Journal of Process Control 17 (2007) 83–92

www.elsevier.com/locate/jprocont

Generalized predictive control using recurrent fuzzy neural

networks for industrial processes
a,b a,*
Chi-Huang Lu , Ching-Chih Tsai
a
Department of Electrical Engineering, National Chung-Hsing University, 250 Kuo-Kuang Road, Taichung 40227, Taiwan
b
Department of Electrical Engineering, Hsiuping Institute of Technology, 11 Gungye Road, Dali City, Taichung 412, Taiwan

Received 22 February 2006; received in revised form 1 August 2006; accepted 2 August 2006

Abstract

This paper presents a design methodology for predictive control of industrial processes via recurrent fuzzy neural networks (RFNNs).
A discrete-time mathematical model using RFNN is constructed and a learning algorithm adopting a recursive least squares (RLS)
approach is employed to identify the unknown parameters in the model. A generalized predictive control (GPC) law with integral action
is derived based on the minimization of a modified predictive performance criterion. The stability and steady-state performance of the
resulting control system are studied as well. Two examples including the control of a nonlinear process and the control of a physical
variable-frequency oil-cooling machine are used to demonstrate the effectiveness of the proposed method. Both results from numerical
simulations and experiments show that the proposed method is capable of controlling industrial processes with satisfactory performance
under setpoint and load changes.
 2006 Elsevier Ltd. All rights reserved.

Keywords: Generalized predictive control; Process control; Recurrent fuzzy neural network; Variable-frequency oil-cooling machine

1. Introduction ear systems that operate over a wide range of operating

conditions. Second, significant progress has been made in
In process control, many physical processes confront the development of model-based controller design for non-
with nonlinear dynamics, in which the relationships linear systems. Those techniques employ nonlinear models
between the controlled and manipulated variables depend directly without the need for local linearization about an
heavily on the operating conditions. For instance, the operating point. Finally, the developments of powerful
industrial processes including high purity distillation computers have made on-line implementation of these non-
columns, highly exothermic chemical reactions, pH neu- linear model-based controllers possible [1].
tralizations, batch systems cooling processes, and biochem- The generalized predictive control is one of the most
ical processes may exhibit such behaviors. In the past powerful and useful model-based control methods for a
decade, the control of nonlinear system has received con- wide class of nonlinear dynamic systems [2–6]. The basic
siderable attention from both academia and industry. principle of the GPC is to generate a sequence of control
The interest in the design and analysis of nonlinear control signals at each sample interval that optimizes the control
system is due to several factors. Two of these factors are effort, in order to follow exactly the reference trajectory.
especially worthy of mentioning. First and foremost, linear Many researchers have presented controllers involving
controllers usually perform poorly when applied to nonlin- the use of nonlinear GPC. Nicolao et al. [7] presented a sta-
bilizing predictive control with nonlinear ARX models for
*
Corresponding author. Tel.: +886 4 22859351; fax: +886 4 22856232.
nonlinear discrete-time systems. A class of nonlinear PID
E-mail addresses: [email protected] (C.-H. Lu), cctsai@dragon. controllers has been derived for nonlinear systems using
nchu.edu.tw (C.-C. Tsai). the nonlinear GPC approach in [8]. Kanev and Verhaegen

0959-1524/$ - see front matter  2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2006.08.003
84 C.-H. Lu, C.-C. Tsai / Journal of Process Control 17 (2007) 83–92

