Electrical Power and Energy Systems 87 (2017) 43–51

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Voltage stability monitoring of power systems using reduced network

and artificial neural network
Syed Mohammad Ashraf ⇑, Ankur Gupta, Dinesh Kumar Choudhary, Saikat Chakrabarti
Department of Electrical Engineering, Indian Institute of Technology Kanpur, Kanpur, India

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents network reduction based methodologies to monitor voltage stability of power sys-
Received 13 January 2016 tems using limited number of measurements. In a multi-area power system, artificial neural networks
Received in revised form 26 September (ANNs) are used to estimate the loading margin of the overall system, based on measurements from
2016
the internal area only. Information regarding the important measurements from the external areas is con-
Accepted 16 November 2016
sidered in measurement transformation through the network reduction process, to enhance the estima-
tion accuracy of the ANNs. A Z-score based bad or missing data processing algorithm is implemented to
make the methodologies robust. To account for changing operating conditions, adaptive training of the
Keywords:
Artificial neural network
ANNs is also suggested. The proposed methods are successfully implemented on IEEE 14-bus and 118-
Continuation power flow bus test systems.
Network reduction Ó 2016 Elsevier Ltd. All rights reserved.
Voltage stability

1. Introduction scenario. In order to monitor the voltage security in real-time, it

is necessary that the process of measurement and estimation of
The electrical power demand is increasing day-by-day, and the the state variables and the analysis be performed within a desired
generation is limited. Deregulated and open electricity markets time frame. Conventionally, remote terminal units (RTUs) have
are norms of the day, and it is essential to maintain reliable and good been used to collect measurements from various locations in a
quality of the electricity supply. All this has resulted in large inter- power system. The refresh rate of RTU measurements is typically
connected power networks that are operated under heavily loaded a few seconds. With the advent of phasor measurement units
conditions, and are often close to their stability limits. Power system (PMUs), it is now possible to obtain the measurement data at a
voltage instability is now one of the challenging problems faced by sub-second rate [10–12]. In the presence of conventional measure-
the utilities. Modern day energy management systems (EMSs) have ments (from RTUs), the voltage phasors at the buses are obtained
strong focus on online voltage stability monitoring [1–3]. from the traditional supervisory control and data acquisition
The maximum power loadability limit of the transmission net- (SCADA) based state estimator (SE), typically every few minutes.
work is one of the widely used indices to represent the voltage Real-time monitoring of the power system, therefore, is not possi-
security of a power system [4]. It is critical for the utilities to track ble with the help of conventional SCADA based measurements.
how close the transmission network is, to its maximum loading Because of high refresh rate and better accuracy, PMUs are increas-
limit, so that in case of emergency, proper control actions can be ingly being deployed in modern power systems. This paper pre-
taken. A large amount of literature exists on the use of analytical sents an algorithm for fast monitoring of the voltage stability of
methods for voltage stability monitoring [5–8]. The conventional the system, utilizing PMU measurements. A reduced network con-
P-V curves are extensively used by the utilities for determination taining the buses observed by the PMUs is used to develop the pro-
of the maximum permissible loading [9]. Continuation power flow posed methodologies.
(CPF) method is frequently used for obtaining the P-V curves [4–6]. Many works have been reported in the literature, exploring the
The loadability limit is determined by increasing the system load in capability of artificial neural network (ANN) for voltage stability
a particular direction, representing the most probable stressing monitoring [13–17]. ANN is used in these works, typically to estab-
lish a relationship between a voltage stability indicator and the
measurable power system parameters affecting the indicator.
⇑ Corresponding author. These methodologies require a large number of ANN inputs which
E-mail addresses: [email protected] (S.M. Ashraf), [email protected] significantly increases the ANN size and diminishes its accuracy.
(A. Gupta), [email protected] (D.K. Choudhary), [email protected] Refs. [18–20] discuss the assessment of voltage stability using
(S. Chakrabarti).

http://dx.doi.org/10.1016/j.ijepes.2016.11.008
0142-0615/Ó 2016 Elsevier Ltd. All rights reserved.
44 S.M. Ashraf et al. / Electrical Power and Energy Systems 87 (2017) 43–51

