Electrical Power and Energy Systems 93 (2017) 97108

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Features-clustering-based earth fault detection using singular-value

decomposition and fuzzy c-means in resonant grounding distribution
systems
Mou-Fa Guo a,b, Nien-Che Yang a,
a
Department of Electrical Engineering, Yuan Ze University, 135, Yuan-Tung Road, Chung-Li, Taoyuan 32003, Taiwan
b
Department of Electric Power Engineering, Fuzhou University, Fuzhou 350116, China

a r t i c l e i n f o a b s t r a c t

Article history: The transient zero-sequence current of each feeder in a resonant grounding system is characterized by
Received 8 November 2016 nonlinearity and nonstationarity when a single-phase-to-ground fault occurs. Because there is a signifi-
Received in revised form 11 February 2017 cant difference between the fault transient zero-sequence current waveforms of the fault feeder and the
Accepted 14 May 2017
sound feeders, a new fault feeder detection method is presented, based on a time-frequency matrix (TFM)
and polarity distribution matrix (PDM) singular values clustering algorithm. By applying a Hilbert-Huang
transform band-pass filter and waveform transformation to the transient zero-sequence current wave-
Keywords:
form of each feeder, the TFM and PDM can be constructed, which are decomposed by singular-value
Resonant grounding system
Fault feeder detection
decomposition (SVD). Moreover, the normalized singular values of the TFM and PDM are merged together
Transient zero-sequence current and are used to form the amplitude-polarity feature matrix (APFM). Thus, the feature quantities including
Hilbert-Huang transform band-pass filter the amplitude and polarity information of each fault transient zero-sequence current waveform are
Time-frequency matrix obtained. Then, fuzzy c-means clustering is applied to the APFM so as to detect the fault feeder by divid-
Singular-value decomposition ing the fault feeder and sound feeders into two categories without a certain threshold setting.
Fuzzy c-means clustering Simulations were carried out via PACAD/EMTDC and a physical system under various kinds of fault con-
ditions and factors including asynchronous sample, two-point-grounding fault, and arc fault. Simulated
results show that the proposed method has the characteristics of high accuracy and reliability in earth
fault feeder detection.
2017 Elsevier Ltd. All rights reserved.

1. Introduction characteristics. The problem of fault feeder detection [27] has

The line voltage of resonant grounding systems [1] is symmet-
rical when a permanent single-phase-to-ground fault occurs. It is
allowed to continue running for approximately 1 to 2 h according
to the technical standards. In order to prevent the sound phase
voltage from increasing and the insulation from weakening, which
can cause the fault to worsen, the fault feeder must be determined
and removed as soon as possible. When a single-phase-to-ground
fault occurs in a resonant grounding system, it results in a complex
fault situation, faint fault current, and the influences that come
into being because of the feeder structure, feeder parameters,
and nonlinear characteristics of the voltage and current trans-
former, as well as the electromagnetic interference. There are sig-
nificant differences in fault transient zero-sequence currents in the
areas of spectrum features, energy distribution, and attenuation
Corresponding author.
E-mail address: ncyang@saturn.yzu.edu.tw (N.-C. Yang).
not been well solved for a long time. Thus, it is necessary to inves-
tigate a single-phase-to-ground fault detection method for reso-
nant grounding systems.
The research on fault feeder detection based on transient sig-
nals, whose amplitude is several times that of the steady signal,
has received much attention in recent years. Prony [8], wavelet
transforms [9], and the Hilbert-Huang transform (HHT) [10] have
been adopted to analyze the transient zero-sequence current sig-
nal. A single characteristic quantity of the fault signals, such as
amplitude, polarity, energy, energy direction, mutation amount,
waveform correlation coefficient, harmonic current, earth capaci-
tance, and information entropy, is always taken as the criterion
for the fault feeder detection. Previous studies have achieved a cer-
tain effect [1014]. However, it is difficult to achieve high accuracy
and reliability in fault feeder detection under all kinds of earth
faults. Cui et al. [10] presented an approach for fault detection
and direction determination for transient/intermittent faults.
Zero-sequence voltage and current are used as the input signals

