Sunteți pe pagina 1din 14

Appl Intell (2011) 34: 127140 DOI 10.

1007/s10489-009-0185-8

An intelligent multiple models based predictive control scheme with its application to industrial tubular heat exchanger system
A.H. Mazinan N. Sadati

Published online: 16 June 2009 Springer Science+Business Media, LLC 2009

Abstract The purpose of this paper is to deal with a novel intelligent predictive control scheme using the multiple models strategy with its application to an industrial tubular heat exchanger system. The main idea of the strategy proposed here is to represent the operating environments of the system, which have a wide range of variation in the span of time by several local explicit linear models. In line with this strategy, the well-known linear generalized predictive control (LGPC) schemes are initially designed corresponding to each one of the linear models of the system. After that, the best model of the system and the LGPC control action are precisely identied, at each instant of time, by an intelligent decision maker scheme (IDMS), which is playing the so important role in realizing the nalized control action for the system. In such a case, as soon as each model could be identied as the best model, the adaptive algorithm is implemented on the both chosen model and the corresponding predictive control schemes. In conclusion, for having a good tracking performance, the predictive control action is instantly updated and is also applied to the system, at each instant of time. In order to demonstrate the effectiveness of

the proposed approach, simulations are carried out and the results are compared with those obtained using a nonlinear GPC (NLGPC) scheme as a benchmark approach realized based on the Wiener model of the system. In agreement with these results, the validity of the proposed control scheme can tangibly be veried. Keywords Fuzzy adaptive predictive control scheme Nonlinear generalized predictive control scheme Multiple models strategy Intelligent decision maker scheme Tubular heat exchanger system

1 Introduction The linear model based predictive control (LMBPC) scheme has been extensively used in many control areas and academic centers, since it has a good performance, as long as we are using an explicit linear model of the system. In most applications of the LMBPC family, such as linear model algorithmic control (LMAC), linear dynamic matrix control (LDMC), linear generalized predictive control (LGPC) and other related techniques, the process is represented over its operating environment by using an explicit linear model [1 10]. In this paper, the LGPC scheme is used for controlling an industrial tubular heat exchanger system. This controller is realized based on the explicit use of process model to predict the controlled variables over a specied range of time horizon. In this scheme, an optimal control is obtained by optimizing an objective function that minimizes the control effort and the error between the predicted output and the set point, during the prediction and control horizon. As we know, the LGPC method is realized based on a single xed linear model or slowly adaptive model of the system. Here, it assumes that the operating environment is either

A.H. Mazinan ( ) Islamic Azad University (IAU), South Tehran Branch, Tehran, Iran e-mail: mazinan@azad.ac.ir N. Sadati Electrical and Computer Engineering Department, University of British Columbia, Vancouver, Canada e-mail: sadati@ece.ubc.ca N. Sadati Electrical Engineering Department, Sharif University of Technology, Tehran, Iran e-mail: sadati@sina.sharif.edu

128

A.H. Mazinan, N. Sadati

time invariant or slowly time variant in the span of time. In this case, the LGPC method based on the linear models are well behaved for the linear processes, but when the operating environment region is extended, the nonlinearity of the process cannot be ignored. In practical applications such as the tubular heat exchanger system, due to the coefcients variation, the system needs to operate in multiple operating environments, which may change abruptly from one to another [11]. An appropriate strategy to improve the LGPC scheme, while we are having a nonlinear system is to use the multiple models control strategy, if the models are approximately available for different operating environments. In fact, the main idea of multiple models control approach is to determine the best model, so to activate the corresponding controller. The multiple models control strategy has been mentioned by several researchers such as Madani, Guerci, Ning, Wang, Gang and others [1226]. In the control strategy proposed, the best model identication mechanism and the nalized control action generation are realized by a new intelligent decision maker scheme (IDMS), where the identication mechanism presented is realized in association with the both fuzzy-based adaptive Kalman lter and fuzzy-based weight generation approaches and also the nalized control action generation is realized based on the soft switching technique. In line with this strategy, as soon as the best model of the system is quite identied by the proposed IDSM, the adaptive algorithm is implemented on the chosen model and the corresponding LGPC controller. Hereinafter, for having a good tracking performance, both in desired set point variation and in system coefcients variation, the nalized LGPC control action is applied to the system, at each instant of time. In fact, the system with wide and rapid variation in coefcients could easily be controlled in the strategy proposed. The remainder of this paper is organized as follows. The tubular heat exchanger system modeling is presented in Sect. 2. The proposed multiple models control strategy and the nonlinear GPC approach are presented in Sects. 3 and 4, respectively. The simulation results and the concluding remarks are nally given in Sects. 5 and 6, respectively.

Fig. 1 Diagram of a tubular heat exchanger system

area of the tube (m2 ), : uid density (kg/m3 ), v: uid velocity (m/s), x: incremental element in tube (m), Tx : temperature of x (K), d: internal diameter of the tube (m), U : overall heat transfer coefcient (W/m2 K), Cp : specic heat capacity (J/kg K). The dynamics of the heat exchanger system is described by the partial differential equations (PDEs). Thus, it is truly used as an innite dimensional system. In this case, the temperature distribution of an incremental element x, along x, based on the principle of conservation of energy, at the time t, could be given as Cp x T = Cp v(Tx Tx+ t
x ) + U d

x T (2.1)

where Cp x T denotes the accumulation of energy in t x, Cp vTx denotes the convection ow of the energy into x, Cp vTx+ x also denotes the convection ow of the energy out of x and nally U d x T represents the heat transfer to x. It should be assumed that for modeling the tubular heat exchanger system, the uid velocity variation should be negligible, i.e., to be independent of x. Also the uid temperature of the shell tube should be constant. Now, by assuming x x, the obtained PDEs describing the system could be written as Cp T T = Cp v + U d T t x (2.2)

2 Tubular heat exchanger system modeling The heat exchanger system is a process that is used to change the temperature distribution of two materials, when they are in direct or indirect contacts [2732]. It has both the inner and the shell tubes with concurrent reactions. The uid ows through the inner tube and its temperature is varied by another uid which ows concurrently around it. The temperature and the ow rate of the uid not only change with respect to time but also change along the axial direction, as shown in Fig. 1. In order to model the heat exchanger system, the following parameters are now dened : section

In such a case, by using Tt and Ts as the temperature parameters in the inner tube and the shell tube, respectively, we can have the following 1 Tt = vt Tt + at Ts at Tt ; at = U d ( sec. ) t x t t Cpt (2.3) Ts = vs Ts + as Tt as Ts ; as = U d ( 1 ) t x s s Cp sec.
s

