Combining Machine Learning and Survival Statistics to Predict Remaining Service Life of Watermains

Abstract

Distribution systems throughout North America are deteriorating and pipe breaks are increasing. To deal with these infrastructure crises, utilities have begun to adopt proactive pipe replacement models to determine which pipes to replace and when. This paper develops the first random survival forest watermain pipe replacement model that incorporates survival analysis techniques into a machine learning framework, avoiding major limitations associated with other popular models. The random survival forest (RSF) model (C-index=0.880) employed in this paper outperforms the Weibull proportional hazard survival model (C-index=0.734) and the random forest machine learning model (C-index=0.807). The results indicate that by adopting the RSF model, a utility avoids costly early pipe replacement, with a case study suggesting a reduction in pipe replacement and repair costs by 14% over the next 50 years. Overall, the findings indicate that by adopting the RSF algorithm, which incorporates right-censored break data, a utility would be able to more accurately predict future pipe breaks, spread out pipe replacement over a longer range of years, and identify financial savings available from a more effective pipe replacement strategy.