Received: 1 March 2021 Revised: 22 June 2021 Accepted: 29 July 2021 The Journal of Engineering

DOI: 10.1049/tje2.12068

ORIGINAL RESEARCH PAPER

A power system transient stability assessment method based on

active learning

Yuqiong Zhang1 Qiang Zhao1 Bendong Tan2 Jun Yang2

1
China Electric Power Research Institute, Beijing, Abstract
China
Due to the wide deployment of phasor measurement unit, the real-time assessment of
2
School of Electrical Engineering and Automation, transient stability based on machine learning shows great potential for development. In
Wuhan University, Wuhan, China
order to solve the problem of time-consuming data generation of offline training in such
methods and the difficulty of quickly updating the model after the grid changes, the paper
Correspondence
Bendong Tan,Mississippi State University, Starkville, proposes a method for transient stability assessment (TSA) of power systems based on
39759, USA. active learning. Firstly, different operation conditions and different faults are considered
Email: [email protected]
to perform short-time simulation (simulation to the instant of fault clearance) to generate
unlabelled samples. After the careful selection of critical TSA features, a part of samples are
Funding information
Research on the stability of power system with randomly selected for long-term simulation to label the transient status of these samples,
high penetration of power electronic equipment, and random forest is further trained to construct TSA model. Finally some data is selected
Grant/Award Number: JS71-20-004 in the remaining unlabelled samples with higher information entropy to label and retrain
the model until the model accuracy no longer changes. The simulation on the test power
system shows that the method proposed in this paper can effectively reduce the time of
offline simulation, and greatly improve the efficiency of model, and is also robust to wide-
area noise.

KEYWORDS
active learning, phasor measurement unit, transient stability assessment

JEL CLASSIFICATION
Generation, transmission and distribution

1 INTRODUCTION function [3] is another method for quickly analyzing the tran-
sient stability of power system. It uses the characteristics that
Transient stability refers to the ability of the power system to whether power system can absorb the energy added after the
transit to a stable state after suffering large disturbances [1]. fault to determine the transient stability. It has the advantage of
With penetration of high proportion of renewable energy and fast calculation speed, but cannot be applied to large-scale actual
rapid development of electricity market reform, the operation power system because of its conservatism and the difficulty of
of power system is near the limitation of stability. Therefore, constructing the energy function. At present, synchronous mea-
quickly and accurately assessment of the transient stable state is surement technology plays an important role in power system
of great significance to the safe and stable operation of power monitoring and control [4, 5].
system. The wide-area measurement system (WAMS) plays an impor-
The traditional method of assessing transient stability is to tant role in the secure analysis because of its function of syn-
use time domain simulation (TDS) [2] to analyze the tran- chronously sampling states of power system. In order to use
sient characteristics of power system by solving a set of high- synchronous measurement technology to achieve fast and accu-
dimensional differential-algebraic equations, but its heavy com- rate transient stability assessment (TSA), researchers have pro-
putation is not suitable for online analysis. Transient energy posed a large number of data-driven transient stability analysis

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is
properly cited.
© 2021 The Authors. The Journal of Engineering published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology

J. Eng. 2021;2021:715–723. wileyonlinelibrary.com/iet-joe 715

716 ZHANG ET AL.

methods for power system. Especially in recent years, with the 150
rapid development of deep learning and artificial intelligence Gen 1 Gen 6
Gen 2 Gen 7
technology, machine learning is widely used in power system Gen 3 Gen 8
security analysis, such as support vector machine (SVM) [6, 7], 100 Gen 4 Gen 9
Gen 5 Gen 10
decision tree (DT) [8–10], neural network (NN) [11, 12] are

Rotor angle (deg)

