Vous êtes sur la page 1sur 4

0-7803-8560-8/04/$20.

00 2004IEEE
MODEL IDENTIFICATION OF THERMAL PROCESS IN POWER
PLANT
Liu Changliang, Liang Weiping, Sun Wanyun, Su Jie
North China Electric Power University, Baoding, 071003,China
ABSTRACT
The iaentification methoas of inaustrial process are
aiscussea. The appliea MATLAB program of least
squares methoa is given ana usea to iaentify pulse
transfer function. A kina of improvea genetic
algorithm is introaucea to iaentify the transfer
function ana parameters of nonlinear aynamic moael
of thermal process. In the algorithm, floating-point
coaing, rank-basea selection, elitist reservation ana
grouping methoa are usea, the premature
convergence is restrainea, the global ana local
searching ability is improvea. With the genetic
algorithm, the transfer function of topical thermal
process can be iaentifiea accurately. The parameters
of nonlinear moael can be moaifiea accoraing to
operating aata of power plant, no matter what kina of
input signal is usea, such as step signal, ranaom
operating signal.
1. INTRODUCTION
The model oI thermal process in power plant is a very
important base to improve the control system,
whether the control algorithms are model-based or
not. Generally, the Iollowing three kinds oI model are
used widely in engineering and technological
research: pulse transIer Iunction, transIer Iunction
and nonlinear dynamic model. The pulse transIer
Iunction model can be got according to operating data
with least squares method. The transIer Iunction oI
thermal process is a useIul model that plays an
important part in the design, debugging and
parameters optimization oI control system. There are
many methods to get object transIer Iunction, such as
step response method, Irequency response method.
Because oI the nonlinear and time-change
characteristics, the thermal process in power plant
must be described with diIIerent transIer Iunction
model in diIIerent operating load. To solve this
problem, nonlinear dynamic model is used in some
cases. The nonlinear model oI thermal process in
power plant is mainly setup according to basic
conservation principles oI mass, energy, momentum,
heater transIer equations and knowledge about
thermodynamics and hydrodynamics. To ensure the
validity oI the model, the model parameters must be
modiIied according to operating data.
In this paper, the MATLAB identiIication program oI
pulse transIer Iunction is designed. A kind oI
improved genetic algorithm is introduced to identiIy
transIer Iunction and to modiIy the parameters oI
nonlinear model.
2. PULSE TRANSFER FUNCTION
IDENTIFICATION
With MATLAB Iunction, it is simple to get pulse
transIer Iunction Irom output data and input data.
Assume that the diIIerence equation oI the model is
shown as Iollow.
(1) u|iT|
1 m
I 1)T| u|(i
m
I
... 1)T| - m u|(i
2
I m)T| u|(i
1
I
y|iT|
1 n
g 1)T| y|(i
n
g ... 1)T| - n y|(i
2
g n)T| y|(i


Where u(i) is input, y(i) is output, T is sampling time.
The corresponding pulse transIer Iunction is:
1 n
g ...
1 - n
z
2
g
n
z
1 m
I ...
1 - m
z
2
I
m
z
1
I
G(z)


(2)
The input and output data should accord with
Iollowing equations:
(3) u(1)
1 m
I u(2)
m
I ... 1) - u(m
2
I
m) u(1
1
I y(1)
1 n
g - y(2)
n
g - ... - y(n)
2
g - n) y(1


(4) u(2)
1 m
I u(3)
m
I ... u(m)
2
I
m) u(2
1
I y(2)
1 n
g y(3)
n
g - ... - 1) y(n
2
g - n) y(2

(5) u(k)
1 m
I ... 1) - k u(m
2
I m) u(k
1
I
y(k)
1 n
g 1) y(k
n
g - ... - 1) - k y(n
2
g - n) y(k


The matrix equation is:

1 m
I
1
I
1 n
g
2
g
k
u
k m
u
k
y -
1 - k n
y -
2
u
2 m
u
2
y -
1 n
y -
1
u
1 m
u
1
y -
n
y -
k n
y
2 n
y
1 n
y
.
.
. .
. . . . . .
. . . . . .
. .
. .
.
.
That is: YXG (6)
The unknown parameters in (6) are: g
2
g
n1
,I1,
I2 Im1. Generally, k~nm1, so, accurate result
can,t be got. It can be solve with least squares method.
That is very simple in MATLAB.
GX\Y, (7)
The MATLAB Iunction oI identiIication is:
[num,aen{moaelID(u0,y0,m,n)
klength(u0),Yy0(n1.k),
for i1.length(Y)
X(i,.)[-y0(ni-1.-1.i) ,u0(mi.-1.i){,
ena
GXY,
aen[1 G(1.n){, numG(n1.length(G)),
Sometimes, the value oI G may be very small; the
data type should be deIined as Iormat long e in the
program.
It is shown by simulation results that the order oI
model (m, n) should be assigned same value. that is
very important.
Although the program is very simple, accurate
identiIication results can be got with it. A model
identiIication example oI power plant with least
squares method is shown in Iigure 1.
Fig.1 model identiIication with least squares
This Iigure shows us the dynamic process oI drum
pressure when turbine control valve increase 10.
The curve 1 is the operating data; curve 2 is the
identiIication result. These two curves is largely
overlap.
3. TRANSFER FUNCTION IDENTIFICATION
Generally, step input or M sequence signal is used in
traditional identiIication methods. It is diIIicult to
create these kinds oI signal in power plant.
Furthermore, ideal response curves and accurate order
is required in some methods. So, it is not easy to use
these identiIication methods into power plant.
According to the characteristics oI thermal process,
the improved genetic algorithm can be used to
identiIy transIer Iunction model |1|. In the algorithm,
Iloat coding, rank-based selection, elitist reservation
and grouping calculation are used. In the crossover
and mutation operation, the population is divided into
Iour parts and diIIerent parameters are used Ior every
part. With these methods, the searching perIormance
oI the genetic algorithm is improved. It is shown by
simulation research that the genetic algorithm-based
identiIication method is oI good practicability;
satisIactory results can be got with it, no matter what
kind oI input signal is used.
BeIore programming, the Iollowing items about
genetic algorithm to identiIy model should be
decided.
(1) Optimized model and object Iunction;
(2) Search range oI every gene;
(3) Operation parameters, such as number oI genes,
crossover and mutation probability;
(4) Coding and decoding method;
(5) Select method, such as proportional mode;
(6) Crossover and mutation method;
According to the characteristic oI thermal process in
power plant, the transIer Iunction model can be
written as Iollowing Iorms.
For the selI-balance system:
(8)
For the selI-imbalance system:
(9)
The object Iunction is calculated as:
(10)
Where, y is the output oI identiIication result, y
0
is the
output oI actual process.
The searching range oI the coeIIicients can be set
according to experience, T
n
0.01~10, t0~300,
K0.01~100.
For the multi-variables optimization, the string oI
binary gene is too long to use. With Iloat-point genes,
the number oI variables is not limited, coding and
decoding are not needed, the precision and eIIiciency
can be increased |2|. So, the Iloat coding method is
used in the improved genetic algorithm.
Rank-based selector model is used in this paper. First,
call the simulation Iunction to calculate Q values Ior
every individual and sort them in increase order with
MATLAB Iunction SORT. Second, assign a
probability table to the sorted population and then
create new population with proportional model.
Arithmetic crossover and uniIorm mutation is
adopted. In the crossover and mutation operation, the
population is divided into Iour parts. The Iirst part is
the optimal gene in the population, it is reserved into
next generation directly, the second part is good
genes, they cross with probability P
c1
and mutate with
probability P
m1
in scope S
m1
, the third part is common
genes, they cross with probability P
c2
and mutate with
probability P
m2
in scope S
m2
, the Iourth part is bad
genes, they cross with probability P
c3
and mutate with
probability P
m3
in scope S
m3
, where P
c1
P
c2
P
c3
,
P
m1
P
m2
P
m3
, S
m1
S
m2
S
m3
. The mutation probability
is a relative large value in the algorithm, P
m
0.5~0.9,
but the mutation area is limited by S
m
.
The MATLAB program oI sort select method is
shown as Iollow:
) 1 )...( 1 )( 1 (
) (
2 1
s T s T s T s
Ke
s G
n
ts