Also by dening Tt and Ts as the outlet and the inlet of the system and assuming s = t , we can deduce the following Tt (x, s) s + at at + Tt (x, s) = Ts (x, s) x vt vt (2.4)

Hence, the outlet temperatures in terms of the inlet temperatures and the x, could be deduced as Tt (x, s) = exp x at (s + at ) + Ts (x, s) vt s + at (2.5)

An intelligent multiple models based predictive control scheme with its application to industrial tubular heat

129

The system modeling results should uniformly be divided into small incremental elements, while the boundary conditions are given at x = kN; k = 0, 1, 2, . . . , n. Therefore, the system temperature could be represented as Ttk = Tt (kN, s), Tsk = Ts (kN, s) (2.6)

3 The proposed multiple models strategy The multiple models control strategy presented here is an approach for controlling the complex systems, where the system parameters may abruptly change in the span of a specied variation, at each instant of time. In fact, the control strategy operates in multiple operating environments, which may change from one to another. Here, the system behavior is either nonlinear or linear time variant, and a linear xed model may not really lead to the expected performance. In accordance with Fig. 2, a good approach for designing the linear controllers to deal with the complex systems is to use multiple models control strategy. To realize the strategy presented, some models which cover the different operating environments of the system must rst be identied and an appropriate controller must also be designed for each one of them. It should be noted that the multiple models strategy presented is described to dene some models corresponding to different operating environments of the system, to design the local controllers corresponding to each one of the predened models, to identify the best model of the system, to select the appropriate control action and nally to generate the nalized control action for the system, at each interval of time. To introduce the achieved specication of the proposed multiple model control strategy, we can say that the wide range of system coefcients variation could be covered. In this way, the weight generation mechanism is realized based on the novel fuzzy-based approach to generate accurate weights

where Ttk and Tsk are given as the temperatures at the kth point of the inner tube and the shell tube, respectively. Hence, using (2.5) and (2.6), the outlet temperature at kth point of the inner tube could now be given as Ttk = exp kN at (s + at ) + Tsk vt s + at (2.7)

Now, the system transfer function could also be written as kN Ttk at 1 exp = (s + at ) Tsk s + at vt (2.8)

Moreover, the obtained results could be expressed in terms of the valve pressure, i.e., Ttk , by using Pv Kv Pv U d
0 kN

(Tsk Ttk )dx = Cs

Tsk t

(2.9)

Also by using (2.8) at x = kN , we could deduce the following Tsk Ttk = Tsk s at x + exp (s + at ) s + at s + at vt (2.10)

Here, Tsk is dened as constant temperature with respect to x, i.e., Ts0 = Ts1 = = Tsk = cte (2.11)

Hereinafter, by using (2.9), (2.10) and (2.11), we could have the following Tsk U d kNs at v t + s + at (s + at )2 kN (s + at ) vt + sCs = Kv Pv (2.12) As a consequence, the tubular heat exchanger modeling could be resulted using (2.8) and (2.12) as Ttk k1 (s) = 1 2 + (a + kN a 1 )s + Pv aat s t
vt s+at (s)

1 exp

(2.13)

where Kv , , Pv and Cs are the valve gain, the compressed steam temperature of the shell tube, the valve pressure and nally the shell tube capacitance, respectively. Also (s), k1 k and a are given as 1 exp( kN (s + at )), U v and UCs , vt d d respectively.

Fig. 2 The scheme of the proposed multiple models strategy

130

A.H. Mazinan, N. Sadati

with respect to the system coefcients variation. The weight generation mechanism is stabled with respect to noise and the strategy is applicable in systems with rapid and wide coefcients variation. In the same way, the number of possible models is minimum and nally the proposed strategy could be used to deal with an array of the industrial systems, provided that the system could precisely be models, at each one of the operating environments. In designing this multiple models control strategy, two major separated parts must usually be considered, i.e., the best model identication mechanism and the appropriate nalized control action realization [33, 34]. Identication part is composed of dening some linear models corresponding to the different environments of the system, where it could be realized by a decision system, namely intelligent decision maker scheme (IDMS) in Fig. 2, in order to decide which model is closest to the process. In agreement with this gure, F /A Model#1 , . . . , F /A Model#r are the r xed/adaptive linear models, which are used in parallel with the system through the IDSM. Also F /A Cont#1 , . . . , F /A Cont#r are used as the corresponding linear generalized predictive controllers (LGPCs), in the strategy presented. At each instant of time, a linear model is identied as the best model of the system by the IDMS. Hereinafter, the models and the corresponding controllers could be used on the xed or adaptive via a selector system that is used in the IDMS. As soon as F /A Model#p ; p = 1, 2, . . . , r is identied as the best chosen model, the corresponding output of the selector system, i.e., cmp ; p = 1, 2, . . . , r is used to change the status of the chosen model. After that, ccp ; p = 1, 2, . . . , r is quickly used to change the status of the corresponding controller. In line with this situation, the chosen model, i.e., F /A Model#p becomes an adaptive model and also the F /A Cont#p becomes an adaptive controller, at this point of time. Based on this matter, the adaptive algorithm presented in the following is implemented on the chosen model, i.e., A(q)y(k) = B(q)u(k) + C(q)n(k) e(k) = y(k) T (k 1) T (k
T (k T (k

predictive controller; LGPC, scheme. In line with this matter, as soon as the adaptive algorithm is implemented on the chosen model, the corresponding control gain matrix; K LGP C , must be varied based on the adapted model polynomial, at each instant of time. Now, by denoting Uc = [ uc (k), . . . , uc (k + Nu 1)]T F = [f (k + N1 ), . . . , f (k + N2 )]T R = [r(k + N1 ), . . . , r(k + N2 )]T (3.2)

the manipulated variable of the LGPC can be given as Uc = K LGP C (R F ) K LGP C = (GT G + I )1 GT where G polynomials matrix is described as gN1 gN1 +1 G= . . . g N2 gN1 1 g N1 . . . gN2 1 ... ... . . . ... 0 0 . . . gN2 Nu +1 (3.4) (3.3)

in which the gi s are the coefcients of the matrix, which correspond to the system step response values [3847]. Also N2 N1 + 1, Nu , R and F denote the prediction horizon, the control horizon, the desired set points and nally the free responses, respectively. Now, the rst row of K LGP C matrix is applied to the system, at each instant of time. In such a case, the LGPC control action could be obtained by Hf (q 1 ) = 1 uc (k) = uc (k) 1 q 1 (3.5)

1) (3.1)

