Improved first zone reach setting of artificial neural network-based directional relay for protection of double circuit transmission lines

Published in IET Generation, Transmission & Distribution Received on 5th April 2013 Revised on 17th September 2013 Accepted on 24th September 2013 ISSN 1751-8687 Improved first zone reach setting of artificial neural network-based directional relay for protection of double circuit transmission lines Anamika Yadav, Aleena Swetapadma Department of Electrical Engineering, National Institute of Technology, Raipur, CG, India E-mail: anamika_jugnu@yahoo.com Abstract: This study proposes an artificial neural network-based protection scheme for double circuit transmission line with improved first zone reach setting up to 99% of line length. The proposed scheme involves three stages. The first stage makes the discrimination among normal condition and faults. The second stage identifies the zone/section of the fault from the relay location. If a forward fault is detected in its first zone then the third stage is activated, which classifies the fault type and identifies the faulty phase. The three-phase currents and voltages measured at only one end of the double circuit line are used to calculate discreet Fourier coefficients. Thus, this technique does not require any communication link. The algorithm is presented in detail and extensively tested using the simulink model of a 400 kv, 300 km distributed parameter line simulated in MATLAB 2009a in the time domain. The simulation results show that all types of shunt faults (forward as well as reverse), its zone/section and faulty phase can be correctly identified within a half cycle time. This method is adaptive to the variation of fault type, fault inception angle, fault location, fault resistance, single circuit operation and CT saturation. The main advantage of the proposed scheme is that it offers primary protection to total line length using single end data only and back up protection for the adjacent forward and reverse line section also. Nomenclature ANN DFT FD FD* FC CB R f Φ i t i artificial neural network discrete Fourier transform fault detector fault section estimator fault classifier circuit breaker fault resistance fault inception angle fault inception time 1 Introduction An overhead transmission line is exposed to the environment and the possibility of experiencing faults on the transmission line is generally higher than other main components. When a fault occurs on a transmission line, it is very important to detect it and find its zone in order to make necessary repairs and to restore power as soon as possible. Distance relaying has been widely used for the protection of transmission lines. A distance relay has to perform the dual task of primary and back-up protection. The primary protection should be fast and without any intentional time delay. Back-up protection should operate if and only if corresponding primary relay fails. Distance relays are provided with multiple zones of protection to meet the stringent selectivity and sensitivity requirements. Zone 1 provides the fastest protection with no intentional time delay; the operating time can be of the order of one cycle. It is set to cover 80% of the line length because of the difficulty in distinguishing between faults which are close to remote bus. Zone 2 protection is delayed by co-ordination time interval. Zone 2 is set to protect primary line and also provides back up protection to 50% part of the adjacent line with 0.25 0.4 s delay. Setting of zone 3 is set to cover complete primary and adjacent line and up to 25% of remote line also with additional delay. However, various conditions such as remote in-feed currents, fault-path resistance and shunt capacitance degrades the performance of distance relays [1]. The current differential protection scheme has been successfully applied to protect the entire line. However, the relay settings are difficult to decide due to line-charging currents and unobvious current variation during high resistance faults. Further composite voltage and current measurements were used to improve relay sensitivity [2]. For fast fault clearance to improve system stability, the relaying schemes based on travelling wave are proposed [3, 4]. However, the techniques are difficult to detect close-in and zero voltage faults. Various kind of protection schemes for transmission lines have been proposed in the past for fault detection and classification (phase selection) and distance location [5 12]. However, these techniques do not estimate the direction of fault and its zone. 1

