Fifteenth National Power Systems Conference (NPSC), IIT Bombay, December 2008

Knowledge Based Approach for Transmission line Distance Relay Coordination

Ravikumar B., Thukaram D. and H. P. Khincha
Department of Electrical Engineering Indian Institute of Science Bangalore, India 560012 Email: {ravi, dtram, hpk}@ee.iisc.ernet.in
AbstractThe protection of transmission lines is traditionally performed on-line, new pattern recognition techniques should be implemented quickly and exibly. In this paper, a new approach to enhance the Transmission system distance relay co-ordination is presented. The approach depends on the apparent impedance loci seen by the distance relay during all possible disturbances. In a distance relay, the impedance loci seen at the relay location is obtained by extensive transient stability studies. Support Vector Machines (SVMs), a class of patterns classiers are used in discriminating zone settings (Zone-1, Zone-2 and Zone-3) using the signals to be used by the relay. Studies on a sample 9-bus are presented for illustrating the proposed scheme.

I. I NTRODUCTION System-wide disturbances become a major concern in the power industry again as a result of the Northeast blackout on Aug 14, 2003. The blackout affected an estimated 50 million people, and more than 61,800 MW of load was lost. The blackout had several causes or contributory factors in common with the earlier outages, including inadequate coordination of relays and other protective devices or systems [1]. Distance protection is widely used in the protection of EHV and UHV transmission lines in view of the fact that it can provide fast fault clearance and system co-ordination. It provides both primary and backup protection by its zonal settings correctly coordinated between distance relays. It is desired that a distance relay cover most of the line in its rst zone of protection. Also, it must not operate for faults beyond the remote bus even for the most unfavorable system conditions. The nominal reach of a non-pilot distance relay is set shorter than the protected line length to account for relay overreach caused by system uncertainties. The usual practice is to set the relay to cover 80% - 90% of the line length. Since system uncertainties such as variations in system parameters, load current, charging current, metering errors, etc. are not usually considered in distance relay setting, the relay may unexpectedly fail to operate for an internal fault or mal-operate for an external fault under certain system conditions. Since protection of transmission lines is traditionally performed on-line, new pattern recognition techniques should be implemented quickly and exibly. Application of Articial Neural Networks (ANNs) for transmission line protection is presented in literature [2-5]. ANN application as a pattern classier works well for transmission line protection, and

consequently results in the improvement of conventional digital relays. However, the algorithm used for adjusting the parameters of ANN does not produce robust classication. Recently, SVMs applications to transmission line protection are examined in the literature [6-8]. This paper presents an approach for estimating the coordinated zonal setting of distance relays using SVMs. The training patterns (feature vectors) consist of apparent impedance values observed at a relay following disturbance. In this paper, apparent impedance values are simulated by creating 3-phase faults. During the fault, the apparent impedance values are generated by using Transient Stability Program [9-10]. This program is extensively used for stability studies of several India Power Networks. The time delay settings (TDS) of zone2 and zone-3 have to be calculated such that all faults have to be cleared with in a maximum allowable time delay. II. P ROPOSED S CHEME Power system protection at the transmission level is based on distance relaying. The apparent impedance seen by the distance relay at substation on a transmission line connecting the nodes i and j, and having ow Pij + jQij is given as ZR = Pij Qij |Vi |2 +j 2 2 Pij + Q2 Pij + Q2 ij ij (1)

Where, all quantities refer to positive sequence values. Thus, the quadrant of ZR depends only on the direction of P and Q ows. Suppose a stable system at time t0 is subjected to fault on a transmission line at time t1 . Conventional relay algorithm will detect the fault, fault type, discriminate the zone settings based on preset reach values and TDSs and clear the fault at time t2 . During this period, the possible events can be opening of the faulted line, and/or some generators may fall in out-of synchronism and/or some load rejection can happen. Transient stability program [9-10] is used to obtain the apparent impedance trajectory seen by different relays of the system. This program is extensively used for stability studies of several India Power Networks. The program simulates the impedances seen during three time laps. First time lap T1 consists of impedances seen before occurrence of fault (T1 = t1 t0 Sec.). Second time lap T2 consists of impedances

