See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/286697732

Fast voltage and power flow contingencies

ranking using enhanced radial basis function
neural network

Article in Iranian Journal of Electrical and Electronic Engineering January 2011

CITATIONS READS

6 18

3 authors, including:

Habib Rajabi Mashhadi Modjtaba Rouhani

Ferdowsi University Of Mashhad Ferdowsi University Of Mashhad
105 PUBLICATIONS 828 CITATIONS 39 PUBLICATIONS 275 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Vulnerabilty Assessment of Khorasan's Power System View project

electricity market decomposition View project

All content following this page was uploaded by Modjtaba Rouhani on 03 February 2016.

The user has requested enhancement of the downloaded file.

Fast Voltage and Power Flow Contingencies Ranking using
Enhanced Radial Basis Function Neural Network

D. S. Javan*, H. Rajabi Mashhadi* and M. Rouhani**

Abstract: Deregulation of power system in recent years has changed static security
assessment to the major concerns for which fast and accurate evaluation methodology is
needed. Contingencies related to voltage violations and power line overloading have been
responsible for power system collapse. This paper presents an enhanced radial basis
function neural network (RBFNN) approach for on-line ranking of the contingencies
expected to cause steady state bus voltage and power flow violations. Hidden layer units
(neurons) have been selected with the growing and pruning algorithm which has the
superiority of being able to choose optimal units center and width (radius). A feature
preference technique-based class separability index and correlation coefficient has been
employed to identify the relevant inputs for the neural network. The advantages of this
method are simplicity of algorithm and high accuracy in classification. The effectiveness of
the proposed approach has been demonstrated on IEEE 14-bus power system.

Keywords: Static Security Assessment, Neural Network, Feature Selection, Contingency,

Performance Index.

1 Introduction1 accurate security monitoring method has become a key

Power system security and contingency evaluation is
one of the most important tasks encountered by
planning and operation engineers of bulk power
systems. Security assessment has been historically
conducted in an offline operation-planning environment
in which the steady-state and dynamic performance of
the near-term forecasted system conditions are
exhaustively determined using tools such as full AC
power flow and time domain simulations. These
deterministic computation methods exhaustively
examine many contingencies and search for security
limits using a rigorous approach. Although this
approach provides the most accurate results available, it
is computationally burdensome and, therefore, time
consuming. The present trend toward deregulation has
forced modern electric utilities to operate the systems
under stressed operating conditions closer to their
security limits. Under such fragile conditions, any
disturbance could endanger system security and may
lead to system collapse [1-2]. Therefore, fast and
Iranian Journal of Electrical & Electronic Engineering, 2011.
Paper first received 4 Apr. 2011 and in revised form 25 Jun. 2011.
*The Authors are with the Department of Electrical Engineering,
Ferdowsi University of Mashhad, Mashhad, Iran.
E-mails: dawoodjavan@ieee.org , h_mashhadi@um.ac.ir
**The Author is with Islamic Azad University, Gonabad Branch,
Gonabad, Iran.
E-mail: m.rouhani@ieee.org.
issue to ensure secure operation of the system.
Variety of algorithms were developed for
contingency analysis, the most popular being the
Performance Index (PI) based method [3-4].
Performance index quantifies the severity of
contingencies by calculating their PI values and to rank
them accordingly. To achieve accurate ranking, each PI
value would need to be calculated from the results of a
full AC load flow [5-7]. This process is time consuming
and unsuitable for on-line use.
To overcome these shortcomings, these applications
can be linked to an Intelligent System (IS) [8-13] that
can utilize the accumulated knowledge from previous
calculations (stored in a database), utilize rules set by
experts, and make inferences based on system
conditions. In this way, security (and security limits)
can be found rapidly without extensive simulations. In
recent years, there is growing trend in applying these
methods for the operation and control of a power system
[14-17].
The artificial neural networks (ANNs) are the most
popular method among machine learning methods and
have been proposed for the static and dynamic security
assessment [18-19], which usually encounters to local
minimum and over-fitting problems. Application of
Support Vector Machines (SVM) for security
assessment has been reported in [20-21]. In these
references, the superior performance of SVM over ANN

