e-p f w sul sim to tion system (UDS). The results indicate that the fault location technique has an acceptable accuracy (error < 1.5%) under the whole variety of different systems and fault conditions. Therefore, the proposed approach can have high performance for the evaluation of the fault location and classification. � 2013 Elsevier Ltd. All rights reserved. ides t the loa Navanithan, Soraghan, Siew, Mcpherson and Gale in [4]. However, due to the limitation of the band width of the conventional CT (up to a few GHz) and VT (up to 50 kHz), the accuracy of fault location provided by such a scheme is not satisfactory for a UDS. Also there have been many activities in using power frequency (low fre- quency) for fault location and protection. Aggarwal, Aslan and Johns in [5] present a new technique in single-ended fault location upport automatic ractical approach s using advanced luding sin h-impedan plicability proposed approach. In [8], an accurate and efficient method posed for fault section estimation and fault distance calcula distribution systems, based on frequency spectrum components of fault generated traveling waves. Simulation results of various types of faults on a typical distribution system demonstrate high effi- ciency and accuracy of the proposed method. In [10], an ANFIS (Adaptive Neural Fuzzy Inference System) based fault classification scheme in neutral non-effectively grounded distribution system is proposed. The results show that it has high accuracy and through simulation, the proposed approach exhibits good performance. ⇑ Corresponding author. Tel.: +98 918 8784368; fax: +98 641 2243271. Electrical Power and Energy Systems 55 (2014) 261–274 Contents lists availab n .e l E-mail address:
[email protected] (J. Moshtagh). fault generated traveling wave methods for fault location and pro- tection [3]. The traveling wave current-based fault location scheme have been developed for permanent faults in underground low voltage DSs in which the distance to fault is determined by the time differences measured at the sending end between an incident wave and the corresponding wave reflected from the fault by indicate the possibility of using this method to s fault management system. Ref. [7] introduces a p to power system fault location in power network fault signal processing. The simulation results, inc to ground faults, faults in mixed feeders and hig ing faults, confirm the accuracy and practical ap 0142-0615/$ - see front matter � 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.ijepes.2013.09.011 gle line ce arc- of the is pro- tion in connection among consumers, generation and transmission sys- tems. Typically radial in nature, the distribution system includes feeders and laterals [1]. The distribution voltages in a specific ser- vice territory are likely similar because it is easier and more cost effective to stock spare parts when the system voltages are consis- tent [2]. In recent years, there have been many activities in using However, in such techniques which are based on power frequency signals, some useful information associated with high frequencies in transient condition is missed. Paper [6], presents the application of calculated non-linear volt- age sag profiles and voltage sag measurement at primary substa- tion to locate a fault in distribution networks. The results Keywords: Fault location Ground faults Underground distribution system Wavelet analysis Neural network Fuzzy logic 1. Introduction The distribution system (DS) prov liver power from the substations to he infrastructure to de- ds. These systems have for overhead DSs, which is based on the concept of superimposed components of voltages and currents rather than total quantities and also special filtering technique have been utilized to accurately extract the fundamental phasors from the measured fault signals. 2013 Accepted 24 September 2013 (ANN) and the fuzzy logic system (FLS) is then used to detect the type and the location of the ground high impedance, ungrounded series, ungrounded and ground shunt faults in a practical underground distribu- A new approach to fault location in thre distribution system using combination o and FLS Ali Rafinia, Jamal Moshtagh ⇑ Department of Electrical Engineering, University of Kurdistan, Sanandaj, Iran a r t i c l e i n f o Article history: Received 20 February 2013 Received in revised form 17 September a b s t r a c t This paper presents the re using EMTP software. The based on wavelet analysis Electrical Power a journal homepage: www hase underground avelet analysis with ANN ts of investigation into a new fault classification and location technique, ulated data is then analyzed using advanced signal processing technique extract useful information from the signals. The artificial neural network le at ScienceDirect d Energy Systems sevier .com/locate / i jepes This paper presents a new off-line method fault location based on signal processing using wavelet and ANNs and FLSs in UDS. A practical 20 kV underground power distribution system is simu- lated using the EMTP software; the faulted current and voltage re- sponses are then extracted from the sending end for different faults and fault conditions. The effect of transducers (CTs and VTs) and hardware errors such as anti-aliasing filters and quantization are taken into account; the information processed throughout the fault locator algorithm is thus very close to real-life situation. Finally, the simulated data is processed in order to locate the fault point using ANNs and FLSs. 2. Data simulation In order to obtain the voltage and the current signals under dif- ferent faults and conditions, a practical three-phase UDS shown in Fig. 1 has been considered. In this paper, the simulation of the quantization process is based on 16-bit A/D converter with ±10 V by using MATLAB pro- gram. In order to keep the voltage and current signals in range ±10v, these signals are divided by 2200 and 700 respectively which are 1/10 of maximum amount of voltage and current signals under all conditions. It is apparent that both the steady and transient states of the voltage and current signals can be affected by some important parameters such as the type of fault, inception angle, faulted branch and distance to fault for ungrounded and grounded faults. In order to obtain useful information from signals in the signal pro- cessing stage and mapping the extracted information to the loca- variable resistance including single-phase to ground high imped- ance fault (3 cases: ag-hi, bg-hi, cg-hi), two-phase to ground high impedance fault (3 cases: abg-hi, acg-hi and bcg-hi), three-phase to high impedance fault (abcg-hi), three types of open-circuit fault (oc) including one-phase open-circuit fault (3 cases: a-oc, b-oc and c-oc), two-phase open-circuit fault (3 cases: ab-oc, ac-oc and bc- oc), three-phase open-circuit fault (one case: abc-oc), three types of short-circuit fault (sc) including phase to phase short-circuit fault (3 cases: ab-sc, ac-sc and bc-sc), three-phase short-circuit fault (one case: abc-sc) and three types of grounded short-circuit fault including single phase to ground short-circuit fault (3 cases: ag-sc, bg-sc and cg-sc), two-phase to ground short-circuit fault (3 cases: abg-sc, acg-sc and bcg-sc) and 3-phase