) 1 )...( 1 )( 1 (
) (
2 1
s T s T s T
Ke
s G
n
ts


2
0
) ( y y Q
sort value1 from small to big ana save it
[Jalue, aecinaex{sort (value1),
for i1.m, for f1.N
m
temppop(i,f)pop(aecinaex(i),f),
ena, ena
assign appropriate probability , then calculate
fitness.
PAm0.7, PA[m PAm m*(1 PAm)PAm{,
AAsum(PA), cfitnessPA /AA,
select new inaiviaual gene with proportional
moae,
for i2.m,
cfitness(i)cfitness(i-1)cfitness(i),
ena
for i1.m,
prana(1), inaex1,
while(p~cfitness(inaex)),inaexinaex1,ena
for f1.N
m,
newpop(i,f)temppop(inaex,f),
ena
ena
In the MATLAB program, m is the number oI genes,
Nm is the number oI optimized parameters, Nmn2
in model (8) and (9).
With MATLAB Iunctions c2d, d2c, ss2tI and tI2ss,
pulse transIer Iunction and transIer Iunction can
transIorm each other.
A simulation example is shown as Iollow.
Assume that the transIer Iunction oI thermal process
is:
) 30 1 )( 20 1 )( 10 1 (
6 . 0
) (
60
s s s
e
s G
s

