www.ietdl.

org
Published in IET Generation, Transmission & Distribution Received on 5th November 2011 Revised on 6th May 2012 doi: 10.1049/iet-gtd.2011.0757

ISSN 1751-8687

Improvement in the performance of neural network-based power transmission line fault classiers
S. Seyedtabaii
Electrical Eng. Dept, Shahed University, Tehran, P.O. Box 18155 159, Iran E-mail: stabaii@shahed.ac.ir

Abstract: A power line expert can easily pinpoint the type of fault that may have been occurred in a power transmission line. Transferring the experts intelligence to an articial neural network (NN) makes the classication process fast and available online. Often the phase currents are used as NN inputs for this purpose. Lack of a somehow one-to-one relationship between the type of fault and the phases faulty currents prohibits the underlying network from being adequately trained. In a search for nding a type of feature that establishes a relatively unique link between the type of faults and the phase currents, it is noticed and mathematically proved that the ratios of the phase current jumps enjoy such a valuable advantage to be a prime choice as NN inputs. The inputs let a multi-layer perceptron (MLP) NN with about one node per phase to identify the faults accurately. The scheme works well in the presence of a various number of fault items. The superiority of the method is well realised when it is compared with the results of similar investigations using wavelet, fuzzy and others. The reference data are generated using MATLAB Power System Toolbox. The test samples are more general than those previously used in other investigations.

Introduction

In an electric power system comprising different complex interacting elements, there is always a possibility of disturbance and fault. Lightning, dirt/salt on insulators, line line ashover because of wind, ashover to tree, tower/pole or conductor falls, objects fall on conductors are the source of faults which arc, re, explosion and vibration are their consequences. The transmission line faults are classied into balanced and unbalanced ones. The balanced faults are three phase (3L) and three-phase to ground (3LG) short circuits. Most of transmission line faults are of unbalanced nature. Singleline to ground (LG), line-to-line (2L) and double-line to ground (2LG) faults are categorised as unbalanced faults. High-impedance fault is often an unbalanced type. Since it is important to protect the power system from unavoidable fault aftermath, power line experts are employed. They can easily identify the fault type; but, they are not adequately fast and may not be available online. Transferring the intelligence of the experts to an intelligent machine or an articial neural network (NN) could provide excellent solution. Fault diagnosis involves three major tasks: detection, classication and locating. The automatic protection system must be (i) highly reliable and very sensitive to isolate targeted-fault (ii) adequately selective and (iii) fast. Often the three-phase current amplitude jumps, because of a fault, are used for classication purpose and ground
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 8, pp. 731 737 doi: 10.1049/iet-gtd.2011.0757

involvement is checked through the analysis of iz(n) (instantaneous zero current) and or a separate index derived from the phase currents. The signal feature used may be root-mean-square (rms) value [1], power of fundamental frequency (using discrete Fourier transform (DFT)), transient energy (calculated using wavelet [2] and or DFT) and even exact sample points [3, 4]. During years, different researchers have developed a number of fault classication techniques. Articial NNs, as a powerful pattern recognition tool [5], could be applied to the line currents to rstly learn and secondly identify faulty phases quickly and automatically [6, 7]. Samantaray et al. [8] preprocesses the three phase currents using hyperbolic S transform and uses the difference between the pre- and post-fault energy for classication. The decision maker is a radial basis function (RBF)NN. Feeding an adaptive network and fuzzy inference (ANFIS) system by a set of iz(n) and the pre- and post-fault differences of the rms value of the three phase currents has been investigated by Yeoa et al. [1]. Mahanty and Gupta [3] employ the post-fault samples of the three phase currents, plus fuzzy strategy for fault identication. Mahanty and Gupta [4] also use samples of the three-phase currents and iz(n) as RBF inputs. Multi-layer perceptron (MLP) [6, 7, 9, 10], RBF [4] and probabilistic NN [11] performance have been examined by researchers. Applications of wavelet [2, 12] and waveletNN [11, 13] have also been investigated. Study of fuzzy logic for fault distinction has been reported in [3, 12, 14]. Almost all the methods give acceptable success rate under
731

