1686 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 27, NO.

8, AUGUST 2016

Assessing Short-Term Voltage Stability of

Electric Power Systems by a Hierarchical
Intelligent System
Yan Xu, Member, IEEE, Rui Zhang, Junhua Zhao, Member, IEEE, Zhao Yang Dong, Senior Member, IEEE,
Dianhui Wang, Senior Member, IEEE, Hongming Yang, Member, IEEE,
and Kit Po Wong, Fellow, IEEE

Abstract— In the smart grid paradigm, growing integration results on the New England 39-bus system verify its superiority
of large-scale intermittent renewable energies has introduced in both learning speed and accuracy over some state-of-the-art
significant uncertainties to the operations of an electric power learning algorithms.
system. This makes real-time dynamic security assessment (DSA)
a necessity to enable enhanced situational-awareness against the Index Terms— Ensemble learning, intelligent system (IS),
risk of blackouts. Conventional DSA methods are mainly based power system, smart grid, voltage stability.
on the time-domain simulation, which are insufficiently fast
and knowledge-poor. In recent years, the intelligent system (IS) N OMENCLATURE
strategy has been identified as a promising approach to facilitate ANN Artificial neural network.
real-time DSA. While previous works mainly concentrate on the AVR Automatic voltage regulator.
rotor angle stability, this paper focuses on another yet increas-
ingly important dynamic insecurity phenomenon—the short- DAE Differential-algebraic equation.
term voltage instability, which involves fast and complex load DSA Dynamic security assessment.
dynamics. The problem is modeled as a classification subproblem DT Decision tree.
for transient voltage collapse and a prediction subproblem for LM Large motor.
unacceptable dynamic voltage deviation. A hierarchical IS is LVRT Low-voltage-ride-through.
developed to address the two subproblems sequentially. The
IS is based on ensemble learning of random-weights neural MAPE Mean absolute percentage error.
networks and is implemented in an offline training, a real- NNRW Neural networks with random weights.
time application, and an online updating pattern. The simulation OPF Optimal power flow.
PMU Phasor measurement unit.
Manuscript received January 30, 2014; revised March 23, 2015, SLFN Single-hidden layer feed-forward networks.
April 26, 2015, and May 17, 2015; accepted May 31, 2015. Date of publication SVM Support vector machine.
September 30, 2015; date of current version July 15, 2016. The work in SM Small motor.
this paper was partially supported by China Southern Power Grid research
project (WYKJ00000027), the University of Newcastle through the Faculty ST-VSA Short-term voltage stability assessment.
Strategic Pilot Grant, China State Key Laboratory for Alternate Electrical TDS Time-domain simulation.
Power Systems with Renewable Energy Sources, Australian Research Council TVCI Transient voltage collapse index.
through the Linkage Project under Grant LP120100302, the Faculty of
Engineering and Information Technologies, The University of Sydney, Sydney, TVSI Transient voltage severity index.
NSW, Australia, through the Faculty Research Cluster Program, and National UVLS Under-voltage load shedding.
Natural Science Foundation of China (Key Program 71331001, 71420107027).
The work of Y. Xu was also supported by the University of Sydney Post-
Doctoral Research Fellowship (Corresponding author: Junhua Zhao). I. I NTRODUCTION
Y. Xu, Z. Y. Dong, and H. Yang are with the Hunan Provincial Key
Laboratory of Smart Grids Operation and Control, International College,
Changsha University of Science and Technology, Changsha 410114, China,
and also with the School of Electrical and Information Engineering, University
T HE fundamental goal of an electric power system is to
maintain a secure and an economic process for electricity
generation, transmission, and distribution. However, in the
of Sydney, Sydney, NSW 2006, Australia (e-mail: [email protected]; practical operation phase, a power system is always exposed to
[email protected]).
R. Zhang is with the Centre for Intelligent Electricity Networks, various disturbances, such as sudden load changes, unexpected
University of Newcastle, Newcastle, NSW 2308, Australia (e-mail: element outages, short-circuit faults, and so on. Once occurred,
[email protected]). the disturbance can harm the power system security and
J. Zhao is with the Chinese University of Hong Kong (Shenzhen), Shenzhen,
Guangdong, China (e-mail: [email protected]). may result in catastrophic consequences such as wide-spread
D. Wang is with the Department of Computer Science and Information blackouts that affect the society and the economy significantly.
Technology, La Trobe University, Melbourne, VIC 3086, Australia (e-mail: As one essential security criterion, dynamic security refers
dh.wang@latrobe. edu.au).
K. P. Wong is with the School of Electrical Electronic and Computer to the ability of a power system to survive the dynamic
Engineering, University of Western Australia, Perth, WA 3009, Australia transition from an initial operating equilibrium to a viable
(e-mail: [email protected]). new equilibrium with most limits bounded after being subject
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. to a disturbance [1]. To protect a power system against the
Digital Object Identifier 10.1109/TNNLS.2015.2441706 risk of blackouts, DSA needs to be performed to examine
2162-237X © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
XU et al.: ASSESSING SHORT-TERM VOLTAGE STABILITY OF ELECTRIC POWER SYSTEMS 1687

