Electric Power Components and Systems

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/uemp20

High Impedance Fault Detection Using Multi-

Domain Feature with Artificial Neural Network

Balu K. Sangeeth & V. Vinod

To cite this article: Balu K. Sangeeth & V. Vinod (2023) High Impedance Fault Detection Using
Multi-Domain Feature with Artificial Neural Network, Electric Power Components and Systems,
51:4, 366-379, DOI: 10.1080/15325008.2023.2172091

To link to this article: https://doi.org/10.1080/15325008.2023.2172091

Published online: 01 Feb 2023.

Submit your article to this journal

Article views: 49

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=uemp20
Electric Power Components and Systems, 51(4): 366–379, 2023
# 2023 Taylor & Francis Group, LLC
ISSN: 1532-5008 print / 1532-5016 online
DOI: 10.1080/15325008.2023.2172091

High Impedance Fault Detection Using

Multi-Domain Feature with Artificial
Neural Network
Balu K. Sangeeth and V. Vinod
Department of Electrical Engineering, College of Engineering Trivandrum, Thiruvananthapuram, Kerala, India

CONTENTS
Abstract—High Impedance Fault (HIF) poses a threat to human
1. Introduction life even though the fault current is low in magnitude. Moreover,
the proliferation of power electronic devices give rise to a range
2. Modeling of High Impedance Fault
of harmonics in the system. Hence it is difficult to identify the
3. Modeling of Non-Linear Load HIF by employing classical approach. In this paper an algorithm is
4. Test System for HIF Study developed that uses signatures in both time and frequency domain
5. Proposed Method for HIF Detection to identify the HIF. Since the parameters are inconsistent at every
situation, it is imperative to use a classifying technique such as
6. Results and Discussions artificial neural network to distinguish the fault and normal
7. Conclusion situation. The algorithm is tested in a IEEE 33 bus system with
References the presence of nonlinear loads. The harmonic loads are simulated
as current source and the harmonic content of the loads are taken
with the help of Harmonic analyzer. The relevant features are
judiciously selected to improve the accuracy of the algorithm to
detect the HIF. The multi-domain features selected are the rate of
change of phase current, change in the angle of the sequence
currents and relative energy of Digital Wavelet Transform (DWT)
coefficients. The proposed algorithm is used to distinguish the
condition such as load switching, bolted ground fault, non-linear
current harmonics from HIF.

1. INTRODUCTION
Protection of distribution feeders has always been a chal-
lenge for power engineers. Most of the faults are cleared
with the help of over-current relays. HIF usually have fault
current magnitude in the range of the load currents, which
makes it difficult to detect the fault with a conventional
over-current relay. Due to the uncertainty in the waveform
of HIF currents, the conventional methods face difficulty to
detect the HIF based on thresh hold values. The presence
of high harmonic content in the fault was studied and used
for detection of such faults, but the non-linear loads in the
Keywords: artificial neural network, high impedance fault, feature vector, distribution system make the detection of HIF is more
harmonic analyzer, confusion matrix, non-linear load, Discrete Wavelet
Transform difficult.
Received 13 January 2021; accepted 15 January 2023 In the literature, methods were proposed in time domain,
Address correspondence to V. Vinod, Department of Electrical
Engineering, College of Engineering Trivandrum, Thiruvananthapuram,
frequency domain and combination of both to detect HIF.
Kerala, India. E-mail: [email protected] A detailed review of different techniques including time

