Neural network model for hydrological forecasting based on multivariate phase space reconstruction

Hydrology system is a nonlinear and dynamic system influenced by such factors as meteorological climate and topography, which it is difficult to forecast by conventional methods. In this paper, the forecasting model for runoff, based on multivariate phase space reconstruction, was put forward by combination with neural network and partial least square method (PLS) to make the most of the information provided by the system and improve the forecasting precision. The phase space of multivariate time series was reconstructed by the time delays and embedding dimensions chosen for each univariate time series. Then the partial least square method was used to extract the most important components from the constructed time series as neural network input, and the neural network was used to solve the nonlinear prediction problem of runoff. The detailed steps of the model were given in the paper. Finally, as an example, the model was built to forecast month runoff of upstream in Zhanghe river, and the comparison of the neural network model with univariate time series was given. The result shows the multivariate model improves the prediction accuracy over univariate time series one.