Received April 1, 2020, accepted April 25, 2020, date of publication April 29, 2020, date of current version

June 1, 2020.
Digital Object Identifier 10.1109/ACCESS.2020.2991263

Recurrent Graph Convolutional Network-Based

Multi-Task Transient Stability Assessment
Framework in Power System
JIYU HUANG 1 , LIN GUAN 1 , (Member, IEEE), YINSHENG SU 2, HAICHENG YAO 2,

MENGXUAN GUO 1 , AND ZHI ZHONG 1

1 School of Electric Power, South China University of Technology, Guangzhou 510641, China
2 CSG Power Dispatching and Control Center, Guangzhou 510663, China
Corresponding author: Lin Guan ([email protected])
This work was supported by the China Southern Power Grid Research Project under Grant ZDKJXM20180084.

ABSTRACT Reliable online transient stability assessment (TSA) is fundamentally required for power
system operation security. Compared with time-costly classical digital simulation methods, data-driven deep
learning (DL) methods provide a promising technique to build a TSA model. However, general DL models
show poor adaptability to the variation of power system topology. In this paper, we propose a new graph-
based framework, which is termed as recurrent graph convolutional network based multi-task TSA (RGCN-
MT-TSA). Both the graph convolutional network (GCN) and the long short-term memory (LSTM) unit are
aggregated to form the recurrent graph convolutional network (RGCN), where the GCN explicitly integrate
the bus (node) states with the topological characteristics while the LSTM subsequently captures the temporal
features. We also propose a multi-task learning (MTL) scheme, which provides joint training of stability
classification (Task-1) as well as critical generator identification (Task-2) in the framework, and accelerate
the process with parallel computing. Test results on IEEE 39 Bus system and IEEE 300 Bus system indicate
the superiority of the proposed scheme over existing models, as well as its robustness under various scenarios.

INDEX TERMS Deep graph-based learning, transient stability assessment (TSA), graph convolutional
network (GCN), recurrent graph convolutional network (RGCN), multi-task learning (MTL).

NOMENCLATURE GRU Gated recurrent unit

ABBREVIATIONS J Jaccard similarity
ACC Accuracy JACC Jaccard accuracy
ADM Assistant decision-making JACCU Jaccard accuracy of unstable
BN Batch normalization LN Layer normalization
CCT Critical clearing time MA Miss alarm
CE Cross-entropy ML Machine learning
CNN Convolutional neural network MLP Multilayer perceptron
DL Deep learning MTL Multi-task learning
DT Decision tree PMU Phasor measurement unit
EJS Expand Jaccard similarity RGCN Recurrent graph convolutional network
ELM Extreme learning machine RGCN-MT-TSA Recurrent graph convolutional
FA False alarm network based multi-task transient
FC Full connected stability assessment
GCN Graph convolutional network SAE Stacked auto encoder
SNR Signal to noise ratio
SVM Support vector machine
The associate editor coordinating the review of this manuscript and TDS Time domain simulation
approving it for publication was Xiaorong Xie . TSA Transient stability assessment

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 8, 2020 93283
 J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

SYMBOLS OF RGCN-MT-TSA OTHER SYMBOLS

G, Gc , Gs The sets of generators, critical generators s The set
and significant generators δ(·) The pulse function
A˜00 , X 0 , O0 The block matrix of adjacency, input node ε, ε0 The Gaussian white noise and color noise
features and output convolved features N (0, 1) The Gaussian distribution
c̃ The vector of the confidence of the cate- O(·) The operation complexity
gories or labels
G̃c , G̃s The predicted sets of critical generators and
significant generators I. INTRODUCTION
H l , W l , bl The input, layer-specific weight and bias For the past decades, there has been an increasing demand
matrics of the l th layer of loads and large-scale deployment of low-inertia converter-
p, p0 The learnable and trained parameter based renewable generation in power system [1]. These
w0 , b0 The trained parameter of Task-1 of the changes bring challenges for the security of power system
sharing hidden layers operation, especially for the system stability under faults
w, b The learnable network parameters [2]. Fast online transient stability assessment (TSA) is a
xt , ct , ht The inputs, cell states and hidden states fundamental tool to provide early-warning for instability and
vectors of LSTM at the moment t instruction for preventive control of the system.
Y The node admittance matrix Conventionally, the core of online TSA is the time domain
z The predictive vector for the categories or simulation (TDS), which solves the high-dimensional non-
labels linear differential-algebraic model of power system [3].
G Graph Although, parallel computing [4], improved integration meth-
V, E The set of nodes and edges ods [5], energy functions based direct methods [6], [7] are
A, Ã, Ã0 The adjacency matrix without and with developed to accelerate the simulation, huge computational
self-loop, renormalization trick burden brought by emerging converter modules and expand-
c The vector of the categories or labels ing of system scale are still the first challenge for the online
D, D̃ The diagonal node degree matrix without TSA.
and with self-loop Data-driven machine learning (ML) techniques provide a
δ, δa The thresholds new train of thought, where the TSA model is established
η Transient stability index offline with batches of training samples and applied to rapidly
σ The activation function assess new contingencies online. Deployment of the pha-
|1δ|max The absolute value of the maximum rotor sor measurement units (PMUs) provides fast and accurate
angle of separation between any two gen- dynamic information of the system for ML models. Hence,
erators during the simulation time a variety of ML approaches, such as decision tree (DT) [8],
|1δ|i The absolute value of separation between support vector machine (SVM) [9], multilayer perceptrons
generator i and the reference generator dur- (MLPs) [10] and extreme learning machine (ELM) [11] are
ing the simulation time applied to the online TSA modeling, and the targets cover var-
|yi,j |, |yi,j |max The module of an element of Y and the max ious TSA scenarios, e.g. the stability classification, the crit-
of the modules ical clearing time (CCT) prediction and the critical genera-
C, F The size of input and output features of tor identification etc. However, for these schemes, the input
each bus (node) features should be carefully constructed based on the expert
L The number of edges experience.
M The number of snapshots Deep learning (DL) can extract fine-grained features from
N The number of buses (nodes) big raw data with the help of more hidden layers or even
NG The number of generators sharing of modules. Zhu et al. [12] employ a two-stage TSA
T The length of the observation window method with stacked auto encoder (SAE). James J Q and his
colleagues exploit long short-term memory (LSTM) [13] and
αi , αu , αs The balanced factor of the ith sample, unsta- stacked gated recurrent units (GRUs) [14] in the correlation
ble samples and stable samples learning of temporal series. Another popular technique is the
β1 , β2 The regularization weights convolutional neural network (CNN), which is capable of
β̃i,j The correction factor learning spatial representations of data. For the TSA problem,
fs The sampling frequency time-series data are arranged into multi-channel metrics (i.e.,
t0 , tc The moment of fault occurrence and fault each element in a matrix refers to a vector of observations)
clearance such that CNN can learn the mapping function from the

