A Climate-Informed Early Warning Framework for Urban Water Pipe Leakage: Integrating Environmental Drivers with LSTM Based Risk Prediction

Frequent leakage events in aging urban water distribution networks (WDNs) pose increasing challenges under evolving climate and infrastructure stressors. This study develops a climate-informed early warning framework for leakage risk assessment, applied to K-City, a subtropical highland city in China. To support model design, meteorological variables (temperature metrics, precipitation), soil moisture (SM) and their seasonal impacts on leakage occurrence were first analyzed. Consumption-related dynamics and pressure-flow interactions were further examined to understand their role in system stress. The analysis revealed heightened vulnerability in smaller-diameter pipelines and during dry, post-monsoon periods. Building on these insights, a Long Short-Term Memory (LSTM) model was developed to perform both point prediction of daily leakage numbers (DLNs) and classification of leakage risk into four severity levels. Model performance was benchmarked against Random Forest (RF) and Extreme Gradient Boosting (XGBoost) alternatives, and interpretability was enhanced using Shapley Additive Explanations (SHAP) to assess the contribution of each input variable. A 7-day input lag yielded the best results, enabling early warning of leakage risks up to one week in advance. While point prediction accuracy was moderate, the LSTM model demonstrated robust classification performance, achieving a total accuracy (TA) of 0.78 and an F1-score of 0.70 across all pipe categories and outperforming both RF and XGBoost despite the complexity of multi-class leakage prediction. The proposed framework demonstrates high potential for practical deployment in proactive leakage management, particularly in resource-constrained urban settings.