artificial neural network with reduced set of inputs. The focus in The paper is organized as follows. Section 2 illustrates the ward
these methodologies is on eliminating the redundant measure- reduction technique, Section 3 defines ANN scheme for the pro-
ments and thus reducing the number of variables needed to assess posed methods of voltage stability monitoring with load variations,
the system voltage stability. It is assumed in these methodologies Section 4 proposes a scheme to address bad or missing data, and
that a large number of measurement variables are available. The the problem of updation of ANN weights is discussed in Section 5.
present work considers a more realistic scenario, in which only a The analysis of the results is given in Section 6, and Section 7 con-
few nodes in the system are assumed to have PMUs installed cludes the paper. A description of voltage stability indicator is pre-
because of economic constraints. Only the data measured by these sented in Appendix A.
PMUs is utilized for stability analysis.
The computational burden, as well as, the communication
2. Ward reduction
requirement for the stability analysis of large interconnected
power systems can be reduced by using a reduced power system
One of the important aspects of the proposed methodology is
model. The reduced model also renders the advantage of: (i) mon-
the use of reduced network representation of the power system.
itoring the system by using only a limited number of measuring
This enables the monitoring of the voltage stability of the system
instruments; (ii) eliminating the need for detailed model in electri-
by observing fewer number of nodes and in reduced amount of
cally remote areas; and (iii) monitoring the interconnected system,
computational time. Ward reduction technique [21,25] is adopted
in which participating utilities are reluctant to share vital data.
in this paper for carrying out the network reduction function. The
Usually, a utility’s own system is called the internal area. The rest
construction of Ward’s equivalent starts from the solved model of
of the system is called the external area for the internal subsystem
the entire interconnected power system. The injected current iðiÞ at
under consideration. For running power system analysis functions
each bus i is determined from the bus’s known complex power
in the internal area, the internal subsystem is typically modeled in
injection sðiÞ and voltage v ðiÞ.
detail. The external areas are usually represented by simple mod-
els, referred to as the external equivalent system. Numerous tech- iðiÞ ¼ s ðiÞ=v  ðiÞ ð1Þ
niques for determining the reduced equivalent have been proposed
Gaussian elimination technique is used in this method to get
in the literature [21–23]. In this paper, Ward equivalent technique
the reduced network and current vectors. The nodal equations
is used and it is assumed that measurements of all the nodes in the
describing the power system are given by,
internal area are available via PMUs.
Two schemes for estimating the voltage instability in an inter- Ybus v ¼ i ð2Þ
connected power system are presented in this paper, using the
reduced equivalent of the entire network. The fundamental motive where Ybus is the n  n bus admittance matrix, v is the n  1 vector
of this research is fast and accurate assessment of the voltage sta- of complex voltages at all nodes, and i is the n  1 vector of complex
bility of an entire network based on the data measured by PMUs at currents injected at all nodes.
certain critical nodes. A feed-forward back-propagation network After elimination of the kth node, Ybus is modified as,
(FFBPN) is used to estimate the maximum loadability of the net- Y ik Y kj
work. In the first scheme, complex bus voltages measured by the Y 0ij ¼ Y ij  ; 8i; j ¼ 1; . . . ; n; i; j–k ð3Þ
Y kk
PMUs are used as the input to the FFBPN, and the available load-
ability margin is used as an indicator of the system voltage stabil- where Y 0ij are the elements of the new ðn  1Þ  ðn  1ÞY bus matrix.
ity. In the second scheme, before estimating the system voltage 0
The modified current vector, i , is given by,
stability margin, an FFBPN is first used to estimate the external
network bus voltages by using the internal network bus voltages. 0 Y ik
The proposed networks have the ability to get adaptive training,
i ðiÞ ¼ iðiÞ  iðkÞ; 8i ¼ 1; . . . ; n; i–k ð4Þ
Y kk
when subjected to any new training pattern, following a change
in the system operating condition. The proposed strategy is applied If the network is reduced to r nodes, a new bus admittance
to IEEE 14-bus and 118-bus systems. The complex bus voltages matrix of dimension r  r, and a new current injection vector of
measured by using PMUs may be subject to data packet loss or dimension r  1 is obtained. The modified current vector is con-
bad data. The bad data are recognized using Z-score algorithm verted back into power injections for analysis purpose [21,25–
[24]. The dropped data packets are compensated by using polyno- 27]. This reduced network carries full information of the original
mial curve fitting. A method to update the ANN weights to incorpo- power system at the base case.
rate a new training pattern is also presented.
The key contributions of this paper are the following. 3. Proposed ANN architecture

Utilizing the voltage phasors measured by PMUs located at The feed-forward ANNs employing back-propagation learning
internal buses of the system to determine the system voltage algorithm are used in this work as a mapping tool to estimate
stability. the available loadability margin of the system. The feed-forward

Although changes in the internal area parameters are reflected networks are capable of approximating any measurable function
by PMU measurements, changes in loading condition for exter- to the desired accuracy level [28]. The back-propagation steps
nal area are unaccounted. For more accurate estimation of volt- repeatedly adjust the weights of the connections in the network
age stability, an ANN and ward reduction based method is to minimize the mean squared error between the desired output
presented to account for changes in the loading in the external and the output of the ANN. In this work, the ANNs are trained by
area. using Levenberg-Marquardt algorithm [29].

To further strengthen the reliability in determination of system The maximum loadability limit is considered in this paper as a
voltage stability, a bad data detection and correction method measure of the voltage stability of the power system. If the total
based on Z-score algorithm is presented. system load exceeds this limit, it will result in voltage collapse

The power system may undergo changes after installation of and possibly blackout. The objective here is to determine the point
ANN. To accommodate for these changes in the system, an of maximum loadability so that corrective and/or preventive
adaptive training technique for ANN is also presented. actions can be taken to avoid any voltage instability problem. In
 S.M. Ashraf et al. / Electrical Power and Energy Systems 87 (2017) 43–51 45

