Multi-objective robust optimization of reservoir operation for real-time flood control under forecasting uncertainty

Abstract

Flood control operation is one of the effective measures to reduce flood risks. Since flood forecasting plays a critical role in real-time reservoir flood control operation, it is necessary to involve forecasting uncertainty in the optimal reservoir operation to inform decision-making. It has always been a difficult task to reduce flood risks by adjusting flood control strategies. To tackle this challenge, this study developed a multi-objective robust optimization methodology for real-time reservoir flood control operation, which mainly coped with forecast uncertainty. Three machine learning (ML) models, including back propagation neural network (BP), long short-term memory neural network (LSTM) and extreme learning machine (ELM) were adopted to forecast the short-term reservoir inflow. A stacking ensemble multi-ML model (SEM) was then applied to integrate the forecasting results of the above models. Furthermore, a multi-objective robust optimal operation model (MOROU) integrating both upstream and downstream safety was established to reduce flood risks and evaluate the impact of forecasting uncertainty on the reservoir operation. To maximize the efficiency of reservoir utilization, this study defined a new indicator of reservoir reserved capacity adaptation (RRCA) as one of the optimization indicators. To better solve the complex multi-objective problem, a scenario to point (STP) method was proposed for searching robust solutions of multi-objective optimization models. Methodologies were validated through an application to the Lishimen reservoir, China. Three main conclusions were derived from the study: (1) Three ML models performed well in flood forecasting with BP being slightly better than the other two, and the SEM method was found to be able to incorporate the characteristics of each model, outperforming the individual models; (2) MOROU showed a narrower distribution of flood risk both upstream and downstream and achieved an approximately 1.5% reduction in the maximum value of the highest water level; (3) RRCA was verified to enable to reduce the discharge flow by an average of 4.52% without occupying additional flood control reservoir capacity, confirming that it can be used as an optimization indicator to improve the utilization of the reservoir. The proposed method can provide robust strategies for real-time reservoir flood control under forecast uncertainty.