| With the continuous development of the economy,the consumption demand in the field of energy has been increasing.To meet energy demand while mitigating the impact on global climate change,hydropower energy needs to play a major role as an important clean energy source.In the context of the successive completion of cascaded reservoirs of each basin,the importance of achieving joint optimal dispatch of cascaded reservoirs and fully utilizing the comprehensive benefits of the basin’s flood control and profitability has been highlighted.Therefore,focusing on the overall process of medium-and long-term optimal dispatch and evaluation decision of cascaded reservoirs,the medium-and long-term runoff probabilistic prediction method is studied.Considering runoff uncertainty,the runoff probability distribution is used as the input to establish the medium-and long-term optimal dispatch model.Thus,the medium-and long-term operation rules extraction method is proposed,as well as the evaluation decision model of optimal dispatch scheme is constructed through the dispatch indicators of cascaded reservoirs.The study uses the Xiluodu-Xiangjiaba cascaded reservoir as an example to validate the proposed models and methods,which has achieved some innovative results that would be valuable for future research.(1)A monthly runoff probabilistic prediction model based on dynamic time warping and kernel density estimation as well as a medium-to long-term runoff probabilistic prediction model based on parameter-optimized natural gradient boosting are proposed.Based on the analysis of historical runoff series,the study,by directly predicting the probability distribution function of runoff,obtains the improved accuracy of the deterministic and probabilistic prediction of runoff,summarizes the recommended range of hyperparameter selection for the runoff prediction model,and explores the selection method of the probability distribution function of runoff under different runoff characteristics.(2)A medium-and long-term optimal dispatch model of cascaded reservoirs based on proximal policy optimization is built.On the basis of the relevant power plant parameters and dispatch period constraints of the cascaded reservoirs,a multi-objective dispatch model environment for the cascaded reservoirs under runoff probability distribution is created,and the reinforcement learning algorithm of the policy gradient is used to obtain the optimal dispatch rules.The study shows that the proposed model can,considering the uncertainty of runoff in the form of runoff probability distribution,meet the demands for medium-and long-term multi-objective optimal dispatch in different dispatch periods and different dispatch time scales.(3)A medium-and long-term operation rules extraction model for cascaded reservoirs based on Transformer is constructed.The water level prediction model is established by treating the dispatch characteristic variables of each cascaded reservoir as a time series and using the transformer mechanism.The model is further divided into a separation pattern and an overall pattern based on the inputs and outputs of the model to improve the accuracy of water level prediction.Besides,the effects on model inputs and model construction patterns brought by model training data size and operation rules extraction time scales are analyzed.(4)With the probability distribution of incoming runoff,the dispatch indicators of cascaded reservoirs under runoff uncertainty are proposed.Also,the normal dispatch scheme calculated by the cascaded reservoirs operation rules extraction model is used as a reference to establish a dispatch evaluation and decision model based on Prospect Theory for the stability improvement of the multiple attribute decision method.Furthermore,in conjunction with the overall process of cascaded reservoir dispatch,a set of optimal dispatching and evaluation decision methods is proposed for cascaded reservoirs,with high engineering application value. |
