Int. J. Appl. Math. Comput. Sci., 2006, Vol. 16, No.

1, 7–26

SOFT COMPUTING IN MODEL–BASED PREDICTIVE CONTROL†

P IOTR TATJEWSKI, M ACIEJ ŁAWRY ŃCZUK

Institute of Control and Computation Engineering
Faculty of Electronics and Information Technology
Warsaw University of Technology
ul. Nowowiejska 15/19, 00–665 Warsaw, Poland
e-mail: {P.Tatjewski,M.Lawrynczuk}@ia.pw.edu.pl

The application of fuzzy reasoning techniques and neural network structures to model-based predictive control (MPC) is
studied. First, basic structures of MPC algorithms are reviewed. Then, applications of fuzzy systems of the Takagi-Sugeno
type in explicit and numerical nonlinear MPC algorithms are presented. Next, many techniques using neural network
modeling to improve structural or computational properties of MPC algorithms are presented and discussed, from a neural
network model of a process in standard MPC structures to modeling parts or entire MPC controllers with neural networks.
Finally, a simulation example and conclusions are given.

Keywords: process control, model predictive control, nonlinear systems, fuzzy systems, neural networks

1. Introduction 2. Model-Based Predictive Control

2.1. Predictive Control Principle
The Model-Based Predictive Control (MPC) technique is
the only one among advanced control techniques – usu- The general principle of predictive control is the follow-
ally defined as techniques more advanced than the well- ing: At each consecutive sampling instant k, the control
known PID control – which was tremendously successful inputs u(k) = u(k|k), u(k + 1|k), . . . , u(k + Nu − 1|k)
in practical applications, exerting great influence on the are calculated, assuming u(k + p|k) = u(k + N u − 1|k)
directions of the development of industrial control sys- for p ≥ Nu , where Nu is the control horizon. The ap-
tems as well as research conducted in this area (Allgöwer plied notation ‘u(k + p|k)’ means the prediction of the
et al., 1999; Brdyś and Tatjewski, 2005; Eder, 1999; Ma- control input value for the future time k + p, performed
ciejowski, 2002; Mayne et al., 2000; Morari and Lee, at the time k. The control inputs are calculated in such
1999; Qin and Badgwell, 2003; Rossiter, 2003; Tatjew- a way as to minimize differences between the predicted
ski, 2002). There are several reasons for this success. controlled outputs y(k + p|k) and the foreseen set points
First, predictive control algorithms can take into account for these outputs y sp (k +p|k) over the prediction horizon
in a natural way constraints on both process inputs (con- N (p = 1, 2, . . . , N ). Then, only the first element u(k|k)
trol signals) and process output values (controlled vari- of the calculated sequence is applied to the process, i.e.,
ables), which often decide on the quality, effectiveness u(k) = u(k|k). At the next sample (k + 1), there oc-
and safety of production. They generate control signals curs a new measurement of the process outputs and the
taking on-line into account these constraints and, owing whole procedure is repeated, with the prediction horizon
to a direct use of a model, also internal interactions within of the same length N , but shifted by one step forward (the
the process. Thus, they can naturally be applied to mul- principle of a receding horizon, also called the repetitive
tivariable process control, also when the numbers of the control principle (e.g., Findeisen, 1997; Findeisen et al.,
control inputs and the controlled variables differ. Third, 1980)). This principle is presented in Fig. 1, for a SISO
the principle of operation of these algorithms is compre- (single-input single-output) process.
hensible and relatively easy to explain to engineering and A model of the process used in the MPC algorithm is
operator staff, which is a very important aspect when in- usually only an approximation of reality. Moreover, there
troducing new techniques into industrial practice. is some uncertainty in the uncontrolled inputs, which can
be inaccurately measured or not measured at all. There-
fore, the output predictions usually differ from the (later)
† The work was supported by a statutory research project. measured values. This fact is depicted in Fig. 1 as a dis-
 8 P. Tatjewski and M. Ławry ńczuk

y
y sp

prediction y(k+p|k)

measurement
y(k)

disturbance estimate d(k)

prediction y(k|k-1)
control horizon
prediction horizon

k–1 k k+1 k+2 k+3 k+Nu k+N time

u
Du(k+1|k) control u(k+p|k), p=0,...,N-1

k–1 k k+1 k+2 k+3 k+Nu k+N time

Fig. 1. Principle of predictive control.

turbance d(k), d(k) = y(k)−y(k|k −1), occurring at the for more general or alternative formulations):
process output at sampling instant k, where y(k|k − 1)
is the output value predicted for the sample k at the pre- N
 2
vious sample (k − 1). The dependence of the trajectory J(k) = y sp (k + p|k) − y(k + p|k)
of the process output predicted at sampling instant k on p=N1

the currently measured value y(k), and not on the value N

u −1
y(k|k − 1), signifies the application of a discrete feedback +λ u(k + p|k)2 , (1)
in the control system. p=0

Determination of the control input trajectory over the where the vectors y sp (k + p|k) and y(k + p|k) are of
control horizon is realized in the predictive algorithms on dimensionality n y = dim y equal to the number of the
the basis of a model, by minimizing a cost function de- controlled outputs, while the vector of control increments
scribing the control quality over the prediction horizon. u(k + p|k) is of dimensionality n u = dim u and
A prime component of this function is the cost of devia- 1 ≤ N1 ≤ N (N1 > 1 is reasonable if there is a delay
tions of the predicted outputs from the set points, i.e., the causing no reaction of the outputs at the sampling instants
cost of predicted control errors. Moreover, it is also typ- k + 1, . . . , k + N1 − 1 to the change in the control input
ical to include into the cost function penalties on control at the sampling instant k). The length of the control hori-
changes. As regards the two components, the following zon Nu must satisfy the constraints 0 < N u ≤ N . It
most commonly used quadratic cost function can be for- is usually assumed that N u < N , as this results in a de-
mulated (see, e.g., (Maciejowski, 2002; Tatjewski, 2002) creased dimensionality of the optimization problem, and
Soft computing in model-based predictive control 9

thus leads to a smaller computational load. The role of the

MPC ALGORITHMS
coefficient λ is to scale the second sum of squared con-
trol increments against the first sum representing squared
predicted control errors.
LINEAR NONLINEAR
The optimal trajectory of control inputs is evaluated, (with linear model) (with nonlinear model)
at every sampling instant, by the minimization of the cost
function (1) subject to several constraints:
⎧ ANALYTICAL NUMERICAL SUBOPTIMAL OPTIMAL
⎨ N
 sp
(control law - (constrained with (optimization
 y (k + p|k) linear, explicit QP optimization process model with nonlinear
min linear-quadratic) problem) linearizations process model)
u(k|k),...,u(k+Nu −1|k) ⎩
p=N1

N
ANALYTICAL FUZZY NUMERICAL
u −1
2 (local control laws (optimization
−y(k + p|k)  + λ
2
u(k + p|k) (2) in a TS fuzzy structure) simplified to QP)

p=0
Fig. 2. General classification of basic MPC algorithms.
subject to

y(k + p|k) ∈ Y, p = 1, . . . , N, In a linear case, by applying the principle of super-

position it is possible to present the trajectory of the pre-
u(k + p|k) ∈ U, p = 0, . . . , Nu − 1, dicted outputs y(k + p|k) in the form of a sum of a free
trajectory y 0 (k + p|k) dependent only on the realized
u(k + p|k) = u(k + Nu − 1|k), p = Nu , . . . , N − 1,
(past) process inputs, and a trajectory y(k + p|k) de-
u(k + p|k) = u(k + p|k) − u(k + p − 1|k), pendent only on the decision variables (current and future
control inputs u(k + p|k)), called the forced output tra-
where the values of predicted outputs are calculated us- jectory. Thus we have
ing a dynamic process model, in general, a nonlinear one.
Assuming that it is in the input-output form, the following y(k+p|k) = y 0 (k+p|k)+y(k+p|k), p = N1 , . . . , N.
formulation is fairly general: The above partition is convenient, though not necessary

for the realization of a predictive control algorithm, be-
y(k + p|k) = fp u(k + p − 1|k), . . . ,
cause the values y 0 (k + p|k), as dependent only on the
u(k|k), u(k − 1), . . . , u(k − nB ), y(k), y(k − 1), past of the process, are calculated at a given sampling in-
 stant k only once, and remain then as fixed parameters in
. . . , y(k − nA ), d(k) , p = 1, . . . , N. (3) the following optimization of the control moves. On the
other hand, the increments y(k + p|k), as dependent
The class that the process model belongs to is critical, as on the current and future control moves u(k + p|k),
model equations are equality constraints in the MPC opti- are calculated many times in the process of numerical op-
mization problem with the quadratic cost function. There- timization. As regards the presented decomposition, the
fore, nonlinear models result in nonquadratic, generally cost function (1) can be written in the form
nonconvex optimization problems – thus creating serious
N

