IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 20, NO.

1, JANUARY 2023 47

Graph Convolutional Network-Based Interpretable

Machine Learning Scheme in Smart Grids
Yonghong Luo , Chao Lu , Senior Member, IEEE, Lipeng Zhu , Member, IEEE,
and Jie Song , Senior Member, IEEE

Abstract— Smart grid is a typical application of industrial work well in the practical challenging environment of smart grids,
cyber-physical systems (ICPS) in the electric power industry. but also it helps the dispatchers in smart grids better understand
Due to the exposure to different kinds of uncertainties and and trust the proposed SVS assessment scheme.
unpredictable faults, how to reliably assess the short-term voltage
stability (SVS) of smart grids to prevent the occurrence of Index Terms— Graph convolutional network (GCN), shapelet
large-scale blackouts is still of primary concern. To tackle transform, short-term voltage stability (SVS), smart grid data
this challenging problem, this article develops a novel machine analytics.
learning scheme to achieve accurate and interpretable online SVS
assessment in two steps. First, it utilizes time-series shapelet N OMENCLATURE
transform to extract key dynamics and convert the postfault
time series into flat features. Second, it designs a graph con- Bi as The learned bias in the system layer.
volutional network (GCN) to incorporate these features with DS The single-variable datasets seperated from the
topology information for SVS assessment. The GCN explores the original multivariable dataset.
spatial-temporal dynamics of power system via graph convolution DT The transformed datasets of DS by shapelet
and introduces a system layer to derive the final assessment candidates.
result. Compared with conventional methods, this novel scheme
makes full use of the spatial-temporal information in SVS Pgcn The active power injection data processed by
dynamics, resulting in higher assessment accuracy and stronger graph convolutional layer.
adaptability. Besides, it is capable of discovering certain valuable Pt r The active power injection features processed by
underlying rules and patterns related to SVS. Test results on the shapelet transform.
IEEE 39-bus system and real-world Guangdong Power Grid in P Active power injection time series.
South China verify the effectiveness of the proposed scheme.
Q gcn The reactive power injection data processed by
Note to Practitioners—To achieve accurate and interpretable graph convolutional layer.
online short-term voltage stability (SVS) assessment in the Qtr The reactive power injection features processed
challenging environment of smart grids, this article develops a by shapelet transform.
novel machine learning scheme with full consideration of the Q Reactive power injection time series.
spatial-temporal information in SVS dynamics. First, it utilizes
the time series shapelet transform to convert the postfault time Sb The learned node weights in the system layer.
series into flat features. Second, it designs a graph convolutional Ss ys The final assessment result.
network (GCN) to incorporate these features with topology infor- S The equivalent weights in the system layer.
mation. The full consideration of spatial-temporal information Vgcn The voltage magnitude data processed by graph
in SVS dynamics can improve the assessment accuracy, and the convolutional layer.
integration of topology in the scheme can promote its adaptability
to topology changes. Apart from the decent performances under Vt r The voltage magnitude features processed by
changeable environments, the proposed scheme can provide shapelet transform.
certain valuable underlying rules and patterns related to SVS. V Voltage magnitude time series.
Therefore, not only the proposed scheme for SVS assessment can W The topology of the target region.
θ pi The learned parameters of active power injection
Manuscript received 1 October 2020; revised 11 April 2021; accepted in the i th Chebyshev order of graph convolutional
5 June 2021. Date of publication 2 August 2021; date of current version
6 January 2023. This article was recommended for publication by Editor layer.
S. Hu upon evaluation of the reviewers’ comments. This work was supported θ qi The learned parameters of reactive power injec-
by the National Key Research and Development Program of China under tion in the i th Chebyshev order of graph convo-
Grant 2017YFB0902801. (Corresponding author: Chao Lu.)
Yonghong Luo and Chao Lu are with the State Key Laboratory of Power lutional layer.
Systems, Department of Electrical Engineering, Tsinghua University, Beijing θsj The learned channel weights of the jth output
100084, China (e-mail: [email protected]). channel.
Lipeng Zhu is with the College of Electrical and Information Engineering,
Hunan University, Changsha 410082, China. θ vi The learned parameters of voltage magnitude in
Jie Song is with the Department of Industrial Engineering and Management, the i th Chebyshev order of graph convoll layer.
College of Engineering, Peking University, Beijing 100871, China. B The equivalent bias in the system layer.
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TASE.2021.3090671. L en The length of candidate shapelets.
Digital Object Identifier 10.1109/TASE.2021.3090671 L max The maximal length of candidate shapelets.
1545-5955 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
48 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 20, NO. 1, JANUARY 2023

L min The minimum length of candidate shapelets. Although these efforts can effectively accelerate simulation,
n The number of observed buses. they are still faced with the challenges of accurate modeling
Nin The input channel number of the graph convolu- of the system model and reliable identification of model
tion layer. parameters. The uncertainty and time-varying characteristics
Nout The output channel number of the graph convo- of load model parameters make it very difficult to conduct
lution layer. accurate SVS assessment by numerical simulation.
Sy The equivalent assessment result. The engineering criteria of the second category are mainly
Sn i The i th node representation in system layer. based on the practical experiences of system operators in smart
T The number of time points for stability grids. The widely accepted engineering criteria assess SVS
assessment. status according to the bus voltage magnitude of the observed
T ( L̃) The Chebyshev filter matrix. power system. When the duration of any bus voltage below
a threshold exceeds the preset time [10], [11], the system is
I. I NTRODUCTION assessed as unstable. It is convenient and can provide assess-
ment results in a short time. However, the construction of the
A. Research Background engineering criteria only relies on the operating experiences of