various credible disturbances in advance. Once the disturbance an unacceptable dynamic voltage deviation subproblem that
is identified as insecure, appropriate security control actions corresponds to a regression process. The two subproblems
need to be implemented to protect the power system if the are sequentially handled by two ensemble-based classifier
contingency really occurs. Mathematically, the dynamics of a and predictor models. The developed IS is numerically tested
power system can be modeled as a large set of nonlinear DAEs, on the New England 39-bus system, and compared with
which cannot be analytically solved. The conventional DSA is other popular learning algorithms. The simulation results
therefore based on TDS that solves the DAEs via step-by-step have verified its effectiveness and superiority in accuracy and
integrations. Although accurate, the TDS method is very learning speed over the compared algorithms.
time-consuming and difficult to provide useful information
about the system’s dynamic security characteristics. II. P ROBLEM D ESCRIPTIONS
With continuous load growth yet unmatched infrastructure A. Short-Term Voltage Instability: Mechanism
investments, modern power systems are being pushed to and Phenomenon
operate near their dynamic security limit. In the meantime, The IEEE and CIGRE joint task force defines the voltage
with the evolution to the smart grid paradigm [2], today’s stability as the ability of a power system to maintain steady
power systems are integrated with a growing capacity of voltages at all buses in the system after being subjected to
renewable energy resources (such as wind and solar power), a disturbance from a given initial operating condition [1].
which are generally stochastic and difficult to predict, bringing As a major threat to the power system security, the loss of
significant uncertainties to the operation of a power system. voltage stability can result in progressive voltage drop or rapid
Consequently, system operators are expecting much faster voltage collapse, which can trigger significant load shedding
DSA tools to enable real-time monitoring of the power system or even wide-spread blackouts. Over the past decades, many
dynamic security condition. major blackouts around the world have been found directly
In recent years, the intelligent system (IS) strategy has been associated with this phenomenon [12]–[14].
identified as a promising approach to fulfill this need [3]. In general, voltage instability stems from the attempt of
By learning from a dynamic security database, the nonlinear load dynamics to restore the power consumption beyond
relationship between the power system operating parameters the capability of the combined transmission and generation
(input) and the corresponding security index (output) can be system [13]. According to [1], the voltage stability can be
extracted and reformulated in an IS. During the online applica- divided into the long-term and the short-term phenomena. The
tion phase, the system’s stability can be assessed as soon as the long-term voltage stability involves slower acting equipment,
input is available. The IS-based DSA not only provides much such as tap-changing transformers, thermostatically controlled
faster online decision-making speed, but also outperforms the loads, and generator current limiters [1]. It can be mod-
conventional methods for its less data requirement, stronger eled as algebraic equations and solved by power flow-based
generalization capacity, and extensibility [3], making it an methods [13]. Therefore, its computation speed can be fast
ideal candidate for future smart grid applications [4]. enough to enable (near) real-time implementation. By contrast,
In the literature, a variety of ISs based on learning the short-term voltage stability phenomenon involves complex
algorithms, such as ANN [5], [6], SVM [7], DT [8], and dynamics of load components such as induction motors
data-mining technique [9], have been reported. However, most tending to restore their consumed power in a very short
of them are designed for transient stability assessment, which time frame (say several seconds) [14]. Following a large
focuses on the generator rotor angle performance [5]–[10]. disturbance, the induction motors decelerate dramatically by
Yet, very few deal with other dynamic security criteria such as the voltage dip or may stall if the electrical torque cannot
voltage stability [6]. Besides, it is observed that the reported overcome the mechanical load. This in turn draws a very high
models usually suffer from extensive learning/tuning burden reactive current that affects adversely the voltage magnitudes.
and unsatisfactory accuracy. A comprehensive review and As a consequence, an unacceptable transient voltage perfor-
classification on previous works can be found in [11]. mance (e.g., delayed voltage recovery) may be experienced,
This paper focuses on another yet increasingly and/or transient voltage collapse may occur. With increased
important dynamic security concern—the short-term voltage penetration of induction motor loads (e.g., air-conditioners),
stability [12]–[14], and develops a hierarchical IS for today’s power systems tend to be more vulnerable to the
its real-time assessment. Unlike rotor angle stability, the short-term voltage instability [14].
short-term voltage stability involves a fast and complex
load dynamics, which is even more difficult to capture. The B. Dynamic Load Model
instability can propagate very rapidly and trigger significant As mentioned above, the key driving force of voltage
load shedding or even wide-spread blackouts. Nowadays, due instability is the tendency of dynamic loads to restore their
to the increased penetration of the wind power, the short-term consumed power within a very short time frame. Therefore,
voltage stability is becoming a critical concern in terms of modeling dynamic loads is a necessity for the short-term
LVRT capability for wind turbines. To address the ST-VSA voltage stability study.
problem, we employ a feedforward NNRWs [15] and employ In this paper, we employ a composite load model
the ensemble learning strategy to design a hierarchical IS. CLOD defined in the commercial power system simulation
We model the ST-VSA problem as a transient voltage collapse package Power System Simulator for Engineering (PSS/E)
subproblem that corresponds to a classification process and to represent various dynamic load components. As shown in

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
1688 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 27, NO. 8, AUGUST 2016

Fig. 1. Structure of the composite load model CLOD.

Fig. 3. Dynamic voltage responses without (top) and with (bottom) load
dynamics—a less severe fault.

