2012 China Interational Conference on E[ectricity Distribution (C[CED 2012) Shanghai, 5-6 Sep.

20 [2
L-bJ/JLIN LCnIICIu8CdCn bCII-IunIn PI WIIhICuIuI
ICIWCI8
YU Lei
NCEP university-China
yulei@16 3.com
ZHANG Jianhua
NCEP university-China
E-mail address
J[ANG Cheng
NCEP university-China
jc 2002@ [63 .com
Abstract-Power distribution system installed with D
STATCOM (Distribution Static Synchronous Compensator)
is a large-scale nonlinear, multivariable system. The
traditional PI controller is not content to modern control
requirement because of its fxation parameters. In this paper,
node voltage control strategy of D-STATCOM based on self
adaptive PI controller with neural network is proposed, and
the controller structure is designed. In dynamic process, PI
parameters of the node voltage controller are on-line
adjusted with neural networks. Thus, this controller is good
at robustness and adaptability. The simulation model of the
control system is built in the MATLAB dynamic simulation
platform, based on which the regulating process of note
voltage is simulated. The result verifed the adaptability and
feasibility of the proposed control strategy
Keywords-D-STATCOM; Self-tuning PI control; Neural
Network; Note voltage control
INTRODUCTION
As a compensator device installed in parallel to the
transmission line, D-ST ATCOM was introduced to
reactive power compensation feld with faster response
and higher control precision compared to the traditional
reactive power compensation device [[].
Node voltage control is one of the main fnctions of D
ST A TCOM, and choosing a good control method is
pivotal to show its effect. P[ control is wildly applied to
D-STATCOM because of its perfecting theory, clear
concept, convenient adjustment and easy realization [2].
However the power distribution system installed with D
ST A TCOM is a large-scale nonlinear, indetermination
and multivariable systems; the structure and parameters of
system will change especially when systems failure and
load disturbance happens accidentally. Therefore the
traditional P[ controller has a limited application in some
cases because of its non-adaptive parameters and linearity.
[n order to solve this problem, a self-tuning P[ control
strategy with fzzy logic is proposed [3], due to the P[
parameters are regulated on line using fzzy logic, the
controller with the proposed strategy is adaptive and
robust. But the fzzy variable grade and fzzy rule are
diffcult to confr in some cases. Besides, choosing the
membership fnction is devoid of academic guidance. [n
paper [4], the indirect voltage control method for
ST A TCOM based on self-tuning P[ with neural network
is proposed. This control method has a limited application
in nonlinear systems, even if it could regulate P[
parameters on line, because its PI parameters tuning is
unattached and linear.
By completely considering the characteristic of the
power distribution system and the control method, this
paper propose a node voltage control strategy for D
STATCOM based on self-adaptive PI control with neural
network. Due to the PI parameters could be regulated on
line in the dynamic process, the node voltage of the
power distribution system is fexibly, real-timely and
quickly controlled using the proposed controller in this
paper.
/>B/IYDISTRIBUTION >>BY
->/LIB
D-STATCOM is a device installed in parallel to the
distribution system. Its principle can be described as: it
inject the node Bus a current whose amplitude is
changeable independent of the node Bus voltage, and
phase is perpendicular to the node Bus voltage's phase.
Changing the curent's amplitude is equivalent to change
node Bus voltage, with which it can control the Bus
voltage. However in practice the current is not strictly
perpendicular to the node Bus voltage, the difference is
used for compensating the loss of D-ST A TCOM.
Fig. [ illustrates a power distribution system installed
with D-STATCOM [ 5].
LS|OUOD SySBR
Figure 1. A small power distribution system with D
STATCOM
Where H_ represents a voltage source, D-STATCOM is
installed in parallel to the distribution system Bus B.
Fig.2 illustrates the control block diagram of D
STATCOM in power distribution systems. Where
'voltage calculate' is employed to give the current voltage
amplitude by the metrical voltage of Bus B. The
difference between the current and given voltage
amplitude is compensated by PI controller whose output
CICED2012 Session 3 Paper No CP05 50 Page 114

2012 Chma [nteratlOna[ Conference on E[ectrIclty DIstrIbutIOn (C[CED 2012) Shanghai, 5-6 Sep. 20 [2
is applied to drive the SHPWM as a modulation ratio.
C0U|llHQ I|BH50|0|
`i-
L-
'nbc
Figure 2. Control block diagram of D-STATCOM
The control rule of PI regulator is shown in (1).
M =
(K
p
l
+
l
)(U
re! -
U) (I)
Taking into account the voltage margin of the main
circuit, especially the switching devices can not be too
large. [f the DC link voltage is too high, it will affect the
safety of the main circuit, especially the switching devices.
So the DC capacitor voltage control is needed in order to
ensure that the switching devices are in the range of
security in a variety conditions. From the literature [ 6],
changing the phase angle of voltage emitted by D
ST A TCOM could change the DC bus voltage. Therefore
in this paper | is selected as control variable, regulating
which the DC bus voltage could be limited to a desired
range. The control rule of Be is shown as:
(c
=
(K
p2
+

