Deep Graph Neural Network for Fault Detection and Identification

in Distribution Systems
Quang-Ha Ngo1 , Bang L H Nguyen2 , Jianhua Zhang1 , Karl Schoder3 , Herbert Ginn4 , and
Tuyen Vu1
1
Clarkson University
Posted on 5 Sep 2024 — CC-BY-NC-SA 4 — https://doi.org/10.36227/techrxiv.172555531.16904989/v1 — e-Prints posted on TechRxiv are preliminary reports that are not peer reviewed. They should not b...

2
Duy Tan University
3
Florida State University
4
University of South Carolina

November 02, 2024

1
 Deep Graph Neural Network for Fault Detection
and Identification in Distribution Systems
Quang-Ha Ngo1 , Bang L. H. Nguyen2 , Jianhua Zhang1 , Karl Schoder3 , Herbert Ginn4 , and Tuyen Vu1
1 Clarkson University, 2 Duy Tan University, 3 Florida State University, 4 University of South Carolina

Email: {hanq, jzhang, tvu}@clarkson.edu, bangnguyen@ieee.org, schoder@caps.fsu.edu, ginnhl@cec.sc.edu

Abstract—With the growing integration of renewable energy relay-based and data driven-based approaches. Digital relay-
sources, the penetration of the distributed generation leads based methods rely on protection devices installed at certain
to increasingly dynamic power distribution system topologies. locations to detect faults, typically using impedance, travelling
This poses challenges for traditional fault diagnostic methods in
accurately classifying and locating faults. This paper develops a waveform, and sparse measurements analysis, as described in
deep graph neural network (GNN) for detecting and managing [2]. Travelling wave methods analyze transient waves from
fault events on distribution systems. The 1-D convolutional graph fault points to facilitate the identification and localization of
attention networks (1-D GAT) can exploit the spatial-temporal faults but require high-frequency sensors, substantially raising
features from both voltages and branch currents to enhance the deployment cost [13]. The impedance-based methods iter-
the accuracy of fault diagnostics compared to existing methods.
The effectiveness of the proposed method is evaluated on the atively solve solve nonlinear equations with line steady-state
Potsdam 13-bus and IEEE 123-bus test feeder systems. Results conditions using voltage or current measurements to locate
indicate notable improvements in accuracy and other metrics, faults [14]. Voltage sag-based methods identify voltage drop
achieving a 1-2% increase in fault event detection, an average patterns at certain monitored buses, as faults produce unique
4% improvement in identifying fault types, and an average 5% voltage sag propagation depending on their location [15].
improvement in fault location.
Index Terms—Fault management, distribution system, 1-D Some other digital relay-based approaches like the transient
convolutional, graph attention networks, deep learning. monitor function [16], or the Hibert transform-based detection
[17] are suggested to identify faults by analyzing current
I. I NTRODUCTION signals. Both [18] and [19] point out that the precision of
Modern power grids are beginning to transition to clean impedance-based techniques can be influenced by factors such
renewable energy resources. It has been predicted that by as types of fault, measurement errors, and varying system
2050, 44% of electrical energy in the US will come from parameters. However, these traditional coordination methods
renewables (with more than 80% from solar and wind) [1]. for protective relays are facing challenges related to cost, ac-
While renewable energy resources provide benefits through curacy, and uncertainty due to the intermittent and distributed
the generation of clean electricity, the integration of more nature of these renewable energy sources. This dynamic nature
renewables elevates the risk of faults in power distribution of modern distribution systems necessitates the development
systems due to bad weather, insulation failures, and improper of new data-driven approaches for fault diagnostics.
operations. More than 70% of the main causes of interruptions Unlike the digital relay-based methods, data driven methods
are caused by various faults in the distribution system opera- focus on mining distinctive features from a large amount of
tion [2]. The resulting accidental blackouts can affect critical system measurements and fault data to identify and locate
business operations and production, and even become threats faults in power distribution systems without solving complex
to communities. Therefore, three essential tasks - fault event equations of the physical systems or analyzing digital relays.
detection, classification, and localization - are introduced to [20], [21] collectively introduce and explore the innovation
prevent this issue. Firstly, fault event detection identifies if a to fault classification using a convolutional neural network
fault occurs in the power system or not. Secondly, fault classifi- (CNN) on a 3-bus distribution network. [22] proposes an
cation involves determining specific fault types, such as single adaptive time-frequency memory (AD-TFM) cell that embeds
line-to-ground, line-to-line, or three-line-to-ground faults. This an adaptive wavelet transform into long short-term memory
information is critical for selecting the appropriate protection (LSTM) for fault detection in power distribution systems.
schemes and isolation methods [3]. Thirdly, fault localization [23] mentions a voltage data processing-based approach us-
identifies faulted positions and expedites restoration. Solving ing Gaussian process regression for fault localization and
three tasks as simultaneously as possible is crucial to enable a isolation in AC microgrids. [24] introduces a bidirectional
rapid and effective fault isolation of protection schemes, and LSTM for predicting the voltage stability in hybrid AC/DC
restoration in power systems, thereby minimizing disruptions microgrids. [25] incorporates LSTM and an adaptive neural
and enhancing overall reliability. fuzzy inference system (ANFIS) to accurately detect faults
Existing power system fault detection techniques, proposed in the IEEE 13-node system. [26] proposes a wavelet multi
in the literature [2], [4]–[12], can be categorized into digital resolution analysis and data mining-based approach for fault
detection and classification. [27] uses Fast Fourier Transform
for feature extraction and employs a multilayer perceptron
neural network for fault classification and location in dis- Neighborhood
tribution systems. [28] proposes a deep CNN-transformer
model that utilizes 1-D deep CNNs for feature extraction
and a transformer encoder with self-attention for sequence
learning to fault detection in the IEEE 14-bus distribution Fig. 1. A self-attention process in each graph attention layer [41]
system. Overall, the data driven-based methods aim to learn
the relationship between measurements and output labels by to perform fault event detection, fault type identification, and
minimizing the loss function over the training data. However, fault localizationHistorical
using and three-phase voltagePower andSystem
branch current
simulation data Real-time Measurements
these conventional approaches struggle to effectively capture data of main buses. The main contributions can be summarized
correlations for regression and classification tasks, particularly as follows:
with the increasing complexity of information from unevenly • We propose the 1-Dand
Data acquisition convolutional graph
1-D Graphattention
Attention neural
Feature extraction
distributed and dynamic distribution systems. network to improve the accuracy Network of faultmodels
diagnostics
Motivated by the limitations of existing data-driven ap- using voltages and branch currents asDeploy inputs.
proaches, graph neural network-based fault diagnostic methods • We are Graph
the construction
first to and exploit branch
Potsdamcurrents as multi-
IEEE 123-Bus
Parameter initialization 13-Bus
have been proposed. [29] proposes a multi-receptive field dimensional edge attributes for feature representation,
graph convolutional network to learn feature representations enhancing the capability of diagnosing faults since branch
from multiple neighborhood domains for mechanical fault Deep Graph Neural
currents can Fault event detection,
indeed reflect fault characteristics.
Network Training classification & localization
diagnosis. In [30], a graph convolutional neural network • The proposed method is noise-resilient, as validated with
Training Stage Testing Stage
leverages dynamic voltage measurements as nodal features the Potsdam 13-node system and the IEEE 123-node
to determine the types and locations of faults in an 8-bus system under noise scenarios.
shipboard test network. A semi-supervised graph convolutional The remaining sections are structured as follows. Section II
networks is proposed in [31] to address the limited availability introduces the description of the proposed fault diagnostic
of labeled data for electromechanical system fault diagnosis. scheme. The section III shows the simulation results of our
[32] introduces a combination of the contrastive learning and proposed method, employed to the IEEE 123-node and Pots-
a generative adversarial neural network (GAN) to detect and dam 13-node systems. The final section, Section IV, presents
classify faults in distribution lines. These papers mention that a the conclusions.
graph neural network (GNN) can leverage the graph-structured
data to effectively capture the spatial correlations among power II. M ETHODOLOGY
system components, addressing the limitation of traditional This section introduces the principles and structures of
data-driven methods. the 1-D convolution layer and the graph attention layer. The
Although GNN-based approaches achieve remarkable re- proposed fault diagnostic approach based on the 1-D GAT is
sults for fault location and classification in power systems, then described in detail.
there are still research gaps. The summary of major technical
differences with current GNN-based approaches is outlined in A. 1-D Convolutional Layer
Table I. Voltage phasors and current injections are utilized The 1-D convolution layer is a deep learning technique for
as node features by [33], [34], and [35], while only volt- analyzing sequential data, such as time series, audio signals,
age measurements are considered as node features for the and images over a single dimension [42]. This layer slides
fault detection, classification, and localization under varying a filter over the sequence, extracting features by applying a
conditions in [36]–[39]. However, these approaches neglect convolution operation that captures the temporal dependencies
branch currents, potentially limiting their fault analysis capa- within the data. This makes the 1-D convolutional layer
bility. Incorporating current measurements as edge features is well-suited to learn local patterns efficiently in recognizing
believed to be essential for enhancing fault diagnostic accuracy trends and anomalies in time series [42]. The 1-D convolution
in the future research of these papers because different types operation can be mathematically formulated as follows:
of faults (such as line-to-ground, line-to-line, or three-phase
faults) exhibit distinct patterns in branch currents [33], [34], 
Nl−1

