Comparative Machine Learning Survival Models for Water Distribution Pipeline Failure Prediction in South Africa

Abstract Aging water distribution pipelines and resource constraints demand accurate and transferable failure prediction models, particularly for heterogeneous networks. While machine learning survival analysis (MLSA) has shown promise, existing studies remain limited: most focused on homogeneous systems, overlooked the effect of data censoring, and failed to test key extreme gradient boosting survival embedding (XGBSE) variants. Moreover, survival performance has often been assessed solely through discriminatory metrics such as the concordance index (C-index). This study addresses these gaps by systematically evaluating unexplored XGBSE models (XGBSEKaplanTree, XGBSEBootstrapTrees, XGBSEStackedWeibull) against established MLSA approaches (XGBSEKaplanNeighbors, random survival forest), probabilistic models (Cox proportional hazards, Weibull proportional hazards model), and traditional machine learning methods (extreme gradient boosting, random forest). Model performance was assessed using discrimination (C-index), calibration [integrated Brier score (IBS)], and regression accuracy metrics (mean absolute error, mean absolute percentage error, root mean squared error, mean squared error, and coefficient of determination, <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="upper R squared" display="inline" overflow="scroll"> <mml:msup> <mml:mrow> <mml:mi>R</mml:mi> </mml:mrow> <mml:mrow> <mml:mn>2</mml:mn> </mml:mrow> </mml:msup> </mml:math> ). Training data were sourced from one South African municipality and validated using datasets from another South African and a Canadian utility to evaluate model transferability. Results show that XGBSEBootstrapTrees, trained on an optimal censoring ratio (30%–60%), achieved the best performance (IBS: 0.006–0.016; C-index: 0.958–0.967) while maintaining strong predictive accuracy across regression measures. Variable importance analysis identified four critical predictors—failure season (autumn, summer), number of known previous failures, and pipe age—supporting the development of parsimonious models that generalized well across municipalities. By explicitly addressing previously unexplored XGBSE variants, data censoring, and network heterogeneity, this study advances MLSA research and provides practical, transferable tools for utilities. The developed models enhance pipeline failure prediction, reduce water losses, promote intermunicipal data sharing, and contribute to achieving Sustainable Development Goal 6.