Electrical Power and Energy Systems 147 (2023) 108909

Contents lists available at ScienceDirect

International Journal of Electrical Power and Energy Systems

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

Faulty feeder detection for single phase-to-ground faults in distribution

networks based on patch-to-patch CNN and feeder-to-feeder LSTM
Jiawei Yuan , Zaibin Jiao *
School of Electrical Engineering, Xi’an Jiaotong University, Xianning West Road No. 28, Xi’an 710000, China

A R T I C L E I N F O A B S T R A C T

Keywords: Faulty feeder detection has always been recognized as an essential issue in guaranteeing the stable and reliable
Faulty feeder detection operation of distribution networks. Existing methods fail to fully exploit the spatial correlation and temporal
Feeder to feeder dependencies in the zero-sequence currents, which are exhibited as waveforms on the spatial domain and are
Hybrid CNN-LSTM model
displayed as time-series data on the time domain. In this paper, a novel detection method based on a hybrid
Patch to patch
Spatial–temporal features
model of convolutional neural network (CNN) and long short-term memory (LSTM) neural network is proposed,
and the spatial–temporal information can be comprehensively explored. Deep spatial features are extracted by
using CNN, and they are further processed by using LSTM network. Besides, to maintain the temporal de­
pendencies of the extracted spatial features, the current waveforms are recognized on the patch scale by the
proposed patch-to-patch CNN (PToP CNN). Moreover, a feeder-to-feeder LSTM (FToF LSTM) network is estab­
lished to learn and compare spatial–temporal correlations between feeders. The joint PToP-CNN and FToF-LSTM
can achieve collaborative superiority on mining the fault features in the zero-sequence currents. To verify the
prospects in real installations, the hybrid model is implemented in an embedded device, called NVIDIA Jetson
AGX Xavier, and the real-time detection demonstrates the efficiency of the proposed method.

1. Introduction polarity, are extracted and combined by using the multi-criteria

comprehensive voting method [7]. A multiple evidence estimation
Stable and reliable operation of distribution networks is essential to method is applied in [8], and the detection criterion is constructed based
ensure the continuity and efficiency of power supply. The routine on the combination of transient features of energy, mutation degree, and
operation of the networks is prone to various faults, where the single- polarity. To merge the transient energy and cosine similarity, the Lap­
phase-to-ground (SPG) faults occupy more than 70 % [1]. Serious fail­ lace distribution is used in [9], and the feeder with the maximum
ure and wide-spread outage would occur if the SPG faults cannot be calculated cumulative density function is selected as the faulty feeder.
properly and timely isolated [2]. To avoid the substantial financial loss However, there are two significant drawbacks of these methods. 1) The
and possible human casualties, the faulty feeder with SPG faults needs to selection of fault characteristics heavily depends on experience, and
be detected as soon as possible. they are susceptible to complex fault scenarios and strong background
Currently, the existing faulty-feeder detection methods mainly noise, resulting in poor adaptability and robustness [10]. 2) The
include the characteristic-analysis methods and deep-learning methods. extracted characteristics only reflect the fault features of raw signals in
The characteristic-analysis methods aim to extract the effective fault temporal domain, and they fail to fully reveal the features in spatial
characteristics to describe the fault degrees of feeders, such as instan­ domain, where the waveforms in spatial domain have complete features.
taneous power [3], correlation polarities [4], residual admittances [5], The performance of the multi-feature fusion strategies may get
and statistical features of data distribution [6]. To remedy the short­ degraded, especially when most of the extracted features fail [11].
comings of a single fault characteristic, most methods mainly seek to use The deep-learning methods expect to improve the detection perfor­
the multi-feature fusion strategies to combine several extracted char­ mance by mining and fusing the fault features using intelligent algo­
acteristics. For instance, the instantaneous characteristics of zero- rithms, such as decision trees [12], long short-term memory (LSTM)
sequence currents, such as phase, energy relative entropy, and neural network [13], ensemble learning algorithms [14], and support

* Corresponding author.
E-mail address: [email protected] (Z. Jiao).

https://doi.org/10.1016/j.ijepes.2022.108909
Received 16 August 2022; Received in revised form 3 November 2022; Accepted 15 December 2022
Available online 23 December 2022
0142-0615/© 2022 Elsevier Ltd. All rights reserved.
J. Yuan and Z. Jiao International Journal of Electrical Power and Energy Systems 147 (2023) 108909

