Power Grid Fault Diagnosis Based On Radial Basis Function Neural Network

With the development of modern power systems，power grids are becoming moreexpanding and interconnected. As a consequence, when a fault does occur, its influenceextended and makes power systems more vulnerable. When complicated fault take places,the fault situation will be uncertainty and incompleteness if protective devices (protectiverelays and/or circuit breakers) associated fail to work properly or alarm information distortsduring the signal transmission. As a result, vast quantities of uncertain and incompletealarm information could be displayed on the console in the control center, and give rise tothe difficulty for dispatchers to understand the essence of the raw information quickly, andidentify the main outage cause correctly. Thus, organizing a plan with the power grid faultpresents more requests for higher-level management. Timely and accurate method forpower grid fault diagnosis is of great value in restoring the stable supply.It is difficult that an evolution process of power grid fault is described by using thetraditional mathematic method. Artificial neural network with its own good at dealing withcomplex nonlinear problem and simulating real-world interaction make it has robustness,fault tolerance and generalization ability. Among different neural networks, radial basisfunction (RBF) neural network has the best approximation and the global optimalperformance. Consequently, problems that come from using RBF neural network for powergrid fault diagnosis have launched thorough research in this article.1) Discuss the background and purpose of the studies. Review the major recentdevelopments of power grid fault diagnosis and describe the advantages anddisadvantages of different fault diagnosis methods. The basic features and functions ofartificial neural network are outlined. And on that basis, the essential principles andtraining algorithms of two neural networks, i.e., back-propagation (BP) neural network and RBF neural network are emphasized.2) In view of the disadvantages of traditional training algorithms of RBF neural network,this thesis presents a novel two-layer learning method for RBF neural network using animproved objective function. At the lower layer, the optimal feature number (thenumber of hidden-layer node) is decided automatically by the introduced ten clustervalidity functions of fuzzy c-means clustering algorithm. After determining the numberof hidden-layer node, two key learning parameters, the central vector and hiddennode’s width, are optimized at the higher layer using the self-adaptive weightingcooperative coevolution which based on self-adaptive differential evolution withneighborhood search and self-adaptive differential evolution without archive, and theweights of connections between hidden layer and output layer can be calculateddirectly by least square. The method can not only determine the number ofhidden-layer node efficiently, but also can improve the generalization ability of RBFneural network using cooperative coevolution for optimizing an improved objectivefunction. The results show clearly that the algorithm can improve the performance ofpower grid fault diagnosis.3) Concerning the complexity of training process of traditional RBF neural network, adecay RBF neural network that can uniformly approximate any continuous multivariatefunctions with arbitrary precise without training is constructed. To meet the needs ofpower grid fault diagnosis, single-output decay RBF neural network has been extendedto multi-output one according to the fault characteristic of power grid components. Themodel can speed up the training process and increase the utility of constructing, as wellas improve accuracy and precision of identifying faulted section.