Abstract

When the neural network is used for hydrologic forecast, the training and learning process often needs more data. But in the practical application, within limits of the time and condition of observation, the data’s acquiring becomes limited, too. The sediment concentration time serial is such a typical example. In this paper, a kind of Chaos Artificial Neural Network (CANN) is built for the prediction of the sediment concentration, whose construction mainly includes the following parts: The treatment of the sediment concentration time series, establishment of the optimum embedding window and delay time, state space reconstruction and the neural networks’ studying and prediction. Before forecast, with regard to the sediment concentration time series, the methods of carrying out smoothness, difference dividing, reducing base value and proportion enlarging etc are used. According to the auto-correlation function, after determination of the optimum embedding dimension and delay time, this paper then carries out the phase space reconstruction. The qualified rate and the amplitude mean square error are employed as the standards to evaluate the forecast results. This paper uses the daily average sediment concentration data of Yichang hydrological station to predict the sediment concentration time serial of the same station in the future. The forecast results show that it is feasible to predict the sediment concentration with the CANN, and all of the five groups gain a preferable effect. The predictive outcomes of the different combinations are quite different, and they are to be greatly influenced when the CANN carries out the difference treatment to the sediment concentration time serial. The influence mechanism of the parameter combinations should be further analyzed in the future.

Keywords: Chaos Artificial Neural Network (CANN), sediment concentration, prediction