DBSCAN-Leak, Novel Real-Time Water Leak Detection: A Case Study of SunRise Demonstrator

Abstract

Losses of Non-Revenue Water (NRW) due to undetected leaks threaten the sustainability and operational efficiency of water distribution systems. This study presents an adaptive real-time DBSCAN-Leak algorithm. This algorithm enhances the conventional Density-Based Spatial Clustering of Applications with Noise (DBSCAN) method for anomaly detection in water distribution systems. Unlike traditional approaches that require static thresholds or complex hydraulic models, DBSCAN-Leak relies solely on flow measurements and dynamically adapts to day-of-week and hour-of-day consumption patterns. The framework utilizes a dynamic neighborhood radius and statistical confidence intervals to estimate the leak probability in real-time. The algorithm was validated using the flow data from the Lille University Scientific Campus. Compared to the original DBSCAN, DBSCAN-Leak achieved significant improvements, detecting 16 confirmed leaks versus one, with 100% precision, 98% accuracy, 80% recall, and an F1-score of 89%. These results demonstrate a low-cost, scalable solution that requires no additional hardware. This approach offers encouraging potential for utility-scale implementation and is extensible to other infrastructure networks, such as gas or electricity distribution networks.