On the Use of Artificial Intelligence for High

Impedance Fault Detection and Electrical Safety

Shiyuan Wang, Student Member, IEEE and Payman Dehghanian, Senior Member, IEEE

Abstract—Accidents caused by faults on overhead power lines performance is observed in their detection logic when facing
have been more frequently reported under extreme weather con- the HIF events. Electrical safety studies have shown that
ditions and may strongly threaten the safety and stability of the conventional protection schemes detected and cleared only
power grids, e.g., massive wildfires caused by the electrical arcs
or lines getting in touch with vegetation, relay miss-operations, 17.5% of staged HIFs [3]–[5]. Therefore, HIF detection and
etc. It has been widely recognized that the electric safety concerns localization in electrical power systems yet remain a safety-
engendered by overhead line faults have to be timely and properly threatening challenge for power system protection engineers,
addressed to minimize the subsequent risks and damages. The a fast and accurate solution to which is urgently needed to
existing monitoring devices and protective relays can barely limit the safety risks, prevent power grid operation violations,
detect high impedance faults (HIFs) and are unable to warn
the system operators until serious abnormalities or damages and save human lives [6], [7].
are observed. Aiming at avoiding the damaging consequences Many research works have been conducted in diagnosing
of HIFs, an online monitoring system embedded with machine HIFs, each with some advantages and disadvantages; reference
learning analytics is proposed that ensures a fast and accurate [8] harnessed the high-frequency content in real vegetation
detection of HIFs in power systems. The performance of the fault signatures and proposed a method for the detection of
proposed artificial intelligence engine is tested under a variety of
simulated conditions and the numerical results demonstrate its distinct and very small-current HIFs. In [9], an HIF detection
efficacy and superiority over the state-of-the-art advancements. approach for power distribution networks is suggested using
fuzzy logic control that evaluates the 3rd and 5th harmonics
Index Terms—Convolutional Neural Network (CNN); electrical
safety; event detection; high impedance fault (HIF); feature in the electrical current signals. However, due to the dif-
extraction; wavelet transform (WT). ferent HIFs characteristics compounded by the existence of
harmonics and noises in the power waveforms, their accuracy
and speed performance may be compromised in real-world
I. I NTRODUCTION scenarios. A scheme to detect HIFs using the Time-Time
Electricity has dramatically changed our daily lives and (TT) transform that analyzes and determines the fault wave
fuels our modern society. The electricity grid is constantly patterns is introduced in [10]; this scheme can handle low
exposed, and yet vulnerable, to a wide range of threats, signal to noise ratio (SNR) in power waveforms through a
some foreseeable and some unpredictable and random in threshold selection procedure using unscented transformation
nature. One safety-threatening disruption in power systems (UT). While it can be applied to microgrids with different
is recognized as High Impedance Faults (HIF), the detection ratings and structures, the detection threshold must be tailor-
of which has long remained a challenging concern in the made and adjusted appropriately. Reference [11] presented a
electric industry. HIFs can cause “arcs” or “flashover” from variable-importance-based feature selection method to identify
the wires, through the air, to the neighboring trees, other HIFs from a large pool of signal signatures; this feature
vegetation or equipment, where it can cause fires, injuries, or selection scheme utilizes the discrete Fourier transform (DFT)
even fatalities [1]. A life-threatening example is the constant and Kalman Filter (KF) for harmonics coefficient estimation
exposure and contact of a power line with a tree branch and HIF duration and magnitude measurement.
during high-wind conditions, which can threaten homes in With the rapid advancements in Artificial Intelligence (AI),
residential neighborhoods and spark wildfires in rural areas. many researchers have implemented various HIF detection
Such types of faults are commonly caused by undesired techniques through a variety of machine learning technologies.
contacts with bare energized electrical conductive parts, the References [1], [12], [13] utilized waveform pattern analysis
high-impedance nature of which significantly restricts the flow and Support Vector Machine (SVM) to classify and ultimately
of fault current to a level hard to be detected by the overcurrent detect HIFs in power grids. In [4], semi-supervised learning
protective relays [2]. In particular, the existing commercial and probabilistic learning are used for HIF detection and
microprocessor-based protective relays activate a tripping de- localization, revealing promising detection accuracy, but with
cision when the electrical measurements are observed well a compromised response time of half a second. In [14], an
beyond the detection threshold; however, an unsatisfactory HIF detection approach using empirical mode decomposition
(EMD) combined with an artificial neural network (ANN)
This work was supported in part by the National Science Foundation (NSF) is proposed, where the HIF detection and classification are
under Grant ICER-2022505. achieved through predominant harmonic signatures caused by
S. Wang and P. Dehghanian are with the Department of Electrical and
Computer Engineering, The George Washington University, Washington, DC HIFs in the electrical signals. Discrete wavelet transform is
20052, USA (e-mails: shiyuan1225@gwu.edu; payman@gwu.edu). applied in [15] to monitor the high-frequency components
Digital Object Identifier 10.1109/TIA.2020.3017698
0093-9994 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
 2

