A Stacking-Based Ensemble Learning Framework for Water Distribution Network Condition Prediction and Interpretability

Water distribution network (WDN) pipelines are essential infrastructure, and their failures cause significant economic and operational losses. Existing condition prediction models often struggle with high-dimensional non-linearities, severe class imbalance, and a lack of decision-centric interpretability. To address these challenges, this study presents a Stacking-based ensemble learning framework that integrates Random Forest, XGBoost, and LightGBM base learners with a Logistic Regression meta-learner under a spatially constrained group sampling scheme. Model performance is evaluated alongside an independent CatBoost benchmark using macro-averaged metrics, an operational cost-loss function for high-risk assets, and Copeland ranking, supplemented by global and class-specific SHAP interpretability analyses. Evaluated on a pipeline dataset, the Stacking framework achieved superior overall performance and recorded the lowest operational cost loss by minimizing severe misclassifications of critical pipelines. SHAP analysis identified pipe age, material, physical dimensions, temperature, and soil moisture as the primary risk drivers. This integrated framework delivers a robust, transparent, and decision-centric decision support tool for proactive municipal pipeline maintenance and risk mitigation.