Item

Abstract

The accuracy of flood forecasting plays a critical role in flood control operations. It is challenging to reduce flood risk and inform decision-making by adjusting reservoir scheduling strategies under forecasting uncertainty. This study developed a many-objective robust optimization methodology for real-time reservoir flood control operation. Three different machine learning (ML) models were adopted to forecast short-term reservoir inflow and a stacking ensemble multi-ML model (SEM) was then applied to integrate the results. Furthermore, this study established a robust optimal operation model (MOROU) to reduce flood risks and assess the impact of forecast uncertainty on reservoir operation. To improve the efficiency of reservoir utilization, a new indicator called reservoir reserved capacity adaptation (RRCA) was defined and used as one of optimization objectives in MOROU. A scenario to point (STP) method was proposed for searching for robust solutions to solve complex many-objective problems. Methodologies were validated through an application to the Lishimen reservoir, China. The main findings are: (1) all three ML models performed well in flood forecasting, and the SEM model was validated to be able to combine the characteristics of multiple models. (2) MOROU model showed a much narrower distribution for both upstream and downstream flood risks and succeeded in reducing the highest water level by about 1.5%. (3) It was proven that the RRCA could reduce the reservoir discharge flow by an average of 4.52% without taking additional flood control capacity. The findings have significance for searching for robust solutions of real-time reservoir flood control under forecast uncertainty.