W ITH the widespread deployment of phasor measure-

ment units (PMUs), the rapid development of infor-
mation technologies and data processing techniques, modern
dispatchers. It lacks support from theory or big data, thereby
short of reliability in the SVS assessment task.
Much effort has been paid to stability mechanism-based
smart grids have evolved as advanced industrial cyber-physical assessment, such as energy function [12], bifurcation analy-
systems (ICPS). sis [13], maximum Lyapunov exponent (MLE) method [14],
However, smart grids are often operated in unpredictable and mechanism analysis of induction motor [15]. Energy func-
and challenging environments. With the integration of renew- tion and bifurcation analysis can promote the understanding
able energy resources and the intrinsic time-varying charac- of voltage instability. Nevertheless, they cannot be applied to
teristic of power loads, the operating points of smart grids large practical power grids because of their complexity. The
are changeable, and smart grids may undergo some extreme MLE method [14] is a typical method of the third category.
operating points. Under that circumstance, if large disturbances It utilizes the sign of the MLE value to conduct an SVS
happen in load centers, smart grids may undergo short-term assessment. However, it fails to provide a reliable result as
voltage instability and result in blackouts [1], [2]. Large-scale the sign of MLE value often oscillates around 0. The recovery
blackouts can lead to huge economic loss and adverse social time of equivalent induction motor rotation speed can also be
impact. Therefore, in order to prevent the occurrence of utilized for SVS assessment. However, the accuracy of this
blackouts, it is crucial to correctly carry out reliable online method is greatly affected by the identification of induction
short-term voltage stability (SVS) assessment to take emer- motor parameters, while the identification of induction motor
gency measures in time. parameters is also very difficult.
Voltage stability refers to the ability of a power system to Due to the superior performance of machine learning,
maintain steady voltages at all buses in the system after being the methods of the fourth category have attracted much
subjected to a disturbance from a given initial operating con- attention recently. Some methods take measurements from
dition [3]. SVS mainly involves dynamics of fast-acting load single snapshots of postfault time series as the learning inputs,
components such as induction motors and HVDC converters such as artificial neural network (ANN)-based method [16],
while long-term voltage stability mainly involves slower acting decision tree (DT)-based method [17], and the support vector
equipment such as tap-changing transformers and generator machine (SVM)-based method [18]. However, the temporal
current limiters [3], [4]. Compared with long-term voltage correlation of instability dynamics cannot be well character-
stability, SVS is related to much more rapid and complicated ized with an only a single snapshot of time series, and these
postfault system dynamics, which makes it very difficult to methods may go wrong when the dynamic parameters of
correctly perform SVS assessment in a short time. the smart grid change, such as the load parameters. To cap-
ture the time-varying characteristics of postfault dynamics,
B. Literature Review shapelet-based methods [19], [20], long short-term memory
There are mainly four categories of approaches for power (LSTM)-based method [21], and random vector functional link
system stability assessment, simulation-based methods, engi- (RVFL) [22] take the postfault time series as the learning input.
neering criteria, stability mechanism-based methods, and Based on the postfault time series, these methods can achieve
machine learning-based methods. higher assessment performance than the methods based on
Given the information about the operating points and faults, single snapshots. However, apart from the time-varying char-
the first category attempts to provide the stability results via acteristics, short-term voltage instability also presents spatial
detailed numerical simulations [5], [6]. These methods have distribution characteristics over topology, which has been
high requirements on the computing time and the accuracy neglected for years by most of machine learning-based meth-
of system component modeling. There are some works trying ods. Up to now, there is only one method that incorporates
to accelerate simulation, for example, using a reduced model spatial information into the learning models for SVS assess-
in simulation [7], using highly dedicated hardware for simula- ment [23]. It utilizes the geographical location information of
tion [8], and adopting distributed time-domain simulations [9]. the nodes in a given region to perform spatial interpolation

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
LUO et al.: GRAPH CONVOLUTIONAL NETWORK-BASED INTERPRETABLE MACHINE LEARNING SCHEME 49

and takes the voltage contour as the link to extract spatial In fact, SVS presents prominent spatial-temporal character-
features. However, the geographical locations of the buses may istics in postfault dynamics. The postfault dynamics of SVS
not exactly reflect the electrical distance between buses, and are shown in Fig. 1. As can be seen in Fig. 1, from the
the extracted spatial features based on geographical locations spatial perspective, the affected regions of short-term voltage
may not be accurate enough. Besides, the fixed geographical instability reveal obvious spatial distribution characteristics
locations cannot reflect the topology changes when the smart over topology, and the low-voltage region reveals locality. This
grid is confronted with reconfiguration. Therefore, the assess- phenomenon can be attributed to the fact that reactive power
ment model based on geographical location information may cannot be transmitted over long distances and the voltage of
not be accurate enough. one bus mainly affects its electrically neighboring buses. From
the temporal perspective, the affected region changes as time
C. Contributions goes by. It is due to the inherent dynamic characteristics of
In fact, graph convolutional networks are powerful to deal the components in smart grids.
with data residing on the graph. They can integrate topol- If the complete spatial-temporal dynamics are considered
ogy information in the learning model directly instead of in SVS assessment, it is feasible to achieve SVS assessment
using geographical locations, and they have been successfully with higher performances under the changeable environment.
applied in social networks [24], transportation [25], chem- Besides, the integration of effective spatial information helps
istry [26], and fault location in power distribution systems [27]. to deal with the reconfiguration and topology changes in smart
Therefore, in this article, a novel machine learning scheme grids.
based on a graph convolutional network is developed for SVS
assessment. First, the time-series shapelet transform [19] is B. Overview of the Proposed Scheme for SVS Assessment
adopted to convert time series into flat distance data. The
shapelet transform method can extract key dynamics in time To exploit the complete spatial-temporal dynamics of
series and generate effective distance data for subsequent SVS, a novel SVS assessment scheme is developed with
learning. Second, the graph convolutional network is designed spatial-temporal information incorporation. The proposed
to incorporate topology with the distance data for online SVS scheme consists of two parts, shapelet transform and the graph
assessment. In this way, the proposed scheme fully exploits convolutional network, as shown in Fig. 2. Shapelet transform
the spatial-temporal characteristics of SVS, which helps to is utilized to exploit the temporal information and converts
improve the accuracy and reliability of SVS assessment under time series into flat features. The shapelet transform method
changeable environments. The main contributions of this arti- is capable of extracting key dynamics in time series and it
cle are outlined as follows. generates effective features based on the key dynamics for
1) A graph convolutional network is designed to integrate subsequent learning. Then, the graph convolutional network is
topology in SVS assessment, which helps to adapt to designed to incorporate spatial information with these features
different topology changes in the smart grid. and provide the final assessment result. The graph convolu-
2) This article for the first time develops an interpretable tional network consists of two parts, the graph convolutional
machine learning scheme with full consideration of layer and the system layer. The graph convolutional layer
spatial-temporal information for SVS assessment. integrates topology in the model, which helps to adapt to
3) The proposed scheme is capable of discovering certain topology changes in smart grids. The system layer is designed
underlying rules and patterns related to SVS. to generate system representation and provide the final assess-
4) The proposed scheme is fully tested in real-world ment result.
Guangdong Power Grid and IEEE 39-bus system, which The implementation of the proposed scheme can be divided
validate the effectiveness of the proposed method. into two stages, the offline training stage and the online assess-
ment stage, as shown in Fig. 3. At the offline training stage,
The rest of this article is organized as follows. The scheme
with the help of power system electromechanical transient sim-
overview for SVS assessment is illustrated in Section II.
ulation tools, a sufficient number of samples can be generated
Shapelet transform in the proposed scheme is described in
to form the database. For each sample in the database, the input
Section III. In Section IV, the designed graph convolutional
data consist of voltage magnitude time series V , active power
network is illustrated in detail. In Section V, test results on the
injection time series P, reactive power injection time series
IEEE 39-bus system and real-world Guangdong Power Grid
Q, and topology W. More specifically, topology matrix W is
are utilized to verify the performance of the proposed scheme.
composed of a node admittance matrix. It represents electrical
Finally, conclusions are presented in Section VI.
coupling of power transmission lines, which is very important
II. S CHEME OVERVIEW FOR SVS A SSESSMENT in the evolution of SVS dynamics [2]. The output label is the
A. SVS Assessment Problem final SVS status Ssys , which is represented as [0 1] or [1 0].
Smart grids work in a challenging environment with differ- If short-term voltage instability occurs in the observed system,
ent kinds of uncertainties and unpredictable faults. Besides, the stability status is [0 1], otherwise, the stability status is
they are also usually confronted with reconfiguration and [1 0]
⎧ 
topology changes. Due to the challenging environment and ⎪
⎨ 10 the observed system is stable
strong nonlinearity of smart grids, the postfault dynamics of
Ssys =   (1)
SVS are very complex and it is hard to achieve accurate and ⎪
⎩ 01 the observed system is unstable.
reliable online SVS assessment.
Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
50 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 20, NO. 1, JANUARY 2023