[9] outlined the algorithm for controller reconfiguration of to the solution of the conventional GPC which demands
nonlinear systems, based on a combination of a multiple for large matrix inversion and numerous matrix multiplica-
model estimator and a generalized predictive controller. tions in solving the optimization problems. The second is
Cao [10] proposed a formulation of nonlinear model pre- to establish the convergence of the RFNN model via
dictive control using automatic differentiation. Lyapunov stability theory, and to study the stability of
Neural networks (NNs) have been shown to possess the closed-loop control system. The third is to illustrate
good function approximation capabilities, and have been the uses of this proposed method for control by computer
applied successfully by many researchers [11–15] in model- simulations on a nonlinear discrete-time process and by
ing some poorly understood systems or processes. The experiments on a physical variable-frequency oil-cooling
results in [11–15] demonstrated the feasibility and efficacy machine.
of neural predictive control techniques for identification The remainder of the paper is organized as follows. Sec-
and control of nonlinear dynamic systems. On the other tion 2 presents the RFNN model for a class of nonlinear
hand, fuzzy models of the Takagi–Sugeno (T–S) type have discrete-time processes. In Section 3, the proposed predic-
been proven suitable for the use of nonlinear GPC, because tive control law with integral action is derived, and the
these models are able to accurately approximate complex real-time predictive control algorithm based on the RFNN
nonlinear systems by using data along with a prior knowl- model is presented. Section 4 details the capabilities of the
edge of processes [16]. The studies in [16–21] reported many proposed algorithm for controlling a nonlinear process sys-
successful applications of GPC using fuzzy models. tem utilizing computer simulations. Experimental results
The concept of incorporating fuzzy logics into NNs has for controlling an oil-cooling process to meet the desired
become increasingly important in recent years. In contrast performance specifications are also presented. Section 5
to the pure NNs or fuzzy systems, the fuzzy neural network concludes this paper.
(FNN) combines the capability of fuzzy reasoning in han-
dling uncertain information and the capability of artificial
learning in modeling the processes. In the recurrent fuzzy
2. Nonlinear system modeling using RFNN models
neural networks (RFNNs) [22–25], recurrent neurons in
the form of feedback connections as internal memories
This section aims at developing an RFNN model for a
are used. The RFNN have been shown more suitable for
class of nonlinear discrete-time processes, where the
describing dynamic systems than the FNN, because it can
RFNN structure is shown in Fig. 1. Layer 1 accepts the
deal with time-varying input or output through its own ^
input variable x i ðkÞ; the nodes therein represent input lin-
natural temporal operation. Moreover, such networks
guistic variables. The nodes in layer 2 are respectively
can be functionally interpreted using fuzzy inference mech-
labeled with F ij . In this layer, each node performs a mem-
anism, and they have been shown applicable to many engi- ^
bership function and acts as a unit of memory. F ij is a lin-
neering fields, such as image processing, control, signal
guistic label associated this node, and a Gaussian function ^
processing, robotics, speech recognition and etc. There
is adopted here as the membership function. Notation # ij
are also many existing predictive control strategies based
denotes the feedback gain in the jth term of the ith input
on FNN or RFNN models. For examples, Zhang and ^
linguistic variable x i ðkÞ to the node of layer 2. What is
Kovacevic [26] presented neurofuzzy model-based predic-
worth mentioning is that the input of layer 2 contains the
tive control for the weld fusion zone geometry, Li et al.
memory terms which store the past information of the out-
[27] designed the generalized predictive control using fuzzy
puts from layer 2. Apparently, this is the main difference
neural networks in TS model, and Zhang and Morris [28]
between the FNN and the RFNN. Every node in layer 3
presented a type of nonlinear model-based long-range pre-
is labeled with P, whose output is the product of all the
dictive controller based on the RFNN model. The method
products of the incoming signals. Every node in layer 4 is
presented in this paper is an extension of the predictive
labeled with N; these nodes calculate the ratio of the firing
controllers proposed in [26–28], where the problems of
strength of every rule to the sum of those of all rules. The
the control performance and the computational complexity
outputs from layer 4 are called normalized firing ^ strengths.
associated with the FNN-based or RFNN-based predictive
Every node j in layer 5 is with a node function h j , and every
controllers have not been addressed yet. ^ ^
output from layer 5 has a mathematic expression 1 j h j ,
There are three objectives of this paper. The first is to ^
where 1 j is the normalized firing strength from layer 4.
propose a controller for a class of nonlinear discrete-time
The single node in layer 6 is labeled with R, which com-
processes using GPC with RFNN model, and derive the ^
putes the output y ðkÞ as the summation of all incoming sig-
less computational controller’s expression based on the
nals from layer 5. For a nonlinear dynamic process, the
minimization of a modified predictive performance crite-
following nonlinear autoregressive moving average model
rion. The developed control law with integral action is par-
with an exogenous input (NARMAX model) is considered:
ticularly useful in eliminating steady-state tracking errors
caused from modeling errors or constant external distur-
yðkÞ ¼ f ðyðk  1Þ; yðk  2Þ; . . . ; yðk  ny Þ;
bances. The proposed predictive controller takes a less
computational requirement to overcome a major obstacle uðk  dÞ; uðk  d  1Þ; . . . ; uðk  nu ÞÞ þ nðkÞ; ð1Þ
 C.-H. Lu, C.-C. Tsai / Journal of Process Control 17 (2007) 83–92 85

z −1

(
ϑ11
F11

(
Σ F11

F1L
(

x1 (k)

(
Π ς 1 θ1

(
N

(
y (k)
Σ
(

x n(k)

(
Π ς θL

(
Fn1 N L

Layer 1 FnL Layer 3 Layer 4 Layer 5 Layer 6

Layer 2
Fig. 1. A configuration of the proposed RFNN architecture.

where uðÞ : Z þ ! R and yðÞ : Z þ ! R are the process where

input and output, f ðÞ : Rny þnu dþ1 ! R denotes the non-
linear mapping, ny 2 Z+ and nu 2 Z+ are the orders of out- j ðkÞ
l Y
n

put and input respectively, d 2 Z+ represents the known xj ðkÞ ¼ PL ; j ðkÞ ¼

l lij ðkÞ; ð4Þ
j ðkÞ
j¼1 l i¼1
time-delay of the system, and nðkÞ 2 R is a sequence of
!
zero-mean Gaussian white noise. ðxi ðkÞ þ lij ðk  1Þ#ij  mij Þ2
To design the proposed RFNN-based predictive con- lij ðkÞ ¼ exp  : ð5Þ
2r2ij
troller, the global operation of the process system (1) can
be divided into several local operating regions. Within each
local region, Rj, a linear ARMAX model is used to repre- And, mij, rij and
PL #ij are the tuning parameters. With (4),
sent the local process behavior. In doing so, the process 0 6 xj(k) 6 1, j¼1 xj ðkÞ ¼ 1.
system (1) can be described as The above mentioned RFNN has been proven a univer-
sal approximator from the Stone–Weierstrass theorem and
Rj : IF x1 ðkÞ is F 1j ; . . . ; xi ðkÞ is F ij ; . . . ; and xn ðkÞ is F nj ; the literatures in [30,31,36].
THEN y j ðkÞ ¼ aj ðz1 Þyðk  1Þ þ bj ðz1 Þuðk  dÞ þ nðkÞ;
2.1. Learning algorithm of RFNN
aj ðz1 Þ ¼ a1j þ a2j z1 þ    þ any j zðny 1Þ ;
To derive the learning algorithm for the parameters mij,
bj ðz1 Þ ¼ b0j þ b1j z1 þ b2j z2 þ    þ bnu j znu ;
rij and #ij in the precondition part of the RFNN, the error
ð2Þ function w(k) of the RFNN is defined by
1
where X ðkÞ ¼ ½ x1 ðkÞ x2 ðkÞ    xn ðkÞ T is the input vec- wðkÞ ¼ ðyðkÞ  ^y ðkÞÞ2 ; ð6Þ
2
tor to the RFNN model at time k with xi ðkÞ 2 PL PL
R ði ¼ 1; . . . ; nÞ, Fij is the fuzzy set corresponding to xi(k) where ^y ðkÞ ¼ j¼1 xj ðkÞ^y j ðkÞ ¼ j¼1 xj ðkÞ½aj ðz1 Þyðk  1Þþ
in the jth fuzzy implication (FI), yj(k) is the process bj ðz1 Þuðk  dÞ is the current output at each time instant
model output in the jth operating region, and n = ny + k. By using the backpropagation learning algorithm in
nu  d + 1. [31], the parameter vector W ¼ ½ wT1 wT2   wTL T with
T
Let lij(k) be the membership degree of the xj(k) in wj ¼ ½m1j   mnj r1j  rnj #1j   #n  is recursively
the fuzzy set Fij, and the complete RFNN model for the updated along with W ðk þ 1Þ ¼ W ðkÞ  g owðkÞ oW
where g is a
process system (1) be assumed to have L FIs. Then the positive learning rate. The updating laws for mij, rij, and
RFNN model of the process system (1) can be inferred #ij are expressed by
from [29] as
owðkÞ
X
L mij ðk þ 1Þ ¼ mij ðkÞ  g
yðkÞ ¼ 1 1
xj ðkÞ½aj ðz Þyðk  1Þ þ bj ðz Þuðk  dÞ þ nðkÞ; omij
j¼1 o^y ðkÞ
¼ mij ðkÞ þ gðyðkÞ  ^y ðkÞÞ ; ð7Þ
ð3Þ omij
86 C.-H. Lu, C.-C. Tsai / Journal of Process Control 17 (2007) 83–92

