Abstract

Drinking water networks are among the essential infrastructure in cities worldwide. The failure of water mains jeopardizes this essential service and the safety of water users. However, across North America, the failure rate of older water mains has been increasing. The goal of this study is to compare the accuracy and applicability of machine learning algorithms to predict water main deterioration across Canadian water systems. In previous studies, different approaches were applied to only one or a few utilities. Nevertheless, it is valuable to compare results among various networks with different characteristics and levels of data collection. Accordingly, data was collected from thirteen Canadian water utilities, including Barrie, Calgary, Halifax, Kitchener, Markham, Region of Durham, Region of Waterloo, Saskatoon, St. John’s, Waterloo, Winnipeg, Victoria, and Vancouver. A variety of factors, including intrinsic, environmental, and operational, were used to develop more reliable predictions and assess the relative importance of each factor. Random forest (RF), artificial neural networks (ANN), extreme gradient boosting (XGBOOST), and logistic regression (LR) were applied to predict the probability of failure. Furthermore, RF, ANN, XGBOOST, and ElasticNet regression models were employed to predict age at first failure and the current rate of failures. Results indicated the superiority of XGBOOST over other models in predicting the probability of failure and the current rate of failure. However, for age at first failure, RF performed better. When datasets were significantly imbalanced, the application of the Synthetic Minority Oversampling Technique (SMOTE) provided more accurate predictions. Because these models provide predictions for every pipe in the network, they can be mapped to facilitate the visualization of deterioration. While models created for one utility cannot be accurately applied to other utilities, the same machine learning algorithms can be quickly and effectively adapted to multiple utilities. Overall, these models support robust and data-driven asset management decision-making.