An Interpretable Ensemble Learning Framework for Leakage Localization in Urban Water Distribution Networks

Abstract

Rapid and accurate leakage localization in urban water distribution networks (WDNs) is vital for reducing non-revenue water and supporting sustainable water resource management. This study develops a spatial–hydraulic coupled and interpretable machine learning framework for zonal leakage localization under operational uncertainties. The WDN is first partitioned into zones, followed by leakage-region prediction using a Stacking ensemble integrating Random Forest, XGBoost, and MLP. SHAP analysis is employed to characterize the contribution of individual monitoring nodes to regional classification. The proposed Stacking framework achieves an overall localization accuracy of 0.9667 and a Macro-F1 score of 0.9666, with relatively stable diagnostic performance under pressure noise. Global SHAP analysis identifies several monitoring nodes with substantial contributions to the overall classification, while the observed misclassifications are mainly concentrated between a small number of neighboring regions with similar feature representations. A further test on the Net3 network achieves an overall localization accuracy of 0.9812, providing additional evidence of the applicability of the framework across different network configurations. Overall, the results indicate that the proposed framework provides an accurate and interpretable approach for leakage localization in WDNs.