366
 Sangeeth and Vinod: High Impedance Fault Detection Using Multi-Domain Feature with Artificial Neural Network 367

domain, frequency domain and hybrid domains are pre- time–frequency transform and probabilistic neural net-
sented in paper [1, 2]. In the literature, methods were pro- work[14]. A decision tree based classifier [15–17] is
posed in time domain, frequency domain and combination employed to differentiate the small phase-to-earth and
of both to detect HIF. The multi domain technique shows phase-to-phase HIFs. In [18], a wavelet based approach
that wavelet analysis has gained more popularity over the with ANN (artificial neural network) is used for boundary
years in identifying HIF and better results are obtained by protection of lines. The angle relationship between negative
combining different domains [3]. Time domain based meth- and zero-sequence voltages[19] and the angle difference
ods depend on the non-linear characteristics of voltage and between the third harmonics current and the fundamental
current [4, 5]. Even though the method utilize time domain voltage[20] is taken into consideration to detect the ground
analysis, it is also suitable for non-homogeneous system faults and HIF fault respectively. All these methods require
with unbalance loading. But in noisy conditions, the accur- sampling frequency at the rate of mega samples per
acy of the algorithm to predict HIF tends to be decreases second. The high sampling rate induces radio interference’s
due to numerical differentiation in the calculation of volt- and give rise to erroneous results.
age. The method estimates the location of the fault based An effective technique make use of Digital Fourier
on the measured and estimated value of voltage. An analyt- transform and Kalman filter is proposed in [21]. With the
ical expression for the interaction between the electric arc help of machine learning algorithm, a ranking procedure is
during high impedance fault and transmission line is adopted to select a desired feature pool to achieve better
explained in [6]. This is only suitable for HIF detection in results. However, the method is very complex, need large
transmission line and not suitable for distribution system. feature data set and require more calculations. A method
A model based detection proposed in [7], calculate the [22] based on wavelet transform support vector machine
error between modeled and measured disturbances is con- and particle swarm optimization is utilized to detect HIF.
sidered for arc fault detection. A model based arcing fault Even though the method gives good results, it is only
detection scheme is proposed in [8]. The generalized model applicable for neutral ungrounded system. In [23], adaptive
is effectively utilized to differentiate the arcing fault from neuro fuzzy algorithm is applied to detect and classify the
non-arcing disturbances. The accuracy of the method is HIF. Third harmonic components in the fault current give
very much depend upon the model parameters, which may results for fault detection. Also fundamental components
vary between the field measurements. Fault localization is give results for the classification of HIF. Due to high pene-
also proposed in [9, 10] with the help of smart meters by tration of power electronics based loads, the proposed har-
detecting the even harmonic distortion index. The bulk monics based detection shows erroneous results. In paper
numbers of smart meter available in the distribution system [24], distortion-based algorithm is applied to predict the
is a needy requirement for the success of the method. But HIF, According to the distortion, HIF are classified into
the presence of even harmonics due to the penetration of five types. The accuracy of the method depends upon the
power electronics based loads leads to the failure in HIF variation, imbalance and harmonic from the load side.
detection. Reference [11] utilizes different algorithm Even though the multi domain based and machine learn-
including changes in the value of the energy of a specific ing based techniques give good results, the proliferation of
frequency band to enhance the confidence of fault detec- non-linear loads and DGs (Distributed generators) in the
tion. A transient based scheme discussed in [12] utilized system may lead to undesired results and this has to be
discrete Fourier transform to detect the HIF. High and low addressed for efficient detection of HIF fault. Moreover,
frequency voltage components at various points in the dis- the multi domain based methods find difficulty to predict
tribution system are analyzed for detection. The need of a an upper threshold to detect the HIF fault. Therefore a
communication infrastructure and a dedicated distribution method based on time and frequency domain analysis
network operation center put together the method more along with a good classification technique like artificial
expensive. neural network can be used to address all the aforemen-
The random behavior in the waveform characteristics tioned issues. The feature selection should be very careful
led to the use of feature extraction and classification tech- to improve the accuracy of the proposed method. Also the
niques for the detection of HIF. Inductive reasoning along calculation involved in the proposed scheme should be less
with pattern classification based on entropy minimization complex and has real time implementation capability. The
has been proposed in [13] to detect the fault. High imped- scheme is also very attractive, if the detection is possible
ance fault is detected in power distribution networks using without any communication burden.
368 Electric Power Components and Systems, Vol. 51 (2023), No. 4

In this work, an algorithm to detect high impedance 2. MODELING OF HIGH IMPEDANCE FAULT
fault in the presence of non-linear load on distribution
Fault current during HIF may vary and usually, these faults are
feeders is proposed. The nonlinear loads in the network
accompanied by arcing. The nature of fault current is quite diffi-
such as One ton Air conditioner (AC), AC with V-
cult to predict due to its random nature. One of the most common
STAB stabilizer (AC STB), 10 W CFL, 5 W LED are
signatures accompanied by a HIF is the arc formation with the
analyzed with the help of Yokogawa WT1600 Digital
fault current. Arc is formed when the voltage of the conductor
power meter. The uncertainty in the pattern of the non-
exceeds the breakdown voltage of the medium and vice versa.
linear current waveform to detect the HIF can handled
Apart from the formation of the arc and its low current magni-
efficiently with the help of Artificial Neural Network
tude, HIF has many other physical features such as asymmetry,
(ANN). The efficiency of the ANN to detect the HIF is
very much dependent on the selection of features for randomness, and non-stationary. The asymmetry in the current
training. The multi-domain features extracted are the waveform is occurs due to the change in the magnitude of peak
relative energy of DWT coefficients, the rate of change values for each half cycle of the HIF current waveform. This is
of phase current and also the change in the angle of the due to the fact that positive and negative voltages have different
sequence currents. The relative energy of detailed and breakdown voltages. The main factor in the randomness because
approximate coefficients are calculated with the help of of the random behavior of the conduction/non-conduction inter-
the discrete wavelet transform. These are fed as input vals of the HIF current. The non-stationary current is formed
into the artificial neural network. ANN Classifier is because of the change in the frequency spectrum of the current
used to distinguish the HIF from non-HIF cases. The waveform. Because of the arc formation, high frequency current
efficacy of the proposed algorithm is evaluated through content is also present in the waveform. Figure 1 is the HIF
Electromagnetic transient Programs (EMTP) simula- model [25, 26] used and it connects one phase of the power line
tions on an IEEE 33 Bus system. to the ground. The randomness of the variable for representing
The work presented in the paper is organized into fol- resistor (R1,R2) and DC value of the battery (V1, V2) in the HIF
lowing sections. Sections 2–4 explain the modeling of HIF, model is achieved using Gaussian distribution. The range of the
modeling of non-linear load and the test system under random variable and the time at which the value changes con-
study. Section 5 outlines the proposed method for HIF tinuously are tabulated in Table 1. This model is used to simulate
detection and the various feature used in the method. The a HIF current waveform as shown in Figure 2 and thereby valid-
results and discussions are presented in Section 6. ate the non-stationary behavior of the frequency spectrum.
Section 7 concludes the paper.
3. MODELING OF NON-LINEAR LOAD
In order to simulate the nonlinear loads in the network, the
current waveform of harmonic producing loads needs to be
analyzed. The most commonly used consumer loads such
as One ton Air conditioner (AC), AC with V-STAB stabil-
izer (AC STB), 10 W CFL, 5 W LED are taken into
account. With the help of Yokogawa WT1600 Digital
power meter, harmonic components up to 22nd order are
analyzed. The magnitude and phase of each harmonic com-
ponent is obtained as in Table 2. The harmonic producing

Time at which
Random value changes
variable Value range continuously

V2 3–4 kV 0.25 ms
V1 2–3 kV 0.25 ms
R1 100–700X 0.25 ms
R2 100–700X 0.25 ms

FIGURE 1. HIF model. TABLE 1. HIF model parameters.

