1

AbstractPhasor Measurement Units (PMUs) have been

increasingly widespread throughout the power network. As a
result, several researches have been made to locate the PMUs for
complete system observability. Many protection applications are
based upon the PMUs locations. This paper introduces an
important application in power system protection which is the
detection of single line outage. In addition, a detection of the
outaged line is achieved depending on the variations of phase
angles measured at the system buses where the PMUs are
located. Hence, a protection scheme from unexpected overloading
in the network that may lead to system collapse can be achieved.
Such detections are based upon an artificial intelligence
technique which is the support Vector Machine (SVM)
classification tool. To demonstrate the effectiveness of the
proposed approach, the algorithm is tested using offline
simulation for the 14-bus IEEE system.

Index TermsPhasor measurement units, PSCAD, Support
vector machines, Transmission line measurements.
I. INTRODUCTION
INE outage study provides a measure of the overall effect
on the system due to that line outage. The power flow of
the system is affected by the line outage. Therefore, this may
lead to overloading for certain lines. Also, this causes changes
in the power angles for several buses. A protective action
should be made to prevent large system disturbances (over
loading of lines) leading to cascaded tripping results in a
system collapse. However, connectivity information over the
wide area is very important for system operations [1, 2].
In recent years, PMUs are the only devices which are
deployed over the entire power grid and are capable of
providing geographically dispersed accurate synchronized
measurements [3]. Therefore, they are used for system control,
protection, state estimation and topology estimation [4 - 8].
Consequently, PMUs became a powerful technology in power
system protection.
Mikolinnas and Wollenberg [9] have presented a version of
the Megawatt Performance Indices (PI) whish are function of
bus voltages and line flows and the corresponding limits. They

A. Y. Abdelaziz, S. F. Mekhamer and M. Ezzat are with Electrical Power and
Machines Department, Faculty of Engineering, Ain Shams University, Cairo,
Egypt.
E. F. El-Saadany is with Electrical and Computer Engineering Department,
University of Waterloo, Canada

include all terms in the infinite Taylors series expansion for
all the change in the performance index due to different
outage. Irissari and Sasson [10] have proposed an improved
computational procedure based on DC load flow method.
Vemuri and Usher [11] have presented a unified approach to
find sensitivity of performance index for single branch outage,
generation/load outage and combination of them. D. Hazarika
et al. [12] describe an algorithm for determining the line
outage contingency of a line taking into account of line over
load effect in remaining lines and subsequent tripping of over
loaded line(s) leading to system split or islanding of a power
system using fast decoupled load flow analysis. J. Tate and T.
Overbye [13, 14] introduce two researches for single line
outage and double line outage respectively using phasor angle
measurements. This is done by utilizing an optimization
technique for event detection. Authors of [13] presented
accurate results for a line outage case for a certain line in the
37-bus IEEE system.
This paper focuses on the line outage problem and proposes
an artificial intelligence based technique for outage detection.
The outaged line is determined by means of an SVM
classification tool. The task of SVM is to utilize the output
information from the PMUs to determine a status for each line
if it is outaged or not. The PMUs calculate the phasor angles at
the certain buses based on the complete observability
approach discussed in [15]. Section II of this paper describes
how to use the SVM approach in line outage detection and
determination of outaged line. Section III proposes an
approach for detection of the outaged line based on the
variations of the phasor angles at different buses with the aid
of the SVM tool. Section IV introduces a numerical simulation
for the 14-bus IEEE system using an offline simulation
program EMTDC/PSCAD [19] and the mathematical analysis
with the aid of MATLAB. Section V clarifies the extracted
conclusions.
II. SUPPORT VECTOR MACHINE
Support vector machine [16, 17] was originally introduced
by Vapnik and co-workers in the late 1990s. It is a
computational learning method based on the statistical
learning theory. SVM mainly has two classes of applications,
classification and regression. In this paper, only the
application of classification is considered.
The classification problem can be restricted to consideration
of the two-class problem without loss of generality. In this
Line Outage Detection Using Support Vector
Machine (SVM) Based on the Phasor Measurement
Units (PMUs) Technology
A. Y. Abdelaziz, S. F. Mekhamer, M. Ezzat and E. F. El-Saadany
L
978-1-4673-2729-9/12/$31.00 2012 IEEE

pro
(a
wo
T
the
(m
of
sep


E
cor
rep
val
cas


W
per
Ad
ma
thr
lin
loo
can
sat
T

N
Th
nor
|1
ho
aga
|1
ma
oblem, the goa
classifier), wh
ork well on uns
There are many
e data, but the
maximizes the d
each class).
parating hyper-
Given a set of
{ } ,
i i
x y
Each x
i
is a
rresponding y
i