and Long Short Term Memory (LSTM) to detect HIFs,

revealing detection accuracy of 90% under scenarios with
clean (not noisy) measurements. An HIF detection approach
using discrete wavelet transformations and Back Propagation
Artificial Neural Network (BP-ANN) is introduced in [16]
and tested using often-noisy real measurement data from a
substation, where the detection accuracy is reported 76%.
While the accurate and swift HIF detection is a yet to be
solved challenge in the electric industry, the state-of-the-art
literature has demonstrated that AI technologies offer a yet- (a) A single-phase diagram for HIF modeling
untapped potential in detecting HIFs and improving electrical
safety by enabling a timely warming notification to the system
operator and activating trip signals if needed.
Inspired by the promising feature extraction and pattern
recognition capabilities through wavelet transforms and the
wide adoption of machine learning analytics in electric power
grids [17]–[22], we propose an artificial intelligence solution (b) A two anti-diode HIF model
that unlocks the full potential in computing at the edge Fig. 1. A single-phase diagram for the HIF model in [4], [5]
for HIF detection and classification, thereby improving the
electrical safety. An improved HIF model is first proposed for
where, vn1 and in1 are the sending-terminal voltage and
generating a large pool of HIF waveforms. A modified Gabor
current; vf stands for the voltage at the location of the fault;
wavelet transform is next proposed for feature extraction and
and δ is the fault distance from the sending-terminal. However,
pattern recognition from the electrical waveforms during HIF
during an HIF event, vF and δ are the values to be estimated.
scenarios. An efficient Convolutional Neural Network (CNN)
The current flow can be found through
engine is eventually developed to process the extracted features
and make the detection and classification decisions. The pro- dvn1
in1 = ic + in2 + iF , ic = δ (2)
posed solution offers a promising accuracy and response time dt
in detecting and classifying low-intensity HIFs under noise where, ic , in2 , and iF are, respectively, the currents flowing
inferences. The suggested functionality can be installed within through the shunt capacitor, received at the end-terminal, and
the existing sensors, protective relays and other Intelligent observed at the HIF branch.
Electronic Devices (IEDs) with online monitoring capabilities, Under some HIF scenarios, arcs can be observed when
thereby rendering a cost-effective solution technology. the air gap between the power line conductor and the high
This paper is organized as follows: Section II introduces the impedance object is energized. Once the imposed voltage
background information on HIF modeling, wavelet transforms magnitude is higher than the voltage (breakdown voltage),
(WT), and the pattern classification and recognition through there would be arc ignitions across the air gap. On the other
Convolutional Neural Networks (CNNs). Section III presents hand, an arc extinguishes when the fault voltage is lower than
the improved HIF model and the proposed HIF detection and the breakdown voltage. Therefore, the HIF current changes
classification scheme, which is consisted of (i) power wave- during each cycle, making its magnitude follow a non-linear
form feature extraction through a pseudo-continuous quadra- characteristic [23]. Based on the above properties of an HIF
ture wavelet transform (PCQ-WT), and (ii) event detection and event, an HIF model is developed and has been widely used
classification via CNN. Case studies and numerical results are in [1], [4], [5], [9], [11], [14], as shown in Fig. 1(b). The fault
analyzed in Section IV. Section V eventually concludes the voltage in this model can be written in the following format:
paper and provides directions for future developments. ⎧
⎨ RF iF + LF diF + VF P , iF ≥ 0
vF = dt (3)
⎩ R i + L d iF − V , i < 0
II. BACKGROUND AND M OTIVATION F F F
dt FN F

A. HIF Modeling where, RF and LF are the HIF’s resistance and inductance
in series; VF P and VF N are the positive and negative arc
A single-line diagram of an HIF event in a radial distribution
voltages during HIF events, which archives the non-linearity
system is illustrated in Fig. 1(a). The sending node is modeled
of HIFs. The simulated voltage and current waveforms are
by an ideal AC source; Z1 and Z2 are known impedance values
demonstrated in Fig. 2. It can be found that although there is
and can be estimated according to the system topology and
almost no influence on the voltage waveform, the HIF event
operating conditions; R, L and C stand for the per unit length
slightly distorts the current signal while the current magnitude
resistance, inductance, and capacitance of the line. By applying
has remained around the rated range; in other words, the HIF
Kirchhoff’s voltage law to the dashed circle in Fig. 1(a), the
event cannot be detected easily by the conventional protective
following equation is derived:
relays. One should note that, at the beginning of an HIF event,
 
d(in1 − ic ) this HIF model can barely match the characteristics of the
vn1 = δ R(in1 − ic ) + L + vF (1) first period due to the build-up phenomenon [24]. However,
dt
 3

detect through current amplitudes, such events often cause

waveform abnormalities and distortions. One should note that
the orders, magnitudes, and phase angles corresponding to
the HIF-caused harmonics would be totally different under
different combinations of HIF parameters. Thus, detecting
HIF through analyzing certain orders of harmonics will be
extremely challenging; this is even further exacerbated by
(a) HIF Voltage
the existence of noise or other harmonics, altogether could
compromise the performance of the HIF detection schemes.
To deeply investigate the waveform features, we expand the
sampled current by the Fourier series as follows:
H

x(t) = A1 cos (ω1 t + φ1 ) + Ah cos(ωh t + φh )
h=2 (5)