An adaptive phasor measurement unit (PMU)-based fault detection, direction discrimination, classification and location techniques are developed using two-terminal synchronised phasors [13], which requires the communication link for other end data and the reach setting is 80% of line length only. Directional relay based on travelling-wave comparison using intrinsic mode function is developed in [14] which also require the communication link. A positive-sequence directional relaying algorithm for series-compensated line is reported in [15] but it does not classify the type of fault. Some of the artificial neural network (ANN)-based directional relaying techniques have been reported in [16, 17], but these techniques do not classify the fault type and identifies the faulty phase. Some of the transmission line protective relaying techniques developed by researchers in the last 2 years are also discussed here with their shortcomings. Transmission line distance protection using ANN with approximate first order single phase model of transmission line is developed in [18] with 85% reach setting. Back-propagation neural network-based distance relaying scheme for parallel transmission line against single line-to-ground fault only is reported in [19] with reach setting 80%. But the above two techniques neither classify the fault nor it estimate the direction and zone of fault. Support vector machine (SVM) and wavelet transform-based fault classification and distance location on transmission line is reported with reach setting 80% only [20]. Combined wavelet and back-propagation neural network-based fault diagnosis for single circuit line has been developed [21]. However, the transmission system considered in these papers [18 21] is one section line only which does not deal with fault direction and zone estimation and the reach setting is kept 80 85% only. A modular transient directional protection scheme using current signals is reported in [22] and wavelet and Bayesian linear discriminant analysis-based fault classification and zone identification scheme is proposed in [23], but faulty-phase identification function is not implemented in these works. The most recent paper develops an algorithm for directional earth-fault relay [24] with no voltage inputs for single phase to ground fault only; but it does not classify the fault type and identify the faulty phase. The reach setting is usually 80 85% only in all the reported literatures. Thus after reviewing various techniques reported till date, it is felt that there is a great need of development of efficient protection technique for transmission lines for fault detection, classification, direction and zone estimation with improved first zone reach setting. This paper proposes an ANN-based directional relay for protection of transmission line with improved first zone reach setting up to 99% of the line length, which can prove as a better alternative to present primary as well as back-up protection schemes. Double circuit line of different voltage levels for example, 500, 400 and 315 kv are considered in this paper. Results demonstrate that the proposed scheme effectively detects the fault, its direction, zone/section and the fault type within half-cycle time and the reliability of this scheme is not affected by different fault conditions such as fault type, fault distance, fault inception angle, fault resistance and single circuit operation also. Therefore it offers primary protection to total line length, remote back-up protection for the adjacent forward and reverse transmission line section and also the possibility to implement single-pole auto-reclosing/tripping in case of transient/sustained single line-to-ground fault. 2 Power system network simulation The power system network consists of double circuit transmission line of 400 kv, sectionalised into three sections of length 100 km each and fed from sources at both the ends. Three-phase source of 400 kv, 50 Hz is connected to bus B1 with 100 kw and 100 kvar load, short circuit capacity of source is 1250 MVA and X/R ratio is 10. Further this power is transferred through transmission line having three sections each of 100 km length to load end at bus4. At bus B4 a load of 250 kw and another three-phase source of 400 kv, 50 Hz is connected to simulate the remote end in feed which also represents the thevenin equivalent source of the interconnected grid. The power system operating conditions and other parameters are given in the Appendix. The fault breakers are used in every section of line to simulate various shunt faults with a specific fault inception angle and fault resistance. The single line diagram of the power system network under consideration is shown in Fig. 1. In this study, our concern is to design a directional relay thus the section-2 is considered to be primary line to be protected by relays installed at Bus-2 and voltage and currents signals measured at bus-2 are used as input. In section-2; there is fault detector FD-2, section estimator FD-2* and fault classifier FC-2. FD2 discriminates between no fault and fault conditions and FD2* estimates the section/zone of fault. Whereas FC-2 classifies the fault type and identifies the faulty phase so that single pole auto-reclosing can be done. Fig. 1 Single line diagram of proposed power system network 2

3 Proposed ANN-based fault detector, section estimator and fault classifier In this paper, three ANNs have been developed, first network for fault detection, second network for fault section estimation and third network is fault classifier which is activated if the fault is forward and its first zone to classify the fault type. With the help of these networks, all shunt faults are detected; its direction, section and faulty phase can be identified by the use of only one terminal data. Various fault parameter variation for example, fault type, fault location, fault inception angle, fault resistance are studied. The flow chart of the proposed complete protection scheme is shown in Fig. 2. The trip logic sent to the circuit breaker by the proposed relay located at bus-2 of section-2 is shown in Fig. 3. If the fault is in forward direction from bus-2 and in its first zone that is, in line section-2, then the trip signal is issued instantaneously as per zone-1 setting to the CB located at bus-2 of section 2. If the fault is in forward direction and in line section-3 then trip signal is issued after time delay of one cycle time as per zone-2 setting to the CB located at bus-2. Further if the fault is in reverse direction to bus-2 and in line section-1, then no trip signal is sent to CB located at bus-2 or backup protection after some delay can be provided. If the fault is detected as forward fault in section-2 then, the third stage of fault classifier and faulty-phase identifier is activated which classifies the fault type and identifies the faulty phase. The design process of the ANN-based relay goes through the following steps: (a) Selection of networks inputs and outputs. (b) ANN architecture and training with suitable training data set. (c) Evaluation/testing of the trained ANN using test patterns. 3.1 Selection of networks inputs and outputs The variations in the measured voltage and current signals in time series are very discernable and explicit under different fault conditions. The principle of variation of current and voltage signals before and after the fault incidence is used and a fast and reliable ANN-based relay is designed. When a fault occurs, different frequency components of signals appear, and the magnitude of DC is attenuated as time progresses. Furthermore, some of the non-fundamental frequency components change for different fault locations. It is thus necessary to pre-process the input data and extract the useful features for training the ANN. The performance of the neural network depends upon the input and output features. Three-phase voltage and current signals are sampled at 1 khz sampling rate and then passed through Fig. 2 Flow diagram of the proposed complete primary protection scheme 3