 Sangeeth and Vinod: High Impedance Fault Detection Using Multi-Domain Feature with Artificial Neural Network 369

loads are modeled in the PSCAD with the help of available 4. TEST SYSTEM FOR HIF STUDY
current dependent sources. Different current sources of har-
The test system used for the simulation is a modified IEEE
monic frequency up to 22nd order are added to simulate the
33 Bus system of 12.66 kV as shown in Figure 4. The
current waveform pattern of actual loads. The general
detailed parameter values of the IEEE 33 Bus system are
equation for the simulated current waveform is
tabulated in Appendix A. Star-Delta step down transformers
i ¼ ia þ i1 sin x1 t þ i2 sin x2 t þ       þ i22 sin x22 t are employed to transform the voltage from 12.66 kV to
(1) 415 V. The combinations of modeled non-linear loads are
connected at different buses of IEEE 33 distribution system
The comparison of the current waveform of modeled loads
for further study of HIF fault detection. Also, the HIF model
with actual loads are shown in Figure 3.
is used to create HIF fault at various buses in the system.
The harmonic level of the system was limited in accordance
with the standard prescribed by the IEEE 519. Instead of cal-
culating the THD, the IEEE 519 prefer the TDD in case of
current harmonics. Total Current Demand Distortion (TDD)
is defined as the total harmonic current distortion against the
full load (demand) level of the electrical system. So TDD
gives better insight about the impact of harmonic distortion
in our system. The loads are injected randomly in to the sys-
tem keeping in mind the limit of TDD and respective indi-
vidual harmonic order. The harmonic limits are kept within
the limit at PCC (metering point) with a slight increase in
harmonic levels for 100 percentage load. The TDD value for
different load conditions in accordance with IEEE 519 stand-
FIGURE 2. HIF fault current. ards are given in Table 3.

CFL lamp LED bulb AC AC STB

hn
Magnitude Phase Magnitude Phase Magnitude Phase Magnitude Phase

1 0.281 0 0.002 0 10.627 0 10.637 0

2 0.001 29.85 0.002 17.13 1.526 96.7 1.666 18.9
3 0.244 178.85 0.001 111.16 0.979 125.08 0.923 157.05
4 0.003 119.93 0.002 61.85 0.128 149.14 0.103 22.48
5 0.183 8.226 0.002 153.38 0.894 78.29 0.912 136.37
6 0.003 59.45 0.002 89.39 0.172 116.26 0.115 142.91
7 0.134 154.68 0.002 144.67 0.24 167.63 0.151 169.91
8 0.004 8.18 0.003 188.63 0.096 3.53 0.081 126.67
9 0.107 43.86 0.0006 149.45 0.041 20.59 0.004 106.23
10 0.002 158.87 0.0007 35.54 0.016 75.6 0.029 86.95
11 0.076 117.7 0.0007 143.89 0.078 59.41 0.069 97.07
12 0.001 123.01 0.0007 99.98 0.022 121.94 0.006 39.23
13 0.057 85.5 0.0008 3249 0.012 34.7 0.006 28.83
14 0.004 33.92 0.0002 37.01 0.006 161.74 0.012 76.93
15 0.045 56.72 0.0002 21.04 0.031 103.15 0.021 74.57
16 0.003 58.51 0.0002 55.57 0.004 6.78 0.012 16.39
17 0.048 163.44 0.0002 0.6 0.03 158.77 0.015 86.66
18 0.005 69.2 0.0002 99.27 0.008 142.91 0.005 10.89
19 0.04 0.81 0.0002 34.55 0.031 50.02 0.038 155.03
20 0.009 150.52 0.0002 92.97 0.002 114.66 0.008 46.27
21 0.038 150.71 0.0002 13.02 0.013 56.61 0.024 161
22 0.005 50.59 0.0001 129.15 0.003 102.72 0.008 31.15
TABLE 2. Magnitude and phase of various harmonic order (hn).
370 Electric Power Components and Systems, Vol. 51 (2023), No. 4

FIGURE 3. Comparison of actual and simulated current waveforms of various non-linear loads.
 Sangeeth and Vinod: High Impedance Fault Detection Using Multi-Domain Feature with Artificial Neural Network 371

