ECAI 2017 - International Conference – 9th Edition

Electronics, Computers and Artificial Intelligence

29 June -01 July, 2017, Targoviste, ROMÂNIA

Genetic Algorithm for System Modelling

Olteanu Marius, Paraschiv Nicolae Koprinkova Petia, Todorov Yancho

Department of Automatic Control, Institute of Information and Communication
Computers and Electronics Technologies,
Petroleum-Gas University of Ploiesti Bulgarian Academy of Sciences
Ploiesti, Romania Sofia, Bulgaria
[email protected] [email protected]
[email protected] [email protected]

Abstract – The process of modeling of a real system simulation with known parameters and with added
usually implies an iterative approach where an initial noise that resembles measurement uncertainties. Only
model is incrementally modified in order to increase its intervals are known for the model parameters, this way
accuracy with regard to available experimental data. the chromosomes are limited to valid range of values.
Many approaches are discussed in the literature, among It is presumed that initial conditions are known with
which some are based on Artificial Intelligence good accuracy.
techniques. Continuous improvement in the
mathematical model adequacy is very similar with the II. PROCESS MODEL
principal of genetic evolution applied in computational
intelligence methods. The present paper investigates the
application of genetic algorithm technique to optimize The simulated process consists of a closed vessel,
the model parameters for an industrial process. in which a mixture of propane (chemical formula
C3H8) and propylene (C3H6) in the liquid and gas
Keywords – genetic algorithm, chemical process, phase is compressed adiabatically by an increase in the
system identification liquid level. The mixture is encountered in oil
refineries where it usually represents the feed for a
I. INTRODUCTION binary distillation column that separates propylene as
Process modeling represents a current activity for the top product and propane as the bottom residue.
many researchers in almost any scientific area. In In Fig. 1, a schematic view of the closed vessel
order to build a model, usually two approaches are together with the main process variables are presented.
followed: building an analytic model based on the
natural laws (from physics, chemistry etc.) or an
experimental model having a more or less complex
structure (statistical models, Artificial Intelligence
based models, etc.). Following one of the approaches,
the step of model parameters identification is an
important one. There are many techniques developed
for the identification of parameters for linear and also
nonlinear models.[l, 2] In the case of a more complex
analytic model, the identification methods are seldom
used due to the complexity of the optimization
problem. The present paper proposes a method of
identification based on genetic algorithms.
Genetic algorithms are known for solving in a Figure 1. Closed vessel process
heuristic manner very complex optimization problems
(functions with many variables, many local optima, The variables have the following meaning:
nonlinearities etc.), by repeatedly applying genetic
operators to the current population of solutions.[3,4] In - Fin, Fout – input/output flow [m3]
the proposed method, using as accurate experimental - H – level of liquid [m]
data as possible, many simulations are made, such that
every chromosome encodes a possible model (the - x, y – light component (propylene)
actual chromosome of the Genetic Algorithm encodes concentration for liquid and gas respectively
the parameters of the model). [mol.fr.]
In order to characterize the efficiency and the - P, V, T – state variables for the gas phase
success of the method, the “real” process is actually a (pressure [bar], volume [m3], temperature [])
978-1-5090-6458-8/17/$31.00 ©2017 IEEE
 2 Authors Names

The process model can be separated in a sub- liquid equilibrium based on Lagrange interpolating
model for the gas – obtained by applying a non ideal polynomials.[7]
gas equation and one for the closed vessel based on
material balance. The model is designed to be more
flexible, such that although the compression of the gas
is considered adiabatic, further improvements to be
easily added. [5]
A. Gas model Figure 3. Input/ouput view of the „equil‟ routine
Because of the relative high pressure of the
mixture that is used in industry (approx. 20 bar), the Initially, the mixture is considered to be at
ideal equation of state cannot be applied (the equilibrium, in conclusion a routine that takes as
compressibility factor Z ~ 0.76 for x=0.6, P=15.8 bar). inputs the concentration x in liquid phase and the
An equation of state often used because of its pressure of the system (Pg=18.5 bar) is used. The
simplicity and relatively good accuracy is Redlich- routine uses real experimental data published in
Kwong equation[6]: literature.
B. Closed vessel model
Increasing the level of liquid in the vessel results in
a compression of the gaseous phase that will generate
where: a higher pressure and also a temperature increase.
Considering an adiabatic compression, all the
- R is the gas constant mechanical work of the gas will transform in sensible
- a and b are parameters calculated with the heat.
following relations:

The symbols Tc and Pc represent the critical

temperature and respectively critical pressure, while
Ωa, Ωb are constant values.
Figure 4. Model for the closed vessel

The following table summarizes all the model

parameters.