dynamic voltage response in the New England 39-bus system

following the short-circuit disturbances with different severity
degrees [25].
The voltage trajectories are obtained by TDS [25]. For each
figure, the upper and lower windows, respectively, show the
Fig. 2. Dynamic voltage responses without (top) and with (bottom) load voltage response without the load dynamics (i.e., the dynamic
dynamics—a very severe fault. load model is not used, and a constant power load is used),
and with load dynamics (i.e., the dynamic load model is used).
Fig. 1, this load model is an aggregation of LM, SM, discharge For each figure, it can be seen that without load dynamics,
lighting, transformer exciting current, and constant power load, the bus voltages recover very fast to the normal level and
as well as the voltage-dependent load, all of which are fed the system is stable. In Fig. 2, a very severe disturbance is
from many real-world substations. However, other dynamic applied, and the lower window shows that transient voltage
load models can also be used if necessary. collapse occurs ∼1.6 s. The voltage collapse is an unaccept-
The mathematical model of induction motors is also able situation for a practical power system, since it can lead to
represented by DAEs. Hence, the whole mathematical model more serious consequences, such as cascading failures and/or
of the power system for the short-term voltage stability study blackouts. For Fig. 3, a less severe disturbance is applied,
composes a very large set of nonlinear DAEs that can be and the lower window shows that some buses (in particular,
compactly formulated as follows: bus 20 and bus 34) experience a prolonged voltage depression
dx and delayed voltage recovery. In practice, the delayed voltage
= f(x, y, p, λ) (1)
dt recovery can trigger UVLS devices that curtail a significant
0 = g(x, y, p, λ) (2) amount of loads.
D. Short-Term Voltage Stability Indices
where set (1) corresponds to the differential equations of the
system components, including generators, motors, dynamic As shown in Figs. 2 and 3, the short-term voltage stability
loads, as well as their control systems, and so on; and set (2) study mainly focuses on: 1) fast voltage collapse (see Fig. 2) or
corresponds to the algebraic equations of the network and 2) unacceptable postdisturbance voltage deviation (see Fig. 3).
static loads. Vector x denotes the state variable (e.g., generator To evaluate the short-term voltage performance, some
angles and dynamic bus voltages); vector y denotes algebraic industrial criteria, such as North American Electric Reli-
variables (e.g., static load voltages and angles); vector p stands ability Corporation criteria, has been proposed [24].
for controllable parameters (e.g., AVR set-points); and vector λ However, they only define the unacceptable voltage-dip
represents uncontrollable parameters (e.g., load levels). magnitude and the duration time, and can only provide a binary
The DAEs cannot be analytically solved. Rather, one answer to the stability status, namely, stable or unstable.
can solve it via step-by-step integrations. In practice, the In this paper, we model the ST-VSA problem as the transient
ST-VSA is realized based on TDS, which is, however, very voltage collapse subproblem and an unacceptable dynamic
computationally burdensome. voltage deviation subproblem. The first subproblem checks the
existence of a short-term equilibrium of the postcontingency
C. Simulation Study system and the second subproblem quantifies the degree of
Using the composite load model CLOD, we examine stability.
the impact of load dynamics on the short-term voltage 1) Transient Voltage Collapse Index: The evaluation of
performance of a power system. Figs. 2 and 3 show the transient voltage collapse is a binary question, that is,

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
XU et al.: ASSESSING SHORT-TERM VOLTAGE STABILITY OF ELECTRIC POWER SYSTEMS 1689

Fig. 5. Structure of an SLFN.

Fig. 4. Illustration of the concept of TVSI. Fig. 2 (lower window), and the TVSI is 0.37 for Fig. 3
(lower window).
whether or not the system loses the short-term equilibrium
following a disturbance. Consequently, the TVCI can be a III. P ROPOSED I NTELLIGENT S YSTEM
binary index, namely, 1 or 0. The ST-VSA can be converted as a classification subprob-
2) Transient Voltage Severity Index for Unacceptable lem for transient voltage collapse and a regression subproblem
Dynamic Voltage Deviation: After the system withholds a for unacceptable dynamic voltage deviation. The latter depends
short-term equilibrium, the system may still experience unac- on the result of the former. In this paper, we develop a
ceptable voltage deviation, such as prolonged voltage depres- hierarchical IS for the real-time ST-VSA. The IS is based on
sion and delayed voltage recovery. Actions of motor protective the ensemble learning strategy where NNRWs are employed.
devices can trip a significant amount of loads. In contrast
to transient voltage collapse, the severity of an unacceptable A. Neural Nets With Random Weights
dynamic voltage deviation can be varying depending on
NNRWs were originally proposed in [15] which suggested
the severity of a disturbance. Hence, a continuous index
a fast implementation of learning for SLFNs (see Fig. 5).
should be used to quantify such phenomenon. In practice,
Given a training data set with N instances in total,
the degree of severity in voltage deviation can be reflected
ℵ N = {(x j , t j )|x j ∈ R n , t j ∈ R m } Nj=1 , where x j is the n × 1
by the magnitude of the voltage violation and the associated
input vector and t j is a m ×1 target vector, the output function
duration time. This paper employs a straightforward index
of the SLFN with Ñ hidden nodes is
called TVSI proposed in [25] to quantify this phenomenon
 N T 
Ñ
i=1 t =Tc TVDIi,t f Ñ (x j ) = βi · ϑ(wi · x j + bi ) = t j , j = 1, 2, . . . , N
TVSI = (3)
N × (T − Tc ) i=1

where N is the total number of buses in the system, T is (5)

the considered transient time frame, Tc is the fault clearing where ϑ is the activation function, wi is the weight vector
time, and TVDI is the transient voltage deviation index, connecting the ith hidden node and the input nodes, βi is the
calculated by weight vector connecting the i th hidden node and the output
⎧ nodes, bi is the bias of the i th hidden node, and wi ·x j denotes
⎨|Vi,t − Vi,0 | , if |Vi,t − Vi,0 | ≥ μ
TVDIi,t = Vi,0 Vi,0 ∀t ∈ [Tc , T ] the inner product of wi and x j .
⎩ NNRW is completely different from the traditional iterative
0, otherwise
learning based ANNs as it learns via randomly selecting the
(4)
input weights and biases for hidden nodes w and b, and
where Vi,t denotes the voltage magnitude of bus i at time t, analytically determining the output weights β via direct matrix
which is obtained from TDS, and μ is the threshold to calculations [15]. Equation (5) can be rewritten into a compact
define unacceptable voltage deviation level, which can be set format as
according to the industrial criteria, e.g., 20%.
Hβ = T (6)
As shown in Fig. 4. TVSI only accounts for the buses
with unacceptable voltage violation during the transient period, where H is called the hidden layer output matrix of the
and it measures not only the voltage violation magnitude but network.
also the associated duration time in the whole system level. On a training data set, given the activation function and
It can provide a quantitative comparison of the system tran- the hidden node number, the NNRW learning process can be
sient voltage performance following a disturbance. A smaller briefly summarized as three major steps.
TVSI value means the transient voltage performance is better. Step 1: Randomly generate the input weights w i and bi ,
Both of the indices can be calculated via performing TDS for i = 1, . . . , Ñ .
to obtain the voltage trajectories, while the former is a binary Step 2: Calculate the hidden layer output matrix H.
value and the later is a continuous value. Taking the simulation Step 3: Calculate the output weights matrix β = H† T,
results in Section II-C, for example, the TVCI is 0 for where H† can be Moor–Penrose generalized inverse of H.