Iranian Journal of Electrical & Electronic Engineering, Vol. 7, No. 4, Dec. 2011 273
in terms of accuracy, speed and distribution of high-risk
cases has been presented for security evaluation of a
large-scale power system. However, SVM learning
algorithms suffer from exceeding time and memory
requirements, if the training pattern set is very large.
SVM based binary classification for static and transient
security evaluation is proposed in [22] and levels of
power system security are just classified in two classes
of secure and insecure. Also, literatures have reported
the use of decision trees [23] for design of classifier.
In this paper, an enhanced Radial Basis Function
Neural Network based on Growing and Pruning
algorithm (GPRBFNN), which automatically selects the
optimal units (neurons) and distributions, has been
presented. The proposed method ranks the
contingencies expected to cause steady state line power
and bus voltage violations. Results of the contingency
rankings by the proposed method are compared with
those obtained by the classical performance index
method (Newton Raphson). Different status for system
security levels, such as secure, correctively secure, alert,
correctable emergency and non-correctable emergency
have been considered and the proposed neural network
has been used to classify the power system security.
Moreover, an approach based on the class separability
index and correlation coefficient [24] is used to select
the relevant features for the GPRBFNN. The
performance of the presented method has been
demonstrated on IEEE 14-bus system and found to be
fitting for online ranking of contingencies.
2 Methodology
The comprehensive block diagram of the proposed
method is presented in Fig. 1. A large number of load
patterns are generated haphazardly in a wide range of
load variation at each bus and for different contingency
cases. The input features are preferred using the feature
selection method to reduce the dimensionality of the
input and the size of the neural network (block I). The
selected input is normalized (block II). Proposed
algorithm has been utilized to determine the number of
nodes in the hidden layer and cluster position (block
III). The input data are straightly fed to these hidden
units and the weights between the hidden layer and the
output layer are adjusted using supervised learning
(block IV) so that the outputs of the RBFNN provide
accurate values of the voltage and power flow ranking
in each of the selected critical contingencies.
RBFNN is used for solving problems such as pattern
classification and nonlinear functional approximation
[25]. The form of the radial basis function, O j (x) , is
strictly positive and symmetric with a unique maximum
at the center. The most commonly used form is the
Gaussian function according to:
|| x C || 2
Q j ( x) exp (1)
j
2 j
274 Iranian Journal of Electrical & Electronic Engineering, Vol. 7, No. 4, Dec. 2011
Voltage and line flow ranking
..
Supervised learning of
IV RBFNN
Calculation center and width
III
Total ....
patterns
II Training set data normalized
P1 P2 Pm V1 V2 Vj Q1 Q2 Qm
I Input feature selection
Large number
of inputs P P2 Pn V1 V2 Vk Q1 Q2 Qn
1
Fig. 1 Conceptual diagram of a supervised learning model for
voltage and power line contingency ranking.
T
where, Cj cj1 ,cj2 ,,c jm is the field centers matrix
and j is the radius in the function.
3 Proposed Algorithm of Enhanced Radial Basis
Function Neural Network
In the traditional RBFNN, if the number of neurons
in hidden layer is too small, the generalized output
vectors may be in low accuracy. Conversely, too large a
number may cause over-fitting of the input data, as well
as upsetting the global generalization performance. A
problem significantly found in RBFNN design,
however, is selecting the appropriate number and
positions of the radial basis functions in the hidden layer
space. If the number of RBFN neurons is not selected
properly, the network may give out poor global
generalization capability, slow training speed, and large
memory space request. Still, a second problem, from a
classification point of view, is the boundary patterns
where clusters of each class contain data from other
classes. In the proposed algorithm, the boundary region
separates various classes and the patterns of each class
lie in their corresponding clusters.
The incremental training algorithm is outlined as
follows:
a) Suppose the classes A and B follow m and n
patterns respectively.
b) Calculate the distance between m patterns in
class A and n patterns in class B:
a 1 b1 a1 b n
Pi (2)
a m b1 a m b n
c) Find the minimum distance from pattern ai to
patterns of class B in every row of Pi :
 Dis(aa 1 , B) min( a 1 b1 , , a 1 b n preesented algoriithm, Steps (bb) to (d) seek the optimum m
cennter and radius:
(3)
a i A & b j B R a Max i M
Min i norm 2 (aa i , b j1,2,...,n ) (7)
Dis(aa m , B) min( a m b1 , , a m b n
i

