Prediction of Leaks in Water Distribution Networks Using Machine Learning and Lessons Learned

<p dir="ltr">This paper proposes the development of a machine learning prediction tool (MLPT) based on lessons learned from leak incidents using real data from a water distribution network (WDN) in a city in Spain. The main purpose is to promote preventive maintenance and consequently avoid incidents that compromise the sustainability of the system. The proposed tool starts with the data collection in an Incident Hub, which records in detail all water leaks detected in its WDN over a three-year period. Heatmaps for the study parameters (in this case, frequency of occurrence of leaks in specific components) were subsequently generated in order to capture their spatial behavior. These maps were further pre-processed using a migration process to account for uncertainty and better extract information. Each map was flattened, and the vectors were positioned in their respective time-space, creating a matrix that allowed spatial-temporal trends to be inferred through day-to-day interpolation. Samples of 16 days time-windows (15 days prior and the current evaluated day) were obtained from the entire spatial-temporal interpolated matrix and used as inputs to train a U-Net inspired MLPT. As outputs, this model predicts the leak probability heatmaps of the current day (as a control) and the future next 15 days. The performance assessment of the model showed its suitability for short-term leak predictions as the correlation between the forecasted and real outputs is, on average, 0.9778. This forecasting tool is thus able to increase the preparedness of WDN by informing the operating teams, in advance, of areas that need special attention.</p><p dir="ltr">This paper was presented at the 21st Computing and Control in the Water Industry Conference (CCWI 2025) at the University of Sheffield (1st - 3rd September 2025).</p>