You are on page 1of 6

1

Abstract This paper presents a method for power swing and


fault diagnosis of power system based on Support Vector
Machine (SVM) classifier. The method adopts Least Square
Support Vector Machine (LS-SVM) classifier to identify the
power swing and fault types. The power swing blocking elements
are based on monitor the rate of change of the impedance , the
power swing center voltage, the positive current and zero
sequence component. The process of training the LS-SVM using a
K-folded cross validation process for determining the values of
parameter

and parameter in RBF kernel parameter that
will give minimum classification error. The proposed method can
successfully detect power swing and provide power swing
blocking for accurate distance protection during power swing.

Keywords Power Swing , Least Square Support Vector
Machine, Distance Relay
I. INTRODUCTION
HEN fault occur in power transmission system,
digital distance relays detect the fault with changing
impedance in power transmission line and initiate
tripping of appropriate circuit breakers[1]. Normally
oscillation in power system will be damped and generators will
be able to return to a normal state condition. During a power
swing period, the distance relay shall be calculated impedance
of power system based on measured voltage and current that
also oscillate in amplitude and phase with a power swing
frequency. The calculated impedance may be very small, that it
will lead to the fault detection to zone-1 of distance relay and
lead to miss-operation of digital distance relay. To prevent the
distance relay from having miss-operation from power swing,
a power swing blocking (PSB) function is required. The power
swing blocking function of conventional method cant able to
distinguish fault with high grounding resistance and cant able

Authors would like to thank the financial support form Mahanakorn
University of Technology, Thailand and Nakhonpathom Rajabhat University
(NPRU) , Thailand.
W. Kampeerawat

is with the department of Electrical Power Engineering,
Mahanakorn University of Technology, Thailand. (e-mail: warayut
@mut.ac.th).
W. Buangam is with the department of Electrical Power Engineering,
Mahanakorn University of Technology, Thailand. (e-mail:
rotnaja@yahoo.com).
S. Chusanapiputt is with the department of Electrical Power Engineering,
Mahanakorn University of Technology, Thailand. (e-mail:
songsak@mut.ac.th).

to clear the blocking of distance relay for symmetrical fault the
occur during power swing[2].
In recent years, many intelligent power swing diagnosis
methods for distance protection have emerged, such as: based
on an adaptive neuro-fuzzy inference system(ANFIS)[2] and
based on neural networks[3]. Presentation in this paper, a new
approach for novel power swing blocking distance relay of
transmission line using Support Vector Machine (SVM) is
main focus. SVM is used as an intelligent tool to indentify the
power swing that it is estimating and finding the power swing
blocking (PSB). The input signals to SVM include the
changing of positive sequence impedance, positive and
negative sequence currents and power swing of center voltage.
Using the EMTPDC/PSCAD program, a power system with
two three phase machines has been simulated for the study of
power swing problem. Training patterns have been generated
by simulating disturbance on the power system and including
power swing with and without fault and different type of fault
without power swing.
This paper approach LS-SVM based power swing
blocking can be detected with different power swing and
blocking distance relay from tripping during swing. This is
also applicable to power swing with high swing frequencies.
II. POWER SWING DETECTION
Digital Distance Relay detection the power swing based
on changing rate of electrical variable, such as derivative
impedance(dZ/dt), derivative resistance(dR/dt) or derivative
voltage(dU.cos/dt) [4]. Impact of impedance when the power
system fault conditions with the power swing the changes at
different rates. When a power swing occurs, the impedance
will change slowly compared to the fault conditions that occur
on power systems. Method to detect the power swing
following ways.
A. The Decreased Impedance Method
This principle of detection impedance during power swing.
The impedance changes slowly. Compared to the impedance
when the fault occurs on power systems. This method is able
to distinguish slow power swings, earth fault with high fault
resistance, but it cannot detect fast power swing, when the
fault during a power swing occurs by the blocking relay will
not reset[1].