widely used to build an online transient stability evaluation sys-
50
tem. Besides, deep learning is also employed to improve the
performance of TSA, while automatic feature engineering is
achieved [13].What these data-driven methods have in common 0
is to establish the mapping relationship between the transient
state and the measurement data. Once the fault is removed, the
transient stability of power system can be accurately determined. -50
This leaves sufficient time for emergency control, which is of
great significance to the safety and stability of power system.
At present, data-driven methods are focus on handling the -100
0 1 2 3 4 5
problems of missing data and unbalanced class. For missing data Time (s)
issues, pioneering work along this path is done by [14] through
recovering missing data by learning the machine learning model. FIGURE 1 Stable case of rotor angle curve of New England 39-bus
Ref. [15] handles missing data by using surrogate in DTs, while power system after a disturbance
the refs. [16, 18] develop observability-constraints based feature
sets searching algorithm to reduce the impacts of missing data.
350
For the unbalanced class problem, this issue is mainly addressed Gen 1 Gen 6
by oversampling [19, 21]. A deep learning based framework is 300 Gen 2 Gen 7
built in ref. [19] to perform ADASYN so that the unstable sam- Gen 3 Gen 8
250
Gen 4 Gen 9
ple can be enhanced, while ensemble method in refs. [20, 21] 200 Gen 5 Gen 10
cost-sensitive method is respectively proposed to deal with this.
Rotor angle (deg)

150
However, the offline training of the above TSA methods
based on data needs to generate a large amount of transient 100
data under different fault conditions, and the actual power sys- 50
tem includes more than thousands of buses, which means it 0
takes a lot of time to generate these data through simulation.
-50
In addition, when the topology of power system changes, it is
also necessary to regenerate the data. Once a fault occurs, the -100
TSA model cannot be not updated in time due to the time- -150
consuming data generation, which is likely to cause catastrophic -200
consequences. 0 1 2 3 4 5
In response to the above problems, the paper proposes an Time (s)
active learning (AL)-based TSA. By introducing random forest
FIGURE 2 Unstable case of rotor angle curve of New England 39-bus
(RF) with the probability discrimination layer, it can be trained
power system after a disturbance
with only a small amount of long-term simulation data marked
with transient status, and actively select short-term simulation
data without marking transient status according to uncertainty develop efficient TSA methods for such response-based pro-
to do iterative learning. The simulation on the New England tection system, for example, predicting transient status as soon
39-bus power system shows that the method proposed in this as possible. However, not all measurements are useful for
paper can greatly reduce the time to generate data and update TSA.
the model in time, thereby achieving safe and stable operation Taking New England 39-bus power system as an example, the
of the power system. rotor angle curves after a disturbance are presented in Figure 1,
where the number after “Gen” indicates the index of the corre-
sponding generator. When power system is subject to a distur-
2 METHODOLOGY bance, the unbalanced torque will be imposed on the generator
rotor due to unbalanced power. Generally, in the steady state
2.1 Critical feature selection for transient of generator, mechanical power equals to the electrical power.
stability assessment However, the unbalanced power is introduced by the instantly
changed electrical power (i.e. the load can vary immediately)
With the gradual deployment of the synchrophasor measure- while the mechanical power keep constant in the transient pro-
ment devices, wide area protection and control systems have cess owing to the large time constant of the governor. As a
attracted significant attention [22]. Therefore, it is urgent to result, some generator, such as Gen 8 shown in Figures 1 and 2,
ZHANG ET AL. 717

