‘Applied Soft Computing 26 (2016) 295-305 Contents lists available at ScienceD) Applied Soft Computing journal homepage: www.elsevier.com/locate/asoe ELSEVIER Distribution network fault section identification and fault location using wavelet entropy and neural networks Ons Adeyemi Charles Adewole*, Raynitchka Tzoneva, Shaheen Behardien enter Sustation auromaton and Energy Management Systems CSAEMS). Cape Pensa University of Teceloy Cape Town, South Aca ARTICLE INFO ABSTRACT Received 30 January 2014 Faullocationin power system distribution networks is especially dificult Because ofthe exstenceofsev- cal latealsjtap-offs in distribution networks. This implies thatthe calculated faut point can be wrongly estimated fo bein any ofthe lateral, This paper proposes a new hybrid method combining Discrete ‘Wavelet Transform (DWT) and artificial neural nework (ANN) for fault section icenbifieation (FS) and Fault location (1) in power system distribution networks. DWT was used in the analysis and extraction ofthe characteristic features from fault transient signals of the thee phase line current measurements obtained 2 a single substation relaying point, rather than the double-ended approach used in the exist- ing literature. Entropy Per Unit (EPU) indices are afterwards computed from the DWT decomposition, and are used as inpat to mul-layer ANN models serving as Ft classifiers ad FL predictors respectively. 3 proposed hybrid method is tested using a benchmark IEE 24-node test feeder. Comparisons, veri cation, and analysis made using the experimental results obtained from the application of the method showed very good performance for different fault types fault locations, fault inception angles, and fault Fesistances. The proposed hybrid method is unique because of the pre-processing stage done with the DDWT-EPU indices, the use of only line current measurements omasingle relaying point andthe ivsion ofthe FSI and FL tasks int sulb-problems with respective ANN models. Discrete waveform transform 1 2016 Elsevier BY, All rights reserved, 1. Introduction Electric power interruptions are characterized by standard reli- ability indies. These indices measure the system performance and reflect the status of the overall performance/Quality of Service (QoS) of a particular network at the load points. The most com- ‘manly used indices in utility companies are the System Average Interruption Frequency Index (AIF and the System Average Inter- ruption Duration Index (SAIDI) SAIFlis the mean of the number of interruptions a customer experiences, and it is measured in units of interruptions/customer over a year. SAIDIis the system average interruption duration per customer per year. This index is com- monly feferred to as customer minutes of interruption. In order to maintain a good QoS and facilitate the timely restoration of power, there is the need for effective fault diagnosis methods to improve the reliability of distribution networks when faults occur. ‘The identification and location of faults is difficult in distribution networks because of the non-homogeneity of lines along a typ- Corresponding autor Email eddresses: adewolex@cpuc aca, yermyadevleeyahon cok IAC Adee 1558-4945/0 2016 sewer BY. Alleghs reserved ical distribution feeder, Also, the presence af laterals (branches), distributed loads, and frequent modifications inthe feeder config- ‘uration, makes distribution networks distinct. The early detection and diagnosis of faults would expedite service restoration and help reduce downtime. Fault section identification (FSI) is the identification of the faulted section in the distribution system, while fault loca- tion (FL) is the estimation of the distance to the fault point from the substation relaying point. When faults occur in dis- ‘uibution lines, the transient features are distinct and the high frequency components in the transients can be employed to reveal fault characteristics. Diagnostic methods are broadly classified into impedance/other fundamental frequency methodologies, high frequency components travelling wave methodologies, and know!- edge based methodologies (1.2). Fault location in transmission networks is widely reported in the literature, {mpedance-based ‘methods for transmission networks have been proposed in Rets, [3-6], while Refs. [7.8] made use of artificial neural networks (ANS). Similarly, Bhowmik et al. [9] used a method based on ‘wavelet analysis and ANN, Hongchun and Xiangfei | 10) proposed a method based on rough sets and ANN, while Sadeh and Afradi [1] used an adaptive network-based fuzzy inference system, However, ‘methods designed for transmission networks ate prone to errorsAC Adewele etal Applied Soft Computing 45 (20 because ofthe dissimilarity between transmission and distribution networks. This is because distribution lines are non-homogeneous with different conductor sizes. Also, the presence of laterals/tap- ‘offs, radial operation, inaccuracies in the configuration of the feeder, unbalanced phases, varying fault resistance, and load taps along distribution lines make