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
1690 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 27, NO. 8, AUGUST 2016

NNRW can be applied for both data regression and pattern

classification, where the real-time requirement with domain
applications is concerned with. Compared with the traditional
learning algorithms, NNRW builds a learner model through
solving a standard linear regression problem; therefore, its
learning speed can be thousands times faster and requires
much less computation memory. Another important merit of
NNRW is its efficient tuning mechanism: given an activation
function, only the hidden neuron nodes number needs to be
tuned, which can be efficiently achieved via a linear validation
procedure. Besides, NNRW can avoid difficulties, such as
stopping criteria, learning rate setting, learning epochs, and
local minima, that can be commonly encountered by the
traditional algorithms.
In the literature, NNRW has been comprehensively validated
on a number of benchmark and real-world problems from
both classification and regression areas, showing the feasibility
for data modeling in real-time. However, the random weight
neural networks perform unstably because each resulting
model can be regarded as an instance of a random variable.
Some supervised mechanisms should be researched, so that
the selected random weights to initialize the neural networks
can achieve a better modeling performance (at least better than Fig. 6. Proposed ensemble model of NNRWs.
the mean of the random models).
of each NNRW can be diverse. In result-aggregation stage, for
B. Neural Network Ensemble
regression problem, the average value of the individual outputs
The simplicity and efficiency properties of the NNRW is used, and for classification problem, the majority class of the
make it a promising candidate for tackling large-scale learning individual outputs is adopted as the final classification result.
tasks. However, as an undesired side effect, the randomly It is important to note that, in addition to inherently selecting
selected input parameters of the NNRW may lead to stability random input layer parameters for single NNRWs, the ensem-
issues, meaning that the classification/regression results are not ble model also assigns random hidden node numbers to the
consistent for different runs. NNRW individuals; this, on the one hand, aims to expand the
As extensively investigated and demonstrated in the internal diversity of the ensemble model, which is paramount
literature, ensemble learning scheme, such as bagging and for accuracy improvement, and on the other hand, is motivated
boosting [17], is an effective strategy to increase the accuracy by the observation that the NNRW usually has an optimal
of a single learner. By combining a set of different single range rather than a fixed number of hidden nodes (since under
learners, they can compensate for each other, and the whole different random input parameter sets, the optimal number of
model can reduce aggregated variance and bias, and therefore hidden nodes can be varied) [10]. In doing so, the randomness
tends to increase accuracy over the individuals [17]–[19]. can be generalized, leading to enhanced robustness.
As a general strategy open to any learning methods, the
feasibility and efficiency of the ensemble learning, however,
depends greatly on the characteristic of the individual learners. C. Application for Real-Time ST-VSA
Since NNRW is stochastic in nature, and has a very fast In the offline stage, a voltage stability database is prepared
learning speed, it is an ideal candidate for ensemble learning. by performing comprehensive TDS considering various
Besides, benefiting from the ensemble effect, the robustness scenarios and/or collecting fault records. Then, the IS is
of single NNRWs can be improved by the whole model. trained by the database. In the real-time application stage,
The effectiveness of building ensemble for NNRWs has been once the input is available, the short-term voltage stability
verified in [10]. can be assessed by the IS rapidly, and the results are used
Based on the idea proposed in [10], this paper designs a to support control decision-makings. Also in the online stage,
new ensemble model of NNRWs for the ST-VSA problem. with updated system information (e.g., network structure, load
As shown in Fig. 6, in the ensemble model, M single NNRWs composition, and so on), a new stability database can be gen-
are combined in parallel, all of which have common input erated to update/enrich the IS. Note that this online updating
variables. In the initialization stage, the ensemble model aims to keep the IS adaptive to unexpected system changes.
randomly assigns a hidden node number to each NNRW within The proposed hierarchical IS for the real-time ST-VSA is
an optimal range, [L 1 , L 2 ], which is determined from a pre- conceptually described in Fig. 7.
tuning procedure. In the training stage, the single NNRWs are The IS consists of an ensemble classifier and an ensemble
trained according to Step 1–Step 3 and export their individual predictor based on the proposed NNRW ensemble. They
outputs. Given the stochastic nature of the NNRW, the output are pretrained in the offline stage with forecasted operating

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
XU et al.: ASSESSING SHORT-TERM VOLTAGE STABILITY OF ELECTRIC POWER SYSTEMS 1691

the irrelevant/redundant features can be removed and the total

size of the training data can be significantly reduced, resulting
in faster training speed and higher accuracy. For the power
system DSA applications, an automatic feature selection can
identify the significant operating variables that best describe
the dynamic security characteristics of the system, and hence
provide useful information for security control [21]. Besides,
using only the significant features as the DSA model input can
reduce the parameter measurement requirements.
Zhang et al. [21] present a comprehensive survey and a
review of feature selection techniques for intelligent DSA and
introduce an alternative algorithm called RELIEF. Unlike other
feature selection algorithms, the idea behind RELIEF is to
evaluate the quality of single features according to how well
their values distinguish among instances near each other.
It not only considers the difference in features’ values and
classes, but also the distance between the instances, so good
features can gather similar instances and separate dissimilar
ones, providing qualitative estimation on the importance of the
features. RELIEF has been successfully used in our previous
works and yield satisfactory results [11], [21]. This paper also
applies the RELIEF for the automatic feature selection for the
proposed ST-VSA model.
The basic RELIEF is to iteratively update the weight of
each feature as follows [23]:

Fig. 7. Proposed hierarchical IS for the ST-VSA. W [X]i+1 = W [X]i − diff(X, Ri , H )/m
+ diff(X, Ri , M)/m, i ≤ m (7)
conditions (e.g., day ahead) and corresponding voltage col-
where X denotes a feature, Ri is the instance sampled in the
lapse status and TVSI. For the online application, the real-
i th iteration, H is the nearest instance from the same class
time system measurements (including generator outputs, bus
as Ri (called nearest hit), while M is the nearest instance
loads, bus voltage magnitudes, bus angles, and so on) are
from the different classes with Ri (called nearest miss), and
collected by Supervisory Control And Data Acquisition or
m is the number of sampled instances guaranteeing the weights
PMU devices. Once the input measurements are obtained, they
between [−1, 1]. Function diff(X, R, R  ) calculates the
are first dropped to the classifier which determines the transient
difference between the values of feature X for two instances
voltage collapse condition of an unknown operating condition
R and R 
with respect to a certain disturbance, if voltage collapse
|value(X, R) − value(X, R  )|
is ascertained, remedial control actions should be taken to diff(X, R, R  ) = . (8)
protect the power system from the voltage collapse should max(X) − min(X)
the disturbance really occurs; otherwise, the measurements From the statistics point of view, the weight of feature X is
are then dropped to the predictor to evaluate the TVSI of the an approximation of the difference of probabilities
operating condition. If the predicted TVSI is unacceptable,
W [X] = P(diff. value of X|nearest inst. from the diff. class)
e.g., less than a certain threshold, remedial actions will be
armed. During the online application phase, the classifier and − P(diff. value of X|nearest inst. from the same class).
predictor can also be updated using the latest online operating (9)
information.
To deal with noises, incomplete data, and multiclass
Due to the very fast training speed of the NNRW, the
problems, RELIEF was extended with the following weight
IS model can be updated online. The updating can be done
updating equation [23]:
periodically, e.g., hourly or half-daily, depending on practical
needs. The data for updating can be obtained from TDS results 
k
on the real-time system operation points. The simulation W [X]i+1 = W [X]i − diff(X, Ri , H j )/(m · k)
results can be saved as new training data. Such new training j =1
⎡
data can be used as a new set or attached to the initial training 
k
P(C)
database to retrain the IS. + ⎣
1 − P(class(Ri ))
C =class(Ri )
⎤
D. Automatic Feature Selection 
k

Automatic feature selection aims to select a most important · diff(X, Ri , M j (C))⎦ (m · k) (10)
subset from the whole input set for training the model. Thus, j =1

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
1692 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 27, NO. 8, AUGUST 2016

where C is a class label, P(C) is the prior probability of TABLE I

class C, and k is an user-defined parameter. S TRUCTURE OF THE S TABILITY D ATABASE
Instead of finding the nearest one hit and miss,
(10) finds k nearest hits and misses to average their
contribution in updating the weight, thereby reducing the risk
of miss estimation; the introduction of P(C) aims to estimate
the ability of separating each pair of classes.
To identify a set of significant features for training the
NNRW ensemble, the features with higher W values can be
selected since they can better differentiate instances among
classes. As supplementary benefits, the selected features can
provide insight into the critical variables that impact the
short-term voltage most, which is valuable for designing
preventive controls [15].

IV. S IMULATION R ESULTS

The developed hierarchical IS is tested on the New England
10-machine 39-bus system (Fig. 8), which is a popular test
system for stability study [20].
The numerical simulation is conducted on a 64-bit
PC with 3.1-GHz CPU and 4-GB RAM. TDS is performed
using commercial software PSS/E [20], and the proposed
algorithms are realized in the MATLAB platform. For the
test system, the dynamic load parameter is assumed as:
1) LM-20%; 2) SM-30%; 3) discharge lighting-5%;
4) transformer exciting current-5%; 5) constant power
load-30%; and 6) voltage-dependent load-10%. Fig. 8. One-line diagram of the New England 10-machine 39-bus system.

A. Database Generation
To train the proposed IS model, a comprehensive
stability database is necessary. In practice, this can be obtained
by performing simulations on various scenarios and/or fault
recordings. In this paper, we artificially generate such a
database through OPF calculation and TDS [11]. A total of
700 different operating states covering a variety of different
load/generation patterns ranging from 80% to 120% of the
base loading level are generated through OPF calculation.
A three-phase short-circuit fault is considered first (denoted as
fault #1). The fault is applied to bus 15 with a duration time
of 0.2 s. For each operating condition, the following candidate Fig. 9. Histogram plot of the TVSI of nonvoltage collapse samples.
features are selected as the input vector: 1) P, Q load, voltage
magnitude, and angle of each load bus; 2) P, Q generation of while 0 not. Only the samples with 0 TVCI have a TVSI. The
each generator; and 3) total P, Q load and generation of the histogram plot of the TVSI of the samples is given in Fig. 9,
whole system. For each operating condition, there are a total where it can be seen that the range [0.3, 1.5] occupies the
of 136 candidate features. major proportion.
TDS is then performed on each generated operating Note that during the training phase the classifier is trained by
condition to determine their TVCI and TVSI. The total all of the samples and the TVCI is used as the training target,
simulation time is 5 s and the simulation step is 0.01 s. The while the predictor is only trained by the samples whose TVCI
illustration of the dynamic response of this system under the is 0, and only the TVSI is used as the training target.
assumed fault is already given in the lower window of Fig. 3. Given the 700 samples, we randomly pick up 25%,
In calculating the TVSI, μ is set to 20%. For the 700 generated i.e., 175 samples as the testing set, leaving the remaining
samples, 256 of them (36.6%) experience voltage collapse and 525 samples serve as the training set.
444 (63.4%) do not experience voltage collapse.
The structure of the short-term voltage stability database is B. Feature Selection
shown in Table I. We use the RELIEF algorithm to select the significant
Note that the input features are normalized into the range features. The RELIEF values of the 136 features are shown
of [0, 1]. For TVCI, 1 represents the voltage collapse occurs in Fig. 10, where it can be seen that different features can

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
XU et al.: ASSESSING SHORT-TERM VOLTAGE STABILITY OF ELECTRIC POWER SYSTEMS 1693

Fig. 10. RELIEF value for each candidate feature.

Fig. 12. Tuning profile for the predictor.

