Fault locating in large distribution systems by empirical mode

decomposition and core vector regression

Benyamin Khorramdel
a
, Hesamoddin Marzooghi
b
, Haidar Samet
a,
, Meisam Pourahmadi-Nakhli
a
,
Mahdi Raoofat
a
a
Department of Power and Control Engineering, School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran
b
School of Electrical and Information Engineering, The University of Sydney, Sydney, New South Wales, Australia
a r t i c l e i n f o
Article history:
Received 20 November 2011
Received in revised form 30 November 2013
Accepted 11 January 2014
Keywords:
Core vector regression
Empirical mode decomposition
Distribution network
Energy spectrum
Fault locating
Path characteristic frequency
a b s t r a c t
This paper proposes an intelligent fault locating method using a new signal analysis technique called
Empirical Mode Decomposition (EMD) and Core Vector Regression (CVR) for large distribution systems.
The conventional fault locators are based on the measurement of post-fault line impedance suffering
from the factors such as path fault impedance, system conguration and line loading, so that they have
low accuracy. On the other hand, because of the vast range of resistances, the negative impact of damping
factors affects the performance of travelling wave-based fault locators in large distribution systems. To
overcome these problems, this paper uses a minimum measuring device to meet the acceptable observa-
tion of transient waves and presents a novel method for locating phase to ground faults in a large distri-
bution system using CVR. Inspecting the energy content of transient voltage around the path
characteristic frequencies by EMD can provide a suitable fault pattern to CVR. Training of the proposed
algorithm needs little time and small amount of memory in comparison with the existing methods. Pre-
sented algorithm is examined on IEEE 34-bus test system which shows satisfactory results. Then, the
results are compared with the method of recent papers based on Articial Neural Networks (ANNs).
2014 Elsevier Ltd. All rights reserved.
1. Introduction
Power quality is an important issue for power systems, espe-
cially power distribution systems. An accurate fault locating tech-
nique can decrease interruption time and costs, and it is essential
to nd the fault point. In [16] power frequency of voltage and cur-
rent fault waves is used to nd the fault location. Refs. [79] devel-
oped different methods based on combined impedance-travelling
wave methods and combined wavelet-fuzzy approach to nd the
fault location on transmission lines. Travelling wave-based meth-
ods are more efcient for fault locating in distribution systems in
comparison with traditional one, which are based on measurement
of post-fault line impedance [1022].
Traditional methods which estimate the path impedance with
fundamental voltages and currents at the main feeder are imprac-
tical for distribution networks due to the complex nature of large
distribution networks with huge number of radial branches.
Time-Domain Reectometry (TDR) is a well-known technique
used for locating faults on transmission lines [11,12]. TDR is based
on the distinction between the reected wave from the fault point
and the remote bus. In large distribution networks TDR signals
cannot be interpreted easily for locating faults because of the
reections of many Tee connections. The disruptive characteristics
of distribution lines as well as many reection factors cause the
monitoring of transient waves to be impossible with TDR, thereby
inefcient to locate them [22]. Besides, the accuracy of this method
is strongly inuenced by noise sources (e.g. corona and partial dis-
charge) around the equipments. A new approach for complexity
reduction of TDR-based methods in distribution networks is signal
processing-based methods; such as Continuous Wavelet Transform
(CWT) proposed in [18,19]. In [4,5,1517] Articial Neural Net-
work (ANN) which is a well-known machine learning technique
was introduced for fault locating on distribution lines, which re-
duced the problems of traditional methods considerably.
High-frequency fault waveforms are used as an important fea-
ture for ANN in [1517]. In [16] maximum values of Discrete
Wavelet Transform (DWT) of phases A, B, C and zero sequence
for post-fault voltage and current waveforms are used as ANN in-
put features. Although this method minimizes the inuence of load
conditions, it adds the effect of inception angle because of using
http://dx.doi.org/10.1016/j.ijepes.2014.01.028
0142-0615/ 2014 Elsevier Ltd. All rights reserved.

Corresponding author. Tel.: +98 711 6133455; fax: +98 711 2303081.
E-mail addresses: benyamin.khorramdel@gmail.com (B. Khorramdel), hesamod-
din.marzooghi@gmail.com (H. Marzooghi), samet@shirazu.ac.ir (H. Samet), mei-
sam.pourahmadi@gmail.com (M. Pourahmadi-Nakhli), raoofat@shirazu.ac.ir (M.
Raoofat).
Electrical Power and Energy Systems 58 (2014) 215225
Contents lists available at ScienceDirect
Electrical Power and Energy Systems
j our nal homepage: www. el sevi er . com/ l ocat e/ i j epes
the maximum values of transient components. Combination of
Support Vector Machine (SVM) and ANN are used for fault locating
in radial distribution systems while steady state three phase volt-
ages and currents are considered as inputs [5].
In [17] frequency and amplitude of the transient currents are
used as ANN inputs. This method minimizes inception angle im-
pact via elimination of low-frequency modes. Based on the energy
content of signals around the Path Characteristic Frequencies
(PCFs) of travelling waves, fault locating is done in [1820].
Inspecting the peak values of energy spectrum around the PCFs
in radial distribution systems is another approach for locating of
faults [18,19]. In [22] a new method has been proposed for fault
locating in small radial distribution systems based on the energy
of transient voltage signals. That reference used wavelet transform
and ANN for specifying the location of faults based on PCFs. Due to
complex nature of large radial distribution systems with different
Tee connections, aforementioned methods are not accurate enough
because of ANN structure and existence of just one measuring
device.
Furthermore, a new signal processing technique which is used
in this paper called Empirical Mode Decomposition (EMD) has re-
ceived much attention in a number of research areas [23]. EMD can
decompose a signal based on interpolation points criterion into a
number of Amplitude and Frequency Modulated (AM/FM) with
zero mean, called Intrinsic Mode Functions (IMFs). Some recent
studies on EMD showed specic aspects of its performance [23
25]. In [25] genetic algorithm optimization approach is used to
have good criteria for selection of interpolation points. Another
type of EMD with a good performance called doubly-iterative
EMD is introduced in [24,26] which estimates interpolation points
based on the extracted optimized criteria in [25], leading to
improve overall decomposition performance. As EMD is used for
feature extraction of the transient voltage signal in this study, it
is described in details in Section 3.1.
A better solution in comparison with ANN can be provided by
the Least Squares Support Vector Machine (LS-SVM). SVM and
LS-SVM are preferable, because their ability is independent of
input data dimension [27]. Due to these advantages, SVM and
LS-SVM are being more admissible than ANN for fault locating.
SVM algorithm solves a quadratic programming (QP); if m is the
number of training data, then training time and space complexity
will be commensurate with m
3
and m
2
respectively [28]. SVM
and LS-SVM algorithms suffer from long training time and large
amount of memory requirement for large dataset.
Refs. [2932] described Core Vector Machine (CVM) and Core
Vector Regression (CVR) as fast machine learning methods. CVM
is a combination of computational geometry techniques with
SVM training which can be used for with any linear and nonlinear
kernels.
The training time and space complexity of this algorithm is
commensurate with the number of training data (m). Although
CVM is nearly the same as SVM in terms of performance, it is faster,
and needs less training time and memory, and produces much few-
er support vectors when the number of training dataset augments
[29]. These merits make CVM more admissible than SVM, LS-SVM
and ANN for fault locating in large distribution systems.
This paper aims to extend the presented concepts in [22] to de-
rive a more regular pattern by using a new fault locator based on
EMD and CVM. The transient voltage waves of both single-end
and double-end networks for different faults are analyzed to show
the efciency of the proposed method for different conditions in
distribution systems. The energy of sub signals (IMFs) extracted
by EMD is computed and used as inputs for CVM. Based on the con-
cept of the EMD, different IMFs are extracted from the main tran-
sient voltage waves for fault locating in a large radial distribution
network. Also the results of CVR method are compared with ANN
to show the effectiveness of the proposed technique over the exist-
ing studies with ANN.
This paper is structured as follows. In Section 2, travelling wave
theory and PCFs are described. Section 3 introduces fault feature
extraction based on EMD. CVM and CVR algorithms are described
briey in Section 4. This is followed by CVR and ANN based fault
locators in Section 5. The results are discussed in Section 6.
2. Travelling wave theory and PCFs
After occurring faults, travelling waves propagate along
networks paths in both directions from the fault points, and are
reected until they arrive to the fault locations. Travelling waves
can be theoretically assessed by solving the voltage and current
equations along the lines considering the long line model. Solving
voltage equations can lead to following equations [33]:
[V(x)[ = e
x[c[
[V

