Forecasting of river flow data with a general regression neutral network

Abstract

This paper proposes a simple one-parameter neural network model, General Regression Neural Network (GRNN), for forecasting chaotic time series. The approach employs the theory of phase-space to reconstruct the evolution trajectory of motion, which is used as the input. In contrast to the nonlinear prediction method (NLP), where the weight of the projected state is the same, the GRNN uses unequal weights. The nearer projected state is weighted heavier than the remotely projected state, a reasonable approximation in the phase-space. The performance of the GRNN is first verified on an artificial chaotic time series and then on a real hydrological time series. The results indicate that GRNN's performance is comparable to that of NLP.