to ground short-cir- cuit fault (one cases: abcg-sc) also 3 fault resistance in the case of ground short-circuit fault (0.1X, 1X and 10X), three inception angles (including 90�, 135� and 180�) and 17 distances of fault from recording point (including branch 1:0 m, 100 m, 1300 m, 3850 m, 5100 m, 7900 m, 11,200, 12,900 m; branch 2: 6200 m, 10,200 m; branch 3:8800 m, 9500 m; branch 4:12,000 m, 12,900 m, 13,800 m, 14,800 m and branch 5: 15,500 m, 16,500 m) are simulated. Figs. 2 and 3, show the three-phase voltages and the three- phase currents after applying ground high impedance fault on phases b and c, open-circuit fault on phases a, b and c in 5100 m from the source, respectively. These faults are applied in two de- grees of 90 and 135 for the system phases. Figs. 4 and 5, show the three-phase voltages and the three-phase currents for the two-phase fault (bcg-sc). As shown in Fig. 4, the ini- tial disturbance in case of fault at 90� is muchmore than 180�. Fig. 5 shows the three-phase current signals after two-phase to ground 262 A. Rafinia, J. Moshtagh / Electrical Power and Energy Systems 55 (2014) 261–274 tion of fault in artificial intelligent (AI) stage, it is necessary to obtain voltage and current signals, in different fault types and dif- ferent conditions in the data simulation stage. In this respect, three types of ground high impedance fault (hi) with electrical arc and Fig. 1. Practical 3-phase undergr high impedance fault event, as the current amplitude in two-phase fault increase significantly. Also the initial current amplitude in the case of fault at 90� is much more than 180�. ound distribution network. r an A. Rafinia, J. Moshtagh / Electrical Powe Figs. 6 and 7 depict the three-phase voltages and the three- phase currents respectively, in the cases of location at 100 m and 5100 m. Fig. 6 shows the voltage signals and it can be seen that the initial distortions are much higher and the transients die down much more slowly in the case of longer distance faults. Fig. 7 shows the current signals and it can be observed that the initial Fig. 2. Voltage and current signals, bc-hi fault, L = 5100 m, Inception angle = 90�. Fig. 3. Voltage and current signals, abc-oc fault, L = 5100 m, inception angle = 135�. Fig. 4. Voltage signals, bcg-sc fault, Rf = 0.1X, L = 5100 m, inception angle = 90 and 180�. Fig. 5. Current signals, bcg-sc fault, Rf = 0.1X, L = 5100 m, inception angle = 90 and 180�. d Energy Systems 55 (2014) 261–274 263 distortions are much smaller compared to the voltage signals. Importantly, the currents of the three faulted phases increase after occurring the fault but are much smaller in the case of location at 5100 m to 100 m fault. 3. Feature extraction using wavelet Transient signal analysis has been extensively used in fault location and condition monitoring of power system lines and cables. The time and frequency information can be calculated using techniques such as Fast Fourier Transform (FFT), Short Time Fou- rier Transform (STFT) and Wavelet Transform (WT). FFT and STFT techniques yield good information on the frequency content of the transient, but the time at which a particular disturbance in the signal occurred is lost. In this paper, a new approach based on feature extraction using the WT is presented. WT possesses some unique features that make it very suitable for this particular application. It maps a given function from the time domain into time-scale domain. Unlike the basis function used in Fourier analysis, the wavelets are not only localized in frequency but also in time. This localization allows Fig. 6. Voltage signals, abc-sc fault, L = 100 m and 5100 m, inception angle = 135�. decaying and oscillating type of high and low frequency voltage and current signals. One of the most popular mother wavelets suit- able for a wide range of applications used is Daubichies’s wavelet. In this respect, db4 wavelet with 8 level of decomposing of signals has been considered herein [13]. Table 1 gives the frequency band information for the different scale of the wavelet analysis, where the sampling frequency is Fs = 100 kHz. There are very useful features in the signals, particu- larly in the approximate part of the signals, as its frequency band is r and Energy Systems 55 (2014) 261–274 the detection of the time of occurrence of abrupt disturbances, such as fault transients. 3.1. Wavelet Transform In the case of WT, the analyzing function, which is called wave- lets, will adjust their time-widths to their frequency in such a way that higher frequency wavelets will be very narrow and lower fre- quency ones will be broader. This property of multi-resolution is particularly useful for analyzing fault transients which localize high frequency components superposed on power frequency sig- nals (Manago and Abur [11]). WT of sampled waveforms can be ob- tained by implementation the discrete WT which is given by the following equation: DWTðf ;m;nÞ ¼ 1ffiffiffiffiffiffi m p Xf ðkÞh� n� ka m 0 m � � ð1Þ Fig. 7. Current signals, abc-sc fault, L = 100 m and 5100 m, inception angle = 135�. 264 A. Rafinia, J. Moshtagh / Electrical Powe a0 k a0 where the parameters am0 and ka m 0 are scaling and translation con- stant respectively, k and m being integer variables and h is the wavelet function which may not be real, as assumed in the above equation for simplicity. In a standard discrete WT (DWT), the coef- ficients are sampled from the continuous WT on a dyadic grid, a0 = 2, yielding a00 ¼ 1, a�10 ¼ 1=2, etc. Actual implementation of the (DWT) involves successive pairs of high-pass and low-pass fil- ters at each scaling stage of the WT. At each detail, there is a signal appearing at the filter output at the same sample rate F and scaling by two (a0 = 2), Eq. (2) shows the association of each scale 2m with a frequency band containing distinct components of signals. Frequency band of scale 2m ¼ F=2mþ2!to F=2mþ1 ð2Þ In this paper the original signals have been sampled at 100 kHz and passed through a DWT; thus according to Eq. (2) the frequency band for detailed and approximate signals are; 25 kHz to 50 kHz at detail-1, 12.5 kHz to 25 kHz at detail-2, etc. [12]. 3.2. Choice of mother wavelet Choosing of mother wavelets plays an important role in localiz- ing and depends on a particular application. Researches, in the study of underground power distribution transients are particu- larly interested in detecting and analyzing short duration, fast 3.3.1. Feature extraction in fault classification The process comprises two stages including; (1) Fault classifica- tion (2) Fault location [15–17]. At first, the original signals are Table 1 Different scale of wavelet analysis. Detail Frequency band Approximation Frequency band D1 25–50 kHz A1 0–25 kHz D2 12.5–25 kHz A2 0–12.5 kHz D3 6.25–12.5 kHz A3 0–6.25 kHz D4 3.125–6.25 kHz A4 0–3.125 kHz D5 1.5625–3.125 kHz A5 0–1.5625 kHz D6 0.78125–1.5625 kHz A6 0–0.78125 kHz DC to 195.3125 Hz. Thus in this study, approximate-8 signal is con- sidered for further feature analysis and this pattern is used for all each signal including the faulted and healthy phases. 