[ e
x[c[
[V

[ (1)
The rst term is the propagated voltage wave and the second
term is the reected voltage wave. The propagation constant ma-
trix [c
2
] is computed by multiplying of the impedance of lines
and admittance matrices.
[c
2
[ = [Z[[Y[ (2)
where, in a single phase line, c is calculated as:
c = a jb =

(R jLx)(G jCx)
q
(3)
where a and b are attenuation and phase constants respectively.
For a better theoretical explanation of propagation role in distri-
bution systems and a reasonable approximation, various faults are
modeled. Fig. 1 shows IEEE 34-bus test distribution system which
is simulated by ATP/EMTP simulator. The data of IEEE 34-bus
distribution system have been noted in Appendix A. ATP/EMTP
simulator considers the skin effect and the effect of all three lines
of this system on faulted line.
In distribution systems, fault-originated waves are attenuated
along the paths with the factors such as transmission and reec-
tion coefcients at junctions as well as the resistance lines.
Depending on the network topology and fault location, the number
of effective paths leading to measuring points differs in length and
total attenuation in the rst reciprocation. In this section it is
shown that how the frequency and attenuation of paths can intro-
duce fault characteristic pattern in a typical distribution system.
In this section the concept of the proposed methodology is
demonstrated by developing propagation equations of a desired
fault location. Suppose a fault occurs at 0.2 km between buses
824 and 828 (i.e. case 1). Fault simulations in the time domain
indicated that recorded transient signals include characteristic fre-
quency (f
p
) related to each path. Travelling along the paths with
speed m cause each wave at measuring points has a transient fre-
quency in range of path characteristic [19].
f
p
=
m
n
p
L
p
(4)
where v is the travelling speed which is close to light speed, L
P
is the
length of the Pth path, and n
p
is the number of times that a wave has
to travel along Pth path until it obtains again the same polarity at
the measuring point. For the paths reach to loads n
p
= 2; and for
junctions and fault points n
p
= 4. Therefore, sixteen paths are iden-
tied for the rst case fault. Table 1 shows all paths that travelling
waves propagate through them with the characteristic frequencies
leading to the measuring points.
According to Table 1 paths #1#7 are associated to measuring
device of bus 800 and paths #8#16 are associated to measuring
device of bus 854. Forward and backward travelling waves which
216 B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225
are indexed with +i and i can be written for some sample buses as
follows.
Assuming V
+800
as forward travelling wave toward all branches,
then we can calculate the forward incident waves at some typical
buses, such as 810, 822 and the rst fault point.
V
810
= V
800
T
t
e
cL
7
(5)
V
822
= V
800
T
2
t
e
cL
1
(6)
V
fp1
= V
800
T
3
t
e
cL
6
(7)
Considering V
+854
for measuring device at bus 854, forward inci-
dent waves at buses 834, 848 and the rst fault point are calculated
as:
V
834
= V
854
T
2
t
e
cL
13
(8)
V
848
= V
854
T
3
t
e
cL
9
(9)
V
fp1
= V
854
e
cL
12
(10)
where L
P
is the length of Pth path, and T
t
is the transmission coef-
cient of Tee points. Also backward travelling waves to measuring
buses 800 and 854 can be expressed as follows:
V
800#7
= V
810
T
t
e
cL
7
(11)
V
800#1
= V
822
T
2
t
e
cL
1
(12)
V
800#6
= V
fp1
T
3
t
e
cL
6
(13)
V
854#13
= V
834
T
2
t
e
cL
13
(14)
V
854#9
= V
848
T
3
t
e
cL
9
(15)
V
854#12
= V
fp1
e
cL
12
(16)
where V
800#p
and V
854#p
are backward waves at measuring buses
through Pth path.
To express reection coefcient of a given bus, by deriving V
+800
from Eqs. (5)(7) and V
+854
from Eqs. (8)(10), reection coefcient
at ith bus for Pth path (C
i#p
) can be computed by Eqs. (11)(16).
For example for buses 810 and 834, Eqs. (17) and (18) are the
results.
V
800#7
V
800
=
V
810
V
810
T
2
t
e
2cL
7
; C
810#7
=
V
810
V
810
(17)
V
854#13
V
854
=
V
834
V
834
T
4
t
e
2cL
13
; C
834#13
=
V
834
V
834
(18)
Therefore, it can be said that attenuation of V
+i
along the Pth path
(n
i#p
) which depends on reection/transmission coefcients and
path length (L
P
) produces wave (V
#p
) [18]:
V
#p
= n
i#p
V
i
(19)
n
i#p
= C
i#p
T
m
t
e
2cL
P
(20)
where m is the twice of the Tee points number along Pth path be-
tween measuring bus and ith bus.
To clarify the propagation concept and related energy that trav-
elling wave takes as it propagates, it is assumed that there is per-
fect match at tee points as if the travelling wave sees similar
impedance at the all three entrances.
Impedance mismatch between the two different environments
where the wave propagates, determines the value of reection
for the point i and transmission coefcient T
i
of the travelling wave
at each point along the wave journey.
C
i
=
Z Z
0
Z Z
0
(21)
where Z
0
is the surge impedance of line and Z is the impedance
which travelling wave faces it.
Due to high frequency range of travelling waves at the TDR do-
main, C
i
for practical dynamic loads (Z Z
0
) and for fault resis-
800
806 808 812 814
810
802 850
818
824 826
816
820
822
828 830 854 856
852
832
838
862
840
836 860
834
842
844