Fig. 1. Postfault dynamics of short-term voltage instability with illustrative voltage distribution maps. (a) Voltage distribution map of t = 0 s after fault
clearance. (b) Voltage distribution map of t = 0.5 s after fault clearance. (c) Voltage distribution map of t = 1.0 s after fault clearance. These illustrative
voltage distribution maps are depicted with spatial voltage interpolation [28], and they are the snapshots of the postfault voltage magnitude time series in the
IEEE 39-bus test system. The legend on the right side of these figures illustrates the correspondence between colors and voltage values.

utilized to conduct an SVS assessment. If the assessment result

indicates that the observed system is stable, it continues to
monitor. Otherwise, emergency measures are needed to tackle
the unstable issue and prevent the occurrence of large-scale
blackout.

III. S HAPELET T RANSFORM

Fig. 2. Proposed scheme for SVS assessment.
Shapelet transform [30] is adopted in the proposed scheme
to exploit the temporal information and convert the time series
dataset into a flat distance dataset for subsequent learning.
Shapelets are the special time series subsequences that can best
distinguish different classes, and the extracted shapelets can be
regarded as the key dynamics of the postfault time series. The
implement of shapelet transform includes three steps.
Step 1: The generation of candidate shapelets. All possible
subsequences in the time-series dataset are extracted as the
shapelet candidates.
Step 2: The evaluation of the candidate shapelets. All the
shapelet candidates need to be evaluated and the shapelet
candidates which can best distinguish different classes are
chosen as the extracted shapelets.
Step 3: The transformation with the extracted shapelets. The
time-series dataset can be converted into a distance dataset
Fig. 3. Implementation of the proposed scheme for SVS assessment.
according to the Euclidean distance between time series and
shapelets.
After preparing the time series database, shapelet transform More specifically, in the first step, suppose there are N
is employed to convert the time series dataset into a distance samples in the time series dataset, each sample is composed
dataset with the extracted shapelets. Then, the graph convolu- of multivariate time series. Split the multivariate time series
tional network can be trained with gradient descent algorithms, dataset into multiple single-variable time series datasets
such as Adam algorithm [29]. When the testing accuracy of
DS = { DS1 , DS2 , . . . , DSi , . . . , DSd } (2)
the proposed method reaches a preset threshold, the shapelets
and the trained graph convolutional network will be saved where DSi ∈ R N ×T , d = 3n. Then, extract shapelets
for online assessment. At the online assessment stage, when from these datasets, respectively. Assuming the length of the
a fault occurs, postfault PMU measurements and topology candidate shapelets is between L min and L max , slide the time
are obtained as the inputs. Then, the extracted shapelets are window with length L en (L min ≤ L en ≤ L max ) in each time
utilized to convert the postfault PMU measurements into dis- series of single-variable dataset DSi to extract all the possible
tance data. After that, the trained graph convolutional model is subsequences as the shapelet candidates.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
LUO et al.: GRAPH CONVOLUTIONAL NETWORK-BASED INTERPRETABLE MACHINE LEARNING SCHEME 51

In the second step, all the shapelet candidates will be

evaluated by the ability to distinguish different classes. Taking
one shapelet candidate sc, for example, the single-variable
time series dataset can be converted into a distance vector
according to the Euclidean distance between the shapelet
candidate sc and the time series from the dataset DSi . Suppose
ds is one instance from DSi , the Euclidean distance between
the shapelet candidate sc and ds is defined as
D(sc, ds) = min (D(sc, ds L )) (3)
ds L ∈CanL

where sc is the shapelet candidate with length L en , Can L

Fig. 4. Designed graph convolutional network for SVS assessment.
is the subsequences set of ds with length L en , and ds L
is the subsequence extracted from subsequences set Can L .
D(sc, ds L ) is the Euclidean distance between sc and ds L . In general, the designed graph convolutional network con-
After calculating the distance values between sc and all tains two parts: the graph convolutional layer and the system
the instances in DSi , the single-variable time series dataset layer. The illustration of the designed graph convolutional
DSi can be converted into a distance vector DTi . With the network for SVS assessment is shown in Fig. 4.
distance vector DTi , the dataset can be classified into two 1) Graph Convolutional Layer: In this article, spectral
subsets according to the split point sp. The information gain theory-based graph convolution is adopted. It studies the prop-
is utilized to measure the ability of shapelet candidate sc for erties of graph by eigenvalues and eigenvectors of the normal-
distinguishing different classes under split point sp. Iterate ized Laplacian matrix. The normalized Laplacian matrix L is
through all possible split points in distance dataset DTi to the mathematical representation of graph. It has real symmetric
search for the largest information gain. The largest information semidefinite properties and can be factored as eigenvalues and
gain is regarded as the ability of shapelet candidate sc to dis- eigenvectors
tinguish different classes. Iterate through all possible shapelet
candidates in time series dataset DSi , the shapelet candidate L = In − D−1/2 W D−1/2 (4)
with the largest information gain is regarded as the resulting L = UU T (5)
shapelet of DSi .
In the third step, with the extracted shapelets in multi- where W is the adjacency matrix with weights, D is a diagonal
ple single-variable time-series datasets DS, the multivariate matrix of the corresponding nodal degrees of W, and In
time-series dataset can be transformed into a distance dataset is an identity matrix.  is a diagonal matrix composed of
according to the Euclidean distance between the shapelets and eigenvalues and U is the matrix of eigenvectors ordered by
the postfault time series V , P, Q. Vt r , Pt r , Q t r are the eigenvalues.
transformed features for the subsequent learning. The graph convolution of the signal x with a filter g is
defined as
IV. G RAPH C ONVOLUTIONAL N ETWORK 
Graph convolutional networks can integrate topology infor- x ∗ Gg = U U T x UT g (6)
mation in the assessment model and incorporate spatial