wh
here, R a and a i are the seelected radii and
i
a center sett
d) Determmine the maxximum distancce from (4); thhen, forr class A. Stepps (e) to (g) coontrast the paattern selectedd
considder R a j and a j as the radiu
us and centerr for as the centre with
w other paatterns of claass A and iff
the firrst neuron andd cluster of claass A: disstances are smmaller than thhe selected rad dius, they aree
R a j maax(Dis(a1 , B), , Dis(a m , B))) plaaced in one clluster and thee algorithm iss repeated forr
(4)
thee remainder of patterns in class A whicch are outsidee
e) Calcullate the distannce between patterns of cllass thee cluster [Figg. 2(a) and FFig. 2(c)]. Therefore,
T thee
A withh each other: prooposed algoritthm is repeateed until all pattterns of classs
a 1 a1 a1 a m A sit
s in their opttimum clusterrs [Fig. 2(b) and
a Fig. 2(d)]..
Finnally, by perfforming the pproposed alg gorithm, threee
Qi (5)
opttimum clusters can be foound for each h of the twoo
a m a1 a m a m
claasses of Fig. 2.
f) Sort thhe distance froom other patteerns in class A
A: The construction of GPR RBFNN, in itts most basicc
forrm, involves three
t layers wwith entirely different
d roless
a j a 1 , a j a 2 , , a j a m (6)
as shown in Fig g. 3. The firstt layer is an input
i layer off
g) Remove the patteerns of class A for smaaller whhich each nodee correspondss to an attribu ute of an inputt
distannces than R a j and
a label it ass the first clusster. patttern. The second layer iss a hidden layer and thee
trannsformation from
f the inpuut layer to thee hidden layerr
For thhe remainderr of the patterns of classs A is nonlinear, whereas
w it iis constructed using thee
whichh have larger distances than R a j , contiinue prooposed algoritthm. The thirrd layer givess the networkk
the prrocess until all patterns in class A are outtput vector wh hich is a lineaar combination of the basiss
eliminnated. funnction outputs. Thus, for the GPRBFN NN having l--
h) Repeaat the algorithm
m for the n paatterns of classs B. neu urons, the reelationship bbetween an n-dimensional
n l
Y y1 , y 22 ,, yn
T
Fig. 2 graphically illustrates
i the procedure of outtput vector and
a an m--
clustering sttages for ann artificial data
d set. In the

(a)) (b)

(c) (d)

Fig. 2 Classificcation of an artiificial dataset. P

Patterns of classs A and B are shown with boldd dots and starss.

Javan et al: F
Fast Voltage and
a Power Fllow Contingenncies Ranking
g 275
 lacck of discrimiination in whiich the perforrmance indexx
forr a case with w many ssmall violations can bee
com mparable in value
v to the iindex for a case
c with onee
hug ge violation, is known ass masking efffect. Maskingg
efffect, to some extent, can bee avoided by using higherr
ordder performance indices. However, to t avoid thee
misranking, prop per selection oof weights forr performancee
inddices is requireed. Some of thhe efforts in reducing thesee
t works of [26-27]. To overcome
efffects include the o thee
maasking, one caan try to raisee the exponen nt used in thee
pen nalty function
ns.

4.1
4 Voltage and a Power Linne Performance Indices
The voltagee performancce index is selected too
m lack due to out-of limits bus voltagess
quaantify system
and
d measures th he severity off contingencies. There aree
sev
veral types off performancee indices in literature
l [28]]
butt following voltage perfo formance index has beenn
acq
quired in this paper:
p
M
LV
W f
PII V i i (12)
Fig. 3 Structurre of GPRBF neeural network. i 1 M Vlim

Vi Vi , Vi Vimax
max

dimensional input vectorr X x1 , x 2 ,, x m fi Vmin V ,

T
cann be Vi Vimin (13)
expressed as: i i
l
here, fi is a fu
wh unction of lim uses only, Vi
mit violated bu
y i (x) w ij O j (x) (8)
j1 us, Vlim thee
is the post contingent voltagge at the ith bu
By minimiziing the Sum of Square Error
E (SSE), the on limit, Viminn and Vimax are
volltage deviatio a the lowerr
weight matriix (W) is definned as:
P andd upper limit of voltage m magnitudes att ith bus, Wi
SSE d p W O p
2
(9) (=11) is the weighting coefficiient, and M is the order off
p 1
thee exponent. If the enumerateed PIV value is i greater thann
where, d p iis a desirablee output and P is numberr of zerro, then the coorresponding contingency is recognizedd
patterns. In thhe matrix form
m: as critical or inseecure; otherw
wise, it is securre.
SSE (D W O) (D W O)
T
(1 0) An index forr quantifying tthe extent of line
l overloadss
maay be defined in terms of M MW performan nce index:
where, D d1 , d 2 , , d P n P and O o1 , o 2 , , o P lP 2n
NL
Wl Pl
are desirablee matrix and output matriix of the hiddden PIMW i lim (14)
layer respecctively. Finallly, objective function is the l 1 2n Pl
minimized SSE that can bee expressed ass: whhere, Pl and Pllim are the M MW flow of line l and thee
Min W D W O)T (D
(D ( W O) MW W capacity off line l, NL is tthe number off system lines,,
W D O (O O )1
T T
( 11) Wli and n are real nonnegat ative weightinng factor( =1))
andd exponent of o penalty fuunction, respeectively. Thiss
4 Performance Index ind
dex PIMW has a small valuue, when all line flows aree
The standdard approachh for steady state
s contingeency witthin their limiits and a highh value when there are linee
checking is to run a loaad flow for th he post-transiient oveerloads. Thus, it provides a measure of the severity off
steady state status following each outaage. Some of the line overloads fo or a given statte of the poweer system.
outages may lead to system m constraint violations
v suchh as It should be pointed out that the weig ghting factorss
load bus vooltages outsiide their norrmal limits and Wi and Wli can c be regardded as tuning g parameters..
transmission line (transfoormer) overloads. The systtem Thhese factors are selected oon the basis of o experiencee
performance index is desccribed as a peenalty functionn to witth the system m and on the rrelative importance placedd
penalize gravvely any violaation of bus vo
oltage limitatiions on the various liimit violationss. Moreover, it is observedd
or line flow cconstraints. fro
om simulation that for M=44 and n = 2, masking
m effectt
By addiing the penaalty function ns together, the hass been removeed for the IEE EE 14-bus test system.
method mayy suffer from a masking efffect. In fact, the