1) = [y(k 1), . . . , y(k n), u(k 1), . . . , u(k n), e(k), . . . , e(k n)]

1) = [a1 , . . . , an , b1 , . . . , bn , c1 , . . . , cn ] 1 (k) = P 1 (k 1) + (k 1) T (k 1) P where A(q), B(q) and C(q) are given as the ARMAX model polynomial over q, y(k) is given as the output variable of the model, u(k) is given as the control variable, n(k) is given as the uncorrelated white noise with zero mean, e(k) is also given as the error modeling and nally (k) is given as the model parameters adaptation [3537]. As it can be seen from the proposed strategy, the F /A Cont#p ; p = 1, 2, . . . , r is realized by concept of the linear generalized

In the same way, the rest of the models are identied as the xed linear models and the K LGP C gain is initially obtained based on the corresponding xed model polynomial. Realization of the multiple models approach based on the IDMS that is realized in association with the best model identication mechanism and the nalized control action generation is now described in the proceeding sections, where the best model identication mechanism is organized based on the fuzzy-based adaptive Kalman lter and the fuzzy-based weight generator approaches. 3.1 Intelligent decision maker scheme The intelligent decision maker scheme (IDMS) is the so important subsystem of the control strategy proposed, due to the fact that it is used to realize several precise tasks, as long as the system coefcients and the desired set point are abruptly varied in the span of time. In reality, the major responsibility of the IDMS presented here could briey be introduced as follows

An intelligent multiple models based predictive control scheme with its application to industrial tubular heat

131

Realizing the best model identication mechanism, using the fuzzy-based adaptive Kalman lter and the fuzzybased weight generator approaches. Selecting the models and the corresponding controllers status in the xed or the adaptive situations. Generating the nalized control action, using the soft switching technique. stabling the system performance under both the system coefcients and the desired set point variations. As it can be seen from the proposed multiple models control strategy, the IDMS has the several inputs and outputs, where the desired set point, the nalized control action; u, cmp and nally ccp ; p = 1, 2, . . . , r are used as the output signals of the IDMS. Also ymp and ucp ; p = 1, 2, . . . , r are used as the input signals of the IDMS in association with the proposed control strategy. Based on these input-output signals, all the mentioned tasks must appropriately be implemented through the IDMS, where the details of them are now given in the best model identication mechanism and the nalized control action generation sections as follows. 3.2 The best model identication mechanism The best model identication mechanism is realized to identify the best model of the system, at each instant of time, when we are suddenly encountered with both the system coefcients and the desired set point variations. In fact, the main idea of the proposed mechanism is to identify both the best model of the system (BM) and the deviated models from the best model of the system (DFBM), as long as the wide range of variation in the system coefcients and in the desired set point could be taken place. The mechanism proposed here, as a subsystem of the intelligent decision maker scheme (IDMS), has the so important role in this control strategy. The closed loop stability of the strategy may unacceptably be changed with system coefcients variations, provided that the best model identication mechanism cannot work correctly with respect to time. Meanwhile, as long as these parameters are rapidly changed in a wide range of variation, the applicability of the best model identication mechanism may actually be deteriorated. It means that the closed loop stability cannot truly be accepted, in such a case. In the mechanism presented, the controller weight parameters; wp,k , p = 1, 2, . . . , r, k = 1, 2, . . . , should accurately be varied to appropriate value, i.e., weight parameters must be adapted to the ones, as soon as the corresponding model state estimation error; ep,k , is close to the acceptable minimum values. Hereinafter, as soon as each one of the models of the system could be identied as the best chosen model, the corresponding output of the selector system, i.e., cmp ; p = 1, 2, . . . , r can cause to change the chosen model from xed to adaptive parameters. In addition, ccp ;

p = 1, 2, . . . , r can cause to change the corresponding controller from xed to adaptive parameters. To demonstrate the proposed mechanism in details, the fuzzy-based adaptive Kalman lter (FAKF) and also fuzzybased weight generators (FWG) must rst be organized. In this way, FAKF#p ; p = 1, 2, . . . , r is used to obtain the model states estimation error; ep,k , p = 1, 2, . . . , r, where FWG is also used to generate the appropriate controller weight signals of the corresponding ep,k . In fact, to realize the proposed approach, the FAKF and the FWG must appropriately be realized, when the statistical behavior of the system inputs and outputs could exactly be known. Now, the best model identication mechanism can briey be described as follows ep,k Min ep,k Max wp,k 1; wp,k 0; i.e., i.e., (3.6) F /A Model#p BM (adaptive model). F /A Model#p DFBM (xed model). Realization of the best model identication mechanism based on the FAKF and the FWG is now described in the proceeding sections. 3.2.1 The fuzzy-based adaptive Kalman lter approach The fuzzy-based adaptive Kalman lter; FAKF, has been used to estimate the model states, as long as the linear model state spaces, i.e., Ap , Bp and Cp ; p = 1, 2, . . . , r could be given. In association with this matter, the linear single inputsingle output model of the system can be described as xk+1 = Ak+1,k xk + Bk uk + wk y k = Ck x k + v k (3.7)

where xk is given as the system state vector, yk is given as the scalar measurement, Ak+1,k is given as the system state matrix, Ck is given as the measurement vector, wk (o, Q) is given as the process Gaussian white noise, vk (o, R) is also given as the measurement Gaussian white noise and nally R and Q are given as the scalar measurement noise covariance and the system noise covariance matrix, respectively. Now, the adaptive Kalman lter approach could be described as xk = xk,k1 + Kk ek Pk,k1 = Ak,k1 Pk1 AT k,k1 + Qk Pk = (I Kk Ck )Pk,k1 To realize the approach presented, at the kth instant of time, the posterior covariance matrix; Pk,k1 , and the prior covariance matrix; Pk , must instantly be obtained. In fact, the
T T Kk = Pk,k1 Ck (Ck Pk,k1 Ck + Rk )1

(3.8)

132

A.H. Mazinan, N. Sadati

system states as the output of the approach should be estimated using the Kalman gain; Kk , and the following model states estimation error ek = yk yk (3.9)

In the same way, the error estimation covariance matrix; Pe , and the error estimation mean value; Me , could be dened as
T Pe = Ck Pk,k1 Ck + Rk k T i=kNe +1 ei ei

Me =

where Qk and Rk must be adapted using a fuzzy-based system. The parameters mentioned could accurately be varied to the appropriate values, so that the state estimation errors could be close to the desired values. To realize the FAKF, the fuzzy-based system presented determines the value of , where Qk and Rk must be followed as Rk = R 2(k+1) , Qk = Q 2(k+1) (3.10)