New Swing-blocking Methods for Digital
Distance Protection Using Support Vector
Machine
W. Kampeerawat

, W. Buangam, S. Chusanapiputt
W
2010 International Conference on Power System Technology
978-1-4244-5939-1/10/$26.002010 IEEE
2

B. The V
s
cos Algorithm
This algorithm evaluates the change in V
s
cos as function
of time. The criterion which the response by the power swing-
blocking relay depends is the steady reduction in the absolute
value of V
s
cos[1].
C. Decreased Resistance Method
When a power swing occurs, the resistance of the system
will be changed continuously. When fault occurs, the
resistance component of the system will rapidly changing. This
method can even detect three-phase fault during power
swing[1].However, it is difficult to choose the setting value for
the derivative of the resistance since it is very small and the
scheme may fail to trip high resistance short circuit fault that
occur during power swing [4].
III. LEAST SQUARE SUPPORT VECTOR MACHINE
Support Vector Machine(SVM) is a powerful tool for
classification problems.For binary-class classifications, SVM
constructs an optimal separating hyperplane between the
positive and negative classes with the maximal margin. Least
Square Support Vector Machine (LS-SVM)[5] is a
reformulations of the standard SVM. Its mathematical
formulations are described in this section. Given a training
data set, { }
1
,
N
k k
k
x y
=
,where
n
k
x represents the k-th input
pattern and
k
y is the k-th output pattern, the LS-SVM
aims at constructing a classifier of the form,

( )
1
( ) ,
N
k k k
k
y x sign y x x b
=
(
= +
(

(1)
Where
k
is positive real constant and bis a real constant.
For
( ) . , .
one typically has the following a choices:

( ) ,
T
k k
x x x x = is the linear SVM

( )
2
2
2
, exp
k
x x
k
x x




`

)
= is the Radial Basis Function kernel
(RBF SVM)

( ) ( )
, 1
d
T
k k
x x x x = + is the polynomial SVM of degree
The LS-SVM to the SVM classifier is done by
formulating the classification problem as,

( )
2
, , 1
1 1
, ,
2 2
min
N
T
k
wb e k
J wb e w w e
=
= +

(2)
Subject to equality constraints,
( ) 1 , 1,.....,
T
k k k
y w x b e k N ( + = =

(3)
Where ( )
k
x is a nonlinear function which maps the input
space into a higher dimensional space. By using the Mercers
Theorem, this function is related to ( , )
k
x x as follows,
( ) ( ) ( , )
T
k k
x x x x = (4)
Equations (1) and (2) to Karush-Kuhn-Tucker systems and can
be written as the solution to the following set of liner
equations,

1
0 0
1
T
T
b Y
Y ZZ I

( ( (
=
( ( (
+
(5)
Where

| |
| | | | | |
1 1 1
1 1
( ) ;....; ( ) , ;....; ,
1 1;....;1 , ;....; , ;....;
T T
N N N
N N
Z x y x y Y y y
e e e


( = =

= = =

Mercers Theorem can be applied again to the matrix
1
T
k
ZZ = where,

1 1 1
1 1
( ) ( )
( , )
T
k k k
k k
y y x x
y y x x

=
=
(6)

Hence, the solution to the classifier as given in equation (1)
can be found by solving the linear set of equations (5) and (6)
instead of using quadratic programming for solving the
equation as is the case with SVM, The LS-SVM network
developed in this work uses the LS-SVM MATLAB
Toolbox[5]
IV. POWER SYSTEM STUDIED
The power system for this study was simulated using
EMTPDC/PSCAD [6] shown in Fig. 1. Distance relay
installed at bus S shown in Fig.1 was included for the
simulation of the power swing in the case of the power swing
with and without the fault during power swing. The power
system network consists of two areas of 230 kV connected by
200 km long transmission line. The power system parameters
are as follow: for transmission line, it has zero-sequence
impedance Z
0
=0.346+j1.066 / km , positive sequence
impedance Z
L1
=0.072+j0.416 / km , for source
impedances: Z
s
=41.7 [1], source voltage: E
S
,E
R
=230 kV.
The power system model is simulation at 1.0 kHz sampling
frequency an 50Hz frequency. The voltage and current signal
for power swing and fault conditions are retrieved at the
relaying point and input to the LS-SVMs for power swing,
fault and ground detection.

Distance
Relay
V
I
A B
BUS A BUS B
Z
S
Z
R


Fig. 1 A power systems model studied
V. IDENTIFICATION OF POWER SWING AND FAULT
A. Extracting Features
This section the details with training and testing results
obtained from the corresponding LS-SVMs for currents ,
voltages , and impedances, it is the change rates of current,
voltages, and/or impedances as inputs to LS-SVM for
identifying power swings and fault type[2]. Ref. [2]. using the
3
changing rate of different measured quantities present in Eqs.
(7) (10) .Where nn is the sample index and Nis Number of
samples/cycle
p
Z ,
p
I ,
n
I and
PSC
V correspond to the
calculated positive sequence impedance, the positive and
negative sequence currents and the power swing center
voltage, respectively.

