See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/291321108

NEURO-FUZZY BASED MODEL PREDICTIVE CONTROL

Conference Paper · May 2000

CITATION READS
1 98

2 authors, including:

Ebrahim Abdulla Mattar

University of Bahrain
185 PUBLICATIONS 446 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

EEG Classification and Decoding View project

Sensor Fusion Mobile Robot Navigation System View project

All content following this page was uploaded by Ebrahim Abdulla Mattar on 24 December 2016.

The user has requested enhancement of the downloaded file.

 NEURO-FUZZY BASED MODEL PREDICTIVE CONTROL
E. A. Al-GALLAF AMIN M. SUULTAN
Department of Electrical and Electronics Engineering, Bahrain Aluminum Company
University of Bahrain, P.O. Box 13184 State of Bahrain
State of Bahrain. email : [email protected]

ABSTRACT major criticism is that these models are mathematically

It has been found, that the Model Predictive Control opaque, and there is no formal mathematical
(MPC) principle has recently been incorporated with representation of the system's behavior. This prevents
fuzzy models. This provides design methods for neuro- the application of conventional empirical modeling
fuzzy model based controllers because predictive techniques to fuzzy systems, making model validation
methods have several advantages that make them good and comparison hard to perform. Furthermore, due to
candidates for industrial applications. This research the vagueness and subjectivity of natural language
concentrates on aspects of neuro-fuzzy based MPC for statements, fuzzy systems based on qualitative
multivariable systems. Mathematical formulation of knowledge alone are unlikely to adequately model
linearized MPC is utilized to introduce the concept of simple system. To circumvent these inadequacies, as in
neuro-fuzzy based MPC scheme, then neuro-fuuzy conventional empirical modeling, available data should
MPC is constructed based on a modeled pH reactor. be used to adjust and validate the model's behavior.
Neuro-fuzzy MPC has been constructed by neural Efforts to combine both empirical and qualitative
network representation of fuzzy models. Hence the modeling have lead to the development of neuro-fuzzy
neuro-fuzzy based MPC algorithm derivation for a modeling techniques. Such techniques allow both
MIMO constrained system is presented. The research linguistic system description and empirical data to be
includes simulation results of the neuro-fuzzy model fully utilized during system identification cycle.
for the nonlinear pH reactor.
Neuro-fuzzy models are basically adaptive fuzzy
KEYWORDS : Identification of Multivariable, systems developed from exploiting similarities between
Neuro-fuzzzy Modeling Systems, Model Predictive fuzzy systems and certain forms of neural networks,
Control namely Associative Memory Networks which are a
particular form of a generalized linear model. This is
where the neuro in neurofuzzy is derived. The behavior
[1] INTRODUCTION of these models can either be represented by a set of
Traditionally, fuzzy controllers have been designed humanly understandable rules or by a generalized linear
without an explicit model of the process being model, making them an ideal framework for performing
controlled. However, in neuro-fuzzy systems, nonlinear system identification.
mathematical models are explicitly used [1]. It has been
found, that predictive control principle has recently been [2] FUZZY MODELING
incorporated with fuzzy models. This provides design Fuzzy models are useful for describing
methods for neuro-fuzzy model based controllers, since processes where the underlying physical
predictive methods have several advantages that make mechanisms are not completely known and
them good candidates for industrial applications. An where a process behavior is understood in
important requirement for any system identification qualitative terms. An important property of
technique is the ability to exploit available priori fuzzy models is their capability to represent
knowledge. Many conventional approaches rely on nonlinear dynamic systems. Therefore, fuzzy
depth physical knowledge describing the system. models can also be applied to systems that
However, for complex ill-defined systems such are well understood but due to the
knowledge is unavailable or limited. In these instances, nonlinearities untraceable with standard
an expert can often describe the behavior of the system linear methods. Rule-based structure of fuzzy
using natural language. models allows for integrating heuristic
knowledge with information obtained from
Since Zadeh's first paper [2], fuzzy algorithms behavior process measurements.
has been used to build models based on such humanistic
descriptions. Despite the apparent success of fuzzy The global operation of a nonlinear process is
systems, there are many aspects of their behavior which divided into several local operating regions.
are unsuited to system identification and modeling. The
Within each region Ri , a reduced order linear NEUROFUZZY ARCHITECTURE
model in ARMAX form is used to represent For simplicity, we assume the fuzzy inference system
the process behavior. This is not restrictive, under consideration has two inputs x(k-1) and y(k-R)
and any appropriate model forms can be and one output y(t). For instant, if the rule base
used. Fuzzy sets are used to define the contains two fuzzy if-then rules of Takagi and Sugeno's
process operating conditions such that fuzzy type [4], then a rule can be written as:
dynamic model of a nonlinear process can be
Rule 1:
described in the following way :
Ri : If operatingconditioni If x(k-1) is A1 and y(k-2) is B1 , then
f1  p1 x(k  1)  q1 y (k  2)  r1
 
