through finite differences in the
sequential approach has been strongly negative. To obtain gradients using
finite difference typically involves differencing the output of an integration
routine with adaptive step size. Although tightly bounded by the integration
algorithm, the integration error is unpredictable. Differencing such an inex-
act quantity greatly degrades the quality of the finite difference derivatives.
For this reason, using finite difference derivatives in the sequential approach
cannot be recommended. See Gill et al. [27] for a further discussion of this
issue.
In contrast, simultaneous model solution and optimization includes both
the model states and controls as decision variables and the model equations
are appended to the optimization problem as equality constraints. This can
greatly increase the size of the optimization problem, leading to a tradeoff
between the two approaches. For small problems with few states and a
short prediction horizon, the sequential method using sensitivity equations
for derivatives is probably the better approach. For larger problems, we
have found that the simultaneous approach generally is more robust, i.e.,
the optimization software is less likely to fail. For problems with state or
output variable constraints in continuous time, the sequential approach is
more complicated. An approach to incorporating state variable constraints
in continuous time was suggested by Sargent and Sullivan [66] that requires
an additional state variable to be defined for each state or output variable
constraint.
The remainder of this section discusses implementation issues for MPG
using discrete-time and continuous-time models and the simultaneous opti-
mization/model solution approach.
254 CHAPTER 5. MODEL PREDICTIVE CONTROL

5.4.1 Discrete-Time Models

In this section, we consider the solution of MPC problems posed solely in
discrete time. Since the computational issues discussed here are primarily
concerned with solving the open-loop, constrained control problem, we will
temporarily abandon the double time index. The model is given by Equa-
tions 5.1 and 5.2 and the objective function is given by
N-1
J
= I g ( x N) — 2JrefI Q (I g ( x j) — Yref I Q + I
u
j—
u u
refI R + I ° j I s)
j=o
(5.15)
For the unconstrained problem, or problems with only control constraints,
it is possible to substitute the model equations directly into Equation 5.15
and to use only the controls as optimization variables (sequential approach).
As noted previously, we have observed significantly better performance from
optimization codes if each state and control is considered an independent op-
timization variable, with the model equations appended as nonlinear equality
constraints:

u
xi — f( x o, o) = 0

u
x2 — f( x i, i) = 0

X u
N — f ( X N-1, N-1) = 0

Considering states and controls as independent decision variables, the size

of the vector of decision variables is n(N + 1) + Nm. Most optimization
software requires the vector to be passed as a single argument. We have
found it most convenient to partition the entire vector of decision variables
1
r T
as L x 0 xl ... ^ T
N uT T
0 u l ... UN-1 J .

MPC is characterized in part by its ability to incorporate constraints.

In chemical process applications, constraints are most commonly expressed
as simple bounds on the problem variables:

Ymin C Yj C Ymax

C U C
Umin i umax
Au Au
min < Au i < max (5.16)
5.4. COMPUTATIONAL ISSUES 255

The control constraints of Equation 5.16 can be expressed as a set of linear

inequality constraint

UO

[ _c ] I ul < [ c
lb ]
uN-1

in which C, cu b and c ib are given by

I
—I I
C_ —I I

—I I

u
-1 + Au max — u -1 — Dumin
Du
max —Dumin
Cub— Clb =

Du
max —Aumin

The form of the above constraint illustrates an example of another commonly

used constraint, the general linear inequality constraint of the form

Dx k < d

Hu k < h

These constraints are applied to each predicted state or control in the hori-
zon. The complete set of constraints may then be expressed using the Kro-
necker matrix product

xi no
X2 ul
(IN ®D) < 1N ® d ( INCH) <1N(h

XN uN-1
256 CHAPTER 5. MODEL PREDICTIVE CONTROL

in which IN is the N-dimensional identity matrix and 1 N is an N- dimensional

vector whose elements are equal to 1.
Sometimes it provides better closed-loop performance to use a control
horizon M that is less than the prediction horizon N, as demonstrated in
the simulation results for the fluidized-bed reactor. For M < N, we add
additional linear equality constraints U N_1_M = U N_Al = • .. = uN-1.
The linear bounds discussed above are sufficiently general to include
simple bounds on the decision variables. Even so, it is usually preferable
to avoid including them in the general formulation for linear constraints,
for two reasons: Since most optimization software permits simple bounds
to be included in the argument list of the calling routines, it simplifies the
programming task to use the available program interface. More importantly,
some implementations include special considerations for simple bounds [27]
that can reduce the computational effort necessary to solve the nonlinear
program.
Most optimization software contain routines to use finite differences to
compute gradients. We strongly recommend against use of finite difference
gradients except perhaps for the smallest and simplest problems. Gradients
(and Jacobians) should be provided if possible to optimization routines to
assure good results. For the objective function of Equation 5.15, the gradient
may be expressed by

2Gp Q [g(xo) — Yref]

2GTQ [g(x i ) — yref]

2GNQ [g(xN) — Yref]

VJ=
2R (up — 2G ref) + 2S(2n 0 — u-1 — ul)
2R (ul — U ref) + 2S(2u1 — UO — U2)

2R ( U N_1 — u ref) + 2S ( U N-1 — uN -2)

in which G k is the Jacobian matrix of g(x) evaluated at (x k , Uk).

The Jacobian of the equality constraints representing the system equa-
tions is sparse and has the following general form:
5.4. COMPUTATIONAL ISSUES 257

I —F.,o
—F. ,1I —F1
—F ,2
I
—F
.,3 I —Fu,3

The Jacobian of the inequality constraints is also sparse with the follow-
ing structure:

D
D
D
D
H
H
H
H
I
—I I
—I I
—I
—I
I —I
I —I
I —I

The above examples represent an MPG problem using a prediction hori-

zon of 5, with linear constraints on state, control variable and control incre-
ments at each time step.

5.4.2 Continuous-Time Models

For continuous-time systems, we consider here nonlinear models represented
as differential equations of the general form

x = f (x, u, t) (5.17)

A controller is sought which minimizes some performance index

tf
J=O[x(tf),tf]+ L [x(t),u(t),t] dt (5.18)
to

We can find necessary conditions for a minimum using variational techniques,

but in general these do not form a practical basis for implementation. We
seek an approximate solution using piecewise constant controls with intervals
258 CHAPTER 5. MODEL PREDICTIVE CONTROL

that correspond to some appropriate sampling frequency of the process. Op-

timal controls for the system of Section 5.2 were obtained using the method
of orthogonal collocation on finite elements, a technique based on numeri-
cal integration using Gaussian quadrature. We present here a brief tutorial
description of the method. See the following for a more detailed theoretical
discussion: [4, 15, 12, 77, 78].
Consider the solution of minimizing the objective function of Equa-
tion 5.18 over a single time element. In the isothermal CSTR of Section 5.2,
the time interval was 0.002 hours or 7.2 seconds. At the beginning of the
time interval the state (or state estimate) is known and we seek a single
control, constant over the interval, that minimizes the objective function.
The orthogonal collocation technique assumes that the solution to Equa-
tion 5.17 can be approximated by an interpolating polynomial, expressed as
the sum of Lagrange polynomials [14], with nodes located at the roots of
an orthogonal family of polynomials. By this choice of nodes, integration
using the orthogonal collocation technique is exact for polynomials of degree
2n — 1, although only a polynomial of degree n is used for the interpolation.
See Davis and Rabinowitz [15] for a complete explanation of this remarkable
result.
In Section 5.2, all computations were performed using either two or three
internal collocation points and one on the boundaries of each finite element.
Three internal collocation points are illustrated in Figure 5.8. If the element
is scaled so that t l = 0 and t 5 = 1 then the node points are given by
{t l , t 2i t 3 , t 4 , t 5 } = {0,0.1127,0.5,0.8873, 1.0}, which are the roots of the
third order Legendre polynomial, augmented by the element endpoints. If
ti 0 0 or t 5 ^ 1, then a linear change of variable is necessary to transform
the time interval to [0, 1]. When the dynamic model is time-invariant, the
only correction necessary is to scale the right-hand side of Equation 5.17 by
t5 - tl.

With these choices of interpolating polynomials and node points, deriva-

tives and integrals can be represented by linear combinations of the state
values at the node points. Let the jth state at node i be represented by
x. We admit the notation xT to indicate the entire state at node i. With
constant control, an approximate solution to the differential equation can be
obtained by solving the following nonlinear equation:

AoX = F(X, u) (5.19)

5.4. COMPUTATIONAL ISSUES 259

tl t2 t3 t4 t5

Figure 5.8: Collocation element.

in which

x x
11 12 x1,n 4 it

(x u
X2 ,l X2 , 2 ''' X2,n fT 2,* , t2)

x= X3 , l X3,2 x3,7, F (X, u ) _ x u

f T ( 3 *, , t3)

X
x
4,1 4,2 X4,n f T ( x 4,* u
^ t4)

x
5,1 x5,2 fT ( x5 ,*' u
' t5)

The matrix A 0 contains the first derivative collocation weights:

1 0 0 0 0
—5.32 3.87 2.07 —1.29 0.68
A0 = 1.5 —3.23 0 3.23 —1.5
—0.68 1.29 —2.07 —3.87 5.32
1 —1.88 2.67 —14.79 13

(The subscript 0 is included to emphasize that the first line of the collocation
weight matrix is added to account for the initial condition.) The objective
260 CHAPTER 5. MODEL PREDICTIVE CONTROL

function of Equation 5.18 is approximated by

5
J W. , t 5 ) + E c k L (x , u, t k ) ( 5.20)

in which the coefficients c k are the quadrature weights l 0, 8 1 , } cor-

1 , , 8
responding to the third order Legendre polynomial. For other choices of
orthogonal polynomials and node points, the collocation weight matrices
and quadrature weights may be computed using the FORTRAN programs
of Villadsen and Michelsen [77] or using root-finding techniques using the
orthogonal polynomials [15]. In the examples of this chapter, collocation
and quadrature weights were computed using the octave [18] interface to
the Villadsen and Michelson routines.
With these choices, the control is obtained as the solution to the following
nonlinear program

min J
w,x

subject to:

AoX = F(X, u) (5.21)

which represents a discretized form of the MPC problem with horizon length 1.
For horizon lengths greater than 1, we can extend the orthogonal collo-
cation method to a set of finite elements in time, with the solution of the
problem approximated by polynomials defined on each element. This sit-
uation is illustrated in Figure 5.9. Note that we have chosen an indexing
scheme that consecutively numbers nodes across finite element boundaries.
It would perhaps be more natural to adopt an additional index to indicate
to which finite element a state or control belongs; however, this would result
in triple indices for states, controls and outputs, which are not supported in
most programming languages or any known optimization software implemen-
tations. The structure shown in Figure 5.9 corresponds to the collocation
scheme used for the isothermal CSTR example of Section 5.2. The fixed-bed
reactor example used one fewer interior collocation point. The matrix A 0 in
 5.4. COMPUTATIONAL ISSUES 261

Element 1
-h- Element 2 —I
mo
— Element N —I

^ ^ 1 1 I I ^

t l t2 t3 t4 t5 t7 t9 t
t6 t8 4(N-1)+1 t4N+1

Figure 5.9: Collocation on finite elements.