Fig. 3 Trip logic sent to the circuit breaker by the proposed relay located at bus-2 of section-2 the analogue filter. The analogue filter used here is a Butterworth low-pass filter of second order with pass band edge frequency of 480 rad/s which is kept so as to remove higher order harmonics from the signal. Subsequently, one full cycle discrete Fourier transform is used to process the time series voltages and currents, and only the fundamental frequency component has been considered. Now these signals are normalised in range 1 to + 1. The network inputs chosen here are the magnitudes of the fundamental components (50 Hz) of three-phase voltages and six currents measured at the relay location at one end of the line. Thus the network inputs for fault detector, section estimator and fault classifier are total nine as given in (1). Ten post-fault samples of fundamental components of three-phase voltage and phase current signals are extracted to form the input matrix of ANN-based fault detector, section estimator and classifier as given in (2) [ ] X = V af, V bf, V cf, I a1f, I b1f, I c1f, I a2f, I b2f, I c2f (1) V af (t), V af (t+1),..., V agf (t+9) V bf (t), V bf (t+1),..., V bf (t+9) V cf (t), V cf (t+1),..., V cf (t+9) [ X = V ] I a1f (t), I a1f (t+1),..., I a1f (t+9) pf = I I b1f (t), I b1f (t+1),..., I b1f (t+9) pf (2) I c1f (t), I c1f (t+1),..., I c1f (t+9) I a2f (t), I a2f (t+1),..., I a2f (t+9) I b2f (t), I b2f (t+1),..., I b2f (t+9) I c2f (t), I c2f (t+1),..., I c2f (t+9) As the basic task of detector module is to distinguish between fault and no fault situation, only one output is chosen which may be 0 if no fault or 1 if there is fault Y1 = [ D] (3) The fault section identification module has three outputs corresponding to the three Sections 1 3, so that it can be identified which of the sections is faulty Y2 = [ S1, S2, S3] (4) Further the fault classification module determines the type of fault along with the faulty phase selection, thus seven outputs corresponding to three phases of double circuit line as well as ground were considered as outputs provided by the network to determine which of the six phases A1, B1, C1, A2, B2 and C2 and/or ground G are present in the fault loop. Based on the fault type which occurs on the system, output should be 0 or 1. Thus the network outputs are Y 3 = [ A1, B1, C1, A2, B2, C2, G] (5) 3.2 ANN architecture and training After the selection of input and output, the next move is to determine the number of layers and the number of neurons per layer. Number of neurons in the hidden layer is decided by investigating randomly with 5, 10,, 20 neurons. Then Table 1 Detection time for fault detection, section estimation and classification modules ANN module Percentage of test cases Detection Percentage accuracy (correct answers) fault detection module (no. of test cases 5051) section estimation module (no. of test cases 3012) fault classification module (no. of test cases 2100) 64.8 less than quarter cycle 100 (5051 cases) 27.9 less than half cycle 8.3 in between half cycle and one cycle 65.99 less than quarter cycle 99.999 (3011 cases) 24.82 less than half cycle 9.15 in between half cycle and one cycle 33.15 less than quarter cycle 99.997(2099 cases) 61.0 less than half cycle 11.35 in between half cycle and one cycle 4

the transfer function is determined. The commonly used functions are logsig, tansig, purelin, satlin and so on. During the training process different fault cases and variations are considered so that network can properly learn the fundamental problem and can be able to respond for variations in different parameters. All 10 types of shunt faults (LG, LLG, LL and LLL) in the three line sections at 170 different fault locations between 0 and 100% of line length with three fault resistance (0, 50 and 100 Ω) and eight fault inception angles (0 360 ) have been simulated. The total number of fault cases used for training and testing are 40 800. It has been found that the network with nine input neurons, 20 neurons in first hidden layer, 20 neurons in second hidden layer and one neuron in output layer each having tansig activation function for ANN-based fault detector (9 20 20 1) is capable of minimising the mean-square error (mse) to a final value of 1 10 6. Similarly ANN-based section estimator and fault classifier were also developed and in the Section 4 the test results are discussed in detail. 4 Results and discussions The ANN-based directional fault detector, section estimator and fault classifier/faulty-phase identifier are required to be tested for operation for all types of fault and variations in fault and network parameters through which the network was not trained previously. Extreme cases like faults near the protection zone boundary including high-fault resistance were also included in the validation data set. The network was tested by presenting about 10 163 different types of faults (LG, LLG, LL and LLL) in the three sections with varying fault locations (L f =0 99 km), fault inception angles (Φ i =0 360 ) and fault resistance (R f =0 100 Ω). The quantitative analysis of performance of the complete scheme in relation to the percentage of correct answers and the detection time for different types of fault test cases ( 10 163) is shown in Table 1. As can be seen from Table 1, the detection, section estimation and classification modules give high percentage of accuracy and correct answers and the detection time is less Fig. 4 Waveform of input variables and test result/output a Waveform input variables b detector FD-2 during AG fault in S2 at 40 km with R f =0Ω, Φ i =0 (t i = 60 ms) c section estimator FD-2* during AG fault in S2 at 40 km with R f =0Ω, Φ i =0 (t i = 60 ms) 5

than half cycle for most of the test cases and it is less than one cycle for few cases only which belongs to single circuit operation of double circuit line. The proposed relay operation/fault detection time and the reach setting are evaluated during faults near boundary in the Section 4.1. The proposed scheme is also tested under different faulted conditions such as faults near boundaries with high-fault resistance, varying faults inception angle, single circuit operation and varying transmission line configuration of different voltage level are analysed and test results are given in the next subsections. The relay also provides back-up protection to the forward and reverse adjacent line section. 4.1 Detection time/time of operation of relay The most important factor that is considered while designing any relay is its time of operation which should be less than one cycle time. The operating time of the conventional digital distance relaying scheme is around one cycle. To evaluate the operating time of the proposed relay, consider a single line-to-ground fault between phase A and ground, at 40 km away from bus-2 in section-2 with R f =0Ω, Φ i = 0 at 60 ms. The inputs applied to the proposed ANN-based relay are fundamental components of three-phase voltage and currents of both the circuit. In this fault situation, Fig. 4 shows the waveform of the inputs and the output of ANN-based fault detector FD-2 and section estimator FD-2*. All the inputs and outputs are plotted in time domain that is, with respect to time in milliseconds in X-axis and in left-hand side (LHS) of Fig. 4 the waveforms of inputs variables are shown and right-hand side (RHS) shows the output of ANN-based fault detector FD-2 and section estimator FD-2*. As expected, the fundamental component of faulty: A phase current waveform of the circuit-1 increases after an AG fault occurs at 60 ms. It can be observed that the phase A current of circuit-2 also increases because of mutual coupling between the two circuits. However, the magnitude of healthy circuit phase current is much smaller as compared to the faulty circuit current. In RHS of Fig. 4, the uppermost plot shows the output D of ANN-based fault detector FD-2 and next three plots are showing outputs of ANN-based section estimator FD-2* that is, S1, S2, S2. Here, we observe that all four outputs are low up to 60 ms time showing that there is no fault. On the occurrence AG fault at 60 ms, the output D of ANN-based fault detector FD-2 and output S2 of ANN-based section estimator FD-2* goes high at 66 ms time with other outputs S1 and S2 remaining low and unaffected. Hence, FD-2 and FD-2* has detected the fault in section-2 at 66 ms. The relay operating time can be calculated as follows Inception time of fault = 60 ms. detection time = 66 ms. Hence time of operation = (66-60) ms = 6 ms (less than half-cycle time). In spite of the fact that ANN uses one full cycle DFT for estimation of fundamental components of three-phase current and voltage as input, we obtain a high speed response time (half cycle). This is because the DFT Fig. 5 Output of ANN module for fault classification and faulty phase identification (FC) during AG fault in section-2at 40 km with R f =0Ω, Φ i =0 (t i = 60 ms) 6

