ABSTRACT
Water loss due to persistent leakages in water distribution networks remains a substantive problem around the world, all the more so given noticeable trends of increasing global water scarcities. In this paper, we present a data-driven leak localization approach leveraging a connected Geographical Information System together with an autoencoder to perform anomaly detection on time-variable sensor data. Data-driven approaches are able to circumvent many of the uncertainty issues associated with model-based approaches, but they usually require significant amounts of high-quality data, reflecting many different leak scenarios, to perform well. Our approach obviates this requirement by relying only on leakless data during model training. We examine the efficacy of this approach on 19 realistic leak experiments conducted in the field. Based on these evaluations, we were able to achieve average search costs as low as 2.2 kilometers, for a total network length of 215 kilometers.
Acknowledgements
We thank T. Van Daele, P. J. Haest, J. Debaenst, and Hydroscan for research assistance, and we thank De Watergroep for sharing data.
Disclosure statement
No potential conflict of interest was reported by the authors.
Notes
2. We multiply each pressure reading with the appropriate conversion factor (10.1974 ) and then add the corresponding sensor’s elevation.