 o i
Then yi (k )  aij y(k  j)  biju(k  j )(i  1,2,, r ) Rule 2:
j j
(1) If x(k-1) is A2 and y(k-2) is B2 , then
f 2  p 2 x(k  1)  q 2 y (k  2)  r2
The final model output is obtained through the center of
gravity defuzzification as follows [3]: where p, q, and r are constants and called parameter set.
That is, the if parts of the rules are same as in the
r
ordinary fuzzy if-then rules, then parts are linear
  y (k )

i i combinations of the input variables. The employed
 i 1
y (k )  r
(2) neuro-fuzzy architecture is shown in Fig. [1], where
 i
node functions in the layers are described below :
th
i 1 Layer 1: Each i node in this layer is a square node
In Equ (1) and Equ (2), y is the process output, u is the with a node function

process input, yi is the prediction of process output in
the i th operating region, (r) is the number of fuzzy Oi1   Ax x(k  1) (5)
operating regions, (i) and (o) are the time lags in the th
input and the output, respectively,  i is the where x(k-1) is the input to i node, and Ai is the
linguistic label (small , large, .. etc.) associated with this
membership function for the i th model, and finally, aij
node function. In other words, Oi1 is the membership
and bij are the ARMAX model parameters. The
function of Ai and it specifies the degree to which the
membership function for an operating region is given x satisfies the quantifier Ai . Usually we choose
constructed in a number of ways. One approach to
calculate its membership function as follows :  Ai ( x(k  1)) to be bell-shaped with maximum equal to
1 and minimum equal to 0, such as :
 i  min  h (x),  m (y )  (3)  Ai x((k  1)) 
1
(6)
 x(k  1)  c  2 
 i   h (x),  m (y ) (4) 1   i  
 b
 ai  

In Equ (3),  i is a membership function in i th
or
operating region,  h (x) is membership function of x
  x(k  1)  c  2 
being ‘high’, and  m (y ) is the membership function of  Ai ( x(k  1))  exp   i 
(7)
a  
y being ‘medium’. In Equ (4), we calculate the gradient   i  
required for gradient based network training. where {ai , bi , ci } is the parameter set. As values of
these parameters change, membership shaped functions
vary accordingly, thus exhibiting various forms of
[3] NEURAL NETWORK membership functions on the linguistic label Ai .
REPRESENTATION OF FUZZY Layer 2: Every node in this layer is a circle node which
MODELS multiplies incoming signals and sends their product out.
Fuzzy model described in section (2) can be represented For instance,
by a special type of network topology which is termed y i   Ai x(k  1)   Bi y (k  2) i  1,2 (8)
here a neuro-fuzzy. Fuzzy reasoning is capable of
handling uncertain and imprecise information while a Each node output represents the firing strength of a rule.
neural network is capable of learning from examples.
Neuro-fuzzy intend to combine the advantages of both
fuzzy reasoning and neural networks.
Layer 3: Every node in this layer is a circle node. The a weight vector which minimizes the following cost
function:
i th node calculates the ratio of the i th rule's firing N
1

strength to the sum of all rules' firing strengths: 
J N (w )  [ y ( k )  y ( x ( k ), w )] 2 (13)
N k 1
yi 
yi  i  1,2 (9) As the model, y (x(k ), w ) , is nonlinear with respect to
y1  y 2  y i the weights, linear optimization techniques cannot be
applied. Instead the popular Truncated Newton
Layer 4: Every node i in this layer is a square node with nonlinear optimization algorithm is employed. As an
a node function alternative a form of back fitting could be applied to
models, but given the slow convergence times of back
O14  yi f i  y i ( pi x(k  1)  qi y (k  2)  ri ) (10) fitting the direct approach is preferred.