vector machine (SVM) [15]. In [16], a fuzzy measure fusion criterion is 2. CNN-LSTM hybrid classification model
utilized to calculate the similarity between the fault data in the real-time
detection and historical sample set, and the feeder can be classified as 2.1. Framework overview
faulty feeder and healthy feeder accordingly. Because the waveforms
have complete features, the time-series data can be converted into Since CNNs are expert in mining the hidden information in images
waveform images and recognized by using convolutional neural and LSTMs are good at processing the time-series data [20], we propose
network (CNN). In [17], CNN is directly used to adaptively extract and a faulty-feeder detection method based on the hybrid CNN-LSTM model.
combine the fault features in zero-sequence currents, and the faulty Moreover, it is significant for the faulty-feeder detection to capture the
feeder can be identified based on its output results. In [18], the sampled spatial–temporal characteristics and conduct correlation comparisons
data of zero-sequence currents are added with each other to achieve the between different feeders [11]. Thus, we propose a PToP CNN to extract
fused waveforms, and CNN is applied to find the fused waveforms that the spatial features from zero-sequence currents, where the temporal
do not contain the faulty feeder. Although the zero-sequence currents dependencies of the extracted features at time scales are maintained,
can be recognized on the waveform scale, the CNN fails to capture the and an FToF LSTM network to learn and compare the spatial–temporal
temporal features in the currents, which results in insensitivity to features between different feeders. The framework and workflow of the
changes in the temporal domain. proposed faulty-feeder detection method based on the PToP-CNN and
In fact, raw fault signals are displayed in the numerical form on the FToF-LSTM is presented in Fig. 1.
temporal domain and exhibited in waveforms on the spatial domain, Firstly, the initial data of zero-sequence currents in the numerical
which implies that they have complete spatial–temporal features. That form are obtained and preprocessed to generate the waveform images.
is, raw signals need to be fully exploited on the comprehensive aspect, Secondly, each image is split into patches and is recognized on the patch
including the spatial domain and temporal domain, thus improving the scale by using CNN, where the learned high-level features based on these
adaptability to variable fault conditions. Because the CNN is skilled at patches can ensure the temporal invariance. Subsequently, the extracted
extracting spatial features but poor in capturing temporal dependencies, features of each patch are concatenated, called PToP features, and the
while the LSTM network shows the opposite features [19], the spa­ PToP features of each feeder can be obtained. Finally, the established
tial–temporal features in fault signals can be obtained when the CNN FToF LSTM is used to compare the extracted PToP features between
and LSTM are combined in the learning process. different feeders, and the faulty feeder can be detected based on the
To improve the detection performance, a hybrid model of patch-to- output result of the LSTM. For the practical application, the proposed
patch (PToP) CNN and feeder-to-feeder (FToF) LSTM, namely PToP- CNN-LSTM model can identify the faulty feeder in an accurate and
CNN-FToF-LSTM, is proposed in this paper, and the joint learning of prompt manner, thereby facilitating further fast postfault actions.
spatial–temporal correlations can be conducted. The main contributions
can be summarized as follows. 2.2. Image creation

(1) The spatial features in the zero-sequence currents can be When an SPG fault occurs [21], the first half-cycle zero-sequence
extracted by CNN, and the temporal features can be captured by currents of feeders are sampled, and they are processed by using (1).
LSTM. By the cooperation of CNN and LSTM, the spa­
tial–temporal features in the currents can be fully exploited inor
i0k (d)
, k ∈ [1, N] (1)
0k (d) =
without information loss, and the superiority on feature extrac­ maxDd=1 {|i01 (d)|, ⋯, |i0N (d)| }
tion guarantees the good detection performance under variable
fault conditions. where N is the number of feeders, and D is the total sampling points in
(2) The hybrid CNN-LSTM model can not only extract spa­ the first half cycle. i0k is the sampled current of the k-th feeder, and inor
0k is
tial–temporal features in the currents, but also compare the the normalized current.
spatial–temporal correlations between feeders. The final detec­ After data processing, the zero-sequence current of each feeder is
tion results depend on the correlation comparison between normalized into [-1, 1]. To generate the images converted from the
feeders, rather than the fault features of a single feeder. In fact, numerical data quickly, the binary images can be obtained by using the
the correlation comparison enables the proposed method to be waveform encoding method [22], where the value is 1 if the waveform
highly robust to the complex fault scenarios and large localized cross the corresponding section. Finally, the normalized data of each
distortions in signals. feeder is converted into a binary image with size of 128 × 256.
(3) The PToP CNN and FToF LSTM are trained jointly, which implies
that both the spatial and temporal features can be learned 2.3. Patch-to-patch CNN
simultaneously. Furthermore, to verify the real-time perfor­
mance, the established PToP-CNN-FToF-LSTM is implemented in With the booming development of CNNs in computer vision, CNNs
the embedded device NVIDIA Jetson AGX Xavier. The experi­ have been widely used in feature extraction, pattern recognition, image,
mental results indicate that the proposed method has good and video recognition. CNNs mainly consist of the convolution layer,
detection performance under various fault scenarios, thus pooling layer, and fully-connected (FC) layer, where the convolution
demonstrating its considerable prospects in practical distribution layer is the core of CNNs. The convolution layer possesses of a lot of
networks. kernels with certain receptive fields. As the number of kernels increases
and convolution layers are stacked, CNNs can extract and learn spatial
The rest of this article is organized as follows. Section 2 describes the information from the input images.
proposed hybrid CNN-LSTM model. Section 3 presents the conducted CNNs are commonly assumed to be invariant to small image trans­
verification cases and discussions. Section 4 compares the performance formations [23], such as image translations, scaling, and deformations.
of the proposed method against other existing methods. Section 5 con­ That is to say, the recognition results are unchanged when the characters
cludes this article. are slightly translated, rescaled, or deformed. However, the invariance
of CNNs to image transformations would weaken the sensitivity on the
changes in the images. In particularly, for the CNN-LSTM-based detec­
tion scheme, the extracted high-level features of CNN loss the temporal
dependencies in raw images, which may easily lead to the mis­
alignments of the extracted features at time scales. Thus, we propose a

2
J. Yuan and Z. Jiao International Journal of Electrical Power and Energy Systems 147 (2023) 108909

D N

N
N

Fig. 1. Flowchart of the proposed CNN-LSTM-based detection method.