owðkÞ Note that the above recursive least-squares equations can

rij ðk þ 1Þ ¼ rij ðkÞ  g
orij be easily derived by the method in [32].
o^y ðkÞ 2.2. Convergence of RFNN
¼ rij ðkÞ þ gðyðkÞ  ^y ðkÞÞ ; ð8Þ
orij
owðkÞ The subsection develops a convergence theorem for
#ij ðk þ 1Þ ¼ #ij ðkÞ  g selecting an appropriate learning rate g. If a small value
o#ij
is given to the learning rate, the convergence of the RFNN
o^y ðkÞ will be guaranteed; but the convergence speed may be very
¼ #ij ðkÞ þ gðyðkÞ  ^y ðkÞÞ ; ð9Þ
o#ij slow. On the other hand, if a larger learning rate is chosen,
the RFNN may become unstable. Accordingly, choosing
where an appropriate learning rate is very critical. The following
o^y ðkÞ lj ðkÞ olij ðkÞ
o^y ðkÞ o theorem shows how this learning rate should be chosen.
¼
omij lj ðkÞ olij ðkÞ omij
o Theorem 1. Assume that the parameters mij, rij, #ij and hj for
ðxi ðkÞ þ lij ðk  1Þ#ij  mij Þ the RFNN model (3) are updated along with (7), (8), (9), (14)
¼ ð^y j ðkÞ  ^y ðkÞÞxj ðkÞ ; ð10Þ and (15). Then the output of the RFNN is convergent,
r2ij
provided that its learning rate g satisfies the following
o^y ðkÞ lj ðkÞ olij ðkÞ
o^y ðkÞ o condition:
¼
orij lj ðkÞ olij ðkÞ orij
o b
g¼  2  2  2  ; ð16Þ
ðxi ðkÞ þ lij ðk  1Þ#ij  mij Þ 2 PL Pn o^y ðkÞ o^y ðkÞ o^y ðkÞ
¼ ð^y j ðkÞ  ^y ðkÞÞxj ðkÞ ; ð11Þ j¼1 i¼1 omij
þ orij
þ o#ij
r3ij

o^y ðkÞ lj ðkÞ olij ðkÞ

o^y ðkÞ o where 0 < b < 2, and o^y ðkÞ=omij ; o^y ðkÞ=orij ; o^y ðkÞ=o#ij are
¼ obtained from (10)–(12), respectively.
o#ij lj ðkÞ olij ðkÞ o#ij
o
ðmij  xi ðkÞ  lij ðk  1Þ#ij Þ Proof. Let a Lyapunov function candidate be chosen as
¼ ð^y j ðkÞ  ^y ðkÞÞxj ðkÞ 2
r2ij ‘ðkÞ ¼ ðyðkÞ  ^y ðkÞÞ ¼ e2 ðkÞ and define d‘(k)  ‘(k + 1) 
‘(k). Then d‘(k) becomes
 lij ðk  1Þ: ð12Þ
d‘ðkÞ ¼ e2 ðk þ 1Þ  e2 ðkÞ ¼ ðdeðkÞ þ eðkÞÞ2  e2 ðkÞ
Next, the RLS method is adopted to identify the 2
unknown parameters in the consequent part of the RFNN. ¼ 2eðkÞdeðkÞ þ ðdeðkÞÞ : ð17Þ
Thus, ^y ðkÞ can be rewritten as By the method in [33], de(k) can be represented as
T
^y ðkÞ ¼ f ðkÞh; ð13Þ de(k) = [oe(k)/oW]TdW where h dWih W(ki + 1)   W(k) =
o^y ðkÞ
T
o^y ðkÞ o^y ðkÞ2
g (ow(k)/oW). With the fact oW oW
¼  oW  , (17)
where
becomes
T
fðkÞ ¼ ½ fT1 ðkÞ fT2 ðkÞ    fTL ðkÞ  ;     !2
o^y ðkÞ2 o^y ðkÞ2
h ¼ ½ hT1 hT2 
T
hTL  ; d‘ðkÞ ¼ 2ge ðkÞ
2 
 oW  þ geðkÞ oW 
 

fj ðkÞ ¼ ½ xj ðkÞyðk  1Þ    xj ðkÞyðk  ny Þ  2  2 !

 o^
y ðkÞ   o^
y ðkÞ 
¼ ge2 ðkÞ 
 oW  2  g oW 
 
xj ðkÞuðk  dÞ    xj ðkÞuðk  nu Þ T ;
 
