612 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 9, NO.

2, MARCH 2023

Fast Single-phase Fault Location Method Based on

Community Graph Depth-first Traversal for
Distribution Network
Jian Dang, Member, IEEE, Member, CSEE, Yunjiang Yan, Rong Jia, Member, IEEE, Xiaowei Wang,
Member, IEEE, and Hui Wei

Abstract—With the increasing complexity of distribution net- and improving the reliability of power supply [2]. Single-
work structures originating from the high penetration of re- phase ground fault accounts for more than 70% of all faults
newable energy and responsive loads, fast and accurate fault in the distribution network and is the most common fault in
location technology for distribution networks is a prerequisite
for rapid isolation of faults and restoration of the power supply. the distribution network [3]. Impedance-based methods [4],
In this paper, a fault location method based on community graph [5], traveling wave-based methods [6], [7], and S-injection
depth-first traversal is proposed for fast location of single-phase methods [8] are commonly used in fault location research.
ground faults in distribution networks. First, this paper defines Impedance-based methods are widely used in fault location
the fault graph weight of the vertices in the distribution network research because they are easy to implement. Reference [4]
graph model, which can be used to reflect the topology of the
vertices and fault points as well as the fluctuation of the vertices’ uses an impedance-based method to screen out all possible
currents. Then, the vertices on the graph model are clustered fault locations, and determines the actual fault point by
by using an improved parallel louvain method (IPLM). Finally, simulating the same fault. Reference [5] uses an algorithm
the community formed by IPLM is used as the smallest unit for iterative state estimation to overcome measurement data
for depth-first traversal to achieve fast and accurate location inaccuracies. The accuracy of the impedance-based methods
of the fault section. The paper develops a distribution network
graph model of IEEE 33-bus system on the graph database for is susceptible to a variety of factors such as fault resistance,
testing. And three other methods are selected for comparison distributed power access, etc. Compared with the impedance-
with IPLMDF. The test results show that IPLMDF can achieve based methods, the traveling wave-based methods are not
fast and accurate fault location when half of the nodes in the affected by the fault resistance and system operation mode,
distribution network are equipped with D-PMUs. When some of but are more sensitive to line branches. Reference [6] proposes
the D-PMUs lose time synchronization, it is still possible to locate
the fault section, and at the same time, the locating results can a method based on traveling wave arrival computational fault
be avoided by falling into local optimal solutions. time difference (CFTD), which is not affected by traveling
wave speed inaccuracies. In reference [7], the PCA-SVM
Index Terms—Depth-first traversal, fault location, graph model is constructed to extract the curve cluster characteristics
computation, single-phase fault. of the fault current traveling wave. In addition to the methods
based on impedance and traveling wave, an active method is
proposed in reference [8]. For a resonant grounded system,
whose neutral point grounds are via an electromagnetic hybrid
I. I NTRODUCTION Petersen coil (EHPC), the identification basis of fault feeder
and fault section can be provided by adjusting the EHPC
N OWADAYS, with the growing load and the penetration
of large amounts of renewable energy, the challenge
of power supply reliability in the distribution network is
complex impedance [8]. This method requires supplementary
devices and extra operations, which increase potential risks
becoming significantly important. It is crucial for electric and costs [9]. Although the above three types of methods have
utilities to minimize outage time [1]. Accuracy and rapidity been improved, they are all constrained to varying degrees
have become two major metrics in fault location research. by the imperfection of the distribution network monitoring
Therefore, a fast and accurate fault location method is a pre- system.
requisite for repairing faults, rapidly restoring power supply, By virtue of the synchronized, low-latency and high-resolu-
tion measurement data, the application of the distribution-level
Manuscript received August 31, 2020; revised December 15, 2020; accepted phasor measurement unit (D-PMU) in distribution networks
March 18, 2021. Date of online publication July 9, 2021; date of current continues to increase, and many scholars have carried out
version July 23, 2022. This work was supported by the National Natural
Science Foundation of China (Grant Nos. 52009106, 51779206), the National research on distribution network fault location based on D-
Key R&D Program of China (No. 2018YFB1500800), and the Natural Science PMU [9]–[14]. Reference [9] identifies the fault section by
Fund Youth Project of Shaanxi Province (2019J-130). calculating the transient energy of the zero-sequence current
J. Dang, Y. Yan, R. Jia (corresponding author, email: [email protected]),
X. Wang and H. Wei are with Xi’an Key Laboratory of Intelligent Energy, in selected frequency bands for each section. Reference [10]
Xi’an University of Technology, Shaanxi 710048, China. proposes a method to achieve fault location without line pa-
DOI: 10.17775/CSEEJPES.2020.04650 rameters, which solves the effect of inaccurate line parameters
2096-0042 © 2020 CSEE
DANG et al.: FAST SINGLE-PHASE FAULT LOCATION METHOD BASED ON COMMUNITY GRAPH DEPTH-FIRST TRAVERSAL FOR DISTRIBUTION NETWORK 613