2
)(V;c-re! - VdJ - ( Y
(2)
Where K
p
and Ki is the proportion ratio and integral
ratio of PI controller, respectively. U
re
! and U is the
given and current values of Bus B voltage amplitude,
respectively. Vde-re
j
and Vdc is the given and current
values of DC link voltage, respectively. M IS
the modulation ratio and | is initial phase of PWM.
>-TUNING LIIIYI\/
IYI>
Artifcial Neural Networks because of its strong non
linear ftting and relatively simple learing algorithm has
a growing number of applications in industrial controls.
[n this paper, two neural networks are employed to
achieve the adaptive PI control. The one is applied to
model the power system with a D-STATCOM, and the
other to adjust the parameters of PI controller for D
STATCOM on line. The control system structure is
shown in Fig.3.
u
Figure 3. Control block diagram of neural network
self-tuning PI
where NNI is the identifcation network and NNT is the
tuning network.
Neural Network Identifcation (NNI)
When applying neural network in system identifcation,
the frst thing we concered is the architecture of the
neural network, and the pivotal step is to decide the input
vector of the neural network. Afer that, frst option is the
feed forward neural network (FNN) and the back
propagation training algorithm to solve the system
identifcation problem.
In this paper, a neural network is employed to model
the power system with D-STATCOM. The neural
network is of the feed forward type and has three layers
consisting of an input layer with twelve inputs, a single
hidden layer with sigmoid activation fnctions consisting
of twenty neurons and an output layer with two outputs.
The inputs of the NNI are Be , M, U and Vde at
times t, t-I and t-2 and outputs U and Vdc at time t+ I .
For training NNI, inputs of plant were perturbed using
Pseudorandom Binary Signals (PRBS). The amplitudes of
PRBS signals applied to U , Vde , M and Be are
limited to 10% of their main signal. The network structure
is shown in Fig. 4.
Power
U t
distribution
system
A
U(t + 1)
A
V t+ 1)
Figure 4. Structure of identification network
Where DEL represents 'delay'.
In this paper, the objective fnction including a
CICED2012 Session 3 Paper No CP05 50 Page2/4

2012 Chma InteratIOnal Conference on ElectrIcIty DIstrIbutIOn (CICED 2012) Shanghai, 5-6 Sep. 20 I 2
discount factor for NNI is shown as:
Jr = {A[(t) -
U(t)]
+ (1 - A) [dc(t) - Vdc(t)]}
2
(3)
Where Z (0.5 < Z< 1) is a discount factor
Based on the Steepest Descent Algorithm, the adjusting
rule ofNNI is shown as follows [7]:
aJr
w
ICt+l
)=w1Ct) -7I-
(4)
8
w
ICt)
aJr
w2(t+l)=w2(t)) -7I
-
(5)
8
w2(t)
Where W
I
is the 2 ? 2 weight matrix belonging to hidden
and output layer, W
2
is the 20 ? 12 weight matrix
belonging to input and hidden layer, 71 (0 < 7
1
< 1) is the
learing rate.
When the training is complete, the plant outputs
A A
U(t +
1) and
Vdc(t +
1) , and the Jacobian parameters i.e.
au /aM and av d/ aec can be obtained which are
applied in the PI parameters tuning network.
Neural Netork Tuning CNNT)
In this paper, a neural network is employed to adjust
the PI parameters on line. The neural network is of the
feed forward type and has three layers consisting of an
input layer with six inputs, a single hidden layer with
sigmoid activation fnctions consisting of twenty neurons
and an output layer with four outputs.
The inputs of the NNT are U and Vdc at times t, t-I
and t-2 and outputs
K
p
I'
K
p
2'
K
il'
K
i2 at time t+ 1.
The NNT is training on line using the differences between
the reference values and prediction values at time t+ I
outputting by NNI, and the Jacobian parameters i.e.
au /aM and av d/ aec at time t indirectly given by NNI.
The network structure is shown in Fig. 5.
Figure 5. Structure of PI parameters tuning network
In this paper, the objective fnction including a
discount factor for NNT is given as:
J1 = {r[(
U
rej -'(t +
1)]
+
(1 -r)[(Vdcrej - c(t +
1)]}
2
(6)
Where r ( 0.5 < r < I ) is a discount factor
Based on the Steepest Descent Algorithm and PI
algorithm, the adjusting rule ofNNT is shown as follows
[ 8 ]:
w3
Ct
+l
)= w3
Ct) -
A A
(
a
Jr aU(t+l)
+
a
Jr aVdc(t+l)
) (8)
72
a(t+l)
a
w
4
c
t)
a