Equation 5.21 become a partitioned matrix that is almost block diagonal:

r- I

'A '
I of

'A'
L _I

A
L _I

r -I
'A'
L _I
r---I

L
A'_I
(5.22)

The overlap is due to the presence of the points on the boundaries of two
adjacent finite elements. The state vector includes contributions from states
262 CHAPTER 5. MODEL PREDICTIVE CONTROL

in every multiple finite element and so becomes

X1,1 X1,2 ''' x1,n

X
2,1 X2,2 ''' X2,n

X
3,1 X3,2

X4,n
X=I

x x
4N-2,1 4N-2,2 x4N-2,n
X x
4N-1,1 4N-1,2 x4N-1,n

The F vector is constructed as in the one-element case, except that the value
of the control must change across element boundaries:

x nit

u
f T l x 2* 0 , t2)

f T ( x 3*, 0 t3)
u

f T ( X 4*, u 0 t4)
u
f T ( x 5*, 0 t5)

f T ( x 6 u1,t6)
F X
( , n 1, n 2, ... , u N-1) _ f T ( x 7*, n 1, t7)
x
f T ( 8*, 1, t8)
u

fT
( xg , u
1, t9)

f T ( x 10*, U 2, t10)

fT (x
(4N-2)* uN
- 1 ^ t4N-2)

f T( x(4N- 1)*)uN- 1^t4N-1) -

5.5. FEEDBACK AND STATE OBSERVERS 263

Taking c = 1 N ®c (the Kronecker product of 1 N and c) the objective function

of Equation 5.18 is approximated by
N-1 4j+5 T
x ,t L x n
( (4N-1)* 4N-1) + E ck ( k *^ j, tk
j=0 k=4j+1
T
U UT
Taking U = [ui , 2 , ... , _1] , we can write the orthogonal collocation
approximation as
AX = F (X, U) (5.23)

Since most optimization codes pass the optimization variables as a single ma-
trix, the matrix equation of Equation 5.23 must be appropriately indexed. In
our examples, we were more interested in coding convenience that in optimal
performance, so we reconstructed the state matrix X as shown above within
the objective function function. Naturally, choices of indexing schemes are at
the discretion of the programmer. For on-line application, we would seek the
fastest evaluation method possible and would probably include such items
as sparse matrix multiplication and optimization of sparse systems.
Since optimization codes require a single vector of decision variables,
the task of re-indexing can become tedious. The Jacobian matrix of the
equality constraints is quite sparse, becoming more sparse with larger prob-
lems, and the overlapped, quasi-block diagonal structure of A is repeated in
the Jacobian. This is illustrated more clearly in Figure 5.10, which shows
the non-zero elements for the fluidized-bed reactor example with prediction
horizon 3, using 5 internal collocation points in each finite element.

5.5 Feedback and State Observers

Like the LQ-optimal controller, the state-space formulation of nonlinear
MPG discussed in this chapter uses state feedback. The current state (or
state estimate) is used as an initial condition at each iteration to predict
the future behavior and select an optimal control sequence. Model predic-
tive control does not prescribe a specific form for the state observer. This
section discusses several possibilities and their effects on closed-loop system
behavior, including steady-state tracking in the presence of modeling errors
and disturbance rejection.
If the full state is measured at each sampling point, an obvious possibility
for feedback would be simply to reset the states to the current observations.
264 CHAPTER 5. MODEL PREDICTIVE CONTROL

ID]

20

30

40

50

0 10 20 30 40 50 60

Figure 5.10: Sparse Matrix Structure of Constraint Jacobian.

5.5. FEEDBACK AND STATE OBSERVERS 265

Cq
1.4

1.25

a
^ 1.1

0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

Time

Figure 5.11: Isothermal CSTR: Feedback through resetting of the state.

The updated states then become the initial conditions for the model predic-
tions. For the case without model error, this method works well as shown
in Figure 5.1. In the presence of modeling error, simply resetting the state
usually produces steady-state offset. We simulated this controller using the
model of Section 5.2 and a +50 percent error in the concentration of species
A in the the feedstream, X F . The resulting system output is shown in Fig-
ure 5.11. The setpoint was 1.0 and the objective was the same is in Figure 5.5
with horizon length 5. We noted the presence of steady-state offset for hori-
zon lengths 6 through 9 as well. This controller produces offset for any finite
horizon. Since resetting the state provides such poor results, it is generally
not used in practice. We next describe two other approaches with better
properties: conventional MPC feedback and MPC with steady-state target
optimization.

5.5.1 Conventional MPC Feedback

By far the most common feedback method in MPC is to compare the mea-
sured output of the process to the model prediction at time k to generate a
disturbance estimate d k = Yk — y k, in which Yk and y k represent the process
measurement and model prediction, respectively. In forming the MPC ob-
jective function, the disturbance term is then added to the output prediction
over the entire prediction horizon to produce the modified objective

N-1
J = E IlYref — (Yk+jik — dk) II Q + Ilouk+j^kII
j=0
266 CHAPTER 5. MODEL PREDICTIVE CONTROL

This procedure assumes that differences observed between the process output
and the model prediction are due to additive step disturbances in the output
that persist throughout the prediction horizon. Although simplistic, this
error model offers several related practical advantages:

• It accurately models setpoint changes which often enter feedback loops

as step disturbances.

• It approximates slowly varying disturbances. Since errors in the model

can appear as slowly varying output disturbances, it provide robustness
to model error.

• It provides zero offset for step changes in setpoint.

Coupled with a linear step response model, feedback through the estimation
of a step disturbance has been extensively applied in industrial applications.
The stability and robustness have been analyzed for linear, unconstrained
systems by Garcia and Moran i [ 23]. More details for linear systems with
an emphasis on robust design using linear models may be found in the the
monograph by Morani and Zafiriou [53].
Figure 5.12 illustrates the effect of feedback using the disturbance term
with the isothermal CSTR model. The simulation used the same model
mismatch, and the same objective function weights as in Figure 5.11. The
results for horizon lengths 6 through 9 are included in Figure 5.12. The
MPC controller was not stabilizing for the nominal system with horizons
less than 6. Instability was observed in the simulations using these horizon
lengths for both cases with and without model mismatch.
Feedback through differencing model prediction and measurement does
not require a state-space description. Linear MPC methods used in industry
do not use a state-space model and incorporate the disturbance directly
into the MPC objective function. The feedback mechanism is sometimes
obscured by this treatment. A state-space description, even if not used in
the controller, allows a more natural interpretation of the conventional MPC
feedback procedure. Using an augmented state-space model, conventional
MPC feedback can be shown to be a particular form of a state observer for
the resulting system:
n
Xk+1 = f ( x k, k) (5.24)
d = d k(5.25)
k +1

Y k = 9(x k ) + dk(5.26)
5.5. FEEDBACK AND STATE OBSERVERS 267

♦♦ N=6
N ♦ N = 7 ---
3
^♦ N=8 ----

2.5

0 0.06 0.12 0.18 0.24

Time

1.4
N=6
♦
N=7
N=8
1.2 N=9....-.

Sir

^ 1

0 0.06 0.12 0.18 0.24

Time
25
:♦ N=6

- ♦♦ N = 8 - - _ _
20
•• • ♦ N = 9 ........

O
15
0
U
10
0

Time

Figure 5.12: Isothermal CSTR: Conventional MPC feedback with model

mismatch.
268 CHAPTER 5. MODEL PREDICTIVE CONTROL
4]T,

Taking the augmented state to be [xk we can use a variety of tech-

niques, including the extended Kalman filter or a moving horizon observer [55,
57, 56, 65], to estimate both the state and the disturbance. A general state
observer for the augmented system of Equation 5.24 through 5.26 has the
form

xk+1 = .f ( x k, u k) + h x ( x k, dk, Yk, uk)

d
k+1 = dk + hd ( x k, d k, y k, uk)

in which the x k and d k represent the observer estimates. To obtain the usual
MPC observer, choose

h
x ( x k, d k, y k, u k) = 0

h
d (
x
k, dk, Yk, u k) = Yk —g(k) —dk

This leaves an open-loop state observer for the original states x k and a
deadbeat observer for the disturbance dk . In the absence of information to
construct a disturbance model, this approach is a good choice in view of its
favorable features. With additional information and engineering judgment,
it is possible to obtain better state estimates using an observer that updates
x k as well as dk . Whether the extra effort leads to better control depends
upon the specific observer design and knowledge about the disturbances.
Several researchers have presented results for linear systems showing better
control using a full state estimator [26, 46, 55, 64] in the presence of noisy
output measurements and model mismatch.
Since the conventional MPC feedback uses a deadbeat observer of the
output disturbance, it is sensitive to random fluctuations in the output. This
is illustrated in Figure 5.13, which reproduces the situation of Figure 5.12
except that a zero-mean Gaussian disturbance with variance 0.09 has been
added to the output measurement. Figure 5.13 show the system response
for horizon lengths 7 and 9. The initial portion of the figures shows the
expected faster response of the longer horizon lengths. The faster response
comes at the expense of poorer disturbance rejection, i.e. noise in the input.

