The prediction of chaotic time series is a vital problem in nonlinear dynamical system. Radial Basis Function Neural Network (RBFNN) has been widely adopted in nonlinear dynamical system identification because of its simple topological structure, fast learning and strong extrapolating capability. The major problem in applying RBFNN is the selection of the number of hidden neurons. In this paper, we adopt the Localized Generalization Error Model (L-GEM) to select number of hidden neurons of RBFNN for chaotic time series prediction. The effectiveness of the L-GEM is evaluated by using two benchmarking chaotic time series datasets: Mackey-Glass series and Lorenz series. Simulations results show that the proposed method provides a better prediction performance in comparison with the RBFNN trained with a cross validation method.