PToP CNN to maintain the time-series dependencies of the extracted ⎧

xp := (x1 , x2 , ⋯xt , ⋯, x32 )
features. ⎪
⎨
[ ]
Suppose the binary image of k-th feeder is Xk, the matrix Xk has two ceil(p/4) − 1 T ceil(p/4) T (3)
⎪
⎩ t ∈ tf + , tf +
dimensions, where there are 128 rows (corresponding to the amplitude) 8 2 8 2
and 256 columns (corresponding to the sampling time). Then, we use the
sliding window method to split the image into patches, where the width where tf is the determined fault time [21], and ceil(.) denotes the round-
of a sliding window is W and the height is H. In this paper, the rows of up function.
the matrix Xk are split into four segments (H = 32), and the columns are For the given patch xp, it corresponds to one section of the image Xk
split into eight segments (W = 32), as shown in Fig. 2, which implies that at the sampling time t. When the CNN is used to extract deep information
there are 32 (4 × 8) patches for the image Xk. Each segmented patch xp from any patch, the extracted information can represent the fault fea­
has the size of 32 × 32, as shown in (2). tures at the time t. Thus, the CNN can not only extract and learn the
spatial features, but also can maintain the feature dependencies at time
xp = S(Xk ), p ∈ [1, 32] (2) scales.
After the images are split into patches, a CNN model is dedicated to
where S(.) denotes the sliding window method.
extract spatial features from the patches. For a 2-D convolution layer,
We can arrange the patches according to the time sequence:
the kernels can extract the feature maps from the input patches, as
shown in (4).
( )
s− 1 ∑
∑ s− 1
ci,j = δ wm,n xi+m,j+n + b (4)
m=0 n=0
W
where wm,n and b are the learned weights and bias of the kernel,
H ... respectively, and s is the size of the kernel. xi,j and ci,j are the input patch
and extracted feature map, respectively. δ denotes the activation func­
tion, and the ReLu function is applied.
Notably, since the images are split into patches with small size, we
...

...

can use CNN with kernels that have little parameters to extract the fault
features, thus significantly reducing the computational cost. Further­
more, the FC layer is discarded. The network structure is shown in Fig. 3.
... The CNN module incorporates six layers, including three convolution
layers, one max pooling layer, one global average pooling (GAP) layer,
and one flatten layer. Among them, the first and second convolution
layers only have 16 kernels, and the third convolution layer has 32
kernels. To downsample the feature maps, the max pooling layer is
added after the first convolution layer, and the second and third

W W
W
W
W

H H H
H H

Fig. 2. Process of obtaining patches from raw image. Fig. 3. Structure of CNN module.

3
J. Yuan and Z. Jiao International Journal of Electrical Power and Energy Systems 147 (2023) 108909

convolution layer have the kernels with stride 2, where both the spatial of the forget gate, respectively. Wi and bi denote the weight and devia­
dimensions of feature maps are halved accordingly. After that, a GAP tion of the input gate, respectively. Wo and bo denote the weight and
layer is used to obtain the feature maps with size of 1 × 1, and a flatten deviation of the output gate, respectively. gt is the intermediate variable,
layer is applied to flatten all the features into a single vector. In this and Wg and bg denote its learned weight and deviation respectively. σ
paper, the output of each patch is the vector with size of 1 × 32, and the denotes the sigmoid function, and tanh denotes the hyperbolic tangent
extracted features of each image can be represented with a 1024 (32 × function. ⊗ means the element-wise multiplication.
32) feature vector, as shown in (5). Since the extracted features by using the PToP CNN retain the de­
pendencies at time scales, the LSTM module is used to handle the tem­
∏
32
( )
Vk = Z xp (5) poral correlation of the features. Besides, it is necessary for faulty-feeder
p=1 detection to conduct correlation comparison between different feeders
[11]. To conduct correlation comparison between feeders, we take the
where Z(.) denotes the process of the feature extraction by using CNN, feature vector Vk as the input xt of the LSTM module. That is to say, the
∏
and denotes the concatenation operation. Vk is the extracted feature LSTM module receives the feature vector {V1, …, Vk, … } from the
vector of the k-th feeder. preceding PToP CNN as the input {x1, …, xt, … }. This type of LSTM can
Therefore, the extracted features of an input image by using the PToP be called as FToF LSTM.
CNN will be transformed into a feature vector containing 1024 elements. In this paper, the established FToF LSTM comprises four layers,
Besides, the feature vector maintains the temporal dependencies, which including two LSTM layers and two dropout layers, where the LSTM
can be further learned by using LSTM for final detection of the faulty layer and dropout layer are alternately stacked. Furthermore, 256
feeder. neurons are adopted in the LSTM layer, and a dropout rate 50 % is
applied in the dropout layer to reduce the overfitting. Subsequently, the
2.4. Feeder-to-feeder LSTM second dropout layer is followed by a FC layer with softmax activation,
and the faulty feeder can be identified according to the classification
As a special kind of recurrent neural network (RNN), the LSTM results. Since a distribution network has many feeders, the identification
network is suitable to process time-series data due to the powerful of the faulty feeder can be considered as a multiclass classification task.
learning capability for long-term dependencies [24]. Owing to the The loss function can be defined as follows:
recurrent architecture, the calculation of the hidden layer nodes can not
1 ∑Q ∑ N
only learn the input of the current layer, but can also retain the L= − P∗ log(Pn,q ) (7)
Q q=1 n=1 n,q
knowledge of the nodes in the previous layers. Moreover, the LSTM can
overcome the drawbacks of RNN on long sequences by using forger gate,
input gate, and output gate. The forget gate is applied to discard where Pn,q is the probability that the n-th feeder in the q-th sample is
redundant information, the input gate is used to select the important classified as the faulty feeder by the CNN-LSTM model, P∗n,q is the actual
information to be stored, and the output gate is utilized to identify the probability of the corresponding faulty-feeder classification. Q is the
output information. A typical LSTM network structure is unrolled in number of samples, and N is the total feeder number.
Fig. 4.
The specific calculation formulas for each gate function of the LSTM 2.5. Training of CNN-LSTM model
cell are shown as follows:
⎧ t [ ]
f = σ(Wf [ ht− 1 , xt] + bf ) In this study, we train the CNN-LSTM model with fault data obtained
⎪
⎪
⎪ t
⎪
⎪ i = σ(Wi ht− 1 , xt + bi ) from PSCAD simulation as it is difficult to collect extensive practical
⎪
⎨ t [ ]
g = tanh(Wg ht− 1 , xt + bg ) data. Three distribution network models with four feeders are estab­
[ t− 1 t ] (6) lished by using PSCAD, as shown in Fig. 5. The models consider different
⎪ t
⎪ o t= σ (W
⎪ o h , x + bo )
⎪
⎪ t t− 1 t t neutral grounding modes and feeder types, and the detailed parameters
⎩ C =t f ⊗t C + i ⊗
⎪ g
h = o ⊗ tanh(Ct ) are shown in Table 1. Among them, the parameters of the overhead lines
are as follows: R1 = 0.17 Ω/km, L1 = 1.21 mH/km, C1 = 0.0097 μF/km,
where ht denotes the hidden layer output at time t, and xt denotes the R0 = 0.23 Ω/km, L0 = 5.48 mH/km, and C0 = 0.006 μF/km, and the
input at time t. ft, it, and ot are the outputs of the forget gate, input gate, parameters of the cable lines are as follows: R1 = 0.098 Ω/km, L1 =
and output gate, respectively. Wf and bf denote the weight and deviation 0.274 mH/km, C1 = 0.351 μF/km, R0 = 0.246 Ω/km, L0 = 0.955 mH/