http://dx.doi.org/10.1016/j.ijepes.2017.05.014
0142-0615/ 2017 Elsevier Ltd. All rights reserved.
98 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108
to extract the fault direction via the direction of instantaneous
power. However, the method in [10] is not suitable for the situa-
tion of polarity inversion that occurs in the transient process.
Zhang et al. [11] proposed an improved phase-locked loop to
extract the fifth harmonic zero-sequence current for detecting
the fault feeder. With the development of modern power electron-
ics technology, a power distribution network contains many elec-
tronic devices that are sources of system harmonics. Moreover,
the fifth harmonic zero-sequence fault current is weak, and will
be drowned in the harmonics generated by these harmonic source
devices. Abdel-Fattah and Lehtonen [12] introduced an algorithm
based on the transient estimation of the earth capacitance of the
background network during the first 1=8 of the power frequency
period. However, the effects of feeder impedances and the Petersen
coil must be compensated by the proposed compensation factors
for overcoming the influence of different transient frequencies.
This results in a method that is difficult to use in practice. Elkalashy
et al. [13,14] investigated a technique based on a probabilistic
method applied to transient features extracted from the zero-
sequence currents using the discrete wavelet transform (DWT).
The detail coefficient in DWT is used for detecting the fault feeder.
An adaptive threshold is used as the input to the probability func-
tion for estimating the fault feeder. Unfortunately, the technique in
[13,14] is invalid in resonant grounding systems.
The fault transient zero-sequence current waveform of each
feeder in a resonant grounding system contains information of
the polarity, amplitude, and shape. Moreover, it can be character-
ized by nonlinearity and nonstationarity when a single-phase-to-
ground fault occurs. Because there is a significant difference
between the fault feeder and sound feeders, a good approach to
fault feeder detection could be through waveform similarity recog-
nition [1517]. Guo et al. [15] developed a fault feeder detection
approach with a stretching transformation applied to transient
zero-sequence currents by using the ratio of the equivalent earth
capacitance between two feeders. A phase-section analysis and
Euclidean distance calculation are used to obtain the characteristic
matrix of the transient zero-sequence current. A fuzzy k-means
clustering algorithm is applied to detect the fault feeder without
a threshold setting. In this case, the short feeder in the resonant
grounding system is sound, and its small earth capacitance will
cause many zero crossings in the transient zero-sequence current.
As a result, the correlation with other sound feeders will decrease,
reducing the reliability of the fault feeder detection method in [15]
on the basis of the correlation coefficient of the undecomposed
transient waveforms. For this reason, the method in [15] is difficult
to use in practical engineering. In order to improve the degree of
similarity and the fault feeder detection margin, the macroscopical
changing trend and time-frequency local feature of transient zero-
sequence current must be taken into account. Guo et al. [16] inves-
tigated a technique with wavelet packet decomposition and recon-
struction as well as signal enhancement applied to process the
fault transient zero-sequence current waveform of each feeder.
By calculating the waveform difference feature matrices in each
time window, the relative entropy is obtained, which is helpful
to define the difference degree for the same time window among
feeders. Equal-weight voting is adopted to count the preliminary
fault feeder detection results in all time windows. Then, the fault
feeder can be detected according to the result of equal-weight vot-
ing and a certain threshold setting. Therefore, the number of time
windows and the threshold setting for the fault feeder detection
criterion are required. Guo et al. [17] presented an approach based
on the similarity of the time-frequency spectrum matrix of the
fault transient zero-sequence current signals. Regarding the time-
frequency spectrum matrix as the pixel matrix of a digital image,
a method for image similarity recognition combined with a
comprehensive similarity coefficient is used for the fault feeder
detection. Unfortunately, the threshold setting for the fault feeder
detection criterion according to many simulated results is inevita-
ble. A novel single-phase-to-ground fault protection scheme was
presented for neutral ineffectively earthed power systems [18]
and bus-connected powerformers [19,20]. Two cluster centers of
the non-fault group and fault group are obtained by applying a
cluster algorithm to the selected features of historical data. The
distances between the detected feature sample and the cluster
centers of the non-fault and the fault are calculated. The detected
feature sample belongs to the cluster with shorter relative distance
to itself. The proposed protection scheme with an explicit physical
mechanism has many advantages over conventional ones, such as
the absence of a threshold setting, updatable historical data, and
high sensitivity and robustness.
In this study, a new fault feeder detection method is presented,
based on a novel amplitude and polarity similarity recognition
algorithm for fault transient zero-sequence current waveforms.
By means of adopting empirical mode decomposition (EMD) to
the fault transient zero-sequence current waveforms of all feeders,
corresponding intrinsic mode function (IMF) components can be
acquired. Then, the Hilbert time-frequency spectrum of the tran-
sient zero-sequence current can be obtained by applying the Hil-
bert transform to the IMF components. By applying the HHT
band-pass filter to the IMF components of transient zero-
sequence current in each feeder, the components of different fre-
quency bands can be extracted; then, the time-frequency matrix
(TFM) is constructed. The fault transient zero-sequence current
waveforms of each feeder are divided into several equal intervals
along the time axis. Each sampling point in one interval is reset
to zero or one according to its value. Then, the polarity distribution
matrix (PDM) for each feeder is constructed from those changed
sampling points in each interval. Moreover, the normalized singu-
lar value of the TFM and PDM are merged together. Thus, the fea-
ture quantities including the amplitude and polarity information of
each fault transient zero-sequence current waveform are obtained.
In addition, fuzzy c-means (FCM) clustering is applied to these nor-
malized singular values so as to detect the fault feeder without a
certain threshold setting. In this way, this approach can resolve
the problems of conventional fault feeder detection methods,
including suitability for all fault conditions and difficulty in thresh-
old setting.
2. Proposed fault feeder detection algorithm
After a single-phase-to-ground fault occurrs in a resonant
grounding system, the transient zero-sequence currents in all feed-
ers will be acquired at the substation. The transient component of
zero-sequence currents in a fault feeder is composed of the free
transient oscillating component of capacitive currents in sound
feeders and the dc transient component of inductive current via
an arc-suppression coil. The amplitude of the high-frequency oscil-
lating damped component and dc damped component via an arc-
suppression coil may not be canceled out. By superposing the dif-
ferent amplitudes and similar frequencies, the amplitude of the
transient zero-sequence current in a fault feeder may rise. The
transient component of zero-sequence current in the sound feeders
is composed only of the free high-frequency oscillating damped
component of the capacitance grounding current. As a result of
the excitation from the same zero-sequence voltage source, the
transient zero-sequence currents in different sound feeders have
the same variation trend. The transient component of zero-
sequence current has similar free oscillating frequencies, but
different amplitudes because of the different parameters of sound
feeders. Moreover, the direction of the transient zero-sequence
currents of the fault feeder and sound feeders is opposite.
 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108 99

Therefore, the amplitude and polarity difference of transient zero-
sequence currents between a fault feeder and sound feeder are
greater than those of transient zero-sequence currents between
different sound feeders.
In this study, the HHT band-pass filter and wave transformation
are used to construct the TFM and PDM, respectively. The singular
values of the TFM and PDM can be obtained as the feature quanti-
ties that characterize the amplitude and polarity information of the
fault transient zero-sequence current signals. Then, the FCM clus-
tering can be applied to the normalized singular values so as to
detect fault feeder without a certain threshold setting.
2.1. Obtaining TFM via the HHT band-pass filter
The HHT is suitable for analyzing the partial dynamic actions
and characteristics of signals, which are nonlinear and nonstation-
ary. The HHT has two main steps, including the EMD and the Hil-
bert transform. The EMD is an adaptive decomposition method.
The original signal is decomposed into a series of IMF components
with different frequencies and bandwidths. The order of IMF com-
ponents varies because the original signals can be different. The
IMF component of the original signals may also be different.
The simulated model of a resonant grounding system in [17]
was constructed with PSCAD/EMTDC. The simulation result of
single-phase-to-ground fault was analyzed with MATLAB. A
single-phase immediate earth fault with the inception phase angle
u p=2 was assumed to occur in the system. The transient zero-
sequence current waveforms of a fault feeder and a sound feeder
are shown in Fig. 1(a). Moreover, the time-amplitude spectrums
of the second-order IMF components obtained by the EMD are
shown in Fig. 1(b).
To restrain the end effect of the EMD, the extreme point contin-
uation method is applied [21]. That is, the maximum and mini-
mum points are continued by a mirror image. The end effect in
the edge can be shifted to the two sides after signal continuation.
The time-frequency spectra of the second-order IMF component
of transient zero-sequence currents in a fault feeder and sound
feeders are shown in Fig. 1(c).
As can be seen in Fig. 1(c), the frequency of the second-order IMF
component of the transient zero-sequence current ranges mainly
from 0 kHz to 2 kHz in the fault feeder, whereas in the sound feeder,
it varies mainly from 0 kHz to 1 kHz. That is, for the same-order IMF
component of the transient zero-sequence current, the frequency
widths in the two kinds of feeders are different.
When the single-phase-to-ground fault occurs, there exists an
amplitude distinction between the transient zero-sequence cur-
rent waveforms of a fault feeder and sound feeders. By extracting,
analyzing, and comparing the components of each sub-frequency
Amplitude(kA)
band, the objective of the fault feeder detection, which is based
on the amplitude characteristic, can be accomplished. If the IMF
components of the transient zero-sequence current of the corre-
sponding orders obtained by the EMD are used directly for compar-
ison and analysis, a possible failure of the fault feeder detection
may result from the low comparability caused by the inconsistent
range of frequencies. To solve this problem, the HHT band-pass fil-
ter is used to restructure the instantaneous frequencies that char-
acterize the original signals so as to enhance the signals
comparability. By applying the Hilbert transform to various orders
of IMF components of the transient zero-sequence currents
obtained by the EMD, the time-frequency spectrum can be
obtained, of which each point corresponds to the respective ampli-
tude of the IMF component.
A suitable frequency width Df is selected to divide the instan-
taneous frequencies of the spectrum into equal intervals. The prin-
cipal components of the frequency spectrum of fault transient
zero-sequence currents vary with fault location, grounding resis-
tance, and type of distribution network. When a single-phase-to-
ground fault occurs, it will not be the same in different feeders
too. In order to make the principal components of the spectrum
of fault signals fall within the same frequency band under different
fault conditions, the maximum frequency of the principal compo-
nents should be chosen as the suitable frequency width (Df). This
will make it comparable. As a consequence, the frequency range
of section i is i  1Df ; iDf . The instantaneous current amplitude
points of all IMF components beyond this section are set to zero,
whereas those points of all IMF components within the range
remain unchanged. By structuring the reserved points, the compo-
nent of the original signals in the frequency range of section i can
be obtained. Similarly, the components in other sub-frequency
bands can also be obtained.
In order to better depict the local feature of the fault transient
zero-sequence current waveform of each feeder in the time
domain and frequency domain, a TFM of the transient zero-
sequence current waveforms is constructed via the advantage of
the HHT band-pass filter. Assuming that the number of sampling
points of transient zero-sequence current in each feeder is n, the
waveform of transient zero-sequence current is decomposed into
m sub-frequency bands through the HHT band-pass filter, and
the data point of each sub-frequency band is eij (i 1; 2; . . . ; m
and j 1; 2; . . . ; n). Then, the TFM can be obtained as
2 3
e11 e12    e1n
6e e22  e2n 7
6 21 7
E6
6 .. .. .. .. 7 7 1
4. . . . 5
em1 em2  emn

