Radial Basis Function (RBF) neural network plays an important role in nonlinear time series prediction. In this paper, we propose a diagonal generalized RBF neural network model, utilize the Bayesian Ying-Yang (BYY) harmony learning algorithm for the selection of number of hidden units and the setting of initial parameters, and construct a synchronous LMS learning algorithm for parameter learning. Moreover, the diagonal generalized RBF neural network is applied to nonlinear time series prediction, with the advantages of high prediction accuracy and fast speed.