Construction of a leak detection and pressure regulation system for municipal water distribution networks integrating digital twin technology

Abstract

Abstract Aging municipal water networks lose a substantial share of treated water to leakage, yet detection and pressure control are still pursued on separate models, so neither can inform the other. This paper builds an integrated system in which both functions read from and write to one continuously synchronized digital twin. The twin couples an open-source hydraulic kernel, a pressure-dependent demand closure and emitter-based leakage to an ensemble Kalman filter and a weighted least-squares recalibration step, so that heads, flows, Hazen–Williams roughness coefficients and demand multipliers are updated from multi-source telemetry rather than frozen at commissioning. Detection normalizes twin residuals against diurnal demand patterns and combines a sliding-window CUSUM statistic with a Mahalanobis joint test; the residual field then feeds a graph attention network that scores each junction and aggregates node posteriors into ranked suspect pipe segments. Regulation solves a bi-objective program over valve settings and pump speeds, weighting nodes by their leak posterior so that head is reduced hardest where water actually escapes, subject to service-head, switching and resilience constraints. Results are reported network by network rather than pooled alone. The pooled F1 reaches 0.916 and the Top-5 hit rate 0.816, against 0.836 and 0.573 for the strongest baseline, with per-network F1 ranging from 0.903 on C-Town to 0.932 on Net3 and mean detection latency falling to roughly four hours. On the official BattLeDIM 2019 scenarios, held out as an independent validation set, the same model retains an F1 of 0.818 but sheds about twenty points of Top-5 hit rate, which bounds how far the in-house training distribution generalizes. Closing the loop on C-Town raises leakage reduction from 24.6% under uniform depressurization to 37.5% while holding service compliance at 99.4%. Ablations identify state assimilation as the load-bearing component. Code and result records accompany the paper; scalability beyond a few thousand nodes, transient hydraulics and the capital cost of dense instrumentation remain open limitations.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