Sunteți pe pagina 1din 6

G. J. Nicholas Ang, L. Huang and C.

Min Lim, "Artificial Neural Network-Based Controllers for A Continuous Stirred Tank Heater
Process," 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), Singapore, 2018, pp. 1414-1419.

Artificial Neural Network-Based Controllers for A Continuous


Stirred Tank Heater Process
Guang Jun Nicholas Ang, Lingzhang Huang, and Choo Min Lim

Abstract— Two linear Artificial Neural Network (ANN)- the CSTH is to mix cold water and hot steam together, with a
based controllers for the control of a nonlinear system are constant inflow of hot water, to achieve a user-defined
presented. The optimization and pole placement control design reference temperature and reference level. With the
strategies are used. For both strategies, linear ANNs are used incorporation of constraints and nonlinearities, the CSTH
to model a Continuous Stirred Tank Heater (CSTH) process. model of [2] is an appropriate and realistic process for testing
Third-order discrete-time models or Adaline ANNs are used of controllers.
whose parameters are updated at every sampling time to
provide adaptability to the controllers. Simulation tests show B. Decoupling Technique
that the linear ANN-based controllers can overcome strong The CSTH pilot plant model is a Two-Input Two-Output
nonlinear coupling effects, reject step disturbances, and (TITO) multivariable control system with strong interactions.
provide adequate damping to the subsystem interactions. Both Decoupling design techniques have been used to simplify the
ANN-based control design strategies are shown to be control of TITO systems. One of such techniques is based on
interchangeable with each other for the control of the CSTH
using linear ANN models whose parameters are estimated
process.
and used in designing the outputs of two controllers. The
I. INTRODUCTION overall system has been shown to be adaptable to changes
and reject disturbances from the interactions within the
The Proportional Integral Derivative (PID) Controller is system [3].
commonly used in industrial control systems. Although a PID
controller provides optimized control for linear systems with C. Proposed Control Scheme Configuration
almost fixed parameters, it may not be well-suited for
nonlinear systems. A PID controller has been observed to
exhibit low robustness when the controlled system has tight
couplings, multiple nonlinearities, hard constraints, and
disturbances [1–10].
Other types of controllers have been proposed to
overcome the above-mentioned problems. This paper
presents another approach in which two Adaptive Linear
Element (Adaline) Artificial Neural Network (ANN)-based
controllers are applied. The optimization of system variables
and pole placement control design strategies are chosen for
the study. Furthermore, a nonlinear Continuous Stirred Tank
Heater (CSTH) process is used to ensure that the coupling
effects, multiple nonlinearities, and hard constraints are Fig. 1. Proposed CSTH TITO Control Scheme Configuration
considered in the proposed studies [1, 2]. Comparison tests w1(k) : desired water level (mA)
between the two approaches are performed and the w2(k) : desired water temperature (mA)
effectiveness of both control design techniques are u1(k) : flow input (mA)
investigated using nonlinear simulation studies in the u2(k) : steam valve (mA)
y1(k) : actual water level (mA)
Simulink environment. The interchangeability of both ANN- y2(k) : actual temperature (mA)
based controllers and some guidelines for exploiting them for CSTH: see [1]
effective control are also presented.
II. CSTH PROCESS Figure 1 shows the TITO control configuration proposed
for study. Signals (w1,u1,y1) and (w2,u2,y2) are passed into the
A. Introduction Level Controller (LC) and Temperature Controller (TC),
The CSTH is an experimental rig developed by the respectively. The estimated model parameters (ai and bi) are
Department of Chemical and Materials Engineering at the updated at every sample in each ANN block based on the
University of Alberta [1]. Experimental data of the CSTH measured input data u(k) and output data y(k), as defined in
pilot plant is modelled in the Simulink environment which is (1). These parameters are then passed on to the respective
available on the CSTH website [2]. The control objective of controllers to compute the control signals. The control
signal u1(k) is applied to the Flow Controller (FC) to control
Guang Jun Nicholas Ang*, Lingzhang Huang, and Choo Min Lim are the flow rate in order to achieve the desired water level.
with the Ngee Ann Polytechnic, School of Engineering, 535 Clementi Rd,
Singapore 599489
(*corresponding author e-mail: anggjnicholas@gmail.com).

DOI: 10.1109/ICARCV.2018.8581301. Available: https://ieeexplore.ieee.org/document/8581301 ©2018 IEEE


G. J. Nicholas Ang, L. Huang and C. Min Lim, "Artificial Neural Network-Based Controllers for A Continuous Stirred Tank Heater
Process," 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), Singapore, 2018, pp. 1414-1419.