93284 VOLUME 8, 2020

J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

inputs to the stability labels [15], [16]. GUPTA et al. [17] carried out on IEEE 39 Bus system and IEEE 300 Bus system
consider a new description of measured generator data as to validate the generalization and robustness of the proposed
an image with each value in the data matrix represented as scheme.
color intensity. The visual dissimilarity of images of stable Generally, this paper is highlighted with the following
and unstable cases is then distinguished by CNN, which is contributions:
trained simultaneously for both stability classification and 1) The adjacency matrix of GCN is designed to represen-
critical generators (i.e generators most affected under the tatively describe the graph topology of power system
disturbance) identification. In [18], the authors adopt discrete and effectively reflect the inherent physical character-
Fourier transform to obtain spectrum from the fault-on gener- istics.
ator trajectories and arrange them into 2D images, such that 2) A block-diagonal sparse matrix is constructed with
CNN can achieve good performance in refined CCT regres- each block corresponding to the adjacency matrix of
sions. Shi et al. [19] construct larger images with variables of a graph. Such an attempt supports batch-wise process
all buses and verify the effectiveness of CNN on instability of graph data and fully utilize parallel computing.
mode (e.g., caused by insufficient synchronizing or damping 3) A cost-sensitive cross-entropy function is designed to
torque) prediction. Aimed at a large scale of contingency deal with category-imbalanced problem in critical gen-
screening, Yan et al. [20] introduce cascade CNNs in stability erator identification.
probability prediction for early TDS termination without 4) A soft sharing scheme is proposed to accelerate the
losses of accuracy, based on continuously refreshing them- multi-task training.
selves with the increase of labeled TDS outputs. Nonethe-
The rest of this paper is organized as follows. Section II
less, all above DL models are not specialized in exploring
introduces the design of RGCN. Section III presents the data
of observations with explicit topological graph correlation,
preprocessing and the application of the RGCN-MT-TSA
where power system is such an interconnected network of
framework with offline training tricks. Section IV demon-
generators and loads [21]. A series of studies [22]–[24] estab-
strates cases study on two different benchmark systems and
lish that there exists a close relationship between topology
various scenarios. The conclusion is discussed in Section V.
and transient stability. As a result, changes of power system
topology, which is frequently triggered by maintenance or
II. RECURRENT GRAPH CONVOLUTIONAL NETWORK
faults, may deteriorate the performance of TSA models based
In this paper, we propose a novel aggregating network
on SAE, CNN or recurrent methods.
structure, named as the recurrent graph convolutional net-
Correspondingly, graph convolutional network (GCN)
work (RGCN) and shown in Fig. 1. RGCN consists of four
develops an explicit way of integrating topological structure
cascading modules, where GCN and LSTM are hierarchical
into the convolution algorithm [25]. GCN has been proved
modules while the time pooling and classifier are single ones.
extremely useful for graph analysis tasks in a wide variety
GCN and LSTM play a critical role in addressing graphical
of application areas, such as knowledge graph learning [26],
and temporal feature extraction. Then the time pooling mod-
text classification [27] and recommender system prediction
ule aggregates features from the whole time steps and the
[28]. The basic idea behind GCN is to distill the high-
classifier provides final discrimination.
dimensional information about a node’s graph neighborhood
Both the GCN and the LSTM adopt hierarchical stacked
into a vector representation with dimension reduction. With
structure containing also the normalization layers and full
this in mind, GCNs are also employed in the field of power
connected (FC) layers. We select two types of normaliza-
system recently, to deal with fault location and load shed-
tion layers, i.e. the batch normalization (BN) and the layer
ding [29], [30]. Specially, under the context of TSA, James
normalization (LN). Details of these layers and modules are
J Q et al. [31] designs a GCN model for recovery of the
introduced as follows.
missing PMU data and indicate lower errors than existing
implementation [14].
In this paper, we propose a new recurrent graph convo-
lutional network (RGCN) for spatio-temporal feature inte-
gration. RGCN adopts cascading architecture where the
improved GCN modules process measurements at nodes con-
sidering the power system graph structures firstly, and then
the LSTM modules accomplish the temporal fusion. Based on
the RGCN, we further design a multi-task TSA framework,
named as the RGCN-MT-TSA in the paper. Multi-task learn-
ing (MTL) is exploited for joint training of two subtasks, i.e.
stability classification (Task-1) and critical generator identi-
fication (Task-2). The proposed framework provides early-
warning based on the results of both tasks such that they
can verify each other spontaneously. Comprehensive tests are FIGURE 1. Cascade architecture of RGCN.

VOLUME 8, 2020 93285

 J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

A. GRAPH CONVOLUTIONAL NETWORK LAYER

CNN performs neighborhood information aggregating on the
input data, which is elaborately designed for image-type of
signals in Euclidean space. However, fixed filters face diffi-
culty in addressing graph data with irregularity connections
of nodes. Instead of convolution with geographical neigh-
borhoods, graph convolutional filters concentrate more on
the correlation over graph edges provided by an adjacency
matrix. Let G = (V, E) denotes an undirected graph where
V ∈ RN denotes the set of nodes and E ∈ RL denotes the set
of edges. Then we have the adjacency matrix A and define a
renormalization trick as: FIGURE 3. An illustration of LSTM cell.
−1/2 −1/2
Ã0 = D̃ ÃD̃ (1) C. FULL CONNECTED LAYER
where à = A + I N represents an adjacency matrix with self- For an FC layer, the output is a linear transformation of the
connections with Pits diagonal node degree matrix D̃ calcu- inputs. Multiple FC layers [10] map the temporal output of
lated by D̃i,j = LSTM to prediction results at each time step.
j Ãi,j . The graph-based propagation for
a GCN layer that maps a N × C input feature matrix to
a new N × F output matrix is performed by a nonlinear D. NORMALIZATION LAYER
function [32]: In order to reduce the impact of the parameters of all preced-
ing layers to the inputs of current layer in the training process,
H l+1 = σ (f (Ã0 , H l )) = σ (Ã0 H l W l + bl ) (2) we adopt the normalization layers, mainly including batch
l
where H denotes the input matrix of the l th
layer, i.e., output normalization (BN) and layer normalization (LN), to fix the
of the (l − 1)th layer. W l ∈ RC×F , bl ∈ RN ×F are layer- means and variances of each layer’s inputs. The effectiveness
specific weight and bias matrics, respectively. σ is an acti- of BN has been widely proved in conventional CNN [35],
vation function, where we adopt ReLu(·) in the rest of this and it is introduced between GCN layers similarly in this
paper. paper. It is suggested by Kim et al. [36] that LN, i.e., a
Fig. 2 demonstrates the feed-forward propagation, where normalization for neurons of the same layer, is preferred to
each node obtains information from first-order neighbors recurrent architecture like LSTM and FC. The normalization
under (2), i.e., a message passing mechanism [33], and then promotes the robustness to noises and thus, the model benefits
updates its representation. As the layers stacked, nodes incre- from good performance without fine-tuning parameters of
mentally aggregate more and more ‘‘message’’ from further dropout layer [37] or L1 regularization.
reaches of the graph. To alleviate the problem of overfitting,
all the GCN modules share the same parameters. E. TIME POOLING MODULE
With sequential inputs of T steps and static outputs, we adopt
a global mean pooling approach, called time pooling module
to merge per-time step predictions [z1 , z2 , . . . , zT ] into a
single prediction z as:
T
1X
z= zt (3)
T
t=0
where z denotes a vector of predictive values for the cate-
gories or labels.

