FAULT CLASSIFICATION AND LOCATION IN SERIES COMPENSATED TRANSMISSION LINES

Abstract

Series compensation is increasingly identified as an important part of power system, which helps in full utilization of existing transmission infrastructure by increasing the power transfer capability of existing lines. Moreover, variable series compensation has additional advantages such as improvement of stability of system by damping the power swings. However, advantages of series compensation pose extra challenges for the protection of compensated line. Fixed compensation by capacitor has conventional problems such as nonlinear operation of protective equipment of capacitor, current reversal, and voltage reversal. These problems are further aggravated if compensation is provided by thyristor controlled series capacitor (TCSC), because of different operating modes and variable compensation of TCSC. Fault classification, section identification, and location of fault are few of the protection tasks that become difficult for series compensated line due to above discussed problems. To perform these protection tasks efficiently, this thesis presents new methods using four popular single-hidden layer feedforward networks (SLFNs), namely Support Vector Machine (SVM), Relevance Vector Machine (RVM), Extreme Learning Machine (ELM), and Kernel Extreme Learning Machine (KELM). These SLFNs are chosen for the task as they are getting popularity due to their ability to map high degree of nonlinearity, which is well suited for highly nonlinear nature of the problems addressed. Performances of the four SLFNs are demonstrated on a digitally simulated two-area system and a 12-Bus system. Although, models of the four SLFNs are same in architecture, their performance based on accuracy, training time, testing time, and number of nodes in hidden layer are different. Therefore, depending on these performance criteria different SLFNs are suitable for different applications e.g. real time or non-real time. While task of fault location is non-real-time, fault classification and section identification can be real-time or non-real-time depending on application. This thesis suggests the suitability of the SLFNs for different tasks. Similar to any learning machine, for training, cross-validation, and testing of an SLFN, large amount of data is required. Therefore, repeated simulation of the two-area system and 12-Bus system were simulated by considering wide variation in system and fault condition. To make this procedure of data generation fast and convenient, thesis proposes two methods of enhancing batch mode capabilities of an existing electromagnetic transient simulation program. Effectiveness of these two methods is illustrated in two cases, in one case nearly two person-weeks and in another case more than one person-moth were saved. SVM and RVM have obtained very high accuracy (more than 99.50%) for fault classification and sparse nature of RVM deemed it suitable for real-time application of fault classification. SVM, RVM, and KELM have achieved very high accuracy (more than 99%) for ii fault classification and faulty section identification, their appropriateness for applications is discussed in detail. Furthermore, another new approach of modifying arrangement of input samples to KELM is proposed for improving accuracy of fault location. Although, KELM‟s testing time is comparatively large, fault location accuracy of KELM is the highest and the time required for training and tuning-parameters is small. Therefore, KELM is found to be most suitable for offline application of fault location. Effect of variation of different parameters on KELM‟s fault location accuracy is also analyzed.