Comparative analysis of machine learning techniques for predicting water main failures in the City of Kitchener

The resilience of water main networks highly depends on the capacity for identifying and fixing structural failures in the system as fast as possible. Given the buried nature of such systems, this will be hard and costly through manual or semi-automated inspections. In this paper, a data-driven method is applied to predict the failure of water mains in the City of Kitchener. Six machine learning prediction models were developed under two scenarios: global models, which consider the three dominant material types in the network; and the homogenous model, which considers only cast-iron pipes. The water main’s condition score, length, and criticality score were the most influential factors on the pipe failure. The random forest models outperformed the other machine learning models with an accuracy of 97.3% and an F1-score of 80.4%; the homogenous modeling showed superior performance than the global one with an F1-score of 86.0%. The results showed that more than 72% of breaks could have been potentially prevented by monitoring and upgrading only 8% of the network. The superiority of the developed models lies in their ability to predict pipe failures with the least number of false alarms.