problems for on-line applications, as required in MPC al-  sp 
J(k) =  y (k + p|k) − y 0 (k + p|k)
gorithms. A classification of basic classes of MPC algo-
rithms is presented in Fig. 2. It should be treated as a p=N1

rather simplified one, i.e., many classes of nonlinear op-

2 N
u −1
timal MPC algorithms (with optimization using nonlinear 
− y(k+p|k) +λ
2
u(k+p|k) , (4)
process models) can be further distinguished, but this is p=0
beyond the scope of this paper.
where free and forced output trajectories are given by the
linear models
2.2. MPC with Linear Process Models  T
y 0 (k + p|k) = y(k) + gp u(k − 1) · · · u(k − nB )
So far, MPC algorithms with linear process models have
 T
been of the greatest importance. First of all, the range + hp y(k − 1) · · · y(k − nA ) , (5)
of direct applications of these algorithms is fairly wide. 
Secondly, they constitute a basis for the construction of y(k + p|k) = mp u(k|k) u(k + 1|k)
relatively simple and efficient nonlinear algorithms using T
linearizations of nonlinear models. · · · u(k+p−1|k) , p = N1 , . . . , N, (6)
 10 P. Tatjewski and M. Ławry ńczuk

y
y sp

prediction y(k+p|k)

y(k+p|k) = y 0 (k+p|k) + Dy(k+1|k)

Dy(k+1|k) - 1st element of the forced trajectory

measurement
y(k)
free trajectory y 0 (k+p|k)
d(k)
(prediction with Du(k+p-1|k)=0, p=1,...,N)

prediction y(k|k-1)
control horizon
prediction horizon

k–1 k k+1 k+2 k+3 k+Nu k+N time

u
Du(k+1|k) control u(k+p|k), p=0,...,N-1

control with Du(k+p-1|k) = 0, p=1,...,N

k–1 k k+1 k+2 k+3 k+Nu k+N time

Fig. 3. Predictive control with a linear process model.

(this is a SISO case, in MIMO cases row vectors g p , hp where ku and ky are the (column) vectors of feedback
and mp become matrices). The principle of predictive coefficients (SISO case).
control for a linear process model is shown in Fig. 3 (com- We shall describe MPC algorithms formulated as
pare it with Fig. 1). control laws given by an explicit analytical formula (or
Owing to the linearity of the model, the optimization a set of formulas) as explicit MPC algorithms. This class
problem (2) is a strictly convex quadratic programming of algorithms consists of two basic groups:
problem. Therefore, it has a unique, global minimum • simple, unconstrained algorithms given by
which can be fast and reliably computed numerically in constraint-free control laws (7),
a constrained case. In a constraint-free case the solution
can be computed analytically as a linear feedback control • more sophisticated, constrained algorithms given by
law, i.e., in the following general form (for the set point a number of constrained linear control laws, each
y sp constant over the prediction trajectory): valid for a different set of active constraints, and a
switching mechanism (originally defined as explicit
u(k) = u(k|k) linear quadratic algorithms for constrained systems

 (Bemporad et al., 2002)) – they are in fact nonlinear
= k e y sp (k) − y(k) algorithms.
 T
+ (ku )T u(k − 1) · · · u(k − nB ) On the other hand, the MPC algorithms solving numeri-
 T cally, at each sampling instant, constrained QP problems
+ (ky )T y(k) · · · y(k − nA ) , (7) will be described as numerical MPC algorithms (see also
Soft computing in model-based predictive control 11

d(k)
y sp(k) + D u(k) 1 u(k) +
_
K e
_ -1 I PROCESS
+ 1-z + y(k)
D -1
å K ju D u(k-j)
j =1

Fig. 4. Structure of the DMC control law.

v(k)
sp
y (k) + D u(k) 1 u(k)
_ ke _ _ PROCESS
+ 1-z -1 y(k)
nB
å
j=
k uj D u(k-j)
1

nA -1 _ D y(k) _ z-1
å
j =0
k jyD y(k-j) +

Fig. 5. Structure of the GPC control law.

Fig. 2). Unconstrained linear feedback control laws for mulated in the following form (Tatjewski, 2002):
two most popular MPC algorithms using input-output type

process models will be briefly shown as examples. u(k) = k e y sp (k) − y(k)
 T
When the process model is in the form of a dis- +(ku )T u(k − 1) · · · u(k − nB )
crete step response, then the linear MPC algorithm is the
 T
well-known, most popular in process control applications +(ky )T y(k) · · · y(k−nA +1) , (10)
DMC (Dynamic Matrix Control) algorithm (Cutler and
Ramaker, 1979; Maciejowski, 2002; Rossiter, 2003; Tat- illustrated in Fig. 5. By comparison with Fig. 4, it can
jewski, 2002). The feedback control law (7) takes then the easily be seen that nA measured past outputs are added as
following form of an explicit DMC algorithm: additional feedback information, but there are fairly fewer
feedback instances from past inputs (usually n B  D).


u(k) = k e y sp (k) − y(k) The explicit MPC algorithm (control law; DMC,
 T GPC or other) can be easily implemented, as defined by
+(ku )T u(k−1) · · · u(k−nB +1) , (8) the appropriate formula, if there is no risk that the gener-
ated control signal will exceed limits defined by physical
possibilities of the process actuator, i.e., if constraints on
where nB = D denotes the process dynamics horizon,
the process input are always inactive, as they have been ig-
i.e., the number of step response coefficients s i present
nored when constructing the laws. However, this is often
until their values become constant, (s i = const for i ≥
not the case and the constraints must be taken into account
D). The structure of the control law (8) is shown in Fig. 4.
in order to get practically applicable controllers. This can
When the process model is in the form of a discrete be done effectively only if the following two principles
transfer function or, equivalently, the difference equation hold:
(ARX-type model):
• the past control values u(k − j) which are fed
nA nB back into the controller structure must be those corre-
 
y(k) = − aj y(k−j)+ bj u(k−j −1)+d(k), (9) sponding to the process past constrained inputs, not
j=1 j=0 the past values generated by the constraint-free con-
trol law (7),
then we obtain the popular GPC (Generalized Predictive • an anti-windup scheme must be added.
Control) algorithm (Camacho and Bordons, 1999; Clarke
et al., 1987; Maciejowski, 2002; Rossiter, 2003; Tatjew- This is illustrated for a case of the DMC algorithm in
ski, 2002). The feedback control law (7) can then be for- Fig. 6, where models of input constraints (rate of change
 12 P. Tatjewski and M. Ławry ńczuk

CONTROLLER z-1I
+
_
u int d
y sp e +
Du + 1
u u +
K I PROCESS
+ _ _ + 1-z -1 + y
D -1 _ z-1 I
å K ju D u(k-j)
j =1 Du +

Fig. 6. Structure of the explicit DMC controller taking into account constraints on the rate of change and
amplitude of the controller output, with an additional anti-windup scheme (upper feedback loop).

and amplitude saturations) are elements of the controller then the optimization problem (2) is certainly not linear-
structure within its inner feedback loop, and an additional quadratic; it is generally a nonconvex and even multi-
anti-wind up feedback loop has been added (compare this modal one. For such problems there are no sufficiently
with Fig. 4). In (Tatjewski, 2002), simulation examples fast and reliable numerical optimization procedures, i.e.,
are presented comparing the results of MPC control with procedures yielding always an optimal point and within
an explicit controller in structures where input constraints predefined time limit – as is required in on-line control
are and are not taken into account. These results show applications. Therefore, many attempts have been made
that implementation as in Fig. 4 leads to drastically de- to construct simplified (and generally suboptimal) nonlin-
creased control quality (or even instability) when input ear MPC algorithms avoiding full on-line nonlinear opti-
constraints are active, while implementation in the struc- mization, first of all using model linearizations or multiple
ture as in Fig. 6 works usually very well, often compara- linear models in fuzzy structures, see Fig. 2. The resulting
bly to numerical MPC implementation (with constrained predictive control algorithms apply soft computing tech-
optimization at each sampling instant). A further discus- niques to a great extent. On the other hand, there are also
sion of this point is beyond the scope of the paper, and the many designs of predictive algorithms based on nonlinear
reader is referred to the cited reference for more details. optimization and also using soft computing, mainly those
Obviously, explicit MPC algorithms can be recom- applying artificial neural network techniques.
mended in constraint-free cases, also when input con- Therefore, the presentation of predictive control al-
straints can become active, but a simple and fast (short gorithms using soft computing techniques will be done
sampling periods) controller is needed – in this case con- within the following groups:
troller implementation must be augmented taking into ac-
count the constraints, as has just been discussed. When • MPC algorithms using fuzzy reasoning:
implementing numerical MPC algorithms with linear – Multi-model explicit algorithms in the fuzzy
models, significantly more powerful controllers are re- Takagi-Sugeno (TS) structure (Babuška et al.,
quired, especially for shorter sampling periods, but the 1999; Marusak and Tatjewski, 2002; Tatjewski,
control is optimal (in the sense of the applied cost func- 2002),
tion).
– Algorithms with on-line linearization of a fuzzy
There is no need for soft computing methods in MPC
TS model (usually with the same structure of
algorithms with linear models. But these algorithms and
all fuzzy rule consequents) and QP optimiza-
their philosophy are basic for more involved MPC algo-
tion (Babuška et al., 1999; Tatjewski, 2002).
rithms with nonlinear plant models, when soft computing
methods play an important role. • MPC algorithms using artificial neural networks:

– Algorithms with nonlinear optimization and

3. Soft Computing in Nonlinear MPC a neural network process model or a neural
Algorithms network prediction model (Ławry ńczuk, 2003;
Ławryńczuk and Tatjewski, 2001a; 2001b;
The general formulation of the MPC optimization prob- 2004; Najim et al., 1997; Nørgaard et al.,
lem has been previously defined as (2), where predic- 2000; Ortega and Camacho, 1996; Temeng et
tions of controlled outputs are calculated using an input- al., 1995; Trajanoski and Wach, 1998; Yu and
output process model (3). When this model is nonlinear, Gomm, 2003),
Soft computing in model-based predictive control 13

y(k) u(k)
D u (k)
1
DMC 1 w1(k)
D u2(k)
y sp(k) e(k) DMC 2 w2(k) + D u(k) 1
+ I PROCESS
+ _
+
1-z -1 u(k) y(k)

...

...
D u r(k)
DMC r w r(k)
Fuzzy
reasoning

Fig. 7. Structure of the explicit (unconstrained) fuzzy DMC controller.

– Algorithms with on-line linearization of a point) of the process to local subdomains, in

neural network model and QP optimization which the process can be approximated by lin-
(Ławryńczuk, 2003; Ławry ńczuk and Tatjew- ear models, for control purposes.
ski, 2002; 2004; Nørgaard et al., 2000), (b) Divide the range of variability of each vari-
– Neural network modeling applied to reduce able of rule antecedents into overlapping fuzzy
computational complexity and to approximate sets (in relation to process non-linearity), defin-
the controller (optimizer) (Liu and Daley, 1999; ing the number of these sets and assigning the
Liu et al., 1998; Parisini and Sacone, 2001; shape and values of the parameters of the mem-
Parisini et al., 1998; Parisini and Zoppoli, bership function of each set. This creates a
1995; Piche et al., 2000; Vila and Wagner, division of the whole domain into a number
2003; Wang and Wan, 2001). of overlapping subdomains (multidimensional
fuzzy subsets).
(c) Formulate a linear process model for each sub-
4. MPC Algorithms Using Fuzzy Reasoning domain constructed so far (each subdomain is
defined by one fuzzy rule).
Techniques
2. For each linear model (submodel) of the process, de-
4.1. Multimodel Explicit Algorithms in the Fuzzy TS
sign an unconstrained explicit MPC controller oper-
Structure
ating correctly in the assigned subdomain. Create a
A simple and effective way to generalize a linear ex- TS fuzzy MPC controller – a TS fuzzy system with
plicit MPC algorithm (a linear control law) to the case rule antecedents created in the previous point (divi-
of a nonlinear process model is to apply fuzzy reasoning sion into fuzzy sets as in the TS fuzzy model of the
techniques. In particular, fuzzy systems with the Takagi- process) and functional consequents which are the
Sugeno (TS) structure are most suitable for the design of designed explicit MPC control algorithms.
fuzzy controllers (Tatjewski, 2002). This design can be
3. Perform the analysis of the obtained TS fuzzy algo-
divided into the following stages:
rithm. If the results are not satisfactory, return to
1. Design a TS fuzzy model of the nonlinear process the previous points and correct the design of the TS
considered, i.e., with fuzzy rules with functional con- fuzzy process model and/or local linear MPC con-
sequents which are linear dynamic models of the trollers.
process in particular local subdomains. The range
The resulting fuzzy MPC controller structure, in the case
of variability of the input and output signals should
of r fuzzy rules (r subdomains) with local DMC con-
cover the entire operation domain of the designed
trol laws as rule consequents, is depicted in Fig. 7, where
controller:
w̃i (k) denotes the normalized activation level of the i-th
(a) Define variables occurring in rule antecedents. fuzzy rule at the k-th sampling instant, i = 1, . . . , r.
The values of these variables will describe The presented approach leads to a very simple con-
the membership of the current state (operating troller implementation if each local process model and
 14 P. Tatjewski and M. Ławry ńczuk

thus each local MPC controller are precisely of the same the controller output signal – we get in this way MPC-
structure (have the same number of parameters). For ex- NSL (MPC-Nonlinear with Successive Linearization) al-
ample, if each of r explicit DMC controllers is of the gorithms (Maciejowski, 2002; Tatjewski, 2002).
form (8)

 However, this is not the most clever use of a lin-
ui (k) = kie y sp (k) − y(k)
earized process model in the MPC algorithm. To see the
 T reason, let us recall the general principle of predictive con-
+(kui )T u(k−1) · · · u(k−nB +1) ,
trol with a linear process model, depicted in Fig. 3. The
i = 1, . . . , r, (11) predicted process output trajectory is divided here into two
parts: the free output trajectory dependent on past process
then fuzzy reasoning results in the following fuzzy con- inputs and outputs, and the forced output trajectory de-
troller output: pendent on controller decision variables – actual and fu-
ture process input moves. The difference between the two
r
 trajectories is fundamental from the computational point
u(k) = w̃i (k)ui (k)
of view: at each sampling instant the former is calculated
i=1
only once (and serves as a set of constant parameters in

r
 the optimization problem) whereas the latter must be eval-
= w̃i (k)kie y sp (k) − y(k) uated every time the future control moves are changed,
i=1
i.e., several times during the optimization process, thus its

r properties define the properties of optimization. There-
+ w̃i (k)(kui )T fore, if the free trajectory is calculated using the nonlin-
i=1 ear process model but the forced trajectory is calculated
 T using the linear (linearized) process model, then the opti-
× u(k−1) · · · u(k−nB +1) , (12) mization problem is still a QP (Quadratic Programming)
problem that can be solved fast and reliably. This is the
that is, the nonlinear fuzzy controller equation is struc- basic idea of probably the most clever use of linearization
turally identical to the linear DMC equation (8), and only
in nonlinear MPC structures – the idea of the MPC-NPL
the values of gain coefficients (given in square brackets (MPC with Nonlinear Prediction and Linearization) algo-
in (12)) are time-varying, being time-varying weighted rithm (Garcia, 1984; Tatjewski, 2002). It is explained in
sums of local controller gains. Fig. 8, where the free output trajectory calculated at the
When constraints on process inputs are to be taken k-th sampling instant is denoted by Y 0 (k),
into account, then the structure of the presented fuzzy TS
predictive controller, as shown in Fig. 7 or in a simplified
version following from (12), should be modified exactly  T
Y 0 (k) = y 0 (k + 1|k) y 0 (k + 2|k) · · · y 0 (k + N |k) ,
in the same way as the structure from Fig. 4 was modified (13)
to obtain that of Fig. 6.
whereas the nonlinear model linearization (over the free
trajectory) is used only to calculate the dynamic matrix
4.2. Algorithms with On-line Linearization of Fuzzy
M(k) describing the relation between the future control
TS Models
moves and the corresponding elements of the forced tra-
The explicit fuzzy TS MPC controller presented in the
previous section was designed as an unconstrained con-
troller, and process input constraints can be taken into ac-
CONTROLLER
count only in a suboptimal way as discussed at the end d(k)
sp
of that section. Moreover, constraints on process outputs Y (k) Optimization D u(k) 1 u(k)
with a I PROCESS
could not be taken into account. As has been mentioned 1-z -1 y(k)
linear model
earlier, numerical MPC algorithms numerically solving at z-1 I
each sampling instant an optimization problem yield op-
0
Y (k) M(k) u(k-1)
timal constrained solutions, but, in general, there are no Nonlinear prediction of Y (k)
0