X  T  !!
o^y ðkÞ2 L
o^y ðkÞ o^y ðkÞ
hj ¼ ½ a1j    an y j b0i    bnu j T : 2
¼ ge ðkÞ   2g
oW  j¼1
owj owj
Define the error
P covariance matrix P(k) = (HT(k)H(k))1  2  2
T
where HðkÞ ¼ k f ðkÞfðkÞ, and assume that the matrix o^y ðkÞ X L X n
o^y ðkÞ
2
¼ ge ðkÞ   2g
H(k) has full rank for "k P 0. Given an initial error oW  j¼1 i¼1
omij
estimate h(0) and an initial error covariance matrix P(0),  2  2 !!
the least-squares estimate of H(k) satisfies the following o^y ðkÞ o^y ðkÞ
þ þ : ð18Þ
recursive equations: orij o#ij

P ðk  1ÞfT ðkÞðyðkÞ  fðkÞhðk  1ÞÞ To ensure a convergent learning, the learning rate g is set
hðkÞ ¼ hðk  1Þ þ ; ð14Þ according to (16) so as to have d‘ðkÞ ¼ ge2 ðkÞko^y ðkÞ=
1 þ fðkÞP ðk  1ÞfT ðkÞ
oW k2 < 0, which implies that the output of the RFNN con-
P ðk  1ÞfT ðkÞfðkÞP ðk  1Þ
P ðkÞ ¼ P ðk  1Þ  : ð15Þ verges. With such a convergent RFNN, Theorems 2 and 3
1 þ fðkÞP ðk  1ÞfT ðkÞ in Section 3 will show that the controlled tracking error
 C.-H. Lu, C.-C. Tsai / Journal of Process Control 17 (2007) 83–92 87

converges asymptotically to zero and the closed-loop sys- Dzp ep ðz1 Þaðz1 ÞyðkÞ ¼ Dzp ep ðz1 Þbðz1 Þzd uðkÞ þ 
nðkÞ;
tem is stable. h ð23Þ
where
3. Derivation and analysis of predictive control law
nðkÞ ¼ Dzp ep ðz1 ÞnðkÞ:
3.1. Predictive control law with integral action
By using (21) and (22), (23) becomes
This subsection is devoted to the derivation of the pre- yðk þ pjkÞ ¼ fp ðz1 ÞyðkÞ þ gp ðz1 ÞDuðk þ p  dÞ þ 
nðkÞ:
dictive control based on the RFNN model in order to
improve capabilities of tracking performance and distur- ð24Þ
bance rejection for this class of nonlinear industrial pro- Thus, the p step-ahead output prediction ^y ðkÞ is obtained
cesses. The control law is derived so as to minimize the from (25)
following cost function:
^y ðk þ pjkÞ ¼ fp ðz1 ÞyðkÞ þ gp ðz1 ÞDuðk þ p  dÞ: ð25Þ
X
Np
J ðkÞ ¼ ð^y ðk þ pjkÞ  ui rðk þ pÞÞ
2 To reduce the computational load, the control horizon
p¼d Nu is chosen as unity. Using (25) and setting
Du(k + 1) =    = Du(k + Np) = 0, the cost function J in
X
dþN u 1

þ ðqðz1 ÞDuðk þ p  djkÞÞ ;

2
ð19Þ (19) can be equivalently expressed in the following form:
p¼d
J eq ðkÞ ¼ ðFyðkÞ þ GDuðkÞ þ K  URÞT ðFyðkÞ
2
where r(k) is an input reference signal, ui is the feed for- þ GDuðkÞ þ K  URÞ þ ðqðz1 ÞDuðkÞÞ ; ð26Þ
ward gain and D = 1  z1. Moreover, q(z1) is a selected
weighting polynomial and qðz1 Þ ¼ q0 þ q1 z1 þ    þ where
qN p þnu 1 zðN p þnu 1Þ . Np and Nu respectively denote the out-
T
F ¼ fd ðz1 Þ fdþ1 ðz1 Þ    fN p ðz1 Þ ;
put horizon and the control horizon.
To derive the predictive control law and to find the p G ¼ ½ gd;0 gdþ1;0    gN p ;0 T ;
step-ahead prediction of y(k), the RFNN model (3) is "
rewritten as Pu 1
dþn Pu
dþn
K¼ gd;q Duðk  qÞ gdþ1;q Duðk  qÞ
q¼1 q¼1
aðz1 ÞyðkÞ ¼ 
 bðz1 Þuðk  dÞ þ nðkÞ; ð20Þ #T
N p þn
Pu 1
where  gN p ;q Duðk  qÞ ;
q¼1
X
L
aðz1 Þ ¼ 1  
 a1 z1  a2 z2      
any zny ; ai ¼
 xj ðkÞaij zi ; U ¼ diagfud udþ1  uN p g;
j¼1