where yi is the output of layer 3, and { pi , qi , ri } is the

parameter set. Parameters in this layer will be referred
[4] NEURO-FUZZY MODEL-BASED
to as consequent parameters. PREDICTIVE CONTROL
Layer 5: The node in this layer is a circle node. It
computes the overall output as the summation of all In the last section we presented neuro-fuzzy systems. In
incoming signals, i.e., this section we concentrate on employing these systems
y f  in model predictive control. Consider the following n
O15  overall output 
i

yi f i  i i i
yi 
(11) inputs and m outputs model :
i
n11 n12 n1n
Thus we have constructed an adaptive network which is
  c

y1(k)  c11( j)u1(k  j)  c12( j)u2 (k  j)  1n ( j)un (k  j)
functionally equivalent to a fuzzy inference system,
j1 j 1 j 1
achieved by fuzzy system alone.
n21 n22 n1n

c c c

TRAINING OF NEUROFUZZY SYSTEM y2 (k)  21( j)u1(k  j)  22( j)u2 (k  j)  2n ( j)un (k  j)
From the designed neuro-fuzzy architecture shown in j 1 j 1 j1

Fig[1], it is observed that, given values of premise nm1 nm2 nmn

c c c

parameters, the entire output is expressed as a linear ym (k)  m1( j)u1(k  j)  m2 ( j)u2 (k  j)  mn( j)un (k  j)
combinations of the consequent parameters. More j1 j 1 j 1
 (14)
precisely, output y in Fig. [1] can be rewritten as:
 y1  y2  y m 1   
ym  y1  y 2   y m 1 where ( y1 ,, y m ) are process outputs, (u1 , u 2 ,, u n )
y 1 y 2 y1 y 2 y m  y m 1
    are control variables, c k1km ( j )' s and
y m  y1 y1  y 2 y 2   y m 1 y m 1
 nk1km ' s(k1  1,2;k 2  1,2; j  1,,nknkm ) are the
y m  ( y1 x ( k  1)) p1  ( y1 y ( k  2 )) q1  ( y1 ) r1
parameters and orders of the model. The model
 ( y 2 x ( k  1)) p 2  ( y 2 y ( k  2 )) q 2  ( y 2 ) r2  parameters c knkm ( j )' s are constant for each given
(12)
(k1 , k m , j ) , since Equ (14) is a linear time-invariant
which is linear in the consequent parameters model.
( p1 , q1 , r1 , p2 , q2 and r2 ) . As a result, we have:
At (k), the controller needs to determine the control
S  set of total parameters action (u1 (k ),, u n (k )) based on feedback
S1  set of premise parameters ( y1 (k ),, y m (k )) to drive a process to reach the desired
S 2  set of consequent parameters outputs (r1 ,, rm ) .

The consequent parameters thus identified are optimal In predictive control, prediction equations should be
(in the consequent parameter space) under the condition developed to predict the outputs. Equ (20) can directly
that the premise parameters are fixed. yield the following recursive prediction equations, for a
two inputs with two outputs system.
PARAMETER IDENTIFICATION
A model's weights are conventionally identified by
performing maximum likelihood estimation. Given a
training data set Z N  {y (k ), x(k )}kN1 the task is to find
 n11

c
 
y1 (k  1)  y1 (k )  11 ( j , u 2 (k  j ))u1 (k  j ) [5] MODEL PREDICTIVE CONTROL
j 1
ADAPTATION FOR NONLINEAR
n11
PROCESSES
 cj 1
11 ( j , u 2 ( k  1  j ))u1 (k  1  j )
The principal components of a linearized MPC system
n12
are a reference neuro-fuzzy model, process output
 cj 1
12 ( j , u1 (k  j ))u 2 (k  j ) predictor, optimization routines or an adaptation
mechanism and MPC controller. In this sense, the
n12 MPC embodies the linearized model parameters of the
 cj 1
12 ( j , u1 (k  1  j ))u 2 (k  1  j ) overall system. In practice, many nonlinear processes
are approximated by reduced order models, possibly
n21 linear, which are clearly related to the underlying
c
 