Table 2 Response for double line-to-ground faults near far end boundaries with high fault resistance and variable inception angle location, km inception detection Relay operation time, ms 81 60 65 5 82 60 65 5 83 62.5 67 5.5 84 62.5 67 5.5 85 65 70 5 86 65 70 5 87 67.5 75 7.5 88 67.5 75 7.5 89 70 75 5 90 70 75 5 91 72.5 76 3.5 92 72.5 76 3.5 93 75 80 5 94 75 80 5 95 77.5 85 7.5 96 77.5 85 7.5 97 80 85 5 98 80 85 5 99 60 68 8 computes the fundamental components values continuously since start of simulation at discrete intervals as per sampling time as shown in LHS of Fig. 4. ANN detects the changes (increase) in fundamental components of current and decrease in fundamental components of voltages just after the inception of fault at 60 ms and gives high output in the corresponding phase and ground at 66 ms as shown in RHS of Fig. 4. Thus ANN-based relay detects and classify the fault in high-speed time that is, within half-cycle time. Further as the fault is forward and in section-2 (in first zone of bus-2 relay), the ANN-based fault classifier and faulty phase identifier FC-2 is activated as per proposed scheme shown in Fig. 2. The output of ANN module for fault classification and faulty phase identification (FC) is shown in Fig. 5, all seven outputs are low (0) before the fault inception time 60 ms, and thereafter the faulted phase A1 and G goes high (1) at 67 ms time. Thus the fault is classified as A1G line-to-ground fault at 67 ms and the fault phase is A of circuit-1. 4.2 Evaluating the reach setting of the relay under faults near boundary with high fault resistance and varying fault inception angle The conventional/modern digital distance relays the first zone reach setting is usually kept as 80% of line length. Thus faults near to remote end bus that is, in particular between 80 and 100 km from the relay location are not detected instantaneously; however, they are detected after some delay as per zone two timings. To study the performance of the relay for fault near to remote end bus with high fault resistance, various double line-to-ground faults with R f = 100 Ω has been simulated with varying fault location in step of 1 km (80 99 km) and fault inception angle in step of 45 (0 360 ) or fault inception time after three cycles (60 80 ms). The response of the protection scheme is summarised in Table 2, the relay operation time is maximum 8 ms, and thus the proposed scheme is able to detect the forward faults and its zone within half cycle time in all the cases. The Fig. 6 Test result a detector FD-2 during ABG fault in S2 at 99 km from bus-2 with R f = 100 Ω, Φ i =0 (t i = 60 ms) b section estimator FD-2* during ABG fault in S2 at 99 km from bus-2 with R f = 100 Ω, Φ i =0 (t i = 60 ms) 7

Fig. 7 Test result of FC-2 during ABG fault in section-2 at 99 km from bus-2 with R f = 100 Ω, Φ i =0 (t i = 60 ms) farthest end fault case at 99 km from the relay location at bus-2 with R f = 100 Ω at 60 ms is studied graphically and test results of FD-2 and FD-2* are shown in Fig. 6. Also the test results of fault classification module are shown in Fig. 7 which correctly classify the fault as LLG and identify the faulty phase and ground as A1B1G. It can be concluded that the reach of the relay is approximately 99% in the first zone. 4.3 Impact of single circuit operation of double circuit line The double circuit transmission line may be operated as single circuit line during scheduled maintenance of the one circuit or when one circuit was disconnected due to fault. The performance of the proposed relay has been checked during this network changes also. The test result during single circuit operation and fault in other healthy circuit are shown in Table 3, it can be seen that the relay correctly identifies the fault within one cycle time. 4.4 Impact of varying transmission line configuration of different voltage level The proposed scheme is tested for varying transmission line configurations of different voltage levels such as 315 kv 60 Hz, 400 kv 50 Hz and 500 kv 60 Hz and the details are given in the Appendix. As an example test results of ANN-based detector FD-2 and section estimator FD-2* in case the transmission line is operating at voltage level 315 kv 60 Hz and subjected to AG fault at 40 km from Table 3 Test result during single circuit operation and fault in other healthy circuit Circuit open created in circuit location, km inception time, ms detection time, ms Relay operation time, ms 2 1 3 60 71 11 1 2 3 60 77 17 2 1 53 60 77 17 1 2 53 60 78 18 2 1 99 60 76 16 1 2 99 60 78 18 8

