Detection and localization limits of sparse pressure-flow monitoring: a machine-learning evaluation in water distribution networks

ABSTRACT

Machine learning is increasingly proposed for leak detection and localization in water distribution networks. Yet performance varies with evaluation practice, and the spatial precision attainable from sparse telemetry remains unquantified. We measure the full inference hierarchy, from detection and severity through zone and nearest sensor to the pipe, on a calibrated pressure-driven Moroccan district network with ten pressure and five flow sensors. Evaluation uses 648 scenarios and five tree-ensemble classifiers under a leakage-controlled protocol. Performance decays monotonically with spatial granularity: detection reaches an area under the precision–recall curve of 0.901; severity and zone reach 0.774 and 0.700; nearest-sensor accuracy falls to 0.423 (0.185 system level); and top-5 pipe accuracy is 0.122 over 171 candidates. Three diagnostics attribute the losses to the sensing layout rather than the algorithms. System-level accuracy factorizes as conditional skill times coverage; a zone oracle lifts top-5 accuracy only to 0.170; and pipe ranking stays near its full-layout value across six sensor configurations. Feature attribution shows a hydraulically faithful detector dominated by flow-deviation features. Results provide a calibrated reference for sparse pressure-flow monitoring; pipe-level inference is reframed as a pre-localization step for acoustic survey, with sensing coverage the binding priority.