5.5.2 MPC with Steady -State Target Optimization

We next investigate an alternative feedback method that does not directly

incorporate the disturbance term in the MPC objective and can penalize
5.5. FEEDBACK AND STATE OBSERVERS 269

3.4
♦ N=7
\\ N = 9 ---
2.8

2.2

I
1.6
0 0.05 0.1 0.15 0.2

Time
1.4 I I I
N=7

I
N=9 ---
1.2

08
0 0.05 0.1 0.15 0.2

Time
I I I I
25
N=7
N= 9 ---
20

0 15

r, 10
O
M
5
0 0.05 0.1 0.15 0.2

Time

Figure 5.13: Isothermal CSTR: Response of system to Gaussian noise using

conventional MPG feedback.
270 CHAPTER 5. MODEL PREDICTIVE CONTROL

both u and Du. In this approach, we use a preprocessing step in which

new steady-state targets for the state and control, x t and u t , are computed
based on the disturbance estimate, dk , and the output setpoint, yref. The
new targets for the state and control are then incorporated directly into the
objective as a penalty on (x — x t ) and (u — u t ). This method is discussed
by Muske and Rawlings [55] in the context of linear systems. The method
presented here is a direct nonlinear analog.
To apply this method we solve the augmented system of Equations 5.24
through 5.26 at steady state to obtain target values for the internal states,
controls or output variable. These are used as setpoints in the MPC objective
function, exactly as we did in the first simulation result of Figure 5.1. Fig-
ure 5.1 (no model error, no disturbances) shows excellent speed of response
without offset. We can duplicate some aspects of that excellent performance
using the alternate feedback method.
Using the conventional MPC observer (open-loop state observer with
deadbeat disturbance estimator), we can choose steady-state target values
by solving the following nonlinear program at each iteration:

min Il u
ref — ut n
I l
(5.27)
Ut,xt

subject to:

Xt = f (xt, ut)
Yref = 9 ( x t) -^- dk
Xt E X

ut E U

In the above nonlinear program uref represents the input setpoint, for prob-
lems in which there are extra degrees of freedom in the input vector, i.e.
more inputs than output setpoints. The input setpoint is selected with
the understanding the the controller will enforce it as a setpoint only in a
minimum-norm sense. The solution {u t , x t } provides target values for the
states and controls that account for the current estimate for the output dis-
turbance dk . At each iteration, a new estimate of d k is obtained from the
state estimator and the target values are readjusted accordingly.
If the above nonlinear programming problem does not have a feasible
solution, then the output setpoint cannot be reached exactly; an alternative
5.5. FEEDBACK AND STATE OBSERVERS 271

NLP may be used to compute the control:

min I
y
re£ — ( 9( x t) +dk)IQ (5.28)
Ut,Xt

subject to:

Xt = f ( x t, ut)

Xt EX
Ut EU
In this case, we cannot simultaneously find values of {x t , u t } to satisfy the
steady-state system equations at the desired setpoint. This can happen,
for example, if an especially large disturbance enters the system, or if the
desired setpoint minus the disturbance does not correspond to a steady state
of the model equations. We always assume that the model has at least one
real steady-state solution that satisfies the input and output constraints, so
Equation 5.28 always has a feasible point.
This technique can give excellent results. Figure 5.14 shows the response
of the isothermal CSTR with model mismatch using a 1 percent weight on
(U-Ut) 2 in relation to the weight on the output. The value of u t was adjusted
at each sampling time using the nonlinear program of Equation 5.27 with
u
ref = 25, the nominal steady-state value. Unlike the previous examples that
used a penalty on Du, the response is rapid. It satisfies both the primary
and secondary control objectives: It brings the system to the setpoint y = 1
without offset and it does so at a low feed rate. These results indicate that
this is the clear method of choice for control of the reactor of Section 5.2.
The curves shown in Figure 5.14 were computed with N = 5. Slightly slower
response was observed with shorter horizon lengths, but the differences were
very small and would not appear as distinct curves if plotted in Figure 5.14.
Even with horizon length 1, the response was very similar, and the time
necessary to compute the MPC control was much less. Recall that this
system was unstable for short horizons using conventional MPC feedback.
Up to this point, all of the isothermal CSTR examples shown have started
from the same initial conditions. It is necessary to check the behavior of the
controller over a range of initial conditions before deciding on a specific
controller design. Figure 5.15 shows trajectories of the system from vari-
ous initial conditions. The system was controlled using feedback through
steady-state target optimization with horizon length 3. The initial condi-
tions were selected to lie along the perimeter of the rectangular region of
272 CHAPTER 5. MODEL PREDICTIVE CONTROL

3.2

2.7
a1
5
ci:
2.2[
0 0.02 0.04 0.06 0.08
Time
N
1.3
N
1.2

1.1

71 1
N
0 0.02 0.04 0.06 0.08 0.1 0.1
Time

24
41

20
0
16
0
A
12
0 0.02 0.04 0.06 0.08 0.1 0.12
Time

Figure 5.14: Isothermal CSTR: System response using target optimization.

5.5. FEEDBACK AND STATE OBSERVERS 273

1.6
G}.;---------R.----------El: ----- ---- ^: -------- ^: ---------C7.
1.4

II
1.2

- ,

41 0.8
0--------------------- -

0.6
C -- ""^7- --- G'---------- G----------- C3-----------{'1
0.4
1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 3.4
State Variable x1

Figure 5.15: Isothermal CSTR: System response from various initial condi-
tions.

the state-space shown in the diagram. The initial conditions and the final
value of the controlled system are indicated with boxes. Since the controller
is implemented at discrete points in time, the system trajectories can cross,
unlike the situation with continuous-time feedback systems in two dimen-
sions. This figure shows that the region of attraction for the MPC controller
in the presence of model error is not negligible and that the controller would
be expected to perform well from a variety of initial conditions.
Using both the conventional MPC feedback and feedback through target
optimization, we compared the ability of the MPC controller to reject zero-
mean Gaussian disturbances in the output. As in Figure 5.13, we used
discrete, additive, zero-mean Gaussian noise in the output of the process
with variance 0.09. To allow direct comparison of various methods, each
simulation used the same noise sequence, which we obtained by re-seeding
the random number generator with the same seed value.
Figure 5.16 shows the response of the two feedback methods. The tar-
get optimization feedback method with horizon length 3 is compared to the
conventional DMC controller using horizon length 7. The target optimiza-
274 CHAPTER 5. MODEL PREDICTIVE CONTROL

tion feedback is insensitive to the horizon length and provides virtually the
same response for horizon lengths from 1 to 5. During the course of the
simulations, about 15 percent of the initial nonlinear programming prob-
lems of Equation 5.27 had no feasible point (which never occurred without
the random disturbances). In these cases, the NLP of Equation 5.28 was
used to determine the reference values. Figure 5.16 shows that feedback
through target optimization with no weighting on Du is more sensitive to
disturbances than the conventional MPC feedback method with the weight-
ing on Au. This effect clearly shows up in the input. To attenuate the
effect of the zero-mean disturbances, one can choose to weight Du in the
controller optimization, but there are other choices as well. For example, we
can incorporate a moving average filter to estimate the disturbance term:
dk = 0.9 d k _ 1 +0.1 (Yk — g (x k )). This reduces the occurrence of infeasibility
in the initial NLP to about 5 percent. The effect of this filter is illustrated
in Figure 5.17. As in the previous figure, the conventional MPC feedback
controller with horizon length 7 is included for reference. From Figure 5.17,
we can see that feedback through target optimization combined with the fil-
ter for the disturbance term has greatly improved the disturbance rejection
ability of the controller, while retaining the ability of the feedback method
to provide robustness and fast response.

Figure 5.17 shows that an MPC controller with a short horizon using
target optimization feedback can perform as well as or better than conven-
tional MPC feedback with longer horizons in terms of robustness, speed of
response and disturbance rejection; however, the feedback through the tar-
get optimization solution is much faster to implement. Using horizon length
9 and conventional MPC feedback, the simulations took about 20 hours to
simulate 120 time steps on an IBM-compatible 80486/33 MHz computer. At
each sampling time, this method required the solution of a nonlinear pro-
gramming problem involving 99 decision variables. In contrast, using feed-
back through target optimization, the initial NLP needed to compute the
targets involves 3 decision variables and requires an insignificant amount of
time to compute. The control computation itself involved as few as 6 deci-
sion variables (for N = 1) to 18 variable (for N = 3) and could easily have
been computed in real time using a sampling interval of 7.2 seconds.
5.5. FEEDBACK AND STATE OBSERVERS 275

a 3

0 0.05 0.1 0.15 0.2

Time
Cq
1.4

1.2

1
O I I I I`-'
08
0 0.05 0.1 0.15 0.2
Time
80
4Z 60

40
0
0 20
U
0
0 0.05 0.1 0.15 0.2
Time

Figure 5.16: Isothermal CSTR: Effect of output disturbances using feedback

through target optimization. Solid line: feedback through target optimiza-
tion. Dashed line: conventional MPC feedback.
276 CHAPTER 5. MODEL PREDICTIVE CONTROL

2.5 Ir

15
0 0.05 0.1 0.15 0.2
Time

1.4

1.2

O 0.8
0 0.05 0.1 0.15 0.2
Time

26

20

14
0
U Q

0 0.05 0.1 0.15 0.2

Time

Figure 5.17: Isothermal CSTR: Effect of output disturbances using feedback

through target optimization. Solid line: feedback through target optimiza-
tion. Dashed line: conventional MPG feedback.
5.6. SELECTION OF TUNING PARAMETERS 277

5.6 Selection of Tuning Parameters

As illustrated by some of the previous examples, the choices of control hori-
zon, weight matrices and feedback methods have a profound effect on MPC
nominal stability and robustness. Ideally, we would prefer to know a priori
what range of tuning parameters would provide nominal stability and robust-
ness. We would then be able to select the tuning parameters to meet specific
process objectives, without being concerned that a poor choice could drive
the system to instability. Implementable sufficient conditions that guaran-
tee nominal stability [55] and robustness [1] of MPC are known for linear
systems. Unfortunately, the known sufficient conditions for nonlinear sys-
tems, such as those described in Section 5.7 (infinite horizon, zero final-state
constraint), are usually much too strong to offer practical implementation
guidance, and we must resort to a set of heuristics based on the extrapolation
of linear systems results, numerical simulations and experiments.
The most significant tuning parameters that must be selected are pre-
diction horizon, control horizon, sampling interval, and penalty weight ma-
trices. In addition, the control engineer must decide whether to penalize Au
or whether to use a filter in the feedback loop. As part of the numerical
implementation, the discretization scheme and software must be selected.
Some guidelines for selection of tuning parameters are provided below. In
the following discussion, we assume that an explicit model is available and
that the model has been acceptably verified.

• Sampling Interval : For stable, minimum phase systems, stability does

not depend on the sampling interval; however, to ensure good closed-
loop performance, the sampling interval should be small enough to
capture adequately the dynamics of the process, yet large enough to
permit the on-line computations necessary for implementation. Large
sampling interval can result in ringing (excessive oscillations) between
sample points. An example of this phenomenon is provided by Garcia
and Morani [23] using a linear system.
For unstable systems, robustness depends critically on the sampling
interval chosen. Recall that the choice of sampling interval for the
fluidized-bed reactor problem of Section 5.2 was based on the doubling
time of the unstable mode of the nominal system. The nominal system
showed stable behavior under MPC, yet the MPC controller is unstable
with model mismatch. We applied a +20 percent error in the pre-
exponential factor of the first reaction. Simulation results are presented
278 CHAPTER 5. MODEL PREDICTIVE CONTROL

in Figure 5.18. Using the original sampling interval of 1.09, we see

that the reactor experiences a temperature runaway near the end of
the third sampling period. This constitutes a catastrophic failure.
(Although our integrator bravely attempted to integrate the model
past the runaway, we have little confidence in either the model or the
integrator beyond this point. Results shown after the runaway must
be viewed skeptically.)
After receiving these simulation results, we cut the sampling interval
in half to 0.545 and observed similar unstable behavior, with the onset
of the runaway occurring slightly later. Halving the sampling interval
again to 0.2725 provided the second curve indicated in Figure 5.18,
which demonstrates good control without any tendency toward run-
away.
For unstable systems, there exists an inverse relationship between the
model error and the maximum allowable size of the sampling interval.
In the example of Figure 5.18, with the true pre-exponential factor al
20 percent higher than the model, we observe a temperature runaway
between sampling periods. Since feedback is only incorporated at the
sampling points, the sampling interval must be chosen sufficiently small
to detect any unexpected increase in temperature before it becomes a
runaway situation. This is a difficult design issue that has not been
fully explored for nonlinear systems. For stable linear SISO systems,
Zafiriou and Morani [79] have provided some criteria for sampling in-
terval selection that later appeared in their book [53]. The interested
reader is referred to these works as well as to standard texts on digital
control and digital signal processing for further discussion of sampling
intervals for linear systems.

