Professional Documents
Culture Documents
Energy Conversion and Management: Zhihuan Chen, Xiaohui Yuan, Hao Tian, Bin Ji
Energy Conversion and Management: Zhihuan Chen, Xiaohui Yuan, Hao Tian, Bin Ji
Energy Conversion and Management: Zhihuan Chen, Xiaohui Yuan, Hao Tian, Bin Ji
_
_
_
X
b11 0
0 b22
0 b32
0 b42
b51 b52
_
_
_
_
u 7
where
a11
ex eg
Ta
0
; a14
ey
Ta
0
; a15
eh
Ta
0
a21
kd
Td
2
; a22
1
Td
a31 ki bp kp
kd
Td
_ _
1
_ _
; a32 ki bp; a33 ki bp
a41
kd kp Td
Td Ty
; a42
1
Ty
; a43
1
Ty
; a44
1
Ty
a51
eqx eg ex
Ta
0
eqh
eqy
Ty eqh
kp
kd
Td
_ _
; a52
eqy
eqh Ty
;
a53
eqy
eqh Ty
a54
1
eqh
eqy
Ty
eqx ey
Ta
0
_ _
; a55
1
eqh
eqx eh
Ta
0
1
Tw
_ _
b11
1
Ta
0
; b22
kd
Td
2
; b32 ki ki bp kp
kd
Td
_ _
;
b42
kd kp Td
Ty Td
b51
eqx
eqh Ta
0
; b52
eqy
eqh Ty
kp
kd
Td
_ _
3. The improved gravitational search algorithm
3.1. Brief introduction of GSA
GSA is a newly developed stochastic search algorithm based on
the physical law of gravity and the law of motion. In this new ap-
proach, a set of agents has been proposed to nd optimum solution
by analogy of Newtonian laws. Agents are considered as objects
and their performance are measured by their masses, and these ob-
jects attract each other by the gravity force, while this force causes
a global movement of whole objects towards the objects with
Generator system Hydraulic system
Speed
governor
Servo-
mechannism
Penstock
Water
turbine
Generator
Load
Fig. 1. Block diagram of water turbine regulation system.
Fig. 2. Mathematical simulation of water turbine regulation system.
308 Z. Chen et al. / Energy Conversion and Management 78 (2014) 306315
heavier masses [23]. The following describes how GSA works for
the problem to be solved.
Usually, the target problem in the real world can be turned into
a minimization or maximization mathematical optimization prob-
lem, and then the GSA is used to solve this optimization problem.
Without loss of generality, in this section, we take minimization
problem as an example, which can be expressed as follows:
min f x x 2 R
n
8
where R
n
is the feasible domain in n dimensions of target problem.
Assumed there are N agents, the position of the ith agent is de-
ned as follows:
X
i
x
1
i
; . . . ; x
d
i
; . . . ; x
n
i
for i 1; 2; . . . ; N 9
where x
d
i
is the position of the ith agent in the dth dimension.
Masses are computed after calculating the tness of agents as
follows:
m
i
t
fit
i
tworstt
besttworstt
M
i
t
m
i
t
N
j1
m
j
t
_
_
10
where M
i
(t) and t
i
(t) represent mass and tness value of the ith
agent at time t, and best(t) and worst(t) respectively specify the
strongest and the weakest agent with regard to their tness route,
which is dened as follows (the denition of best(t) and worst(t)
for maximization optimization problem is on the contrary):
bestt min
j2f1;...;Ng
fit
j
t
worstt max
j2f1;...;Ng
fit
j
t
_
_
_
11
According to Newton gravitation theory, the force acting on the
ith agent from the jth agent at time t is calculated as follows:
F
d
ij
t Gt
M
i
t M
j
t
R
ij
t n
x
d
j
t x
d
i
t 12
where G(t) is gravitational constant at time t, R
ij
(t) is the Euclidian
distance between ith agent and jth agent (i.e. R
ij
(t) = kx
i
(t), x
j
(t)k
2
).
n is a small constant which is set for avoiding the divisor equal to
zero. (There usually uses R
ij
(t) instead R
2
ij
t in GSA because re-
searches show that the performance of R
ij
(t) is better than R
2
ij
t in
most cases.)