[36]. In [40], a physics-preserved graph network is introduced Ckl = σ 
X
l−1
(wik ∗ Cil−1 ) + blk  , (1)
to enhance fault location accuracy under limited data, but the i=1
range of fault types is limited. Regarding fault detection and
identification, the work in [33]–[35], [39], [40] only focus where Ckl is the output of the k th neuron at the layer l and
on one or two tasks. In summary, current GNN-based fault Cil−1 is the output of the ith neuron at previous layer l − 1.
l−1
diagnostic methods have notable limitations including: Limited Nl−1 represents the set of feature maps. wik are the trainable
fault diagnosis capability, and Low resolution of fault types. weights from the i neuron at layer l − 1 to the k th neuron at
th

In this paper, a novel fault diagnostic deep graphical learn- layer l. The operator * indicates the convolution operation and
ing method for distribution systems is developed. The target is b stands for the network bias. σ(·) is the activation function.
 TABLE I
C OMPARISON OF T ECHNICAL A SPECTS W ITH E XISTING GNN- BASED R ESEARCH

Aspects Chen et al. [33] Hu et al. [34] De et al. [35] Bang et al. [36] Xu et al. [37] Chan et al. [38] Tong et al. [39] Li et al. [40] Ours
Fault Detection ✗ ✗ ✗ ✓ ✓ ✓ ✓ ✗ ✓
Fault Classification ✗ ✓ ✗ ✓ ✓ ✓ ✓ ✗ ✓
Fault Localization ✓ ✓ ✓ ✓ ✓ ✓ ✗ ✓ ✓
Branch Currents
Considered as Inputs ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓
Noise resilient ✓ ✓ ✗ ✗ ✗ ✗ ✓ ✗ ✓
Types of Fault LG, LL LG, LL LG LG, LL, LLG LG, LL LG, LL, LLG LG, LL LG, LL LG, LL, LLG
LLG LLG, LLL LLL, LLLG LLL LLL, LLLG LLG, LLL LLG LLL, LLLG

Following each 1-D convolutional layer, a nonlinear acti- and attention mechanism. The components involved in the
vation layer is applied. The layer is crucial to enable the information aggregation can be described as follows.
modeling of complex relationships between inputs and outputs 1) Self-attention mechanism: This enables nodes in a graph
by introducing non-linearity. Using an activate layer also to weigh the importance of their neighboring nodes
increases the flexibility and power of neural networks to model when updating their own representation. This allows
complex relationships within the data effectively. Among the each node to dynamically adjust its representation based
widely used activation functions are the Rectified Linear Unit on the information from its neighbors, capturing the re-
(ReLU), Sigmoid, and Tanh, each offering different character- lational dependencies within the graph. The coefficients
istics suited to specific tasks. The ReLU function, which is ei,j are computed across pairs of nodes i,j based on
employed in this paper could be written as follows: their features as:

 ei,j = a(W T hi , W T hj ), (3)