fast and reliable numerical procedures for solving opti- Linearization to get
the local dynamic matrix M(k)
mization problems with nonlinear process models. One of
the most natural suboptimal solutions to this problem is to
linearize the nonlinear process model at each sampling in- Fig. 8. Structure of the MPC-NPL controller.
stant, and then to use a linear MPC algorithm to evaluate
Soft computing in model-based predictive control 15

jectory, c.f. (6), ⎡ ⎤ Each equation (18) will be then a consequent of a cor-
m1 responding fuzzy rule of the nonlinear fuzzy TS process
⎢ ⎥ model. Then, the linearization of this model at the k-th
⎢ m2 ⎥
M=⎢
⎢ ..
⎥.
⎥ (14) sampling instant will simply be given by
⎣ . ⎦ nA

mN y(k) = − aj (k)y(k − j)
j=1
A computationally simple linearization can be ob-
tained when the nonlinear process model is a fuzzy TS nB

model with all rule consequents of the same structure, i.e., + bj (k)u(k − j − 1) + d(k), (19)
when all local linear process models are of the same kind j=0
and structure. Denoting the vector of variables of a gen- where
eral process model by x, x = [x 1 , x2 , . . . , xn ]T , where r

the elements of x can consist of process inputs and out- aj (k) = w̃i (k)aij , j = 1, . . . , nA , (20)
puts (current and delayed), the rules R i of a TS fuzzy i=1
model can be written is the following form: r

bj (k) = w̃i (k)bij , j = 0, . . . , nB , (21)
i
R : IF x1 (k) is Ai1 and · · · and xn (k) is Ain i=1

and the matrix M(k) will be calculated as in the standard

THEN y i (k+1) = ai0 +ai1 x1 (k)+· · ·+ain xn (k).
GPC algorithm using the coefficients a j (k) and bj (k),
The output of the fuzzy model is known to be given by the instead of aj and bj .
formula If local linear models are given by the local step re-
r  n
 sponse coefficients sij , j = 1, . . . , D, i = 1, . . . , r (r
y(k + 1) = w̃i (k) ai0 + aij xj (k) , (15) stands for the number of fuzzy rules), then at each sam-
i=1 j=1 pling instant the matrix M(k) will be composed of the co-
where r denotes the number of rules and w̃ i (k) are nor- efficients of a “fuzzy” step response given by the formula
malized activation levels of the fuzzy model at the sam- r

pling instant k (at a point x(k)). Due to the linearity and sj (k) = w̃i (k)sij , j = 1, . . . , D. (22)
identical structure of all consequents of the rules, we can i=1
present (15) as follows:
n
 5. MPC Algorithms Using Artificial Neural
y(k + 1) = a0 (k) + aj (k)xj (k), (16)
j=1
Networks
where 5.1. Neural-Network Model of the Plant
r

aj (k) = w̃i (k)aij , j = 0, . . . , n. (17) Let the single-input single-output (SISO) process under
i=1
consideration be described by the following nonlinear
discrete-time equation:
In this way we obtain a fuzzy model in the form of a lin-

ear model with variable coefficients, with values depend- y(k) = g u(k − τ ), . . . , u(k − nB ),
ing on the current process state. If at the current step the 
process state is x(k), then the simplest and most natural y(k − 1), . . . , y(k − nA ) , (23)
way of generating a local linear model to be used in the where g : RnA +nB −τ +1 −→ R, g ∈ C 1 , τ ≤ nB . In the
optimization problem of a predictive algorithm is to take sequel, it is assumed that the feedforward neural network
the model in the form (16). with one hidden layer and linear output (Haykin, 1999) is
For example, let us assume that local models of used as the function g in (23). The structure of the neural
process dynamics (identified at a number of process equi- network is depicted in Fig. 9. The output of the model can
librium points) are in the form of the difference equation be expressed as
nA
 K
i

y (k) = − aij y(k − j) y(k) = w2 (0) + w2 (i)vi (k)
j=1 i=1

nB
 K


+ bij u(k − j − 1) + d(k). (18) = w2 (0) + w2 (i)ϕ zi (k) , (24)
j=0 i=1
 16 P. Tatjewski and M. Ławry ńczuk

Fig. 9. Structure of the neural network.

⎡ ⎤
where zi (k) is the sum of inputs and and v i (k) is the out- y(k + N1 |k)
put of the i -th hidden node, respectively, ϕ : R −→ R ⎢ ⎥
Y (k) = ⎢ ⎥,
..
is a nonlinear transfer function, K is the number of hid- ⎣ . ⎦
den nodes. Recalling the input arguments of the general y(k + N |k)
nonlinear model (23), one has ⎡ ⎤ (26)
u(k|k)
⎢ ⎥
U (k) = ⎢ ⎥,
Iu
 ..
zi (k) = w1 (i, 0) + w1 (i, j)u(k − τ + 1 − j) ⎣ . ⎦
j=1 u(k + Nu − 1|k)
nA
 it is convenient to express the cost function (1) in the fol-
+ w1 (i, Iu + j)y(k − j). (25) lowing form:
j=1
2
J(k) = Y sp (k) − Y (k)
The weights of the network are denoted by w 1 (i, j),  2
i = 1, . . . , K, j = 0, . . . , nA + nB − τ + 1, and w2 (i), + λ (I + J N O )U (k) + U N O  , (27)
i = 0, . . . , K, for the first and second layers, respec-
tively. The number of the network’s input nodes depend- where the matrices I (unit matrix) and J N O are of di-
ing on the input signal u is I u = nB − τ + 1. The mension Nu × Nu , and the vector U N O is of length Nu ,
total number of weights is (n A + Iu + 1)K + K + 1 = ⎡ ⎤
(nA + nB − τ + 2)K + K + 1. 0 ··· 0 0
⎢ ⎥
⎢ −1 · · · 0 0 ⎥
JNO = ⎢
⎢ .. .. .. .. ⎥,
⎥
5.2. MPC Algorithms with Nonlinear Optimization ⎣ . . . . ⎦
(MPC-NO) and Neural Network Models 0 · · · −1 0
⎡ ⎤ (28)
In general, there are two methods of using neural mod- u(k − 1)
⎢ ⎥
els in MPC schemes with nonlinear optimization. In the ⎢ 0 ⎥
first approach, the gradients of the cost function J(k) are U NO = ⎢⎢ .. ⎥.
⎥
⎣ . ⎦
approximated numerically and the nonlinear optimization
problem (2) is solved on-line. In the second approach, the 0
structure of the neural model is exploited (Ławry ńczuk,
2003; Ławry ńczuk and Tatjewski, 2001a; 2001b; Nør- Differentiating (27) with respect to a future control se-
gaard et al., 2000). Defining the vectors quence, U (k), results in
⎡ ⎤  T
∂J(k) ∂Y (k)

y sp (k + N1 |k) =2 Y (k) − Y sp (k)
⎢ ⎥ ∂U (k) ∂U (k)
Y sp (k) = ⎢ ⎥,
..
⎣ . ⎦
T
y sp (k + N |k) + 2λ I + J N O U (k). (29)
Soft computing in model-based predictive control 17

The matrix of dimension (N − N u + 1) × Nu, containing where Iuf (p) = max{min{p − τ + 1, Iu }, 0} is the
partial derivatives of the predicted output with respect to number of the network’s input nodes depending on fu-
future control is ture control signals and I yp (p) = min{p − 1, nA } is
⎡ ⎤ the number of the network’s input nodes depending on
∂y(k + N1 |k) ∂y(k + N1 |k)
⎢ · · · output predictions. Because typically N u < N (hence
∂u(k|k) ∂u(k + Nu − 1|k) ⎥
∂Y (k) ⎢ ⎢ .. . ..
⎥
⎥ u(k + p|k) = u(k + Nu − 1|k) for p ≥ Nu ), it can be
=⎢ . .. . ⎥. noticed that
∂U (k) ⎢ ⎥
⎣ ∂y(k + N |k) ∂y(k + N |k) ⎦
··· Iuf (p)

∂u(k|k) ∂u(k + Nu − 1|k)
u(k − τ + 1 − j + p|k)
(30) j=1

The predictions y(k + p|k) for p = N 1 , . . . , N are cal- = INu (p) u(k + Nu − 1|k)
culated from the general prediction equation
Iuf (p)

y(k + p|k) = y(k + p) + d(k), (31) + u(k − τ + 1 − j + p|k), (35)
j=INu (p)+1
where the quantities y(k + p) are calculated from the
model. The above formulation uses the “DMC type” dis- where INu (p) = min{max{p − τ − Nu + 1, 0}, Iu }.
turbance model (Maciejowski, 2002; Tatjewski, 2002), in Taking into account (35), Eqn. (34) can be written as
which the unmeasured disturbance d(k) is assumed to be
constant over the prediction horizon. Its value is estimated zi (k + p) = w1 (i, 0)
from the equation
INu (p)

d(k) = y(k) − y(k|k − 1) + w1 (i, j)u(k + Nu − 1|k)
 K
 j=1

= y(k) − w2 (0) + w2 (i)vi (k) . (32) Iuf (p)

i=1
+ w1 (i, j)u(k − τ + 1 − j + p|k)
From the neural model (24) one has j=INu (p)+1