III. OPTIMIZATION CONTROL DESIGN Remark 2: The scaling factor r ensures that the energy
required is minimized and within reasonable limits [4].
A. Problem Formulation
As the CSTH TITO system comprises two subsystems, Remark 3: The scaling factor q controls the rate of change of
the linear dynamics of each subsystem can be adequately output response, p controls the gradient of the rate of change
described by a third-order discrete-time model of the form of output response, and r controls the input energy. These
scaling factors were proposed earlier in [3, 4] to provide
3 3 additional parameters for tuning. Moreover, these scaling
y (k )   ai y (k  i )   bi u (k  i ) (1) factors allow the characteristics of the subsystems and their
i 1 i 1
interactions to be analyzed and studied.
where y(k) and u(k) are the system output and the control Remark 4: Following the concepts and motivated by the
signal, respectively [3, 4]. In (1), (ai and bi) are the estimated experimental successes of [3–7], the proposed control design
discrete-time output and input model parameters, is expected to enhance the stability of the overall system due
respectively. This model is an Adaline ANN [5]. to the additional scaling factors introduced in (2).

B. Optimization Control Scheme C. Overview


The performance index leading to a one-step ahead
prediction control of each subsystem is defined as follows:

J   y (k  1)  w(k )   q  y (k  1) 
2 2

(2)
 p  y (k  1)  r u (k )  u (k  1) 
2 2

where q' and p' are the gains of the first derivative and
second derivative of y(·), respectively [3]. Equation 2 is
further simplified with the following approximations:

y (k  1)   y (k  1)  y (k )  / To
(3)
y (k  1)   y (k  1)  2 y (k )  y (k  1)  / To
Fig. 2. Test System 1: Diagram of the Proposed TITO Control Scheme
with the optimization control design strategy [1–3].
to yield
J   y (k  1)  w(k )   q  y (k  1)  y (k ) 
2 2
Figure 2 shows Test System 1 in which the optimization
control design strategy is applied to the CSTH process. LC
 p  y (k  1)  2 y (k )  y (k  1) 
2
(4) and TC are ANN-based controllers, while FC is a PID
 r u (k )  u (k  1) 
2
controller. Standard Operating Point 2 is chosen as the
operating point for CSTH process where HW valve (mA) is
where q = q'/To2 and p = p'/To2. By minimizing (4), the 5.5. The Recursive Least Squares Method with adaptive
optimal control signal for each subsystem is of the form directional forgetting is chosen as the on-line identification
method for the ANN [11]. The sampling time, TO, chosen for
each controller is 0.1 s.
 y (k )(q  2 p)  (1  q  p) s(k ) 
u (k )  K1  
  w(k )  py (k  1)
D. Simulation Tests
 (5)
 K 2 u (k  1) 
Step changes are applied to the reference level and
reference temperature of the system with periods 355 s and
where 655 s, respectively.
1
K1  b1 b12 (1  q  p )  r 
1
(6)
K 2  r b12 (1  q  p )  r 
and
3 3
s (k )   ai y (k  i  1)   bi u (k  i  1). (7)
i 1 i2
Remark 1: The scaling factor of the first derivative of y(·), q,
places emphasis on the gradient of the output response, thus
enhancing the damping characteristic of the system. The
scaling factor of the second derivative of y(·), p, places
emphasis on the gradient of the rate of change of output
signal. As such, the dependency on the scaling factor q is
reduced, allowing easier tuning of the controller [3].

DOI: 10.1109/ICARCV.2018.8581301. Available: https://ieeexplore.ieee.org/document/8581301 ©2018 IEEE


G. J. Nicholas Ang, L. Huang and C. Min Lim, "Artificial Neural Network-Based Controllers for A Continuous Stirred Tank Heater
Process," 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), Singapore, 2018, pp. 1414-1419.

Fig. 5. The estimated Model Parameters of the Level ANN and the
Temperature ANN when both step changes are applied, and for scaling
factors q1=25, r1=0.03, p1=25, q2=25, r2=1.0 and p2=26. Parameter a1 of
both sets of updating parameters saturated at around -1.

