Control of a Heat Exchanger Using Takagi-Sugeno

Fuzzy Model
Anna Vasiþkaninová Monika Bakošová
Slovak University of Technology in Bratislava Slovak University of Technology in Bratislava
Faculty of Chemical and Food Technology Faculty of Chemical and Food Technology
Radlinského 9, 812 37 Bratislava, Slovakia Radlinského 9, 812 37 Bratislava, Slovakia
[email protected] [email protected]

Abstract—Possibility to use the Takagi-Sugeno fuzzy model to applied in numerous applications and attracted numerous
design suitable controllers for heat exchanger control was studied and researches such as [4], [5], [6].
studied. The controlled process was a shell-and-tube heat
exchanger and its Takagi-Sugeno fuzzy model was obtained using Heat exchangers are standard equipment in industry. They
experimental data. The control objective was to keep the output are devices specifically designed for the efficient heat transfer
temperature of the heated stream at a reference value. The from one fluid to another fluid over a solid surface. The heat
controlled output is the output temperature of the heated stream exchangers can be represented as nonlinear systems [7] with
- petroleum and the control input is the volumetric flow rate of interval parametric uncertainty. The control of them is a
the heating stream - water. Designed controllers were tested complex process due to the non-linear behaviour and
using simulations in the Matlab/Simulink environment. The complexity caused by many phenomena such as leakage,
simulation results confirm that designed controllers are suitable friction, temperature-dependent flow properties, contact
for successful control of heat exchangers. Using of the designed resistance, unknown fluid properties, etc. [8]. Various control
controllers can lead to smaller consumption of the heating strategies were developed for overcoming all mentioned
medium. Comparison with classical PID control demonstrates the problems.
superiority of the proposed control especially in the case, when
the controlled process is affected by disturbances. In the presented paper, the problems of the set point
tracking and the disturbance rejection in polynomial pole
Keywords—heat exchanger, Takagi-Sugeno fuzzy model, PID placement controller and Hf control techniques are
control, polynomial control design, Hf control investigated. The controllers are designed using Takagi-
Sugeno fuzzy model. The control responses obtained by the
I. INTRODUCTION polynomial and Hf controllers have smaller overshoots and
energy consumption compared with conventional PID control.
Fuzzy control has long been applied to industry with
several important results. Originally introduced as model-free II. TAKAGI-SUGENO FUZZY PLANT MODEL
control design approach, model-based fuzzy control has
gained widespread significance. Fuzzy control has proven to A dynamic Takagi-Sugeno fuzzy model is described by a
be a successful control approach to many complex nonlinear set of fuzzy “if …then” rules with fuzzy sets in the
systems or even nonanalytic ones. Takagi-Sugeno fuzzy antecedents and dynamic linear time-invariant systems in the
models represent fuzzy dynamic models or fuzzy systems. consequents.
This brings a twofold advantage. Any model-based technique
The ith rule of the Takagi-Sugeno fuzzy model is
including a nonlinear one can be applied to the fuzzy dynamic
considered in the form [9]:
models and the controller itself can be considered as a fuzzy
system. Fuzzy control has long been applied to industry with Plant rule i:
several important theoretical results and successful results. In
[1] a survey on recent developments of analysis and design of if z1(t) is M 1i and ... and zs(t) is M si then
fuzzy control systems focused on industrial applications na nb
reported after 2000 is presented. Over the last decades, there  yi t  =  ¦ ani  j y ( j ) (t ) ¦ bni  j u ( j ) (t )  
has been a numerous research on nonlinear modelling. It’s j 1
a
j 0
b

commonly known that an efficiency of the designed controller

is based on the model representing the plant or system itself.
for the continuous-time case,
Ever since Takagi and Sugeno [2] and Zadeh [3] presented an
ingenious method to represents a nonlinear system as a group if z1(t) is M 1i and ... and zs(t) is M si then
of linear time invariant (LTI) models mixed with nonlinear
na nb
functions to form a fuzzy model representation, thus the used
of Takagi-Sugeno (TS) fuzzy model approach has been  yi t  =  ¦ ani  j y(t  j) ¦ bni  j u(t  j) 
a b

j 1 j 0

646
978-1-4799-3528-4/14/$31.00 ©2014 IEEE
 for the discrete-time case, where i=1,..N. Further, A. Design of the controller
z(t) = [z1(t), z2(t), ..., zn(t)]T  Rn are the premise parameters Assume that the plant is described by Takagi-Sugeno
(the measurable plant variables), u(t)=[u1(t), fuzzy model (2). Let the controller be a system expressed by
u2(t), ... , um(t)]T Rm is the input to the plant, y(t)=[y1(t), y2(t), its transfer function [9]
..., yp(t)]T Rp is the output of the plant, M ij are fuzzy sets. 