K
 Iu

 
y(k + p) = w2 (0) + w2 (i)ϕ zi (k + p) . (33) + w1 (i, j)u(k − τ + 1 − j + p)
i=1 j=Iuf (p)+1

As regards prediction over the horizon N , the quantities Iyp (p)

zi (k + p) and, consequently, y(k + p) will depend on 
+ w1 (i, Iu + j)y(k − j + p|k)
some control signal values applied to the plant at previ-
j=1
ous sampling instances, future control signals, i.e., deci-
sion variables of the control algorithm, measured values nA

of the plant output signal and future output predictions. + w1 (Iu + j)y(k − j + p). (36)
From (25) one has j=Iyp (p)+1

zi (k + p) = w1 (i, 0) Taking into account (31) and (33), the entries of the matrix
∂Y (k)/∂U (k) given by (30), i.e., the partial derivatives
Iuf (p)
 of the predicted output signal with respect to future con-
+ w1 (i, j)u(k − τ + 1 − j + p|k)
trols are calculated from
j=1
K
Iu
 ∂y(k + p|k)  dϕ(zi (k + p)) ∂zi (k + p)
= w2 (i) .
+ w1 (i, j)u(k − τ + 1 − j + p) ∂u(k + r|k) dzi (k + p) ∂u(k + r|k)
i=1
j=Iuf (p)+1 (37)
Iyp (p)
Obviously,

+ w1 (i, Iu + j)y(k − j + p|k) ∂zi (k + p) ∂y(k + p|k)
j=1 = = 0, r ≥ p−τ +1. (38)
∂u(k + r|k) ∂u(k + r|k)
nA

+ w1 (Iu + j)y(k − j + p), (34) It can be noted that decision variables of the algorithm
j=Iyp (p)+1 affect only the first, second and fourth sums in (36). It can
 18 P. Tatjewski and M. Ławry ńczuk

be also noted that only some of the output predictions are Defining a linearization point as a vector composed
influenced by future controls. Hence of past input and output values:

INu (p)
 T
∂zi (k + p)  ∂u(k + Nu − 1|k) x(k) = u(k−τ ) . . . u(k−nB ) y(k−1) . . . y(k−nA )
= w1 (i, j) (41)
∂u(k + r|k) j=1
∂u(k + r|k)
and using the Taylor series expansion at this point, the
Iuf (p) linearized model has the form
 ∂u(k − τ + 1 − j + p|k)
+ w1 (i, j)

∂u(k + r|k) y(k) = g x(k)
j=INu (p)+1
nB

Iypf (p)


 ∂y(k − j + p|k) + bl x(k) u(k − l) − u(k − l)
+ w1 (i, Iu + j) , (39) i=1
j=1
∂u(k + r|k)
nA



− al x(k) y(k − l) − y(k − l) . (42)
where Iypf (p) = max{min{p−τ, nA}, 0} is the number
i=1
of network input nodes depending on output predictions
which are affected by future controls. Obviously, Taking into account the structure of the neural model and
 the corresponding equations (24) and (25), the coefficients
∂u(k + p|k) 0 if p = r, of the linearized model are calculated from
= (40)
∂u(k + r|k) 1 if p = r,


 ∂g x(k)
al x(k) = −
whereas the derivatives of the predicted output signals ∂y(k − l)
with respect to future controls have to be calculated re- K
 dϕ(zi (x(k)))
cursively. =− w2 (i) (i, Iu + l),
i=1
dzi (x(k))
The discussed method of calculating the gradients of
the predicted output trajectory with respect to future con- l = 1, . . . , nA (43)
trols is used not only for obtaining the gradients of the
cost function J(k), but also for finding the gradients of and
output constraints if they have to be taken into account.


In some nonlinear optimization algorithms, for example, bl x(k)
SQP (Bazaraa et al., 1993), the analytical Hessian matrix
⎧
can be used. Unfortunately, it requires much more com- ⎪
⎪ 0 if l = 1, . . . , τ − 1,
putational effort than the calculation of gradients. That is ⎪
⎪
⎪
⎪
why in the presented solution the optimization routine is ⎪
⎪
⎪
⎨
∂g(x(k))
provided with analytical gradients while the Hessian is ap-
= ∂u(k − l) (44)
proximated, as is done in most SQP practical implemen- ⎪
⎪  K
⎪
⎪ dϕ(z i (x(k)))
tations. ⎪
⎪ = w2 (i) (i, l − τ + 1)
⎪
⎪ dzi (x(k))
The extension of the presented MPC-NO algorithm ⎪
⎩ i=1
if l = τ, . . . , nB .
with neural networks to systems with many inputs and
many outputs (MIMO) is discussed in (Ławry ńczuk,
Let
2003) and (Ławry ńczuk and Tatjewski, 2001a). As the


model of the process n y (= dim y) MISO nonlinear mod- al (k) = al x(k) , bl (k) = bl x(k) (45)
els (neural networks) are used. and redefine the variables


y(k) := y(k) − g x(k) ,
5.3. Linearization-Based MPC Algorithm with
y(k − i) := y(k − i) − y(k − i), l = 1, . . . , nA ,
a Neural Network Model
u(k − i) := u(k − i) − u(k − i), l = 1, . . . , nB .
As was mentioned in Section 4.2, one of the most precise (46)
and numerically effective combinations is a nonlinear free The linear approximation of the model (23), obtained at
response prediction together with the use of a linearized the sampling instant k, can be expressed as
model for forced response calculations, i.e., for on-line
optimization in the form of a QP problem. A(k, z −1 )y(k) = B(k, z −1 )u(k), (47)
Soft computing in model-based predictive control 19

where The extension of the presented MPC-NPL algorithm with

neural networks to systems with many inputs and many
A(k, z −1 ) = (1 + a1 (k)z −1 + · · · + anA (k)z −nA ), outputs (MIMO) is discussed in (Ławry ńczuk, 2003).

B(k, z −1 ) = (b1 (k)z −1 + · · · + bnB (k)z −nB ). (48) The stability of the presented MPC-NPL and NO
algorithms with neural models is achieved by properly
It can be noted that the linearization point given by (41) tuning the weighting coefficient λ in the cost function
and hence the coefficients a l (k), bl (k) are not influenced J(k). Both of these MPC-NPL and MPC-NO algorithms
by the most recent output value y(k), which is available can be combined with the stabilizing dual-mode approach
to measurements. It may be crucial in the case of fast (Ławryńczuk and Tatjewski, 2004), (Ławry ńczuk, 2003)
processes. Therefore, it is then recommended to use developed by Michalska and Mayne (1993). The control
law is calculated as the solution to the standard MPC op-

x(k) = u(k − τ + 1) . . . u(k − nB + 1) y(k) timization problem if the current state of the system lies
outside a neighborhood of the origin while a local, usually
T
. . . y(k − nA + 1) . (49) linear, controller is used otherwise. An additional inequal-
ity constraint is imposed on the predicted terminal state.
If τ = 1, for linearization purposes one may set u(k) = In this approach merely feasibility, rather than optimality,
u(k − 1) or u(k) = u(k|k − 1). The MPC-NPL al- is sufficient to guarantee stability.
gorithm using the linearization point (41) will be named
NPL1 whereas the one which uses (49) – NPL2.
5.4. Reducing Computational Complexity in MPC
The nonlinear free response y 0 (k + p|k), p = with Neural Networks
1, . . . , N , is calculated recursively from the general pre-
diction equation (31), taking into account the output of The MPC-NO algorithm is computationally demanding
the neural model given by (33) and the DMC disturbance and the computation time is much longer than that of
model (32): linearization-based algorithms. Moreover, the main dif-
K
ficulty comes from the fact that the optimization problem