( ) ( )
p p p
Z Z n Z n N = (7)
( ) ( )
p p p
I I n I n N = (8)
( ) ( )
n n n
I I n I n N = (9)
( ) ( )
PSC PSC PSC
V V n V n N = (10)
( )
p
Z n is calculated based on processed phasors of voltage
and currents.

( )
( )
( )
p
p
p
V n
Z n
I n
= (11)
Where ( )
p
V n and ( )
p
I n are the positive sequence of voltage
and current phasors at the relay location, respectively.An
approximation of the PSCV can obtained through the use of
locally available quantities as follows:
cos
PSC s
V V (12)
Where
s
V is the magnitude of locally measured voltage
( )
s
V , and is the angle difference between
S
V and the
locally measured current.
B. Training of LS-SVM Classifier
Using the MATLAB Toolbox [7] the support vectors are
calculated. Fig. 2 shows the process of training the LS-SVM.
Block diagram is a K-folded cross validation process for
determining the values of parameter and parameter in
RBF kernel parameter that will give minimum classification
error[8] and the number of support vectors . Based on the
power swing and fault types, three LS-SVMs are developed to
identify the power swing , fault types and ground detection.
With all training samples of three types, the first LS-SVM (LS-
SVM-1) is trained to separate normal state from power swing.
With samples of the fault samples, the second LS-SVM (LS-
SVM-2) is train separates fault detection. With samples of the
zero-sequence component, the third LS-SVM (LS-SVM-3) is
trained to separates ground detection. Thus, the multi-layer
SVM classifier is obtained. The basic principle of power swing
diagnosis based on multi-layer LS-SVM classifier is show in
Fig.3. Fig.4. show the power swing blocking function is based
on three functions to detect a power swing , fault and ground
fault.



I nput of training samples
Pre-processing
Normalization
K-folded cross validation
to determine the best value
of parameters and
Calculating the support vectors
And the optimal hyperplane
Training accomplished

a
b
c
d
e
f


Fig. 2 Block diagram of the training process.
VI. SIMULATION RESULT
This section the details with the power swing and fault
classification using LS-SVMs. The power swing is caused by a
three phase fault at location on AB line. Difference type of
shunt fault are L-G, L-L, L-L-G, L-L-L, and L-L-L-G[9][10].
The simulation of power swing , training and testing of the
scheme of power swing and fault with LS-SVMs classifier as
follows.

A. Power Swing Classifier (LS-SVM-1): The scheme using
changing rate of impedance (
p
Z
) and the power swing center
voltage (
PSC
V ) signal for power swing detection. When input
of LS-SVM-1 is a normal state sample, output of LS-SVM-1 is
set to -1 and power swing occurred, output of LS-SVM-1 is set
to 1. In the proposed LS-SVM-1 schemes, which consists of
over 500 training and testing patterns. The process of training
the LS-SVM-1using the 10-folded cross validation process for
determining the values of parameter =0.76602 and
parameter =0.5836 in RBF kernel parameter. The example of
this classifier results is shown in Fig. 4.
B. Fault Classifier (LS-SVM-2) : The scheme using fault
current signal
A
I ,
B
I and
B
I for fault detection[9][10]. When
input of LS-SVM-2 is a normal state sample, output of LS-
SVM-2 is set to -1 and fault occurred, output of LS-SVM-2 is
set to 1. In the proposed LS-SVM-2 schemes, which consists
of over 500 training and testing patterns. The process of
training the LS-SVM-2 using the 10-folded cross validation
process for determining the values of parameter = 0.0598
and parameter = 0.84953 in RBF kernel parameter .


4
LS-SVM-1
Power System
Simulation
Feature Extract and Preprocessing
p
Z
PSC
V
Power Swing
Normal or Fault
LS-SVM-2
Normal
Zero Sequence
Component
I a,I b,I c
LS-SVM-3
I p,I n,I z
Fault Classification Ground Fault
detection
Power Swing BlockingLogic Function
PBS
I a,I b,I c


Fig. 3 The Scheme of Power Swing and Fault with LS-SVMs Classification.

