An Evolutionary ANN Based on Rough Set and Its Application in Power Grid Fault Diagnosis

Abstract

In order to overcome the inherent flaws of artificial neural networks (ANN), such as long training time, slow convergence and low diagnosis accuracy, a novel evolutionary ANN combining with rough set (RS), named as RSANN, is suggested, and it's proposed to apply in power grid fault diagnosis. The ANN used is a three-layer back-propagation (BP) neural network. RS can reduce the dimensionality of attributes and find out the core attributes through its reduct. The attribute in this research is the information of circuit breakers (CBs) tripping and protection relays action, which is used to diagnose power grid fault. In RSANN, the RS is applied to serve for pretreatment unit which can deal with uncertain or incomplete information, and the core attributes are applied to optimize both topology and connection weights of ANN so as to simplify network structure and improve learning quality. Therefore, the disadvantages such as the incompleteness or error of ANN input data are resolved well through RSANN, and it has rapid reasoning, powerful error tolerance ability. In the end, the simulation experiment in power grid fault diagnosis shows the availability and accuracy of this method.