Vous êtes sur la page 1sur 3

Proceedings of the 37th IEEE Conference on Decision & Control Tampa, Florida USA December 1998

TP09 16:40

Fuzzy model-based predictive control


Mohamed Laid Hadjili, Vincent Wertz and Gbrard Scorletti'
Centre for Systems Engineering and Applied Mechanics Universite catholique de Louvain 4, av. G. Lemaitre B-1348 Louvain-la-Neuve, Belgium E-mail: hadjili@csam.ucl.ac.be

Abstract
In this paper, we focus on an extension of the predictive control approach. Predictive control has been developed to control Linear Time Invariant plants, described by ARIMAX models. We discuss the extension of this method in the case when the behavior of the plant is modeled using fuzzy modeling. 1 Introduction Generalized predictive control (GPC) is a popular control method, especially for industrial processes (see e.g. [l,2 and references therein). It combines the predic1 tion of the future behavior of the plant with feedback control. A process model is used to predict the future behavior of the plant over a given time horizon, in the case when the current state and the future perturbation inputs are known. From this prediction, a control command is computed by minimizing a quadratic cost function allowing a tradeoff between performance and control cost. In this approach, the most important issue is to get a model of the plant. The plants can be highly non linear and time-varying which are often too complicated for first principle modeling and classical linear identification. In this context, fuzzy models appear as an appealing modeling tool in the case of complicated systems, where conventional approaches failed or their application is too intricate (see e.g. [3]). Predictive control for fuzzy dynamic models was recently proposed [4, 51. All methods are derived using the TS fuzzy model. In the first paper, Finite Impulse Response (FIR) models are considered. Babuska investigates a specific class of TS models, where the consequent takes constant values (the so-called singleton fuzzy models). In this work, we investigate the problem of predictive control for TS fuzzy models where the consequents are ARX models and we focus our interest on the way to compute multistep ahead predictions for these fuzzy models.

2 Fuzzy models Fuzzy models are based on a set of IF-THEN rules: IF antecedent THEN consequent. In this paper, we focus on a special case of the Takagi-Sugeno (TS) fuzzy model where the offsets are zeros [6](the so-called homogeneous TS model) as controller synthesis and stability analysis are easier in that case [7, 51. The one-step ahead predictive TS dynamic fuzzy model is described by a set of N, fuzzy rules as follows:
R,: IF
THEN
Y is Ai ~ ( k1) = OTY

(1)

is

where Y = {y(k), y(k - l),...,u ( k ) , u ( k- l),...} the set of input/output values, R = { A I , + , ...,Bl,i, B2,i, ...} is the set of Az,i, membership functions associated to the ith rule and 0 = ai,^ : ai,n bi,l * . . bi,,] is the parameter vector of the ith LTI sub-model. We define the degree of fulfillment of the ith fuzzy rule
s.s:

( 2)

where wi is the firing strength of the ith fuzzy rule:


n

where 0 5 wi

5 1 and C z lwi > 0


al (k)y(k -

The TS fuzz1 model can be reprTented as:


y(k

+ 1) =
I=1

z + 1) +
1=1

61 ( k ) u (k - 2 + 1)
NC

(4)

NC

where a l ( k ) = E X i ( k ) a i , l , & ( k )
i=l

= E X i ( k ) b i , l are
i=l

time-varying parameters. The Fuzzy model eq.(4) defines a linear time-varying (LTV) model, since the parameters Til(k) and bl(k) are time varying. This makes the computation of multistep predictions more complicated. By assuming that the time variation is slow enough, LTI models can still be used for future predictions as described in the next section. An alternative to this is to construct

IThis paper presents research results of the Belgian Programme on Interuniversity Poles of Attraction, initiated by the Belgian State, Prime Minister's Office for Science, Technology and Culture. The scientific responsibility rests with its authors. 0-7a03-4394-819a$10.00 1998 IEEE 2927

a new TS fuzzy model with identical antecedents but with the consequent consisting of several predictors up to K N based on informations available at time k. These predictors can be written as follows:

G( IC + j + 1)

=
1=1

6 (k)( k : y

- z + 1)
(5)

1=1

1=1

where j = 1,2,.., N, diction),


NC
i=1
NC

(N,defines the horizon of pre=

a:'(lc) = C X i ( k ) a : , , , % ( I C )
$(IC)
=
i= 1

CXZ(IC)b:,, and
i= 1

NC

X i ( k)g$

are the time-varying parameters.

Compute the incremental control action for each rule. 0 The global control value is the center average defuzzification of incremental actions computed separately for each rule. 3.2 Controller C 2 0 Compute the degree of fulfillment at step k by eq.(2) and subsequently the LTV model. Assume that X i is constant over the prediction horizon. Compute the optimal j-step ahead predictors for the LTI model (4) with constant Xi. 0 Compute the control increment value which minimizes the quadratic cost function. 3.3 Controller C3 0 Construct the TS fuzzy model (5) with multistep predictions of the output. 0 Compute the control increment value which minimizes the quadratic cost function.
4 Simulation In this example we design a predictive controller based on the linear TS (LTS) fuzzy model for a glass furnace [8]. We use subtractive clustering to parameterize membership functions (see figure 1) and least squares algorithm to train linear sub-models parameters. The plant is described by 2 fuzzy rules with 3 inputs (y(lc),u(lc),u(k- 1)) and one output y ( k + l ) . We

3 Fuzzy model-based Predictive Control

A straightforward extension of predictive control based on fuzzy models is difficult, because the combination of the N, LTI sub-models is an LTV model as expressed by eq.(4). The main difficulty is to generate future predictions of the system using a model which is linear but time varying. Future values of time-varying parameters depend on the predictive control inputs.To deal with this problem, we propose two strategies:
1. Compute the control action for each rule using GPC technique based on the considered LTI sub-model. The control action computed for each rule minimizes, independently of the degree of fulfillement of this rule, the associated quadratic criterion. The global control action is the convex combination of the control actions computed for each fuzzy rule.