information with temporal features, which helps to achieve where denotes the Hadamard product. If a filter is denoted
higher assessment accuracy and stronger adaptability. There as gθ = diag(U T g), the graph convolution can be simplified
are mainly two categories of approaches to implement graph as
convolution [24], [31], spectral theory-based approaches, and
spatial-based approaches. Developed from the graph signal x ∗ Gg = U gθ U T x. (7)
processing domain, spectral theory-based graph convolution
is to realize graph convolution by introducing filters from All spectral theory-based graph convolutional networks fol-
the perspective of graph signal processing. Spatial-based low this definition, but the filters gθ are different [32]–[34].
graph convolution is implemented as aggregating features Due to the less computational complexity and locality of
from neighbors, which is similar to the traditional convo- Chebyshev filter [33], ChebNet is adopted in SVS assessment.
lutional neural network. Spatial-based graph convolution is ChebNet defines filters as Chebyshev polynomials of the
more flexible, but it lacks a theoretical basis. Due to the solid eigenvalue diagonal matrix
theoretical basis in spectral theory-based graph convolution, K
spectral theory-based graph convolution is adopted as the gθ = θi Ti ()
˜ (8)
i=0
graph convolutional layer of the designed graph convolutional
network. where  ˜ = 2/ιmax − In , Ti ()˜ = 2T˜ i−1 ()
˜ − Ti−2 ()
˜
After that, the system layer is designed to generate the node with T0 ()˜ = 1, T1 ()
˜ = . ˜ ιmax is the maximal eigenvalue,
representations of the buses after the graph convolutional layer θi is the learnable parameter, and K is the Chebyshev order.
and synthesize the node representations to provide the final The graph convolution of the transformed voltage magnitude
SVS assessment result. Vt r , active power injection Pt r and reactive power injection

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
52 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 20, NO. 1, JANUARY 2023

Q t r with filter gθ can be represented as column denotes the stable probability, and the second column
 K denotes the unstable probability. The corresponding category
Vt r ∗ Ggθ = U θ v i Ti ()
˜ U T Vt r (9) of the column with the highest probability value is regarded
i=0
K as the predicted category. The cross-entropy loss function is
= θ v i Ti ( L̃)Vt r (10) adopted as the loss function of the proposed network.
 i=0

Pt r ∗ Ggθ = U
K
θ pi Ti ()
˜ U T Pt r (11) To facilitate the analysis, the elements of matrix Sb T in the
i=0 first column minus the elements of matrix Sb T in the second
K
= θ pi Ti ( L̃) Pt r (12) column are regarded as the equivalent node weights S in
 i=0
K
the system layer. The element of matrix Bi as in the first
Q t r ∗ Ggθ = U θ q i Ti ()
˜ UT Qtr (13) column minus the element of the matrix Bi as in the second
i=0
K column is regarded as the equivalent bias B in the system
= θ q i Ti ( L̃) Q t r (14) layer. Consequently, the SVS status can also be obtained by
i=0
multiplying the node representations SN and the equivalent
where θ v i , θ pi , θ qi are the learned parameters in the graph
parameters S, then plus the equivalent bias B
convolutional layer, Ti ( L̃) = 2 L̃Ti−1 ( L̃) − Ti−2 ( L̃), T0 ( L̃) is
an identity matrix, T1 ( L̃) = L̃, L̃ = 2L/ιmax − In , L = In − S y = SN ∗ S + B (17)
D −1/2 W D −1/2 , and W represented the topology of the target
where Sy is the equivalent assessment result. When Sy is
region. T ( L̃) is Chebyshev filter matrix, which can be regarded
positive, the assessment result of the target region is stable,
as the key to graph convolution. The locality of Chebyshev
otherwise, it is unstable.
filter matrix T ( L̃) ensures that each bus updates its node infor-
mation according to the information of its neighboring buses, V. C ASE S TUDY
and it is consistent with the spatial distribution characteristics
The proposed methodology is tested on the IEEE 39-bus
of SVS. After the graph convolution, the activation function,
system and real-world Guangdong Power Grid. The results
rectified linear unit (ReLU), is utilized to model the nonlinear
of the IEEE 39-bus system are mainly to illustrate the inter-
characteristics of SVS.
pretability and effectiveness of our method. The discovered
Nin is the input channel number of graph convolution, here
underlying rules and patterns related to SVS are illustrated
Nin = 3. With Nout filters in graph convolution, there are
with the IEEE 39-bus system. The test results on the real-world
Nin Nout (K + 1) parameters in the graph convolutional layer.
Guangdong Power Grid are utilized to further illustrate the
Vgcn , Pgcn , Q gcn are the data processed by the graph convolu-
superiority of our method under unpredictable faults, change-
tional layer. The data processed by different output channels
able operating points, changeable load parameters, and topol-
of the graph convolutional layer are further denote as V g j ,
ogy changes. Samples are generated by PSD-BPA software,
P g j , Q g j , j = 1, 2, . . . , Nout . In Fig. 4, the illustration of
which is a widely used time-domain simulation tool in China.
graph convolutional network for SVS assessment is simplified
The shapelet transform is conducted with MATLAB. The
by setting Nout = 1.
graph convolutional network is conducted with the tensor flow.
2) System Layer: After the graph convolutional layer,
The overall flat form and detailed setups are illustrated in
the system layer is designed to provide the SVS status of the
Appendix.
target region.
First, the node representations of the buses are generated A. IEEE 39-Bus System
based on the data processed by the graph convolutional
1) System Description and Simulation Setting: The one-line
layer. The summation is applied on V g j , P g j , Q g j . Then,
diagram of the IEEE 39-bus system is shown in Fig. 5.
the weighted summation is applied on the output channels,
A receiving-end regional subsystem formed of bus 15, bus 16,
and the processed data are regarded as node representations
bus 17, bus 21, bus 24, and bus 27 is utilized for SVS
of the buses
assessment.
Nout
To simulate the unpredictable faults and changeable
Sn i = θ s j (V g j + P g j + Q g j ) (15)
load parameters in smart grids, 1300 cases are generated
j =1
with PSD-BPA software by setting different fault locations,
where θ s j is the learned channel weight of the jth output fault clearance time, and proportions of induction motor.
channel, and Sn i is the i th node representation in the system Three-phase short-circuit faults are imposed on the system.
layer. 1000 samples are utilized as the training dataset and the
Second, a dense layer plus SoftMax function is utilized remaining 300 samples for testing. The simulation lasts 10 s,
to synthesize the node representations of different buses and and the output label is provided according to the 10-s postfault
provide the final assessment result. SoftMax function can temporal data. The 1-s temporal data after fault clearance,
output the probability values for different classes [35], [36] including voltage magnitude, active power injection, and reac-
tive power injection time series, are used as the input time
Ssys = SoftMax SN ∗ Sb T + Bi as (16)
series of the assessment module. The node admittance matrix
where SN = [Sn 1 Sn 2 , . . . , Sn n ], Sb are the learned node of topology is extracted to form the input topology matrix.
weights, Sb ∈ R2×n , and Bi as is the learned bias, Bi as ∈ R1×2 . As for the setting of hyperparameters in the graph convolu-
Ssys is the final SVS assessment result, Ssys ∈ R1×2 . The first tional network, the batch size is set to 100, the training epoch

