Constructive MoG neural networks for pollution data forecasting

Abstract

Time series forecasting problems can be solved by considering them as function approximation problems whose inputs are determined by using past samples of the sequence to be predicted. However, it is possible to show that such an approach can lead to ill-posed data driven modeling problems, especially when the time series to be predicted are characterized by a chaotic behavior. By considering the system generating the sequence to be predicted, the usual approach is intended to synthesize directly the function linking the current sample to a set of past ones. However, it is much more effective to model the transfer state function of this system. In the paper, the function approximation problem is approached by using a neural network based on a mixture of Gaussian (MoG) components. We demonstrate that the proposed technique is particularly suited when dealing with the prediction of environmental data sequences, often characterized by a chaotic behavior.