AN OVERVIEW OF INDUSTRIAL MODEL

PREDICTIVE CONTROL TECHNOLOGY

S. Joe Qin
Department of Chemical Engineering
The University of Texas at Austin
Austin, TX 78712
Thomas A. Badgwell
Department of Chemical Engineering
Rice University
Houston, TX 77251
Abstract
This paper provides an overview of commercially available Model Predictive Control (MPC)
technology, based primarily on data provided by MPC vendors. A brief history of industrial
MPC technology is presented rst, followed by results of our vendor survey of MPC control and
identication technology. A general MPC control algorithm is presented, and approaches taken
by each vendor for the dierent aspects of the calculation are described. Identication technology
is then reviewed to determine similarities and dierences between the various approaches. MPC
applications performed by each vendor are summarized by application area. The nal section
presents a vision of the next generation of MPC technology, with an emphasis on potential
business and research opportunities.
Keywords
Industrial Survey, Model Predictive Control

Introduction The primary purpose of this paper is to present

Model Predictive Control (MPC) refers to a class of an overview of commercially available MPC technol-
algorithms that compute a sequence of manipulated ogy. A brief history of MPC technology development
variable adjustments in order to optimize the future is presented rst. A general MPC controller formula-
behavior of a plant. Originally developed to meet the tion is then described to provide a basis for discussion
specialized control needs of power plants and petroleum of the commercial products. Results of our industrial
reneries, MPC technology can now be found in a survey are then presented. The survey is not exhaus-
wide variety of application areas including chemicals, tive in that several well-known companies either were
food processing, automotive, aerospace, metallurgy, not asked to participate, chose not to participate, or
and pulp and paper. responded too late to be included in this paper. Nev-
Several authors have published excellent reviews ertheless we believe that the products discussed here
of MPC theoretical issues, including the papers of are suciently representative to allow us to draw con-
Garca et al. (Garca, Prett and Morari, 1989), Ricker clusions regarding the current state of MPC technol-
(Ricker, 1991), Morari and Lee (Morari and Lee, 1991), ogy. Signicantly unique features of each oering are
Muske and Rawlings (Muske and Rawlings, 1993) and outlined and discussed. MPC applications to date by
Rawlings et al. (Rawlings, Meadows and Muske, 1994). each vendor are summarized by application area. The
The other papers in the present session by Mayne nal section presents a view of the next generation of
(Mayne, 1996) and Lee (Lee, 1996) summarize the very MPC technology, emphasizing potential business and
latest technical developments in MPC control theory. research opportunities.
Froisy provides a vendor's perspective on industrial
MPC technology and summarizes likely future devel- A Brief History of Industrial MPC
opments (Froisy, 1994). A recent survey of MPC tech- This section presents an abbreviated history of indus-
nology in Japan provides a wealth of information on trial MPC technology. Control algorithms are empha-
application issues from the point of view of MPC users sized here because relatively little published informa-
(Ohshima, Ohno and Hashimoto, 1995). tion is available on the identication technology.
1
2 MODEL PREDICTIVE CONTROL
The development of modern control concepts can unconstrained linear plant, it had little impact on
be traced to the work of Kalman in the early 1960's, control technology development in the process indus-
who sought to determine when a linear control system tries. The most signicant of the reasons cited for
can be said to be optimal (Kalman, 1960a; Kalman, this failure include (Richalet, Rault, Testud and Pa-
1960b). Kalman studied a Linear Quadratic Regula- pon, 1978; Garca et al., 1989) :
tor (LQR) designed to minimize an quadratic objec-
tive function. The process to be controlled can be de-  constraints
scribed by a discrete-time, deterministic linear state-  process nonlinearities
space model:
 model uncertainty (robustness)
xk+1 = Axk + Buk (1)  unique performance criteria
yk = Cxk  cultural reasons (people, education, etc.)
The vector u represents process inputs, or manipu- It is well known that the economic operating point
lated variables; vector y describes process output mea- of a typical process unit often lies at the intersection
surements. The vector x represents process states. Fig- of constraints (Prett and Gillette, 1980). A successful
ure 1 provides a schematic representation of a state industrial controller must therefore maintain the sys-
space model. The state vector is dened such that
knowing its value at time k and future inputs allows tem as close as possible to constraints without violating
one to predict how the plant will evolve for all future them. In addition, process units are typically complex,
time. Much of the power of Kalman's work relies on nonlinear, constrained multivariable systems whose dy-
the fact that this general process model was used. namic behavior changes with time due to such eects
The objective function to be minimized penalizes as changes in operating conditions and catalyst aging.
squared input and state deviations from the origin and Process units are also quite individual so that devel-
includes separate state and input weight matrices Q opment of process models from fundamental physics
and R to allow for tuning trade-os: and chemistry is dicult to justify economically. In-
deed the application areas where LQG theory had a
1
X?
more immediate impact, such as the aerospace indus-
J = kxk+j k2Q + kuk+j k2R (2) try, are characterized by physical systems for which it is
j =1 technically and economically feasible to develop accu-
where the norm terms in the objective function are rate fundamental models. Process units may also have
dened as follows: unique performance criteria that are dicult to express
in the LQG framework, requiring time dependent out-
kxk2Q = xT Qx (3) put weights or additional logic to delineate dierent
Implicit in this formulation is the assumption that all operating modes. However the most signicant rea-
variables are written in terms of deviations from a de- sons that LQG theory failed to have a strong impact
sired steady-state. The solution to the LQR problem may have been related to the culture of the industrial
was shown to be a proportional controller, with a gain process control community at the time, in which in-
matrix K computed from the solution of a matrix Ri- strument technicians and control engineers either had
catti equation: no exposure to LQG concepts or regarded them as im-
practical.
uk = ?Kxk (4) This environment led to the development, in indus-
try, of a more general model based control methodology
The innite prediction horizon of the LQR algo- in which the dynamic optimization problem is solved
rithm endowed the algorithm with powerful stabiliz- on-line at each control execution. Process inputs are
ing properties; it was shown to be stabilizing for any computed so as to optimize future plant behavior over
reasonable linear plant (stablizable and detectable) as a time interval known as the prediction horizon. In
long as the objective function weight matrices Q and R the general case any desired objective function can be
are positive denite. A dual theory was developed to used. Plant dynamics are described by an explicit pro-
estimate plant states from noisy input and output mea- cess model which can take, in principle, any required
surements, using what is now known as a Kalman Fil- mathematical form. Process input and output con-
ter. The combined LQR controller and Kalman lter is straints are included directly in the problem formula-
called a Linear Quadratic Gaussian (LQG) controller. tion so that future constraint violations are anticipated
Constraints on the process inputs, states and outputs and prevented. The rst input of the optimal input se-
were not considered in the development of LQG theory. quence is injected into the plant and the problem is
Although LQG theory provides an elegant and solved again at the next time interval using updated
powerful solution to the problem of controlling an process measurements. In addition to developing more
 Industrial Model Predictive Control Technology 3
exible control technology, new process identication IDCOM
technology was developed to allow quick estimation The rst description of MPC control applications was
of empirical dynamic models from test data, substan- presented by Richalet et al. at a 1976 conference
tially reducing the cost of model development. This (Richalet, Rault, Testud and Papon, 1976) and later
new methodology for industrial process modeling and summarized in a 1978 Automatica paper (Richalet
control is what we now refer to as Model Predictive et al., 1978). They described their approach as Model
Control (MPC) technology. Predictive Heuristic Control (MPHC). The solution
In modern processing plants the MPC controller software was referred to as IDCOM, an acronym for
is part of a multi-level hierarchy of control functions. Identication and Command. The distinguishing fea-
This is illustrated in Figure 2, which shows a conven- tures of the IDCOM approach are:
tional control structure on the left for Unit 1 and a
model predictive control structure on the right for Unit  impulse response model for the plant, linear in in-
2. Similar hierarchical structures have been described puts or internal variables
by Richalet et al. (Richalet et al., 1978) and Prett and
Garca (Prett and Garca, 1988). At the top of the  quadratic performance objective over a nite pre-
structure a plant-wide optimizer determines optimal diction horizon
steady-state settings for each unit in the plant. These  future plant output behavior specied by a refer-
may be sent to local optimizers at each unit which run ence trajectory
more frequently or consider a more detailed unit model
than is possible at the plant-wide level. The unit opti-  input and output constraints included in the for-
mizer computes an optimal economic steady-state and mulation
passes this to the dynamic constraint control system for
implementation. The dynamic constraint control must  optimal inputs computed using a heuristic itera-
move the plant from one constrained steady state to an- tive algorithm, interpreted as the dual of identi-
other while minimizing constraint violations along the cation.
way. In the conventional structure this is accomplished Richalet et al. chose a black-box process represen-
by using a combination of PID algorithms, Lead-Lag tation of the process, shown in Figure 3. From this
(L/L) blocks and High/Low select logic. It is often dif- point of view the process inputs in uence the pro-
cult to translate the control requirements at this level cess outputs directly. Process inputs are divided into
into an appropriate conventional control structure. In manipulated variables (MV's) which the controller ad-
the MPC methodology this combination of blocks is justs, and disturbance variables (DV'S) which are not
replaced by a single MPC controller. available for control. Process outputs are referred to as
Although the development and application of MPC controlled variables (CV's). They chose to describe the
technology was driven by industry, it should be noted relationship between process inputs and outputs using
that the idea of controlling a system by solving a se- a discrete-time Finite Impulse Response (FIR) model.
quence of open-loop dynamic optimization problems For the single input, single output (SISO) case the FIR
was not new. Propoi, for example, described a mov- model looks like:
ing horizon controller in 1963 (Propoi, 1963). Lee and
Markus (Lee and Markus, 1967) anticipated current
MPC practice in their 1967 optimal control text: yk+j =
XN hi uk+j?i (5)
i=1
One technique for obtaining a feedback This model predicts that the output at a given time
controller synthesis from knowledge of open- depends on a linear combination of past input values;
loop controllers is to measure the current the summation weights hi are the impulse response co-
control process state and then compute very ecients. The sum is truncated at the point where
rapidly for the open-loop control function. past inputs no longer in uence the output; this repre-
The rst portion of this function is then used sentation is therefore only possible for stable plants.
during a short time interval, after which a The nite impulse response was identied from
new measurement of the function is computed plant test data using an algorithm designed to min-
for this new measurement. The procedure is imize the structural distance between the plant and
then repeated. model impulse responses in parameter space. The re-
There is, however, a wide gap between theory and sulting iterative algorithm makes small adjustments to
practice. The essential contribution of industry was to the coecients at each step in such a way that the
put these ideas into practice on operating units. Out structural distance continuously decreases. The algo-
of this experience came a fresh set of problems that has rithm was shown to converge to unbiased parameter es-
kept theoreticians busy ever since. timates in the face of noisy output measurements. The
control problem was solved using the same algorithm
4 MODEL PREDICTIVE CONTROL
by noting that control is the mathematicaldual of iden- main fractionator column, a power plant steam gen-
tication. In the identication problem one knows the erator and a Poly-Vinyl Chloride (PVC) plant. All of
outputs and the inputs and wishes to estimate the co- these examples are constrained multivariable processes.
ecients; in the control problem one knows the de- The main fractionator example involved controlling key
sired output trajectory and the model coecients and tray temperatures to stabilize the composition of heavy
the goal is to estimate the required inputs. Because and light product streams. The controller adjusted
the output prediction appears as a dot product of in- product owrates to compensate for inlet temperature
put and coecient vectors, the same algorithm can be disturbances and to maintain the level of a key inter-
used to nd either one. The iterative nature of the nal tray. The power plant steam generator problem
control algorithm allows input and output constraints involved controlling the temperature and pressure of
to be checked as the algorithm proceeds to a solution. steam delivered to the turbine. This application is in-