0.2 2.5
0.1
Fault feeder
0 2
Amplitude (kA)

Frequency(kHz)

0.1 -0.1 Fault feeder

Sound feeder
0.208 0.21 0.212 0.214 0.216 1.5

0 Time (s)
Amplitude(kA)

1 Sound feeder
0.02
0 0.5
-0.1
-0.02
0
0.206 0.216 0.226 0.236 0.208 0.21 0.212 0.214 0.216 0.208 0.21 0.212 0.214 0.216
Time (s) Time (s) Time (s)
(a) (b) (c)
Fig. 1. Transient zero-sequence current waveforms of a fault feeder and a sound feeder: (a) Original waveforms; (b) Time-amplitude spectrums of second-order IMF component;
(c) Time-frequency spectrums of second-order IMF component.
100 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108

The rows of matrix E denote the reconstructed waveforms of zero-sequence current waveform, of which the numerical value
transient zero-sequence currents in each sub-frequency band, and represents the amplitude and the polarity features, respectively,
the columns denote the sampling times of transient zero-sequence in the waveform. The singular value is a relatively stable algebraic
currents. Matrix E depicts the amplitude feature of the transient feature of the matrix with favorable robustness and generalization
zero-sequence current waveform in each sub frequency band. ability, which can adapt to various complicated situations when
the single-phase-to-ground fault occurs in a resonant grounding
2.2. Obtaining PDM via wave transformation system. Assuming that the number of feeders is M, the element
of the amplitude feature matrix (AFM) and the polarity feature
Normally, the frequency of the fault transient zero-sequence cur- matrix (PFM) for each feeder is g ij (i 1; 2; . . . ; M and
rent ranges from 0 kHz to 3 kHz. Noise whose frequency is higher j 1; 2; . . . ; N) and hij (i 1; 2; . . . ; M and j 1; 2; . . . ; K), respec-
than 3 kHz can be removed via the HHT band-pass filter mentioned tively. Then, the AFM can be obtained as
previously. Then, a wave transformation method is taken to obtain 2 3
g 11 g 12    g 1N
the PDMs of the fault transient zero-sequence current waveforms.
6g g 22  g 2N 7
The sampling points of the transient zero-sequence current in each 6 21 7
G6
6 .. .. .. ..
7
7 4
feeder are divided into k equal parts along the time axis. If the value 4. 5
. . .
of the sampling point in each part is less than zero, it will be reset to
zero; otherwise, it will be reset to one, so the elements of each PDM
g M1 g M2    g MN
are f ij (i 1; 2; . . . ; k and j 1; 2; . . . ; n=k, n is an integer multiple of ^ can be obtained as
The elements of the normalized AFM G
k). Then, the PDM of the transient zero-sequence current in each fee-
g ij  minG
der can be obtained as g^ij 5
2 3 maxG  minG
f 11 f 12    f 1n=k
6f ^ g is an element of G; min(.) and max(.)
6 21 f 22    f 2n=k 7
7 where g^ij is an element of G; ij
6 7
F 6. .. . . .. 7 2 are the minimum and maximum operators, respectively; and the
6 .. . . . 7 ^ is 0; 1. The PFM can be obtained as
4 5 range of elements in the matrix G
f k1 f k2    f kn=k 2 3
h11 h12    h1K
The matrix F depicts the polarity feature of the transient zero- 6h h22  h2K 7
6 21 7
sequence current waveform in each sub-time slot. H6
6 .. .. .. .. 7
7 6
4. . . . 5
2.3. Calculation of PDM and TFM singular values hM1 hM2    hMK

The element of the normalized PFM H ^ can be obtained by

By considering a real matrix A aij mn , there exist two unitary
referring to (5).
matrices (V mm and U ) and a diagonal matrix ^ with the normal-
 p p p nn  In the combination of the normalized AFM G
Dll diagf k1 ; k2 ; . . . ; kr ; 0;    ; 0g; l minm; n according
^ the element of the amplitude and polarity feature
ized PFM H,
to the singular-value theory, such that
matrix (APFM) is qij (i 1; 2; . . . ; M and j 1; 2; . . . ; N K). Then,
X
r p X
r p
A VDU T ki v i uTi ki Ai 3 the APFM can be obtained as
2 3
i1 i1
q11 q12    q1N q1N1 q1N2    q1NK
p p p 6
where k1 ; k2 ; . . . ; kr are defined as singular values of real matrix 6 q21 q22    q2N q2N1 q2N2    q2NK 7
7
6 7
A; the rank of A is rA minm; n; V v 1 ; v 2 ; . . . ; v m denotes Q 6. .. .. . .. .. . . .. 7 7
6 .. . . .. . . . . 7
T
the unit feature vector of AA ; and U u1 ; u2 ; . . . ; un denotes the 4 5
unit feature vector of AT A. qM1 qM2    qMN qMN1 qMN2    qMNK
The singular-value decomposition (SVD) means that matrix
Amn with rank r is decomposed into a weighted sum of r matrices 2.4. Fault feeder detection based on FCM clustering
with m  n order and rank 1. The weight is the singular values
p
ki of sub-matrix Ai . Ai and its corresponding singular values As an unsupervised dynamic clustering, FCM clustering is an
p
ki can be obtained by taking advantage of the decomposition effective method of pattern recognition. A flexible fuzzy partition
of the matrix A. When the elements of the matrix A are distorted of data points can be achieved by attaching it to certain clustering
or disturbed, the variable quantity of the quadratic sum of matrix centers according to the degree of membership [22]. The degree of
A will not be more than the norm of the perturbation matrix, membership is the relation strength between the data points and
and the singular values will change only small amount. That is, certain clustering centers. X fx1 ; x2 ;    ; xm g is defined as the
its stability is superior. Therefore, the singular value of a matrix sample space, and m denotes the sample capacity. Each sample
is relatively stable, which is usually used for characterizing a signal xi fxi1 ; xi2 ;    ; xin g is an n-dimensional vector. X is divided into
in pattern recognition. c categories, where 2 6 c 6 m and every sample belongs to differ-
Applying SVD to each row of the matrix E and F can cut down ent categories to a certain extent. uij denotes the degree of mem-
the feature dimension and reduce calculation. Elements of each bership of the sample j attached to the category i in X. If it satisfies
row of matrix E have one singular value, so each transient zero-
8 c
sequence current waveform has N singular values fg 1 ; g 2 ;    ; g N g, > X
>
> uij 1
where N is the number of sub-frequency bands. Elements of each >
>
>
>
row of matrix F have one singular value, so each transient zero- < i1
sequence current waveform has K singular values fh1 ; h2 ;    ; hK g, uij 2 0; 1 8
>
>
where K equals the number of divided parts. Among them, the >
> Xn
>
> uij 2 0; m
number of singular value reflects the number of different patterns >
:
j1
included in the TFM (E) and the PDM (F) of the transient
 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108
then U cn fuij g can be called the membership matrix. The
FCM clustering algorithm minimizes the objective function
through an iterative method, that is to say
X
m X
c
minimize J fcm U; V upij kxj  v i k2
j1 i1
where J fcm U; V is the objective function, and U is the membership
matrix; V is the clustering center; v i is the clustering center of cat-
egory i; p is the fuzzy weighted exponent, generally p > 1, set here
as p 2; and kxj  v i k denotes the Euclidean distance from the
sample xj to the clustering center v i . When there is convergence
of iterations, an optimal clustering center V fv i g and a member-
ship matrix U fuij g of X can be obtained.
In the engineering application of fault feeder detection, the
APFM (Q) is the characteristic parameters of the fault transient
zero-sequence current in different feeders, which are divided into
two categories by applying FCM clustering to obtain a membership
matrix in (10).
 