present a non-m
lue of 1 to rep
se. The points x
Where w is n
rpendicular di
dding the offs
argin. In its a
rough the origi
nearly separabl
oks for the sep
n be formulate
tisfy the follow
.
i
x
.
i
x
These can be co
(
i
y x
Now consider t
hese points lie
rmal w and
b|/||w||. Simi
lds lie on the h
ain w, and
1b|/||w||. It i
argin is simply
al is to separat
hich is induced
seen examples,
y possible line
ere is only on
distance betwe
This linear cl
-plane.
f training data
}, 1,..., , i n =
a d-dimension
is a numerical
matching train
present a matc
x which lie on
. w x
ormal to the
istance from t
set parameter
absence, the h
in. ||w|| is the
le case, the su
parating hyper-
d as follows (s
wing constraints
. 1 w b + +
1 w b f +
ombined into o
. ) 1
i
w b +
the points for w
on the hyper-
perpendicula
larly, the poin
hyper-plane H
2
perpendicular
is noticed that
2/||w|| which
e the two clas
d from availab
, i.e. it generali
ear classifiers
ne that maxim
een it and the n
lassifier is term
{ } , 1,1
i
y
nal real vecto
l decision has
ning data to the
ching training
the hyper-plan
0 b + =

hyper-plane a
the hyper-plan
b allows us
hyper-plane is
Euclidean no
upport vector a
plane with larg
suppose that all
s):
1
i
for y = +
1
i
for y =
one set of inequ
0
, for all
which the equa
-plane H
1
: x
i
.
ar distance f
nts for which th
2
: x
i
. w + b =
r distance fr
H
1
and H
2
are
is observed in
ses by a funct
ble examples a
izes well.
that can separ
mizes the mar
nearest data po
med the optim
},
d
i
x R

or R
d
, and
the value of -1
e study case o
data to the stu
ne satisfy:

and b/||w|| is
ne to the orig
to increase
s forced to p
rm of w. For
algorithm simp
gest margin. T
l the training d
1


ualities:
l i
ality in (3) hol
w + b = 1 w
from the ori
he equality in
1, with norm
from the ori
e parallel and
Fig. 1.

ion
and
rate
gin
oint
mal
(1)
the
1 to
or a
udy
(2)
the
gin.
the
pass
the
ply
This
data
(3)
(4)
(5)
lds.
with
gin
(4)
mal
gin
the
Fig. 1. M
samples f

Traini
support
maximu
can be
to solve
root.
Fortun
||w|| wi
account
optimal
convex
Minim
Subje


Where
constan
maximi
number
increase
better.
In SV
the gen
linearly
generali
determi
original
product
A con
in this
mappin
An inne
input sp


This a
plane i
polynom


Maximum margin
from two classes.
ing points for w
t vectors. The
um margin by
found. The op
e because it de
nately it is pos
ith 0.5||w||
2
wi
t the noise with
l hyper-plane
quadratic prog
mize
mi
||w
ct to

y
i

e
i
is measur
nt C > 0 det
ization and tra
r of samples. M
es the margin 2
VM, the optima
neralization ab
y separable, th
ization abilit
ined optimally
l input space
t space called t
nstruction of th
high dimens
ng of the input
er product in f
pace as in (13).
k
allows the algo
in the transfor
mial kernel fun
( , k x
hyper-plane and
which the equa
e pair of hype
y minimizing |
ptimization pro
epends on ||w|
ssible to alter th
ithout changin
h slack variable
can be found
gramming (QP
in
w||

1
2
||w||
2
+
( . )
i i
x w b +
0, for i =
ring the degre
termines the t
aining error m
Minimizing the
2/||w||, so that
al hyper-plane
bility. But if
e obtained cla
ty although
y. Thus to enha
is mapped int
the feature spac
he linear separ
sional feature
data to feature
feature space h
.
( ) , ( ), x x x =
orithm to fit th
rmed feature
nction used wit
) (
, x x =
margins for a SV
ality in (5) hol
er-planes whic
|w||, subject to
oblem presente
|| which involv
he equation by
g the solution.
e and error p
by solving th
P) problem:
+C

n
=1

1 ,
i


1,..., n =

ee of misclassi
trade-off betw
minimization a
optimization v
the classificat
is determined
the training d
assifier may no
the hyper-
ance linear sep
to a high-dim
ce.
rating hyper-p
space, after
e space as show
has an equivale
, ( ) x

e maximum-m
space. In thi
th a degree of n
)
1
n
x +

2
VM trained with
lds are called
ch gives the
o constraints
ed is difficult
ves a square
y substituting
. Taking into
enalty C, the
he following
(6)
(7)
ification, the
ween margin
and n is the
variable ||w||
tion becomes
to maximize
data are not
ot have high
-planes are
paration, the
mensional dot
plane is done
a nonlinear
wn in Fig. 2.
ent kernel in
(8)
margin hyper-
is work, the
n:
(9)