distribution networks distinct. There- Tore, fault diagnosis in distribution networks is beginning to attract the attention of researchers, Methods based on impedance calculations [12-19], wavelet and travelling wave theory (20-28], computational intelligence (29-34), and hybrid methods [35,36], have been proposed for distribution networks fault location, Most of the existing meth- ‘ods make use of current and voltage quantities which add to the computational effort and processing time. This also implies that Voltage Transformers (VTs) and Current Transformers (CTs) would bbe required for real world implementation in substations. However, the substations in distribution networks only have CTs (current measurement), usually for overcurrent protection relays (ANSI 50/51). Furthermore, the impedance and fundamental frequency ‘based method suffer from limitations due tothe loading of the line, fault path resistance, admittance matrix manipulation, harmonic components in the current source parameters, measurement error, and load imbalance. There is also the possibility of multiple esti- mation due to the existence of multiple possible fault locations in the distribution network The shortcomings of the travelling ‘wave methodology include the high implementation casts, com= plexity, the need for high sampling rate, sophisticated measuring ‘equipment, and the need for synchronizationjeommunication for double-ended methods. Thisis further exacerbated by the disconti- nities caused by the numerous sub-feeders that are characteristic of radial distribution systems. The discontinuities may be between, the end of the line and the fault point, and these add up the reflec tions to the transient waves from the fault, The techniques in the knowledge-based method are regarded as complex and the compu tational costs may be high in applications involving a high number of faults, From the foregoing, it can be seen that many diagnostic meth- ‘ods have been developed and proposed, but a perfect, dependable, and secure method is still needed. In this paper, a new 2-stage ‘method based on the hybrid integration of wavelet transform and artificial neural network s proposed for fault section identification and fault location in power system distribution networks. The pro- posed fault section identification stage is initiated before the fault location stage in order to solve the distribution network multiple estimation problem whereby multiple laterals corresponds to the calculated fault point. The contribution of this paper can be sum= ‘marized as follows: (i) the development of a non-iterative 2-stage predictive hybrid method comprising of the identification of the faulted section and the fault location in disteibution systems, (i) the application ofa frequency-time wavelet energy entropy technique designed to capture the transient information that are present in the signal during faults, ii the use of single-ended measurements only, and (iv) the development of an accurate hybrid method for PS] and FL using line current measurements only ‘This paper is divided into the following sections: Section 2 describes tie proposed method based on wavelet transform and neural network. The implementation ofthe fault section identifica tion and fault location algorithms are outlined in Section 3, Section 4 discusses the results of this approach, while ection 5 summarizes the conclusion. 2, The proposed wavelet transform-neural network ‘method The aim of the proposed method is the identification of the faulted section and the location of the point where the fault is, in a given power system distribution network This is cartied out after the fault has been detected in the distribution network and the type of fault has been classified as proposed by the authors in Refs. [37.28]. The proposed hybrid method is developed to sat- isfy the following criteria: (i) take into account the specific nature of power system distribution networks ane the scenarios that are likely to oceut, (i) to use line current measurements obtained from the three phases and zero sequence currents measured in the dis tribution network as the input data to the proposed method, (it) the extraction of significant information that can represent the fun ‘damental characteristics of the measured signal froma disturbance ‘waveform, and (iv) the ability ofthe proposed method to take fast decision an the basis ofthe extracted disturbance waveform, The steps involved in the implementation of the proposed hybrid method are detailed below: i)the specificfeature extraction of the fault signal obtained through the application ofthe Discrete Wavelet Transform (DWT) and the computation of the Wavelet Energy Entropy (WEE), and (i) fast decision making for the iden- tification of the faulted section and the location of the fault point through the training of artificial neural network models using the extracted DWT WEE and the computed Entropy Per Unit (EPU) as inputs. Fig, 1 illustrates the above-mentioned steps, and the theoretical background of the methods involved are given in the proceeding