Application of Wavelet Neural Network for Electric Field Estimation

Abstract

Estimation of electric stress along the electrode and the insulator surfaces using wavelet neural networks (WNNs) has been carried out in this work. The application example considered here is an electrode-spacer arrangement which is used in gas-insulated substations (GISs). Four different WNNs using Gaussian, Morlet, Mexican hat and Shannon wavelet have been used to estimate electric stress over the electrode-insulator arrangement under study. Electric field computations have been carried out by applying boundary element method (BEM). The WNN is trained using a training set comprising of 71 data, and consequently, the trained network is tested with a testing set consisting of 15 data. Root mean squared error is a metric used for ascertaining the accuracy of the trained network while the testing accuracy is determined with the help of mean absolute error (MAE). For a given wavelet function, three parameters of the network, viz., the number of wavelons (N_w), number of iterations (N_it) and the learning factor ($\gamma_{k}$) are exhaustively varied, and the combination of these three parameters that yields least value of RMSE is determined, and the corresponding network is considered as the optimum WNN architecture. Hence, for the given application example corresponding to four wavelet functions, there will be four optimum WNN architectures. Among these four optimum architectures, the one that produces the least MAE will be considered as the best estimator for the application example under study.