Power Swing
Detector
Three Phase Fault
Detector
Ground Fault
Detector
OR
AND PSB

Fig. 4 Power Swing Blocking (PSB) Logic Function
0 0.1 0.2 0.3 0.4 0.5
-200
-100
0
100
200
Power Swing Signal
Time(sec)
V
o
l
t
a
g
e
(
k
V
)
0 0.1 0.2 0.3 0.4 0.5
-2
-1
0
1
2
LS-SVM-1 Logic output
Time(sec)
L
o
g
i
c

o
u
t
p
u
t

Fig. 5 Power swing waveform and logic of LS-SVM1 output signal
when the power swing occur.

Fig.5. presents the decision of the LS-SVM-1 when
detection the power swing occurred. The output logic of LS-
SVM-1 set to 1. And Fig.5. presents the decision of the LS-
SVM-1 when detection fault occurred. The output logic of LS-
SVM-1 set to -1.
0 0.1 0.2 0.3 0.4 0.5
-1
0
1
2
Three Phase (Phase A) Fault Signal
Time(sec)
C
u
r
r
e
n
t
(
k
A
)
0 0.1 0.2 0.3 0.4 0.5
-2
-1
0
1
2
LS-SVM-1 Logic output
Time(sec)
L
o
g
i
c

o
u
t
p
u
t

Fig. 6 Three phase fault current waveform and logic of LS-SVM1 signal
when the fault phase A, B and C.

C. Ground Detection (LS-SVM-3) : The scheme using zero-
sequence component fault current signal for ground fault
detection[9][10][11].The ground detection is done separately
by training and testing LS-SVM-3. The value of the zero-
sequence component of the fundamental fault current signal
input of LS-SVM-3 is the fault without involving ground
sample, output of LS-SVM-3 is set to -1 and the fault
involving ground sample, output of LS-SVM-3 is set to 1. In
the proposed LS-SVM-3 schemes, which consists of over 500
training and testing patterns. The process of training the LS-
SVM-3 using the 10-folded cross validation process for
determining the values of parameter = 1.73337 and
parameter = 19.3377 in RBF kernel parameter .
TABLE I
TESTING RESULT FOR POWER FAULT AND GROUND FAULT DETECTION
No.
LS-SVM-2 LS-SVM-3
TYPE OF
FAULT
PHASE
A
PHASE
B
PHASE
C
GROUND
1
1 -1 -1 1 A-G
2
-1 1 1 1 B-G
3
-1 -1 1 1 C-G
4
1 1 -1 1 A-B-G
5
-1 1 1 1 B-C-G
6
1 -1 1 1 A-C-G
7
1 1 1 1 A-B-C-G
8
1 1 -1 -1 A-B
9
1 -1 1 -1 A-C
10
-1 1 1 -1 B-C

5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
0.5
1
Three Phase Fault Signal
Time(sec)
C
u
r
r
e
n
t
(
k
A
)
0 0.1 0.2 0.3 0.4 0.5
-1
0
1
Logic Fault PhaseA
Time(sec)
L
o
g
i
c

o
u
t
p
u
t
0 0.1 0.2 0.3 0.4 0.5
-1
0
1
Logic Fault PhaseB
Time(sec)
L
o
g
i
c

o
u
t
p
u
t
0 0.1 0.2 0.3 0.4 0.5
-1
0
1
Logic Fault PhaseC
Time(sec)
L
o
g
i
c

o
u
t
p
u
t

Fig. 7 Logic of LS-SVM2 signal when fault phase A , B and C

0 0.1 0.2 0.3 0.4 0.5
0
0.1
0.2
0.3
0.4
Zero-Sequence Component Current Fault Signal
Time(sec)
C
u
r
r
e
n
t
(
k
A
)
0 0.1 0.2 0.3 0.4 0.5
-2
-1
0
1
2
LS-SVM-3 Logic output
Time(sec)
L
o
g
i
c

o
u
t
p
u
t
Fig. 8 Zero sequence component current waveform and logic of LS-SVM3
signal when ground fault occur.