3.3. Feature extraction using statistical relations In this paper, the statistical relations were used to obtain a suit- able increasing or decreasing model to be used as input for the training of the neural network. The situation in fuzzy logic is also similar to that was employed for neural networks. Therefore, this approach tried to use mathematical concepts and relationships to determine the best performance in finding a suitable fault location and classification. The definitions of some statistical concepts are as follows [14] (The equations were also excluded to avoid being bulky paper): � Mean: The arithmetic mean (typically referred to as the mean) is the most common measure of central tendency. The mean is the only common measure in which all the values play an equal role. The mean serves as a ‘‘balance point’’ in a set of data. � Median: The median is the value that splits a ranked set of data into two equal parts. The median is not affected by extreme val- ues, so you can use the median when extreme values are pres- ent. The median is the middle value in a set of data that has been ordered from lowest to highest value. � Mode: The mode of a set of data is the value which occurs most frequently. Like the median and unlike the mean, extreme val- ues do not affect the mode. � Skewness: In probability theory and statistics, skewness is a measure of the extent to which a probability distribution of a real-valued random variable ‘‘leans’’ to one side of the mean. � Correlation coefficients: The coefficient of correlation measures the relative strength of a linear relationship between two numerical variables. � Central moment: In probability theory and statistics, a central moment is a moment of a probability distribution of a random variable about the random variable’s mean; that is, it is the expected value of a specified integer power of the deviation of the random variable from the mean. D7 390.225–781.25 Hz A7 0–390.225 Hz D8 195.3125–390.625 Hz A8 0–195.3125 Hz passed through a DWT then 8-detailed and 1-approximate signals are extracted. With regard to statistics option in wavelet and data processing on approximate signals of the voltage and current phases, it was observed that some useful information can be extracted from stan- dard deviation (STD) of approximate-8 signals in fault classifica- tion, since the amount of STD for every input data with dimension 6 (three voltage phases and three current phases) has an obvious relationship with the type of fault and faulted phases. STD equation is as the following equation: STDðvÞ ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 n� 1 Xn i¼1 ðv i � �vÞ2 vuut ð3Þ where vi is the ith sample of signal v, n is the number of samples and �v is the average of the samples. In order to classify the fault type, both the voltage and current phase signals are employed. Firstly, the original signals are passed through a DWT, then eight detailed and one approximate signals are extracted. Since the number of samples for the original signal is 4001, thus the numbers of coefficients for the decomposed sig- nals are 2004, 1005, 506, 256, 131, 69, 38, 22 and 22 for details at levels 1–8 and approximate at level 8, respectively. After recognizing the type of fault, STD of voltage phases, cur- rent phases are used for all types of faults to recognize the faulted phases and whether it is a grounded or ungrounded fault. Figs. 8–11 show the STD of voltage phase and current phase sig- nals for the 51 conditions in the case of abg-hi, abc-oc, abc-sc faults and the 153 conditions in the case of ag-sc. Also, these figures show such data which is used in the fault classification associated with the type of fault and faulted phases. Each figure comprises two graph associated with voltage phases and current phases. Each graph shows three waveforms related to the three phases and each waveform depicts the STD of approximate-8 of signal for the all conditions dealt with in the previous section. Also, each waveform contains 3 separate the parts. Each part corresponds to the 17 loca- tions and the same inception angle. As it can be seen, there is a sig- nificant difference between the faulted phases and healthy phases. In this stage, the feature of STD has been used because STD of voltage in the faulted phases increases and also STD of current in the faulted phases decreases in each part as the fault distance in- creases from the measurement point while it is more or less con- stant for healthy phases. For example, in Fig. 11 although the difference between the maximum STD and minimum STD de- creases, in particular the level of STD of the faulted phase ap- proaches those in the healthy phases as Rf increases, there is still a discernable difference in the STD levels between the faulted and healthy phases. 3.3.2. Feature extraction in finding the faulty branch Since the proposed network has five branches, the location of the fault should be considered in two stages;(1) selecting a branch sign A. Rafinia, J. Moshtagh / Electrical Power and Energy Systems 55 (2014) 261–274 265 Fig. 8. STD of approximate-8 Fig. 9. STD of approximate-8 sign al in the case of abg-hi fault. al in the case of abc-oc fault. (1) Ratio of voltage approximate skewness to STD of the current approximate at level 8. sign 266 A. Rafinia, J. Moshtagh / Electrical Power and Energy Systems 55 (2014) 261–274 (2) Ratio of the square of skewness current approximate to square of skewness voltage approximate at level 8. (3) The absolute value of central moment voltage approximate for the elements of 3 at level 8 (see Fig. 12). It should be mentioned that these six factors are employed only for the faulted phases in the case of a phase to phase fault, but only phase ‘a’ is considered in the case of three phase fault. In order to obtain more accurate results, the signals are normalized according to the following equation: Xnormed ¼ X � XminXmax ð4Þ 3.3.3. Feature extraction to find the fault distance from the source Similar to previous studies, it can be seen that there are very useful features in the decomposed voltage and current signals using the DWT. These features are considered to determine the which fault has been occurred, (2) locating fault (distance from the source). Thus, three features have been used. These three features are as follows: Fig. 10. STD of approximate-8 location of all type of faults and only used faulty phases. These fea- tures are: Fig. 11. STD of approximate-8 sig 3.3.3.1. Ground high impedance fault (hi). (1) Ratio of voltage approximate variance to current approxi- mate variance at level 8. (2) Ratio of absolute square of voltage approximate mode to absolute square of current approximate mode at level 8. (3) Ratio of central moment of voltage approximate for the ele- ments of 2 to central moment of current approximate for the elements of 2 at level 8. 