Predictive analytics for water main breaks using spatiotemporal data

Water main breaks are a common recurring problem in water distribution networks, resulting in cascading effects in the whole system and the interconnected infrastructures such as transportation. Having integrated the physical features of pipes such as diameter and environmental factors like precipitation, we propose predictive models based on spatiotemporal data and machine learning methods. In this study, the dataset is the main breaks recorded from 2015 to 2020 in the city of Tampa, Florida. First, a spatial clustering is conducted to identify vulnerable areas to breaks. A time series analysis is also carried out for the temporal data. The result of these analyses informed the machine learning algorithms as independent variables. We then compared the predictive models based on information-based and rank-based criteria. Obtained results indicated that Boosted Regression Tree (BRT) model was superior to the others. Finally, we present predicted normalized failure rates for the water distribution network to inform rehabilitation and fortification decisions at the municipality level.