Application example of neural networks for time series analysis: Rainfall-runoff modeling

The study presented in this paper represents a variable selection method based on the stochastic exploration of the input space referring to the input variables neighborhood. The ANN structure properly instructed to perform the required mapping, had been submitted to the known input variables, but contaminated with noise purposefully to reduce the initial number of used variables through the determination of the relevancy of each variable initially used. The proposed method is comparatively analyzed with standard stepwise regression method widely used in the variable selection procedure. The proposed method showed better performances in comparison to stepwise regression method. The problem associated with this method may be the time required for the high dimensional input space exploration. (C) 1998 Elsevier Science B.V. All rights reserved.