You are on page 1of 9

Electric Power Systems Research 53 (2000) 197 205

www.elsevier.com/locate/epsr

Fast voltage contingency screening and ranking using cascade


neural network
S.N. Singh a,*,1, L. Srivastava b, J. Sharma c
a
Department of Electrical Engineering, Hong Kong Polytechnic Uni6ersity, Hung Hom, Kowloon, Hong Kong
b
Department of Electrical Engineering, M.I.T.S., Gwalior, India
c
Department of Electrical Engineering, Uni6ersity of Roorkee, Roorkee, India

Received 28 January 1999; accepted 12 May 1999

Abstract

Voltage contingency selection and ranking is performed to choose the contingencies that cause the worst voltage problems. In
this paper, a cascade neural network based approach is proposed for fast voltage contingency screening and ranking. The
developed cascade neural network is a combination of a filter module and a ranking module. All the selected contingency cases
are applied to the filter module, which is trained to classify them either in critical contingency class or in non-critical contingency
class using a modified BP algorithm. The screened critical contingencies are passed to the ranking module (four-layered
feed-forward ANN) for their further ranking. The effectiveness of the proposed method is demonstrated by applying it for
contingency screening and ranking at different loading conditions for IEEE 30-bus system and a practical 75-bus Indian system.
Once trained, the cascade neural network gives fast and accurate screening and ranking for unknown patterns and will be suitable
for on-line applications at energy management centre. 2000 Elsevier Science S.A. All rights reserved.

Keywords: Security analysis; Contingency selection and ranking; Cascade neural network; Subordinate and dominant class; Imbalance training set

1. Introduction and background flow, linearized load flow, one iteration of AC load
flow, local solution methods etc. to identify cases caus-
Contingency selection is carried out for quickly iden- ing limit violations. The network monitored quantities
tifying those contingencies which may cause out-of- are first calculated for all the contingencies and ranking
limit violations so as to reduce the number of is done based on the results of the approximate
contingencies that need to be analysed by full AC load solutions.
flow while assessing the power systems security. Two Most of the work on contingency selection algorithm
popularly used methods for contingency selection are utilises the second order performance indices which, in
ranking methods and screening methods. general, suffers from masking and misranking effects.
The lack of discrimination, in which the performance
Ranking methods involves ranking of contingencies in
index for a case with many small violations can be
approximate order of severity. Contingencies are
comparable in value to the index for a case with one
ranked based on the value of a scalar performance
huge violation, is known as Masking effect. By most of
index (PI) which measures the system stress in some the operational standards, the contingency with one
manner. In ranking method, the performance indices huge violation is much more severe than that with
are explicitly expressed in terms of network variables many small violations. Misranking of contingencies are
and are directly evaluated. Several PI based methods mainly due to the inaccuracies in the model used for
have been suggested and tested for voltage security computing the performance indices and is characterised
analysis [19]. Screening methods use approximate net- by errors in the computed order of relative severity of
work solutions such as distribution factors, DC load various contingencies. Masking effect to some extent,
can be avoided by using higher order performance
* Corresponding author. indices. However, to avoid the misranking, proper se-
1
On leave from the university of Roorkee, Roorkee, India. lection of weights for performance indices are required.

0378-7796/00/$ - see front matter 2000 Elsevier Science S.A. All rights reserved.
PII: S 0 3 7 8 - 7 7 9 6 ( 9 9 ) 0 0 0 5 9 - 0
198 S.N. Singh et al. / Electric Power Systems Research 53 (2000) 197205