5. PROPOSED METHOD FOR HIF DETECTION Therefore the execution time including the acquisition
time should be limited to 250l s. The difference in cur-
The uncertainty in the pattern of the current waveform
rent magnitude calculated between the consecutive
is one of the main challenge encountered during HIF.
cycles form the features (DA, DB and DC) is suitable as
Artificial Neural Network (ANN) is an advanced tool to
an excellent feature for ANN training.
identify the HIF by analyzing the current waveform in a
detailed manner. The efficiency of the ANN to detect
the HIF is very much dependent on the selection of fea- 5.2. Angle Between Negative and Zero Sequence
tures for training. The features used in the training pro- Current
cess and the corresponding symbols are given in Table Suppose a single line to ground fault is occur at phase
4. The details of the selected features are explained in A of the feeder with no load. Due to fault, the phase A
the following subsections. voltage will fall to near zero and the magnitude of the
phase current will shoot up. Also, the fault current IA
can be expressed as combination of symmetrical com-
5.1. Magnitude Change in Phase Currents
ponents. The negative sequence current IA2 and zero
In a distribution feeder the magnitude of the current sequence current IA0 in the unbalanced system should
changes with the fault. This rate of change of fault cur- aid the positive sequence current IA1 and also aligned in
rent is used to identify the initiation of the fault. The the same phase as that of IA1 : Therefore considerable
magnitude of the fundamental frequency is extracted change is happened in the negative and positive current
with the help of Fourier transform. Here, the Fourier phasor after the fault. Moreover the positive sequence
extraction algorithm is used to find the magnitude of the current phasor after the fault is not deviated much.
currents of Phase A, Phase B, and Phase C. This feature Therefore, the angle difference between negative and
is selected based on the possibility of reduction in cur- zero sequence current is selected as a prominent feature
rent during HIF. Fourier algorithm is performed at 80 to identify the ground faults. The magnitude and phase
samples per cycle or a sampling frequency of 4 kHz. of the positive component I1 /h1 , negative component
I2 /h2 and zero component I0 /h0 currents are found by
using 2.
2 3 2 3 2 3
I0 /h0 1 1 1 Ia /ha
6 7 6 7 6 7
4 I1 /h1 5 ¼ 4 1 a a2 5  4 Ib /hb 5 (2)
I2 /h2 2
1 a a Ic /hc

The negative sequence angle and zero sequence angle are

monitored continuously and the difference between both
the angles is taken as 3. This is given also as one of the
input feature vectors to Artificial Neural Network. Phase A
is taken as the reference point and the difference between
the angles are fed to the ANN network accordingly.
/h2  /h0 (3)

Maximum harmonic current distortion in percent of load current

Individual harmonic order
Load 3<¼ 11 < ¼ 17 < ¼
percentage h < 11 h < 17 h < 23 TDD

50% 3.73 .96 .53 3.889

70% 4.1 1.09 .64 4.38
100% 11.81 3.44 1.32 12.39

FIGURE 4. Test system for HIF study. TABLE 3. Harmonics limit at different load conditions.
372 Electric Power Components and Systems, Vol. 51 (2023), No. 4

Features Symbols insignificant during transformer and capacitor switching.

Therefore, the change in angle is considered as a vital
Change in the magnitude DA, DB, DC feature for HIF detection.
of phase A, phase B,
and phase C currents
Angle difference between /h2  /h0 5.3. Relative Energy of Digital Wavelet Transform
negative and zero (DWT) Coefficients
sequence current
Wavelet transform [27] gives time-frequency resolution of
Relative energy of A2, RelAA2, RelAD2
D2, D1 of phase A the given signal. The DWT of a signal is given by
phase B and phase C  
1 X k  naj
RelAD1, RelBA2 DWTðj, nÞ ¼ pffiffiffi xðnÞ  g (4)
RelBD2, RelBD1 a n aj
RelCA2, RelCD2
RelCD1. Here, x(n) represents the discrete input signal and g(n) rep-
TABLE 4. Different features for ANN training. resents the mother wavelet. Daubechies and Coiflet are the
popular mother wavelet employed to detect the presence of
fast transients. Due to the resemblance of the arc fault
Figure 5 depicts the change in sequence angle before
waveform, Daubechies 4 (db4) has been selected as the
and after the HIF at 6th Bus for 70 % load. As per
Figure 5(a), the angle difference between the positive mother wavelet. The signal can be decomposed into signals
sequence current before fault (IA1 ) and after fault (IA1F ) of different frequency ranges according to the level of
is negligible. Similarly, the angle difference is insignifi- decomposition and sampling frequency FS : For the first
cant for the other phases. Therefore the angle of positive level decomposition the signals in the range FS =2 to FS =4
sequence current is not a potential candidate to select as is known as detailed coefficients ðd1 Þ and signals in the
a feature for ANN training. However, the change in range of 0 to FS =4 is termed as approximate coefficient
angle is noticeable for both negative and zero sequence ða1 Þ: The approximate coefficients are decomposed to cal-
currents. With respect to Figure 5(b, c), angle difference culate further level of DWT coefficients. The energy of the
between the IA2 and IA0 is found as 380 (132 to 170) detailed and approximate coefficients are used to detect the
before the fault. After the fault, the angle difference presence of high impedance fault. Figure 6 shows the ori-
between the IA2 and IA0 altered to 233:80 (173.7 to ginal current signal, d1 and d2 coefficients for a HIF at
60.81). During no-load situation, the angle difference is 2500 sample instant. The corresponding variation of ampli-
found to be 2400 : The deviation in angle from 2400 hap- tude in d1 and d2 coefficients can easily visible after 1250
pens due to the presence of load current. Due to the and 625 sample instant in the first and second level of
absence of zero sequence current, the change in angle DWT. This sustained amplitude variation of coefficients is
between the negative and zero sequence current is considered as an excellent indication to detect the HIF