846
848
858
Measurement
Measurement
864
Case 1
(Fault Point 1)
Case 2
(Fault Point 2)
a
b
Fig. 1. IEEE 34-bus distribution system.
Table 1
PCFs in case 1.
Path L
P
n
P
f
P
(kHz) Path L
P
n
P
f
P
(kHz)
1(800822) 18 2 8.30 9(854848) 8.5 2 17.60
2(800816) 12 4 6.25 10(854840) 8.5 2 17.60
3(800824) 15.1 4 4.96 11(854838) 9.7 2 15.46
4(800808) 6 4 12.50 12(854FP1) 4.8 4 15.60
5(800826) 15.9 2 9.43 13(854834) 6.8 4 11
6(800FP1) 15.3 4 4.90 14(854836) 8.2 4 9.14
7(800810) 7.8 2 19.20 15(854856) 4 2 37.50
8(854864) 5.4 2 27.70 16(854858) 5 4 15
B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225 217
tance very less than Z
0
(Z Z
0
) is about +1 (same polarity) and 1
(inverse polarity) respectively [15]. For the Tee connections with
the perfect match assumption the reection and transmission coef-
cient are as follows [34]:
C
i
=
1
3
; T
i
=
1 C
i
2
Thus, for paths lead to the Tee terminations and fault points, travel-
ling waves attain inverse polarity to the measuring buses at the rst
trip.
Due to these facts and considering repetitive trend of attained
waves, there will be too chaotic transient wave to be interpreted
using TDR technique. But the fact is that each wave at measuring
points leaves part of its energy in the range of characteristic fre-
quency so that the energy concentration around the PCFs can give
a useful pattern that reects the fault pattern [18,19].
For large distribution systems due to a large number of Tee
points and branches with long lengths, these patterns would be
chaotic for the faults which occur at the end of the system, so inac-
curate distinction is possible. So, the high percentage of the errors
in fault location can be reduced by placing another measuring
point at the end of the system to evaluate PCFs of related paths.
These patterns can be extracted from recorded transient waves
at measuring points by an appropriate tool described in the next
section. Four different faults are simulated on IEEE 34-bus test dis-
tribution system, and Fig. 2 shows the voltage transient waveforms
of the simulated faults.
In the rst fault due to the proximity of fault point to measuring
bus 800, the fault leaves a transient waveform in which there is a
signicant energy content that fullls the energy based locating
method. Another fault is simulated between buses 858 and 864
with lower energy percentage. As far as faults occur away from
bus 800, they leave weak transient waves with lower energy per-
centage at measuring bus 800; such as faults between buses
862838 and 860836. These cases leave weak transient wave-
forms that measuring bus 800 cannot distinguish what has been
taken place in the system.
Having the minimum number of measuring devices in the sys-
tem in order to cope with faded waveforms attained by faults in far
points, we used another measuring device placed at bus 854 based
on a heuristically search. This measuring device helps faults that
are far from bus 800 to be effectively recorded at nearer place
(i.e. bus 854). In the next section the energy of fault originated sig-
nals is assessed in the frequency domain to fulll our methodology.
3. Extraction of fault features based on EMD
The used methodology in this paper is described by inspecting
the energy percentage and frequency content of the faults. Suppose
two faults occur separately on the lines 824828 (0.2 km from bus
824) and 800808 (5 km from bus 800) as case 1 and case 2,
respectively. These faults are simulated with ATP/EMTP under
certain fault conditions (i.e. R
f
= 150 X, incepted at a = 2p/3 ele-
c.rad and double-end network) as a typical sample. However, in
the case study, different cases run to examine the proposed meth-
od comprehensively. The voltage fault transient signals recorded at
observation points consist of both time constant low frequency
component and time varying high frequency of transient distur-
bances [19].
As described previously, based on Eq. (4) characteristic frequen-
cies are the exclusive feature of the fault locations and correspond-
ing to different locations of the faults. Tables 1 and 2 show PCFs for
cases 1 and 2.
5 5.2 5.4 5.6 5.8 6 6.2
0.5
1
1.5
2
x 10
4
Time (ms)
V
o
l
t
a
g
e

(
v
o
l
t
)
Location Fault : 806-808
Location Fault : 858-864
Location Fault : 862-838
Location Fault : 860-836
Fig. 2. Recorded voltages at measuring device bus 800.
Table 2
PCFs in case 2.
Path L
P
n
P
f
P
(kHz) Path L
P
n
P
f
P
(kHz)
1(854FP2) 15.1 4 4.96 9(854858) 5 4 15
2(854810) 15.9 2 9.43 10(854864) 5.4 2 27.70
3(854822) 14.1 2 10.64 11(854834) 6.8 4 11
4(854808) 14.1 4 5.31 12(854848) 8.5 2 17.60
5(854816) 8.1 4 9.26 13(854836) 8.2 4 9.14
6(854824) 5 4 15 14(854840) 8.5 2 17.60
7(854826) 5.8 2 25.86 15(854838) 9.7 2 15.46
8(854856) 4 2 37.50 16(800FP2) 5 4 15
0 1 2 3 4 5 6 7 8 9 10
x 10
4
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
N
o
r
m
a
l
i
z
e
d

E
n
e
r
g
y

S
p
e
c
t
r
u
m
Bus 854
Bus 800
#15
#8
#4
#2
#3
#6
#9
#10
#11
#16
#12
FFT (IMF1)
0 1 2 3 4 5 6 7 8 9 10
x 10
4
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
N
o
r
m
a
l
i
z
e
d

E
n
e
r
g
y

S
p
e
c
t
r
u
m
Bus 854
Bus 800
#12
#7
#14
#13
#1
#5
FFT(IMF2)
0 1 2 3 4 5 6 7 8 9 10
x 10
4
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
N
o
r
m
a
l
i
z
e
d