• Prediction Horizon : For linear systems, several researchers have pro-

vided selection criteria for the prediction horizon that assure stabil-
ity [23, 62, 67]. The question has not been resolved for nonlinear sys-
tems, but we can use the simulation results of Section 5.2 to illustrate
some general principles.
With conventional MPC feedback, Figure 5.12 shows that longer hori-
zons tend to produce more aggressive control action, more overshoot,
and faster response. From Figure 5.13, we see also that longer predic-
tion horizons are more sensitive to disturbances, although this effect
can be partially mitigated by including a filter in the feedback loop.
5.6. SELECTION OF TUNING PARAMETERS 279

0.6
OT=1.09
0.4 \ DT=0.2725 ---
Q \
F3
0.2 -- -------------------------------
0 --------- ---------- ----------- --------------------- ----------- ---------

0 5 10 15 20 25 30 35 40

------ ----- ---- ----- ----- ----- -----

0.4
0T = .09
0.2 OT=0.21725 ---

0 ------ -- - - -- --------- ---------- ------- ------ ------

0 5 10 15 20 25 30 35 40

2
0T= 1.09
1.6 OT = 0.2725 ---

a 1.2
0.8

0 5 10 15 20 25 30 35 40

4
° 0 v- ------
0
-4 OT = 1.09
OT = 0.2725 ---
-8
0 5 10 15 20 25 30 35 40
Dimensionless Time

Figure 5.18: Fluidized-bed reactor: Effect of sampling interval.

280 CHAPTER 5. MODEL PREDICTIVE CONTROL

Nominal stability is strongly affected by the horizon length, as illus-

trated by Figure 5.5, in which we observed a critical horizon length.
Prediction horizons shorter than the critical value produced an unsta-
ble closed-loop system.
Based on linear systems results [23] and our simulations, the. effect
of increasing prediction horizon diminishes as N becomes large, and
the advantages of longer horizons are outweighed by the increase in
computations required.
With feedback through the nonlinear programming technique, we ob-
served that prediction horizon had almost no effect. Whether this
applies more generally is unknown and is topic of current research.

• Control Horizon : Linear systems results [23] indicate that shortening

the control horizon relative to the prediction horizon tends to produce
less aggressive controllers, slower system response and less sensitivity
to disturbances. The effect is very similar to that of increasing the
penalty on control action in. the MPG objective function. We investi-
gated this effect using the fluidized-bed reactor model of Section 5.2.
For the nominal system, Figure 5.7 showed the effects of using pre-
diction horizon N = 4 and control horizons M = 1 and M = 2.
Comparing Figure 5.7 with Figure 5.6, we observe that the time scale
of the system response is much longer for I1 = 1 or II = 2 compared
to Al = 4 with less aggressive control action. Unfortunately, in the
case with model mismatch, this is exactly what is not needed, as in-
dicated in Figure 5.19. Slowing the control action permits the reactor
temperature to run away for M = 1. For Ir1 = 2, the control action is
quick enough to prevent the runaway.

• Penalty Weights : For choice of weighting matrices, Bryson and Ho [6]

suggest applying weights that are inversely related to the maximum
acceptable range of the variables being penalized. We have gener-
ally found that this technique rarely provides good results because it
over-penalizes control action. To provide good performance by the op-
timization algorithm, we recommend applying a penalty to setpoint
deviations that places the objective in the interval 1-100 for the range
of expected conditions, and then applying small penalties (less than
10 percent of the output penalty) on control or control increments, to
achieve good closed-loop performance.
5.6. SELECTION OF TUNING PARAMETERS 281

0.6 M =1
M=2 --
0.4 `
`.
0.2

0
0
IIIIII EEEEEI
5 10 15 20

-- ------------------------ M =1------
0 .6 M=1
0.4 M=2---
w
0.2
0 -------------------- ---- _------- --------------------- ---------------------
0 5 10 15 20

2
M=1

C^

a
1.6

1.2 : ----
J\\
------ ------------------
M = 2 ---

^,

0.4
0 5 10 15 20

4
0 0
-4 M=1
O M=2 ---
-8
-12
0 5 10 15 20

Dimensionless Time

Figure 5.19: Fluidized-bed reactor: Effect of control horizon.

282 CHAPTER 5. MODEL PREDICTIVE CONTROL

In the simulations results presented in this chapter, we typically ap-

plied a penalty weight of 100 on the integral square deviation of the
output from the setpoint. Relative to the penalty on setpoint devia-
tions, in our simulation studies we applied a 0.1 to 1 percent penalty
to the control or control increments. Naturally, these weight must be
adjusted if the terms which contribute to the objective have widely
differing magnitudes arising from dimensional differences. See Gill, et
al. [27] for a discussion of scaling issues in optimization.

• Collocation approximation : For the simulations using the isothermal

CSTR model, we used 3 internal collocation points in the interior of
each element, as well as points on the element boundaries. For the
fixed-bed reactor, we generally used two interior collocation points to
reduce the computation time. In one case, we repeated the fixed-
bed reactor simulation in the presence of model mismatch using both
1 and 4 interior collocation points. The higher order approximation
gave virtually identical results as the case with 2 interior collocation
points. The lower order approximation resulted in a reactor runaway.
Clearly, there are relationships between the size of the sampling inter-
val, model error, the degree of the interpolating polynomial, and the
performance and stability of MPC using orthogonal collocation. These
relationships have not been quantified, so we can provide only a set of
heuristic guidance based on our own experience. We suggest at least
two interior collocation points be used, and no more than about 5.
Unless the model predicts rapidly varying solutions within one finite
element, a smaller number is desirable to reduce computation time.
• Filter in feedback loop : In Section 5.5, we showed that filtering the
feedback signal when using steady-state target optimization provided
good disturbance rejection and fast system response. The choice and
effect of the disturbance filter is strongly system dependent. In the ex-
ample of Section 5.5, we (naively) chose a very simple moving average
filter with relatively slow response and observed excellent results. Ap-
plying the same procedure to the unstable fixed-bed reactor problem,
the filter can produce catastrophic failure, depending on the speed of
the linear filter chosen.
As in Section 5.5, we used the filter d k = a f dk _ 1 +(1—a f ) ( Yk — g (xk))
in which ca f = 0.9. The same filter applied to the fixed-bed reactor
problem produced a runway in simulations that included a zero-mean
5.7. STABILITY AND PERFORMANCE 283

Gaussian disturbance in the output. Figure 5.20 shows the effect of

the disturbance, both with and without the filter. The variance of the
Gaussian disturbance is 0.01. As before, we do not place any confi-
dence in the post-runaway simulation results; however, the computed
temperature following the runaway illustrates that the algorithm with
the filter slows and smooths the computed control input. We previ-
ously observed that increasing the control penalty or decreasing the
control horizon relative to the prediction horizon also slowed control
action and also resulted in reactor runaway.
We attempted to avoid the runaway situation by using a faster filter.
Using a f = 0.5 produced the simulation results shown in Figure 5.21.
We have succeeded in avoiding the runaway, the filter is effective in
reducing noise sensitivity in the control signal, and slightly less effec-
tive in reducing variation in the temperature, which is the controlled
output. If we consider data in the final portion of the time interval of
Figure 5.21, the controller is essentially in a regulator mode to main-
tain the setpoint in the presence of disturbances. A rough measure of
the effectiveness of the filter is the ratio of standard deviation with the
filter to the standard deviation without the filter. After subtracting
linear trends, the filter with a f = 0.5 produced standard deviation
ratios of 0.538, 0.720 and 0.577 for the control, temperature and feed-
back (d k ) signals, respectively. (For comparison, the same filter used
to attenuate a Gaussian noise signal alone produces a ratio of 0.578.)
These results indicate that a filter in the feedback loop can be effective
in reducing the controller sensitivity to output disturbances; however,
like other methods that slow system response, the filter can induce
unstable closed-loop behavior with unstable processes in the presence
of uncertainty. Simulation studies are essential to detect such behavior
before implementation.

5.7 Stability and Performance

5.7.1 Nominal Stability

The first property of a control system that should be satisfied is nominal

stability, i.e., stability for systems free from modeling errors or disturbances.
For linear systems without constraints, stability can be verified by check-
ing the eigenvalues of the closed-loop system. For nonlinear systems, there
284 CHAPTER 5. MODEL PREDICTIVE CONTROL

I I I
0.6 No filter --
Filter constant 0.9 — — — ' ' '
Q 0.4 ^. ''^

0.2 ^

0 5 10 15 20

0.6 - t
No filter
0.4 I Filter constant 0.9 ---
^
0.2 l
----....1---- --^-'
0 5 10 15 20

1.4 '
\ No filter
Cd 1.2 N Filter constant 0.9 ---

G,, 1 '
0.8 -----

0.6
0 5 10 15 20

No filter
4
O Filter constant 0.9 ---

o 0 v v - ^1 _ ,Vt LL

-4
0 5 10 15 20

Dimensionless Time

Figure 5.20: Fluidized-bed reactor: Disturbances and filtering effects.

5.7. STABILITY AND PERFORMANCE 285

0.6 No filter
Filter constant 0.5 ---
0.4

0
0 5 10 15 20

0.7

Filter constant 0.5 — —

0.5
0 5 10 15 20

°;

O o
' No filter
}^..^ 0.7 Filter constant 0.5 ---

E^ 0.6

0 5 10 15 20

No filter
4 Filter constant 0.5 — — —
O

-4 I
0 5 10 15 20

Dimensionless Time

Figure 5.21: Fluidized-bed reactor: Disturbances and filtering effects.