FIGURE 5. Phasor diagram showing the change in sequence angle before and after the HIF at 6th Bus for 70 % load.
(a) Positive sequence current, (b) negative sequence current, and (c) zero sequence current.
 Sangeeth and Vinod: High Impedance Fault Detection Using Multi-Domain Feature with Artificial Neural Network 373

from other disturbances such as SLG, capacitor switching, relative energy changes are minimal for switching distur-
load switching, and transformer energization. bances. This is due to the absence of sustained presence of
The energy spectrum can vary with the amount of loads, DWT coefficients. The relative energy changes during SLG
it is difficult to predict an upper threshold to detect the fault is appreciable and reduce to a lower value within two
HIF fault. Moreover, the range of the energy variation is cycle or 160 samples. Also, for a HIF of 800 ohm, the rela-
very high depends upon the load current and the fault dis- tive energy changes are not considerable due to the low
tance from the source. To bring the threshold in a normal- magnitude of fault current. Moreover, the presence of high
ized manner, relative energy is taken into consideration. harmonic loads makes it difficult to predict the HIF under
Therefore the relative energy of each decomposition coeffi- all circumstances. Therefore, the need of ANN is very
cient with respect to the total energy is taken as a reliable essential to classify the HIF from other disturbances under
threshold to differentiate HIF from other disturbances. The all prevailed worst situations.
relative energy of d1 coefficient for A phase is calculated
as per (5). Similarly, the relative energy of all other coeffi- 5.4. Artificial Neural Network
cients of all the other phases is also calculated.
ANN is one of the powerful tools in classification and pat-
kþðDkcycle =2jÞ
X tern recognition. The ANN architecture shown in Figure 8
Ed1 ¼ d12 ðnÞ is a feed-forward network with an input layer, hidden
n¼k
layer, and output layer. The selected feature vectors (13
kþðDkcycle =4jÞ
X numbers) for the detection of HIF act as the input of the
Ed2 ¼ d22 ðnÞ
n¼k (5) ANN. The number of neurons in the hidden layer is
kþðDkcycle =4jÞ selected such that the trained data will give very less or
X
Ea2 ¼ a22 ðnÞ zero error. This layer determines how well an ANN can
n¼k classify the given signal. The mapping is done from a 13
ETOTAL ¼ Ed1 þ Ed2 þ Ea2 dimension to an output layer with one dimension (one neu-
RelAD1 ¼ Ed1 =ETOTAL ron). The non-linear tan sigmoid function in accordance
where Dkcycle is the number of samples per cycle and j is with (7) is used in the hidden layer and the sigmoid func-
the energy level. The relative energy of detailed coefficient tion represented by (8) is used in the output layer. The
d1 and d2 under different cases happened at 6th bus are sum of the products of synaptic weight corresponding to
shown in Figure 7. It is clear from the figure that the each neuron is given as (6). The synaptic weights are

FIGURE 6. Original current signal and detailed coeffi- FIGURE 7. (a) Relative energy of d1 coefficient and (b)
cients for HIF at 2500 sample instant for linear loads. relative energy of d2 coefficient.
374 Electric Power Components and Systems, Vol. 51 (2023), No. 4

adjusted with conjugate gradient back propagation and 5.6. Flow Chart of the HIF Detection Scheme
steepest descent method. The training or supervised learn- The flow chart of the HIF protection scheme is shown in
ing of the network is performed in an offline mode. Figure 9. Phase currents Ia, Ib, and Ic are sampled at 80
samples per cycle. The Fourier extraction technique is
uðxÞ ¼ h1b B1 þ h11 x1 þ h21 x2 þ h31 x3 þ    þ h13
1 x13 (6) employed to calculate the magnitude and phase angle of
the currents. Thereafter, the algorithm calculates the magni-
tude change in phase currents and the angle between the
2 negative and zero sequence currents. The scheme directly
rðuÞ ¼ tanhðuÞ ¼ 1 (7)
1 þ e2u calculates the nine different relative energy values corre-
sponding to all the phases from the sampled current sig-
nals. All these 13 relevant features are given as input to
1 calculate the output of ANN. The constants and bias
/ðhÞ ¼ (8) needed for calculating the value of ANN outputs are
1 þ eh
obtained during training. If the output value is above 0.5,
HIF is detected. Otherwise, if the output value is less than
0.5, indicate the non-detection of HIF.
5.5. ANN Training
The efficacy of the proposed scheme to detect the HIF is
6. RESULTS AND DISCUSSIONS
entirely dependent upon the proper training of the ANN.
The datasets for training are created by simulating the HIF The data needed for training the ANN is taken by creating
at various buses in the system. The output corresponding different cases at different buses. Datasets are created by
to the HIF is mapped to ”ONE” during training. The test simulating the HIF, non-faulty, switching conditions and
cases related to capacitor switching, transformer switching, Single Line to Ground (SLG) fault cases. HIF cases are
SLG fault and normal operation under non-linear load are simulated for fault impedance of 20 ohm and 800 ohm and
mapped to ”ZERO” in the output of the ANN. All the the faults are created at buses 3, 6, 11, 18, 22, 25 and 33.
above test cases are simulated with 50%, 70% and 100% These data are mapped to the output 1 in the ANN for the
load to incorporate the situations arising in a practical dis- algorithm to detect the HIF situations. Transformer switch-
ing, SLG fault and normal operation under nonlinear loads
tribution feeder.
are considered as non-HIF situations and the data are
mapped to the output zero in ANN. The above test cases
are simulated with 100, 70 and 50 percentage load to
incorporate all the situation in a practical distribution
feeder. A total of 300 cases are tested including 240 sam-
ples for training, 30 for validation and 30 for testing. The
input features magnitude are of different ranges and they
needed to be normalized. The normalized data is fed into
the neural network model for training.
The conjugate gradient descent algorithm is utilized to
generate the ANN model and associated parameters. As
shown in Figure 10, the algorithm was converged at 68th
iteration with the mean squared error value of .34273 after
6 consecutive increase in the cross entropy value. The con-
jugate gradient descent reached a gradient of .0473 on con-
vergence. The error histogram shown in Figure 11 clarified
that for the target of 1 there was a maximum deviation of
.4779 under permissible range and there were a total of 7
cases are mapped out of the range with max deviation of
0.9503. similarly for the target of zero the max deviation
FIGURE 8. ANN architecture. under permissible range was .4669 and there were 19
 Sangeeth and Vinod: High Impedance Fault Detection Using Multi-Domain Feature with Artificial Neural Network 375