1 Ne

(3.11)

where Ne denotes the size of the mean value window. 3.2.2 The fuzzy-based weight generator approach The fuzzy-based weight generator (FWG) approach presented here is used to generate the appropriate weight signals; wp,k , p = 1, 2, . . . , r, as long as we are encountered with variation in the system coefcients and also in the desired set point, abruptly. Based on this approach, the FWG must be positioned in sequence with the FAKF. It means that the FWG is realized based on the their input data; ep,k , p = 1, 2, . . . , r, that are generated of FAKF. To prevent the random weight variation, the stability of the proposed FGW could relatively be guaranteed by using the following performance index
k 2 Jp,k = ep,k (t) + j =0 2 exp((k j ))ep,j ;

In this way, Q and R are given as the constant matrices and also must be chosen either equal or greater than one. Now, for having high accuracy in the fuzzy-based system, the fuzzy set parameters are initially obtained from the GA algorithm [4855]. The obtained fuzzy sets are also shown in Fig. 3, where Z, S, M and L denote the zero, the small, the medium and nally the large fuzzy sets, respectively. Also the rules of the fuzzy-based Kalman lter is tabulated in Table 1.

0; , > 0

(3.12)

Fig. 3 The fuzzy sets of FAKF scheme Table 1 The fuzzy rule based of Kalman lter Pe Z Me Z S L S Z L S Z L M L Z M Z

Here, ep,k denotes the pth model state estimation error, at the kth instant of time. The above performance index aims to use the model states estimation errors, using the past to the present time. In other words, while disturbance takes place abruptly in the system, the performance index could reject the noisy data from ep,k , where the past data are used in this approach. In such a case, the performance index parameters are given as; , and , where and are the weighting factors on the instantaneous measures and the long term accuracy, respectively. In addition, is a forgetting factor, which assures the boundedness of the criterion for the bounded ep,k . Also in order to prevent the rapid unwanted changes in the mechanism presented, it is better to apply the achieved weights in the minimum time delay to the control strategy. Also an effective way to increase the weight generation accuracy is to apply the weights average in periods of time to this control strategy. In such a case, the unwanted changes have no effect on the system performance. Now, the approach presented here is realized based on a novel fuzzy-based algorithm, given by Dening the some performance indexes, i.e., Jp,k ; p = 1, 2, . . . , r, k = 1, 2, . . . , . Determining the minimum value of performance indexes and also the maximum acceptable value of performance indexes, i.e., Jmin and Jmax , respectively.

An intelligent multiple models based predictive control scheme with its application to industrial tubular heat

133

Dening the acceptable, the conditionally acceptable and the unacceptable fuzzy sets, i.e., AFS, CAFS and UAFS, respectively, for each one of the performance indexes. If the performance indexes, i.e., Jp,k is obtained in the AFS, the corresponding model; F /A Model#p , p = 1, 2, . . . , r, should now be identied as the best chosen model of the system and the algorithm stopped, otherwise the rest of the algorithm must be followed. Dening the some decision maker parameters; (p) = Jp,k Jmin ; p = 1, 2, . . . , r. Dening the fuzzy sets corresponding to the acceptable decision and the unacceptable decision; ADFS, UDFS, respectively, for each one of the decision maker parameters; (p). Identifying the best predened model of the system; F /A Model#p ; p = 1, 2, . . . , r in the following fuzzy rule based system IF Jp,k is AFS THEN F /A Model#p IF Jp,k is CAFS AND (p) is ADFS THEN F /A Model#p BM BM

where the fuzzy sets of the performance indexes; Jp,k , and the decision maker parameters; (p), are given in Figs. 4 to 5, respectively. Also the fuzzy sets of the controller weight parameters; wp,k , are given in Fig. 6. Regarding the proposed fuzzy-based weight generator approach, the best model of the system could instantly be identied for the system, as soon as the system coefcients variation is abruptly implemented on the control strategy proposed. 3.3 The nalized control action generation The nalized control action generation is realized by the IDSM scheme, in this proposed control strategy, as mentioned before. In reality, this subsystem is used to generate an appropriate control action for the system. In this way, as soon as the system coefcients are abruptly varied, the control action must be adapted to the appropriate value. In

IF Jp,k is CAFS AND (p) is UDFS THEN F /A Model#p DBFM DFBM

IF Jp,k is UAFS THEN F /A Model#p

Dening the fuzzy sets corresponding to low value and high value; LVFS, HVFS, respectively, to generate the controller weights parameters. Calculating the controller weights parameters based on the predened models; F /A Model#p , in the following fuzzy rule based system IF F /A Model#1 AND, . . . , AND F /A Model#r DFBM BM AND F /A Model#2 DFBM

Fig. 4 The scheme of the fuzzy sets of the performance indexes

THEN w1,k is HVFS, w2,k is LVFS, . . . , wp,k is LVFS IF F /A Model#1 AND, . . . , AND F /A Model#r DFBM DFBM AND F /A Model#2 BM
Fig. 5 The scheme of the fuzzy sets of the decision maker parameters

THEN w1,k is LVFS, w2,k is HVFS, . . . , wp,k is LVFS . . . IF F /A Model#1 AND, . . . , AND F /A Model#r BM
Fig. 6 The scheme of the fuzzy sets of the controller weight parameters

DFBM AND F /A Model#2

DFBM

THEN w1,k is LVFS, w2,k is LVFS, . . . , wp,k is HVFS

134

A.H. Mazinan, N. Sadati

fact, the system coefcients variation must be compensated by an appropriate control action, at each instant of time. On the other hand, the success of the proposed strategy is quite seen in proposed appropriate control action generation. As introduced in the multiple models strategy before, the F /A Cont#p ; p = 1, 2, . . . , r as the local controllers are realized based on the concept of the LGPC scheme, in this paper. Here, the controller designing is realized based on the predened explicit linear model of the system. In this case, the nalized control action could be realized by a soft switching technique, i.e., the linear combination of the local controllers, given by
r r

uk =
p=1

wp,k ucp,k ,
p=1

wp,k = 1

(3.13)

where r is given as the number of appropriate local LGPC controllers, wp,k is given as the appropriate weight of the pth local LGPC controller, at the kth instant of time, that is given by the best model identication mechanism, ucp,k is also given as the pth local LGPC output and nally uk is given as the nalized control action. Based on this strategy, the control action is adapted to appropriate value and also is applied to the system, at the kth instant of time.