u11 u12    u1M
U 10
u21 u22    u2M
where uij is the degree of membership of the feeder j attached to the
category i, and M is the number of feeders. The feeder attached to
category 1 or category 2 can be confirmed through the membership
matrix U, among which the fault feeder is divided into one cate-
gory alone. The fault feeder can be detected without a certain
threshold setting.
3. Simulation and verification
PSCAD/EMTDC and physical system tests proved the effective-
ness of the proposed earth fault detection method. The physical
system based on an analog model is a three-phase system that
can be used for modeling medium-voltage 10 kV feeders connected
to the model bus bar. The analog model is designed on the basis of
the similarity principle. The feeders are modeled with a three-
phase P type equivalent circuit. The model bus bar was supplied
via an equivalent transformer from a low-voltage 3  380 V net-
work. The neutral point of the physical system was grounded by
a designed Peterson coil via a zigzag transformer. Single-phase-
to-ground faults were simulated by controlling a designed device
based on a fast electronic switch. The zero-sequence current and
voltage-acquisition devices are designed around STM32F407VGT6
and AD7606 chips. The sampling frequency of all signals was
10 kHz. MATLAB was used to implement the fault feeder detec-
tion algorithm.
As shown in Fig. 2, the structure of the PSCAD/EMTDC
simulation model is the same as that of the physical system. The
Peterson coil is located at the busbar. OL denotes the length of the
overhead feeder, CL denotes the length of the cable feeder, and
CTn n 1; . . . ; 5 denotes the zero-sequence current transformer
of the five feeders.
Yd11 is taken as the connection mode of the main transformer
winding. In the PSCAD/EMTDC simulation model, the capacity of
the 110 kV=10 kV main transformer is 31.5 MVA, the open-circuit
loss of the main transformer is DP0 19:5 kW, the short-circuit loss
is DPk 118:9 kW, the no-load current percentage is I0 % 0:12,
and the short-circuit voltage percentage is U k % 14:99. In the
physical system, the capacity of the 0:4 kV=0:4 kV main trans-
former is 20 kVA, and the parameters of main transformer are
DP0 203 W, DPk 229:4 W, I0 % 1:88, and U k % 10:47.
In the PSCAD/EMTDC simulation model, the zero and positive
sequence distributed parameters of the overhead feeders were
given as
101
R0 0:23 X=km; C 0 0:008 lF=km;
L0 5:478 mH=km; R1 0:17 X=km;
9 C 1 0:00969 lF=km; L1 1:21 mH=km
Moreover, the distributed parameters of the cable feeders were
given as
R0 2:7 X=km; C 0 0:28 lF=km;
L0 1:019 mH=km; R1 0:27 X=km;
C 1 0:339 lF=km; L1 0:255 mH=km
As shown in Fig. 3, the three-phase P model was taken as the
model of the feeders in the physical system. The hollow circular
spiral inductors were designed as the self-inductance and mutual
inductance for all feeders. The parameters of the overhead feeders
were given as
r 0:19 X=km; Lmm 2:632 mH=km;
Lmn 0:255 mH=km; C p 0:00485 lF=km;
C g 0:688 lF=km
Moreover, the parameters of the cable feeders were given as
r 1:08 X=km; Lmm 0:51 mH=km;
Lmn 0:255 mH=km; C p 0:169 lF=km;
C g 2:413 lF=km
The zigzag grounding transformer was made up of three single-
phase saturated transformers. In the PSCAD/EMTDC simulation
p p
model, the capacity of each 10=2 3 kV=10=2 3 kV transformer
was 0.3 MVA, and the parameters of each one were given as
DP0 0:68 kW, DPk 2:84 kW, I0 % 2:32, and U k % 2:47. In
p p
the physical system, the capacity of each 0:4=2 3 kV=0:4=2 3 kV
transformer was 0.5 kVA, and the parameters of each one were
given as DP0 1:7 W, DPk 7:1 W, I0 % 2:3, and U k % 2:5.
In the PSCAD/EMTDC simulation model and physical system,
the load was replaced by an equivalent impedance, 100 j40 X.
The maximum degree of overcompensation of the arc-
suppression coil was 10%. The inductor parameter of the arc-
suppression coil acquired by calculation was 0:6853 H. In general,
the active power loss of the inductive loss in the arc-suppression
coil was approximately 2:5  3:5%. In this study, the value of 3%
was adopted, and the resistance parameter acquired by calculation
was 6:4592 X.
The physical system based on these parameters is shown in
Fig. 4.
An earth fault in feeder 3 was chosen as an example to illustrate
the process of fault feeder detection in PSCAD/EMTDC; the fault
was simulated as 1 km away from the bus bar with a grounding
resistance of 100 X and at the 45 point of the phase voltage. The
transient zero-sequence current i0 waveforms of one fault feeder
and four sound feeders are shown in Fig. 5(a). The similarity
between the transient zero-sequence current waveforms of a fault
feeder and sound feeders is lower than that between two sound
feeders. The zero-sequence current sampling frequency for all
feeders was 10 kHz. The fault transient zero-sequence current
within the first 1=2 periodical waveform was analyzed, whose data
in this period is considered as the most concentrating and salient
part of the transient feature in the transient process.
Whether the zero-sequence voltage is out of limit or not is
the basis to start the fault feeder detection algorithms. The HHT
102 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108

Fig. 2. Simulation model of a 10 kV resonant grounding system.