in distribution networks. However, this method cannot be vertices and the relationship between data can be represented
applied to single-phase ground faults in neutral ungrounded by edges. The edges can send computing messages between
systems. There are also neural network-based fault location the connected vertices [15].
methods [11], [12]. Reference [11] proposes an adaptive The grid is a natural graph structure. The devices in the grid
convolutional neural network-based method, but this type of can be abstracted as vertices. The information of devices can
method usually faces the problem of insufficient sample data. be stored in the attributions of the vertex. Transmission lines
References [9]–[11] focus too much on the study of D- can be described as edges, and even the various relationships
PMU measurement information and do not make full use of between devices can be abstracted as edges. When the data
the topology information of the distribution network. Refer- of grid is stored in the form of graph data, the topological
ence [12] proposes a fault location method based on a graph structure of the power system can be preserved and easily
convolutional neural network, which incorporates the network expanded [16].
topology and is more adaptable to changes in the network Graph computation is a type of node-based parallel com-
topology. Reference [13] proposes a minimum spanning tree- puting, which means each vertex has its own calculation
based method, which identifies the fault section based on logic and all calculations on each vertex are independent
the network topology and the phase of zero sequence cur- to other vertices [17]. Pregel, developed by Google, is the
rent. However, the methods in references [12], [13] need to earliest graph computation system. Pregel proposed a vertex-
configure D-PMUs on each node in the network to ensure centric programming model and its calculation mode is a bulk
the accuracy of the localization results, which reduces the synchronous parallel computing model (BSP).
feasibility of these methods. Reference [14] proposes a fault Graph computation can be applied to many aspects within
location method with only installed D-PMU at the head and the power system. Reference [18] realizes power flow calcula-
end of the feeder, but this method cannot be used to locate a tion of the power system by graph computation. Reference [19]
single-phase ground fault. uses graph computation to do the real-time power system net-
Power network topology is a type of natural graph struc- work topology analysis. They use the idea of graph partition-
ture, yet this feature is underutilized in existing studies. ing, and propose the parallel topology analysis algorithm based
Graph databases and graph computation have good application on the graph computing framework. Reference [20] presents a
prospects in mining the value of network topology informa- graph-computation based EMS (Energy Management System,
tion, and currently there is no fault location method based on EMS) which takes advantage of both nodal and hierarchical
graph computation. To address the above problems, this paper parallelism to achieve high-speed graph-based state estimation,
proposes a fast single-phase fault location method based on power flow, and contingency analysis applications. In this
community graph depth-first traversal. The community graph paper, we take graph computation to calculate the voltage and
formed by the improved parallel louvain method ensures that current of bus without D-PMU installed.
the faulted section can be identified in 2–3 queries, regardless
of where in the network the fault occurs. The proposed method B. Distribution Network with D-PMU Graph Modeling
can rapidly and accurately locate the fault section in the case In the graph database, Schema specifies the types and
of high resistance grounding. Moreover, the proposed method attributes of the vertices and edges of the graph data, as well
only requires D-PMUs to be configured on half of the nodes as the relationship between various vertices. In order to fully
in the distribution network model. And it can still rapidly and describe the state of the distribution network in the event of
accurately locate the fault section when some of the D-PMUs a fault, this paper establishes a Schema as shown in Fig. 1
are disturbed and lose their time synchronization. on the graph database TigerGraph, which contains different
The rest of this paper is organized as follows. The graph types of vertices and edges. Each vertex contains six attributes:
database and graph computation are introduced and the dis- ID, upstream vertex, downstream vertex, fault graph weight,
tribution network graph modeling is described in Section II. D-PMU state, and community ID. Each edge contains two
The definition of fault graph weight is given and a depth-first attributes: NAME and fault graph weight. Among the above
traversal based on fault graph weight is proposed in Section III. attributes, the role of ID is to ensure the uniqueness of each
The parallel louvain method is introduced and the community vertex. The role of upstream and downstream nodes is to
graph depth-first traversal is proposed in Section IV. The determine the location of the vertices on the network topology.
accuracy, robustness and calculation speed of the proposed The fault graph weight is used to describe the strength of the
method are verified on the IEEE 33-bus system graph model fault on the nodes and lines. The definition of the fault graph
in Section V. Conclusions are provided in Section VI. weight will be introduced in the next section. The D-PMU
state is a two-dimensional array of two elements with 0 or
1 to describe whether a node is configured with D-PMU and
II. D ISTRIBUTION N ETWORK G RAPH M ODELING
whether D-PMU is synchronized. The NAME of each edge is
A. Graph Database and Graph Computation named as “Upstream Vertex ID - Downstream Vertex ID.”
A graph databases has emerged in recent years. It is one type Based on the of set attributes, the relevant information in
of non-relational database that can be used to process massive the distribution model is converted into data text for vertices
data in parallel. A graph database is formed based on the and edges, respectively. By loading them into the Schema,
idea of graph theory. Vertices and edges are the two elements the corresponding distribution network graph model can be
making up the graph database. Data can be stored in the obtained. The detailed flow is shown in Fig. 1.
614 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 9, NO. 2, MARCH 2023

Z25 Z5

Z12

1 2 Z24 5 Verrex Data

P+jQ P+jQ
Z13

4
3
P+jQ Z4 P+jQ
Z36 Edge Data

Z6

6 P+jQ Schema

2
Edge Attributes Data Type node
1 5 Name STRING Vertex Attributes Data Type
Fault graph weight FLOAT ID INT
3 Upstream vertex INT
Downstream vertex INT
6 Fault graph weight FLOAT
D-PMU status JSONARRAY
Community ID INT

Fig. 1. Distribution network diagram modeling process.

h j h i
In this paper, D-PMUs are configured as follows: no D-PMU
is configured at the terminal node, and only one of any two i
connected nodes in the distribution model is configured with
a D-PMU.S.ch a configuration principle ensures that nodes h
and j on both sides of each node i that is not configured (a) (b)
with D-PMU are configured with D-PMU. if the D-PMU
Fig. 2. Minimum monitoring unit.
configuration state attribute of vertex i is represented by ki , it
can be expressed as:
U̇h − U̇i
I˙i =
(
1 Node with D-PMU Configurated (4)
ki = (1) Zhi
0 Node without D-PMU Configurated Figure 2(b) corresponds to the case where node i is a branch
For each edge eij of the distribution network graph model, end node, and node i has no downstream node j. The voltage
the following equation is satisfied. information of node i is calculated from (3) to (5), and the
current information is calculated in the same way.
ki + kj = 1 (2)  
Ph + jQh
∗ 
U̇i = U̇h − Ih − Zhi (5)
Equation (2) ensures that each line in the distribution net- Uh
work contains a D-PMU, and the information of unconfigured In this paper, a modified IEEE 33-bus system graph model
D-PMU nodes can be calculated from two adjacent configured is built on the graph database as shown in Fig. 8, which
D-PMU nodes to satisfy the global observable condition. is more intuitive than using the adjacency matrix to reflect
Based on the above configuration principles, the distribution the connection between nodes, and the topology of the graph
model can be simplified to the set of minimum monitoring model is consistent with the actual topology to ensure that
units shown in Fig. 2. Node i is the unconfigured D-PMU the graph calculation can be performed directly on the actual
node, and node h and node j are the nodes upstream and topology of the grid. The green nodes are the installed D-PMU
downstream of node i that are configured with D-PMU, nodes and the black nodes are the unconfigured D-PMU nodes.
respectively.
Figure 2(a) corresponds to the case where node i is not a III. D EPTH - FIRST T RAVERSAL BASED ON FAULT
node at the end of the line, in which case the computation of G RAPH W EIGHT
node i information is divided into two parts: forward solving A. Definition of Fault Graph Weight
voltage and backward solving current, as shown in (3) and (4).
Single-phase ground fault is the most common fault type in
U̇i = U̇j + I˙j · Zij (3) distribution networks. When a single-phase ground fault oc-
DANG et al.: FAST SINGLE-PHASE FAULT LOCATION METHOD BASED ON COMMUNITY GRAPH DEPTH-FIRST TRAVERSAL FOR DISTRIBUTION NETWORK 615

curs in a small-current grounding system (neutral ungrounded Ia

system or resonant grounding system), the symmetry of the 500 Ib
three-phase line to ground is broken and the fault point forms Ic
a circuit only through the capacitance to ground. Taking the
450
neutral point ungrounded system as an example, when a single-
phase grounding fault occurs, the simplified zero-sequence