276 Iranian
I Journnal of Electrical & Electron
nic Engineerinng, Vol. 7, No. 4, Dec. 2011

4.2 Security Index
In order to quantify the concept of secure and
insecure operating states, five severity levels have been
determined. The four critical levels are correctively
secure, alert, correctable emergency and non-correctable
emergency and are represented by the hypotheses {H1,
H2, H3, H4}, and one non-critical level is shown by
{H0}. In this way, term of {H4} represents which all
loads supplied but operating limits are violated. These
cannot be corrected without loss of load. For ranking the
severity of contingencies, single line and generator
outages corresponding to each load pattern are also
simulated by full AC load flow to calculate the voltage
and power line performance index. Based on the
calculated performance index, the voltage and power
flow security status can be determined, according to
Table 1.
works well with linearly separable classes, but not well
established on non-linearly separable classes [31].
A usual procedure includes exploiting the crisp
secure-insecure classification provided by class
separability index ( F ) and correlation coefficient [24] is
used to select appropriate training feature for the
GPRBFN. F is defined by expression according to:
mi(S) mi(I)
Fi (S) (15)
i i(I)
where, m i(S) and m i( I ) are the mean values of variable
i in the secure and insecure classes, respectively, i
(S)
and i(I) are the corresponding standard deviations.
Variables with greater F have more discriminating
power, and are chosen as relevant attributes. Correlation
coefficient between the variables with greater F is
calculated using the expression below:

5 Generation of Training Data Cij

The RBFNN learning methods are found on our
knowledge about the behavior of the system, performed
from a large number of off-line simulations. In this way,
load patterns were generated randomly by changing the
load at each bus in a wide range. Training data should
be able to represent the whole operating space of power
system. For this purpose, for each load bus a
corresponding upper and lower limit ( PLmax , PLmin ) is
defined which represent boundary of daily load
variation. Considering this point, Fig. 4 shows the
methodology used in generating data for training and
testing of the proposed neural network. For each data
set, in order to obtain an estimate of the generalization
error, five-fold cross-validation was used, i.e., five
experiments were conducted, just one of the folds used
for testing.
6 Feature Selection
Feature selection is performed to recognize those
features that contribute most to the discrimination
ability of the neural network. Only these features are,
then, used to train the neural network and the rest are
discarded. By selecting only the proper features of the
data, training time and size of the neural network can be
reduced [29]. Many feature selection algorithms are
available in the literature such as fisher discrimination
analysis [30] and entropy maximization. The main
problem with the existing feature algorithms is that it
E x i x j E x i E x j
(16)
i j
7 Results and Discussion
To demonstrate the effectiveness of the proposed
GPRBFNN model, it has been tested for static security
evaluation of IEEE 14-bus system to rank the
contingencies expected to cause steady state power flow
and bus voltage violations at different loading
conditions (%60 to %140 of base load) and
contingencies. Load flow analysis has been carried out
using the Power System Analysis Toolbox (PSAT) [32].
Training and testing results are simulated on Pentium IV
2.8 GHz, 512 MB RAM, 40GB Hard drive.
The performance of the GPRBFN is determined by
evaluating the following measures during training and
testing phases:
1) Mean of Absolute Percentage Error (MAPE):
NNOi DOi
APEi (%) 100 (17)
DOi
1 N
APEi
MAPE(%) (18)
N i 1
where N, NNOi and DOi are the number of samples in
the data set, neural network output and desired output
obtained from off-line simulation, respectively.

