Author(s): Ivana Lucin; Sinisa Druzeta; Boze Lucin; Zoran Carija

Linked Author(s):

Keywords: Leak localization; Water distribution network; EPANET; Random forest; Optimization

Abstract: Precise leak localization for urban WDN is needed to reduce pipe repair costs. The problem is that studies mostly use simplification where only network nodes can be identified as leak locations, i. e. precision of leak localization is predefined with WDN discretization. In this research, a methodology that couples the machine learning (ML) method and optimization technique with additional pipe segmentation is presented. ML is used to narrow down the potential leak location and an optimization method is used for more precise leak location identification. EPANET is used to generate a large number of synthetic leak scenarios. This data is used for the training of random forest classifier which is used to identify the most probable leak locations. Additional refinement is created around the most probable leak locations to increase the precision of leak localization. These nodes are then extracted and used for the optimization technique. To tackle the multimodality of the considered problem, independent optimizations are conducted for each potential leak location. In this way, a list of the most probable solutions is obtained and the complexity of the optimization problem is reduced with leak size as the only optimization variable. Results showed that ML models can narrow down the leak location, however, accuracy decreases for larger WDN and when demand uncertainties are considered. Additional pipe segmentation provides an opportunity for more precise leak localization, however, multimodality of the considered problem is still present for optimization methods and is even more prominent when demand uncertainties are considered.

DOI: https://doi.org/10.3850/978-90-833476-1-5_iahr40wc-p1010-cd

Year: 2023