TABLE II
P ERFORMANCE IN T ERMS OF A CCURACY (FAULT #1)

range occurs when 30 to 60 hidden nodes are used. For the

Fig. 11. Tuning profile for the classifier. predictor, a similar trend is also observed (with continuously
added hidden nodes, the prediction error decreases sharply and
varying capabilities to distinguish the operating conditions in then increases gradually), and when 15 to 50 hidden nodes are
terms of the status of the short-term voltage stability. A large used, the resulting MAPE is at the lowest range. Therefore,
RELIEF value indicates a stronger capability. The top three these hidden node numbers are used to train the ensemble
features are the voltage angle of bus 24, Q load of bus 27, model.
and voltage angle of bus 26, all of them are voltage related Also from Figs. 11 and 12, it can be observed from
variables of the system. In the following test, we use the the tuning process that a single NNRW can achieve
top 20 features as the model inputs. maximum 95.5% classification accuracy and minimum 4.56%
prediction MAPE. Besides, it is worth mentioning that the
C. Model Tuning training of a single NNRW is very fast, it only costs 0.01 s to
learn the tuning data set.
According to the NNRW theory, a certain hidden node
number should be predetermined before training. This cor-
responds to the optimal hidden node range [L 1 , L 2 ] for the D. Performance Test
NNRW ensemble. Since the hidden node has a direct impact In this paper, we use 100 individual NNRWs to build the
on the accuracy of a single NNRW, the range [L 1 , L 2 ] needs ensemble model. Once the model is well trained, it can be
to be well tuned in advance. For this purpose, the training applied online. As soon as the input variables (significant
data set is randomly divided into two nonoverlapped sets, features) are available (e.g., measured by PMUs), the short-
one for training and the other for validation. [L 1 , L 2 ] can term voltage stability can be determined instantaneously.
be determined at the range where the lowest validation error The testing set is used to validate the trained IS model. For
is obtained. To measure the performance of the proposed IS, comparison purpose, the single NNRW is also tested. Besides,
percentage classification accuracy is used for the classifier, and some state-of-the-art single learning algorithms, including
the MAPE is used for the predictor Radial basis function neural network, SVM and DT, are
d also tested with the same training and testing data sets. The
|yi − ŷi |
MAPE = i=1 (11) performance of the algorithms is presented in Tables II and III.
|yi | · d As Table II shows, the proposed IS outperforms the others
where yi and ŷi are the actual and predicted TVSI value, in both classification and prediction performance. Importantly,
respectively, and d is the total number of the testing samples. it should be noted that the proposed IS significantly increases
The tuning profiles are shown in Figs. 11 and 12. It can the performance over the single NNRW, achieving a very high
be seen that for the classifier, with continuously increased classification accuracy and a very low prediction error.
hidden nodes, the classification accuracy increases sharply, As Table III shows, the training efficiency of the proposed
but then reduces gradually. The highest classification accuracy IS is very high, although 100 NNRWs are used, the overall

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
1694 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 27, NO. 8, AUGUST 2016

TABLE III TABLE VI

P ERFORMANCE IN T RAINING E FFICIENCY (FAULT #1) TVSI P REDICTION MAPE

TABLE IV
FAULT C ONTINGENCIES FOR E XTENDED T EST

TABLE V
V OLTAGE C OLLAPSE C LASSIFICATION A CCURACY

training time is 28 s, depending on the number of hidden

nodes, one single NNRW only cost ∼0.3 s for training, this
owes to the unique learning process of the NNRW—randomly
selecting the input weights and analytically determining
Fig. 13. Average performance on the three contingencies. (a) Classification
the output weights. Compared with other learning models, for voltage collapse. (b) Prediction for TVSI.
DT is fastest since its learning procedure does not require
time-consuming iterative process. SVM and RBFNN are To summarize the overall performance, the average perfor-
generally slower due to the time-consuming optimization mance metrics on the three contingencies for each model are
process during the model training. Given the very high compared in Fig. 13. In average, the proposed IS model still
training speed, the online updating of the model can be outperforms the others in both classification and prediction
readily achieved, e.g., periodically and half-daily. Once more subproblems, owing to the ensemble effect.
TDS results are available, they can be attached to the original
training data to retrain the IS model. For all of the models, V. C ONCLUSION
the classification/prediction process is very fast, which costs
around 1 ms for one single test sample, indicating a real-time Today’s power systems are integrated with growing capacity
decision-making speed for ST-VSA. of renewable energies, such as wind and solar power. These
renewable energies are intermittent in nature, and hence can
introduce significant uncertainties and complexities to secure
E. Extended Accuracy Test and reliable operation of the power system. In order to protect
In the above test, only one fault contingency is consid- the power system against dynamic insecurity that could lead
ered. In practice, it is usually needed to assess more fault to blackouts, the real-time DSA is becoming necessary in
contingencies depending on their likelihood to occur. For a the control center. The IS strategy is a promising approach
comprehensive test of the accuracy, another two contingencies to enable the real-time DSA. Although widely researched in
are tested here. The two contingencies have previously been the literature, most works focus on the rotor angle stability
studied in [25] and their details are given in Table IV. Given problem. This paper concentrates on another yet increas-
the 700 generated operating states, the two contingencies are ingly important dynamic insecurity phenomenon—the short-
simulated yielding to another two databases. The databases term voltage instability, which involves a fast and complex
are then used to train and test the proposed IS model and load dynamics. We decompose the whole problem into a
compared with the state-of-the-art algorithms through the same transient voltage collapse subproblem that corresponds to a
procedure of the above test. The classification and prediction classification process, and an unacceptable dynamic voltage
performance are shown in Tables V and VI, respectively. deviation subproblem that corresponds to a regression process.