the Wiener model of the system is to remove the nonlinearity of the system, when the inverse of the nonlinear function of the Wiener model could be used in sequence with the system, as shown in this strategy. Based on this approach, the linear part of the Wiener model could be used as the model of the system. Moreover, for designing a control strategy, control engineers normally need to have an initial model of the system, which would give them a scope about the structure of the system under investigation. This initial model allows them to use a simulation platform, where control strategy could be tested before being transferred into the real time environment. Hence in most cases, the structure of the model including linearity and nonlinearity could be known. It means the various structures lead to the same linearity and nonlinearity effect. In this case, for realizing the Wiener model of the system, the N nonlinear model output; ym (k), could be represented, when L the linear model output; ym (k), is obtained. Here, by using the recursive least square (RLS) identication algorithm, the linear part of the Wiener model could be identied. In addition, the nonlinear part of the Wiener model could also be expressed as
N y (k) = f (yL (k)) m m N = y (0) + 0 tanh(0 (yL (k) yL (0))) m m m

(4.1)

4 Nonlinear GPC approach In this section, a nonlinear GPC (NLGPC) approach for controlling an industrial tubular heat exchanger system is proposed [56, 57]. As we know, the LGPC approach is a wellknown control strategy used both in industrial and academic environments, for deriving the linear systems. So nonlinear systems cannot appropriately be controlled by this approach. Here, we need to modify the traditional LGPC approach in its present form, where it could be used for controlling the nonlinear systems as well. The strategy is shown in Fig. 7, L N where u(k), ym (k) and ym (k) denote the control action, the linear model output and the nonlinear model output, respectively. For realizing the NLGPC approach, we rst need to obtain linear and nonlinear parts of the Wiener model of the system that is shown in Fig. 8. The main purpose of realizing

where 0 and 0 denote the nonlinear model coefcients. We know that the LGPC approach must be used with linear system and the NLGPC approach must also be used with nonlinear system, so we have to nd the Wiener model of the system that is shown, in this approach. Now, by having the obtained results, the nonlinearity of the system could be removed using the inverse of the nonlinear function; NLM1 , as shown in the proposed strategy. As we know, a sequence of the future model outputs using the system modeling could be obtained by the following j -step ahead predictor of the LGPC algorithm
L ym (k + j ) = Hj (q 1 ) u(k 1) + Gj (q 1 ) u(k + j 1) L + Fj (q 1 )ym (k); j = N1 , . . . , N2

(4.2)

where Fj (q 1 ), Hj (q 1 ) and Gj (q 1 ) are all given as Fj (q 1 ) = [FN1 (q 1 ), . . . , FN2 (q 1 )]T H (q 1 ) = [HN1 (q 1 ), . . . , HN2 (q 1 )]T j Gj (q 1 ) = [GN1 (q 1 ), . . . , GN2 (q 1 )]T

(4.3)

Fig. 7 The nonlinear GPC approach in controlling the heat exchanger

Fig. 8 The Wiener model scheme of the heat exchanger

An intelligent multiple models based predictive control scheme with its application to industrial tubular heat

135

Here, the gi (q 1 )s are denoted as the coefcients of Gj (q 1 ) matrix polynomials, which correspond to the system step response values, given by
j

Gj (q 1 ) j gN (q 1 ) j 1 1 g N1 +1 (q ) = . . . gN2 (q 1 )
j

gN1 1 (q 1 )
j

... ... . . . ...


j

0 0 . . . gN2 Nu +1 (q 1 )

gN1 (q 1 ) . . .
j

gN2 1 (q 1 )
j

Calculating a sequence of future nonlinear model outputs based on the proposed nonlinear function, using (4.1). Identifying the linear part of the Wiener model of the system based on the RLS identication algorithm and also calculating a sequence of future linear model by the GPC algorithm, using (4.2). Obtaining the NLGPC manipulated variables by optimizing the proposed cost function, using (4.8). Obtaining the NLGPC control action based on the discrete lter, using (4.9).

(4.4) Meanwhile, N2 N1 + 1 and Nu are given as the prediction horizon and the control horizon, respectively. Afterward, Fj (q 1 ), Hj (q 1 ) and Gj (q 1 ) could be obtained using the following Diophantine equation, i.e., 1 = Ej (q 1 )A(q 1 ) (q 1 ) + q j Fj (q 1 ) where we could have Ej (q 1 )B(q 1 ) = Gj (q 1 ) + q j Hj (q 1 ) (4.6) (4.5)

5 Simulation results To consider the applicability of the proposed approach, a tubular heat exchanger system, which has the many industrial environments such as food processing, automotive, aerospace, metallurgy, pulp and paper, fertilizers, chemicalspetrochemicals and cement is considered for simulation. As it can be seen, the uid of the inner tube at x = Ln = 2.5 m, t = 0.15 m2 and vt = 0.1 m/s has a wide variation in the span of time, as long as the water, steam, engine oil, mineral oil, palm oil, white oil, vegetable oil, dry air, milk, liquid metal, petroleum jelly, petroleum resin and other related liquids could be used as the uid of the inner tube. Moreover, the steam is used as the uid of the shell tube, in this simulation. In such a case, the inner tube uid is used as an outlet of the system and the shell tube uid is also used as an inlet of the system. Here, the inner tube temperature should be adjusted by commanded valve pressure; Pv , on the shell tube. Here, the tracking performance of the proposed scheme, called by the authors as an intelligent multiple models based adaptive predictive control scheme (IMMBAPC), using both the desired set points between 0C and 44C and the following parameters variation, is considered [58]. t (k) = t + t (k), Ut (k) = Ut + Ut (k) t t (k) t , Ut Ut (k) Ut Cpt Cpt (k) Cpt ,
Cpt = 0.14 kJ/kg K, Cpt (k) = Cpt + Cpt (k),

Here, A(q 1 ) and B(q 1 ) could be obtained using the RLS identication algorithm, as the CARIMA model of the system, and are adapted, at each instant of time. Hereinafter, by using the obtained results, we could calculate a sequence, i.e., j = N1 , . . . , N2 , of future nonlinear part of the Wiener model output as
N y (k + j ) = f (yL (k + j )) m m N = y (0) + 0 tanh(0 (yL (k + j ) yL (0))) m m m

(4.7) Afterward, the manipulated variable; u(k), could be obtained by optimizing the following cost function
N2

JN LGP C =
j =N1

N (y (k + j ) r(k + j ))2 m Nu

(5.1)

j =1

u2 (k + j 1)

(4.8)

t = 711 kg/m3 , Ut = 7.90 W/m2 K

where r(k) and denote the desired set point and control weight coefcients, respectively. In this strategy, the NLGPC control action; u(k), is nally obtained by using the following discrete lter Hf (q 1 ) = 1 u(k) = u(k) 1 q 1 (4.9)

where q 1 denotes the delay term. As a consequence, according to Fig. 7, realization of the proposed NLGPC scheme could be summarized as