FIGURE 3 The structure of random forest

TABLE 1 Algorithm of ReliefR R1 and R2 are different samples and p(C ) is ratio of class C
Inputs: Training dataset, sampling times m, the number of in all samples. Note that, the larger the feature importance, the
the nearest samples n. stronger capacity for transient stability identification. Besides, a
Outputs: Importance of each feature W. positive importance is beneficial for TSA while a negative one
1. Initialization: W = 0;
is not.
2. For i = 1 to m do
1) Randomly choose a sample R; 2.2 Random forest
2) Choose n near hits H j ( j = 1, 2, … , n) from the same
classes, and choose n near misses M j ( j = 1, 2, … , n) RF, as shown in Figure 3, is a combined classifier composed of
from the different classes; multiple DT, so DT is firstly introduced. DT [8] can be consid-
3) For A = 1 to N all features do ered as a classification tree model, including root nodes, internal
nodes and leaf nodes. Among them, the root node contains the
∑
n
di f f (A, R, H j ) entire dataset, and each internal node is a judgment criterion. It
W[A] = W[A] − +
j =1
m×n will allocate the dataset to two or more child nodes according
(1)
∑ p(C ) ∑k
di f f (A, R, M j ) to the test results of the judgment criterion, including the final
.
C ≠class(R)
1 − p(C ) j =1
m×n classification category.
end
RF trains multiple tree classifiers to combine the predictions
of multiple DTs, and finally obtains the prediction results by
end
voting. The basic idea is to use Bootstrap sampling to extract k
samples from the original training set, secondly to establish k
DT models to obtain k classification results, and finally to vote
on all results to determine which category they ultimately belong
will lose synchronism with other generators while the remaining to. Its model function is:
generators are still synchronous. This means there are some crit-
ical features indicating the transient stability. For example, when ∑
k
a fault occurs at a bus, the measurements at the local area near H (x) = arg max I (hi (x) = Y ), (2)
Y
to the bus would change sharply while the area far from the bus i=1
may just vary slightly. Thus, it is important to extract critical fea-
where kis the index of DTs; Y is the output variable (target vari-
tures to reduce redundancy before building TSA model, which
able); I is the indicative function; H (x) represents the combined
is achieved by ReliefF [23] in this paper, and this is the key issue
classification model; hi (x)represents the classification model of
to improve the classification performance [24].
the i-th DT. RF increases the difference between classification
ReliefF is a kind of feature weighting algorithm. It assigns
models by constructing different training sets, thereby improv-
weights to features according to their relevance to classes. Its
ing the extrapolation and prediction ability of combined clas-
detailed implement is shown in Table 1.
|R [A]−R2 [A]| sification models. Therefore, the RF solves the problem of
Where di f f (A, R1 , R2 ) = 1 , overfitting well, integrating multiple weak classifiers into a
max(A)−min(A)
718 ZHANG ET AL.

strong classifier. Besides, to analysis the uncertainty of samples, TABLE 2 Algorithm of active learning
the output of RF is probability of class. The predicted class Inputs:
probabilities of an input sample are computed as the mean pre- (1) Initial training set L={xi , yi },i = 1, 2, … , l , which
dicted class probabilities of the trees in the forest. The class contains l labelled samples;
probability of a single tree is the fraction of samples of the same (2) Pool of candidates U ={x j }, j = l + 1, l + 2, … , l + u,
class in a leaf. which contains u unlabelled samples;
(3) Number of a batch of the most useful samples q;
Outputs: Trained classifier C.
2.3 Active learning
1: Repeat
In the existing TSA methods, a large amount of transient data is 2: Train C with current training set L;
often required to ensure the effectiveness of the model. Gener- 3: For each candidate in U do
ally speaking, the more faults considered, the richer the operat- 4: Calculate E for each candidate according to (4);
ing conditions, the better the model generalization ability. How- 5: end for
ever, in the actual operation of power system, the scale of
6: Rank the candidates in U according to the score of E;
power system is very large. It is very time-consuming to use
the power system analysis software to obtain a large amount 7: Select the samples whose score ranks top q;
of time-domain simulation data. In addition, due to the con- 8: Lebel the q samples and add them into L;
tinuous change of operating conditions, it is impossible to meet 9: Until Classifier C converges
the requirements of rapid configuration by regenerating a large
amount of simulation data to train the model. Therefore, this
paper uses AL to overcome these problems. If the amount of
the training sample is too large, the simulation time will also 3 ACTIVE LEARNING-BASED
be relatively large. AL first randomly selects a small part of the TRANSIENT STABILITY ASSESSMENT
long-term simulated labelled transient data for training RF, then
queries the most useful unlabelled short-time simulation sam- 3.1 Input features of the model
ples through a certain algorithm for labelling, and then trains
the model with the query samples to improve accuracy. Physical variables strongly related to transient stability accord-
The model of AL can be defined as: ing to existing literature [15–17] are selected, so generator rotor
angle, generator active and reactive power, bus voltage magni-
A = (C , Q, S , L, U ). (3) tude and angle, active and reactive power of branch, active and
reactive power of load at the time of fault clearance is used to
Among them, C is a classifier, Lis used to train the C. Qis construct input features. In order to ensure stable training of the
a query function for querying information with a large amount model during application, the feature input is normalized:
of information from the unlabelled sample pool U , and S can
label the samples in U with the correct labels. The model starts xi j − min(xi j )
i
learning with a small number of initial labelled samplesL, selects x∗i j = . (5)
max(xi j ) − min(xi j )
one or a batch of the most useful samples through a certain i i
query functionQ, and asksS (in this paper, S is the power system
simulation software) for labels, and then uses the acquired new Among them, xi j is the j-th feature of the i-th sample. After
knowledge to train the classifier and perform the next round of normalization, the value range of all input data is between [0,1].
queries. AL is a cyclic process until a certain stopping criterion is When it is applied to TSA, it can be denormalized according to
reached. When the evaluation accuracy of the model no longer Equation (5).
improves, this cycle can be ended.
In general, the query function Qcan select samples with large
information entropy. Assuming that the probability that a sam- 3.2 Offline training
ple is stable and unstable on the model is p1 andp2 , respectively,
then the information entropy of the sample can be calculated: Before the model is applied to TSA, it needs to be trained in the
data under different fault conditions. In order to save the time
∑
2 of offline training, a training method of TSA model based on
E= p1 lg p2 , (4) AL is proposed in this paper shown as Figure 4.
i=1
1. Considering different operation levels and different fault
E is information entropy, p1 + p2 = 1. Therefore, the strategy conditions, a large amount of transient data used to train
for query function Q is to find samples with highest information the model is generated and divided into a training dataset
entropy, because the rich amount of information contained in and a test dataset. This process only requires a long period
these samples is useful for training models. Table 2 shows the ofN1 samples randomly obtained from the training set, and
detailed algorithm for AL. these samples are simulated and marked with transient status.
ZHANG ET AL. 719

