Professional Documents
Culture Documents
00
Printed in Great Britain. Pergamon Press pie
(~ 1989 International Federation of Automatic Control
Survey Paper
Al~tmd--We refer to Model Predictive Control (MPC) as by the integration of all aspects of automation of the
that family of controllers in which there is a direct use of an decision making process (Garcia and Prett, 1986;
explicit and separately identifiable model. Control design
methods based on the MPC concept have found wide Prett and Garcia, 1988).
acceptance in industrial applications and have been studied
• M e a s u r e m e n t s . T h e gathering and monitoring of
by academia. The reason for such popularity is the ability of
MPC designs to yield high performance control systems process measurements via instrumentation.
capable of operating without expert intervention for long • C o n t r o l . T h e manipulation of process degrees of
periods of time. In this paper the issues of importance that freedom for the satisfaction of operating criteria. This
any control system should address are stated. MPC typically involves two layers of implementation: the
techniques are then reviewed in the light of these issues in
order to point out their advantages in design and single loop control which is performed via analog
implementation. A number of design techniques emanating controllers or rapid sampling digital controllers; and
from MPC, namely Dynamic Matrix Control, Model the control performed using real-time computers with
Algorithmic Control, Inferential Control and Internal Model relatively large CPU capabilities.
Control, are put in perspective with respect to each other
and the relation to more traditional methods like Linear • Optimization. The manipulation of process
Quadratic Control is examined. The flexible constraint degrees of freedom for the satisfaction of plant
handling capabilities of MPC are shown to be a significant economic objectives. It is usually implemented at a
advantage in the context of the overall operating objectives rate such that the controlled plant is assumed to be at
of the process industries and the 1-, 2-, and oo-norm steady state. Therefore, the distinction between
formulations of the performance objective are discussed. The
application of MPC to non-linear systems is examined and it control and optimization is primarily a difference in
is shown that its main attractions carry over. Finally, it is implementation frequencies.
explained that though MPC is not inherently more or less • Logistics. T h e allocation of raw materials and
robust than classical feedback, it can be adjusted more easily scheduling of operating plants for the maximization of
for robustness.
profits and the realization of the company's program.
335
336 C.E. GARdA et al.
the process outside the physical limitations of the transparent specification of performance criteria such
equipment. as constraints. Our experience has demonstrated that
• Product quality. Consumer specifications on Model Predictive Control (MPC) techniques provide
products must be satisfied. the only methodology to handle constraints in a
• Human preference. There exist excessive levels of systematic way during the design and implementation
variable oscillations or jaggedness that the operator of the controller. Moreover, in its most general form
will not tolerate. There can also be preferred modes MPC is not restricted in terms of the model, objective
of operation. function and/or constraint functionality. For these
reasons, it is the only methodology that currently can
In addition, the implementation of such integrated reflect most directly the many performance criteria of
systems is forcing our processes to operate over an relevance to the process industries and is capable of
ever wider range of conditions. As a result, we can utilizing any available process model. This is the
state the control problem that any control system must primary reason for the success of these techniques in
solve as follows. numerous applications in the chemical process
On-line update the manipulated variables to satisfy industries.
multiple, changing performance criteria in the face of In this paper the MPC methodology is reviewed and
changing plant characteristics. compared with other seemingly identical techniques.
The whole spectrum of process control method- We particularly emphasize the unconstrained version
ologies in use today is faced with the solution of this of MPC since it is only in this form that it is possible
problem. The difference between these methodologies to compare it with other schemes. Then the several
lies in the particular assumptions and compromises existing forms of constrained MPC are reviewed,
made in the mathematical formulation of performance concluding with the non-linear MPC approaches.
criteria and in the selection of a process represen- Although the issue of model uncertainties in MPC
tation. These are made primarily to simplify the techniques is not dealt with in this paper, some
mathematical problem so that its solution fits the comments on robustness of MPC are included.
existing hardware capabilities. The natural mathe-
matical representation of many of these criteria is in HISTORICAL BACKGROUND
the form of dynamic objective functions to be The current interest of the processing industry in
minimized and of dynamic inequality constraints. The MPC can be traced back to a set of papers which
usual mathematical representation for the process is a appeared in the late 1970s. In 1978 Richalet et al.
dynamic model with its associated uncertainties. The described successful applications of "Model Predictive
importance of uncertainties is increasingly being Heuristic Control" and in 1979 engineers from Shell
recognized by control theoreticians and thus are being (Cutler and Ramaker, 1979; Prett and Gillette, 1979)
included explicitly in the formulation of controllers. outlined "Dynamic Matrix Control" (DMC) and
However, one of the most crucial compromises made reported applications to a fluid catalytic cracker. In both
in control is to ignore constraints in the formulation of algorithms an explicit dynamic model of the plant is
the problem. As we explain below these simplifi- used to predict the effect of future actions of the
cations can deny the control system of its achievable manipulated variables on the output (thus the name
performance. "Model Predictive Control"). The future moves of the
It is a fact that in practice the operating point of a manipulated variables are determined by optimization
plant that satisfies the overall economic goals of the with the objective of minimizing the predicted error
process will lie at the intersection of constraints subject to operating constraints. The optimization is
(Arkun, 1978; Prett and Gillette, 1979). Therefore, in repeated at each sampling time based on updated
order to be successful, any control system must information (measurements) from the plant.
anticipate constraint violations and correct for them in Thus, in the context of MPC the control problem
a systematic way: violations must not be allowed while including the relative importance of the different
keeping the operation close to these constraints. The objectives, the constraints, etc. is formulated as a
usual practice in process control is to ignore the dynamic optimization problem. While this by itself is
constraint issue at the design stage and then "handle" hardly a new idea, it constitutes one of the first
it in an ad hoc way during the implementation. Since examples of large-scale dynamic optimization applied
each petro-chemical process (or unit) is unique we routinely in real time in the process industries.
cannot exploit the population factor as in other The MPC concept has a long history. The
industries (e.g. aerospace). That is, we cannot afford connections between the closely related minimum
extreme expenses in designing an ad hoc control time optimal control problem and Linear Program-
system that we know will not work in another process ming were recognized first by Zadeh and Whalen
and therefore its cost cannot be spread over a large (1962). Propoi (1963) proposed the moving horizon
number of applications. Due to the increase in the approach which is at the core of all MPC algorithms.
number of applications of this type (resulting from the It became known as "Open Loop Optimal Feedback".
need to achieve integration), this implies an enormous The extensive work on this problem during the 1970s
burden both in the design and maintenance costs of was reviewed in the thesis by Gutman (1982). The
these loops. In our experience, these costs more than connection between this work and MPC was
offset the profitability of any ad hoc control system. discovered by Chang and Seborg (1983).