& The Institution of Engineering and Technology 2012

www.ietdl.org
their specic test conditions. However, the reason behind their arrangement is not explicitly clear other than the performance of the algorithms. Initially, it seems that the power line fault classication is a simple task. Nevertheless, this does not come true. Consequence of a fault is phase current jump. The amplitude of the jump greatly depends on factors such as fault resistance and location. This means that a diverse set of amplitude jump values may be the result of occurrence of a certain type of fault. Adversely, in an NN, each class is required to be dened by the value of inputs while just a limited uncertainty is also allowed. Therefore an NN, as it is, is not appropriate for power line fault classication. In response, four types of solutions have been introduced. Resorting to a heuristic max level for NN inputs [15], shifting to use RBF [4], relying on just discrete wavelet transform (DWT) detail coefcients [16] and NN input preparation using fuzzy methods are among them. These are more elaborated in Section 5 where the unique alternative method is also presented. In this paper, MLP inputs are refurnished for the classication of the faults. A unique phase current manipulation policy based on the ratios of the three phase current jumps is discussed and implemented here that provides perfect NN classication success rate under various test situations. This is accomplished without any classication data compromise and involvement of heuristic parameters. The immediate advantage of the scheme is decrease in the size of the network to the limit of about one node per phase and increase in the network training speed. This policy is tested in the presence of considerable amount of faulty input patterns. Fault conditions that are adopted here have more complex consequences than those already studied. In Section 2, the test power system and fault consequences are discussed. Input signal (probable) denoising and feature extraction are elaborated in Section 3. Section 4 details fault detection. The main fault classication discussion comes in Section 5 and, lastly, the conclusion is presented in Section 6.
Table 1
The test power transmission line model parameters R (V/km) zero seq. positive seq. 0.2296 0.0114 L (mH/km) 2.9 0.9 C (mF/km) 0.00859 0.01343

Fig. 2 Each pattern shows the faulty phase current jump under various test conditions (fault resistance, phase angle and so on) at a single location

2.1

Data set

Line currents of 1260 faulty cases were collected using MATLAB-SIMULINK Power System Toolbox under the following conditions: Fault location: at steps of 20 km. Fault resistances: Rf 0.1, 5, 10, 50 and 200 V. Load angle: u 2308, 258, 08, 58, 108, 208 and 308. Fault inception angle: 08, 908, 1408 and 1608. This collection containing the three phase instantaneous currents, ia(n), ib(n) and ic(n) are called the set, . To exhibit a clear picture of the current amplitude jump because of a fault, Fig. 2 is presented. This gure shows the change in the phase A current because of a phase A to ground (AG) fault. Under various conditions, there exist different amounts of jump in the phase current amplitude that all must be interpreted as an AG type of fault in spite of their remarkable diverse values.

2 Power system model and fault consequences

For test, a 230 kV, 200 km power transmission line is simulated under various types of faults with different system conditions and parameters using MATLAB Power System Toolbox. The one-line diagram of the studied system is shown in Fig. 1. The transmission line is terminated in two sources of 230 MVA each at both ends of the line. The load angle (the second source phase angle) is u. R and L of the both sources are 0.529 V and 16.8 mH. The line is assumed to be symmetric (transposed) and modelled by its impedance sequences as given in Table 1. The sampling rate is considered 20 samples per cycle. No extra amplitude normalisation is adopted. Currents are just measured at source 1.