Because the control law not linear and could not be teresting because the process response time varied in-
expressed as a transfer function, Richalet et al. refer to versely with load on the system. This nonlinearity was
it as heuristic. In today's context the algorithm would overcome by executing the controller with a variable
be referred to as a nonlinear MPC controller. sample time. Benets for the main fractionator appli-
The MPHC algorithm drives the predicted future cation were reported as $150,000/yr, due to increasing
output trajectory as closely as possible to a reference the owrate of the light product stream. Combined en-
trajectory, dened as a rst order path from the cur- ergy savings from two columns in the PVC plant were
rent output value to the desired setpoint. The speed reported as $220,000/yr.
of the desired closed loop response is set by the time
constant of the reference trajectory. This is important DMC
in practice because it provides a natural way to control Engineers at Shell Oil developed their own indepen-
the aggressiveness of the algorithm; increasing the time dent MPC technology in the early 1970's, with an
constant leads to a slower but more robust controller. initial application in 1973. Cutler and Ramaker pre-
Richalet et al. make the important point that dy- sented details of an unconstrained multivariable con-
namic control must be embedded in a hierarchy of plant trol algorithm which they named Dynamic Matrix
control functions in order to be eective. They de- Control (DMC) at the 1979 National AIChE meet-
scribe four levels of control, very similar to the struc- ing (Cutler and Ramaker, 1979) and at the 1980
ture shown in Figure 2: Joint Automatic Control Conference (Cutler and Ra-
 Level 3 - Time and space scheduling of production maker, 1980). In a companion paper at the 1980 meet-
ing Prett and Gillette (Prett and Gillette, 1980) de-
 Level 2 - Optimization of setpoints to minimize scribed an application of DMC technology to an FCCU
costs and ensure quality and quantity of produc- reactor/regenerator in which the algorithm was modi-
tion ed to handle nonlinearities and constraints. Neither
paper discussed their process identication technology.
 Level 1 - Dynamic multivariable control of the Key features of the DMC control algorithm include:
plant
 Level 0 - Control of ancillary systems; PID control  linear step response model for the plant
of valves.  quadratic performance objective over a nite pre-
diction horizon
They point out that signicant benets do not
come from simply reducing the variations of a con-  future plant output behavior specied by trying to
trolled variable through better dynamic control at level follow the setpoint as closely as possible
1. The real economic benets come at level 2 where
better dynamic control allows the controlled variable  optimal inputs computed as the solution to a least-
setpoint to be moved closer to a constraint without vi- squares problem
olating it. This argument provides the basic economic The linear step response model used by the DMC
motivation for using MPC technology. This concept algorithm relates changes in a process output to a
of a hierarchy of control functions is fundamental to weighted sum of past input changes, referred to as in-
advanced control applications and seems to have been put moves. For the SISO case the step response model
followed by many practitioners. Prett and Garca., for looks like:
example, (Prett and Garca, 1988) describe a very sim-
ilar hierarchy in their 1988 text Fundamental Process
Control.
Richalet et al. describe applications of the MPHC yk+j =
X
N?1
si
uk+j?i + sN uk+j?N (6)
algorithm to a Fluid Catalytic Cracking Unit (FCCU) i=1
 Industrial Model Predictive Control Technology 5
The move weights s are the step response coecients.
i timization and dynamic control is quite similar to the
Mathematically the step response can be dened as structure described by Richalet et al. and has since be-
the integral of the impulse response; given one model come standard in industrial control system design.
form the other can be easily obtained. Multiple out- The DMC algorithm had the job of moving from
puts were handled by superposition. By using the step the system from one optimal steady-state to another.
response model one can write predicted future output Although the LP solution provided optimal targets
changes as a linear combination of future input moves. for process inputs and outputs, dynamic disturbances
The matrix that ties the two together is the so-called could potentially cause the DMC algorithm to move in-
Dynamic Matrix. Using this representation allows the puts away from their optimal steady-state targets in or-
optimal move vector to be computed analytically as the der to keep outputs at their steady-state targets. Since
solution to a least-squares problem. Feedforward con- moving one input away from its optimal target may be
trol is readily included in this formulation by modifying much more expensive than moving another, the con-
the predicted future outputs. In practice the required trol system should determine this trade-o in a ratio-
matrix inverse can be computed o-line to save com- nal way. The DMC algorithm was modied to account
putation. Only the rst row of the nal controller gain for such trade-os by including an additional equation
matrix needs to be stored because only the rst move for each input in the process model. The new equation
needs to be computed. required that the sum of all moves for a particular in-
The objective of a DMC controller is to drive the put should equal the total adjustment required to bring
output as close to the setpoint as possible in a least- that input to its optimal steady-state target. This al-
squares sense with a penalty term on the MV moves. lowed the inputs some freedom to move dynamically
This results in smaller computed input moves and a but required that the steady-state input solution be
less aggressive output response. As with the IDCOM satised in a least-squares sense, with trade-os deter-
reference trajectory, this technique provides a degree mined by the appropriate objective function weights.
of robustness to model error. Move suppression factors Prett and Gillette described additional modica-
also provide an important numerical benet in that tions to the DMC algorithm to prevent violation of
they can be used to directly improve the conditioning absolute input constraints. When a predicted future
of the numerical solution. input came suciently close to an absolute constraint,
Cutler and Ramaker showed results from a fur- an extra equation was added to the process model that
nace temperature control application to demonstrate would drive the input back into the feasible region.
improved control quality using the DMC algorithm. These were referred to as time variant constraints. Be-
Feedforward response of the DMC algorithm to inlet cause the decision to add the equation had to be made
temperature changes was superior to that of a conven- on-line, the matrix inverse solution had to be recom-
tional PID lead/lag compensator. puted at each control execution. Prett and Gillette de-
In their paper Prett and Gillette (Prett and veloped a matrix tearing solution in which the original
Gillette, 1980) described an application of DMC tech- matrix inverse could be computed o-line, requiring
nology to FCCU reactor/regenerator control. Four only the matrix inverse corresponding to active time
such applications were already completed and two ad- variant constraints to be computed on-line.
ditional applications were underway at the time the pa- The initial IDCOM and DMC algorithms repre-
per was written. The overall FCCU control system was sent the rst generation of MPC technology; they had
implemented in a multi-level hierarchy, with a nonlin- an enormous impact on industrial process control and
ear steady-state FCCU model at the top. At the start served to dene the industrial MPC paradigm.
of each optimization cycle, parameters in the nonlin-
ear model were estimated so as to match model predic- QDMC
tions with measured steady-state operating data. The The original IDCOM and DMC algorithms provided
calibrated nonlinear model was then perturbed numer- excellent control of unconstrained multivariable pro-
ically to generate partial derivatives of each process cesses. Constraint handling, however, was still some-
output with respect to each process input (the matrix what ad-hoc. Engineers at Shell Oil addressed this
of partial derivatives is known as the Jacobian matrix weakness by posing the DMC algorithm as a Quadratic
in numerical analysis). The partial derivatives were Program (QP) in which input and output constraints
then used in a Linear Program (LP) to compute a new appear explicitly. Cutler et al. rst described the
economic optimal operating point for the FCCU, sub- QDMC algorithm in a 1983 AIChE conference paper
ject to steady-state process constraints. The optimal (Cutler, Morshedi and Haydel, 1983). Garca and Mor-
process input and output targets were then passed to shedi published a more comprehensive description sev-
a DMC algorithm for implementation. As soon as the eral years later (Garca and Morshedi, 1986).
DMC controller moved the unit to the new steady state Key features of the QDMC algorithm include:
the optimization cycle was repeated. This separation
of the control system into constrained steady-state op-  linear step response model for the plant
6 MODEL PREDICTIVE CONTROL
 quadratic performance objective over a nite pre- non-minimum phase dynamics are present. When
diction horizon output constraint violations are predicted to occur,
one can attempt to minimize the violation in a least-
 future plant output behavior specied by trying to squared sense. This is the soft constraint concept il-
follow the setpoint as closely as possible subject to lustrated in the middle plot of Figure 10. Garca and
a move suppression term Morshedi described an approximate implementation of
 optimal inputs computed as the solution to a the soft constraint concept using a setpoint approxi-
quadratic program mation, illustrated at the bottom of Figure 10. The
setpoint approximation idea is to guess a priori where
Garca and Morshedi began with a clear and con- the optimal solution will require a constraint violation,
cise presentation of the unconstrained DMC algorithm, and penalize this deviation by adding a setpoint that
including an interesting discussion of tuning. Their ex- forces the output to stick to the constraint boundary.
perience showed that the DMC algorithm was closed One way to guess where output violations will occur is
loop stable when the prediction horizon was set long to examine output predictions based on the optimal in-
enough to include the steady-state eect of all com- put solution from the previous time step. Because it is
puted input moves. This is supported by a rigorous dicult to guess where the true solution at the current
proof presented by Garca and Morari (Garca and time step will require a constraint violation, the set-
Morari, 1982) which shows that the DMC algorithm point approximation method is generally sub-optimal.
is nominally stabilizing for a suciently large predic- A true soft constraint can be implemented by
tion horizon. adding a slack variable to an inequality constraint,
Garca and Morshedi then show how the DMC and then adding the slack variable to the objective
objective function can be re-written in the form of function to be minimized. This approach has been
a standard QP. Future projected outputs can be re- studied by several researchers, including Ricker et al.
lated directly back to the input move vector through (Ricker, Subrahmanian and Sim, 1988), Zariou and
the dynamic matrix; this allows all input and output Chiou (Zariou and Chiou, 1993), and Genceli and
constraints to be collected into a matrix inequality in- Nikolaou (Genceli and Nikolaou, 1993). Zheng and
volving the input move vector. Although the QDMC Morari (Zheng and Morari, 1995) recently analyzed
algorithm is a somewhat advanced control algorithm, an innite horizon MPC algorithm with soft output
the QP itself is one of the simplest possible optimiza- constraints implemented in this way. They show that
tion problems that one could pose. The Hessian of the global asymptotic stability can be guaranteed provided
QP is positive denite for any reasonable problem and that the plant is not unstable.
so the resulting optimization problem is convex. This Garca and Morshedi wrapped up their paper by
means that a solution can be found readily using stan- presenting results from a pyrolysis furnace applica-
dard commercial optimization codes. tion. The QDMC controller adjusted fuel gas pres-
The default QDMC algorithm requires strict en- sure in three burners in order to control stream tem-
forcement of input and output constraints at each point perature at three locations in the furnace. Their test
of the prediction horizon. Constraints that are strictly results demonstrated dynamic enforcement of input
enforced are referred to as hard constraints. This is constraints and decoupling of the temperature dynam-
illustrated in Figure 10. In practice Garca and Mor- ics. They reported good results on many applications
shedi report that hard output constraints are typically within Shell on problems as large as 12x12 (12 pro-
required to be satised over only a portion of the hori- cess outputs and 12 process inputs). They stated that
zon which they refer to as the constraint window. The above all, the QDMC algorithm had proven particu-
constraint window generally starts at some point in the larly protable in an on-line optimization environment,
future and continues on to steady state. They report providing a smooth transition from one constrained op-
that if non-minimum phase dynamics are present, per- erating point to another.
formance is improved by pushing the constraint win- The QDMC algorithm can be regarded as repre-
dow farther into the future. This amounts to ignoring senting a second generation of MPC technology, com-
hard output constraints during the initial portion of prised of algorithms which provide a systematic way to
the closed loop response. It is interesting that Rawl- implement input and output constraints. This was ac-
ings and Muske recently arrived at the same solution complished by posing the MPC problem as a QP, with
to ensure feasibility of their innite-horizon algorithm the solution provided by standard QP codes.
(Rawlings and Muske, 1993) after a careful theoretical
analysis. They show that output constraints can be IDCOM-M, HIECON, and PCT