286 CHAPTER 5. MODEL PREDICTIVE CONTROL

is no simple equivalent criterion. Since closed-loop solutions to MPC con-

trol problems for nonlinear systems are almost never available, the Second
Method of Lyapunov has been extensively used to analyze the stability of
MPC. With sufficiently strong conditions on the nonlinear program that
provides the MPC controller, we can view MPC as a method for generating
Lyapunov functions for nonlinear systems.
Since Lyapunov analysis assumes such a prominent role, we repeat the
basic theorem below. The following is adapted from Vidyasagar [76]:

Theorem 5 (Lyapunov Stability Theorem)) Let x k be a sequence in R'

indexed by k E {0, 1, 2 ...}. The dynamic process Xk is asymptotically stable
if there exist a function V : R n - J2+ and three functions a, 3 and 'y of
class K such that
a
(II x kII) <_ V(x) < 3 (II x kUI) (5.29)

AV
( x k) V
( x k+ l ) -
V
( x k) <_ - 1r(II x J) (5.30)

Class K functions are a specific class of positive definite functions. Using

the definition of Vidyasagar [76], a class K function a -+ R+ is contin-
uous and strictly increasing and satisfies a(0) = 0. If a, , 3 and -y have the
properties of class K functions only in a neighborhood the origin, then the
Lyapunov stability theorem provides only a local result.
Model predictive control provides an objective function that satisfies the
sufficient conditions of Theorem 5 if we include the final state constraint
x k+Nlk = 0. For simplicity, we will use a quadratic stage cost Il x
k+jlkllQ +
Il uk+jlkII R with Q and R positive definite. Without loss of generality, we can
assume that f (0, 0) = 0 through a simple change of coordinates. Using the
MPC objective function

N-1
J*(x k ) = min u
(II x k+j^kll Q + Il k+j^kIIR) + IQ
Uk+Jik j=0

we see that J*(Xk) > II x k1IQ and the quadratic function

a
(II x kII) = am(Q) Il xkli2

(in which ami,,(Q) is the smallest singular value of Q), satisfies the lower
bound of Equation 5.29. Note that Xk and n k are equivalent to x k ^ k and
u k1k , respectively.
5.7. STABILITY AND PERFORMANCE 287

Equation 5.30 requires a bound for the difference between values of the
Lyapunov Function for subsequent states. Using the MPC objective func-
tion as our candidate Lyapunov function, Equation 5.30 is also satisfied by
including the final state constraint x k+Nlk = 0. If we represent a sequence
of controls that corresponds to the optimal J*(x k ) by
( * * *
j u klk , U k* +llk , ... 2Gk+N-1^k

and let J' be defined by l

*
J' = J (x k ) — (UU x kUQ + jukj^R) (5.31)
N-1
= (I^ x k+jlkiiQ + Iluk+jJkIIR)
j=1

The value of J' is simply J * ( x k ) without the contribution of the first stage. In
this case there is also no contribution from the final state to the summation
since it is constrained to be zero.
After the control uk^ k is implemented, the state moves to x k+llk =
f Since we are considering the nominal system, the state moves
( x k , uk^ k ) .

exactly as predicted. At the new time k + 1, the optimal control moves

computed at time k still constitute a feasible sequence to which we can ap-
pend a zero control without affecting the value of J'. This feasible sequence
{ u k+11k ... uk+N-11k, 0} may or may not be optimal at time k + 1, but it
,

provides an upper bound for the optimal objective J*(xk+l) so that we have
if
> J* ( x k+l) . Combining this inequality with Equation 5.31, we obtain

J* x J* x (5.32)
( k) — ( k+l) > IlxkllQ
Q
> min(Q) Ixk112 (5.33)

which satisfies the inequality of Equation 5.30 of Theorem 5 after a change in

sign. Notice that it was not necessary to consider the effects of constraints.
For the nominal system, the existence of a solution to the nonlinear program
at some initial time is sufficient for the existence of subsequent feasible so-
lutions. Figure 5.22 provides an intuitive basis for the stability arguments
above. After the control at time k is selected and implemented, a new
problem is defined over the subsequent "window" as shown in the figure.
The new objective includes state and control contributions on the interval
[k+1, k+N+ 1]. The value of the optimal objective at k + 1 must be less
288 CHAPTER 5. MODEL PREDICTIVE CONTROL

• State Trajectory

O State Prediction

• 0 O
0 O _J L Control Implemented
O
Control Predicted

k-1 k k+1 k+N

Prediction Horizon at Time k ►{

Hr— Prediction Horizon at Time k + 1 • 0

Figure 5.22: Model predictive control concept.

than that at k, because of the final state constraint. The decrease is lower
bounded by 1 1 x k I 1 Q, since the contribution of the state at k is no longer
included.
To apply the Lyapunov Stability Theorem to MPC, it remains to be
shown that there exist an upper bound ,C3(^IxIl) on the MPC objective. The
upper bound is closely related to the null-controllability of the system. In
general, a priori evaluation of null controllability is difficult for nonlinear
systems; the case of constrained nonlinear systems is even more difficult.
Suppose, however, that we knew that the origin could be reached using
feasible controls from any state in some specific subset S of R' and the
state trajectory under this sequence of controls remained in the feasible set.
Then the MPC objective function could then be evaluated using the feasible
control sequence and would provide an upper bound on the optimal MPC
objective. The argument proceeds as follows: Let x k be some state in S such
that the control sequence { u k^k, u k+11 k, ... u k+N-ij k } produces a final state
x
k+NIk = 0. We can construct a (non-optimal) MPC objective

N-1
J(xk)
_ E, I lXk+jlkIIQ+I
J uk+jlk (5.34)
j=0

Again, there is no contribution of the final state because of the zero end-
state. If J(x k ) is continuous, then it is a class 1C function, since it is lower
bounded by HxkU Q which is a class 1C function. The arguments above may
5.7. STABILITY AND PERFORMANCE 289

be formalized in the following theorem:

Theorem 6 (Stability of MPC) For the nominal system

X
k+1 = f(xk,uk)

f (0, 0) = 0

if the MPC objective function is continuous, then the origin is an asymptot-

ically stable fixed point under MPC with a region of attraction that consists
of the subset of I1 that is constrained null controllable.

A proof of the theorem follows from the previous discussion.

It is possible to generalize the previous discussion by substituting a gen-
eral class K stage cost for the quadratic functions used above. In this case,
asymptotic stability is retained. It is also possible to arrive at a similar sta-
bility result using an infinite horizon in which the final state constraint may
be viewed as a finite parameterization of the more general infinite horizon
problem. Details are provided by Meadows et al. [49]. Similar results were
also reported by Keerthi and Gilbert, who considered both the infinite hori-
zon and the final state constraint [34]. Stronger results are available. If the
stage cost is quadratic, as discussed above, then MPC produces exponential
stability. The result follows from the fact that the bounding class IC func-
tions a, ,Q and 'y are quadratic [68]. See also Vidyasagar [76] for a continuous
time proof.
We have not directly addressed the question of constrained null con-
trollability. As a practical matter, null controllability is usually verified in
the process of solving the nonlinear program for the MPC controller. If no
feasible point can be found, then for practical purposes, the system is not
null-controllable. A numerical algorithm is not equivalent to a mathematical
proof, but as a practical matter, if optimization software cannot find a fea-
sible point for some initial state of interest, the system is a poor candidate
for control with MPC. This is an aesthetically unsatisfactory approach to
the verification of nonlinear controllability; however, it must be noted that
the same problem exists with respect to linear MPC controllers subject to
constraints, and has not prevented their widespread implementation.
The topic of nonlinear constrained controllability is an active research
area which has not yet produced a complete theory. For further discussion
of nonlinear controllability concepts see the following: [28, 29, 58, 70, 71, 76].
290 CHAPTER 5. MODEL PREDICTIVE CONTROL

5.7.2 Steady-State Performance

In addition to nominal stability, an evaluation of a control method should

consider robustness to model error. In this section we state some steady-
state results that hold for MPC controllers applied to arbitrary plants
rather than the nominal plant. The price we pay for the generality of con-
sidering arbitrary plants is that the results are in the form of alternative
conditions, i.e. either the closed loop does not approach a steady state or
the plant goes to a steady state with certain guaranteed properties, such as
zero offset. The value of these results is that they characterize the steady
state that the closed loop exhibits for any plant, in terms of the nominal
model only. The steady-state properties can therefore be examined without
knowledge of the plant.
We are concerned in this section with the target optimization feedback
form of MPC. Some of the methods we use subsequently can also be applied
to develop analogous results for the conventional MPC feedback scheme, but
we do not investigate that here. Our opinion is that the target optimization
provides a more general framework to handle a larger variety of process mod-
els than can be addressed with the conventional MPC feedback, so we focus
our attention there. First we summarize the three parts of the controller
discussed in Section 5.5.2

State estimator. The state estimator takes the current control and plant
measurement and updates the state and disturbance estimates via
x
k+1 = .f ( x k, uk)
d
k+1 = Yk — 9(xl^)

Steady-state target optimization. The steady-state target optimiza-

tion takes d k and Yref and produces as output x t , u t through solution of

i I ref — u t
m n u (5.35)

subject to:

Xt = f ( x t, ut)

Yref = 9( x t) + dk
Xt E X

ut E U
 5.7. STABILITY AND PERFORMANCE 291

Notice that we consider explicitly any controller input and state (output)
constraints in the target optimization as well. If this optimization problem
does not have a feasible point, the following nonlinear program is used to
compute the targets
d
min bref — ( g ( x t) + k)11 2 (5.36)
subject to:

Xt = f ( x t, ut)

Xt E X

Ut E U

We always assume that this optimization problem has a feasible point, i.e.
the model has at least one steady-state solution satisfying the constraints.

MPC control law. Finally, the MPC controller takes x t , u t and x k and
produces the control input u k from the solution of
N-1
min E ll xk
+jl k —
x
tllQ + Ii ttk
+jl k — utII R (5.37)
Uk+jlk j=0

subject to:

xklk = Xk
x
k+Nlk = Xt

xk+jlk E X

uk+jlk E U

The first optimal input is then applied to the plant, Uk = uklk.

The first result states that if the closed-loop system goes to steady state,
then the steady-state plant output and input satisfy the target optimization
problem.

Theorem 7 (Steady-state properties) If

1. The steady-state model, x = f (x, u), has a solution, x E X, u E U.

2. The MPC control law is nominally asymptotically stabilizing.
292 CHAPTER 5. MODEL PREDICTIVE CONTROL

3. The closed-loop goes to a steady state, y k -4 Ys, u k -+ u s, Xt xts,

,ut -+ uts.

Then either H u ref — u sII R is minimized and Ys = y ref if Equation 5.35 has a
feasible solution at steady state or H yref — y SIIQ is minimized.