X
L R ¼ ½ rðk þ dÞ rðk þ d þ 1Þ    rðk þ N p Þ T :
bðz1 Þ ¼ 
 b1 z1 þ 
b2 z2 þ    þ 
bnu znu ; 
bi ¼ i
xj ðkÞbij z :
j¼1 Because the cost function J(k) is quadratic in Du(k), a min-
imum solution for Du(k) is easily obtained from
Next, given a polynomial ep(z1), polynomials fp(z1) and 
oJ eq ðkÞ 
gp(z1) can be defined as follows: ¼ 0; ð27Þ
oðDuðkÞÞDuðkÞ¼Du ðkÞ
1 ¼ Dep ðz1 Þ
aðz1 Þ þ zp fp ðz1 Þ; ð21Þ
which yields the following equation:
 1 Þ;
gp ðz1 Þ ¼ Dep ðz1 Þbðz ð22Þ
GT ðFyðkÞ þ K  URÞ þ ðGT G þ qðz1 Þq0 ÞDuðkÞ ¼ 0: ð28Þ
where From Eq. (28), it leads to
1
D¼1z ; GT ðFyðkÞ  URÞ þ GT K þ ðGT G þ q20 ÞDuðkÞ
ep ðz1 Þ ¼ 1 þ ep;1 z1 þ ep;2 z2 þ    þ ep;p1 zðp1Þ ; þ q0 ðq1 z1 þ q2 z2 þ    qN p zN p ÞDuðkÞ ¼ 0;
fp ðz1 Þ ¼ fp;0 þ fp;1 z1 þ fp;2 z2 þ    þ fp;ny zny ;
GT ðFyðkÞ  URÞ þ ðGT G þ q20 ÞDuðkÞ
1 1 2 ðpþnu 1Þ
gp ðz Þ ¼ gp;0 þ gp;1 z þ gp;2 z þ    þ gp;pþnu 1 z :
X
Np
þ GT K þ q0 qi Duðk  iÞ ¼ 0;
Premultiplying (20) by Dzpep(z1) gives the following i¼1
equality:
GT ðFyðkÞ  URÞ þ ðGT G þ q20 ÞDuðkÞ þ HðkÞ ¼ 0;
88 C.-H. Lu, C.-C. Tsai / Journal of Process Control 17 (2007) 83–92
1  
where Bðz ÞrðkÞ nðkÞ
! lim EfyðkÞ  rg ¼ lim E þ lim E
X
Np k!1 k!1 Aðz1 Þ k!1 Aðz1 Þ

HðkÞ ¼ gd;0 gi;1 þ q0 q1 Duðk  1Þ nðk  1Þ
i¼d  lim E  lim Efrg
k!1 Aðz1 Þ k!1
!  
X
Np 1
Bðz ÞrðkÞ  EfnðkÞg
þ gdþ1;0 gi;2 þ q0 q2 Duðk  2Þ þ    ¼ þ
i¼d
Aðz1 Þ z¼1 Aðz1 Þ z¼1

! Efnðk  1Þg
X
Np
 r
þ g2d1;0 gi;d þ q0 qd Duðk  d Þ Aðz1 Þ z¼1

i¼d
bðz1 Þzd ðGT G þ kÞ1 Xr
 
! 
X
Np ¼ 1 T 
 r
D 1  1 d T
aðz Þ þ bðz Þz ðG G þ kÞ G F z¼1
þ g2d;0 gi;dþ1 þ q0 qdþ1 Duðk  d  1Þ þ    
i¼dþ1
bðz1 Þzd ðGT G þ kÞ1 Xr 

! ¼  r
X
Np
bðz1 Þzd ðGT G þ kÞ1 GT F z¼1

þ gN p 1;0 gi;N p 1 þ q0 qN p  1 Duðk  N p  1Þ P 
Np
i¼N p 1 g
p¼d p;0 p f ð1Þ r
¼ PN p  r ¼ 0;
þ ðgN p ;0 gN p ;N p þ q0 qN p ÞDuðk  N p Þ: ð29Þ p¼d g p;0 fp ð1Þ

In order to obtain the simplified control law, we set

k ¼ q20 and set q1 ; q2 ; . . . ; qN p satisfy H(k) = 0. Hence, the which completes the proof of Theorem 2. h
current control increment is expressed by
Theorem 3. Assume that the RFNN model (3) is controlled
GT ðUR  FyðkÞÞ by (30), the upper bounds for ny, nu and d are known, the esti-
DuðkÞ ¼ ; k > 0: ð30Þ
GT G þ k mated parameters of mij, rij, #ij and hj are bounded, and all
setpoints are constant. Let limp!1E{r  y(k + p)} be zero,
3.2. Steady-state performance and stability analyses as can be referred to Theorem 2. Then the closed-loop system
is stable.
To guarantee a successful identification, the persistent
excitation (PE) signals are used [32]. The RFNN model Proof. By choosing a Lyapunov function candidate
has been shown stable via the Lyapunov stability theory X
n

in Section 2. Thus the estimates mij, rij, #ij and hj are con- LðkÞ ¼ lim ðEfr  yðk þ p  1ÞgÞ2 ð33Þ
n!1
p¼1
vergent and uniformly bounded. The following theorems
state that the aforementioned control law (30) has zero we have
steady-state tracking errors and the resulting closed-loop
X
n
system is stable. Lðk þ 1Þ ¼ lim ðEfr  yðk þ pÞgÞ
2
n!1
p¼1
Theorem 2. Assume that the predictive control law (30) is
applied to the RFNN model (3), the upper bounds for ny, nu X
n
2 2
¼ lim ðEfr  yðk þ p  1ÞgÞ  ðEfr  yðkÞgÞ
and d are known, the estimated parameters of mij, rij, #ij and n!1
p¼1
hj are bounded, and all setpoints are constant, i.e. r(k) = r.
Then the closed-loop system has the following property: þ lim ðEfr  yðk þ pÞgÞ2
p!1

lim EfyðkÞ  rg ¼ 0: ð31Þ ¼ LðkÞ  ðr  Ef^y ðkÞ þ nðkÞgÞ2

k!1
2
¼ LðkÞ  ðr  ^y ðkÞÞ :
Proof. Substituting the control law (30) into (3) with the
well-estimated parameters mij, rij, #ij and hj yields the Since L(k + 1) 6 L(k), we can conclude by Lyapunov’s sta-
resulting closed-loop system bility theorem that the closed-loop system is stable. h

