Prediction of water main failures with the spatial clustering of breaks

Due to limited budgets and an aging system, infrastructure managers have increasingly sought cost-effective means to evaluate asset condition. This is a particular challenge for water distribution systems due to the vast amount of buried and unseen pipelines. A spatial clustering of pipe breaks fits well into a wider asset management framework with the aim of identifying regions with abnormally high failure rates. The information about spatial clusters identified using historical breaks, if and where they exist, can potentially improve predictions on the location of future breaks. In this research, we present three algorithms (poisson based, density based, and locally weighted density based) for scanning and clustering pipe break data and demonstrate their application on a real pipeline network. We also explore whether the use of spatial clusters as an explanatory variable can improve the accuracy of pipe break machine learning models. Empirical findings show that the locally weighted density scan provides the greatest precision for finding high breakage zones. The application of these clusters generally improves the performance of predictive models by helping them prioritize high risk pipes with greater accuracy.