A
Lmm r
B Lmn
Lmn
C
Cp Cp
Cg Cg

(a) (b)

Fig. 3. Equivalent P model of feeder: (a) Principle diagram; (b) Physical model.

Fig. 4. Physical system: (a) Before installation; (b) After installation; (b) Earth fault test.

band-pass filter was applied to the first 1=2 periodical waveforms matrix, row 1 to row 5 represent the first to fifth feeder, respec-
of the fault transient zero-sequence current in the five feeders. For ^ was obtained as
tively. According to (5), the normalized AFM G
the distribution networks shown in Fig. 2, the maximum frequency 2 3
of the principal components of the fault signals was approximately 0:0836 0:0513 0:0123 0:0114 0:0114
6 7
500 Hz. In order to maintain a certain margin, 600 Hz was chosen 6 0:4600 0:0850 0:0998 0:1425 0:0844 7
6 7
as the suitable frequency width (Df). Normally, the frequency ^ 6
G 61 0:1221 0:0891 0:1268 0:7680 7 11
7
range of a fault transient zero-sequence current changes from 6 7
4 0:1194 0:0780 0:0303 0:0240 0:0240 5
0 kHz to 3 kHz, so the number of sub-frequency bands was set to
0:5139 0:0582 0:0580 0:0573 0:0573
5. Thus, the waveform of the transient zero-sequence current
was decomposed as five different sub-frequency bands, which The first 1=2 periodical waveforms of the fault transient zero-
may compose the TFM of the corresponding feeder. The amplitude sequence current in the five feeders were divided into five equal
singular-value vector with dimension, 1  5, of each transient intervals along the time axis. The PDM of the corresponding feeder
zero-sequence current waveform is acquired by way of applying is obtained by the way of waveform transformation. The polarity
the SVD to the elements of each row of the TFM. Then, five ampli- singular-value vector with dimension 1  5 of each transient
tude singular-value vectors of the transient zero-sequence current zero-sequence current waveform is acquired by way of applying
merge into a feature matrix G, whose dimension is 5  5. In this the SVD to the elements of each row of the PDM. Then, five polarity
 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108 103

-3
x 10
L1 L1
1

Normalized singular value

Fault feeder L2 L2 1

Membership degree
L3 L3
2 0.8
Amplitude (kA)

L4 L4
L5
L5
0.6 C1
C2
0 0.4 0.5

0.2
-2
0
0
0.204 0.214 0.224 0.234 2 4 6 8 10 1 2 3 4 5
Time (s) Column number Feeder number
(a) (b) (c)

Fig. 5. Fault detection procedure: (a) Transient zero-sequence current waveforms; (b) Elements of APFM and its cluster center; (c) Elements of membership matrix.

singular-value vectors of the transient zero-sequence current The historical fault signals and the parameters of the feeders or
merge into a feature matrix H, whose dimension is 5  5. In this power distribution systems are not needed. The number of feeders
matrix, row 1 to row 5 represent the first to fifth feeder, respec- can be set in the fault detection program to adapt to distribution
^ was obtained as
tively. By referring to (5), the normalized PFM H systems with different numbers of feeders. When an event occurs
2 3 that switches a feeder on or off, the switching status can be
0:2236 0:2236 0:6325 0:9220 1 acquired by the fault detection systems and used to set the number
6 7
60 0 0:5000 17 1 of feeders for the fault detection program.
6 7
^
H6 17 12
6 0:9747 1 1 1 7
6 7
4 0:3162 0:4472 0:5916 0:8944 1 5
3.1. Typical earth fault
0 0 0:6325 0:9220 1

By combining the normalized AFM G ^ with the normalized PFM A single-phase-to-ground fault usually occurs at the moment
^ when the phase voltage reaches its peak point. However, the wire
H, the APFM Q can be obtained. Q is set as the input for the FCM
on the power supply side may drop out when a single-phase-line-
clustering. Because only the fault state and the sound state should
break fault occurs in a bare overhead wire. Furthermore, where
be distinguished in fault feeder detection, the category number of
underground pipes are under construction, cables may be
the FCM clustering was set to 2, and the weighted exponent p
destroyed because of human factors. For these reasons, a single-
was set to 2. The termination factor e of iteration was set to
phase-to-ground fault may sometimes occur at other moments,
105 , and the maximum iterations kmax was set to 100. The solu- for example, when the phase voltage reaches its zero point. As
tion procedure converged after seven iterations. The result of the shown in Table 1, single-phase-to-ground faults under the condi-
objective function J fcm U; V was 0:1244; the elements of APFM tions of different feeders, different fault positions, different
and its clustering center are shown in Fig. 5(b), where C1 and C2 grounding resistance values, different fault inception angles, and
denote the cluster center of the features of the fault feeder and different compensation degrees are simulated. In Table 1, Lm is
non-fault feeders, respectively. The membership matrix U was the number of the fault feeder, X f is the distance from the fault
obtained as point to the bus bar, Rf is the grounding resistance, h is the fault
  inception angle, and p is the compensation degree of the arc-
0:0231 0:0313 1 0:0627 0:0324
U 13 suppression coil. Under the fault conditions shown in case 1 of
0:9769 0:9687 0 0:9373 0:9676
Table 1, the fault feeder detection results are entirely correct in
The rows of matrix U denote the sound or fault category, and both the PSCAD/EMTDC and physical system tests.
column 1 to column 5 represent the first to fifth feeders, respec-
tively. All the maximum elements of each column in the same
row belong to one category. As shown in Fig. 5(c), feeder 3 is a cat- 3.2. Anti-interference capability
egory alone, whereas the other feeders belong to another category.
That is, feeder 3 is the fault feeder. Noise interference signals in the environment may distort the
In practical engineering applications, the detection time, com- waveform of the fault transient zero-sequence current so that the
putational power, and normal switching events must be consid- results of fault feeder detection are affected. Noise whose fre-
ered. The processor, RAM, and operating system of the personal quency is higher than 3 kHz can be removed by the HHT band-
computer in this study were a 2.4-GHz Intel CoreTM i7-5500U, pass filter. The SVD is a way of extracting characteristic parame-
8.00 GB, and Win64, respectively. The run time of the MATLAB ters, which can keep the signal features relatively steady under dis-
fault detection algorithm program was approximately 0.351 s. turbance and noise interference. To test this, a single-phase-to-
The sub-program of the HHT band-pass filter consumed approxi- ground fault was simulated in feeders 3 and 5 under the distur-
mately 0.201 s. These results satisfied the requirements of practical bance of white Gaussian noise whose signal-to-noise ratio was
engineering applications. High-performance CPUs and digital sig- 20 dB; only frequencies in the range 0 kHz to 4 kHz were consid-
nal processors (DSP) are now widely used, and the computational ered in this case. The PSCAD/EMTDC test results are correct
power requirements of the proposed method are not especially entirely under the fault conditions shown in case 2 of Table 1. Thus,
large. The program can be effectively achieved. the HHT band-pass filter and the SVD give the proposed fault fee-
When a single-phase-to-ground fault occurs, only the acquired der detection method superior anti-noise-interference ability. In
voltage signal of the bus bar and current signal of each feeder are other words, one can accomplish fault feeder detection correctly
required to detect the fault feeder without a threshold setting. under the circumstance of noise interference.
104 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108

Table 1
Conditions of earth fault cases.