Based on the law of motion, the agent acceleration a
d
i
t is cal-
culated as follows:
a
d
i
t
j2kbest; ji
rand
j
F
d
ij
t
M
i
t
j2kbest; ji
rand
j
Gt
M
j
t
R
ij
t n
x
d
j
t x
d
i
t 13
where kbest is the set of rst K agents with the best tness value
and biggest mass, which is a function of time, initialized to K
0
at
the beginning and decreased with time, rand
j
is a random number
in the interval [0, 1]. (kbest is set to decrease the unnecessary calcu-
lation and enhance the efciency of optimization).
Thenthe next velocity of an agent is consideredas a fractionof its
current velocity added to its acceleration. Therefore, the next veloc-
ity and newly position of an agent can be computed as follows:
v
d
i
t 1 rand
i
v
d
i
t a
d
i
t 14
x
d
i
t 1 v
d
i
t x
d
i
t 15
where v
d
i
t and x
d
i
t are the velocity and position of ith agent at
time t in the dth dimension, respectively. rand
i
is a random number
between 0 and 1. It is to give a randomized feature to the search.
It must be pointed out that the gravitational constant G(t) is
important in determining the performance of GSA and dened as
follows:
Gt G
0
exp b
t
t max
_ _
16
where G
0
is the initial gravitational constant, b is a constant, t is the
current iteration, t_max is the maximum iteration.
The principle of GSA is shown in Fig. 3.
3.2. Improvements on GSA
A major of meta-heuristic optimization algorithms searching
the best solution due to the balance of two related concepts: explo-
ration and exploitation [37,38]. Exploration seeks to understand
the connectivity relationship of the search space, which is helpful
to the global optimal solution; exploration hunts for better optimal
solutions in adjacent area of the visited domain, which can
strengthen the convergence capability of local search. So an excel-
lent algorithm should improve the exploration ability in the rst
stage and then enhance the exploitation ability in the second stage
with the iterations increasing. In this section, some strategies
which can enhance the exploration and the exploitation ability of
algorithms are introduced.
3.2.1. Combination with particle swarm optimization
The reason why birds are able to nd foods mainly owes to the
capability of communication and memory among the ock. PSO
algorithm is the imitation of this capability. In the standard GSA,
the movement direction of each agent is determined by the total
force that other better agents act on it and lacking of communica-
tions between the agents. So in this paper, we try to improve the
searching ability of GSA by introducing the communication and
memory characteristics of PSO. The newly moving which is obedi-
ent to the law of gravity and received guide of memory and social
information is dened as follows:
v
d
i
t 1 r1 v
d
i
t a
d
i
t c1 r2 P
d
ibest
t x
d
i
t
c2 r3 G
d
best
t x
d
i
t 17
x
d
i
t 1 x
d
i
t v
d
i
t 18
Generate initial population
Evaluate the fitness for each agent
Update G, best and worst of the population
Calculate M and a for each agent
Update velocity and position
Meeting end of criterion
No
Return best solution
Yes
Fig. 3. Principle of GSA algorithm.
Z. Chen et al. / Energy Conversion and Management 78 (2014) 306315 309
where r
1
, r
2
and r
3
are random variables in the range [0, 1], c
1
and c
2
are learning genes in the range [0, 2], P
ibest
(t) is the best position that
ith agent has ever suffer until time t, G
best
(t) is the past global best
position in the agents at time t.
This new strategy which is a hybrid and universal format of GSA
and PSO has been conrmed to own a faster convergence speed
(i.e. strong exploitation capability) than standard PSO and GSA
[3]. Besides, the special moving mechanism of GSA in Eq. (17) pro-
vides a slower motion of agents in the search space and hence a
more precise search [23].
3.2.2. Elastic-ball strategy
It is normal to encounter the agents whose movement positions
are beyond boundary during the evolutionary process for many
stochastic algorithms. We often handle it with the following
strategy:
if x
d
i
t > Ubd x
d
i
t Ubd or
if x
d
i
t < Lbd x
d
i
t Lbd 19
where Ub(d) and Lb(d) are upper limit and lower limit in the dth
dimension.