Fig
T
dis
vo
are
me
con
com
app
app
dep
vo
A
of
Fo
20
vec
loa
Ea
out
Th
T
has
wh
of
i, j
lin
and
ele
I
fun
SV
ext
T
rep
res
des
I
is p
Th
loc
the
sam
Th
equ
T
for
pro
loa
con

g. 2 Mapping the tr
The line outag
stribution of th
ltages phase a
e the buses ne
easurements a
nsideration. T
mplete observ
plied to calc
proach uses th
pends on succ
ltage magnitud
A training of th
the line outag
r the proposed
lines connec
ctors is 20 sing
ading cases at
ach vector cont
tage case. The
hese vectors for
The y vector o
s the values of
hich the test da
-1 represents a
and k in the X
ne. Hence, whe
d k
th
elements
ements in the y
n the testing o
nction is applie
VM functions
tracted through
To show the e
peated at diffe
sults as a fun
scribing the pr
n order to test
performed usin
he 14-bus IEEE
cated upon the
e buses 2, 7, 1
mples are take
hen, the angles
uations introdu
The training of
r each line to r
oposed approa
ading cases at
nsidered in the
Normal loadin
An increase o
raining data into a
III. PROPOSE
ge in a transm
he lines power
angles are achi
ear the outage
all over the n
The PMUs are
vability [15]. T
culate the ph
he Rockefeller
cessive three s
de and angle. It
he SVM is ma
e scenarios at
d study case (14
ting the buse
gle line outage
t different bus
tains the result
eses vectors a
rm a matrix X.
of the SVM fo
f 1 and -1. The
ata matches the
a non-matchin
X matrix repre
en studying the
in the y vect
y vector should
of the studied
ed for each line
are collected
h the results of
effect of the e
erent values of
nction of C. F
oposed approa
IV. SIMULAT
the proposed a
ng EMTDC/PS
E system is sho
e complete ob
0 and 13 respe
en at the buses
s of voltage ph
uced in Append
f SVM require
record the ang
ch suggests m
t different bu
e SVM training
ng case as the
of 25% in loadi
higher-dimension
ED APPROACH
mission netwo
r flow. Hence,
ieved. The mo
ed lines. There
network shoul
e located at c
Then, a mathe
hasors angles.
and Udren alg
amples to calc
t is introduced
ade using all th
several differe
4 buses IEEE s
s. So, the num
e cases multipli
ses, (i.e. 120 t
ts of phase ang
are the x vecto

or a certain lin
e value of 1 rep
e training data
ng data (e.g. if
sent an outage
e outage of tha
or should equ
equal to -1)
system, an SV
e and the predic
d. Hence, the
f the testing of
error penalty
f C, to get the
Fig. 3 introduc
ach.
TION RESULTS
algorithm, an o
SCAD simulati
own in Fig. 4, a
servability app
ectively. There
s where the PM
hasors are calc
dix A.
es many study
gles variations
making such tra
uses. The foll
g using the poly
standard IEEE
ing at bus 4.
al feature space.
ork causes a
variations in
ost affected bu
efore, large sc
d be taken in
certain buses
matical model
. The propos
gorithm [18] t
culate the pha
in Appendix A
he possible ca
ent loading cas
system), there
mber of train
ied by 6 differ
training vector
gles for each l
ors of the SV
e is a vector t
presents a case
. While the va
the rows numb
e case of a cert
at line, the i
th
,
al to 1 and ot
VM classificat
cted values of
outaged line
SVM.
C, the testing
e accuracy of
ces a flow ch
offline simulat
ion package [1
and the PMUs
proach in [15]
efore, the volta
MUs are locat
culated using
y cases of outa
in all cases. T
aining at differ
owing cases
ynomial kernel
E system states.
re-
the
uses
cale
nto
for
l is
sed
that
asor
A.
ases
ses.
are
ing
rent
rs).
line
VM.
that
e in
alue
ber
tain
j
th
,
her
ion
the
is
g is
the
hart
ion
19].
are
] at
age
ted.
the
age
The
rent
are
l.
.
An i
An i
An i
An i
















































Fig. 3 Flo

Fig. 4 IEE
increase of 25%
increase of 25%
increase of 25%
increase of 25%
owchart describing
EE 14 bus system.
Sampl
SVM
Line 1
Output
y
1
YES
% in loading at
% in loading at
% in loading at
% in loading at
g the proposed app

END

Input:
les of PMUs meas
START
Collect output res
Extract outaged l
Calculate the
phase angles
Form a vector
containing the an
SVM
Line 2
Output
y
2
Need to vary Pen
error 'C'
NO
t bus 6.
t bus 9.
t bus 10.
t bus 12.
proach.
surements
sults
ine
e
s
r x
ngles
T
SVM
Line 20
Output
y
20
nalty
3