Let us use product as t-norm operator of the antecedent deg q

part of rules and the centre of mass method for defuzzification.
q ( z , h)
¦ q ( h) z
i
i

The final output of the fuzzy system is inferred as follows.  GC z, h  = i 0

deg p
 
p ( z , h)
N
¦ ȝi zt  yi (t ) N
¦ p ( h) z
i 0
i
i

y t  = i=1
N
= ¦ hi zt  yi (t )
 ¦ ȝi zt  i=1
 
The coefficients qi, pi are some functions of the vector of
weights of rules h.
i=1
N § na nb · The goal is to assign the same characteristic polynomial of
= ¦ hi z t ¨  ¦ a ni a  j y (t  j )  ¦ bni b  j u (t  j ) ¸ the closed loop system for all h  H, i.e.
¨ ¸
i=1 © j 1 j 0 ¹
 a( z, h) p( z, h)  b( z, h)q( z, h) c( z)  
where
ȝi z t  where c(z) is a Schur polynomial of an appropriate degree.
 hi z t  = N
  The well-known condition of solvability of (9) is that the
¦ ȝi zt  greatest common divisor of polynomials a(z,h) and b(z,h)
i=1 divides the polynomial c(z) for all h  H, i.e.

s
  ȝi z t  = –M
j=1
i
j
z j t     g (( z, h), b( z, h)) c( z )  

The degrees of polynomials r(z,h), q(z,h) and c(z) can be

N found comparing degrees of polynomials of both sides of
  ¦ h zt  = 1 
i=1
i  equation (9):

deg c ! 2 deg a  1
By applying Z-transform to (3) the transfer function of the
model of plant is obtained [9]:  deg q  deg a  1  
deg p ! deg c  deg a
N

Y ( z , h)
¦ ȝi zt G ip ( z ) Substituting (6) and (8) to (9) the following equation is
G p z , h  = i=1
N
= obtained:
U ( z , h)
¦ ȝi zt 
i=1
N na deg p N nb deg q

¦ ¦ ¦ ¦ ¦ ¦q
N nb
 a ij z j pjz j  b ij z j j
c( z ) 
N ¦ hi zt ¦ b ij z j
i 1
hi
j 0 j 0 i 1
hi
j 0 j 0
jz

 ¦ hi zt G ip ( z ) i=1
N
j 0
na
 
i=1
¦ hi zt ¦ a ij z j All solutions of (12) can be expressed as
i=1 j 0
q ( z , h) q 0 ( z , h )  f ( z ) a ( z , h)
N   
¦ hi zt bi ( z) b( z , h)
p ( z , h) p 0 ( z, h)  f ( z )b( z, h)
i=1
N
a ( z , h) where q0(z,h) and p0(z,h) are the particular solutions of (12)
¦ hi z t a i ( z) and f(z) is an arbitrary polynomial.
i=1

where G ip (z ) denotes the transfer function of the

consequent of the ith rule.

2014 15th International Carpathian Control Conference (ICCC) 647

 III. SIMULATION AND RESULTS model was identified using the Strejc method [12] from the
step responses in the form of the nth order plus time delay
A. Process Description transfer function in (14).
Consider a co-current shell-and-tube heat exchanger [10],
where petroleum is heated by hot water through a copper tube
K
(Fig.1).  G= e  Ds  (14)
IJs + 1n
hot water
q3(t), T3in T3(z,t) Because of several step responses intervals for values of
T2(z,t) the gain K, the time constant W and the time delay D were
inlet outlet
obtained (Table 2) and the heat exchanger can be represented
petroleum petroleum as a system with interval parametric uncertainty. The system
q1, T1in q1, T1out
d1 d 2 d 3 T1(z,t) order n = 3. The mean values of the parameters are considered
to be nominal.