TABLE III
TESTING RESULT FOR POWER SWING, FAULT AND GROUND FAULT
Type of
Classifier
Event Simulated
N. F. G. PS
PS&
F
PS&G
Power
swing
-1 -1 -1 1 1 1
Fault -1 1 -1 -1 1 -1
Ground
Fault
-1 -1 1 -1 -1 1
PBS
Logic
-1 -1 -1 1 -1 -1
N:Normal , F:Fault , G:Ground Fault, PS: Power Swing

For the test of the classifier's ability, the test system was
simulated in 6 events ,i.e. normal state (N), fault without
ground (F), Ground Fault(G), Power swing (PS), Power swing
and fault in the same time (PS&F) and Power swing and
ground fault in the same time (PS&G). The classifications'
result of LS-SVM are shown in Table I. The LS-SVM can
successfully classify the event of power swing, fault and
ground fault. The classification results lead to correct output of
Power Swing Blocking for distance protection.
VII. CONCLUSIONS
This paper presents a scheme to detect power swing using
LS-SVM. It can detect power swing and the fault during a
power swing. The method adopts three-layer LS-SVMs
classifier to identify power swing and fault situation. The
proposed method can successfully distinguish power swing
from fault events and provide accurate PBS output for distance
protection during power swing. It is possible to clear blocking
if fault occur during a power swing.

VIII. REFERENCES

[1] H. Khoradshadi-Zadeh, Evaluation and performance comparison of
power swing detection algorithms Power Engineering Society General
Meeting, 2005. IEEE
[2] H. Khorashadi-Zadeh and Z. Li, A novel power swing blocking
scheme using adaptive neuro-fuzzy inference system Electric Power
Systems Research, Vol. 78, Issue 7, July 2008, pp 1138-1146.
[3] H. Khorashadi Zadeh S. Hosseini Naveh, B. T. Hosseini and S. H.
Abdol Khalegh Zadeh A NEW METHOD OF POWER SWING
BLOCKING FOR DISTANCE PROTECTION USING ANN AUPEC
2005 Australasian Universities Power Engineering Conference25
th
.
[4] Z. Shengshi, G. Qui and Z. Li, New swing-blocking methods for
digital distance protection Power Systems Conference and Exposition,
2004, pp.587-591.
[5] J.A.K Suykens, J. Vandewalle, Least Square Support Vector Machine
Classifier Neural Processing Letter, vol. 9(3), pp 293-300, August
1999.
[6] PSCAD/EMTDC Users Manual, Manitoba HVDC Research Center,
Winnipeg, MB, Canada, 2005.
[7] K. Pelckmans, J.A.K. Suykens, T. van Gestel, J. de Brabanter, L. Lukas,
B. Hamer, B. de Moor, and J. Vandewalle, LS-SVMlab Toolbox Users
Guide, ver.1.5 , http://www.esta .kuleuven.ac.be/be/sista/lssvmlab
[8] P.G.V. Axelberg, I.Y.H Gu, and H.J. Bollen Support Vector Machine
for Classification of Voltage Disturbances IEEE TRANSACTIONS
ON POWER DELIVERY VOL.22, NO. 3. JULY 2007, pp 1297-1303.
[9] S.R. Samantaray, and P.K. Dash, Transmission line distance relaying
using machine intelligence technique Generation, Transmission &
Distribution, IET Vol. 2, Issue 1, Jan. 2008 pp:53 61.
[10] P.K. Dash, S.R. and G.Panda, Fault Classification and Section
Identification of an Advanced Series-Compensated Transmission Line
Using Support Vector Machine IEEE TRANSACTION ON POWER
DELIVERY, Vol.22, NO. 1, JANUARY 2007, pp: 67-73.
[11] G. Lv, H. Chaeng, H. Zhai, and L. Dong Fault diagnosis of power
transformer based on multi-layer SVM classifier Electric Power
Systems Research Vol. 75, Issue 1, July 2005, pp. 9-15.










6
IX. BIOGRAPHIES

Warayut Kampeerawat, received his B.E
and M.E degree in Electrical Engineering
from Khon Kaen University, Khon Kaen,
Thailand, in 2005 and 2007 respectively.
Currently, he is a lecturer at Mahanakorn
University of Technology.




Wirot Buangam, received the B.Eng.
(EE) from Mahanakorn University of
Technology, Thailand in 1999, the
M.Eng. degree from King Mongkuts
University of Technology North Bangkok
, Thailand in 2007 respectively. He is
presently a D.Eng. student in the
department of Electrical Power
Engineering, Mahanakorn University of
Technology, Thailand.

Songsak Chusanapiputt received the
B.Eng. (EE) from Kasetsart University,
Thailand in 1983, the M.Eng. and D.Eng.
degree from Chulalongkorn University,
Thailand in 1985 and 1991 respectively.
He is presently a lecturer in the
department of Electrical Power
Engineering, Mahanakorn University of
Technology, Thailand.

You might also like