All off-boundary agents are gathered in the boundary after such
processing, which will generate a huge force compelling other
agents to move forward boundary in accordance with law of grav-
ity and the uniform distribution of agents is disrupted, which is
great harmful to the global exploration, especially when there
are local optimums around the boundary. Therefore, a new treat-
ment named elastic-ball strategy is used in this paper. This novel
strategy imitates the characteristic that the elastic-ball reects
back excessive information if there is a barrier blocking its moving
direction and described as follows:
if x
d
i
t > Ubd outside x
d
i
t Ubd x
d
i
t Ubd outside
20
if x
d
i
t < Lbd outside Lbd x
d
i
t x
d
i
t Lbd outside
21
After dealing with the two steps, most agents rebound to the
feasible domain. The few rest agents against the boundary will
be reset position as follows:
if x
d
i
t > Ubdjjx
d
i
t < Lbd x
d
i
t rand Ubd
Lbd Lbd
22
This novel elastic-ball strategy has overcome the shortcomings
that pulling the off-boundary agents back to the boundary directly
maybe mislead evolution steps and enhanced the varieties of
searching agents at the same time, which is helpful to the global
exploration.
3.2.3. Mutation operator based on chaotic behavior
Premature phenomenon and local convergence are the com-
mon problems for many intelligence algorithms and one of the
effective approaches in the current is brought in a mutation oper-
ator to overcome these unhealthy performances. In this paper, a
chaotic mutation operator which is able to visit all points in a
specied range without any repeat is incorporated into the im-
proved algorithm. The chaotic mutation searches optimal by
means of regularity, ergodicity and intrinsic stochastic properties
of chaotic motion and can nd out global optimum in great prob-
ability. By mutating the best particle based on chaotic sequences,
the current best agents will leap out local tracking and looking for
a better available solution in global. In order to ensure the muta-
tion operator would not make the tness of populations worse,
only if obtained agent by mutation is better than the worst agent,
it will replace the worst one in next generation. The chaotic pro-
cedure is described as follows:
(1) Create a newparticle X
new
x
1
new
; . . . ; x
d
new
; . . . ; x
n
new
and let it
be the best particle Gbest in populations as: X
new
= Gbest.
(2) Convert the position of X
new
into a chaos vector e as follows:
ed
0
x
d
new
Lbd
Ubd Lbd
d 1; 2; . . . ; n 23
where e(d) is the element of chaos vector e in the dth dimension.
(3) Search the global area by Logistic map as follows:
ed
0
u ed 1 ed 24
where e(d)
0
is the Logistic map result of vector variable e(d), u is a
control variable in which u = 4, the mapping is in full state.
(4) Convert the obtained chaos vector e(d)
0
into the position as
follows:
X
new
d Lbd ed
0
Ubd Lbd d 1; 2; . . . ; n 25
(5) Calculate tness of obtained new position X
new
and compare
it with tness of the worst particle Gworst in the population.
If X
new
s is better, replace Gworst by X
new
; else remap and
relocate new Logistic vector with iterative calculation of
Generate initial population
Evaluate agents fitness and store Pbest and Gbest
Update best(t), worst(t), and M(t) in the population
Calculate gravitational constant and acceleration for each agent
Update the velocity and position for each agent
Meeting end of criterion
Yes
Elastic-ball strategy
is invoked
Chaotic mutation is conducted for the best agent
Yes
No
Beyond boundarys
Evaluate the fitness for each agent s new position
Update Pbest and Gbest in the population
Return best solution
No
No
Fig. 4. Flowchart of the newly IGSA algorithm.
310 Z. Chen et al. / Energy Conversion and Management 78 (2014) 306315
Eqs. (23) and (24) several times until the tness of X
new
is
better than the tness of Gworst or the mapping is covered
with every corner in the searching range.
Through chaotic mutation, the best agent has a bigger possibil-
ity to leap out the current local optimal domain if there are any
other better unsearched regions (i.e. strong exploration ability).
In addition, excellent characters of chaos movement such as
searching feasible areas without repeat are also contributed to glo-
bal exploitation.
By means of three strategies added in the algorithm, the im-
proved GSA (IGSA) is summarized as the following and is illus-
trated by the diagram in Fig. 4. In this work, IGSA will be used to
solve the model parameter identication problem and is applied
in parameter identication of WTRS. The performance of IGSA
and effectiveness of the improvements in this section will be dem-
onstrated through results in the later identication experiments.
Step 1: Initialization. Randomly initialize the agent position and
velocity.
Step 2: Fitness evolution. Calculate the tness of agents by their
initial position, storing current position of each agent as
the best history record position of the agent (i.e.