Fig. 8 Test result a detector FD-2 for 315 kv, 60 Hz double circuit transmission line subjected to AG fault in S2 at 40 km from bus-2 with R f =0Ω, Φ i =0 (t i = 60 ms) b section estimator FD-2* for 315 kv, 60 Hz double circuit transmission line subjected to AG fault in S2 at 40 km from bus-2 with R f =0Ω, Φ i =0 (t i =60ms) bus-2 at 60 ms is evaluated and shown in Fig. 8. As the fault occurs in section-2, the fault is forward thus output D of FD-2 and S2 output of FD-2* is high (0.995) at 65 ms time and all other are low. Thus the direction and section of the fault has been correctly estimated within half-cycle time. 4.5 Back-up protection to the 99% adjacent forward line section-3 If the fault occurs in the adjacent forward line that is, in section-3, the conventional relay provides back-up protection to 20% of the adjacent line section in zone 2 setting with some delay. However, the proposed scheme provides back-up protection to the complete adjacent forward line. This has been proved through number of fault simulations in the line section-3 and test results are shown in Table 4. One of the test results is evaluated in graphical form to show the detection time. A phase to phase AB fault is considered in section-3 at 99 km from bus-3 or 199 km from bus-2 at 60 ms. The test results shows that the fault detector and the section estimator have detected the fault as forward in line section-3 as depicted in Fig. 9. 4.6 Response to reverse fault in section-1 The performance of the proposed relay located at bus-2 has been tested for fault in reverse direction from the relay that is, in section-1 and results are given in Table 5. The fault locations are shown negative as these are reverse fault in LHS from bus-2. Maximum time required by the relay to detect the reverse fault is 5 ms for fault located at 97 km in LHS from bus-2 or 3 km from bus-1 as shown in Fig. 10. It can be seen that the proposed scheme can detect the reverse fault also within half-cycle time. 4.7 Effect of current transformer (CT) saturation To consider the effect of CT saturation, six CT s are used to measure current flowing through the 400 kv double circuit transmission line network using the saturated transformer block of Simpowersystem toolbox of MATLAB. The CT s Table 4 Response for phase to phase fault in adjacent forward line section-3 location, km inception detection Relay operation time, ms 101 60 65 5 103 60 65 5 113 60 65 5 123 60 66 6 133 60 66 6 143 60 67 7 153 60 67 7 163 60 68 8 173 60 68 8 183 60 69 9 193 60 69 9 199 60 69 9 9

Fig. 9 Test result a detector FD-2 during AB fault in S3 at 199 km from bus-2 with R f =0Ω, Φ i = 0, t i =60ms b section estimator FD-2* during AB fault in S3 at 199 km from bus-2 with R f =0Ω, Φ i = 0, t i =60ms are rated 2000 A/5 A and 25 VA. The CT is assumed to saturate at 8 pu. As the primary current of CT is very high during single line-to-ground fault at zero fault inception angle because of high value of dc offset current, this causes saturation of the CT and thus the current in the secondary of CT is decreased. Three-phase current waveforms during an AG fault with and without CT saturation are shown in Fig. 11. Table 5 section-1 location, km Test results for reverse fault at different location in inception detection Relay operation time, ms 97 60 65 5 87 60 63 3 77 60 62 2 67 60 62 2 57 60 62 2 47 60 62 2 37 60 62 2 27 60 62 2 17 60 62 2 7 60 62 2 The proposed scheme is tested considering CT saturation during an AG fault in ckt-1 in section-2 at 99 km from bus-2 with R f = 0.001 Ω, Φ i =0 (t i = 60 ms). The test result of ANN module for fault detection and fault section estimation with CT saturation are shown in Fig. 12. It can be seen that, the performance of the proposed scheme is not affected by CT saturation as the fault is detected and section is identified in 8 ms time. Further the output of ANN module for fault classification and faulty phase identification (FC) is shown in Fig. 13, all seven outputs are low (0) before the fault inception time 60 ms, and thereafter the faulted phase A1 of ckt.-1 and G goes high (1) at 70 ms time. Thus the fault is classified as A1G line-to-ground fault in 10 ms time and the faulty phase is A of circuit-1. Further some of the test results with CT saturation of different types of fault in section-2 at 99 km are given in Table 6, which shows the relay operation time lies in between 5 and 12 ms. 5 Comparison with some other existing/ artificial intelligence-based schemes The results of the simulation studies presented in the preceding section clearly shows that the proposed scheme 10

Fig. 10 Test result a detector FD-2 during reverse AG fault in S1 at 97 km from bus-2 with R f =0Ω, Φ i = 0, t i =60ms b section estimator FD-2* during reverse AG fault in S1 at 97 km from bus-2 with R f =0Ω, Φ i = 0, t i =60ms Fig. 11 Three-phase current waveforms during an AG fault at 60 ms in ckt.-1 with and without CT saturation 11

Fig. 12 Test result a detector FD-2 with CT saturation during AG fault in S2 at 99 km from bus-2 with R f = 0.001 Ω, Φ i =0 (t i = 60 ms) b section estimator FD-2* with CT saturation during AG fault in S2 at 99 km from bus-2 with R f = 0.001 Ω, Φ i =0 (t i = 60 ms) Fig. 13 Test result of FC-2 with CT saturation during AG fault in section-2 at 99 km from bus-2 with R f = 0.001 Ω, Φ i =0 (t i = 60 ms) 12