Bðz1 ÞrðkÞ nðkÞ nðk  1Þ

yðkÞ ¼ 1
þ 1
 ; ð32Þ 3.3. Real-time predictive control algorithm
Aðz Þ Aðz Þ Aðz1 Þ
where Aðz1 Þ ¼ D bðz1 Þzd ðGT G þ kÞ1 GT F
aðz1 Þ þ  A real-time control algorithm is employed recursively to
and Bðz Þ ¼ bðz Þz ðG G þ kÞ1 X with X ¼ GT U
1  1 d T
update the RFNN and therefore the predictive controller’s
½ 1 1    1 T Þ. Using the properties E{n(k)} = E{n parameters. Fig. 2 is a schematic diagram that presents the
|fflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflffl}
N p dþ1 structure of the predictive control using the RFNN model.
(k  1)} = 0, Djz=1 = 0 and choosing the feed forward gain The computations at each time instant k can be outlined as
ui to satisfy ui = fi(z1)jz=1, one obtains the following steps:
 C.-H. Lu, C.-C. Tsai / Journal of Process Control 17 (2007) 83–92 89

yðkÞ ¼ 0:9722yðk  1Þ þ 0:3578uðk  1Þ

 0:1295uðk  2Þ  0:3103yðk  1Þuðk  1Þ
RFNN
 0:04228y 2 ðk  2Þ þ 0:1663yðk  2Þuðk  2Þ
Process Process  0:03259y 2 ðk  1Þyðk  2Þ  0:3513y 2 ðk  1Þuðk  2Þ
Set-Point
Input Output þ 0:3084yðk  1Þyðk  2Þuðk  2Þ
GPC Process
þ 0:1087yðk  2Þuðk  1Þuðk  2Þ: ð35Þ

The simulation was performed for two sets of reference in-

puts r(k) and the parameters of the predictive controller
specified by

Fig. 2. Generalized predictive control using the RFNN. 1; 0 < k 6 400;
rðkÞ ¼ N p ¼ 50; k ¼ 50:
0; 400 < k 6 800;

From the process model (35), it clearly indicates that the

Step 1. Set r(k), ny, nu, d, L, k and Np.
output order, the input order, the time-delay and input
Step 2. Measure the current plant output y(k).
vector of RFNN model can be easily specified by ny = 2,
Step 3. Update mij, rij, #ij and hj of the RFNN using (7),
nu = 2, d=1 and X(k) = [y(k  1)y(k  2)u(k  1)
(8), (9), (14) and (15) and an adaptive learning rate
u(k  2)]T. The key parameter of the RFNN model is cho-
(ALR) from (34).
sen as L = 3, which were shown effective under computer
Step 4. Compute the increment control output Du(k) via
simulations.
(30).
Fig. 3 shows the output response and the control signal,
Step 5. Output the control signal u(k) to the controlled
respectively, of the predictive control under setpoint
process.
changes. The result in Fig. 3(a) indicates that the proposed
Step 6. Repeat steps 2–5.
control system has a short setting time and a small maxi-
mum overshoot.
Note that the control signal u(k) in Step 5 must be lim-
In order to investigate disturbance rejection capability
ited to satisfy the input constraint of the controlled process.
of the proposed controller, the mathematical model (35)
To guarantee the ALR inside the stable region, the ALR
was perturbed by a disturbance v(k), where v(k) = 0.05
for training the weights of the RFNN is obtained from
for 200 6 k < 600 and v(k) = 0.2 for k P 600. Fig. 4 shows
Theorem 1 and given by
the simulation results which reveal that the proposed con-
1 troller demonstrates a good disturbance rejection
g¼  2  2  2  : ð34Þ
PL Pn o^y ðkÞ o^y ðkÞ o^y ðkÞ capability.
j¼1 i¼1 omij
þ orij
þ o#ij The setpoint tracking responses from a velocity-type
PID controller (Kp = 1, KI = 0.2 and KD = 0.01) and a

1.5
4. Illustrative examples
1
In this section, two illustrative examples are provided to
demonstrate the performance of the proposed RFNN- 0.5
y

based predictive control. The examples also show the effect

0
of setpoint changes and load disturbances on the control
systems employing the proposed controller. -0.5
0 100 200 300 400 500 600 700 800
Sample Number
4.1. Simulation of a laboratory-scale liquid-level process
4

A simulated laboratory scale liquid-level system of Sales 2

and Billings [34,35] is considered. The simulated system is
composed of a DC pump to feed water into a conical flask 0
u

that, in turn, feeds square tank, giving the system second-

-2
order dynamics. The input is the voltage to the pump
motor and the system output is the height of the water in -4
the conical flask. The aim, under simulation conditions, 0 100 200 300 400 500 600 700 800
Sample Number
is to follow some demanding trajectory for the water level.
The process model was identified as follows: Fig. 3. Performance of the RFNN-based predictive controller. (a)
Simulated setpoint tracking response. (b) Control signal.
90 C.-H. Lu, C.-C. Tsai / Journal of Process Control 17 (2007) 83–92

1.5 1.5

1 1
y

0.5 0.5

y
0 0 PID controller
GPC controller
-0.5 -0.5
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Sample Number Sample Number

4
1.5
2 1

y
0 0.5
u

0
-2 PID controller
-0.5 GPC controller
-4
0 100 200 300 400 500 600 700 800 0 100 200 300 400 500 600 700 800
Sample Number Sample Number

Fig. 4. Performance of the proposed controller in the presence of load Fig. 5. Performance of the PID controller and the GPC controller. (a)
disturbances. Simulation results for setpoint tracking. (b) Setpoint tracking responses
under external disturbances.