P
best
i
t x
i
t for t = 1) and position of best agent gbest(t)
(which owns the biggest tness value in the population)
as the best position in global (i.e. G
best
(t) = gbest(t) for
t = 1).
Step 3: Update best(t), worst(t), and M
i
(t) for i = 1, 2, . . . , N.
Step 4: Calculate gravitational constant in the current iteration
and acceleration for each agent.
Step 5: Update agents velocity and position with Eqs. (17) and
(18).
Step 6: Judge whether the new position of the agent is beyond
the boundary. If the new position is against the bound-
ary, elastic-ball program is invoked.
Step 7: Evaluate the tness in accordance with each agents new
position.
Step 8: Compare the obtained tness of new position x
i
(t + 1)
with tness of P
best
i
t while i changes from 1 to N. If
x
i
(t + 1) has a better tness value, replace position of
P
best
i
t by x
i
(t + 1).
Step 9: Compare tness of the agents with G
best
(t). If there is an
agent has a better tness than tness of G
best
(t), G
best
(t)
will be replaced by the position of the agent.
Step 10: Chaotic mutation is conducted for the best agent.
Step 11: Repeat Step 3 to Step 10 until the stop criteria reached.
4. The parameter identication strategy
For a system with known model structure but unknown
parameters, the parameter identication problem can be
converted into an optimization problem. The unknown
parameters vector for WTRS is usually set as a particle in swarm
or a gene in chromosome and a performance function measuring
how well the model response ts the system response is built to
optimize.
4.1. Objective function based on WTRS system
The model of WTRS has been illustrated by Fig. 2. The turbine
speed x, guide vane opening y and water torque m
t
are observed
output state variables. Although three outputs are selected as state
output variables, their contribution and importance in solving the
problem of parameter identication are different. At the same var-
iation of parameters, the more signicant the output of identied
model deviates from that of original system, the more important
the output in the objective function will be. And the weight for this
output is heavy. In this way, weights of different errors outputs are
designed, according to the signicance of deviation. The improved
object function is dened as:
C
IOF
h
N
k1
n
j1
wjz
j
k
^
z
j
k
2
26
where parameter vector h kp ki kd Ty Tw Ta
0
eg , out-
put of actual system z x y m
t
, z
_
x
_
y
_
m
_
t
is simulated
output of identied model, N is the number of samples, n is dimen-
sion of system output vector, in this work, n = 3, the weight vector
w w
1
w
2
w
3
.
The weights are calculated according to following steps:
(1) Set value of vector parameters h
i
, i = 1, 2, . . . , m (m is the
dimension of parameter vector h) in the WTRS system, and
obtain system output z
j
(k), k = 1, 2, . . . , N, j = 1, 2, . . . , n.
(2) Loop A: j = 1:n.
Loop B: i = 1:m.
Vary the ith parameter, h
new
= h
i
(1 + D%), and obtain sys-
tem output ^z
ji
k.
Calculate w
ji
kz
ji
^z
ji
k
2
.
The jth weight w
j
= average(w
ji
).
End Loop B.
End Loop A.
Original
system
Original
system
Measured
output
x
Estimated
system
Estimated
system
Fitness
Evaluator
Fitness
Evaluator
y
t
m
y
x
t
m
Simulated
output
IGSA-based
identifier
IGSA-based
identifier
System input
Identified parameter
(
IOF
C
Fig. 5. Diagram of IGSA based WTRS parameter identication.
Table 1
Transfer coefcients of turbine under two running condition.
Working condition Transfer coefcients in the water turbine system
ex ey eh eqx eqy eqh
No-load 1.0567 0.9080 1.4191 0.0574 0.7887 0.4571
Load 1.4673 0.7713 1.7179 0.4901 0.8184 0.7257
Z. Chen et al. / Energy Conversion and Management 78 (2014) 306315 311
4.2. Parameter identication strategy
As shown in Fig. 5, the original and estimated systems are sup-
plied with a same excitation inputs and their outputs are given as
inputs to the tness evaluator, where the tness is calculated. The
tness function C
IOF
^
h is then used by IGSA-based identier to
identify the unknown parameter vector
^
h. By minimizing of tness
function through IGSA, the outputs of estimated system approxi-
mates to the outputs of original system, while the unknown
parameters trend to be equal with real values.