In conclusion, economics demand that the control Since the rediscovery of MPC in 1978 and 1979, its
systems must be designed with no ad hoc fixups and popularity in the Chemical Process Industries has
Model Predictive Control: theory and practice---a survey 337
operations.
In academia MPC has been applied under a_ ~ lCl,z_, (6)
controlled conditions to a simple mixing tank and a
heat exchanger (Arkun et al., 1986) as well as a where /~, are the impulse response coefficients, the
coupled distillation column system for the separation magnitudes of which vanish as i---, ~. Thus, in the time
of a ternary mixture (Levien, 1985; Levien and domain we have the truncated impulse response model
Morari, 1987). Parrish and Brosilow (1985) compared
MPC with conventional control schemes on a heat y(k ) = ~ ISl,u(k - i) (7)
exchanger and an industrial autoclave. i~l
Most of the applications reported above are and with the definitions
multivariable and involve constraints. It is exactly
these types of problems which motivated the = t-t, - H , _ 1 (8)
development of the MPC control techniques. Largely
independently a second branch of MPC emerged, the H, = ~ ~ (9)
main objective of which is adaptive control. Peterka's i-1
predictive controller (Peterka, 1984), Ydstie's the truncated step response model
extended-horizon design (Ydstie, 1984) and EPSAC
n-I
developed by DeKeyser and van Cauwenberghe
y(k) = ~. H, Au(k - i) + H,u(k - n) (10)
(1982), DeKeyser et al. (1985) are in this category as
well as Clarke's generalized predictive control where
algorithm (Clarke et al., 1987a, b). These develop- Au(k) = u(k) - u(k - 1) (11)
ments are essentially limited to SISO systems with
extension to the MIMO case conceptually straightfor- and H~ are the step response coefficients. Depending
ward but very involved when the details are on the time delay structure of the system the leading
considered. The constrained case is not considered in step response coefficient matrices may be zero or have
zero elements.
any detail in these papers. Because of the different
underlying philosophy these algorithms are outside
the focus of this paper. Nevertheless some cross MPC ALGORITHM FORMULATIONS
references will be useful at times because these The name "Model Predictive Control" arises from
algorithms were largely developed for the non- the manner in which the control law is computed (Fig.
adaptive case with the adaptation added in an ad hoc 1). At the present time k the behavior of the process
manner based on recursive least squares (or similar) over a horizon p is considered, Using a model the
parameter estimates. The stability and robustness of process response to changes in the manipulated
the adaptive scheme was generally not analyzed. variable is predicted. The moves of the manipulated
338 C.E. GARCiA et al.
past future
¢1
target
.................................................. 0 ........ 0 .q ...... 0 0 ........ 0
^ o
y (k + Ilk) o
y(k) o
o o
o
| u(k+l)
i
II,
-1 kh kh k÷p
horizon
variables are selected such that the predicted response The prediction of output (13) involves three terms on
has certain desirable characteristics. Only the first the right-hand side. The first term includes the present
computed change in the manipulated variable is and all future moves of the manipulated variables
implemented. At time k + 1 the computation is which are to be determined so as to solve (12). The
repeated with the horizon moved by one time interval. second term includes only past values of the
We will demonstrate how DMC and MAC are manipulated variables and is completely known at
derived. All other MPC algorithms which have been time k. The third term is the predicted disturbance d
proposed are very similar. which is obtained from (7). ~(k + I l k ) is assumed
constant for all future times (1-> 0). At time k it is
Dynamic Matrix Control (D M C ) estimated as the difference between the measured
The manipulated variables are selected to minimize output y(k) and the output predicted from the model.
a quadratic objective In block diagram notation (14) corresponds to a
P
model P in parallel with the plant P (Fig. 2(a)) with
min ~ IIf(k + 11 k) - r(k +/)ll 2, the resulting feedback signal equal to d(k I k).
Au(k)...lXu(k+rn- l) I=l Equations (12)-(15) define a Quadratic Program
+ IIAu(k + 1 - 1)112, (12) which is solved on-line at every time step. This
I "controller" is represented by block Q in Fig. 2(a).
~(k + l I k) = ~ HiAu(k + l - i) Though computationally more involved than
iffil
standard linear time invariant algorithms, the flexible
n--1
constraint handling capabilities of MPC are very
+ ~ HiAu(k+l-i)
i=l+l attractive for practical applications: a stuck valve can
be simply specified by the operator on the console as
+H,u(k + l - n ) + c l ( k + l l k ) (13)
an additional constraint for the optimization program.
a(k +ll k ) = a(klk)=y , ( k ) The algorithm will automatically adjust the actions of
n-I
all the other manipulated variables to compensate for
- ~ H, A u ( k - i ) + H , u ( k +l-n) (14) this failure situation as well as possible. In an
unexpected emergency which a traditional fixed-logic
p scheme might find difficult to cope with, MPC will
C'~(k + I l k ) + CL,u(k + l - 1) + c ' ---0; keep the process operating safely away from all
I~l
constraints or allow the operator to shut it down in a
/ ' = 1 nc (15)
smooth manner.