397

Fifteenth National Power Systems Conference (NPSC), IIT Bombay, December 2008

seen during fault (T2 = t2 t1 Sec). After the fault has been cleared and we are observing the system till time t3, then the third time lap T3 consists of impedances seen during post-fault time (T3 = t3 t2 ). The feature vector consists of the apparent impedances seen by the relay located on a line during fault time T2 . Let a conventional fault detection algorithm detects the fault at time td (t1 <td <t2 ), from td to next one or two cycles (te) fault data information will be captured to form a feature vector. The size of the feature vector now will depend on the simulations time step during the gap te-td. In this paper, we are mainly concentrating on discriminating the zones based on the available knowledge obtained from the simulations which is an important aspect of protecting a transmission line. Fig. 1(a), shows the collection of apparent impedances values observed by the relay and Fig. 1(b), shows the training/testing of the SVM. Fig. 1(a) and Fig. 1(b) collectively show the implementation of the proposed method at a relay location. Support Vector Machines are used to capture the underlying concept/model between reach of different zones and the impedance swing trajectory characteristics. In this paper, this information is intelligently utilized for identifying the different zonal settings of a relay.
i j

by Vapnik et. al [11-12] and are described in more detail by in B. Sch lkopf et. al.[13-14] The roots of this approach, o the so-called support vector (SV) methods of constructing the optimal separating hyperplane for pattern recognition. The SV technique was generalized for nonlinear separating surfaces in [14], and it was further extended for constructing decision rules in the non separable case. The training task involves optimization of a convex cost function conveying to a technique without local minima. A. Support Vector Classication The problem of classication consists of estimating a function f : RN {1} using l i.i.d input-output training data (X1 , Y1 ), ..., (Xl , Yl ) RN x{1} from a data set D such that f classies correctly on unobserved data (x, y) (i.e., f (x) = y for examples (x, y) generated from the some underlying probability distribution P(x,y)). In other words, the loss function L can be dened by (2) L(yi , f (xi )) = |1 yi f (xi )|+ (2)

R1

Where |val|+ = max{0, val} val R A brief review of support vector classication (SVC) [15-18] is presented in this section; when data is linearly separable there exists a vector w RN and a scalar b R such that yi (w.xi + b) 1 for all patterns in the training set (i=1...l). The optimal hyper plane separates points lying on opposites classes yielding to the maximum margin separation. A separating hyper plane which generalizes well can be found by solving the following quadratic programming (QP) problem for (i=1...l): 1 w 2 2 Subjected to yi (w.xi + b) 1 Minimize

(3) solved by

Feature Vector
1(a)
(R, X) td

This constrained optimization constructing a Lagrangian p (w, b, ) = 1 2

l

problem

is

i=1

i (yi (w.xi + b) 1)

(4)

Relay Observations
(during Fault) (td< tk < te) (R, X) tk

SVM

Zone Discrimination
(Zone1 / Zone2 / Zone3)

(R, X) te

1(b)

Fig. 1: (a) Collection of features from relay input to generate data for
SVM; (b) Discrimination of feature data using SVM for identifying the zones.

The Lagrangian has to be minimized with respect to the primal variables w and b and maximized with respect to the dual variables i . The Karush-Kuhn-Tucker (KKT) conditions lead to nd the solution vector in terms of the training patterns l w = i=1 i yi xi for some i 0. Notice that i = 0 only for a subset of the training patterns, precisely those few vectors that lie on the margin, called the support vectors (SVs). Under certain conditions, a kernel function K(.,.) can be found such that K(xi , xj ) = xi .xj . An SVM uses then the convolution of the scalar product to build, in input space, the nonlinear decision function
l

III. B RIEF R EVIEW ON S UPPORT V ECTOR M ACHINES Support Vector Machines are a new learning-by-example paradigm spanning a broad range of classication, regression, and density estimation problems. They were rst introduced

Feature Vector

f (x) = sgn(
i=1

i yi K(xi , xj ) + b)

(5)

Where b is found from the primal constraints and is computed by i (yi (w.xi + b) 1) = 0, i = 1...l such that i = 0. When the training data is not linearly separable, a separating

398

Fifteenth National Power Systems Conference (NPSC), IIT Bombay, December 2008

hyper plane does not exist. Besides, when real data sets are used, SVMs can t noise and outliers leading to poor generalization. Thus, a hard margin classier is no longer adequate. Introducing a soft margin, the learning task is essentially the same as indicated in (3) except for the introduction of the penalty term C and the slack variable . The classier tries then to separate the data by minimizing the objective function Minimize 1 2 w
2