3.3.3.2. Short-circuit fault (sc). Three features similar to section a. 3.3.3.3. Ground short-circuit fault (g-sc). Three features similar to section a b. 3.3.3.4. Open-circuit fault (oc). (1) Maximum of square of mean correlation coefficients between voltage and current approximate at level 8. (2) STD of absolute the median values for elements along the dimension of current approximate by 2 at level 8. (3) Mean of absolute the median values for elements along the dimension of current approximate by 2 at level 8. (4) Square of central moment of voltage approximate for the elements of 2 at level 8. al in the case of abc-sc fault. Figs. 13–15 show the results for 51 fault conditions, associated with three-phase to ground, ungrounded three-phase and phase-a nal in the case of ag-sc fault. Fig. 12. Three parameters used in the finding faulty branch stage. A. Rafinia, J. Moshtagh / Electrical Power and Energy Systems 55 (2014) 261–274 267 to ground faults. Figs. 13 and 14 depict three graphs according to 51 fault conditions associated with 17 locations and three incep- tion angles, solely for ‘a’ phase in the case of the abcg-hi and a-sc faults. The reason for only selecting phase ‘a’ is the similarity be- tween the three phases in terms of the considered parameters. Fig. 15 shows four features in the case of abc-oc. Each figure shows three graphs themed by the three aforementioned parameters. As mentioned before, each 17 consecutive conditions are based on the same inception angle but different locations. Also Fig. 16 de- picts the behavior of three features for the faulted phase ‘a’ to ground according to the aforementioned 153 fault conditions. All graphs comprise 9 parts related to three inception angles and three values of Rf and each part shows their behavior depending on the 17 fault locations. Each three consecutive parts correspond to the same Rf but a different inception angle. For single and two-phase fault the same process is used. It should be considered that for two-phase fault the features are related to the two phases. Fig. 13. Three parameters used in finding the dista 4. Artificial intelligence techniques 4.1. Artificial neural network ANNs have emerged as a powerful pattern recognition tech- nique and act on data by detecting some form of underlying orga- nization not explicitly given or even known by human experts and it possesses certain features which are not attainable by the con- ventional methods. In this respect, this paper describes a new method for accurate fault location based on the ANNs technique. The successful development of ANNs approaches depends on the successful learning of the correct relationship or mapping between the input and output patterns by the ANNs [19–20]. In order to achieve this, practical issues surrounding the design and testing of an ANN such as the best network size, generalizing versus mem- orization feature extraction and scaling of signals have been ad- dressed and examined. nce of fault stage in the case of abcg-hi fault. r an 268 A. Rafinia, J. Moshtagh / Electrical Powe There are many types of ANNs but the most commonly used are the multi-layer feed-forward networks, as, a three-layer network (input, one hidden and output layers). Because of this, a fully con- Fig. 14. Three parameters used in finding the dis Fig. 15. Four parameters used in finding the dista Fig. 16. Three parameters used in finding the dis d Energy Systems 55 (2014) 261–274 nected three-layer feed-forward ANNs with Levenberg–Marquardt (LM) learning algorithm has been used in the complete fault clas- sification and fault location networks. tance of fault stage in the case of a-sc fault. nce of fault stage in the case of abc-oc fault. tance of fault stage in the case of ag-sc fault. Table 2 depicts the specifications of employed ANNs in pro- posed fault location technique. Where N1 = number of training data, N2 = dimension of input layer, N3 = number of neuron in hid- den layer, N4 = number of neuron in output layer, F1(x) = transfer function in hidden layer and F2(x) = transfer function in output layer. The NNtf recognizes the type of fault and faulted phase’s un- grounded and grounded faults. The NNbf represents the branch of fault. The NNag-hi and the other similar architectures mentioned in Table 2 determine the distance of fault from the source in the case of HI, SC and OC fault. 4.2. Fuzzy logic FLS is a convenient way to map an input space to an output space with a set of common-sense rules. In almost every case, FLS can be replaced by other options, such as linear, non-linear, Neural-Network, expert system and genetic algorithm, but FLS is faster and more adaptable. Furthermore, it is conceptually easy to understand as well as is flexible, tolerant to imprecise data and is based on natural language. Also it can be built on the expe- rience of experts and can be blended with conventional control techniques. Because of previously mentioned points, in this paper, FLS has been selected to localize and identify different types of fault. Different steps are needed in the practical design of FLS in power engineering application [22]. satisfactory to merely employ a single ANN or FLS and attempt to train and design them with a large amount of data. A much bet- ter approach is to separate the problem into three parts: 1. To em- ploy FLSs and train some ANNs to classify the faults, as they indicate on which type of fault (open circuit, short circuit or high impedance) and which phase(s) the fault is and whether there is ground involved in a particular fault, irrespective of the actual fault location at this stage; 2. In order to achieve a good generalization, to use separately trained ANNs and designed FLSs (one for each type of fault and faulty phase(s)) to accurately locate the actual fault branch and 3. Using some other features, the distance of the fault from the source is measured on the underground distribution system. Fig. 17 show the fault location scheme based on ANNs and FLSs. Inputs to the neural network and the fuzzy logic for fault clas- sification, faulty branch location and the fault distance from the source are given in Sections 3.3.1, 3.3.2 and 3.3.3 respectively. 4.3. Fault classification based on ANNs and FLSs The fault type classification technique is based on training three-layer ANNs by the LM learning algorithm and also after an extensive series of studies, hyperbolic tangent and linear transfer functions were selected as the hidden and the output layer neu- rons, respectively. Table 4 shows desired ANN and FLS outputs for the fault type classification (Single-phase, two-phases and three-phases). The outputs of the ANN and FLS comprise of seven variables HI, SC, OC, A, B, C and G; of these, HI, SC and OC is asso- ciated with the type of fault (a value close to unity indicates the NNab-oc 51 4 10 1 TanSig TanSig NNac-oc 51 4 10 1 TanSig TanSig NNbc-oc 51 4 10 1 TanSig TanSig FLb-oc 51 204 17 1 trapmf trapmf A. Rafinia, J. Moshtagh / Electrical Power an NNabc-oc 51 4 10 1 TanSig TanSig NNa-sc 51 3 8 1 TanSig TanSig NNb-sc 51 3 8 1 TanSig TanSig NNc-sc 51 3 8 1 TanSig TanSig NNab-sc 51 3 8 1 TanSig TanSig NNac-sc 51 3 8 1 TanSig TanSig NNbc-sc 51 3 8 1 TanSig TanSig NNabc-sc 51 3 8 1 TanSig TanSig NNag-sc 153 3 8 1 TanSig TanSig NNbg-sc 153 3 8 1 TanSig TanSig NNcg-sc 153 3 8 1 TanSig TanSig NNabg-sc 153 3 8 1 TanSig TanSig NNacg-sc 153 3 8 1 TanSig TanSig Table 3 shows the specifications of employed FLSs in proposed fault location technique. Where M1 = number of input data, M2 = number of input MF, M4 = number of output MF, M5 = num- ber of output variables, G1(x) = input MFs and G2(x) = input MFs. The FLtf recognizes the type of fault and faulted phase’s un- grounded and grounded faults. The FLbf represents the branch of fault. The FLag-hi and the other similar architectures determine the distance of fault from the source in the