By comparing Figs. 3b–3d, it can be seen that the


temperature output is easily affected by the level step
changes. Short rise time and fall time are observed in the
level output response. Similarly, the estimated model
parameters for the temperature loop take a longer time to
settle down as compared to level loop. This is expected as
the level loop has a shorter time constant.
E. Result Analysis
Fig. 3. The Input-Output Responses of Test System 1 for scaling factors
q1=25, r1=0.03, p1=25, q2=25, r2=1.0 and p2=26. (a) without step changes, It can be seen from Figs. 3 – 5 simulation test results that:
(b) temperature-step disturbance, (c) level-step disturbance, (d) 1) The ANN-based optimization control design strategy is
temperature-step and level-step disturbances. able to stabilize the nonlinear CSTH TITO process. Figure 5
shows the model parameters changing and converging
rapidly after the application of step changes.
2) An oscillatory response to step change disturbances is
observed in the level output response. However, the use of
constantly updating ANN parameters at every sampling time
ensures that the oscillatory output signal is rapidly damped,
see Fig. 5a. As a result, quick adaptability and good damping
are observed in the level output of Figs. 3c and 3d.
3) The PID controller chosen as the FC in Test System 1
is observed to overshoot the reference signal u1 due to the
sudden and large change in flow reference magnitude,
resulting in oscillations, see Fig. 4. The PID controller is
unable to effectively control the flow rate because of the
fast-running ANN and control design technique that uses a
Fig. 4. The Flow Input-Output Response with the PID controller as the short sampling time for model parameters estimation.
FC with tuning parameters kp = 2.5; ki = 2.5; kd = 4 for Test System 1.

DOI: 10.1109/ICARCV.2018.8581301. Available: https://ieeexplore.ieee.org/document/8581301 ©2018 IEEE


G. J. Nicholas Ang, L. Huang and C. Min Lim, "Artificial Neural Network-Based Controllers for A Continuous Stirred Tank Heater
Process," 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), Singapore, 2018, pp. 1414-1419.

4) The larger value of r2 decreases the proportion of the B. Simulation Tests


error, the rate of change of output and the gradient of the rate
of change of output in (2). This hinders the ability of the
temperature subsystem from rejecting the disturbances of the
level subsystem interaction, see Figs. 3c and 3d. The
temperature output requires extra time to settle due to the
nonlinearities and unknown quantities such as heat transfer
through the heating coils reported in [1].
5) Greater fluctuations of model parameters are observed
when the weightage of tuning parameter r decreases, see Fig.
5. As r2 > r1, the larger value of r2 limits the energy of the
control signal, thereby reducing fluctuations of model
parameters.
6) When the scaling factor of ẏ(·) or ÿ(·) is removed, the
system becomes unstable (results not shown). Hence, the
anticipatory properties of ẏ(·) and ÿ(·) are necessary in
ensuring overall system damping and stability. It should be
noted that ẏ(·) and ÿ(·) need not be measured due to the use
of the approximations given in (3). Therefore, the number of
hardware is reduced.
IV. POLE PLACEMENT CONTROL DESIGN
A. Overview
The pole placement controller with on-line identification
has been shown to be effective for the control of a nonlinear
servo system and a semi-batch reactor process [12–14]. In
this test system, the pole placement control design is
implemented into the proposed study. Two ANN-based third-
order pole placement two-degrees-of-freedom (2DOF)
controllers are used for Test System 2. The formulation and
design algorithm of the pole placement controllers are
presented in [15]. Equation (1) is chosen as the discrete-time
model and for on-line identification of the ANN.
Figure 6 shows Test System 2 in which the pole placement
control design strategy is applied to the CSTH process. LC
and TC are ANN-based pole placement controllers, while FC
is a PID controller. The same on-line identification method of
the ANN and tuning parameters of the PID in Test System 1
are used in Test System 2 for fair comparison.

Fig. 6. Test System 2: Diagram of the Proposed TITO Control Scheme


with the pole placement control design strategy [13]. Fig. 7. The Input-Output Responses of Test System 2 in which ANN-
T01,2 = 0.5, 0.7 based pole placement controllers are applied. (a) without step disturbances,
ζ1,2 = 2.0, 3.5 (b) temperature-step disturbance, (c) level-step disturbance, (d)
ω1,2 = 0.1, 0.2 temperature-step and level-step disturbances.

DOI: 10.1109/ICARCV.2018.8581301. Available: https://ieeexplore.ieee.org/document/8581301 ©2018 IEEE


G. J. Nicholas Ang, L. Huang and C. Min Lim, "Artificial Neural Network-Based Controllers for A Continuous Stirred Tank Heater
Process," 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), Singapore, 2018, pp. 1414-1419.