Fig. 4. Structure of LSTM module.

4
J. Yuan and Z. Jiao International Journal of Electrical Power and Energy Systems 147 (2023) 108909

CNN-LSTM model if the distribution networks have no more than three

feeders.

3. Case study

In this section, we perform large amounts of simulation by using

PSCAD software, where different topologies, parameters, and fault
conditions are considered. Besides, we also conduct the field test at a
practical substation, focusing on the detection under intermittent arc
grounding faults and high impedance faults (HIFs). Both the sampled
data of zero-sequence currents obtained from PSCAD simulation and
field test are utilized as the test data in the following test.
Besides, it is necessary to verity the real-time detection performance
of faulty-feeder detection methods in real devices, especially for the
deep-learning methods. In this study, the trained model is implemented
Fig. 5. PSCAD model in the training dataset. on a professional AI-powered device in the test process, called NVIDIA
Jetson AGX Xavier, as shown in Fig. 6. The software platform of the
Jetson module is Ubuntu (version 20.04), Python (version 3.8), and
Table 1
Pytorch (version 1.12).
Parameters of the established training models.
Furthermore, the performance of faulty feeder detection is evaluated
Training models Sg Feeder type
in different aspects. In addition to the accuracy metric, the evaluation
Model 1 Open Overhead lines (l1 – l4) metrics of precision, recall, and F1 score are also applied, and they are
Model 2 Closed Overhead lines (l1 – l4) defined as:
Model 3 Closed Cable lines (l1 – l2) / Overhead lines (l3 – l4)
⎧
⎨ precision = TP/(TP + FP)
recall = TP/(TP + FN) (8)
km, and C0 = 0.166 μF/km. ⎩
F1 = 2 × precision × recall/(precision + recall)
Considering SPG faults occurred under different fault conditions,
various fault locations (Lf), fault distances (Df), initial phases (θf), and where TP, FN and FP are the total number of the positive samples that
grounding resistances (Rf) are simulated in the training models, as are predicted as positive, the positive samples that are predicted as
summarized in Table 2. Totally 3,060 (816 × 3 + 204 × 3) sets of fault negative, and the negative samples that are predicted as positive,
data are obtained by using ‘Multiple Run’ module in PSCAD, and the respectively.
sampling frequency is 20 kHz.
In fact, the SPG faults obtained in PSCAD simulation are insufficient
for training the CNN-LSTM model. To obtain sufficient fault data, three 3.1. PSCAD simulation
data augmentation methods are applied: flip image vertically, masking,
and rearrange the sequence of input images. For an SPG fault, four raw Fig. 7 shows a radial distribution network with two power sources,
images can be obtained, and they can be flipped vertically to expand the two Buses, and ten feeders established in PSCAD simulation. As shown in
dataset. Besides, one of the images of the healthy-feeder currents are Fig. 7, three feeders are connected to Bus I, and the connected source
masked randomly and the remaining images are input to the CNN-LSTM adopts the ungrounded mode, called the ungrounded network. Seven
model. Moreover, the images can be input to the model in different se­ feeders are linked to Bus II, and the corresponding source is connected to
quences, and six arrangements for three input images and 24 arrange­ the ground via a Peterson coil, namely resonant grounded network.
ments for four input images are considered. Therefore, by using the Furthermore, multiple hybrid feeders with the combination of overhead
three data augmentation methods, a significant increase in the number feeders and cable feeders are considered, including different connection
of fault data can be achieved. Finally, 183,600 (3060 × 2 × 6 + 3060 × modes of feeders. The parameters of the overhead feeders are: R1 = 0.33
2 × 24) sets of fault data are obtained, and they are transmitted to the Ω/km, L1 = 1.31 mH/km, C1 = 7.0 nF/km, R0 = 1.041 Ω/km, L0 = 3.96
CNN-LSTM model for more effective training. The learning environment mH/km, and C0 = 4.0 nF/km, and the parameters of cable feeders are as
is deployed on a personal computer with an Inter CoreTM processor i7- follows: R1 = 0.0791 Ω/km, L1 = 0.264 mH/km, C1 = 373.0 nF/km, R0
6700 (CPU), 3.40 GHz, 16.00 GB RAM, and NVIDIA GeForce GTX = 0.227 Ω/km, L0 = 0.926 mH/km, and C0 = 166.0 nF/km.
1050, and Pytorch (version 1.11.0) is adopted as the software platform. In addition to different feeder types and parameters, the test model
Notably, the PToP CNN and FToF LSTM are trained jointly, where the considers the changing fault conditions of SPG faults, as shown in
extracted features by using PToP CNN are input to the FToF LSTM for Table 3. Besides, Gaussian white noise with the signal-to-noise-ratio
final identification. Besides, the trained CNN-LSTM model has the su­ (SNR) of 30 dB is added to each zero-sequence current in the test pro­
periority in the independency from the changing topologies, which can cess. Moreover, to evaluate the detection performance under low sam­
be directly applied to other distribution networks. To guarantee the pling frequency, the zero-sequence currents in the PSCAD simulation are
detection efficiency in practical applications, only four feeders with the sampled at just 10 kHz frequency.
largest amplitudes of zero-sequence currents will be selected for faulty Typical zero-sequence currents in PSCAD simulation are shown in
feeder detection if the distribution networks have more than four Fig. 8, and the corresponding attended regions on the currents visualized
feeders, and the input images of feeders will be directly feed into the using the gradient-weighted class activation mapping (Grad-CAM)
method [25] are also provided, where the important regions are colored
in red. Evidently, both zero-sequence currents of faulty feeder and
Table 2 healthy feeder are varied significantly under different fault conditions,
Fault conditions simulated in each training model. leading to changing oscillation characteristics in waveforms. Further­
more, there are noticeable burrs in the current waveforms due to the
Lf Df θf / ◦
Rf / Ω Number
noise interference. Despite the changing characteristics and noise
l1 – l4 10 %, 50 %, 90 % 0 ~ 345.6 per 21.6 20, 100, 500, 1000 816 interference, the hybrid CNN-LSTM model can still learn and compare
Bus / 0 ~ 351 per 7.02 20, 100, 500, 1000 204
discriminative features in waveforms, where both the differences

