Abstract

As essential infrastructure, water distribution systems provide water to the vital needs of individuals, businesses, and industries. Watermain failure jeopardizes water systems' ability to deliver clean water safely. The main target of this study was to identify the most influential factors on watermain failure across Canada. Dimensionality reduction approaches were applied to watermain data from thirteen Canadian cities, Barrie, Calgary, Region of Durham, Halifax, Kitchener, Region of Markham, Region of Waterloo, Saskatoon, St. John’s, Vancouver, Victoria, Waterloo, and Winnipeg. While previous studies have focused on small datasets of a few cities at a time, the present study compares various factors in different networks with different characteristics. Multiple physical, historical, protection, operational, and environmental factors were compared. Two target attributes were defined, current rate of failure and break status. A correlation analysis was applied to each city to identify the relationships between different attributes and the targets. Four dimensionality reduction approaches were employed to evaluate the impacts of different factors on the targets and identify the most important factors The four approaches are Factor Analysis of Mixed Data (FAMD), Categorical PCA (CATPCA), Random Forest Recursive Feature Elimination (RF-RFECV), and Extreme Gradient Boosting Recursive Feature Elimination (XGBOOST-RFECV). Results indicate CATPCA is more reliable than other approaches. Furthermore, protection activities were found to be more important than physical and historical attributes in most utilities. Thus, the collection of protection data should be prioritized for utilities with higher rates of protection activities, especially if they have already collected data on fundamental physical and historical attributes. While few utilities collect data on environmental, operational, and certain physical factors such as roughness, dead-end, restrained, and pipe depth, these were also found to be important and should be further investigated. These findings create the foundation for a new data collection framework for predicting main breaks.