Some of the efforts in reducing these effects include Vankayala and Rao [18] used a method based on a
the works of Halpin et al. [6] and Schafer et al. [7]. The coupled scheme (ANN and expert system) for power
post-contingency PI is presented by Dabbaghchi and system security enhancement, combining ANN based
Irissari [8] as a sum of the base case performance index contingency screening with an expert system (ES) based
and three correction terms. Out of the three correction preventive control. Ghosh and Chowdhury in [19] de-
terms only one term needs to be calculated for every signed a four-layer neural network for line-flow contin-
outage which includes branch independent and base gency ranking. A regression-based correlation
case dependent terms. Singh et al. [9] proposed a new technique was used for feature selection to train the
method for optimal selection of weights along with ANN using back propagation algorithm.
higher order voltage performance index which elimi- Mansour et al. [20] applied a feed-forward ANN for
nates misranking and masking effects. The method is dynamic security contingency screening and ranking.
based on the concept of tuning the weights (with their Stability energy margin, which reflects the ranking, was
minimum possible values) such that the relative severity selected as output of the ANN. The training error
of the contingency predicted along with the approxi- decreases along with the number of iterations while the
mate distribution factors method matches with the full testing error decreases at first, bounces around, and
AC power flow method. then starts to increase. The optimal learning and gener-
Hao et al. [10] explored the use of distributed com- alization are achieved close to the global minimum of
puting capability, coupled with efficient bounding and the testing error [20]. Brandwajn et al. [21] extended the
sparse vector methods to improve the speed of contin- existing concepts of contingency severity measures used
gency screening applications. The first iteration of fast in static security analysis into the area of dynamic
decoupled method was used to compute the network security assessment and applied an MLP model for
solution and rank the contingencies according to classification of contingencies into stable/unstable class.
voltage and thermal violations. Parallel self-organising hierarchical neural networks
In [11], Wehenkel proposed an approach to voltage (PSHNN) are also use for the various applications of
security assessment exploiting non-parametric regres- the power systems [27,28].
sion technique to extract simple /reliable models of the Though ANN has been applied for contingency
severity of a contingency, defined as the difference screening and ranking for dynamic and steady state
between pre- and post-contingency load power margins security assessment in a number of references, no effort
(LPM). Ejebe et al. [12] proposed a curve fitting based has been made to apply cascade neural network (CNN)
approach for contingency screening and ranking for for voltage contingency screening and ranking, to the
voltage stability analysis. In [13], Chadalavada et al. best of authors knowledge. The conventional methods
developed a contingency screening system for the selec- of voltage contingency ranking are either time consum-
tion and ranking of dynamic security assessment using ing or inaccurate specially, at different loading
a sparse transient energy function (STEF) based filter- conditions.
ing scheme. In a case, when only a four-layered feed-forward
With the development of artificial intelligence and ANN was applied for the screening and ranking of
neural network in recent years, there is growing interest contingencies, it was observed that the normalised val-
in applying these approaches to screen and rank contin- ues of voltage PIs reproduced by ANN were close to
gencies. Sobajic and Pao [14] proposed a rule-based the desired values of PIs. But some contingencies,
contingency screening for both single and multiple line- which were non-critical, were shown by the ANN as
outages. In [15], Hsu et al. presented a fuzzy set-based critical ones and were ranked above some critical lines.
contingency ranking. Fischel et al. [16] used a three Thus a single MLP model was unable to distinguish the
layer neural network with back-propagation learning to non-critical contingencies. So an approach has been
classify contingencies into those leading to secure and investigated, in this paper, to filter out the non-critical
insecure states. The inputs used are the bus susceptance contingencies first and rank only the critical contin-
matrix and the real power injection at each bus. In the gency cases.
test results on a small system, they found that the In this paper, a cascade neural network (CNN) based
neural network gave only a few errors in classifications. method is developed for voltage contingency screening
In [17], a Hopfield network was proposed for contin- and ranking. The proposed cascade neural network is a
gency screening, using an optimisation method to find combination of filter module and a ranking module. A
the weights and thresholds of the ANN. The optimisa- filter module is designed for classification task and
tion method used is linear programming techniques to trained using a modified BP algorithm [22] so that the
maximise the probability of correct classification of module learns faster and provides efficient classification
contingencies. This implementation classifies contingen- for contingencies (PIs) in critical/non-critical class. In
cies according to the number and type of limit the proposed architecture of CNN, the filter module is
violations. a three-layered feed-forward neural network with single
S.N. Singh et al. / Electric Power Systems Research 53 (2000) 197205 199

output while the ranking module is a four-layered fi = V LOWER


i Vi for Vi B V LOWER
i
neural network. fi is a function of limit violated (LV) buses only, Vi is
The pre-outage real and reactive line-flows and termi- the post contingent voltage, V UPPER is the upper and
i
nal voltages of the contingent element and total de- V LOWER
i is lower limit of voltage, and wi is weighted
mand, along with a topology number in the form of function at ith bus, and M is order of exponent.
bi-polar digits (+ 1, 1), corresponding to line-outage Assuming wi = 1 and M= 4, masking effect is re-
are selected as input features for training of the cascade moved. Here summation is carried out only for limit
neural network. The proposed neural network based violated buses. A number of line outages, having no
technique is applied for contingency screening and bus voltage violations (non-critical contingencies) give 0
ranking at different loading conditions in IEEE 30-bus value of voltage performance index for all the load
system and a practical 75-bus Indian system. patterns. Such lines are not selected for training of the
neural network. It is observed that the list of critical
contingent cases is dynamic in nature depending upon
2. Voltage performance index the loads at various buses, i.e. a critical contingency
may be a non-critical one at some other loading condi-
The severity of the voltage contingency is evaluated tion. Similarly, the ranking order of critical contingency
(using full AC load flow) as voltage performance index cases may be different at different loading conditions.
[2] given by