FIGURE 2. Feed-forward propagation with message passing mechanism. F. CLASSIFIER

In our application, two different targets, i.e. stability classi-
fication and critical generator identification are addressed.
B. LONG SHORT-TERM MEMORY LAYER These two targets belong to different fields in ML,
As one of the most popular variants of recurrent neural net- i.e., binary-category classifier and multi-label classifier
work (RNN), LSTM overcomes the problem of vanishing respectively.
gradients in deep RNN, which is designed to pass the infor- • Binary-category classifier
mation of previous time steps to subsequent ones. In Fig. 3, The softmax function is utilized and we obtain the con-
an LSTM cell [34] typically comprises three gates: input, fidence c̃i for the category i as:
forget and output gates. xt , ct and ht represent the vectors
ezi
of input, cell and hidden states, and σ refers to the sigmoid c̃i = P z (j = 1, 2) (4)
function σ (x) = 1/(1 + e−x ). ie
i

93286 VOLUME 8, 2020

J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

where c̃1 + c̃2 = 1. The system is predicted as unstable MTL is essentially a multi-objective optimization,
when c̃1 > 0.5, which is labeled as [1, 0]T . Otherwise, i.e., multiple loss functions are simultaneously minimized
the system is stable and labeled as [0, 1]T . based on gradient descent. In the context of DL, hard and
• Multi-label classifier soft parameter sharing [38], [39] are the most commonly
For multi-label classification, each sample is simulta- used settings for MTL. The former requires all the tasks to
neously associated with a set of labels. We assign a share the same subset of the hidden layers and thus effectively
0/1 binary code for each label to represent False/True. alleviates the chance to overfit. However, the drawback is
Then the problem can be actually decomposed to multi- that we might solve the multi-objective programming directly
ple related binary-category learning. We adopt the sig- to obtain the common representation that captures multiple
moid function that limits z to c̃(c̃i ∈ (0, 1)), where c̃ tasks. Another practical way is to merge the weighted loss
denotes a vector of the confidence of all labels. Define functions and thus optimize the single-objective problem.
a threshold δ, a label i is predicted to be true one when The assignment of the weights among the tasks implements
its confidence c̃i ∈ (0, δ], and the final output is a binary a direct effect on the generalization of all the tasks. Here,
vector. we adopt the latter setting.
Going back to the critical generator identification prob- As shown in Fig. 4, task-1 and task-2 have their own mod-
lem, each label is corresponding to a generator and els and parameters, but we regularize the distance between the
the set of true labels from output refers to that of the parameters of the two models to encourage their parameters
predicted critical generators G̃c . Furthermore, if there is to be similar. Considering that the operation complexity for
a set of labels whose confidence belongs to (δ, δa ) with Task-2 is significantly larger than Task-1 with when there
δ < δa < 1, we say this set refers to the set of predicted are tens of labels, Task-1 will be trained at first and its spa-
significant generators G̃s . tial feature extractor, GCN modules, are then transferred to
Task-2 as an initial setting. A regularization term is merged in
III. THE RGCN-MT-TSA FRAMEWORK the loss function of Task-2 to minimize the distance between
A. GENERAL INTRODUCTION OF THE FRAMEWORK its parameters and the trained parameters of Task-1. With the
In this paper, we propose a multi-task TSA solution based on benefit of such a generalization design, we use the implicit
RGCN. The framework is shown in Fig. 4. experience in Task-1 as guidance for the parameters optimiza-
tion of Task-2 and the multi-objective problem is simplified
to a two-stage single-objective optimization.

2) INPUT VECTOR
We choose three physical variables of each bus (node) to form
the input space of GCN. They are the bus voltage magnitude,
the bus relative phase and the rotor speeds of generators
connected to the power plant bus, i.e., the derivative of rotor
angles with respect to time. For the load buses, their values of
the third variable (rotor speed) are uniformly set to zero. The
observation time window of the model inputs starts from the
moment of fault occurrence t0 and ends at the fault clearance
period tc (including t0− and tc+ ). Denoting length of the
observation window as T and the sampling frequency as
fs , then the number of snapshots of above variables will be
M = T fs + 1. Therefore, for a power system with N buses,
FIGURE 4. The flowchart of the proposed RGCN-MT-TSA framework.
we need M RGCNs and each RGCN has an input feature
matrix of size N × 3.
1) MULTI-TASK DESIGN In the online TSA, the input data can be obtained from
Our TSA task is composed of two subtasks, i.e. the stability either PMUs or TDS.
classification (Task-1) and the critical generator identification
(Task-2). Conventionally, different tasks may have distin- 3) ADJACENCY MATRIX
guished parameters or even architectures. However, individ- For any graph G = (V, E) describing the structure of power
ual designed blocks the sharing of knowledge in the training system, the nodes refer to the buses, while the edges refer to
process. In fact, for the transient stability problem, the judg- the transmission lines. Typically, the element at (i, j) of the
ment of instability has strong, or even causal, links to the adjacency matrix A is defined as follows:
behavior of the critical generators. Hence, we follow the
(
0 Vi , Vj ∈ V, (Vi , Vj ) ∈
/E
conceptional idea of multi-task learning (MTL) in our design ai,j = (5)
1 Vi , Vj ∈ V, (Vi , Vj ) ∈ E
such that the model shares the representation of the related
tasks and performs better on the target tasks. where (Vi , Vj ) denotes the edge from i to j.

VOLUME 8, 2020 93287

 J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

Note that the power disturbance caused by fault will prop-

agate through transmission lines and the distribution of bus
power along the lines is approximately proportional to the
admittance of the transmission lines. Thus we take the advan-
tage of the node admittance matrix Y and define the domain-
related adjacency matrix of G in the GCN as follows:
   FIGURE 5. The logic diagram of the ADM module.
  yi,j
 
yi,j ∈ Y , i 6 = j


ai,j = yi,j max (6) with the vector c = [1, 0]T . Otherwise, it is considered as a
yi,j ∈ Y , i = j

stable one and labeled with c = [0, 1]T .
0
  For an unstable sample, particularly, there exist leading
where yi,j  denotes the module of an element of Y with the
 instability generators. Let NG denote the number of gener-
maximum yi,j max as the reference value for normalization.
ators, then we define G as the set of generators and c ∈ RNG
Substitute (23) in (7), we obtain the renormalized matrix
is a binary vector that denotes the stability status of all the
with its element as:
 generators. The status of a generator i is expressed as:
1
ã0i,j = q ãi,j = β̃i,j
−1
ãi,j
 (
0 |1δi | > 180◦



d̃ d̃


s i,i j,j c(i) = (9)
1 |1δi | ≤ 180◦



 P   s P  
k yi,k yj,k
β̃i,j = 1 +   1+  k  (7)
where |1δi | is the absolute value of separation between gener-

 yi,j  yi,j 
 max max
ator i and the reference generator during the simulation time.
 P  
k yi,k

  
d̃i,i = y  (i 6 = k)