sections, 2.1, Wavelet transform Although. the Fourier transform (FT) has been used in several fields for signal processing, its application for transient signals is, limited by the fact that fault transients are non-stationary and there is the need to analyze them at various transitions as the signal changes, Short-Time Fourier Transform (SIFT) was introduced to correct the shortcoming of the FT. However. a fixed time window is used. ‘Atechnique such as wavelet transform (WT) capable of multi- ple resolutions in time and frequency, witha flexible window size sullable for non-stationary signals s thereby required. ‘The windowing in WT automatically uses short time intervals {or high frequency components, and ong time intervals for low fre- quency components determined by using the scaling and shifting techniques, Discrete Wavelet Transform (DWT) isa variant of WT, and it is a versatile signal processing technique widely applied in, ‘many engineering and scientific fields. One area in which the DWT has been particularly successful is for transient analysis in power systems [9)~42], This is because it aequites the transient features and accurately analyzes them in both the time anc frequency con- texts at different frequency bands, and with different resolutions by decomposing the signal into a coarse approximation and detail components. It employs two sets of functions, called the scaling function @ and the wavelet function ¥, which are related to low pass and high pass filters, respectively, The mathematical expression for DWT is given by Refs. (40,41,43,44): bwin, wr (SF) a where fk) is the discrete signal represented as a function of its coefficients, yt.) i the mother wavelet (window function), m and nare the time scale parameters. kis both the discrete time and the ‘number of coefficients in the DWT, 2” is the vatiablefor scaling, 2” is the variable for shifting, and 1/V/2™ isthe energy normalization component to ensure the same scale as the mother wavelet. The original signal sequence f(x) can also be represented by @ ‘sum of all components, That is, the sum of all the details and the208 AC Adewoe etal Ape soft Computing 462016) 295-105 Fare |, Extraction [2205] a modes approximation components at the last level of decomposition. For ‘example. for two levels of decomposition. the representation is $008) =e Dy(k) + eVAn(k) = 13 (0) + €2D2(K) + e9A(k) @ ‘where Dy and D; are the detail components ofthe frst and second ‘decomposition evels respectively, and A, are the approximation ‘components of the first and second decomposition levels, cis the coefficient of the detail or approximation components at the jth ‘decomposition level, and ¢1Ai(K) = ¢2D2(h) + GA2(R). In the implementation of Multi Resolution Analysis (MRA) for DWT, the scaling and wavelet functions are given by 20} Salt) = 246 (2 tn) @ Vona(t) = 240¥ (2-Me— ni) @ where G(s the scale function, and ym nf) is the wavelet func tion of time t ‘Wavelets are localized in both time (through translation oper- ation) and frequency (through dilation operation) The first scale covers a broad frequency range at the high frequency end of the spectrum and the higher scales cover the lower end of the fre- ‘quency spectrum. The decomposition of the signal starts by passing it through a set of filters. The approximations are the high-scale, low-frequency components ofthe signal produced by filtering with a low-pass filter (h), The details are the low-scale, high-frequency ‘components of the signal produced by a high-pass filter (g). The bandwidth of these two filters is equal but (fis the reversed vversion of (g), After the first decomposition ofthe signal, the sam- pling frequency is decreased by half. To obtain the components of the next DWT level, te low-pass filter output (approximation) is decomposed. Daubechies-4 (db4) is one of the most used wavelet in power system disturbance analysis and it was chosen to analyze the fault signals in this paper because of its orthogonality, compact support inthe time domain, and for its good performance in power system studies as reported by Refs. [25.40 22. Wavelet energy spectrum entropy Fault signals are known to contain transients and harmonics. ‘The direct use of these signals would result to classifiers and pre- dlictors with poor performance. Thus, feature extraction should be carried out in order to identify attributes (features) which best represent the characteristics of the fault signal, Wavelet Energy Entropy (WEE), which evolved from Shannon entropy [39~41] is capable of giving the energy information in signals or systems. Since fault signals have high frequency components, itis more dis- tinctive to use the energy of the detail coefficients to extract the ‘characteristics ofthe fault signals, “The wavelet energy of a signal at a scale j and an instant k is siven as [39-41 Fa= BOP 6) {At the jth scale, the instants are k= 1,2, 3, .., N, where Nis the numberof instants (coefficients) in the jth scale, Lis the number of {decomposition levels. The summation ofthe signal wavelet energy spectrum atthe jth