order to reduce computational burden and fast assessment of the 3.2. Method 2
loadability limit, measurements from the retained buses in the
reduced network are used. Two separate methods using different Although the methodology discussed in the previous section is
ANN configurations are investigated in this paper. easy to implement, its performance may deteriorate in case of sig-
Both of these methods use the same training data set. The train- nificant change in the loading conditions in the buses not retained
ing data comprises of a set of measured voltage phasors at the in the reduced network. To overcome this limitation, in the second
buses that are directly or indirectly observable by PMUs installed method, i.e., in Method 2, the bus voltage phasors in the external
in the system, and the corresponding loading margins. A large system (part of the system outside the reduced network) are first
number of different operating conditions are generated by ran- estimated by using the available PMU measurements. An ANN,
domly altering the base case real and reactive power of the loads designated as ANN 2, shown in Fig. 2, is used for this purpose. It
in the entire system in the range of ±30%, and correspondingly estimates the voltage phasors of the external buses by using the
adjusting the generation. The output data set comprises of the voltage phasors measured in the reduced network (i.e., the internal
maximum loadability limits for these operating conditions, which area) as the input. This makes available the information regarding
are obtained by running CPF up to the nose point. the voltage phasors at all buses in the system. Some of these volt-
The objective here is to estimate the loadability margin based age phasors are directly observable by the PMUs; the remaining
on a limited number of PMU measurements. The original system are estimated by ANN 2.
may not be completely observable with PMU measurements. A Using these measured and estimated voltage phasors, known
subset of the buses is first specified, keeping only the buses that network parameters, and topology; the Ward reduction method
are directly or indirectly observable by PMUs. Both the methods discussed in Section 2 is used to obtain the power injections in
discussed below utilizes the voltage phasors measured at these the internal area. The main motivation for this step is to include
buses, for estimating the loadability margin of the entire system. the information regarding the power injections in the external
The main advantage of using PMU measurements is their high area, in the power injections in the internal area or the reduced
refresh rate and accuracy, compared to the conventional network. Due to the transformations of the power injections as a
measurements. result of network reduction, the information regarding the power
injections in the external area is now embedded in the power
injections in the internal area [25,31].
3.1. Method 1
An ANN, referred as ANN 3, is now trained with the real and
reactive power injections in the internal area as the input, obtained
In Method 1, the loadability margin of the entire system is esti-
after transformation resulting from the network reduction step
mated, based on the voltage phasor measurements from the buses
discussed above. For each pattern of the input data, the output is
observed by the PMUs. Since only a small number of PMU mea-
the corresponding maximum loadability margin, which is obtained
surements are used (the system is not required to be completely
by running CPF for the entire network. Fig. 3 describes the config-
observable by PMUs), and the PMU measurements are obtained
uration of ANN 3, which has a single hidden layer with the number
at sub-second rate (50 or 60 times per second, for 50 and 60 Hz
of neurons determined by (5).
systems, respectively), the loadability margin can be estimated at
a sub-second rate.
An ANN, hereafter called as ANN 1, is trained to estimate the
4. Bad data detection and correction
maximum loadability margin based on the voltage phasors mea-
sured by the PMUs in the system. There is a single hidden layer
Once trained properly, the ANNs described above can be imple-
in ANN 1. The number of neurons in the hidden layer is chosen
mented in actual power systems to estimate the system loadability
according to the following thumb rule [30].
margin. Measurements from PMUs are to be used for these loading

 margin estimators. PMU measurements, however, may be sub-
2
nh ¼ niv þ nov ð5Þ jected to the problems of bad data, loss of data packets, communi-
3
cation link failure, noise, etc. In case of such erroneous data, the
where nh is the number of neurons in the hidden layer, niv is the estimated loading margin may not be accurate. Pre-processing of
number of elements in the input vector, and nov is the number of the PMU data i.e., processing of the measurement data coming
elements in the output vector. from the PMUs before utilizing it as input data for ANNs, is there-
The inputs to ANN 1 are the complex voltage phasors measured fore required to ensure the reliability and robustness of the ANN-
by the PMUs, and the output is the system loadability limit. Fig. 1 based estimators.
gives a schematic representation of ANN 1. Since the inputs to the A commonly encountered type of bad data in field measure-
ANN are the complex voltages measured by the PMUs, which are ments is the outliers. An outlier is an observation that appears to
available at sub-second rate, the calculation of system loadability deviate markedly from other observations in the sample. A number
limit is also possible to complete at sub-second rate. of techniques are reported in the literature to spot the outliers in

Fig. 1. Block diagram of ANN 1. Fig. 2. Block diagram of ANN 2.

46 S.M. Ashraf et al. / Electrical Power and Energy Systems 87 (2017) 43–51

2 32 3 2 3
X
n X n a0 X n

6 n xðkÞ  x m
ðkÞ 76 7 6 yðkÞ 7
6 76 7 6 k¼1 7
6 k¼1 k¼1 76 7 6 7
6 76 7 6 7
6 X X X 7 6 7 6 7
6 n n n
76 7 6 X n
7
6 xðkÞ x ðkÞ   
2
x ðkÞ
mþ1 7 6 a 1 7 6 xðkÞyðkÞ 7
6 76 7 6 7
6 k¼1 k¼1 k¼1 76 7 ¼ 6 k¼1 7
6 76 7 6 7
6 76 . 7 6 7
6 .
.. .
.. .
. 7 6 . 7 6 .
. 7
6  . 7 6 . 7 6 . 7
6 76 7 6 7
6 76 7 6 7
6X n Xn Xn 76 7 6 X n 7
4 54 5 4 5
Fig. 3. Block diagram of ANN 3. x ðkÞ
m
x ðkÞ   
mþ1
x mþm
ðkÞ x ðkÞyðkÞ
m

k¼1 k¼1 k¼1 am k¼1

ð11Þ
the measured data [24,32–34]. A simple and fast way to identify Eq. (11) is solved to determine the coefficients of the curve that
the outliers is the Z-score [24]. The Z-score of an observation, Z i , best fits the given data. The polynomial, thus derived, is used to
is defined as, estimate the missing data points or to replace the outliers.

