A novel hybrid neural network based on phase space reconstruction technique for daily river flow prediction

The main purpose of this study is to construct a new hybrid model (PSR-ANN) by combining phase space reconstruction (PSR) and artificial neural network (ANN) techniques to raise the accuracy for the prediction of daily river flow. For this purpose, river flow data at three measurement stations of the USA were used. To reconstruct the phase space and determine the input data for the PSR-ANN method, the delay time and embedding dimension were calculated by average mutual information and false nearest neighbors analysis. The presence of chaotic dynamics in the used data was identified by the correlation dimension methods. The results of the PSR-ANN, pure ANN and gene expression programming (GEP) models were inter-compared using the Nash-Sutcliffe and root-mean-square error criteria. The inter-comparisons showed that the proposed PSR-ANN method provides the best prediction of daily river flow. Moreover, the ANN model showed higher ability than the pure GEP in estimation of the river flow.