FIGURE 10. Cross-entropy vs iterations during training.

FIGURE 11. Error histogram.

the relative energy for each of the DWT coefficients. The

change in the current magnitude levels for the SLG fault is
FIGURE 9. The proposed algorithm.
relatively high in magnitude when compared with the HIF
faults. The current change that occurs during load switch-
cases that did not map correctly with a maximum error of ing is very low in magnitude. In the case of HIF, as shown
.8448. in Figure 12(f), the output of the ANN tends to increase
Figure 12 shows the influence of various features used above the threshold of 0.5 and therefore able to identify
to produce the desired ANN output to detect the HIF. The the HIF. During SLG and load switching, the output of the
different cases considered are HIF of 20 ohm and 800 ohm, ANN is below the threshold level of 0.5 and hence identify
SLG fault, and load switching. All the events are simulated the disturbances as non-HIF. Therefore the proposed algo-
at 620th sampling instant. As per Figure 12(a), the angle rithm gives a fair response to detect the HIF even during
between negative and zero sequence current has been extreme situations.
changed to a value nearer to 2400 : But in the case of load Confusion matrix is used for evaluating the HIF detec-
switching, the angle between negative and zero sequence tion performance of the proposed method. In the matrix,
current will more or less remain the same. The correspond- the column represents the output class and rows indicate
ing angle difference in the case of SLG fault is 1400 : the target class. Out of 300 different cases considered, 80%
During HIF, there is a considerable amount of change in is taken for training, 10% is used for validation and the
376 Electric Power Components and Systems, Vol. 51 (2023), No. 4

FIGURE 12. Different features and ANN output.

remaining 10% is for testing. Figure 13 shows the classifi- current, during fault at a location far away from the source.
cation accuracy including the training, validation, and test- Even though the signatures are faded, the ANN output is
ing cases. As shown in Figure 13, 144 numbers of HIF still above the threshold value of 0.5 and able to identify
cases are correctly classified and 10 numbers of HIF cases the HIF.
are unclassified. Similarly, 139 numbers of HIF cases are
classified correctly and 7 numbers of HIF cases are not
7. CONCLUSION
identified correctly. Therefore, the total accuracy of the
proposed method has been calculated as 94.33% ( 144þ139
300 ).
Due to the proliferation of harmonics producing loads, the
The HIF of 20 ohm and 800 ohm are created at various detection of HIF is a major challenge in distribution side
buses from bus number 1 to 33. ANN output of different of the power system. Considering this, the proposed work
buses under 50%, 70%, and 100% of load are shown in collects the current waveform of different harmonic pro-
Figure 14(a–c). The ANN output is nearly equal to one for ducing loads using Yokogawa WT1600 Digital power
the buses nearer to the source. The signatures of the con- meter. The work also identifies the relevant features
sidered features are diminished due to the reduction in extracted to detect the HIF. The features considered are the
 Sangeeth and Vinod: High Impedance Fault Detection Using Multi-Domain Feature with Artificial Neural Network 377

shows the accurate detection of HIF and also differentiates

the HIF from load switching, SLG, etc.