Harmonic Components
(b) HIF Current where, h is the order of harmonics; H is the maximum order of
Fig. 2. Simulated voltage and current waveforms (with per unit values) harmonics limited by the sampling rate; Ah , ωh , and φh are the
captured at the sending-terminal based on [4]. h-th order harmonic component’s instantaneous magnitude,
frequency, and phase angle, respectively. Also, h = 1 stands
for the fundamental frequency component. One can see, in
following a few fundamental cycles, the HIF stabilizes and the
Fig. 2, that the fundamental magnitude is affected by both
fault parameters—RF , LF , VF P and VF N —are approximated
HIF and load variation events. Therefore, it is clear that the
as constant values. Therefore, the HIF current waveform and
second term in (5) carries valuable information in assessing
the related shoulder phenomenon can still be characterized
the HIF impacts, thus the main focus in the proposed data
using the model presented in Fig. 1(b).
mining and pattern extraction process.
With the Wide installation of Phasor Measurement Units
(PMUs) in the power grid, micro-PMUs (μ-PMUs) in power
distribution systems [25] and many other IEDs with PMU C. WT and Pseudo-Continuous WT
functionalities, HIF detection can be achieved through such de- As the combinations of harmonic components contain valu-
vices with high-precision and high-resolution measurements; able information on HIFs, it would be preferred to implement
that is, the case presented in Fig. 1(a) can be easily solved time-frequency analysis on the current waveforms. Short time
as long as both sending and receiving terminals are equipped Fourier transform (STFT) [28] or DFT through a sliding
with such devices that can ensure the availability of high- window [11] can be approached assisting the HIF detection
precision high-resolution electrical measurements. However, a process; this is because the STFT offers a high frequency
full observation is very hard to achieve in every segment of the resolution; however, this is achieved at the cost of a high
power network due to the cost limitations [26] and only the computational complexity [29], [30] and it is hard to provide
power waveforms from the upstream (sending-end) terminal accurate detection performance utilizing the spectrum alone.
are usually measured; meanwhile, the performance of the PMU On the other hand, wavelet transform (i) offers a promis-
surveillance and monitoring network can be compromised ing computing efficiency compared to the STFT in time-
by communication failures, latencies and cyber-attacks [27]. frequency analysis, and (ii) has been widely applied for feature
Therefore, it makes it very challenging for the HIF detection extractions in many real-world applications [17], [31], [32],
and classification scheme to operate as desired since the HIF including research on HIF detection [16], [33] and [34].
current is viewed incremental by the upstream terminal, not Figure 3 illustrates a performance comparison of the STFT
large enough to violate the tripping thresholds in the protective and WT. It can be observed, from the spectrum in Fig. 3(b)(f),
relays, and are often mistakenly corresponded to the common that the HIF current waveform is featured with more harmonic
load increments in the network. components at larger than 5th orders, where the performance
of the HIF detection method in [9] will be easily compromised
B. Power Waveform Modeling by such harmonic pollution or noise interferences. Meanwhile,
comparing the scalograms in Fig. 3(c)(d) with Fig. 3(g)(h),
With the current waveforms carrying valuable information
one can note that the results from STFT provide very detailed
on the underlying phenomenon, one can evaluate the current
frequency information, and the scalograms from HIF-affected
waveform in each phase to examine the existence of HIFs.
waveform reveal very obvious features compared to those from
First, the current waveform can be generalized by
the harmonic-polluted waveforms; nevertheless, the accuracy
x(t) = Acos (ωt + φ) (4) in frequency measurements is compromised in WT. Therefore,
a combination of WT and STFT is generally much desired.
where, A, ω, and φ are the magnitude, frequency, and phase However, this could highly increase the computational burden
angle in each phase. Although HIF phenomena are hard to of the detection mechanism, and in particular, challenge the
 4

(a) Harmonic Waveform (b) STFT (c) Morlet (d) Daubechies 20

(e) HIF Waveform (f) STFT (g) Morlet (h) Daubechies 20

Fig. 3. Comparison of the STFT vs. WT with: HIF-affected waveform and harmonic-injected waveform starting at t=20ms with harmonic orders h= 3 and
5 and magnitudes of 0.08pu and 0.08pu, respectively.

online applicability of the HIF detection schemes. In order to D. Convolutional Neural Networks
archive a low-computing complexity and yet accurate design, Convolutional Neural Networks (CNNs) are among deep
we select WT alone in this paper for power waveform feature machine learning techniques that have been proven very ef-
extraction and HIF detection. fective in processing image-related tasks [38]–[40]. Therefore,
The wavelet transform can be seen as the computation once the scalograms from WT are generated, CNNs can be
process of the similarities between the signal of interest x(t) used as the HIF event detection engine through classifying
and the selected wavelet. x(t) is here the measured single- the scalograms. The CNN training process mainly involves
phase waveform. The WT calculation is defined as follows: feature/representation learning for feature detection and clas-
sification. In general, the execution of CNN for HIF detection
∞
1 t−b and classification is achieved through a set of cross-correlation
X(ω|a, b) = √ x(t)Ψ* ( )dt (6)
a −∞ a assessments as follows:

sp (m, n) = I u (m + v, n + w)K p (v, w), (8)
where a, b are the scaling factor and time shift; Ψ(t) is the
u v w
selected wavelet (called mother wavelet when a = 1 and
p
b = 0), and “*” denotes the complex conjugate operator. where s (m, n) stands for the convolutional layer’s output at
With different values of a and b selected, Ψ( t−b position (m, n) and p-th channel; the u-th convolutional kernel
a ) becomes
the ”daughter wavelets” of Ψ(t) [35]–[37], and each value of is marked as K u ; and I u denotes the image/data volume in
a corresponds to a pseudo frequency. Typically, the WT uses the u-th channel. A complex convolutional layer is comprised
discrete scaling factors ak =2i , where i is integer. Thus, the of a set of simple layers [41], as expressed in the following:
extracted feature from each pseudo frequency is finite. To have
I l = pool (σ(s)) , (9)
more information extracted and waveform feature redundancy
in the scalogram, a set of linearly-increasing real numbers can where I l stands for the l-th layer’s output volume; σ(·) stands
be assigned to i, and the WT then becomes pseudo-continuous for the nonlinear operation of the active function; and pool(·)
(PCWT). Here, the PCWT is defined as follows: is a pooling (down-sample) operation in the pooling layer.
The abstraction ability of the network generally increases with
W −1
1  nTs − bk the number of stacked convolutional layers [42]. The final
X[ω|ak , bk ] = √ x[n]Ψ* [ ] (7) representations in a CNN’s last layer are usually reshaped to
a n=0 ak
vectors and fed into the fully-connected (FC) layers.