Fault cases Lm h Xf (km) Rf X p (%)

Case 1 2 90 2 1000 10
2 0 3 0 10
2 45 3 500 5
4 90 1 5 8
4 0 2 100 5
4 60 1.5 200 8
Case 2 3 0 1.5 5 10
3 90 2 1000 5
5 0 5 5 8
5 90 6 500 10
Case 3 1 0 5 200 8
1 90 10 0 10
4 0 1 100 5
4 45 2 500 8
Case 4 2, 2 0 2 500 10
45 3 0 10
3, 3 0 1.5 5 8
60 2.5 500 8
Case 5 2, 3 0 3 5 10
0 2 100 10
2, 3 90 3 500 8
90 2 200 8
Case 6 3 0 1 8
3 30 2 5
4 45 2 8
4 90 1 10

3.3. Asynchronous sampling
As shown in Fig. 4(c), the device of the fault feeder detection is
normally designed by adopting a distributed framework for the
convenience of installation in engineering applications. In this
way, the data acquisition modules of zero-sequence voltage and
currents are installed into different switch cabinets separately,
and fault transient signals of different feeders are collected and
analyzed through the mode of communication. As a result, high-
speed data acquisition can be accomplished to meet the demand
of fault feeder detection via fault transient signals. At this time,
the problem of asynchronous sampling of each signal is inevitable.
The elements of APFM will be affected by asynchronous sampling.
Benefitting from the characteristics of SVD and the FCM algorithm,
the proposed fault detection algorithm can overcome the influence
of slightly asynchronous sampling. In this examination, a single-
phase-to-ground fault was simulated to occur in feeders 1 and 4,
and the sampling time of the zero-sequence current of feeders 1,
2 and 3 lagged 0:001 s behind that of feeders 4 and 5. Under the
fault conditions shown in case 3 of Table 1, the proposed algorithm
can detect the fault feeder correctly in both the PSCAD/EMTDC
and physical system tests. More examples under the conditions
of different fault inception angles (FIAs), grounding resistance val-
ues, fault locations, compensation degrees, and asynchronous
times were simulated and tested; the results show that the
allowed asynchronous time varies with changes in fault conditions.
When the FIA is 90, the maximum asynchronous time reaches
0.0011 s. When the FIA is 0, the maximum asynchronous time
reaches 0.0022 s. The reason is that asynchronous sampling has a
greater impact on the polarity feature of the high-frequency signal.
3.4. Two-point-grounding fault
If the single-phase-to-ground fault with arc is not dealt with in
time, the resulting over-voltage may result in another earth fault.
This fault may occur in an insulation weakening point of the same
phase in the fault feeder or in another feeder. In the case of severe
weather such as a strong typhoon, branches of trees will easily
touch bare overhead wires, which may also cause a two-point-
grounding fault. To test this case, feeders 2 and 3 were chosen as
an example. In this examination, two simultaneous two-point-
grounding faults were simulated to occur on the same-phase.
One occurred in different positions of a feeder, and the other
occurred on different feeders, as simulated by the PSCAD/EMTDC.
The fault conditions are shown in case 4 and case 5 of Table 1. Ben-
efitting from the FCM clustering, when the single-phase-to-ground
fault occurs in two points of a system, whether simultaneously or
not, the proposed method still can accomplish the fault feeder
detection correctly. The proposed methods in [1114] take only a
single-point-grounding fault into account, and are not suitable
for two-point-grounding fault detection.
3.5. Arc fault
An intermittent arc fault is more likely to occur with a complex
process in a real distribution power system. In the PSCAD/EMTDC
simulation study, an arc model was obtained via some assump-
tions and simplifications, which can emulate the arc with a compli-
cated physical process and massive influence factors [23]. The
commonly used arcing models include the models of Cassie, Mayr,
Schwarz, and cybernetics. At present, the last three simulation
models are suitable for an arc fault that occurs in a resonant
grounding system with the characteristic of low current. In this
study, the cybernetics model was selected because it can set up
the arc length directly. Therefore, the impact of arc length on arc
impedance can be investigated, as well as the steady and transient
features of a grounding arc. In the test, there were two carbon rods.
One connected to a phase, and the other connected to ground. As
shown in Fig. 6, the distance of the two carbon rods was controlled
by an microprocessor control unit (MCU). Thus, the arc fault simu-
lation was carried out in the physical system. Arc faults that occur
in feeders 3 and 4 under the different circumstances shown in case
6 of Table 1 were simulated in PSCAD/EMTDC and a physical sys-
tem test. The results of the fault feeder detection are correct.
 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108 105

it will be determined as fault feeder. Here, this method is defined as

method 2.
Method 1 and method 2 are considered comprehensively. Three
feeders with the first three maximum normalized I0 vaules and
their corresponding signs of P were selected. According to the
selected signs of P in the feeders, the fault feeder will be detected
[24]. This comprehensive method is defined as method 3.
Because of the complexity of the earth fault mechanism, envi-
ronments, and random parameters, the conventional methods
based on the amplitude or polarity feature individually are difficult
to achieve high accuracy and reliability in fault feeder detection
under all kinds of earth faults. No matter when the normalized
AFM or the normalized PFM is used as the feature quantities of
fault transient zero-sequence current signals independently, earth
fault detection cannot be accomplished correctly in some special
Fig. 6. The arc fault simulation in physical system.
cases, which will cause wrong detection results after applying
4. Comparisons FCM to them. Benefitting from SVD and FCM, the APFM consisting
of the AFM and PFM can solve this kind of problem. Here, the
For the conventional method based on the transient amplitude methods based on AFM, PFM and APFM are defined as method 4,
feature [24], the true RMS of the transient zero-sequence current method 5, and method 6, respectively.
can be obtained as The following four earth fault types were used to verify the
v validity and robustness of fault detection methods 16. The
u N
u1 X 2 experiment of the third type of earth fault was simulated under
I0j t i 14 PSCAD/EMTDC, and that of the others was tested under the phys-
N k1 0jk
ical system. According to the designed similarity ratio, the ampli-
tude of the transient zero-sequence current under the physical
where i0 is the fault transient zero-sequence current, N is the num-
system test was transformed. The transient zero-sequence cur-
ber of sampling points, j is the feeder number, and k is the sampling
rents, fault detection procedure, and results of the four earth fault
point number. If the normalized I0 in one feeder is a maximum, then
types are shown in Figs. 710, respectively. The first type of earth
that feeder is determined as a fault feeder. Here, this method is
fault is a fault that occurs at the end of feeder 5 with metallic
defined as method 1.
grounding and an inception angle of 90, with the assumption that
For the conventional method based on the transient polarity
the first 12 data points of the waveform in feeder 5 were not sam-
feature [24], the polarity relationship of two transient zero-
pled. As shown in Fig. 7(a), the amplitude of the transient zero-
sequence currents can be obtained as
! sequence currents in the sound feeders and fault feeder on this
X
N occasion oscillate violently, and the polarity of the transient
Pjm sign i0jk  i0mk 15 zero-sequence currents in feeder 5 has changed slightly. This
k1
results in an error in polarity judgment, and makes the fault detec-
where j; m is the feeder number. With feeder m as the reference, if tion more difficult. It is easy to see that the fault detection results
the sign of P in one feeder is different from that of the other feeders, of methods 1 and 6 are correct, and the others are wrong.

