Leakage Detection in Water Distribution Networks using a Physics-Aware Neural Network and Leakage Localization using a Rule-based System

Water Distribution Networks (WDNs) suffer significant water and economic losses due to undetected leaks, requiring reliable and computationally efficient monitoring systems. This study proposes a Physics-Aware Artificial Neural Network (PA-ANN) that
embeds hydraulic constraints directly into the learning objective by enforcing conservation of mass and energy using the Equation of Continuity and the Hazen-Williams energy equation.

The model is evaluated on the LeakDB benchmark using the Hanoi Water Distribution Network dataset comprising of ∼17.5
million data-points. Compared to the base ANN model, the proposed model achieves 93.03% accuracy, 96.59% precision, and an F-1 score of 85.98%, while reducing false positives by 62.98%.

In addition, a rule-based localization system reduces average fault isolation time by 63.6%, while maintaining 79.19% localization
accuracy. The results demonstrate that embedding physical constraints into the model architecture significantly improves detection
reliability, and reduces operational false-alarms.