An Explainable Machine Learning Framework for Leakage Diagnosis in Urban Water Distribution Networks: Implications for Water–Energy–Carbon Management

Urban water distribution network leakage not only causes significant water loss but also increases energy consumption and carbon emissions during water supply operations. To support intelligent leakage management, this study proposes an explainable machine-learning framework based on pressure-window features extracted from hydraulic simulation data. A total of fifteen statistical and temporal features are constructed from pressure signals, and an XGBoost classifier is developed to identify leakage conditions. SHAP (SHapley Additive exPlanations) is employed to quantify feature importance and improve model interpretability, enabling transparent understanding of the decision-making process.Experimental results based on the LeakDB Hanoi benchmark demonstrate that the proposed framework achieves accurate leakage diagnosis while identifying several key pressure-related indicators that dominate model predictions. In addition to leakage detection, a simplified Water–Energy–Carbon assessment framework is introduced to qualitatively illustrate how improved leakage diagnosis may contribute to reducing water losses, pumping energy consumption, and associated carbon emissions. Furthermore, sensor deployment considerations are discussed to provide practical guidance for monitoring network construction.Although the current study is validated using the LeakDB benchmark, future work will incorporate real-world water distribution network data to further verify model robustness and engineering applicability. More comprehensive quantitative analyses will also be conducted to establish explicit relationships between leakage diagnosis performance, sensor deployment strategies, and Water–Energy–Carbon benefits under different operational scenarios. These improvements are expected to enhance the generalizability and practical value of explainable machine learning for smart water management.