y 2 (k  1)  y 2 (k )  21 ( j , u 2 (k  j ))u1 ( k  j ) process characteristics. However, these models may
j 1 only be valid within certain specific operating ranges.
n21 When operating conditions change, different model may
 c j 1
21 ( j , u 2 ( k  1  j ))u1 (k  1  j ) be required to be employed or the model parameters
may need to be adapted. This is specially true for linear
n22 representations.
 c j 1
22 ( j , u1 ( k  j ))u 2 (k  j )
Our approach to the modeling of nonlinear processes
n22
(via neuro-fuzzy system) is to divide the whole
 c j 1
22 ( j , u1 (k  1  j ))u 2 ( k  1  j ) envelope of process operation into several operating
regions, hence to use a local reduced order model to
approximate the process in each region. The role of the
(15) MPC mechanism is to select the model and the tuning
parameters of the controller in response to the error
where (k  1) and with initials between the outputs of neuro-fuzzy reference model and
plant [5].

y 1 ( k )  y1 ( k )
 (16) LINEARIZED FUZZY MODELS
y 2 (k )  y 2 (k )
A Takagi-Sugeno-Kang (TSK) fuzzy system [6] is
where notations c11 ( j , u 2 (k  j )), , emphasize that constructed typically from the following rules :
c11 ( j ),  , are dependent on u2 (k  j )) . Since we are IF u1 (k  1)isC1l andand u n (k  n)isCnl ,
dealing with constrained linear MPC formulation, to (18)
achieve a good control, it is required the following cost THEN y l (k  1)  C0l  C1l u1 (k  1)  Cnl u n (k  n)
function to be minimized:
1 where Cil are fuzzy sets, Cil are constants, and l = 1,2,
min J  u ( k ) T H u ( k )  G ( k  1) T u ( k )
u(k ) 2 u ... , M. That is, the IF parts of the rules are the same as
subject to (17) in the ordinary fuzzy IF-THEN rules, but the THEN
C u ( k )  c ( k  1) parts are linear combinations of the input variables.

u u Given an input u  (u1 ,,u n )T , the output y (k ) of a
u min  u ( k )  u max fuzzy system is computed as the weighted average of
where the y l ' s , that is,
H   pm  T   pm  T  (the QP Hessian matrix )
u


M
and ylw l
 l 1
G ( k  1)   pm  T e( k  1) (theQP gradient vector ) y (k )  (19)

M l
w
l 1
Solution of Equ (17) by a QP algorithm at each l
where the weights w are computed as
sampling interval (k) produces an optimal set of moves n
u(k) which satisfies the constraints. Since H is likely
u
wl    c (x )
i 1
l
i i (20)
to be fixed at all sampling intervals, a parametric QP
algorithm is used to reduce on-line computation time. If an output of a fuzzy model appears as one of its input,
we obtain a dynamic TSK fuzzy system. A dynamic
TSK fuzzy system is constructed from the following flow rate ( f1 ), the flow rate ( f 2 ) and concentration
rules : ( c2 ) of sodium hydroxide and the pH in the reactor.
Such nonlinear relationship suggests that the process
IF y ( k )is A1p and and y ( k  n  1)is Anp and u ( k )is B p operation can be partitioned into several regions based
THEN y ( k  1)  a1p y ( k )  a np y ( k  n  1)  b p u ( k ) on the reactor pH. Hence, we first divided the process
into three operating regions: pH low, pH medium, and
(21) pH high.

where Aip and B p are fuzzy sets, aip and b p are

constants, p=1,2,...,N, u(k) is the input to the system, TRAINING
and x(k )  (x(k ), x(k  1),, x(k  n  1))T is the state To generate training and testing data, random
perturbations are added to u1 , u 2 and u 3 . Five hundred
vector of the system. Hence output of the dynamic
fuzzy system is computed as : data points are generated as shown in Fig. [3]. Training
is terminated when the number of iterations exceeds

N
x p ( k  1) v p 100. After training the neuro-fuzzy system, a typical
p 1
x ( k  1)  (22) identified fuzzy model for three regions are :