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
LUO et al.: GRAPH CONVOLUTIONAL NETWORK-BASED INTERPRETABLE MACHINE LEARNING SCHEME 53

the reception field of the buses grows bigger. The locality of

the reception field is consistent with the spatial distribution
characteristics of SVS, and it is beneficial to promote the
performance of the assessment model.
The absolute values of learned parameters in the graph
convolutional layer are shown in Fig. 9. Some valuable rules
can be discovered from these parameters. The absolute value
of θ v 0 is the largest. The largest parameter θ v 0 is multiplied by
an identity matrix in graph convolution as shown in formula
(10). It means that the voltage magnitude of single nodes
can best reflect the stability of the system. The absolute
values of θ p2 , θ q2 are, respectively, the second-largest and
the third-largest values. As shown in (12), (14), these parame-
ters correspond to the second-order Chebyshev filter matrix.
As most elements in the second-order Chebyshev filter matrix
are nonzero, the reception field of each bus covers almost
the entire target region. Therefore, each bus updates its node
Fig. 5. One-line diagram of the IEEE 39-bus system. information according to the information from almost the
entire target region. The second-largest parameter θ p2 and
is set to 1500, the step in one epoch is set to 10, and the the third-largest parameter θ q2 illustrate that the overall active
learning rate is set to 0.001. Chebyshev order K is set to 2. power injection and reactive power injection of the system can
The number of output channels in graph convolution Nout is have considerable impacts on the stability of the system.
set to 1. 4) System Layer—The Influence of Different Buses on SVS:
2) Shapelet Transform—The Discovered Patterns Related After processed by the graph convolutional layer, the system
to SVS: Shapelet transform is adopted to exploit the tem- layer is designed to generate node representations and provide
poral information and convert the time-series dataset into the final assessment result. As the number of output channels
the distance dataset with the extracted shapelets. Limited by in graph convolution Nout is set to 1, the learned channel
the space, this article only visualizes the partial shapelets weight θ s j is not analyzed here. The node representations
corresponding to the voltage magnitude of bus 15, bus 16, are processed by the dense layer, and these parameters are
and bus 17 in Fig. 6. extracted. The dense layer can be equivalent to Sy = −(5.68 ∗
The extracted shapelets are the key dynamics of the voltage Sn 1 + 10.95 ∗ Sn 2 + 4.61 ∗ Sn 3 + 1.68 ∗ Sn 4 + 2.14 ∗ Sn 5 +
magnitude time series related to short-term voltage insta- 11.96 ∗ Sn 6 ) + 8.07. The absolute values of these equivalent
bility and can best distinguish stable and unstable cases. node weights S can represent the importance degree of one bus
The extracted shapelets can also be regarded as the typical on SVS. The bigger the absolute value of the node weights,
patterns of short-term voltage instability. As shown in Fig. 6, the more important the corresponding bus on SVS.
the extracted shapelets are subsequences of the original time In fact, it is difficult to theoretically calculate the influence
series with low voltage magnitude. As voltage instability is of each bus on voltage instability and demonstrate the effec-
closely related to persistent low voltage, this phenomenon is tiveness of the findings. To illustrate the correctness of the
consistent with the traditional theory of short-term voltage learned weight values, this article manages to calculate the
instability. After extracting shapelets, the multivariate time sensitivities of voltage magnitude to reactive power load by
series can be converted into the distance data by shapelet changing the reactive power of the load buses. Here, the main
transform, and the distance data will be processed by a impact of one bus on the system can be represented by the
subsequent graph convolutional network. largest absolute value of its sensitivities to other buses, and the
3) Graph Convolutional Layer—The Influence of Different quantified impact is regarded as the equivalent sensitivity of
Variables on SVS: In the graph convolutional layer, it can the bus on the system. As shown in Fig. 10, the absolute values
incorporate the topology with the distance data for online SVS of the learned weights have a consistent relationship with the
assessment. Chebyshev filter matrix T ( L̃) is the key element equivalent sensitivities, which demonstrates the effectiveness
to realize graph convolution. The Chebyshev filter matrices of the learned node weights.
are processed by taking their absolute values to facilitate 5) Effectiveness of the Proposed Scheme: To illustrate the
analysis. The processed Chebyshev filter matrices of different effectiveness of the proposed scheme, the training process
orders and their corresponding reception fields are visualized of the graph convolutional network and the performance
in Figs. 7 and 8. comparisons with other typical methods are illustrated in this
As shown in Figs. 7 and 8, the nonzero elements of section.
the Chebyshev filter matrix correspond to the buses in the To illustrate the effectiveness of the proposed graph convo-
reception field. Each bus only exchanges information with the lutional network with the training process, the distributions of
buses in its reception field. The locality of the Chebyshev filter training loss value and training accuracy are shown in Fig. 11.
matrix ensures that buses in the reception field of one bus are As shown in Fig. 11, the performance of the proposed network
their neighboring buses. With the increase of Chebyshev order, has been quickly improved after 300 training epochs. With the

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
54 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 20, NO. 1, JANUARY 2023

Fig. 6. Shapelets and the original time series in their corresponding single-variable time-series datasets. (a) Shapelet and original time series of bus 15.
(b) Shapelet and original time series of bus 16. (c) Shapelet and original time series of bus 17. The black lines in these figures represent shapelets, blue lines
represent the time series of stable samples, and red lines represent the time series of unstable samples.

Fig. 7. (a) Processed first-order Chebyshev filter matrix and the correspond-
ing reception field of bus 15. (b) Processed first-order Chebyshev filter matrix.
The corresponding reception field of bus 15. The numbers in brackets of the Fig. 9. Absolute values of parameters in the graph convolutional layer.
buses correspond to the numbers of Chebyshev filter matrix.