FIGURE 4 Overall block diagram

The remaining samples only need to be simulated until the

first cycle of fault removal (transient status is not marked). FIGURE 5 New England 39-bus power system
The transient status can be defined as [12]:

360 − |Δ𝛿max |
𝜂= (6)
360 + |Δ𝛿max | obtained, the critical features are selected and the transient
status of power system can be assessed;
In the formula, |Δ𝛿max | is the maximum absolute value of 2. When the topology of power system changes, a large num-
the rotor angle difference between any two generators after ber of short-term simulation (simulation to the first cycle
TDS. The sample is judged to be transiently stable if 𝜂 ≥ 0, after fault clearance) transient data generated for retraining
and the sample is judged to be transiently unstable if𝜂 < 0. the model under different operating levels and different fault
2. The critical features are selected by ReliefR, and a RF is conditions can be considered, and step (3) and step (4) in
trained based on theN1 samples obtained from random sam- Section 3.2 can be repeated to quickly update the model.
pling. The trained model performs probabilistic assessment
on the unlabelled samples, and can be used to calculate the
information entropy of each unlabelled sample based on the 4 CASE STUDY
assessment results;
3. TheN2 unlabelled sample with the highest information 4.1 Dataset generation
entropy is selected to be simulated for a long time and
label the transient status, and the model is retrained with The New England 10-machine 39-bus power system [25],
the labelled samples already in the training dataset, then the shown in Figure 5, is used to construct the dataset and verify
assessment accuracy rate of the test dataset is calculated; the model proposed in the paper. The New England 39- bus
4. The retrained model is used to assessment the stable proba- power system is composed of 39 buses, 46 branches, 10 gen-
bility of the unlabelled samples, and the step (3) is repeated erators and 19 loads. The reference power is 100MVA and the
until the assessment accuracy does not change. reference voltage is 345 kV.
The power system analysis software PSS/E is used to write a
transient batch program to generate a large number of transient
3.3 Online application datasets to train the model proposed in the article. Consider-
ing different faults and operation conditions, 5775 samples are
Using RF trained by offline training, the model can be deployed finally generated, and the specific TDS parameters were set as
online to assess the transient status of the power system after follows:
disturbance. The overall block diagram is shown in Figure 2.
Specifical steps are as follows: 1. 11 different load levels (75%, 80%, 85%, 90%, 95%, 100%,
105%, 110%, 115%, 125% of the baseline) are considered.
1. After the fault is cleared, once the data of generator rotor At the same time, output power of generators is also adjusted
angle, generator active and reactive power, bus voltage mag- in the same proportion;
nitude and angle, active and reactive power of branch, active 2. Under each operation condition, three-phase short-circuit
and reactive power of load at the time of fault clearance are ground faults are applied at four positions of the bus and the
720 ZHANG ET AL.