2. Firstly construct the LTV model by convex combination of the N, LTI sub-models. Compute the control action using GPC criterion minimizing directly a global criterion based on the LTV model. In order to compute the N-step ahead predictions from this LTV model, we investigate two alternatives: Assume that the rate of variation of the degree of fulfillment can be neglected over the considj ) = &(IC) for ered prediction horizon: &(IC j = 1,2, ...,Nzl. For the control computation, the LTV model becomes an LTI model. A direct approach is to construct a fuzzy model that expresses the N-step ahead behavior of the process as in eq.(5). As a consequence, the degree of fulfillment is, by definition, constant for all predictions up to N-step ahead. So we propose three algorithms for fuzzy predictive controller design:

Figure 1: Two rules-based fuzzy model constructed via Subtractive clustering


design a predictive controller based on this TS fuzzy I model with the following parameters: N = l , N y = Nu = 3. The weighting constants are: a1 = a2 = 0.2. The simulation results with C1 are given in figure 2. The second algorithm C2 is used for the same process with the same design parameters. Figure 3 presents r e sults of simulation using this controller. The controller C3 is designed based on a 3-step ahead prediction fuzzy model. The computa.tiona1 complexity is higher in this

3.1 Controller C1
0

Compute the optimal j-step ahead predictors for each rule.

2928

20

o .

eo

eo
a,"*

100

120

1.0

?(LO

1.70

Figure 2: LTS fuzzy model response for predictive controller C1, ym: model response; r: reference signal

m:") 20

o .

so

00

100

3 2 0

01 .

,eo

3 - 0

case than for the first two algorithms. Figure 4 presents the LTS fuzzy model response for predictive controller c3.

trols obtained for each rule, using a fuzzy controller model similar to the fuzzy model of the process. Note however that in this approach, even though the controls computed for each sub-model are optimal, in the sense that they minimize the GPC criterion, the combination of controls is not optimal itself. The second approach aims at finding a control which globally minimizes the GPC criterion. Since the TS fuzzy model is effectively LTV, a solution is to assume that the degree of fulfillment of the individual rules are constant over the prediction horizon. Hence, we perform the minimization of the cost function for the global TS model, assuming that the parameters of the LTV model are constant and, as GPC is a receding control strategy, only the first control increment is applied. The simplifying assumption is thus not as dramatic as it could appear at first sight. The third proposed algorithm is based on a TS fuzzy model where each rule directly provides a set of predictions up to N,-step ahead. The global predictions are then obtained using convex combinations of the values given by the sub-models, and the control action is computed by minimizing the criterion using these global predictions. The last two approaches aim at minimizing the cost function globally instead of minimizing it independently for each rule. One should expect better performance in these two cases, although the computational complexity is of course higher. The stability analysis of these control schemes will be outlined in a future work.

References
D. W. Clarke, C. Mohtadi, and P. S. Tuffs. Generalized predictive control-part i: The basic algorithm. Automatica, 23(2):137-148, 1987. R. R. Bitmead, M. Gevers, and V. Wertz. Adaptive [2] optimal control, the thinlcing man's GPC. Prentice Hall International, 1990. [3] T. Takagi and M. Sugeno. Fuzzy identification of systems and its application to modeling and control. IEEE 7kans. Sys. Man and Cyber., 15(1):11&132, 1985. Y. Nakamori, K. Suzuki, and T. Yamanaka. Model [4] predictive control using fuzzy dynamic models. In IFSA '91, volume Engineering, pages 135-138, July 1991. [5] R. Babuska. Fuzzy modeling and identification. PhD thesis, Delft University, Netherlands, 1996. C. Fantuzzi and R. Rovatti. On the approximation [6] capabilities of the homogeneous takagi-sugeno model. In Fifth IEEE International Conference on Fuzzy Systems, New Orleans, USA, pages 1067-1072, 1996. [7] J. Zhao. Fuzzy logic in modeling and control. PhD thesis, CESAME, UCL University, Belgium, CESAME, UCL, Belgium, 1996. [e] M. L. Hadjili, A. Lendasse, V. Wertz, and S. Yurkovich. Identification of fuzzy models for a glass furnace process. In Conference on Control Application, %a t e , Itally, Sep. 1998.
[l]

am-

Figure 4: LTS fuzzy model response for predictive controller C1, ym: model response; r: reference signal

5 Conclusion In this work we have discussed different ways to synthesize a predictive controller based on a TS fuzzy model. The main difficulty is to generate future predictions of the system using a model which is linear but time varying. To deal with this problem, three solutions have been proposed. The first one relies on the fact that each sub-model is effectively time invariant. Predictive controllers for each of the submodels are then computed using classical techniques and a global control action is subsequently computed as a combination of the con-

2929

Vous aimerez peut-être aussi