~(k + l I k) = predicted value of y at time k + l based
on information available at time k
a~(k + l I k) = predicted value of additive disturbances
at process output at time k + l based on Model Algorithmic Control (MAC)
information available at time k MAC is distinctive from DMC in three aspects.
ym(k) = measurement of y at time k
Au(k + l) = u(k + l) - u(k + l - 1) (1) Instead of the step response model involving
H,, i= 1, n = model step response matrix coefficient Au, an impulse response model involving u is
n = truncation order employed. If the input u is penalized in the quadratic
nc = number of constraints objective, then the controller does not remove offset.
p = horizon length (in general p >> n) This can be corrected by a static offset compensator
m = number of manipulated variable moves (Garda and Morari, 1982). If the input u is not
in the future ( A u ( k + l ) = O Vl>-m; penalized then extremely awkward procedures are
re<p) necessary to treat non-minimum phase systems
Ilx lib = x r Qx (Mehra and Rouhani, 1980).
F , BI = weighting matrices (2) The number of input moves m is not used for
C~,, C~,, c j = constant matrices. tuning (m = p).
Model Predictive Control: theory and practice---a survey 339
d
~'Y r ~ y
(b)
(a)
d
]] f-----~----~ ~
l~ -~..L.-J-Vl - d
• ~ :: :mr :, , ~; ® e
(d)
(0
Flo. 2. (a) IMC structure. (b) Equivalent classic feedback structure. (c) Two-degrees-of-freedom IMC
structure. (d) Equivalent two-degrees-of-freedomclassic feedback structure.
(3) The disturbance estimate (14) is filtered. Let assumption a linear time invariant controller is found.
y(k) be the measurement and )~(k) the model Garcia and Morari (1985b) have shown how to obtain
prediction. Then the disturbance estimate is defined the controller transfer function from the linear least
recursively as squares solution.
t The relation between x and y is called affine when FIG. 3. Alternate representation of Smith Predictor
yfA+Bx. controller.
M o d e l Predictive Control: t h e o r y and practice---a survey 341
feedback control law. Because of the finite horizon in f~ll0ws a P0isson distribution (p. 56 of Chang (1961)).
(12) most conditions which guarantee that (12)-(14) The deterministic equivalent of (24) is d = constant, as
will lead to a stabilizing controller are only sufficient. was assumed for MPC. It is only meaningful to define
Thus, at this time the tuning of MPC has to proceed an optimal control problem for the system (23)-(25) if
largely by trial and error with these sufficient it is stabilizable and detectable.
conditions as guidelines.
For simplicity, the theorems below (Garcia, 1982;
Garcia and Morari, 1982) are formulated for SISO
{[a 0i] [01/
systems without delays. Equivalent results for M I M O is clearly not stabilizable. By differencing (23)-(25)
systems have been derived (Garcia and Morari, become (Prett and Garcia, 1988)
1985b) but are somewhat more complicated because
specific non-singularity conditions have to be imposed. + 1)]
y(k + I) J
Theorem 1. For Ft ~ 0, B, = 0, selecting m = p -< n
yields the model inverse control Q(z) = z - ' P ( z ) - L =[A A °I[~(k)I [w,(k)I
Lw2(k)J
This output deadbeat control law is only stable if y(k + 1) = y(k + 1) + w3(k + 1) (27)
all zeros of P(z) are inside the unit circle. Even when
where w(k)=[w,(k), w2(k), w3(k)] T is again a se-
it is stable it is generally very aggressive and can lead
quence of zero mean vector stochastic variables the
to intersample rippling if P(z) has zeros close to
( - 1 , 0). This control law can be acceptable if the correlation functions of which can be derived from
V. )~ is an additional state. If we assume that if,, = 0
sampling time is relatively large.
and 6,3 = 0 then w(k) is again uncorrelated and the
TheOrem 2. There exists a finite B * > 0 such that for
B, >-B*(l = 1. . . . . m) the control law is stable for all
m ~ l , p - > l and F t > 0 .
[0
variance matrix of ff,(k) and w(k) becomes
0 Vo . (28)
0 0
This theorem implies that penalizing control action
can stabilize the system regardless of the other We will also assume that the covariance matrix of the
parameter choices. state estimate (Ax, ~)r at the initial point is
Au and Ay to be nonsingular which is clearly a very and R3 constant and extending the horizon to infinity.
reasonable requirement. In summary, conditions 1 Unconstrained MPC is time invariant because only the
and 2 are satisfied in any practical situation. first control move obtained from the finite horizon
Detectability. System (26) problem is applied at each time step.
(4) The time invariant LQC is tuned by adjusting
t[ Az I° Ill the constant weighting matrices R2 and R3. MPC can
be tuned by adjusting R2 and R 3 but a more frequently
is detectable if and only if {A, CA} is detectable. used tuning parameter is the number of moves m. The
(This can be proved using the arguments by Morari horizon p is usually selected such that p -> n. There
and Stephanopoulos (1980).) Because ,4 is stable, appears to be consensus in the literature that it is
{,4, CA } is detectable. easier to pick a reasonable m a priori than R2 and R3.
Optimal control problem. Rewrite problem (12) for (5) If the control problem is really of a stochastic
system (26)-(29) in a form similar to the one used on nature, the implicit assumptions in MPC are quite
p. 539 of Kwakernaak and Sivan (1972) restrictive (no measurement and state excitation
min
{"
E ~ [y(k + l)
noise) and might lead to performance which is inferior
to what could be obtained by a full L Q G approach.
Au(k)...Au(k+p-l) I=l
If--as is usually the case---the covariance matrices are
- r(k + l)]TR3(l)[y(k + l) - r(k + l)] de facto tuning parameters then the MPC perform-
ance should be no better or worse than LQG.