C l

i
i=1

Subjected to yi (w.xi + b) 1 i , 0 i C/l, i 0

(6)

for i=1...l. In this sense, it acts by controlling the classier capacity and the number of training errors. In other words, l the task is now to minimize the sum of errors i=1 i in 2 addition to w . Again this optimization problem can be transformed into a QP problem. The value of C can be found by experimentation in a validation set and cannot be determined from either the model or the data set. In this paper one-against-one method is used for the extension from the binary two-class problem to n classes for multiclass classication, because of its less training time over oneagainst-all. B. Kernel Choice The use of kernel methods [19] provides a powerful way of obtaining nonlinear algorithms capable of handling nonseparable data sets in the original input space. Different types of kernels used for training the SVMs are Linear Kernel Qij = K(xi , xj ) = xT .xj , Polynomial Kernel Qij = i ((xi .xj ) + r)Degree , Radial Basis Function (RBF) kernel Qij = exp( xi xj 2 ), where related with the kernel width and Sigmoidal Kernel Qij = tanh((xi .xj ) + r) where and r are kernel parameters. The RBF kernel has less numerical difculties. Moreover, we must note that the sigmoid kernel is not valid (i.e. not the inner product of two vectors) under some parameters [20]. In this paper we suggest that RBF kernel is a reasonable rst choice for SVM training. C. SVM Model Selection: In any predictive learning task, such as classication, an appropriate representation of examples as well as the model and parameter estimation method should be selected to obtain a high level of performance of the learning machine. Under the SVMs approach, the usually parameters to be chosen are the following: 1) The penalty term C which determines the trade-off between the complexity of the decision function and the number of training examples misclassied; and 2) Kernel function parameters IV. S YSTEM S TUDIES AND R ESULTS The single line diagram of the system considered for illustrating the proposed scheme is shown in Fig. 2. The system has 3 generators, 3 transformers, 8 transmission lines and

loads are presented at busses 5, 6 and 7. Transient Stability program [9-10] is used to simulate the 3-ph fault on the system. The transient stability program is developed by considereing IEEE standard models for representing the Turbines[2122], Generators[23-24], Loads[25-26], and Excitation control actions [27-28]. Unlike the rules developed for identifying the zone-1, zone-2 and zone-3 settings of the conventional algorithm, the proposed approach classies the zones based on the connectivity of the lines. Zone-1/Class-1 represents the primary protection of its own line on which the relay of interest is located. Zone-2/Class-2 represents the lines that are connected to the receiving end of the primary line. The Zone-3/Class-3 represents the lines that are next adjacent to primary line. In this paper, the relay R13 at bus 6, mounted for primarily protecting the line 6-9 is considered for applying the proposed scheme. This relay will be acting as a zone-2 back up for 9-8 line, 3-9 transformer and in its zone-3 backup, the lines 8-7(1) and 8-7(2) are covered.
1 4
R 15 R2 R3

5 R 1

R4 R5 R6

R 16

R7

R 14

R 9 R 10

R 8

R 13

R 12

R 11

Fig. 2: Single line diagram of the sample system considered.

TABLE I: G ENERATION OF T RAINING AND TEST PATTERNS

Line Zone/Class label Fault distance from relay location Training At distances of 10%,20%...90% of their total line length(Total 9) Testing At distances of25%,45%,65% and 85% of their total line length(Total 4) Faultresistance values varied over 0, 3, 7, 15, 30 and 60

6-9

9-8 8-7(1) 8-7(2)

2 3 Faultresistance values varied over 0, 2, 5, 10, 20, 25, 40, 45, 50, 55, 65 and 75

To illustrate the algorithm, only 3-phase faults are considered in this paper. A normal operating system at time 0 seconds is subjected to 3-Ph fault on line time 0.1Sec and cleared the fault by isolating the line from operation at time 0.2 seconds.