TABLE I. MATHEMATICAL MODEL PARAMETERS

Figure 2. Gas block model Unit

Parameter Value

Knowing the value of the liquid level H, the Vessel area -

volume can be computed using the physical 1.767 m2
Av
dimensions of the vessel (area of the vessel: Av=1.767 Vessel height -
2.0 m
m2 and the total height of the vessel: Hv=2m). Hv
Introducing the volume in the EOS, it results the Input flow - Fin 0.05 m3/s
current pressure of the gas: Pc.
Output flow -
0.002 m3/s
In order to compute the correct molar volume, the Fout
total number of moles (mg – moles) has to be known Initial pressure
15.8 bar
in advance. Because in reality it is very hard to - Pc
Initial liquid
quantify the number of moles of gas, the Genetic level - H
1.0 m
Algorithm will be used to estimate an appropriate Propane liquid
value. concentration – 0.60 molar fr.
x
The value of the light component concentration (x Propane gas
– molar fractions) represents the other parameter that concentration - 0.6276 molar fr.
will be estimated by the Genetic Algorithm. In y
practice, an exact concentration of a mixture is hard to Initial
obtain (in many situations, only laboratory tests are temperature - 40.47 °C
Tc
available at hourly intervals). Molar mass -
1385.2 moles
Because the mixture of propane and propylene, mg
especially at high pressures deviates significantly from
ideality, the authors have developed a mathematical The sensible heat that results from the compression
routine that can be used to solve the problem of vapor- is given by the following relation:
 PaperName 3

structure /member/
real x1
where: real x2
- Cv – represents the specific heat of the gas, real fitval
computed with the relation:
end structure

Cv1 and Cv2 correspond to the specific constant

volume heat of propylene and propane.
The simulation runs until the liquid level equals
Hfinal = 1.7m, a total time of tmax = 25.8 sec. results. In
the following figures (Fig.4 and Fig.5) there are
represented the pressure and temperature increase in
the specified time interval.

Figure 7. General flowchart for Genetic Algorithm

Figure 5. Temperature increase of the gas The two encoded values are used for the
representation of the estimated liquid propylene
The final temperature of the gas is 76.66 deg.C, concentration (x) and for the estimation of molar mass
and also the final pressure is 36.07 bar absolute. (mg). Fitness function was chosen as the MSE (Mean
Squared Error) between the experimental data
(simulated data) and the actual values of the
population members:

Where:
- npop represents the population number
The genetic operators used were:
- tournament selection (a random selection of
Figure 6. Pressure increase of the gas
two members of the population from which
the best is chosen – having the smallest
fitness)
- crossover for real encoding – two members of
III. GENETIC ALGORITHM OPTIMIZATION the population were chosen and their x1 and 2
In many optimization applications, the flexibility values were crossed by arithmetical mean
and versatility of the Genetic Algorithms are used as
main advantages. - mutation – a random individual was chosen
and its value was randomly changed in a
The main sections of a general Genetic Algorithm predefined interval
are presented in the following figure (Fig. 7). One of
the most important problems that have to be first The intervals used for the two parameters of
solved is the encoding of the solution – or the the model were:
chromosome representation. - mol.fr.
For the optimization of the presented mathematical - moles
model, a FORTRAN structure having three fields had
been used. For the stopping criteria, a number of generations
has been used.
 4 Authors Names

The entire simulation and also the genetic

algorithms were programmed successfully in Oracle
Solaris Studio FORTRAN (v.12.4) under Linux V. CONCLUSIONS
operating system.
The values obtained for many Genetic Algorithm
IV. RESULTS executions are in good agreement with the real data
The following table lists the usual parameters from the experimental simulations. More parameters
utilized with the Genetic Algorithm. can be encoded in the genetic algorithm which
represents the most important aspect of the technique.
TABLE II. GENETIC ALGORITHM PARAMETERS The complexity of the model will not introduce
additional constraints, the only problem is the
Parameter Value simulation running time.
The method could be further improved by
Population size 400 introducing a parallel genetic algorithm that can speed
Generation
up the execution of the many necessary simulations.
50
number
Crossover
0.85
probability
REFERENCES
Mutation
0.1
probability
[1] L. Ljung, System Identification: Theory for the User. Prentice
Hall, 1987.
The computer used for the simulations was Intel [2] P. Eykhoff, Identificarea sistemelor, vol.1, Editura Tehnica,
Core 2 Duo E7200 based, running at 2.53GHz for Bucuresti, 1977.
which a simulation time was approximately 1 minute. [3] M. Mitchell, An Introduction to Genetic Algorithms, MIT
In the following table there are presented simulation Press, 1999.
results with the mentioned parameters. [4] Z. Michalewicz, Genetic Algorithms + Data Structures =
Evolution Programs, Springer-Verlag, 1996.
TABLE III. SIMULATION RESULTS [5] R.G.E. Franks, Modelarea si simularea in ingineria chimica,
Editura Tehnica, Bucuresti, 1979.
Best [6] R.H.Perry, D.Green, Perry‟s Chemical Engineers‟ Handbook,
Genera x value mg value Fitness
member 6th Edition, International Student Edition, McGraw-Hill,
tion value
index 1984.
14 12 [7] M. Olteanu, Cercetari privind utilizarea algoritmilor evolutivi
0.603547 1384.48 17.9640 in conducerea automata avansata a proceselor de fractionare,
30 101 Teza de doctorat, Universitatea Petrol-Gaze din Ploiesti,
0.611205 1381.13 17.9681 2014.
5 201
0.599570 1384.96 17.9653