SVM
Training
4
Each of the six loading cases has an internal 20 cases of line
outage for the 20 lines in the system. Therefore, 120 vectors
are available to perform the required training of SVM.
The training vector consists of a number of parts arranged
together. Each part clarifies the angle variations at the buses
where the PMUs are located. The next step is to apply a
different loading case to test the SVM including all possible
line outages. Then, the results are tabulated and the accuracy
is calculated. It is proposed to perform such testing at
conditions that differ from the training conditions.
Firstly, to study the importance of complete observability of
the power system, the training and testing are performed by
considering two, three and four PMUs. Table I gives detailed
results for a study case where the load at bus 14 is increased
25% than the normal loading in the standard IEEE 14-bus
system and two PMUs located at buses 2 and 10 are used.
Table II gives similar results while three PMUs at buses 2, 7,
and 10 are utilized. And Table III shows the same analysis
with four PMUs located at the buses 2, 7, 10, and 13.
Secondly, In order to ensure the superiority of the SVM
classification tool in the field of line outage detection, several
study cases have been studied including the complete
observability principle. The study cases are given as follow:
Study case (1): An increase of 25% in loading at bus 14.
Study case (2): An increase of 25% in loading at bus 3.
Study case (3): An increase of 40% in loading at bus 9.
The targets of studying more study cases are; (1) to get an
approximated accuracy of the proposed approach, (2) to select
the appropriate range of penalty error for each line outage
detection case.
The results of the first study case are given in Table III,
while the other study cases are presented in Table IV and
Table V. The results are compared, and the comparison is
introduced in Fig. 5.
It is clear from Tables I, II and III that for each line, there is
a detectable range of penalty error "C" for the line outage.
When two PMUs are used, many lines are subjected to lack of
classification, either no detection or a miss of classification,
the accuracy in this case reaches a 16 correct classification
over 20 line outage cases. The remaining four cases are for
lines far away from the PMUs locations that results in a wrong
detection. The accuracy is improved when using three PMUs
and the correct detections reach 19 detections over 20 line
outage cases. Finally, when four PMUs are used and located
upon the complete observability approach, a complete
detection scenario is achieved and the correct detections reach
20 detections over 20 line outage cases and with a more wide
range of penalty error.
It is clear from Fig. 5 that for each line, there is a detectable
range of penalty error "C" for the line outage. Some lines are
classified clearly even the value of C is small which means
narrow range of classification or large which means wide
margin of classification (e.g. line connecting bus 6 and bus
13), where the result is correct for all tested range of C. While
other line needs a specified range of C (e.g. line connecting
bus 6 and bus 12), it requires a greater value of C starting from
1x10
4
. On the opposite, some lines need small values of C
(e.g. line connecting bus 9 to bus 14), it requires lower value
of C up to 10, after that a miss-classification may occur and
the result becomes confusing.
TABLE I
RESULTS OF STUDY CASE (1) WITH 2 PMUS LOCATED AT BUSES 2 AND 10

Study Case (1): An increase of 25% at bus 14 and PMUs are located at buses 2 and 10

Outage
Case
Result at
C = 1
Result at
C = 10
Result at
C = 100
Result at
C = 1000
Result at
C = 1x10
4

Result at
C = 1x10
5

Result at
C = 1x10
6

Result at
C = 1x10
7

Line 1-2 No outage 1-2 1-2 1-2 1-2 1-2 1-2 1-2
Line 1-5 No outage No outage 1-5 1-5 1-5 1-5 1-5 1-5
Line 2-3 No outage No outage 2-3 2-3 2-3 2-3 2-3 2-3
Line 2-4 No outage 2-4 2-4 2-4 2-4 2-4 2-4 2-4
Line 2-5 No outage No outage No outage No outage 2-5 2-5 2-5 2-5
Line 3-4 No outage No outage No outage No outage No outage No outage 3-4 3-4
Line 4-5 No outage No outage No outage No outage 4-5 4-5 4-5 4-5
Line 4-7 No outage No outage No outage No outage No outage No outage 7-9 7-9
Line 4-9 No outage No outage No outage No outage No outage No outage No outage 4-9
Line 5-6 No outage No outage No outage No outage No outage 5-6 5-6 5-6
Line 6-11 6-11 6-11 6-11 6-11 6-11 6-11 6-11 6-11
Line 6-12 No outage No outage No outage No outage No outage No outage No outage No outage
Line 6-13 No outage No outage No outage No outage No outage No outage 6-13 6-13
Line 7-8 No outage No outage No outage No outage No outage No outage No outage No outage
Line 7-9 No outage No outage No outage No outage No outage No outage 7-9 7-9
Line 9-10 No outage No outage No outage No outage No outage No outage 9-10 9-10
Line 9-14 No outage 9-14 9-14 9-14 , 1-5 9-14 , 1-5 9-14 , 1-5
9-14, 7-9,
1-5, 6-13
9-14, 7-9,
1-5, 6-13
Line 10-11 No outage No outage No outage No outage No outage 10-11 10-11 10-11
Line 12-13 No outage No outage No outage No outage No outage No outage No outage No outage
Line 13-14 No outage No outage No outage No outage No outage No outage 13-14 13-14
5
TABLE II
RESULTS OF STUDY CASE (1) WITH 3 PMUS LOCATED AT BUSES 2, 7, AND 10

Study Case (1): An increase of 25% at bus 14 and PMUs are located at buses 2,7, and 10

Outage
Case
Result at
C = 1
Result at
C = 10
Result at
C = 100
Result at
C = 1000
Result at
C = 1x10
4

