RBF Neural Network Using Improved Differential Evolution for Groundwater Table Prediction

Abstract

Radial basis function (RBF) neural network is increasingly used to predict groundwater table, which often shows complex nonlinear characteristic. But the traditional RBF training algorithm based on gradient descent optimization method can only obtain the partial/local optimums solution sometimes. Furthermore, man-made selecting the structure of RBF neural network has blindness and expends much time. Therefore differential evolution (DE) algorithm was adopted to automatically search the weight of output layer, the center of RBF and the width of network. In order to improve the population's diversity and the ability of escaping from the local optimum, a self-adapting crossover probability factor was presented. Furthermore, a chaotic sequence based on logistic map was employed to self-adaptively adjust mutation factor, which can improve the convergence of DE algorithm. Study case shows that, compared with groundwater level prediction model based on traditional RBF neural network, the new prediction model based on DE trained RBF neural network can greatly improve the convergence speed and prediction precision.