x, x ⩾ 0
F (x) = (2) where W is a weight matrix for the linear transformation
0, x < 0
applied to every node. hi and hj denote feature vectors
To this end, 1-D convolutional layers are essential to the at node i and node j. a represents a concatenation of
sequential pattern recognition that enables efficient and accu- two node representations.
rate processing of time-series data, enhancing the accuracy of 2) Normalization: To make coefficients comparable across
fault detection capabilities. various nodes, the Softmax function are employed to
normalize them over available choices of j.
B. Graph Attention Layer E(ei,j )
αi,j = sof tmaxj (ei,j ) = P|V | , (4)
i
In 2018, [41] introduced a graph attention networks (GAT), k=1 E(ei,k )
employing a self-attention strategy to perform node classi- where E represents an exponential function. |Vi | denotes
fication of graph-structured data. GAT employs a dynamic the set of neighbor vertices of node i, and αi,j specifies
weighting mechanism that assigns varying importance levels the weighting factor of node j’s features to node i.
to node features, rather than relying on predetermined weights 3) Aggregation: The aggregated features from each node is
from node degrees. In this way, each node’s representation is concatenated to obtain a new node feature h′i as
computed as a weighted sum of its neighbors’ representations  
during the information aggregation process, and the adjacency X
matrix is dynamically adjusted during training to optimize h′i = σ  αi,j W hj  , (5)
performance. This adjustment allows the neural network to j∈Vi
make more accurate predictions by focusing on the most where σ applies the LeakyReLU non-linearity, an acti-
important information. vation function.
As can be seen in Fig. 1, each graph attention layer
takes node embeddings as input and produces transformed C. Fault Diagnostics via a 1-D GAT Neural Network
embeddings as output. The input of the layer is a collection This work aims to develop a novel 1-D graph attention
of node features h = {h1 , h2 , · · · , hN }, where N represents neural network model for fault management, with triple tasks
the total number of nodes or buses, and hn denotes a node of detection, classification, and localization. To effectively
feature vector, for instance, a bus phasor voltage at the enhance the model’s information aggregation capabilities and
nth node. The attention layer produces a new set of node feature representation, branch currents, the most distinct pa-
features h′ = {h′1 , h′2 , · · · , h′N }. According to [43], a standard rameter to reflect fault characteristics, are incorporated into
graph attention layer includes the self attention, normalization, the proposed model as multi-dimensional edge attributes. The
 Spatial Information
Investigated System

PMU PMU
2 3 4 5 6 7
Graph-structured Dataset

DGU 5

Reshape
13

PMU PMU PMU

1 12 11 10 9 8

DGU 1 DGU 2 DGU 3 DGU 4

Temporal Information

ID(60, 32, edge index, ID(13, 16,

edge attributes) kernel size = 6) ID(432) ID(256) ID(128) OD(1)
Batch Fault Event
Detection
GATConv

Dropout

Dropout
Conv1d

Flatten

Dense
Dense

ReLu

ReLu
Dense
OD(11)
ReLu

ReLu
Fault Type
Classification
OD(13)
Fault Location
DataBatch(x=13, 1, 20, 3), edge index
= [2, 13], edge attrs = [13, 1, 20, 3]) OD(13, 32) OD(16, 27) OD(256) OD(128) Sigmoid or Softmax

Fig. 2. The 1-D Convolutional Graph Attention Network Architecture

training process, depicted in Fig. 2, can be summarized as feature of bus i and the edge attribute as a branch current
Target node
follows. Initially, the spatial and temporal information is col- between bus i and bus j are formulated as
lected from the power system and pre-processed to transform Neighborhood

into graph format. The graph datasets are divided into batches
   
Va,1 Va,2 · · · Va,K Ia,1 Ia,2 · · · Ia,K
which are propagated through the 1-D GAT model. The Xi =  Vb,1 Vb,2 · · · Vb,K  , Eij =  Ib,1 Ib,2 · · · Ib,K  (7)
configuration of 1-D GAT includes one graph attention layer, Vc,1 Vc,2 · · · Vc,K Ic,1 Ic,2 · · · Ic,K
one 1-D convolutional layer, and three fully connected layers The information of each data is considered as a graph
with decreasing number of features. The inclusion of a 1-D G = (X : {X1 , X2 , · · · XM } , E, A, E = {E1 , E2 , · · · EN }),
convolutional Historical
layer allows
and 1-D GAT to Powerincorporate
System a temporal where X represents a set of node features, M denotes the total
simulation data Real-time Measurements
pattern recognition capability into the graph-structured data number of buses in the system, E refers to the collection of
analysis. The input data for fault diagnosis includes the node edges, and each edge connects two nodes [44]. E consists of
features X,Datatheacquisition and
edge indices E, the adjacency
1-D Graph Attention
matrix A, and the vector of edge features, N is the total number of edges, and
Feature extraction Network models
the edge features E. The learning goal is to develop a model A denotes the adjacency matrix, representing the connectivity
Deploy
that can effectively capture and generalize the relationship of the distribution network.
Graph construction and Potsdam
between the node and edge attributes,
Parameter initialization
and theIEEEdesired
13-Bus
123-Bus output, The fault types are categorized into 11 labels, including
illustrated as Eq. (6a). The loss function L(ŷ, y), described single line-to-ground (AG, BG, CG), line-to-line (AB, BC,
as Eq. (6b),Deep considers
Graph Neuralthe probability distribution
Fault event detection, over the CA), double line-to-ground (ABG, BCG, CAG), triple line-
classes predicted
Networkby the model and
Training classification
compares & localization
them with the to-ground (ABCG), and triple short-circuit (ABC). The yevent
Training
true distribution. The Stageinput-output model Testing
of theStage1-D GAT and denotes fault occurrence with binary values, 1 for presence
the loss function can be expressed as and 0 for absence. The ytype ranges from 0 to 10, indicating
corresponding 11 fault labels. Similarly, the fault location label
ŷ = F (X, E, A, E, W ) (6a) is ylocation = i, where i = 1, 2, · · · M if there is a fault at the
XN ith bus.
L(ŷ, y) = − yi log(yˆi ) (6b) Algorithm 1 and 1 describe the training and testing proce-
i=1 dure for fault detection, classification, and localization. The
training procedure involves iteratively sampling batches of
where W is the trainable weights, y is the ground-truth label graph data, using the proposed 1-D GAT model to identify
and ŷ is the output label. yi represents the true probability of labels, computing the loss, and updating model parameters.
class i (0 or 1) and yˆi represents the predicted probability of The testing procedure outlines where the trained model’s
class i. The label vectors y and ŷ could correspond to yevent parameters are loaded, the model is applied to the test data,
for fault binary event, ytype for fault categories, or ylocation and evaluation metrics are computed. The training of graph
for fault locations. datasets utilizes the stochastic gradient descent (SGD) as an
Each three-phase bus voltage and branch current are respec- optimizer. The binary cross entropy (BCE) serves as the loss
tively denoted as Va,k , Vb,k , Vc,k and Ia,k , Ib,k , Ic,k , where K function for fault event detection, while fault classification and
is the length of the period and k = 1, 2, ..., K. The node location utilize the cross entropy.
 III. S IMULATION R ESULTS 2 3 4 5 6 7