Result at
C = 1x10
5

Result at
C = 1x10
6

Result at
C = 1x10
7

Line 1-2 No outage 1-2 1-2 1-2 1-2 1-2 1-2 1-2
Line 1-5 No outage 1-5 1-5 1-5 1-5 1-5 1-5 1-5
Line 2-3 No outage No outage 2-3 2-3 2-3 2-3 2-3 2-3
Line 2-4 2-4 2-4 2-4 2-4 2-4 2-4 2-4 2-4
Line 2-5 No outage No outage No outage No outage 2-5 2-5 2-5 2-5
Line 3-4 No outage No outage No outage 3-4 3-4 3-4 3-4 3-4
Line 4-5 No outage No outage No outage 4-5 4-5 4-5 4-5 4-5
Line 4-7 4-7 4-7 4-7 4-7 4-7 4-7 4-7 4-7
Line 4-9 No outage No outage No outage No outage No outage No outage 4-9 4-9
Line 5-6 No outage No outage No outage No outage 5-6 5-6 5-6 5-6
Line 6-11 6-11 6-11 6-11 6-11 6-11 6-11 6-11 6-11
Line 6-12 No outage No outage No outage No outage No outage No outage No outage 6-12
Line 6-13 No outage No outage No outage No outage No outage No outage 6-13 6-13
Line 7-8 No outage No outage No outage No outage No outage No outage No outage 7-8
Line 7-9 7-9 7-9 7-9 7-9 7-9 7-9 7-9 7-9
Line 9-10 No outage No outage No outage No outage No outage 9-10 9-10 9-10
Line 9-14 9-14 9-14 , 3-4
9-14, 3-4,
1-5
9-14, 3-4,
1-5
9-14, 3-4,
1-5
9-14, 3-4,
1-5
9-14, 3-4,
1-5, 6-13
9-14, 3-4,
1-5, 6-13
Line 10-11 No outage No outage No outage 10-11 10-11 10-11 10-11 10-11
Line 12-13 No outage No outage No outage No outage No outage No outage No outage No outage
Line 13-14 No outage No outage No outage No outage No outage No outage 13-14 13-14


TABLE III
RESULTS OF STUDY CASE (1) WITH 4 PMUS LOCATED AT BUSES 2, 7, 10, AND 13 (COMPLETE OBSERVABILITY)

Study Case (1): An increase of 25% at bus 14 and PMUs are located for complete observability

Outage
Case
Result at
C = 1
Result at
C = 10
Result at
C = 100
Result at
C = 1000
Result at
C = 1x10
4

Result at
C = 1x10
5

Result at
C = 1x10
6

Result at
C = 1x10
7

Line 1-2 No outage 1-2 1-2 1-2 1-2 1-2 1-2 1-2
Line 1-5 No outage 1-5 1-5 1-5 1-5 1-5 1-5 1-5
Line 2-3 No outage 2-3 2-3 2-3 2-3 2-3 2-3 2-3
Line 2-4 2-4 2-4 2-4 2-4 2-4 2-4, 10-11 2-4, 10-11 2-4, 10-11
Line 2-5 No outage No outage No outage No outage 2-5 2-5 2-5 2-5
Line 3-4 3-4 3-4 3-4 3-4 3-4 3-4 3-4 3-4
Line 4-5 No outage No outage No outage 4-5 4-5 4-5 4-5 4-5
Line 4-7 4-7 4-7 4-7 4-7 4-7 4-7 4-7 4-7
Line 4-9 No outage No outage No outage No outage No outage No outage 4-9 4-9
Line 5-6 No outage No outage No outage No outage 5-6 5-6 5-6 5-6
Line 6-11 6-11 6-11 6-11 6-11 6-11
6-11,
10-11
6-11,
10-11
6-11,
10-11
Line 6-12 No outage No outage No outage No outage 6-12 6-12 6-12 6-12
Line 6-13 6-13 6-13 6-13 6-13 6-13 6-13 6-13 6-13
Line 7-8 No outage No outage No outage No outage No outage No outage 7-8 7-8
Line 7-9 7-9 7-9 7-9 7-9 7-9 7-9 7-9 7-9
Line 9-10 No outage No outage No outage No outage 9-10 9-10 9-10 9-10
Line 9-14 9-14 9-14 9-14, 3-4 9-14, 3-4 9-14, 3-4 9-14, 3-4 9-14, 3-4 9-14, 3-4
Line 10-11 No outage No outage No outage 10-11 10-11 10-11 10-11 10-11
Line 12-13 No outage No outage No outage No outage No outage No outage No outage 12-13
Line 13-14 13-14 13-14 13-14 13-14 13-14 13-14 13-14 13-14


6
TABLE IV
RESULTS OF STUDY CASE (2) WITH 4 PMUS LOCATED AT BUSES 2, 7, 10, AND 13 (COMPLETE OBSERVABILITY)

Study Case (2): An increase of 25% at bus 3 and PMUs are located for complete observability

Outage
Case
Result at
C = 1
Result at
C = 10
Result at
C = 100
Result at
C = 1000
Result at
C = 1x10
4