Proof: At steady state (Assumption 3), the estimator and target optimiza-
tion yield

xs = f ( x s, us)
d
s = y
s — 9(xs)
x
ts = f ( x ts, uts)
xts E X

ti ts E U

The target optimization is well defined because of Assumption 1. We next

show that because the controller is nominally asymptotically stabilizing (As-
sumption 2), x s = x ts , i.e. the state estimate converges to the steady-state
target. Consider the control problem with steady-state input and state tar-
gets (Assumption 3). The state estimate evolves according to

x
k+l = flxk,us)

Because the controller is nominally asymptotically stabilizing and the state

estimator evolves according to the nominal model, the stage cost in Equa-
tion 5.37 goes to zero, which implies i S = x ts and u ,s = u t,s . From the
disturbance estimate equation

Ys = 9( x ts) + ds

Therefore, if at the steady state Equation 5.35 is being solved then the steady
input minimizes I u ref — u sII R and ys = y ref . If this target problem has no
feasible solution at steady state, then Equation 5.36 applies and the steady
plant output minimizes I I y ref — y s I I Q
Because of the constraints, plant-model mismatch, and plants with more
setpoints than inputs, it may not be possible to satisfy the conditions neces-
sary for zero steady-state offset. We can add further restrictions to provide
conditions that do ensure zero steady-state offset.
5.8. DISCUSSION 293

Corollary 1 (Zero steady-state offset) If the conditions of Theorem 7

are satisfied and there exists x t , u t satisfying

xt = f ( x t, ut)
Yref = 9( x t) + dk
xt EX
ut E U

for every Yref and dk , then there is zero steady-state offset, ys = Yref
Proof: The added conditions of the corollary ensure that the target opti-
mization in Equation 5.35 has a feasible point. Therefore, zero steady-state
offset follows from Theorem 7.

5.8 Discussion
MPG has its roots in the development of optimal control theory in the late
1950's, in which the "state-space" approach was considered a modern al-
ternative to the frequency-domain techniques that had dominated control
theory and practice since the 1930's. Kalman's landmark paper [30] consoli-
dated and extended much of the then current work in the field, and provided
rigorous definitions for the now-common concepts of controllability and ob-
servability. Although Kalman reported work for linear systems, many of the
key concepts may be extended to nonlinear systems.
It was recognized early that optimality in the linear-quadratic problem
is not sufficient for stability. Kalman [30] provided sufficient conditions for
stability for this case. Several other workers considered formulations that
could provide a stabilizing LQ optimal controller. One of these was Klein-
man [36], whose "easy way" to stabilize a linear system was equivalent to a
finite-horizon LQ optimal control that is subject to the final state constraint
x(T) = 0. Kleinman followed up his continuous-time result with a discrete-
time analog [37] in which he provided a criterion for selection of an initial
condition for the matrix Riccati Difference Equation (RDE) that arises from
the LQ optimal control problem in discrete time.
Thomas and coworkers[72, 73, 74] recognized that the zero final state con-
straint could be incorporated into the LQ optimal regulator by inverting the
Riccati Differential (or Difference) Equation and integrating (or iterating)
the result from an initial condition of zero. Conceptually, this is equivalent to
294 CHAPTER 5. MODEL PREDICTIVE CONTROL

an infinite final state weighing matrix in a finite-horizon LQ optimal control

problem. Kwon and Pearson provided a firm theoretical basis for using the
zero final-state constraint to provide stability for time-varying and invariant
linear systems, including results for continuous time [42] and discrete time
systems [43]. To implement MPG in continuous time requires that a solution
of the optimization be available continuously. For linear systems, this means
that the solution of matrix Riccati differential equation must be available.
Kwon et. al [41] investigated the time-varying case and showed that a sta-
bilizing solution to the matrix Riccati difference equation could be updated
based on the system equations, without the need for a new solution at each
time. Although not presented, the authors pointed out that the results are
also applicable to discrete-time systems.
By the late 1970's, a substantial body of research results had been de-
veloped that provided a solid theoretical foundation for implementation of
model predictive control. At about the same time, the price instability of
feedstocks and energy in the process industries demanded higher efficiency
from existing plants. Rather than control processes in midrange of opera-
tions where traditional linear controllers would suffice, it became necessary
to operate closer to limits using models that could be generated quickly
with a minimum of modeling effort. In 1978, Richalet et. al [63] presented
IDCOM and in 1979, Cutler and Ramaker [13] presented Dynamic Matrix
Control (DMC), both of which responded to these new demands. These ap-
proaches used either step or impulse response models in discrete time whose
coefficients can be determined from plant step test data. Since processes are
assumed to be time-invariant, the plant dynamics can be described by a ma-
trix of constant coefficients, hence the acronym DMC. The DMC controller
used the linear system represented by the dynamic matrix to predict future
process output based on current and future control action. The control was
selected that minimized a quadratic objective based on control and process
variables predicted by the model. Constraints on the control variable were
explicitly incorporated by the optimization algorithm in subsequent formu-
lations [24].
In the chemical engineering literature, model predictive control came to
be identified with the methods of DMC, including linear step response mod-
els, finite prediction horizons, and the use of the DMC disturbance model.
DMC assumes that the difference between current processes measurements
and the model prediction is due to an unmodeled disturbance that maintains
the same value throughout the prediction horizon; to obtain a prediction, one
must use that model equations and add to the predicted variables the value
5.8. DISCUSSION 295

of the current disturbance estimate. The effect is extremely useful, since it

eliminates steady-state offset for step changes in setpoint, as discussed in
Section 5.5.
The basic DMC algorithm has undoubtedly been a phenomenal success
by any measure; however, its origins as a somewhat ad hoc method necessar-
ily left some questions unanswered. For example, with the introduction of
output variable constraints, some stable processes can become closed-loop
unstable. The multivariable nature of DMC allows for a large number of
tuning variables in the objective function. The length of the prediction hori-
zon is a tuning parameter. Often, the control horizon is different from the
prediction horizon—yet another tuning parameter is available.
The influential paper by Garcia and Moran i [ 23] strengthened the trend
toward identification of DMC and its descendants with MPC in general. For
the unconstrained case, Garcia and Moran i provided some stability results
and suggested tuning parameters for the DMC algorithm.
The step response model used in DMC admits only stable processes.
Seeking to extend the applicability of the step response models and to in-
corporate the large body of existing results for state space models, Li and
coworkers [46] and Lee et al. [44] produced MPC algorithms that converted
the step response models into a state-space form. The new formulations
were based on the step response models and incorporated the step response
coefficients. Although Eaton and Rawlings [19] and Lee et al. extended the
basic method to include integrating processes, formulations that depend on
step response coefficients are unsuitable for general unstable processes.
In parallel, Clarke and coworkers opted for a transfer function model of
the process and formulated the approach known as generalized predictive
control (GPC) [9, 10, 11]. Part of the aim of this development was to be
able to handle unstable plants.
Rawlings and Muske [54, 55, 62] departed from the step response models
common in chemical engineering practice to use the standard linear state
space model to formulate a stabilizing constrained MPC method. Their
control formulation guaranteed nominal stability using any valid tuning pa-
rameters.
Inequality constraints in any MPC method lead to a non-linear feed-
back control law. A natural extension of existing results for linear systems
involves MPC with explicitly nonlinear models. In the chemical engineer-
ing literature, Patwardhan and coworkers [60] and Eaton and Rawlings [19]
propose a simultaneous optimization-model solution approach that provides
a continuous-time MPC controller. Although no stability proofs were pro-
296 CHAPTER 5. MODEL PREDICTIVE CONTROL

vided, numerical results indicated stability using chemical reactor example

problems. Patwardhan, et al., also demonstrate robustness to parameter
error.
Chen and Shaw [8] provide an early discussion of model predictive con-
trol with the end point constraint for control of nonlinear systems. A rigor-
ous and comprehensive approach to the nonlinear problem in discrete time
was presented by Keerthi and Gilbert in a series of papers in the mid-
1980's [32, 33, 34, 35] . They consider a general nonlinear, time-varying
dynamic system with a general cost function defined on either a finite or
infinite horizon. Finite horizon problems include the zero final state con-
straint. Sufficient conditions are provided that ensure asymptotic stability
of the nonlinear system using an MPC control law. These include a modified
controllability condition on the nonlinear system. Keerthi and Gilbert also
provided sufficient conditions for convergence of finite horizon results to the
infinite horizon case. Meadows and coworkers [49] present similar results for
the time-invariant case and show that any feedback linearizable system can
be stabilized using MPC. In the same work, they also show that there exist
discrete-time systems that require discontinuous feedback laws for stability
and that nonlinear MPC can generate such control laws.
The work of Keerthi and Gilbert and Meadows et al. provides sufficient
conditions for stability of nonlinear MPC. The key difficulty in implementa-
tion arises from the final state constraint; it is difficult or even impossible to
find the set of states for which the final state constraint is feasible a priori.
As we argued previously, we can shift the burden of verifying feasibility to
the optimization routines, but even so, the range of feasible states may be
unacceptably small.
With these limitations in mind, Mayne and Michalska [47, 52] offer an
approach that retains the stability property of MPC using the final state
constraint, while expanding the set of initial feasible states. Rather than
adopt the strict final state equality constraint XN = 0, Mayne and Michal-
ska propose using a "dual-mode" controller whose final state is constrained
to lie within a region of the state space W C R that contains the ori-
gin. Outside the region, a nonlinear MPC controller is used that includes
the final state constraint XN E W. Within W, a stabilizing controller is
used whose design is based on the system's linearization at the origin. The
region W is defined as the largest elliptical region in the state space in
which the linearized controller produces a decreasing (and therefore asymp-
totically stable) Lyapunov function. It can be determined off-line through
the solution of a single initial nonlinear programming problem. Mayne and
5.9. FUTURE DIRECTIONS 297

Michalska presented detailed analyses of stability and robustness, with spe-

cial emphasis on providing an implementable algorithm. A disadvantage of
the Mayne-Michalska algorithm is that it requires a system model that has a
stabilizable linearization at the origin, but this condition is usually satisfied
in systems of engineering interest.

5.9 Future Directions

Nonlinear MPC offers a number of interesting challenges. Among these we
can list a few:

• Robustness : We have seen that MPC used with the deadbeat distur-
bance estimator is robust to model structural and parameter uncer-
tainty. To date, most results in this area indicate good performance in
the presence of model uncertainty; however, a rigorous analysis that
can provide sufficient conditions for robustness has not been completed.

• Estimator/Controller Interactions : Unlike the linear systems result,

for general nonlinear systems there is no separation principle. Con-
troller performance is intimately linked to estimator design. In the
CSTR example, without state feedback it was difficult to achieve the
desired steady state even though the setpoint was satisfied. Meadows
et al. [50] and Scokaert and Rawlings [68] showed that with constrained
linear systems, if the closed-loop dynamic system satisfies a Lipschitz
condition, a weak form of separability can be demonstrated. More
work is needed for the general nonlinear case.

