Unsupervised Identification of Post-Rainfall Behavior in a Medium Reservoir Using Feature-Engineered Clustering

Understanding reservoir behavior after rainfall is essential for effective water management, especially in medium-sized reservoirs with limited data. This study uses a daily time-series dataset comprising 2,922 days of observations from 2014 to 2023 and proposes a novel unsupervised learning approach to identify operational patterns during post-rainfall periods, which are critical for deciding whether to store or release water. The method integrates rainfall-regime segmentation with feature engineering to emphasize hydrologically significant periods. Real-world data from a medium-sized reservoir were analyzed, focusing on inflow, storage, outflow, and usable water variables. A post-rainfall subset was derived to capture the system's recovery behavior following rainfall. Clustering of these post-rainfall days revealed three distinct operational modes: (1) flood response with high inflow and release, (2) recharge with moderate inflow and retention, and (3) idle or drought with minimal activity. Internal validation demonstrated strong performance (Silhouette coefficient = 0.565, Davies–Bouldin Index (DBI) = 0.550), outperforming traditional approaches. These findings highlight the value of context-aware clustering in revealing interpretable, operation-relevant patterns, providing a practical decision-support tool for medium reservoirs under data limitations.