Result at
C = 1x10
5

Result at
C = 1x10
6

Result at
C = 1x10
7

Line 1-2 No outage 1-2 1-2 1-2 1-2 1-2 1-2 1-2
Line 1-5 No outage 1-5 1-5 1-5 1-5 1-5 1-5 1-5
Line 2-3 No outage No outage 2-3 2-3 2-3 2-3 2-3 2-3
Line 2-4 2-4 2-4 2-4 2-4 2-4 2-4 2-4 2-4
Line 2-5 No outage No outage No outage No outage 2-5 2-5 2-5 2-5
Line 3-4 3-4 3-4 3-4 3-4 3-4 3-4 3-4 3-4
Line 4-5 No outage No outage No outage 4-5 4-5 4-5 4-5 4-5
Line 4-7 4-7 4-7 4-7 4-7 4-7 4-7 4-7 4-7
Line 4-9 No outage No outage No outage No outage No outage 4-9 4-9 4-9
Line 5-6 No outage No outage No outage No outage 5-6 5-6 5-6 5-6
Line 6-11 6-11 6-11 6-11 6-11 6-11 6-11 6-11 6-11
Line 6-12 No outage No outage No outage No outage 6-12 6-12 6-12 6-12
Line 6-13 6-13 6-13 6-13 6-13 6-13 6-13 6-13 6-13
Line 7-8 No outage No outage No outage No outage No outage No outage 7-8 7-8
Line 7-9 7-9 7-9 7-9 7-9 7-9 7-9 7-9 7-9
Line 9-10 No outage No outage No outage No outage 9-10 9-10 9-10 9-10
Line 9-14 9-14 9-14 9-14 9-14 9-14 9-14 9-14 9-14
Line 10-11 No outage No outage No outage 10-11 10-11 10-11 10-11 10-11
Line 12-13 No outage No outage No outage No outage No outage No outage No outage 12-13
Line 13-14 13-14 13-14 13-14 13-14 13-14 13-14 13-14 13-14
TABLE V
RESULTS OF STUDY CASE (3) WITH 4 PMUS LOCATED AT BUSES 2, 7, 10, AND 13 (COMPLETE OBSERVABILITY)

Study Case (3): An increase of 40% at bus 9 and PMUs are located for complete observability

Outage
Case
Result at
C = 1
Result at
C = 10
Result at
C = 100
Result at
C = 1000
Result at
C = 1x10
4

Result at
C = 1x10
5

Result at
C = 1x10
6

Result at
C = 1x10
7

Line 1-2 No outage 1-2 1-2 1-2 1-2, 10-11 1-2, 10-11 1-2, 10-11 1-2, 10-11
Line 1-5 No outage 1-5 1-5 1-5 1-5, 2-5 1-5, 2-5 1-5, 2-5 1-5, 2-5
Line 2-3 No outage 2-3 2-3 2-3 2-3 2-3 2-3 2-3
Line 2-4 2-4 2-4 2-4 2-4 2-4, 2-5
2-4, 2-5,
7-8
2-4, 2-5,
4-9, 7-8,
2-4, 2-5,
4-9, 7-8,
Line 2-5 No outage No outage No outage No outage 2-5 2-5 2-5, 7-8 2-5, 7-8
Line 3-4 No outage 3-4 3-4 3-4 3-4, 10-11 3-4, 10-11 3-4, 10-11 3-4, 10-11
Line 4-5 No outage No outage No outage 4-5 4-5 4-5 4-5 4-5
Line 4-7 4-7 4-7 4-7 4-7 4-7, 10-11 4-7, 10-11 4-7, 10-11 4-7, 10-11
Line 4-9 No outage No outage No outage No outage No outage No outage 4-9 4-9
Line 5-6 No outage No outage No outage No outage 5-6 5-6, 10-11
5-6,10-11,
7-8, 9-10
5-6,10-11,
7-8, 9-10
Line 6-11 6-11 6-11 6-11 6-11 6-11 6-11 6-11, 4-9 6-11, 4-9
Line 6-12 No outage No outage No outage No outage 6-12 6-12 6-12 6-12
Line 6-13 6-13 6-13 6-13 6-13 6-13 6-13 6-13 6-13
Line 7-8 No outage No outage No outage No outage No outage No outage 7-8 7-8, 10-11
Line 7-9 7-9 7-9 7-9 7-9 7-9 7-9 7-9, 4-9 7-9, 4-9
Line 9-10 No outage No outage No outage No outage No outage 10-11 10-11, 7-8 10-11, 7-8
Line 9-14 No outage 9-14 9-14 9-14 9-14, 10-11 9-14, 10-11 9-14, 10-11 9-14, 10-11
Line 10-11 No outage No outage No outage No outage No outage 10-11 10-11 10-11
Line 12-13 No outage No outage No outage No outage No outage No outage No outage 12-13
Line 13-14 13-14 13-14 13-14 13-14 13-14 13-14
13-14,
9-10
13-14,
9-10



Fig

Fig
g. 5 Comparison of
g. 6 Phase angle va
P
h
a
s
e

a
n
g
l
e

a
t

B
u
s

2
P
h
a
s
e

a
n
g
l
e

a
t

B
u
s

1
0
f different study ca
ariation for outage
0 0.1
-20
-15
-10
-5
0
0 0.1
-20
-15
-10
-5
0
ases using polynom
of line 6-13 at the
0.2 0.3
Ti me i n seconds
0.2 0.3
Ti me i n seconds
mial kernels.
e instant t = 0.3 sec
0.4 0
s
0.4 0
s

