A Machine-Learning Framework for Real-Time Non-Revenue Water Prediction in Aging Urban Distribution Networks

Abstract

Non-revenue water (NRW) remains one of the most significant operational and economic challenges affecting aging urban water distribution systems because leakage, infrastructure deterioration, hydraulic instability, and inefficient maintenance practices contribute to substantial water losses and reduced system performance. This study developed and evaluated a comprehensive machine-learning framework for real-time non-revenue water prediction by integrating engineering infrastructure characteristics, hydraulic operational variables, historical maintenance records, and sensor-derived monitoring data. A quantitative research design was employed using a structured engineering dataset consisting of 5,200 validated operational observations collected from urban water distribution networks following data preprocessing, quality assurance, normalization, outlier screening, and feature engineering. Twelve predictive models were comparatively evaluated, including Linear Regression, Logistic Regression, Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbor, Artificial Neural Network, Deep Learning, Gradient Boosting, XGBoost, LightGBM, and an anomaly detection model. Model performance was assessed using prediction accuracy, precision, recall, F1-score, receiver operating characteristic analysis, area under the curve (AUC), root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), cross-validation, bootstrap validation, and residual diagnostics. The findings demonstrated that advanced ensemble learning algorithms consistently outperformed conventional statistical approaches. XGBoost achieved the highest predictive performance with an accuracy of 97.5%, precision of 0.976, recall of 0.974, F1-score of 0.975, AUC of 0.996, RMSE of 0.054, MAE of 0.039, and an explained variance (R²) of 0.987. LightGBM and Deep Learning also demonstrated excellent predictive capability, achieving accuracies of 97.2% and 97.1%, respectively, while Random Forest achieved 95.4% accuracy. Statistical analyses identified pressure variability (β = 0.421), burst frequency (β = 0.392), pipeline age (β = 0.368), repair history (β = 0.341), and flow imbalance (β = 0.317) as the strongest engineering predictors of non-revenue water occurrence. Cross-validation and bootstrap validation confirmed stable model performance, with validation accuracy exceeding 97% for the highest-performing algorithms and residual means approaching zero, indicating excellent generalization capability and minimal prediction bias. Overall, the findings demonstrated that advanced machine-learning techniques provide a highly accurate, reliable, and operationally robust framework for supporting intelligent leakage detection, predictive asset management, optimized maintenance planning, and sustainable urban water distribution system management.