+ Au(k + 1 - 1)TR2(I)Au(k + 1 - 1)} (32) (6) It is well known that the L Q G controller leads
to a closed-loop stable system as the horizon is
where E { . } is the expected value operator, R2 > 0, extended to infinity. A similar result can be proven for
R3 > 0 and R2(l) = oo for l > m + 1. unconstrained MPC (Garcia and Morari, 1982,
The unique optimal solution is the feedback law 1985b).
(7) Both controllers reject sustained step-like
Au(k) = F~(k)A2(k) + FE(k)[)~(k)- r(k)] (33)
disturbances, i.e.include implicitlyintegral action.
where the formula for F(k) can be found on p. 494 of
Kwakeruaak and Sivan (1972) and A2(k) and )~(k) are Internal Model Control ( IMC)
the state estimates obtained from the optimal observer The motivation behind the development of IMC
(p. 530 of Kwakernaak and Sivan (1972)) was to combine the advantages of the different
unconstrained MPC schemes and to avoid their
A£(k + 1) = AA.~(k) + BAu(k) (34) disadvantages: easy on-line tuning via adjustment of
)~(k + 1) = CAA2(k) + y(k) + CBAu(k). (35) physically meaningful parameters and without any
concern about closed-loop stability; good performance
Note that because of the specific noise assumptions without intersample rippling; ability to cope with
the optimal observer is time invariant. According to inputs other than steps. IMC utilizes the controller
(34) the states Ax are estimated in open-loop fashion structure shown in Fig. 2(a) (and (c)) and uses a filter
from the process model. Equation (35) is identical for tuning in a similar manner as MAC. The design
with the corresponding part of (26) except that procedure which is described in detail by Morari and
the measurement y(k) appears instead of )~(k): the Zafiriou (1989) and Zafiriou and Morari (1986b)
estimate )~(k + 1) is the measurement y(k) plus the consists of two steps.
effect of the manipulated variables as expressed
through the model. In block diagram form the (1) First, O is designed to yield a good response
estimator (34) and (35) corresponds to a model in (usually in the least square sense) without regard for
parallel with the plant as is characteristic of MPC. constraints on the manipulated variables and
robustness.
Comparison of unconstrained MPC and LQC (2) The IMC controller Q is found by augmenting
(1) The control laws resulting from the two O with a low-pass filter F(Q = QF) the parameters of
computational procedures are essentially equal. The which are adjusted either off-line or on-line to reduce
only difference arises from the fact that the state space the action of the manipulated variable and to improve
model and the truncated step response model are not the robustness (the controller Q is "detuned").
identical (but can be made arbitrarily close to each
other). The controller Qo(z) which minimizes E~-o e~ for a
(2) The MPC computation requires the solution of a disturbance input d is derived by Zafiriou and Morari
(possibly large) linear least squares problem. LQC (1986b) (also see Morari and Zafiriou (1989))
involves the solution of some recursive matrix Qo(z) = z(PM(z)dM(z))-I{(PA(z)-1dM(z)z-'}, (36)
relations. The order of the system which determines
the dimension of the matrices involved is generally where the plant P is factored into an all pass PA and a
much less than the truncation order of the step minimum phase part PM. We have
response model. P = PA" PM (37)
(3) For implementation it is desirable to use a time
invariant control law. In the general LQC context = z (1-
time invariance is achieved by making the weights R2 /_, (1 ~ ~;"~ (38)
Model Predictive Control: theory and practice---a survey 343
IMC design procedure generates PID controllers with particular time) and temporally (over the honzon
one adjustable parameter. length). Apparently, a QP of the type (12)-(15) has
(3) The advantage of unconstrained MPC (with the been included in MAC for quite a while (Mehra et
possible exception of IMC) over other LQ techniques al., 1982) but the first detailed descriptions of solution
is still to be demonstrated with more than case procedures in the open literature were provided in the
studies. Contrary to other techniques, however, the context of DMC by Garcia and Morshedi (1984) and
MPC ideas can be extended smoothly to generate Ricker (1985).
non-linear time-varying controllers for linear systems The advantage of the 2-norm is that at least for the
with constraints. unconstrained case an explicit expression for the
control law can be found and its stability, robustness
CONSTRAINED MPC
and performance characteristics can be analyzed with
As emphasized in the introductory sections of this standard tools. A disadvantage is that for MIMO
paper, constraints are always present in any real life
systems with many future moves m and tong horizon p
process control situation. Their importance has storage requirements can be quite formidable.
increased because supervisory optimizing control
schemes frequently push the operating point toward
the intersection of constraints (Arkun, 1978; Prett and 1-Norm objective function
Gillette, 1979). Most of today's control implemen- Instead of the 2-norm the spatial and temporal
tations handle constraints through split range control- 1-norm can be used to express the objective
lers, overrides and more general min-max selectors
with some logic. These schemes are difficult to design,
debug, explain to the operating personnel and rrfin ~, ~ w,, lYi(k + I I k) - r,(k + l)l
A u ( k ) . , . A u ( k + m - - I ) I--1 i--1
maintain. For an example of a mildly complicated
scheme of this type the reader is referred to Bristol w,~ -- 0. (50)
(1980). The main attraction of MPC is that the
engineer/operator can enter the constraints in a direct With the additional inequality constraints the
manner and that the algorithm will automatically find resulting Linear Program (LP) is computationally
the best solution satisfying all of them. simpler than the QP discussed previously. The long
tradition of LP in optimal control was reviewed earlier
Structure in this paper. A 1-norm version of DMC was
Recall that the IMC structure (Figs 2(a) and (c)) is introduced by Morshedi et al. (1985).
effectively open loop when P = P and that it is A severe disadvantage of the 1-norm formulation is
internally stable if and only if P and Q are stable. This that even in the unconstrained case a simple
result holds trivially even when P and Q are nonlinear fundamental analysis of stability and performance is
or when the inputs to P are subject to saturation not possible because an explicit closed-form ex-
constraints as long_as the constrained inputs are also pression for the (nonlinear) control law does not exist.