Table 1 Security levels for classification of line overloads and out-of limits bus voltages.
Security level Line flow (%) PIMW Voltage Deviation (%) PIV
Secure (H0) <100 <0.25 <10 <0.04
Correctively Secure (H1) 100-120 0.25-0.51 10-20 0.04-0.79
Alert (H2) 120-150 0.51-1.26 20-30 0.79-4.00
Correctable Emergency (H3) 150-180 1.26-2.62 30-40 4.00-12.60
Non-correctable Emergency (H4) >180 >2.62 >40 >12.60

Javan et al: Fast Voltage and Power Flow Contingencies Ranking 277
 Load Variation Class I: <0.25 Class II: 0.25-0.51 Class III: 0.51-1.26
(60% to 140% base case)

Class IV: 1.26-2.62

Calculate the
K-fold Cross-Validation
power flow
(training and testing patterns)
performance Class V: >2.62
index (PIMW)
IEEE 14-Bus Power System Full AC Load
Flow Calculate the
voltage Class V: >12.6
performance
Contingency Set index (PIV)
Class IV: 4.00-12.6

Class I: <0.04 Class II: 0.04-0.79 Class III: 0.79-4.00

Fig. 4 Methodology for the for data generation contingency analysis using neural network.

2) Mean of Square Error (MSE): the same for different load scenarios such as the line
1 N outage between bus 1 and bus 2. In addition, it can be
MSE = NNOk DOk
2
(19) seen that GPRBFNN method has a considerable
N k =1
performance than classical RBFNN. As it can be
3) Classification Accuracy (CA): observed from Table 3, the outage of L1 is most severe
No. of samples classified correctly
CA(%) = 100 (20) in all of these cases. The bold outages indicate the
Total No. of samples in data set misclassified contingencies.

7.1 IEEE 14-Bus System Table 2 Class separability index F for IEEE 14-bus system.
IEEE 14-bus is the test system which contains 2 Feature No. Variables Separability Index F
generators, 14 buses, 20 lines and 3 condensers. This 1 V9 0.74
2 Phase14 0.33
power system is shown in Fig. 5.
3 Qload13 0.99
Input features are selected by a statistical method
4 Qload11 0.80
based on class separability index and correlation 5 Qsc8 0.65
coefficient technique. Separability factor of every 6 Pline7-9 0.60
variable of the power system was calculated as per (15) 7 Pline6-13 0.34
and (16), respectively. In Table 2, 12 features were 8 Qline4-9 0.36
selected, having high separability factor and low 9 Qline5-6 0.37
correlation coefficient among them. 10 Qline6-3 0.68
A total of 800 (contingencies) load flow cases are 11 Qline7-8 0.62
selected for the load scenarios using different lines and 12 Qline12-13 0.36
generation outages to compute the voltage performance
indices and determine the security levels as pointed out
in Table 1. After training, the optimum size of the
GPRBFNN was found to be (12-284-5) which each of
the security status of secure, correctively secure, alert,
correctable emergency and non-correctable emergency
has 42, 105, 94, 40 and 3 neurons (clusters) in the
hidden layer, respectively. GPRBFNN was tested using
different sets of load flow cases that were not used
during the training process. For testing, 80 load flow
cases are selected using different transmission lines and
generator outages.
It can be observed from Table 3 that the column
corresponding to the voltage contingency ranking
obtained by GPRBFNN for different loading conditions
is almost same as that obtained by the Newton Raphson
Fig. 5 IEEE 14-bus test system.
(NR). The ranking order of similar contingencies is not

278 Iranian Journal of Electrical & Electronic Engineering, Vol. 7, No. 4, Dec. 2011
Table 3 IEEE 14-bus system-Voltage ranking at different loading conditions.
Loading Line/Gen Contingency Ranking
No PIV
P (MW) Q (MW) From/To Load Flow (NR) RBFNN GPRBFNN
1 360.48 50 1-2 5.72 H3 H0 H3
2 412.87 60 2-3 0.43 H1 H1 H1
3 26861 80 9-14 0.89 H2 H4 H2
4 423.37 55 G8 0.11 H1 H1 H0
5 353.32 73 1-2 14.65 H4 H2 H4
6 306.65 40 G1 0.53 H1 H1 H1
7 310.60 62 6-11 0.92 H2 H2 H2
8 299.34 85 10-11 0.51 H1 H0 H1
9 378.33 63 5-6 0.80 H2 H0 H2
10 324.39 78 G6 0.31 H1 H1 H1