A. Case Studies
DGU 5 : Fault points
This paper focuses case studies involving the Potsdam 13- 13 : PMU measurements
node and the IEEE 123-node systems, illustrated in Figs. 3
and 4. The Potsdam 13-node system, designed by our lab, is
powered by five inverter-based generators (IBGs), including 1 12 11 10 9 8

one photovoltaic array, two hydro plants, and two fossil fuel-
based backup generators. This microgrid can operate in grid-
DGU 1 DGU 2 DGU 3 DGU 4
connected as well as islanded modes, ensuring uninterrupted
Fig. 3. The Potsdam 13-node system diagram
power supply even during outages. The nominal voltage is
32 29 250 350
13.2 kV at 60 Hz. The microgrid is simulated using Matlab and 2
33
PMU 303
251
4 51 5 PMU 6
111 110 112 7113 114
28 50 151 300
Opal-RT, with load and IBG parameters set according to those 31
25 47
49
109 107
48 DGU 5 46
outlined in [45]. Input measurements are captured at locations 27
26
45
64
108
106 104 451

43 103
identified by blue squares. The bus voltages and branch 13 23 44 65
105
450
102
63 100
42 41
currents are measured at a rate of 1 kHz. For the proposed fault PMU
24
21 PMU
66 PMU
101
99 71
40 98
diagnostic approach utilizing the deep graph neural network, 1
22
12 11 38
39
10
62
9
197 70
8
36 97 69
operational data including load-change scenarios and fault 20
19
18
135
35
68
75
160 67
conditions are gathered for training and testing procedures. 37
60
73
74

57
The graph structure representing the system’s topology is DGU 111 14
59
58 DGU 2 DGU 372 DGU 4
85
61 79
610
constructed, taking into account all 13 buses in the system. 10
9
52
53 54 77
78
2 152 55 56
Regarding the IEEE 123-node system, we perform the sim- 7 8 13
76
80
84
94
ulation in the Opal-RT, with parameter settings in accordance 149 1 34
96 76
90 88 81
150 12 92
17
with [46]. Fault positions are created at different three-phase
15 86 83
buses across the system, as illustrated in Fig. 4. Measurement 3 5 6
95
93 91 89
87 82

16
data, collected only from the buses, are indicated by blue 4 195

: Fault points : μPMU measurements

squares on the diagram. The graph structure of dataset is
Fig. 4. The IEEE 123-node feeder system diagram
constructed based on the connectivity of 51 primary buses (1,
7, 8, 13, 18, 21, 23, 25, 28, 29, 30, 35, 40, 42, 44, 47, 49, 50,
at 13 different bus locations in the Potsdam microgrid system
51, 52, 53, 54, 55, 57, 60, 62, 63, 64, 65, 67, 72, 76, 77, 78,
under 150 randomly load scenarios. We gathered a graph
80, 81, 82, 83, 86, 87, 89, 91, 93, 97, 98, 99, 100, 101, 105,
dataset comprising 55,000 fault cases for the training set,
108, 197) and 46 branches.
along with 9,350 cases for the test set. We add 10,000 non-
B. Data Simulation and Model Training fault cases for fault event detection. Regarding the 123-node
The dataset information utilized in the training process for system, similarly, we collected a set of 16,500 data samples,
both the Potsdam 13-node and IEEE 123-node systems is encompassing 11 fault types and 3 resistance values across
presented in Table II. The dataset consists of 64,350 samples 50 randomly generated load-change cases. 13,200 load-change
covering 11 different fault types with three different fault cases are generated to form the train set and 3,300 cases are
resistance values (0.1, 1, and 10 Ω). These faults are simulated selected for the test set. Additionally, 4,125 non-fault cases
are included for fault event detection. A one-second voltage,
illustrating a fault occurrence in between, is shown in Fig.
Algorithm 1: Training of 1-D GAT model. 5. For each sample, a 20-millisecond segment in the fault
Data: D (training data), G (graph data), L (loss duration is extracted and used as the input interval for model
function), W (model parameters) training.
Result: Optimal parameters of the 1-D GAT model
1 Initialize W with random values;
2 for epoch ← 1 to nepochs do Algorithm 2: Testing of 1-D GAT model
3 for B ∈ D do 1 . Data: G (graph data), W ∗ (optimized parameters)
4 Sample a batch B from D; Result: Fault detection, classification and localization
5 Predict the output ŷ using 1-D GAT model; 2 Load the optimal parameters W ∗ ;
6 Calculate the probability of each label; 3 for data in testset do
7 Compute the loss L ; 4 Predict the output by the 1-GAT model;
8 Update the parameter vector W ; 5 Calculate the probability of each label;
9 end 6 end
10 end 7 Compute the accuracy and other metrics;
∗
11 return Optimized parameters W 8 return fault event, fault type, and fault location;
 Fault type: ABCG, R=0.1Ω, B7 Fault type: ABCG, R=0.1Ω, B7
4000 Zoom
4000 1 0 0

3000 3000

2000 2000
Vc
Vb 8 0
1000 1000

Voltage ABC (V)

Voltage ABC (V)

A c c u r a c y (% )
Va

0 0

1000 1000
6 0 M L P
2000 2000
C N N
G C N
3000 3000
G A T
1 -D G A T
4000 4000 4 0
0 200 400 600 800 1000 400 405 410 415 420 0 1 0 2 0 3 0 4 0 5 0
Time (ms) Time (ms)
E p o c h s
Fig. 5. The three-phase voltage data in the IEEE 123-node Fig. 6. The training accuracy curves with machine learning models