fed to the model P so that P = t5 is preserved. While Brosilow and co-workers (Brosiiow et al., 1984:
input saturation causes complex stability problems for Brosilow and Zhao, 1986) circumvent this problem in
the classic feedback structure (Figs l(b) and (d)), it the following manner. They first design a controller
has no effect whatsoever on the stability of MPC. If for the unconstrained process by some standard
input saturation is unlikely during normal operation procedure. Then they minimize the 1-norm of the
and one is mainly concerned about stability under error not between the process output and setpoint as
emergency conditions, the controller should simply suggested by (12) but between the predicted
be designed for the unconstrained system as discussed constrained output and the ideal unconstrained
earlier in the paper and then implemented in the form output. In this formulation a term for penalizing the
suggested in Figs 2(a) and (c). control action is not necessary, which eliminates some
Because of the quasi open-loop structure, however, tuning parameters. If the constraints are not active the
the unconstrained controller Q is unaware of the input optimal value for the objective function found by the
constraints and the performance can suffer badly when LP is zero and the process output is equal to what
the inputs saturate. Qualitatively, the inputs should be would be obtained by the linear time invariant
held longer at the constraints than what the controller controller which was designed in the first step.
in Figs l(a) and (c) does, in order to compensate for Obviously when the constraints become active an
the restricted control action. Specifically, Q should be analysis of the properties of the control algorithm
designed with the constraints included explicitly in the becomes impossible. However, because the response
MPC formulation. As discussed by Campo and Morari is kept close to the response of the unconstrained
(1986) and summarized next there are a number of system it is reasonable to expect that its properties
possibilities for defining an appropriate objective (stability, robustness, etc.) remain preserved.
function. In general Q becomes nonlinear and time
varying and no stability analysis procedure is
available when the constraints are active. oo-Norm objective function
Campo and Morari (1986) adopt Brosilow's idea but
2-Norm objective function use a different norm. The unconstrained controller is
The objective function (12) employs the 2-norm designed by the IMC procedure. Then the oo-norm of
spatially (for the output and input vectors at a the error between the predicted constrained and the
Model Predictive Control: theory and practice---a survey 345
ideal unconstrained process output is minimized ~ e Solution to this problem when (54) is not present
and g(x, u) is a function of u only can be found in all
rain max w~, I)~i(k+ l I k) - ~(k +/)1. (51) classical references on optimal control (Athans and
Au(k)...Au(k +ra-I) I=l,P
i=l,r Falb, 1966; Lee and Markus, 1967). The variational
In practice, the ~-norm is particularly meaningful methods become extremely complex when inequalities
when peak excursions from desired trajectories are to involving the states are present. Small examples are
be avoided. The 1- and 2-norms tend to keep average shown by Bryson et al. (1963) and Denham and
deviations small but allow large peak deviations. Also, Bryson (1964). Because of the computational
in general, the oo-norm results in an LP which can be complexity these methods are not suitable for on-line
solved more efficiently than the 1-norm LP. use.
It is more promising to discretize the control vector
Summary with respect to time and to convert (52)-(55) into a
Because of our deep understanding of linear non-linear programming problem ("mathematical
systems it is preferable to employ an algorithm which programming approach" (Bryson and He, 1975)). It is
is linear and for which a closed-form expression can also possible to use a "black box" simulation model
be derived and analyzed as long as the constraints are instead of (53) and to compute the gradients
not active. QDMC and the formulation by Brosilow necessary for the mathematical program numerically.
have these characteristics. The experience with This has been done successfully for the on-line
different norms in conjunction with the Brosilow optimization of gas pipeline networks by Marqu6s
formulation has been too limited to date to draw any (Marqu6s, 1985; Marqu6s and Morari, 1986).
definite conclusions.
Linearized optimal control
NON-LINEAR MPC Garcia (1984) linearized a problem similar to
While we can deal with mild nonlinearities just by (52)-(55) and solved the linear problem by the
detuning linear controllers, it is likely that in the DMC/QDMC approach. The linear model is updated
presence of strong nonlinearities, non-linear control- as the state of the process changes and used to obtain
lers offer distinct advantages. the step response coefficients. The non-linear model is
Even though there any many important unresolved used to perform the model prediction. That is, the
details the conceptual extension of the IMC structure non-linear differential equations are integrated on-
to non-linear systems is straightforward. Nevertheless line, in parallel with the process, while the controller
only Frank (1974) appears to have recognized its is a linear QDMC controller. The application to a
potential. batch reactor produced excellent results.
For non-linear systems the assumption of an
unmeasured additive disturbance acting at the process Inversion
output is usually artificial. Indeed, for non-linear The implicit objective of (12)-(14) is to make the
systems the issues of model error (robustness) and system output y track the reference trajectory r, while
unmeasured disturbances become indistinguishable. keeping the manipulated variables u at a reasonable
In general, all blocks in Fig. 2(a) or (c) can be level. In the I M C context the penalty term on the
nonlinear. The process model is a simulation program inputs is omitted. Instead, r, is processed through the
where the non-linear differential equations are solved I M C filter so that r can be tracked exactly without
on-line in parallel with the process. putting too much strain on the inputs. Thus the
As shown by Economou et al. (1986) the stability control algorithm should determine u such that
properties of the IMC structure carry over to
nora-linear systems when the appropriate definitions Pu = r (56)
are made. Indeed, as for linear systems, all stabilizing where r is the filtered reference. In other words the
controllers for the non-linear plant P are generated right inverse of p is to be determined
from stable IMC controllers Q.