LSTM-based method [21], RVFL-based method [22], and

shapelet transforms combined with traditional convolu-
tional neural network (Shapelet-CNN). The conventional
shapelet-based method utilizes shapelet transform and DT
to assess stability. The LSTM-based model consists of an
LSTM layer, a dense layer, and the sigmoid function. The
RVFL-based method is a randomized learning method in the
form of a hidden-layer feedforward network, and the inputs
are the flattened postfault time series. Shapelet-CNN utilizes
shapelet transform and a traditional convolutional network to
evaluate stability.
As the dataset of the IEEE 39-bus system considers differ-
Fig. 8. (a) Processed second-order Chebyshev filter matrix and the corre-
ent faults, the performances of different stability assessment
sponding reception field of bus 15. (b) Processed second-order Chebyshev methods under unpredictable faults are shown in Table I.
filter matrix. The corresponding reception field of bus 15. The numbers in As shown in Table I, the proposed method has the highest
brackets of the buses correspond to the numbers of Chebyshev filter matrix.
testing accuracy compared with the other methods. This is
due to the combination of temporal data and spatial infor-
increase of training epochs, the training loss value gradually mation in the proposed scheme. The proposed method uti-
decreases, and the training accuracy increases. At the end of lizes the network topology and temporal trajectory as the
the training process, the training loss value stables near zero, input information, which fully exploits the spatial-temporal
and the training accuracy stables near 98%, which illustrates characteristics of short-term voltage instability and results in
the effectiveness of the proposed network. decent performances. The other four methods only utilize the
To comprehensively evaluate the performances of the pro- temporal trajectory as the learning input. Insufficient use of
posed scheme, it is compared with other stability assessment spatial-temporal information results in relatively poor perfor-
methods, namely the conventional shapelet-based method [19], mance. In addition, the LSTM-based method, RVFL-based

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
LUO et al.: GRAPH CONVOLUTIONAL NETWORK-BASED INTERPRETABLE MACHINE LEARNING SCHEME 55

TABLE I
P ERFORMANCE C OMPARISON OF D IFFERENT A PPROACHES
U NDER U NPREDICTABLE FAULTS

for power systems, data residing on a graph makes it hard

for CNN to work well on power system SVS assessment.
Graph convolutional network (GCN) can be considered as the
generalization of CNN. It is designed for graph data and can
exploit the characteristics of the graph. Therefore, GCN is
more suitable for power system SVS assessment and shows
better performance compared with other methods. Besides,
Fig. 10. Absolute values of node weights in the system layer and the the proposed method has decent interpretability, which has
equivalent sensitivities. (a) Absolute values of node weights in the system been illustrated in Sections V-A2–V-A4.
layer. (b) Equivalent sensitivities of the buses.
B. Guangdong Power Grid
1) System Description and Simulation Settings: The pro-
posed method is also tested on the real-world Guangdong
Power Grid, which is a typical receiving-end system of China
Southern Power Grid and more likely to suffer from short-term
voltage instability. The backbone structure of Guangdong
Power Grid is shown in Fig. 12. There are 101 high-voltage
buses in Guangdong Power Grid, and they are all utilized for
online stability assessment. The loads in the test system are
represented by an induction motor and static loads.
To simulate the challenging environment in practical smart
grids, cases are generated with full consideration of unpre-
dictable faults, changeable load parameters, different operating
points, and topology changes. 4800 cases are generated with
PSD-BPA software by setting different fault locations, differ-
ent fault clearing times, and different induction motor ratios.
240 different fault locations in Guangdong Power Grid are
randomly selected to impose three-phase short-circuit faults.
The fault clearing time is set to 0.1, 0.2, 0.3, 0.4, 0.5 s. The
Fig. 11. Accuracy and loss value in training process. (a) Training accuracy. induction motor ratio is set to 0.3, 0.5, 0.7, 0.9. Besides,
(b) Loss value in training process. 1500 cases are generated by setting different operating points,
random fault locations, random fault clearing time, and topol-
method, and Shapelet-CNN also lack interpretability. The ogy changes. These 6300 cases are randomly mixed to test the
RVFL method has the highest training accuracy, but the performances of the proposed scheme. 5500 cases are utilized
lowest testing accuracy. The reasons can attribute to two for training, and the remaining cases are for testing.
aspects. On the one hand, this method does not conduct As for the setting of hyperparameters in the graph convolu-
feature extraction to obtain effective features. On the other tional network, the batch size is set to 100, training epoch is
hand, this method is in the form of a single hidden layer set to 1500, the step in one epoch is set to 55, and the learning
feedforward network. The shallow structure and lack of feature rate is set to 0.001. Chebyshev order K is set to 3. The output
extraction can result in overfitting, which further leads to channel in the graph convolution Nout is set to 50, and the
poor performance. As for the Shapelet-CNN method, CNN is distance data transformed by shapelets are normalized.
more suitable for grid data, such as pictures or videos. Local 2) Performance Under Changeable Operating Points,
connection and global sharing mechanism in CNN make it Unpredictable Faults and Topology Changes: To comprehen-
better to find the local features in figures or video. However, sively evaluate the performances of the proposed scheme in

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
56 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 20, NO. 1, JANUARY 2023

Fig. 12. Backbone structure of real-world Guangdong power grid.

TABLE II TABLE III

P ERFORMANCE C OMPARISON U NDER C HANGEABLE O PERATING P OINTS , S ETUPS OF E XPERIMENTS
U NPREDICTABLE FAULTS AND T OPOLOGY C HANGES

topology in the proposed scheme. On the one hand, SVS

presents spatial characteristics and the introduction of topology
can improve the accuracy of the learning model. On the
other hand, when the topology changes, the proposed scheme
can adapt to the changes in topology. However, the other
methods do not incorporate topology information in their
models, which results in poor performances. The RVFL-based
method presents the worst performance. This is because the
RVFL-based method is in the form of a single hidden-layer
Fig. 13. Illustration of overall validation platform.
feedforward network. The shallow structure makes it more
susceptible to disturbance.
real-world Guangdong Power Grid under changeable oper-
ating points, unpredictable faults and topology changes,
it is compared with the conventional shapelet-based method VI. C ONCLUSION
(shapelet+DT) [19], LSTM-based method [21], and RVFL Smart grids are exposed to different kinds of uncertainty and
method [22]. unpredictable faults. Taking SVS assessment in smart grids as
As shown in Table II, the proposed method also has the a paradigm, this article develops a novel machine learning
highest testing accuracy in Guangdong Power Grid com- scheme to fully exploit the spatial-temporal information in
pared with the other three methods. It can attribute to the SVS dynamics. It first utilizes the shapelet transform to extract
spatial-temporal incorporation and the introduction of network key dynamics and convert multivariate time series into distance

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
LUO et al.: GRAPH CONVOLUTIONAL NETWORK-BASED INTERPRETABLE MACHINE LEARNING SCHEME 57