Yi  Y
Zi ¼ ; ð6Þ 5. Adaptive training of the ANNs
r
where Y i , Y, and r denotes the observation at the ith instant, the In practice, the weights obtained after completion of the ANN
sample mean, and the sample standard deviation, respectively. training are saved and used during the operational phases as con-
The mean and standard deviations are calculated from the stant values. However, the parameters of a power system do not
observed values over a window consisting of certain number of remain static. This implies that the training dataset may not be
past samples. The necessary condition for being an outlier is, adequate to handle the unexpected changes in the system param-
eters. The weights are required to be dynamic and changing in an
jZ i j > Z th ; ð7Þ
adaptive manner, based on the recent changes of the system.
where Z th is the threshold value for the Z-score. The variation in Y i To address the issue, an adaptive weight update strategy is pro-
should be within a specific percentage of the first sample in the posed in this paper. The initial training is carried out in conven-
observed window. The condition described (7) ensures that the Z- tional manner, and the ANN is deployed in the field. When the
score at the ith instant is Z th units of standard deviations away from prediction starts, the weights are updated based on the error in
its mean, i.e. the measurement at the ith instant is significantly dif- the previous performance. Once the complete data is available
ferent from the other observed values during the inspected period. for the system, the loadability limit is calculated using CPF and
Once the measurement is determined as a possible outlier, proper the error in estimation is computed. If the error is in excess of a
care should be taken so that it does not affect the case study. pre-specified threshold value, the parameters, and the error are
Loss of packets of PMU data can also be treated as outliers, and recorded. If the errors are found in excess for more than n predic-
the method described above can be used to detect such events. A tions (n depends on the system), the n recorded errors are used to
simple and efficient way to replace missing data is to use the poly- train the ANN again, and updated weights are obtained. This means
nomial curve fitting technique [35,36]. The idea behind polynomial that an ongoing training for the ANNs is implemented in the field
curve fitting is to fit a polynomial to a set of n data points by uti- to account for the changes in the operating conditions of the sys-
lizing least squares error approach. Consider a general form of tem. This is best explained with the following flowchart (Fig. 4):
polynomial of mth order,
X
m 6. Results
f ðxÞ ¼ a0 þ a1 x1 þ    þ am xm ¼ a0 þ aj x j ð8Þ
j¼1 The proposed methodologies are tested on standard IEEE 14-
bus and 118-bus systems [37]. The test systems are partitioned
The precise form of function f ðxÞ is determined based on the
into internal and external areas. The partitioning of the system is
values of the coefficients, a0 ; . . . ; am . The curve which gives mini-
based on a sensitivity analysis which identifies the most significant
mum error between data, y, and the fit, f ðxÞ, is the most appropri-
measurements that have to be retained in the reduced network
ate. The residual, R, is obtained by,
[38]. The partitioning of IEEE 14-bus system is demonstrated in
X
n
2 Fig. 5 [38]. Tables 1 and 2 present the internal and external buses
R2 ¼ ½yðkÞ  fa0 þ a1 xðkÞ þ a2 x2 ðkÞ þ    þ am xm ðkÞg ð9Þ
considered in IEEE 14-bus and 118-bus systems, respectively. All
k¼1
the buses in the internal area are directly or indirectly observable
where y(k) is the desired output at the kth instant, x(k) is the input by placing the PMUs at critical nodes. The nodes in the internal
at kth instant, and m is the order of the polynomial. area which have PMUs installed are mentioned in Table 3. The
The partial derivatives obtained from (9) are, time-series data for the studied cases were generated using PSAT
X X software [39]. Simulation is performed using a variable time step
@R2 n m
¼ 2 ½yðkÞ  fa0 þ aj x j ðkÞg ¼ 0 integration option and later on data is interpolated at every 0.1 s.
@a0 k¼1 j¼1 To simulate the phasor measurement noise, a random noise with
@R2 Xn Xm standard deviation r = 0.01 p.u. is also added onto the obtained
¼ 2 ½yðkÞ  fa0 þ aj x j ðkÞgxðkÞ ¼ 0 ð10Þ time series data.
@a1 k¼1 j¼1 Utilizing the real and reactive loads and the real power genera-
@R2 Xn Xm
tion of the base case, 240 random sets of loads and corresponding
¼ 2 ½yðkÞ  fa0 þ aj x j ðkÞgxm ðkÞ ¼ 0
@am k¼1 j¼1
generations are created with in a limit of ±30% of the base case
loading. The CPF analysis is carried on all the 240 sets to determine
Eq. (10) can be re-written and a Vandermonde matrix is a set of maximum loadability points for the system. The test sys-
obtained as shown below [35,36], tems are then divided into internal and external areas.
 S.M. Ashraf et al. / Electrical Power and Energy Systems 87 (2017) 43–51 47

Table 1
Distribution of the IEEE 14-bus system.

Internal system buses External system buses

1, 2, 3, 4, 5 6, 7, 8, 9, 10, 11, 12, 13, 14

Table 2
Distribution of the IEEE 118-bus system.

Internal system buses External system buses

24, 45–77, 116, 118 1–23, 25–44, 78–115, 117

Table 3
Nodes in the internal subsystem having PMUs installed.

Test system Number of Location of PMUs

PMUs
IEEE 14-Bus System 1 2
IEEE 118-Bus 10 47, 49, 52, 56, 62, 64, 68, 70, 71, 76
System