The proposed algorithm is performed for 7.2 Computation Burden

contingencies ranking of line overload severity and the
optimum size of the GPRBFNN was found to be (12-
215-5) which each of the security classes has 31, 53, 72,
49 and 10 neurons (clusters) in the hidden layer.
In Table 4, the IEEE 14-bus system was tested for
20 single line outage cases and five generator outage
cases to identify the most severe lines or generators
contingencies on line overloads at 120% loading. Such
critical contingencies should be quickly identified for
further detailed evaluation or, where possible, corrective
measures taken.
In all of the obtained results, the preference of
GPRBFNN method in fast and accurate ranking of the
contingencies expected to cause steady state line power
and bus voltage violations is obvious respect to the
classical RBFNN method.
Table 5 shows the obtained results of the GPRBFNN
and RBFNN for steady state security evaluation in a
wide range of loading. The proposed method has
enough speed to evaluate the power system security in
ranking of voltage violated buses and line overload
severities. As we know, the RBF neural network is a
famous tool for pattern classification and nonlinear
functional approximation. However, the results of
Tables imply that in almost all cases, the classical
RBFNN gives the largest errors due to their limitations
in optimum clustering.
One of the main shortcomings attributed to classical
radial basis function neural network (RBFNN) is their
long training process. This is basically because of the
iterative nature of training of such networks. Inability in
choosing proper units center and width (radius) and
also numbers of optimal neurons in hidden layer causes
to take long time for training process. In this neural
network, neurons are being added to the hidden layer till
the classical algorithm converges to a tuning error.
However, using the proposed algorithm, the
enhanced neural network determines automatically
appropriate centers and radii.
As shown in Table 5, the training process of RBF
neural network for voltage and power flow
contingencies ranking takes 40.60 and 33.13 seconds,
respectively. On the other hand, the training process of
the proposed method (GPRBFNN), for the same data
and on the same computer, takes 9.8 seconds. The main
reason is the non-iterative nature of the GPRBFNN,
which also brings more stability over the classical
RBFNN.
8 Conclusion
The analytical method for contingency evaluation
consists of the computation of the performance indices
by running full AC load flow for all contingencies
which is both time consuming and susceptible to miss
ranking effects.

Table 4 IEEE 14-bus system-PIMW ranking of generator and power line outages at 120% loading.
Ranking of contingencies according to severities (line & generator outages)
Loading Methods Secure Correctively Alert Correctable emergency non-Correctable
(%) secure emergency
G8 L14 G3,G6,L19 G2,L13,L9,L15,L11,L5,L20,L8 L1,L3,L10,L2,L4,L7
Load Flow
L17,L6,L12,L16,L18
120 L3, L5 G8 G3,G6,L19 G2,L13,L9,L15,L11,L14 L1,L2, L7
RBFNN
L8 L4,L10,L18 L17,L6,L12,L16,L20
G8 L14 G3,L15,G6,L19 G2,L13,L9,L11,L5,L20,L8,L17 L1,L3,L10,L2,L4,L7
GPRBFNN L6 ,L12,L16,L18

Javan et al: Fast Voltage and Power Flow Contingencies Ranking 279
 Table 5 IEEE-14 bus system- obtained results of voltage and power flow ranking performance at different loading conditions.
Voltage Performance Index Ranking Power Flow Performance Index Ranking
Train Set Test Set Train Set Test Set