classical GPC controller (Np = Nu = 20 and k = 3) are

given in Fig. 5. Fig. 5(a) shows that the PID and GPC con- condenser, an expander, an evaporator, a pump and a heat
trol systems have good response in the absence of external exchanger. Such a machine has been designed to have a
loads. In the presence of disturbances v(k) for the duration cooling capability from 1200 kcal/h at 30 Hz to
k > 600, Fig. 5(b) discloses that PID controller has an 2900 kcal/h at 90 Hz. The controller has two platinum
unstable tracking performance, and the GPC controller temperature-sensing modules connected to resistance-to-
produces big overshoot. The results in Figs. 4 and 5 clearly voltage (R–V) transducers within an accuracy of and has
indicate that the proposed controller outperforms the con- an AD/DA board inserted in the slot of a personal
ventional PID controller and the classical GPC controller. computer (PC). There are 32-channel 12-bit analog-to-dig-
These results also reveal the usefulness of the proposed ital converters and 2-channel 12-bit digital-to-analog
RFNN-based controller, especially, for this class of nonlin- converters ranging from 0 to 5 V. The 586 compatible PC
ear process systems. is responsible for editing, compiling, and running C
program codes for the controller. The sampling period in
this control system adopted here is 3 s. Fig. 6 shows a sche-
4.2. Real-time application to a variable-frequency matic diagram of the PC-based temperature control
oil-cooling process system.
The first experiment was conducted to test the setpoint
The variable-frequency oil-cooling machine is indeed a tracking capability of the proposed method. The four
temperature controlled process extensively used in the parameters and the input vector obtained from the experi-
manufacturing industry. It has been widely used to provide mental input/output data of the oil-cooling system are
appropriate oil coolant temperature for high-speed ny = 2, nu = 13, d = 12, L = 3 and X(k) = [y(k  1)y
machine tools, such as cutting, milling, and drilling (k  2)u(k  12)u(k  13)]T. The experiment was operated
machines. The oil-cooling process has a complicated ther- under two temperature setpoints as well as the parameters
mal dynamics caused by the variable-frequency compressor of the predictive controller being specified
and the heat exchange subsystem, and generates a nonlin-
ear and time-delay dynamic behavior. In order to have
20  C; 0 < k 6 600
good manufacturing performance with the machine tools, rðkÞ ¼ ; N p ¼ 80; k ¼ 1800:
the temperature controller must precisely adjust the 18  C; 600 < k 6 1200
rotational speed of the induction motor. The oil cooling
process is equipped with a three-phase 220 V rotary com- Consequently, the control signal u(k) for the process is
pressor driven by a variable-frequency induction motor, a given by
PN p PN p PN p
p¼d g p;0 up rðk þ pÞ  p¼d gp;0 fp;0 yðkÞ  p¼d gp;0 fp;1 yðk  1Þ
uðkÞ ¼ uðk  1Þ þ PN p 2 ; ð36Þ
p¼d g p;0 þ k
 C.-H. Lu, C.-C. Tsai / Journal of Process Control 17 (2007) 83–92 91

R-V Transducer

Pt 100

Expander

Condenser Evaporator Heat Exchanger

Compressor

A/D

D/A
IBM 586 Compatible PC Oil Tank
Pump
Variable-Frequency
Induction Motor

Fig. 6. A schematic diagram of the PC-based temperature control system.

where 25
!r ! !
X
pd X
L X
L
gp;0 ¼ xj ðkÞa1j ðkÞ xj ðkÞb0j ðkÞ ;
24
r¼1 j¼1 j¼1
!r
X
p X
L
up ¼ 1; f p;0 ¼ 1 þ xj ðkÞa1j ðkÞ ; 23
Temperature (ºC )

r¼1 j¼1
!r
X
p X
L
22
fp;1 ¼  xj ðkÞa1j ðkÞ :
r¼1 j¼1

ð37Þ 21

Fig. 7 presents the setpoint tracking response of the PC-

based temperature control system. The steady-state errors 20
of the variable-frequency oil-cooling process remained
within ±0.5 C.
19
0 200 400 600 800 1000 1200
3 sec/sample

28 Fig. 8. Step response of the proposed controller with the fixed load of
1000 W for k P 570.

26
The second experiment was performed to demonstrate
the temperature regulation in the presence of a heat load
24
change of 1000 W during the time duration k P 570.
Temperature ( º C )

Fig. 8 displays the temperature response of the proposed

22 controller with the setpoint r(k) = 20 C. The result shows
that the proposed predictive control has acceptable distur-
bance rejection ability. Figs. 7 and 8 present the usefulness
20
of the proposed controller for such industrial processes
under setpoint and load changes.
18

5. Conclusions
16
0 200 400 600 800 1000 1200 This paper has proposed a systematic design methodol-
3 sec/sample
ogy to develop an RFNN-based predictive control with
Fig. 7. Setpoint tracking response of the proposed controller. integral action for a class of nonlinear process systems.
92 C.-H. Lu, C.-C. Tsai / Journal of Process Control 17 (2007) 83–92