To overcome the system coefcients variation, we need to dene several system operating environments and to identify the corresponding models. For the number of models in the multiple models control strategy presented, it is better to dene the least operating environments, while the system performance is not ignored. Here, due to the results obtained from experiments, the optimal number of models was obtained to be three (p = 3) for these simulations. By using this result, the system operating environments, i.e.,

136

A.H. Mazinan, N. Sadati

EV#p ; p = 1, 2, 3, will cover the whole of the system coefcients variation, given below: t (k) = t = 0 C (k) = Cpt = 0 pt Ut (k) = Ut = 0 (k) = t +t = 208 t 2 Cpt +Cpt = 2.02 Cpt (k) = 2 Ut +Ut Ut (k) = = 2000 2 t (k) = t = 416 (k) = Cpt = 4.04 Cpt Ut (k) = Ut = 4000

Now, to validate the models, the ith model error; ei (k), with respect to the system output; y(k), are expressed as ei (k) = y(k) y i (k); i = 1, 2, 3 (5.6)

EV#1 , i.e., M#1 :

(5.2)

EV#2 , i.e., M#2 :

(5.3)

EV#3 , i.e., M#3 :

(5.4)

where t , Cpt and Ut are given in 0C and t , Cpt and Ut are also given in 100C, respectively. Now, using the RLS identication method, the following CARIMA models of the system corresponding to different system operating environments; EVs, could be obtained, where the results of the identication process are now tabulated by Table 2. Ai (q 1 )y i (k) = B i (q 1 )u(k 1) +
i Ai (q 1 ) = 1 + a1 q 1 i + + ap q p ; e(k) ; (q 1 )

The results, as tabulated by Table 3 can verify the validity of the chosen models. Also the LGPC prediction horizon and the control horizon are given as; N2 N1 + 1 = 3 and Nu = 3, respectively. These control parameters are obtained based on the system performance with respect to the recursive computational operation of the predictive controller. In this way, the system performance could be improved, provided that these parameters are appropriately chosen. Here, Fig. 9 shows the tracking performance of the proposed IMMBAPC, while Fig. 10 represents the performance of the nonlinear GPC (NLGPC). These results are obtained, when we are suddenly encountered with both the system operating environment and the desired set point variations, at several points of time. In these simulations, the system coefcients are abruptly varied at several points of time, i.e., at 9, 23, 38, 45, 60, 67, 76, 82, 90, 112, 126, 134 s and nally at 186 s, respecTable 3 The models validation Model error M#1 3.82E03 5.14E02 M#2 7.41E03 9.89E02 M#3 5.12E03 6.46E02

i = 1, 2, 3 (5.5)

e2 (V ) |e| (V )

p=4

i i i B i (q 1 ) = b0 + b1 q 1 + + bp q m ; q = 4

Here, y i (k), u(k) and e(k) denote the ith model output variable, the control action variable and nally the random sequence number, respectively. Also (q 1 ) is taken as 1 q 1 .

Table 2 The coefcients of the CARIMA models k 1 1 1 1 2 2 2 2 3 3 3 3 j 1 2 3 4 1 2 3 4 1 2 3 4


k aj k bj

0.9933 0.4343 0.0069 0.4219 0.9947 0.4327 0.0083 0.4204 0.9960 0.4313 0.0097 0.4189

0.2506E03 0.3519E03 0.5283E03 0.1830E03 0.2469E03 0.3426E03 0.5208E03 0.1738E03 0.2434E03 0.3336E03 0.5135E03 0.1637E03

Fig. 9 The scheme of IMMBAPC tracking performance

Fig. 10 The scheme of NLGPC tracking performance

An intelligent multiple models based predictive control scheme with its application to industrial tubular heat

137

tively, while the desired set point is varied at 0 and 112 s, respectively. In accordance with Fig. 11, F /A Model#1 , F /A Model#2 and F /A Model#3 are identied as the best chosen model several times, by the intelligent decision maker scheme (IDSM) presented in this control strategy. Here, the models behavior have the important roll in the performance of the IDSM. Hereinafter, F /A Cont#p ; p = 1, 2, 3 are used as the dominant adaptive predictive controller, at the corresponding time, i.e., when the F /A Model#p ; p = 1, 2, 3 are identied as the best chosen model of the system. As it can be seen from Fig. 11, the F /A Model#p ; p = 1, 2, 3 are identied as the best chosen model of the system, as long as theses models could be relatively close to system behavior that is abruptly inuenced by variation in the system coefcients and also in the desired set point. In line with these results, F /A Model#1 is identied as the best chosen model of the system from 1 to 2 s, from 24 to 27 s, from 46 to 49 s, from 68 to 72 s, from 82 to 86 s, from 113 to 116 s, from 118 to 126 s, from 130 to 134 s, from 140 to 118 s and nally from 192 to 200 s, respectively. In this case, F /A Model#2 is also identied as the best chosen model of the system from 6 to 7 s, from 10 to 11 s, at 40 s, from 62 to

63 s, at 77 s, at 92 s and nally at 127 s, respectively. In addition, F /A Model#3 is identied as the best chosen model of the system from 3 to 5 s, at 9, at 39, at 61, at 76 and nally at 90 s, respectively. In the IMMBAPC approach presented, the local control actions are shown in Fig. 12, where these signals are represented after multiplication of corresponding weights. Furthermore, the nalized control action is shown in Fig. 13, where this signal is obtained by proposed IDSM, in this control strategy. It should be noted that the proposed IDMS enable to generate the accurate weights and the accurate control action

Fig. 12 The scheme of IMMBAPC control action signals

Fig. 11 The scheme of IMMBAPC weight signals

Fig. 13 The scheme of IMMBAPC nalized control action signal

138

A.H. Mazinan, N. Sadati

signals, as long as the system coefcients are suddenly varied, at each instant of time. Consequently, the simulation results are compared with those obtained using the nonlinear LGPC (NLGPC) scheme, where for realizing this scheme the linear and the nonlinear parts of the Wiener model of the system are organized as n L L y (k) = y L (0) + ap (k)ym (k p) m m p=1
n