 that has to be solved on-line is usually nonconvex and it
y 0 (k + p|k) = w2 (0) + w2 (i)ϕ zi0 (k + p) + d(k).
i=1
may terminate in local minima. That is why linearization-
(50) based MPC approaches attract so much attention and are
The quantities zi0 (k
+ p) are determined from (34) as- widely used in industry (Henson, 1998; Morari and Lee,
suming no changes in control signals from the sampling 1999). On the other hand, to reduce the computational
instant k and replacing the predicted output signal from complexity, a few neural-network based alternatives have
k + 1 by the corresponding values of the free response: been suggested. In general, these approaches can be di-
vided into two groups: in the first one, special structures
u(k + p|k) := u(k − 1), p ≥ 0, of neural models are used to make the optimization prob-
(51) lem simpler (convex), while in the second, explicit ap-
y(k + p|k) := y 0 (k + p|k), p ≥ 1. proximate MPC algorithms (without on-line optimization)
combined with neural networks are used.
Hence
5.4.1. Neural Network Based MPC with On-Line
zi0 (k + p) = w1 (i, 0)
Optimization
Iuf (p)
 Even though linearization-based MPC algorithms result in
+ w1 (i, j)u(k − 1)
the quadratic optimization problem which poses no nu-
j=1
merical difficulty, in some specific applications, namely,
Iu
 in the case of large-scale or fast-sampling systems, it is
+ w1 (i, j)u(k − τ + 1 − j + p) necessary to reduce the computational burden. An inter-
j=Iuf (p)+1 esting neural approach is presented in (Wang and Wan,
2001). A structured neural network that implements the
Iyp (p)
 gradient projection algorithm is developed to solve the
+ w1 (i, Iu + j)y 0 (k − j + p|k) constrained QP problem in a massively parallel fashion.
j=1 Specifically, the structured network consists of a projec-
nA
 tion network and a network which implements the gradi-
+ w1 (Iu + j)y(k − j + p). (52) ent projection algorithm. The projection network consists
j=Iyp (p)+1 of specially structured linear neurons. A training algo-
 20 P. Tatjewski and M. Ławry ńczuk

rithm is formulated for which the convergence is guaran- The coefficients of the dynamic model, v i and wi , which
teed. The networks are trained off-line, whereas the con- depend on the current state of the plant are calculated from
trols are calculated on-line from the networks without any the neural static model. The resulting MPC optimization
optimization. task is not convex. Nevertheless, the model is relatively
In addition to linearization-based MPC schemes simple and the approach is reported to be successful in
which use the QP approach, it is also possible to develop a many industrial applications.
specially structured neural model to avoid the necessity of
nonlinear optimization. In (Liu et al., 1998), affine non- 5.4.2. Neural Network Based MPC without On-Line
linear models of the following form are considered: Optimization


 In recent years approximate MPC algorithms have at-
y(k) = F y(k) + G y(k) u(k − τ ), (53)
tracted much attention (Bemporad et al., 2002; Johansen,
2004). The key idea is to calculate control signals on-line
where F (·) and G(·) are nonlinear functions and
without any optimization. The main advantage of this ap-
 T proach is its speed. On the other hand, the control law
y(k) = y(k − 1) . . . y(k − nA ) . (54)
must be precomputed off-line and stored somehow in the
controller’s memory. Because, in general, a very large
A set of nonlinear affine predictors of the following struc-
amount of computer memory may be necessary to store
ture is used in the MPC algorithm:
the control law, it is justified to find its approximation to
y(k + τ + p|k) efficiently compute control on-line.
p Bemporad et al. (2002) show that for linear systems

 

= Fp x(k) + Gpj x(k) u(k + j|k), (55) with linear constraints and a quadratic cost function the
j=0 optimal control action is a continuous and piecewise affine
function of the state. The state space is partitioned into
where p = 0, . . . , N . The quantities F p (x(k)) and polyhedral sets and a control law is calculated for every
Gpj (x(k)), which depend on the current state of the plant set. This results in a search tree which determines the set
 T to which a given state belongs. The tree is computed off-
x(k) = y(k) . . . y(k − nA ) u(k − 1) . . . u(k − τ ) , line, but the classification and calculation of the control
(56) signals is performed on-line. The main advantage is the
are calculated by neural networks. The key idea is that the fact that the necessity of on-line optimization is avoided.
present and future controls, i.e., the decision variables of Unfortunately, the number of polyhedral sets can be huge,
the optimization problem occur linearly in the predictor’s which affects the overall computational efficiency. Neural
equation (55). The predictor depends in a nonlinear way networks can be effectively used to improve the situation
only on the past values of input and output signals. Hence, (Haimovich et al., 2003). The solution to the constrained
the resulting MPC optimization problem is convex. In linear MPC optimization problem is pre-computed off-
some cases, however, such an approach to modeling may line in an explicit form as a piecewise affine state feedback
turn out to be insufficient to capture the nonlinearity pre- law defined on polyhedral regions of the state space (Be-
cisely enough. In (Liu et al., 1998), the corresponding mporad et al., 2002). The problem of determining on-line
on-line training algorithm and the MPC scheme based on in which polyhedral region the state lies is solved by using
such predictors are discussed. a neural network, i.e., for a given state of the system the
network determines the corresponding region. An appar-
Neural networks are also used in commercially avail- ent advantage of this approach is that regions that have the
able software packages, e.g., Process Perfecter from Pavil- same control law are joined. The structure of the closed-
ion Technologies Inc. (Piche et al., 2000). In this inter- loop system is depicted in Fig. 10. Unlike the approxi-
esting approach, to simplify the identification task and
solve the MPC optimization problem on-line, a neural net-
work is used to capture the steady-state properties of the
process. A second-order quadratic-in-the-input dynamic
model is used,

δy(k)
2 
 
2
= vi δu(k−i)+wi (δu(k−i)) −ai δy(k−i) . Fig. 10. Approximate neural MPC scheme with
i=1 neural selection of the control law.
(57)
Soft computing in model-based predictive control 21

mate neural controller which replaces the whole MPC al- then approximated by n u = dim u single-output feedfor-
gorithm as in (Parisini et al., 1998; Parisini and Zoppoli, ward neural networks with one hidden layer with nonlin-
1995), the neural network is not used for finding the actual ear transfer functions and linear output units. Networks
control directly. Comparing to other solutions, the advan- are trained off-line. Bounds on the approximation are es-
tage of the neural approach is the fact that the complexity tablished. The main parameters of the algorithm are de-
of the neural network grows linearly with the number of termined partly analytically (the scalar a and the penalty
control laws. The multilayer feedforward neural network matrix P ) and partly by solving specific global optimiza-
with two hidden layers is used. tion problems (the number of hidden units in the neural
Another version of the neural approach for the “ex- network). Because neural networks work as approxima-
plicit” constrained MPC scheme for linear systems is de- tors, nonoptimal control values may be obtained. It can be
scribed in (Hoekstra et al., 2001). It is known that the proved, however, that, being suboptimal, the algorithm re-
solution to the standard MPC optimization problem is a mains stable. Training neural networks is a serious issue,
continuous function of the state, the reference signal, the because a sufficiently large set of patterns is necessary to
noise and the disturbances. Hence, it can be approximated cover the whole operation domain.
arbitrarily well by a feedforward neural network. In order An interesting hybrid approximate neural predictive
to develop a analytic predictive controller, at first the stan- controller is presented in (Parisini and Sacone, 2001). A
dard one, an optimization-based one, has to be developed. hybrid scheme consists of two control layers: a continuous
Closed-loop simulations have to be performed to gather one, in which a set of neural approximate MPC regulators
examples to cover the operation domain. Because the ob- are used, and a supervisory discrete-event one, aimed at
tained controller is only an approximation of the real one, choosing the best control action (from those proposed by
precise satisfaction of constraints cannot be guaranteed. the regulators) which should be applied to the plant. The
A trivial solution is to saturate the output of the neural choice depends on the current condition of the system and
network. The constraints cannot be modified for a given on external events. Such an approach may be efficient
neural network. for complex systems with many decision variables. The
In (Parisini et al., 1998; Parisini and Zoppoli, 1995), stability of the overall hybrid system is guaranteed.
an interesting nonlinear stabilizing approximate neural A predictive neural-network controller of uncertain
predictive controller is suggested. The structure of the systems is investigated in (Vila and Wagner, 2003). In this
closed-loop system is depicted in Fig. 11. The overall pro- solution neural networks are used both for modeling and
cedure consists of two steps. At first, a stabilizing MPC minimization. More specifically, to increase the accuracy
algorithm is developed. The formulation used allows us of the prediction, the model consists of N separate net-
to take into account constraints imposed on system in- works. Both the system and controller neural networks are
puts and states. Stability is enforced by a proper termi- first selected off-line by a statistical Bayesian procedure in
nal penalty term in the cost function, and no stabilizing order to make the predictive controller work effectively.
terminal equality constraints are necessary: An off-line selection procedure of the neural predictor’s
and neurocontroller’s architecture are discussed, and the
N
 −1

 issue of the stochastic stability of the closed loop is con-
J(k) = h x(k + p|k), u(k + p|k)
sidered as well.
p=0

+ a x(k + N |k)2P , (58)

5.5. Applications and Exemplary Simulation Results
where h(x(k + p|k), u(k + p|k)) is the cost function (in
general, a nonquadratic one), a is a positive scalar and P MPC algorithms with neural network models of different
is a positive definite symmetric matrix. The algorithm is structures have been applied to a wide class of processes,
for example, a combustion system (Liu and Daley, 1999),
a pneumatic servo system (Nørgaard et al., 2000), a mo-
bile robot (Ortega and Camacho, 1996), an industrial
packed bed reactor (Temeng et al., 1995), an insulin deliv-
ery problem (Trajanoski and Wach, 1998), a multivariable
chemical reactor (Yu and Gomm, 2003), traffic control
on freeways (Parisini and Sacone, 2001), and a biologi-
cal depolluting treatment of wastewater (Vila and Wag-
ner, 2003). In the following part of the article, simulation
examples of a control system with MPC using neural net-
Fig. 11. Approximate neural MPC scheme. works will be given.
 22 P. Tatjewski and M. Ławry ńczuk

Fig. 12. High-purity high-pressure ethylene-ethane distillation column control system.