6.1. Proposed ANN architecture

The Neural Network Toolbox Version 8.2.1 provided with

MATLAB 8.4 [40] was used to develop the neural network models.
The data obtained from the CPF analysis was called in a MATLAB
script for training, testing, and validation purposes. In the first sce-
nario, the set of complex voltages measured by PMUs serve as
input for ANN 1, and the set of maximum loadability points serve
as the targets. 70% of the data set is used for training, 15% for val-
idation, and 15% for testing. The number of neurons in all the ANNs
is determined by first applying (5) and then through rigorous test-
ing using trial and error. ANN 1 comprises of a single hidden layer.
For IEEE 14-bus system, the number of neurons in input, hidden,
Fig. 4. Flow chart for adaptive update of ANN weights.
and output layers are 10, 8, and 1, respectively. For IEEE 118-bus
system, the number of neurons in input, hidden, and output layer
is 72, 38, and 1, respectively.
In the second scenario, the set of real and imaginary parts of the
internal area complex voltage are fed as input for ANN 2 and the
set of real and imaginary parts of external area complex voltage
are its targets. Real and reactive power injections corresponding
to the reduced system are fed as input to ANN 3, and the set of
maximum loadability points are the targets. In both ANN 2 and
ANN 3, 70% of the data is used for training, 15% for validation
and 15% for testing. ANN 2 and ANN 3 also comprise of a single hid-
den layer. For IEEE 14-bus system, the number of neurons in input,
hidden, and output layers are 10, 22, and 18, respectively. For IEEE
118-bus system, the number of neurons in input, hidden, and out-
put layer is 72, 87, and 164, respectively. ANN 3 has a structure
similar to that of ANN 1.
For testing of these ANNs, the dataset of 240 samples is used as
input. Figs. 6 and 7 show the performance of ANN 1 for IEEE 14-bus
and 118-bus, respectively. Figs. 8 and 9 show the performance of
ANN 3 for IEEE 14-bus and 118-bus, respectively. The target is
the actual set of maximum loadability point and output is the set
of maximum loadability point determined form ANNs.
In the first scenario, for IEEE 14-bus system, mean squared error
is determined as 3.370exp-04 and maximum error is 7.770exp-02.
For IEEE 118-bus system, mean squared error is determined as
1.249exp-03 and maximum error is 1.757exp-01, as can be seen
in Table 4. In the second scenario, for IEEE 14-bus system, mean
squared error is determined as 3.265exp-04, and maximum error
Fig. 5. Internal and external areas of IEEE 14-bus system. is 8.926exp-02. For IEEE 118-bus system, mean square error is
48 S.M. Ashraf et al. / Electrical Power and Energy Systems 87 (2017) 43–51

4 Table 4
maximum loadability
Percentage error in
Results for Method 1.

3 Test system Mean squared error Maximum error

IEEE 14-bus 3.370exp-04 7.770exp-02
2 IEEE 118-bus 1.249exp-03 1.757exp-01

1
Table 5
0 Results for Method 2.
0 50 100 150 200 250
Test system Mean squared error Maximum error
Test case
IEEE 14-bus 3.265exp-04 8.926exp-02
Fig. 6. Percentage error in maximum loadability for IEEE 14-bus system using ANN IEEE 118-bus 8.015exp-04 3.059exp-01
1.

better. However, for IEEE 118-bus system, scenario 2 outperforms

scenario 1.
The reason for better performance of scenario 2 for bigger sys-
tem is that, with the system getting bigger, and lesser number of
monitored or retained nodes, it is better to incorporate the loading
details of external system via network reduction process. The max-
imum loadability point for smaller systems with moderate number
of retained or monitored nodes can found out by both the scenarios
with similar accuracies.

6.2. Bad data detection and correction

The Z-score algorithm discussed earlier is used for detection of

Fig. 7. Percentage error in maximum loadability for IEEE 118-bus system using bad data. For testing of the algorithm, random error (<35%) is intro-
ANN 1.
duced into the complex voltage measurement coming from PMUs,
for preselected buses, at a particular instant. The normal variation
in voltage at the buses is considered to be ±5%. The output
sequence, thus obtained is tested for outliers by Z-score algorithm.
For IEEE 14-bus system test case, random errors of 32.7%,
31.56%, 30.36%, and 30.91% were introduced in the voltage magni-
tude and angle measurements coming from bus number 4 and 5,
respectively; at the 13th instant of a 30 sample observation win-
dow. In case of IEEE 118-bus test system, random errors of
34.16%, 33.09%, 32.6%, 34.32%, 30.49%, and 34.54% were introduced
in the voltage magnitude and angle measurements coming from
bus number 45, 50, and 51, respectively; at the 23rd instant of
the observation window. The plots of the Z-score test are presented
in Figs. 10–13. It is observed that the Z-score of the 13th and 23rd
Fig. 8. Percentage error in maximum loadability for IEEE 14-bus system using ANN instant of IEEE 14-bus and IEEE 118-bus, respectively, are very high
3. in comparison to the Z-score of previous instances, indicating an
outlier. Based on the observation, the threshold value of the Z-
score for bad data detection was selected to be 2.5. Both angle
and magnitude of the power systems can be used for outlier
detection.
Once bad data is detected, it is corrected using curve fitting
technique. In this work, a second order curve fit is applied to data

Fig. 9. Percentage error in maximum loadability for IEEE 118-bus system using
ANN 3.

determined as 8.015exp-04, and maximum error is 3.059exp-01, as

can be seen in Table 5. It can be seen that both the mean squared
error and the maximum error for IEEE 14-bus system are compara- Fig. 10. Plot of Z-score obtained from bus voltage angle against PMU sampling
ble in both the scenarios, with scenario 2 coming out to be slightly instant for IEEE 14-bus system.
 S.M. Ashraf et al. / Electrical Power and Energy Systems 87 (2017) 43–51 49

collected in the observation window. A tabulated output of the

curve fitting technique is presented in Tables 6 and 7.