The hyperparameters used in the model’s training process, C. Comparative Results

displayed in Table III , provide insights into the configuration’s
impact on performance. The selection of a batch size is 32 The effectiveness of machine learning models by consider-
based on the consideration of the number of nodes within ing 4 metrics: Accuracy, Recall, F1-score, Precision. The com-
each graph. The training epochs for fault event detection, parative performance metrics of five machine learning models
classification, and localization are respectively 50, 130, and on two different systems, the Potsdam 13-node microgrid and
150 based on the distinct requirements of each task and the the IEEE 123-node system, are presented in Tables IV and V.
number of output labels. To address the issue of overfitting Fault event detection: It can be seen from both tables
the model, a random dropout rate of 10% is applied in the that the proposed model outperforms the others (MLP, CNN,
hidden layers during the training process. During the training GCN, and GAT) in terms of all metrics for fault detection.
process, the learning rate is initialized at 0.02 and subsequently Specifically, the proposed model achieves 99.56% and 99.61%
change to 0.001. The training performance of five neural accuracy, 99.18% and 99.76% recall, 99.55% and 99.61%
network models for fault event detection on the IEEE 123- F1-score, and 99.92% and 99.46% precision, indicating its
node system are illustrated in Fig. 6, with our proposed method superior ability to correctly identify the presence of faults,
exhibiting convergence in accuracy around 99.5% after 50 respectively. Alternative neural network architectures achieve
training epochs. Each training takes approximately 30 minutes. accuracies ranging from approximately 97.72% to 98.85%.
The models are built in the Pytorch environment and trained Fault type classification: The proposed model again demon-
on a single Intel Core i7-10700 CPU Processor (2.9 GHz and strates its superiority with the highest scores across all metrics
16 GB RAM). on both systems. On the Potsdam 13-node microgrid, the
proposed model scores an accuracy of 96.97% and a precision
of 97.01%, surpassing the other models. Similarly, on the
TABLE II IEEE 123-node system, it achieves an accuracy of 96.64%
DATASET I NFORMATION FOR T RAINING P ROCESS and a precision of 96.83%. This indicates that the proposed
model is not only adept at detecting the presence of faults but
Elements Value Number
also accurately classifying their types, which is essential for
AG, BG, CG, AB,
implementing appropriate corrective measures.
Fault type BC, CA, ABG, BCG, 11
CAG, ABC, ABCG
Fault resistance 0.1, 1.0, 10 (Ω) 3
TABLE III
Potsdam 13-node system H YPERPARAMETER C ONFIGURATION
Fault position Buses: 1, 2, 3, 4, 5, 6, 13
7, 8, 9, 10, 11, 12, 13 Hyperparameters Value
Load scenarios randomly 50 Batch Size 32
Total fault cases: 64,350 | Train set: 55,000 | Test set: 9,350 Epochs for Detection 50
Total non-fault cases: 10,000 | Train set: 8,000 | Test set: 2,000 Epochs for Classification 130
IEEE 123-node system Epochs for Localization 150
Fault position Buses: 7, 18, 25, 51, 10 Hidden Layer Activation ReLU
53, 62, 80, 89, 97, 101 Output Layer Activation Sigmoid or Softmax
Load scenarios randomly 150 Optimizer SGD
Total fault cases: 16,500 | Train set: 13,200 | Test set: 3,300 Initial Learning Rate (α) 0.02
Total non-fault cases: 4,125 | Train set: 3,300 | Test set: 825 Dropout Rate 0.1
 AG 858 0 0 0 0 0 0 0 0 0 0 B1 734 3 0 0 0 0 0 0 0 2 0 0 0
BG 0 820 0 0 1 0 1 0 0 0 0 B2 0 685 0 1 0 0 0 0 0 0 0 0 1
CG 0 0 856 0 2 1 0 0 1 0 0 B3 3 2 714 2 1 0 0 0 0 0 0 0 0
B4 0 0 0 626141 4 0 0 0 0 0 0 0
AB 0 0 0 840 0 0 1 0 0 0 1 0 0 0 135584 2 0 0 0 0 0 0 0
B5
BC 0 0 0 0 828 0 0 0 0 0 0 B6 0 0 0 0 0 664 3 0 0 0 0 0 0
true label

true label
CA 0 0 0 0 1 867 0 0 0 0 0 B7 0 0 0 0 0 2 698 6 0 0 0 0 0
ABC 0 0 0 0 0 0 751 0 0 0 125 B8 0 0 0 0 0 0 5 717 0 0 0 0 0
ABG 0 0 0 0 0 0 0 837 0 0 0 B9 2 0 0 0 0 0 0 0 730 4 2 0 0
BCG 0 0 0 0 0 0 0 0 832 0 0 B10 0 1 0 0 0 0 0 3 1 739 0 0 0
0 0 0 0 0 0 0 0 0 892 0 B11 0 0 0 0 0 0 0 0 0 0 70313 3
CAG 0 0 0 0 0 0 0 0 0 0 10708 6
B12
ABCG 0 0 0 0 0 0 152 0 0 0 686 B13 0 2 0 0 0 0 0 0 0 1 1 12675
CG
AB

CA
AG
BG

BC

C
G
G
AB G
CG
AB

B1
B2
B3
B4
B5
B6
B7
B8
B19
B10
B11
B12
3
AB

CA
BC

B
predicted label predicted label
Fig. 7. The fault classification confusion matrix for Potsdam microgrid Fig. 9. The fault location confusion matrix on Potsdam microgrid

AG 268 0 0 0 0 0 0 0 0 0 0 B7 331 0 0 0 0 0 0 0 0 0
BG 15 303 0 0 0 0 0 0 0 0 0 B18 1 315 0 1 0 0 0 0 0 0
CG 4 0 296 0 2 1 0 0 0 2 0 B25 3 1 343 0 0 0 0 0 0 0
AB 7 0 4 302 0 0 0 0 0 0 0 B51 3 3 0 338 0 0 0 1 0 2
BC 0 0 0 0 278 0 0 0 0 0 0

true label
B53 2 0 0 0 297 0 0 0 0 0
true label

CA 4 0 0 0 3 296 0 0 0 0 0
B62 1 0 0 0 2 337 0 0 0 0
ABC 0 0 0 0 0 0 276 0 3 1 10
B80 0 0 0 0 0 0 319 1 0 0
ABG 11 1 0 0 0 0 0 274 0 0 0
BCG 7 0 0 0 1 0 0 0 320 0 0 B89 0 0 0 0 0 0 0 323 0 0
CAG 12 0 0 0 0 0 1 0 1 297 0 B97 0 0 0 1 0 0 0 0 316 17
ABCG 0 1 0 0 0 0 22 0 0 0 278 B101 0 0 0 1 0 0 1 0 24 316