5
J. Yuan and Z. Jiao International Journal of Electrical Power and Energy Systems 147 (2023) 108909

Fig. 6. Experimental system based on the NVIDIA Jetson AGX Xavier.

l l

l l

l l l

l l

l l

l l l

l l l

l l

Fig. 7. PSCAD model in the test dataset.

between faulty-feeder currents and healthy-feeder currents and simi­

Table 3
larities between healthy-feeder currents are effectively covered. For
Fault conditions simulated in the test model.
instance, Fig. 8 (d) exhibits the zero-sequence currents after an SPG fault
Fault Ungrounded network Resonant grounded occurred in feeder 1. The zero-sequence currents have obvious transients
conditions network
in the initial fault time, and then they decrease sharply with oscillation.
Lf lI1 – lI3 / Bus I lII1– lII7 / Bus II Specifically, the zero-sequence current of feeder 1 is larger than that of
Df 3 % / 23 % / 73 % / Bus I 5 % / 35 % / 85 % / Bus II healthy feeders and the polarity is opposite in the initial fault time, while
θf / ◦ 0 ~ 360 per 3.6 0 ~ 360 per 9
Rf /Ω 5 / 35 / 147 / 243 / 621 / 1238 1 / 10 / 80 / 230 / 680 / 1470
the zero-sequence currents of healthy feeders are similar to each feeder,
/ 2000 / 2000 and the currents decay to zero gradually over time. Because the CNN-
Number 6,363 (Line fault) / 707 (Bus 6,027 (Line fault) / 287 (Bus LSTM model can learn well to exploit and compare the spa­
fault) fault) tial–temporal correlations between the currents, where the extracted
features of feeder 1 is different from the other feeders, the feeder 1 can

6
J. Yuan and Z. Jiao International Journal of Electrical Power and Energy Systems 147 (2023) 108909

be accurately identified as the faulty feeder. Finally, the bus fault and
i line fault can be distinguished, and the faulty feeder can be correctly
identified.
The detailed detection performance under various fault conditions is
i shown in Table 4. It is obvious that the proposed method achieves good
i

overall results in terms of accuracy, precision, recall, and F1 score.

i i Furthermore, the average detection time for an SPG fault in the
i i
embedded device is 34.45 ms (27.11 ms for image creation and 7.34 ms
for image recognition), which can satisfy the requirement for timely
t
detection in practical distribution networks. Therefore, the proposed
i i method has high detection performance in all aspects.
i i

3.2. Field test

i
In this study, we conducted the field test at Luohe substation of
i

Henan in China, as shown in Fig. 9. There are six feeders in the sub­
i station, including three practical feeders (feeder 1–3) and analog feeders
(feeder 4–6), and the grounding modes can be changed through switch
t
Sg.
i i Because the practical SPG faults are usually accompanied by the
i i i electric arcs, there are huge intermittence, randomness, and asymmetry
i i in the zero-sequence currents. Besides, HIFs with non-ideal conductors
increase the nonlinearity in the current waveforms. Therefore,
i
compared with PSCAD simulation, the field test considers new fault
i

conditions of arc grounding faults and HIFs, including transient arcing

i faults, steady arcing faults, short gap arcing faults, HIFs with fixed re­
sistances, and single-phase line-broken faults falling on the concrete
t floor. Typical zero-sequence currents and their corresponding visuali­
zation results are described in Fig. 10.
i i
i i i It can be observed that both the zero-sequence currents of faulty
i i feeder and healthy feeder are significantly distorted at the zero-crossing
i point due to arc extinction and reignition, as shown in Fig. 10 (a), Fig. 10
i (b), and Fig. 10 (c), resulting in noticeable distortions in waveforms.
i