made feasible by relaxing them for a nite time j1, and As MPC technology gained wider acceptance, and
they derive an upper bound for j1 . problems tackled by MPC technology grew larger and
Garca and Morshedi report another option for more complex, control engineers implementing second
handling output constraints that may be useful when
 Industrial Model Predictive Control Technology 7
generation MPC technology ran into other practical cations into a consistent set of relative weights. In
problems. The QDMC algorithm provided a system- some cases it does not make sense to include these
atic approach to incorporate hard input and output variables in the same objective function; driving the
constraints, but there was no clear way to handle an inputs to their optimal targets may lead to larger vi-
infeasible solution. For example it is possible for a feed- olation of output soft constraints, for example. Even
forward disturbance to lead to an infeasible QP; what when a consistent set of relative weights can be found,
should the control do to recover from infeasibility? The care must be taken to avoid scaling problems that lead
soft constraint formulation is not completely satisfac- to an ill-conditioned solution. Prett and Garca com-
tory because it means that all constraints will be vi- mented on this problem in their text on Fundamental
olated to some extent, as determined by the relative Process Control (Prett and Garca, 1988):
weights. Clearly some output constraints are more im-
portant than others, however, and should never be vio- The combination of multiple objectives
lated. Wouldn't it make sense then to shed low priority into one objective (function) does not allow
constraints in order to satisfy higher priority ones? the designer to re ect the true performance
In practice, process inputs and outputs can be lost requirements.
in real time due to signal hardware failure, valve satu- These issues motivated engineers at Adersa and
ration or direct operator intervention. They can just as Setpoint, Inc. to develop a new version of the IDCOM
easily come back into the control problem at any sam- algorithm. The version marketed by Setpoint was
ple interval. This means that the structure of the prob- called IDCOM-M (the M was to distinguish this from
lem and the degrees of freedom available to the control a single input/single output version called IDCOM-S),
can change dynamically. This is illustrated in Figure 4, while the Adersa version was referred to as HIECON
which illustrates the shape of the process transfer func- (Hierarchical Constraint Control). The IDCOM-M
tion matrix for three general cases. The square plant controller was rst described in a paper by Grosdidier
case, which occurs when the plant has just as many et al. (Grosdidier, Froisy and Hammann, 1988). A sec-
manipulated variables (MV's) as controlled variables ond paper presented at the 1990 AIChE conference de-
(CV'S), leads to a control problem with a unique solu- scribes an application of IDCOM-M to the Shell Funda-
tion. In the real world, square is rare. More common is mental Control Problem (Froisy and Matsko, 1990) and
the fat plant case, in which there are more MV's avail- provides additional details concerning the constraint
able than there are CV's to control. The extra degrees methodology. Distinguishing features of the IDCOM-
of freedom available in this case can be put to use for M algorithm include:
additional objectives, such as moving the plant closer
to an optimal operating point. When valves become  linear impulse response model of plant
saturated or lower level control action is lost, the plant
may reach a condition in which there are more CV's  controllability supervisor to screen out ill-
than MV's; this is the thin plant case. In this situa- conditioned plant subsets
tion it will not be possible to meet all of the control
objectives; the control specications must be relaxed  multi-objective function formulation; quadratic
somehow, for example by minimizing CV violations in output objective followed by a quadratic input ob-
a least-squared sense. jective
Fault tolerance is also an important practical issue.  controls a single future point in time for each out-
Rather than simply turning itself o as signals are lost, put, called the coincidence point, chosen from a
a practical MPC controller should remain online and reference trajectory
try to make the best of the sub-plant under its control.
A major barrier to achieving this goal is that a well  a single move is computed for each input
conditioned multivariable plant may contain a number
of poorly conditioned sub-plants. In practice an MPC  constraints can be hard or soft, with hard con-
controller must recognize and screen out poorly condi- straints ranked in order of priority
tioned sub-plants before they result in erratic control The IDCOM-M controller retains the linear im-
action. pulse response plant model used by the original ID-
It also became increasingly dicult to translate COM algorithm. The IDCOM-M controller allows the
control requirements into relative weights for a single capability to include purely integrating responses, how-
objective function. Including all the required trade- ever. These are assumed to describe the response of the
os in a single objective function means that relative rst order derivative of the output with respect to time.
weights have to be assigned to the value of output set- The IDCOM-M algorithm includes a controllabil-
point violations, output soft constraint violations, in- ity supervisor which decides, based on the current set
puts moves, and optimal input target violations. For of available inputs and outputs, which outputs can be
large problems it is not easy to translate control speci- independently controlled. The selection is based on
8 MODEL PREDICTIVE CONTROL
the eective condition number of the plant gain ma-  Input Optimization
trix; a list of controllability priorities is used to deter-
mine which outputs to drop from the problem if an Grosdidier et al. provide simulation results for
ill-conditioned set is encountered. a representative FCCU regenerator control problem.
An important distinction of the IDCOM-M algo- The problem involves controlling ue gas composition,
rithm is that it uses two separate objective functions, ue gas temperature, and regenerator bed tempera-
one for the outputs and then, if there are extra degrees ture by manipulating feed oil ow, recycle oil ow and
of freedom, one for the inputs. A quadratic output ob- air to the regenerator. The rst simulation example
jective function is minimized rst subject to hard input demonstrates how using multiple inputs can improve
constraints. Each output is driven as closely as possi- dynamic performance while reaching a pre-determined
ble to a desired value at a single point in time known optimal steady-state condition. A second example
as the coincidence point. The name comes from the demonstrates how the controller switches from control-
fact that this is where the desired and predicted val- ling one output to controlling another when a measured
ues should coincide. The desired output value comes disturbance causes a constraint violation. A third ex-
from a rst order reference trajectory that starts at the ample demonstrates the need for the controllability su-
current measured value and leads smoothly to the set- pervisor. When an oxygen analyzer fails, the controlla-
point. Each output has two basic tuning parameters; bility supervisor is left with only ue gas temperature
a coincidence point and a closed loop response time, and regenerator bed temperature to consider. It cor-
used to dene the reference trajectory. rectly detects that controlling both would lead to an
In many cases the solution to the output optimiza- ill-conditioned problem; this is because these outputs
tion is not unique. When additional degrees of free- respond in a very similar way to the inputs. Based
dom are present a second input optimization is per- on a pre-set priority it elects to control only the ue
formed. A quadratic input objective function is mini- gas temperature. When the controllability supervisor
mized subject to equality constraints that preserve the is turned o the same simulation scenario leads to er-
outputs found in the output optimization. The inputs ratic and unacceptable input adjustments.
are driven as closely as possible to their Ideal Resting Engineers at Promatics addressed similar issues
Values (IRV's) which may come, for example, from an in the development of their PCT algorithm (Predictive
overlying steady-state optimizer. By default the IRV Control Technology). This alogrithm also uses con-
for a given input is set to its current measured value. straint prioritization to recover from infeasibility. An-
The input optimization makes the most eective use of other interesting feature is known as predict-back, in
available degrees of freedom without altering the opti- which unmeasured disturbances that enter lower level
mal output solution. PID loops are estimated and used for feedforward con-
The IDCOM-M calculation is greatly simplied by trol. This feature is very useful when a lower level
computing a single move for each input. This input PID loop has slow dynamics, where the predict-back
blocking assumption results in a loss of performance but estimate helps the MPC controller respond to an un-
provides additional robustness to modeling errors. In measured disturbance much faster.
practice this has been an acceptable trade-o. Badg- The IDCOM-M algorithm is one of several that
well has analyzed the robustness properties of input- represent a third generation of MPC technology; oth-
blocking for the SISO case (Badgwell, 1995). ers include the PCT algorithm sold by Promatics,
In the IDCOM-M context, constraints are divided the RMPC controller developed by Honeywell, and the
into hard and soft categories, with the understanding PFC algorithm developed by Adersa. This genera-
that hard constraints must be ranked in order of prior- tion distinguishes between several levels of constraints
ity. When the calculation becomes infeasible, the low- (hard, soft, ranked), provides some mechanism to re-
est priority hard constraint is dropped and the calcu- cover from an infeasible solution, addresses the issues
lation is repeated. One can specify several constraints resulting from a control structure that changes in real
to have the same priority, and it is possible to require time, and allows for a wider range of process dynamics
that the control turn itself o and notify the opera- and controller specications.
tor if constraints above a given priority level cannot be
enforced. Survey of MPC Technology Products
Grosdidier et al. (Grosdidier et al., 1988) describe
the ow of a typical calculation: Commercial MPC technology has developed consider-
ably since the introduction of third generation tech-
 Determine available process inputs and outputs nology nearly a decade ago. We recently surveyed ve
MPC vendors in order to assess the current status of
 Determine the list of controllable outputs (control- commercial MPC technology. We believe that this list
lability supervisor) is representative in that the technology sold by these
 Output Optimization companies represents the industrial state of the art;
 Industrial Model Predictive Control Technology 9
we fully recognize that we have omitted some major yk = g (xk ) + k (8)
MPC vendors from our survey. Some companies were
not asked to participate, some chose not to partici- where uk 2 <mu is a vector of mu MV's, yk 2 <my
pate, and some responded too late to be included in is a vector of my CV's, xk 2 <n is a vector of n state
the paper. Only companies which have documented variables, vk 2 <mv is a vector of mv measured DV's,
successful MPC applications were asked to participate. wk 2 <mw is a vector of mw unmeasured DV's or noise.
It should be noted that several companies make use The PFC algorithm is the only one considered in
of MPC technology developed in-house but were not this survey that allows for nonlinear and unstable linear
included in the survey because they do not oer their internal models. Nonlinear dynamics can be entered
technology externally (Shell, Exxon, etc.). These MPC in the form of the nonlinear state space model shown
packages are either well-known to academic researchers above. The PFC algorithm uses transfer functions or
(e.g., QDMC from Shell Oil) or not known at all for ARX models to describe linear unstable dynamics. The
proprietary reasons. remaining MPC products are designed based on Linear
The companies surveyed and their product names Time-Invariant (LTI) process models with stable or in-
and acronyms are listed in Table 1. Initial data were tegrating dynamics. Nonlinearities may be accounted
collected from industrial vendors of MPC technology for in an approximate way by using a local linear model
using a written survey. Blank copies of the survey or nonlinear transformation of a specic CV. The SMC-
are available upon request from the authors. Sur- Idcom algorithm, for example, allows the model gains
vey information was supplemented by published pa- to be adjusted on-line.
pers, product literature (DMC Corp., 1994; Setpoint, A very general discrete-time LTI model form is the
Inc., 1993; Honeywell, Inc., 1995), and personal com- linear state space model:
munication between the authors and vendor represen-
tatives. Results of the survey are summarized in Ta- xk+1 = Axk + Buuk + Bv vk + Bw wk (9)
bles 2, 3 and 4. In presenting the survey results our yk = Cxk + k (10)
intention is to highlight the important features of each
algorithm; it is not our intent to determine the superi- An equivalent transfer function model in the form of
ority of one product versus another. The choice of an matrix fraction description(Kailath, 1980) can be writ-
appropriate MPC package for a particular application ten as:
is a complex question that must be answered on a case
by case basis; such issues are beyond the scope of this yk = [I ? y (q?1)]?1[u(q?1)uk + v (q?1)vk +
paper. w (q?1)wk ] + k (11)
We focus here on the main aspects of the con-
trol and identication technology. We fully understand where q?1 is a backward shift operator. The output
that a sound industrial oering must address many error identication approach (Ljung, 1987) minimizes
needs not necessarily related to the mathematics of the measurement error k , which results in nonlinear
the algorithms; these include software and hardware parameter estimation. Multiplying [I ? y (q?1 )] on
compatibility, user interface requirements, personnel both sides of the above equation results in an autore-
training, and conguration and maintenance issues. It gressive model with exogenous inputs (ARX),
should also be clear that the descriptions given here yk = y (q?1)yk + u(q?1)uk + v (q?1)vk +
are necessarily incomplete, since every MPC product
has proprietary features. With this understanding in w (q?1)wk + k (12)
mind, we rst discuss the process models at the core where
of MPC technology and then describe the details of a k = [I ? y (q?1 )]k (13)
typical MPC calculation. Subsequent sections describe
how dierent MPC vendors approach the dierent as- The equation error identication approach minimizes
pects of implementing MPC technology. k , which is colored noise even though the measure-
ment noise k is white. For a stable system, a Finite
Process Models Impulse Response (FIR) model can be derived as an
approximation to the transfer function model:
The mathematical form of the process model denes
the scope of an MPC algorithm. Tables 3 and 4 show
yk =
XN Huiuk?i + XN Hvivk?i +
u v
that a wide variety of model forms are used in industrial
MPC algorithms. All of the control and identication
algorithms described here use time-invariantmodels. A
XiN=1Hwi wk?i + k i=1
w
(14)
i=1
general nonlinear discrete-time state space model may
be described as This model form is used by the SMC-Idcom, HIECON,
and OPC algorithms. Typically the sample time is
xk+1 = f (xk; uk; vk; wk) (7) chosen so that from 30 to 120 coecients are required
10 MODEL PREDICTIVE CONTROL
to describe the full open loop response. An equivalent Table 3 summarizes how each of the MPC vendors has
velocity form is useful in identication: accomplished this translation.
XN u
Hui
uk?i +
XN v v
Figure 5 illustrates the ow of a representative
MPC calculation at each control execution. The rst