case of HI, SC and OC fault. Table 2 Specifications of employed ANNs. ANN N1 N2 N3 N4 F1(x) F2(x) NNtf 2142 9 6 7 TanSig Linear NNbf 50 3 10 5 TanSig Linear NNag-hi 51 3 8 1 TanSig TanSig NNbg-hi 51 3 8 1 TanSig TanSig NNcg-hi 51 3 8 1 TanSig TanSig NNabg-hi 51 3 8 1 TanSig TanSig NNacg-hi 51 3 8 1 TanSig TanSig NNbcg-hi 51 3 8 1 TanSig TanSig NNabcg-hi 51 3 8 1 TanSig TanSig NNa-oc 51 4 10 1 TanSig TanSig NNb-oc 51 4 10 1 TanSig TanSig NNc-oc 51 4 10 1 TanSig TanSig NNbcg-sc 153 3 8 1 TanSig TanSig NNabcg-sc 153 3 8 1 TanSig TanSig In order to find the best topology for accurate fault location un- der all practically encountered different system and fault condi- tions, an extensive series of studies have revealed that it is not FLc-oc 51 204 17 1 trapmf trapmf FLab-oc 51 204 17 1 trapmf trapmf FLac-oc 51 204 17 1 trapmf trapmf FLbc-oc 51 204 17 1 trapmf trapmf FLabc-oc 51 204 17 1 trapmf trapmf FLa-sc 51 153 17 1 trapmf trapmf FLb-sc 51 153 17 1 trapmf trapmf FLc-sc 51 153 17 1 trapmf trapmf FLab-sc 51 153 17 1 trapmf trapmf FLac-sc 51 153 17 1 trapmf trapmf FLbc-sc 51 153 17 1 trapmf trapmf FLabc-sc 51 153 17 1 trapmf trapmf FLag-sc 153 459 17 1 trapmf trapmf FLbg-sc 153 459 17 1 trapmf trapmf FLcg-sc 153 459 17 1 trapmf trapmf FLabg-sc 153 459 17 1 trapmf trapmf FLacg-sc 153 459 17 1 trapmf trapmf FLbcg-sc 153 459 17 1 trapmf trapmf FLabcg-sc 153 459 17 1 trapmf trapmf Table 3 Specifications of employed FLSs. FLS M1 M2 M3 M4 G1(x) G2(x) FLtf 2142 12,852 7 7 trimf trimf FLbf 50 150 5 5 trimf trimf FLag-hi 51 153 17 1 trapmf trapmf FLbg-hi 51 153 17 1 trapmf trapmf FLcg-hi 51 153 17 1 trapmf trapmf FLabg-hi 51 153 17 1 trapmf trapmf FLacg-hi 51 153 17 1 trapmf trapmf FLbcg-hi 51 153 17 1 trapmf trapmf FLabcg-hi 51 153 17 1 trapmf trapmf FLa-oc 51 204 17 1 trapmf trapmf d Energy Systems 55 (2014) 261–274 269 type of fault), a value close to unity for any of the A, B and C corre- sponds to the appropriate a, b or c phases being faulty, respec- tively. A near unity of G signifies that ground is involved in a fault. They are all driven from the ANNs and FLSs designed to classify the fault type and the input data is generated the same way as it is mentioned in detail in previous two sections. ANNs (or FLSs) are trained (or designed) with different training data to cater for all types of commonly encountered faults. Feature Extraction Fault Classification Ground High Impedance Fault Fault Branch V,I Fault Dictance Fault Location Open-Circuit Fault Short-Circuit Fault ag- sc bg- sc cg- sc Ground Short-Circuit Fault abg -sc bcg -sc acg -sc abcg- sc ab- sc bc- sc ac- sc abc- sc a- oc b- oc c- oc ab- oc bc- oc ac- oc abc- oc ag- hi bg- hi cg- hi abg -hi bcg -hi acg -hi abcg- hi Fig. 17. Schematic diagram of fault location technique. 270 A. Rafinia, J. Moshtagh / Electrical Power and Energy Systems 55 (2014) 261–274 4.4. Finding the fault branch based on ANNs and FLSs There are five branches in the network with different length; accordingly, the proper features are extracted for ANN and FLS in- put. Three features that pointed in Section 3 are employed as input of the ANN and FLS. The five outputs represented in Table 5 shows the faulted branch. In this table a value close to unity indicates the branch of fault. 4.5. Fault distance from the source based on ANNs and FLSs As mentioned before, separate ANNs and FLSs are designed to accurately calculate fault distance from the source for each type of fault under all practically encountered different fault conditions. Table 4 Desired ANNs and FLSs outputs for fault classification. Type fault HI OC SC A B C G ag-hi 1 0 0 1 0 0 1 bg-hi 1 0 0 0 1 0 1 cg-hi 1 0 0 0 0 1 1 abg-hi 1 0 0 1 1 0 1 acg-hi 1 0 0 1 0 1 1 bcg-hi 1 0 0 0 1 1 1 abcg-hi 1 0 0 1 1 1 1 a-oc 0 1 0 1 0 0 0 b-oc 0 1 0 0 1 0 0 c-oc 0 1 0 0 0 1 0 ab-oc 0 1 0 1 1 0 0 ac-oc 0 1 0 1 0 1 0 bc-oc 0 1 0 0 1 1 0 abc-oc 0 1 0 1 1 1 0 a-sc 0 0 1 1 0 0 0 b-sc 0 0 1 0 1 0 0 c-sc 0 0 1 0 0 1 0 ab-sc 0 0 1 1 1 0 0 ac-sc 0 0 1 1 0 1 0 bc-sc 0 0 1 0 1 1 0 abc-sc 0 0 1 1 1 1 0 ag-sc 0 0 1 1 0 0 1 bg-sc 0 0 1 0 1 0 1 cg-sc 0 0 1 0 0 1 1 abg-sc 0 0 1 1 1 0 1 acg-sc 0 0 1 1 0 1 1 bcg-sc 0 0 1 0 1 1 1 abcg-sc 0 0 1 1 1 1 1 5. Analysis of test results In order to analyze the accuracy of the proposed method, two groups of test data (group-1 for ground faults and group-2 for un- grounded faults) which are different from that was used for train- ing of ANN and input of FLS have been selected. Then, the sensitivity of the method to the fault and system parameters such as: (1) Fault location (100 m, 1600 m, 1925 m, 4475 m, 5650 m, 6500 m, 8200 m, 8350 m, 9150 m, 9550 m, 9950 m, 12,050 m, 13,750 m, 14,300 m and 16,000 m), (2) Branch (1, 2, 3, 4 and 5), 3. Fault angle (62 deg, 117 deg and 158 deg) and 4. Fault resistance (only for ground short-circuit fault including 0.5X, 5X and 15X) with adding of distributed generation (DG), Load variation includ- ing load shedding, load increasing and load unbalancing was eval- uated. Table 6 shows the various fault conditions employed during the test procedure. In order to research and investigate any fault location/protec- tion technique, it is vitally important to employ data simulation under a whole variety of different systems and fault conditions, for the development of the new technique. This paper gives a sum- mary of all the different systems and fault conditions studies for generating the requisite data and the proposed method that is able to detect any type of fault. The idea is to include all fault data and problems of the distribu- tion system in fuzzy logic and neural networks as inputs. There- fore, whenever a fault occurred in the new distance, the location of the fault could be determined. Each of these features was gener- ated from the combination of several statistical concepts that would change with increasing distance from the source of the fault and also the changing of the fault type. Fig. 18, shows basic configuration of the ANN and FL based fault location technique for analysis of test results. It should be noted that neural networks and fuzzy logic are not in interaction and each of them can solve the fault location issues separately. There- fore, the neural network obtained better performance in finding the fault distance from the source as compared to fuzzy logic. The fuzzy logic also shows its advantage in classification and loca- tion of the branch fault. But in general, the neural networks seems to be more efficient than fuzzy logic due to their accurate determi- nation of the fault location and also the designation of neural net- work is more easier than fuzzy inference systems (with increasing membership functions, the designation of fuzzy inference systems are too complex and time consuming). 