E
n
e
r
g
y

S
p
e
c
t
r
u
m
Bus 854
Bus 800
#14
#4
#2
#12
FFT(IMF3)
(a)
(b)
(c)
Fig. 3. FFT analysis of IMFs observed by buses 854 and 800 in case 1.
218 B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225
These theoretical frequencies also can be observed by Fast
Fourier Transform (FFT) analysis of extracted IMFs from recorded
transient signals at measuring buses 800 and 854. It is proven in
this paper that FFT can provide meaningful pattern for the pro-
posed methodology.
In order to extract essential information based on the PCFs of
the faults, EMD is used to extract different IMFs and then FFT is
implemented to show the frequency content and the energy
spectrum percentage. Figs. 3 and 4 show the normalized energy
spectrum of IMFs versus the specied frequencies in cases 1 and
2. They illustrate that the energy density of IMFs around the fre-
quency of different paths is higher than other appeared
frequencies.
By comparing the energy content of IMFs from Fig. 3, the effect
of proximity of faults to measuring devices is shown.
As it can be seen in Fig. 3, by FFT analysis of the rst IMF, var-
ious paths frequencies are identied. Those frequencies which cor-
respond to characteristic frequency of faulty paths have high
energy content than other frequencies, as it is shown in Table 1.
For instance, frequencies for paths #12 and #6 are15.6 kHz and
4.9 kHz, respectively. The paths #9, #10, #11 and #16 for recorded
signal from bus 854 and the paths #2 and #3 for recorded signal
from bus 800 have energy content near peak values. This status
is also true for a fault near measuring bus 800. Fig. 4 shows FFT
analysis of IMFs for case 2 with similar pattern to case 1. The
characteristic frequencies of all paths presented in Table 2, can
be specied in Fig. 4, where it has consecutive peaks of energy.
The energy content of IMFs for bus 800 is higher than bus 854
due to the vicinity of the second case fault to bus 800. Also, the
frequencies 4.96 kHz and 15 kHz related to paths #1 and #16 with
high energy content are shown. Although exact detection of char-
acteristic frequencies is not the goal of this paper, as it is demon-
strated in Figs. 3 and 4, the approximate value of appeared
frequencies is shown for different paths.
In this paper IMF1 to IMF3 are considered to describe how PCFs
with the energy percentage of recorded signals can give a proper
pattern to detect fault locations. It should be noted that IMFs have
different frequency content; their frequency decreases while their
order increases, so IMFs with high order are not proper to be used
in fault locating. Because, the range of PCFs varies approximately
from 1 kHz to 750 kHz for the faults at 44.8 km and 0.1 km from
bus 800. Thus, the rst ve IMFs are considered to be used to com-
plement the methodology.
3.1. Empirical mode decomposition method
The approach of EMD introduced in [35] as an algorithm for
processing multi component signals is based on interpolation
points criterion that decomposes theminto a number of amplitude
and frequency modulated with zero mean, called Intrinsic Mode
Functions (IMFs) [35,36]. Due to the interesting feature of EMD
[25], it can be applied to nonlinear or non-stationary signals and
provides useful results. Also the extracted signals are monocompo-
nent and narrowband, and can provide meaningful frequencies
estimated by Hilbert transform or any other techniques [37].
EMD expresses the main signal based on the expanding of basic
functions which are signal-dependent. These functions are esti-
mated by an iterative procedure called sifting.
Decomposition of a multi component signal f(t) into K number
of zero average and narrowband (IMFs) sub signals g
i
(t) plus a non-
zero mean and low order signal r(t) is done by EMD method as
follows.
f (t) =
X
K
i=1
g
i
(t) r(t) (22)
Each IMF or g
k
(t) is obtained by an iterative process called sift-
ing applied to the residual multi component signal as in the follow-
ing equation:
f
k
(t) =
f (t); k = 1
f (t)
P
k1
i=1
g
i
(t); k P2
(
(23)
An estimation of the kth IMF during the nth iteration of the sift-
ing process is calculated by (24).
1 2 3 4 5 6 7 8 9 10
x 10
4
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
N
o
r
m
a
l
i
z
e
d

E
n
e
r
g
y

S
p
e
c
t
r
u
m
Bus 800
Bus854
#16
#1
#4
#2
#5
#13
#3
#11
#8
FFT(IMF1)
#6
#9
#15
#10
#12
#14
#7
0 1 2 3 4 5 6 7 8 9 10
x 10
4
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
N
o
r
m
a
l
i
z
e
d

E
n
e
r
g
y

S
p
e
c
t
r
u
m
Bus 800
Bus 854
#16
#1
#4
#6
#9
#15
#2
#5
#13
#12
#14 #7
#10
#8
FFT(IMF2)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 10
4
0
0.2
0.4
0.6
0.8
1
Frequency (Hz)
N
o
r
m
a
l
i
z
e
d

E
n
e
r
g
y

S
p
e
c
t
r
u
m
Bus 800
Bus 854
#2
#5
#13
#16
#12
#14
FFT(IMF3)
(a)
(b)
(c)
Fig. 4. FFT analysis of IMFs observed by buses 854 and 800 in case 2.
Fig. 5. Quantities of the EMD method for a multi component signal.
B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225 219
g
k
n
(t) = g
k
n1
(t) m
k
n
(t) (24)
where g
k
j
(t) is the temporal estimation of the kth IMF at the jth iter-
ation and m
k
j
(t) is an estimation of the local mean of g
k
j
(t).
According to Fig. 5, the mean envelope m
k
j
(t) is considered as
the average of two envelopes, an upper and a lower envelope,
which enfold the kth IMF g
k
j
(t). Generally, the envelopes are com-
posed based on the following expression.
Firstly, some time instances s
u
= [s
u1
, . . . , s
uM
] and s
l
= [s
l1
, -
. . . , s
lL
] called nodes and the corresponding IMF values g
k
j
(s
u
) and
g
k
j
(s
l
) called interpolation points should be specied. Nodes for
upper and lower envelopes might be determined according to
being the maximum dg
k
j
(t)=dt = 0; d
2
g
k
j
(t)=dt
2
< 0

and mini-
mum dg
k
j
(t)=dt = 0; d
2
g
k
j
(t)=dt
2
> 0

extrema of g
k
j
(t) for upper
and lower envelopes. Interpolation points are determined based
on a predetermined scheme (e.g. natural cubic spline interpola-
tion). As can be seen in Fig. 5, utilizing these points, two upper
and lower envelopes I
su
(t) and I
s
l
(t) are composed. Finally, the cur-
rent estimation of the local mean is given by Eq. (25).
m
k
j
(t) =
I
su
(t) I
s
l
(t)
2
(25)
The sifting process will nd the kth IMF after N sifting iterations
when a stopping criterion is fullled g
k
(t) = g
k
N
(t)