yk =
XNi=1w
Hw
wk?i +
k
H
v +
i=1 i k?i
(15)
step is to read the current values of process inputs
(DV's and MV's) and process outputs (CV's). In addi-
i=1 i tion to their numerical values, each measurement car-
An alternative model form is the nite step response ries with it a sensor status to indicate whether the sen-
model (FSR) (Cutler et al., 1983), given by: sor is functioning properly or not. Each MV will also
carry information on the status of the associated lower
Xk ui
k?i + Xk vi
k?i + level control function or valve; if saturated then the
yk =
Xki=1 wi
k?i + k i=1
S
S
u
w 
S
(16)
v MV will be permitted to move in one direction only. If
the MV controller is disabled then the MV cannot be
used for control but can be considered a measured dis-
i=1
turbance (DV). The following sections describe other
where S0 = 0 and Sj = SN for j > N . Note that aspects of the calculation in greater detail.
the summation goes from the initial time to the cur-
rent time k. The FSR model is used by the DMC and Output Feedback
RMPCT algorithms. The FIR model is related to the The model update step is where feedback enters the
FSR model through: loop. All of the industrial MPC algorithms surveyed
Hi = Si ? Si?1 (17) here use the same form of feedback for stable processes,
based on comparing the current measured process out-
The SMC-Idcom and RMPCT algorithms also pro- put ykm to the current predicted output yk :
vide the option to enter Laplace transfer function mod-
els. All of the algorithms allow control of processes bk = ykm ? yk (18)
with integrating dynamics, either by modeling the time
derivative of the output response or by using a modied The bias bk term is added to the model for use in
feedback procedure. subsequent predictions:
yk+j = g (xk+j ) + bk (19)
A General MPC Control Calculation
MPC controllers are designed to drive the process from This form of feedback is equivalent to assuming an
one constrained steady state to another. They may output disturbance that remains constant for all future
receive an optimal steady-state operating point from an time (Morari and Lee, 1991; Lee, Morari and Garca,
overlying optimizer, as shown in Figure 2, or they may 1994). Rawlings et al. (Rawlings et al., 1994) show that
compute an optimal operating point using an internal this method of feedback removes steady-state oset,
steady-state optimizer. The general objectives of an which provides theoretical support for its use.
MPC controller, in order of importance, are: Variations of this basic feedback approach are used
for the case of integrating dynamics. In RMPCT, for
1. prevent violation of input and output constraints example, both a model bias and the rate of change of
the bias are used for CV's that have integrating ele-
2. drive the CV's to their steady-state optimal values ments. DMC uses a rotation factor to combine a con-
(dynamic output optimization) tribution from the bias term with a contribution from
3. drive the MV's to their steady-state optimal val- the rate of change of the bias term.
ues using remaining degrees of freedom (dynamic Additional practical details of the bias term cal-
input optimization) culation should be noted. The bias calculation may
be ltered to remove high frequency noise; the RM-
4. prevent excessive movement of MV's PCT and SMC-Idcom algorithms provide this option.
In some cases the CV measurement may not be avail-
5. when signals and actuators fail, control as much able at each control execution; this may happen, for
of the plant as possible. example, when the CV measurement is provided by an
The translation of these objectives into a mathe- analyzer. In this case one can skip the bias update for
matical problem statement involves a number of ap- the aected CV for a number of control intervals. A
proximations and trade-os that dene the basic char- counter is provided to disable control of the CV if too
acter of the controller. Like any design problem there many executions go by without feedback. The DMC
are many possible solutions; it is no surprise that there and SMC-Idcom algorithms provide this feature.
are a number of dierent MPC control formulations.
 Industrial Model Predictive Control Technology 11
Determining the Controlled Sub-process The OPC algorithm does not distinguish between
Once the model has been updated the controller must critical and non-critical CV's; if a CV fails, its mea-
determine which MV's can be manipulated and which surement is replaced with a model estimate.
CV's should be controlled. In general, if the measure- In most MPC products, sensor faults are limited
ment status for a CV is good, and the operator has en- to complete failure that goes beyond pre-specied con-
abled control of the CV, then it should be controlled. trol limits. Sensor faults such as signicant bias and
An MV must meet the same criteria to be used for con- drifting that are within normal limits are generally not
trol; in addition, however, the lower level control func- detected or identied in these products.
tions must also be available for manipulation. If the Removal of Ill-conditioning
lower level controller is saturated high or low, one can
add a temporary hard MV constraint to the problem to At any particular control execution, the process en-
prevent moving the MV in the wrong direction. If the countered by the controller may require excessive in-
lower level control function is disabled, the MV cannot put movement in order to control the outputs inde-
be used for control. In this case it should be treated as pendently. This problem may arise, for example, if
a DV. From these decisions a controlled subprocess is two outputs respond in an almost identical way to the
dened at each control execution. In general the shape available inputs. Consider how dicult it would be to
of the subprocess changes in real-time as illustrated in independently control adjacent tray temperatures in a
Figure 4. distillation column, or to control both regenerator and
The RMPCT and DMC algorithms provide an ad- cyclone temperature in an FCCU. It is important to
ditional mechanism to prevent low level control satu- note that this is a feature of the process to be con-
ration by including the low level control outputs (e.g., trolled; any algorithm which attempts to control an
valve position) in the control formulation as additional ill-conditioned process must address this problem. For
CVs. These CV's are then forced to stay within high a process with gain matrix G, the condition number
and low saturation limits by treating them as range of GT G provides a measure of process ill-conditioning;
or zone control variables. In this conguration, the a high condition number means that small changes in
number of MV's is typically less than the number of the future error vector will lead to large MV moves.
CV's, which include both range CV's and setpoint Although the conditioning of the full control prob-
CV's. However, the number of setpoint CV's is typ- lem will almost certainly be checked at the design
ically less than or equal to the number of MV's. If a phase, it is nearly impossible to check all possible sub-
range CV is well within the saturation constraints, it processes which may be encountered during future op-
has no eect on the objective function. In this case, the eration. It is therefore important to examine the con-
control conguration reduces to a typical fat or square dition number of the controlled sub-process at each
plant. It should be noted that other CV's which need control execution and to remove ill-conditioning in the
not to be controlled tightly at a setpoint value (e.g., internal model if necessary. Three strategies are cur-
surge tank level) are also treated as range CV's. When rently used by MPC controllers to accomplish this; sin-
these CV's are within constraint limits, no MV action gular value thresholding, controlled variable ranking
is required for these CV's. This releases additional de- and input move suppression.
grees of freedom to drive remaining CV's or MV's to The Singular Value Thresholding (SVT) method
their targets. used by the RMPCT controller involves decomposing
The DMC, RMPCT, and SMC-Idcom algorithms the process model using a singular value decomposi-
distinguish between a critical CV failure and a non- tion. Singular values below a threshold magnitude are
critical CV failure. If a non-critical CV fails, the DMC discarded, and a process model with a much lower con-
controller completely removes it from the control calcu- dition number is then reassembled and used for control.
lation. The RMPCT and SMC-Idcom algorithms con- The neglected singular values represent the direction
tinue control action by setting the failed CV measure- along which the process hardly moves even if a large
ment to the model predicted value, which means there MV change is applied; the SVT method gives up this
is no feedback for the failed CV. If the non-critical CV direction to avoid erratic MV changes. This method
fails for a specied period of time, RMPCT drops this solves the ill-conditioning problem at the expense of ne-
CV from the control objective function. glecting the smallest singular values. If the magnitude
If a critical CV fails, the DMC and RMPCT con- of these singular values is small comparing to model un-
trollers turn o immediately. The SMC-Idcom algo- certainty, it may be better to neglect them anyway. Af-
rithm, however, tries to maintain control for the part ter thresholding, the collinear CV's are approximated
of the process that is not aected by the critical CV with the principal singular direction. In the case of two
failure. The SMC-Idcom algorithm also allows a sen- collinear CV's, for example, this principal direction is a
sor to be temporarily turned o for calibration without weighted average of the two CV's. Note that the SVT
interrupting control. approach is sensitive to output weighting. If one CV
12 MODEL PREDICTIVE CONTROL
is weighted much more heavily than another, this CV The LP is used primarily to enforce steady-state in-
will represent the principal singular direction and the put and output constraints and to determine optimal
results will be approximately equivalent to the priority steady-state input and output targets for the thin and
approach used in SMC-Idcom. fat plant cases of Figure 4. Input optimization in the
The SMC-Idcom algorithm addresses this issue fat plant case is accomplished using economic factors
using a user-dened set of CV controllability ranks. which describe the cost of using each input. In the thin
When a high condition number is detected, the con- plant case the output error trade-os are evaluated us-
troller drops low priority CV's until a well-conditioned ing slack variable weights.
sub-process remains. The sub-process will be con- The RMPCT and PFC algorithms provide an addi-
trolled without erratic input movement but the low tional level of exibility by allowing for both linear and
priority CV's will be uncontrolled. Note, however, that quadratic terms in the steady-state objective function.
if a low priority CV is dropped because it's open loop They also includes hard input and output constraints.
response is close to that of a high priority output, it will The SMC-Idcom algorithm solves the local steady-
follow the high priority CV and will therefore still be state optimization problem using a sequence of
controlled in a loose sense. In the case of two collinear quadratic programs. CV's are ranked by priority such
CV's having no dierentiable priority, it may be desir- that control performance of a given CV will never be
able to use an weighted average of the two. sacriced in order to improve performance of a lower
Controllers that use input move suppression, such priority CV. The prediction error can be spread across
as the DMC and OPC algorithms, provide an alter- a set of CV's by grouping them together at the same
native strategy for dealing with ill-conditioning. Input priority level. The calculation proceeds by optimizing
move suppression factors increase the magnitude of the the highest priority CV's rst, subject to hard and soft
diagonal elements of the matrix to be inverted in the output constraints on the same CV's and all input hard
least squares solution, directly lowering the condition constraints. Subsequent optimizations preserve the fu-
number. The move suppression values can be adjusted ture trajectory of high priority CV's through the use
to the point that erratic input movement is avoided of equality constraints. Likewise inputs can be ranked
for the commonly encountered sub-processes. In the in priority order so that inputs are moved sequentially
limit of innite move suppression the condition num- towards their optimal values when extra degrees of free-
ber becomes one for all sub-processes. There probably dom permit.
exists a set of nite move suppression factors which
guarantee that all sub-processes have a condition num- Dynamic Optimization Objectives
ber greater than a desired threshold value. In the case At the dynamic optimization level, an MPC controller
of two collinear CV's, the move suppression approach must compute a set of MV adjustments that will drive
gives up a little bit on moving each CV towards its tar- the process to an optimal steady-state operating point
get. The move suppression solution is similar to that of without violating constraints. All of the controllers dis-
the SVT solution in the sense that it tends to minimize cussed here can be described (approximately) as mini-
the norm of the MV moves. mizing the following dynamic objective function:
Local Steady-State Optimization
The DMC, SMC-Idcom, and RMPCT controllers split =
XP k yk+jk2Q + XM?1k
k+jk2S +
XjM=1?1k uk+jk2R j=0
J e j
u j
the control calculation into a local steady-state opti-
mization followed by a dynamic optimization. Opti- j =0 e j
(20)
mal steady-state targets are computed for each input
and output; these are then passed to a dynamic opti- subject to a model constraint:
mization to compute the optimal input vector. This
should not be confused with the more comprehensive xk+j = f (xk+j ?1 uk+j ?1) 8 = 1
; j ;P