I (A)
network is as shown in Fig. 3. 400

Reference zero sequence current direction

350
Actual zero sequence current direction
Z
300

F 0 1000 2000 3000 4000 5000 6000

Time (ms)
(a)
Z Z
Ea Ia
14
Ib
Z Rf Z
Ic
12

10
Fig. 3. Simplify zero sequence network diagram.

I (A)
8
With the zero-sequence network shown in Fig. 3, it can be
concluded that the zero-sequence current of the node located
6
before the fault point on the fault branch lags the zero-
sequence voltage by 90◦ , and the zero-sequence current of
4
the other nodes leads the zero-sequence voltage by 90◦ [15].
Therefore, by judging the phase relationship between the zero- 0 1000 2000 3000 4000 5000 6000
sequence current and zero-sequence voltage of each node, the Time (ms)
(b)
position relationship between each node and the fault point
can be derived. Fig. 4. Three-phase current variation.
Let ∆θ be the phase difference between the zero-sequence
current and zero-sequence voltage. Consider the interference
v
u X It,i,k − I¯i,k 2
ut0 +T

of background noise, set ±10◦ as the reasonable deviation Si,k = t (8)
that may occur between ∆θ and the theoretical value when t=t
Tf − 1
0
rounding ∆θ. (7) is the rounding function of ∆θ.
In (8), k is the fault phase in the three phases, f is the
∆θ = θI0 − θU0 (6) adopted frequency of the D-PMU, T is the time window of
 0.05 s, t0 is the starting moment, and It,i,k is the amplitude
1 ∆θ ∈ (−80◦ , −100◦ )
of the fault phase current at node i at moment t.

σi (∆θ) = 0 ∆θ ∈ (−10◦ , 10◦ ) (7) The fault graph weight of node i contains not only the phase
−1 ∆θ ∈ (80◦ , 100◦ )

difference of the zero-sequence current and voltage, but also

the amplitude fluctuation of the fault phase current, which is
Set the line between nodes 11 and 12 in the IEEE 33-bus ′
denoted as Si,k .
system to have an a-phase short circuit to ground at 0.2 s
′
with a fault resistance of 10 Ω and excise the fault at 0.4 s. Si,k = Si,k · σi (∆θ) (9)
Fig. 4(a) and (b) show the variation of the three-phase currents
recorded by the D-PMU at node 1 and node 17 at the time B. Depth-first Traversal based on Fault Graph Weight
of the fault. After the occurrence of the fault, the three-phase Most of the existing studies on fault location in distribution
current at node 1 rapidly increases, with the fault phase current networks, whether based on transient or steady-state quanti-
significantly increasing. The non-fault phase current at node ties, rely on a large number of measurement points in the
17 rapidly increases, while its fault phase current rapidly drops network, lacking the use of network topology information and
and is close to zero. At the first node, the phase with the largest requiring high time synchronization of measurement points in
sudden change in current value is the fault phase, and the fault the network. The long power supply radius of the distribution
phase is recorded as k. network, many line branches and high background noise will
Set a time window T = 0.05 s and calculate the standard not only increase the cost but also make it difficult to guarantee
deviation of the current of the fault phase at each node within the time synchronization of measurement points in the network
a time window. when a large number of measurement points are arranged.
616 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 9, NO. 2, MARCH 2023

Under these conditions, the accuracy of the fault location will node becomes non-viewable, and node 8 and node 10 are only
be affected. viewable through the D-PMUs of nodes 7 and 11. The fault
To address the above issues, this paper proposes a depth- graph weight of node 9 can be estimated using the fault graph
first traversal based on fault map weights for fault location. weights of node 8 and node 10, which is estimated by taking
The fault graph weight, as an attribute of the vertices in the the average of the fault graph weights of neighboring nodes.
distribution network graph model, reflects both the fluctuation ′
X
′

of node currents after a fault and the relationship between Ss,k = AV G St,k est ∈ G(v, e) (12)
nodes and fault points in the topology. Depth-first traversal on Equation (12) is the formula of estimation of the fault graph
the distribution network graph model can make full use of the weight, where s represents the node whose fault graph weight
topology information of the network. is 0 due to the loss of synchronization of D-PMU, and node t
In this paper, starting from the first node, the branch where is the vertex adjacent to node s. The graph model built in this
′
the node Si,k > 0 is located is traversed in priority, and the paper is an undirected graph, so node t represents the upstream
′
traversal is stopped when a node j is visited with Sj,k < 0. and downstream nodes of node s.
The line connected by node j and its upstream node i is the Since nodes 8–11 are all nodes located in front of the fault
fault section. point, the estimated fault graph weight of node 9 will be
In the test, it was found that the fault graph weight of the numerically different from the original case, but the positive
node without D-PMU installed on the fault section was always and negative attributes remain the same and will not affect the
greater than 0. Therefore, when the last visited node i is an result of fault location. If the D-PMU of node 7 and node 11
uninstalled D-PMU node, further judgment is required using also lose their temporal synchronization, the estimated fault
(10). graph weights of node 6 to node 12 can only rely on the
′ ′ ′
Si,k information of node 5 and node 13. Since both S5,k and S13,k
δ= ′ (10) are greater than 0, the accuracy of the fault location results
Sh,k
can still be guaranteed.
In (10), node h is the upstream node of node i. When δ < When the D-PMU at node 13 also loses temporal synchro-
0.9, the fault section is between node h and node i, otherwise, nization, it is necessary to use the fault graph weight of node
the fault section is between node i and node j. 15 for estimation. In this case there will be S5,k ′
> 0 and
In the problem of time synchronization of measurement ′
S15,k < 0. The positivity and negativity of the fault map
units, each node in the distribution network graph model has weights of all nodes in the interval from node 6 to node 14
the D-PMU state attribute. This attribute is a custom two- are difficult to determine, and the fault location results will be
dimensional array [D-PMU configuration, D-PMU synchro- affected.
nization], and the various types of this variable are shown in Therefore, the minimum constraint for accurate fault lo-
Fig. 5. cation is that the temporal accuracy of the D-PMU at the
connecting fault branch on the non-fault branch, the D-PMU
[0,0] [0,1] at the first node on the fault branch, and the D-PMU at the
[1,1] [1,0] two nodes closest to the fault point must not fail.