Signal feature extraction

In the case of one terminal fault analysis, the three-phase voltages and currents are measured and available. The instantaneous zero current, iz(n), is calculated as follows 1 iz (n) = (ia (n) + ib (n) + ic (n)) 3 (1)

where ia(n), ib(n) and ic(n) are the instantaneous phase currents. iz(n) is a signal that provides a good exhibition of LG and 2LG type of faults; however, it comes short in representing the 2L, 3L and 3LG ones. Thus, the current jump between the pre- and post-fault cycles, id(n) is calculated for each phase current. The current jump for phase A is written as follows ida (n) = ia (n) ia (n (N 1)) (2)

Fig. 1 Power system model

732 & The Institution of Engineering and Technology 2012

where ia(n) is the instantaneous phase A current and N is the number of samples per cycle. This signal adequately projects the faults that can be detected by examining iz(n). The
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 8, pp. 731 737 doi: 10.1049/iet-gtd.2011.0757

www.ietdl.org
current jump signal behaves so much the same as iz(n), since it is zero in the pre-fault times and show jump in the post-fault occasions. The rms of the post-fault current jump of the three phases, idp and zero current are calculated as follows i dp = 1 M 1 M
n0 +M 1

i2p (n), d
n=n0 n0 +M 1

p [ {a, b, c} (3)

iz =

i2 (n) z
n=n0

where n0 is the fault initial time and M is the number of available post-fault samples. The protection system is required to respond accurately and appropriately to a fault in a minimum time. Therefore post-fault available data are necessarily limited. In case that post-fault samples, M is less than N, the samples by replication are extended to reach N, M N, in a manner that full sinus to be formed. If decision has to be made in a one-cycle time, then M is 20 and no replication of data is required.

In spite of these various situations, iz(n) can adequately represent occurrence of LG and 2LG type of faults, no matter how the line current varies. In the case of 2L, 3L and 3LG fault types, negative, positive or no change in the current level may also be observed. Here, iz(n) no longer can reveal the fault since its value remains zero. In substitute, current jump idp(n) (2) of one of the phases could be employed to bridge the gap. In most situations, faulty signal is very strong and is immune to noise and interference. However, there are also circumstances that weak post-fault signal may be degraded by noise (low-value parts of the graph of Fig. 2). Thus, for preventing false alarm, resulting from measurement noise and line oscillation, the smoothed alternative signals are preferred. The smoothing is performed at each time index isz (n) = isz (n 1) + (|iz (n)| |iz (n N/2)|) (4)

where isz(n) is the smoothed version of the jump in iz(n). Then, its rms is calculated for an amount of time (half a period) considered for the fault detection decision making 2 N
n(N /2)+1

ifz (n) =

i2 (k) sz
k=n

(5)

Fault detection

Our aim is to detect fault and classify it in less than a half of a cycle. Fig. 3 exhibits some typical type of fault consequences. Huge increase in current, no change in current and even decrease in phase current amplitude are exhibited. With the initiation of a fault, line oscillation is also observed. Naturally, it is believed that fault coincides with increase in the phase current. However, decrease in the current was recorded at positive load angles, u, when the fault resistance exceeds a certain value. Fault at 160 km with Rf 30 V and +208 load angle ensues no current amplitude change situation and decrease in the current is exhibited under Rf 30 and u +308. The positive load angle, the source 2 leading the source 1, is reected in the line voltage current phase difference. In the case of positive u, the phase difference is around 1808, meaning that power factor is negative and the direction of the average power is towards the source.

where N is the number of samples per cycle. With similar calculations ifda(n) is derived from ida(n). For the conrmation of fault, a threshold level has to be set. NN provides a robust level of setting for fault/normal discrimination. For the given samples , ifz(n) and ifda(n) are calculated at n n0 + N/2 2 1 where n0 is the fault initiation time. Fault detection is conducted using a perceptron NN of just one node [5]. A vector of 2 19 656 elements, containing ifz(n0 + N/2 2 1) and ifda(n0 + N/2 2 1), forms the input of the NN, and a vector of 1 19 656 containing 0s (normal) and 1s (faulty) is the target vector. After the network training, the test results have been shown in Fig. 4, which indicates perfect success rate, 9828 normal and 9828 faulty cases all have been classied accurately. It took about 35 epoch iterations to fully converge. The number of training iterations varies from execution to execution as the initial condition of the network is randomly selected. No normalisation is required since network neurons non-linear function inherently poses such a capacity. Log-sigmoid (logsig) non-linear function in the hidden layer of the NN outperforms hyperbolic tangent sigmoid (tansig) function since it ts more to the type of data that is used here for the fault detection. The output layer non-linear function is set to tansig.