Besides, there are long time for the currents to be zero when the arc is
extinguished, especially under steady arcing faults. Moreover, because
i the amplitude of the faulty-feeder current is close to or even smaller than
that of the healthy-feeder currents under HIFs, as shown in Fig. 10 (d)
t
and Fig. 10 (e), it is rather difficult to distinguish the faulty feeder and
healthy feeders. In addition, the currents also have large distortions
i caused by the non-ideal conductor concrete floor under single-phase
line-broken faults. Thus, both the faulty-feeder current and healthy-
feeder current have huge distortions under intermittent arc grounding
i
faults, and the faulty-feeder current may be even smaller than healthy-
i

i i
i i feeder current under HIFs, which would easily lead to the failure of
i i i detection methods based on the certain fault features. However, because
i
the CNN-LSTM model can always focus on the meaningful spa­
t tial–temporal features and conduct effective correlation comparison
accordingly, which is highly robust to distortions in waveforms, the
i faulty feeder can still be accurately identified.
A total of 36 sets of fault data was recorded by fault recorder in field
test, and the sampling frequently is 20 kHz. The detection results are
i listed in Table 5. Evidently, the CNN-LSTM model trained under com­
i

i i mon fault conditions is still applicable under real complex arc grounding
i i
i i faults and HIFs, thus demonstrating its strong generalization capability
i i
and remarkable application prospects.
t
4. Comparison
Fig. 8. Zero-sequence currents and their corresponding Grad-CAM visualiza­
tion results in PSCAD simulation (a) lI2, Df = 0.09 km, θf = 21.6◦ , Rf = 2 kΩ, Currently, the existing faulty-feeder detection methods can be
ungrounded network; (b) lI3, Df = 2.19 km, θf = 334.8◦ , Rf = 35 Ω, ungrounded
network; (c) Bus I, θf = 205.2◦ , Rf = 1238 Ω, ungrounded network; (d) lII1, Df = Table 4
0.25 km, θf = 72◦ , Rf = 1 Ω, resonant grounded network; (e) lII6, Df = 1.5 km, θf Detection performance of the proposed method in PSCAD simualtion.
= 315◦ , Rf = 80 Ω, resonant grounded network; (f) Bus II, θf = 135◦ , Rf = 2 kΩ,
Models Accuracy Precision Recall F1
resonant grounded network.
Ungrounded network 99.96 % 99.96 % 99.94 % 99.95 %
Resonant grounded network 100 % 100 % 100 % 100 %

7
J. Yuan and Z. Jiao International Journal of Electrical Power and Energy Systems 147 (2023) 108909

L L L L L L

Fig. 9. Network structure of the field test.

divided into characteristic-analysis methods and deep-learning feeder detection in resonant grounding distribution systems. There­
methods. To verify the superiority of the proposed method over the fore, the identification results of faulty-feeder detection by using the
existing methods, one characteristic-analysis method and two deep- proposed method and other three methods are compared under line
learning methods are selected for comparison: faults in the distribution networks with resonant grounding.
To evaluate the robustness to noise interference, we compare the
(1) Method in [8]: The extracted fault features of transient energy, detection accuracy of the proposed method and other three methods
kurtosis, and cross-correlation distance are combined by using under noise with different SNRs ranging from 1 dB to 30 dB in PSCAD
multiple evidence estimation method, and the fault trust degree simulation. The detection results are shown in Fig. 11. It can be clearly
of each feeder can be calculated based on the fusion results. The seen that the proposed method outperforms all the compared methods
feeder with maximum fault degree is selected as the faulty feeder. under different SNR levels. Especially, the proposed method still has
Evidently, the method in [8] belongs to the characteristic- high detection accuracy under strong noise interferences, such as SNR
analysis methods. with 1 dB. Therefore, large amounts of experiment results demonstrate
(2) Method in [26]: By using continuous wavelet transform (CWT), that the proposed method has superior robustness against noise
the time–frequency gray scale images of zero-sequence currents interference.
are obtained and input to the CNN for classification, which cor­ In addition to detection accuracy over the five datasets in PSCAD
responds to the deep-learning methods. The identification result simulation, we also evaluate the classification metrics, such as precision,
of the faulty feeder is ‘1’, and it is ‘0’ for the healthy feeder. recall, and F1 score, as shown in Table 6, and the evaluation metrics in
(3) Method in [27]: In [27], the authors proposed a deep-learning the field test are also provided. It is obvious that the proposed method
method based on the cooperation of variational mode decompo­ has superior detection performance in terms of all the evaluation metrics
sition (VMD) and LSTM network. Three intrinsic mode functions on the six datasets. There is little detection performance deterioration of
(IMFs) are extracted from zero-sequence currents by using VMD the proposed method when the noise increases, while the other three
and learned by LSTM network accordingly. Similarly, the faulty methods exhibit the opposite performance. Therefore, the proposed
feeder can be detected based on its output results. method has remarkable improvements in the detection performance of
faulty-feeder identification, which can be attributed to the overall
It is noteworthy that the method in [8] directly select the feeder with exploitation of the spatial–temporal features in the zero-sequence cur­
the maximum calculated fault degree as the faulty feeder, which implies rents and comprehensive correlation comparison between feeders.
that the bus fault cannot be identified. For the method in [26] and [27],
they fail to compare the correlations between currents in the detection 5. Conclusion
process, which is also not suitable for distinguishing the bus faults and
line faults. Furthermore, these methods mainly focus on the faulty- To fully exploit and compare the spatial–temporal features in the