ORCID
V. Vinod http://orcid.org/0000-0001-6057-1188

REFERENCES
[1] A. Ghaderi, H. E. L. Ginn, III and H. A. Mohammadpour,
“High impedance fault detection: a review,” Electr. Power
Syst. Res., vol. 143, pp. 376–388, 2017. DOI: 10.1016/j.
epsr.2016.10.021.
[2] G. N. Lopes, V. A. Lacerda, J. C. M. Vieira and D. V.
Coury, “Analysis of signal processing techniques for high
impedance fault detection in distribution systems,” IEEE
Trans. Power Deliv., vol. 36, no. 6, pp. 3438–3447, 2021.
DOI: 10.1109/TPWRD.2020.3042734.
[3] M. Biswal, S. Ghore, O. P. Malik and R. C. Bansal,
“Development of time-frequency based approach to detect
high impedance fault in an inverter interfaced distribution
FIGURE 13. Confusion matrix.
system,” IEEE Trans. Power Deliv., vol. 36, no. 6,
pp. 3825–3833, 2021. DOI: 10.1109/TPWRD.2021.3049572.
[4] S. H. Mortazavi, Z. Moravej and S. M. Shahrtash, “A searching
based method for locating high impedance arcing fault in distri-
bution networks,” IEEE Trans. Power Deliv., vol. 34, no. 2, pp.
438–447, 2019. DOI: 10.1109/TPWRD.2018.2874879.
[5] B. Wang, J. Geng and X. Dong, “High-impedance fault
detection based on nonlinear voltage–current characteristic
profile identification,” IEEE Trans. Smart Grid, vol. 9, no.
4, pp. 3783–3791, 2018. DOI: 10.1109/TSG.2016.2642988.
[6] V. Torres, S. Maximov, H. F. Ruiz and J. L. Guardado,
“Distributed parameters model for high-impedance fault
detection and localization in transmission lines,” Electr.
Power Compon. Syst., vol. 41, no. 14, pp. 1311–1333, 2013.
DOI: 10.1080/15325008.2013.823256.
[7] M. Mitolo, R. Musca and G. Zizzo, “A cost-effective solu-
tion for clearing high-impedance ground faults in overhead
low-voltage lines,” IEEE Trans. Ind. Appl., vol. 55, no. 2,
pp. 1208–1213, 2019. DOI: 10.1109/TIA.2018.2884927.
[8] W. Zhang, Y. Jing and X. Xiao, “Model-based general arc-
ing fault detection in medium-voltage distribution lines,”
IEEE Trans. Power Deliv., vol. 31, no. 5, pp. 2231–2241,
2016. DOI: 10.1109/TPWRD.2016.2518738.
[9] A. Radhakrishnan and S. Das, “Location of high impedance
faults using smart meters in distribution systems with DGs,
FIGURE 14. ANN output during HIF at (a) 50%, (b) power electronic loads and electric arc furnaces,” 2019 IEEE
70%, and (c) 100% load. Milan PowerTech, pp. 1–6, 2019. DOI: 10.1109/PTC.2019.
8810403.
[10] S. Chakraborty and S. Das, “Application of smart meters in
high impedance fault detection on distribution systems,”
relative energy of DWT coefficients, the rate of change of
IEEE Trans. Smart Grid, vol. 10, no. 3, pp. 3465–3473,
phase current and also the change in the angle of the 2019. DOI: 10.1109/TSG.2018.2828414.
sequence currents. ANN technique is applied to detect the [11] C. L. Benner and B. D. Russell, “Practical high-impedance
HIF, even during extreme situations. The simulated result fault detection on distribution feeders,” IEEE Trans. Ind.
378 Electric Power Components and Systems, Vol. 51 (2023), No. 4