10-3 10-3
20 20

15 15
Feature importance

Feature importance
10 10

5 5

0 0

-5 -5
0 50 100 150 200 0 50 100 150 200
Feature rank Feature rank

FIGURE 6 Unranked feature importance of New England 39-bus power FIGURE 7 Ranked feature importance of New England 39-bus power
system system

TABLE 3 Parameter settings

branch (20%, 40%, 60% and 80% from the end of branch), Model Parameter
respectively; SVM Kernel function: Polynomial function;
3. For each specifically fault, the fault occurrence time is 1 s penalty coefficient: 10e6
after the simulation starts, and the fault duration is set to 0.1, RF The number of DTs:100
0.3 or 0.5 s;
NN The number of hidden neurons:100
4. Randomly samplingN1 sample for long TDS, the total sim-
Method proposed in The number of DTs:100
ulation time is 10 s, and the simulation step is 0.01 s; the
this paper
remaining samples are simulated in a short time to the first
cycle after fault clearance.

In order to prevent the model from overfitting, all samples 4.3 Model performance comparison
are divided into a training set and a test set according to the ratio
of 4: 1. The training set is used for offline training, while the test The method proposed in the paper is verified on the New Eng-
set verify the model proposed in the article. In addition, all sim- land 39-bus power system, and the optimal parameters of the
ulations are implemented based on the Intel Core i5-7300HQ model are determined by grid search. At present, the commonly
CPU and 8GB of RAM. The machine learning model is imple- used TSA methods are: SVM, RF, NN and these three models
mented in the Scikit-learn library of Python. are used to compare with the method proposed in the paper, the
various method parameter settings are shown in Table 3.
Table 4 shows the results of each model. It can be found
4.2 Critical feature selection that accuracy of the method proposed in this paper is the high-
est in all models, and simulation time is the least in the pro-
The careful feature selection for critical features can reduce
the redundancy in all the features, and also decrease the com- TABLE 4 Performance comparison of each model
plexity of the TSA model. In this paper, the critical features
are selected according to the importance of each feature cal- The
number of
culated by ReliefR, as shown in Figures 6 and 7. It can be seen Simulation labelled
that almost all features have the positive importance. To reduce Model Accuracy time (s) samples
the redundancy in the features, feature importance is ranked
SVM 99.31 46,200 4620
in descending order, and the features belong to the cumula-
tive importance of the top 95% are selected. As a result, 197 RF 99.39 46,200 4620
features are selected from 288 features, which means 31.6% of NN 96.19 46,200 4620
the amount the features are removed. Consequently, the redun- Method proposed 99.57 19,680 1500
dancy in all the features and the complexity of the TSA model in this paper
have been significantly reduced.
ZHANG ET AL. 721

posed method. Compared with the original RF, the accuracy TABLE 5 Performance comparison of each model
is improved and simulation time is less. The simulation time The
is reduced by about 57.4%, and the number of samples to be number of
marked is reduced by about 67.5%. This is because the method Simulation labelled
proposed in this paper uses information entropy to evaluate Model Accuracy time (s) samples
the uncertainty to find useful information in unlabelled sam- SVM 98.87 46,200 4620
ples. On one hand, it reduces the simulation time, and on the RF 98.96 46,200 4620
other hand, it filters the redundant information in the samples,
NN 96.54 46,200 4620
thereby improving the model performance.
Method proposed 99.48 21,780 2100
in this paper

4.4 Analysis of the influence of wide-area

noise
In the previous analysis, it is assumed that all system measure- samples needs to be increased by 400, resulting in a slight
ments are accurate, but the actual WAMS inevitably affected increase in simulation time, compared with the number of
by noise. Therefore, Section 4.4 will analyze the impact of labelled samples in other models, it still has a huge advantage
wide-area noise on the performance of different models. in simulation time.
According to the standard set by IEEE C37.118 [26] for phasor
measurement unit (PMU), the error of PMU measurement
vector should a total vector error less than 1%. The specific 4.5 Analysis of the influence of
expression is [15]: high-dimension data

| X̃ ∠𝜃̃ − X ∠𝜃 | To further verify the capacity of the proposed method on

| | < 1% (7)
| X ∠𝜃 | handling high-dimension data in power system, the proposed
| | method is performed on the power system in South Carolina of
̃ United States. The power system consists of 56 generators and
whereX̃ ∠𝜃represents the measurement vector obtained from
500 buses, as shown in Figure 8.
the WAMS, and X ∠𝜃represents the true value of the measure-
The fault duration of the power system is set to three types:
ment vector of the WAMS after subtracting noise.
0.1, 0.3 and 0.5 s. The power flow level is considered as 70%,
It can be seen from Table 5 that various methods are affected
80%, 90%, 100% and 110% of the rated value. The fault loca-
to different degrees when they are disturbed by noise, but the
tion is applied on all buses and on 20%, 40%, 60% and 80% of
methods proposed in this paper can use information entropy to
distance away from the bus of each branch, and finally 34,725
filter noise in training data, so the impact of noise is reduced and
samples were generated, including 1243 unstable samples. The
accuracy can reach 99.48%. Although the number of labelled