Fig. 3 An LG fault in a three-phase system; no change (top-left) and increase (top-right) in the faulty phase current. Decrease in the faulty phase current (bottom-left) and its corresponding iz(n) (bottom-right)
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 8, pp. 731 737 doi: 10.1049/iet-gtd.2011.0757

Fig. 4 Success rate of the proposed fault detection scheme

733

& The Institution of Engineering and Technology 2012

www.ietdl.org
Table 2
output Fault type lines involved no. fault AG BG CG AB AC BC ABG ACG BCG ABC, ABCG 1 0 0 0 0 0 0 0 0 0 0 A 0 1 0 0 1 1 0 1 1 0 1 5 outputs B 0 0 1 0 1 0 1 1 0 1 1 C 0 0 0 1 0 1 1 0 1 1 1 G 0 1 1 1 0 0 0 1 1 1 0 Five-bit fault denition codes using as the network

Fig. 5 Left: non-zero post-fault iz (t) during a 3LG fault in an unsymmetrical power line (the post-fault value for a 3L type of fault is zero) Right: phase current in a 3LG fault

Our conclusion to this section is that the pre-processed ifz(n) and ifda(n) are the sufcient indicator of power transmission line fault. In addition, a perceptron NN classier with just one node in the hidden layer can detect fault easily and accurately.

which it later can be subtracted from the power of the post-fault signal for denoising purposes. Alternatively, wavelet or DFT may be used for the similar action. When the noise is ignored, the outputs of the three methods are the same and the simpler, rms, is the choice.

Fault classication

5.2

NN-based fault classication

Once the fault detection has been accomplished, its classication is initiated. Each fault is dened by a 5 bit code shown in Table 2. From the table, the code for the normal situation is 10 000 and for the AG fault is 01001. As long as the line is modelled transposed and the zeropositive sequences represent its parameters, no difference between 3L and 3LG faults can be noted in the post-fault time. Nevertheless, when the fault is removed, phase currents behave remarkably different in the two cases that may be used for differentiating between them. However, when a non-transposed model such as R= C L= = 83.4 73.2 71.6 1.6 0.7 0.6 11.66 2.13 0.58 73.2 85.7 73.2 , 0.7 1.6 0.7 , 2.13 12.12 2.13 71.6 73.2 83.4 0.6 0.7 1.6 0.58 2.13 11.66 mH/km nF/km mV/km is assigned to the line that takes into account unsymmetrical conguration of conductors against the earth, earth wire and each other. Now, iz could be used for 3L and 3LG fault separation. In a non-transposed model, the post-fault iz is zero for a 3L fault while is non-zero for a 3LG fault as it has been depicted in Fig. 5. 5.1 Phase current signal features

Consequence of a fault is phase current jump. The amplitude of the jump greatly depends on the factors such as fault resistance and fault location. Meaning that, a specic type of fault is represented by a diverse set of current jump amplitudes. From Fig. 2, it is evident that the range of the current jump for even one class of fault is very wide. In most NNs, each class is required to be dened by its input values while allowing a limited uncertainty. However, in this case the uncertainty is wide and an NN, as it is, is not capable of power line fault classication. To circumvent the difculty several remedies have been implicitly introduced in the literatures. 1. In some papers [15], a heuristic max level is set for the NN inputs to alleviate the confusion emerging from having various current amplitude jumps for just one type of fault. Those used MLP and the rms currents, as NN inputs, have followed this scheme. A network of at least 12 nodes, when a max level of 1.5 was imposed to the set , correctly classied the faults. However, the network performance sensitivity to the choice of max level remains problematic. 2. Other researchers have tried to shift to RBF [4], a network that by adding extra nodes can cover a wide range of input space, compatible with what fault classication data are. This has its own costs, unnecessary number of nodes and obligation in providing sufcient training data to cover all sections of the input space. 3. Those used wavelet transform and ignored high-energy approximation coefcient and banked on just the lowenergy detail coefcients, simply are trying to escape the trap and to avoid the wide span of the input data by disregarding the strong component of fault consequences [16]. The cost of this method is susceptibility to noise which makes it less attractive. 4. Incorporating fuzzy methods in NN input preparation is a viable choice. It is also a heuristic approach similar to the item 1, except that it is conducted in a procedural manner [3].
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 8, pp. 731 737 doi: 10.1049/iet-gtd.2011.0757

