Heterogeneous local model networks for time series prediction

Abstract

The approaches of local modeling have emerged as one of the promising methods of time series prediction. By use of the divide-and-conquer method, local models can exploit state-dependent features to approximate a subset of training data accurately. However, the generalization performance of local model networks is subject to the proper selection of model parameters. In this paper, we present a new method for local model construction for the noisy time series prediction. The proposed method uses the principal component analysis (PCA) and cross-validation technique to construct an optimal input vector for each local model. A heuristic learning rule is also proposed to update the mixture of experts network structure, which determines the confidence level of local prediction model. The proposed method has been tested with noisy Mackey–Glass time series and Sunspot series.