RBFNN GPRBFNN RBFNN GPRBFNN RBFNN GPRBFNN RBFNN GPRBFNN

No. Neuron 431 284 431 284 512 215 512 215

MSE 0.163 0.009 0.83 0.012 0.093 0.010 0.563 0.015

MAPE (%) 2.93 1.07 12.03 3.77 2.10 0.69 9.98 1.150

CA (%) 93.14 97.12 88.14 95.20 94.03 98.12 84.30 96.03

Time (s) 40.60 9.8 1.03 0.52 33.13 9.8 1.14 0.65

Radial basis function neural network based on
Growing and Pruning algorithm has been proposed to
evaluate static security of IEEE 14-bus power system.
The proposed neural network approach to contingency
analysis consists of training the ranking module for
different contingencies corresponding to voltage
violated buses and power flow (MW) violated lines. The
ranking module provides accurate levels of performance
indices (PIV, PIMW) for unknown load patterns, which
are compared with Newton Raphson method. In the
presented algorithm, optimum centers and
corresponding radii are obtained. So, the training of
RBFN based on Growing and Pruning algorithm
requires less computation time and the testing accuracy
of RBFN has been made higher by applying it. Test
results reveal that the proposed network involve less
time and is suitable for online static assessment under
uncertain loading conditions and is expected to perform
similarly on even larger systems and handle even
greater number of contingencies than reported here.
References
[1] Shahidehpour M., Tinney W. F. and Fu Y.,
Impact of security on power systems operation,
Proc. IEEE, Vol. 93, No. 11, pp. 2013-2025, Nov.
2005.
[2] Mahadev P. M. and Christie R. D., Envisioning
Power System Data: Vulnerability and Severity
Representations for Static Security Assessment,
IEEE Trans on Power Systems, Vol. 9, No. 4, pp.
1915-1920, Nov. 1994.
[3] Ejebe G. C. and Wollenberg B. F., Automatic
contingency selection, IEEE Trans. Power
Apparatus Syst, Vol. PAS-98, pp. 97-109, Jan.
1979.
[4] Jadid S. and Jalilzadeh S., Effective transient
stability assessment based on composite indices,
Iranian Journal of Electrical & Electronic
Engineering (IJEEE), Vol. 1, No. 2, pp. 74-80,
2005.
[5] Scott B. and Alsac O., Fast decoupled load
flow, IEEE Trans on PAS, Vol. PAS-93, pp.
859-869, May 1974.
[6] Mori H., Tanaka H. and Kanno J., A
Preconditioned Fast Decoupled Power Flow
method for contingency screening, IEEE Trans.
on Power System, Vol. 11, pp. 357-362, Feb.
1996.
[7] Brandwajn V. and Lauby M. L., Complete
bounding method for AC contingency screening,
IEEE Trans. on Power System, Vol. 4, No. 1, pp.
724-729, May 1989.
[8] Tsang I., Kocsor A. and Kwok J., Large-scale
maximum margin discriminate analysis using
core vector machines, IEEE Trans. Neural
Network, Vol. 19, pp. 610-624, 2008.
[9] Kim H. and Singh C., Power System
Probabilistic Security Assessment using Bayes
Classifier, Electric Power Systems Research,
Vol. 74, No. 1, pp. 157-165, 2005.
[10] Ramos G., Sanchez J. L., Torres A. and Rios M.
A., Power Systems Security Evaluation Using
Petri Nets, IEEE Trans on Power Delivery, Vol.
25, No. 1, pp. 316-322, 2010.
[11] Diao R., Vittal V. and Logic N., Design of a
Real-Time Security Assessment Tool for
Situational Awareness Enhancement in Modern
Power Systems, IEEE Trans. on Power System,
Vol. 25, No. 2, pp. 957-965, 2010.
[12] Aghamohammadi M. R., Static Security
Constrained Generation Scheduling Using
Sensitivity Characteristics of Neural Network,
Iranian Journal of Electrical & Electronic
Engineering (IJEEE), Vol. 4, No. 3, pp. 104-114,
2008.
[13] Aghamohammadi M. R., Hashemi S. S. and
Ghazizadeh M. S., Improving Voltage Stability
Margin Using Voltage Profile and Sensitivity
Analysis by Neural Network, Iranian Journal of
Electrical & Electronic Engineering (IJEEE),
Vol. 7, No. 1, pp. 33-41, 2011.