scale is: B= fe i= TT 6) ‘The relative wavelet energy is given as |40) y= BE mB ‘The wavelet energy ratio (Eq. (7)) represents the energy distri- bution. Therefore, the wavelet energy entropy (WEE) is given by ” (40.41): Were =-Y Php l0E Pip. §= TT @) where p « {4,8 C} are phases ofthe distribution network. ‘he wave Ena Ber Un Index proposes Shvoste (Spm) ‘ (= wr) * ( > oe) . ‘ . (9) Eq, (9) forms the basis for the feature extraction and selection, process. The extracted features are used as inputs to the ANN mod- els trained to identify the faulted section and the location of the fault point respectively 23. Fault section identification ANN design Four categories of faults are possible ina distribution network. These are () single line-to-ground (1 Ph.g) faults, (ji) two phase (2 Ph) faults (it) two phase-to-ground (2 Ph.g) faults, and (iv) three phase (3 Ph.) faults. In view of the above, individual ANN model are ‘rained for the identification of the faulted section depending on the type of fault detected in the system. Using an ANN model for ‘each fault type increases the learnability ofthe ANN model, reduces the size ofthe hidden layer neuron, and improves the accuracy of the ANN model. Thus, the proposed FSI algorithm consists of four bback-propagation ANN models, During the training of the ANNs, each ANN acquires knowledge about the problem from the training dataset, and stores the knowledge acquired using synaptic weights, between the neurons, Exploratory experimentation involving various numbers of hid- den layers, hidden layer neurons, learning rate, and activation functions was carried out, Sigmoid activation functions (tansig) areAC Adewele et al Apple soft Computing 45 (2018) 25-108 209 vuonoas yes ig. 2, Architectre of the ANN model for he fat ection ieenfestion ak used for the hidden and output layers of the FSI ANN models. The logsimoid and purelin activation functions were also experimented upon, but were found to give poor performances. Fiz? shows the structure ofthe ANN model designed for the FSI task 2.4. Fault location ANN design Similar approach to the FSI task was followed in the design of the fault location (FL) algorithm. The FL is decomposed into four regression tasks corresponding to the four classes of faults that can ‘occur in distribution networks. Thus, ANN models for fault location, are developed for 1 Ph.g,2 Ph, 2 Phg, and 3 Ph, faults respectively. The designed ANN models take in sets of four inputs of EPU/WEE of the three phase and zero sequence line currents, and one output node that gives the estimated distance of the fault in kilometres (cm). The sigmoid activation function (tansig) was used in the hid- den layer, while the lineat activation function (pureli) was used in the output layer. ‘The performance criteria used in the selection of the optimal ANN model are the neural network size (number of hidden layer neurons), correlation coefficient, error histogram, and testing with untrained datasets. The accuracy ofthe untrained datasets is deter- ined by the percentage error of the ANN accuracy, Fig. 3 shows the structure ofthe ANN madel designed for the FL task. 25. Proposed algorithm The algorithms for the implementation ofthe proposed 2-stage hybrid method are shown in Fig. 4. The faut signal waveforms are decomposed using DWT. EPU is afterwards computed and used as the input to the ANN model. The particular ANN model to use is determined through the use of the faull-type classification algo- rithm proposed by the authors in Refs (37.35]. The algorithms for the proposed hybrid method are implemented as described in Sec- tion 3 ofthis paper, Hidden Layer ‘Output Layer uopeoo) ynea 3. Implementation of the proposed method 3.1. Modelling: IEEE 34- Node Renchumark Test Feeder ‘The testing and validation of the proposed hybrid method was carried out using the IEEE 34 Node Test feeder 45). The IEEE 34 [Node Test Feederisa long feeder with unbalanced loading and nom- inal voltage of 24.9 kV. The total feeder load equals 2060 kW and the 2S00KVA. (ig. 5 shows the IEEE 24 Node Test Feeder. Modelling ofthis distibution network was done in DigSILENT PowerFactory. Binary notations have been inserted on the test feeder to denote the various fault sections as shown in Fig. 5. 3.2, Simulations using the test feeder FlectroMagnetic Transient (EMT) simulations consisting of different fault types at different locations with various fault resis tances (R) and fault inception angles (@j) were performed. A ‘waveform window of 10 cycles obtained from the EMT simulations ‘was used. Voltage sags and current swells are usually experienced at the faulted phase(s) during fault conditions. The magnitude of these sags or overcurrents depends on the type ofthe fault and the system parameters. Three phase line currents and zero sequence current of the various simulation waveforms were exported to MATLAB for signal processing. Tables | an¢ 2 give the parameters Table ANN raining simulate Frulccesstance(@) 030,300 0.20,100 ule inception ange Single phase-o-round (1 Paez} hwo pace We so-so (2 he