5.5.1. High-Purity High-Pressure Ethylene-Ethane Four models of the plant were used. The first one
Distillation Column was used as the real process during the simulations. It
was based on technological considerations (Ławry ńczuk,
The plant under consideration is a high purity, high
2003). An identification procedure was carried out, and
pressure (1,93 MPa) ethylene-ethane distillation column
as a result two linear models for different operating points
shown in Fig. 12 (Ławry ńczuk, 2003). The feedstream
and a neural one were obtained. For the empirical models
consists of ethylene (approx. 80%), ethane (approx. 20%),
we have nA = 1, τ = nB = 3. The horizons were set to
and traces of hydrogen, methane and propylene. The
N = 10, Nu = 3, the weighting coefficient λ = 2. In all
product of the distillation is ethylene which can con-
the simulations it is assumed that at the sampling instant
tain up to 1000 ppm (parts per million) of ethane. The
k = 1 the set-point value is changed from 100 ppm to
main problem is to develop a supervisory controller which
350 ppm, 600 ppm and 850 ppm. Because of some tech-
would be able to increase relatively fast the impurity level
nological reasons, the following constraints were imposed
when composition changes in the feedstream are relatively
on the reflux ratio: r min = 4.051, rmax = 4.4571.
small. Reducing the purity of the product, of course taking
into account technological limits, results in a decreased At first, MPC algorithms based on two linear mod-
energy consumption. The production scale is very large, els were developed. The first linear model is valid for a
the nominal value of the product stream being equal to “low” impurity level and the resulting control algorithm
43 tons/h. The column has 121 trays and the feedstream works well in this region, but exhibits unacceptable os-
is delivered to the tray number 37. cillatory behavior for medium and big setpoint changes,
Two fast single-loop PID controllers (denoted by LC) as is shown in Fig. 13. On the contrary, the second lin-
are used to maintain the levels in the reflux tank and the ear model captures the process properties for a “high” im-
bottom product tank. Yet another PID controller (denoted purity level and the closed-loop response is fast enough
by TC) is also used to control the temperature on the tray for the biggest setpoint change, but very slow for smaller
number 13. The PID controllers comprise the basic con- ones, as is shown in Fig. 14.
trol layer. As far as the supervisory MPC algorithm is con-
cerned, the control loop has one input variable r, which is Simulation results of MPC-NPL algorithms with a
the reflux ratio r = R/P , where R is the reflux stream neural network are depicted in Fig. 15. Both algorithms
delivered to the column by the top tray and P is the prod- work well for all three setpoint changes. The NPL1 algo-
uct stream taken from the tray number 110, and one out- rithm is slightly slower than NPL2. Simulation results of
put variable z, which represents the impurity level in the the MPC-NO algorithm with a neural network are shown
product (ethylene). The sampling time of the MPC al- in Fig. 16. Compared with suboptimal linearization-based
gorithm is relatively long (a slow composition analyzer), algorithms, nonlinear optimization leads to faster closed-
equal to T p = 40 min. loop responses.
Soft computing in model-based predictive control 23

sp
r z , z [ppm]
4.35
850
4.3

4.25
600

4.2

4.15
350

4.1

4.05 100
1 5 10 15 20 25 1 5 10 15 20 25
k k
Fig. 13. Simulation results of the ethylene-ethane distillation column with the
MPC algorithm based on a linear model for a “low” impurity level.

r sp
4.35
z , z [ppm]
850
4.3

4.25
600

4.2

4.15
350

4.1

4.05 100
1 5 10 15 20 25 1 5 10 15 20 25
k k
Fig. 14. Simulation results of the ethylene-ethane distillation column with the
MPC algorithm based on a linear model for a “high” impurity level.
sp
r z , z [ppm]
4.35
850
4.3

4.25
600

4.2

4.15
350

4.1

4.05 100
1 5 10 15 20 25 1 5 10 15 20 25
k k
Fig. 15. Simulation results of the ethylene-ethane distillation column with MPC-NPL1 (dashed
line) and MPC-NPL2 (solid line) algorithms with neural networks.
24 P. Tatjewski and M. Ławry ńczuk

sp
r z , z [ppm]
4.35
850
4.3

4.25
600

4.2

4.15
350

4.1

4.05 100
1 5 10 15 20 25 1 5 10 15 20 25
k k
Fig. 16. Simulation results of the ethylene-ethane distillation column with the MPC-NO algorithm with a neural network.

sp
r z , z [ppm]
4.35
850
4.3

4.25
600

4.2

4.15
350

4.1

4.05 100
1 5 10 15 20 25 1 5 10 15 20 25
k k
Fig. 17. Simulation results of the ethylene-ethane distillation column with the MPC-NPL2 (dashed line)
and MPC-NO (solid line) algorithms with neural networks and the constraint Δrmax = 0.03.

sp
r z , z [ppm]
4.35
850
4.3

4.25
600

4.2

4.15
350

4.1

4.05 100
1 10 20 30 40 50 1 10 20 30 40 50
k k
Fig. 18. Simulation results of the ethylene-ethane distillation column with the MPC-NPL2
algorithm with a neural network and unmeasured disturbances.
Soft computing in model-based predictive control 25