The identiIication results are shown in Iigure 2 and


table 1.
Fig.2 model identiIication oI genetic algorithm
Tab.1 IdentiIication results
N T
1
T
2
T
3
K t
1 9.99 18.74 31.69 0.602 60.2
2 18.21 19.58 22.16 0.593 58.5
3 9.01 19.94 31.23 0.602 60.6
4 10.47 16.25 33.21 0.599 61.0
5 12.44 17.07 30.97 0.602 59.9
Ave 12.0 18.3 29.9 0.6 60
In Fig.2, curve 1 is random input signal that simulate
the operating data, curve 2 is object response, curve 3
is identiIication result, and they overlap in the Iigure.
Table 1 shows us Iive identiIication results in
zero-pole Iormat. Because the time constant is large,
the high Irequency part oI system is not sensitive. The
pole value is not same at every time. With MATLAB
Iunction ZP2TF, the results can be transIerred into
polynomial Iormat.
4. NONLINEAR MODEL IDENTIFICATION
Generally, the nonlinear dynamic model can be setup
according to basic conservation principles |3|. To
ensure the validity, the model parameters must be
modiIied according to operating data in power plant.
Genetic algorithm is a kind oI valid method to
identiIy or modiIy the parameters oI nonlinear model.
To show the identiIication process, a simple nonlinear
dynamic model oI boiler superheater is given here.
Assume that the system in Iigure 3 is an approximate
superheater.
Fig.3 the principle drawing oI superheater
Energy conservation Ior steam, metal and gas:
2 2 1
2
Q H D H D
at
aH
M
s s s s
s
s
(11)
2 1
Q Q
at
m
aT
m
C
m
M (12)
1
)
2 1
(
2
Q
g
T
g
T
pg
C
g
W
at
g
aT
pg
C
g
M (13)
Heat transIer equation:
) (
1
A
1 m
T
g
T Q (14)
) (
2
A
2 s
T
m
T Q (15)
Where,
M
s
, M
g
, M
m
: mass oI steam, gas and metal;
C
pg
, C
m
: speciIic heat oI gas and metal;
D
s
, W
g
: mass Ilow oI steam and gas;
H
s1
, H
s2
: enthalpy oI steam at inlet / outlet;
T
g1
, T
g2
: temperature oI gas at inlet / outlet;
T
m
: metal temperature;
T
s
, T
g
: average temperature oI steam and gas;
A
1
, A
2
: convection heat transIer coeIIicient;
65 . 0
2 1 1
A
g
W K K
,
8 . 0
4 3 2
A
s
D K K
The model can be programmed as S-Iunction oI
Simulink, then the superheater system (include spray
water system) can be conIigured in Simulink. In the
model, some parameters such as K1, K2, K3, K4
mainly aIIect static outputs oI the model, most oI
them can be calculated according to design data;
some parameters such as MmCm, Ms, MgCpg mainly
aIIect dynamic characteristics oI the model, they
should be determined by means oI identiIication. The
Simulink model oI last superheater system is shown
in Iigure 4.
Fig.4 Simulink model oI superheater system
The inputs are spray water pressure Pa, temperature
ta and valve position Ua2; steam pressure Ps,
temperature ts1 and Ilow Ws1 at inlet; gas
temperature Tg1 and Ilow Wg. The outputs are spray
water Ilow Wa, steam pressure Ps, temperature Ts2 at
outlet. HPT2 is the S-Iunction to calculate water or
superheat steam enthalpy; Tank is the S-Iunction oI
attemperator; Heater is the S-Iunction oI superheater.
The improved genetic algorithm is used to modiIy
parameters oI the dynamic model, such as MmCm,
Ms, MgCpg. These parameters should be deIined as
global variables when programming in MATLAB.
With Iunction |t,x,y|sim(]simulinkmodel,), the
Simulink model outputs can be got easily. The inputs
oI the Simulink model can be got with Simulink
Iunction From Workspace or From File. The
operating data or test data oI power plant should be
transIerred into MATLAB mat Iiles.
As an example, Iigure 5 and 6 show us the response
oI steam temperature when spray water Ilow increase
10. Curve 1 is model output, 2 is output oI actual
system. The parameters is not modiIied in Iigure5, it
is modiIied with improved genetic algorithm in
Iigure6.
Fig.6 response oI steam temp. aIter modiIication
It is shown in Iigure 5 and 6 that the dynamic model
output can overlap with output actual process aIter
the model parameters are modiIied with genetic
algorithm.
With MATLAB Iunction, the calculation speed is
slow iI the model is large. It can be improved iI
substitute S-Function with Windows DLL.
5. CONCLUSION
The identiIication methods Ior thermal process in
power plant are discussed. A kind oI improved
genetic algorithm is given, which can be used to
identiIy transIer Iunction or modiIy parameters oI
nonlinear model. It is shown by simulation research
and actual identiIication results that the genetic
algorithm-based identiIication method is oI good
practicability; satisIactory results can be got with it.
The visible advantage oI this method is that the
operating data can be used to identiIy transIer
Iunction model or modiIy nonlinear model
parameters; no special tests are needed in power
plant.
6. ACKNOWLEDGEMENTS
This paper is supported by doctor degree Ioundation
oI North China Electric Power University.
7. REFERENCE
|1| Liu C.L, et al, ,The Application OI Genetic
Algorithm In Model IdentiIication., IEEE
Tencon02, Beijing, 2002, pp. 1261~1264.
|2| Liu C.L, et al, ,Model IdentiIication OI Power
Plant Thermal Process Based On Genetic
Algorithm., Proceeaings of the Chinese Society
of Electrical Engineering, Vol.23, No.3, March,
pp. 170~174
|3| Liu C.L, et al, ,Nonlinear Boiler Model oI
300MW Power Unit For System Dynamic
PerIormance Studies., IEEE ISIE, Jol.2, Pusan,
Korea, 2001, pp. 1296~1300
Fig.5 response oI steam temp. beIore modiIication

Vous aimerez peut-être aussi