Tree phase singe phaceto-ground "Phe two phase (2 Phi Two pase-to-ground (2 Poth and Taree phase Fault ype200 AC Adewoleet al Aptied seft Computing 452018) 295-305 #7] F ei] |a a Bogs i BEL WEEE fe] PRE AN ie Ta Ses etenon ig. 4, Functional Bock eagram ofthe propased hybrid dagnestc method 01 used in the generation ofthe dataset for the training and test ‘the F5I and FL ANN models respectively 3.3, Feature extraction and selection ‘The three phase and zero sequence line currents from a current ‘transformer located at the substation relaying point of the distribu- tion network given in Fi, 5 is exported to the MATLAB computation platform. The waveforms are sampled at 128 samples per cycle, and decomposed using Daubechies db level-6 into their detail and ‘approximation coefficients respectively “The high pass (g) and low pass (h) filters ofthe dba have four coefficients. These are [4.45 $1 0.1294, gp 0.2241, g) =0.8365, g¢=0.4830 (10) tabte2 ultresstance(@)—05,.25,5 os.2ss uultineeponangle 0.30, 45,60, 90, 030-45 60,90 ult type Single phasetoground single phaseto-aroune (Phe) Twe phase (2 Pai Two phaseo-ground (2 phase-t-ground (2 Phos) and Three phase Phos) ane Thre phe (1-2 Two phase Ph).tWwo TPR ANN ANN Pg ANN ANN 1000 4 2 wos ne 1] dag seo axor010 age ooo. 0110 a rar hy = 0.4820, fy = 0.8365, fy = -0.2241, hy = 0.1294 (11) Level-5 DWT detail was selected and the Entropy Per Unit ‘obtained was computed for the 3 phase and zero sequence line ‘currents using Eq. (9). These features were selected because the coefficients obtained from the level-5 DWT decomposition gave the best performance in all the simulations carried out. This is supported by the fact that at level-5, the dominant non-frequency ‘transient thats generated by the faultsis observable within the fre- ‘quency range of 120-240 Hz based on the Nyquist criterion. Also, the wavelet energies and entropy obtained at level-s were dis- tinctive and consistent, Hence, it was unnecessary to utilize the coefficients from the other scales. lable 3 gives the frequency range for the various decomposition levels based on the sampling fre- {quency used in the decomposition of the fault signal ‘The identification ofthe faulted section and the location of the fault are then predicted based on the type of the fault, Table 4 ives the size of the dataset generated for the training and test- Db 480-960 be ascasoAC Adewele et al Apple Soft Computing 45 (2018) 25-108 ao Tabled ables Sma ofthe ening and tes tse forFSL and Fl |ANN parameters forthe PSL and FL tasks, OPE Suectwniningdataset Sueoliet dates Parmetens I 7 Single phaseto-giound 500 100) Tnpatneurens a 4 400 100 Output neurons 1 400 100 lecen layer neuron 55:28 (ave) ing of the FSI and Fl. ANN models using time-domain simulations in DIgSILENT PowerFactory software. 3:4. ANN implementation In MATLAB, the implementation of the neural networks is pet= formed by means of matrix manipulation of the inputs. outputs, synaptic weights and bias vectors. The training datasets presented tothe ANN sequentially using batch-mode supervised learning, The ‘weights andthe bias ofthe network are updated based on the error between the network output and the target. The batch-mode iter- ations continue until the training error obtained is minimized. A search for the number of hidden layer neurons was done to deter~ ‘mine the number of hidden layer neurons fr the 1 Ph-g.2 Ph. 2 Phi-g, and 3 ph, ANN models. Also, the number of epochs was deter ‘mined by experimentation. The number of epoch forthe training, was set between the range of $00 and 3500. An epoch of 500 means, the weights are updated withthe learning nule continuously until, the input dataset has been presented 500 times. The training is, repeated for 10 trials, ancl a Mean Squared Error (MSE) rate of e-03, ‘was used. In the early stopping method, the dataset is randomly, divided into three pats: training set (70x of the dataset), valida- tion set (15% of the dataset), and test set (15% of the dataset), The datasets selected forthe neural network training are normalized to the range (1, 1) using Eq. (12) below: [min (y)_ Trax = min OO) where min (I) and max (0) refer to the minimum and maximum values ofthe input attribute Levenberg-Marquardt (L-M) algorithm was used in updating the weight and bias ofthe ANN models. The derivation of -Maigo- rithm starts withthe Gauss-Newton (G-N) algorithm. Thisis given, by [46.47 b=? 