B7
8
5
1
3
2
0
9
7
01
CG
AB

CA
AG
BG

BC

C
G
G
AB G
CG

B1
B2
B5
B5
B6
B8
B8
B9
AB
AB

CA
BC

B1
predicted label predicted label

Fig. 8. The fault classification confusion matrix for IEEE 123-node system Fig. 10. The fault location confusion matrix on IEEE 123-node system

Fault location: We can see the proposed model maintains its Similarly, confusion matrices of 1-D GAT for fault location
lead, especially noteworthy on the IEEE 123-node system with are shown in Figs. 9 and 10. From confusion matrices, it
an accuracy of 98.03% and a precision of 98.06%. While the can be seen that the diagonal elements are substantially high,
performance gap between the proposed model and others like implying good performance for fault location. The proposed
the GAT and GCN is narrower in fault location compared to model can localize the fault positions with 13 buses and 10
detection and classification, the proposed model still exhibits buses for Potsdam microgrid and IEEE 123 bus, respectively.
the best overall capability in pinpointing the exact location E. Impact of Measurement Noises
of faults within the systems. This precision in fault location
is crucial for quick responses and minimizing downtime in To evaluate the effects of noises, we add the noises into the
electrical grids. voltage and branch current measurements before training and
testing. This noise is characterized by a normal distribution
with a mean of zero, with noise ratios of 3%, 6%, and 10%.
D. Confusion Matrices for Classification and Location
The impact of additional noises on the accuracies of both the
The confusion matrices of 1-D GAT for fault type clas- Potsdam 13-node and the IEEE 123-node systems is presented
sification in the Potsdam 13-node and the IEEE 123-node in Table VI. The results indicate that low noise levels, such
systems are illustrated in Figs. 7 and 8, respectively. The diag- as 3% and 6%, have a small impact on the performance, with
onal elements represent correct classifications, while the off- accuracy decreases from 0.5% to 2% for each fault scenario.
diagonal elements indicate mis-classifications between certain However, under 10% of noise, the accuracies drop noticeably.
fault types. These matrices reveals that the proposed model Under three noise scenarios in the IEEE 123-node system,
demonstrated a high degree of accuracy in classifying and the performance of GAT and 1-D convolution GAT is com-
predicting the nature of the fault data. However, it is hard for pared in Table VII. The proposed method outperforms GAT
the model to distinguish the ABC and ABCG faults because by 1.3% to 3% as noise increases from 3% to 10%. The
two fault types have a similar transient behavior. To this end, proposed 1-D convolution GAT approach is more robust and
the proposed models can learn and correctly classify all the noise-resilient than the traditional GAT method, particularly
unseen remaining fault types. in scenarios with higher levels of noise or disturbances in the
 TABLE IV
FAULT DETECTION METRICS ON P OTSDAM 13- NODE MICROGRID

Fault event Fault type Fault location

Method Accuracy Recall F1-score Precision Accuracy Recall F1-score Precision Accuracy Recall F1-score Precision
MLP 98.77 97.72 98.76 99.82 88.89 89.17 85.85 84.16 86.03 85.94 85.05 86.61
CNN 98.50 97.59 98.48 99.38 92.26 92.14 90.54 94.49 90.28 90.52 90.43 90.69
GCN 99.11 98.42 99.10 99.78 94.86 94.85 94.87 95.01 91.36 91.39 91.41 91.45
GAT 99.47 99.00 99.46 99.93 95.54 95.57 95.58 95.58 92.10 92.07 92.05 92.09
Proposed 99.56 99.18 99.55 99.92 96.97 96.98 96.99 97.01 96.01 96.13 96.11 96.10

TABLE V
FAULT DETECTION METRICS ON IEEE 123- NODE SYSTEM

Fault event Fault type Fault location

Method Accuracy Recall F1-score Precision Accuracy Recall F1-score Precision Accuracy Recall F1-score Precision
MLP 97.94 98.52 97.95 97.94 94.06 94.01 94.00 94.16 93.21 93.35 93.26 93.41
CNN 98.02 98.93 98.03 97.14 94.00 94.02 93.97 93.98 94.42 94.48 94.48 94.50
GCN 98.72 98.94 98.73 98.52 94.06 94.09 94.02 94.16 96.76 96.88 96.81 96.94
GAT 98.85 98.39 98.84 99.30 95.39 95.40 95.37 95.65 96.97 97.02 97.02 97.04
Proposed 99.61 99.76 99.61 99.46 96.64 96.69 96.63 96.83 98.03 98.07 98.06 98.06