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
XU et al.: ASSESSING SHORT-TERM VOLTAGE STABILITY OF ELECTRIC POWER SYSTEMS 1695

The so-called TVSI is used to quantify the unacceptable [18] M. Alhamdoosh and D. Wang, “Fast decorrelated neural network ensem-
dynamic voltage deviation. Then, a hierarchical IS is devel- bles with random weights,” Inf. Sci., vol. 264, pp. 104–117, Apr. 2014.
[19] K. Meng, Z. Y. Dong, K. P. Wong, Y. Xu, and F. J. Luo, “Speed-up the
oped to address the two subproblems sequentially. The IS is computing efficiency of power system simulator for engineering-based
based on the ensemble learning strategy and NNRW algorithm. power system transient stability simulations,” IET Generat., Transmiss.
It can be implemented in an offline training, a real-time appli- Distrib., vol. 4, no. 5, pp. 652–661, May 2010.
[20] Y. Xu, Z. Y. Dong, L. Guan, R. Zhang, K. P. Wong, and F. Luo,
cation, and an online updating pattern. The RELIEF algorithm “Preventive dynamic security control of power systems based on pat-
is used to select the most significant features as the training tern discovery technique,” IEEE Trans. Power Syst., vol. 27, no. 3,
input. The proposed IS is validated on the New England pp. 1236–1244, Aug. 2012.
[21] R. Zhang, Y. Xu, Z. Y. Dong, and D. J. Hill, “Feature selection for
39-bus system and compared with some other state-of-the- intelligent stability assessment of power systems,” in Proc. IEEE PES
art learning algorithms. The simulation results show that the General Meeting, San Diego, CA, USA, Jul. 2012, pp. 1–7.
hierarchical IS can effectively increase the performance over [22] K. Kira and L. A. Rendell, “A practical approach to feature selection,”
in Proc. 9th Int. Workshop Mach. Learn., 1992, pp. 249–256.
a single NNRW and outperform the compared algorithms in [23] I. Kononenko, “Estimating attributes: Analysis and extensions of
both classification and prediction. RELIEF,” in Proc. Eur. Conf. Mach. Learn., 1994, pp. 171–182.
In the future, it would be interesting to expand the proposed [24] D. J. Shoup, J. J. Paserba, and C. W. Taylor, “A survey of current
practices for transient voltage dip/sag criteria related to power system
model to other stability assessment problems, e.g., frequency stability,” in Proc. IEEE PES Power Syst. Conf. Expo., Oct. 2004,
stability assessment. Besides, more efforts can be devoted to pp. 1140–1147.
the ensemble strategy and results aggregation scheme. [25] Y. Xu, Z. Y. Dong, K. Meng, W. F. Yao, R. Zhang, and K. P. Wong,
“Multi-objective dynamic VAR planning against short-term voltage
instability using a decomposition-based evolutionary algorithm,” IEEE
R EFERENCES Trans. Power Syst., vol. 29, no. 6, pp. 2813–2822, Nov. 2014.

