2310 IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 22, NO. 12, DECEMBER 2011 Improved GART Neural Network Model for Pattern Classification and Rule Extraction with Application to Power Systems Keem Siah Yap, Member, IEEE, Chee Peng Lim, and Mau Teng Au, Member, IEEE Abstract— Generalized adaptive resonance theory (GART) is a neural network model that is capable of online learning and is effective in tackling pattern classification tasks. In this paper, we propose an improved GART model (IGART), and demonstrate its applicability to power systems. IGART enhances the dynamics of GART in several aspects, which include the use of the Laplacian likelihood function, a new vigilance function, a new match- tracking mechanism, an ordering algorithm for determining the sequence of training data, and a rule extraction capability to elicit if-then rules from the network. To assess the effectiveness of IGART and to compare its performances with those from other methods, three datasets that are related to power systems are employed. The experimental results demonstrate the usefulness of IGART with the rule extraction capability in undertaking classification problems in power systems engineering. Index Terms— Fuzzy inference systems, generalized adaptive resonance theory, pattern classification, rule extraction. I. I NTRODUCTION O VER the last two decades, artificial neural networks (or simply neural networks) have been developed for solving pattern classification problems in many different domains [1]–[4]. In power systems engineering, a number of successful applications using neural networks have been reported. These include the use of neural networks for short-term power load or price forecasting [5], [6], and fault diagnosis for transformers. The fuzzy ARTMAP (FAM) neural network was used for fault diagnosis of an operational transformer [7]. Other neural network-related fault detection models for transformers include an abductive network model (a hybrid fuzzy logic and neural network model), a hybrid genetic algorithm and wavelet neural network, and a multilayer feedforward neural network [8]–[10]. In [11], a neural network was used to provide an early warning of combustion-related faults in a diesel engine. In [12] and [13], the fuzzy min-max and the FAM networks were used for condition monitoring of the circulating water system in a power generation plant. A fault detection method in circuit breakers using wavelet Manuscript received December 4, 2010; revised August 11, 2011; accepted October 16, 2011. Date of publication November 4, 2011; date of current version December 13, 2011. K. S. Yap is with the College of Graduate Studies, Universiti Tenaga Nasional, Kajang 43009, Malaysia (e-mail: yapkeem@uniten.edu.my). C. P. Lim is with the School of Computer Sciences, University of Science Malaysia, Penang 11800, Malaysia (e-mail: cplim@cs.usm.my). M. T. Au is with the Power Engineering Centre, Universiti Tenaga Nasional, Kajang 43009, Malaysia (e-mail: mtau@uniten.edu.my). Digital Object Identifier 10.1109/TNN.2011.2173502 packet and the multilayer feed-forward neural network with the backpropagation learning algorithm was described in [14]. Other applications of neural networks in power systems engineering include assessment of power quality [15], load shedding [16], power system stabilizers [17], and active power filters [18]. While there are many different types of neural networks, not many of them have the ability of knowledge extraction. Most neural networks suffer from the limitation whereby their reasoning process is considered as a “black box” [19]–[21], and there is no logical explanatory facility to justify their predictions. To overcome this limitation, researchers have pro- posed neural network models with rule extraction capabilities [12], [13], [21], [22]. According to [12], a neural network with a rule extraction capability has two major advantages: 1) explanation of results—the extracted rules can be used as a self-explanatory module to its predictions, and 2) data exploration and feature revelation—the ability of extracting rules into the if-then form [21] that may help users to identify if an input attribute has been always (or seldom) assigned to the same linguistic value (e.g., big, small, etc.). The objectives of this paper are twofold. First, a new improved generalized adaptive resonance theory (GART) neural network with a rule extraction capability is proposed. In our previous work, the theoretical development of GART neural network has been reported [23]. It is a hybrid neural network model combining the adaptive resonance theory (ART) [24], [25] network and the generalized regression neural network (GRNN) [26] that is capable of incremental learning. In this paper, the capability of GART is further improved (hereafter denoted as IGART) with several features that are essential for practical applications. These include improved dynamics of GART with a Laplacian likelihood function [27], a new vigilance function, and a match-tracking mechanism. An ordering algorithm for determining the presentation sequence of the training samples, and a series of post processing procedures, i.e., network pruning and rule extraction, are also incorporated into IGART. The second objective is to demonstrate the effectiveness of IGART for solving problems related to power systems engineering. Three case studies in three main areas of the power engineering industry, i.e., generation, transmission, and distribution, are examined. It is envisaged that researchers and practitioners from different areas in the power engineering industry can obtain ideas and techniques from this paper, 1045–9227/$26.00 © 2011 IEEE