TABLE VI the branch currents as input features, enhance the efficiency

T HE I MPACT OF N OISES ON 1-D GAT P ERFORMANCE to detect, classify, and localize the faults. By combining 1-D
convolutional layers and graph attention layers, our proposed
Fault event Fault type Fault location
framework effectively captures spatial-temporal features from
Noises PD13 123bus PD13 123bus PD13 123bus voltage and current data. The 1-D convolutional layers focus
No noises 99.56 99.61 96.97 96.64 96.01 98.03 on temporal features, while the graph attention layers employ
3% 99.29 99.54 95.78 95.97 95.49 97.52 node attention mechanism and edge attribute aggregation to
6% 99.21 99.47 94.95 95.42 92.66 95.63 exploit the spatial features. This helps the proposed model to
10% 98.81 98.41 92.08 93.69 90.34 90.36 incorporate the edge information into feature representations,
enhancing the model’s robustness. Experimental results on
both the IEEE 123-node and Potsdam 13-node systems demon-
TABLE VII
N OISE S CENARIO P ERFORMANCE C OMPARISON IN IEEE 123- NODE strate high accuracy in fault detection, classification, and
localization. Compared to bench-mark metrics, the proposed
Fault event Fault type Fault location model achieves considerably better performance, even in noise
scenarios.
Noises GAT Proposed GAT Proposed GAT Proposed
No noises 98.85 99.61 95.39 96.64 96.97 98.03
3% 98.68 99.54 94.33 95.97 95.42 97.52 R EFERENCES
6% 97.93 99.47 93.48 95.42 92.76 95.63
10% 96.77 98.41 92.28 93.69 88.35 90.36 [1] EIA, Annual Energy Outlook 2022 , Report. Addison-Wesley Profes-
sional, 2022.
[2] S. S. Gururajapathy, H. Mokhlis, and H. A. Illias, “Fault location and
detection techniques in power distribution systems with distributed gen-
data. The distinction lies in the proposed model’s integration eration: A review,” Renewable and sustainable energy reviews, vol. 74,
pp. 949–958, 2017.
of 1-D convolutional layers alongside the GAT architecture, [3] L. Vu, T. Vu, T. L. Vu, and A. Srivastava, “Multi-agent deep reinforce-
enhancing its ability to handle noise-induced disruptions more ment learning for distributed load restoration,” IEEE Transactions on
effectively compared to the traditional GAT method. However, Smart Grid, vol. 15, no. 2, pp. 1749–1760, 2024.
scalability and potential voltage measurement losses in certain [4] K. Chen, C. Huang, and J. He, “Fault detection, classification and
location for transmission lines and distribution systems: a review on
areas remain as limitations to address. the methods,” High voltage, vol. 1, no. 1, pp. 25–33, 2016.
[5] P. K. Ganivada and P. Jena, “A fault location identification technique for
IV. C ONCLUSIONS active distribution system,” IEEE Transactions on Industrial Informatics,
This paper presents a novel approach employing a deep vol. 18, no. 5, pp. 3000–3010, 2022.
[6] M. M. Zaben, M. Y. Worku, M. A. Hassan, and M. A. Abido, “Machine
graph attention network structure for fault diagnosis in dis- learning methods for fault diagnosis in ac microgrids: A systematic
tribution systems. The main contribution is the integration of review,” IEEE Access, vol. 12, pp. 20 260–20 298, 2024.
 [7] A. Zidan, M. Khairalla, A. M. Abdrabou, T. Khalifa, K. Shaban, [27] Y. Aslan and Y. E. Yağan, “Artificial neural-network-based fault location
A. Abdrabou, R. El Shatshat, and A. M. Gaouda, “Fault detection, for power distribution lines using the frequency spectra of fault data,”
isolation, and service restoration in distribution systems: State-of-the- Electrical Engineering, vol. 99, pp. 301–311, 2017.
art and future trends,” IEEE Transactions on Smart Grid, vol. 8, no. 5, [28] J. B. Thomas, S. G. Chaudhari, S. K. V., and N. K. Verma, “Cnn-based
pp. 2170–2185, 2017. transformer model for fault detection in power system networks,” IEEE
[8] H. Jiang, J. J. Zhang, W. Gao, and Z. Wu, “Fault detection, identification, Transactions on Instrumentation and Measurement, vol. 72, pp. 1–10,
and location in smart grid based on data-driven computational methods,” 2023.
IEEE Transactions on Smart Grid, vol. 5, no. 6, pp. 2947–2956, 2014. [29] T. Li, Z. Zhao, C. Sun, R. Yan, and X. Chen, “Multireceptive field graph
[9] V. Rizeakos, A. Bachoumis, N. Andriopoulos, M. Birbas, and A. Birbas, convolutional networks for machine fault diagnosis,” IEEE Transactions
“Deep learning-based application for fault location identification and on Industrial Electronics, vol. 68, 2020.
type classification in active distribution grids,” Applied Energy, 2023. [30] R. A. Jacob, S. Senemmar, and J. Zhang, “Fault diagnostics in shipboard
[10] H. Mirshekali, R. Dashti, A. Keshavarz, A. J. Torabi, and H. R. Shaker, power systems using graph neural networks,” in 2021 IEEE 13th
“A novel fault location methodology for smart distribution networks,” International Symposium on Diagnostics for Electrical Machines, Power
IEEE Transactions on Smart Grid, vol. 12, no. 2, pp. 1277–1288, 2021. Electronics and Drives (SDEMPED), vol. 1. IEEE, 2021, pp. 316–321.
[11] W. Deng, K. T. Nguyen, K. Medjaher, C. Gogu, and J. Morio, “Rotor [31] X. Zhao, M. Jia, and Z. Liu, “Semisupervised graph convolution deep
dynamics informed deep learning for detection, identification, and lo- belief network for fault diagnosis of electormechanical system with
calization of shaft crack and unbalance defects,” Advanced Engineering limited labeled data,” IEEE Transactions on Industrial Informatics,
Informatics, vol. 58, p. 102128, 2023. vol. 17, no. 8, pp. 5450–5460, 2021.
[12] S. Shi, B. Zhu, A. Lei, and X. Dong, “Fault location for radial [32] J. G. Fornás, E. H. Jaraba, A. L. Estopiñan, and J. Saldana, “Detec-
distribution network via topology and reclosure-generating traveling tion and classification of fault types in distribution lines by applying
waves,” IEEE Transactions on Smart Grid, vol. 10, 2019. contrastive learning to gan encoded time-series of pulse reflectometry
[13] F. Liu, L. Xie, K. Yu, Y. Wang, X. Zeng, L. Bi, and X. Tang, “A signals,” IEEE Access, vol. 10, pp. 110 521–110 536, 2022.
novel fault location method based on traveling wave for multi-branch [33] K. Chen, J. Hu, Y. Zhang, Z. Yu, and J. He, “Fault location in power