Parameter identication accuracy is measured by parameter er-
ror (PE):
PE
jh
i
^
h
i
j
h
i
100% i 1; 2; . . . ; m 27
and average parameter error (APE):
APE
1
m
m
i1
jh
i
^
h
i
j
h
i
100% 28
where h
i
is the parameter elements of h in original system,
^
h
i
is the
parameter elements of
^
h in estimated system, m is the size of h.
5. Experiments and results analysis
In this section, the WTRS is simulated in MATLAB, and the pro-
posed IGSA is applied to identify the parameters of simulated sys-
tem. The model of WTRS is illustrated in Fig. 2. Seven key
parameters are chosen to be estimated in simulation experiments
which are kp, ki, kd, Ty, Tw, Ta
0
and eg.
Inexperiments, simulationmodel of WTRS are excitedunder two
condition, loadconditionandno-loadcondition. Under no-loadcon-
dition, step disturbance of speed is employed to excite the system,
andunder loadcondition, stepdisturbanceof loadis employedtoex-
cite the system. The parameters of simulated model are set values
adopted in a Chinese hydroelectric station as follows:
Table 2
Comparison of average identication results of different methods under no-load condition.
Identied parameters h
i
System real value Average of identied parameters (30 trials)
PSO GSA IGSA
^
h
i
PE ^
h
i
PE ^
h
i
PE
kp 5.5912 6.0418 0.0806 5.9529 0.0716 5.5919 0.0007
ki 1.0611 0.0749 0.9295 0.5954 0.4477 1.0596 0.0033
kd 3.2800 4.1630 0.2692 3.5520 0.0921 3.2805 0.0006
Ty 0.1000 0.1472 0.4724 0.1180 0.1917 0.1000 0.0012
Tw 1.5000 1.4712 0.0197 1.4708 0.0283 1.5000 0.0004
Ta
0
12.000 17.159 0.4299 14.035 0.1762 12.005 9.0E-5
eg 0.4433 0.5071 0.1613 0.4148 0.0665 0.4431 0.0016
Table 3
Mean best cost and mean APE of 30 times under no-load condition.
PSO GSA IGSA
Mean best cost 86.319 20.782 0.0117
Mean APE 0.3375 0.1534 0.0013
0 20 40 60 80 100
0
500
1000
1500
2000
2500
3000
3500
4000
C
o
s
t
Iteration
PSO GSA IGSA
Fig. 6. Comparison of average iteration process under no-load condition.
20 40 60 80 100
0
20
40
60
80
100
120
140
160
180
200
C
o
s
t
Iteration
PSO GSA IGSA
Fig. 7. Local magnication of average iteration process under no-load condition.
0 5 10 15 20 25 30
0
1
2
Time
S
p
e
e
d
original system estimated system
0 5 10 15 20 25 30
-10
0
10
Time
T
o
r
q
u
eoriginal system estimated system
0 5 10 15 20 25 30
0
5
10
Time
G
u
i
d
e
v
a
n
e
original system estimated system
Fig. 8. Comparison of system outputs using IGSA under no-load condition.
312 Z. Chen et al. / Energy Conversion and Management 78 (2014) 306315
The gains of PID in steady-state working condition are 5.5912,
1.0611, and 3.2800 (i.e. kp = 5.5912, ki = 1.0611, kd = 3.2801). The
characteristic parameter of penstock in the hydroelectric station
is calculated to approximate equal to 1.5 (i.e. Tw = 1.5). The inertia
time constant and the adjusting parameter of generator and load
are equal to 12 and 0.4433 respectively through the measurement
(i.e. Ta
0
= 12, en = 0.4433). The major servomotor time constant is
nearly approach 0.1 according to the historical data (i.e. Ty = 0.1).
The feedback coefcient and the differential time constant is set
to 0.04 and 0.28 according to the ever experience (i.e. bp = 0.04,
Td = 0.28).
The parameters of water turbine under different conditions are
shown in Table 1.
Based on the above discussion, the parameter vector (i.e.
h kp ki kd Ty Tw Ta
0
eg ) in original system is set as
h 5:5912 1:0611 3:2801 0:1 1:5 12 0:4433. The simu-
lation time is set to be 30 s which is long enough to make sure the
system change to be stable from a transient process. The sampling
time is set to be 0.01 s which is fast enough to capture system
dynamic process. The outputs in vector z, which contains turbine
speedx, guide vane opening y andturbine torque m
t
, are all sampled.