As mentioned in Section 1, various features of phase currents such as rms, power of fundamental frequency, detail coefcients of DWT and even time samples have been used for the fault classication. Among them, rms includes the overall energy of the post-fault signal and is a stronger of all and smoother than time samples. Therefore the rms of the phase current jump, id (3) and zero current, iz (3) of the samples in the set are calculated. M is set to 40; two periods of the 50 Hz power line signal sampled at 1 Khz. In the real-world case, the variance of the noise can be accurately estimated from the zero-mean pre-fault iz or id ,
734 & The Institution of Engineering and Technology 2012

www.ietdl.org
The advantage of the method that comes next is that the size of the network is minimum; no heuristic parameters are involved and it is more or less compatible with what an expert in power transmission line fault classication follows. 5.3 Proposed NN-based fault classication

To nd an appropriate strategy for accommodating the wide span of the fault data amplitude, one may refer to an expert linguistic expression of fault as this: involvement of a phase with fault is ascertained when change in its current amplitude exceeds certain (condition dependent) threshold while the other phases not. In this expression, there is a talk about the ratio among currents not their exact values. Some phase currents may jump above a certain level while some may not. This directs us towards a type of normalisation that does not require a priori knowledge of the power transmission line specs. Based on this argument, the network inputs are normalised according to the following equation idx = idx , max {ida , idb , idc } x [ {a, b, c} (6)

Fig. 6 Proposed ensemble of MLP networks structure

The consequence of the suggested normalisation for an AG type of fault (an LG fault that phase A is involved) at two different locations is as follows ida = 10 ida = 1, idb = 1 idb = 0.1, idc = 1 idc = 0.1, ida = 0.5, ida = 1 idb = 0.05 idb = 0.1 idc = 0.05, idc = 0.1

Fig. 7 Network training speed using (i) the phase currents input and an 11 node NN, (ii) the normalised data and a four-node NN, (iii) the normalised data and an eight-node NN

As it is clear, the current jumps are replaced by their ratios. This is done at each fault instance independent of the others. While this method attenuates the currents jumps of strong faults, it boosts the current jump of high-impedance ones, in other word; the method has non-linear sensitivity in favour of high-impedance faults. The advantage of the scheme can easily be veried mathematically. Assume an AG type of fault. The change in the faulty phase current can be expressed by [17] ida = kif where k is the complex distribution factor and if is the fault current. The location of fault is expressed by k and other parameters such as fault resistance, load angle and so on are reected in if . Mutual coupling between lines may be represented by low value bs. idb = bb ida , idc = bc ida

