A Gradient RBF Network Based on Multi-Innovation Algorithm for Nonstationary Time Series Prediction

Time series of natural phenomena including human behaviours usually show irregular fluctuations. In this paper, a fast tunable gradient radial basis function (GRBF) network based on the multi-innovation algorithm for modeling and predicting nonstationary and nonlinear time series is proposed. In the process of constructing the frame of the network, the orthogonal least squares algorithm is used to simplify the frame of the GRBF network. During the online adaption, the node with the minimal effection is replaced with an appropriate node while the network cannot track the changes of the time series. The algorithm based on the multi-innovation theory is used to refresh the weight vector. Lorenz time series is used to verify the online prediction performance of the proposed method.