5.1. Comparison of identication method under no-load condition
In this part of experiments, IGSA, GSA and PSO have been em-
ployed to identify the parameters in the dynamic model of WTRS.
A step disturbance of given speed is adopted to excite the system.
The model described in Section 2 is simulated as the original sys-
tem, and experiments of parameter identication are conducted.
In simulation, to perform fair comparison in tness evaluation,
population size of PSO, GSA and IGSA are all 80. The maximum
generation is set to be 100 in the three algorithms. For GSA,
G
0
= 30, b = 10; For PSO, w = 0.6, c1 = c2 = 2; For IGSA, c1 = c2 = 2,
G0 = 30 and b = 10. In order to overcome the randomness of three
heuristic algorithms, 30 trials are tried and average results are
obtained.
Table 4
Comparison of average identication results of different methods under load condition.
Identied parameters h
i
System real value Average of identied parameters (30 trials)
PSO GSA IGSA
^
h
i
PE ^
h
i
PE ^
h
i
PE
kp 5.5912 8.7602 0.5668 6.4076 0.1538 5.6170 0.0060
ki 1.0611 1.4695 0.3938 1.1587 0.0988 1.0654 0.0052
kd 3.2800 4.1421 0.3530 3.5340 0.0928 3.2827 0.0044
Ty 0.1000 0.0500 0.5416 0.0950 0.1591 0.0997 0.0122
Tw 1.5000 1.4294 0.0471 1.4680 0.0240 1.4981 0.0021
Ta
0
12.000 17.685 0.4737 13.448 0.1262 12.043 0.0046
eg 0.4433 0.4312 0.0377 0.4482 0.0415 0.4458 0.0082
Table 5
Mean best cost and mean APE of 30 times under load condition.
PSO GSA IGSA
Mean best cost 0.5528 0.0923 0.0017
Mean APE 0.3448 0.0995 0.0061
0 20 40 60 80 100
0
5
10
15
20
25
C
o
s
t
Iteration
PSO GSA IGSA
Fig. 9. Comparison of average iteration process under load condition.
20 40 60 80 100
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
C
o
s
t
Iteration
PSO GSA IGSA
Fig. 10. Local magnication of average iteration process under load condition.
0 5 10 15 20 25 30
0
0.1
0.2
Time
T
u
r
b
i
n
e
s
p
e
e
d
original system estimated system
0 5 10 15 20 25 30
-2
-1
0
Time
T
u
r
b
i
n
e
t
o
r
q
u
e
original system estimated system
0 5 10 15 20 25 30
-1.5
-1
-0.5
0
Time
G
u
i
d
e
v
a
n
e
original system estimated system
Fig. 11. Comparison of system outputs using IGSA under load condition.
Z. Chen et al. / Energy Conversion and Management 78 (2014) 306315 313
Comparing identication accuracy of different methods under
no-load, PE is used and listed in Table 2. Simulation results consid-
ering best cost and APE are listed in Table 3. In Tables 2 and 3, it is
seen that compared with GSA and PSO, IGSA achieve better param-
eter identication accuracy. In Table 2, the mean cost and mean
APE are as small as 0.0117 and 0.0013, respectively, much smaller
than those achieved by other methods, which conforms the valid-
ity of improvement measures added.
The convergence of algorithms is compared in Fig. 6, which
exhibits the average convergence of 30 times, it is seen that IGSA
could converge on the optimal quickly compared with other meth-
ods. Although in the incipient evolution period, the convergence of
PSO algorithm is better, but 20 iteration later, Fig. 7 which is the
local magnication gure of Fig. 6 exhibits IGSA has a faster con-
vergence speed while the other algorithms are tracking in a local
optimal soon.
Fig. 8 shows the estimated system outputs of WTRS obtained by
using the average identied parameters with IGSA, and then com-
pared with original system outputs, where guide vane opening,
turbine torque and turbine speed are compared. It is obvious origi-
nal curves and estimated curves are very closely, which means
parameter identication is effective and obtained higher accuracy.
5.2. Comparison of identication method under load condition
In order to verify the validity of IGSA in parameters identica-
tion of WTRS, different running conditions of WTRS is considered.
In this part, the system is under load condition, and a step distur-
bance of load is employed to excite the system. The parameters
regarding to water turbine under load condition is shown in
Table 1, the parameter vector set in original system is not changed,
and parameters of PSO, GSA and IGSA are also keep unchanged.