References
1. Mazinan AH, Sadati N (2009) Fuzzy predictive control based multiple models strategy for a tubular heat exchanger. Appl Intell. doi:10.1007/s10489-009-0163-1 2. Mazinan AH, Sadati N (2008) Fuzzy multiple models predictive control of tubular heat exchanger. In: Proc of IEEE world congress on computational intelligence, pp 18451852 3. Mazinan AH, Sadati N (2008) Multiple modeling and fuzzy predictive control of a tubular heat exchanger system. Trans Syst Control 3:249258 4. Mazinan AH, Sadati N (2008) Fuzzy multiple modeling and fuzzy predictive control of a tubular heat exchanger system. In: International conference on application of electrical engineering, pp 77 81 5. Mazinan AH, Sadati N (2008) Fuzzy multiple modeling and fuzzy predictive control of a tubular heat exchanger system. In: International conference on robotics, control and manufacturing technology, pp 9397 6. Hong X, Harris CJ (2002) A mixture of experts network structure construction algorithm for modelling and control. Appl Intell 16:5969 7. Flores A, Saez D, Araya J, Berenguel M, Cipriano A (2005) Fuzzy predictive control of a solar power plant. IEEE Trans Fuzzy Syst 1:5868 8. Yager RR (1992) A general approach to rule aggregation in fuzzy logic control. Appl Intell 2:333351 9. Sousa JMDC, Kaymak U (2001) Model prediction control using fuzzy decision functions. IEEE Trans Syst Man Cybern, Part B, Cybern 1:5465 10. Rashidi F, Mazinan AH (2004) Modeling and control of three phase boost rectiers via wavelet based neural network. Trans Syst 3:494497 11. Xia L, DeAbreu-Garcia JA, Hartley TT (1991) Modeling and simulation of a heat exchanger. In: Proc. of the IEEE international conference on system engineering, pp 453456 12. Ho TB, Nguyen TD, Shimodaira H, Kimura M (2003) A knowledge discovery system with support for model selection and visualization. Appl Intell 19:125141 13. Thiaw L, Malti R, Madani K (2003) A multiple models approach for nonlinear systems identication: Comparison between ANN based and conventional implementation. In: Proceeding Book of International Conference on Neural Networks and Articial Intelligence (ICNNAI 2003), Minsk, Byelorussia, pp 210214. ISBN 985-444-571-2 14. Madani K, Chebira A, Rybnik M (2003) Data driven multiple neural network models generator based on a tree-like scheduler. In: Mira J, Alvarez JR (eds) Computational methods in neural modeling. Lecture notes in computer science, vol 2686. Springer, Berlin, pp 382389. ISBN3-540-40210-1 15. Chebira A, Madani K, Rybnik M (2003) Non linear process identication using a neural network based multiple models generator. In: Mira J, Alvarez JR (eds) Articial neural nets problem solving methods. Lecture notes in computer science, vol 2687. Springer, Berlin, pp 647654. ISBN 3-540-40211-X 16. Thiaw L, Rybnik M, Malti R, Chebira A, Madani K (2004) A comparative study between a multi-models based approach and an articial neural network based technique for nonlinear systems identication. Comput Int Sci J 3(1):6674. ISSN 1727-6209 17. Bouyoucef E, Chebira A, Rybnik M, Madani K (2005) Multiple neural network model generator with complexity estimation and self-organization abilities. Int Sci J Comput 4(3):2029. ISSN 1727-6209 18. Madani K, Thiaw L (2005) Multi-model based identication: application to nonlinear dynamic behavior prediction. In: Saeed, K., Mosdorf, R., Pejas, J., Hilmola, O.-P., Sosnowski, Z., (ed) Image

bp (k)u(k p) + p=1 N y (k) = y N (0) + tanh( (y L (k) y L (0)))


m m 0 0 m m

(5.7)

L N Here, ym (0), ym (0), 0 , 0 and n are given as 0.7, 0.5, 1.6, 0.5 and 4.0, respectively. In addition, both ap (k) and bp (k) must be obtained using the RLS algorithm, at each instant of time. By using both the IMMBAPC and the NLGPC schemes in several simulations, with the same conditions, the performance improvement of the IMMBAPC scheme is easily observed. In these cases, the NLGPC scheme does not perform well, when changes in the system coefcients and in the desired set point are suddenly taken place. In fact, it is shown that the IMMBAPC approach could track appropriately the desired step points in the control strategy presented.

6 Conclusion A novel multiple models strategy using the well-known linear generalized predictive control (LGPC) scheme is proposed to control an industrial tubular heat exchanger system. In the approach presented here, the best model identication mechanism and also the nalized control action generation are realized by an intelligent decision maker scheme (IDSM). In line with this strategy, the best model identication mechanism is organized in agreement with the fuzzybased adaptive Kalman lter and the fuzzy-based weight generator approaches. The applicability of the strategy presented is summarized in controlling the system with rapid and wide range of variation in the coefcients and also in the desired set point. Here, the control strategy is implemented on the system and the results are compared with those obtained using a nonlinear GPC (NLGPC) scheme realized based on the Wiener model of the system. The achieved results can verify the validity of the proposed control strategy. As it can be seen from these simulation results, the multiple models control strategy presented outperforms the NLGPC scheme in an satisfactory manner.