Fig. 10. The Input-Output Responses of the Level control loop when
Fig. 8. The Flow Input-Output Responses with the PID controller as the
different ζ and ω are used.
FC with parameters kp = 2.5; ki = 2.5; kd = 4 for Test System 2.
y1,1 = when ζ = 2 and ω = 0.1;
y1,2 = when ζ = 2 and ω = 0.2;
y(1,3)/6 = when output/6, ζ = 2 and ω = 0.3;
y1,4 = when ζ = 1 and ω = 0.3;

C. Result Analysis
It can be seen from Figs. 7–10 simulation test results that:
1) The ANN-based pole placement control design
technique can stabilize and provide sufficient damping to the
CSTH TITO system. Figure 9 shows the model parameters
changing to overcome the step change disturbances,
indicating that the ANN adapts to damp the output response.
2) Due to the high damping factor required to overcome
the nonlinearity of the CSTH process, long rise time and fall
time are observed.
3) Different sampling times for the level and temperature
ANN-based controllers are chosen. A longer sampling time
and smaller natural frequency ω allow the PID FC to follow
the flow input, u1, thereby improving the level output
response.
4) The pole placement control design response is bound to
the dynamic behavior of the second-order continuous
characteristic polynomial of s2 + 2ζωs + ω2 [12, 13]. This is
to be expected as the design objective is to achieve pole
placement through specifying ζ and ω.

Fig. 9. The estimated Model Parameters of the Level and Temperature V. COMPARISON AND DISCUSSION
ANNs of Test System 2 when both step change disturbances are applied.
Parameter a1 of both sets of updating parameters saturated at around -1.
A longer time is taken for the model parameters to update
in the pole placement control design strategy than in the
optimization control design strategy, due to a longer
High damping factors of ζ > 1 are required to adequately
sampling time, higher ζ, and a smaller ω, compare Figs. 5a
damp the nonlinear system, thereby reducing overshoot and
and 9a. However, the pole placement control design strategy
slowing output response. For the output responses of Fig. 7d provides better damping than the optimization control design
to be achieved, ζ1 ≥ 2 and ζ2 ≥ 3.5. However, increasing ζ strategy. As a result, the rise and fall time of the output
results in severe overdamping and ω must be increased to response are observed to be longer. Since the temperature
quicken the rise and fall time of the output response. output has little or no influence on the level output, a
sufficiently damped level output response will result in a
Increasing ω results in a faster rise and fall time but an better temperature output response.
increasingly oscillatory response. It can be observed from
Fig. 10 that ζ has to be decreased, inevitably increasing The pole placement control design strategy is bounded to a
second-order characteristic response. On the other hand, the
overshoots, albeit a necessity in preventing instability in the
optimization control design strategy is not bounded to any
system as ω increases.
closed-loop polynomial characteristic response.

DOI: 10.1109/ICARCV.2018.8581301. Available: https://ieeexplore.ieee.org/document/8581301 ©2018 IEEE


G. J. Nicholas Ang, L. Huang and C. Min Lim, "Artificial Neural Network-Based Controllers for A Continuous Stirred Tank Heater
Process," 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), Singapore, 2018, pp. 1414-1419.

