2360 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 27, NO. 4, OCTOBER 2012 Accurate Single-Phase Fault-Location Method for Transmission Lines Based on K-Nearest Neighbor Algorithm Using One-End Voltage Mohammad Farshad, Student Member, IEEE, and Javad Sadeh, Member, IEEE Abstract—In this paper, some useful features are extracted from voltage signals measured at one terminal of the transmission line, which are highly efficient for accurate fault locating. These features are the amplitude of harmonic components, which are extracted after fault inception through applying discrete Fourier transform on one cycle of three-phase voltage signals and then are normalized by a transformation. In this paper, the location of single-line-to- ground faults as the most probable type of fault in the transmission networks is considered. The SLG fault locator, which is designed based on the simple algorithm of k-nearest neighbor ( -NN) in re- gression mode, estimates the location of fault related to the new input pattern based on existing available patterns. The proposed approach only needs the measured data from one terminal; hence, data communication between both ends of the line and synchro- nization are not required. In addition, current signals are not used; therefore, the proposed approach is immune against current-trans- former saturation and its related errors. Tests conducted on an un- transposed transmission line indicate that the proposed fault lo- cator has accurate performance despite simultaneous changes in fault location, fault inception angle, fault resistance, and magni- tude and direction of load current. Index Terms—Fault location, Fourier transform, k-nearest neighbor, single-line-to-ground fault, transmission line. I. INTRODUCTION A CCURATE fault locating of permanent and temporary faults is of high importance from the aspects of quick re- pairs and troubleshooting, identification of weak points of the transmission line, and adoption of required measures for de- creasing fault occurrence probability at those locations [1]. Ex- isting approaches for fault locating in transmission lines can be categorized into two main groups of approaches based on hard computing and soft computing. In recent years, there have been considerable efforts to improve and develop the fault-locating approaches based on hard computing using analytical models or traveling-wave theory. Parallel to these efforts, there has been a substantial amount of research on implementing soft computing techniques to solve the fault-location problem, due to the flexi- bility and capability of these techniques in facing the complexity of the problem. Learning machines are considered as the tools Manuscript received March 14, 2012; revised June 11, 2012; accepted July 27, 2012. Date of publication September 11, 2012; date of current version September 19, 2012. Paper no. TPWRD-00267-2012. The authors are with the Electrical Engineering Department, Faculty of Engi- neering, Ferdowsi University of Mashhad, Mashhad 91779-48944, Iran (e-mail: m.farshad@ieee.org; sadeh@um.ac.ir). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TPWRD.2012.2211898 of performing soft computations in the fault-location problem. In the learning-based approaches, it is possible to train machine according to existing real patterns or patterns generated using reality-based simulation techniques. In this mode, the learning algorithm is responsible for the task of discovering hidden rules and complicated relationships between features of the patterns. Selecting or extracting appropriate features and imple- menting an efficient learning algorithm are two pivotal issues in the fault-locating approaches based on machine learning. Low sensitivity to changes in the effective parameters, such as magnitude and direction of prefault current, fault resistance and fault inception angle, and high correlation with the fault location can be mentioned as some positive characteristics of the extracted features for fault locating. Furthermore, using only measured data from one end of the line can bring about a decrease in expenses resulting primarily from requirements for transmitting and synchronizing measured data of both terminals. In addition, among single-ended measurement data, voltage signals have some advantages over current signals. Using the current signals may be associated with a decrease in fault locating accuracy level caused by the existence of a significant dc component, saturation of current transformer, and high sensitivity of the extracted features to magnitude and direction of prefault current. In [2], an approach was presented for ground fault locating in transmission lines, which was based on high-frequency voltage transients measured at one end of the lines. The presented approach in [2] requires a very high sampling frequency and its fault locating accuracy decreases noticeably with a decrease in the sampling frequency. Several learning tools and methods have been implemented for fault locating, including the multilayer perceptron neural network (MLPNN) [3]–[12]; radial basis function neural net- work (RBFNN) [13]–[15]; support vector machine (SVM) [16], [17]; extreme learning machine (ELM) [17]; Elman recurrent network [18]; fuzzy inference system (FIS) [19]–[22]; fuzzy neural network (FNN) [20]–[25]; and adaptive structural neural network (ASNN) [26]. These tools have a limitation on the dimension of input space, which can be trouble when a large number of potential features like high-frequency components of current or/and voltage signals are used. They are not practically capable of appropriate learning patterns with a high number of features due to enlargement of the structure and an extreme increase of the number of learning parameters [10]. On the other hand, dimension reduction using linear transforms or experimental selection of some features may result in elimi- nation of some useful information. Using energies of limited 0885-8977/$31.00 © 2012 IEEE