Developing an Open Repository of Water Main Break Prediction Models in Kitchener

This study presents an open repository of predictive machine learning models for water main breaks as a way to help manage water supply networks proactively. Lack of standardized datasets has been a challenge in previous research, a problem which is addressed in the present study through provision of a benchmark dataset that features pipe dimensions, age, proximity to previous breaks, and climatic variables, among other elements. The repository allows for model testing and comparison with machine learning algorithms such as XGBoost and LightGBM. Implemented in Python and available on GitHub, this project promotes a collaborative approach towards the enhancement of urban infrastructure management through accurate prediction of water main breaks, leading to fewer interruptions in service. Findings show that, while random splits work well in training and testing, their performance is poor when it comes to future prediction. Conversely, time-based splits maintain a good consistency between training and testing phases, but they lack the capacity to predict future periods.