399

Fifteenth National Power Systems Conference (NPSC), IIT Bombay, December 2008

And the time of simulation is taken up to 1.0Sec after the fault is cleared. In this case, the time laps T1 is 0.1Sec, T2= (0.2-0.1) =0.1Sec and T3= (1.0-0.2) =0.8Sec. Fig.3 shows the observations made by Relays R2 , R14 , R13 and R10 during T1 , T2 and T3 for 3-ph faults on line 6-9 at 50% away from bus6. Fig. 4 shows the observations by relays at various locations of the system for 3-phase faults simulated at various locations on the lines 6-9. Fig. 5 shows the apparent impedance movement observed by the relay R13 located at 6 (on lines 6-9) during 3-phase fault at various locations of the system. As such, Relay observations for different fault conditions are different. Also Fig. 3 to Fig. 5 contains the typical conventional zone-1, zone2 and zone-3 reach settings of distance step protection modeled by mho characteristics [29-30]. Fig. 6 shows the observations made by relay R13 for variation in fault resistance value over 0, 10, 20, 25, 40, 45 and 50 and fault location varied over 10% to 90% in 10% distance variations. Fig.6 indicates that the fault resistance pull the loci away from the origin, which affect the zone-1 settings of a conventional relay. From the gures 3, 4 5 and 6, it is clearly observed that, a knowledgebase can be developed based on the data available for various types of faults, fault location, fault resistance and fault types. In this paper, we are concentrating only on 3-phase faults for generating the feature vectors, which can be further, extended to various other types of faults. A. SVM Training and Testing Training patterns are generated by simulating the 3-ph faults at different locations on the transmission lines. The faults are created at distances of 10% to 90% insteps of 10% of their overall transmission length. During the simulation, the fault resistance values are varied between the values 0 to 75. Over all, the training patterns are generated 4 transmission lines over 9 locations with varying 12 impedance values. The number of training patterns is 5 X 9 X 12 = 432 patterns; The input patterns training and test patterns are normalized to [-1, +1] before inputting to the SVMs. For the normal scaling method, if the maximum and minimum values of the ith attributes are Mi and mi respectively, then, scaling to [-1, +1] means x1=2(x-mi)/(Mi-mi)-1. Test patterns are generated by creating at 25%, 45%, 65% and 85% distances of their overall transmission line length. Fault resistance values are varied over the values 0, 3, 7, 15, 30 and 60. The number of test patterns is 4 lines X 6 fault resistances X 4 locations = 96 patterns per each fault type. The details of simulations carried for generating the training and test patterns are given in Table I. B. SVM classiers SVMs training require selection of the cost function (C) and kernel function parameters, which inuence the ensuring model performance. In our simulations, we have considered radial basis function (RBF) as kernel function. RBF kernel is advantageous in complex non-separable classication problems due to its ability of nonlinear input mapping. So selecting a good RBF kernel required to select the

0.5
R2

T1

0.5
X (p.u)

1
T3

1.5

T2

2 1.2

0.8

0.6

R (p.u) (a)

0.4

0.2

0.2

0.4

(a)
0.2 0.1 0
T1

0.1
T2

0.2
X (p.u)

0.3 0.4 0.5 0.6 0.7 0.8 1.4 1.2 1 0.8 0.6 0.4 0.2 0 0.2
T3

R (p.u) (b)

(b)
0.5
R13

0
T2 T1

0.5

1
X (p.u)

1.5

2
T3

2.5

3 0.9

0.8

0.7

0.6

0.5

R (p.u) (c)

0.4

0.3

0.2

0.1

0.1

(c)
0.2
R10 T2

0.2
T3

X (p.u)

0.4

0.6

0.8
T1

1.2 1.2

0.8

0.6

R (p.u) (d)

0.4

0.2

0.2

(d)

Fig. 3: (a d) : Observations made by Relays R2 , R14 , R13 and R10 during Times T1 , T2 and T3 for 3-ph faults on line 6-9 at 50% away from bus6.

400

Fifteenth National Power Systems Conference (NPSC), IIT Bombay, December 2008

1.5

0.5

0.5

R15 R16 R2 R1 R14 R13 R4 R6 R9 R10 R11 R12

1.5 1

0.8

0.6

0.4

0.2

0 R (p.u)

0.2

0.4

0.6

0.8

Fig. 4: Observations made by Relays at various locations on the

system for 3-ph faults on line 6-9 at various location.

0.5 0.4 0.3 0.2 0.1 X (p.u) 0 0.1 0.2 0.3 0.4 0.5 1 45 46 57 (1) 69 78 (1) 89

(= 1/2 2 , : kernelwidth) is important parameter to be chosen. In this paper, LIBSVM [31-32] is used for training the support vector machines in classication and regression modules. In a rst series of experiments we run the classier with several values of C and somehow trying to guess which combination of parameters might be the best for a good model. Larger C corresponds to less number of SVs as well as higher testing accuracy although over-tting cannot thus be avoided. Further explanation is required for these results taking into account both C and parameters. For non-separable data, the penalty term C is able to reduce the training errors in the working data set. Therefore, the margin is an indicator of the generalization accuracy. In the absence of a method to compute the best tradeoff between the regularization term and the training errors, the balance sought by the SVMs technique is hard to nd. Thus, a larger C corresponds to assign a higher penalty of training errors and clearly over-tting occurs. On the other hand, when the kernel parameter becomes higher, the greater the variety of the decision boundaries that can be formed, originate a more complex model. The added exibility decreases initially the generalization error as the model can better t the data.