. The number of
sifting process might be chosen either statically or dynamically
[38,39].
The estimation of g
k
(t) after N iterations is described mathemat-
ically as follows:
v For the rst iteration: g
k
1
(t) = f
k1
(t) m
k
1
(t).
v For the second iteration: g
k
2
(t) = g
k
1
(t) m
k
2
(t).
v For the Nth iteration: g
k
N
(t) = g
k
N1
(t) m
k
N
(t).
For the rst iteration the residual signal f
k1
(t) is used as initial IMF
estimation for g
k
1
(t). Therefore, it can be concluded that according
to such a reformulation, the kth IMF can be obtained directly from:
g
k
N
(t) = f
k1
(t)
X
N
j=1
m
k
j
(t) (26)
Figs. 6 and 7 show the results of EMD to extract the rst three
IMFs from the observed transient signals at buses 854 and 800 in
cases 1 and 2. As it is shown, the frequency content of IMFs de-
creases when its order increases.
4. CVM and CVR algorithms
SVM has been proposed to solve classication, data mining and
other machine learning problems with good generalization perfor-
mance and ease of implementation in several application elds
[29]. SVM training is formulated as a quadratic programming prob-
lem. This QP problem has been optimized by numerical solvers.
However, the complexity of a QP problem is about O(m
3
) in re-
quired time and O(m
2
) in required space, where m is the number
of training samples. Therefore, SVM is computationally expensive
for large data sets.
4.1. CVM
CVM is a technique which provides equivalence between Mini-
mum Enclosing Ball (MEB) and QP problem of SVM [3033]. In
-5000
0
5000
-5000
0
5000
0 1 2 3 4 5 6 7 8 9 10
-4000
0
4000
Time (ms)
V
o
l
t
a
g
e
IMF1
IMF2
IMF3
Fig. 6. Extracted IMFs from transient voltage observed by bus 854 with EMD method (case 1).
-5000
0
5000
-5000
0
5000
V
o
l
t
a
g
e
0 1 2 3 4 5 6 7 8 9 10
-5000
0
5000
Time (ms)
IMF1
IMF2
IMF3
Fig. 7. Extracted IMFs from transient voltage observed by bus 800 with EMD method (case 2).
220 B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225
CVM by reformulating SVMs QP as a MEB problem, an efcient
(1 + e)-approximation algorithm has been applied to MEB problem
to obtain a close to optimal solution of SVM. The approximate MEB
algorithm expands the current core-sets size by including the fur-
thest point away from the current center.
4.1.1. MEB
Given a nite set of data points S = {x
i
e R
d
}, the MEB of S (spec-
ied as MEB(S)) has been dened as the smallest ball that contains
all the points in S.
Let k be a kernel function with features of map , where
k(x
1
; x
2
) = (x
i
)(x
j
)). Then the primal MEB problem in kernel
feature space to nd B(C, R), where C is the center of B and R is
the radius of B. At tth iteration, B is expanded with the points which
are farthest from C
t
and fall outside the (1 + e)-approximation of B.
This problem can be formulated as follows:
B(C; R) = minR
2
(27)
s:t: |c /(x
i
)|
2
6 R
2
\i
The Lagrange dual of this problem can be expressed as follows
[30]:
max
a
i
X
m
i=1
a
i
k(x
i
; x
j
)
X
m
i;j=1
a
i
a
i
k(x
i
; x
j
) (28)
s:t:
X
m
i=1
a
i
= 1; a
i
P0 i = 1; . . . ; m
When the involved kernel fullls k(x, x) = K a constant, any SVMs
QP can be identied as an MEB. Then, the Lagrange dual problem
of the MEB problem can be rewritten as the following QP problem:
min
a
i
X
m
i;j=1
a
i
a
j
k(x
i
; x
j
) (29)
s:t:
X
m
i=1
a
i
= 1; a
i
> 0 i = 1; . . . ; m
Therefore, CVM algorithm can be summarized, as follows:
v Step 1: Initialize S
0
= {u(Z
0
)}, C
0
= u(Z
0
) and R
0
= 0.
v Step 2: End if no u(Z) falls outside B(C
t
, (1 + e)R
t
).
v Step 3: Find Z
t
so that u(Z
t
) is farthest away from C
t
. Set
S
t+1
= S
t
{u(Z
t
)}.
v Step 4: Find MEB (S
t+1
).
v Step 5: t = t + 1 and go to step 2.
4.1.2. Approximate MEB
After reformulating SVMs QP problem as a MEB problem, CVM
can solve SVM problem with the (1 + e)-approximate MEB
problem.
The MEB algorithm can terminate within O
1
e

, iterations. There-
fore, the time and space complexities of CVM is O
m
e
2

1
e
4

and O
1
e
2

,
respectively.
4.2. CVR
CVR is described based on CVM concept, in this section. Here,
each u(x
i
) increases by an extra D
i
e R to form [u(x
i
)
/
, D
i
]
/
and then
the MEB for these increased points is calculated based on the con-
straint that compels the last coordinate of the balls center to be
zero. Fig. 8 depicts the center-constrained MEB problem [30].
The MEB equations can be written, as follows:
min R
2
(30)
s:t: |C /(x
i
)|
2
D
2
i
6 R
2
i = 1; . . . ; m (31)
where D = D
2
1
; . . . ; D
2
m

/
P0.
Modications on CVM algorithm which has been presented in
Section 4.1.1, lead to CVR algorithm, as follows:
v Step1: Initialize S
0
= {u(Z
0
)}, C
0
= u(Z
0
) and R
0
= 0.
v Step2: End if no u(Z) falls outside B(C
t
, (1 + e)R
t
). Compute dis-
tance between C
t
and any point Z
i
using.