where Ts is the sampling time, and W is the number of data

sampled in the buffer. Choosing a proper type of wavelet and III. P ROPOSED F EATURE E XTRACTION AND HIF
adjusting the parameters correctly are very critical as they will D ETECTION BY CNN S
affect the covered range for the pseudo frequency and the As each single-phase current waveform captured from the
PCWT performance. Besides, the computing complexity can upstream terminal in a radial power distribution system carries
be reduced and the response time efficiency can be improved information of the downstream terminal, we mainly focus on
by using well-designed parameters since the online feature detecting low-intensity HIFs which is one most challenging
extraction process requires uncongested real-time data streams. task in electric industry. The workflow of the proposed HIF
 5

Fig. 4. Series of integrated functions in the proposed HIF detection system.

detection technology is demonstrated in Fig. 4; it functionally

consists of the following four modules:
1) Signal Acquisition: The proposed framework shares the
same input waveforms of a typical PMU (or protective
relay with PMU functionality) from the Analog to Dig-
ital (A2D) processing module. It avoids any additional
Fig. 5. The proposed improved HIF model.
A2D modules in the front-end, and makes the proposed
architecture a economically-viable sensor solution. The
waveform data is stored in the cache for the subsequent pattern recognition processes [43]–[45]. The modified Gabor
analytical processes. wavelet in this paper is expressed as follows:
2) Feature Extraction: This module applies pseudo-  
continuous quadrature wavelet transform (PCQ-WT) to (t − b)2
Ψ(t) = exp (jωc (t − b)) · exp − (10)
the cached waveform data and generates scalograms.  α02
Periodic Component 
The scalograms are matrices that contain signal signa- Gaussian Envelope
tures corresponding to the HIF events in the power grid.
The scalograms are then quantized to digital images in where wc is the central frequency which determines the pseudo
order to compress the data size. frequency of the feature extraction in the waveform. This can
3) Event Detection and Classification: The images ob- be proved by the Fourier transform of this Gabor wavelet as
tained in the previous stage are fed into a CNN that, √ α2
FΨ (ω) = α0 π · exp(−jωb) · exp(− 0 (ω − ωc )2 ) (11)
with a detection confidence, classifies whether there is 4
an HIF event. Finally, the detected event will be reported when ω = ωc , the magnitude of FΨ (ω) is at its maximum;
to the local protection device or the control center. in other words, the component with frequency ωc in the
As one can see, the proposed HIF detection system only sampled waveform will have the maximum correlation with
requires software-level modifications to the existing sensors this wavelet. Thus, the feature is extracted and highlighted
(e.g., PMUs, protective relays, etc.), with no additional hard- in the scalogram. Another advantage of this modified Gabor
ware investments. Here, we utilize a physics-guided machine wavelet is that it has a predictable narrow-bandwidth (11)
learning technique, as the overall detection system only re- due to the exponential operation, which makes it easier to
quires offline training. The training process is under the determine its pseudo frequency. By properly selecting α0 in
guidance of the pre-recorded and/or simulated HIF waveforms PCQ-WT, a desired bandwidth of the pseudo frequencies is
as the training dataset. covered at ωc , while the time shift b plays no magnitude impact
on (11). By substituting t − b with t−b a , the Gabor wavelet
transform of each harmonic component in (5) turns into
A. Improved HIF Model
∞
Ah j(ωh t+φh ) −j(ωh t+φh )  * t−b
The waveform assessment based on the HIF model pre- X(ωh |a, b)= e +e Ψ ( a )dt
sented in Fig. 1(b) is adequate to some extent; however, to −∞ 2
enrich the proposed solution’s knowledge on a variety of HIF (12)
waveforms, we propose a comprehensive HIF model shown in According to the Hubbard–Stratonovich transformation [46]:
Fig. 5, where the fault resistance and inductance are assigned  ∞ 
to the positive and negative branch. For simplicity, we name α 2 1 y2
exp(− x ) = exp − 2α − jxy dy (13)
ZF P and ZF N as positive and negative arc impedances, 2 2πα −∞
respectively. One advantage of this proposed model is its The WT of Gabor wavelet in (12) becomes
flexibility to approximate different HIF conditions including
Ah jωh (φh +b) √ α2
0 2
those studied in [1], [4], [5], [9], [11], [14]. X(ωh |a, b) = e · aα0 πe− 4 (aωh −ωc )
2
Ah −jωh (φh +b) √ α20 2
+ e · aα0 πe− 4 (aωh +ωc ) .
B. Proposed PCQ-WT Based Feature Extraction 2
In order to achieve a low computing complexity solution ≈0
(14)
with high-fidelity feature extraction performance, a modified
complex Gabor wavelet from [43] is adopted in this paper. It can be seen that the second term on the right side of (14) can
Gabor wavelets have been widely used in two-dimension (2-D) be neglected as the exponent in the last exponential operation
 6

is a large negative. When ω0 = ωc /a, X(ωh |a, b) reaches its TABLE I

maximum value, indicating that the dominant feature of the PARAMETER S PECIFICATIONS FOR G ENERATING THE T EST WAVEFORMS
selected frequency is extracted. Here, we make
System Setting
aα0 = ωc /γ, (15) Vbase * 13.8kV Sbase 0.5 MW

where γ is a constant. Accordingly, and based on (10), the Fs 7.68kHz Line Length 2km
length of the Gaussian window in Gabor wavelet also adapts SLoad 0.5 - 1.5 pu pfload 0.8 - 1
different frequencies. HIF Model Setting
The proposed Gabor wavelet in the discrete time domain is
Parameter Range Parameter Range
   