8
J. Yuan and Z. Jiao International Journal of Electrical Power and Energy Systems 147 (2023) 108909

Table 5
i i
i i i Detection performance of the proposed method in field test.
i i Models Accuracy Precision Recall F1

Ungrounded network 100 % 100 % 100 % 100 %

i
Resonant grounded network 100 % 100 % 100 % 100 %
i

i i
i i i
i i

i
i

i i
i i i
i i

i
Fig. 11. Comparison of the detection accuracy under different SNRs.
i

i currents, this paper proposes a novel detection method based on the

cooperation of the PToP CNN and FToF LSTM. The main conclusions are
t drawn as follows.

(1) The PToP CNN can learn the spatial features of zero-sequence
i
currents in waveforms, and the FToF LSTM can mine the tem­
poral dependencies in time-series data. By using the hybrid CNN-
i LSTM model, the spatial–temporal correlations between the
currents can be extracted and compared, and the attended re­
i

i i
i i gions visualized using the Grad-CAM method prove the effec­
i i i tiveness of the feature extraction and comparison.
(2) The hybrid CNN-LSTM model is implemented in the embedded
t device NVIDIA Jetson AGX Xavier, demonstrating its feasibility in
actual applications. Totally 13,384 sets of fault data obtained
from PSCAD simulation and 36 sets of practical data collected
i
from field test are utilized to verify the detection performance.
The detection results show that the proposed method has ach­
i ieved over 99.9 % in terms of all the evaluation metrics, and it can
identify the faulty feeder within 34.45 ms on average in the
i

i i
i i embedded device, demonstrating its good real-time detection
i i i performance.
(3) Owing to the comprehensive exploitation and comparison of the
t spatial–temporal features in the currents, the proposed method
has superiority robustness against noise interference compared
i i with the existing detection methods, and it still has 98.84 %
i i i
detection accuracy under strong noise with SNR of 1 dB. There­
i i
fore, the proposed method has considerable prospects in practical
i distribution networks.
i

CRediT authorship contribution statement

i
Jiawei Yuan: Conceptualization, Methodology, Software, Valida­
t tion, Formal analysis, Investigation, Data curation, Writing – original
draft, Writing – review & editing. Zaibin Jiao: Writing – original draft,
Fig. 10. Zero-sequence currents and Grad-CAM visualization results in field Writing – review & editing, Supervision.
test (a) transient arcing fault, Sg is open; (b) steady arcing fault, Sg is open; (c)
short gap arcing fault, Sg is open; (d) HIFs with 1 kΩ, Sg is closed; (e) HIFs with
2 kΩ, Sg is closed; (f) single-phase line-broken fault, Sg is closed. Declaration of Competing Interest

The authors declare that they have no known competing financial

9
J. Yuan and Z. Jiao International Journal of Electrical Power and Energy Systems 147 (2023) 108909

Table 6
Detection performance of the compared methods on the six datasets.
Datasets Models SNR/dB Metrics Method [8] Method [26] Method [27] Proposed method

PSCAD Resonant grounded network 30 Accuracy 85.75 % 96.73 % 99.14 % 100 %

Precision 90.84 % 96.87 % 99.14 % 100 %
Recall 85.75 % 96.73 % 99.14 % 100 %
F1 85.85 % 96.74 % 99.14 % 100 %
20 Accuracy 81.09 % 91.29 % 99.44 % 100 %
Precision 88.77 % 91.72 % 99.44 % 100 %
Recall 81.09 % 91.29 % 99.44 % 100 %
F1 81.46 % 91.39 % 99.44 % 100 %
10 Accuracy 73.30 % 61.76 % 99.17 % 100 %
Precision 84.33 % 67.35 % 99.18 % 100 %
Recall 73.30 % 61.76 % 99.17 % 100 %
F1 72.86 % 63.25 % 99.17 % 100 %
5 Accuracy 70.22 % 47.88 % 98.34 % 99.85 %
Precision 82.57 % 53.33 % 98.40 % 99.94 %
Recall 70.22 % 47.88 % 98.34 % 99.89 %
F1 69.04 % 49.02 % 98.35 % 99.92 %
1 Accuracy 68.91 % 37.65 % 91.87 % 98.84 %
Precision 82.04 % 41.92 % 92.61 % 99.71 %
Recall 68.91 % 37.65 % 91.87 % 99.16 %
F1 67.29 % 37.79 % 92.07 % 99.43 %
Field test Resonant grounded network / Accuracy 44.44 % 44.44 % 72.22 % 100 %
Precision 50.00 % 59.72 % 60.00 % 100 %
Recall 41.67 % 44.44 % 40.00 % 100 %
F1 45.00 % 43.10 % 46.00 % 100 %