hot water TABLE II. IDENTIFICATION OF THE PROCESS DYNAMICS

z q3(t), T3out
z+dz
l Wmin [s] 15.5
Wmax [s] 23.5
Wmean [s] 19.5
Fig. 1. Scheme of the shell-and-tube heat exchanger
Kmin [K m-3 s] 4.2×104
Kmax [K m-3 s] 6.5×104
The controlled variable is the outlet petroleum temperature Kmean [K m-3 s] 5.35×104
T1out. Among the input variables, the hot water flow rate q3(t) Dmin [s] 1.0
is selected as the control variable. The mathematical model of Dmax [s] 3.0
the heat exchanger is derived under several simplifying Dmean [s] 1.5
assumptions. The coordinate z measures the distance of a
modelled section from the inlet. The fluids move in a plug
velocity profile and the petroleum, tube and water Takagi-Sugeno Fuzzy Model
temperatures T1(z,t), T2(z,t) and T3(z,t) are functions of the
The Takagi-Sugeno plant model was approximated by the
axial coordinate z and the time t. The petroleum flow rate is
following fuzzy rules [9] using experimental data
constant, whereas the water flow rate q3(t) is a function of
time t only. The petroleum, water and tube material densities Ri: if y(t-1) is M 1i and u(t-1) is M 2i then
Ui as well as the specific heat capacities CPi, i = 1, 2, 3, are
assumed constant.  yt  = ai y(t  1)  bi u(t  1)  

The simplified nonlinear dynamic mathematical model of

the heat exchanger is described by three partial differential where Ri (i=1, ..., 4) denotes the ith rule, M ij , j = 1, 2, are
equations [10]. Parameters and steady-state inputs of the heat fuzzy sets and they are defined by the Gaussian functions
exchanger are given in Table 1, where the superscript s
denotes the steady state and the subscript in denotes the inlet
[11].
 i
V i i
exp ¨ 
j 
§ x  ci 2
¨ j  ·
¸
M j (x j j,cj ) ¸ (16)
TABLE I. HEAT EXCHANGER PARAMETERS AND INPUTS ¨
©
i
2V j  
2
¸
¹
Variable Unit Value Variable Unit Value
N 1 5 U3 kg m-3 1000 The membership functions are
L m 10 CP1 J kg-1 K-1 2100
§  y  33.782 ·
¸ , M12 exp¨   y  33.89 ¸ ,
d3 m 0.05 CP2 J kg-1 K-1 418 § 2·
d12 m 0.025 CP3 J kg-1 K-1 4186 M11 exp¨ 
d23 m 0.028 q1 m3 s-1 3.7723×10-4 ¨
© 21.742 ¸¹ ¨
© 21.742 ¸¹
h J s-1 m-2 K-1 750 q3ins m3 s-1 1.1111×10-4
h J s-1 m-2 K-1 1480 T1ins K 308.52 § u  0.0000662 ·
U1 kg m-3 810 T2ins K 317.76 M 21 exp¨  ¸,
U2 kg m-3 8960 T3ins K 324.82 ¨ 20.0000282 ¸
© ¹
§ u  0.000133 ·
2
For the identification, the step changes ±10%, ±20%, M 22 exp¨  ¸.
¨ 20.0000282 ¸
±30% of the inlet volumetric flow-rate of the heating water © ¹
were generated.
Rule viewer that simulates the entire fuzzy inference
According to these step changes, the heat exchanger is a process is shown in Fig. 2.
time-delay nonlinear system with asymmetric dynamics. The