 The set of critical generators Gc = {i ∈ G |c(i) = 0 } is a
i,j max
subset of G and in particular, Gc is an empty set if and only if
where d̃i,i denotes the degree of the node i with node j as the system is transient stable.
0
one of its neighbors. ãi,j , ã0i,j are elements of A0 and Ã
5) ASSISTANT DECISION MAKING MODULE
respectively, while the correction factor β̃i,j is related to the
degrees of the node itself and its neighbor. For online application of TSA, there may be conflict in the
In terms of the massage passing, above setting enables results of task-1 and task-2. We define an assistant decision-
the neighbors with larger degrees (i.e., physically the sum making (ADM) module to provide logically consistent esti-
of its normalized admittance) being assigned with smaller mation. According to Fig. 5, ADM first analyzes the signal set
weighted edges so as to prevent them from occupying a including two single signals (i.e., S1 or S2 from Task-1 and S3 ,
large amount of the neighborhood ‘‘message’’. On one hand, S4 or S5 from Task-2), and simultaneously provides the post-
it reduces the difficulty for GCN to focus on critical buses. fault stability status as well as the sets of generators based on
On the other hand, the neighborhood features might be parallel computing. The detailed logic diagram of ADM is
implicitly utilized as additional information when certain demonstrated in Fig. 5, where ‘‘Critical’’, ‘‘Uncertain’’ and
nodes face the data integrity problem with slightly poor mea- ‘‘Secure’’ refer to the final signals of stability status, while
surement or communication quality. G̃c , G̃s denote the sets of critical and significant generators.
It is worth noting that for the 1st to the (M − 1)th RGCNs, When ADM receives the set (S2 , S3 ), the system enters a
the adjacency matrix follows the power system pre-fault state of emergency with the signal ‘‘Critical’’. Meanwhile,
topology, while the adjacency matrix for the M th RGCN will G̃c is provided for dispatchers to implement critical control.
reflect the post-fault topology. For instance, if the fault line The stability status of the system is uncertain, however, when
between bus i and j is tripped, let ã0i,j = 0. receiving the set (S1 , S3 ) or (S2 , S4 ). In this case, ADM
outputs G̃c as the collection of potentially critical generators
4) GROUND TRUTH DATA and G̃s to be monitored with signals (S1 , S3 ), or only G̃s if
G̃c remains unknown with (S2 , S4 ). TDS is also suggested,
The ground truth for our method, i.e., the stability status of the
if necessary, for further determination. Except for the above
system, is decided by the post-fault rotor angle waveforms for
cases, the system is definitely secure.
a longer period. It can be practically obtained by introducing
the transient stability index η:
B. CRITICAL TRICKS FOR OFFLINE TRAINING
180◦ − |1δ|max
η= (8) 1) GRAPHICAL PARALLEL COMPUTING
180◦ + |1δ|max For a GCN module, the adjacency matrix Ã0 , as well as the
where |1δ|max is the absolute value of the maximum rotor matrix of its degrees are typically sparse matrices. It is easily
angle of separation between any two generators during the perceived that the sparseness has a marked increase in a larger
simulation time. When |1δ|max > 180◦ , i.e., η < 0, one or graph. According to (2), a node can only communicate with
more generators lose their synchronization. We defined this their first-order neighbors in a single propagation, while any
as a transient unstable sample, and label the stability status two non-connected nodes, i.e., with no available path between

93288 VOLUME 8, 2020

J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

them, cannot gain ‘‘message’’ from each other. Therefore, TABLE 1. Confusion matrix for Task-1.
we consider batches of graphs as subgraphs of one or more
large graphs, with the characteristic that any two nodes
belonging to two different subgraphs are still separated from
each other in the synthetic graph. The parallel computing
process for n subgraphs is illustrated as Fig. 6.

we can generalize (28) to L-label learning:

X X
Loss2 = − αi ( ci,j log c̃i,j + (1 − ci,j )
i j
× log(1 − c̃i,j )) + LossL2 +Loss0 L2 (13)

Here, an extra regularization term is added to maintain

the similarity between both tasks, which is defined as:
1
FIGURE 6. Parallel computing process. Loss0L2 = β2 (kw − w0 k2 + kb − b0 k2 ) (14)
2

When the process is GPU accelerated and the block matrix where w0 and b0 are trained parameters of the sharing
does not suffer from memory leak, a single sparse opera- hidden layers of Task-1, while β2 is another regulariza-
tion with complexity O(nL) in a convolution operation as tion weight.
(2) is converted to n parallel operations with O(Li ). Hence, All the loss functions mentioned above are optimized
The max complexity of the sparse operation drops down to with Adam algorithm [40], which is one of the most
O(Lmax ). In our model, we assume a sample as a graph-based commonly-used optimization algorithms for DL.
series of size T , and a block-diagonal matrix for a batch with
m samples is C. PERFORMANCE METRICS
00 0 0 0 0 0 Taking the difference between the tasks into account,
à = diag([Ã1,1 , Ã1,2 , . . . , Ã1,T , . . . Ã2,T , . . . , Ãm,T ]) (10) we designed two categories of metrics to measure the per-
formance of the model.
corresponding to a block matrix of node features X 0 and an
output convolved sparse matrix O0 .
1) CONFUSION MATRIX BASED METRICS
Based on the confusion matrix in Tab. 1, (16) to (19) explain
2) COST-SENSITIVE CROSS-ENTROPY FUNCTION
the specific metrics for our model evaluation, including ACC,
• Task-1
miss alarm (MA) rate, false alarm (FA) rate and G-mean.
Cross-entropy (CE) is widely adopted as the cost func-
tion for classification tasks. However for the problem TP + TN
with imbalanced samples, the stable (negative) samples ACC = (15)
TP + FP + FN + TN
attract too much attention and as a result, the unstable FP
(positive) samples suffer a loss of fit and generalization. MA = (16)
TN + FP
Here, we adopt the cost-sensitive cross-entropy (CSCE) FN
function with L2 regularization term as: FA = (17)
FN + TP
p
X X
Loss1 = − αi ( ci,j log c̃i,j ) + LossL2 (11) G−mean = (1 − MA)(1 − FA) (18)
i j

where αi is the balanced factor. Normally, αi of the where ACC denotes the proportion of the correctly predicted
unstable samples has a bigger value to encourage higher samples. MA represents the proportion of the correct results
accuracy (ACC) for them. [ci,1 , ci,2 ] is the annotated in all unstable samples, which reflects the reliability of assess-
categories, while [c̃i,1 , c̃i,2 ] denotes the softmax function ment with a higher risk priority than FA. FA represents the
outputs of the ith sample. It follows that proportion of the correct results of the stable ones, which
is used to monitor excessive alarm. Furthermore, G-mean is
1 a comprehensive index for the classification of imbalanced
LossL2 = β1 (kwk2 + kbk2 ) (12)
2 samples.
with w and b as learnable network parameters. β1 is the
regularization weight. 2) SET SIMILARITY BASED METRICS
• Task-2 Distinguished from Task-1 with scalar based evaluation,
Assume [ci,1 , ci,2 , . . . , ci,L ] to be annotated labels and Task-2 predicts the set of critical generators. Here, we intro-
[c̃i,1 , c̃i,2 , . . . , c̃i,L ] to be the sigmoid function outputs, duce the Jaccard similarity to evaluate the distance between

