Journal of Korean Society of Agricultural Engineers ISSN:1738-3692(Print) 2093-7709(Online) 한국농공학회논문집

The operation of agricultural reservoirs during flood seasons requires timely pre-release strategies to comply with restricted water levels and mitigate downstream risks. Effective decision-making depends on accurate inflow, storage, and release data; however, such data are often limited. Reinforcement Learning (RL) offers an advantage in such uncertain environments by learning optimal policies without predefined answers. In this study, we developed a reservoir operation simulation technique using the Proximal Policy Optimization (PPO) algorithm and trained the model. The trained agent demonstrated stable convergence, with reward values concentrated in high-performance ranges and minimal policy fluctuations in later episodes. Simulations applying a maximum allowable release of 2,475 m³/s-based on observed historical operations-resulted in a 1.2% (623 m³) lower total release and a 10% (247.5 m³/s) lower peak release compared to the actual case. Final water levels remained within flood-season restricted water level. In a more conservative scenario reflecting Yedang Reservoir’s guideline (1,645 m³/s max release), the RL model achieved a 2.7% (1,439 m³) lower total release and a 33.5% (830 m³/s) reduction in peak release compared to the real operation. Although peak release duration increased from 7 to 15 hours, minimum and maximum water levels matched the observed case, and the final level (21.63 EL.m) was close to the target flood limit (21.5 EL.m). These results suggest that flood response could have been successfully managed even with stricter release constraints.