You are on page 1of 8

2012 International Conference on Computing, Electronics and Electrical Technologies [ICCEET]

ANN Based MARC Controller Design for an Industrial Chemical Process

G.Glan Devadhas Dr.S.Pushpakumar D.M.Mary Synthia Regis Prabha

Associate professor Principal Associate Professor
Dept of EIE Heera College of Engineering Dept of EEE
N.I. Centre for Higher Education N. I. Centre for Higher Education

This work aims to develop an Artificial Neural PID control strategy and the advanced Model
Network based Model Reference Adaptive Control Reference Adaptive control strategy. The
strategy for an Industrial Process. The inherent ineffectiveness of both these strategies to certain
advantage of the Neural Network to adapt in real application specific areas is presented. The
time together with the effectiveness of Model concentration and temperature control of the process
Reference Adaptive control is utilized in the control is done by using the model reference adaptive system
of Industrial Chemical process. which gives more accurate response by minimizing
Key words: CSTR process, MARC, ANN, Adaptive the error than using the generalized PID controller.
An intelligent system [13], [8] is a system that is able II.PROCESS DESCRIPTION
to make decisions that would be regarded as The schematic sketch of the Continuous­
intelligent if were done by humans. In dealing with Stirred-Tank Reactor (CSTR)[6,7] is shown in
highly sophisticated process control, where much figure.l
stability and accuracy are expected probably
perturbations may interface and disturb the output
variable. This disturbance is caused as the
perturbations are getting fed back through the error
detector to the controller. Intelligent controllers
adapt them they correct decisions automatically for
future .Neural network is such an intelligent system.
A neural network based model reference adaptive
control [17] is proposed to resolve the problem of
concentration control in CSTR process. This
proposed method can adjust the controller parameters
in response to changes in plant and disturbance in
real time by referring to the reference model that �
Product. out
specifies properties [15] of the desired control Figure 1 Schematic Sketch of the Continuous Stirred
system. Hence the optimization of the controller is Tank Reactor (CSTR)
significant. It enhances stability over parameters and The reactor vessel volume is V. An agitator
makes the system ultimately immune to is used to make the reactor contents be well mixed.
perturbations. The resulting system [14]can be The inlet reactant with feed rate F, concentration Cin
viewed as a sort of adaptable control methodology. and temperature Tm is fed from the top of the vessel.
The system is optimized [1], [9], [10] using the The product stream is gathered at the exit. If the
neural supervisory control. The set of input and reactor contents are perfectly mixed, then the
output are predetermined using the model reference compositions are uniform everywhere in the reactor
adaptive system. The set of inputs and outputs thus and the product stream is at the same concentration
obtained are used for training the neural network .As as the mixture in the reactor. In other words, the
a solution to the mentioned problems, a control composition concentration CA and temperature T at
strategy employing Model Reference Adaptive which reaction takes place are the same as the
control with a peripheral Neural Network Interface is composition concentration and temperature of any
proposed exit stream. Therefore, progress of reaction can be
This paper presents here a comparative study of monitored by observing the product stream
the performance of the process with the convention concentration CA and the reactor temperature T. The

978-1-4673-0210-4112/$31.00 ©20l2 IEEE 375

2012 International Conference on Computing, Electronics and Electrical Technologies [ICCEET]

exothermic chemical reaction taking place in a CSTR Fig 2 General Block Diagram of MRAController
produces a high heat of reaction. To maintain the The block diagram in fIg 2 shows the structure of a
reactor temperature at its proper value, the jacket is MRAC system that composed of a plant containing
cooled or heated by the coolant with temperature Te. the unknown parameters, a reference model, and a
It is assumed that the product concentration CA will feedback control law containing adjustable
settle down to the desired value if the reactor parameters.
temperature T is controlled properly. Therefore, the Adaptive controller [10], [11] involves modifying the
main objective in this research work is to properly control law used by a controller to cope with the fact
adjust the temperature of coolant To of the CSTR that the parameter [5] of the system being controlled
control system such that the reactor temperature T are slowly time varying or uncertain. The adaptive
can follow the desired values even in the case of control does not need prior information about the
disturbances in inlet reactant temperature Tin. bounds on these uncertain or time- varying
The component balance for the reactor can parameters.
be given as, The parameters are adjusted by using the MIT rule.