Jt+l)
8
w
4Ct)
Where W 3 is the 4 ? 20 weight matrix belonging to
hidden and output layer, W
4
is the 20 ? 6 weight matrix
belonging to input and hidden layer, '2 ( 0 < 7
2
< 1) is
the learing rate.
>B\/JIIH>\>
In order to verif the adaptability and feasibility of the
proposed controller, the simulation model shown in Fig.l
is built in MA TLAB. The simulation parameters are
shown in Table 1.
Table 1. Parameters of the simulation model
Parameters
Bus H_ (u)
DC /tnkvo/tage(u)
equtva/enttnductanceL, (u)
equtva/entreststance Rs (u)
synchront:attonangu/arve/octty
/earntngrate 11
/earntngrate 12
dtscountjactor A
dtscountjactor r
valuc
lLO
0.5
0.0058
0.06
314
0.1
0.3
0.7
0.7
Fig.6 and Fig.7 are the regulation process of node Bus
voltage when the power distribution system suffer small
disturbance. Evidently, the self-tuning PI controller is
superior to the traditional ones in the response speed.

c
1.2
c.a
D

.C
"
A
.7
1rad|||Onal ~| Onl|Ol
I
I
SO||-Iun|ng Pl nko|

Time(sec)
CICED2012 Session 3 Paper No CP05 50 Page3/4

2012 Chma [nteratlOna[ Conference on E[ectrIclty DIstrIbutIOn (C[CED 2012) Shanghai, S-6 Sep. 20 [2
Figure 6. Dynamic regulation process when node Bus
voltage increases
1.A
1Z
,
0.8
i

"
0.4
.
FOBQIIVO Pl CODIfOl
-
Se|f-!uning Pl OCDl

Time(see)
Figure 7. Dynamic regulation process when node Bus
voltage decreases
[n order to verif the adaptability of the proposed
controller, the value R shown in Table [ is decreased to
O.OS. Fig. 8 and Fig.9 are the regulation process of node
Bus voltage when the power system suffer small
disturbance. Obviously, the proposed self-tuning P[
controller is more adaptive than the traditional ones with
faster response speed and less overshoot

12 _
` \
n
c6

.O
m
c
ICODl

l U CDDI|D|

Time(see)
Figure 8. Dynamic regulation process when node Bus
voltage increases
1
1Z J|BOIODBl P| clm|
1 .-

/
BlUHIH Pl COD/Ol
U.A
0.2

Tlme(see)
Figure 9. Dynamic regulation process when node Bus
voltage decreases
LIIL\>II
Simulation studies have been carried out to evaluate the
proposed controllers. Conclusions are shown as follows:
a. The proposed self-tuning PI controller is speedier
than the traditional PI controller in the response
speed.
b. When the system parameters change, the traditional
PI controller can not maintain the dynamic
performance which is originally designed. However
the proposed controller is not sensitive to alterable
parameters, with greater adaptability and robustness
when the operation condition changes.
IL>
[1] LTU Yi-ting. Research on Intelligent Control Strategy of D
STATCOM[D]. Wuhan Institute of Technology, 2011.
[2] XUE Van-hong. The Study on PSO and ADRC Theroy for Control
System of D-STATCOM [D]. Guangxi University, 2011.
[3] TANG Jie, LUO A, OU Jian-bo. Voltage Control Strategy of D
STATCOM Based on Fuzzy-PI Controller [J]. Transactions of
China Electro technical Society, 2008, 23(2):120-126.
[4] OU Jian-bo, LUO A, TANG Jie. A Direct Voltage Control
Method of STATCOM Based on Neuron Selfadaptive PID[J].
Automation of Electric Power Systems, 2007, 31(12):77-80.
[5] MENG Xing-ping, WANG Hui, ZHAO Liang. Controlling Study of
D-ST A TCOM Based on Reinforcement Learning Adaptive PID[C].
IEEE International Conference on Automation and Logistics,
Changchun China, 2009.
[6] TANG Jie, DENG Zhi-yong, YANG Zhi-hong. Double closed loop
control for distribution static synchronous compensator based on
state PI feedback control[J]. Journal of Central South
University(Science and Technology), 20 I 0, 41 (6): 2282-2287.
[7] Han Liqun. Artitcial Neural Network[M]: Theory, Design and
application. Beijing, Chemical Industry Press, 2007.
[8] Liu Jinkun. Advanced PID Control and MA TLAB
simulation[M]. Publishing House of Electronic Industry, 2003.
CICED2012 Session 3 Paper No CPOSSO Page4/4