FIGURE 8 South Carolina 500-bus system wiring

diagram
722 ZHANG ET AL.

0.05 TABLE 6 Performance comparison of each model

The
0.04 number of
Simulation labelled
Feature importance

Model Accuracy time (s) samples

0.03
SVM 99.57 223,060 22,306
RF 99.70 223,060 22,306
0.02
NN 98.69 223,060 22,306

0.01 Method proposed 99.73 75,959 5000

in this paper

-0.01
0 200 400 600 800
Feature rank
quently, the proposed method performs better than other meth-
FIGURE 9 Unranked feature importance of South Carolina 500-bus ods even in high-dimension data of power system.
system

5 CONCLUSIONS
0.05
To solve the problem of time-consuming data generation of
0.04 offline training caused by traditional data-driven TSA meth-
ods and the difficulty of quickly updating the model after
Feature importance

0.03 the topology of power system changes, the paper proposes a

TSA method based on AL, which effectively reduces the time
0.02 of offline simulation, greatly improve the efficiency of model
deployment. The main conclusions obtained by verifying on the
New England 39-bus power system are as follows:
0.01

1. Compared with existing methods, AL can reduce the num-

0 ber of samples labelled in transient status, thereby reducing
simulation time and improving model training efficiency;
-0.01 2. The method proposed in the paper uses information
0 200 400 600 800
entropy to select unlabelled data, which can query samples
Feature rank
with uncertainties and exclude redundant samples, thereby
improving the accuracy of the model;
FIGURE 10 Ranked feature importance of South Carolina 500-bus
system 3. The method proposed in the paper is insensitive to wide-area
noise, and can eventually achieve stable convergence, so it is
highly robust.
final 24,308 samples are used for training, and the remaining
samples are used for testing. The feature input of the algorithm The paper proposes a selecting method for the training sam-
is still composed of generator rotor angle, generator active and ples of the TSA model. Once the topology of power system
reactive power, bus voltage magnitude and angle, active and changes, the model still needs to be retrained. Therefore, in the
reactive power of branch, active and reactive power of load at future, how to construct global features and combine AL to
the time of fault clearance, whose feature dimension is much fine-tune the model to achieve online learning of the model is
higher than that of New England 39-bus power system. Fig- the focus of future research.
ures 9 and 10 show that almost all features have the posi-
tive importance, resulting in good performance in TSA. From ACKNOWLEDGEMENTS
Table 6, It can be seen that the proposed method has the high- This research was funded by Research on the stability of power
est accuracy while achieves the shortest simulation time. It is system with high penetration of power electronic equipment,
because the number of labelled samples used in the proposed and the grant number is JS71-20-004.
method is 5000, however, other methods need 22,306 samples.
Note that the simulation time means the time to perform time DATA AVAILABILITY STATEMENT
simulation to generate transient samples. Therefore, the simula- The data used to support the findings of this study are available
tion time of SVM, RF and NN are the same in Table 6. Subse- from the corresponding author upon request.
ZHANG ET AL. 723