ωc Ts (n − bk ) Ts2 (n − bk )2 ZFP 0.02 pu - 1.5 pu ZFN 0.02 pu - 1.5 pu
Ψ[n|ak , bk ] = exp j exp − ,
ak a2k α02 pfFP 0-1 pfFN 0-1
(16)
VFP 0.08 pu - 0.65 pu VFN 0.08 pu - 0.65 pu
Applying different discrete scaling factors ak and time shift
Fault Location 1% - 99%
bk , we achieve the proposed PCQ-WT as follows,
* :A signal to noise ratio of 40dB is added to the AC voltage source.
W
 −1
Ts (n − bk )
X(ωk |ak , bk ) = x[n]Ψ* [ ]
n=0
ak
W −1   feature extraction will not be compromised. The scalograms
 ωc T 2 (n − bk )2 fed into the CNN are cropped from 10ms to 50ms (of the total
= x[n]exp −j Ts (n − bk ) − s 2 2 .
n=0
ak ak α 0 60ms WT output)—i.e., observation+ wavelet length— which
(17) has 40ms (308 samples) duration.
The proposed compact CNN configuration for scalo-
If the pseudo frequencies of interest and the Gabor wavelet
gram classification is as follows: Input layer (256×308);
bank are designed properly, a set of PCQ-WTs in form of a
Convolution (Conv.) layer (32×5×11); Max-pooling layer
vector Xω can be generated conveying waveform features in a
(3×3); Conv. layer(32×5×5); Max-pooling layer(3×3);
certain frequency range. During both transient and steady state
Conv. layer(32×5×5); Full-connected (FC) layer(200×1); FC
operating modes, WT time-frequency analysis is conducted
layer(3×1). Conventional images have homogeneous units on
along time, and a scalogram stream can be then achieved.
the horizontal and vertical axes, while the scalograms axes
carry different information regarding the HIF events on either
C. Proposed CNN Configuration for HIF Detection time or frequency. Therefore, a wide kernel in the first Conv.
With the PCQ-WT extracted features in form of scalo- layer that can extract more information from the transitions
grams available, the HIF detection problem is converted to along the time axis is used. The stride of the first layer
a supervised scalograms classification problem. However, the is (2,3), and the remaining Conv. layers use strides with
classification of the high-dimensional 2-D scalograms is chal- a size of (1,1). Besides, batch normalization [47] is used
lenging. specifically, every frame of the obtained scalogram in the last FC layer. In the last Conv. layer and the first
has hundreds by hundreds (scales × time) pixels; it is very FC layer, Dropout [48] is used to prevent over-fitting. All
challenging to process such high dimensional data through activation functions in the CNN are Rectified Linear Unit
the conventional pattern classification approaches. Here, we (ReLU). We choose cross-entropy as the loss function. One
convert the PCQ-WT scalograms into 2-D images and propose can see that our proposed CNN is not that ”deep” compared to
a compact CNN architecture to classify the HIFs concealed in the regular image classification CNNs; the suggested compact
the scalograms by the PCQ-WT. The proposed CNN has a CNN architecture further reduces the computing complexity
simple architecture for HIF detection, yet achieving a very in HIF detection.
fast processing time for online applications.
IV. C ASE S TUDY AND N UMERICAL E XPERIMENTS
D. PCWT and CNN Parameter Setting
A. Test Scenarios Configuration
The sampling rate F s for the signal and the feature ex-
traction is 7680Hz which provides 128 samples per nominal The parameter specifications used for generating the test
fundamental cycle. The observation window for the PCQ- waveforms are listed in Table. I. We, in particular, focus on
WT is set to 308 samples (40ms). The time shift b for the HIF event detection since conventional faults can typically be
modified Gabor daughter wavelets is 10ms (77 samples) for detected by existing protective devices. Three test scenarios
simplicity. The scaling factor a for the proposed PCQ-WT (HIF event, load change event, and normal operation event)
is chosen as 2i , where i is sampled uniformly from 0 to 8. are simulated. For each HIF simulation, all parameters are
The central frequency ωc = 1152 and the pseudo frequency randomly selected in the designated ranges. In each simulated
will roughly reach up to the 19th order of harmonics. All waveform, only one event occurs at a random point in time.
the Gabor wavelets have 20ms duration. By this design, the The waveform generation system is developed according to
computational complexity in computing the scalograms is Fig. 1(a) and the improved HIF model in Fig. 5. Gaussian
reduced, while the pseudo frequency bandwidth coverage for noises with SNR of 40dB are added to the AC voltage
 7

(a) HIF with SNR 30dB (b) Load Increase with SNR 30dB
Fig. 6. Test waveform simulation results: (a) HIF event; (b) Load increase event; both simulated events start at t=20ms.