2014 15th International Carpathian Control Conference (ICCC) 648

 The PID controller parameters tuned using the Chien-
Hrones-Reswick method are kp = 1.46u10-4, ti = 19.5 s, td =
0.75.
2) Polynomial Control Design
The models of the controlled process G and the tuned
controller C are given in the form of transfer functions

b( s ) q(s)
 G=  C =  (18)
a(s) p(s)
Fig. 2. Fuzzy inference system
where polynomials p(s) and q(s) are calculated from the
The consequent parameter values are ai = [0.99 0.99 0.99 Diophantine equation [12]
0.99]T, bi = [962 934 836 839]T. Using (7), four discrete
transfer function were obtained, respectively four continuous
transfer functions by applying inverse Z-transform. The step  a ( s ) p ( s )  b( s ) q ( s ) c( s )  (19)
responses of the outlet temperature are shown in Fig. 3, where
experimental measured data and responses on the input for various cases of the polynomial c(s).
changes ±10%, ±20%, ±30% from the nominal transfer
function and from the TS fuzzy model are compared. The closed-loop system containing the controlled system
and controller (18) calculated from (19) cannot guarantee the
zero steady-state control error. The controller formed by the
q(s)/p(s) and 1/s assures that for the step changes of the
setpoint w and the step changes of the disturbance d the
control error e approaches asymptotically to zero if q(s) and
p(s) are solutions of the equation

 a( s ) sp( s )  b( s ) q ( s ) c( s )  

where c(s) is a stable polynomial.

The discrete models of the controlled system were
Fig. 3. Step responses of the outlet temperature on the step changes of the hot obtained from the Takagi-Sugeno fuzzy model. The
water flow rate. continuous-time models were obtained using the inverse Z-
transform. The controllers were calculated for each model
using (20) and the final controller was obtained from mean
B. Control of the heat exchanger values of four corresponding values in the form
1) PID control
PID controllers are described by the transfer function in
(17) 0.0041s  0.0001 1
 C=  (21)
s  70.066 s

§ 1 · 3) H’ Control
 C = k p ¨¨1 + + td s ¸¸  (17)
ti s Consider the plant model with the feedback as shown in
© ¹
Fig. 4. The signal w is the external input, z is the error signal,
which ideally should be zero, u is the control input, and y is
where kp is the proportional gain, ti the integral time, td the the observed feedback output. The block G is the generalized
derivative time and these parameters were tuned using the plant, and C the controller [15].
Cohen-Coon and Chien-Hrones-Reswick methods [13]. The
controller parameters were tuned for the model with the mean
(nominal) values of the identified parameters.
The transfer function parameters in (14) used for the PID
controller design are: the gain K = 5.35u104 , the time constant
W = 19.5 s and the time delay D = 1.5 s. The other two
parameters obtained from identification are the effective time
delay tu = 15.68 s and the effective time constant tn = 71.95 s. Fig. 4. General feedback system.

The PID controller parameters tuned using the Cohen- Suppose G(s) can be partitioned as follows [16]
Coon formulas are kp = 1.19u10-4, ti = 35.44 s, td = 5.49.

2014 15th International Carpathian Control Conference (ICCC) 649

 §z· § Gzw Gzu ·§ w · § w· §  WS S ·
 ¨¨ ¸¸ ¨ ¸¨ ¸ G¨ ¸  (22 ¨ ¸
¨ G yw ¸¨ ¸
G yu ¹© u ¹ ¨u¸
© y¹ © © ¹ F (G, C ) f ¨ WT T ¸ (27)
¨ W CS ¸
© CS ¹ f
The closed loop system z F (G, C )w has the transfer
function F(G,C) given by S and T defined in (28) are the sensitivity and the
complementary sensitivity functions, respectively, WS is the
performance weighting function which limits the magnitude of
 F (G, C ) 
Gzw  Gzu I  CGyu 1 CGyw  (23 the sensitivity function and WT is the robustness weighting
function to limit the magnitude of the complementary
Assume G has following state-space representation sensitivity function.

x (t ) Ax(t )  Bw w(t )  Bu u (t ) S (s) ( I  G (s)C ( s)) 1

(28)
 z (t ) C z x(t )  D zw w(t )  D zu u (t )   T (s) ( I  G( s)C ( s)) 1 G( s)C ( s)
y (t ) C y x(t )  D yw w(t )  D yu u (t )