nonlinear optimization that takes place above the MPC yk+j = g (xk+j ) + bk 8 =1
j ;P

algorithm in the plant control hierarchy (see Figure

2). The local steady-state optimization uses a linear and subject to inequality constraints:
steady-state model which may come from linearizing a yj  yk+j  yj 8 =1
comprehensive nonlinear model at each control execu- j ;P

tion or may simply be the steady-state version of the u  uk+j  u  8 =0 ?1

j ;M

linear dynamic model used in the dynamic optimiza-
u 
uk+j 
u 8 = 0 ? 1 j ;M

tion. The objective function 20 involves three con icting

The DMC controller uses an LP to do the local contributions. Future output behavior is controlled by
steady-state optimization. The optimization is carried penalizing deviations from a desired response, dened
out subject to hard input and soft output constraints. by eyk+j , over a prediction horizon of length . Future
P
 Industrial Model Predictive Control Technology 13
input behavior is controlled using input penalties de- This provides a similar eect to the move suppression
ned by euk+j over a control horizon of length . Rapid
M used in DMC.
input changes are penalized with a separate term in- Using a single objective function means that trade-
volving the moves
uk+j . The relative importance of os between the three contributions must be resolved
the objective function contributions are controlled by using the the relative weight matrices Qj , Sj , and Rj .
setting the time dependent weight matrices Qj , Sj , and The HIECON algorithm resolves con icting dynamic
Rj ; these are assumed to be positive semi-denite. control objectives by solving a sequence of multiple op-
The solution is a set of M input adjustments: timization problems. The decision is made, a priori,
that CV errors are more important than MV errors. A
uM = (uk uk+1
; uk+M ?1)
; : : :; (21) quadratic output optimization problem is solved rst,
similar to 20 but including only the eyk+j terms. For the
The rst input uk is injected into the plant and the thin and square plant cases this will provide a unique
calculation is repeated. solution and the calculation terminates. For the fat
The DMC, SMC-Idcom, RMPCT, PFC and OPC plant case there are remaining degrees of freedom that
algorithms use a single quadratic objective function can be used to optimize the input settings. For this
similar to 20. The DMC and OPC algorithms penalize case the controller solves a separate quadratic input
the last input in order to drive the system towards the optimization problem, similar to 20 but including only
optimal steady state: the euk+j terms. The input optimization includes a set
of equality constraints that preserve the future output
Rj = 0; 8 ?1 (22) trajectories found in the output optimization. This
j < M
eliminates the need to set weights to determine the
RM ?1 6 0
= trade-o between output and input errors, at the cost
If the nal input weight is large enough and the of additional computation.
process is stable, it is approximately equivalent to hav- Constraint Formulations
ing a terminal state constraint. If the dynamic so-
lution is signicantly dierent from the LP targets, There are basically two types of constraints used in
which means the terminal states are not eectively con- industrial MPC technology; hard and soft. These are
strained, the DMC controller will be turned o. This illustrated in Figure 10. Hard constraints, shown in
setting may provide nominal stability for DMC con- the top of Figure 10 are those which should never
troller; further analysis is necessary to rmly establish be violated. Soft constraints, shown in the middle
this result. The nal input weights are also applica- of Figure 10 are those for which some violation may
ble to integrating processes where the derivative of the be allowed; the violation is generally subjected to a
integrator is driven to zero. quadratic penalty in the objective function.
The SMC-Idcom, RMPCT, HIECON and PFC All of the MPC algorithms allow hard MV maxi-
controllers do not penalize input moves directly. The mum, minimum, and rate of change constraints to be
SMC-Idcom, HIECON and PFC algorithms use a pre- dened. These are generally dened so as to keep the
dened reference trajectory to avoid aggressive MV lower level MV controllers in a controllable range, and
moves. The RMPCT controller denes a funnel, which to prevent violent movement of the MV's at any single
will be described later in the paper, and nds the opti- control execution. The PFC algorithm also accommo-
mal trajectory and optimal MV moves by minimizing: dates maximum and minimum input acceleration con-
straints which are useful in mechanical servo control
(uM yr ) = arg min
XP r 2
applications. The SMC-Idcom, RMPCT, HIECON
;
j =1kyk+j ? yk+j kQ + and OPC algorithms perform rigorous optimizations
subject to the hard input constraints. The DMC
kuk+M ?1 ? ussk2R (23) and PFC algorithms, however, enforce input hard con-
subject to the funnel constraints. The relative pri- straints in a sub-optimal manner in the sense that the
ority of the two terms is set by the two weighting ma- solution does not generally satisfy the Karush-Kuhn-
trices. In the case that the rst term is completely Tucker (KKT) conditions. In the DMC algorithm,
satised, which is typical due the funnel formulation, when an input is predicted to violate a maximum or
the CV error will vanish and the minimization is in fact minimum limit it is set equal to the limit and the cal-
performed on the second term only. In this case the re- culation is repeated with the MV removed. The PFC
sults will be similar to having two separate objectives algorithm performs the calculation without constraints
on CV's and MV's. In the case of an innite number and then clips the input values if they exceed hard con-
of solutions, which is also typical due to "relaxing" the straints. Both of these techniques will prevent violation
trajectory, a minimum norm solution to the MV's is of hard input constraints but will, in general, involve a
obtained due to the use of singular value thresholding. loss of performance that is dicult to predict.
14 MODEL PREDICTIVE CONTROL
The SMC-Idcom, RMPCT, HIECON and PFC al- causes CV's which violate or almost violate constraints
gorithms enforce output soft constraints in an opti- to be driven towards the zone boundaries.
mal manner as part of the dynamic optimization. The The SMC-Idcom, HIECON, and PFC algorithms
DMC and OPC algorithms, however, use a setpoint provide a CV reference trajectory option. The refer-
approximation to enforce soft constraints. This is il- ence trajectory idea is to bring the CV up to the set-
lustrated at the bottom of Figure 10. Setpoints are point more slowly, in order to avoid overshoot. A rst
dened for each soft constraint, resulting in objective order curve is drawn from the current CV value to the
function penalties on both sides of the constraint. The setpoint, with the speed of the response determined by
output weight is adjusted dynamically, however, so a trajectory time constant. Future CV deviations from
that the weight becomes signicant only when the CV the reference trajectory are penalized. In the limit of
comes close to the constraint. When a violation is pre- a zero time constant the reference trajectory reverts
dicted the weight is increased to a large value so that back to a pure setpoint; for this case, however, the
the control can bring the CV back to its constraint controller would be sensitive to model mismatch un-
limit. As soon as the CV is within the constraint limit, less some other strategy such as move suppression is
the LP target is used as the setpoint instead. also being used. A drawback of the reference trajec-
The SMC-Idcom and RMPCT algorithms con- tory formulation is that it penalizes the output when
sider hard output constraints only in the local steady- it happens to drift too quickly towards the setpoint,
state optimization. This prevents problems which as might happen in response to an unmeasured dis-
may occur due to enforcing output constraints early turbance. If the CV moves too quickly due to model
in the prediction horizon of the dynamic optimization mismatch, however, the reference trajectory is bene-
(see the discussion in Zariou's paper, for example cial in that it will slow down the CV and minimize
(Zariou, 1990)). Only the HIECON and PFC algo- overshoot. The reference trajectory can be interpreted
rithms consider hard output constraints in the dynamic mathematically as a lter in the feedback path, similar
optimization. Hard output constraints are ranked in to the robustness lter recommended by IMC theory
order of priority so that low priority constraints can (Morari and Zariou, 1989). In general, as the refer-
be dropped when the problem becomes infeasible. The ence trajectory time constant increases, the controller
PFC algorithm also considers hard output constraints is able to tolerate larger model mismatch.
in the steady-state optimization. The RMPCT algorithm attempts to keep each CV
within a user dened zone, with setpoints dened by
Output and Input Trajectories setting the maximum and minimum zone limits equal
Industrial MPC controllers use four basic options to to each other. When the CV goes outside the zone,
specify future CV behavior; a setpoint, zone, reference the RMPCT algorithm denes a CV funnel, shown at
trajectory or funnel. These are illustrated in Figure the bottom of Figure 6, to bring the CV back within
6. The shaded areas correspond to the eyk+j and euk+j its range. The slope of the funnel is determined by a
terms in 20. All of the controllers provide the option to user dened performance ratio, dened as the desired
drive the CV's to a xed setpoint, with deviations on time to return to the zone divided by the open loop
both sides penalized in the objective function. In prac- response time. A weighted average open loop response
tice this type of specication is very aggressive and may time is used for multivariable systems.
lead to very large input adjustments, unless the con- The SMC-Idcom algorithm uses a variation of the
troller is detuned in some fashion. This is particularly funnel when a zone CV falls out of it's range. In this
important when the internal model diers signicantly case a funnel is dened using a one-sided reference tra-
from the process. The DMC and OPC algorithms use jectory that terminates just inside the zone boundary.
move suppression factors for this purpose. A potential advantage of the funnel is illustrated in
All of the controllers also provide a CV zone con- Figure 7. Consider the case of a reference trajectory,
trol option, designed to keep the CV within a zone shown on the left side of Figure 7. If a disturbance
dened by upper and lower boundaries. One way to causes the predicted future CV to reach the setpoint
implement zone control is to dene upper and lower more quickly than the reference trajectory allows, the
soft constraints. Other implementations are possible, controller will take action to bring the CV back down
however. The DMC algorithm, for example, uses a dy- to the dened trajectory. In the same situation the
namic weighting strategy to implement the zone using funnel shown on the right side of Figure 7 would take
objective function penalties. When the CV is predicted no action at all.
to lie within its zone, its weight is set to zero so the All of the MPC algorithms surveyed here provide
controller will ignore it. If the CV is near one edge of MV setpoints to drive the inputs towards their opti-
the zone, its weight increases gradually depending on mal values when there are sucient degrees of freedom.
how close it is to the constraint. If the CV violates The SMC-Idcom algorithm also provides an option to
the constraint, the weight is set to a large value. This ramp the MV along a straight line to it's optimal value.
 Industrial Model Predictive Control Technology 15
Output Horizon and Input Parameterization 1. From the stated control objectives, dene the size
Industrial MPC controllers generally evaluate future of the problem, and determine the relevant CV's,
CV behavior over a nite set of future time intervals MV's, and DV's
called the prediction horizon. This is illustrated at the 2. Test the plant systematically by varying MV's and
top of Figure 8. The nite output horizon formulation DV's; capture and store the real-time data show-
is used by all of the algorithms discussed in this paper. ing how the CV's respond
The length of the horizon P is a basic tuning parameter
for these controllers, and is generally set long enough to 3. Derive a dynamic model from the plant test data
capture the steady-state eects of all computed future using an identication package
MV moves. 4. Congure the MPC controller and enter initial
The SMC-Idcom and PFC controllers allow the op- tuning parameters
tion to simplify the calculation by considering only a
subset of points in the prediction horizon. This concept 5. Test the controller o-line using closed loop simu-
is illustrated at the bottom of Figure 8. These points lation to verify the controller performance.
are called coincidence points because the desired and
predicted future outputs are required to coincide at 6. Download the congured controller to the desti-
these points. A separate set of coincidence points can nation machine and test the model predictions in
be dened for each output, which is useful when one open-loop mode
output responds quickly relative to another. The full 7. Commission the controller and rene the tuning
nite output horizon can be selected as a special case. as needed.
Industrial MPC controllers use three dierent
methods to parameterize the MV prole; these are il- All of the MPC packages surveyed here provide
lustrated in Figure 9. The DMC, RMPCT, and OPC software tools to help with the control design, process
algorithms compute a set of future moves to be spread identication and closed loop simulation steps. A sig-
over a nite control horizon, as shown at the top of Fig- nicant amount of time is currently spent at the closed
ure 9. The length of the control horizon M is another loop simulation step to verify acceptable performance
basic tuning parameter for these controllers. Increased of the control system. Typically tests are performed to
performance is obtained as M increases, at the expense check the regulatory and servo response of each CV,
of additional computation. and system response to violations of major constraints
The SMC-Idcom and HIECON algorithms com- is veried. The nal tuning is then tested for sensi-
pute a single future input move, as shown in the middle tivity to model mismatch by varying the gain and dy-
of Figure 9. This greatly simplies the calculation for namics of key process models. However, even the most
these algorithms, which is helpful because they solve thorough simulation testing usually cannot exhaust all
a series of optimization problems at each control ex- possible scenarios.
ecution. The use of a single move involves a sacrice Controller tuning is always required not only for
of closed loop performance that is dicult to quantify, stability, but also for trade-os between performance
however. and robustness. It is true that a high performance
The PFC controller parameterizes the input func- controller will not be robust with respect to model
tion using a set of polynomial basis functions. A pos- mismatch, but a low performance controller is not nec-
sible solution is illustrated at the bottom of Figure essarily robust. Most MPC products provide tuning
9. This allows a relatively complex input prole to "knobs" to de-tune the controller. In DMC and OPC,
be specied over a large (potentially innite) control two types of tuning parameters are used: (1) move sup-
horizon, using a small number of unknown parameters. pression factors, which are weights on
u and (2) equal
This may provide an advantage when controlling non- concern error factors, which are the inverse of output
linear systems. Choosing the family of basis functions weights. The move suppression factors change the ag-
establishes many of the features of the computed input gressiveness of the controller, while the equal concern
prole; this is one way to ensure a smooth input sig- error factors normalize the importance of each CV.
nal, for example. If a polynomial basis is chosen then The SMC-Idcom, HIECON, and PFC controllers
the order can be selected so as to follow a polynomial use the time constant of the reference trajectory as the
