Failure-risk classification in water distribution networks using hydraulic modelling and machine learning

Failure-risk prioritization in water distribution systems requires information on both pipe characteristics and the hydraulic consequences of pipe isolation. This field-based case study integrates utility records, EPANET-based shutdown simulations, and supervised machine learning to classify pipe-segment records into low-, medium-, and high-risk categories. The analysis included 512 records. To avoid post-event causality and spatial memorization, the final predictor set was restricted to four physically interpretable variables available under normal operation: pipe diameter, pipe length, mean endpoint pressure, and absolute endpoint pressure difference. Pipe, street, and node identifiers, repair duration, and all shutdown-consequence variables were excluded from the predictor matrix. A majority-class baseline, Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM) were evaluated using stratified pipe-section-level train/validation/test splitting, randomized hyperparameter search with five-fold stratified pipe-section cross-validation, class-wise metrics, ordinal mean absolute error (MAE), quadratic weighted kappa, and severe-error rate. SVM achieved the highest held-out accuracy (0.598) and macro-F1 (0.533), whereas RF showed the most favourable ordinal error profile among the evaluated models, with ordinal MAE 0.451, quadratic weighted kappa 0.385, and a severe two-level error rate of 2.0%. In a supplementary post-selection five-fold stratified pipe-section cross-validation on the complete record dataset, RF achieved the highest mean macro-F1 (0.499 ± 0.076). The results support internal consequence-based screening within the analysed network, but they do not establish transferability to other utilities. Limitations include the single-network design, the use of demand-driven EPANET simulations, and the absence of normal-operation pipe-flow data in the verified analytical dataset.