Abstract: Aiming at some problems in multivariate series prediction of complex systems, such as data samples excess, information redundancy and so on, the radial basis function (RBF) network is improved from two aspects: the learning samples selection and the clustering centers optimization. Based on multivariate time series of complex systems, a linear correlation function and a nonlinear correlation function are used respectively to detect the linear correlations and the nonlinear correlations in the multivariate states firstly. A small data set which includes effective information of the system is defined. Then based on the small data set, K-means clustering algorithm is applied to adjusting hidden layer's clustering centers of the RBF neural network. A local search procedure is introduced to optimize the clustering centers. Network weights are determined by inputting other training samples. Simulation results show that compared with the learning method based on conventional RBF network, the improved method determines the clustering centers of the network more effectively and gets better prediction accuracy when they have same numbers of hidden layer's nodes.