N
vp
p 1
p R1 : If pH is low:
where x (k  1) is given in Equ (21) and
y ( t )  0 . 1837 y ( t  1)  0 . 0223 u1 ( t  1) 
n
v  p
  Aip [x(k p
 i  1)] B [u(k )] (23) 0 . 0444 u 2 ( t  1)  0 .0337 u 3 ( t  1)
i 1 R2 : If pH is medium:
Hence, one can express the Least-Squares solution of y ( t )  0 . 201 y ( t  1)  0 . 1912 u1 ( t  1) 
MPC equations as the following quadratic minimization (25)
problem, with linear plant model obtained from the 0 . 2777 u 2 ( t  1)  0 . 2712 u 3 ( t  1)
neuro-fuzzy system : R3 : If pH is high:
y ( t )  0 . 3028 y ( t  1)  0 . 0097 u1 ( t  1) 

1 m

T T m
 1

T T
minJ   p u(k)  e(k 1)   p u(k)  e(k 1)  u(k)  u(k) 0 . 0098 u 2 ( t  1)  0 . 0243 u 3 ( t  1)
u(k ) 2 2
m T T m T 1 m T T 
u(k | k)  [( p )   p   ] ( p )   [r(k 1| k)  y(k 1| k)] MEMBERSHIP FUNCTIONS
(24) The identified final membership functions are plotted in
Fig. [4] to Fig. [7]. It is interesting to observe that the
[6] SIMULATION OF NEUROFUZZY sharp changes of the training data surface around the
origin is accounted for by the moving of the
BASED MODEL PREDICTIVE membership functions toward the origin. Changes of
CONTROL membership shape from initial to final shows that the
neuro-fuzzy system has in fact made a correct
The proposed technique has been successfully applied understanding of how the fuzzy system list the fuzzy
to a multivariable nonlinear system, a pH neutralization. rules.
Results of simulated identification of the nonlinear
dynamic plant using neuro-fuzzy are presented. Hence, MODEL TESTING AND VALIDATION
results of employing neuro-fuzzy model of Fig. [2] in The neuro-fuzzy model is tested by some testing data. It
predictive control systems are discussed at a later stage. can be seen that the neuro-fuzzy model is very accurate,
it has an RMS error of 0.0988 for the training data. The
NEUROFUZZY MODEL OF A NONLINEAR simulated model output is compared with the actual
NEUTRALIZATION pH REACTOR SYSTEM output in Fig. [8].
A pH neutralization reactor is a multivariable nonlinear
process. The process gains differ dramatically at SURFACE PLOT OF FUZZY RULES AND MPC
different pH ranges. The steady state relationship RESPONSE
between acid flow rate ( f1 ) and pH (y) in the reactor
shows a nonlinear relation. It was found that the In order to asset the behavior of the obtained fuzzy
process gain is very high in the medium pH region, models, three dimensional plots of the fuzzy rules are
while it is quite low in both low and high pH regions. viewed as shown in Fig. [8]. In this sense, the predictor
(model) output is obtained via the summation of
A neuro-fuzzy model is developed to model the identified fuzzy regions. Each fuzzy region over the
nonlinear dynamic relationships between the acetic acid entire 3D fuzzy surface represents a linear region. For
instant, in Fig. [8], effect of pH (k-1) and sodium flow
at (k) on the pH (k) is obtained via the designation of
the fuzzy membership functions (i.e. small, medium and
large). A typical rule can then be written as linear
summation as follows :
If acid fl(k-1) is small AND sdm fl(k-1) is large THEN
pH(k) is large.
To validate the constructed fuzzy model output, Fig.
[9] shows a comparison between the actual process
output and model output. From the figure it is clear
how the process has been modeled with minimum error.
The pH Neutralization Plant system being examined by
the neuro-fuzzy is a multivariable pH neutralization
process. In this respect, Fig. [10] shows the associated
MPC simulation for the process, at two required set-
points. As the set-point changes, the process model also
changes via the neuro-fuzzy system. In this MPC
simulation, Equ (25) has been employed to compute the
associated control signals via the utilization of the
linearized process model. This depends entirely on the
region within which the process is operating. Fig.[1]: Neuro-fuzzy employed in MPC system.