distribution network,” Electric Power Systems Research, vol. 224, p. distribution systems via deep graph convolutional networks,” IEEE
109753, 2023. Journal on Selected Areas in Communications, pp. 119–131, 2019.
[34] J. Hu, W. Hu, J. Chen, D. Cao, Z. Zhang, Z. Liu, Z. Chen, and
[14] J. Tavoosi, M. Shirkhani, A. Azizi, S. U. Din, A. Mohammadzadeh, and
F. Blaabjerg, “Fault location and classification for distribution systems
S. Mobayen, “A hybrid approach for fault location in power distributed
based on deep graph learning methods,” Journal of Modern Power
networks: Impedance-based and machine learning technique,” Electric
Systems and Clean Energy, vol. 11, no. 1, pp. 35–51, 2022.
Power Systems Research, vol. 210, p. 108073, 2022.
[35] J. T. de Freitas and F. G. F. Coelho, “Fault localization method for power
[15] S. Turizo, G. Ramos, and D. Celeita, “Voltage sags characterization distribution systems based on gated graph neural networks,” Electrical
using fault analysis and deep convolutional neural networks,” IEEE Engineering, vol. 103, no. 5, pp. 2259–2266, 2021.
Transactions on Industry Applications, vol. 58, no. 3, 2022. [36] B. L. H. Nguyen, T. V. Vu, T.-T. Nguyen, M. Panwar, and R. Hovsapian,
[16] I. Sadeghkhani, M. E. Hamedani Golshan, A. Mehrizi-Sani, J. M. “Spatial-temporal recurrent graph neural networks for fault diagnostics
Guerrero, and A. Ketabi, “Transient monitoring function–based fault in power distribution systems,” IEEE Access, vol. 11, pp. 46 039–46 050,
detection for inverter-interfaced microgrids,” IEEE Transactions on 2023.
Smart Grid, vol. 9, no. 3, pp. 2097–2107, 2018. [37] Z. Chen, J. Xu, T. Peng, and C. Yang, “Graph convolutional network-
[17] T. Cui, X. Dong, Z. Bo, and A. Juszczyk, “Hilbert-transform-based based method for fault diagnosis using a hybrid of measurement and
transient/intermittent earth fault detection in noneffectively grounded prior knowledge,” IEEE transactions on cybernetics, vol. 52, no. 9, pp.
distribution systems,” IEEE Transactions on Power Delivery, vol. 26, 9157–9169, 2021.
no. 1, pp. 143–151, 2011. [38] D. Chanda and N. Y. Soltani, “A heterogeneous graph-based multi-
[18] L. Cui, Y. Liu, L. Wang, J. Chen, and X. Zhang, “High-impedance task learning for fault event diagnosis in smart grid,” arXiv preprint
fault detection method based on sparse data divergence discrimination arXiv:2309.09921, 2023.
in distribution networks,” Electric Power Systems Research, vol. 223, p. [39] H. Tong, R. C. Qiu, D. Zhang, H. Yang, Q. Ding, and X. Shi, “Detection
109514, 2023. and classification of transmission line transient faults based on graph
[19] Z. Yao, Y. Liu, J. Chen, J. Ji, M. Zhang, and Y. Gong, “Active high- convolutional neural network,” CSEE Journal of Power and Energy
impedance fault detection method for resonant grounding distribution Systems, vol. 7, no. 3, pp. 456–471, 2021.
networks,” IEEE Access, vol. 12, pp. 10 932–10 945, 2024. [40] W. Li and D. Deka, “Ppgn: Physics-preserved graph networks for real-
[20] M. Alqudah, M. Pavlovski, T. Dokic, M. Kezunovic, Y. Hu, and time fault location in distribution systems with limited observation and
Z. Obradovic, “Fault detection utilizing convolution neural network on labels,” arXiv preprint arXiv:2107.02275, 2021.
timeseries synchrophasor data from phasor measurement units,” IEEE [41] P. Velickovic, G. Cucurull, A. Casanova, and A. Romero, “Pietrolio
Transactions on Power Systems, vol. 37, no. 5, pp. 3434–3442, 2022. and yoshua bengjio. 2018. graph attention networks,” in International
[21] W. Li, D. Deka, M. Chertkov, and M. Wang, “Real-time faulted line Conference on Learning Representations, ICLR, 2018.
localization and pmu placement in power systems through convolutional [42] M. N. I. Siddique, M. Shafiullah, S. Mekhilef, H. Pota, and M. Abido,
neural networks,” IEEE Transactions on Power Systems, vol. 34, no. 6, “Fault classification and location of a pmu-equipped active distribution
pp. 4640–4651, 2019. network using deep convolution neural network (cnn),” Electric Power
[22] Q. Li, H. Luo, H. Cheng, Y. Deng, W. Sun, W. Li, and Z. Liu, “In- Systems Research, vol. 229, p. 110178, 2024.
cipient fault detection in power distribution system: A time–frequency [43] C. Zhang, J. J. Q. Yu, and Y. Liu, “Spatial-temporal graph attention
embedded deep-learning-based approach,” IEEE Transactions on Instru- networks: A deep learning approach for traffic forecasting,” IEEE
mentation and Measurement, vol. 72, pp. 1–14, 2023. Access, vol. 7, pp. 166 246–166 256, 2019.
[23] A. Srivastava and S. Parida, “A robust fault detection and location [44] Q.-H. Ngo, B. L. Nguyen, T. V. Vu, J. Zhang, and T. Ngo, “Physics-
prediction module using support vector machine and gaussian process informed graphical neural network for power system state estimation,”
regression for ac microgrid,” IEEE Transactions on Industry Applica- Applied Energy, vol. 358, p. 122602, 2024.
tions, vol. 58, no. 1, pp. 930–939, 2022. [45] B. L. Nguyen, T. V. Vu, J. M. Guerrero, M. Steurer, K. Schoder, and
[24] Y. Seyedi, H. Karimi, and J. Mahseredjian, “A data-driven method for T. Ngo, “Distributed dynamic state-input estimation for power networks
prediction of post-fault voltage stability in hybrid ac/dc microgrids,” of microgrids and active distribution systems with unknown inputs,”
IEEE Transactions on Power Systems, vol. 37, no. 5, 2022. Electric Power Systems Research, vol. 201, p. 107510, 2021.
[25] C. F. Mbey, V. J. Foba Kakeu, A. T. Boum, and F. G. Y. Souhe, “Fault [46] K. P. Schneider, B. A. Mather, B. C. Pal, C.-W. Ten, G. J. Shirek, H. Zhu,
detection and classification using deep learning method and neuro-fuzzy J. C. Fuller, J. L. R. Pereira, L. F. Ochoa, L. R. de Araujo, R. C. Dugan,
algorithm in a smart distribution grid,” The Journal of Engineering, vol. S. Matthias, S. Paudyal, T. E. McDermott, and W. Kersting, “Analytic
2023, no. 11, p. e12324, 2023. considerations and design basis for the ieee distribution test feeders,”
[26] S. Baloch, S. S. Samsani, and M. S. Muhammad, “Fault protection in IEEE Transactions on Power Systems, vol. 33, no. 3, 2018.
microgrid using wavelet multiresolution analysis and data mining,” IEEE
Access, vol. 9, pp. 86 382–86 391, 2021.