Multivariate phase space reconstruction based on combination of nonlinear correlation degree and ICA

Abstract

In view of multi-input variables in complex system, this paper is to study a new methodology to reconstruct a multivariate input phase space based on combination of independent component analysis (ICA) and nonlinear correlation degree. Firstly, a concept of nonlinear correlation degree is introduced to compute the correlation between different time series. The variables which have stronger correlation with the output are selected as components of input vector. Then, C-C method is used to construct an initial input vector including different time states of selected input variables. Furthermore, the FastICA method is expanded to extract the effective independent information, aiming to reduce the dimension of initial input vector. Finally, RBF network is trained to make prediction for multivariate series, and the simulation results show the effectiveness of the method. © 2012 Springer-Verlag GmbH.