|C
t
u(Z
i
)|
2
D
2
i
q
.
v Step 3: Find Z
t
so that u(Z
t
) is farthest away from C
t
. Set
S
t+1
= S
t
{u(Z
t
)}.
v Step4: Find MEB (S
t+1
).
v Step5: t = t + 1 and go to step 2.
5. CVR based fault locator
As mentioned before, the energy percentage of IMFs is selected
as fault locators inputs, due to effective feature which reects the
fault behavior based on the energies around the PCFs.
In this section, a CVR-based algorithm is proposed for fault
locating in large distribution networks. Based on the proposed
algorithm the CVR-based fault locator shown in Fig. 9 is trained
with extracted energies from voltage transient signals, and a num-
ber that can be used to determine the fault location is received as
output.
This algorithm has three steps:
v Step 1: Database creation
Database creation is the rst step in the proposed algorithm. All
data are created using valid simulation results obtained by ATP/
EMTP and MATLAB software. The database covers the whole ex-
pected range of faults in the simulated distribution system. The
various fault locations, fault resistances and fault inception angles
are used for creation the database. This database has randomly
split into training (i.e. 70%) and (i.e. 30%) testing datasets. The
training dataset is used for the CVR training and testing dataset
is used for performance assessment of the trained CVR.
v Step 2: CVR training
The second step in the proposed algorithm is CVR training.
Choosing optimal parameters and appropriate kernel function is
one of the principal aspects of the CVR application. Regulation
parameter (l) and scale parameter (C) are CVR parameters which
should be optimally chosen. Fig. 8. Center-constrained MEB problem.
B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225 221
v Step 3: Performance assessment
The third step includes performance assessment of the algo-
rithm. Performance of the proposed algorithm is assessed using
the following indices [40]:
v Max error (ME)
This index is dened, as follows:
ME = Max[D
act;i
D
pre;i
[ i = 1; . . . ; T (32)
where D
act,i
is the actual value of the distance between the fault
location and slack bus for the ith point in testing dataset and D
pre,i
is the output of CVR model and T is the number of testing datasets.
v Mean square error (MSE)
MSE index is used to quantify the differences between actual
values of fault locations (distances) and output of the used CVR
model. Therefore, the average of the squared error is measured,
as follows:
MSE =
1
T
X
T
i=1
(D
act;i
D
pre;i
)
2
(33)
Having excellent performance of the CVR, this index tends to be
closer to 0.
v Squared correlation coefcient (SCC)
This index is applied to specify variability in a dataset. SCC is an
appropriate index for showing the correlation between predictive
and actual values.
SCC = 1
SS
err
SS
tot
(34)
SS
err
=
X
T
i=1
(D
act;i
D
pre;i
)
2
SS
tot
=
X
T
i=1
(D
act;i
D
act
)
2
where D
act
=
1
T
P
T
i=1
D
act;i
. The range of this index is between 0 and 1.
High ability to predict output values leads it to be closer to 1.
The results of CVR method are compared with ANN to show the
effectiveness of the proposed technique over the existing studies
with ANN.
The models which are based on ANN are accurate and fast en-
ough, however, there are some disadvantages such as requiring
large amount of memory, being trapped on locally minimum, over
tting and long training time.
Simulations with ANN are performed with two different struc-
tures; single-hidden layer structure with 15 neurons and double-
hidden layer with 12 and 4 neurons, respectively. Both structures
have 10 neurons in input and one neuron for output layer. The
number of neurons is selected based on trial and error.
Hyperbolic tangent sigmoid (Tansig) function (35) is used for
the hidden layers and a linear transfer function (Purelin) applied
for the output layer.
u(s) =
1
1 e
s
(35)
These ANN are trained using the Bayesian regularization back
propagation algorithm. As for each fault ve IMFs are separately
provided by EMD for transient signals of two measuring devices
800 and 854, we have (2 5) data as input. The assessed ANN-
based fault locator can be shown in Fig. 9 with ANN block instead
of CVR.
For establishing the fault locator, the fault location must be
coded to a range of real numbers reecting the possible locations
of all existing branches. To achieve the coded outputs for interpret-
ing where faults occur, the output values are allocated in the space
of [0:44.8]. According to Table 3, place designation is performed
from slack bus (i.e. bus 800) to the bus at the end of the test system
(i.e. bus 840) respectively by dividing the system to 15 lines with
different lengths. The sum of lines length is equal to the total
length of the system (i.e. 44.8 km). For instance if the fault locator
output is the real number 5.5, it means that the fault point is on the
line 800808 (0 < 5.5 < 6) and its distance from bus 800 is 5.5 km.
Also the number 12.9 means that the fault occurs between buses
Fig. 9. The block diagram of the proposed fault locator.
Table 3
Place designation of all locations of the network.
Location Output Location Output
800808 [0 6] 854858 [32.7 37.7]
808816 [6 12] 858864 [37.7 38.1]
810808 [12 13.8] 858834 [38.1 39.9]
816822 [13.8 19.8] 834848 [39.9 41.6]
816824 [19.8 22.9] 834836 [41.6 43]
824826 [22.9 23.7] 836838 [43 44.5]
824854 [23.7 28.7] 836840 [44.5 44.8]
854856 [28.7 32.7]
Table 4
Comparison of CVR results with ANN results for single-end network using EMD
method.
Indices ME MSE SCC Training time (s)
CVR 0.497 0.114 0.994 5.85
ANN (double-layer) [22] 0.956 0.390 0.972 422
ANN (single-layer) 1.210 0.430 0.961 202.5
Table 5
Comparison of CVR results with ANN results for double-end network, using EMD
method.
Indices ME MSE SCC Training time (s)
CVR 0.513 0.105 0.998 6.2
ANN (double-layer) [22] 0.900 0.290 0.980 470
ANN (single-layer) 0.932 0.350 0.972 228
222 B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225
810 and 808 (12 < 12.9 < 13.8), and it has the distance of 0.9 km
from bus 810. There may be more than one point with equal dis-
tances from the slack bus; by output coding these points can be
truly distinguished to locate accurate place of faults. For example,
in Fig. 1 the points a and b with equal distances from slack bus
are shown. If they happen at 0.2 km from bus 834, according to Ta-
ble 3 outputs will be 41.8 for fault point a and 40.1 for fault point
b.
Thus, by processing energy contents of two transient signals re-
ceived at buses 800 and 854 and output coding, fault locators nd
just one output for each fault.