VOLUME 8, 2020 93289

 J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

sets of integers. Given any two set si , sj ∈ N, Jaccard similar- TABLE 2. RGCN construction in IEEE 39 Bus system.
ity is defined as:
   
si ∩ sj  s i ∩ s j 
J(si , sj ) =  =     (19)
si ∪ sj  |si | + sj  − si ∩ sj 
where J ∈ [0, 1] and J(si , sj ) = 1. Here, we consider the
sample correct only when J(G̃c , Gc ) = 1. Similar to ACC
and MA, we define Jaccard accuracy (JACC) of all samples
as well as Jaccard accuracy of unstable (JACCU) samples.
In terms of the parameter similarity between Task-1 and
Task-2, we prefer expand Jaccard similarity (EJS) instead of
Jaccard similarity that considers the difference in value and
direction of an ordered set, e.g. a vector. Given vectors vi , vj ∈
R, EJS is calculated by:
v ·v
EJS(vi , vj ) = i j (20)
kvi k + vj − vi · vj
Here we take account of two parameter sets p0 = [w0 , b0 ] for
Task-1 and p = [w, b] for Task-2, and the similarity of the
sharing layers can be expressed as:
1
EJS(p, p0 ) = (EJS(w, w0 ) + EJS(b, b0 )) (21)
2
where EJS ∈ [0, 1]. The closer the similarity is to 1, the better
the two sets of parameters satisfy the similarity constraint.

IV. CASE STUDY

We set up the RGCN-MT-TSA model for two test systems
with different sizes: the IEEE 39 Bus system and the IEEE
300 Bus system. The TDS platform PSD-BPA is applied
to generate the training and testing samples. The proposed FIGURE 7. Comparison among αu of different values.
model is implemented in Pytorch [41] developed by Face-
book. All the tests are fulfilled on a computer with Intel 2) MODEL SETTING AND METHODS COMPARISON
Core i7-9700 3.0GHz CPU,16GB RAM and GTX 1660Ti 6G The samples from ‘‘Base’’ and ‘‘N−1’’ cases form the train-
GPU. ing set (60%), while the ‘‘N−2’’ cases are used for validation
(20%) and test (20%). This can verify the adaptability and
A. IEEE 39 BUS SYSTEM generalization of our method on various topologies. We set
1) TEST SYSTEM AND TDS SETTING the input time window T as 0.1s with reference to the shortest
The IEEE 39 Bus system has 39 buses, 10 generators, 19 loads fault duration of the contingencies and the PMU sampling
and 46 transmission lines. All generators use 6th -order model, frequency fs as 100Hz. Hence, the total of input time steps
with excitation system of IEEE model type I and speed con- are M = 11. Test results are analysed in various aspects with
trol system of IEEE G1. TDS data are generated according to proposed model setting as follows.
the principles as follows: • Design of the RGCN-MT-TSA
• ‘‘N−1’’ and ‘‘N−2’’ cases are generated with the trans- The learning rate for training is set to 0.001, while the
mission lines of the basic cases randomly switched off. batch size is 256. Regularization weights β1 , β2 are 5e-4,
Those cases with islands are rejected. 0.01 and the balanced factor of the stable αs is set as 1.
• All the loads change within 75% to 120% of basic load Note that the ratio of the stable samples to the unstable
level, while the generator outputs are randomly adjusted samples is about 6:1. We intialize the balanced factor of
until the feasible power flow can be obtained. the unstable αu as 6. Assume δ equal to 0.5, we compare
• Contingency: three-phase to ground faults at the begin- groups of hyperparameter settings of RGCNs on vali-
ning and the end of any transmission line, and cleared dation set and definitely select the best model with the
after 0.1s with tripping of the faulted line. detailed setting as Tab. 2.
• Labels of contingencies are determined according to the To go a step further, we search a more desirable value for
TDS lasting 4s. αu according to the metrics comparison in Fig. 7. As the
Finally, 29500 samples are generated with 25168 stable value of αu shows an growing trend, there are insignif-
ones and 4332 unstable ones. icant changes in ACC but distinct rises in JACCU.

93290 VOLUME 8, 2020

J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

TABLE 3. Metrics comparison of methods in IEEE 39 Bus system. TABLE 4. Metrics comparison of composite methods in IEEE 39 Bus
system under ‘‘N−3’’ cases.

and G-mean of RCNN have a remarkable rise by approx-

imately 1% and 0.8%. From Fig. 8 a and Fig. 8 b,
the overlap of samples belonging to different categories
decreases if we replace SAE with CNN. These results
reflect that CNNs are more efficient modules to discover
critical spatial information. Furthermore, in Fig. 8 b
there are clusters of samples along the classification
boundary, i.e., the white area, while most samples are
Nonetheless, we can not ignore that JACC suffer a loss far from the boundary in Fig. 8 c. We can thus infer that
when αu > 10. In this case, with high balanced factors, additional representation, graphically learned by GCN,
the model over-emphasizes the unstable samples and contributes to the compactness of the samples belonging
thus has a worse performance on the stable samples. to the same category but enhances the distance between
To balance the trade-off between JACC and JACCU, those belonging to different categories on the contrary.
we ultimately set αu as 10 such that our model can meet It explains clearly why our method has the best perfor-
the requirement of both tasks. mance that ACC and G-mean are more than 99.4% with
• Single Method Experiment a gain of 1%, when GCN takes the place of CNN.
Tab. 3 shows the performance of different ML meth- In order to evaluate on the extent of topology changes
ods on the test set. Obviously, the generalization of that the TSA model can be tolerated, we set up 50
the typical shallow networks, ω-SVM, SVM and MLP, ‘‘N−3’’ operation conditions and randomly generate
are the weakest with ACC less than 97%, which can 2004 new samples to form an extra test set. Three
be significantly improved by DL methods, e.g., SAE, composite models are tested on the set with the results
SSAE, CNN and LSTM. The DL models benefit at least listed in Tab. 4. To face the greater topology changes,
0.3% increase in ACC and G-mean. Among them, CNN the ACC of the non-topological learning models, RSAE
can simultaneously capture the spatio-temporal features and RCNN, decline dramatically to be below 90%.
and perform the best on the prediction of unstable sam- Nevertheless, the proposed method maintains good per-
ples. The MA is reduced by 0.34% and 0.91% com- formance with both the ACC and G-mean over 95%.
pared with SAE and LSTM, respectively. The latter two An explanation may be that with the topology informa-
methods are comparable on ACC and G-mean, while tion (adjacency matrix) integrated, the GCN modules
the weight-sharing based LSTM has a lighter storage are capable of extracting the characteristics similarity
burden, 1.2MB, only about 1/25 of that of SAE. between untrained topology variations and the known
• Ablation Experiment ones. The results indicate that the topological learning
To check the necessity and effectiveness of spatial method benefits stronger adaptability and robustness
extraction and GCN modules, we construct another two than traditional methods in case of topology changes.
composite models based on the proposed method by Fig. 9 a depicts the training curves, where our proposed
substituting GCN for SAE and CNN as spatial extrac- model converges quickly and smoothly. ACC of our
tors. Then we have a recurrent convolutional neural method grows up sharply to 95% for less than 200 itera-
network (RCNN) and a recurrent stacked auto encoder tions (nearly 60% of RCNN), without pretraining. As for
(RSAE). Multi-dimensional hidden layer activations of an iteration, RSAE and RCNN are trained for a relatively
the spatial extractors are visualized with t-SNE [42] as short time, while RGCN requires a response time of
Fig. 7, where a circle refers to a sample and the back- up to 15810ms without parallel computing. Here the
ground color intensity denotes the values of confidence. response time refers to the prediction time. The training
V0 and V1 represent the compressed 2D activations. and test time with our batch-wise method falls down
Among the models, the spatial extractors of RSAE and dramatically to 1/176, 1/1486 of the previous method.
RCNN require to be pretrained [12] for 160 iterations. Furthermore, a single sample is tested for merely 35µs,
As shown in Tab. 3, the performance of RSAE changes almost half of that of RCNN, meeting the rapidity of the
slightly compared with the single methods, while ACC online application in practice.