[7] CONCLUDING
Within this article we have introduced the basic
structure of a neuro-fuzzy system and how the inter- acid_fl(k-1)
layers are configured to model a pH neutralization
reactor plant. The modeling neuro-fuzzy network has Neuro-fuzzy
shown great results in terms of reducing the prediction System
sdm_fl(k-1)
Root Mean Squared error. Once neuro-fuzzy models
are obtained, we validated the neuro-fuzzy models that
operate over linearized plant region of operation. In this
respect, the neuro-fuzzy models have been employed to sdm_conc(k-1)
model linearized operating regions of the plant under
linear model predictive control. The employed
algorithm produced fast settling time and convergence,
in addition to a good neuro-fuzzy structure for modeling pH(k-1)
the nonlinear behavior of the plant under control.
Fig.[2]: Neuro-fuzzy architecture for pH neutralization.
[8] REFERENCES 10
Input#1 acid flow sequence for training
[1] Zhi Qiao Wu and Chris J. Harris, “Modeling and 0
Adaptive Filtering of Nonlinear System using Neural -10
Network,” University of Southampton, Technical 0 100 200 300 400 500
Report. Input#2 sdm flow sequence for training
10
[2] Zadeh L.A., “Fuzzy sets. Information and Control,”
0
8:338-353, 1965.
-10
[3] J.Zhang, and A.J. Morris, ”Fuzzy Neural Networks 0 100 200 300 400 500
for Nonlinear Systems Modeling,” IEE Proc.-Control Input#3 sdm conc.sequence for training
Theory Appl., Vol. 142, No. 6, November, 1995. 10
[4] Jang J-S.R., “ANFIS: Adaptive-Network-Based 0
Fuzzy Inference System,” IEEE Trans. on System Man -10
0 100 200 300 400 500
and Cybernetics, 23(3):665-685, 1995.
Output pH sequence for training
[5] S. Reay and M. W. Dunnigan, ”Learning Issue in 5
Model Reference Based Fuzzy Control,” IEE Proc.- 0
Control Theory Appl., Vol. 144, No. 6, Nov., 1997. -5
[6] Takagi T. and Sugeno M., “Fuzzy identification of 0 100 200 300 400 500
systems and its application to modeling and control,”
IEEE Trans. on System Man and Cybernetics, 15:116- Fig. [3] : Training sets for pH neutralization system.
132, 1985.
View publication stats

Final membershipfunctions #1“acidfl”

1
0.9
0.8
0.7
0.6 pH(k)
0.5
0.4
0.3
0.2
0.1
0-3 pH(k-1)
-2 -1 0 1 2 3 4
sdm fl
sd
Fig [4] : Final membership functions of the acid f1.
Fig [8] 3-D plot of fuzzy rules, and relation between
pH(k-1), sdmf1 and the process output pH(k).
final membershipfunctions#2“sdmfl”
1 Predicted Output
0.9 2
0.8
0.7 1
0.6
0.5
0
0.4
0.3
0.2 -1
0 50 100 150 200 250
0.1
Actual Output
0
-0.35 -0.3 -0.25 -0.2 -0.15 -0.1 -0.05 0 2

Fig [5] : Final membership functions of the sdm f1 in 1

pH Reactor plant.
0
final membership functions #3 “sdmconc”
1 -1
0.9 0 50 100 150 200 250
0.8
0.7
0.6
0.5 Fig.[9]: Neuro-fuzzy model and actual pH output.
0.4 Plant Output
0.3 3
0.2
0.1 2
0
-0.35 -0.3 -0.25 -0.2 -0.15 -0.1 -0.05 0 pH ___
1
Fig [6] : Final membership functions of the sdm. conc in
pH Reactor plant.
0
0 2 4 6 8 10
final membership functions #4 “pH” Time
1 Manipulated Variables
0.9 2
0.8 u1 ___
0.7 u2 - - -
1
0.6 u3 _ _ _
0.5
0.4 0
0.3
0.2
0.1 -1
0 2 4 6 8 10
0-1 -0.5 0 0.5 1 1.5 Time
Fig. [10] : Neuro-fuzzy model predictive control for the
Fig. [7] : Final membership functions for the pH pH Neutralization Reactor plant. Reactor system with a
Reactor plant. set-point of two.