The optimization control design strategy provides more designers with the flexibility to exploit the ANNs for
user-defined terms and parameters for tuning than the pole effective control.
placement control design strategy, where ζ and ω are the
fundamental tuning parameters. The optimization control The presented study is an extension to our previous work
design strategy controls and minimizes the errors, the rate of [3], where the authors proposed a semi-decoupling control
change of outputs, the gradient of the rate of change of technique and an enhanced self-tuning control scheme that is
outputs, and the control signals. Unlike the pole placement computationally less intensive than other advanced methods.
control design strategy, observations are made easier because In view of the above results, the studies reported in both
of the visible and apparent correlation between the input and papers will be developed into teaching resources for
output responses, and the updating parameters, compare the educational purposes.
level output in Figs. 3c or 3d with Figs. 4 and 5a. This allows
designers to have more control over a process or study an ACKNOWLEDGMENT
unknown process and its characteristics better, as reported in The authors are grateful to Dr. Chalupa, for his Self-
[3]. However, the optimization control design technique tuning Controllers Simulink Library of the pp3c2dof used
requires an initial set of values in the tuning parameters for for simulation tests on the pole placement control design of
tuning to begin. As the r parameter limits the energy of the the CSTH process.
control signal, the r parameter is the most significant
parameters in each performance index. Hence, it is advisable
to set the desired r scaling factors first before tuning the other REFERENCES
parameters. Setting the wrong r, q, and p parameters will lead [1] N. F. Thornhill, S.C. Patwardhan, and S. L. Shah, “A continuous
to an unstable system. stirred tank heater simulation model with applications,” Journal of
Process Control 18, pp. 347-360, 2008.
Nonetheless, both control design strategies can be used [2] N. F. Thornhill, S.C. Patwardhan, and S. L. Shah, THE STIRRED
interchangeably. If fast rise and fall time are required for the TANK HEATER SIMULATION. [online] Available at: http://personal-
control of the level response, notwithstanding the initial pages.ps.ic.ac.uk/~nina/CSTHSimulation/index.htm [Accessed 27
April 2018].
oscillatory response, designers can choose to alternate [3] Guang Jun Nicholas Ang, Wei Ze Lim, Choo Min Lim, "A Proposed
between the two control design strategies. The results will not Two-input Two-output Self-tuning Control Scheme", 2018 IEEE
be significantly affected because both control designs use the Symposium on Computer Applications & Industrial Electronics
same third-order discrete-time model or Adaline ANN with (ISCAIE), pp. 102-107, 2018.
the decoupling technique applied, see Fig. 11. [4] C. M. Lim, and T. Hiyama, “Self-tuning control scheme for stability
enhancement of multimachine power systems”, IEE PROCEDINGS,
137 (4), pp. 269-275, July 1990.
[5] D. H. Nguyen, and B. Widrow, “Neural Networks for Self-Learning
Control Systems”, IEEE Contr. Syst. Mag., vol. 10, no. 3, 1990.
[6] Ratneshwar Jha and Chengli He, “Neural-network-based adaptive
predictive control for vibration supression of smart structures”, Smart
Materials and Structures, Vol. 11, pp. 909-916, 2002.
[7] Choo Min Lim and Swee Siang Quck, “Experimental Results on Self-
Tuning Control Scheme”, IEEE Transactions on Energy Conversion,
Vol. 10, No. 2, pp. 210-217, June 1995.
[8] C. M. Lim, and Li Qing, “An Enhanced Adaptive Neural Network
Control Scheme for Power Systems”, IEEE Transactions on Energy
Conversion, Vol. 12, No. 2, June 1997.
[9] John G. Kuschewski, Stefen Hui, and Stanislaw H. Zak, “Application
of Feedforward Neural Networks to Dynamical System Identification
and Control”, IEEE TRANSACTIONS ON CONTROL SYSTEMS
TECHNOLOGY, VOL. 1, NO. 1, MARCH 1993.
Fig. 11. The Input-Output Responses of the ANN-based optimization
[10] Seul Jung, Hyun-Taek Cho, and T.C. Hsia, “Neural Network Control
controller for the level output and the ANN-based pole placement controller
for the temperature output. for Position Tracking of a Two-Axis Inverted Pendulum System
Experimental Studies”, IEEE TRANSACTIONS ON NEURAL
NETWORKS, VOL. 18, NO. 4, JULY 2007.
VI. CONCLUSION [11] P. Chalupa, “Predictive Control Using Self-tuning Model Predictive
Controllers Library”, 17th International Conference on Process
The proposed control design schemes have been applied to Control 2009, pp. 419-425, 2009.
enhance the performance of the nonlinear CSTH plant. [12] D. Novosad, L. Macku, “Pole placement controller with compensator
Simulation results have shown that: adapted to semi-batch reactor process”, 18th International Conference
on Process Control, pp. 383-387, June 2011.
(i) The controllers can perform effectively in the [13] Bobál, V., Böhm, J., Fessl, J., Machácek, J., “Digital Self-tuning
presence of both level and temperature disturbances. Controllers: Algorithms, Implementation and Applications,” Springer-
Furthermore, the controllers provide sufficient damping to Verlag London, May 2005.
the nonlinear CSTH plant. [14] P. Chalupa, V. Bobal, J. Novak, and P. Dostal, “Using of Self-tuning
(ii) Both controllers can be used interchangeably without Controllers Simulink Library for Real-Time Control of Nonlinear
Servo System”, Proceedings 23rd European Conference on Modelling
degrading the system performance. and Simulation, ISBN: 978-0-9553018-8-9.
(iii) The scaling factors of performance indices and the [15] H. Badihi, Y. Zhang, H. Hong, “Design of a Pole Placement Active
tuning parameters of the pole placement scheme provide Power Control System for Supporting Grid Frequency Regulation and
Fault Tolerance in Wind Farms”, 19th The International Federation of
Automatic Control World Congress, pp. 4328-4333, August 2014.

DOI: 10.1109/ICARCV.2018.8581301. Available: https://ieeexplore.ieee.org/document/8581301 ©2018 IEEE

S-ar putea să vă placă și