Identication experiments are repeated 30 times, average indices
of results are taken into consideration
Tables 4 and 5 show the identication accuracy achieved by dif-
ferent algorithms, where average PE, APE and best cost are taken
into consideration. The results show clear that compared with
PSO and GSA, IGSA performs best on all indices, which means the
proposed IGSA is effective. The average convergence process of
PSO, GSA and IGSA are compared in Figs. 9 and 10, which shows
that IGSA possesses excellent ability in obtaining optimal value
of cost function compared with other algorithms. Identied out-
puts using the average parameters based on IGSA are compared
with the original system outputs in Fig. 11, which exhibits the
identied system meet original system perfect.
6. Conclusion
In this paper, the problem of parameter identication of WTRS
is studied. A simulation model associated with engineering experi-
ence is brought in. An improved gravitational search algorithm is
proposed to the parameters identication of WTRS. Simulation re-
sults are provided to validate the effectiveness of the identied
method. It is shown that IGSA is capable of solving the problem
of parameters identication. Comparing to GSA and PSO, IGSA per-
forms the best with high accuracy and stability. Meanwhile, the
identication method is not focus on the specic input signal,
which is easy to implement in the system simulation and conve-
nient to transplant into other parameter identication problems.
Acknowledgment
The authors gratefully acknowledge the nancial supports from
National Natural Science Foundation of China under Grant No.
51379080.
References
[1] Kishor N, Singh SP, Raghuvanshi AS. Dynamic simulations of hydro turbine and
its state estimation based LQ control. Energy Convers Manage
2006;47(18):311937.
[2] Fang H, Chen L, Shen Z. Application of an improved PSO algorithm to optimal
tuning of PID gains for water turbine governor. Energy Convers Manage
2011;52(4):176370.
[3] Li C, Zhou J. Parameters identication of hydraulic turbine governing system
using improved gravitational search algorithm. Energy Convers Manage
2011;52(1):37481.
[4] Jiang C, Ma Y, Wang C. PID controller parameters optimization of hydro-
turbine governing systems using deterministicchaotic-mutation evolutionary
programming (DCMEP). Energy Convers Manage 2006;47(9):122230.
[5] Fang J, Zheng D, Ren Z. Computation of stabilizing PI and PID controllers by
using Kronecker summation method. Energy Convers Manage
2009;50(7):18217.
[6] Eker I. Governors for hydro-turbine speed control in power generation: a SIMO
robust design approach. Energy Convers Manage 2004;45(13):220721.
[7] Khodabakhshian A, Hooshmand R. A new PID controller design for automatic
generation control of hydro power systems. Int J Electr Power
2010;32(5):37582.
[8] Li C, Zhou J, Xiao J, Xiao H. Hydraulic turbine governing system identication
using T-S fuzzy model optimized by chaotic gravitational search algorithm.
Eng Appl Artif Intel 2013;26(9):207382.
[9] Kishor N, Singh SP. Simulated response of NN based identication and
predictive control of hydro plant. Expert Syst Appl 2007;32(1):23344.
[10] Kishor N. Nonlinear predictive control to track deviated power of an identied
NNARX model of a hydro plant. Expert Syst Appl 2008;35(4):174151.
[11] Xiao Z, Wang S, Zeng H, Yuan X. Identifying of hydraulic turbine generating unit
model based on neural network. In: Proceeding of IEEE international intelligent
system design and applications, Jinan, China; January 2006. p. 1137.
[12] Gao L, Dai Y, Xia J. Parameter Identication of hydro generation system with
uid transients based on improved genetic algorithm. In: Fifth international
conference on natural computation (ICNC), TianJian, China; August 2009. p.
398402.
[13] Liu C, Liu N, Sun X, Cui J. The research and application on parameter
identication of hydraulic turbine regulating system based on particle swarm
optimization and uniform design. In: Proceedings of international conference
on computer science and information technology (ICCSIT), Chengdu, China;
July 2010. p. 6058.
[14] Bai J, Xie A, Yu X, Zhou L. Simulation model of water turbine speed control
system and its parameters identication based on resilient adaptive particle
swarm optimization algorithm In: Proceedings of IEEE AsiaPacic power and
energy engineering conference (APPEEC), Chengdu, China; March 2010. p. 14.