PI= % (wi /M)( fi )M (1) 3. Cascade neural network approach


i  LV

Where function fi is defined as The block diagram of the proposed cascade neural
network, consisting of a filter module and a ranking
fi = Vi V UPPER
i for Vi \V UPPER
i module is shown in Fig. 1. A filter module which is a
three-layered feed-forward ANN with single clamped
output as shown in Fig. 2, is used to classify the
contingencies in critical or in non-critical class. The
filter module is trained using modified BP algorithm
[22] such that its target output is high (0.9) when
presented with a sample from non-critical class (C1),
and low (0.1) when presented with a sample from
critical class (C2).The outputs of the filter module
greater than 0.9 are clamped to 0.9, similarly outputs
smaller than 0.1 are clamped to 0.1 to reduce the
Fig. 1. Cascade neural network for contingency selection and rank-
ing. likelihood of the network getting stuck in local minima.
The training set for filter module contains many
more exemplars for critical class C2 than for class C1.
When training a network with standard back-propaga-
tion for such a two-class problem, in which one domi-
nant class contains far more exemplars than a
subordinate class (i.e. the training set is imbalance),
the rate of convergence of net output error is very low.
This is because the negative gradient vector computed
by back-propagation for an imbalanced training set
does not initially decrease the error for the subordinate
class as shown in Fig. 3. Consequently in the initial
iteration, the net error for the exemplars in the subordi-
Fig. 2. Filter module (three-layered ANN).
nate class increases significantly. The subsequent rate of
convergence for the exemplars of the subordinate class
is very low. To solve this problem, a modified al-
gorithm [22] is used that calculates a direction in
weight-space which is downhill for class C1 as well as
for class C2 in each iteration. For such a two-class
problem it is proposed that the computed descent vec-
tor 6 should point in a downhill direction for both
Fig. 3. Relationship between gradient vectors. classes as shown in Fig. 4, i.e. 6 satisfies
200 S.N. Singh et al. / Electric Power Systems Research 53 (2000) 197205

ered neural network (model B) was selected. The prob-


lem of determining the optimal number of hidden
neurons is a crucial one. The number of hidden neurons
must be sufficiently large to realize a certain function.
Several structures were considered with different num-
bers of hidden neurons to determine the best
configuration.

Fig. 4. Direction of gradient vectors in modified algorithm. 3.3. Training set generation

6.9E1(W)B 0, and 6.9E2(W) B0 (2) For generating the training set, a large number of
load patterns are generated by perturbing load ran-
where E1(W) refers to the error due to C1 and E2(W) domly at all the buses. Full AC load flow is performed
refers to the error due to C2. for each case to compute pre-outage line-flows and
Direction of 6 is set so that it bisects the angle terminal voltages at the contingent line, and total load.
between 9E1(W) and9E2(W): For each load pattern, AC load flow is also performed
9E1(W) 9E2(W) for all the single line-outages to compute voltage PIs.
.6= .6 (3) In realistic systems, the training set generation may be
9E1(W) 9E2(W)
a difficult task. The past experience can be used for
and thus the rate of learning can be accelerated by one ignoring a number of line-outage cases, which were
order of magnitude for such two-class problems. The always observed as non-critical, and training set may be
learning rate can be further improved by using adaptive generated by considering the remaining lines. Though,
learning rate with this approach. power systems do not always operate under n1 con-
tingencies, the n 1 contingency is the most common
3.1. Feature selection event in power systems and so only single-line outages
are considered here.
Total demand, pre-contingency real and reactive line-
flows in contingent element and pre-contingency termi- 3.4. Data normalization
nal voltages of contingent element are selected as input
features for training of the cascade neural network. Higher valued input variables may tend to suppress
Some other feature selection techniques [27,29] may be the influence of smaller ones. To overcome this prob-
used for optimal selection of inputs for training and lem, the neural networks were trained with normalized
testing of neural network. Since the same cascade ANN input data, leaving it to the network to learn weights
has been trained to screen and to rank a number of associated with the connections emanating from these
single line-outage contingencies, a topology number in inputs. The raw data are scaled in the range 0.10.9 for
the form of a string of bi-polar digits (+1, 1) is also use by neural networks to minimize the effects of
applied as input to the CNN. magnitude between inputs [23]. In case of output vari-
able, if it assumes value close to unity (1) or zero (0), it
3.2. Number of layers and hidden neurons causes difficulty in training as the value unity or zero
are practically never realized by the activation or
For filter module, a three-layer neural network [22] threshold function [24]. A way to overcome this
was designed and trained so that non-critical contingen- difficulty is to normalize the variable (x) to remain
cies may be filtered out. However, for ranking modules, between some suitable range (say 0.1 and 0.9).In this
different ANN architectures using one and two hidden case, each input or output parameter x is normalized as
layers were tried to select the best architecture. The xn before being applied to the neural network according
performance of a four-layer feed-forward neural net- to
work was compared with a three-layer neural network
0.8 (xxmin)
of equal complexity for ranking the critical contingen- xn = +0.1 (4)
cies. This was observed that performance of the four- (xmax xmin)
layered ANN (model A), with sigmoidal function in where xmax and xmin are the maximum and minimum
both of its hidden layer was poor as compared to its values of data parameter x. The input data are nor-
equivalent three-layer ANN. But, when tanh function malised between values 0.9 and 0.1. Similarly, output
was used in its first hidden layer and sigmoidal in the data (PIs) are normalized for each load pattern be-
second layer, the four-layered ANN (model B) was tween 0.9 and 0.1 (0.1 for non-critical and 0.9 for most
proved to be superior in terms of training time and critical case for each load pattern). For some line
error performance. So, for ranking module, a four-lay- outages, load flow solution does not converge at some
S.N. Singh et al. / Electric Power Systems Research 53 (2000) 197205 201