setpoint signal with no lag. This feature is important main tuning parameter. Smaller time constants de-
for mechanical servo control applications. mand more aggressive control, while larger time con-
stants result in less aggressive action. One may start
Control Design and Tuning with the trajectory time constant equal to the open
The MPC control design and tuning procedure is gener- loop time constant of the CV, then rene the tuning
ally described as follows (DMC Corp., 1994; Honeywell, based on performance/robustness trade-os.
Inc., 1995; Setpoint, Inc., 1993):
16 MODEL PREDICTIVE CONTROL
A performance ratio is provided in RMPCT which sition works. A few packages such as DMI and SMC-
is dened as the ratio of the required closed loop set- Test allow several MV's to change simultaneously with
tling time to the weighted-average open loop settling uncorrelated signals for dierent MV's. For DMI, the
time. The performance ratio is used to determine the plant test is run 24 hours a day with engineers mon-
length of the funnel, which is somewhat similar to the itoring the plant. Each MV is stepped 8 to 15 times,
settling time of a setpoint trajectory. A performance with the output (CV) signal to noise ratio at least six.
ratio equal to one means that the closed loop settling The plant test may take up to 10 to 15 days, depend-
time is equal to the open loop settling time. A perfor- ing on the time to steady state and number of variables
mance ratio less than one results in a more aggressive of the unit. Two requirements are imposed during the
controller. Only one tuning parameter per CV needs test: (i) no PID conguration or tuning changes are
to be specied. Independent tuning is available in RM- allowed; and (ii) operators may intervene during the
PCT for feedforward control, which allows the user to test to avoid critical situations, but no synchronizing
achieve faster response in feedforward control than in or correlated moves are allowed. One may merge data
setpoint tracking. from multiple test periods, which allows the user to
The RMPCT package provides a min-max design cut out a period of data which may be corrupted with
procedure in which the user enters estimates of model disturbances.
uncertainty directly. Tuning parameters are computed If the lower level PID control tuning changes sig-
to optimize performance for the worst case model mis- nicantly then it may be necessary to construct a new
match. Robustness checks for the remaining MPC con- process model. A model is identied between the input
trollers are performed by closed loop simulation. and output, and this is combined by discrete convolu-
All of the MPC packages provide a way to test - tion with the new input setpoint to input model. The
nal controller performance by closed loop simulation. SMC-Model package provides a convenient interface for
It is particularly important to check the response of such calculations.
the nal controller with respect to key constraint vi- It appears that PRBS or PRBS-like stepping sig-
olations. It is almost impossible, however, to test all nals are the primary test signals used by the identi-
possible situations of active constraints for a realistic cation packages. The GLIDE package uses a binary
problem. The problem becomes even more intractable signal in which the step lengths are optimized in a ded-
if one wishes to test performance in the presence of icated way. Others use a step test or pulse test in com-
model mismatch. This is one place where academic re- bination with PRBS (e.g., OPC, SMC-Test, and RM-
search can help industrial MPC practice signicantly. PCT). As a special case, OPC allows one to use oper-
Rawlings and Muske (Rawlings and Muske, 1993), for ating data as the ultimate source to build steady-state
example, have shown that their innite horizon MPC models for quality variables (Lines, Hartlen, Paquin,
algorithm is nominally stabilizing in the presence of Treiber, de Tremblay and Bell, 1993). These variables
constraints if the initial optimization problem is feasi- may be measured by an analytical sensor or laboratory
ble. Additional work is needed to extend this result to analysis. The following steady-state model is used for
the case of an imperfect plant model. Vuthandam et al. analyzer variables:
(Vuthandam, Genceli and Nikolaou, 1995) have done
this already for a modied QDMC algorithm; their re-
sults should be useful to many industrial practitioners. yk = yk?1 + K(uk ? uk?1) + Kf (ykm ? yk?d) (24)
Identication Test Protocol where K is the steady-state gain matrix identied from
the operating data. The vector ykm represents current
Test signals are required to excite both steady-state analyzer measurements which have a time delay repre-
(low frequency) and dynamic (medium to high fre- sented by d. Kf is an adjustable feedback gain matrix
quency) dynamics of a process. A process model is then to correct the model estimation. The entire relation
identied from the process input-output data. DMCC may be thought of as a steady-state Kalman lter. The
believes that the plant test is the single most impor- same mechanism can be applied to dynamic models as
tant phase in the implementation of DMC controllers. well.
To prepare for a formal plant test, a pre-test is usually
necessary for three reasons: (i) to step each MV and Model Forms and Parameter Estimation
adjust existing instruments and PID controllers; (ii) to
obtain the time to steady state for each CV; and (iii) The model forms used in identication are generally
to obtain data for initial identication. more diversied than those used for control. Table 4
Most identication packages test one (or at most lists various model forms used in dierent MPC prod-
several) manipulated variables at a time and x other ucts. The identication approaches in the MPC prod-
variables at their steady state. This approach is valid ucts are mainly based on minimizing the following least
as long as the process is assumed linear and superpo-
 Industrial Model Predictive Control Technology 17
squares criterion, options to identify FIR and ARMAX models directly
J =
XL kyk ? ykmk
2
(25)
from data. The ARMAX model is used on a single-
input-single-output basis. Note that OPC estimates
k=1 the noise dynamics by using moving average terms in
the model. These models are nally converted into FIR
using either an equation error approach or an output models for control.
error approach (Ljung, 1987). The major dierence It appears that dierent philosophies exist in
between the equation error approach and the output choosing the model forms for identication. DMC be-
error approach appears in identifying ARX or transfer lieves that complex dynamics can only be identied
function models. In the equation error approach, past with high-order FIR models and worries little about
output measurements are fed back to the model in Eqn. overparametrization. A few other products are more
12, concerned with overparametrization which may induce
unrealistic variance, and use low order transfer func-
yk = y (q?1)ykm + u (q?1 )uk + v (q?1)vk (26) tion models as an alternative.
Model uncertainty bounds are provided in several
while in the output error approach, the past model products such as OPC, RMPCT, and SMC-Model.
output estimates are fed back to the model, In GLIDE, continuous transfer function models are
identied directly by using gradient descent or Gauss-
Newton approaches. Then model uncertainty is iden-
yk = y (q?1 )yk + u (q?1 )uk + v (q?1)vk (27) tied by a global method, which nds a region in the
The equation error approach formulates a linear parameter space where the tting criterion is less than
least squares problem, but the estimates are biased a given value. This given value must be larger than
even though the measurement noise  in Eqn. 11 is the minimum of the criterion in order to nd a feasible
white. The output error approach is unbiased given region.
white measurement noise. However, the ARX model Most products apply nonlinear transformations to
parameters appear nonlinearly in the model, which re- variables that exhibit signicant nonlinearity. For ex-
quires nonlinear parameter estimation. One may also ample, a logarithm transformation is often performed
see that the equation error approach is a one-step ahead on composition variables.
prediction approach with reference to ykm , while the
output error approach is a long range prediction ap- Applications Summary
proach since it does not use ykm .
Using FIR models results in a linear-in-parameter Table 2 summarizes the reported applications experi-
model and an output error approach, but the estima- ence of each MPC vendor surveyed. Note that this is a
tion variance may be in ated due to possible over- count of completed MPC applications reported by each
parametrization. In DMI, a least squares method is vendor; for a particular problem one vendor may use
used to estimate FIR model parameters in velocity a single large controller while another may use several
form (Eqn. 15). The advantage of using the veloc- smaller controllers. In some cases a single controller is
ity form is to reduce the eect of a step-like unmea- designed and then subsequently used in thousands of
sured disturbance (Cutler and Yocum, 1991). However, copies; this may happen with an automobile applica-
the velocity form is sensitive to high frequency noise. tion, for example. Note also that this is a count of MPC
Therefore, pre-smoothing for the process data is done applications performed by the vendors themselves; this
in DMI before tting models to the data. The FIR co- does not include in-house applications performed by li-
ecients are then converted into FSR coecients for censees of vendor technology. Vendors were given wide
control. latitude to report the numbers in any distribution that
RMPCT adopts a three-step approach: (i) iden- they wished to use.
tify an FIR model using least squares; (ii) t the FIR The total number of reported MPC applications is
coecients to a low-order ARX model to smooth out currently over 2200. All of the vendors report a consid-
induced variance due to possible overparametrization erable number of applications in progress so it is likely
in the FIR model. The output error approach is used that this number will continue to increase rapidly. Ta-
to t the ARX model and a Gauss-Newton method is ble 2 shows that MPC technology can now be found
used to estimate the parameters; and (iii) convert the in a wide variety of application areas. The majority of
ARX models into Laplace transfer functions. When the applications (67 %) are in rening, one of the original
model is used in control, the transfer function models application areas, where MPC technology has a solid
are discretized into FSR models based on a given sam- track record of success. A signicant number of appli-
pling interval. The advantage of this approach is that cations can also be found in petrochemicals and chem-
one has the exibility to choose dierent sampling in- icals, although it has taken longer for MPC technology
tervals than that used in data collection. OPC provides to break into these areas. Signicant growth areas in-
18 MODEL PREDICTIVE CONTROL
clude the chemicals, pulp and paper, food processing, of a noisy output signal into the feedback path. It is
aerospace and automotive industries. not possible, in general, to represent an unstable pro-
The DMC Corporation reports the largest total cess using an impulse response model. All of these
number of applications, at 600, with the other four problems can be overcome by using an auto-regressive
vendors reporting between 350 and 450 applications parametric model form such as an ARX or state-space
each. Table 2 shows that DMC, Setpoint, Honeywell model.
and Treiber Controls are highly focused in rening and Sub-optimal solutions to the dynamic optimization
petrochemicals, with a handful of applications in other 20 are used in several of the packages, presumably in
areas. Adersa appears to have the broadest range of order to speed up the solution time. This seems di-
experience with applications in the food processing, cult to justify for the case of rening and petrochemical
aerospace and automotive areas, among others. applications, where the controllers run on the order of
The bottom of Table 2 shows that largest appli- once each minute, unless it can be shown that the sub-
cations to date by each vendor, in the form of (out- optimal solution is always very nearly optimal. For
puts)x(inputs). The numbers show a dierence in phi- high speed applications where the controller must exe-
losophy that is a matter of some controversy. The cute in a few milliseconds, such as tracking the position
DMC Corporation recommends solving a large con- of a missile, it may not be feasible to solve a QP at ev-
trol problem with a single controller whenever possible; ery control execution. For this case a good sub-optimal
they report an olens application with 603 outputs and solution may be the only option.
283 inputs. Other vendors prefer to break the problem The bias update feedback technique used by indus-
up into meaningful sub-processes. This is an issue that trial MPC controllers can be interpreted as assuming
needs further attention from the academic community. an output disturbance that remains constant for all
future time (Morari and Lee, 1991; Muske and Rawl-
Limitations of Existing Technology ings, 1993). This is probably the best assumption that
can be used in the absence of detailed disturbance in-
Morari and Lee pointed out several problems with ex- formation, but better feedback is possible if the distri-
isting technology in their CPC-IV paper (Morari and bution of disturbances can be characterized more care-
Lee, 1991). Muske and Rawlings have also pointed fully. Morari and Lee have shown how to extend MPC
out limitations of existing industrial MPC technology technology to achieve better feedback while still retain-
(Muske and Rawlings, 1993). These include: ing the step response model (Morari and Lee, 1991).
Muske and Rawlings have demonstrated how better
 impulse and step response models are over- performance can be achieved by using a state space
parameterized and limit application of the algo- model and an optimal state observer (Muske and Rawl-
rithm to strictly stable processes ings, 1993).
 sub-optimal solution of the dynamic optimization Tuning MPC controllers for stable operation in the
presence of constraints may be dicult, even when the
 sub-optimal feedback (constant output distur- process model is perfect. This is why so much eort
bance assumption) is spent on closed loop simulation prior to commis-
sioning a controller. Rawlings and Muske addressed
 tuning is required to achieve nominal stability this issue directly in the development of their innite
 model uncertainty is not addressed adequately horizon MPC algorithm (Rawlings and Muske, 1993).
They used the LQR controller of Kalman (Eqn. 2)
Impulse and step response models are known to as a starting point. The Kalman LQR is stable for
be over-parameterized. The dynamics of a rst order any reasonable choice of tuning parameters, due to the
process, for example, can be described by three num- use of an innite prediction horizon. They developed
bers using a parametric model (gain, time constant and an innite horizon MPC controller that is guaranteed
deadtime). An impulse response will typically require to be closed loop stable, for the nominal case, in the
from 30 to 120 coecients to describe the same dynam- presence of constraints. Feasibility of the initial QP is
ics. These diculties can be overcome at the identica- enough to guarantee constrained stabilizability. When
tion step by rst identifying a low order model and then the QP is infeasible, Muske and Rawlings propose giv-
calculating the impulse response coecients. However ing up on all of the constraints for a short period of
the impulse response model still requires more storage time, much like the constraint window idea used by the
space than is necessary. A potentially more signicant QDMC algorithm (Garca and Morshedi, 1986). Other
problem with the impulse and step response models is options for recovering feasibility are also possible, such
that they are limited to strictly stable processes. While as dropping low priority constraints. Zheng and Morari
it is certainly possible to modify the algorithms to ac- (Zheng and Morari, 1995) recently analyzed an in-
