You are on page 1of 5

Modeling of Heat Transfer Process by Using

Discrete Fractional-Order Neural Networks

Dominik Sierociuk Ivo Petras
Institute of Control Institute of Control and Informatization
and Industrial Electronics of Production Processes BERG Faculty
Warsaw University of Technology Technical University of Kosice
Koszykowa 75, Warsaw, Poland B. Nemcovej 3, Kosice, Slovakia
Email: Email:

Abstract- This paper deals with a discrete fractional-order The three equivalent definitions used for the general frac­
neural network and its application for modeling of a heat tional differintegral are the Griinwald-Letnikov definition, the
transfer process. Proposed neural network is implemented in
Riemann-Liouville and the Caputo definitions [10], [l3]. For
Matlab/Simulink and applied to the heater which is a part of the
equipment Armfield PCT40. Results obtained from experiments
our purpose we will adopt the Griinwald-Letnikov definition:
presented in the article show that the fractional-order neural Definition 1: The Griinwald-Letnikov definition for k =
t a, where a is a real constant, is given as:
network properly modeled the unknown dynamics. i"

Nowadays the fractional calculus is attracted many scien­

lim h
L( �
j (�)x(t
� jh)
tists and engineers. However the fractional calculus is more
. I
than 300 years old, only in last few decades appeared the hm -h [:,."'x(t),
h�O '"
publications with practical applications of this mathematical
phenomenon. Some of them can be found in books [4], where [.J means the integer part.
[10], [11], [l3]. Moreover, there are works with application •
of the fractional calculus in control theory [9] and even in To present fractional-order discrete time neural network we
fractional-order nonlinear systems [12]. In these both areas have to introduce discrete fractional order nonlinear system.
of applications, the neural networks play an important role In this paper the following definition of the fractional order
[6], [7]. It is possible to apply the fractional calculus in difference is used [10], [13]:
modeling of dynamical systems via the neural networks. The Definition 2: Fractional order difference is given as follows:
preliminary results of the discrete fractional order networks k
as well as comparison to the integer order network dynamic
were presented in [16]. In this paper the extension for this idea
[:,."'Xk = L( �
j (�)Xk-j,
and application to the real plant is presented. The extension
is that in the model the also differences of the input signal is where aE lR, is a fractional order and kEN is a number of
included. sample for which the difference is obtained.
This article is organized as follows: In Sec. II a discrete •
fractional-order nonlinear systems are introduced. Sec. III In our case the artificial neural network is used to modeling
describes a discrete time fractional-order neural networks. the fractional order nonlinear systems. Using fractional order
In Sec. IV the experimental setup and obtained results are difference the following nonlinear discrete fractional order
presented. Sec. V concludes this article with some remarks system in state-space is defined:
and ideas for further work. Definition 3: [l5]The nonlinear discrete fractional order
system in a state-space representation is given by the follow­

Fractional calculus itself is a generalization of integration

and differentiation to non-integer order fundamental operator
aD't, aE lR, where a and t are the bounds of the operation.
The continuous integro-differential operator is defined as

a> 0, Yk
a=O, where aE lR is a system order, 10 and hO are the nonlinear
a < O. functions.

978-1-4577-0914-2111/$26.00 ©2011 IEEE 146

�r1 I
The system which we take into consideration is given as
z-n I
the following relation:

g ( A(n-l)e> Ae>
Yk+n-l,···,L.l. Yk+l,Yk I

This system can be modeled using the artificial neural

Y z�(n-l)Q(z) I
network presented in the next section.
neural H �-nQ(z) P'r+-
11 I
z-n I

The discrete time fractional order neural networks are a

very efficient tool for modeling the fractional order non-linear
systems, especially when the analytical model is not accessible
H z-n+l�Q(z) I

or is too complicated to use it directly. The main idea is to

replace the non-linear function gO in the model given by
equation (1) by the static neural network. This neural network
H z�(n-l)Q(z) I

has to be trained with using fractional order differences of

the outputs and inputs signals. These differences have to be
obtained by the prefiltering process.
By time shift of the equation (1) we obtain the following

g ( A(n-l)e>
Yk-l,oo.,L.l. Yk-n+l,Yk-n
Fig. 2.
Scheme for on-line simulation of discrete fractional order neural

L.l. Uk-l, 0 0 . , L.l. Uk-n+l, Uk-n (2)
Fig. 2 presents the signal processing scheme for simulation
of the nonlinear discrete fractional order system based on the
neural network obtained in the training process.