Algorithms Shapelet Transform Algorithms Graph Convolutional Network Learning

Input: multivariate training dataset DS, N cases in the Input: The transformed features Vtr , Ptr , Q tr , corresponding
dataset, d variables in the dataset, Mi shapelet candidates labels ltrue objective function f (θ ), network parameters θ that
for single-variable dataset DSi , number of splitting points v need to be trained
Output: transformed data by the searched shapelets Output: Well-trained graph convolutional network
(1) for i = 1 to d do (1) Set hyperparameters α = 0.001, β1 = 0.9, β2 = 0.999,
(2) set I G max = 0  = 10−8 , emax , and exponential decay rate ι ∈ [0, 1]
(3) for j = 1 to Mi do (2) Initialize network parameters θ
(4) Calculate the distance between shapelet candidate (3) Initialize 1st moment vector m 0 = 0, 2st
sc j and N cases in single-variable moment vector v 0 = 0
dataset DSi according to formula (3) (4) for k = 1 to emax do
(5) set I G j = 0 (5) β1,k = β1 ιk−1
(6) for k = 1 to v do (6) gt = θ f k (θk−1 )
(7) Calculate the information gain I G spk (7) m k = β1,k · m k−1 + (1 − β1,k ) · gk
under split point spk (8) v k = β2 · v k−1 + (1 − β2 ) · gk2
(8) if I G spk > I G j (9) m k = m k /(1 − β1k )
(9) I G j = I G spk (10) v k = v k /(1 − β2k ) √
(10) endif (11) θk = θk−1 − α · m k /( v k + )
(11) endfor (12) endfor
(12) if I G j > I G max (13) The graph convolutional network with trained
(13) I G max = I G j parameters θ for online assessment
(14) shapi = sc j
(15) endif
(16) endfor graph convolutional network learning. Data preprocess and
(17) Calculate the distance di si between shapelet shapelet transform is conducted with MATLAB. The graph
shapi and N cases in single-variable dataset convolutional network is conducted with tensor flow. All the
DSi according to formula (3) simulation and tests are conducted on a workstation, which is
(18) endfor equipped with Intel Core i7-6800K CPU (3.4 GHz*12) and
(19) The transformed distance values di si form NVIDIA GeForce GTX 1080 GPU.
the transformed features Vtr , Ptr , Q tr for subsequent 2) Detailed Setups: The summarized setups of the experi-
graph convolutional network learning ment are shown in Table III.

B. Algorithms of Shapelet Transform and Graph

data. Then, it designs a graph convolutional network to incor- Convolutional Network Learning
porate distance data with spatial information for online SVS
The algorithms of shapelet transform and graph convo-
assessment. The test results of the IEEE 39-bus system indicate
lutional network learning are as follows. Adam algorithm
that the proposed method is capable of discovering certain
is adopted as the graph convolutional network learning
valuable underlying rules and patterns related to SVS. Test
algorithm.
results of the IEEE 39-bus system and real-world Guangdong
Power Grid illustrate that the proposed method can achieve
R EFERENCES
higher performances than other methods under changeable
operating points, unpredictable faults, and topology changes. [1] AEMO. (2016). Preliminary Report-Black System Event in
South Australia. Accessed: Sep. 28, 2016. [Online]. Available:
https://www.aemo.com.au/Media-Centre/-/media/BE174B1732CB4B3A
A PPENDIX BB74BD507664B270.ashx
[2] T. V. Cutsem and C. Vournas, Voltage Stability of Electric Power
A. Illustration of Overall Platform and the Detailed Setups Systems. Norwell, MA, USA: Kluwer, 1998.
[3] P. Kundur et al., “Definition and classification of power system stability
1) Illustration of Overall Platform: The illustration of over- IEEE/CIGRE joint task force on stability terms and definitions,” IEEE
all validation platform is shown in Fig. 13. The power system Trans. Power Syst., vol. 19, no. 3, pp. 1387–1401, Aug. 2004.
is modeled in PSD-BPA software, and samples are simulated [4] N. Hatziargyriou et al., “Definition and classification of power system
stability revisited & extended,” IEEE Trans. Power Syst., vol. 36, no. 4,
in the software for subsequent assessment model construction. pp. 3271–3281, Jul. 2020.
PSD-BPA software is widely used by power grid companies [5] K. Meng, Z. Y. Dong, K. P. Wong, Y. Xu, and F. J. Luo, “Speed-up the
in China, which is capable of analyzing the dynamic behavior computing efficiency of power system simulator for engineering-based
power system transient stability simulations,” IET Generat., Transmiss.
of a power system when it is subjected to various disturbances Distrib., vol. 4, no. 5, pp. 652–661, May 2010.
in a certain operation condition. The dynamics of the postfault [6] V. Martinez and A. Juan, “Real-time simulation technologies in engi-
power system can be described as a set of differential algebraic neering,” in Transient Analysis of Power System: Solution Techniques,
Tools and Applications. Piscataway, NJ, USA: Wiley, 2014, pp. 72–99.
equations (DAEs). [7] D. Osipov and K. Sun, “Adaptive nonlinear model reduction for fast
The implement of the assessment module mainly con- power system simulation,” IEEE Trans. Power Syst., vol. 33, no. 6,
sists of three parts: data preprocess, shapelet transform, and pp. 6746–6754, Nov. 2018.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.
58 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 20, NO. 1, JANUARY 2023