VOLUME 8, 2020 93291

 J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

FIGURE 8. Hidden layer activations visualization of the spatial extractors. (a) RSAE. (b) RCNN. (c)RGCN.

3) ROBUSTNESS ANALYSIS
Once the offline model is applied online, we should consider
the input damage problem led by loads fluctuation or poor
measurement. On one hand, the existing research tends to
simplify the distribution of error in the sampling and cal-
culation stage, as ideal Gaussian white noise ε ∼ N (0, 1).
However, white noise might be converted to color noise [43]
in the low-pass filtering of PMUs. The pulse expression of a
low-pass filter is defined as:
15
1 X
h(t) = δ(t − i) (22)
15
i=0
where δ denotes the pulse function. A series of Gaussian color
noise ε0 is generated as:
ε0 = h ∗ ε (23)
Generally, signal to noise ratio (SNR) is used to calculate
distance between noise and signal:
kxk
SNR = 20 lg 0 (24)
ε
SNR of small values refers to high signal distortion. On
the other hand, communication error or signal interference,
etc., usually result in data missing or abnormal values in the
sampling stage. We simulate this scenario by assuming values
of data drop to zero or soar to two times of themselves with
an assumed probability.
Extensive performance of the models for both tasks in
above multiple scenarios is listed as Tab. 5. Under the ideal
scenario, the changes in the loss functions make no significant
FIGURE 9. Convergence metrics comparison of the composite methods.
difference to ACC of Task-1, while we find the improvement (a) Training curves. (b) Response time.
of 0.68% in MA with CSCE. JACCU of Task-2 rises by over
5% with the parameter similarity ( EJS) of more than 0.99. Considering wide area abnormal values of 1% to 3%,
All the metrics remain practically unchanged when consid- the proposed method with message passing based GCN mod-
ering the color noise of big values. Assuming an extreme con- ules is only mildly affected by individual abnormal buses.
dition SNR = 20dB where the noise reaches 10% of the orig- The max loss of metrics in both tasks keeps less than 2%
inal input, MA and FA slightly increase to almost 1%. Due to compared with the ideal scenario. On the whole, the proposed
distinctively more prediction objectives than Task-1, JACCU method results in desirable performance under the designed
suffers a drop of around 2%. Nonetheless, our model covers scenarios and fulfills the requirements of adaptability and
97% of the unstable samples under large noise interference. robustness.

93292 VOLUME 8, 2020

J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

TABLE 5. Robustness metrics of Task-1 and Task-2.

B. IEEE 300 BUS SYSTEM TABLE 6. Metrics comparison of methods in IEEE 300 Bus system.

1) TEST SYSTEM AND TDS SETTING

We employ our method to a system of larger scale and
higher complexity called IEEE 300 Bus system in this sub-
section. It comprises 300 buses, 69 generators, 203 loads and
411 transmission lines. All generators use 6th -order model,
with parameters of the control systems obtained from the
practical grid. Similarly, 31062 samples are generated includ-
ing 26966 stable ones and 4096 unstable ones. Test results are
listed in Tab. 6.

2) TEST RESULTS
In contrast with IEEE 39 Bus system, the input scale grows at
a geometric progression and as a result, ACC of the shallow
networks declines by 8.93% to 13.49%. Deep networks have
ACC over 90% and LSTM performs worst among them. It is Fig. 10 d refers to true stable samples while the others refer to
inferred that a single temporal method generalizes poorly true unstable samples. Hidden activations of generator nodes
to data with abundant spatial and topological characteris- are similarly compressed as 2D vectors with t-SNE. Here a
tics. In terms of the composite methods, our method is the circle represents a generator and its color intensity is related
most reliable one regardless of the system scale. Here ACC to the confidence. The set of predicted critical generators is
and G-mean both maintain about 99%. We then transfer the highlighted by a solid oval, and that of predicted significant
pretrained GCN modules of Task-1 to more fine-grained generators is circled by a dashed one.
Task-2, where the numbers of generators to be predicted From details in Fig.10 a and Fig.10 d, the systems are pre-
are almost 7 times of those in the previous system. JACCU dicted to be unstable in Task-1, while the whole generators are
and JACC of the proposed method are 97.32%, 97.98% and considered to be stable in Task-2 with all of their confidence
meanwhile, the similarity regularization is satisfied with EJS over 0.9. Therefore, ADM determines the system status as
of 0.992. ‘‘Secure’’ and avoids false alarms. In terms of ‘‘Uncertain’’
cases in Fig. 10 b and Fig. 10 e, there still exists conflict in the
C. VISUALIZATION VERIFICATION OF RGCN-MT-TSA predictions of both tasks, where the model of Task-2 detects
FRAMEWORK the set of critical or significant generators. It is expected to
Assume δa equal to 0.9, we apply the proposed framework be concerned more about these generators while utilizing
online based on parallel computing of Task-1 and Task-2. TDS to further reduce harmful MA phenomena. When both
The average time of a batch assessment in IEEE 39 Bus tasks predict the systems and the generators to be unstable as
system, as well as IEEE 300 Bus system, is respectively Fig. 10 c and Fig. 10 f, ADM indicates the state of emergency
16ms and 62ms. It follows that ADM generates three signals, and prompt critical control can be implemented based on the
e.g., ‘‘Secure’’, ‘‘Uncertain’’ and ‘‘Critical’’. The following set of critical generators. Generally due to the visualization,
typical examples described in Fig. 10 are selected to verify dispatchers might efficiently recognize the set of generators
the effectiveness of the designed signals, where Fig. 10 a and to be controlled based on the color intensity and clusters of

VOLUME 8, 2020 93293

 J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

FIGURE 10. Typical visualization results of RGCN-MT-TSA framework. (a)(d) Secure. (b)(e) Uncertain. (c)(f) Critical.