An intelligent multiple models based predictive control scheme with its application to industrial tubular heat analysis, computer graphics, security systems and articial intelligence applications, pp 365375. ISBN 83-87256-86-2 Thiaw L, Madani K (2006) Self-organizing multi-model based identication: Application to nonlinear dynamic systems behavior prediction. Image Process Commun J 10(2):6374. ISSN 1425-140X Madani K, Thiaw L (2007) Self-organizing multi-modeling: A different way to design intelligent predictors. Neuro Comput 70(16 18):28362852. ISSN 0925-2312 Murray-Smith R, Johansen TA (1997) Multiple model approaches to modeling and control. Taylor & Francis, London. ISBN 0-74840595-X Guerci J, Feria E (1991) Multi-model predictive transform estimation. In: Proc of aerospace and electronics conference, pp 119 125 Ning L, Shao-Yuan L, Yu-Geng X (2004) Multi-model predictive control based on the Takagi-Sugeno fuzzy models: a case study. In: Proc of IEEE conference on information science, pp 247263. Wang N (2002) A fuzzy PID controller for multi-model plant. In: Proc of IEEE conference on machine learning and cybernetics, pp 14011406 Qi-Gang G, Dong-Feng W, Pu H, Bi-Hua L (2003) Multi-model GPC for steam temperature system of circulating uidized bed boiler. In: Proc of IEEE international conference on machine learning and cybernetics, vol 2, pp 906911 Sadati N, Bagherpour M, Ghadami R (2005) Adaptive multimodel CMAC-based supervisory control for uncertain MIMO systems. In: Proc of the 17th IEEE international conference on tools with articial intelligence, Hong Kong, China, Nov 2005, pp 457 461 Bakhshandeh R (1994) Multiple inputs-multiple outputs adaptive predictive control of a tubular heat exchanger system. MSc Thesis, Electrical Engineering Department, Sharif University of Technology [in Persian] Skrjanc I, Matko D (2000) Predictive functional control based on fuzzy model for heat-exchanger pilot plant. IEEE Trans Fuzzy Syst 8:705711 Matko D, Kavsek-Biasizzo K, Skrjanc I, Music G (2000) Generalized predictive control of a thermal plant using fuzzy model. In: Proc of the American control conference, vol 3, pp 20532057 Abe N, Seki K, Kanoh H (1996) Two degree of freedom internal model control for single tubular heat exchanger system. In: Proc of the IEEE international symposium on industrial electronics, vol 1, pp 260265 Fazlur Rahman MHR, Devanathan R (1994) Feedback linearisation of a heat exchanger. In: Proc of the 33rd IEEE international conference on decision and control, vol 3, pp 29362937 Fazlur Rahman MHR, Devanathan R (1994) Modeling and dynamic feedback linearization of a heat exchanger. In: Proc of the third IEEE international conference on control applications, vol 3, pp 18011806 Sadati N, Talasaz A (2004) Robust fuzzy multimodel control using variable structure system. In: Proc of IEEE conference on cybernetics and intelligent systems, vol 1, pp 497502 Sadati N, Ghadami R, Bagherpour M (2005) Adaptive neural network multiple models sliding mode control of robotic manipulators using soft switching. In: Proc of the 17th IEEE international conference on tools with articial intelligence, pp 431438 Chang BR, Tsai H (2007) Composite of adaptive support vector regression and nonlinear conditional heteroscedasticity tuned by quantum minimization for forecasts. Appl Intell 27:277289 Liang K, Yao X, Newton CS (2001) Adapting self-adaptive parameters in evolutionary algorithms. Appl Intell 15:171180 Neri F, Toivanen J, Makinen RAE (2007) An adaptive evolutionary algorithm with intelligent mutation local searchers for designing multidrug therapies for HIV. Appl Intell 27:219235

139

19.

20.

21.

22.

23.

24.

25.

26.

27.

28.

29.

30.

31.

32.

33.

34.

35.

36. 37.

38. Saez D, Cipriano A (1997) Design of fuzzy model based predictive controller and its application to an inverted pendulum. In: Proc of the sixth IEEE international conference on fuzzy systems, vol 2, pp 915919 39. Hadjili ML, Wertz V, Scorletti G (1998) Fuzzy model-based predictive control. In: Proc of IEEE decision and control, vol 3, pp 29272929 40. Huang S, Tan KK, Lee TH (2002) Applied predictive control. Springer, London 41. Clarke DW (1988) Application of generalized predictive control to industrial processes. IEEE Control Syst Mag 4955 42. Sousa JM (2000) Optimization issues in predictive control with fuzzy objective functions. Int J Intell Syst 15:879899 43. Zamarreno JM, Vega P (1999) Neural predictive control application to a highly non-linear system. Eng Appl Artif Intell 12:149 158 44. Gadkar KG, Doyle FJ III, Crowley TJ, Varner JD (2003) Cybernetic model predictive control of a continuous bioreactor with cell recycle. Biotechnol Prog 19:14871497 45. Saha P, Krishnan SH, Rao VSR, Patwardhan SC (2004) Modeling and predictive control of MIMO nonlinear systems using WienerLaguerre models. Chem Eng Commun 8:10831120 46. Franco E, Sacone S, Parisini T (2004) Practically stable nonlinear receding-horizon control of multi-model systems. In: Proc of IEEE Conference on Decision and Control, 3, pp 32413246 47. Ding Z, Leung H, Chan K (2000) Model-set adaptation using a fuzzy Kalman lter. In: Proc of the third international IEEE conference on information fusion, vol 1, pp 29 48. Shiu SCK, Li Y, Zhang F (2004) A fuzzy integral based query dispatching model in collaborative case-based reasoning. Appl Intell 21:301310 49. Zhang Y, Chi Z, Liu X, Wang X (2007) A novel fuzzy compensation multi-class support vector machine. Appl Intell 27:2128 50. Chen S, Chen S (2005) A prioritized information fusion method for handling fuzzy decision-making problems. Appl Intell 22:219 232 51. Li JH, Lim MH, Cao Q (2005) A qos-tunable scheme for ATM cell scheduling using evolutionary fuzzy system. Appl Intell 23:207 218 52. Sun S, Zhuge F, Rosenberg J, Steiner RM, Rubin GD, Napel S (2007) Learning-enhanced simulated annealing: method, evaluation, and application to lung nodule registration. Appl Intell 28:8399 53. Lee KK, Yoon WC, Baek DH (2006) A classication method using a hybrid genetic algorithm combined with an adaptive procedure for the pool of ellipsoids. Appl Intell 25:293304 54. Karr CL, Wilson E (2003) A self-tuning evolutionary algorithm applied to an inverse partial differential equation. Appl Intell 19:147155 55. Lee Z (2008) A robust learning algorithm based on support vector regression and robust fuzzy cerebellar model articulation controller. Appl Intell 29:4755 56. Are MM, Montazeri A, Poshtan J, Jahed-Motlagh MR (2006) Nonlinear model predictive control of chemical processes with a Wiener identication approach. In: Proc of IEEE conference on industrial technology, pp 17351740 57. Rueda A, Cristea S, Prada CD, Keyser RD (2005) Non-linear predictive control for a distillation column. In: Proc of 44th IEEE conference on decision and control, pp 51565161 58. Cengel YA, Turner RH (2004) Fundamentals of thermal uid sciences, 2nd edn. McGraw-Hill, New York

140 A.H. Mazinan was born on May 4, 1969, in Tehran, Iran. He received the B.Sc. degree in Electronic Engineering from the Islamic Azad University (IAU), Karaj Branch, Iran, in 1992, the M.Sc. degree in Control Engineering from the IAU, South Tehran Branch, Iran, in 1995 and nally the Ph.D. degree in Control Engineering from the IAU, Science and Research Branch, Iran, in 2009, respectively. He is now with Electrical Engineering Department of the IAU, South Tehran Branch as a fac-

A.H. Mazinan, N. Sadati ulty member, since 1996. His current research activities include predictive control, estimation theory, fuzzy logic, neural network, genetic algorithm and their applications in multiple modeling and in hybrid control systems. N. Sadati http://ee.sharif.edu/~sadati/.

S-ar putea să vă placă și