dCA Fi ( ) k -E/RT A.MIT RULE


dx - V CAi - CA - CA
The energy balance by, A system which has one adjustable parameter
Fp Cp (Tin - T) + Vk e-E/RT cA.dH + Q
pV Cp '!:!:. (2)
.'A: is considered .The closed loop response is

Most over damped processes can be sufficiently well

specifIed by a model whose output is Ym. The error is
approximated by a fIrst order system with a time lag 'e' between the output y of the closed loop system
and output Ym of the reference model. The variable
element as follows.
control parameter A is adapted in such a way the
G(S) = K e_LS I1+Ts
' )

(3) cost function j(A) is

Where K, T, L denotes the system gain, the time J(A) = � e2 is minimized .

constant and the time lag respectively. The MIT rule is:
dA D] De
= (5)
The controlled object is given by the following dt Y DA = ye DA

G(s) = 0.5 e·45s /1+100s (4) PROCESS IN SIMULINK

The discrete -time model corresponding to the above

The fIgure 3 shows the model reference adaptive
equation is considered where the sampling interval is
controller with reference to our application.
Ts=1O.0[s], and the system is disturbed by Gaussian
white noise with mean and variance 0.001.

The Model Reference Adaptive Control System

(MRAC) [16] is an adaptive servo system in which
the desired performance is expressed in terms of
reference model, which gives desired response to the
reference signal.

Fig 3. Schematic of MRAC Reference to the

Process in Simulink
The process is assumed to have a know structure,
although the parameters are unknown. The process
transfer function is given as:
G(s) = 5:5 (6)

2012 International Conference on Computing, Electronics and Electrical Technologies [ICCEET]

The model plant is referred by the transfer function

as: Gm(s) = 5+0.5
The controller is used to adjust the parameters. They
exists different set of controller parameters for which
the controller task can be achieved. The control law
is usually linear in terms of the adjustable

The adaptation unit as shown in fig 4 adjust the

parameters of the control law. In MRAC systems the
adaptation law searches for the parameters such that
the response of the plant under adaptive control Fig 5 Display unit of MRAC
becomes the same as that of the reference model. The The plant response and the model response are
adaptation mechanism is designed to guarantee the depicted in display 1 and the plant response, model
stability of the system and convergence of the response along with the error tracking is shown in
tracking error to zero. the display unit 2.Fig 5 shows the simulink block of
display unit


Neural network [18] is a kind of intelligent
----1l1>i y U control methodology. This methodology is strictly
based on the heuristic type algorithm. Hence neural
----1l1oly network have the capabilities to adapt themselves to
new environments and conditions. Like wise to
----1l>! y
make the system quite adaptable to new conditions,
familiarizing the network to specific standard
ADAPTAT I ON UN I T conditions is necessary which is usually termed as
Fig 4.Block of Adaptation Unit training .This training ensures appreciable responses
D.DISPLAY UNIT of the system to non-familiar situations, noise with
different characteristics, in this case.
In the display unit we have the plant response,
model response and the error. Initially the error can
be found out .Once the system gets adjusted by the
adaptable mechanism the error get tracked and

--- -.J IEirror


Fig 6 Schematic of ANN MARC

The appropriate neural network is supposed to
simple one, ideal for non-deterministic data and more
reliable with an error feedback. The choices include
feed forward back propagation neural network,

2012 International Conference on Computing, Electronics and Electrical Technologies [ICCEET]

Hopfield neural network, SOM neural network etc.

but the neural network with back propagation
dominates over other s and is best suited for the
specific requirements discussed. The algorithm
meant to make the system adaptable to new
conditions is usually named as training algorithm and
which is preceding the discussion. 1 '

Fig 8 Response of Model Reference Adaptive

The response of the model of MRAC is depicted in
Fig.9.With the help ofcontroller the system get adapt
to the model as the time varies

Fig 7 Schematic of NN with noise factor Reference

to the Process in Simulink

The model reference adaptive system

was originally proposed to solve a problem in which
Fig 9 Model Response of Model Reference
the performance specifications are given in terms of
Adaptive Controller
reference model .This model tells how the output
The response of Plant of MRAC is depicted in
ideally should respond to the command signal.
Fig.10. Initially the plant response and model
The model reference adaptive system is
response are not identical as there may be error.
used to control the level of crude oil in the tank. The
level of crude oil the tank is to maintain in a
particular level to optimize the refming process. By
adjusting the gain the error is minimized.
The response of MRAC is given in
Fig8.Intially there is a large error between the plant
output and the model output. As the time moves on
the system get adapt to the reference model and the

error decreases.