Fig. 5. All cases of D-PMU status. IV. C OMMUNITY G RAPH D EPTH - FIRST T RAVERSAL
In distribution network fault location, real-time is an impor-
In the above figure, [0, 0] and [1, 1] are two normal states, tant requirement. Fig. 6 shows the fault section search process
there is no [0, 1] state because the node which is not configured based on depth-first traversal This method can locate the fault
with D-PMU will not have a D-PMU time synchroniza- section after a few traversal steps when the fault point is close
tion [1, 0] is the state where the node D-PMU loses time to the power node. If the fault point is near the end of the line
synchronization, and the judgment condition is as follows: and the fault branch is long, multiple traversals are required to
locate the fault section, which makes it difficult to guarantee

 ki = 1

i ̸= j the real-time nature of the algorithm.
σi (∆θ) = 0 (11)
 v i , vj ∈ G(v, e)
σj (∆θ) ̸= 0

f
From the previous analysis, we can conclude that the node  
located in front of the fault point on the fault branch has a fault 1 2 3 N-2 N-1 N
graph weight greater than 0, and the rest of the nodes have a
fault graph weight less than 0. When the time synchronization Fig. 6. Depth-first traversal query process.
of some measurement units fails, the fault location results will
not be affected as long as the positive and negative signs of To ensure fast location, this paper combines depth-first
the fault graph weights of all nodes in the network can be traversal with graph clustering. The louvain method is a
guaranteed to be unchanged. locally greedy graph clustering algorithm that divides the
Suppose that when a failure occurs in the network of Fig. 8, graph into multiple communities. Fault section search based on
after the D-PMU of node 9 loses its time synchronization, the a community graph ensures that fault sections can be located
DANG et al.: FAST SINGLE-PHASE FAULT LOCATION METHOD BASED ON COMMUNITY GRAPH DEPTH-FIRST TRAVERSAL FOR DISTRIBUTION NETWORK 617

by a few traversal steps even when the fault point is near the In references [22], parallel computing is added to the
end of the line. louvain method, and the parallel louvain method (PLM) is
proposed. PLM initially treats each node as an independent
A. Parallel Louvain Method community. In each iteration, all vertices in the network are
The Louvain mothed was proposed in 2008 by Blondel visited in parallel, the modularity gain ∆Q generated by
et al. of Louvain University to process graph data [21]. This moving each vertex to an adjacent community is calculated,
algorithm performs the exploration of community discovery on and the vertex is moved to the community with the largest ∆Q.
the graph model based on the gain of modularity, dividing the Each time a vertex is moved from its original community to
graph into several communities according to the edge weights, another community, it must be guaranteed that the modularity
and its objective function is to maximize the modularity of gain is positive, i.e., ∆Q > 0. Node-parallel computation
the whole network. Community is a type of subgraph formed improves the computational speed of the Louvain method in
after clustering. The modularity is a measure of the strength each iteration, but may produce negative modularity gains
of connections between nodes, usually denoted by the letter during vertex community attribution changes. This is due
Q, which is calculated as follows: to the fact that all vertices calculate the modularity gain of
1 X

ki kj
 moving themselves to an adjacent community in parallel, and
Q= Aij − δ (ci , cj ) (13) by the time a vertex is moved to an adjacent community, the
2m i,j 2m
structure of that community may have changed and the actual
Equation (13) is the formula of Modularity Q, where Aij modularity gain may be negative. The negative modularity
gain can be corrected by multiple iterations. The community
representsP the weight of the edge between node i and node j,
m=P 1 attribution of each vertex no longer changes after one iteration
2 ij Aij is half the sum of weights of all edges, and
kj = j Aij is the sum of the weights of the edges connected indicates that the vertices have been moved to the best com-
to node j, cj is the community to which vertex j is assigned, munity at this point and the first phase of PLM is complete.
when ci = cj , δ(ci , cj ) = 1 otherwise 0. The second phase of the Louvain’s algorithm is computed in
It is easy to see by the above equation that if all the vertices the same way as the first phase, but the object of computation
in the network are independent communities, the modularity is changed from the vertices in the network to the community.
of the whole network is 0. If the modularity of the whole Each community is considered as a “super node, “ and the
network is to be maximized, it is necessary to maximize Aij connection weights among the “super nodes” are calculated,
within the same community in (13). and the first stage of PLM is executed with the “super node”
If the weight of an edge is viewed as the closeness of the as the smallest unit.
connection between two vertices, then this process can be
understood from the perspective of the null domain as each B. Community Graph Depth-first Traversal
vertex tends to be assigned to the community where the vertex In the graph model of this paper, the weight wij of each
with which it is most closely connected is located. The core of edge is the fault graph weight of vertex j downstream of this
Louvain method is based on this process, so that the vertices edge, i.e.,
with strong ties are in the same community, and the weak ′
ties between vertices are diluted to achieve the clustering of wij = Sj,k ∀eij ∈ G(v, e) (16)
complex networks.
In this paper, we run PLM on the distribution network
The Louvain method is divided into two phases, in each of
diagram model to obtain the community graph. Then the
which the modularity gain resulting from the vertex movement
depth-first traversal is performed based on the community
is calculated iteratively. The first stage of Louvain method cal-
graph.
culates the modularity gain due to individual node movement,
In this paper, we first visit the first vertex of the network
and the second stage calculates the modularity gain due to
and record the ID of the community where the first vertex is
community merging.
"P located, noted as f id. Then, the most downstream vertex in
2 #
the community with community ID fid that has a fault graph
P
in +2ki,in tot +ki
∆Q = − − weight greater than 0 is located on the topology which is
2m 2m
"P  P 2  2 # visited, and denoted as s. Visit the downstream vertex of vertex
in tot ki s, denoted as t. If the fault graph weight of t is greater than
− − (14)
2m 2m 2m 0, assign the community ID where vertex t is located to f id.
The above steps are repeated until the failure graph weight of
Equation (14) is the formula for calculating the modularity vertex t is less than 0 after a certain iteration, the loop ends
gain ∆Q, where Σin is the sum of the weights of all edges and vertex s and vertex t are output. At this point, if a D-
in Cj , Σtot is the sum of the edge incident to vertices in Cj , PMU is configured on vertex s, the fault section is the line
ki,in is the weight of the edge connected vertex i and Cj . By represented by the edge connecting vertex s and vertex t. If
simplifying the above equation, we can obtain the simplified vertex s is a vertex without D-PMU configured, it is necessary
modularity formula as follows: to use (10) to further determine the exact fault section.
 P 
1 tot ki Based on the results of the graph clustering in PLM,
∆Q = ki,in − (15)
2m 2m the fault location process in this paper is improved from
618 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 9, NO. 2, MARCH 2023