A general technique for the design of the u = P-lr. (57)
non-linear controller Q is not available to date. Three
attempts reported in the literature will be discussed From this point of view the key questions for the
next. design of non-linear MPC involve the existence,
uniqueness and numerical construction of the inverse
Non-linear optimal control of the non-linear operator P. For specific static
In analogy with (12)-(15) one can define the nonlinearities, saturation constraints and muitiplica-
general non-linear problem tire nonlinearities the construction of (approximate)
stable inverses is discussed by Frank (1974). Hirschorn
min G[x(t,)] + F[x(t), u(t)] dt (52) (1979) derived conditions for the existence of the
inverse of quite a general class of non-linear
subject to operators and also described an analytic procedure for
=fix(t), u(t)], X(to) = Xo (53) their inversion. Though mathematically rigorous, the
method implies the use of higher order derivatives and
h(x, u) = 0 (54)
is therefore very sensitive to noise and/or numerical
g(x, u) <- o. (55) errors. It is unsuitable for on-line use.
346 C.E. GARCiA et al.
Economou and co-workers (Economou et al., 1986; Despite the progress during the last decade there
Economou, 1986) experimented with various methods are quite a few unresolved research issues.
(contraction mapping, Newton's method) to solve the
• Linear systems. T h e tuning procedures necessary
non-linear operator equation (56) for u. Simulation
to achieve robust performance for ill-conditioned
experiments were very successful. Indeed, stable
MIMO systems (for an example see the study of
performance of the non-linear control scheme was
Skogestad and Morari (1986)) are very complex. We
found under conditions where any linear controller
lack the physical understanding of the connections
would make the system unstable.
between uncertainty and performance to suggest
In an extension of this work, Economou (1986) and
Economou and Morari (1985) applied the ideas of simpler techniques.
• Constraints. T h e LP and QP algorithms which
non-linear operator inversion directly to the design of
have been employed for MPC are basically "off-the-
feedback controllers.
shelf" with minor modifications. For large-scale
ROBUSTNESS applications more efficient tailor made algorithms will
By robustness we mean roughly that the quality of be needed. In the constrained algorithm model
performance of the feedback system is preserved uncertainty is neglected. Thus, there is no guarantee
when the dynamic behavior of the real plant is that the real process variables will satisfy the
different from that assumed in the model. The constraints when the model variables do. Also, when
importance of the robustness problem has become there are constraints on both the inputs and states it is
recognized universally in the last decade and much possible that the L P / Q P is unable to find a feasible
progress has been made toward its systematic solution solution when a disturbance pushes the process
(at least for linear systems: very little is known about outside the usual operating region. It is unclear how
non-linear systems). A detailed review and discussion to recover gracefully in these situations. Some ideas
have been presented by Gutman (1982), Garcia and
is beyond the scope of this paper. The reader is
Morshedi (1984) and Brosilow and Zhao (1986).
referred to Morari and Doyle (1986).
• Non-linear systems. T h e surface has barely been
In its linear form, unconstrained MPC is equivalent
to classic feedback and therefore all the developed scratched. All important questions like nominal
robustness analysis procedures can be applied. stability/model uncertainty and constraint handling
Whether MPC is robust or not depends on how it is are unanswered. The computational requirements are
designed. MPC is not inherently more or less robust expected to be very high and the power of special
than classic feedback as has been falsely claimed computer architectures (e.g. hypercube) will have to
(Mehra et al., 1982). Also, the large number of be utilized.
parameters in a step or impulse response model as
Acknowledgement--Partial support from the National
opposed to the three or four parameters in a Science Foundation and the Department of Energy are
parsimonious transfer function model does not add gratefully acknowledged. The authors would like to thank E.
any robustness. By that argument one could include Zafiriou for some helpful comments and criticism.
some parameters in the model which have no effect on
the input-output description at all and obtain even
REFERENCES
more robustness. Arkun, Y. (1978). Design of steady-state optimizing control
On the other hand there is no doubt that MPC can structures for chemical processes. Ph.D. Thesis, University
be adjusted more easily for robustness than classic of Minnesota.
feedback. For example, the filter in MAC and IMC Arkun, Y., J. Hollett, W. M. Canney and M. Morari (1986).
Experimental study of internal model control. Ind. Engng
has a very direct effect on robustness (Garcia and Chem. Process Des. Dev., 25, 102-108.
Morari, 1982, 1985a, b). This fact might be respon- Athans, M. and P. L. Falb (1966). Optimal Control: an
sible for the misconceptions regarding MPC Introduction to the Theory and its Applications. McGraw-
robustness. Hill, New York.
Bristol, E. H. (1980). After DDC. Idiomatic control. Chem.
Engng Prog. 76(11), 84-89.
CONCLUSIONS Brosilow, C. B. (1979). The structure and design of Smith
Process control is driven by the need to respond Predictors from the viewpoint of inferential control. Proc.
competitively to a rapidly changing marketplace. Joint Aut. Control Conf., Denver, Colorado.
Contrary to aerospace control the dynamics of the Brosilow, C. B., G. Q. Zhao and K. C. Rao (1984). A linear
underlying system are simple (overdamped, no high programming approach to constrained multivariable
control. Proc. Am. Control Conf., San Diego, California,
frequency resonances). The difficulties arise from pp. 667-674.
model uncertainty and the requirement to satisfy Brosilow, C. B. and G. Q. Zhao (1986). A linear
dynamically a number of constraints on states and programming approach to constrained multivariable
inputs, which define the economically optimal process control. Adv. Control Dyn. Syst., 24.
Bryson, A. E., W. F. [)enham and S. E. Dreyfus (1963).
operating condition. MPC is an attractive tool for Optimal programming problems with inequality
addressing these issues. Tuning for robustness is direct constraints--I. Necessary conditions for extremal solu-
and the constraints are considered explicitly in the tions. A I A A J., 1, 2544-2550.
algorithm. The future outputs predicted by the Bryson, A. E. and Y. Ho (1975). Applied Optimal Control.
internal model could be displayed to the operator to Hemisphere, Washington, DC/Wiley, New York.