Fig 10 Plant Response of Model Reference

Adaptive Controller
The error variation of the MRAC is shown in Fig
.11.1ntially there is large error . Later when the
system automatically adapts to the reference model.
The error reduces.

2012 International Conference on Computing, Electronics and Electrical Technologies [ICCEET]

. 02
, 0

••• - -+-- �e-- - -+-- �f__- _+_- ___!
.< .

Fig 13 Neural Network Trained to 50 Samples

Fig 11 Error Response of Model Reference Adaptive

The controller output of the Model Reference
Adaptive Controller is depicted in Fig.12

Fig 14 Neural Network Trained to 500 Samples

Fig 12 Controller Response of Model Reference

Adaptive Controller "'�.--+---+--�-�----

The MRAC fails to track the noise or external Fig 15 Neural Network Trained to 5000Samples
disturbances that may affect the process on the later
stages as well during real time applications. To V.ANALYSIS OF NEURAL NETWORK
minimize those errors and to have better control over The Neural Network identified for the proposed
the system we introduce artificial neural network control strategy is Feed Forward with back
based on model reference adaptive controller. propagation. The training of the Neural Network has
been done with a sufficiently large dataset which is
necessary for ensuring the effectiveness of the
NEURAL NETWORK Fig.16 shows the response of neural network
scheme adapted with gain variation. Initially large
To make the system immune to possible noise amount of noise will be incorporated to the system as
conditions, familiarizing the system to specific the neural system start to initialize its tracks all the
standard conditions is worthy. The reliability, error, it can even track and reduce all the error that
accuracy, finally the entire performance itself may occur during the whole process.
improves when the number of standard noise
samples known to the system increases. When only
50 samples the error is reduced. Again when the
neural network is trained with 500 samples the error
reduces. The above points are brought out through a
comparative study among the responses shown in fig
13,14 and 15

2012 International Conference on Computing, Electronics and Electrical Technologies [ICCEET]

10 1 .1 I
1 '1 I

� - - - - - - - - - =

Fig 16 Response of Neural Network Fig 19 Error Response of Neural Network

The model response of the neural network is
The neural based system is now
analyzed with output variation. Large external
disturbances are given to the system and are tested.
Fig 20 depicts the response of neural network
with output variations

Fig 17 Model Response of Neural Network

The plant response of the neural network is depicted

in fig 18. Initially the system contain large
Fig 20 Response of Neural Network with Output
Fig 21 depicts the response of model of neural
network with output variations.

Fig 18. Plant Response of Neural Network Fig 21 Model Response of Neural Network with
The error response of the neural network is Output Variation
depicted in fig 19. Initially the system contains large
disturbances as there is variation between the plant Fig 22 depicts the response of plant of neural
and model. The systems automatically get adjusted network with output variations
and the errors get minimized.

