Failure Prediction in Large Diameter Trunks Using Machine Learning Models

Failures in large-diameter trunk mains can disrupt bulk water transfer and incur high repair costs, yet the low frequency of historical breaks makes proactive management challenging. This study develops a predictive-maintenance framework that applies three tree-based machine-learning classifiers—Random Forest, eXtreme Gradient Boosting and LightGBM—to estimate the likelihood of failure (LoF) for individual pipes. Predictor variables include intrinsic attributes (length, age, diameter, material) and hydraulic conditions (pressure, velocity) for a 670-km UK trunk-main network. The models were trained on 1996–2014 data, validated on 2015–2019 records and tested on 2020–2024 failures. All algorithms achieved area-under-ROC scores above 0.80; LightGBM reached 0.85, while Random Forest yielded the highest F1-score. Replacing the 5 % of pipes with the greatest LoF would have prevented 31 % of observed failures in the test period. The approach offers water utilities a robust, data-driven tool for prioritising capital investment and enhancing network resilience under current regulatory and sustainability performance targets. This paper was presented at the 21st Computing and Control in the Water Industry Conference (CCWI 2025) at the University of Sheffield (1st - 3rd September 2025).