loading condition. Such line-outages are placed at the of the weak systems (highly stresses) in India. The
top of the ranking list with PI value as 0.91. training set was generated as discussed in Section 3.3.
Training of filter and ranking modules was accom-
3.5. Solution algorithm plished on Digital Alpha Server 400/6/166. The RMS

'
error is calculated as
The solution algorithm for voltage contingency 1 P
screening and ranking can be summarized in following RMS Error= % [T(i )O(i )]2 (6)
Pi = 1
steps: Where P is the number of patterns, T(i ) is the target
(a) A large number of load patterns are generated by output and O(i ) is the actual output of ANN for ith
perturbing load randomly at all the buses. pattern. The test results of two sample systems using
(b) Full AC load flow is performed for each case to the proposed cascade ANN based method is presented
compute pre-outage line-flows, terminal voltages at the below.
contingent line, and total load, and also for single
line-outages to compute voltage PIs using (1). 4.1. IEEE 30 -Bus system
(c) The line-outage cases having zero values of PI for
all the load patterns are not considered for training of For IEEE 30-bus system 13 lines were selected in
the neural network. contingency list for screening and ranking by the cas-
(d) The input data as well as output (voltage PIs) are cade neural network. For voltage contingency screening
normalised between 0.9 and 0.1, using (4). and ranking 910 patterns showing 70 load scenarios
(e) The normalised input data along with the topol- were generated by varying the loads randomly at all the
ogy number corresponding to the contingent line are buses in the range of 50160% of their base case values
used for training of the filter module. and utilising the pre-contingency data and the voltage
(f) During training, two gradient vectors 9E1(W) and performance indices corresponding to 13 single line-
9E2(W) corresponding to the subordinate class and the outages. Inputs used to train the CNN were the total
dominant class respectively are computed. load demand (real and reactive), pre-contingency real
(g) The direction of descent vector 6 which points in and reactive line-flows and terminal voltages in the
a downhill direction for both classes is calculated using contingent element along with the four-digit line-outage
(3) and the weights are modified as code (topology number), thus 10 in numbers.
Seven hundred and eighty patterns corresponding to
W(k +1)=W(k) l6
60 load scenarios, each having 13 cases of single line-
where W(k + 1) is the weight in (k + 1)th iteration, outages were used to train the filter module (10-6-1) of
W(k) is the weight in kth iteration and l is the learning the cascade neural network. The filter module with six
rate. neurons in the hidden layer was trained to filter out
(h) The magnitude of the descent vector 6 is calcu- non-critical contingencies and only critical contingen-
lated as cies were selected for their further ranking by the
ranking module. Thus out of 780 training patterns, the
6 = 9E1(W)+ 9E2(W) (5)
screened critical contingencies were 352 corresponding
(i) The adaptive learning rate is also used to speed up to 60 load scenarios. These 352 patterns were used to
the training of the module. train the ranking module, a four-layered (10-6-7-1)
(j) Training of the filter module is continued till all ANN (model B) using standard back-propagation
the training patterns are classified correctly. algorithm.
(k) Only those contingency cases belonging to critical During testing phase, remaining 130 patterns were
class are applied to train the ranking module. tested for screening through the trained filter module.
(l) Training of the ranking module is continued using The trained module correctly screened all the critical
the standard BP algorithm till the error for testing cases. These 54 critical contingency cases corresponding
patterns [20] is obtained minimum. to ten load scenarios were then tested by the trained
ranking module for their ranking.
The superiority of modified BP algorithm depends
4. Results and discussion upon the unbalancing of training patterns [22]. In case
of IEEE 30-bus system, 352 patterns were in critical
To establish the effectiveness of the proposed cascade class while 428 were in non-critical class. Training of
neural network based method, it has been tested for BP (10-6-1) model and modified BP (10-6-1) model was
voltage contingency screening and ranking of IEEE continued till all the patterns were classified correctly.
30-bus system [25] and a practical 75-bus Indian system As shown in Table 1, the modified BP (MBP) algorithm
[26] representing 400 and 220 kV buses of the U.P. was found to be about 2.6 times faster as compared to
State Electricity Boards (India) network which is one standard BP algorithm.
202 S.N. Singh et al. / Electric Power Systems Research 53 (2000) 197205

