基于降雨预报信息的梯级水电站不确定优化调度研究I:聚合分解降维

Stochastic optimization operation for cascade hydropower reservoirs by using precipitation forecasts I. Using aggregation-decomposition methodology to reduce high computational costs

基于降雨预报信息的水库群预报优化调度有利于提高水库群水电站发电效益。本文首先采用聚合分解思想将梯级水库群来水量和库容聚合等效为单库,从而简化水库群径流过程的描述和降低高维计算空间,使随机动态规划模型(SDP)在梯级水库群的应用中可以考虑更多的信息来提高模型效率;然后在径流预报中考虑美国全球预报系统(GFS)发布的未来10d降雨预报信息,来提高中期径流预报精度;最后在考虑径流预报不确定性的基础上建立了聚合分解贝叶斯随机动态规划模型(AD-BSDP)。同时与传统调度图、聚合来水量的随机动态规划模型(AF-SDP)和聚合来水量、库容的聚合分解随机动态规划模型(AD-SDP)进行对比分析,其结果表明,考虑预报信息不确定性的AD-BSDP模型比其他模型具有更高的效率和稳定性。

This paper presents an aggregation-decomposition Bayesian stochastic dynamical optimization mod- el （AD-BSDP） to reduce the complexity of the model used in cascade reservoirs and to make the solution computationally tractable. Firstly an aggregation flow stochastic dynamical optimization model （AF-SDP） is constructed in this study by considering aggregate inflow and storages of individual reservoirs as state vari- ables. Secondly an aggregation-decomposition stochastic dynamical optimization model （AD-SDP） is devel- oped to reduce computation burden by aggregating inflows and storages of individual reservoirs. Then the AD-BSDP model is developed by embedding a Bayesian approach in AD-SDP, and the Bayesian approach is used to address forecast uncertainty and inflow uncertainty. This paper use a 10-day precipitation value of the Quantitative Precipitation Forecasts （QPF） by the Global Forecast System （GFS） as input data to run hydrological rainfall-runoff model forecasting the 10-day average inflow of Hun River. And the AD-BS- DP model runs iteratively until if reaches steady stage, using backward recursion, to generate the 10-day steady-state operating policy incorporate inflow forecasts. Finally performance of the AD-BSDP is compared with General Operation Chart, AF-SDP, AD-SDP to evaluate the value of the GFS-QPF in hydropower generation and the efficiency of the models. The simulation results demonstrate that the GFS-QPF is benefi- cial to the Hun River hydropower system and the AD-BSDP model is efficiency than the others.