Table 6 Test results with CT saturation of different types of fault in section-2 location, km type inception detection Relay operation 99 AG 60 68 8 99 BG 60 68 8 99 CG 60 71 11 99 ABG 60 72 12 99 BCG 60 69 9 99 CAG 60 69 9 99 AB 60 69 9 99 BC 60 71 11 99 CA 60 71 11 99 ABC 60 65 5 for fault detection, section estimation and fault classification is very effective under wide variations in the operating and fault conditions. SVM has been utilised to compare the results obtained with proposed ANN-based scheme employing Marquardt Levenberg algorithm. In order to compare the two methods the same data created for the proposed scheme have been used. It has been found that the SVM-based scheme does not work properly during single circuit operation and thus, the data set of single circuit operation has been removed. Various SVM models have been developed in training phase, if there are n classes then n SVM models are developed, one model for each class. Quantitative comparison of results obtained with similar analysis done using ANN and SVM is given in Table 7. It can be seen that the average response time of the proposed scheme using ANN is lower than that of SVM and accuracy in terms of percentage of correct answers is higher than that of SVM. The number of networks developed during training is one each for fault detection, fault section estimation and fault classification using ANN as compared to two SVMs for fault detection, four SVMs for fault section estimation and seven SVMs for fault classification. It is to mention here that, the time required for training although offline is much higher in case of SVM around 3 4 h. Thus it can be concluded that ANN-based scheme performs better than SVM. Quantitative comparison of the proposed scheme with some other artificial intelligence-based schemes [23, 25 28] is given in Table 8. It can be seen that in earlier reported techniques the reach setting is usually 80 85% only except in one scheme-based SVM and wavelet for single circuit line it is 95%. None of these techniques provides complete protection scheme for fault detection, fault-section estimation and fault classification and most of the schemes are non-directional developed for single circuit line. However, the proposed technique is developed for double circuit line in which the zero sequence mutual coupling between the faulty and healthy phase causes mal-operation of conventional protection scheme. The schemes [23, 25, 26] are not developed in time domain and they requires fault to be detected first by some other means and then the scheme can be applied. The response time of the proposed scheme is less than half cycle time and the inputs required by the proposed scheme are much lesser and simple and wide variation in different fault parameters/cases has been considered as compared to other schemes. The proposed scheme is more advantageous as compared with other AI-based schemes in terms of following reasons: (a) Reach setting provided by the proposed scheme is up to 99% of line length. (b) Simplicity in calculation of ANN inputs. (c) Wide variation in different fault parameters/cases. (d) Back-up protection to the 99% of adjacent forward and reverse line sections. (e) Fast response time of less than 10 ms for fault detection, section estimation and classification in time domain. (f) Not effected by CT saturation. 6 Advantages of the proposed scheme a. The performance of ANN-based detector, section estimator and fault classifier is completely unaffected by variation in fault location, fault type, fault resistance and fault inception angle. b. The reach setting provided by the proposed scheme is up to 99% of line length which is greater than presently used protection schemes, thus providing primary protection greater portion of line length as compared to conventional scheme. c. The relay also provides back-up protection to the 99% of adjacent forward and reverse line sections. d. The operating time is less than one cycle as it detects the changes in voltage and currents during the pre-fault and post-fault period itself. e. There is no requirement of communication link for operation. Table 7 Quantitative comparison of results obtained using ANN and SVM Parameters Using ANN in time domain Using SVM in time domain detection section estimation classification detection section estimation classification relay operation time for fault at 4 5 5 5 6 6 near end 0.1 km, ms relay operation time for fault at 8 8 5 5 9 7 farthest end 99 km, ms average relay operation time, 5 5 6 5 7 7 ms accuracy, % 100 99.99998 99.997 99 95 98 number of networks developed during training one ANN one ANN one ANN two SVMs four SVMs seven SVMs 13