vertex-based depth-first traversal to community-based depth- of the fault graph weights of all vertices in the community is
first traversal, which reduces the number of iterations of depth- equal to the sum of the absolute values of the fault graph
first traversal. Therefore, the effect of PLM clustering will weights. The two constraints are “or” relationships, and either
directly affect the speed of fault location in this paper. In one of the two constraints can be satisfied to fuse two neigh-
testing, it was found that PLM graph clustering is not always boring communities into one community. This improvement
ideal, and at the most, two or three vertices with fault graph significantly increases the size of the community, reducing the
weights greater than 0 will form a community. In this paper, number of communities and the number of iterations of the
we expect the community formed by vertices with fault graph fault section query process.
weights greater than 0 to be as large as possible, so that the In this paper, phase A ground fault occurs at f3 on the
number of iterations in the fault localization process will be distribution network model shown in Fig. 8, with a fault resis-
further reduced. tance of 100 Ω. The weights of each edge in the distribution
To address the above situation, this paper analyzes the network graph model are obtained according to (17), and
reasons for the unsatisfactory clustering effect of PLM. There the direct depth-first traversal query, PLM-based depth-first
is ∆Q > 0 constraint for each vertex community attribution traversal query, and IPLM-based depth-first traversal query are
change in PLM. Since the distribution network graph model is used respectively, and the positioning results and the number
built with reference to the radial distribution network topology of iterations are shown Table I.
in this paper, this results in nearly 80% of the vertices with TABLE I
degree 2, and the connection relationship between the vertices C OMPARISON OF T HREE S EARCH M ETHODS
is relatively simple. This means that for 80% of the vertices, Location Maximum
Search method Iterations
the only community attribution available is to the community result community size
where the upstream and downstream vertices are located. With Direct depth-first traversal query 14,15 14 1
PLM-based depth-first traversal 14,15 5 5
strict adherence to the ∆ Q > 0 constraint, PLM forms many query
small communities with limited improvement in query speed. IPLM-based depth-first traversal 14,15 2 13
Based on the above analysis, this paper improves the PLM query
and proposes an improved parallel louvain method (IPLM). In
the first phase of PLM, the graph model has been changed in a
round of community attribution in terms of vertices to form a V. C ASE S TUDIES
preliminary community distribution. The second phase of PLM A. Accuracy Verification of the Proposed Method
is based on the “super node” unit and then the community In this paper, the IEEE 33-bus system graph model is built
attribution change, compared with the first phase, the second on the graph database, and the green nodes in Fig. 8 are D-
phase has fewer variables and the results are more easily PMU nodes installed and the black nodes are unconfigured
converged. In this paper, the constraint on the change of vertex D-PMU nodes. Phase A ground fault was set on six lines in
community attribution in PLM Phase II is modified and its the distribution network diagram model, as shown in Fig. 7.
constraint is changed from ∆ Q > 0 as shown in (17). In order to consider the influence of fault resistance on the
X
′
X
′ accuracy of fault location, this paper sets the fault resistance
∆Q > 0 or Si,k = Si,k , vi ∈ Ci (17)
with resistance values of 100 Ω, 500 Ω and 1000 Ω on all
In this paper, a constraint is added to PLM Phase II: the sum six fault lines. Since the fault occurs at different locations

The first phase

5 6 12 5 6 12 5 6 12 5 6 12
1 1 1 1
7 11 7 11 7 11 ... 7 11
2 2 4 2 2
4 4 4
9 9 9 9
3 8 10 3 8 10 3 8 10 3 8 10

The second phase

5
5 6 12 5 6 12
1 1
7 1 2 7
11 9 4 5 11
2 2
4 4
9 9
3 8 10 3 3 8 10
5 2

Fig. 7. PLM process.

DANG et al.: FAST SINGLE-PHASE FAULT LOCATION METHOD BASED ON COMMUNITY GRAPH DEPTH-FIRST TRAVERSAL FOR DISTRIBUTION NETWORK 619

f4
19 20 21 22

f1 f2 f3
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

f6
26 27 28 29 30 31 32 33
f5
23 24 25

Fig. 8. The modified IEEE 33-bus system.