L1
1.5 L1 L1
Method 11 Method 22 Method 33
Normalized singular value

0.1 L2 Method Method Method L2 1 L2

L3 1 L3 L3
Normalized value

0.8
Amplitude (kA)

L4 L4 L4
L5 0.5 L5 L5
0.6 C1
0 0 C2
0.4
-0.5
0.2
-1
-0.1 0
-1.5 1 2 3 4 5
0.206 0.216 0.226 0.236 L1 L3 L5 L1 L3 L5 L2 L5
L2 L4 L2 L4 L3 Column number
Time (s)
(a) (b) (c)
L1 3.5
Normalized singular value

1 L1
Normalized singular value

L2 1 L2
L3 L3 3
Membership degree

0.8 L4 0.8 L4
L5 L5 2.5 APFM
0.6 C1 0.6 C1
C2 C2
2
0.4 0.4 1.5 PFM

0.2 0.2 1
0.5 AFM
0 0
0
1 2 3 4 5 2 4 6 8 10 1 2 3 4 5
Column number Column number Feeder number

(d) (e) (f)

Fig. 7. Fault detection procedure of the first type of earth fault: (a) Zero-sequence current waveforms; (b) Fault detection results of method 1, 2, 3; (c) Elements of AFM and its
cluster center; (d) Elements of PFM and its cluster center; (e) Elements of APFM and its cluster center; (f) Fault detection results of method 4, 5, 6.
106 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108

L1 L1 L1

Normalized singular value

Method 1 Method 2 Method 3 L2 1
L2 L2
0.02 1 L3

Normalized value
L3 L3
Amplitude (kA)

L4 L4 0.8 L4
L5 0.5 L5 L5
0.6 C1
0 0 C2
0.4
-0.5
0.2
-0.02 -1
0
-1.5
0.202 0.212 0.222 0.232 L1 L3 L5 L1 L3 L5 L2 L5 1 2 3 4 5
L2 L4 L2 L4 L3 Column number
Time (s)
(a) (b) (c)
L1 3.5
Normalized singular value

1 L1

Normalized singular value

L2 1 L2
L3 3

Membership degree
0.8 L3
L4 0.8 L4 2.5 APFM
L5 L5
0.6 C1 0.6 C1 2
C2 C2
0.4 0.4 1.5 PFM

0.2 0.2 1
0.5 AFM
0 0
0
1 2 3 4 5 2 4 6 8 10 1 2 3 4 5
Column number Column number Feeder number

(d) (e) (f)

Fig. 8. Fault detection procedure of the second type of earth fault: (a) Zero-sequence current waveforms; (b) Fault detection results of method 1, 2, 3; (c) Elements of AFM and its
cluster center; (d) Elements of PFM and its cluster center; (e) Elements of APFM and its cluster center; (f) Fault detection results of method 4, 5, 6.

L1
1.5 L1
Method 1 Method 2 Method 3 L1
Normalized singular value
L2 L2 1 L2
0.01 L3 1 L3
Normalized value

L3
Amplitude (kA)

L4 L4 0.8 L4
L5 0.5 L5 L5
0.6 C1
0 C2
0
0.4
-0.5
0.2
-1
-0.01 0
-1.5
0.202 0.212 0.222 0.232 L1 L3 L5 L1 L3 L5 L2 L5 1 2 3 4 5
L2 L4 L2 L4 L3
Time (s) Column number
(a) (b) (c)

L1 L1 3.5
Normalized singular value
Normalized singular value

1 L2 1 L2
L3 L3 3
Membership degree

0.8 L4 0.8 L4
L5 L5 2.5 APFM

0.6 C1 0.6 C1
2
C2 C2
0.4 0.4 1.5 PFM

0.2 0.2 1
0 0 0.5 AFM

1 2 3 4 5 2 4 6 8 10 0
Feeder number
Column number Column number
(d) (e) (f)

Fig. 9. Fault detection procedure of the third type of earth fault: (a) Zero-sequence current waveforms; (b) Fault detection results of method 1, 2, 3; (c) Elements of AFM and its
cluster center; (d) Elements of PFM and its cluster center; (e) Elements of APFM and its cluster center; (f) Fault detection results of method 4, 5, 6.

The second type of earth fault is under the fault conditions The third type of earth fault is a two-point single-phase-to-
Lm 3, X f 1 km, Rf 5 X, h 0 , and p 10%, with the assump- ground fault occurring on the same phase of feeders 3 and 5.
tion that the polarity of the zero-sequence CT in feeder 3 is With the metallic grounding and the same inception angle of
reversed in field because of an error connection. As shown in 90, one fault occurs at the end of feeder 5, and the other occurs
Fig. 8(a), the transient zero-sequence current of fault feeder 3 is in feeder 3 and is 1 km away from the bus bar. As shown in
on one side of the time axis within the first periodic waveform. Fig. 9(a), the changes in amplitude and polarity of the transient
This will easily result in the failure of the fault detection method zero-sequence currents make the fault detection more difficult.
based on the polarity feature. The fault detection results of meth- So far, the existing earth fault detection methods have not con-
ods 1, 4, 5, and 6 are correct, and the others are wrong. sidered this situation. It can be seen that the fault detection
 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108 107

1.5 L1

Normalized singular value

0.35 L1 Method 2 Method 3
L1 1 L2
Method 1 L2
L2 1 L3
L3 0.8

Normalized value
L3 L4
L4
Amplitude (kA)

L4 L5
0.15 L5 0.5 L5
L6
0.6 L6
L6 C1
0 C2
0.4
-0.05 -0.5
0.2
-1
0
-0.25 -1.5
0.206 0.216 0.226 0.236 L1 L3 L5 L1 L3 L5 L2 L6 1 2 3 4 5
L2 L4 L6 L2 L4 L6 L3 Column number
Time (s)
(a) (b) (c)
3.5
L1 L1

Normalized singular value

Normalized singular value

1 L2
1 L2 3

Membership degree
L3 L3
0.8 L4 0.8 L4 2.5 APFM
L5 L5
0.6 L6 0.6 L6 2
C1 C1
0.4 C2 0.4 C2 1.5 PFM

0.2 1
0.2
0.5 AFM
0 0
0
1 2 3 4 5 2 4 6 8 10 1 2 3 4 5 6
Column number Column number Feeder number
(d) (e) (f)

Fig. 10. Fault detection procedure of the fourth type of earth fault: (a) Zero-sequence current waveforms; (b) Fault detection results of method 1, 2, 3; (c) Elements of AFM and
its cluster center; (d) Elements of PFM and its cluster center; (e) Elements of APFM and its cluster center; (f) Fault detection results of method 4, 5, 6.

