An Experimental Study On Early Leak Localization In Drinking Water Networks using pressure measurements

Abstract

[EN] Leaks represent a major issue impacting the management and efficiency of Drinking Water Networks (DWN) in cities worldwide. According to the Development Bank of Latin America, by 2018 the losses in DWN range from 40 to 60% in the region. In Europe, the OECD reports a wider range with few losses in cities like Amsterdam (4%) up to 37% in Naples. With this context, some regional policies have emerged like the 2020 european drinking directive “Right2Water”, that aims to encourage major suppliers (more than 50000 users), to develop tools to measure and reduce leakages by 2025. Considering this situation, we introduce here a systematic approach for leak management that combines field data, hydraulic models (HM) and machine learning.Model-based and data driven methods have been of great interest for leak location methodologies in DWN. This research will design energy-efficient and cost-efficient leak localization hotspots in the DWN. The approach is intended for sectorized DWN, equipped with a SCADA system and where a calibrated hydraulic model (e.g. EPANET) is available. This latter serves to evaluate the sensitivity of the system to leaks and identify potential points for pressure measurements in order to optimise the number of installed sensors. Given a detected leak in the network, a multiclass classifier using pressure data is developed to reduce the inspected pipe length for the leak location. The leakage localization method is implemented combining multiple individual classifiers using ensemble learning methods and a reduced number of decision variables. The methodology is tested on a real case study from a Colombian site. The method faces challenges in (a) collecting correctly labelled real leak data, and (b) modelling and calibrating hydraulic models. Those challenges are being addressed. The outcome shows that the length of pipes inspected can be reduced by one third with high performance in accuracy with few sensors required (low capital expenditures) and low computational effort (low energy and low operational expenditures).