Fig. 6
all phas
near bu
decreas
accurate
Fig. 7
effect o
line 7-
accordin
outage.
outage o
from Fi
For th
than the
and (2).
case (2)
a miss c
grey sh
1x10
4
. T
40% ov
cases es
permits
bus 10.
approac
detectio
The c
value o
for som
classific
case of
process
subrout
"C" tha
to get c
of the S
cases ov
conds for study cas

0.5 0
-20
-15
-10
-5
0
P
h
a
s
e

a
n
g
l
e

a
t

B
u
s

7
0.5 0
-20
-15
-10
-5
0
P
h
a
s
e

a
n
g
l
e

a
t

B
u
s

1
3
6 clarifies that
sors at all buse
us to the outag
e of about 4
e classification
7 shows that th
on the phase an
-8 is normally
ng to the load
However, thi
of line 7-8 for
ig. 5.
he study case (
e standard syst
. The differenc
), the SVM fail
classification a
hadow. It dete
The reason of
ver loading le
specially near
a miss classif
Even a miss c
ch still has 19
on which mean
conclusion is th
f C increases t
me cases the inc
cations. There
f line outage
ing for all
tine operates w
at to be adjuste
correct decision
SVM classifica
ver 60 cases w
se (1).
0.1
T
0.1
T
the outage of
es. But, the gr
ged line (i.e.
4
o
). Therefore,
n in that case.
he outage of lin
ngle variations
y loaded wit
d flow study
is may lead to
r a great range
(3), the bus nu
tem that differs
ce between the
ls in the detect
as shown in T
ects line 10-1
f this miss clas
ets the study c
the heavy loa
fication for the
classification o
9 correct dete
ns a percentage
hat the overall
to have much
crease in C allo
efore, to have
e, it is impor
lines with pa
with the suitable
ed through diff
ns. Consequen
ation tool for th
with a percentag
0.2 0.3
Time in seconds
0.2 0.3
Time in seconds
line 6-13 has
reatest effect is
bus 13 phase
this result h
ne 7-8 is almo
. This happens
th a very sm
before the ac
o a miss dete
of penalty erro
umber 9 is 40%
s from both stu
e two study ca
tion of line 9-1
able V by bold
1 instead of 9
ssification may
case far from
aded bus (i.e.
e line connectin
occurs, the acc
ections versus
e of 95 % of ac
l accuracy incr
classification
ows an appeara
correct decis
rtant to make
arallel subrou
e or optimum p
fferent study ca
ntly, the obtain
hese study case
ge of 98.3%

0.4 0.5
0.4 0.5
7
an effect for
s at the most
angle has a
helps for an
ost having no
s because the
mall loading
ction of line
ection of the
or as noticed
% overloaded
udy cases (1)
ses is that in
0 and makes
ded font and
9-10 at C =
y be that the
the training
bus 9). This
ng bus 9 and
curacy of the
s one wrong
curacy.
reases as the
margin. But,
ance of miss-
sion for any
e a parallel
utines. Each
penalty error
ases in order
ned accuracy
es reaches 59

Fig

T
de
the
res
(i.e
en
pro
err
co
inc
Th
wh
tra
be
wi
SV
ne
cla
un
sev
T
alg
Th
[1]
g. 7 Phase angle va
The paper p
etermination of
e technology o
sults when the
e. the PMUs
ntire power netw
ocessing for e
ror. Therefore
orrect detection
cluding three s
he research ca
hen the regula
aining of SVM
e so far from th
ill be achieved
VM related to t
etwork dispatch
assifications m
nloaded or ligh
vere action. Fu
The operating
gorithm [18]
he amplitude an
US-Canada Pow
14, 2003 the Bl
ariation for outage
V. CON
presents the
f the outaged li
of PMUs. The p
e concept of co
are located in
work). The pa
each line with
e, the approac
ns over 60 pos
study cases for
an be applicab
ar daily loadin
. Hence, the ac
he training case
d. It is suggest
the variations o
h center. The
may occur du
ht loaded line w
uture work will
VI. APP
g equations o
,
nd phase angle
VII. REF
wer System Outag
lackout in the Unit
0 0.1
-6
-5.5
-5
-4.5
-4
P
h
a
s
e

a
n
g
l
e

a
t

B
u
s

2
0 0.1
-17
-16.5
-16
-15.5
-15
P
h
a
s
e

a
n
g
l
e

a
t

B
u
s

1
0
of line 7-8 at the i
NCLUSION
SVM classif
ine in a power
proposed appr
omplete observ
n certain buse
per suggests a
the optimum
ch reaches an
ssible cases of
the 14-bus IEE
ble for any ne
ng cases are c
ctual case of lin
es. As a result,
ed to get a dy
of network loa
paper also sho
ue to a case
which cannot b
l consider large
PENDIX A
f the Rockef
e of the sample
,
FERENCES
ge Task Force. Fi
ted States and Can
0.2 0.
Time in second
0.2 0.
Time in second
instant t = 0.3 seco
fication tool
network based
oach has accur
vability is appl
es to observe
pplying a para
range of pena
accuracy of
f outage scenar
EE power syste
etwork especia
considered in
ne outage will
, a good detect
ynamic training
ding stated by
ows that the m
of an outage
be considered a
er systems.
feller and Ud
d signal are:

nal Report on Au
nada, 2004.
.3 0.4
ds
.3 0.4
ds
onds for study case
in
d on
rate
lied
the
allel
alty
59
rios
em.
ally
the
not
tion
g of
the
miss
of
as a
dren

ugust
[2] US-C
Impl
[3] B. S
of ph
inco
Engi
[4] Wei
Estim
Tran
[5] Geo
bad
mea
1524
[6] A. M
estim
Aug
[7] M. K
Proc
[8] E. L
iden
Pow
[9] T. A
selec
no. 2
[10] G. D
meth
Syst
[11] S. V
algo
pp. 3
[12] D. H
anal
Ener
[13] J. E.
mea
pp. 1
[14] J. E.
angl
Proc
[15] S. C
mea
on P
[16] S. A
Spri
[17] I. St
Spri
[18] Q. H
math
[19] PSC
0.5 0
-15
-14.5
-14
-13.5
-13
P
h
a
s
e

a
n
g
l
e

a
t

B
u
s

7
0.5 0
-17
-16.5
-16
-15.5
-15
P
h
a
s
e

a
n
g
l
e

a
t

B
u
s

1
3
e (1).
Canada Power S
lementation of the
Singh, N. Sharma,
hasor measuremen
orporated with F
ineering, Science
iqing Jiang, Vijay
mator Utilizing
nsactions on Powe
orge N. Korres, an
data processing
asurements", Electr
4, March 2011.
Monticelli, Model
mation, IEEE Tr
g. 1993.
Kezunovic, Moni
c. 39th Hawaii Int.
Loureno, A. Cost
ntification method
wer Syst., vol. 21, n
A. Mikolinnas ,
ction algorithm, I
2, pp. 608617, Fe
D. Irissari, and A.
hod for online se
t., vol. PAS-100, n
Vemuri, and R. E.
orithms, IEEE Tr
346354, Feb. 198
Hazarika, S. Bhuya
lysis including the
rgy Systems, vol.
. Tate, and T. J. O
asurements, IEEE
16441652, Nov. 2
. Tate, and T. J. Ov
le measurements,
ceeding, pp. 15, J
Chakrabarti, and
asurement units fo
Power Systems, vo
Abe, Support Vec
inger-Verlag, Lond
teinwart, and A. C
inger, 2008.
H. Wu, Z. Lu, T.
hematical Moroph
CAD Users Guide
0.1 0
Ti
0.1 0
Ti
ystem Outage Ta
e Task Force Reco
A. Tiwari, K. Ver
nt units (PMUs) in
FACTS controlle
and Technology, v
Vittal, and Geral
Synchronized
er Systems, Vol. 22
nd Nikolas M. Ma
g for system i
tric Power System
ling circuit breake
rans. Power Syst.
itoring of power s
. Conf. System Sci
ta, K. Clements, an
directly based on
no. 4, pp. 192019
B. F. Wollenbe
IEEE Trans Powe
eb. 1981.
. M. Sasson, An
ecurity analysis,
no. 4, pp. 183818
. Usher, On line
rans Power Appar
83.
an, and S.P. Chow
e system islanding
28, no. 4, pp, 232
Overbye, Line out
E Transactions on
2008.
verbye, Double li
IEEE Power & E
July 2009.
E. Kyriakides,
or power system o
ol. 23, no. 3, pp. 14
tor Machines for
don, Ltd, 2005.
Christmann, Suppo
Y. Ji, Protective r
hology, Springer-V
e Ver. 4.2, Manitob
.2 0.3
me in seconds
.2 0.3
ime in seconds

ask Force. Final
mmendations, 200
rma, and S. Singh
n electric power sy
ers, Internationa
vol. 3, no. 3, pp. 64
d T. Heydt, "A D
Phasor Measure
2, No. 2, pp. 563-5
anousakis, "State
including PMU
ms Research, 81 (20
ers in weighted lea
, vol. 8, no. 3, p
system topology in
iences, Jan. 2006.
nd R. Cernev, A
n collinearity tests
929, Nov. 2006.
erg, An advance
er Apparatus Syst.
automatic conting
IEEE Trans Po
844, Apr. 1981.
automatic conting
ratus Syst., vol. P
wdhury, Line outa
g scenario Electr
2243, May 2006.
tage detection usin
n Power Systems,
ine outage detectio
Energy Society G
Optimal placem
bservability, IEE
4331440, Aug. 20
Pattern Classifica
ort Vector Machin
relaying of power
Verlag London Lim
ba Research Cente
0.4 0.5
0.4 0.5
8
Report on the
06.
h, Applications
ystems network
al Journal of
482, 2011.
Distributed State
ements", IEEE
571, May 2007.
estimation and
and SCADA
011), pp. 1514-
ast squares state
pp. 11431149,
n real-time, in
A topology error
s, IEEE Trans.
ed contingency
, vol. PAS-100,
gency selection
ower Apparatus
gency selection
PAS-102, no. 2,
age contingency
rical Power and
ng phasor angle
vol. 23, no. 4,
on using phasor
General Meeting
ment of phasor
EE Transactions
008.
ation. England:
nes, New York:
r systems using
mited 2009.
er, April 2005.