[8] I. Nagel, L. Fabre, M. Pastre, F. Krummenacher, R. Cherkaoui, and [30] J. Lines, L. M. Davis, J. Hills, and A. J. Bagnall, “A shapelet transform
M. Kayal, “High-speed power system transient stability simulation using for time series classification,” in Proc. 18th ACM SIGKDD Int. Conf.
highly dedicated hardware,” IEEE Trans. Power Syst., vol. 28, no. 4, Knowl. Discovery Data Mining, 2012, pp. 289–297.
pp. 4218–4227, Nov. 2013. [31] P. W. Battaglia et al., “Relational inductive biases, deep learning,
[9] S. Huang, Y. Chen, C. Shen, W. Xue, and J. Wang, “Feasibility study and graph networks,” 2018, arXiv:1806.01261. [Online]. Available:
on online DSA through distributed time domain simulations in WAN,” http://arxiv.org/abs/1806.01261
IEEE Trans. Power Syst., vol. 27, no. 3, pp. 1214–1224, Aug. 2012. [32] J. Bruna, W. Zaremba, A. Szlam, and Y. Lecun, “Spectral networks
[10] NERC/WECC Planning Standards. Accessed: Apr. 15, 2015. and locally connected networks on graphs,” 2013, arXiv:1312.6203.
[Online]. Available: https://www.wecc.biz/library/Library/Planning [Online]. Available: https://arxiv.org/abs/1312.6203
%20Committee%20Handbook/WECC-NERC%20Planning [33] M. Defferrard, X. Bresson, and P. Vandergheynst, “Convolutional neural
%20Standards.pdf networks on graphs with fast localized spectral filtering,” in Proc. Adv.
[11] Z. Liu, “Voltage stability and preventive control in China,” in Proc. IFAC Neural Inf. Process. Syst., 2016, pp. 3837–3845.
Symp. Power Plants Power Syst. Control, Seoul, South Korea, 2003, [34] T. N. Kipf and M. Welling, “Semi-supervised classification with graph
pp. 39–42. convolutional networks,” 2016, arXiv:1609.02907. [Online]. Available:
http://arxiv.org/abs/1609.02907
[12] K. L. Praprost and K. A. Loparo, “An energy function method for deter-
[35] A. Krizhevsky et al., “ImageNet classification with deep convolutional
mining voltage collapse during a power system transient,” IEEE Trans.
neural networks,” in Proc. Adv. Neural Inf. Process. Syst., vol. 25, 2012,
Circuits Syst. I, Fundam. Theory Appl., vol. 41, no. 10, pp. 635–651,
pp. 1097–1105.
Oct. 1994.
[36] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolu-
[13] G. M. Hung, L. Zhao, and X. Song, “A new bifurcation analysis for tional networks,” in Proc. Eur. Conf. Comput. Vis., 2014, pp. 818–833.
power system dynamic voltage stability studies,” in Proc. IEEE Power [37] M. G. Brown, M. Biswal, S. Brahma, S. J. Ranade, and H. Cao,
Eng. Soc. Winter Meeting Conf., New York, NY, USA, Jan. 2002, “Characterizing and quantifying noise in PMU data,” in Proc. IEEE
pp. 882–887. Power Energy Soc. Gen. Meeting, Jul. 2016, pp. 1–5.
[14] S. Dasgupta, M. Paramasivam, U. Vaidya, and V. Ajjarapu, “Real-time
monitoring of short-term voltage stability using PMU data,” IEEE Trans.
Power Syst., vol. 28, no. 4, pp. 3702–3711, Nov. 2013. Yonghong Luo received the B.S. degree in electrical
[15] Y. Dong, X. Xie, K. Wang, B. Zhou, and Q. Jiang, “An emergency- engineering from Chongqing University, Chongqing,
demand-response based under speed load shedding scheme to improve China, in 2017. She is currently pursuing the Ph.D.
short-term voltage stability,” IEEE Trans. Power Syst., vol. 32, no. 5, degree with Tsinghua University, Beijing, China.
pp. 3726–3735, Sep. 2017. Her research mainly deals with data-driven power
[16] D. Q. Zhou, U. D. Annakkage, and A. D. Rajapakse, “Online monitoring system voltage stability assessment and control.
of voltage stability margin using an artificial neural network,” IEEE
Trans. Power Syst., vol. 25, no. 3, pp. 1566–1574, Aug. 2010.
[17] E. Voumvoulakis, A. Gavoyiannis, and N. Hatziargyriou, “Decision trees
for dynamic security assessment and load shedding scheme,” in Proc.
IEEE Power Eng. Soc. Gen. Meeting, Jun. 2006, p. 6.
[18] F. R. Gomez, A. D. Rajapakse, U. D. Annakkage, and I. T. Fernando, Chao Lu (Senior Member, IEEE) received the B.S.
“Support vector machine-based algorithm for post-fault transient stabil- and Ph.D. degrees in electrical engineering from
ity status prediction using synchronized measurements,” IEEE Trans. Tsinghua University, Beijing, China, in 1999 and
Power Syst., vol. 26, no. 3, pp. 1474–1483, Aug. 2011. 2005, respectively.
[19] L. Zhu, C. Lu, and Y. Sun, “Time series shapelet classification based He is currently an Associate Professor with the
online short-term voltage stability assessment,” IEEE Trans. Power Syst., Department of Electrical Engineering, Tsinghua
vol. 31, no. 2, pp. 1430–1439, Mar. 2016. University. His research interests include power sys-
[20] L. Zhu, C. Lu, Z. Y. Dong, and C. Hong, “Imbalance learning machine- tem analysis and control, and intelligent control
based power system short-term voltage stability assessment,” IEEE applications.
Trans. Ind. Informat., vol. 13, no. 5, pp. 2533–2543, Oct. 2017.
[21] J. J. Q. Yu, D. J. Hill, A. Y. S. Lam, J. Gu, and V. O. K. Li, “Intelligent
time-adaptive transient stability assessment system,” IEEE Trans. Power
Lipeng Zhu (Member, IEEE) received the B.S.
Syst., vol. 33, no. 1, pp. 1049–1058, Jan. 2018.
degree in electrical engineering from the Huazhong
[22] Y. Zhang, Y. Xu, R. Zhang, and Z. Y. Dong, “A missing-data tolerant University of Science and Technology, Wuhan,
method for data-driven short-term voltage stability assessment of power China, in 2012, the M.S. degree in electrical engi-
systems,” IEEE Trans. Smart Grid, vol. 10, no. 5, pp. 5663–5674, neering from Wuhan University, Wuhan, in 2015,
Sep. 2019. and the Ph.D. degree in electrical engineering
[23] L. Zhu, C. Lu, I. Kamwa, and H. Zeng, “Spatial–temporal feature Tsinghua University, Beijing, China, in 2018.
learning in smart grids: A case study on short-term voltage stability He worked as a Post-Doctoral Fellow/Senior
assessment,” IEEE Trans. Ind. Informat., vol. 16, no. 3, pp. 1470–1482, Research Assistant with The University of Hong
Mar. 2020. Kong, Hong Kong, from 2018 to 2021. He has
[24] Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, “A com- been appointed as an Associate Professor at Hunan
prehensive survey on graph neural networks,” 2019, arXiv:1901.00596. University, Changsha, China, since 2021. His research interests mainly include
[Online]. Available: http://arxiv.org/abs/1901.00596 spatial-temporal energy data analysis, Internet of Things (IoT) data manage-
[25] B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional ment and analytics in smart grids, synchrophasor measurement technologies,
networks: A deep learning framework for traffic forecasting,” in Proc. and data-driven power system stability and control.
27th Int. Joint Conf. Artif. Intell., Jul. 2018, pp. 3634–3640.
[26] J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, Jie Song (Senior Member, IEEE) received the B.S.
“Neural message passing for quantum chemistry,” in Proc. 34th Int. degree in applied mathematics from Peking Univer-
Conf. Mach. Learn. (ICML), vol. 70, 2017, pp. 1263–1272. sity, Beijing, China, in 2004, and the M.S. and Ph.D.
[27] K. Chen, J. Hu, Y. Zhang, Z. Yu, and J. He, “Fault location in power degrees from the Department of Industrial Engineer-
distribution systems via deep graph convolutional networks,” IEEE ing, Tsinghua University, Beijing, in 2007 and 2010,
J. Sel. Areas Commun., vol. 38, no. 1, pp. 119–131, Jan. 2020. respectively.
[28] J. D. Weber and T. J. Overbye, “Voltage contours for power system She is an Associate Professor with the Department
visualization,” IEEE Trans. Power Syst., vol. 15, no. 1, pp. 404–409, of Industrial Engineering and Management, Peking
Feb. 2000. University. Her research interests are simulation
[29] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization, stochastic modeling in the application
optimization,” 2014, arXiv:1412.6980. [Online]. Available: areas of logistics, healthcare and production.
https://arxiv.org/abs/1412.6980 Dr. Song is a member of INFORMS.

Authorized licensed use limited to: Shanghai Jiaotong University. Downloaded on September 26,2024 at 10:25:30 UTC from IEEE Xplore. Restrictions apply.