C,aldwell, J. M. and G. D. Martin (1987). On-line analyzer
give him confidence in the algorithm and to allow predictive control. Sixth Annual Control Expo Conf.,
transparent on-line tuning. Rosemont, Illinois, 19-21 May.
Model Predictive Control: theory and practice---a survey 347
Campo,P. J. and M. Morari (1986). =-Norm f0rmulation of tuning guidelines. Ind, Etigng Chem. Process Des. Dev.,
model predictive control problems. Proc. Am. Control 24, 484-494.
Conf., Seattle, Washington, pp. 339-343. Garcia, C. E. and D. M. Prett (1986). Advances in industrial
Chang, S. S. L. (1961). Synthesis of Optimum Control model--predictive control. In M. Morari and T. J.
Systems. McGraw-Hill, New York. McAvoy (Eds), Chemical Process Control---CPC 111, pp.
Chang, T. S. and D. E. Seborg (1983). A linear 249-294. CACHE and Elsevier, Amsterdam.
programming approach to multivariable feedback control Oiloi, W. (1959). Zur Theorie und Verwirklichung einer
with inequality constraints. Int. Y. Control, 37, 583-597. Regelung fiir Laufzeitstreken nach dem Prinzip der
Clarke, D. W. and C. Mohtadi (1987). Properties of erg~nzenden Riickfiihrnng. Ph.D. Thesis, TH Stuttgart.
generalized predictive control. Proc. lOth IFAC World Grosdidier, P. (1987). Setpoint, Inc., Houston, Personal
Congress, Munich, Germany, Vol. 10, pp. 63-74. communication.
Clarke, D. W., C. Mohtadi and P. S. Tufts (1987a). Gutman, P. O. (1982). Controllers for bilinear and
Generalized predictive controi--I. The basic algorithm. constrained linear systems. Ph.D. Thesis, Lund Institute of
Automatica, 23, 137-148. Technology.
Clarke, D. W., C. Mohtadi and P. S. Tufts (1987b). Hirschorn, R. M. (1979). Invertibility of Multivariable
Generalized predictive control--II. Extensions and inter- nonlinear control systems. IEEE Trans. Aut. Control, 24,
pretations. A utomatica, 7,3, 149-160. 855-865.
Cutler, C. R. and B. L. Ramaker (1979). Dynamic matrix Horowitz, I. M. (1963). Synthesis of Feedback Systems.
eontrol---a computer control algorithm. AIChE National Academic Press, London.
Mtg, Houston, Texas (1979); also Proc. Joint Aut. Control Kwakernaak, H. and R. Sivan (1972). Linear Optimal
Conf., San Francisco, California (1980). Control Systems. Wiley-Interscience, New York.
Cutler, C. R. and R. B. Hawkins (1987). Constrained Lee, E. B. and L. Markus (1967). Foundations of Optimal
multivariable control of a hydroeracker reactor. Proc. Am. Control Theory. Wiley, New York.
Control Conf., Minneapolis, Minnesota, pp. 1014-1020. Levien, K. L. (1985). Studies in the design and control of
DeKeyser, R. M. C. and A. R. van Cauwenbcrghe (1982). coupled distillation columns. Ph.D. Thesis, University of
Typical application possibilities for self-tuning predictive Wisconsin, Madison.
control. IFAC Symp. on Identification and System Levien, K. L. and M. Morari (1987). Internal model control
Parameter Estimation, Washington, DC, Vol. 2, pp. of coupled distillation columns. A.LCh.E. Jl, 33, 83.
1552-1557. Marquts, D. (1985). On-line optimization of large scale
DeKeyser, R. M. C. and A. R. van Cauwenberglie (1985). dynamic systems. Ph.D. Thesis, University of Wisconsin,
Extended prediction self-adaptive control. IFAC Syrup. on Madison.
Identification and System Parameter Estimation, York, Marquts, D. and M. Morari (1986). Model predictive control
U.K., pp. 1255-1260. of gas pipeline networks. Proc. Am. Control Conf.,
DeKeyser, R. M. C., Ph. G. A. VandeVelde and F. A. G. Seattle, Washington, pp. 349-354.
Dumoriter (1985). A comparative study of self-adaptive Martin, G. D., J. M. Caldwell and T. E. Ayral (1986).
long-range predictive control methods. IFAC Symp. on Predictive control applications for the petroleum refining
Identification and System Parameter Estimation, York, industry. Japan Petroleum Institute--Petroleum Refining
U.K., pp. 1317-1322. Conf., Tokyo, Japan, 27-28 October.
Denham, W. F. and A. E. Bryson (1964). Programming Matsko, T. N. (1985). Internal model control for chemical
problems with inequality constraints--II. Solution by the recovery. Chem. Engng Prog., 81(12), 46-51.
steepest ascent. AIAA J., 2, 25-26. Mehra, R. K. and R. Rouhani (1980). Theoretical
Eeonomou, C. G. and M. Morari (1985). Newton control considerations on model algorithmic control for
laws for nonlinear controller design. Proc. 1EEE Conf. on nonminimum phase systems. Proc. Joint Aut. Control
Decision and Control. Fort Lauderdale, Florida, pp. Conf., San Francisco, California.
1361-1366. Mehra, R. K., R. Rouhani, J. Eterno, J. Richalet and A.
Economou, C. G. (1986). An operator theory approach to Rault (1982). Model algorithmic control: review and
nonlinear controller design. Ph.D. Thesis, California recent development. Engng Foundation Conf. on Chemi-
Institute of Technology. cal Process Control II, Sea Island, Georgia, pp. 287-310.
Economou, C. G., M. Morari and B. Palsson (1986). Morari, M. and G. Stephanopoulos (1980). Minimizing
Internal model control---5. Extension to nonlinear unobservability in inferential control schemes. Int. J.
systems. Ind. Engng Chem. Process Des. Dev., 2,$, Control, 31, 367-377.