impediment in the training path meaning that the clusters of data have been arranged in a manner that they can be easily separated. This has been depicted in Fig. 7. While i da , i db and i dc have mutual interactions in representing a faulty phase, the mission of iz is binary (ground/ no ground). This separable duty in the fault classication signals steers one towards enjoying the advantage of the ensemble of networks, allocating one network to the faulty phase recognition, and another one for the ground detection. This network has been illustrated in Fig. 6. By this provision, no extra normalisation for iz , the only input of the network 2, is required, since non-linear function of NN neurons basically do the same and no training impediment occurs. To evaluate the performance of the proposed algorithm, the following steps are taken: Obtain samples of ia , ib , and ic for each 1260 faulty cases. Calculate ida(n), idb(n) and idc(n) using (2) and iz(n) using (1) for all samples. Obtain ida , idb and idc and iz using (3). Determine i da , i db and i dc based on (6) for every sample. Build an MLP network of 4 nodes. Just one node per class: normal, phase A, phase B and phase C. Generate a matrix of i da , i db and i dc as the network input pattern. Generate target matrix based on Table 2. Train the network. Build a separate 1 node perceptron NN. Train the network based on iz as input and a target vector derived from Table 2. Fig. 7 compares the training steps of the MLP under various working congurations. First, the phase currents are applied to an MLPNN. The training of the MLP with 11 nodes, 4 inputs (current jumps, ids and iz) and 4 outputs does not succeed. After 400 epoch iterations, convergence is not in the horizon (graph 1). Extensive evaluations of various network sizes, all, were ended in complete failure. However, when the proposed algorithm is tried, perfect
735

The ratios of the current jumps, as dened, render the subsequent values as the indicator of the class of the fault ida = 1, idb = bb , idc = bc

While the amplitudes of the phase currents have been disregarded, their manifestations are preserved. In this respect, no classication data compromise and or heuristic parameters involvement are observed. By the normalisation, the data are shaped in a very appropriate manner perfectly suitable for clustering of the faults. There is also graphical evidence here that it conrms clearly the aforementioned claim. The slope of the training speed of the network devised with the ratio of currents shows that there is no
IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 8, pp. 731 737 doi: 10.1049/iet-gtd.2011.0757

& The Institution of Engineering and Technology 2012

www.ietdl.org
6 Conclusion
This paper discusses the problem of power transmission line fault classications. Past investigations based on NN are briey reviewed. It is perceived that those various methods trying to avoid the difculty emerging from the wide span of the three-phase current jumps because of a fault, that all are going to be the indicator of just one specic fault under different situations. This means that there is no one-to-one relationship between the fault and the current jump. This makes the training of the NN impossible. The methods presented such as RBFNN, fuzzy backed NN are realised to be in response to the existing complication. To bypass the problem, an algorithm is suggested that, instead of using the phase currents as NN inputs, employs their specic ratios. The immediate upshot of the algorithm is: (i) accurate MLP training without any heuristic parameter involvement; (ii) reduction in the size of network to the minimum of one node per phase; (iii) enhancing the effect of high-impedance faults; (iv) lowering the impact of strong faults in a way that all have relatively equal shares in NN training. Simulations using a relatively large fault sample data-set prove the effectiveness of the suggested method. The fault classication algorithm is further enriched when the idea of the ensemble of networks is incorporated into the algorithm.

Fig. 8 Success rate in accurate classication of faulty lines

Acknowledgments

This work has been partially supported by the Research Department of Shahed University, Tehran, Iran.
Fig. 9 Left: training curve of 1 node network used for ground fault detection, Right: success rate in ground involved fault classication

References

success rate is achieved. A network of 4 nodes, 3 inputs (ids) and four outputs could solve the classication puzzle accurately. The fall of the training error has been illustrated in Fig. 7 (graph 2). Increase in the number of neurons does not provide a considerable improvement, Fig. 7 (graph 3). The very negligible error of the NN test output is another indicator of the accuracy of the classication. By redening the fault coding (shown in the horizontal axis of Fig. 8) and re-executing the algorithm, perfect faulty line classication is achieved as Fig. 8 exhibits. Position (1,1) of the matrix shows that 9828 normal instances have all been correctly classied and so for the 1260 AG faulty samples, position (2,2) and the other cases. The training steps and the success rate matrix of the one node NN responsible for ground detection have also been illustrated in Fig. 9. The results show that all 7560 ground involved faults are truly identied. The training speed of the NN in each execution is not the exact duplicate of the previous ones. In each execution different training path is passed. Thus, the presented training curves only show what the network may behave and by no means reect the exact behaviour of the network under various executions. Convergence is achieved, but the number of iterations may differ. As mentioned earlier, the 3L and 3LG type of faults cannot be differentiated in the simulations, where a symmetrical model for the line is assumed through modelling the line by zero and positive sequences. If a non-transposed model for power lines is implemented, the algorithm can easily categorise them.
736 & The Institution of Engineering and Technology 2012