interests or personal relationships that could have appeared to influence [12] Samantaray SR. Ensemble decision trees for high impedance fault detection in
power distribution network. Int J Elect Power Energy Syst 2012;43(1):1048–55.
the work reported in this paper.
[13] Shadi MR, Ameli MT, Azad S. A real-time hierarchical framework for fault
detection, classification, and location in power systems using PMUs data and deep
Data availability learning. Int J Elect Power Energy Syst 2022;134:107399.
[14] Yuan JW, Hu YJ, Jiao ZB. Faulty feeder detection method for SLG faults in
distribution networks based on comprehensive fault characteristics across entire
Data will be made available on request. frequency spectrum. Int J Elect Power Energy Syst 2022;140:107835.
[15] Srivastava A, Parida SK. A robust fault detection and location prediction module
References using support vector machine and gaussian process regression for AC microgrid.
IEEE Trans Ind Appl 2022;58(1):930–9.
[16] Yu K, Zou H, Zeng XJ, Li YY, Li H, Zhuo C, et al. Faulty feeder detection of single
[1] Yuan JW, Jiao ZB, Feng G, Chen M, Xu MM. Study on fault line detection methods phase-earth fault based on fuzzy measure fusion criterion for distribution
based on multi-feature fusion in distribution systems. IET Gener Transm Distrib networks. Int J Electr Power Energy Syst 2021;125:106459.
2021;15(5):860–9. [17] Du Y, Liu YD, Shao QZ, et al. Single line-to-ground faulted line detection of
[2] Gao J, Wang XH, Wang XW, Yang AJ, Yang H, Wei XX. A high-impedance fault distribution systems with resonant grounding based on feature fusion framework.
detection method for distribution systems based on empirical wavelet transform IEEE Trans Power Del 2019;34(4):1766–75.
and differential faulty energy. IEEE Trans Smart Grid 2022;13(2):900–12. [18] Gao JH, Guo MF, Chen DY. Fault line detection using waveform fusion and one-
[3] Cui T, Dong XZ, Bo ZQ, Juszczyk A. Hilbert-transform-based transient/intermittent dimensional convolutional neural network in resonant grounding distribution
earth fault detection in noneffectively grounded distribution systems. IEEE Trans systems. CSEE J Power Energy Syst 2021;7(2):250–60.
Power Del 2011;26(1):143–51. [19] Bao TZ, Zaidi SAR, Xie SQ, Yang PF, Zhang ZQ. A CNN-LSTM hybrid model for
[4] Yuan JW, Hu YJ, Liang YH, Jiao ZB. Faulty feeder detection for single line-to- wrist kinematics estimation using surface electromyography. IEEE Trans Instrum
ground fault in distribution networks with DGs based on correlation analysis and Meas 2021;70:1–9.
harmonics energy. IEEE Trans Power Del 2022. https://doi.org/10.1109/ [20] Ren L, Dong JB, Wang XK, Meng ZH, Zhao L, Deen MJ. A data-driven auto-CNN-
TPWRD.2022.3203992. LSTM prediction model for lithium-ion battery remaining useful life. IEEE Trans
[5] Liu BW, Ma HZ, Xu HH, Ju P. Single-phase-to-ground fault detection with Ind Inform 2021;70(5):3478–87.
distributed parameters analysis in non-direct grounded systems. CSEE J Power [21] Yuan JW, Wu T, Hu YJ, Jiao ZB. Faulty feeder detection based on image
Energy Syst 2019;5(1):139–47. recognition of voltage-current waveforms in non-effectively grounded distribution
[6] Liu PH, Huang C. Detecting single-phase-to-ground fault event and identifying networks. Int J Elect Power Energy Syst 2022;143:108434.
faulty feeder in neutral ineffectively grounded distribution system. IEEE Trans [22] Yuan JW, Jiao ZB. Faulty feeder detection based on image recognition of current
Power Del 2018;33(5):2265–73. waveform superposition in distribution networks. Appl Soft Comput 2022;130:
[7] Wang XW, Gao J, Wei XX, Zeng ZH, Wei YF, Kheshti M. Single line to ground fault 109663.
detection in a non-effectively grounded distribution network. IEEE Trans Power [23] Azulay A, Weiss Y. Why do deep convolutional networks generalize so poorly to
Del 2018;33(6):3173–86. small image transformations? J Mach Learn Res 2019;20:1–25.
[8] Wei XX, Yang DC, Wang XW, Wang B, Gao J, Wei KW. Faulty feeder detection [24] Zuo Y, Wu YL, Min GY, Huang CQ, Pei K. An intelligent anomaly detection scheme
based on fundamental component shift and multiple-transient-feature fusion in for micro-services architectures with temporal and spatial data analysis. IEEE Trans
distribution networks. IEEE Trans Smart Grid 2021;12(2):1699–711. Cognit Commun Netw 2022;6(2):548–61.
[9] Wei XX, Wang XW, Gao J, Yang DC, Wei KW, Guo L. Faulty feeder detection for [25] Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-cam: Visual
single-phase-to-ground fault in distribution networks based on transient energy explanations from deep networks via gradient-based localization. IEEE
and cosine similarity. IEEE Trans Power Del 2022;37(5):3968–79. International Conference on Computer Vision (ICCV) 2017; 618–626.
[10] Wang XW, Wei XX, Gao J, Song GB, Wu L, Liu JW, et al. High impedance fault- [26] Guo MF, Zeng XD, Chen DY, Yang NC. Deep-learning-based earth fault detection
detection method based on the variational mode decomposition and Teager-Kaiser using continuous wavelet transform and convolutional neural network in resonant
energy operators for this distribution network. IEEE Trans Smart Grid 2019;10(6): grounding distribution systems. IEEE Sens J 2018;18(3):1291–300.
6041–54. [27] Zhai EJ, Shu ZY, Wang J, Huang ZP. Fault line selection method of small current
[11] Yuan JW, Jiao ZB. Faulty feeder detection based on fully convolutional network grounding system based on VMD-LSTM. Adv Technol Electr Eng Energy 2021;40
and fault trust degree estimation in distribution networks. Int J Elect Power Energy (1):70–80.
Syst 2022;141:108264.

10