1 2) wit wh Tay tet a3) where etistheerrorcostfunction,g = 1, 2, .., Bistheiteration index, Fis the maximum number of epochs (iterations), w is the weight vector. The Hessian matrix H and the Jacobian matrix are related by’ as) a4) Im order to make the Hessian matrix invertible, an approxima- ton factor known asthe identity matrix is added to it. Such that: Has} e ut as) By combining Eqs. (13) and (15), the update rule is given as: WH yy GT an ytet as) where jis the combination coefficient. ‘The performance of the trained: FSI ANN model is tested using. the test dataset. Performance analysis is carried out based on the size of the trained neural network, confusion matrix, classification accuracy, and ability to generalize to untrained datasets. Table 5 presents some of the parameters used in the design of the ANN models for the FSI task Training algorithm Sealed conjugate Levenberg-Marquarét aracient(s¢) tM) Aatiexton foncson fans, tans (hieeen layer Aeston fonction toni reli wer update methad_Bateh-mede Bateh-mede Binary notation represeating the vious line sections. For the FSI task, the ANN models use four inputs obtained from the Entropy Per Unit (EPU) of the three phase and zero sequence line current DWF coetficients. The output of the FSI ANN models is, ‘the fault section in the distribution network denoted by the binary notations in Fig. 5. The binary notation used is given in Table 6. Similarly, for the FL task. the ANNs use four inputs obtained from the Entropy Per Unit (EPU) of the three phase and zero sequence line eurrent DWT coefficients, The output of the FL ANN models is the location of the fault in kilometres, The absolute relative error (RE) is calculated using Eq. (77), ang it isa function of the actual faut location and the total length ofthe line. Unlike a transmission network, it would be erroneous to use the total length of the distribution network or the total length of the main feeder in Eq, (17). In this regard, the length of the lat- eral where the fault occurted is taken as the length of the line in computing the absolute relative error. The length for the various line segments i calculated from the beginning of the feeder (node £800) to the end of the lateral. Table 7 gives the total length of the various segments in the test feeder. |Actual Location — Estimated Location} RE, Tah ore x100% (17) table? Feederitera nats 3 (Geert 6-822 4 (teal 820826 bere 5 Literal 540856 8394 6 Lateral 2890 55520 7 teal 54864 54599 * Literal 40808 1350 ° (teal 8-838 Seon22 AC Adewoleet al Ape oft Computing 452016) 295-105 “Ol "| mm Wh Th wae “ae | ae — “Se tle] jee spelen] spite 7 te alle i Fon) a a0 a a ~ i ys ee] of A bed df yan opp | ofp thf | ole pe oa aa“ cn 7900, mea So one 1900, Fig 6, Wavelet decomposition ofA ult line 905-808 using} db mother wavelet) db2mathes wavelet} bs mother wavelet and Coie-3 mother wavelet 4, Results and discussion Extensive simulation studies have been carried out to test and validate the proposed hybrid method. However, only some of the results of the simulation studies carried out ate reported due to lack of space. Fig, 5 shows the DWT decomposition using various ‘mother wavelets for a1 Ph (phase A-to-ground) fault atthe line segment 806-808, where SLG-A is the single line-to-ground fault for phase A of the distribution network, Line 806-808 is at 10% of the length of the main feeder. Spikes are noticeable at the instant ‘of the fault inception, ‘Table 8 shows the wavelet energy spectrum entropy computed for the above-mentioned fault using Eq, (9). From Fig. 6 it ean be seen that the detail signal at level-S gave the best reflection of Sales we we Wik Wik Tee asa 37482 ais 35551 Level 2osa7 Sm aia zia0e Levels desea 25038 3.078 aims Levelt dist aan 44197 aema aves 37374 Sioi 45261 55526 the original transient signal before and after the inception of the fault Also, exploratory investigations cartied out for various levels, ‘of decomposition like in Table 8 showed that the results obtained at level-5 ofthe DWT decomposition demonstrated consistency for the various scenarios investigated. Ths is indicated by the low rel- ative wavelet energy spectrum entropy computed for the faulted phase. For the FSI task, the ANN training using the Scaled Conjugate Gradient (SCG) training algorithm was faster and gave better per~ ormance than the Levenberg-Marquarde(L-M) training algorithm, The confusion matrices obtained during the training ofthe 1 Ph. and 3h, ANN models are given in Fig. 7-The green squares give the cor- rect response per prediction class for the ANN models, while the red ‘squares indicate the incorrect responses, The accuracy per output, lassi given by the grey squares, and the blue squares present the ‘overall accuracy of the ANN model. The processor and CPU usages ‘obtained during the training of the FSI ANN model for 1 Ph is shown in Fig. . “Tables 9 and 10 present a comparative analysis of the perfor- mance of the ANN models obtained using the EPU and WEE inputs for the training of 1 Ph and 3 Ph. FSI ANN models respectively From the tables, it can be seen that the ANN model trained using EPU inputs generally gave better prediction accuracy, had faster computation time, and used less processor and memory comparedAC Aer /Apled Sa cmqung (01) 296-305 so conta a ontaon a a se =o ‘st cia o © (7 enfsn maces Phe ANN ede (4016-4 a8 ()3, SAN 20-4) mas ‘» g hoz Eon z } hot a bog ‘eons 128550 asia TRBEOO T5606 Fe. Pacer (PU) waged emery wag por Pg S.A nee 20204 raw ream mde hed ox and WEE LP) Tanaris —ERipiion WEE wedsion Bina Wetnen eh men Wetman Une FU Witmen Kanon” Meats” seman — rpucton menage menunuge wurde eo eG wa rae ve ams of AW meds on and WEE pst 3) Zane” eat)” open Sempottin —anovinge eneenge use =a oat Be Bone Ma Be Oh tea 3 mae OO BS att 2 BSB rae yee or AW model Tape a ae Ts Cea Taran Side pisciogwnd io O01 ma 2 iweptue ‘eal covet s 3 Toph serene oa comme x BS Teese pil seus % 3304 ‘AC Adewoe etal Ape Soft Computing 45 (2016) 286-205 Comparative analysis of ANN models bared on PU ane WEE pute or (1 Ph) TAN suucve —ErUsepieson Wibregresion EfUmean Wikmean EhUmen _Wifmen _ EPUmenGP WEFmenGPD “ times fines) 8) st) : 8.96860 osna bee 405 jaoai paso 70 hes Comparative amass of ANN dale see on FU ané WEF np for FL Ph fecticientCeliciest computation computation memoty usage emery wage usage (2 sage) tines line's co) MB) ‘Smary of esuls for FLANN made Sige phase-e-ground ‘a0 958 jon Twophase ass 09976 1301070 Two pase-o-gzound pererey B00 o5s/010 he phase 104 8.9999 038/001 Performance analysis ofthe Fan FL algorithms. a mae en ie os Be 25 so oro aa0sse oro diane 1 ae 25 o om 534525 om 523596 re Be 05 a too 562737 100 secae7 os? ce 25 30 101 3251 101 sas700 ced ce so 30 101 srs23i 101 387008 1 rs a5 00 100 sears? 1000 Saas? 187 Me 05 30 101 Srasz0 101 Sass ost re 05 30 iio 7.4807 iio siasis ost Bc os 30 tia Siaso7 oo $8339 130 cae 05 x0 too siasa7 too sian oa ae 100 so ou sasais 1001 sags? 160 tothe ANN models trained using WEE inputs. The final ANN models ‘chosen for the FSI task are given in Table I. The format used for the final structutes ofthe ANN models Table 11 is Input-Hiddenlayer- Output. For example, for single phase-to-ground fats, the FSI ANN model consists of 4 neurons in the input layer, 2 hidden layers with 10 neurons each, and 4 neurons in the output layer respectively ‘The final ANN models were selected based on the prediction accu racy obtained, computation time, and processor}memory usages. Its important that the chosen ANN models have lower footprints in terms of computation time and processor/memory usages. Tis is necessary for real-time implementation, field deployment, and ANN retraining, For the FL task, the ANN models trained using WEE inputs gave more accurate results compared to ANN models trained using EPU inputs as shown in Tables 12 and 13 for the 1 Ph and 3 Ph. ANN ‘models respectively. However, the computation time and mean processor usage were greater for the WEE-based ANN models com= pared to the EPU-based ANN models. The final ANN models chosen for the FL task are given in Table 14 ‘Table 15 presents the performance test of the method using untrained dataset for fault locations, fault inception angles, and fault resistances different from those used in training the ANN ‘models. Very promising results were obtained. The FSI ANN model for 1 Phg faults misclassified adjacent faulted sections (lateral {836-838 and lateral 836-840) at high fault resistance as shown, by the text in bold in Table 15. However. the predicted fault loca- tion with 1.60 relative percentage error was still within acceptable margin based on the specifications of commercially available fault locators implemented in substation Intelligent Electronic Devices (GEDs)/protection relays, which are usually specified witiy a max-AC Adewele eal Apple Soft Computing 45 (2018) 25-108 205 ‘anlers Peformance comparison with sme existing methods x Fiopesedmeioa 180) Innpeeance-basd 13] 155 Impeeane-sased 12] 601 Impeeancesased 15] i Inpeeance-sased 13] 135 ANN a inpedance-based 25) 437 ANNSased 32, ec considered imum relative percentage error of 5.0 as given in Refs. (48,49 ‘Table 16 gives a comparative analysis of the proposed method with some existing methods inthe literature 5, Conclusion This paper has developed a hybrid 2-stage method for distribu- ton network fault section identification (FSI) and fault location (FL) based on the coefficients from level-5 detail coefficients obtained from DWT decomposition using db4 mother wavelet. Wavelet Energy Spectrum Entropy (WEE) and Entropy Per Unit (EPU) indices are computed from the DWT detail coefficients. These indices are used in training artificial neural network models for the FSI and FL tasks respectively. Comparison of ANN models trained using ‘the EPU and WEE indices is carried out in terms of the prediction accuracy, computation time, processor usage, and memory usage. In order to validate the proposed hybrid method, it is applied to the IEEE 34-node benchmark test feeder, The proposed method can easily be implemented practicaly using actual data obtained from Digital Fault Recorders (DER) or intelligent Electronic Devices EDs), Although, data obtained from DFRS/IEDS are usually noisy, the noise would be filtered out as a result of the DWT decom- position. The ANN models trained using EPU indices were shown, to require less computer memory, less processor usage, and gave faster computation speed ‘The robustness of the proposed hybrid method to various fault resistance, fault types, fault inception angle, and fault location was demonstrated, Accurate results irrespective of fault locations fault resistances, fault inception angles, and fault types were obtained for the FSI and FL tasks respectively. The proposed 2-stage hybrid ‘method is suitable for