on the line, the fault current at the time of the fault will be TABLE II
ACCURACY V ERIFICATION OF THE P ROPOSED M ETHOD
different; therefore, this paper sets up single-phase grounding
and earthing faults at the first 10%, 50% and last 10% of the The actual Fault
Ground
total length of the above six lines to verify the fault location location location DF PLMDF IPLMDF CS
resistance
of fault on the line
accuracy of the algorithm in the whole section of the line. 4,5 Top 10% 100 Ω 4,5 4,5 4,5 3,5
For different fault situations on the above six lines, 4,5 500 Ω 4,5 4,5 4,5 3,5
the depth-first traversal method (DF), PLM-based depth-first 4,5 1000 Ω 4,5 4,5 4,5 3,5
4,5 50% 100 Ω 4,5 4,5 4,5 3,5
traversal method (PLMDF), IPLM-based depth-first traversal 4,5 500 Ω 4,5 4,5 4,5 3,5
method (IPLMDF) and cuckoo algorithm-based fault location 4,5 1000 Ω 4,5 4,5 4,5 3,5
4,5 Last 10% 100 Ω 4,5 4,5 4,5 3,5
(CS) mentioned in the reference [23] are used in this paper, 4,5 500 Ω 4,5 4,5 4,5 3,5
respectively. The first three methods are implemented on the 4,5 1000 Ω 4,5 4,5 4,5 3,5
graph database TigerGraph and are based on the network 13,14 Top 10% 100 Ω 13,14 13,14 13,14 13,15
13,14 500 Ω 13,14 13,14 13,14 13,15
topology for fault section location. In the CS, assuming that 13,14 1000 Ω 13,14 13,14 13,14 13,15
all the above fault currents can be detected by the D-PMU, (7) 13,14 50% 100 Ω 13,14 13,14 13,14 13,15
13,14 500 Ω 13,14 13,14 13,14 13,15
is modified to (18), and the result generated by (18) is used 13,14 1000 Ω 13,14 13,14 13,14 13,15
as the basis for locating the CS. 13,14 Last 10% 100 Ω 13,14 13,14 13,14 13,15
( 13,14 500 Ω 13,14 13,14 13,14 13,15
1 ∆θ ∈ (−80◦ , −100◦ ) 13,14 1000 Ω 13,14 13,14 13,14 13,15
σi (∆θ) = (18) 14,15 Top 10% 100 Ω 14,15 14,15 14,15 13,15
0 ∆θ ∈ / (−80◦ , −100◦ ) 14,15 500 Ω 14,15 14,15 14,15 13,15
14,15 1000 Ω 14,15 14,15 14,15 13,15
Table II shows the results of the tests. From the results in 14,15 50% 100 Ω 14,15 14,15 14,15 13,15
14,15 500 Ω 14,15 14,15 14,15 13,15
Table II, it can be seen that the three fault localizations based 14,15 1000 Ω 14,15 14,15 14,15 13,15
on the graph calculation proposed in this paper can directly 14,15 Last 10% 100 Ω 14,15 14,15 14,15 13,15
14,15 500 Ω 14,15 14,15 14,15 13,15
and accurately locate the faults occurring at f1 , f2 , f3 , f4 and 14,15 1000 Ω 14,15 14,15 14,15 13,15
f6 . When a fault occurs at f5 , the output of the first three 20,21 Top 10% 100 Ω 20,21 20,21 20,21 19,21
methods in this paper is 25,0. Because 25 vertices are line 20,21 500 Ω 20,21 20,21 20,21 19,21
20,21 1000 Ω 20,21 20,21 20,21 19,21
end vertices in the distribution network diagram model and 20,21 50% 100 Ω 20,21 20,21 20,21 19,21
there are no ID 0 vertices in the model, the locating result is 20,21 500 Ω 20,21 20,21 20,21 19,21
20,21 1000 Ω 20,21 20,21 20,21 19,21
still accurate. 20,21 Last 10% 100 Ω 20,21 20,21 20,21 19,21
Comparing CS with the first three methods, only the fault 20,21 500 Ω 20,21 20,21 20,21 19,21
occurring at f5 in the output result is accurate, and the rest 20,21 1000 Ω 20,21 20,21 20,21 19,21
24,25 Top 10% 100 Ω 25,0 25,0 25,0 24,25
of the output results contain both fault and normal sections, 24,25 500 Ω 25,0 25,0 25,0 24,25
with a larger localization range. This is due to the fact that the 24,25 1000 Ω 25,0 25,0 25,0 24,25
24,25 50% 100 Ω 25,0 25,0 25,0 24,25
CS relies only on the overcurrent signal with the measurement 24,25 500 Ω 25,0 25,0 25,0 24,25
unit for fault location, and the smallest unit of output is in the 24,25 1000 Ω 25,0 25,0 25,0 24,25
section between the two measurement units. Therefore, with 24,25 Last 10% 100 Ω 25,0 25,0 25,0 24,25
24,25 500 Ω 25,0 25,0 25,0 24,25
the same number of D-PMUs configured, the DF, PLMDF, 24,25 1000 Ω 25,0 25,0 25,0 24,25
and IPLMDF proposed in this paper have better positioning 28,29 Top 10% 100 Ω 28,29 28,29 28,29 28,30
28,29 500 Ω 28,29 28,29 28,29 28,30
effects than CS. 28,29 1000 Ω 28,29 28,29 28,29 28,30
28,29 50% 100 Ω 28,29 28,29 28,29 28,30
B. Robustness Verification of the Proposed Method 28,29 500 Ω 28,29 28,29 28,29 28,30
28,29 1000 Ω 28,29 28,29 28,29 28,30
The conclusions in the previous subsection have verified 28,29 Last 10% 100 Ω 28,29 28,29 28,29 28,30
the positioning accuracy of DF, PLMDF, and IPLMDF in all 28,29 500 Ω 28,29 28,29 28,29 28,30
28,29 1000 Ω 28,29 28,29 28,29 28,30
D-PMU normal operating cases.
620 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 9, NO. 2, MARCH 2023

In this paper, a phase A ground fault with the fault resistance fault point lose synchronization, and none of the above four
of 1000 Ω is set at f1 − f6 50% of the line, and some D- methods can locate the exact fault section.
PMUs in the network lose time synchronization. We test the In summary, when some D-PMUs lose time synchronization
fault location effectiveness of the above four methods when in the network, IPLMDF has the best positioning effect,
some D-PMUs lose time synchronization. Table III shows the followed by DF and PLMDF, and CS has the worst.
results of the tests.
Table III compares the accuracy of the four fault-setting unit C. Comparison of the Calculation Speed of the Proposed
methods in the event that some of the D-PMUs in the network Method
lose their time synchronization. In the previous two sections, the accuracy of DF, PLMFDF,
Among the 13 cases mentioned above, DF and PLMDF and IPLMDF were verified in the normal case and in the partial
can accurately locate the fault section in the first 9 cases, D-PMU out-of-step case. In this section, we want to compare
and IPLMDF can accurately locate the fault section in the the computational speed of the above four methods.
first 10 cases. CS does not converge the output results in Since the number of iterations of the above four methods
the 5th and 9th cases and outputs incorrect results in the 3rd is primarily affected by the location of the fault section, the
case compared to the first three methods. In the 10th case, results of the midpoint of the line where the fault occurred
the results of DF do not converge, PLMDF falls into local and the fault resistance value of 1000 Ω are selected as being
optimum, IPLMDF accurately locates the fault, and CS does representative in this paper.
not accurately locate the fault section. In Fig. 10, the horizontal coordinates are the fault locations
The 10th case community diagram is shown in Fig. 9. Vertex f1 − f6 and the vertical coordinates are the number of
6 is a branch vertex on the graph model with downstream iterations. Fig. 9 shows the number of iterations for each case
vertices 7 and 26. until the D-PMU of vertex 7 loses synchro- in Table III, the horizontal coordinate is the serial number
nization, the fault graph weight of vertex 7 is less than 0 and corresponding to each case, and the vertical coordinate is the
does not cause misjudgment. After the loss of synchronization
of the D-PMU at vertex 7, the estimated fault map weight CS DF PLMDF IPLMDF
obtained is greater than 0. At this point, the failure graph 16
weights of both vertices downstream of vertex 6 are greater 15
than 0, so the results of DF do not converge. However, vertex 14
6 and vertex 7 are in the same community, vertex 26 is in 13
another community, which causes the PLMDF to fall into a 12
local optimum during the query process, and the output fault 11
section is 7,8. The community is larger in IPLMDF, vertex 6, 10
Iterations

