Enhancing Dam Operation Strategies Using Deep Reinforcement Learning and Real-Time Observation Data

Código

I-EBHE0089

Autores

Kim, Sunmin, Shibata, Masaharu, Tachikawa, Yasuto

Tema

WG 1.12: Development & application of river basin simulators

Resumo

In recent years, the severity of rainfall-related disasters has increased, underscoring the need to enhance flood control functions of dam reservoir operations. However, the challenges posed by the limited accuracy of current forecasting technologies and the absence of clear operational rules during abnormal flood events remain significant. This study investigates the application of deep reinforcement learning (DRL) to support decision- making in water release operations during dam flooding, leveraging real-time data rather than uncertain predictions. This research shifts away from predictive methodologies, instead utilizing real-time data on inflow, discharge, water levels, and other relevant meteorological variables, such as precipitation, up to the current moment at the dam site. This real-time data is employed to estimate the timing of floods at the dam site. The primary objective is to develop a DRL model that provides decision support for water release operations. The model is implemented and evaluated at the Hiyoshi Dam within the Yodo River system. For model training, data from three past flood events are utilized. The model's performance and learning capabilities are assessed using these training events as well as two additional past flood events not used in training. Notably, the Hiyoshi Dam experienced a significant flood characterized by abnormal conditions, requiring critical decision-making during disaster prevention operations. The study focuses on optimizing neural network settings and reward structures to enhance model performance. The well-trained model demonstrates proficient operation across all flood events, including those not used in training, showcasing its general applicability. Additionally, the model's robustness is validated by examining its behavior under various conditions, such as changes in input variables, flood event start times, and initial water storage levels. The model exhibits general-purpose knowledge acquisition without overfitting. This research contributes to the advancement of dam operation strategies by employing deep reinforcement learning and real- time data, thus facilitating effective decision-making during dam flooding and mitigating the impact of rainfall- induced disasters.