Table 8 Quantitative comparison with some other artificial intelligence-based schemes Parameters Adaptive wavelet and Bayesian classifier for single circuit line [23] Wavelet-neuro-fuzzy combined approach for single circuit line [25] PSO and ANN-based approach for single circuit line [26] SVM and wavelet-based technique for single circuit line [27] Data-mining model for FACTS-based single circuit line [28] Proposed technique for double circuit line reach setting not mentioned 83.33% of line length 80% of line length 95% of line length 85% of line length 99% of line length accuracy: not done not done not done 100% 99 99.51% 100% fault detection accuracy: 100% not done not done 100% not done 99.99998% fault section estimation accuracy: 59.80 100% not mentioned done 99.54 100% not done not done 99.997% fault classification accuracy: not done not mentioned done 99.54 100% not done not done 99.997% faulty phase identification directional directional non-directional non-directional non-directional non-directional directional (forward/ re-verse) response time 15 ms for fault detection inputs required fault type fault locations fault inception angle fault resistance it requires fault to be detected by other means and then post fault data of 2.058 ms is required for fault classification and section/ zone estimation 1024 inputs: 1024 1 vector of single phase current (2.0183 ms data window) LG, LLG, LL and LLL 37 locations for each fault section. 12 Φ i :0 330 with step of 30 fixed 40 Ω for LG and 5 Ω for LLG it requires fault to be detected by other means and response time is not mentioned/ applicable as the work is not carried out in time domain. summation of third level detail coefficients of three phase currents LG, LLG, LL and LLL 14 locations: 0 100 km with step of 10 and150, 200, 250 km (line length 300 km) 10 Φ i :0 180 with step of 20 It requires fault to be detected by other means and then post fault data of two cycles is required for fault classification. 512 samples of three phase currents to form six inputs: sum of seventh level detail coefficient of three phase currents and sum of absolute values of seventh level detail coefficient of three phase currents LG, LLG, LL and LLL 64 locations: 0.25 300 km 5 ms for fault detection (internal or external) 75 inputs: three phase currents signal energy at five frequency bands corresponding to details coefficients of wavelet: ¼ cycle data samples. 60 inputs: 10 samples of each phase voltage and current LG, LLG, LL and LLL LG, LLG, LL and LLL 15 locations four locations: 25, 45, 55 and 85% of single line length (300 km) 90 Φ i :0 360 Φ i : 10 90 3 Φ i :30, 45 and 60 <10 ms for fault detection, classification and section estimation in time domain. In case of single circuit operation; it is <20 ms nine inputs: three phase voltage and three phase currents fundamental components of ckt-1 and 2 LG, LLG, LL and LLL 170 locations : 0.1 99% of line length 8 Φ i : 0, 45, 90, 135, 180, 225, 270, 360 2 R f : 0.001 100 Ω 21 R f :0 200 Ω R f : 0.01 50 Ω 5 R f :0 200 Ω 3 R f :0Ω, 50Ω, 100 Ω fault cases 5328 1000 1 209 600 1500 38 400 for TCSC and 43 200 for UPFC 40 800 7 Conclusions ANN-based directional fault detection, section estimation and classification scheme is found to be very effective under various fault situations and single circuit operation also. The main contribution of the proposed technique is the improvement in the reach setting of the first zone of protection of transmission line up to 99% of the line length. The reliability of this scheme is not affected by different fault conditions such as fault type, fault distance, fault-inception angle, fault resistance and CT saturation. Results demonstrate that the proposed scheme effectively detects the fault direction, its zone/section, and the fault type within half cycle time. It offers primary protection to total line length, back-up protection for the adjacent forward and reverse transmission line and also the 14