Appl., vol. 33, no. 3, pp. 635–640, 1997. DOI: 10.1109/28. [22] D. Jinxiao, L. Yuanlong and X. Zhongyi, “Research on sin-
585852. gle-phase high impedance fault identification of neutral
[12] W. C. Santos, F. V. Lopes, N. S. D. Brito and B. A. Souza, ungrounded system,” 2020 IEEE/IAS Industrial and
“High-Impedance Fault Identification on Distribution Commercial Power System Asia (I&CPS Asia), pp. 851–
Networks,” IEEE Trans. Power Deliv., vol. 32, no. 1, 855, 2020. DOI: 10.1109/ICPSAsia48933.2020.9208552.
pp. 23–32, 2017. DOI: 10.1109/TPWRD.2016.2548942. [23] M. S. A. Aziz, M. A. M. Hassan and E. A. Zahab, “High-
[13] C. J. Kim and B. D. Russell, “A learning method for use in impedance faults analysis in distribution networks using an
intelligent computer relays for high impedance faults,” IEEE adaptive neuro fuzzy inference system,” Electr. Power
Trans. Power Deliv., vol. 6, no. 1, pp. 109–115, 1991. DOI: Compon. Syst., vol. 40, no. 11, pp. 1300–1318, 2012. DOI:
10.1109/61.103728. 10.1080/15325008.2012.689418.
[14] S. R. Samantaray, B. K. Panigrahi and P. K. Dash, “High [24] M. Wei, W. Liu, H. Zhang, F. Shi and W. Chen,
impedance fault detection in power distribution networks “Distortion-based detection of high impedance fault in dis-
using time-frequency transform and probabilistic neural tribution systems,” IEEE Trans. Power Deliv., vol. 36, no.
network,” IET Trans. Gener., Transm. Distrib., vol. 2, 3, pp. 1603–1618, 2021. DOI: 10.1109/TPWRD.2020.
pp. 261–270, 2008. DOI: 10.1049/iet-gtd:20070319. 3011930.
[15] D. P. S. Gomes, C. Ozansoy and A. Ulhaq, “High-sensitiv- [25] S. R. Nam, J. K. Park, Y. C. Kang and T. H. Kim, “A mod-
ity vegetation high-impedance fault detection based on sig- eling method of a high impedance fault in a distribution sys-
nal’s high-frequency contents,” IEEE Trans. Power Deliv., tem using two series time-varying resistances in EMTP,”
vol. 33, no. 3, pp. 1398–1407, 2018. DOI: 10.1109/ 2001 Power Engineering Society Summer Meeting.
TPWRD.2018.2791986. Conference Proceedings (Cat. No.01CH37262), pp. 1175–
[16] K. Sekar and N. A. K. Mohanty, “Data mining-based high 1180, vol. 2, 2001. DOI: 10.1109/PESS.2001.970231.
impedance fault detection using mathematical morphology,” [26] J. Doria-Garcıa, C. Orozco-Henao, R. Leborgne, O. D.
Comput. Electr. Eng., vol. 69, pp. 129–141, 2018. DOI: 10. Montoya and W. Gil-Gonzalez, “High impedance fault mod-
1016/j.compeleceng.2018.05.010. eling and location for transmission line,” Electr. Power Syst.
[17] A. Nakho, K. Moloi and Y. Hamam, “High impedance fault Res., vol. 196, p. 107202, 2021. DOI: 10.1016/j.epsr.2021.
detection based on HS-transform and decision tree 107202.
techniques,” 2021 Southern African Universities Power [27] Z. Xiangjun, K. K. Li and W. L. Chan, “Wavelet analysis
Engineering Conference/Robotics and Mechatronics/ based protection for high impedance ground fault in supply
Pattern Recognition Association of South Africa systems,” Proceedings of the International Conference on
(SAUPEC/RobMech/PRASA), pp. 1–5, 2021. DOI: 10. Power System Technology, pp. 275–279, vol. 1, 2002. DOI:
1109/SAUPEC/RobMech/PRASA52254.2021.9377236. 10.1109/ICPST.2002.1053548.
[18] N. Zhang and M. Kezunovic, “Transmission line boundary
protection using wavelet transform and neural network,”
IEEE Trans. Power Deliv., vol. 22, no. 2, pp. 859–869,
2007. DOI: 10.1109/TPWRD.2007.893596. BIOGRAPHIES
[19] D. C. Yu and S. H. Khan, “An adaptive high and low
Balu K. Sangeeth, MTECH in power system, and affili-
impedance fault detection method,” IEEE Trans. Power
Deliv., vol. 9, no. 4, pp. 1812–1821, 1994. DOI: 10.1109/ ation is College of Engineering, Trivandrum. Research
61.329514. interests are power system protection, high impedance fault
[20] S. Huang, L. Luo and K. Cao, “A novel method of ground detection, modelling of power system components, machine
fault phase selection in weak-infeed side,” IEEE Trans. learning application in power systems.
Power Deliv., vol. 29, no. 5, pp. 2215–2222, 2014. DOI: 10.
1109/TPWRD.2014.2322073.
V. Vinod, PhD, and affiliation is College of Engineering,
[21] Q. Cui, K. El-Arroudi and Y. Weng, “A Feature Selection
Method for High Impedance Fault Detection,” IEEE Trans. Trivandrum. Research interests are power system protec-
Power Deliv., vol. 34, no. 3, pp. 1203–1215, 2019. DOI: 10. tion, micro-grids, ac traction systems, compensation of cur-
1109/TPWRD.2019.2901634. rent transformer saturation, HVDC protection.
 Sangeeth and Vinod: High Impedance Fault Detection Using Multi-Domain Feature with Artificial Neural Network 379

APPENDIX A: IEEE 33 BUS-LINE AND BUS DETAILS

Sl no: Sending End Recieving End R (pu) X (pu) PL QL

1 1 2 0.0922 0.0470 100.0 60.0

2 2 3 0.4930 0.2511 90.0 40.0
3 3 4 0.3660 0.1864 120.0 80.0
4 4 5 0.3881 0.1941 60.0 30.0
5 5 6 0.8190 0.7070 60.0 20.0
6 6 7 0.1872 0.6188 200.0 100.0
7 7 8 0.7114 0.2351 200.0 100.0
8 8 9 1.0300 0.7400 60.0 20.0
9 9 10 1.0440 0.7400 60.0 20.0
10 10 11 0.1966 0.0650 45.0 30.0
11 11 12 0.3744 0.1238 60.0 35.0
12 12 13 1.4680 1.1550 60.0 35.0
13 13 14 0.5416 0.7129 120.0 80.0
14 14 15 0.5910 0.5260 60.0 10.0
15 15 16 0.7463 0.5450 60.0 20.0
16 16 17 1.2890 1.7210 60.0 20.0
17 17 18 0.7320 0.5740 90.0 40.0
18 2 19 0.1640 0.1565 90.0 40.0
19 19 20 1.5042 1.2554 90.0 40.0
20 20 21 0.4095 0.4784 90.0 40.0
21 21 22 0.7089 0.9373 90.0 40.0
22 3 23 0.4512 0.3083 90.0 50.0
23 23 24 0.8980 0.7091 420.0 200.0
24 24 25 0.8960 0.7011 420.0 200.0
25 6 26 0.2030 0.1034 60.0 25.0
26 26 27 0.2842 0.1447 60.0 25.0
27 27 28 1.0590 0.9337 60.0 20.0
28 28 29 0.8042 0.7006 120.0 70.0
29 29 30 0.5075 0.2585 200.0 600.0
30 30 31 0.9744 0.9630 150.0 70.0
31 31 32 0.3105 0.3619 210.0 100.0
32 32 33 0.3410 0.5302 60.0 40.0