It is assumed: Dzw=0 and Dyu=0, (A, Bu) is stabilizable, where

(Cy, A) is detectable, DTzuDzu and DywDTyw are invertible,
DTzwCz=0, DywBTw=0, (Cz, A) and (A, Bw) have no
unobservable modes on the imaginary axis. 1 1
 S ( jZ )  , T ( jZ )   (29)
Various techniques are available for the design of the Hf- WS ( jZ ) WT ( jZ )
controller. As it is known the robust controller is designed to
bring the f-norm of the plant to its minimum, three weight The weights WS, WT, WCS are the tuning parameters, and it
functions are added to the control system for loop shaping typically requires some iterations to obtain weights which
>17@. The loop shaping assures obtaining the desired frequency yield a good controller >17@. A good starting point is to choose
response of the plant. The classical feedback system structure
with weighting is shown in Fig. 5 >17@.
s Z0
z1  Z0 s
W1(s)  WS M , WT M ,W const .  (30)
CS
z3
s  Z0H Hs  Z0
W3(s)
z2
w e
C(s)
u
G(s) W2(s)
where H < 1 is the maximum allowed steady state offset, Z0
is the desired bandwidth and M is the sensitivity peak (for
-
example H = 0.01 and M = 2). For the controller synthesis, the
inverse of WS is an upper bound on the desired sensitivity loop
shape, and WT1 will effectively limit the controller output u
which is symmetric to WS around the Z = Z0.
Fig. 5. Control system for the synthesis of Hf controller
The objective is to find the controller C(s) in the form of a
The generalized plant F is given rational function and to make the closed loop system stable
satisfying expression F (G, C ) f Tzw f  J for a given
J ! 0.
z1 WS w  Gu  § z1 · § WS  WS G ·
¨ ¸ ¨ ¸ The Hf-controllers for the heat exchanger were calculated
z2 WT Gu ¨z ¸ ¨ 0 WT G ¸ § w·
 o¨ 2¸ ¨0 W ¸ ¨¨ ¸¸  (25) for each TS fuzzy model and the final controller was obtained
z3 WCS u z
¨ 3¸ ¨ CS ¸ ©u ¹ from mean values of four corresponding values in the form
e w  Gu  ¨e ¸ ¨1  G ¸
© ¹ © 
 ¹
Gn
0.0013s 2  0.000028s  1.1 u 10 7
 C (s)  
A possible state space realization can be written as s 3  10.12s 2  0.041s  2.58 u 10 5

The closed-loop poles with the controller in (31) are:

§ WS · §  WS G · §  WS S · [-0.0008, -0.005, -9.9994, -0.1182, -0.0167, -0.0051], i.e. the
¨ ¸ ¨ ¸ ¨ ¸
F (G, C ) ¨0 ¸  ¨ WT G ¸ C (1  GC )
1
¨ WT T ¸ (26) closed loop system is stable and J = 0.0165.
¨0 ¸ ¨ W ¸ ¨ W CS ¸
© ¹ © CS ¹ © CS ¹