vertex 7 and vertex 26 are merged into the same community 9

8
and the output results in the global optimal solution, i.e., 28,29.
7
In the last three cases, the two D-PMUs closest to the
6
5
PLM 4
19 20 21 22 IPLM 3
2
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
1
26 27 28 29 30 31 32 33 0
f1 f2 f3 f4 f5 f6
23 24 25 Fault location
f5
Fig. 10. Comparison of calculation speed of four methods under normal
Fig. 9. The 10th case community graphs. conditions.

TABLE III
ROBUSTNESS V ERIFICATION OF THE P ROPOSED M ETHOD
Serial number The actual location of fault ID of the D-PMU out of step DF PLMDF IPLMDF CS
1 4,5 28 4,5 4,5 4,5 3,5
2 13,14 9 13,14 13,14 13,14 13,15
3 13,14 9,11 13,14 13,14 13,14 7,9
4 14,15 9,17 14,15 14,15 14,15 13,15
5 14,15 3,7,11,21 14,15 14,15 14,15 5,7,26; 9,11; 13,15; 5,24
6 20,21 5,17,28 20,21 20,21 20,21 19,21
7 24,25 3 25,0 25,0 25,0 24,25
8 24,25 3,21,30 25,0 25,0 25,0 24,25
9 28,29 3,26 28,29 28,29 28,29 28,30; 5,12; 1,3
10 28,29 7,21 7,8; 28,29 7,8 28,29 28,30
11 4,5 3 1,2 1,2 1,2 1,3
12 14,15 13,15 13,14 13,14 13,14 11,13
13 20,21 19 19,20 19,20 19,20 1,3,19
DANG et al.: FAST SINGLE-PHASE FAULT LOCATION METHOD BASED ON COMMUNITY GRAPH DEPTH-FIRST TRAVERSAL FOR DISTRIBUTION NETWORK 621

CS DF PLMDF IPLMDF of the fault sections. In addition, IPLMDF is applicable to the

19 whole section of the line, independent of the fault resistance,
18
17 and has a certain degree of robustness. In the case of partial D-
16 PMU losing time synchronization, it is still possible to quickly
15 and accurately locate the fault section, while avoiding the
14
13
situation of falling into local optimal solutions.
12 Currently, the IPLMDF proposed in this paper can only be
11 used to locate single-phase ground faults. In our next study, it
Iterations

10
is planned to extend the application of IPLMDF to all types
9
8 of short-circuit faults.
7
6
5 R EFERENCES
4
3 [1] M. Gholami, A. Abbaspour, M. Moeini-Aghtaie, M. Fotuhi-Firuzabad,
2 and M. Lehtonen, “Detecting the location of short-circuit faults in
1 active distribution network using PMU-based state estimation,” IEEE
0 Transactions on Smart Grid, vol. 11, no. 2, pp. 1396–1406, Mar. 2020.
1 2 3 4 5 6 7 8 9 10 11 12 13
Serial number [2] M. M. Saha, J. Izykowski, and E. Rosolowski, “Fault Location on Power
Networks,” New York: Springer, 2010, pp. 1–4.
[3] J. Liu, X. Q. Zhang, W. Shen, L. Quan, and Z. H. Zhang, “Performance
Fig. 11. Comparison of calculation speed of four methods under some D-
testing of single phase to earth fault location for distribution network
PMUs out of step.
with neutral point non-effectively grounded systems,” Automation of
Electric Power Systems, vol. 42, no. 1, pp. 138–143, Jan. 2018.
[4] M. Daisy and R. Dashti, “Single phase fault location in electrical
number of iterations. distribution feeder using hybrid method,” Energy, vol. 103, pp. 356–
From Figs. 10 and 11, we can conclude that: CS has unstable 368, May 2016.
performance due to the random process, and the number of [5] S. Jamali and A. Bahmanyar, “A new fault location method for distri-
iterations is generally more than 10; DF is most sensitive to bution networks using sparse measurements,” International Journal of
Electrical Power & Energy Systems, vol. 81, pp. 459–468, Oct. 2016.
the topological location of the fault section, and it can quickly [6] R. Chen, X. Yin, X. G. Yin, Y. L. Li, and J. Y. Lin, “Computational
locate the fault section when it is close to the power node, fault time difference-based fault location method for branched power
but requires more iterations to locate the fault section if it is distribution networks,” IEEE Access, vol. 7, pp. 181972–181982, Feb.
2019.
close to the end of the line; PLMDF is less affected by the [7] H. C. Shu, J. L. Liu, X. C. Tian, “Fault Location of Radial Distribution
topological location of the fault section than DF Compared Network Based on Sudden Changes of Fault Traveling Wave Along Line
with DF, PLMDF is less affected by the topological location and Model Matching,” Automation of Electric Power Systems, vol. 44,
no. 9, pp. 158–163, May. 2020.
of the fault section, but still requires 5 or more iterations to [8] P. Wang, B. C. Chen, H. Zhou, C. H. Tian, and B. Sun, “Fault location
locate the fault if the fault section is near the end of the line; in resonant grounded network by adaptive control of neutral-to-earth
IPLMDF has the most stable performance and usually requires complex impedance,” IEEE Transactions on Power Delivery, vol. 33,
no. 2, pp. 689–698, Apr. 2018.
only 2–3 iterations to directly and accurately locate the fault [9] X. W. Wang, Wang, H. X. Zhang, F. Shi, Q. W. Wu, V. Terzija,
section. W. Xie, and C. Fang, “Location of single phase to ground faults in
By comparing the above three aspects, it can be intuitively distribution networks based on synchronous transients energy analysis,”
IEEE Transactions on Smart Grid, vol. 11, no. 1, pp. 774–785, Jan.
determined that IPLMDF outperforms the other three methods 2020.
both in terms of fault tolerance and computation speed. [10] C. B. Wang, Z. H. Yun, H. X. Zhang, F. Shi, P. Ling, and B. P. Xie,
“Parameter-free fault location algorithm for multi-terminal overhead
transmission line of distribution network based on µMPMU,” Power
VI. C ONCLUSION System Technology, vol. 43, no. 9, pp. 3202–3209, Sep. 2019.
In this paper, a distribution network graph model is built on [11] J. F. Liang, T. J. Jing, H. N. Niu, and J. B. Wang, “Two-terminal fault
location method of distribution network based on adaptive convolution
the graph database TigerGraph, and the fault graph weight of neural network,” IEEE Access, vol. 8, pp. 54035–54043, Mar. 2020.
the vertices are defined to reflect the topological relationship [12] K. J. Chen, J. Hu, Y. Zhang, Z. Q. Yu, and J. L. He, “Fault location
between the vertices and the fault points, as well as the degree in power distribution systems via deep graph convolutional networks,”
IEEE Journal on Selected Areas in Communications, vol. 38, no. 1, pp.
of fault impact on the vertex currents. In this paper, a fault lo- 119–131, Jan. 2020.
cation method based on community graph depth-first traversal [13] L. Zhang, B. Zhao, L. J. Wang, and X. Q. Fan, “On-line fault location
is proposed for fast location of single-phase ground faults in method for ungrounded system based on graph theory,” Power System
Technology, vol. 43, no. 8, pp. 3025–3031, Aug. 2019.
distribution networks. The main idea of this approach is the use [14] K. Jia, L. Li, Z. Yang, G. K. Zhao, and T. S. Bi, “Research on distribution
of an improved PLM to cluster the vertices in the graph model network fault location based on Bayesian compressed sensing theory,”
to form communities of different sizes. And then perform a Proceedings of the CSEE, vol. 39, no. 12, pp. 3475–3485, Jun. 2019.
depth-first traversal with the community as the smallest unit. [15] Z. X. Zhou, C. Yuan, Z. Y. Yao, J. P. Dai, G. Y. Liu, R. C. Dai, Z.
W. Wang, and G. M. Huang, “CIM/E oriented graph database model
The positioning accuracy, robustness, and computational speed architecture and parallel network topology processing,” in 2018 IEEE
of the four methods, DF, PLMDF, IPLMDF, and CS, were Power & Energy Society General Meeting (PESGM), 2018, pp. 1–5.
compared through simulation tests. With half of the nodes [16] H. Huang, J. P. Dai, Y. Wang, X. Shan, M. Y. Zhai, and G. Y. Liu,
“Graph database based construction and network topology of CIM/E
in the distribution network model configured with D-PMUs, for power grid,” Automation of Electric Power Systems, vol. 43, no. 22,
IPLMDF can quickly and accurately determine the localization pp. 122–129, Nov. 2019.
622 CSEE JOURNAL OF POWER AND ENERGY SYSTEMS, VOL. 9, NO. 2, MARCH 2023