waveform. When evaluating high frequency areas in both

scalograms, one can see that the higher frequency range (scale
from 0 to 96) has less discontinuous patterns compared to the
lower frequency range (scale from 96 to 192). However, these
frequency portions (96 to 192) are discontinuous with ripple
shape carrying significant differences in Fig. 6(a) and Fig. 6(b).
For the HIF scalogram, the ripple pattern is formed at
t = 35ms and becomes stable afterward; thus, one can see
that it will take less than 20ms (PCQ-WT window length) to
reveal the corresponding HIF patterns—some pyramid shape
ripples appear in a group. When there is a load increase
event, the load change feature takes 10ms to emerge (see
the input waveform in Fig. 6(b)) and the impact remains
Fig. 7. Test results of the proposed CNN framework for online HIF detection consistently present. However, for scale from 96 to 192, there
and classification
is no pyramid shape ripple found as those in Fig. 6(a). Also,
only the intensity of the existing ripples slightly increases. One
source to approximate the thermal and measurement noises should notice that the patterns extracted from this simulated
in different conditions. HIF event approximately covers the scale from 100 to 200,
A total of 20,000 samples from the test waveforms are which roughly corresponds to the 12th to 5th order harmonics
simulated for each event; therefore, a total of 60,000 samples in the frequency span.
(wavelet scalogram) are simulated in the MATLAB/Simulink 2) Event Detection: The HIF detection test results are
environment. A total of 48,000 samples are randomly selected summarized in the confusion matrix in Fig. 7, where the true
as the training dataset, 6,000 samples for validation, and 6,000 label stands for the actual tested events, and the predicted
samples for testing datasets. For training the neural network label corresponds to the classification outcomes of the CNN
and increasing the versatility of the CNN, the generated module. The average accuracy of the proposed HIF event
waveforms are manually imposed by white Gaussian noise detection scheme is found 99.95 %. To further examine the
with 30dB SNR. We use Adam [49] as the optimizer, which online event detection and classification performance, we used
has the initial learning rate of 1 × 10−4 , and weight decay of a workstation with a stock eight-core AMD Ryzen 3800X
1×10− 5. We trained the proposed CNN 120 epochs. The best CPU as the computational platform. We transfer the PCQ-
validated model was recorded and tested. WT and CNN modules into MATLAB 2020a to record the
computational time on one single core of the CPU. The overall
time for processing the PCQ-WT and CNN is recorded as
B. Experimental Results and Analysis 6.3±0.6ms, confirming a promising solution to be used in real-
1) Feature Extraction: The extracted features from an HIF time HIF detection applications.
and a load change events are demonstrated in Fig. 6, where the We also test the proposed framework on a recorded wave-
main energy concentration with high intensity is marked by red form of 0.45s duration, the result of which is shown in Fig.
standing for the fundamental frequency component extracted 8. The top heat-map is the event detection results over time,
from the waveforms. In Fig. 6(a), the HIF at t = 20ms where the classification confidence rate is marked with the
increases the main energy concentration at t = 30ms, which color bar. One should note that, during the Normal operating
matches the corresponding magnitude increase in Fig. 3(e). event, the confidence rate is very high and the classification
This phenomenon can also be observed in Fig. 6(b), as a load result is accurate even though the waveform is polluted with
increase will indeed increase the magnitude of the current 30dB Gaussian noise and the distortion is very obvious.
 8

Fig. 8. Online HIF detection on a simulated single-phase current waveform: detected result (top) and original waveform (bottom).

Relatively lower confidence rates always exist during the R EFERENCES

transitions between two different events.
The Load Change event is detected accurately within 10ms [1] M. Sarwar, F. Mehmood, M. Abid, A. Q. Khan, S. T. Gul, and
and it takes another 40ms for the detection scheme to report A. S. Khan, “High impedance fault detection and isolation in power
distribution networks using support vector machines,” Journal of King
a normal condition event. Meanwhile, at the moment of HIF Saud University-Engineering Sciences, 2019.
occurrence, the CNN module classified the first two cycles [2] C. G. Wester, “High impedance fault detection on distribution systems,”
of HIF-contained waveform (from t = 325ms to t = 355ms) in 1998 Rural Electric Power Conference Presented at 42nd Annual
Conference, 1998, pp. c5–1.
mistakenly as the load change event. The reason lies in the [3] J. Tengdin, R. Westfall et al., “High impedance fault detection technol-