Table 1
Comparison of BP and modified BP algorithms for IEEE 30-bus system

S. No. ANN Model Training Testing

Iterations Time (CPU) Max. abs. error RMS error

1 BP Model 33 774 30.29 s 0.209374 pu 5.24971E-02 pu


2 MBP Model 9981 11.76 s 0.180521 pu 3.24628E-02 pu

The performance of different architectures of the 4.2. 75 -bus Indian system


ranking module having equal complexity is compared
in Table 2. It can be observed from Table 2, that the For 75-bus Indian system, 15 cases for single line-
four-layered ANN with tanh function in its first hidden outage were selected in contingency list. The modules of
layer and sigmoidal in second hidden layer (model B) is the CNN were trained to screen and rank these 15
the best model for contingency ranking. single line-outage contingencies. For voltage contin-
Though, the proposed cascade neural network gency screening and ranking, 1350 patterns showing 90
screens and ranks all the testing patterns correctly, the load scenarios were generated by varying the loads
test results corresponding to only one load scenario are randomly at all the buses in the range of 80120% of
presented in Tables 3 and 4, due to limited space. Table their base case values, utilising the pre-contingency data
3 shows the performance of the filter module which and the voltage performance indices corresponding to
screens contingencies, i.e. classifies the contingencies in 15 single line-outages. The input features as in case of
critical/non-critical class while Table 4 presents the IEEE 30-bus system along with the topology number
ranking of critical contingencies screened by the filter
module. Table 3
It can be observed from Table 3 that the filter Voltage PI classification for IEEE 30-bus systema
module screens all the critical cases correctly. Table 4
LO From bus To bus Actual Filter Class
shows that normalised values of desired voltage PI and
those obtained by the ANN (output voltage PI) are PI T O
very close to one another and the ranking order is also
1 13 7 0.100 0.90 0.90 NC
about the same. The line between bus 7 and bus 8, and
2 11 9 0.631 0.10 0.11 C
the line between bus 24 and bus 25 have been ranked as 3 28 10 0.100 0.90 0.83 NC
5 and 7, respectively, by the desired voltage PI ranking, 4 2 5 0.900 0.10 0.10 C
whereas the ranking module output has ranked them as 5 13 12 0.217 0.10 0.10 C
6 1 2 0.910 0.10 0.10 C
7 and 5, respectively. This minor change in ranking
7 1 27 0.100 0.90 0.90 NC
order may not be significant as long as these lines are 8 7 8 0.286 0.10 0.10 C
selected as critical lines. The RMS error for all the test 9 15 18 0.100 0.90 0.88 NC
patterns is 0.0325 pu for filter module and 0.0667 pu 10 8 20 0.165 0.10 0.10 C
11 24 25 0.186 0.10 0.10 C
for the ranking module. The trained cascade ANN is
12 25 10 0.475 0.10 0.10 C
able to screen all the critical contingencies correctly and 13 27 11 0.100 0.90 0.90 NC
at the same time, able to rank them according to their
a
severity except few cases, which are less critical. LO, Line-outage; T, desired output of filter module; O, actual
output of filter module; C, critical class; NC, non-critical class.