X (p.u)

0.8

0.6

0.4 R (p.u)

0.2

0.2

Fig. 5: Impedance characteristics observed by relay R13 (at bus 6 of 6-9 line) for 3-ph faults on various lines with varying distances .

0.2
0ohms 10ohms

Fig. 7: SVM parameter selection using interactive grid search. Choosing the best parameters can be timing consuming if a systematic approach is not used and/or the problem knowledge do not aid for proper selection. Therefore, an interactive grid search model selection has been accomplished for SVM and the generalized accuracy evaluated on the train data. Fig. 7 portrays the generalization graphic contours for the SVM after a ve-fold cross validation, thus, reducing the search space. The efcient heuristic way of searching points in that space with small generalization errors will lead to a good understanding of the hyper parameter space. We can then do a rened search of the (C, ) pairs for proper model selection. Fig. 7 shows the parameter selection using interactive grid search for classier SVM. The range of C chosen for training is [20 , 220 ] and values ranges from [27 , 27 ] are chosen.

0.1

20ohms 25ohms 40ohms

45ohms 50ohms

(p.u) X

0.1

0.2

0.3

0.4

0.5 0.5

0.5

R (p.u)

1.5

Fig. 6: Relay R13 observations for faults on the line 6-9 with varying fault locations and fault resistance values that are used for training in Table I.

