Risk-based water pipe failure prediction through machine learning and hydraulic models

Abstract

Failures in Water Transmission Lines (WTLs) can cause severe disruptions, high repair costs, and extensive damage to surrounding areas. To mitigate these risks, there is a growing interest in proactive management through predictive methods. Traditional methods for identifying high-risk pipelines are often constrained by the scarcity of available data, especially for large-diameter pipes, making accurate failure prediction challenging. This study proposes a hybrid approach combining machine learning techniques with hydraulic modelling to improve the accuracy of pipe failure risk predictions. Using data from 48 water transmission networks in South Korea, including pipe intrinsic properties, environmental conditions, operational factors, and failure history from 2008 to 2023, the study applied tree-based machine learning models (Random Forest, XGBoost, and CatBoost) along with various data sampling techniques to predict the probability of pipe failure. XGBoost with class weighting showed the best performance across key evaluation metrics, including F1-score, F2- score, AUC-ROC and AUC-PR. A hydraulic model was used to assess the impact of pipe isolation, quantifying water supply shortages and secondary damages. An economic analysis was conducted to prioritize pipeline rehabilitation, balancing the cost of repair with risk reduction. The results demonstrate a cost-effective approach to risk-based maintenance planning, enabling utilities to allocate resources efficiently for pipeline rehabilitation while minimizing disruption and water loss.