[15] Kou P, Zhou J, Li C, He Y, He H. Identication of hydraulic turbine governor
system parameters based on bacterial foraging optimization algorithm. In:
Sixth international conference on natural computation (ICNC), Yantai, China;
August 2010. p. 333943.
[16] Kocaarslan I, am E. Fuzzy logic controller in interconnected electrical power
systems for load frequency control. Int J Electr Power 2005;27(8):5429.
[17] Yuan X, Cao B, Yang B, Yuan Y. Hydrothermal scheduling using chaotic hybrid
differential evolution. Energy Convers Manage 2008;49(12):362733.
[18] Yuan X, Su A, Nie H, Yuan Y, Wang L. Application of enhanced discrete
differential evolution approach to unit commitment problem. Energy Convers
Manage 2009;50(9):244956.
[19] Yuan X, Su A, Nie H, Yuan Y, Wang L. Unit commitment problem using
enhanced particle swarm optimization algorithm. Soft Comput
2011;15(1):13948.
[20] Yuan X, Wang L, Yuan Y, Zhang Y, Cao B. A modied differential evolution
approach for dynamic economic dispatch with valve-point effects. Energy
Convers Manage 2008;49(12):3353447.
[21] Yuan X, Wang Y, Xie J, Qi X, Nie H, Su A. Optimal self-scheduling of hydro
producer in the electricity market. Energy Convers Manage
2010;51(12):252330.
[22] Tabatabaei SM, Vahidi B. Bacterial foraging solution based fuzzy logic decision
for optimal capacitor allocation in radial distribution system. Electr Power Syst
Res 2011;81(4):104550.
[23] Rashedi E, Nezamabadi-Pour H, Saryazdi S. GSA: a gravitational search
algorithm. Inform Sci 2009;179(13):223248.
[24] Rashedi E, Nezamabadi-Pour H, Saryazdi S. BGSA: binary gravitational search
algorithm. Nat Comput 2010;9(3):72745.
[25] Behrang MA, Assareh E, Ghalambaz M, Assari MR. Forecasting future oil
demand in Iran using GSA (Gravitational Search Algorithm). Energy
2011;36(9):564954.
[26] Gven U, Snmez Y, Duman S, Yrkeren N. Combined economic and
emission dispatch solution using gravitational search algorithm. Sci Iran
2012;19(6):175462.
[27] Duman S, Gven U, Snmez Y, Yrkeren N. Optimal power ow using
gravitational search algorithm. Energy Convers Manage 2012;59(1):8695.
[28] Duman S, Snmez Y, Gven U, Yrkeren N. Optimal reactive power dispatch
using a gravitational search algorithm. IET Gener Transm Distrib
2012;6(6):56376.
[29] Tsai HC, Tyan YY, Wu YW, Lin YH. Gravitational particle swarm. Appl Math
Comput 2013;219(17):910617.
314 Z. Chen et al. / Energy Conversion and Management 78 (2014) 306315
[30] Khajehzadeh M, Taha MR, ElShae A. A modied gravitational search
algorithm for slope stability analysis. Eng Appl Artif Intel 2012;25(8):158997.
[31] Li C, Li H, Kou P. Piecewise function based gravitational search algorithm and
its application on parameter identication of AVR system. Neurocom
2014;124:13948.
[32] Kumar JV, Kumar DM, Edukondalu K. Strategic bidding using fuzzy adaptive
gravitational search algorithm in a pool based electricity market. Appl Soft
Comput 2013;13:244555.
[33] Han X, Chang X. A chaotic digital secure communication based on a modied
gravitational search algorithm lter. Inform Sci 2012;208:1427.
[34] Ghasemi A, Shayeghi H, Alkhatib H. Robust design of multi-machine power
system stabilizers using fuzzy gravitational search algorithm. Int J Electr
Power 2013;51:190200.
[35] Mallick S, Ghoshal SP, Acharjee P. Optimal static state estimation using
improved particle swarm optimization and gravitational search algorithm. Int
J Electr Power 2013;52:25465.
[36] Cheng X, Zhang J. Water turbine automatic regulation. Beijing: China
Waterpower Press; 2010 [in Chinese].
[37] Alba E, Dorronsoro B. The exploration/exploitation tradeoff in dynamic cellular
genetic algorithms. IEEE Trans Evolut Comput 2005;9(2):12642.
[38] Liu SH, Mernik M, Hrncic D, C