280 Iranian Journal of Electrical & Electronic Engineering, Vol. 7, No. 4, Dec. 2011
[14] Morisoon K., Wangg L., Kundur P., Lin X., G Gao [25
5] Broomheaad D. S. andd Lowe D., Multivariable
e
W., HHe C., Xue F., Xu J., Xu T. and Xue Y., functionall interpolationn and adaptiv ve networks,,
Criticcal requiremments for su uccessful onl nline Complex Syst.,
S Vol. 2, NNo. 1, pp. 321
1-355, 1988.
securitty assessmentt, presented at a the IEEE P PES [26
6] Halpin T. F., Fischl R. and Fink R.,, Analysis off
PSC& &E Conf., pp. 30-39,
3 New York,
Y Oct. 10--13, contingency selection algorithms, IEEE Trans.
2004. on Power Apparatus annd Systems, Vol. V PAS-103,,
[15] Avila--Rosales R., Sadjadpour A., A Gibescu M., No. 5, pp. 938-945, Maay 1984.
Morisoon K., Ham madani H. and Wang L., [27
7] Schafer K.K F. and V Verstege J. F.,F Adaptivee
ERCO OTs implem mentation of online dynam amic proceduree for maskingg effect com mpensation inn
securitty assessmennt, presented d at the Paanel contingency selection aalgorithm, IE EEE Trans. onn
Session of the IEEEE PES, 2003. Power System, Vol. 5,, No. 2, pp. 539-546,
5 Mayy
[16] Wang Y. H., Chann K. W. and Pong T. Y. G., 1990.
Dynaamic voltage security con nstrained optim mal 8] Wood A. J. and W
[28 Wollenberg B. F., Powerr
coordiinated voltagge control using u enhannced Generatioon Operation aand Control, 2nd Ed, Johnn
particlle swarm optimization, IET Genner. Wiley & Sons,
S pp. 15-664, 1996.
Transm m. Distrib, Vool. 5, No. 2, pp.
p 239-248, F Feb. [29
9] Nojun Kw wak N. and Chong-Ho Choi C. H.,,
2011. Input feature
f selec
ection for classificationn
[17] Samanntaray S. R., Kamwa I.. and Joos G., problems, IEEE Tranns. Neural Neetworks, Vol..
Ensem mble decisionn trees for phaasor measurem ment 13, No. 1, pp. 143-159,, Jan. 2002.
unit-baased wide-areea security asssessment in the 0] Jensen C. A., El-Sharkaawi M. A. and
[30 d Marks R. J.,,
operations time frrame, IET Gener. Trannsm. Power sy ystem security ty assessment using neurall
Distribb, Vol. 4, No.
N 12, pp. 1334-1348,
1 D
Dec. networks: feature selection using u fisherr
2010. discriminaation, IEEE Trans. on PowerP System,,
[18] Alessaandri A., Grillo S., Massuccco S., Silvestroo F. Vol. 16, No.
N 4, pp. 757--763, 2001.
and VVimercati G., ANN Appllication For O On- [31
1] Sa Da Co osta J. M. G G. and Munro N., Patternn
Line Power Systtem Security y Assessmen ent, Recognition in Pow wer System m Security,,
Internaational Confference on Power Systeems Internatio
onal Journal of Electrica al Power &
(PMAP PS 2006), pp.1-6, 2006. Energy Syystems, Vol. 6, No. 1, pp p. 31-36, Jan..
[19] Shang J., Zhang J.,, Zhao W. an nd Liu J., AN
ANN 1984.
Based Dynamic Vooltage Security y Assessmentt for [32
2] Milano F.,
F Matlab power sytem m simulationn
a Prractical Pow wer, Intern national Pow wer package, http://www w.power.uwateerloo.ca/~fmill
Engineeering Conferrence (IPEC 2007), pp. 7794- ano/psat. htm.
h
798, D
Dec. 2007.
[20] Moulinn L. S., Alvess Da Silva A. P., El-Sharkaawi,
M. A. and Marks R. R J., Support vector machiines
for trransient stabiility analysiss of large-sccale Dawoodd Seyed Javan n was born inn
power systems, IE EEE Trans. on n Power Systtem, Mashhadd, Iran, in 1984. He receivedd
Vol. 19, No. 2, pp. 818-825,
8 2004
4. the B. Sc. degrees in electricall
[21] Moham mmadi M. and Ghareh hpetian G. B., engineerring from the Azad
A Universityy
of Mashhhad. He is currrently pursuingg
Appliication of corre vector machines for on-lline the M.SSc. degree in thhe Departmentt
voltage security assessment using g a decision-trree- of Electtrical Engineerring, Ferdowsii
based feature selecction algorith hm, IET Genner. Universiity of Mashh had, Mashhad,,
Transm m. Distrib., Vol.
V 3, No. 8, pp. 701-7712, Iran. HHis research interests aree
2009. power
p system
m security asssessment, loaad and pricee
[22] Kalyanni S. and Shannti Swarup K.., Power Systtem forecasting
f in electricity maarkets, neural networks andd
Securiity Assessmennt using Binaary SVM Baased genetic
g algorith
hms.
Patternn Recognitionn, Internatio onal Journall of
Electriical Powerr and Ennergy Systeems
Engineeering, Vol. 2,
2 No. 3, pp.18 89-195, 2009.
Habib RRajabi Mashha adi was born inn
[23] Saeh II. S. and Khaiiruddin A., D Decision Tree for Mashhadd, Iran, in 1967. He receivedd
Static Security Asssessment and d Classificatioon, the B.SSc. and M.Sc. degrees withh
Internaational Confference on Future
F Compuuter honor frrom the Ferdowsi Universityy
and Coommunicationn (ICFCC), pp p. 681-684, 20009. of Masshhad, both in electricall
[24] Matos M. A., Hatziargyriou N. D. D and Pecasloopes engineerring, and the Ph.D. degreee
J. A., Multicontinngency stead dy state secuurity from thee Department off Electrical andd
evaluaation using fuzzy
f clusteriing techniquees, Computeer Engineerin ng of Tehrann
IEEE T Trans. Powerr Syst., Vol. 15
5, No. 1, pp. 1177- Universiity, Tehran, Iraan, under jointt
cooperation
c of
o Aachen U University off Technology,,
183, F
Feb. 2000.

Javan et al: F
Fast Voltage and
a Power Fllow Contingenncies Ranking
g 281
 Germany, inn 2002. He is asa Associate Proofessor of electrrical
engineeringg at Ferdowsii University of Mashhad. His
research intterests are poweer system operaation and plannning,
power systeem economics, and biological computation.

Modjtaba Rou
M uhani receivedd the
B
B.S. and M.S.M degrees in
e
electrical en
ngineering ffrom
F
Ferdowsi univeersity of Mashhad,
M
Mashad, Iran, in
i 1990 and 19993,
r
respectively, an
nd the Ph.D. deggree
i electrical engineering ffrom
in
A
Amir Kabir university of
T
Technology, Tehran,
T Iran, in
2
2001. He is cu urrently with A Azad
I
Islamic univ
versity, Gonaabad
branch, Gonnabad, Iran, whhere he is an assistant Professoor in
control enngineering, eleectrical engineering departm ment,
engineeringg faculty. Hiss research intterests are neeural
networks, computationaal intelligencee and nonlinnear
adaptive coontrol.

282 Iranian
I Journnal of Electrical & Electron
nic Engineerinng, Vol. 7, No. 4, Dec. 2011

View publication stats