5.1. Performance of fault classification and fault branch finding In order to quantitatively evaluate the performance of the fault classification technique, the NNtf was test by four aforementioned test data including over 46 system and fault conditions. It is evi- dent from the results that the number of error decision was zero Table 5 Desired ANN and FLS outputs for faulty branch. Branch number 1 2 3 4 5 Branch1 1 0 0 0 0 Branch2 0 1 0 0 0 Branch3 0 0 1 0 0 Branch4 0 0 0 1 0 Branch5 0 0 0 0 1 characteristics and the types of faults that occur in power system is presented. The combination of signal processing and intelligent methods has been used for fault classification and location. The novelty of this research was using new statistical features and fault detection method compared to previous works. All possible simulations also have been performed. Initially, the radial network with multiple branches is considered and the type of fault, the faulty branch and the fault distance from the source have been calculated. Then, the radial network has been converted to a ring network with the addition of distributed generation (DG). In these conditions both of the proposed methods can accurately determine the type of fault, the faulty branch and the fault distance from the source and the proposed network is also considered as a real network. 5.2.1. Effect of load shedding with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs It is well known that load increasing can adversely affect the A. Rafinia, J. Moshtagh / Electrical Power and Energy Systems 55 (2014) 261–274 271 in relation to the NNtf; therefore it can be concluded that this method is indeed to classify the type of fault and branch of fault and recognizes the faulted phases and have operated with accuracy more than 99%. 5.2. Performance of fault location In order to analyze the accuracy of the proposed method, the data used for testing is different and unseen from that used for training. The trained ANNs and the designed FLSs involved in the third stage of the fault location and classification technique were tested by five aforementioned groups of test data. The error for fault location is expressed as a percentage of the length of the cable, and is given as the following equation: error% ¼ ðactual locationÞ � ðdesired locationÞðcable lengthÞ � 100 ð5Þ Tables 7–10 show the ANN and the FLS performance for the tests carried out on a system in case of high impedance faults, short-circuits faults and open-circuit faults, respectively. Of course, Fig. 18. Basic configuration of the ANN and FL based fault location technique for analysis of test results. the results related to some faults are shown and other results are excluded for brief. In this paper, a comprehensive solution for fault location in distribution networks with respect to various parameters, Table 6 Testing data for fault location based on ANN and FLS. No. TDa Group-1 (g-sc fault) Location (m) H (deg) Rf (X) TD-1 750 62 0.5 TD-2 1600 117 5 TD-3 1925 158 15 TD-4 4475 62 0.5 TD-5 5650 117 5 TD-6 6500 158 15 TD-7 8200 62 0.5 TD-8 8350 117 5 TD-9 9150 158 15 TD-10 9550 62 0.5 TD-11 9950 117 5 TD-12 12,050 158 15 TD-13 13,750 62 0.5 TD-14 14,300 117 5 TD-15 16,000 158 15 a Number of Test Data (TD). Group-2 (hi, scoc) fault) Branch Location (m) H (deg) Branch 1 750 62 1 1 1600 117 1 1 1925 158 1 1 4475 62 1 12 5650 117 12 12 6500 158 12 1,2,3,4 8200 62 1,2,3,4 1,2,3,4 8350 117 1,2,3,4 1,2,3,4 9150 158 1,2,3,4 1,2,4 9550 62 1,2,4 1,2,4 9950 117 1,2,4 1,4,5 12,050 158 1,4,5 4,5 13,750 62 4,5 4,5 14,300 117 4,5 accuracy of conventional fault locators. In this respect, the algo- rithm was tested based on group-1&2 of test data. Table 8 depicts the results. In comparison to the previous case associated, it is evi- dent that the presence of load increasing has only a slight effect on accuracy particularly in the case of one phase, phase to phase fault and three phase, as the maximum and the mean of error associated with four faults increased very slightly to (0.604 & 0.211), (0.697 & 0.287), (1.1 & 0.841) and (0.73 & 0.259) percent for ANN and (0.64 & 0.246), (0.69 & 0.293), (1.2 & 0.851) and (0.7 & 0.294) percent for FLS respectively. These small changes can be directly attributed to the fact that with load increasing, the current changes in the healthy phase in terms of magnitude and distortion. The load shedding significantly affects the fault transient volt- age and current signals and it is vitally important to verify the ef- fect of this parameter on the performance of the proposed technique. Table 7 show the accuracies attained for groups-1&2 of test results. It is clearly evident from the results that the ANNs and FLSs give very accurate evaluation of fault position and the maximum and mean of error correspond to the g-sc, sc, oc and hi faults are for ANN: (0.8 & 0.346), (0.78 & 0.283), (1.01 & 0.812) and (0.512 & 0.225) percent and for FLS: (0.7 & 0.369), (0.8 & 0.292), (1.21 & 0.819) and (0.61 & 0.264) percent, respectively. This study clearly demonstrates that the algorithm is virtually immune to any errors caused by either the load shedding. 5.2.2. Effect of load increasing with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs 5 16,000 158 5 Table 7 Effect of load shedding with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs. No. TD Group-1, G-SC Group-2, SC Group-2, OC Group-2, HI Ma. error% M. error% M. error% M. error% ANN FLS ANN FLS ANN FLS ANN FLS TD-1 0.55 0.46 0.002 0.01 0.8 0.88 0.02 0.05 TD-2 0.8 0.63 0.7 0.67 0.81 0.91 0.025 0.05 TD-3 0.25 0.35 0.42 0.32 0.55 0.5 0.23 0.3 TD-4 0.25 0.45 0.072 0.07 0.905 0.95 0.44 0.4 TD-5 0.65 0.7 0.387 0.37 0.546 0.46 0.11 0.1 TD-6 0.05 0.3 0.337 0.37 0.69 0.9 0.13 0.3 TD-7 0.525 0.08 0.475 0.75 0.89 0.8 0.05 0.09 TD-8 0.04 0.42 0.025 0.05 0.77 0.7 0.45 0.4 TD-9 0.02 0.04 0.22 0.2 0.78 0.8 0.237 0.27 TD-10 0.115 0.149 0.1 0.11 0.63 0.3 0.266 0.26 TD-11 0.492 0.42 0.0 0.01 1.01 1.21 0.361 0.61 TD-12 0.222 0.42 0.35 0.3 0.97 1.07 0.09 0.01 TD-13 0.625 0.62 0.78 0.8 1.0 1.0 0.512 0.52 TD-14 0.325 0.25 0.23 0.2 0.955 0.95 0.221 0.21 TD-15 0.275 0.25 0.145 0.15 0.876 0.86 0.24 0.4 TMb.error% 0.346 0.369 0.283 0.292 0.812 0.819 0.225 0.264 a Mean of errors. b Total mean of errors. Table 8 Effect of load increasing with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs. No. TD Group-1, G-SC Group-2, SC Group-2, OC Group-2, HI M. error% M. error% M. error% M. error% ANN FLS ANN FLS ANN FLS ANN FLS TD-1 0.1 0.11 0.06 0.2 0.75 0.5 0.07 0.17 TD-2 0.015 0.05 0.2 0.1 0.65 0.5 0.12 0.2 TD-3 0.147 0.17 0.01 0.03 1.1 1.0 0.31 0.3 TD-4 0.286 0.26 0.243 0.23 0.87 0.8 0.05 0.08 TD-5 0.132 0.32 0.295 0.25 0.555 0.94 0.02 0.2 TD-6 0.08 0.2 0.028 0.07 0.45 0.98 0.11 0.01 TD-7 0.512 0.52 0.567 0.56 0.9 0.8 0.34 0.4 TD-8 0.124 0.14 0.555 0.5 0.7 0.8 0.45 0.5 TD-9 0.1 0.11 0.27 0.41 0.945 0.95 0.73 0.7 TD-10 0.2 0.22 0.06 0.16 0.98 0.9 0.25 0.6 TD-11 0.604 0.64 0.423 0.43 1.04 1.2 0.24 0.2 TD-12 0.07 0.03 0.697 0.69 0.96 0.8 0.01 0.01 TD-13 0.154 0.15 0.137 0.17 0.89 0.9 0.524 0.54 TD-14 0.235 0.25 0.387 0.37 0.95 0.9 0.261 0.2 TD-15 0.412 0.52 0.243 0.23 0.88 0.8 0.401 0.31 TM. error% 0.211 0.246 0.278 0.293 0.841 0.851 0.259 0.294 Table 9 Effect of load unbalancing with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs. No. TD Group-1, G-SC Group-2, SC Group-2, OC Group-2, HI M. error% M. error% M. error% M. error% ANN FLS ANN FLS ANN FLS ANN FLS TD-1 0.12 0.2 0.04 0.4 1.0 0.91 0.1 0.3 TD-2 0.41 0.4 0.02 0.3 0.88 0.8 0.2 0.1 TD-3 0.02 0.3 0.101 0.01 0.84 0.8 0.1 0.2 TD-4 0.55 0.5 0.4 0.1 0.98 0.9 0.33 0.55 TD-5 0.244 0.44 0.3 0.6 0.7 0.7 0.291 0.21 TD-6 0.323 0.323 0.01 0.1 0.97 0.7 0.268 0.28 TD-7 0.13 0.21 0.33 0.3 1.2 0.92 0.01 0.51 TD-8 0.105 0.45 0.41 0.4 0.86 1.31 0.57 0.57 TD-9 0.542 0.32 0.28 0.4 0.81 1.02 0.286 0.23 TD-10 0.47 0.27 0.66 0.56 0.95 0.9 0.365 0.65 TD-11 