• Nominal Stability : There continues to exist a gap between observed

results and theory in nonlinear MPC. Experimental results, in terms
of both simulations and applications, continue to indicate good perfor-
mance of MPC on a wide class of problems, but tuning the algorithm
to obtain stability is usually necessary. A theory to guide the tuning
process is not yet available. Current theoretical results require infi-
nite horizons or the final state constraint XN = 0 to assure stability,
conditions that can be much too strong (the final state constraint) or
unachievable in practice (infinite horizon).
As an alternative to these two approaches, it is possible to guarantee
stability by using a final state penalty function rather than a con-
straint. Consider the case in which an infinite horizon is used. If this
298 CHAPTER 5. MODEL PREDICTIVE CONTROL

problem could be solved, it would provide a stabilizing controller. If we

had an analytical expression for the optimal infinite horizon objective
function J (x), we could devise an equivalent finite horizon function
that used J^ (x) as a final state penalty function:

N-1
T* u
( x k) = min (Hxk+jlkII Q + Il k+jlkll R) + J^(xN)
Uk+ j l k
j=1

Provided that the stage cost used to compute J is the same as in IV*,
the principle of optimality guarantees that Q* (x) = J (x); therefore,
MPC using the objective function J^(x k ) is stabilizing.
It is probably difficult or impossible to provide a closed-form expression
for Jam; however, if we are able to compute a bound on J^ (x) < J* (x)
such that J* (0) = 0, then the MPC controller computed using J* is also
stabilizing, subject to some technical conditions. Some preliminary
results in this area have already been presented [48] and an article
containing more details concerning this approach is in preparation.

Acknowledgments
The authors wish to thank Dr. Pierre 0. M. Scokaert for helpful discussions.
The financial support of NSF, through grant CTS-9311420, and the indus-
trial members of the Texas—Wisconsin Modeling and Control Consortium
(TWMCC) is gratefully acknowledged. The first author was financially sup-
ported during the writing of this chapter from a fellowship from the Brazil-
ian government, acting through the Conselho Nacional de Desenvolvimento
Cientifico e Tecnologico (National Council of Scientific and Technological
Development). This support is gratefully acknowledged.

5.10 Notation
A matrix of first derivative collocation weights without initial
condition
A matrix of first derivative collocation weight in orthogonal
collocation approximation with multiple finite elements
Ao matrix of first derivative collocation weight, modified to in-
clude the initial condition equation
5.10. NOTATION 299

C matrix used to form linear inequality constraints on Du

Ck vector of quadrature weights for integration in orthogonal
collocation approximation
ck vector of quadrature weights for integration in orthogonal
collocation approximation
C
lb, c
ub upper and lower bounding vectors for linear inequality con-
straints on control increments
D, H matrices used to express linear constraints on state and con-
trol variables
d, h right-hand sides of linear inequality constraints on state and
control variables
dk disturbance term used in conventional MPC feedback
dk disturbance estimate obtained from state observer
F right-hand side of matrix equation representing a differen-
tial equation discretized by orthogonal collocation approxi-
mation with a single finite element
F right-hand side of matrix equation representing a differen-
tial equation discretized by orthogonal collocation approxi-
mation with multiple finite elements
FF,k , Fu k Jacobian matrix of f (x, u) with respect to x or u, evaluated
at ( x k, uk)
f state transition function in dynamic system equation
Gk Jacobian matrix of g(x) evaluated at (x k , uk)
g function relating state vector to output vector
hx , hd correction terms for state update in state observer
J MPC objective function
J. infinite horizon optimal MPC objective function
J* optimal MPC objective function
J. upper bound on infinity horizon optimal MPC objective
function
j indexing variable for time within prediction horizon for dis-
crete system
300 CHAPTER 5. MODEL PREDICTIVE CONTROL

K constant for proportional control in fluidized-bed reactor

model
1C set of functions used in Lyapunov stability theorem
k independent variable representing time in discrete system
k 1 , k2 reaction rate constants in fluidized-bed reactor model
k 1 , k2 , k3 reaction rate constants in isothermal CSTR model
L stage cost, penalty incurred at each time in prediction hori-
zon as function of outputs, controls or control increments
LeA, Le B Lewis number in fluidized-bed reactor model
M MPC control horizon, when different from N
m dimension of control vector
N MPC prediction horizon
n dimension of state vector
Q, R, S weighting matrices for vector norms used in MPC objective
function
q, q -1 shift operators in discrete-time systems
S set or null-controllable initial states
T temperature in fluidized-bed reactor model
T prediction horizon in continuous-time MPG
TF feed temperature in fluidized-bed reactor model
T
set setpoint temperature in fluidized-bed reactor model
AT sampling interval
t independent variable time in continuous-time systems
tf final time in prediction horizon for continuous-time MPC
problems
to initial time in prediction horizon for continuous-time MPC
problems
uk control vector at time k
ut input target vector obtained from target optimization
u* solution of nonlinear program for feedback through steady-
state target optimization
5.10. NOTATION 301

Duk control increments, i.e., Uk — uk-1

m
U subset of R used as general control constraint
V Lyapunov function
W in the algorithm of Mayne and Michalska, the region near
the origin in which a linear control law is stabilizing
wk random input variable in discrete-time stochastic model
X matrix containing elements
X matrix of states x i, j in orthogonal collocation approximation
with multiple finite elements
X subset of R' used as general state constraint
x 1 , x2 state variables in the isothermal CSTR model
X A , XB concentrations in fluidized-bed reactor model
XA,F, XB,F concentrations of species A and B in fluidized-bed reactor
model
XF feed concentration of species A in isothermal CSTR model
in problems discretized with orthogonal collocation, the j-th
element of the state vector at discretization point i
x 2,* in problems discretized with orthogonal collocation, the en-
tire state vector at discretization point i
x i n; t initial condition for state in orthogonal collocation approxi-
mation
xk state estimate obtained from state observer
xk state vector at time k
xt state target vector obtained from target optimization
yk output vector at time k
Yref reference trajectory in MPG objective
yk model output, used when necessary to distinguish model
output from true process output

Greek Symbols

a 1filter constant in experiments with fluidized-bed reactor

302 CHAPTER 5. MODEL PREDICTIVE CONTROL

al , a2 pre-exponential factors in fluidized-bed reactor model

a, /3, (y class IC function used in Lyapunov Stability Theorem
control weight in MPC objective function
,Q residence time in fluidized-bed reactor model
,13 1 ,132 heats of reaction in fluidized-bed reactor model
'yl, ry2, y3 weighting constants in MPC objective function of Equation 5.7
- y l, 'Y2 activation energies in fluidized-bed reactor model
µ feedback control law resulting from MPC
0 final state penalty function
4* finite horizon objective function equivalent to an infinite hori-
zon objective
U smallest singular value
min

T independent variable time in MPC objective function of Equa-

tion 5.7

Subscripts

A, B refers to species A and B in fluidized-bed reactor model

F value in feed stream in example models
k value at time k
jk value at time j, given information up to and including time
k
i, j in combination, refers to j -th element of vector at collocation
point i
i, * in combination, refers to entire vector at collocation point i
init initial condition in orthogonal collocation approximation
lb, ub upper and lower bounding vector for linear constraints
max, min used to indicate upper and lower limits in linear inequality
constraints
min used in conjunction with a min to indicate the smallest singu-
lar value
ref reference values, often indicates setpoint
REFERENCES

set setpoint in fluidized-bed reactor model

x, d in state observer, refers to terms which account for the state
and disturbance estimate, respectively
+ only used in conjunction with J2+ to indicate the subset of
real numbers [0, cc)
0o infinite horizon value

Operators and Special Symbols

11 . 11 vector norm
• Q Q-weighted vector norm
the n-th Cartesian product of the real numbers
R+ the semi-infinite subset of the real numbers [0, oo}
0 Kronecker matrix product

References

[1] T. A. Badgwell. Robust stability conditions for a SISO model predictive

control algorithm. Submitted for publication in Automatica, 1995.

[2] R. E. Bellman. Dynamic Programming. Princeton University Press,

Princeton, New Jersey, 1957.

[3] R. E. Bellman and S. E. Dreyfus. Applied Dynamic Programming.

Princeton University Press, Princeton, New Jersey, 1962.

[4] L. T. Biegler. Solution of dynamic optimization problems by successive

quadratic programming and orthogonal collocation. Comput. Chem.
Eng., 8(3/4):243-248, 1984.

[5] R. R. Bitmead, M. Gevers, and V. Wertz. Adaptive Optimal Control,

The Thinking Man's GPC. Prentice-Hall, Englewood Cliffs, New Jersey,
1990.

[6] A. E. Bryson and Y. Ho. Applied Optimal Control. Hemisphere Pub-

lishing, New York, 1975.
304 CHAPTER 5. MODEL PREDICTIVE CONTROL

[7] M. Caracotsios. Model Parametric Sensitivity Analysis and Nonlinear

Parameter Estimation. Theory and Applications. PhD thesis, University
of Wisconsin, Madison, 1986.

[8] C. C. Chen and L. Shaw. On receding horizon control. Automatica,

16(3):349-352, 1982.

[9] D. W. Clarke and C. Mohtadi. Properties of generalized predictive

control. Automatica, 25(6):859-875, 1989.

[10] D. W. Clarke, C. Mohtadi, and P. S. Tuffs. Generalized predictive

control—Part I. The basic algorithm. Automatica, 23(2):137-148, 1987.

[11] D. W. Clarke, C. Mohtadi, and P. S. Tuffs. Generalized predic-

tive control Part II. Extensions and interpretations. Automatica,
23(2):149-160, 1987.

[12] J. E. Cuthrell and L. T. Biegler. On the optimization of differential-

algebraic systems. AIChE J., 33(8):1257-1270, 1987.

[13] C. R. Cutler and B. L. Ramaker. Dynamic matrix control –a com-

puter control algorithm. In Proceedings of the Joint Automatic Control
Conference, 1980.

[14] P. J. Davis. Interpolation and Approximation. Dover Books on Ad-

vanced Mathematics. Dover Publications, Inc., New York, 1975.

[15] P. J. Davis and P. Rabinowitz. Methods of Numerical Integration. Aca-

demic Press, Orlando, Florida, 2nd edition, 1984.

[16] S. E. Dreyfus. Dynamic Programming and the Calculus of Variations.

Academic Press, New York, 1965.

[17] S. E. Dreyfus and A. M. Law. The Art and Theory of Dynamic Pro-
gramming. Academic Press, New York, 1977.

[18] J. W. Eaton. Octave—a high-level interactive language for numerical

computations. Technical Report CCSR-93-003, The University of Texas
Center for Control and Systems Research, Austin, Texas 78712, 1993.

[19] J. W. Eaton and J. B. Rawlings. Model-predictive control of chemical

processes. Chem. Eng. Sci., 47(4):705-720, 1992.
REFERENCES 305

[20] J. W. Eaton, J. B. Rawlings, and L. H. Ungar. Stability of neural net

based model predictive control. In Proceedings of the 1994 American
Control Conference, pages 2481-2485, 1994.

[21] S. S. E. H. Elnashaie. Multiplicity of the steady states in fluidized-bed

reactors. III. yield of the consecutive reaction, A-*B-*C. Chem. Eng.
Sci., 32:295-301, 1977.

