Data-driven Approaches For Reservoir Operation

Reservoir is the water conservancy project for flood control and water flow regulation.It can not only reduce the flood damage,but also provide huge benefits.With the popularity of monitoring systems and the rapid development of hydrology and water resource theories,mass hydrological observations and model data can be provided.From the perspective of data-driven,this paper focuses on five major problems existed in the reservoir operation:unidirectional real-time reservoir operation without feedbacks,uncertainty of reservoir operating rules,uncertainty of flood spatial distribution,“curse of dimensionality”-“multiobjective”-“time-consuming” problems existed in the large-scale multi-reservoir operation system and the multivariate flood nonstationarity.The main contents and results are summarized as follows:(1)Closed-loop control system for real-time reservoir operation is proposed based on data assimilation.Based on the optimal release provided by the real-time optimization model,the reservoir storage can be updated through the assimilation of the real-time water level observations by Constrained Ensemble Kalman Filter(CEn KF).Taken the real-time flood control operation of Three Gorges Reservoir as an example,the results show that the closed-loop control system,which combines the experience of reservoir managers and the optimization results,can not only handle the model error and observation error together,but also improve the flood control and the hydropower generation benefit effectively.(2)Robust ensemble reservoir operating rules are derived by Bayesian Model Averaging(BMA).Based on the optimal release trajectory from the reservoir deterministic optimization model,three individual reservoir operating rules are derived by Implicit Stochastic Optimization method(ISO),including piecewise linear regression(PL-REG),surface fitting(SURF),least squares support vector machine(LS-SVM).Then,BMA is used to derive the robust ensemble reservoir operating rule and the operation interval.The flood control of Baise Reservoir in Xijiang River Basin is taken as an example.The results show that the BMA operating rules outperform three individual reservoir operating rules.Besides,the operation interval can provide flexible decision-making for reservoir managers.(3)The “Flood classification-Aggregation-Decomposition”(FAD)operating rules are proposed for multi-reservoir flood control operation.The projection pursuit method is used for flood classification to investigate the uncertainty of flood spatial distribution.Based on the classified floods,the FAD operating rules are derived.Taken multi-reservoir flood control operation in Xijiang River Basin as a case study,the results show that the floods in Xijiang River Basin can be classified into three flood types,including middle and upstream(MU),middle and downstream(MD)and whole basin(WB)floods.The operation results indicate that FAD operating rules can reduce the flood damage in the mainstream and tributaries simultaneously,and it can improve the ability to deal with the uncertainty of flood spatial distribution.(4)Efficient techniques are proposed for large-scale multi-reservoir operation system with multiple objectives.Firstly,the aggregation-decomposition and sensitivity analysis methods are proposed for dimensional reduction.Then,the Weighted Non-dominated Sorting Genetic Algorithm II(WNSGA II)is applied to improve the Pareto frontier searching efficiency through the directed searching towards the non-dominated region by the weighted crowding distance.Finally,the surrogate model,Gaussian process regression,is trained to replace the time-consuming simulation model to improve computational efficiency.The Weighted Multi-Objective Adaptive Surrogate Model Optimization(WMOASMO),which integrates WNSGA II with surrogate model,are applied in the large-scale multi-reservoir operation in Xijiang River Basin.The results show that the proposed methods can increase hydropower generation and reduce ecological damage effectively.The techniques can solve the large-scale multi-reservoir operation system with multiple objectives efficiently from the perspective of “variable-objective-model”.(5)The flood risk calculation methods and adaptive design solutions are investigated under flood multivariate nonstationarity.Firstly,flood multivariate nonstationarity is analyzed based on the long series of flood peak,volume and duration.Then,C-vine CopulaMonte Carlo and Hazard Function Analysis methods are proposed to calculate the flood risk and structure-based return period of reservoir and downstream protection objects under flood multivariate nonstationarity.Finally,the effects of reservoir storage redistribution(transferable storage)on reservoir flood damage and water-use benefit are investigated under flood nonstationarity.Taken Shelbyville reservoir in Illinois,USA as an example,the structure-based return period of reservoir and downstream protection objects follow the Weibull distribution.Besides,the relationship between transferable storage and reservoir flood damage and water-use benefit can give some guidelines for reservoir adaptive design solutions under flood nonstationarity.