In practice, big changes in the manipulated variable that the research area concerned with using soft comput-
r are not allowed because of technological and safety rea- ing techniques in predictive control is still open and many
sons (high pressure, big production scale). That is why new results appear every year.
an additional constraint Δr max = 0.03 was used. Fig-
ure 17 compares simulation results of the MPC-NPL2 and
MPC-NO algorithms with a neural network. Although the References
constraint significantly slows closed-loop responses, it can
still be noticed that the MPC-NO algorithm is somewhat
Allgöwer F., Badgwell T.A., Qin J.S., Rawlings J.B. and
faster. Wright S.J. (1999): Nonlinear predictive control and mov-
Simulation results of the MPC-NPL2 algorithm ing horizon estimation – An introductory overview, In: Ad-
and unmeasured stochastic disturbances are presented in vances in Control – Highlights of ECC’99 (P.M. Frank,
Fig. 18. Such disturbances are unavoidable in industry. ed.).— London: Springer, pp. 391–449.
Babuška R., Sousa J.M. and Verbruggen H.B. (1999): Predic-
tive control of nonlinear systems based on fuzzy and neural
models. — Proc. European Control Conf., Karlsruhe, Ger-
6. Conclusions many, (on CD-ROM).
The subject of the paper was applications of soft com- Bazaraa M.S., Sherali J. and Shetty K. (1993): Nonlinear Pro-
puting methods to model-based predictive control tech- gramming: Theory and Algorithms. — New York: Wiley.
niques. The resulting algorithms make it possible to Bemporad A., Morari M, Dua V. and Pistikopoulos E.N. (2002):
effectively control highly nonlinear, multidimensional The explicit linear quadratic regulator for constrained sys-
processes, usually subject to constraints which result from tems. — Automatica, Vol. 38, No. 1, pp. 3–20.
technological and safety reasons. The algorithms consid- Brdyś M.A. and Tatjewski P. (2005): Iterative Algorithms for
ered can be easily implemented and used on-line. Multilayer Optimizing Control. — London: Imperial Col-
lege Press/World Scientific.
As far as MPC algorithms using fuzzy techniques
are concerned, explicit (unconstrained) multimodel algo- Camacho E.F. and Bordons C. (1999): Model Predictive Con-
rithms exploiting the fuzzy Takagi-Sugeno (TS) structure trol. — London: Springer.
and algorithms with on-line linearization of a fuzzy TS Clarke D.W., Mohtadi C. and Tuffs P.S. (1987): Generalized
model were described. In the first case the values of con- predictive control—I. The basic algorithm. — Automatica,
trol variables can be computed by a set of explicit formu- Vol. 23, No. 2, pp. 137–148.
las, which makes the application very easy. In the sec- Cutler R. and Ramaker B.L. (1979): Dynamic matrix control –
ond case (MPC-NPL algorithms), decision variables are A computer control algorithm. — Proc. AIChE National
obtained by solving a QP optimization problem, which al- Meeting, Houston.
lows taking constraints explicitly into account. Eder H.H. (1999): MBPC benefits and key success factors. —
A family of MPC algorithms using artificial neural Proc. 5-th European Control Conf., ECC’99, Karlsruhe,
networks (i.e., the most popular multilayer perceptron) Germany, (on CD-ROM).
were also described. In comparison with fuzzy models, Findeisen W., Bailey F.N., Brdys M., Malinowski K., Tatjew-
neural structures do not suffer from the “curse of dimen- ski P. and Wozniak A. (1980): Control and Coordination
sionality”, which is troublesome in multivariable cases. in Hierarchical Systems. — Chichester: Wiley.
Three classes of solutions were considered. Algorithms Findeisen W. (1997): Control Structures for Complex Processes.
with nonlinear optimization are potentially very precise, — Warsaw: Warsaw University of Technology Press, (in
but they hinge on the effectiveness of the optimization Polish).
routine used. Because, in practice, convergence to a global Garcia C.E. (1984): Quadratic/dynamic matrix control of non-
optimum cannot be guaranteed, MPC-NPL algorithms linear processes: An application to a batch reaction
with on-line linearization of a neural network model were process. — Proc. AIChE Annual Meeting, San Francisco.
also presented. A number of approaches aimed at re- Haimovich H., Seron. M.M., Goodwin G.C. and Agüero J.C.
ducing the computational complexity of the optimization (2003): A neural approximation to the explicit solution of
problem or to approximate the whole control algorithm constrained linear MPC. — Proc. European Control Conf.,
were also reviewed. Cambridge, UK, (on CD-ROM).
In the paper, emphasis was put on computational ef- Haykin S. (1999): Neural Networks – A Comprehensive Foun-
ficiency. It was shown that the application of fuzzy rea- dation. — Englewood Cliffs, NY: Prentice Hall.
soning and neural networks leads to efficient realizations Henson M.A. (1998): Nonlinear model predictive control: Cur-
of nonlinear predictive control algorithms, suitable for ef- rent status and future directions. — Comput. Chem. Eng.,
fective on-line applications. It should be also mentioned Vol. 23, No. 2, pp. 187–202.
 26 P. Tatjewski and M. Ławry ńczuk

Hoekstra P., van den Boom T.J.J. and Botto M.A. (2001): De- Nørgaard M., Ravn O., Poulsen N.K. and Hansen L.K. (2000):
sign of an analytic constrained predictive controller using Neural Networks for Modelling and Control of Dynamic
neural networks. — Proc. European Control Conf., Porto, Systems. — London: Springer.
Portugal, (on CD-ROM). Ortega J.G. and Camacho E.F. (1996): Mobile robot navigation
Johansen T.A. (2004): Approximate explicit receding horizon in a partially structured static environment, using neural
control of constrained nonlinear systems. — Automatica, predictive control. — Contr. Eng. Pract., Vol. 4, No. 12,
Vol. 40, No. 2, pp. 293–300. pp. 1669–1679.
Liu G.P. and Daley S. (1999): Output-model-based predictive Parisini T. and Sacone S. (2001): Stable hybrid control based on
control of unstable combustion systems using neural net- discrete-event automata and receding-horizon neural reg-
works. — Contr. Eng. Pract., Vol. 7, No. 5, pp. 591–600. ulators. — Automatica, Vol. 37, No. 8, pp. 1279–1292.
Liu G.P., Kadirkamanathan V. and Billings S.A. (1998): Predic- Parisini T., Sanguineti M. and Zoppoli R. (1998): Nonlinear sta-
tive control for non-linear systems using neural networks. bilization by receding-horizon neural regulators. — Int. J.
— Int. J. Contr., Vol. 71, No. 6, pp. 1119–1132. Contr., Vol. 70, No. 3, pp. 341–362.
Ławryńczuk M. (2003): Nonlinear model predictive control al- Parisini T. and Zoppoli R. (1995): A receding-horizon regula-
gorithms with neural models. — Ph.D. thesis, Warsaw Uni- tor for nonlinear systems and a neural approximation. —
versity of Technology, Warsaw, Poland. Automatica, Vol. 31, No. 10, pp. 1443–1451.
Ławryńczuk M. and Tatjewski P. (2001a): A multivariable Piche S., Sayyar-Rodsari B., Johnson D. and Gerules M. (2000):
neural predictive control algorithm. — Proc. IFAC Ad- Nonlinear model predictive control using neural networks.
vanced Fuzzy-Neural Control Workshop, Valencia, Spain, — IEEE Contr. Syst. Mag., Vol. 20, No. 3, pp. 56–62.
pp. 191–196. Qin S.J. and Badgwell T.A. (2003): A survey of industrial
Ławryńczuk M. and Tatjewski P. (2001b): A nonlinear predic- model predictive control technology. — Contr. Eng. Pract.,
tive control algorithm for processes modelled by means Vol. 11, No. 7, pp. 733–764.
of neural networks. — Proc. 7-th Int. Conf. Methods Rossiter J.A. (2003): Model-Based Predictive Control. — Boca
and Models in Automation and Robotics, Mi˛edzyzdroje, Raton, FL: CRC Press.
Poland, pp. 489–494.
Tatjewski P. (2002): Advanced Control of Industrial Processes.
Ławryńczuk M. and Tatjewski P. (2002): A computationally Structures and Algorithms. — Warsaw: Akademicka Ofi-
efficient nonlinear predictive control algorithm based on cyna Wydawnicza EXIT, (in Polish, revised and extended
neural models. — Proc. 8-th Int. Conf. Methods and English edition in preparation).
Models in Automation and Robotics, Szczecin, Poland,
Temeng K.O., Schnelle P.D. and McAvoy T.J. (1995): Model
pp. 781–786.
predictive control of an industrial packed bed reactor using
Ławryńczuk M. and Tatjewski P. (2004): A stable dual-mode neural networks. — J. Process Contr., Vol. 5, No. 1, pp. 19–
type nonlinear predictive control algorithm based on on- 27.
line linearisation and quadratic programming. — Proc.
Trajanoski Z. and Wach P. (1998): Neural predictive control
10-th Int. Conf. Methods and Models in Automation and
for insulin delivery using the subcutaneous route. — IEEE
Robotics, Mi˛edzyzdroje, Poland, pp. 503–510.
Trans. Biomed. Eng., Vol. 45, No. 9, pp. 1122–1134.
Maciejowski J.M. (2002): Predictive Control. — Harlow: Pren-
Vila J.P. and Wagner V. (2003): Predictive neuro-control of un-
tice Hall.
certain systems: Design and use of a neuro-optimizer. —
Marusak P. and Tatjewski P. (2002): Stability analysis of non- Automatica, Vol. 39, No. 5, pp. 767–777.
linear control systems with unconstrained fuzzy predictive Wang L.X. and Wan F. (2001): Structured neural networks
controllers. — Arch. Contr. Sci., Vol. 12, pp. 267–288. for constrained model predictive control. — Automatica,
Mayne D.Q., Rawlings J.B., Rao C.V. and Scokaert P.O.M. Vol. 37, No. 8, pp. 1235–1243.
(2000): Constrained model predictive control: Stability Yu D.L. and Gomm J.B. (2003): Implementation of neural net-
and optimality. — Automatica, Vol. 36, No. 6, pp. 789– work predictive control to a multivariable chemical reac-
814. tor. — Contr. Eng. Pract., Vol. 11, No. 11, pp. 1315–1323.
Michalska H. and Mayne D.Q. (1993) (2000): Robust receding
horizon control of constrained nonlinear systems. — IEEE
Trans. Automat. Contr., Vol. 38, No. 11, pp. 1623–1633.
Received: 18 October 2005
Morari M. and Lee J.H. (1999): Model predictive control: Revised: 21 December 2005
Past, present and future. — Comput. Chem. Eng., Vol. 23,
No. 4/5, pp. 667–682.
Najim K., Rusnak A., Meszaros A. and Fikar M. (1997): Con-
strained long-range predictive control based on artificial
neural networks. — Int. J. Syst. Sci., Vol. 28, No. 12,
pp. 1211–1226.