fact that a load change event (Fig. 6(b)) reveals similar patterns ogy report of psrc working group d15,” 1994.
at the very beginning when compared to the HIF event (Fig. [4] Q. Cui and Y. Weng, “Enhance high impedance fault detection and
location accuracy via mu-PMUs,” IEEE Transactions on Smart Grid,
6(a)); therefore, it is very hard to make a correct detection and vol. 11, no. 1, pp. 797–809, 2019.
classification decision at the exact time when the HIF event [5] L. U. Iurinic, A. R. Herrera-Orozco, R. G. Ferraz, and A. S. Bretas,
happens. However, when the waveform magnitude stabilizes, “Distribution systems high-impedance fault location: A parameter esti-
the CNN module shows a high confidence rate in classifying mation approach,” IEEE Transactions on Power Delivery, vol. 31, no. 4,
pp. 1806–1814, 2015.
the HIF event correctly. Also, one can see that the maximum [6] S. Wang, “On the use of artificial intelligence for electrical safety,”
peak value of the current waveform affected by an HIF event Poster Presentation at 2020 Elec. Safety Workshop, Reno, NV USA,
is observed 1.25 pu which will barely trip the protective relays March 2-6, 2020.
[7] S. Maximov, V. Torres, H. Ruiz, and J. Guardado, “Analytical model
in the absence of the proposed HIF detection scheme. for high impedance fault analysis in transmission lines,” Mathematical
In summary, the proposed solution provides satisfactory Problems in Engineering, vol. 2014, 2014.
results in detecting low-intensity HIFs under noisy measure- [8] D. P. S. Gomes, C. Ozansoy, and A. Ulhaq, “High-sensitivity vegeta-
tion high-impedance fault detection based on signal’s high-frequency
ments and can distinguish it with load change events. More- contents,” IEEE Transactions on Power Delivery, vol. 33, no. 3, pp.
over, the overall detection accuracy is desirable (99.98%) at a 1398–1407, 2018.
promising speed of within 40ms (33ms delay plus 6.3±0.6ms [9] A. P. Kujur and T. Biswal, “Detection of high impedance fault in
processing time). distribution system considering distributed generation,” in 2017 Interna-
tional Conference on Innovative Mechanisms for Industry Applications
(ICIMIA), 2017, pp. 406–410.
V. C ONCLUSION [10] O. Gashteroodkhani, M. Majidi, M. Fadali, M. Etezadi-Amoli, and
E. Maali-Amiri, “A protection scheme for microgrids using time-time
This paper proposed an effective scheme that leverages ar- matrix z-score vector,” International Journal of Electrical Power &
tificial intelligence advancements for detecting HIFs in power Energy Systems, vol. 110, pp. 400 – 410, 2019. [Online]. Available:
grids and improving electrical safety. The proposed solution http://www.sciencedirect.com/science/article/pii/S0142061518322270
functionally integrates a PCQ-WT feature extraction tool using [11] Q. Cui, K. El-Arroudi, and Y. Weng, “A feature selection method for
high impedance fault detection,” IEEE Transactions on Power Delivery,
a modified Gabor wavelet and a compact CNN-based event vol. 34, no. 3, pp. 1203–1215, 2019.
detection technique. Experiment results demonstrated that the [12] S. Bhongade and S. Golhani, “HIF detection using wavelet transform,
proposed analytics successfully achieved an ultra-fast (within travelling wave and support vector machine,” in 2016 Int. Conf. on
Electrical Power and Energy Systems (ICEPES), 2016, pp. 151–156.
40ms) and accurate (99.95% accuracy) HIF detection perfor- [13] K. Moloi, J. A. Jordaan, and Y. Hamam, “High impedance fault detection
mance even under noisy measurements. technique based on discrete wavelet transform and support vector
The proposed function could be embedded within the exist- machine in power distribution networks,” in 2017 IEEE AFRICON,
2017, pp. 9–14.
ing PMUs and/or other IEDs that are capable of capturing and [14] H. Lala and S. Karmakar, “Detection and experimental validation of
processing the power waveforms. Future work can be focused high impedance arc fault in distribution system using empirical mode
on (i) enriching the proposed approach with a variety of HIF decomposition,” IEEE Systems Journal, pp. 1–12, 2020.
models and a further strong pool of HIF waveforms, (ii) inves- [15] K. V. Shihabudheen, B. Kunju, I. Ahammed, A. Guruvarurappan, J. Jose,
D. Keerthana, and P. B. Revathi, “Detection of high impedance fault
tigating HIF intensity measurements through CNNs or other using machine learning techniques,” in TENCON 2019 - 2019 IEEE
machine learning algorithms, and (iii) hardware-in-the-loop Region 10 Conference (TENCON), 2019, pp. 2117–2122.
performance analysis when the proposed HIF detection and [16] Ming-Ta Yang, Jhy-Cherng Gu, Jin-Lung Guan, and Chau-Yuan Cheng,
“Detection of high impedance faults in distribution system,” in 2005
localization function is embedded into the PMUs, protective IEEE/PES Transmission Distribution Conference Exposition: Asia and
relays, and/or other IEDs. Pacific, 2005, pp. 1–5.
 9