practical deployment in distribution network control centres because of its simplicity and accuracy, and can be used to aid engineers and control centre dispatchers in fault diag nosis. It should be noted that since the proposed hybrid method uses only line current signals available {rom current transformers already existing in substations, there is no added cost involved in {he roll-out (implementation) ofthe proposed hybrid method, Possible future extension of the proposed hybrid method could be in the use of synchrophasor measurements from Phasor Mea- surement Units (PMUs). Also, machine learning classifiers and predictors based on decision trees/ensembles of decision trees can, be applied to identify the faulted section and the location of the fault in the power system distribution network. Acknowledgements This research work is funded by the South Aftican National Research Foundation (NRF) THRIP Grant TP2011061100004 “CSAEMS development and growth’. The authors are grateful for the financial support. References (34 a ows alow Ft ean Pe New rnc ain neg Foe eda fe aldeeion sd sees cote ee ea San at (2 FUSSY many Yen Renin Marais caged Teun peter een ee nce (a Lani NESS Rodi oak caw ‘spl’ pet aa om 15] AT Johns, 5, Jamal. Accurate fault location technique for pawer transmission (sweet? Varta any Unset ve en a Le eae aa incrbceurar ies et Fa eee 080) oa (0 Lies ler hea re Baie Lap ke Sante eit it east ulead aes witness stra Eager gestae (ys ap ce tid clang alse asmison intel eget nuit inner coo on Foe en Fe eae ncn cist od babsentnes tin pn etek vcd faa ere Soe [12] AA Giegis. CM. Fallon, DL Lubkerman, A faut location technique fer rucal Coat patie Sard Fa tenor ero Beane ote ee Sk es at mag tis ROTEAEE Soclaet eet ia eee ci lect mee sienna obese ra a (om Cede Mares Vrain fil ‘sti i emt per pens re eset 118] RH. Salim, M, Resener, AD. Filomena, KK. Caine De Oliveira AS. Breas, toy Aaa tie tts. aes Fa can ranastae tot Maga ou Ane cn enue aa rbson ‘penta oe pa Pe apc Se va Bw Et re attire Technology Conference (2003}468~472, (ay rst ca sc ate ee cepa (st EA aie dove Abeer sen AS, eta Cee eer ee tet restr ietdtad Be eet SO ME as, Tc CE Santina sss leben Reventon news sd So) tees Sh Se Sang Adon (a We a as sede tat tection Sino ores telson naan sels cae ee Se a yy eae208 AC Adewoeet al / Ape soft Computing 45 (2016) 295-105 (s] ¥en tksoneege gusto an eps ale petri, lets Power Sse Res 27 (1993) [30] EA Mfchameg. LD. Rao, Arial neural neswork ful diagnostic system or ‘lecc power dributon feeders Electr Power Sys Res 35 (1993) 110, [at] Dw: chan, c2 tu, Dusibatin system fat centcaton by mapping ot fharacertic vector let Pe Sy Res 57 (2001) 15-23, [aa] MUAisshaher MM Sabra, AS Sale Fait locaton in ling station etwork using areal neural network lets Powe Syst Kes 542003 [33] LS Marsins, VE. Pes, CM.Aegi, Anew acute elt nation method ‘ring space vector grit Froceding of (ath SCC 32002" 1-5 [pa] [ Mora-Rérez cormane-anganta Orde 'aez-Pats Kemeans algriam dnd mc cissbtlons fo loca alts power syste ect Power Spates 792008) 714-724 [35] D Thokaram. 1 Kinc, . Vjaynarasinha Atif mura network and oppor vector machine approach fer locting fut in asa deebion| ‘yates IEE Trane, Power Del 20 (2003) 710-721, (et haa iL Chan one 2 nnn Aten of want fzry netlnetoock in locting singe hie to ound fal (SLC). ‘iscbuon ines let Power Emery Sy 29 (2007 487-505, [37] AC Adewole.tnvestizauon of Metodloes or aut Deection and Diagnosis in lecac Power SytenFretecaon. Mase’ Thess Departmest of eital Engineering, ape Penisula Univers of Tehlog, Sell A aon [36] Ac Adewoe. Teeneva Distsbaton network ak detection and diagnos thing wavelorm energy spectsum entapy and neural nesworas In ke He ag 9 (2004) 165-173, [39] 5 sumaniaay BX Panga, PAC Dash High impedance fault detection in power dsribtian networks using timestequeney taster and obabilc eur etwork ET, Generation, Tranemision nd Distebton 140] hers Ling. hens 8. nan Fal tection and clasietonn E'raneminonne bare on waver snglarentapy, EE rane Power Dee 25 0) oe-2163 Ue Sek ein M Peta, ney and entepy dase feature extraction lecting fal on tassion ies by sie eal peter aed wavelet Deke: decomposition, Exper St Appl 4 (2008 2837-2548, (4a Bi rng Pa Opin ere ston or csiestion a ‘aly dstbances using wavelet packet bse faz kenearestnelghber Sirti et Conran, aaron Dtrnuson 2 (2008 296-30 4a Argh bore DiSiveta CA Nowe Wt Pasone Contncos-vaveles eansorm ort action n stron over Nerwors Dedniton of ther Wavelet ited om Ful orga {44 bnabeches Ten ectoreyon Waele IAM, 1992. TBs eb 34 Noe et eer 2 8S Dabo ye Aras Subconmie'sDatuton Let Feeder Waking Croup acesied 1001.10 {45 Hebemath i: ele gan Neural erent foolbonor Use wi aviab Uses Cide Veron 42064 1471 Ent nap Nea Neon for ater Recognition, Oxon University tres 1998 U4] A Sanat A Jonis M Dawa Sai adaptive gots orf Teal ot power aniston ines asd ontetod of Mone-Caf Proc ih nternational Conference on robbie Meth ApledePower Stems September 2-70 2000 Naples aly, 20 9p. 485-80. 49) Fos buptal ay Contoler nacre 214 Manual Cenerl lecic USA {aecesa350033)