6. Results
In order to show the effectiveness of the proposed method, it is
applied to IEEE 34-bus test system. To have a comprehensive eval-
uation of the system, two cases are performed. For case 1, IEEE 34-
bus test system operates as a single-end network. In case 2, IEEE
34-bus test system operates as a double-end network with embed-
ding another power supply on bus 840 at the end of the system.
The details of the modeling procedure for each step are described
in the following parts:
v CVR training
To select the best kernel function and their parameters for CVR
training a grid search is run. It was observed that the best predic-
tion capability for the CVR-based model gained with the Laplacian
kernel function with l = 0.894 and C = 447,000.
v Performance evaluation
Three indices ME, MSE and SCC which were introduced in Sec-
tion 5, are used to assess the performance of the CVR-based model.
ME, MSE and SCC for the two case studies are shown in Tables 4
and 5. In both cases, it is seen that the proposed CVR-based model
has very short training time in comparison with ANN-based mod-
els. Also, the CVR-based model can predict the fault location for
both case studies precisely.
The training time by CVR is about 6 s in both cases; however it
dramatically increases to 202 s and 422 s for single and double-
layer ANN-based models respectively. Although the double-layer
ANN has longer training time than single-layer one, but the predic-
tion capability is better.
By comparing three indices ME, MSE and SCC it can be observed
that CVR has the best performance. In all cases, CVR approach has
the maximum value for SCC and minimum values for MSE and ME
compared to the ANN-based models. The efcacy of the CVR over
ANN for the fault location prediction can be revealed by Tables 4
and 5.
Also some individual estimation errors are presented in Table 6
for a certain case. It contains 21 test points to evaluate estimation
error of CVR and ANN based fault locators for double end network.
The rst column reects the real location of faults where they have
occurred. The outputs of the CVR based fault locators are shown in
the second columns, and third column presents estimation errors.
The fourth and fth columns show estimation results and estima-
tion errors for ANN-based fault locator. Among these test faults,
the maximum estimation error for CVR and ANN are 393 m for
the fault at 2.2 km from bus 800 and 689 m for the fault at
3.2 km from bus 854, respectively. So, the priority of CVR over
ANN can be revealed by test results of Table 6.
7. Conclusion
In this paper a new method based on EMD and CVR was pre-
sented for fault locating in large distribution systems. This method
will excel at the ANN-based methods and the current methods
based on the measurement of post-fault line impedance and the
reection of travelling waves between different points. The meth-
od of this paper uses the energy spectrum of the recorded transient
voltages at two observation points. It was proven that such an en-
ergy spectrum has cumulated around the path characteristic fre-
quencies. In addition, depending on the network conguration
and fault location the recorded transient voltages include certain
set of path characteristic frequencies with useful feature obtained
by EMD method and FFT technique. Using the energy spectrum of
the different IMFs of transient voltages gives specic information
about location of fault.
This information is used for training of the proposed CVR. This
method was tested on IEEE 34-bus distribution system with both
CVR and ANN methods to have comparative view in different case
studies. Simulation results show the usefulness and accuracy of the
proposed algorithm for fault locating in large distribution systems.
Table 6
Estimation errors for a certain case.
Real location of fault (km) CVR-based fault locator ANN-based fault locator
Estimated location of fault (km) Estimation error (km) Estimated location of fault (km) Estimation error (km)
0.6 0.915 0.3150 1.150 0.5500
2.2 2.593 0.3930 2.840 0.6400
4.8 5.084 0.2840 5.298 0.4980
7.2 7.529 0.3290 7.777 0.5770
8.8 9.122 0.3220 9.364 0.5640
11.5 11.771 0.2710 11.975 0.4750
13 13.257 0.2570 13.450 0.4500
17.1 17.304 0.2040 17.457 0.3570
19.4 19.700 0.3000 19.926 0.5260
22.8 23.080 0.2800 23.291 0.4910
28.3 28.654 0.3540 28.921 0.6210
31.9 32.265 0.3650 32.589 0.6890
34.5 34.784 0.2840 34.998 0.4980
36.7 37.030 0.3300 37.278 0.5780
38.9 39.221 0.3210 39.463 0.5630
39.9 40.207 0.3070 40.438 0.5380
41.8 42.097 0.2970 42.321 0.5210
42.6 42.890 0.2900 43.108 0.5080
43.4 43.625 0.2250 43.794 0.3940
44 44.388 0.3880 44.680 0.6800
44.65 44.984 0.3340 45.236 0.5860
B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225 223
Acknowledgments
The author gratefully acknowledges the technical support given
by Mr. Hadi Fotoohabadi, department of power and control Engi-
neering, Shiraz University, and Mr. Hossein Khorramdel, School
of Electrical and Computer Engineering, Islamic Azad University,
Safashahr branch, Safashahr, Fars, Iran.
Appendix A
Data of the IEEE 34-bus test system are presented in Tables A1
A3. The over head line structure of the test system corresponds to
Fig. A1.
References
[1] Takagi T et al. Development of a new type fault locator using the one-terminal
voltage and current data. IEEE Trans Power Apprat Syst 1982;101(8).
[2] Choi MS et al. A new fault location algorithm using direct circuit analysis for
distribution systems. IEEE Trans Power Del 2004;19(1):3541.
[3] Girigis A, Fallon CM. Fault location technique for radial and loop transmission
systems using digital fault recorded data. IEEE Trans Power Del
1992;7(4):193645.
[4] Coser J, do Vale DT, Rolim JG. Design and training of articial neural networks
for locating low current faults in distribution systems. In: Proc IEEE int conf
intelligent systems applications to power systems (ISAP); 2007. p. 16.
[5] Thukaram D, Khincha HP, Vijaynarasimha HP. Articial neural network and
support vector machine approach for locating faults in radial distribution
systems. IEEE Trans Power Del 2005;20(2):71021.
[6] Gracia J, Mazon AJ, Zamora I. Best ANN structures for fault location in single-
and double-circuit transmission lines. IEEE Trans Power Del
2005;20(4):238995.
[7] Ngu EE, Ramar K. A combined impedance and traveling wave based fault
location method for multi-terminal transmission lines. Int J Electr Power
Energy Syst 2011;33:176775.
[8] Jayabharata Reddy M, Mohanta DK. A wavelet-fuzzy combined approach for
classication and location of transmission line faults. Int J Electr Power Energy