[22] S. S. E. H. Elnashaie, M. E. Abashar, and F. A. Teymour. Chaotic be-

haviour of fluidized-bed catalytic reactors with consecutive exothermic
chemical reactions. Chem. Eng. Sci., 50(1):49-67, 1995.

[23] C. E. Garcia and M. Morari. Internal model control. 1. A unifying review

and some new results. Ind. Eng. Chem. Proc. Des. Dev., 21:308-323,
1982.

[24] C. E. Garcia and A. M. Morshedi. Quadratic programming solution

of dynamic matrix control (QDMC). Chem. Eng. Commun., 46:73-87,
1986.

[25] C. E. Garcia, D. M. Prett, and M. Morari. Model predictive control:

Theory and practice—a survey. Automatica, 25(3):335-348, 1989.

[26] G. Gattu and E. Zafiriou. Nonlinear quadratic dynamic matrix control

with state estimation. Ind. Eng. Chem., 31(4):1096-1104, 1992.

[27] P. E. Gill, W. Murray, and M. H. Wright. Practical Optimization.

Academic Press, London, 1981.

[28] R. Hermann and A. J. Krener. Nonlinear controllability and observ-

ability. IEEE Trans. Auto. Cont., 22(5):728-740, 1977.

[29] A. Isidori. Nonlinear Control Systems. Springer-Verlag, Berlin, 2nd

edition, 1989.

[30] R. E. Kalman. Contributions to the theory of optimal control. Bull.

Soc. Math. Mex., 5:102-119, 1960.

[31] J. C. Kantor. Stability of state feedback transformations for nonlin-

ear systems-some practical considerations. In Proceedings of the 1986
American Control Conference, page 1014, Seattle, Washington, June
1986.
306 CHAPTER 5. MODEL PREDICTIVE CONTROL

[32] S. S. Keerthi. Optimal Feedback Control of Discrete-Time Systems with

State-Control Constraints and General Cost Functions. PhD thesis,
University of Michigan, 1986.

[33] S. S. Keerthi and E. G. Gilbert. An existence theorem for discrete-time

infinite-horizon optimal control problems. IEEE Trans. Auto. Cont.,
30(9):907-909, September 1985.

[34] S. S. Keerthi and E. G. Gilbert. Moving-horizon approximations for a

general class of optimal nonlinear infinite-horizon discrete-time systems.
In Proceedings of the 20th Annual Conference on Information Sciences
and Systems, pages 301-306, Princeton, NJ, 1986. Princeton University.

[35] S. S. Keerthi and E. G. Gilbert. Optimal infinite-horizon feedback laws

for a general class of constrained discrete-time systems: Stability and
moving-horizon approximations. J. Optim. Theory Appl., 57(2):265-
293, May 1988.

[36] D. L. Kleinman. An easy way to stabilize a linear constant system.

IEEE Trans. Auto. Cont., 15(12):692, December 1970.

[37] D. L. Kleinman. Stabilizing a discrete, constant, linear system with

application to iterative methods for solving the riccati equation. IEEE
Trans. Auto. Cont., 19:252-254, June 1974.

[38] C. Kravaris and P. Daoutidis. Nonlinear state feedback control of

second-order nonminimum-phase nonlinear systems. Comput. Chem.
Eng., 14(4/5):439-449, 1990.

[39] C.-C. Ku and K. Y. Lee. Diagonal recurrent neural networks for dy-
namic systems control. IEEE Trans. Neural Net., 6(1):144-156, January
1995.

[40] J. G. Kuschewski, S. Hui, and S. H. Zak. Application of feedforward

neural networks to dynamical system identification and control. IEEE
Trans. Cont. Sys. Tech., 1(1):37-49, March 1993.

[41] W. H. Kwon, A. M. Bruckstein, and T. Kailath. Stabilizing state-

feedback design via the moving horizon method. Int. J. Control,
37(3):631-643, 1983.
REFERENCES 307

[42] W. H. Kwon and A. E. Pearson. A modified quadratic cost problem

and feedback stabilization of a linear system. IEEE Trans. Auto. Cont.,
22(5):838-842, October 1977.

[43] W. H. Kwon and A. E. Pearson. On feedback stabilization of time-

varying discrete linear systems. IEEE Trans. Auto. Cont., 23(3):479-
481, June 1978.

[44] J. H. Lee, M. Moran, and C. E. Garcia. State-space interpretation of

model predictive control. Autornatica, 30(4):707-717, 1994.

[45] A. U. Levin and K. S. Narendra. Control of nonlinear dynamical systems

using neural networks: controllability and stabilization. IEEE Trans.
Neural Net., 4(2):192-206, March 1993.

[46] S. Li, K. Y. Lim, and D. G. Fisher. A state space formulation of model

predictive control. AIChE J., 35(2):241-249, February 1989.

[47] D. Q. Mayne and H. Michalska. Receding horizon control of nonlinear

systems. IEEE Trans. Auto. Cont., 35(7):814-824, July 1990.

[48] E. S. Meadows. Stability and Continuity of Nonlinear Model Predictive

Control. PhD thesis, The University of Texas at Austin, 1994.

[49] E. S. Meadows, M. A. Henson, J. W. Eaton, and J. B. Rawlings. Re-

ceding horizon control and discontinuous state feedback stabilization.
To appear in International Journal of Control.

[50] E. S. Meadows, K. R. Muske, and J. B. Rawlings. Constrained state

estimation and discontinuous feedback in model predictive control. In
Proceedings of the 1993 European Control Conference, pages 2308-2312.
European Automatic Control Council, June 1993.

[51] E. S. Meadows and J. B. Rawlings. Topics in model predictive control.

In R. Berber, editor, NATO Advanced Study Institute on Methods of
Model Based Control Proceedings, Antalya, Turkey, July 1994. Kluwer
Academic Publishers.

[52] H. Michalska and D. Q. Mayne. Robust receding horizon control of

constrained nonlinear systems. IEEE Trans. Auto. Cont., 38(11):1623-
1633, 1993.
308 CHAPTER 5. MODEL PREDICTIVE CONTROL

[53] M. Morani and E. Zafiriou. Robust Process Control. Prentice-Hall,

Englewood Cliffs, New Jersey, 1989.

[54] K. R. Muske and J. B. Rawlings. Linear model predictive control of

unstable processes. J. Proc. Cont., 3(2):85-96, 1993.

[55] K. R. Muske and J. B. Rawlings. Model predictive control with linear

models. AIChE J., 39(2):262-287, 1993.

[56] K. R. Muske and J. B. Rawlings. Nonlinear moving horizon state es-

timation. In R. Berber, editor, NATO Advanced Study Institute on
Methods of Model Based Control Proceedings, Antalya, Turkey, July
1994. Kluwer Academic Publishers.

[57] K. R. Muske, J. B. Rawlings, and J. H. Lee. Receding horizon recursive

state estimation. In Proceedings of the 1993 American Control Confer-
ence, pages 900-904, June 1993.

[58] H. Nijmeijer and A. J. van der Schaft. Nonlinear Dynamical Control

Systems. Springer-Verlag, New York, 1990.

[59] R. E. Nordgrenn and P. H. Merckl. An analytical comparison of a

neural network and a model-based adaptive controller. IEEE Trans.
Neural Net., 4(4):685-694, July 1993.

[60] A. A. Patwardhan, J. B. Rawlings, and T. F. Edgar. Nonlinear model

predictive control. Chem. Eng. Commun., 87:123-141, 1990.

[61] S.-Z. Qin, H.-T. Su, and T. J. McAvoy. Comparison of four neural
net learning methods for dynamic system identification. IEEE Trans.
Neural Net., 3(1):122-130, January 1992.

[62] J. B. Rawlings and K. R. Muske. Stability of constrained receding

horizon control. IEEE Trans. Auto. Cont., 38(10):1512-1516, October
1993.

[63] J. Richalet, A. Rault, J. L. Testud, and J. Papon. Model predictive

heuristic control: Applications to industrial processes. Automatica,
14:413-428, 1978.

[64] N. L. Ricker. Use of quadratic programming for constrained internal

model control. Ind. Eng. Chem. Proc. Des. Dev., 24:925-936, 1985.
REFERENCES 309

[65] D. G. Robertson, J. H. Lee, and J. B. Rawlings. A moving horizon-based

approach for least squares state estimation. Submitted for publication
in AIChE J., 1994.

[66] R. W. H. Sargent and G. R. Sullivan. The development of an efficient

optimal control package. In J. Stoer, editor, Lecture Notes in Con-
trol and Information Sciences 7, pages 158-168, Berlin, 1978. 8th IFIP
Conference on Optimization Techniques, Springer-Verlag.

[67] R. Scattolini and S. Bittanti. On the choice of the horizon in long-range

predictive control—some simple criteria. Autornatica, 26(5):915-917,
1990.

[68] P. O. Scokaert and J. B. Rawlings. Stability of model predictive control

under perturbations. In IFAC Symposium on Nonlinear Control System
Design, Tahoe City, California, 1995.

[69] P. B. Sistu and B. W. Bequette. Model predictive control of processes

with input multiplicities. Chem. Eng. Sci., 6(6):921-936, 1995.

[70] E. D. Sontag. An algebraic approach to bounded controllability of non-

linear systems. Int. J. Control, 39:181-188, 1984.

[71] H. J. Sussmann and V. Jurdjevic. Controllability of nonlinear systems.

J. Duff. Eq., 12:95-115, 1972.

[72] Y. Thomas and A. Barraud. Commande optimale a horizon fuyant.

Revue RAIRO, J(1):126-140, April 1974.

[73] Y. Thomas, D. Sarlat, and L. Shaw. Receding-horizon approach to

the synthesis of nonlinear multivariable regulators. Electron. Lett.,
13(11):329-331, May 1977.

[74] Y. A. Thomas. Linear quadratic optimal estimation and control with

receding horizon. Electron. Lett., 11:19-21, January 1975.

[75] J. G. Van de Vusse. Plug-flow type reactor versus tank reactor. Chem.
Eng. Sci., 19:964, 1964.

[76] M. Vidyasagar. Nonlinear Systems Analysis. Prentice-Hall, Inc., En-

glewood Cliffs, New Jersey, 2nd edition, 1993.
310 CHAPTER 5. MODEL PREDICTIVE CONTROL

[77] J. Villadsen and M. L. Michelsen. Solution of Differential Equation

Models by Polynomial Approximation. Prentice-Hall, Englewood Cliffs
New Jersey, 1978.

[78] J. V. Villadsen and W. E. Stewart. Solution of boundary-value problems

by orthogonal collocation. Chem. Eng. Sci., 22:1483-1501, 1967.

[79] E. Zafiriou and M. Moran. Design of robust digital controllers and

sampling-time selection for SISO systems. Int. J. Control, 44(3):711-
735, 1986.