The setpoint tracking and load disturbance rejection capa- [15] C.H. Lu, C.C. Tsai, Adaptive neural predictive control for industrial
bilities of the proposed method can be improved by the multivariable processes, Journal of Systems and Control Engineering
218 (7) (2004) 557–567.
modified GPC cost function. Two sufficient conditions [16] S. Mollov, R. Babuška, J. Abonyi, H.B. Verbruggen, Effective
have been established for securing both steady-state perfor- optimization for fuzzy model predictive control, IEEE Transactions
mance and closed-loop stability of the system. The real- on Fuzzy Systems 12 (5) (2004) 661–675.
time control algorithm, including both the RLS estimator [17] J.M. Sousa, Optimization issues in predictive control with fuzzy
and the learning law for the RFNN, has both been success- objective functions, International Journal of Intelligent System 15
(2000) 879–899.
fully applied to a simulated and a real process system, [18] M. Mahfouf, D.A. Linkens, M.F. Abbod, Multi-objective genetic
respectively. Through the results from computer simula- optimization of GPC and SOFLC tuning parameters using a fuzzy-
tions and experiments, the proposed control system has based ranking method, IEE Proceedings—Control Theory and
been proven useful and pragmatic for achieving setpoint Applications 147 (3) (2000) 344–354.
tracking either in the absence or in the presence of external [19] J.M.D.C. Sousa, U. Kaymak, Model prediction control using fuzzy
decision functions, IEEE Transactions on System, Man, and Cyber-
disturbances and load changes. netics—part B: Cybernetics 31 (1) (2001) 54–65.
[20] E. Ali, Heuristic on-line tuning for nonlinear model predictive
Acknowledgements controllers using fuzzy logic, Journal of Process Control 13 (2003)
383–396.
The authors greatly acknowledge the financial support [21] A. Flores, D. Sáez, J. Araya, M. Berenguel, A. Cipriano, Fuzzy
predictive control of a solar power plant, IEEE Transactions on
from the National Science Council of Taiwan, ROC, under Fuzzy Systems 13 (1) (2005) 58–68.
Grant NSC93-2213-E-005-008. We also wish to thank the [22] C.F. Juang, C.T. Lin, A recurrent self-organizing neural fuzzy
editor and the reviewers for their many helpful comments inference network, IEEE Transactions on Neural Networks 10 (4)
and suggestions. (1999) 828–845.
[23] F.J. Lin, R.J. Wai, Hybrid Control using recurrent fuzzy neural
network for linear-induction motor servo motor, IEEE Transactions
References on Fuzzy Systems 9 (1) (2001) 102–115.
[24] S.M. Zhou, L.D. Xu, A new type of recurrent fuzzy neural network
[1] M.A. Henson, D.E. Seborg, Nonlinear Process Control, Prentice- for modeling dynamic systems, Knowledge-based Systems 14 (2001)
Hall, New Jersey, 1997. 243–251.
[2] D.W. Clarke, Application of generalized predictive control [25] P.A. Mastorocostas, J.B. Theocharis, An orthogonal least-squares
to industrial processes, IEEE Control Systems Magazine 8 (1988) method for recurrent fuzzy-neural modeling, Fuzzy Sets and Systems
49–55. 140 (2003) 285–300.
[3] J.B. Rawlings, Tutorial overview of model predictive control, IEEE [26] Y.M. Zhang, R. Kovacevic, Neurofuzzy model-based predictive
Control Systems Magazine (4) (2000) 38–52. control of weld fusion zone geometry, IEEE Transactions on Fuzzy
[4] F. Allgower, A. Zheng, Nonlinear Model Predictive Control, Systems 6 (3) (1998) 389–401.
Birkhauser Verlag, Germany, 2000. [27] X. Li, Z. Chen, Z. Yuan, Generalized predictive control using fuzzy
[5] B. Kouvaritakis, M. Cannon, Nonlinear Predictive Control – Theory neural networks in T–S model, in: Proceeding of the 3th World
and Practice, IEE, London, 2001. Congress on Intelligent Control and Automation, Hefei, PR China,
[6] S. Huang, K.K. Tan, T.H. Lee, Applied Predictive Control, Springer- 2000, pp. 885–889.
Verlag, London, 2002. [28] J. Zhang, A.J. Morris, Recurrent neural-fuzzy networks for nonlinear
[7] G.D. Nicolao, L. Magi, R. Scattolini, Stabilizing predictive process modeling, IEEE Transactions on Neural Networks 10 (2)
control of nonlinear ARX Models, Automatica 33 (9) (1997) (1999) 313–326.
1691–1697. [29] C.C. Hsiao, S.F. Su, T.T. Lee, C.C. Chuang, Hybrid compensation
[8] W.H. Chen, D.J. Balance, P.J. Gawthrop, J.J. Gribble, J. O’Reilly, control for affine TSK fuzzy control systems, IEEE Transactions on
Nonlinear PID predictive controller, IEE Proceedings—Control System, Man, and Cybernetics—part B: Cybernetics 34 (4) (2004)
Theory and Applications 146 (6) (1999) 603–611. 1865–1873.
[9] S. Kanev, M. Verhaegen, Controller reconfiguration for non-linear [30] W. Rudin, Principles of Mathematical Analysis, McGraw-Hill, New
systems, Control Engineering Practice 8 (2000) 1223–1235. York, 1976.
[10] Y. Cao, A formulation of nonlinear model predictive control using [31] C.H. Lee, C.C. Teng, Identification and control of dynamic systems
automatic differentiation, Journal of Process Control 15 (2005) 851– using recurrent fuzzy neural networks, IEEE Transactions on Fuzzy
858. Systems 8 (4) (2000) 349–366.
[11] P.F. Tsai, J.Z. Chu, S.S. Jang, S.S. Shieh, Developing a robust model [32] K.J. Åström, B. Wittenmark, Adaptive Control, Addison-Wesley,
predictive control architecture through regional knowledge analysis New Jersey, 1995.
of artificial neural networks, Journal of Process Control 13 (2002) [33] C.C. Ku, K.Y. Lee, Diagonal recurrent neural networks for dynam-
423–435. ical system control, IEEE Transactions on Neural Networks 6 (1)
[12] J.M. Zamarreño, P. Vega, Neural predictive control, application to a (1995) 144–156.
highly non-linear system, Engineering Applications of Artificial [34] K.R. Sales, S.A. Billings, Self-tuning control of nonlinear ARMAX
Intelligence 12 (1999) 149–158. model, International Journal of Control 51 (1990) 753–769.
[13] A.G. Palos, S. Parthasarathy, A.F. Atiya, Neural-predictive process [35] M.S. Ahmed, Neural-net-based direct adaptive control for a class
control using on-line controller adaptation, IEEE Transactions on nonlinear plants, IEEE Transactions on Automatic Control 45 (1)
Control System Technology 9 (5) (2001) 741–755. (2000) 119–124.
[14] J.Q. Huang, F.L. Lewis, Neural-network predictive control for [36] C.J. Lin, C.H. Chen, A compensation-based recurrent fuzzy neural
nonlinear dynamic systems with time-delay, IEEE Transactions on network for dynamic system identification, European Journal of
Neural Networks 14 (2) (2003) 377–389. Operational Research 172 (2006) 696–715.