Syst 2007;29:66978.
[9] Tian Yu, Chunju Fan, Gong Zhendong. A study on accurate fault location
algorithm for parallel transmission line with a teed connection. Int J Electr
Power Energy Syst 2010;32:697703.
[10] Sousa Martins L, Martins JF, Ferna Pires V, Alegria CM. A neural space vector
fault location for parallel double-circuit distribution lines. Int J Electr Power
Energy Syst 2005;27:22531.
[11] Bo ZQ, Weller G, Redfern MA. Accurate fault location technique for distribution
system using fault-generated high-frequency transient voltage signals. IEE
Proc-Gener Transm Distrib 1999;146(1):739.
[12] Magnago FH, Abur A. Fault location using wavelets. IEEE Trans Power Del
1998;13(4):147580.
[13] Rajendra S, Mclaren PG. Traveling-wave techniques applied to the protection
of teed circuits: principle of traveling wave technique. IEEE Trans Power
Apprat Syst 1985;pas-104(12):354450.
[14] Shehab-Elidin EH, Mclaren PG. Traveling-wave distance protection problem
areas and solutions. IEEE Trans Power Del 1988;3(3):894902.
[15] Moshtagh J, Aggarwal RK. A new approach to ungrounded fault location in a
three-phase underground distribution system using combined neural
networks & wavelet analysis. In: Proc IEEE Can conf electrical and computer
engineering (CCECE); 2006. p. 37681.
[16] Ngaopitakkul, Pothisarn C. Discrete wavelet transform and back-propagation
neural networks algorithm for fault location on single-circuit transmission
line. In: Proc IEEE int conf robotics and biomimetics (ROBIO); 2008. p. 16138.
[17] Tawk MM, Morcos MM. On the use of Prony method to locate faults in loop
systems by utilizing modal parameters of fault current. IEEE Trans Power Del
2005;20(1):5324.
[18] Borghetti et al. On the use of continuous-wavelet transform for fault location
in distribution power networks. Electr Power Energy Syst 2006;28:60817.
[19] Borghetti et al. Continuous-wavelet transform for fault location in distribution
power networks: denition of mother wavelets inferred from fault originated
transients. IEEE Trans Power Syst 2008;23(2):3808.
[20] Xiaoli Z et al. Fault location using wavelet energy spectrum analysis of
traveling waves. In: Proc IEEE int power engineering conf (IPEC); 2007. p.
112630.
[21] Navaneethan S et al. Automatic fault location for underground low voltage
distribution networks. IEEE Trans Power Del 2001;16(2):34651.
[22] Pourahmadi-Nakhli M, Safavi AA. Path characteristic frequency-based fault
locating in radial distribution systems using wavelets and neural networks.
IEEE Trans Power Del 2011;26(2):77281.
[23] Kopsinis Yannis, McLaughlin Steve. Improved EMD using doubly-iterative
sifting and high order spline interpolation. EURASIP J Adv Signal Process 2008:
8p, http://dx.doi.org/10.1155/2008/128293 [Article ID 128293].
[24] Kopsinis Y, McLauglin S. Enhanced empirical mode decomposition using a
novel sifting-based interpolation points detection. In: Proceedings of the14th
IEEE/SPWorkshop on statistical signal processing (SSP 07), Madison, WI, USA,
August 2007. p. 7259.
[25] Kopsinis Y, McLaughlin S. Investigation and performance enhancement of the
empirical mode decomposition method based on a heuristic search
optimization approach. IEEE Trans Signal Process 2008;56(1):113.
[26] Rilling G, Flandrin P. One or two frequencies? The empirical mode
decomposition answers. IEEE Trans Signal Process 2008;56(1):8595.
[27] Huo HB, Zhu XJ, Cao GY. Nonlinear modeling of a SOFC stack based on a least
squares support vector machine. J Power Sources 2006;162:12205.
[28] Burges CJC. A tutorial on support vector machines for pattern recognition. Data
Min Knowl Discov 1998;2(2):12167.
[29] Tsang IW, Kwok JT, Lai KT. Core vector regression for very large regression
problems. In: Proceedings of the 22nd conference (int) on machine learning,
Bonn, Germany; 2005. p. 91320.
[30] Tsang IW, Kocsor A, Kwok JT. Large-scale maximum margin discriminate
analysis using core vector machines. IEEE Trans Neural Netw
2008;19(4):61024.
Table A1
Line data.
Line type Resistance (X/
km)
Inner radius (cm) Outer radius
(cm)
J Marti 0.065 0.14 0.8
Line between
buses
Length (km) Line between
buses
Length (km)
800808 6 854858 5
808810 1.8 858864 0.4
808816 6 858834 1.8
816822 6 834846 1.7
816824 3.1 834836 1.4
824826 0.8 836862 1.5
824854 5 836840 0.3
854856 4
Table A2
Load data.
Bus RL-load(3u)
810, 826, 846, 864, 840 R = 0.0196(X), L = 3e4(H)
822, 856, 838 R = 0.04(X), L = 3e4(H)
Table A3
Transformer data.
Bus V
1
/V
2
(kV)
Primary Secondary Type
800 345/
24.9
R = 0.001(X),
L = 0.01(H)
R = 0.002(X),
L = 0.01(H)
D-Y-
1
810, 826, 846, 864 840,
822, 856, 838
24.9/
0.220
R = 0.001(X),
L = 0.01(H)
R = 0.001(X),
L = 0.01(H)
Y-Y
Fig. A1. The over head line structure of the test system.
224 B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225
[31] Tsang IW, Kwok JT, Zurada JM. Generalized core vector machines. IEEE Trans
Neural Netw 2006;17(5):112640.
[32] Tsang IW, Kocsor A, Kwok JT. Simpler core vector machines with enclosing
balls. In: Proceedings of the 24th conference (int) on machine learning,
Corvallis; 2007. p. 9118.
[33] Franklin GA. Using modal analysis to estimate received signal levels for a
power-line carrier channel on a 500-kV transmission line. IEEE Trans Power
Del 2009;24(4):244654.
[34] Papazyan R, Eriksson R. Calibration for time domain propagation constant
measurements on power cables. IEEE Trans Instrum Meas 2003;52(2):4158.
[35] Huang NE, Shen Z, Long SR, et al. The empirical mode decomposition and the
Hilbert spectrum for nonlinear and non-stationary time series analysis. Proc R
Soc Lond A 1998;454(1971):90395.
[36] Cohen L. Time-frequency analysis. Upper Saddle River (NJ, USA): Prentice-Hall;
1995.
[37] Huang NE, Wu Z. An adaptive data analysis method for nonlinear and
nonstationary time series: the empirical mode decomposition and Hilbert
spectral analysis. In: Proc 4th int conf wavelet anal its appl (WAA), Taipa,
Macau; 2005.
[38] Rilling G, Flandrin P, Gonalvs P. On empirical mode decomposition and its
algorithms. In: Proceedings of the 6th IEEE/EURASIP workshop on nonlinear
signal and image processing (NSIP 03), Grado, Italy, June 2003.
[39] Huang NE, Wu M-LC, Long SR, et al. A condence limit for the empirical mode
decomposition and Hilbert spectral analysis. Proc R Soc A
2003;459(2037):231745.
[40] Mohammadi M, Raoofat M, Marzooghi H, Gharehpetian GB. Nonlinear
multivariable modeling of solid oxide fuel cells using core vector regression.
Int J Hydrogen Energy 2011;36:1253848.
B. Khorramdel et al. / Electrical Power and Energy Systems 58 (2014) 215225 225