results of methods 5 and 6 are correct, and the others are fault feeder correctly under various kinds of fault conditions and
wrong. factors.
The fourth type of earth fault is that, as in Fig. 2, another 15 km
long cable with feeder number 6 is added to the busbar. The earth
fault is under the fault conditions Lm 3, X f 2 km, Rf 5 X, Acknowledgments
h 90 , and p 10%. As shown in Fig. 10(a), the fault transient
zero-sequence current amplitude in sound feeder 6 is much larger This work was sponsored in part by National Natural Science
than that of the other sound feeders. It may be close to the tran- Foundation of China under the project of Research of Flexible and
sient zero-sequence current amplitude of fault feeder 3, which will Adaptive Arc-suppression Method for Single-phase Grounding
easily result in the failure of a fault detection method based on the Fault in Distribution Networks. (Project number: 51677030).
amplitude feature individually. The fault detection results of meth-
ods 1, 5, and 6 are correct, and the others are wrong.
The comparisons of earth fault detection results indicate that References
the earth fault detection method developed in this study has real
[1] Conwell RN, Evans RD. The Petersen earth coil. Trans Am Inst Electr Eng
advantages, compared with the similar conventional methods.
1992;41(2):7793.
[2] Welfonder T, Leitloff V, Fenillet R, Vitet S. Location strategies and evaluation of
detection algorithms for earth faults in compensated mv distribution systems.
5. Conclusion
IEEE Trans Power Del 2000;15(4):11218.
[3] Leitloff V, Feuillet R, Griffel D. Detection of resistive single-phase earth faults in
This study successfully designed a new fault feeder detection a compensated power-distribution system. Eur Trans Electr Power 2007;7
method for a resonant grounding system based on the TFM and (1):6573.
[4] Zeng XJ, Li KK, Chan WL, Su S, Wang Y. Ground-fault feeder detection with
PDM singular-values clustering algorithm. The HHT band-pass filter fault-current and fault-resistance measurement in mine power systems. IEEE
and wave transformation are applied to the waveforms of fault Trans Ind App 2008;44(2):4249.
transient zero-sequence currents in the feeders to structure their [5] Lin X, Huang J, Ke S. Faulty feeder detection and fault self-extinguishing by
adaptive Petersen coil control. IEEE Trans Power Del 2011;26(2):12901.
TFMs and PDMs. The AFM and PFM are obtained from the TFM [6] Lin X, Sun J, Kursan I, Zhao F, Li Z, Li X, et al. Zero-sequence compensated
and PDM, respectively, via the SVD. Then, the FCM clustering is admittance based faulty feeder selection algorithm used for distribution
applied to the APFM so as to determine the fault feeder. The effec- network with neutral grounding through peterson-coil. Int J Electr Power
Energy Syst 2014;63(63):74752.
tiveness of the proposed method was verified by numerical and [7] Wang Y, Zhou J, Li Z, Dong Z, Xu Y. Discriminant-analysis-based single-phase
physical simulations by considering all possible fault scenarios. earth fault protection using improved pca in distribution systems. IEEE Trans
The simulated results for various earth faults indicate: (1) The Power Del 2015;30(4):197483.
[8] Oinis C, Patrick B, Michel M. Pronys method: an efficient tool for the analysis
TFM and PDM obtained via the HHT band-pass filter and wave of earth fault currents in Petersen-coil-protected networks. IEEE Trans Power
transformation can fully characterize the amplitude and polarity Del 1995;10(3):123441.
feature information, respectively, of the transient zero-sequence [9] Chaari O, Meunier M, Brouaye F. Wavelet: a new tool for the resonant
grounded power distribution systems relaying. IEEE Trans Power Del 1996;11
current waveform. (2) The singular value is a relatively stable alge-
(3):13018.
braic feature of a matrix, which can not only represent the feature [10] Cui T, Dong XZ, Bo ZQ, Juszczyk A. Hilbert-transform-based transient/
quantities of fault transient zero-sequence current signals but also intermittent earth fault detection in noneffectively grounded distribution
reduce the feature dimension. (3) FCM clustering is applied to systems. IEEE Trans Power Del 2011;26(1):14351.
[11] Zhang ZX, Liu X, Piao ZL. Fault line detection in neutral point ineffectively
detect the fault feeder without a threshold setting in the fault fee- grounding power system based on phase-locked loop. IET Gener Transm
der detection criterion. (4) The proposed method can detect the Distrib 2014;8(2):27380.
108 M.-F. Guo, N.-C. Yang / Electrical Power and Energy Systems 93 (2017) 97108
[12] Abdel-Fattah MF, Lehtonen M. Transient algorithm based on earth capacitance
estimation for earth-fault detection in medium-voltage networks. IET Gener
Transm Distrib 2012;6(2):1616.
[13] Elkalashy NI, Tarhuni NG, Lehtonen M. Simplified probabilistic selectivity
technique for earth fault detection in unearthed MV networks. IET Gener
Transm Distrib 2009;3(2):14553.
[14] Elkalashy NI, Elhaffar AM, Kawady TA, Tarhuni NG, Lehtonen M. Bayesian
selectivity technique for earth fault protection in medium-voltage networks.
IEEE Trans Power Del 2010;25(4):223445.
[15] Guo MF, Zheng XT, Yang GJ, Gao W, Miao XR. A novel fault line
detection method for resonant earthed system based on transient waveform
stretching transformation. Electric Power Automat Equip 2014;34(9):
3340.
[16] Guo MF, Gao Y, Yang GJ. A new approach to fault line detection in resonant
earthed system based on difference recognition of transient waveforms.
Electric Power Automat Equip 2014;34(5):3542.
[17] Guo MF, Liu SD, Yang GJ. A novel approach to detect fault line in distribution
network using similarity recognition based on time-frequency spectrum. Proc
Chinese Soc Electr Eng 2013;32(3):6774.
[18] Zeng XJ, Yu K, Wang YY, Xu Y. A novel single phase grounding fault protection
scheme without threshold setting for neutral ineffectively earthed power
systems. CSEE J Power Energy Syst 2016;2(3):7382.
[19] Wang YY, Zeng XJ, Dong ZY, Huang Y. Novel protection scheme of stator single-
phase-to-ground fault for powerformers. Int J Electr Power Energy Syst
2013;53(53):3218.
[20] Wang YY, Zeng XJ, Dong ZY, Xu Y, Yuan J, Huang Y. Stator single-phase-to-
ground fault protection for bus-connected powerformers based on hierarchical
clustering algorithm. IEEE Trans Energy Convers 2013;28(4):9918.
[21] 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;454:90395.
[22] Rezaee MR, Lelieveldt BP, Reiber JH. A new cluster validity index for the fuzzy
c-mean. Pattern Recogn Lett 1998;19(3):23746.
[23] Darwish H, Elkalashy N. Universal arc representation using EMTP. IEEE Trans
Power Del 2005;20(2):7729.
[24] Xue YD, Feng ZR, Xu BY, Chen Y, Li J. Earth fault protection in non-solidly
earthed network based on transient zero sequence current comparison.
Automat Electric Power Syst 2003;27(9):4853.