circles. It is convenient to sort the importance of generators In future work, we will pay attention to periodic model
and then develop more precise control strategies. update in our framework when facing more complex changes
in the topology. This adaptive framework is expected to
V. CONCLUSION expand on a large practical system with thousands of buses.
In this paper, a multi-task transient stability assessment
framework is proposed to address early-warning of stability REFERENCES
classification and critical generator identification according [1] P. Kundur, J. Paserba, V. Ajjarapu, G. Andersson, A. Bose, C. A. Canizares,
to PMU data. We design a cascade neural network archi- N. D. Hatziargyriou, D. J. Hill, A. M. Stankovic, C. Taylor, T. Van Cutsem,
tecture named RGCN to capture the transient characteristics and V. Vittal, ‘‘Definition and classification of power system stability
IEEE/CIGRE joint task force on stability terms and definitions,’’ IEEE
graphically and temporally, where a state-of-the-art network, Trans. Power Syst., vol. 19, no. 3, pp. 1387–1401, Aug. 2004.
GCN, is creatively used to explicitly extract physical topolog- [2] S. Obuz, M. Ayar, R. D. Trevizan, C. Ruben, and A. S. Bretas, ‘‘Renewable
ical information of the power system. The offline models of and energy storage resources for enhancing transient stability margins:
A PDE-based nonlinear control strategy,’’ Int. J. Elect. Power Energy Syst.,
different tasks are trained in a parallel way, with a new cost- vol. 116, Mar. 2020, Art. no. 105510.
sensitive cross-entropy function to handle the imbalanced [3] P. Kundur, N. J. Balu, and M. G. Lauby, Power System Stability and
problem. A similarity regularization item is designed such Control, vol. 7. New York, NY, USA: McGraw-Hill, 1994.
that the model of Task-1 can be transferred to that of Task- [4] R. Diao, S. Jin, F. Howell, Z. Huang, L. Wang, D. Wu, and Y. Chen,
‘‘On parallelizing single dynamic simulation using HPC techniques and
2 and the training difficulty is alleviated. To evaluate the APIs of commercial software,’’ IEEE Trans. Power Syst., vol. 32, no. 3,
effectiveness and robustness of the proposed method, a series pp. 2225–2233, May 2017.
of case studies as well as comparisons with six different single [5] L. M. Skvortsov, ‘‘A fifth order implicit method for the numerical solution
or aggregating models are comprehensively conducted on two of differential-algebraic equations,’’ Comput. Math. Math. Phys., vol. 55,
no. 6, pp. 962–968, Jun. 2015.
benchmark systems of different scales. Test results indicate [6] T. Athay, R. Podmore, and S. Virmani, ‘‘A practical method for the direct
the desirable performance and reliability of the proposed analysis of transient stability,’’ IEEE Trans. Power App. Syst., vol. PAS-98,
method. Furthermore, our framework provides comprehen- no. 2, pp. 573–584, Mar. 1979.
sive signals and 2D visualization of the generators, which [7] Y. Xu, Z. Y. Dong, R. Zhang, Y. Xue, and D. J. Hill, ‘‘A decomposition-
based practical approach to transient stability-constrained unit com-
helps to improve the false alarm rate as well as implement mitment,’’ IEEE Trans. Power Syst., vol. 30, no. 3, pp. 1455–1464,
more accurate and timely control. May 2015.