0.29 0.2 0.31 0.3 0.69 0.75 0.563 0.53 TD-12 0.34 0.4 0.44 0.4 0.86 0.8 0.52 0.5 TD-13 0.623 0.61 0.651 0.51 0.59 0.9 0.104 0.04 TD-14 0.423 0.23 0.39 0.3 0.94 0.9 0.202 0.02 TD-15 0.22 0.3 0.34 0.5 0.9 0.93 0.11 0.1 TM. error% 0.32 0.343 0.312 0.345 0.878 0.882 0.268 0.319 272 A. Rafinia, J. Moshtagh / Electrical Power and Energy Systems 55 (2014) 261–274 nd F FL TD-9 0.66 0.61 1.25 TD-10 0.85 1.9 0.6 08 r an 5.2.4. Effect of DG with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs Integration of a DG into an existing distribution system has many impacts on the system, with the power system protection being one of the major issues. Short circuit power of a distribution system changes when its state changes [23]. Thus, adding DG can adversely affect the accuracy of conventional fault locators. There- fore, with adding of DG-4MW at the end of branch-5, the ANN and FLS algorithm were tested based on group-1&2 of test data and 5.2.3. Effect of load unbalancing with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs It is apparent that load unbalancing significantly affect the fault transient waveforms. Therefore, it is vitally important to verify the effect of the load unbalancing on the performance of the proposed technique. In these respect groups 1&2 are considered and Table 9 depicts the results. It is clearly evident from the results that the accuracy achieved in fault location is very high; being less than 1% error in all the test cases. TD-11 0.69 0.64 0.43 TD-12 0.91 0.39 0.678 TD-13 0.54 0.59 0.73 TD-14 0.75 1.1 1.86 TD-15 0.49 0.9 0.94 TM. error% 0.798 0.856 0.86 Table 10 Effect of DG with the data of Groups 1&2 on accuracy fault location based on ANNs a No. TD Group-1, G-SC Group-2, SC M. error% M. error% ANN FLS ANN TD-1 0.56 0.91 0.68 TD-2 0.63 1.05 1.2 TD-3 1.41 0.67 0.91 TD-4 0.95 1.21 0.93 TD-5 1.2 0.92 0.95 TD-6 1.1 0.72 0.68 TD-7 0.65 0.83 0.57 TD-8 0.58 0.4 0.5 A. Rafinia, J. Moshtagh / Electrical Powe Table 10 show the results. The results clearly demonstrate that the accuracy achieved in fault location is very high; being less than 1.5% error in all the test cases and shows that the ANNs give accurate evaluation of fault position that is largely independent on the DG adding. Thus it can be concluded that the ANNs and FLSs give accurate assessment significantly independent to the load variation, DG, inception angle and distance. 6. Conclusion In this paper at first, a new method is proposed to analyze power distribution system transient signals based-EMTP by using WT technique. This method offers important advantages over other methods such as FFT and STFT due to good time and frequency localization characteristics. Analysis presented results clearly show that particular wavelet components can be used as the features to locate the fault in UDS. Then an accurate fault location technique based on ANN and FLS is developed, as ANNs and FLSs are trained and designed to classify the fault type and separate ANNs and FLSs are designed to accurately locate the actual ungrounded fault posi- tion on a practical UDS. In this respect, three-layer feed-forward ANNs and the LM algorithm is used to adopt the weights and biases to achieve the desired non-linear mapping from inputs to outputs. Through a series of test and modifications, it is evident that the ANNs and FLSs can classify the type of fault very accurately under different system and fault conditions. In order to illustrate the effectiveness of fault location based on ANNs and FLSs technique, each ANN and FLS is tested with a separate set of unseen data and their performance on the accuracy of the results are presented. The presented results herein, clearly show that the proposed meth- od gives a high accuracy in fault location under a whole variety of different system and fault condition. Finally, it must be concluded that both methods have advanta- ges and disadvantages that should be considered. Neural networks are fast and easy to run which allows the designer a more rapid designation to attain the target. The main disadvantage of neural networks is finding the appropriate features which can guide the neural networks to understanding the relationship between the in- put and the desired output seems to be very difficult in some con- ditions. On the other hand, in fuzzy logic not only the inputs can be designed by user in the best possible way but also the membership functions can be adjusted with more careful planning and manip- LSs. Group-2, OC Group-2, HI M. error% M. error% S ANN FLS ANN FLS 0.72 1.75 1.9 0.87 1.1 0.93 1.65 0.95 0.82 0.8 0.79 1.4 1.33 0.91 1.3 0.66 1.7 0.9 0.95 0.68 0.44 1.5 1.4 0.62 0.72 0.94 1.4 1.8 0.71 0.97 1.6 1.9 2.05 0.74 0.8 0.5 0.9 1.8 0.95 0.7 0.81 2.1 0.9 0.79 0.8 0.86 1.98 1.8 0.85 0.94 0.83 1.04 1.7 0.84 0.98 0.9 0.86 1.3 1.1 1.01 0.7 0.9 1.2 1.5 0.89 0.67 1.5 1.6 1.2 0.81 0.83 1.3 0.8 0.9 0.71 12 1.46 1.43 0.916 0.88 d Energy Systems 55 (2014) 261–274 273 ulation to find the gaps of the individual membership function in appropriateness. One of the disadvantages of fuzzy logic method compared to neural network is in designation technique, the fuzzy systems; generally have lower degrees of desirability due to their large number of inputs and membership functions, as compared to the neural networks. This method is responsible for various topological configura- tions and operating conditions. It is enough to coordinate the neu- ral networks and fuzzy logic according to the distributed network topology. The results clearly show that the neural network is more efficient than fuzzy logic. In the results of section, two combined methods including wavelet-ANN and wavelet-FLS have been compared under various fault and system conditions and their performance has been eval- uated to determine the type and location of the faults. The pro- posed methods recommend the use of an ANN-FLS for the determination of fault types and locations. Therefore, the best method is to use a combination of wavelets, neural networks and fuzzy logic methods. In cases of more complexity and data abun- dance, the designer can use neural networks instead of fuzzy logic and whenever the artificial neural network fails to identify the relations between inputs and outputs, the fuzzy logic method can be used instead to solve the problem. Thus it can be concluded that the proposed approach based on combined WT, ANN and FLS is robust to different case studies; this is a signification advantage and can be directly attributed to the fact that WT technique effectively extracts the very crucial time– frequency features from UDS transient signals and ANN and FLS approach are able to give a very high accuracy in the fault classifi- cation and fault location. In order to avoid further economic and social costs because of load interruptions, the fault diagnosis has to be concluded as soon as possible. Intelligent systems have been successful in dealing with fault diagnosis problems [24]. References [1] Peretto L, Tinarelli R, Bauer A, Pugliese S. Fault location in underground power networks: a case study. IEEE/PES Innovative Smart Grid Technologies (ISGT) 2011;1–6. [2] Kawady TA, Taalab AI, El-Sad M. An accurate fault locator for underground distribution networks using modified apparent-impedance calculation. In: 10th IET