[1] P. Kundur, “Definition and classification of power system stability

IEEE/CIGRE joint task force on stability terms and definitions,” IEEE
Trans. Power Syst., vol. 19, no. 3, pp. 1387–1401, May 2004. Yan Xu (S’10–M’13) received the B.E. and M.E. degrees from the
[2] F. Li et al., “Smart transmission grid: Vision and framework,” IEEE South China University of Technology, Guangzhou, China, in 2008 and
Trans. Smart Grid, vol. 1, no. 2, pp. 168–177, Sep. 2010. 2011, respectively, and the Ph.D. degree from The University of Newcastle,
[3] Z. Y. Dong, Y. Xu, P. Zhang, and K. P. Wong, “Using IS to assess an Callaghan, NSW, Australia, in 2013.
electric power system’s real-time stability,” IEEE Intell. Syst., vol. 28, He was a Research Fellow with the Center for Intelligent Electricity
no. 4, pp. 60–66, Jul./Aug. 2013. Networks, The University of Newcastle, from 2013 to 2014. He is currently the
[4] P. Zhang, F. Li, and N. Bhatt, “Next-generation monitoring, analysis, University of Sydney Post-Doctoral Fellow with the School of Electrical and
and control for the future smart control center,” IEEE Trans. Smart Grid, Information Engineering, University of Sydney, Sydney, NSW, Australia. His
vol. 1, no. 2, pp. 186–192, Sep. 2010. current research interests include power system stability and control, power
[5] Y. Mansour, E. Vaahedi, and M. A. El-Sharkawi, “Dynamic security system planning, smart grid, and intelligent system applications to power
contingency screening and ranking using neural networks,” IEEE Trans. engineering.
Neural Netw., vol. 8, no. 4, pp. 942–950, Jul. 1997.
[6] D. Q. Zhou, U. D. Annakkage, and A. D. Rajapakse, “Online monitoring
of voltage stability margin using an artificial neural network,” IEEE
Rui Zhang (S’12) received the B.E. degree from the University of
Trans. Power Syst., vol. 25, no. 3, pp. 1566–1574, Aug. 2010.
Queensland, Brisbane, QLD, Australia, in 2009, and the Ph.D. degree from
[7] L. S. Moulin, A. P. Alves da Silva, M. A. El-Sharkawi, and
The University of Newcastle, Callaghan, NSW, Australia, in 2014.
R. J. Marks, II, “Support vector machines for transient stability analysis
She is currently a Research Associate with the Center for Intelligent Elec-
of large-scale power systems,” IEEE Trans. Power Syst., vol. 19, no. 2,
tricity Networks, The University of Newcastle. Her current research interests
pp. 818–825, May 2004.
include power system security operation and control, power system planning,
[8] K. Sun, S. Likhate, V. Vittal, V. S. Kolluri, and S. Mandal, “An online
intelligent system techniques, and their applications to power engineering.
dynamic security assessment scheme using phasor measurements and
decision trees,” IEEE Trans. Power Syst., vol. 22, no. 4, pp. 1935–1943,
Nov. 2007.
[9] Y. Xu, L. Guan, Z. Y. Dong, and K. P. Wong, “Transient stability Junhua Zhao (S’04–M’07) received the B.E. degree from Xi’an Jiaotong
assessment on China Southern Power Grid system with an improved University, Xi’an, China, in 2004, and the Ph.D. degree from the University
pattern discovery-based method,” in Proc. Int. Conf. Power Syst. of Queensland, Brisbane, QLD, Australia, in 2007.
Technol., Hangzhou, China, Oct. 2010, pp. 1–6. He is currently an Associate Professor with the Chinese University of
[10] Y. Xu, Z. Y. Dong, J. H. Zhao, P. Zhang, and K. P. Wong, “A reliable Hong Kong (Shenzhen), Shenzhen, Guangdong, China. His current research
intelligent system for real-time dynamic security assessment of power interests include power system analysis and computation, electricity market
systems,” IEEE Trans. Power Syst., vol. 27, no. 3, pp. 1253–1263, management and analysis, nonlinear optimization, statistical methods, data
Aug. 2012. mining, and their applications in power engineering.
[11] R. Zhang, Y. Xu, Z. Y. Dong, K. Meng, and Z. Xu, “Intelligent systems
for power system dynamic security assessment: Review and classifica-
tion,” in Proc. 4th Int. Conf. Electr. Utility Deregulation Restruct. Power
Technol., Weihai, China, Jul. 2011, pp. 134–139. Zhao Yang Dong (M’99–SM’06) received the Ph.D. degree from the
[12] C. W. Taylor, Power System Voltage Stability. New York, NY, USA: University of Sydney, Sydney, NSW, Australia, in 1999.
McGraw-Hill, 1994. He was an Ausgrid Chair Professor and the Director of the Centre for
[13] T. V. Cutsem and C. Vournas, Voltage Stability of Electric Power Intelligent Electricity Networks with The University of Newcastle, Callaghan,
Systems. Norwell, MA, USA: Kluwer, 1998. NSW, Australia. He held academic and industrial positions with The
[14] J. A. Diaz de Leon and C. W. Taylor, “Understanding and solving short- Hong Kong Polytechnic University, Hong Kong, the University of Queensland,
term voltage stability problems,” in Proc. IEEE PES Summer Meeting, Brisbane, QLD, Australia, and Transend Networks, Lenah Valley, TAS,
Chicago, IL, USA, Jul. 2002, pp. 745–752. Australia. He is currently a Professor and the Head of the School of Electrical
[15] W. F. Schmidt, M. A. Kraaijveld, and R. P. W. Duin, “Feedforward and Information Engineering with the University of Sydney. His current
neural networks with random weights,” in Proc. 11th IAPR Int. Conf. research interests include smart grid, power system planning, power system
Pattern Recognit. Methodol. Syst., Aug./Sep. 1992, pp. 1–4. security, load modeling, renewable energy systems, electricity market, and
[16] J. W. Han and M. Kamber, Data Mining: Concepts and Techniques. computational intelligence and its application in power engineering.
San Francisco, CA, USA: Morgan Kaufmann, 2001. Prof. Dong is an Editor of the IEEE T RANSACTIONS ON S MART G RID,
[17] L. K. Hansen and P. Salamon, “Neural network ensembles,” IEEE Trans. the IEEE T RANSACTIONS ON S USTAINABLE E NERGY, and the IEEE P OWER
Pattern Anal. Mach. Intell., vol. 12, no. 10, pp. 993–1001, Oct. 1990. E NGINEERING L ETTERS .

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.
1696 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 27, NO. 8, AUGUST 2016

Dianhui Wang (M’03–SM’05) received the Ph.D. degree from the School Kit Po Wong (M’87–SM’90–F’02) received the M.Sc., Ph.D., and D.Eng.
of Information Science and Engineering, Northeastern University, Shenyang, degrees from the Institute of Science and Technology, University of
China, in 1995. Manchester, Manchester, U.K., in 1972, 1974, and 2001, respectively.
He was a Post-Doctoral Fellow with the School of Electronic and He was with the Department of Electrical Engineering, The Hong Kong
Electrical Engineering, Nanyang Technological University, Singapore, from Polytechnic University, Hong Kong, as a Chair Professor from 2002 to 2011
1995 to 1997. He was a Research Associate and Research Fellow with and the Head from 2002 to 2007. He has been with the School of Electrical,
the Department of Computing, The Hong Kong Polytechnic University, Electronic and Computer Engineering, The University of Western Australia,
Hong Kong, for three years, in 2001. Since 2001, he has been with the Perth, WA, Australia, since 1974, where he is currently a Winthrop Professor.
Department of Computer Science and Computer Engineering, La Trobe He is currently a Conjoint Professor with The University of Newcastle,
University, Melbourne, VIC, Australia, where he is currently a Reader and Callaghan, NSW, Australia. His current research interests include power
an Associate Professor. Since 2010, he has been an Adjunct Professor with system analysis, planning and operations, smart grids, and renewable energy.
the State Key Laboratory of Synthetical Automation for Process Industries, Prof. Wong received three Sir John Madsen Medals from IEAust in
Northeastern University. 1981, 1982, and 1988, the Outstanding Engineer Award from the IEEE
Dr. Wang serves as an Associate Editor of several international journals, Power Chapter Western Australia in 1999, and the IEEE Third Millennium
including the IEEE T RANSACTIONS ON N EURAL N ETWORKS AND L EARN - Award in 2000. He was a Co-Technical Chairman of the IEEE Machine
ING S YSTEMS , the IEEE T RANSACTIONS ON C YBERNETICS , Information Learning and Cybernetics Conference in 2004, and the General Chairman
Sciences, and Neurocomputing. of the IEEE/CSEE PowerCon in 2000. He is the Editor-in-Chief of the
IEEE P OWER E NGINEERING L ETTERS , and was the Editor-in-Chief of the
IEE Proceedings—Generation, Transmission and Distribution.

Authorized licensed use limited to: Oman Virtual Science Library (Masader). Downloaded on August 26,2024 at 11:41:21 UTC from IEEE Xplore. Restrictions apply.