A Methodology for Developing and Benchmarking Burst Detection Tools in Water Distribution Systems

Sustainable and reliable operation of water distribution systems requires timely detection of bursts and effective control of real water losses. The digitalisation of water utilities and the increasing deployment of smart monitoring infrastructures are enabling continuous monitoring and diagnostic support, but the development of operational anomaly detection tools remains constrained by scarce labelled events and by the lack of structured workflows for designing, testing and comparing alternative solutions. This study proposes a methodology for developing and benchmarking burst detection tools in water distribution systems. The framework integrates stochastic-hydraulic synthetic data generation to create or enrich labelled datasets, feature engineering to extract temporal and spatial descriptors from hydraulic signals, baseline modelling of normal system behaviour, residual generation, anomaly identification, and systematic performance evaluation. The methodology is applied to a real water distribution network partitioned into nine district metered areas, enabling a consistent comparison of alternative strategies for normal-behaviour modelling and anomaly detection. Results show that the forecasting-based approach, including multi-horizon prediction, produces more informative residuals than the reconstruction-based approach, while XGBoost provides the best overall trade-off between sensitivity and false alarms. These findings demonstrate the value of the proposed methodology for the data-driven development of operational burst detection tools in smart water distribution systems.