The Temporal Transferability of Parameters of Reservoir Long-Term Optimal Operation Models Based on BP ANN

Abstract

Initial parameters are the main influence factors of training time in a specific artificial neural network (ANN) model of which the structure has been determined. Under the background of long-term reservoir operation of The Three Gorges, Back Propagation (BP) ANN models were built to obtain optimal operation rules. Simulation experiments were carried out to compare the difference of training times between two schemes of initial parameters calibration, which are randomizing generation and transferring parameters calibrated from previous training under similar situation. Using feasible degree and efficiency improving degree to evaluate the temporal transferability of parameters when the new training samples were added with time continuously. Based on the outputs of flood season, dry season and a whole year, the results show that in a certain time span, the temporal transfer of parameters is feasible and the efficiency is improved significantly, seasonal differences are shown in results, the performance of transferability tends to be weaken down with time.