Table 2
Comparison of different neural networks for IEEE 30-bus system

S. No. ANN Model Training Testing

Iterations Total error Time (s) Max. abs. error (pu) RMS error (pu)

1 (10-10-1) Model 4000 5.31967 2.78 0.37534 0.14587


45 141 2.16987 31.59 0.25428 0.08022
2 (10-6-7-1) Model A 4000 4.21075 3.08 0.27589 0.09919
33 437 2.16981 25.21 0.23672 0.07114
3 (10-6-7-1) Model B 4000 4.11628 3.23 0.30332 0.09881
24 289 2.16973 19.12 0.18218 0.06671
S.N. Singh et al. / Electric Power Systems Research 53 (2000) 197205 203

Table 4
Voltage PI ranking for IEEE 30-bus system

LO From bus To bus Desired Output

PI(d) Ranking PI(o) Ranking

2 11 9 0.631 3 0.661 3
4 2 5 0.900 2 0.865 2
5 13 12 0.217 6 0.215 6
6 1 2 0.910 1 0.917 1
8 7 8 0.286 5 0.211 7
10 8 20 0.165 8 0.145 8
11 24 25 0.186 7 0.280 5
12 25 10 0.475 4 0.561 4

Table 5
Comparison of BP and modified BP algorithms for 75-bus Indian system

S. No. ANN Model Training Testing

Iterations Time (CPU) Max. abs. error RMS error

1 BP Model 79 095 123.87 s 0.24816 pu 7.15627E-02 pu


2 MBP Model 10 359 22.62 s 0.21832 pu 4.67246E-02 pu

(in the form of a string of four-bipolar digits) to Performance of BP (10-7-1) model and modified BP
represent the contingent line were used to train the (10-7-1) is compared in Table 5. As shown in Table 5,
CNN. the modified BP algorithm was found to be about 5.5
One thousand two hundred patterns corresponding times faster as compared to standard BP algorithm.
to 80 load scenarios, each with 15 cases of single line The performance of a four-layer neural network (10-6-
outage, were used to train the filter module (10-7-1) of 7-1) is compared with a three-layer neural network
the cascade neural network. The filter module was (10-10-1) for ranking purpose in Table 6. As was in
trained to filter out non-critical contingencies, while case of IEEE 30-bus system, the four-layered ANN
only critical contingencies were applied to train the
with tanh function in its first hidden layer (model B)
ranking module for their ranking.
gave best performance and so used for ranking.
Thus, out of 1200 training patterns, the screened
The cascade neural network performed equally well
critical contingency cases were 817 corresponding to 80
load scenarios. These 817 patterns were used to train for all the test patterns. Due to limited space, test
the ranking module (10-6-7-1). During testing phase, results corresponding to only one load scenario are
the 150 unseen patterns corresponding to ten load presented in Tables 7 and 8. It is found from Table 7
scenarios were tested for screening through the trained that the filter module screens all the critical contingen-
filter module. Out of 150 testing patterns, 86 critical cies correctly.
contingency cases, as screened correctly by the filter It can be observed from Table 8 that normalised
module were tested by the trained ranking module for values of voltage PI obtained by the ANN, PI(o) are
their ranking. very close to PI(d) and the ranking order is also about

Table 6
Comparison of different neural networks for contingency ranking in 75-bus Indian system

S. No. ANN Model Training Testing

Iterations Total error Time (s) Max. abs. error (pu) RMS error (pu)

1 (10-10-1) Model 10 000 10.64218 15.04 0.43615 0.13987


38 544 7.11121 57.17 0.26135 0.08354
2 (10-6-7-1) Model A 10 000 9.97335 16.92 0.30113 0.12686
30 600 7.11106 51.77 0.19003 0.07813
3 (10-6-7-1) Model B 10 000 9.28481 17.82 0.24940 0.10534
20 157 7.11103 35.34 0.17876 0.07531
204 S.N. Singh et al. / Electric Power Systems Research 53 (2000) 197205