A. Experimental Setup of Armfield PCT40

-1 z-n+l�Q(z) I
As it was aforementioned, we used a PC to control the
PCT40 unit. The PC was connected directly to the PCT40
-1 z�(n-l)Q(z)
I training
via 60 pins connector. In our case we have connected the
PCT40 unit to the PC via multifunction 110 laboratory card
MF624. By using a Real Time Toolbox in Matlab/Simulink
�� z-n I
it is possible to communicate with the PCT40 unit and its all
inputs and outputs.
We will use a part of the PCT40 unit, namely heater of
-1 z-n+l�Q(z) I
I water (or other liquid), which has two modes: (i) heating and
(ii) cooling. The heating mode is based on electrical spiral

-1 z�(n-l)Q(z) I
dipped in the water. The cooling is based on pipe spiral dipped
in the water, which allows to flow cold water in the pipe and
makes water-cool. Technical parameters of the heater are [1]:

-1 �nQ(z) I
I • input for heating spiral SSR (inverse logic signal): 0
means heating, 1 means stop heating (binary values);
• input for cooling pipe spiral with cold water SOLENOID:
open or close valve in interval 0 - 100 %;
Fig. 1. Signal prefiitering for teaching the neural network
• output temperature T1: voltage signal within range 0-5 V
The prefiltering process for teaching the neural network is with maximal temperature approximately 80 DC;
presented in Fig. 1. Such neural network can be implemented Note that each module of the PCT40 unit has its own fuse
in Matlab/Simuling using Neural Network Toolbox and Frac­ system, e.g. it is not possible to switch on heating while there
tional State-Space Toolkit [14], [15]. is not water at certain level, etc.

by applying the Laplace transform to the equation (3) we
T(s,x) a2 sT(s,x) - a2 T(O,x) (5)

assuming zero initial conditions on the whole beam the sec­

ond order ordinary differential equation (with respect to the
variable x) is obtained. The solution of such an equation has
the following form:

T(s,x) = Cl ( S ) sinh(axvs) + C2 ( S ) cosh(axVS) (6)

it is easy to find that:

Fig. 3. Photo of experimental setup

using the equation (4) and assumption that the heat flux at the
In Fig. 3 is depicted the experimental hardware setup of
end of the beam is equal to zero (isolated end of the beam)
PCT40 unit with heater module and its connection to the PC
with the Matlab/Simulink through MF624 laboratory card.
H(xj,t) = 0 and H(xj,s) 0, the Cl ( S ) x independent

function is obtained as:

SOLENOID Cold water


By using the relation (6) and (4) the transfer function
describing the relation between the heat flux on the beam
beginning H(t,O) as an input and temperature at the same
point T(t,0) was obtained as follows:

T(s,O) 1
G(s) (7)
H(s,O) avstanh(axjVS)
= =

what for semi-infinite beam xj --+ 00 has the form

Fig. 4. Control scheme of heater T(s,O) 1


H(s,O) avs
= =

Control diagram of the experimental setup (see Fig. 3) is

depicted in Fig. 4. The VS can be interpreted as an half order derivative what
gives the following relation
B. General Mathematical Model of the Heat Transfer
The ideal heating process, without energy loss, of a the H(t,O) = a OD�·5T(t,0). (9)
beam can be described by the following partial differential
equation [5], [8] The results presented above was obtained for ideal, very
special case, but gives strong motivation of using the fractional
(3) order model for modeling the heat process. However the
PCT40 setup has much more complicated geometry what can
with the following boundary conditions: not be easily described analytically. For modeling this relation
the fractional order neural network can be used. Moreover
T(O,x) = 0, T(t,O) = u(t), the dependency between the control signal and the heat flux
generated by heater is also non-linear, because the heat flux
where T(t,x) is a temperature of the beam at time instant
depends on the heat power P U 2 / R and we can control the
t and space coordinate (distance) x, and a is a parameter

voltage on the heater (exactly on/off voltage). Additionally

which depends on beam parameters like heat conductivity and
the cooling flux has a different character, is caused by the
density. The another important equation is a definition oh the
cold water flux in the cooling pipe. The cooling flux is not
heat flux:
a symmetrical to the heat flux and this relation is unknown, this
H(x,t) T(t,x) (4) also can be modeled by the neural network.

C. Results of Fractional-Order Neural Modeling 30
_teaching output signal Y,
In this section the results of modeling the heat transfer 25
_network response for U,
process by using discrete fractional order neural network is
presented. The system has one input which for positive value
turns on the heater (it is binary value) and for negative value
turns on the cold water flux (also binary value). In that way the
input controls both heating and cooling process. The output of
the systems is taken a relative temperature (difference between
measured and initial temperature). All data were collected in 20 40 60 80 100
time [s)
sampling time 1 sec. The neural network has one hidden layer
with 6 neurons with nonlinear tansig activation function and
Fig. 6. Results of fractional order neural modeling for the teaching signals
one linear neuron in output layer (experimentally found). The Ul,Yl
network is trained with 400 epochs by using the Matlab routine
For the neural modeling we assume the following structure 15
of the nonlinear system:
- teaching output signal Y2
(10) _network response for U2

The structure has the final order equal to 0.5 because as it

can be seen in literature, the ideal heat transfer process is a
half-order integrator. The additionally order 0.25 is added in
order to allows better modeling unknowns dynamics of the
real plant model. 20 40
time [s)
60 80 100

The neural network used in the experiment has 4 inputs in

order to allows modeling the relation given by the equation Fig. 7. Results of fractional order neural modeling for the teaching signals
In the experiment we use 3 types of input signal for teaching
process and one type for testing, all of the signals are presented
in the Fig. 5. �14
OJ 12
� 10
f-.---------------, ...................... . - teaching output signal Y3
§ B
- network response for U3
....... teaching input signal U, � 4
. teaching input signal U 2 i. ...
._._.teaching input signal U
3 %��----L-------L-------L ------�L ------�00
6 O B O 1
_ testing input signal Ut