commodate a pure integrator, these modications may nite horizon MPC algorithm which uses soft output
lead to other problems, such as adding the derivative constraints to avoid infeasibility. They show that the
 Industrial Model Predictive Control Technology 19
closed loop system using output feedback is stable for extrapolate well.
a strictly stable plant. An alternative approach would be to use rst-
Model uncertainty is not addressed adequately by principles models developed from well known mass,
current MPC technology. While most of the identica- momentum, and energy conservation laws. However,
tion packages provide estimates of model uncertainty, the cost of developing a reasonably accurate rst-
only one vendor (Honeywell) provides a convenient way principles model is likely to be prohibitive until new
to use this information in the control design. All of the software tools and validation procedures become avail-
MPC algorithms provide a way to detune the control able. Hybrid models that integrate steady state non-
to improve robustness, but the trade-o between per- linear rst-principles models with dynamic empirical
formance and robustness is generally not very clear. models (linear or nonlinear) may prove most promis-
Until this connection is made, it will not be possible ing for the near future. Gain-scheduling with linear
to determine when a model is accurate enough for a dynamic models is an example of this approach.
particular control application. This is one area where From a theoretical point of view using a nonlin-
academic research can help. Vuthandam et al. for ear model changes the control problem from a con-
example, have recently presented robust stability con- vex QP to a non-convex Non-Linear Program (NLP),
ditions for a modied QDMC algorithm (Vuthandam the solution of which is much more dicult. There
et al., 1995). More research is needed in this area. is no guarantee, for example, that the global opti-
mum can be found. Bequette describes several ap-
proaches to solving the general nonlinear MPC problem
Next-Generation MPC Technology
in his review of nonlinear control for chemical processes
MPC vendors were asked to describe their vision of (Bequette, 1991). Although solving the nonlinear MPC
next-generation MPC technology. Their responses problem at each time step is much more dicult, Rawl-
were combined with our own views and the earlier anal- ings et al. (Rawlings et al., 1994) have shown that the
ysis of Froisy (Froisy, 1994) to come up with a compos- nominal Lyapunov stability argument presented for lin-
ite view of future MPC technology. ear models carries over to the general nonlinear case
with minor modications. The most important change
Basic Controller Formulation is that a constraint must be included to zero the states
Because it is so dicult to express all of the relevant at the end of the input horizon. They point out, how-
control objectives in a single objective function, next- ever, that nonlinear MPC may require unexpected in-
generation MPC technology will probably utilize multi- put adjustments. They present an interesting example
ple objective functions. The innite prediction horizon in which a simple nonlinearity in the process model
has benecial theoretical properties and will probably leads to a discontinuous feedback control law. This
become a standard feature. Output and input trajec- implies that tuning nonlinear MPC controllers may be
tory options will include setpoints, zones, trajectories very dicult, particularly for the case of model mis-
and funnels. Input horizons will include options for match.
multiple moves or parameterization using basis func- Adaptive MPC
tions.
A few adaptive MPC algorithms such as the GPC al-
Nonlinear MPC gorithm introduced by Clarke et al. have been pro-
MPC using nonlinear models is likely to become more posed (Clarke, Mohtadi and Tus, 1987) but only a
common as users demand higher performance and new single adaptive MPC algorithm has reached the mar-
software tools make nonlinear models more readily ketplace (STAR from Dot Products (Dollar, Melton,
available. Developing adequate nonlinear empirical Morshedi, Glasgow and Repsher, 1993)). This is de-
models may be very challenging, however. Test sig- spite the strong market incentive for a self-tuning MPC
nals such as PRBS that are adequate for linear mod- controller. This re ects the diculty of doing adaptive
els are not likely to provide adequate excitation of control in the real world. Barring a theoretical break-
nonlinear systems. Also there is no model form that through, this situation is not likely to change in the
is clearly suitable to represent general nonlinear pro- near future.
cesses. Froisy (Froisy, 1994) points out that second or- On the other hand, adaptive and on-demand tun-
der Volterra models may bridge the gap between linear ing PID controllers have been very successful in the
empirical models and nonlinear fundamental models in marketplace. This suggests that adaptive MPC con-
the near future. Genceli and Nikolaou, for example, trollers may emerge for SISO loops as adaptive PID
have studied the use of second order Volterra series technology is generalized to handle more dicult dy-
with a modied QDMC controller (Genceli and Niko- namics.
laou, 1995). However, nonlinear empirical models such
as Volterra series or neural networks do not seem to
20 MODEL PREDICTIVE CONTROL
Robust Stability of MPC els identied from test data. The impact of identica-
With one exception (Honeywell), industrial MPC con- tion theory on process modeling is perhaps compara-
trollers rely solely on brute-force simulation to eval- ble to the impact of optimal control theory on model
uate the eects of model mismatch. Robust stabil- predictive control. It is probably safe to say that MPC
ity guarantees would signicantly reduce the time re- practice is one of the largest application areas of system
quired to tune and test industrial MPC algorithms. identication. The current success of MPC technology
It is likely that the powerful theoretical results re- may be due to carefully designed plant tests.
cently presented for MPC with a perfect model (Muske Another observation is that process identication
and Rawlings, 1993; Zheng and Morari, 1995) will be and control design are clearly separated in current
extended to include model mismatch in the near fu- MPC technology. Eorts towards integrating identi-
ture. This has already been accomplished for a modi- cation and control design may bring signicant benets
ed QDMC algorithm using an impulse response model to industrial practice. For example, uncertainty esti-
(Vuthandam et al., 1995). New robust stability guaran- mates from process identication could be used more
tees will then be combined with uncertainty estimates directly in robust control design. Ill-conditioned pro-
from identication software to greatly simplify design cess structures could be re ected in the identied mod-
and tuning of MPC controllers. els and also used in control design.
Choosing an MPC technology for a given appli-
Robust Identication Schemes cation is a complex question involving issues not ad-
dressed in this paper. It is the opinion of the authors
The use of FIR models results in overparametrization. that for most applications, a knowledgeable control en-
In practice this leads to some kind of engineering mod- gineer could probably achieve acceptable control per-
ication, such as a smoothness factor to "regularize" formance using any of the packages discussed here, al-
the model parameters. Although these modications though the time and eort required may dier. If the
are proprietary, regularization (smoothness) and bi- process is nonlinear or unstable, or needs to track a
ased regression algorithms (ridge regression or partial complex setpoint trajectory with no oset, the PFC
least squares) are common approaches to dealing with algorithm may oer signicant advantages. If a vendor
overparametrization. An alternative approach would is to be selected to design and implement the control
be to identify the process model based on paramet- system, it would be wise to weigh heavily their experi-
ric models, such as transfer function models or state ence with the particular process in question.
space representation, then convert the model into de- Research needs as perceived by industry are mostly
sired form for the controller to use. control engineering issues, not algorithm issues. Indus-
trial practitioners do not perceive closed loop stability,
Conclusions for example, to be a serious problem. Their problems
are more like: Which variables should be used for con-
MPC technology has progressed steadily in the twenty trol? When is a model good enough to stop the identi-
two years since the rst IDCOM and DMC applica- cation plant test? How do you determine the source
tions. Survey data reveal approximately 2200 applica- of a problem when a controller is performing poorly?
tions to date, with a solid foundation in rening and When can the added expense of an MPC controller be
petrochemicals, and signicant penetration into a wide justied? How do you design a control system for an
range of application areas from chemicals to food pro- entire plant? How do you estimate the benets of a
cessing. control system? Answering these questions could pro-
Current generation MPC technology oers signif- vide control practitioners and theoreticians with plenty
icant new capabilities but the controllers still retain, of work in the foreseeable future.
for the most part, an IDCOM-like or a DMC-like per- Several technical advances have not yet been in-
sonality. The SMC-Idcom and HIECON algorithms are corporated into industrial MPC technology. These in-
IDCOM-like controllers which have evolved to use mul- clude using an innite prediction horizon to guarantee
tiple objective functions and ranked constraints. The nominal closed loop stability, and using linear estima-
DMC, RMPCT and OPC algorithms are DMC-like tion theory to improve output feedback. In addition,
controllers that use a single dynamic objective func- robust stability conditions have been developed for a
tion to evaluate control and economic trade-os using modied QDMC algorithm. It would seem that the
weighting factors. The PFC controller inherits some company which rst implements these advances will
of the IDCOM personality but is signicantly dier- have a signicant marketing and technical advantage.
ent in that it can accommodate nonlinear and unstable The future of MPC technology is bright, with all
processes and uses basis functions to parameterize the of the vendors surveyed here reporting signicant ap-
input function. plications in progress. Next-generation MPC technol-
An important observation is that industrial MPC ogy is likely to include multiple objective functions, an
controllers almost always use empirical dynamic mod- innite prediction horizon, nonlinear process models,
 Industrial Model Predictive Control Technology 21
better use of model uncertainty estimates, and better Genceli, H. and Nikolaou, M. (1995). Design of robust con-
handling of ill-conditioning. strained model-predictive controllers with volterra
series, AIChE J. 41(9): 2098{2107.
Grosdidier, P., Froisy, B. and Hammann, M. (1988). The
Acknowledgments IDCOM-M controller, in T. J. McAvoy, Y. Arkun
and E. Zariou (eds), Proceedings of the 1988
The authors wish to thank the vendor representatives IFAC Workshop on Model Based Process Control,
who responded to our MPC survey for their time and Pergamon Press, Oxford, pp. 31{36.
eort. We also wish to thank one anonymous reviewer Honeywell, Inc. (1995). RMPCT Concepts Reference.
and the following people for their constructive criticism Product literature from Honeywell, Inc.
of the initial draft: Jim Downs, Bjarne Foss, Brian Kailath, T. (1980). Linear Systems, Prentice-Hall, Engle-
Froisy, Pierre Grosdidier, Rob Hawkins, Evelio Her- wood Clis, New Jersey.
nandez, Joseph Lu, Jim Rawlings, Jaques Richalet and Kalman, R. E. (1960a). Contributions to the theory of
Steve Treiber. optimal control, Bull. Soc. Math. Mex. 5: 102{119.
Kalman, R. E. (1960b). A new approach to linear ltering
and prediction problems, Trans. ASME, J. Basic
References Engineeering pp. 35{45.
Lee, E. B. and Markus, L. (1967). Foundations of Optimal
Badgwell, T. A. (1995). Robust stability conditions Control Theory, John Wiley and Sons, New York.
for a SISO model predictive control algorithm. Lee, J. H. (1996). Recent advances in model predictive
Manuscript in preparation. control and other related areas, Chemical Process
Bequette, B. W. (1991). Nonlinear control of chemical pro- Control|CPC V, CACHE, Tahoe City, Califor-
cesses: A review, Ind. Eng. Chem. Res. 30: 1391{ nia.
1413. Lee, J. H., Morari, M. and Garca, C. E. (1994). State-
Clarke, D. W., Mohtadi, C. and Tus, P. S. (1987). Gener- space interpretation of model predictive control,
alized predictive control|Part I. The basic algo- Automatica 30(4): 707{717.
rithm, Automatica 23(2): 137{148. Lines, B., Hartlen, D., Paquin, F., Treiber, S., de Trem-
Cutler, C., Morshedi, A. and Haydel, J. (1983). An in- blay, M. and Bell, M. (1993). Polyethylene reactor
dustrial perspective on advanced control, AIChE modeling and control design, Hydrocarbon Process.
Annual Meeting, Washington, D.C. pp. 119{124.
Cutler, C. R. and Ramaker, B. L. (1979). Dynamic matrix Ljung, L. (1987). System Identication: Theory for the
control|a computer control algorithm, AIChE User, Prentice-Hall, Inc., Englewood Clis, New
National Meeting, Houston, TX. Jersey.
Cutler, C. R. and Ramaker, B. L. (1980). Dynamic matrix Mayne, D. Q. (1996). Nonlinear model predictive con-
control|a computer control algorithm, Proceed- trol: An assessment, Chemical Process Control|
ings of the Joint Automatic Control Conference. CPC V, CACHE, Tahoe City, California.
Cutler, C. R. and Yocum, F. H. (1991). Experience with Morari, M. and Lee, J. H. (1991). Model predictive
the DMC inverse for identication, in Y. Arkun control: The good, the bad, and the ugly, in
and W. H. Ray (eds), Chemical Process Control| Y. Arkun and W. H. Ray (eds), Chemical Process
CPC IV, Fourth International Conference on Control|CPC IV, Fourth International Confer-
Chemical Process Control, Elsevier, Amsterdam, ence on Chemical Process Control, Elsevier, Am-
pp. 297{317. sterdam, pp. 419{444.
DMC Corp. (1994). [DMC]TM: Technology Overview. Morari, M. and Zariou, E. (1989). Robust Process Control,
Product literature from DMC Corp. Prentice-Hall, Englewood Clis, New Jersey.
Dollar, R., Melton, L. L., Morshedi, A. M., Glasgow, D. T. Muske, K. R. and Rawlings, J. B. (1993). Model predictive
and Repsher, K. W. (1993). Consider adaptive control with linear models, AIChE J. 39(2): 262{
multivariable predictive controllers, Hydrocarbon 287.
Process. pp. 109{112. Ohshima, M., Ohno, H. and Hashimoto, I. (1995). Model
Froisy, J. B. (1994). Model predictive control: Past, present predictive control: Experiences in the university-
and future, ISA Trans. 33: 235{243. industry joint projects and statistics on mpc appli-
Froisy, J. B. and Matsko, T. (1990). Idcom-m application cations in japan, International Workshop on Pre-
to the shell fundamental control problem, AIChE dictive and Receding Horizon Control, Korea.
Annual Meeting. Prett, D. M. and Garca, C. E. (1988). Fundamental Process
Garca, C. E. and Morari, M. (1982). Internal model con- Control, Butterworths, Boston.
trol. 1. A unifying review and some new results, Prett, D. M. and Gillette, R. D. (1980). Optimization and
Ind. Eng. Chem. Proc. Des. Dev. 21: 308{323. constrained multivariable control of a catalytic
Garca, C. E. and Morshedi, A. M. (1986). Quadratic cracking unit, Proceedings of the Joint Automatic
programming solution of dynamic matrix control Control Conference.
(QDMC), Chem. Eng. Commun. 46: 73{87. Propoi, A. I. (1963). Use of linear programming methods
Garca, C. E., Prett, D. M. and Morari, M. (1989). Model for synthesizing sampled-data automatic systems,
predictive control: Theory and practice|a survey, Automn. Remote Control 24(7): 837{844.
Automatica 25(3): 335{348. Rawlings, J. B., Meadows, E. S. and Muske, K. R. (1994).
Genceli, H. and Nikolaou, M. (1993). Robust stability anal- Nonlinear model predictive control: A tutorial and
ysis of constrained l1 {norm model predictive con- survey, ADCHEM '94 Proceedings, Kyoto, Japan.
trol, AIChE J. 39(12): 1954{1965.
22 MODEL PREDICTIVE CONTROL

