Improving Urban Water Security through Pipe-Break Prediction Models: Machine Learning or Survival Analysis

Abstract

North America’s water distribution systems are aging and incurring increased pipe breaks. These breaks pose a serious threat to urban drinking water security, leading to service interruptions, loss of revenue, and increasing risk of water contamination. Prediction models have been developed to help identify when individual underground water pipes are expected to break, helping utilities develop pipe renewal projects and avoid costly pipe breaks that impact water supply reliability. This paper provides an in-depth comparison of the two leading statistical pipe-break modeling methods: machine-learning and survival-analysis algorithms. A gradient-boosting decision tree machine-learning model and a Weibull proportional hazard survival-analysis model are used to predict time to next break for cast-iron pipes in a major Canadian water distribution system. Results indicate that removal of censored events from the machine-learning model biases the model to predict earlier pipe breaks than occur. Overall, water utilities concerned with short-term security arising from impacts of pipe breaks on water security may favor the machine-learning approach, but the survival-analysis models’ ability to incorporate right-censored data makes it more appropriate for long-term asset management planning.