Probabilistic leak localization in water distribution networks using a hybrid data-driven and model-based approach

Abstract

Abstract In this paper, a hybrid leak localization approach in WDNs is proposed, combining both model-based and data-driven modeling. Pressure heads of leak scenarios are simulated using a hydraulic model, and then used to train a machine-learning-based leak localization model. A key element of the methodology is that discrepancies between simulated and measured pressures are accounted for using a dynamically calculated bias correction, based on historical pressure measurements. Data of in-field leak experiments in operational water distribution networks were produced to evaluate our approach on realistic test data. The results show that the leak localization model is able to reduce the leak search region in parts of the network where leaks induce detectable drops in pressure. When this is not the case, the model still localizes the leak but is able to indicate a higher level of uncertainty with respect to its leak predictions.