93294 VOLUME 8, 2020

J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

[8] M. He, J. Zhang, and V. Vittal, ‘‘Robust online dynamic security assess- [30] C. Kim, K. Kim, P. Balaprakash, and M. Anitescu, ‘‘Graph convolutional
ment using adaptive ensemble decision-tree learning,’’ IEEE Trans. Power neural networks for optimal load shedding under line contingency,’’ in
Syst., vol. 28, no. 4, pp. 4089–4098, Nov. 2013. Proc. IEEE Power Energy Soc. Gen. Meeting (PESGM), Aug. 2019,
[9] F. R. Gomez, A. D. Rajapakse, U. D. Annakkage, and I. T. Fernando, pp. 1–5.
‘‘Support vector machine-based algorithm for post-fault transient stability [31] J. J. Q. Yu, D. J. Hill, V. O. K. Li, and Y. Hou, ‘‘Synchrophasor recovery and
status prediction using synchronized measurements,’’ IEEE Trans. Power prediction: A graph-based deep learning approach,’’ IEEE Internet Things
Syst., vol. 26, no. 3, pp. 1474–1483, Aug. 2011. J., vol. 6, no. 5, pp. 7348–7359, Oct. 2019.
[10] S. A. Siddiqui, K. Verma, K. R. Niazi, and M. Fozdar, ‘‘Real-time mon- [32] T. N. Kipf and M. Welling, ‘‘Semi-supervised classification with graph
itoring of post-fault scenario for determining generator coherency and convolutional networks,’’ in Proc. Int. Conf. Learn. Represent., 2017,
transient stability through ANN,’’ IEEE Trans. Ind. Appl., vol. 54, no. 1, pp. 1–14.
pp. 685–692, Jan. 2018. [33] J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, ‘‘Neural
[11] I. B. Sulistiawati, A. Priyadi, O. A. Qudsi, A. Soeprijanto, and N. Yorino, message passing for quantum chemistry,’’ in Proc. Int. Conf. Mach. Learn.,
‘‘Critical clearing time prediction within various loads for transient stabil- vol. 70, 2017, pp. 1263–1272.
ity assessment by means of the extreme learning machine method,’’ Int. J. [34] S. Hochreiter and J. Schmidhuber, ‘‘Long short-term memory,’’ Neural
Elect. Power Energy Syst., vol. 77, pp. 345–352, May 2016. Comput., vol. 9, no. 8, pp. 1735–1780, 1997.
[12] Q. Zhu, J. Chen, L. Zhu, D. Shi, X. Bai, X. Duan, and Y. Liu, ‘‘A deep [35] S. Ioffe and C. Szegedy, ‘‘Batch normalization: Accelerating deep network
end-to-end model for transient stability assessment with PMU data,’’ IEEE training by reducing internal covariate shift,’’ in Proc. Int. Conf. Mach.
Access, vol. 6, pp. 65474–65487, 2018. Learn., 2015, pp. 448–456.
[13] J. J. Q. Yu, D. J. Hill, A. Y. S. Lam, J. Gu, and V. O. K. Li, ‘‘Intelligent [36] T. Kim, I. Song, and Y. Bengio, ‘‘Dynamic layer normalization for adaptive
time-adaptive transient stability assessment system,’’ IEEE Trans. Power neural acoustic modeling in speech recognition,’’ in Proc. Interspeech,
Syst., vol. 33, no. 1, pp. 1049–1058, Jan. 2018. Aug. 2017, pp. 2411–2415.
[14] J. J. Q. Yu, D. J. Hill, and A. Y. S. Lam, ‘‘Delay aware transient stabil- [37] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and
ity assessment with synchrophasor recovery and prediction framework,’’ R. Salakhutdinov, ‘‘Dropout: A simple way to prevent neural networks
Neurocomputing, vol. 322, pp. 187–194, Dec. 2018. from overfitting,’’ J. Mach. Learn. Res., vol. 15, no. 1, pp. 1929–1958,
[15] Y. Zhou, Q. Guo, H. Sun, Z. Yu, J. Wu, and L. Hao, ‘‘A novel data-driven 2014.
approach for transient stability prediction of power systems considering [38] R. Caruana, ‘‘Multitask learning: A knowledge-based source of inductive
the operational variability,’’ Int. J. Elect. Power Energy Syst., vol. 107, bias,’’ in Proc. Int. Conf. Mach. Learn., 1993, pp. 41–48.
pp. 379–394, May 2019. [39] L. Duong, T. Cohn, S. Bird, and P. Cook, ‘‘Low resource dependency
[16] R. Zhang, J. Wu, Y. Xu, B. Li, and M. Shao, ‘‘A hierarchical self- parsing: Cross-lingual parameter sharing in a neural network parser,’’ in
adaptive method for post-disturbance transient stability assessment of Proc. 53rd Annu. Meeting Assoc. Comput. Linguistics, 7th Int. Joint Conf.
power systems using an integrated CNN-based ensemble classifier,’’ Ener- Natural Lang. Process., 2015, pp. 845–850.
gies, vol. 12, no. 17, p. 3217, 2019. [40] D. P. Kingma and J. Lei Ba, ‘‘Adam: A method for stochastic optimiza-
[17] A. Gupta, G. Gurrala, and P. S. Sastry, ‘‘An online power system stability tion,’’ in Proc. Int. Conf. Learn. Represent., 2015, pp. 1–15.
monitoring system using convolutional neural networks,’’ IEEE Trans. [41] A. Paszke et al., ‘‘PyTorch: An imperative style, high-performance
Power Syst., vol. 34, no. 2, pp. 864–872, Mar. 2019. deep learning library,’’ in Proc. Adv. Neural Inf. Process. Syst., 2019,
[18] L. Zhu, D. J. Hill, and C. Lu, ‘‘Hierarchical deep learning machine for pp. 8024–8035.
power system online transient stability prediction,’’ IEEE Trans. Power [42] L. van der Maaten and G. Hinton, ‘‘Visualizing data using t-SNE,’’ J. Mach.
Syst., vol. 35, no. 3, pp. 2399–2411, May 2020. Learn. Res., vol. 9, pp. 2579–2605, Nov. 2008.
[19] Z. Shi, W. Yao, L. Zeng, J. Wen, J. Fang, X. Ai, and J. Wen, ‘‘Convolu- [43] C. K. Papadopoulos and C. L. Nikias, ‘‘Parameter estimation of exponen-
tional neural network-based power system transient stability assessment tially damped sinusoids using higher order statistics,’’ IEEE Trans. Acoust.,
and instability mode prediction,’’ Appl. Energy, vol. 263, Apr. 2020, Speech, Signal Process., vol. 38, no. 8, pp. 1424–1436, Aug. 1990.
Art. no. 114586.
[20] R. Yan, G. Geng, Q. Jiang, and Y. Li, ‘‘Fast transient stability batch
assessment using cascaded convolutional neural networks,’’ IEEE Trans.
Power Syst., vol. 34, no. 4, pp. 2802–2813, Jul. 2019.
[21] T. Ishizaki, A. Chakrabortty, and J.-I. Imura, ‘‘Graph-theoretic analysis of JIYU HUANG received the B.S. degree in electri-
power systems,’’ Proc. IEEE, vol. 106, no. 5, pp. 931–952, May 2018.
cal engineering from the South China University of
[22] F. Ebrahimzadeh, M. Adeen, and F. Milano, ‘‘On the impact of topol-
Technology (SCUT), Guangzhou, China, in 2019,
ogy on power system transient and frequency stability,’’ in Proc. IEEE
where he is currently pursuing the Ph.D. degree.
Int. Conf. Environ. Electr. Eng., IEEE Ind. Commercial Power Syst. Eur.
(EEEIC/I&CPS Europe), Jun. 2019, pp. 1–5. His research interests include deep learning
[23] Y. Song, D. J. Hill, and T. Liu, ‘‘Characterization of cutsets in networks in power system security and transient stability
with application to transient stability analysis of power systems,’’ IEEE assessment.
Trans. Control Netw. Syst., vol. 5, no. 3, pp. 1261–1274, Sep. 2018.
[24] T. Weckesser, H. Jóhannsson, M. Glavic, and J. Østergaard, ‘‘An improved
on-line contingency screening for power system transient stability assess-
ment,’’ Electr. Power Compon. Syst., vol. 45, no. 8, pp. 852–863,
May 2017.
[25] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ‘‘ImageNet classification
with deep convolutional neural networks,’’ in Proc. Adv. Neural Inf. Pro- LIN GUAN (Member, IEEE) received the B.S. and
cess. Syst., 2012, pp. 1097–1105. Ph.D. degrees in electric power engineering from
[26] R. Ye, X. Li, Y. Fang, H. Zang, and M. Wang, ‘‘A vectorized relational the Huazhong University of Science and Technol-
graph convolutional network for multi-relational network alignment,’’ in ogy, Wuhan, China, in 1990 and 1995, respec-
Proc. 28th Int. Joint Conf. Artif. Intell., Aug. 2019, pp. 4135–4141.
tively.
[27] L. Yao, C. Mao, and Y. Luo, ‘‘Graph convolutional networks for text
She is currently a Professor with the Electric
classification,’’ in Proc. AAAI Conf. Artif. Intell., vol. 33, Jul. 2019,
pp. 7370–7377.
Power College, South China University of Tech-
[28] R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and nology, Guangzhou, China. From 2014 to 2015,
J. Leskovec, ‘‘Graph convolutional neural networks for Web-scale recom- she was a Visiting Scholar with Stanford Univer-
mender systems,’’ in Proc. 24th ACM SIGKDD Int. Conf. Knowl. Discov- sity. She is the author of more than 120 articles
ery Data Mining, Jul. 2018, pp. 974–983. and a Principal Investigator of more than 50 projects. Her research interests
[29] K. Chen, J. Hu, Y. Zhang, Z. Yu, and J. He, ‘‘Fault location in power include application of artificial intelligence technology in electrical engi-
distribution systems via deep graph convolutional networks,’’ IEEE J. Sel. neering, power system security and control, and power system planning and
Areas Commun., vol. 38, no. 1, pp. 119–131, Jan. 2020. reliability.

VOLUME 8, 2020 93295

 J. Huang et al.: Recurrent Graph Convolutional Network-Based Multi-Task TSA Framework in Power System

YINSHENG SU received the B.S. and M.S. MENGXUAN GUO was born in Hunan, China,
degrees in electrical engineering from Shanghai in 1997. She received the B.S. degree in electrical
Jiaotong University, Shanghai, China, in 1999 and engineering from the South China University of
2002, respectively. He is currently a Senior Spe- Technology (SCUT), Guangzhou, China, in 2019.
cialist with China Southern Grid Co. Ltd. His Her research interests include deep learning in
research interest includes electric power system power system security and small-signal stability
operation and control. analysis.

HAICHENG YAO received the B.S. degree in ZHI ZHONG received the B.S. degree in electri-
water conservancy and hydropower engineering cal engineering from Southeast University, Nan-
from the Huazhong University of Science, Wuhan, jing, China, in 2019. He is currently pursuing
China, in 2004, and the M.S. degree from the the M.S. degree in electrical engineering with the
State Grid Electric Research Institute, Shanghai, South China University of Technology (SCUT),
China, in 2007. He is currently a Senior Engineer Guangzhou, China. His research interests include
with China Southern Grid Co. Ltd. His research machine learning, transient stability assessment of
interest includes electric power system operation power system, and the applications of big data in
and control. smart grids.

93296 VOLUME 8, 2020