possibility to implement single-pole tripping in case of single line-to-ground fault as the faulty phase is identified by fault classifier. 8 References 1 Phadke, A.G., Thorp, J.S.: Computer relaying for power systems (Wiley, New York, 1988) 2 Aggarwal, R.K., Johns, A.T.: A differential line protection scheme for power systems based on composite voltage and current measurements, IEEE Trans. Power Deliv., 1989, 4, (2), pp. 1595 1601 3 Mansour, M.M., Swift, G.W.: A multi-microprocessor based travelling wave relay theory and realization, IEEE Trans. Power Deliv., 1986, 1, (1), pp. 272 279 4 Saleem, S.M.A., Sharaf, A.M.: A novel travelling wave based relaying scheme using wavelet transforms for arcing faults detection on series compensated transmission lines. Canadian Conf. Electrical and Computer Engineering (CCECE) 2007, Vancouver, BC, 22-26 April 2007, pp. 575 578 5 Jongepier, A.G., van der Sluis, L.: Adaptive distance protection of a double-circuit line, IEEE Trans. Power Deliv., 1994, 9, (3), pp. 1289 1297 6 Bo, Z.Q., Aggarwal, R.K., Johns, A.T., Li, H.Y., Song, Y.H.: A new approach to phase selection using fault generated high frequency noise and neural networks, IEEE Trans. Power Deliv., 1997, 12, (1), pp. 106 115 7 Agarwal, R.K., Xuan, Q.Y., Dunn, R.W., Johns, A.T., Bennett, A.: A novel fault classification technique of double circuit lines based on a combined unsupervised/supervised neural network, IEEE Trans. Power Deliv., 1999, 14, (4), pp. 1250 1256 8 Dash, P.K., Pradhan, A.K., Panda, G. : A novel fuzzy neural network based distance relaying scheme, IEEE Trans. Power Deliv., 2000, 15, (3), pp. 902 907 9 Mazon, A.J., Zamora, I.J., Minambres, F., Zorrozua, M.A., Barandiaran, J.J., Sagastabeitia, K.: A new approach to fault location in two-terminal transmission lines using artificial neural networks, Electr. Power Syst. Res. J., 2000, 56, (3), pp. 261 266 10 Youssef, O.A.S.: New algorithm to phase selection based on wavelet transforms, IEEE Trans. Power Deliv., 2002, 17, pp. 908 914 11 Pradhan, A.K., Routray, A., Pati, S., Pradhan, D.K.: Wavelet fuzzy combined approach for fault classification of a series-compensated transmission line, IEEE Trans. Power Deliv., 2004, 19, (4), pp. 1612 1618 12 Jiang, J.A., Lin, Y.H., Yang, J.Z., Too, T.M., Liu, C.W.: An adaptive PMU based fault detection/location technique for transmission lines. II. PMU implementation and performance evaluation, IEEE Trans Power Deliv., 2000, 15, (4), pp. 1136 1146 13 Jiang, J.A., Chen, C.S., Liu, C.W.: A new protection scheme for fault detection, direction discrimination, classification, and location in transmission lines, IEEE Trans. Power Deliv., 2003, 18, (1), pp. 34 42 14 Jiandong, D., Qin, Z., Lin, A.: Directional relay based on travelling-wave comparison using intrinsic mode function for EHV transmission lines. 5th Int. Conf. Critical Infrastructure (CRIS), 20 22 September 2010, pp. 1 5 15 Jena, P., Pradhan, A.K.: A positive-sequence directional relaying algorithm for series-compensated line, IEEE Trans. Power Deliv., 2010, 25, (4), pp. 2288 2298 16 Sidhu, T.S., Singh, H., Sachdev, M.S.: An artificial neural network based directional discriminator for protecting transmission lines. Proc. Canadian Conf. Electrical and Computer Engineering, 1993, pp. 205 208 17 Sidhu, T.S., Singh, H., Sachdev, M.S.: An artificial neural network for directional comparison relaying of transmission lines. Proc. Sixth Int. Conf. Developments in Power System Protection, 25 27 March 1997, 434, pp. 282 285 18 Ricardo Caneloi dos, S., Eduardo Cesar, S.: Transmission lines distance protection using artificial neural networks, Int. J. Electr. Power Energy Syst., 2011, 33, (3), pp. 721 730 19 Upendar, J., Gupta, C.P., Singh, G.K.: Comprehensive adaptive distance relaying scheme for parallel transmission lines, IEEE Trans. Power Deliv., 2011, 26, (2), pp. 1039 1052 20 Sami, E.: Support vector machines for classification and locating of fault in transmission line, Appl. Soft Comput., 2012, 12, (6), pp. 1650 1658 21 Ngaopitakkul, A., Bunjongjit, S.: An application of a discrete wavelet transform and a back-propagation neural network algorithm for fault diagnosis on single-circuit transmission line, Int. J. Syst. Sci., 2013, 4, (9), pp. 1745 1761 22 Perera, N., Rajapakse, A.D.: Design and hardware implementation of a modular transient directional protection scheme using current signals, IET Gener. Transm. Distrib., 2012, 6, (6), pp. 554 562 23 Pérez, F.E., Aguilar, R., Orduña, E., Jäger, J., Guidi, G.: High-speed non-unit transmission line protection using single-phase measurements and an adaptive wavelet: zone detection and fault classification, IET, Gener. Transm. Distrib., 2012, 6, (7), pp. 593 604 24 Stojanovic, Z.N., Djuric, M.B.: An algorithm for directional earth-fault relay with no voltage inputs, Electr. Power Syst. Res., 2013, 96, pp. 144 149 25 Jayabharata Reddy, M., Mohanta, D.K.: A Wavelet-neuro-fuzzy combined approach for digital relaying of transmission line faults, Electr. Power Compon. Syst., 2007, 35, (12), pp. 1385 1407. 26 Upendar, J., Gupta, C.P., Singh, G.K., Ramakrishna, G.: PSO and ANN-based fault classification for protective relaying, IET, Gener. Transm. Distrib., 2010, 4, (10), pp. 1197 1212 27 Jafarian, P., Sanaye-Pasand, M.: High-frequency transients-based protection of multiterminal transmission lines using the SVM technique, IEEE Trans. Power Deliv., 2013, 28, (1), pp. 188 196. 28 Samantaray, S.R.: A data-mining model for protection of FACTS-based transmission line, IEEE Trans. Power Deliv., 2013, 28, (2), pp. 612 618 9 Appendix Power system operating conditions and different parameters 9.1 For 400 kv line Nominal source voltage 400 kv, 50 Hz Pre-fault power flow angle at bus-1 (δ) 15 Short circuit capacity 1250 MVA Base voltage 400 kv X/R ratio 10 Line length 300 km (divided to three sections each of 100 km) +ve, ve and zero sequence line resistance in Ω/km [0.0275, 0.275 and 0.21] +ve, ve and zero sequence line inductance in H/km [1.002e 03, 3.268e 03 and 2.0e 03] +ve, ve and zero sequence line capacitance in F/km [13e 09, 8.5e 09 and 5.0e 09] Load at bus-1 100 kw and 100 kvar Load at bus-4 250 kw 15

9.2 For 315 kv line Nominal source voltage 315 kv, 60 Hz Pre-fault power flow angle at bus-1 (δ) 15 Short circuit capacity 1250 MVA Base voltage 400 kv X/R ratio 10 Total line length 300 km (divided to three sections each of 100 km) +ve, ve and zero sequence line resistance in Ω/km [0.027151, 0.25155 and 0.22453] +ve, ve and zero sequence line inductance in H/km [0.00099919, 0.0036373 and 0.0023547] +ve, ve and zero sequence line capacitance in F/km [1.1662e 008, 6.9959e 009 and 2.1088e 009] Load at bus-1 100 kw and 100 kvar Load at bus-4 250 kw 9.3 For 500 kv line Nominal source voltage 500 kv, 60 Hz Pre-fault power flow angle at bus-1 (δ) 15 Short circuit capacity 1250 MVA Base voltage 400 kv X/R ratio 10 Line length 300 km (divided to three sections each of 100 km) +ve, ve and zero sequence line resistance in Ω/km [0.018396, 0.26486 and 0.24619] +ve, ve and zero sequence line inductance in H/km [0.00092959, 0.0032022 and 0.0019996] +ve, ve and zero sequence line capacitance in F/km [1.2571e 008, 7.8555e 009 and 2.0444e 009] Load at bus-1 100 kw and 100 kvar Load at bus-4 250 kw 16