REFERENCES 15. He, M., Vittal, V., Zhang, J.: Online dynamic security assessment with miss-
1. Liu, T., et al.: Non-parametric statistics-based predictor enabling online ing PMU measurements: A data mining approach. IEEE Trans. Power
transient stability assessment. IET Generation, Transmission & Distribu- Syst. 28(2), 1969–1977 (2013)
tion 12(21), 5761–5769 (2018) 16. Zhang, Y., Xu, Y., Dong, Z.Y.: Robust classification model for PMU-based
2. Kundur, P.: Power System Stability and Control, McGraw Hill Education, on-line power system DSA with missing data. IET Generation, Transmis-
New York (2002) sion & Distribution 11(18), 4484–4491 (2017)
3. Vittal, V., et al.: Derivation of stability limits using analytical sensitivity of 17. Zhang, Y., Xu, Y., Dong, Z.Y.: Robust ensemble data analytics for incom-
the transient energy margin. IEEE Trans. Power Syst. 4(4), 1363–1372 plete PMU measurements-based power system stability assessment. IEEE
(1989) Trans. Power Syst. 33(1), 1124–1126 (2017)
4. Zhang, Y., et al.: Wide-area frequency monitoring network (FNET) 18. Zhang, Y., et al. Online power system dynamic security assessment with
architecture and applications. IEEE Trans. Smart Grid 1(2), 159–167 incomplete PMU measurements: A robust white-box model. IET Genera-
(2010) tion, Transmission & Distribution 13(5), 662–668 (2018)
5. De La Ree, J., et al.: Synchronized phasor measurement applications in 19. Tan, B., et al.: A deep imbalanced learning framework for transient stability
power systems. IEEE Trans. Smart Grid 1(1), 20–27 (2010) assessment of power system. IEEE Access 7, 81759–81769 (2019)
6. Hashiesh, F., et al.: An intelligent wide area synchrophasor based system 20. Thirugnanasambandam, V., Jain, T.: AdaBoost classifiers for phasor
for predicting and mitigating transient instabilities. IEEE Trans. Smart measurements-based security assessment of power systems. IET Gener-
Grid 3(2), 645–652 (2012) ation, Transmission & Distribution 12(8), 1747–1755 (2017)
7. Wang, B., et al.: Power system transient stability assessment based on big 21. Zhou, Y., et al.: A novel data-driven approach for transient stability predic-
data and the core vector machine. IEEE Trans. Smart Grid 7(5), 2561– tion of power systems considering the operational variability. International
2570 (2016) Journal of Electrical Power & Energy Syst. 107, 379–394 (2019)
8. Papadopoulos, P.N., Guo, T., Milanović, J.V.: Probabilistic framework for 22. Begovic, M., et al.: Wide-area protection and emergency control. Proc.
online identification of dynamic behavior of power systems with renewable IEEE 93(5), 876–891 (2005)
generation. IEEE Trans. Power Syst. 33(1), 45–54 (2017) 23. Urbanowicz, R.J., et al.: Relief-based feature selection: Introduction and
9. Guo, T., Milanović, J.V.: Probabilistic framework for assessing the accuracy review. J. Biomed. Inf. 85, 189–203 (2018)
of data mining tool for online prediction of transient stability. IEEE Trans. 24. Fan, H. et al.: Post-fault transient stability assessment based on k-Nearest
Power Syst. 29(1), 377–385 (2013) neighbor algorithm with mahalanobis distance. 2018 International Confer-
10. Papadopoulos, P.N., Milanović, J.V.: Probabilistic framework for transient ence on Power System Technology (POWERCON). IEEE, (2018)
stability assessment of power systems with high penetration of renewable 25. Pai, M.A.: Energy Function Analysis for Power System Stability. Springer
generation. IEEE Trans. Power Syst. 32(4), 3078–3088 (2016) Science & Business Media, New York (1989) pp. 223–227
11. Yan, R., et al.: Fast transient stability batch assessment using cascaded con- 26. IEEE Standards Association. (2011). IEEE Standard for Synchrophasor
volutional neural networks. IEEE Trans. Power Syst. 34(4), 2802–2813 Data Transfer for Power Systems. IEEE Std C 37, 18–20 (2011)
(2019)
12. Zhu, Q., et al.: A deep end-to-end model for transient stability assessment
with PMU data. IEEE Access 6, 65474–65487 (2018)
13. Wu, S., et al.: Improved deep belief network and model interpretation How to cite this article: Zhang Y., et al.: A power
method for power system transient stability assessment. Journal of Modern
system transient stability assessment method based on
Power Systems and Clean Energy 8(1), 27–37 (2019)
14. James, J.Q., Hill, D.J., Lam, A.Y.S.: Delay aware transient stability assess- active learning. J. Eng. 2021, 715–723 (2021).
ment with synchrophasor recovery and prediction framework. Neurocom- https://doi.org/10.1049/tje2.12068
puting 322, 187–194 (2018)