INTEGRATING SWITCHED PREDICTION METHOD AND MULTIPLE ARTIFICIAL NEURAL NETWORKS FOR PREDICTING INFLOW AND SEDIMENT INTO RESERVOIRS

抄録

 Accurate predictions of the reservoir inflow and sediment concentration are necessary for real-time reservoir operation. This study used multiple artificial neural networks (ANNs), namely the back propagation networks (BPN) and four types of kernel function of support vector machines (SVM), to predict inflow and sediment concentration. These ANNs were calibrated and validated based on observed data of the Shihmen reservoir for typhoon events from 2008 to 2015. To avoid the risk of selecting multiple ANNs, the switched prediction method (SPM) is proposed to select the optimal predicting module time by time. This paper compares the predictions from SPM with optimal individual predictions and the ensemble means (EM) with respect to the root mean square error. The improvements in SPM compared with optimal individual ANN and EM are 3.8% and 10%, respectively. In conclusion, the uncertainty of the predictions could be effectively reduced by applying the switched prediction method.