Hybrid Machine Learning for Leak Localization in the Drinking Water Grid

Water distribution companies around the world struggle to meet the demands of an ever-increasing population.
In Flanders, approximately 60 million m3 of drinking water is lost every year due to leaks in the drinking water grid. In times of water scarcity, losing so much treated water is unacceptable.

A key approach to reduce leak losses is by combining hydraulic modeling and machine learning to localize leaks in a hybrid model-based and data-driven manner. We present our hybrid leak localization methodology, which also integrates human feedback from drinking water consumers into its predictions.

Our methodology is evaluated on 10 leak experiments in real-time. We show that leaks can be localized down to the street-level, with each leak location prediction needing less than 2 hours of computing time. As a result, our solution enables Flanders to save approximately 24 million m3 of water per year.