Table 7 flow and considering only the voltage violated buses, it


PI Classification for 75-bus Indian system
gives accurate values of voltage PI which are used for
LO From bus To bus Actual Filter Class the training of the cascade neural network. Once the
training of the neural network is successfully accom-
PI T O plished, the estimation of voltage PI (thus ranking) for
unknown load patterns is almost instantaneous. In this
1 16 46 0.100 0.90 0.79 NC
paper, a cascade neural network (combining a filter
2 17 19 0.277 0.10 0.11 C
3 17 23 0.100 0.90 0.86 NC module and a ranking module) is developed for voltage
4 19 26 0.100 0.90 0.90 NC contingency screening and ranking. Test results of two
5 35 36 0.693 0.10 0.11 C sample systems reveal that
6 40 48 0.900 0.10 0.10 C In cascade neural network, the filter module screens
7 74 41 0.100 0.90 0.79 NC
the critical contingencies correctly and the trained
8 74 41 0.100 0.90 0.84 NC
9 26 41 0.187 0.10 0.27 C ranking module gives voltage PI close to the desired
10 48 49 0.183 0.10 0.12 C voltage PI values.
11 49 40 0.614 0.10 0.10 C The RMS error for all the test patterns is 0.0325 pu
12 41 42 0.100 0.90 0.87 NC for filter module and 0.0667 pu for ranking module
13 42 74 0.100 0.90 0.77 NC
in IEEE 30-bus system. For 75-bus Indian system,
14 23 74 0.100 0.90 0.82 NC
15 73 45 0.100 0.90 0.84 NC the RMS errors are 0.0467 pu and 0.0753 pu for
filter and ranking modules, respectively.
The trained cascade ANN is able to screen all the
the same. The line between bus 26 and bus 41, and the critical contingencies and at the same time able to
line between bus 48 and bus 49 have been ranked as 5 rank them correctly according to their severity, ex-
and 6, respectively, by the desired voltage PI ranking, cept few cases having marginal values of voltage PI.
whereas the network output has ranked them as 6 and The number of selected features for ANN training
5, respectively. This minor change in ranking order may were almost fixed and provided effective screening
not be significant as long as these lines are screened as and ranking for both the systems and can be used
critical lines. The RMS error for all the test patterns is for any system irrespective of its size and complexity.
0.0467 pu for filter module and 0.0753 pu for ranking The performance of the cascade ANN is superior to
module. The trained cascade ANN is able to screen all that of a single multi-layer feed-forward ANN.
the critical contingencies and at the same time able to The proposed method is accurate and very fast, so it
rank them correctly according to their severity, except can be implemented for on-line applications.
few cases having nearly equal values of voltage PI.

References
5. Conclusions
[1] G.C. Ejebe, B.F. Wollenberg, Automatic contingency selection,
IEEE Trans. Power App. Syst. 98 (1) (1979) 97 109.
The computation of voltage performance index by [2] F. Albuyeh, A. Bose, B. Heath, Reactive power considerations in
analytical methods takes more time, as some approxi- automatic contingency selection, IEEE Trans. Power App. Syst.
mate AC load flow analysis has to be carried out for 101 (1982) 107 112.
every contingency to calculate its value. These methods [3] M.G. Lauby, T.A. Mikolinnas, N.D. Reppen, Contingency se-
lection of branch outage causing voltage problems, IEEE Trans.
usually suffer from masking and misranking effects. On Power App. Syst. 102 (12) (1983) 3899 3904.
the other hand, in the proposed method, the voltage [4] K.L. Lo, M.A. Bismil, A comparison of voltage ranking meth-
performance index is calculated using the full AC load ods, Elec. Power Syst. Res. 16 (1989) 127 140.

Table 8
PI Ranking for 75-bus system Indian system

LO From bus To bus Desired Output

PI(d) Ranking PI(o) Ranking

2 17 19 0.277 4 0.278 4
5 35 36 0.693 2 0.793 2
6 40 48 0.900 1 0.925 1
9 26 41 0.187 5 0.161 6
10 48 49 0.183 6 0.162 5
11 49 40 0.614 3 0.636 3
S.N. Singh et al. / Electric Power Systems Research 53 (2000) 197205 205