6.3. Adaptive update of ANN weights

To test the adaptive update of ANN weights, some loads are

introduced in the external system. The ANNs trained earlier for
testing of proposed architecture are then retrained using the new
error. In IEEE 14-bus system, an active load of 20 MW is applied
at bus 7, whereas, in IEEE 118-bus test system, an active load of
70 MW is applied at bus 9, and training data set are obtained as
in the previous case. A comparison of ANN output error without
weight update and with weight update is presented in Tables 8
Fig. 11. Plot of Z-score obtained from bus voltage magnitude against PMU sampling
instant for IEEE 14-bus system. and 9.
It is observed that the ANN architecture with adaptive weight
gives better results in comparison to the ANN architecture without
an adaptive weight scheme. The update of ANN weights is not
reflected well in performance of Method 2, as the update of
weights are done for ANN 2 only, and the input of ANN 3 depends
on output of ANN 2. Since ANN 2 will be facing a new scenario, its
output will have some error. However, an improvement is seen in
both the discussed methods.

7. Conclusion

Because of growing interconnections and geographical spread,

today’s power systems are becoming increasingly larger and more
complex. Online voltage stability assessment of such multi-area
power systems in near real-time is becoming a challenging task.
Fig. 12. Plot of Z-score obtained from bus voltage angle against PMU sampling In this paper, a methodology is proposed to monitor a large power
instant for IEEE 118-bus system. system with limited number of measurements. The loadability
limit is considered as the stability margin. Assuming the power
system to be consisting of multiple areas, the security of the entire
system is determined through measurements from one of the
areas, referred as the internal area. Adaptively trained ANNs are
used to estimate the loading margin. Information regarding the
measurements from the external areas is incorporated into the
internal measurements through the process of network reduction.
With this additional embedded information regarding the operat-
ing conditions of the external areas, the estimation accuracy of
the ANNs is improved. Lower computational burden is still main-
tained due to the use of only limited number of measurements
from the internal area, while assessing the voltage stability of the
entire power system. A Z-score based methodology is also imple-
mented to detect and process any bad data in the training data
Fig. 13. Plot of Z-score obtained from bus voltage magnitude against PMU sampling set for the ANNs. The proposed methodologies are successfully
instant for IEEE 118-bus system. tested on IEEE 14-bus and 118-bus test systems.

Table 6
Bad data correction for IEEE 14-bus.

Bus Actual voltage Observed voltage Corrected voltage Actual voltage Observed voltage Corrected voltage
no. magnitude magnitude magnitude angle angle angle
4 1.0149 1.3230 1.0449 9.8474 13.0675 9.5793
5 1.0489 1.3731 1.0429 7.8112 10.2761 8.1841

Table 7
Bad data correction for IEEE 118-bus.

Bus Actual voltage Observed voltage Corrected voltage Actual voltage Observed voltage Corrected voltage
no. magnitude magnitude magnitude angle angle angle
45 1.0313 1.3852 1.0093 13.8800 18.6216 14.6555
50 1.0123 1.3209 1.0124 16.7470 22.2880 17.1741
51 0.9848 1.3250 0.9840 13.6442 18.0923 13.9143
50 S.M. Ashraf et al. / Electrical Power and Energy Systems 87 (2017) 43–51

Table 8
Adaptive update of ANN weights for Method 1.

Test system Without weight update With weight update

Mean squared error Maximum error Mean squared error Maximum error
IEEE 14-bus 5.427exp-02 3.598exp-01 1.613exp-03 5.500exp-02
IEEE 118-bus 3.646exp-02 4.916exp-01 8.392exp-03 2.540exp-01

Table 9
Adaptive update of ANN weights for Method 2.

Test system Without weight update With weight update

Mean squared error Maximum error Mean squared error Maximum error
IEEE 14-bus 2.237exp-01 1.507exp-01 2.989exp-03 7.354exp-02
IEEE 118-bus 8.488exp-02 1.852exp-01 1.286exp-02 2.368exp-01

Appendix A. Voltage stability indicator where PLi, QLi, PGi represent the active load, the reactive load, and
the active generation at bus i, respectively; PLi0, QLi0, PGi0 represent
The loadability limit is used as the voltage instability indicator the nominal values of the active load, the reactive load, and the
in this paper. It indicates, how much the system can be stressed, active generation at bus i, respectively; wi is the power factor angle
from the current operating condition, before it becomes unstable. (assumed to be constant) of the load at bus i; KLi, KGi are constants
Continuation power flow (CPF) is performed to determine the load- specifying the rate of change in load and generation at bus i, respec-
ability limit [4]. A brief discussion on the CPF is included in the fol- tively, as k changes.
lowing to make the paper self-contained. Initially, the parameter k is selected as the continuation param-
The power flow equations can be expressed as, eter. In the vicinity of the nose point, parameterization with
respect to k may not give any solution. In that case, one of the com-
fðx; kÞ ¼ 0 ðA:1Þ
ponents of x may be selected as the continuation parameter. The
n
where f represents the full set of power flow equations; x 2 R is the parameter yk , corresponding to which the element in the tangent
vector of bus voltage magnitudes and angles; and k 2 [0, kcritical] is vector has the maximum absolute value, is selected as the contin-
the varying loading parameter (its maximum value being kcritical) uation parameter [42].
[41]. It is assumed here that all the loads in the system change with
jdyk j ¼ maxfjdy1 j; jdy2 j; . . . ; jdynþ1 jg ðA:8Þ
the common factor.
By varying the parameter k continuously, a path of solutions can where dy 2 R n+1
is the tangent vector along the direction of the pre-
be obtained by solving (A.1), and the nose curve depicting the tra- dictor, and d yk is its kth component.
jectory of solutions can be plotted. The solution corresponding to
kcritical is called the critical point. The nose curve is obtained by A.2. Predictor
employing a predictor-corrector scheme.
The predictor is used to obtain an approximate solution, start-
A.1. Parameterization and selection of continuation parameter ing from the current operating point [42].
Differentiating (A.4), with the state variables corresponding to
In the method of parameterization, the loading parameter is the initial solution, will result in the following equation:
included as an additional variable, and the original set of equations " #
@f @f
  