401

Fifteenth National Power Systems Conference (NPSC), IIT Bombay, December 2008

[10] D Thukaram,Technical Manual on Power System Analysis Department of Electrical Engg.., Indian Institute of Science, Bangalore, India. [11] V. N. Vapnik, The Nature of Statistical Learning Theory, New York: Springer-Verlag, 1995. [12] Vapnik, V., Statistical Learning Theory. Wiley, New York, NY, 1998. [13] Smola and B. Scholkopf. A Tutorial on Support Vector Regression, Neurocolt Tech. Rep. NV2-TR-1998-030. [14] Sch lkopf, C. Burges, and A. Smola, Advances in Kernel Methodso Support Vector Learning Cambridge, MA: MIT Press, 1999. [15] Sastry, P.S., An introduction to Support Vector Machine, Chapter from computing and information sciences: Recent Trends, Narosa Publishing House, New Delhi 2003. [16] Burges, C. J. C, A tutorial on support vector machines for pattern recognition, Data Mining and Knowledge Discovery, vol. 2, no. 2, pp. 955-974, 1998. [17] Joachims, T, Making large-scale SVM learning practical. Advances in Kernel Methods-Support Vector Learning, Cambridge, MA, 1998. MIT Press. [18] Keerthi, S. S., and Lin, C.J., Asymptotic behaviors of support vector machines with Gaussian kernel, Neural Computation 15 (7), pp: 16671689, 2003. [19] Platt, J., Fast Training of Support Vector Machines using Sequential Minimal Optimization, Advance in Kernel Methods: Support Vector Learning, pp. 185-208, MIT Press, Cambridge, MA, 1999. [20] C. Campbell, Kernel methods: A survey of current techniques, Neurocomput., vol. 48, pp. 63-84, 2002. V. C ONCLUSION [21] IEEE Committee Report, Dynamic Models for Steam and Hydro Turbines in Power System Studies, IEEE Trans. Vol PAS-92, pp 1904The use of Support Vector Machines as a powerful tool for 1915, Nov/Dec 1973. classication and its application to distance relay coordination [22] IEEE Working Group Report, Hydraulic Turbines and Turbine Control is presented. Support Vector Machines with RBF kernel is able Models for System Dynamic Studies, IEEE Trans. Vol. PWRS-7, No.1, pp. 167-179, Feb. 1992. to learn the underlying concepts between the reach settings [23] IEEE Committee Rep., Current usage & suggested practices in power of different zones of protection and the apparent impedance system stability simulations for synchronous machines, IEEE Trans. movement during the faulted condition. From the results, Energy Conversion, vol. EC-1, pp. 77-93, Mar. 1986. we conclude that an intelligent system can be useful for [24] K. R. Padiyar, Power System Dynamics Stability and Control, 2nd ed. Hyderabad, India: BS Publications, 2002. discriminating the zonal settings for backup protection of [25] IEEE Task Force Report, Load Representation for Dynamic Perfortransmission lines in an efcient way for varying conditions mance Analysis, Paper 92 WM126-3 PWRD, Presented at the IEEE PES Winter Meeting, New York, Jan 26-30, 1992. of the power systems. This can further extended to topological [26] C. Concordia and S. Ihara, Load Representation in Power System changes of the power network for adaptive control of the Stability Studies, IEEE Trans. Vol. PAS-101, pp 969-977, April 1982. transmission system. [27] IEEE Committee Report, Excitation System Dynamic Characteristics, IEEE Trans. Vol. PAS-92, pp. 64-75, Jan/Feb. 1973. [28] IEEE Committee Report, Excitation System Models for Power System R EFERENCES Stability Studies, IEEE Trans. Vol. PAS-100, pp. 494-509, Feb.1981. [1] Online Available at:,ftp://www.nerc.com/pub/sys/all updl/docs/blackout/ch7- [29] Damborg, M.J.; Ramaswami, R.; Venkata, S.S.; Postforoosh, J.M.;, Computer Aided Transmission Protection System Design Part I: Al10.pdf corithms, IEEE Transactions on Power Apparatus and Systems, Vol [2] T. S. Sidhu, H. Singh, and M.S. Sachdev, Design,implementation and PAS-103, Issue 1,pp:51-59, Jan. 1984. testing of an articial neural network based fault direction discrimination for protecting transmission lines, IEEE Trans. on Power Delivery, vol. 10, [30] Ramaswami, R.; Venkata, S.S.; Damborg, M.J.; Postforoosh, J.M.; Computer Aided Transmission Protection System Design Part II: Imno. 2, pp. 697-706, 1995. plementation and Results, IEEE Transactions on Power Apparatus and [3] D. V. Coury and D. C. Jonge, Articial neural network approach to Systems, Vol. PAS-103, Issue 1, pp:60-65, Jan. 1984. distance protection, IEEE Trans. on Power Delivery, vol. 13, no. 1, pp. [31] Chih-Chung Chang and Chih-Jen Lin, LIBSVM: a library 102-108, 1998. for support vector machines, 2001., Software available at [4] S. K. Chakrabarty, C. V. Nayar, and N. R. Achuthan, Applying pattern http://www.csie.ntu.edu.tw/cjlin/libsvm. recognition in distance relaying-Part-2, IEE Proc. Gen. Trans. Disbn., [32] Online available at: http://www.csie.ntu.edu.tw/cjlin/papers/libsvm.pdf. vol. 139, no. 4, pp. 306-314, 1992. [5] P. K. Dash, A. K. Pradhan, and G. Panda, Application of Minimal Radial Basis Function Neural Network to Distance Protection, Transactions on Power Delivery, IEEE, Vol 16, No 1, pp: 68-74, Jan. 2001. [6] Thukaram, D., Khincha, H. P., and Vijaynarasimha, H. P.: Articial Neural Network and Support Vector Machine Approach for Locating Faults in Radial Distribution Systems, Transactions on Power Delivery, IEEE, 20, (2), April 2005, pp. 710-721 [7] Salat, R., Osowski, S.: Accurate fault location in the power transmission line using support vector machine approach, Power Systems, IEEE Transactions on, Vol. 19 (2),May 2004, pp:979-986 [8] Moulin, L. S., da Silva, A. P. A., El-Sharkawi, M. A., and Marks II, R. J.: Neural networks and support vector machines applied to power systems transient stability analysis, Int. J. Eng. Intell. Syst., vol. 9, no. 4, pp. 205-211, Dec. 2001. [9] D Thukaram,Transient stability program developed at Department of Electrical Engg.., Indian Institute of Science, Bangalore, India.

The highest cross validation accuracy resulted for SVM is 97.22% on the training data with extracted model parameters of C=262144 and = 0.125. Once the SVM is learned with these parameters, all parameters of the trained SVM have been frozen and then used in retrieval mode for testing the capabilities of the system on the data not used in learning. The test data samples have been extracted using the transient stability program as explained earlier. The %testing accuracy is dened by No. of samples correctly classied*100 / total number of samples presented. The obtained model parameters during grid search are merit listed for selecting the best parameters with highest testing accuracy. The parameters C= 262144 and =0.08839 has resulted highest test accuracy of 97.91% with 94 patterns out of 96 test patterns are correctly classied. The results obtained choosing the pair of parameters conveys to the learning model with the smallest capacity and, thus, the highest generalization.

402