[5] Y. Chen, A. Bose, Direct Ranking for voltage contingency cation, Proceedings of the IEEE International Symposium on
ranking, IEEE Trans. Power Syst. 4 (4) (1989) 13351344. Circuits and Systems (ISCAS), New Orleans, LA, USA, 1990,
[6] T.F. Halpin, R. Fischl, R. Fink, Analysis of contingency selec- pp. 2925 2928.
tion algorithms, IEEE Trans. Power App. Syst. 103 (5) (1984) [18] Vankayala, N.D. Rao, Power System security enhancement us-
938 945. ing a coupled ANN-ES scheme, Proceedings of the Fourth
[7] F. Schafer, J.F. Verstege, Adaptive procedure for masking effect Expert System Applications to Power System, Melbourne, 1993.
compensation in contingency selection algorithm, IEEE Trans. [19] S. Ghosh, B.H. Chowdhury, Design of an artificial neural net-
Power Syst. 5 (2) (1990) 539546. work for fast line flow contingency ranking, Elec. Power Energy
[8] I. Dabbaghchi, G. Irissari, AEP automatic contingency selector: Syst. 18 (5) (1996) 271 277.
branch outage impacts on load bus voltage profile, IEEE Trans. [20] Y. Mansour, A.Y. Chang, J. Tamby, E. Vaachedi, B.R. Corns,
Power Syst. 1 (2) (1986) 3745. M.A. El-Sharkawi, Large scale dynamic security screening and
[9] S.N. Singh, S.C. Srivastava, Improved contingency selection ranking using neural networks, IEEE Trans. Power Syst. 12 (2)
algorithm for voltage security analysis, J. Elec. Machines Power (1997) 954 960.
Syst. 26 (8) 1998 (in press). [21] Brandwajn, A.B.R. Kumar, A. Ipakchi, A. Bose, S.D. Kuo,
[10] S. Hao, A. Papalexopoulos, T.M. Peng, Distributed processing Severity indices for contingency screening in dynamic security
for contingency screening applications, IEEE Trans. Power Syst. assessment, IEEE Trans. Power Syst. 12 (2) 1997 11361142.
10 (2) (1995) 838 844. [22] R. Anand, K.G. Mehrotra, C.K. Mohan, S. Ranka, An im-
[11] L. Wehenkel, Contingency severity assessment for voltage secu- proved algorithm for neural network classification of imbalanced
rity using non-parametric regression techniques, IEEE Trans. training sets, IEEE Trans. Neural Networks 4 (6) (1993) 962
Power Syst. 11 (1) (1996) 101111. 969.
[12] G.C. Ejebe, G.D. Irisarri, S. Mokhtari, O. Obadina, P. [23] D. Srinivasan, A.C. Liew, C.S. Cheng, A neural network short-
Restanovic, J. Tong, Methods for contingency screening and term load forecaster, Elec Power Syst. Res. 28 (1994) 227234.
ranking for voltage stability analysis of power systems, IEEE [24] G. Singh, S.C. Srivastava, P.K. Kalra, D.M. Vinod Kumar, A
Trans. Power Syst. 11 (1) (1996) 350356. fast approach to artificial neural network training and its appli-
[13] V. Chadalavada, V. Vittal, G.C. Ejebe, G.D. Irisarri, J. Tong, G. cation to economic load dispatch, Elec. Machines Power Syst.
Pieper, M. McMullen, An on-line contingency filtering scheme (1995) 13 24.
for dynamic security assessment, IEEE Trans. Power Syst. 12 (1) [25] L.L. Freris, A.M. Sasson, Investigation of the load-flow prob-
(1997) 153 161. lem, Proc. IEE 115 (10) (1968) 1459 1470.
[14] D.J. Sobajic, Y.H. Pao, An artificial intelligence system for [26] S.N. Singh, S.C. Srivastava, Corrective action planning to
power system contingency screening, IEEE Trans. Power Syst. 3 achieve a feasible optimum power flow solution, IEE Proc. C 142
(2) (1988) 647 653. (6) (1995) 576 582.
[15] Y.Y. Hsu, H.C. Kuo, Fuzzy set based contingency ranking, [27] L. Srivastava, S.N. Singh, J. Sharma, Parallel self-organizing
IEEE Trans. Power Syst. 7 (3) (1992) 11891195. hierarchical neural network based fast voltage estimation, IEE
[16] R. Fischl, M. Kam, J.C. Chow, R. Ricciardi, Screening power Proc. C 145 (1) (1998) 98 104.
system contingencies using a back-propagation trained multi [28] L. Srivastava, S.N. Singh, J. Sharma, Estimation of loadability
perceptron, Proceedings of the IEEE International Symposium margin using parallel self-organizing hierarchical neural network,
on Circuits and Systems (ISCAS), Portland, OR, USA, 1989, pp. Int. J. Comp. Elec. Eng. (in press).
486 489. [29] L. Srivastava, S.N. Singh, J. Sharma, Comparison of feature
[17] J.C. Chow, R. Fischl, M. Kam, H.H. Yan, S. Ricciardi, An selection techniques for ANN based voltage estimation, J. Elec.
improved Hopfield model for power system contingency classifi- Power Syst. Res. (in press).

You might also like