is augmented by an additional equation specifying the value of the @x @k
dx 0
¼ ðA:9Þ
loading parameter [41–43]. ek dk 1
Let the augmented set of variables be defined as,
T
where dy ¼ ½dxT dk is the tangent to be calculated; ek is an appro-
y :¼ ðx; kÞ 2 Rnþ1 ðA:2Þ
priately dimensioned row vector, with all elements equal to zero
If the kth element of yk is chosen as the continuation parameter, except the kth, which is equal to one, and corresponds to dyk . The
the additional equation can be written as, sign of the state variable that has the greatest rate of change near
yk  g ¼ 0 ðA:3Þ the given solution determines the sign of ek . Proper selection of ek
ensures non-singularity of the Jacobian in (A.9).
where g is an appropriate value for the kth element of y. The prediction for the starting point of the next solution,
The augmented equation combining (A.1), (A.2) and (A.3) is, h iT
  b
y¼ b xT b k , is given by,
fðyÞ
¼ ½0 ðA:4Þ      
yk  g ^
x x dx
¼ þr ðA:10Þ
^
k k dk
To apply the CPF, the power flow equations are reformulated to
introduce a load parameter, k. The load and generation profiles  T
where y ¼ xT k is a solution of the previous iteration or initial
with variation are formulated as [9],
operating point for the current iteration, and r is the step size.
PLi ¼ PLi0 þ kK Li ðA:5Þ
A.3. Corrector
Q Li ¼ Q Li0 þ k tanðwi ÞK Li ðA:6Þ
The corrector is used to obtain a solution point from the starting
PGi ¼ P Gi0 þ kK Gi ðA:7Þ b , by solving (A.4), using slightly modified Newton-Raphson
point, y
 S.M. Ashraf et al. / Electrical Power and Energy Systems 87 (2017) 43–51 51