2014 15th International Carpathian Control Conference (ICCC) 650

 Simulation results obtained using the designed polynomial Takagi-Sugeno model led to the smaller consumption of the
pole placement controller with integral action, Hf controller, heating medium and to the smaller overshoots than the use of
Cohen-Coon PID controller and Chien-Hrones-Reswick PID PID controllers tuned using the Cohen-Coon and the Chien-
controller are shown in Fig. 6. The figure presents the Hrones-Reswick methods. The simulation results confirmed
simulation results of the control of the heat exchanger in the that controllers designed using the Takagi-Sugeno model
task of the set point tracking and in the task of the disturbance represent the possibilities for successful control of heat
rejection. The set point changes from 40 °C to 37 °C at 400 s exchangers. Comparison with classical PID control
and then to 42 °C at 800 s. Disturbances were represented by demonstrates the superiority of the proposed control.
water temperature changes from 75 °C to 80 °C at 200 s, from
80 °C to 76 °C at 600 s and to 79 °C at 1000 s. ACKNOWLEDGEMENT
The simulation results were compared also using the IAE The authors gratefully acknowledge the contribution of the
(integral absolute value of error) criterion defined e.g. in [14] Scientific Grant Agency of the Slovak Republic under the
as follows grants 1/0973/12 and the contribution of the Slovak Research
and Development Agency under the project APVV 0551-11.
f
 IAE ³ e(t ) dt   REFERENCES
0
[1] R.E. Precupa, H. Hellendoorn, “A survey on industrial applications of
fuzzy control,” Computers in Industry, vol. 62, 2011, pp. 213–226.
[2] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its
application to modelling and control,” IEEE Transactions on Systems,
Man and Cybernetics, vol. 15, 1985, pp. 116-132.
[3] L. Zadeh, 1996. “Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems,”
Selected Papers, 1996.
[4] A. O. Hamadameen and Z.M. Zainuddin, “Fuzzy stochastic linear
programming problems with uncertainty probability distribution,”
Australian Journal of Basic and Applied Sciences, vol.7(2), 2013, pp.
846-852.
[5] S. Hameed, F. Shahzad and S. Asgar, “A Fuzzy correlation scheme for
privacy preservation in knowledge-based systems,”Australian Journal of
Basic and Applied Sciences, vol. 6(9), 2012, pp. 562-571.
[6] C. W. Park and Y.W. Cho, “Robust fuzzy feedback linearisation
controllers for Takagi-Sugeno fuzzy models with parametric
Fig. 6. Comparison of the outlet petroleum temperature control. uncertainties,” Control Theory & Applications, IET, vol. 1, 2007,
pp.1242-1254.
The IAE values and the consumption of the heating [7] W.S. Janna, Engineering Heat Transfer, 3rd ed. Tennessee: The
University of Memphis, 2009.
medium are given in Table 3. The smallest values of IAE and
[8] R.W. Serth, Process Heat Transfer Principles and Applications.
the consumption of the heating medium are assured using the Academic Press, Burlington, USA, 2007.
polynomial controller with integral action.
[9] R. Pytelková and P. Hušek, “Synthesis of Stable Takagi-Sugeno Fuzzy
Systems.” Available from http://www.wseas.us/e-library/ conferences/
TABLE III. VALUES OF IAE AND HOT WATER CONSUMPTION tenerife2001/papers/532.pdf.
[10] A. Vasiþkaninová, M. Bakošová, A. Mészáros and J. Klemeš, “Neural
Controller IAE V [m3] network predictive control of a heat exchanger,” Applied Thermal
Polynomial 422 0.2332 Engineering. vol. 31, 2011, pp. 2094-2100
Hf 460 0.2355 [11] A. Vasiþkaninová and M.Bakošová, “Robust control of heat
PID Cohen-Coon 490 0.2382 exchangers,” Chemical Engineering Transactions, vol. 29, 2012; pp.
PID Chien-Hrones-Reswick 674 0.2417 1363-1368.
[12] J. Mikleš and M. Fikar, Process Modeling, Identification, and Control.
Berlin, Springer, 2007.
[13] W.B. Bequette, Process Control: Modeling, Design, and Simulation.New
IV. CONCLUSION Delhi, Prentice-Hall of India, 2003.
[14] B.A. Ogunnaike and W.H. Ray, Process Dynamics, Modelling,
Two controllers were designed using the Takagi-Sugeno and Control. New York, Oxford University Press, 1994.
model of the shell-and-tube heat exchanger, the polynomial
[15] O.H. Bosgra, H. Kwakernaak and G. Meinsma, Design methods for
pole placement controller with integral action and the Hf control systems, DISC lecture notes 2000-2001. Available from
controller. Two PID controllers were also tuned using the web.boun.edu.tr/eskinat/ME687.html.
Cohen-Coon and the Chien-Hrones-Reswick methods for [16] H.T. Toivonen, Robust Control Methods, Process Control Laboratory
comparison. The simulation results obtained using all Abo Akademi University Turku, Finland. Available from
designed controllers were compared calculating the integral [email protected], http://users.abo.fi/htoivone/courses/robust/rob1.pdf.
performance index IAE and the consumption of the heating [17] A. Bansal and V. Sharma, “Design and Analysis of Robust H-infinity
Controller,” Control Theory and Informatics, vol. 3(2), 2013, pp.7-14.
medium. The control response obtained by the polynomial
controller had the smallest overshoots and also the smallest
IAE value. The use of the controllers designed using the

2014 15th International Carpathian Control Conference (ICCC) 651