[17] C. Yuan, Y. Q. Zhou, G. F. Zhang, G. Y. Liu, R. C. Dai, X. Chen, and Rong Jia received the Ph.D. degree from Xi’an
Z. W. Wang, “Exploration of graph computing in power system state University of Technology, Shannxi, China, in 1999.
estimation,” in 2018 IEEE Power & Energy Society General Meeting He is currently a professor with the Department of
(PESGM), 2018, pp. 1–5. Electrical Engineering, Xi’an University of Technol-
[18] G. Y. Liu, K. W. Liu, D. Shi, W. D. Zhu, Z. W. Wang, and X. Chen, ogy, Shannxi, China. His research interest includes
“Graph computation and its applications in smart grid,” in 2017 IEEE power system operations and control, and power
International Congress on Big Data (BigData Congress), 2017, pp. 507– equipment status monitoring.
510.
[19] J. P. Dai, Z. Y. Yao, G. F. Zhang, G. Y. Liu, R. C. Dai, and Z. W.
Wang, “Graph computing-based real-time network topology analysis for
power system,” in 2019 IEEE Power & Energy Society General Meeting
(PESGM), 2019, pp. 1–5.
[20] R. C. Dai, G. Y. Liu, Z. W. Wang, B. W. Kan, and C. Yuan, “A novel
graph-based energy management system,” IEEE Transactions on Smart
Grid, vol. 11, no. 3, pp. 1845–1853, May 2020. Xiaowei Wang received the Ph.D. degree from
[21] V. D. Blondel, J. L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast Xi’an Jiaotong University, Shannxi, China, in 2019.
unfolding of communities in large networks,” Journal of Statistical He is currently an Associate Professor with the De-
Mechanics: Theory and Experiment, vol. 2008, pp. 10008, Oct. 2008. partment of Electrical Engineering, Xi’an University
[22] C. L. Staudt and H. Meyerhenke, “Engineering parallel algorithms of Technology, Shaanxi, China. His research interest
for community detection in massive networks,” IEEE Transactions on includes small current grounding system and fault
Parallel and Distributed Systems, vol. 27, no. 1, pp. 171–184, Jan. 2016. line selection.
[23] X. L. Huang, Z. Y. Xie, and X. Y. Huang, “Fault location of distribution
network base on improved cuckoo search algorithm,” IEEE Access, vol.
8, pp. 2272–2283, Dec. 2019.

Jian Dang received the Ph.D. degree from Xi’an Hui Wei received the B.S. degree in Electrical Engi-
University of Technology, Shannxi, China, in 2018. neering at Xi’an University of Technology, Shannxi,
He is currently an Associate Professor with the China, in 2015. She is currently pursuing the M.S.
Department of Electrical Engineering, Xi’an Uni- degree in Electrical Engineering at Xi’an University
versity of Technology, Shaanxi, China. His research of Technology. Her research interests include knowl-
interests include fault location of distribution net- edge graph applications in power systems, power
works, artificial Intelligence and graph applications equipment state estimation.
in power systems.

Yunjiang Yan received the B.S. degree in Electri-

cal Engineering at Xi’an University of Technology,
Shannxi, China, in 2015. He is currently pursuing the
M.S. degree in Electrical Engineering at Xi’an Uni-
versity of Technology. His research interests include
graph computation application in power system,
distribution network fault location.