·iii -0 . 5 , time [s)

-1 _._._._._._ . _._ '-------1 Fig. 8. Results of fractional order neural modeling for the teaching signals
o 20 40 60 80 100
time [s)

Fig. 5. Teaching and testing signals used in experiment


Figures 6, 7 and 8 present results of neural modeling for the '"

OJ 15
same signals that was used for network training process U1,U2 '§
and U3. As it could be notice the accuracy of the modeling is
very good. E
The final check for the neural network training process 15
is a check for the testing signal different from the learning
signals. This can presents clearly the ability of the network 20 40 60 BO 100
time [s)
to generalize the unknown non-linear function. The result of
this check is presented in Fig. 9 and as it could be seen the Fig. 9. Results of fractional order neural modeling for the testing signals
discrete fractional order neural network properly modeled the Utl,Y tl
unknown dynamics.
In Fig. 5 - Fig. 9 are shown the results obtained via exper­
imental measurements on the heater and the results obtained
via fractional-order neural network modeling. A. Concluding Remarks
In this paper was proposed a discrete fractional-order neural
network. We have shown that such structure of the neural

network is appropriate for modeling of the fractional-order [14) D. Sierociuk, Fractional Order Discrete State-Space System Simulink
Toolkit User Guide,
dynamic systems. In our case we used the data from heater
[15) D. Sierociuk and A. Dzielinski, Fractional Kalman filter algorithm for
and we have used the structure with the final order equal to 0.5. states, parameters and order of fractional system estimation, Int. 1. Appl.
The fractional-order neural network has been implemented in Math. Comput. Sci., vol. 16 (1),2006, pp. 129-140.
Matlab/Simulink. [16) D. Sierociuk, G. Sarwas and A. Dzielinski, Discrete fractional order
artificial neural network, Proceedings of the FDA'10, IFAC, October
18-29,2010, Article no. FDAIO-134.
B. Ideas for Further Work
Since the neural networks are a good candidate for a
generic, parametric, non-linear model and have capabilities for
modeling with a desired accuracy even for modeling of non­
linear fractional-order system [3], it is necessary pay attention
to them. However, there are many open and potential problems,
which should be taken into account.
Some addition extensions and possible applications are:
• neural network estimator
• neural network controller
In further work we will show the way how to implement the
fractional-order neural network estimator and fractional-order
neural network controller for the heater used in this paper.


This work was partially supported by the Polish Min­

istry of Science and Higher Education grant number
4125/B/T02/2009/36 and the European Union in the frame­
work of European Social Fund through the Warsaw Univer­
sity of Technology Development Programme (by Center for
Advanced Studies WUT).
This work was also supported in part by the Slovak
Grant Agency for Science under grants V EGA: 110390110,
110497111, 110746111; grants: APV V-0040-07, SK-FR-0037-
09, SK-PL-0052-09, and CEV I: OPVaV-200812.1I01-S0RO.


[I] Armfield,LtD., PCT40 Multifunction Process Control Teaching System,

Technical documentation,, 2010.
[2) 1. Babic, G. Takac, I. Petras and D. Bednarova, Identification of Model
Parameters and Control of Heater on Laboratory Object PCT40, Proc.
of the ICCC'20ll, Velke Karlovice, Czech Republic, May 25-28,2011.
[3) F. Benoit-Marand, L. Signac, T. Poinot and 1. C. Trigeassou, Identi­
fication of non linear fractional systems using continues time neural
networks, Proceedings of the FDA'06, IFAC,2006.
[4) R. Caponetto, G. Dongola, L. Fortuna and I. Petras, Fractional Order
Systems: Modeling and Control Applications, Singapore, World Scien­
[5) A. Dzielinski, D. Sierociuk and G. Sarwas, Some applications of
fractional order calculus, Bulletin of the Polish Academy of Sciences,
Technical Sciences, Vol. 58, No. 4,2010
[6) 1. Kalkkuhl, K. 1. Hunt, R. Zbikowski and A. Dzielinski, Applications
of neural adaptive control technology, World Scientific, 1997.
[7) B. Krose and P. van der Smagt, An Introduction to Neural Networks,
The University of Amsterdam, 1996.
[8) 1. Mikusi6ski, Operational Calculus, PWN-Polish Scientific Publishers,
Warszawa 1983.
[9) C.A. Monje, Y.Q. Chen, B.M. Vinagre, D. Xue and V. Feliu, Fractional­
order Systems and Controls, Springer,2010.
[10) K. Oldham and 1. Spanier, Fractional Calculus, Academic Press, New
York, 1974.
[II) A. Oustaloup, La Derivation Non Entiere: T heorie, Synthese et Appli­
cations, Paris, Hermes, 1995.
[12) I. Petras, Fractional-Order Nonlinear Systems: Modeling, Analysis and
Simulation, Series: Nonlinear Physical Science, Springer, HEP,2011.
[13) I. Podlubny, Fractional Differential Equations, Academic Press, San
Diego, 1999.