2012 International Conference on Computing, Electronics and Electrical Technologies [ICCEET]

the compensator as well the command to the system

so as to result in an optimized output. The demerit of
PID and MRAC for the CSTR process is analyzed.
The Artificial Neural Network has been applied to
the task of identification of nonlinear systems. The
validity of the scheme is validated through
simulation. The Neural Network algorithm used in
this work is feed forward with back propagation. The
strategy has been simulated using SIMULINK for a
CSTR process modeled as First Order Lag Plus
Delay Process.
Fig 22 Plant Response of Neural Network with REFERENCES
Output Variation 1. Athans. M. and P.L. Falb, "Optimal Control: An
Fig 23 depicts the error response of neural network introduction to the Theory and its Applications",
with output variations. McGraw Hill, New York, 1966
The system gets automatically tracked and minimizes 2. B.W. Bequette, "Nonlinear control of chemical
all the error which may occur during the process. process: a review", Indi.Eng.Chem.Res,
3. Chyi-Tsong Chen and Shih-Tein Peng,
"Intelligent process control using neural fuzzy
techniques", Journal of process control, Vo1.9,
1999, pp.493-503.
� 4. D.E.Seborg, "A perspective on advanced
i ot-��I�.I��+-
� -+����
strategies for process control, Modeling,
1 '1 I
Identification and control", Vo1.15.1994, pp179-
5. E.P. Nahas, M.A. Henson and D.E. Seborg,
"Nonlinear Internal Model Control Strategy for
�:C-o ----,;
O'" ;;--c;;""; �";;;-
" --;:"
--- ";;--c;;o;;-
o '---c""=---;""C;---='" "c--'"'=-�'" "
Neural Network Models", Computers Chemical
Fig 23 Error Response of Neural Network with Engineering, Vo1.16, No.12, pp.l039-1057
Output Variation 6. G.Glandevadhas, Dr.S.Pushpakumar, Robust
The strength of proposed control strategy is in its Temperature controller design for a chemical
ability to automatically adjust parameters so as to process International Journal of Engineering
obtain an optimum response for the process under Science and Technology. Vol. 2(10), 2010,
consideration. The Neural Network identified for the 5831-5837.
proposed control strategy is Feed Forward with back 7. G.Glandevadhas, Dr.S.Pushpakumar Intelligent
propagation. The ineffectiveness of both convention controller design for a chemical process
PID control strategy and the advanced Model International Journal of Engineering Vol. (5),
Reference Adaptive control strategy these strategies 2010, 399-410.
to certain application specific areas is presented. As a 8. G.Stephanopoulos and C.Han, "Intelligent
solution to the mentioned problems, a control systems in process engineering: a review",
strategy employing Model Reference Adaptive Computers Chern. Engg. Vo1.20.l996, pp.743-
control with a peripheral Neural Network Interface is 791.
proposed. The effectives of the proposed control 9. Kirk, D.E., "Optimal Control Theory: An
scheme has been has been verified for different introduction", Prentice Hall, Englewood cliffs,
processes with varying disturbances and N.J., 1970
perturbations. The results strongly validate the 10. Karl J. Astrom and Bjorn Wittenmark,
control strategy. The simulation results are also "Adaptive Control" Pearson Education Press,
presented. 2001
VI. CONCLUSION 11. P.Lakshmi, M.Srinivasan, Winner model based
This work presents an advanced control strategy PID controller for CSTR, International journal
which brings out a novel method for Neural Network of Power system and power electronics
based Supervisory control for automatically adapting Vo1.2.2009 pp7-12

2012 International Conference on Computing, Electronics and Electrical Technologies [ICCEET]

12. Morari and ZafIriou, "Robust Process Control",

Prentice Hall, Englewood Cliffs, New Jersey,
13. G.Glan Deva Dhas, Dr.Pushpakumar
"OptimwnSetting of PID controller for a
chemical Process , American Journal of Applied
Sciences VoI.8(11)1140-1148
14. .Aiping Wang, Hong Wang and Jinhui Wu,
'Modeling Unknown Nonlinear Systems
DefIned on a Unbounded Set via Neural
networks', IEEE, Proceedings of the 15
International Symposium on Intelligent Control,
15. Bradley T. Burchett and Richard A. Layton, 'An
Under Graduate System IdentifIcation
Experiment', American Control Conference,
16. Boonsrirnuang P.,Numsomran A.and
Kangwanrat S. "Design of PI Controller
Using MRAC Technique For Couple-Tanks
Process".World Academy of
Science,Engineering and Technology 59 2009
17. Patuio H.D and Liv D.(2000):Neural Network
Based Model Reference Adaptive Control
Systems-IEEE transaction system Man
Cybern.VoI30,Nol pp 198-204
18. Zhao Yingkai, Lin Jinguo, shu Zhibing,(1997) "
The Design of Neural Network Controller for
Nonlinear Plants," IEEE Int. Conf. on Intelligent
Proc. Sys., Beijing, pp.525-527.