1 Yeoa, S.M., Kima, C.H., Honga, K.S., et al.: A novel algorithm for fault classication in transmission lines using a combined adaptive network and fuzzy inference system, Int. J. Electr. Power Energy Syst., 2003, 25, (9), pp. 747 758 2 Chanda, D., Kishore, N.K., Sinha, A.K.: Application of wavelet multiresolution analysis for identication and classication of faults on transmission lines, Electr. Power Syst. Res., 2005, 73, (3), pp. 323333 3 Mahanty, R.N., Gupta, P.B.D.: A fuzzy logic based fault classication approach using current samples only, Electr. Power Syst. Res., 2007, 77, (56), pp. 501 507 4 Mahanty, R.N., Gupta, P.B.D.: Application of RBF neural network to fault classication and location in transmission lines, IEE Proc. Gener. Transm. Distrib., 2004, 151, (2), pp. 201 212 5 Haykin, S.: Neural networks: a comprehensive foundation (PrenticeHall, NJ, 1999, 2nd edn.) 6 Kezunoic, M.: A survey of neural net application to protective relaying and fault analysis, Eng. Int. Syst., 1997, 5, (4), pp. 185192 7 Santos, R.C., Senger, E.C.: Transmission lines distance protection using articial neural networks, Int. J. Electr. Power Energy Syst., 2011, 33, (3), pp. 721730 8 Samantaray, S.R., Dash, P.K., Panda, G.: Fault classication and location using HS-transform and radial basis function neural network, Electr. Power Syst. Res., 2006, 76, (9 10), pp. 897 905 9 Kezunovic, M., Rikalo, I.: Detect and classify transmission line faults using neural nets, IEEE Comput. Appl. Power, 1996, 9, (4), pp. 4247 10 Silva, K.M., Souza, B.A.: Fault detection and classication in transmission lines based on wavelet transform and ANN, IEEE Trans. Power Deliv., 2006, 21, (4), pp. 20582063 11 Upendar, J., Gupta, C.P., Singh, G.K.: Discrete wavelet transform and probabilistic neural network based algorithm for classication of fault on transmission systems. Proc. IEEE INDICON, 11 13 December 2008, pp. 206211 12 El-Safty, S., El-Zonkoly, A.M.: Applying wavelet entropy principle in fault classication, Int. J. Electr. Power Energy Syst., 2009, 31, (10), pp. 604607 13 He, Z., Gao, S., Chen, X., Zhang, J., Bo, Z., Qian, Q.: Study of a new method for power system transients classication based on wavelet IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 8, pp. 731 737 doi: 10.1049/iet-gtd.2011.0757

www.ietdl.org
entropy and neural network, Int. J. Electr. Power Energy Syst., 2011, 33, (3), pp. 402410 14 Reddy, M.J.: Adaptive-neuro-fuzzy inference system approach for transmission line fault classication and location incorporating effects of power swings, IEE Proc. Gener. Transm. Distrib., 2008, 2, (2), pp. 235244 15 Abdollahi, A., Seyedtabaii, S.: Comparison of Fourier & wavelet transform methods for transmission line fault classication. Proc. Fourth Int. Power Engineering and Optimization Conf., 23 24 June 2010, pp. 579 584 16 Bhowmik, P.S., Purkait, P., Bhattacharya, K.: A novel wavelet transform aided neural network based transmission line fault analysis method, Int. J. Electr. Power Energy Syst., 2009, 31, (5), pp. 213219 17 Wiszniewski, A.: Accurate fault impedance locating algorithm, IEE Proc., 1983, 130, (6), pp. 311314

IET Gener. Transm. Distrib., 2012, Vol. 6, Iss. 8, pp. 731 737 doi: 10.1049/iet-gtd.2011.0757

737

& The Institution of Engineering and Technology 2012