A novel radial basis function neural network for rainfall forecasting based on Kernel Principal Component Analysis

In a radial basis function neural network (RBF network), the number of hidden layer nodes, centers and width are difficult to identify. In order to improve the network performance, in this study, proposed an improvement RBF algorithm that uses fuzzy clustering algorithm to determine the initial width, and can dynamically determine and adjust the center and width of the Gauss kernel function. In this algorithm, first used the fuzzy clustering analysis method to do the initial clustering, with an initial data width equal to the minimum distance between sets; then applied the Orthogonal Least Squares method to train a new data center, and the number of weights, and modify the width; finally used the gradient descent algorithm to train and adjust the center, the weight and the width. By combining these algorithms and further optimization, the generalization performance of the network is much improved. Because of the large number of precipitation affecting factors, pretreated the sample data using the Kernel Principal Component Analysis (KPCA) for feature extraction to reduce dimensionality. As an experiment, applied the model on daily precipitation forecast in the month of May for three districts in Guangxi. The results show that, the model has good generalization performance, and the forecasting accuracy is higher than that of T213 precipitation forecast model, thus this model has certain promotion value.