[17] S. Wang, P. Dehghanian, L. Li, and B. Wang, “A machine learning [38] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
approach to detection of geomagnetically induced currents in power with deep convolutional neural networks,” in Advances in neural infor-
grids,” IEEE Transactions on Industry Applications, vol. 56, no. 2, pp. mation processing systems, 2012, pp. 1097–1105.
1098–1106, 2019. [39] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
[18] S. Wang, P. Dehghanian, and L. Li, “Power grid online surveillance large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
through PMU-embedded convolutional neural networks,” IEEE Trans- [40] L. Li, M. Doroslovački, and M. H. Loew, “Discriminant analysis deep
actions on Industry Applications, vol. 56, no. 2, pp. 1146–1155, 2019. neural networks,” in 53rd Annual Conference on Information Sciences
[19] M. H. Rezaeian Koochi, P. Dehghanian, S. Esmaeili, P. Dehghanian, and and Systems, March 2019, pp. 1–6.
S. Wang, “A synchrophasor-based decision tree approach for identifica- [41] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press,
tion of most coherent generating units,” in Annual Conference of the 2016, http://www.deeplearningbook.org.
IEEE Industrial Electronics Society, Oct 2018, pp. 71–76. [42] Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A
[20] S. Wang, P. Dehghanian, L. Li, and B. Wang, “A machine learning review and new perspectives,” IEEE transactions on pattern analysis
approach to detection of geomagnetically induced currents in power and machine intelligence, vol. 35, no. 8, pp. 1798–1828, 2013.
grids,” in IEEE Industry Applications Society Annual Meeting (IAS), [43] J. V. B. Soares, J. J. G. Leandro, R. M. Cesar, H. F. Jelinek, and
2019, pp. 1–7. M. J. Cree, “Retinal vessel segmentation using the 2-D Gabor wavelet
[21] M. Babakmehr, F. Harirchi, M. Nazir, S. Wang, P. Dehghanian, and and supervised classification,” IEEE Transactions on Medical Imaging,
J. Enslin, “Sparse representation-based classification of geomagnetically vol. 25, no. 9, pp. 1214–1222, Sept 2006.
induced currents,” in Clemson University Power Systems Conference, [44] Z. Zhang, M. Lyons, M. Schuster, and S. Akamatsu, “Comparison
March. 2020, pp. 1–7. between geometry-based and Gabor-wavelets-based facial expression
recognition using multi-layer perceptron,” in Proceedings Third IEEE
[22] B. Shinde, S. Wang, P. Dehghanian, and M. Babakmehr, “Real-time
International Conference on Automatic Face and Gesture Recognition,
detection of critical generators in power systems: A deep learning hcp
April 1998, pp. 454–459.
approach,” in 2020 IEEE Texas Power and Energy Conference (TPEC).
[45] B. S. Manjunath and W. Y. Ma, “Texture features for browsing and
IEEE, 2020, pp. 1–6.
retrieval of image data,” IEEE Transactions on Pattern Analysis and
[23] A. Ghaderi, H. L. Ginn III, and H. A. Mohammadpour, “High impedance Machine Intelligence, vol. 18, no. 8, pp. 837–842, Aug 1996.
fault detection: A review,” Electric Power Systems Research, vol. 143, [46] J. Hubbard, “Calculation of partition functions,” Physical Review Letters,
pp. 376–388, 2017. vol. 3, no. 2, p. 77, 1959.
[24] W. C. dos Santos, B. A. de Souza, N. S. D. Brito, F. B. Costa, and [47] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
M. R. C. Paes, “High impedance faults: From field tests to modeling,” network training by reducing internal covariate shift,” arXiv preprint
Journal of Control, Automation and Electrical Systems, vol. 24, no. 6, arXiv:1502.03167, 2015.
pp. 885–896, 2013. [48] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut-
[25] A. von Meier, E. Stewart, A. McEachern, M. Andersen, and dinov, “Dropout: a simple way to prevent neural networks from over-
L. Mehrmanesh, “Precision micro-synchrophasors for distribution sys- fitting,” The Journal of Machine Learning Research, vol. 15, no. 1, pp.
tems: A summary of applications,” IEEE Transactions on Smart Grid, 1929–1958, 2014.
vol. 8, no. 6, pp. 2926–2936, 2017. [49] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
[26] R. F. Nuqui and A. G. Phadke, “Phasor measurement unit placement arXiv preprint arXiv:1412.6980, 2014.
techniques for complete and incomplete observability,” IEEE Transac-
tions on Power Delivery, vol. 20, no. 4, pp. 2381–2388, 2005.
[27] T. Nguyen, S. Wang, M. Alhazmi, M. Nazemi, A. Estebsari, and
P. Dehghanian, “Electric power grid resilience to cyber adversaries: State
of the art,” IEEE Access, vol. 8, pp. 87 592–87 608, 2020.
[28] E. Mangueira Lima, C. M. dos Santos Junqueira, N. Silva Dantas
Brito, B. A. de Souza, R. de Almeida Coelho, and H. Gayoso Meira Shiyuan Wang (S’18) received the B.Eng. degree in
Suassuna de Medeiros, “High impedance fault detection method based mechanical engineering from University of Science
on the short-time fourier transform,” IET Generation, Transmission and Technology Beijing, China, in 2012; the M.Sc.
Distribution, vol. 12, no. 11, pp. 2577–2584, 2018. and Ph.D. degrees in electrical engineering from The
[29] Y.-C. Su, K.-L. Lian, and H.-H. Chang, “Feature selection of non- George Washington University, Washington, DC,
intrusive load monitoring system using stft and wavelet transform,” in e- USA, in 2014 and 2020, respectively. His research
Business Engineering (ICEBE), 2011 IEEE 8th International Conference interests include power system reliability and re-
on. IEEE, 2011, pp. 293–298. siliency, smart grid and renewable energy, power
grid harmonic analysis, and application of signal
[30] S.-H. Ni, K.-F. Lo, L. Lehmann, and Y.-H. Huang, “Time–frequency
processing in energy analytics.
analyses of pile-integrity testing using wavelet transform,” Computers
and Geotechnics, vol. 35, no. 4, pp. 600–607, 2008.
[31] S. Wang, P. Dehghanian, and B. Zhang, “A data-driven algorithm for
online power grid topology change identification with PMUs,” in 2019
IEEE Power & Energy Society General Meeting. IEEE, 2019, pp. 1–5.
[32] S. Wang, P. Dehghanian, and Y. Gu, “A novel multi-resolution wavelet
transform for online power grid waveform classification,” in The 1st
IEEE International Conference on Smart Grid Synchronized Measure- Payman Dehghanian (S11, M17, SM20) received
ments and Analytics (SGSMA), May 2019, pp. 1–8. the B.Sc. degree in electrical engineering from the
[33] J. C. Chen, B. T. Phung, H. W. Wu, D. M. Zhang, and T. Blackburn, University of Tehran, Tehran, Iran, in 2009, the
“Detection of high impedance faults using wavelet transform,” in 2014 M.Sc. degree in electrical engineering from the
Australasian Universities Power Engineering Conference, 2014, pp. 1–6. Sharif University of Technology, Tehran, Iran, in
[34] Shyh-Jier Huang and Cheng-Tao Hsieh, “High-impedance fault detection 2011, and the Ph.D. degree in electrical engineering
utilizing a morlet wavelet transform approach,” IEEE Transactions on from Texas A&M University, Texas, USA, in 2017.
Power Delivery, vol. 14, no. 4, pp. 1401–1410, 1999. He is currently an Assistant Professor with the
[35] D. P. Mishra, S. R. Samantaray, and G. Joos, “A combined wavelet and Department of Electrical and Computer Engineer-
data-mining based intelligent protection scheme for microgrid,” IEEE ing, The George Washington University, Washing-
Transactions on Smart Grid, vol. 7, no. 5, pp. 2295–2304, Sept 2016. ton, DC, USA. His research interests include power
system protection and control, power system reliability and resiliency, asset
[36] K. Thirumala, M. S. Prasad, T. Jain, and A. C. Umarikar, “Tunable-q
management, and smart electricity grid applications.
wavelet transform and dual multiclass svm for online automatic detection
Dr. Dehghanian was a recipient of the 2013 IEEE Iran Section Best M.Sc.
of power quality disturbances,” IEEE Transactions on Smart Grid, vol. 9,
Thesis Award in Electrical Engineering, the 2014 and 2015 IEEE Region
no. 4, pp. 3018–3028, July 2018.
5 Outstanding Professional Achievement Awards, and the 2015 IEEE-HKN
[37] S. Wang, L. Li, and P. Dehghanian, “Power grid online surveillance Outstanding Young Professional Award.
through PMU-embedded convolutional neural networks,” in IEEE In-
dustry Applications Society (IAS) Annual Meeting, 2019, pp. 1–7.