403-411. Morari, M. and J. C. Doyle (1986). A unifying framework
Enns, D. (1984). Model reduction for control system design. for control system design under uncertainty and its
Ph.D. Thesis, Stanford University. implications for chemical process control. In M. Morari
Frank, P. M. (1974). Entwurf yon Regelkreisen mit and T. J. McAvoy (Eds), Chemical Process Control---CPC
vorgeschriebenem Verhalten. G. Braun, Karlsruhe (1974). II1, pp. 5-52. CACHE and Elsevier, Amsterdam.
Gantmacher, F. R. (1959). The Theory of Matrices, Vol. I. Morari, M. and C. Scali (1989). A comparison of internal
Chelsea, New York. model control and linear quadratic optimal control for
Garcia, C. E. (1982). Studies in optimizing and regulatory SISO systems. Ind. Engng Chem. Res. (submitted).
control of chemical processing systems. Ph.D. Thesis, Morari, M. and E. Zafiriou (1989). Robust Process Control.
University of Wisconsin, Madison. Prentice-Hall, Englewood Cliffs, New Jersey.
Garcia, C. E. and M. Morari (1982). Internal model Morshedi, A. M., C. R. Cutler and T. A. Skrovanek (1985).
control--1. A unifying review and some new results. Ind. Optimal solution of dynamic matrix control with linear
Engng Chem. Process Des. Dev., 21, 308-323. programming techniques (LDMC). Proc. Am. Control
Gareia, C. E. (1984). Quadratic dynamic matrix control of Conf., Boston, Massachusetts, pp. 199-208.
nonlinear processes. An application to a batch reaction Newton, G. C., L. A. Gould and J. F. Kaiser (1957).
process. AIChE Annual Mtg, San Francisco, California. Analytical Design of Feedback Controls. Wiley, New York.
Garcia, C. E. and A. M. Morshedi (1984). Quadratic Parrish, J. R. and C. B. Brosilow (1985). Inferential control
programming solution of dynamic matrix control applications. Automatica, 21, 527-538.
(QDMC). Proc. Am. Control Conf., San Diego, Peterka, V. (1984). Predictor-based self-tuning control.
California (1984); also Chem. Engng Commun., 46, 73-87 Automatica, 7,0, 39-50.
(1986). Prett, D. M. and R. D. Gillette (1979). Optimization and
Garcia, C. E. and M. Morari (1985a). Internal model constrained multivariable control of a catalytic cracking
control--2. Design procedure for multivariable systems. unit. AIChE National Mtg, Houston, Texas; also Proc.
Ind. Engng Chem. Process Des. Dev., 24, 472-484. Joint Aut. Control Conf., San Francisco, California.
Garcia, C. E. and M. Morari (1985b). Internal model Prett, D. M. and C. E. Garcia (1988). Fundamental Process
control----3. Multivariable control law computation and Control. Butterworths, Stoneham, Massachusetts.
348 C.E. GARCiA et al.
Propoi, A. I. (1963). Use of LP methods for synthesizing The multivariable case. IEEE Trans. Aut. Control. AC-21,
sampled-data automatic systems. Automn Remote Control, 319-338.
24, 837. Z,adeh, L. A. and B. H. Whalen (1962). On optimal control
Reid, J. G., R. K. Mehra and E. Kirkwood (1979). and linear programming. IRE Trans. Aut. Control, 7(4),
Robustness properties of output predictive dead-beat 45.
control: SISO case. Proc. IEEE Conf. on Decision and Zafiriou, E. and M. Morari (1985a). Digital controllers for
Control, Fort Lauderdale, Florida. SISO systems. A review and a new algorithm. Int. J.
Richalet, J. A., A. Rault, J. L. Testud and J. Papon (1978). Control, 42, 855-876.
Model predictive heuristic control: applications to an Zafiriou, E. and M. Morari (1985b). Robust digital controller
industrial process. Automatica, 14, 413-428. design for multivariable systems. AIChE National Mtg,
Ricker, N. L. (1985). The use of quadratic programming for Chicago, Illinois.
constrained internal model control. Ind. Engng Chem. Zafiriou, E. and M. Morari (1986a). Design of the IMC filter
Process Des. Dev., 24, 925-936. by using the structured singular value approach. Proc.
Skogestad, S. and M. Morari (1986). Control of ill- Am. Control Conf., Seattle, Washington, pp. 1-6.
conditioned plants: high purity distillation. AIChE Annual Zafiriou, E. and M. Morari (1986b). Design of robust
Mtg, Miami Beach. Florida. controllers and sampling time selection for SISO systems.
Smith, O. J. M. (1957). Closer control of loops with dead Int. J. Control, 44, 711-735.
time. Chem. Engng Prog., 53(5), 217-219. Zafiriou, E. and M. Morari (1987). Digital controller design
Vidyasagar, M. (1985). Control System Synthesis----a for multivariable systems with structural closed-loop
Factorization Approach. MIT, Cambridge, Massachusetts. performance specifications. Int. J. Control, 46, 2087-2111.
Ydstie, B. E. (1984). Extended horizon adaptive control. 9th Zames, G. (1981). Feedback and optimal sensitivity: model
World Congress of the IFAC, Budapest, Hungary. reference transformations, multiplicative seminorms, and
Youla, D. C., J. J. Bongiorno and H. A. Jabr (1976a). approximate inverses. IEEE Trans. Aut. Control, AC-?,6,
Modern Wiener-Hopf design of optimal controllers---I. 301-320.
The single-input-output case. IEEE Trans. Aut. Control, Zirwas, H. C. (1958). Die erg~inzende Riickftihrung als
AC-21, 3-13. Mittel zur schnellen Regelung von Regeistrecken mit
Youla, D. C., H. A. Jabr and J. J. Bongiorno (1976b). Laufzeit. Ph.D. Thesis, TH Stuttgart.
Modern Wiener-Hopf design of optimal controllers---II.