power flow method. The introduction of the additional equation Proc - Gener Transm Distrib 1998;145:355. http://dx.doi.org/10.1049/ip-
gtd:19981977.
specifying yk makes the Jacobian non-singular at the nose point,
[21] Ward JB. Equivalent circuits for power-flow studies. Trans Am Inst Electr Eng
so that CPF analysis continues beyond the nose point. 1949;68:373–82. http://dx.doi.org/10.1109/T-AIEE.1949.5059947.
[22] Irisarri G, Sasson AM, Dopazo JF. Real-time external system equivalent for on-
References line contingency analysis. IEEE Trans Power Appar Syst 1979;PAS-98. http://
dx.doi.org/10.1109/TPAS.1979.31941.
[23] Savulescu SC. Real-time stability assessment in modern power system control
[1] Suzuki M, Wada S, Sato M, Asano T, Kudo Y. Newly developed voltage security centers. John Wiley & Sons; 2009.
monitoring system. IEEE Trans Power Syst 1992;7:965–73. http://dx.doi.org/ [24] Shiffler RE. Maximum Z scores and outliers. Am Stat 1988;42(1):79–80. http://
10.1109/59.207309. dx.doi.org/10.1080/00031305.1988.10475530.
[2] Taylor CW. The future in on-line security assessment and wide-area stability [25] Vardikar M, Chakrabarti S, Kyriakides E. Transformation of measurements for
control. In: 2000 IEEE power eng soc winter meet conf proc (Cat using external network equivalents in state estimation. In: 2013 IEEE power
No00CH37077). p. 1. http://dx.doi.org/10.1109/PESW.2000.849931. energy soc gen meet. IEEE; 2013. p. 1–5. http://dx.doi.org/10.1109/
[3] Zima M, Larsson M, Korba P. Design aspects for wide-area monitoring and PESMG.2013.6672546.
control systems. Proc IEEE 2005;93:980–96. http://dx.doi.org/10.1109/ [26] Grainger. Power system analysis. McGraw-Hill Education (India) Pvt Limited;
JPROC.2005.846336. 2003.
[4] van Cutsem T, Vournas C. Voltage stability of electric power systems. Springer [27] Bergen AR. Power systems analysis. Pearson Education; 2009.
Science & Business Media; 1998. [28] Hornik K, Stinchcombe M, White H. Multilayer feedforward networks are
[5] Ajjarapu V, Christy C. The continuation power flow: a tool for steady state universal approximators. Neural Networks 1989;2:359–66. http://dx.doi.org/
voltage stability analysis. IEEE Trans Power Syst 1992;7:416–23. http://dx.doi. 10.1016/0893-6080(89)90020-8.
org/10.1109/59.141737. [29] Hagan MT, Menhaj MB. Training feedforward networks with the Marquardt
[6] Canizares CA, Alvarado FL. Point of collapse and continuation methods for large algorithm. IEEE Trans Neural Netw 1994;5:989–93. http://dx.doi.org/10.1109/
ac/dc systems. IEEE Trans Power Syst 1993;8:1–8. http://dx.doi.org/10.1109/ 72.329697.
59.221241. [30] Heaton J. Introduction to neural networks with java. Heaton Research Inc.;
[7] Dobson I, Lu L. New methods for computing a closest saddle node bifurcation 2008.
and worst case load power margin for voltage. IEEE Trans Power Syst [31] Ashraf SM, Chakrabarti S. Voltage stability monitoring using reduced network
1993;8:905–13. http://dx.doi.org/10.1109/59.260912. and measurement transformation. In: 2015 Australas univ power eng
[8] Alvarado F, Dobson I, Hu Y. Computation of closest bifurcations in power conf. IEEE; 2015. p. 1–5. http://dx.doi.org/10.1109/AUPEC.2015.7324840.
systems. IEEE Trans Power Syst 1994;9:918–28. http://dx.doi.org/10.1109/ [32] Yin S, Wang G, Yang X. Robust PLS approach for KPI-related prediction and
59.317655. diagnosis against outliers and missing data. Int J Syst Sci 2014;45:1375–82.
[9] Ajjarapu V. Computational techniques for voltage stability assessment and http://dx.doi.org/10.1080/00207721.2014.886136.
control (power electronics and power systems); 2006. [33] Yuan K-H, Marshall LL, Bentler PM. A unified approach to exploratory factor
[10] Klump R, Wilson RE, Martin KE. Visualizing real-time security threats using analysis with missing data, nonnormal data, and in the presence of outliers.
hybrid SCADA/PMU measurement displays. In: Proc 38th annu Hawaii int conf Psychometrika 2002;67:95–121. http://dx.doi.org/10.1007/BF02294711.
syst sci. IEEE; 2005. http://dx.doi.org/10.1109/HICSS.2005.679. p. 55c. [34] Tabachnick BG, Fidell LS. Using multivariate statistics. Pearson Education;
[11] Parashar M, Jianzhong Mo. Real Time Dynamics Monitoring System (RTDMS): 2012.
phasor applications for the control room. In: 2009 42nd Hawaii int conf syst [35] Bishop CM. Neural networks for pattern recognition, vol. 23. Clarendon Press;
sci. IEEE; 2009. p. 1–11. http://dx.doi.org/10.1109/HICSS.2009.362. 1995.
[12] Moslehi K, Kumar R. Smart grid - a reliability perspective. In: 2010 innov smart [36] Weisstein EW. Least squares fitting—polynomial,
grid technol. IEEE; 2010. p. 1–8. http://dx.doi.org/10.1109/ISGT.2010.5434765. <http://mathworld.wolfram.com/LeastSquaresFittingPolynomial.html>; 2015
[13] Dinavahi VR, Srivastava SC. ANN based voltage stability margin prediction. In: [accessed August 3, 2015].
2001 Power eng soc summer meet conf proc (Cat No01CH37262). p. 2. http:// [37] Power Systems Test Case Archive - UWEE n.d. <https://www.ee.washington.
dx.doi.org/10.1109/PESS.2001.970256. edu/research/pstca/> [accessed September 2, 2015].
[14] Chakrabarti S, Jeyasurya B. Generation rescheduling using ANN-based [38] Dongare K, Chakrabarti S, Kyriakides E. Power system state estimation
computation of parameter sensitivities of the voltage stability margin. Eng considering real-time equivalents of the external networks. In: 2013 IEEE
Appl Artif Intell 2008;21:1164–9. http://dx.doi.org/10.1016/j. innov smart grid technol (ISGT Asia). IEEE; 2013. p. 1–6. http://dx.doi.org/
engappai.2008.05.002. 10.1109/ISGT-Asia.2013.6698714.
[15] Bulac C, Tristiu I, Mandis A, Toma L. On-line power systems voltage stability [39] Milano F. Open source power system analysis toolbox. IEEE Trans Power Syst
monitoring using artificial neural networks. In: 2015 9th int symp adv top 2005;20:1199–206. http://dx.doi.org/10.1109/TPWRS.2005.851911.
electr eng. IEEE; 2015. p. 622–5. http://dx.doi.org/10.1109/ [40] MATLAB. Neural network toolbox release 2014b. Natick, Massachusetts,
ATEE.2015.7133884. United States: The MathWorks, Inc; n.d.
[16] Zhou DQ, Annakkage UD, Rajapakse AD. Online monitoring of voltage stability [41] Canizares CA, Mithulananthan N, Berizzi A, Reeve J. On the linear profile of
margin using an artificial neural network. IEEE Trans Power Syst indices for the prediction of saddle-node and limit-induced bifurcation points
2010;25:1566–74. http://dx.doi.org/10.1109/TPWRS.2009.2038059. in power systems. IEEE Trans Circ Syst I Fund Theory Appl 2003;50:1588–95.
[17] El-Keib AA, Ma X. Application of artificial neural networks in voltage stability http://dx.doi.org/10.1109/TCSI.2003.819807.
assessment. Power Syst IEEE Trans 1995;10:1890–6. [42] Xu P, Wang X, Ajjarapu V. Continuation power flow with adaptive stepsize
[18] Nakawiro W, Erlich I. Online voltage stability monitoring using artificial neural control via convergence monitor. IET Gener Transm Distrib 2012;6:673. http://
network. In: 2008 third int conf electr util deregul restruct power technol. dx.doi.org/10.1049/iet-gtd.2011.0573.
http://dx.doi.org/10.1109/DRPT.2008.4523542. [43] Mohn FW, Zambroni De Souza AC. Tracing PV and QV curves with the help of a
[19] Bahmanyar AR, Karami A. Power system voltage stability monitoring using CRIC continuation method. IEEE Trans Power Syst 2006;21:1115–22. http://dx.
artificial neural networks with a reduced set of inputs. Int J Electr Power doi.org/10.1109/TPWRS.2006.876693.
Energy Syst 2014;58:246–56. http://dx.doi.org/10.1016/j.ijepes.2014.01.019.
[20] Popović D, Kukolj D, Kulić F. Monitoring and assessment of voltage stability
margins using artificial neural networks with a reduced input set. IEE