Towards resilient water networks and proactive water management: machine learning frameworks for forecasting, detection, and localization of leakages

Efficient and sustainable water distribution is critical to addressing the global challenges of ageing infrastructure, increasing demand, and climate variability.Modern water distribution systems suffer from significant inefficiencies, with up to 25% of treated water lost to leakages in the UK and a yearly global loss of 32 billion m of treated water.These losses not only strain water resources but also lead to increased energy consumption and elevated greenhouse gas emissions, as additional energy is required to treat and pump water that ultimately does not reach consumers.Overall, this results in resource wastage, operational disruptions, increased energy costs, avoidable carbon emissions, and diminished public trust.To mitigate these issues, an integrated framework incorporating three pillars -forecasting, detection, and localization of leakages -is essential.For forecasting, the proposed framework includes a class-specific attention-based deep learning model which utilizes historical flow data, weather variables (including temperature, humidity, solar radiation, evaporation, and precipitation), and recent consumption patterns of the preceding five days to unifiedly predict future water demand and probability of leakages in a given location aggregated by district-metered area (DMA).These predictions enable utilities to preemptively allocate resources and reduce the burden on overworked infrastructure.For leakage detection and complementing the forecasting, a domaininformed variational autoencoder (VAE) leverages the recorded flow data to associate irregular flow