international conference on developments in power system protection (DPSP); 2010. p. 1–5. [3] Mokhlis H, Li HY, Bakar AHA, Mohamad H. Locating fault using voltage sags profile for underground distribution system. In: IEEE international conference 274 A. Rafinia, J. Moshtagh / Electrical Power and Energy Systems 55 (2014) 261–274 Acknowledgements The authors would like to appreciate Reza Bashiri Khuzestani for editing the article. Appendix A. Various Elements The specifications of the various elements in Fig. 1 are as follows: Source: VL = 20 kV, f = 50Hz, Xs/Rs = 30, Xs = 6X, Rs = 0.2X Cable: XLPE, Three-phase pipe type cable (core + grounded sheath), cable length = 16,500 m Transformers: Load 2: S = 1 MVA Winding1: VL = 20 kV, Rp = 8.1X, Lp = 112 mH Winding2: VL = 0.4 kV, Rp = 0.0032X, Lp = 0.045 mH Loads 1, 3, 4, 5, 6, 7, 8, 12 and 15: S = 0.8 MVA Winding1: VL = 20 kV, Rp = 10.3125X, Lp = 139.43 mH Winding2: VL = 0.4 kV, Rp = 0.00413X, Lp = 0.0557 mH Loads 9, 10, 11, 13, 14 and 16: S = 0.5 MVA Winding1: VL = 20 kV, Rp = 18.72X, Lp = 221.31 mH Winding2: VL = 0.4 kV, Rp = 0.0075X, Lp = 0.0885 mH Load1: VL = 0.38 kV, f = 50 Hz, PL = 440KW, QL = 324 KVAR Load2: The combination of three-phase static and dynamic loads Static Load: VL = 0.38 kV, f = 50Hz, PL = 490 kW, QL = 360 KVAR Dynamic Load: VL = 0.38 kV, f = 50 Hz, P = 200 HP Load3: VL = 0.38 kV, f = 50Hz, PL = 398KW, QL = 298 KVAR Load4: VL = 0.38 kV, f = 50 Hz, PL = 390KW, QL = 288 KVAR Load5: VL = 0.38 kV, f = 50 Hz, PL = 388KW, QL = 286 KVAR Load6: VL = 0.38 kV, f = 50 Hz, PL = 380KW, QL = 280 KVAR Load7: VL = 0.38 kV, f = 50 Hz, PL = 242KW, QL = 178 KVAR Load8: VL = 0.38 kV, f = 50 Hz, PL = 220KW, QL = 164 KVAR Load9: VL = 0.38 kV, f = 50 Hz, PL = 280KW, QL = 210 KVAR Load10: VL = 0.38 kV, f = 50 Hz, PL = 274KW, QL = 204 KVAR Load11: VL = 0.38 kV, f = 50 Hz, PL = 262KW, QL = 200 KVAR Load12: VL = 0.38 kV, f = 50 Hz, PL = 260KW, QL = 190 KVAR Load13: VL = 0.38 kV, f = 50 Hz, PL = 240KW, QL = 180 KVAR Load14: VL = 0.38 kV, f = 50 Hz, PL = 200KW, QL = 150 KVAR Load15: VL = 0.38 kV, f = 50 Hz, PL = 175KW, QL = 130 KVAR Load16: VL = 0.38 kV, f = 50 Hz, PL = 142KW, QL = 68 KVAR on, power system technology; 2010. p. 1–5. [4] Navanithan S, Soraghan J, Siew WH, Mcpherson F, Gale PF. Automatic fault location for undergrounded low voltage distribution networks. IEEE Trans Power Delivery 2001;16(2):346–51. [5] Aggarwal RK, Aslan Y, Johns AT. New concept in fault location for overhead distribution systems using superimposed components. IEE J 1999;146:209–16. [6] Mokhlis H, Li H. Non-linear representation of voltage sag profiles for fault location in distribution networks. Int J Electr Power Energy Syst 2011;33(1):124–30. [7] Bernadic´ A, Leonowicz Z. Fault location in power networks with mixed feeders using the complex space-phasor and Hilbert–Huang transform. Int J Electr Power Energy Syst 2012;42(1):208–19. [8] Sadeh J, Bakhshizadeh E, Kazemzadeh R. A new fault location algorithm for radial distribution systems using modal analysis. Int J Electr Power Energy Syst 2013;45(1):271–8. [9] Zhang J, He ZY, Lin S, Zhang YB, Qian QQ. An ANFIS-based fault classification approach in power distribution system. Int J Electr Power Energy Syst 2013;49:243–52. [11] Magnago H, Abur A. Fault location using wavelet. IEEE Trans Power Delivery 1998;13(4):1475–80. [12] Moshtagh J. Accurate fault location in low voltage underground cables using advantage signal processing and artificial intelligence, PhD thesis, University of bath UK; July 2006. [13] Kartalopoulos Stamatios V. Understanding neural network and fuzzy logic: basic concept and applications. New York: The Institute of Electrical and Electronics Engineering Inc.; 1996. [14] Berenson ML, Levine DM, Krehbiel TC. Basic Business Statistics concepts and applications. twelfth ed. Pearson; 2012 [p. 912]. [15] Moshtagh J, Jalili P, Souraki, HP. A new approach to arcing fault detection in underground distribution system using combined neural network wavelet analysis. In: 5th International conference on technical and physical problems of power engineering, Bilbao, Spain; September 2009. [16] Moshtagh J, Jalili P, Souraki, HP. A new approach to series fault location in underground distribution system using combined neural network wavelet analysis. In: IEEE international conference on electrical power and energy conservation systems. American University of Sharjah, UAE; November 2009. [17] Moshtagh J, Aggarwal RK. A new approach to ungrounded fault location in a three-phase undergrounded distribution system using combined neural networks wavelet analysis. In: IEEE Canadian conference on electrical and computer engineering (CCECE). Ottawa, Canada; May 2006. [18] Moshtagh J, Aggarwal RK. A new approach to modelling high voltage three- phase cable using conformal mapping. In: IEEE bulgarian conference, ELMA2005. Sofia, Bulgaria; September 2005. [19] Aggarwal R, Song Y. Artificial neural networks in power systems. I. General introduction to neural computing. Power Eng J 1997;11(3):129–34. [20] Aggarwal R, Song Y. Artificial neural networks in power systems. II. Types of artificial neural networks. Power Eng J 1998;12(1):41–7. [21] Aggarwal R, Song Y. Artificial neural networks in power systems. III. Examples of applications in power systems. Power Eng J 1998;12(6):279–87. [22] Moshtagh J, Aggarwal RK. A new approach to fault location in a single core underground cable system using combined fuzzy logic wavelet analysis, developments in power system protection. In: Eighth IEE international conference, vol. 1, 2004. p. 5–8. [23] Sarabia AF. Impact of distributed generation on distribution system. M.Sc. thesis. University of Aalborg Denmark; June 2011. [24] Alves da Silva AP, Lima ACS, Souza SM. Fault location on transmission lines using complex-domain neural networks. Int J Electr Power Energy Syst 2012;43(1):720–7. A new approach to fault location in three-phase underground distribution system using combination of wavelet analysis with ANN and FLS 1 Introduction 2 Data simulation 3 Feature extraction using wavelet 3.1 Wavelet Transform 3.2 Choice of mother wavelet 3.3 Feature extraction using statistical relations 3.3.1 Feature extraction in fault classification 3.3.2 Feature extraction in finding the faulty branch 3.3.3 Feature extraction to find the fault distance from the source 3.3.3.1 Ground high impedance fault (hi) 3.3.3.2 Short-circuit fault (sc) 3.3.3.3 Ground short-circuit fault (g-sc) 3.3.3.4 Open-circuit fault (oc) 4 Artificial intelligence techniques 4.1 Artificial neural network 4.2 Fuzzy logic 4.3 Fault classification based on ANNs and FLSs 4.4 Finding the fault branch based on ANNs and FLSs 4.5 Fault distance from the source based on ANNs and FLSs 5 Analysis of test results 5.1 Performance of fault classification and fault branch finding 5.2 Performance of fault location 5.2.1 Effect of load shedding with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs 5.2.2 Effect of load increasing with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs 5.2.3 Effect of load unbalancing with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs 5.2.4 Effect of DG with the data of groups 1&2 on accuracy fault location based on ANNs and FLSs 6 Conclusion Acknowledgements Appendix A Various Elements References