Rawlings, J. B. and Muske, K. R. (1993). Stability of

constrained receding horizon control, IEEE Trans.
Auto. Cont. 38(10): 1512{1516.
Richalet, J., Rault, A., Testud, J. L. and Papon, J. (1976).
Algorithmic control of industrial processes, Pro-
ceedings of the 4th IFAC Symposium on Identica-
tion and System Parameter Estimation, pp. 1119{
1167.
Richalet, J., Rault, A., Testud, J. L. and Papon, J. (1978).
Model predictive heuristic control: Applications
to industrial processes, Automatica 14: 413{428.
Ricker, N. L. (1991). Model predictive control: State of the
art, in Y. Arkun and W. H. Ray (eds), Chemical
Process Control|CPC IV, Fourth International
Conference on Chemical Process Control, Elsevier,
Amsterdam, pp. 271{296.
Ricker, N. L., Subrahmanian, T. and Sim, T. (1988). Case
studies of model-predictive control in pulp and pa-
per production, in T. J. McAvoy, Y. Arkun and
E. Zariou (eds), Proceedings of the 1988 IFAC
Workshop on Model Based Process Control, Perg-
amon Press, Oxford, pp. 13{22.
Setpoint, Inc. (1993). SMC-Idcom: A State-of-the-Art Mul-
tivariable Predictive Controller. Product literature
from Setpoint, Inc.
Vuthandam, P., Genceli, H. and Nikolaou, M. (1995). Per-
formance bounds for robust quadratic dynamic
matrix control with end condition, AIChE J.
41(9): 2083{2097.
Zariou, E. (1990). Robust model predictive control of pro-
cesses with hard constraints, Comput. Chem. Eng.
14(4/5): 359{371.
Zariou, E. and Chiou, H.-W. (1993). Output constraint
softening for SISO model predictive control, Pro-
ceedings of the 1993 American Control Conference,
pp. 372{376.
Zheng, A. and Morari, M. (1995). Stability of model predic-
tive control with mixed constraints, IEEE Trans.
Auto. Cont. 40: 1818{1823.
 Industrial Model Predictive Control Technology 23
Tables and Figures

Table 1. Companies surveyed, their Product Names and Acronyms

Company Acronym Product Name (Function)
Adersa HIECON Hierarchical Constraint Control
PFC Predictive Functional Control
GLIDE (Identication package)
DMC Corp. DMC Dynamic Matrix Control
DMI Dynamic Matrix Identication
Honeywell RMPCT Robust Model Predictive Control Technology
Promatics PCT Predictive Control Technology
Setpoint Inc. SMCA Setpoint Multivariable Control Architecture
SMC-Idcom (Multivariable control package)
SMC-Test (Plant test package)
SMC-Model (Identication package)
Treiber Controls OPC Optimum Predictive Control

Table 2. Summary of Reported MPC Vendor Applications by Areas (estimates based on vendor survey; estimates
do not include applications by companies who have licensed vendor technology)
Area DMC Setpoint Honeywell Adersa Treiber Total
Corp. Inc. Promatics Controls
Rening 360 320 290 280 250 1500
Petrochemicals 210 40 40 - - 290
Chemicals 10 20 10 3 150 193
Pulp and Paper 10 - 30 - 5 45
Gas - - 5 - - 5
Utility - - 2 - - 2
Air Separation - - - - 5 5
Mining/Metallurgy - 2 - 7 6 15
Food Processing - - - 41 - 41
Furnaces - - - 42 - 42
Aerospace/Defense - - - 13 - 13
Automotive - - - 7 - 7
Other 10 20 - 45 - 75
Total 600 402 377 438 416 2233
First App DMC:1985 IDCOM-M:1987 PCT:1984 IDCOM:1973 OPC:1987
SMCA:1993 RMPCT:1991 HIECON:1986
Largest App 603x283 35x28 28x20 - 24x19
24 MODEL PREDICTIVE CONTROL

Table 3. Comparison of Industrial MPC Control Technology

Company DMC Setpoint Honeywell Adersa Adersa Treiber
Corp Inc IAC Controls
Algorithm DMC SMC-Idcom RMPCT HIECON PFC OPC
Model Forms1 SR IR,TF SR,TF,ARX IR SS,TF,ARX IR
L,S,I L,S,I L,S,I L,S,I L,N,S,I,U L,S,I
Feedback2 CD, ID CD, ID CD, ID CD, ID CD, ID CD, ID
Rem Ill-cond3 IMS RCV SVT - - IMS
SS Opt Obj4 L[I,O] Q[O]...Q[I],R Q[I,O] - Q[I,O] -
SS Opt Const5 IH,OS IH,OH,OS,R IH,OH - IH,OH -
Dyn Opt Obj6 Q[I,O,M],S Q[I,O] Q[I,O] Q[O],Q[I] Q[I,O],S Q[I,O,M],S
Dyn Opt Const7 IH,OS IH,OS IH,OS IH,OH,OS,R IC,OH,OS,R IH,OS
Output Traj8 S,Z S,Z,RT,F S,Z,F S,Z,RT S,Z,RT S,Z
Output Horiz9 FH FH,MP FH FH FH,MP FH
Input Param10 MM SM MM SM BF,SM MM
1 Model Form: (IR) nite Impulse Response, (SR) nite Step Response, (TF) Laplace Transfer Function, (SS) State-Space, (ARX)
Auto-Regressive Exogenous input, (L) Linear, (N) Nonlinear, (S) Stable, (I) Integrating, (U) Unstable
2 Feedback: (CD) Constant output Disturbance, (ID) Integrating output Disturbance
3 Removal of Ill-conditioning: (RCV) Ranked Controlled Variables, (SVT) Singular Value Thresholding, (IMS) Input Move Suppres-
sion
4 Steady-State Optimization Objective: (L) Linear, (Q) Quadratic, (I) Inputs, (O) Outputs, (...) multiple sequential objectives, (R)
outputs Ranked in order of priority
5 Steady-State Optimization Constraints: (IH) Input Hard maximum, minimum and rate of change constraints, (OH) Output Hard
maximum and minimum constraints
6 Dynamic Optimization Objective: (Q) Quadratic, (I) Inputs, (O) Outputs, (M) Input Moves, (S) Sub-optimal solution
7 Dynamic Optimization Constraints: (IH) Input Hard maximum, minimum and rate of change constraints, (OH) Output Hard
maximum and minimum constraints, (OS) Output Soft maximum and minimum constraints, (IC) Input Clipped maximum, minimum
and rate of change constraints (sub-optimal), (R) constraints Ranked in order of priority
8 Output Trajectory: (S) Setpoint, (Z) Zone, (RT) Reference Trajectory, Funnel (F),
9 Output Horizon: (FH) Finite Horizon, (MP) Multiple Point
10 Input Parameterization: (SM) Single Move, (MM) Multiple Move, (BF) Basis Functions

Table 4. Comparison of Industrial MPC Identication Technology

Product Test Protocol Model Form Estimation Method Uncertainty Bound
DMC multi-steps (velocity) FIR modied LS No
SMC-Model PRBS, step FIR, ARX output error LS1 Yes
RMPCT PRBS, step FIR, ARX LS, G-N2 Yes
Glide non-PRBS G(s) GD3,G-N, GM4 Yes
OPC PRBS, step, pulse, FIR, ARMAX LS Yes
operating data
1 LS: Least Squares
2 G-N: Gauss-Newton
3 GD: Gradient Descent
4 GM: Global Methods
 Industrial Model Predictive Control Technology 25

Process Process Process

Inputs State States Output Outputs
Dynamics Measurements
u1 x1 y1
u2 x2 y2
u x3 y
3 .. .. .. .. 3
. . . .
um xn ym
u y

Figure 1. State-space process description.

Unit 1 - Conventional Unit 2 - Model Predictive

Control Structure Control Structure
Global Steady-State
Plant-Wide Optimization Optimization
(every day)

Local Steady-State
Unit 1 Local Optimizer Unit 2 Local Optimizer Optimization
(every hour)

High/Low Select Logic

Dynamic
PID PID Model Predictive Control Constraint
L/L
(MPC) Control
(every minute)

SUM SUM

Basic Dynamic
Unit 1 DCS- PID Controls Unit 2 DCS-PID Controls Control
(every second)

FC PC TC LC FC TC LC
PC

Figure 2. Hierarchy of control system functions in a typical processing plant. Conventional structure is shown
at the left; MPC structure is shown at the right.
26 MODEL PREDICTIVE CONTROL

Process Process
Inputs Multivariable Outputs
Process
u1 y1
MV’s u2 y2
u y CV’s
and 3 .. .. 3
. .
DV’s
um ym y
u

Figure 3. Input-output process description used in industrial MPC technology. Process inputs ui consist of two
types; manipulated variables (MV's) and disturbance variables (DV's). Process outputs are the controlled variables
(CV's).

MV’s

MV’s
MV’s
Thin Square Fat
CV’s Plant CV’s CV’s
Plant Plant

Over-determined Unique solution Under-determined

degrees of freedom < 0 degrees of freedom = 0 degrees of freedom > 0

Figure 4. Process structure determines the degrees of freedom available to the controller. Adapted from Froisy
(Froisy, 1994).
 Industrial Model Predictive Control Technology 27

Read MV, DV, CV values from process

Update process model (feedback)

Determine controlled process subset

Remove ill-conditioning

Local Steady-State Optimization

Dynamic Optimization

Output MV’s to process

Figure 5. Flow of MPC calculation at each control execution.

28 MODEL PREDICTIVE CONTROL

quadratic penalty Setpoint

past future

quadratic penalty Zone

past future

Reference Trajectory

quadratic penalty
past future

Funnel
quadratic penalty
past future

Figure 6. Four options for specifying future CV behavior; setpoint, zone, reference trajectory and funnel. Shaded
areas show violations penalized in the dynamic optimization.
 Industrial Model Predictive Control Technology 29

DV DV

SP SP
B B

A A
Funnel

CV CV
(a) MPC based on a trajectory (b) MPC based on a funnel

Figure 7. MPC based on a funnel allows a CV to move back to the setpoint faster than a trajectory would require
if a pulse disturbance releases. A trajectory based MPC would try to move away from the setpoint to follow the
trajectory.

Finite Horizon

past future prediction horizon P

Coincidence Points

past future

Figure 8. Output horizon options. Finite horizon (top) includes P future points. A subset of the prediction
horizon, called the coincidence points (bottom) may also be used. Shaded areas show violations penalized in the
dynamic optimization.
30 MODEL PREDICTIVE CONTROL

Multiple Moves

past future
control horizon M

Single Move

past future

u
Basis Function

past future

Figure 9. Input parameterization options. Multiple move option (top), single move option (middle), basis function
parameterization (bottom).
 Industrial Model Predictive Control Technology 31

Hard Constraint
past future

quadratic penalty

Soft Constraint
past future

quadratic penalty
Setpoint approximation of Soft Constraint
past future

Figure 10. The three basic types of constraint; hard, soft and setpoint approximation. Hard constraints (top)
should not be violated in the future. Soft constraints (middle) may be violated in the future, but the violation is
penalized in the objective function. Setpoint approximation of constraint (bottom) penalizes deviations above and
below the constraint. Shades areas show violations penalized in the dynamic optimization. Adapted from Froisy
(Froisy, 1994).