Author(s): David Modesto; Fernando Cucchietti; Victor Saenger; Bernat Joseph-Duran; Jordi Meseguer; Rafael Gimenez; Gabriela Cembrano; Daniel Cardelus; Catalina Balseiro

Linked Author(s): Bernat Joseph-Duran, Jordi Meseguer

Keywords: Drinking Water Networks; Pressure Control; Water Loss; Reinforcement Learning

Abstract: Pressure control in water supply is one of the most important challenges in real-time operation of drinking water networks (DWNs). Excessive pressure at network nodes (i.e., during lower demand periods) leads to an increase of water losses and may also worsen the condition of pipes increasing the frequency of pipe bursts and new leaks. Real-time determination of appropriate set-points for pressure reducing valves (PRVs) allows to reduce DWN pressure profile and consequently, water losses and other negative impacts of high pressures while ensuring a proper water supply service quality. In the present work a Reinforcement Learning (RL) approach is proposed for real-time pressure control of existing PRVs in the DWN. The goal of the RL agents is to keep pressures at all consumption nodes as low as possible within prescribed limits by means of controlling pressure set-points of existing PRVs. Both training and validation take place against an EPANET model of the considered DWN. The resulting controller has shown to provide an improvement with respect to the current scheduled approach in keeping the pressures at the consumption nodes as low as possible within the desired limits. Moreover, the controller design strategy has shown to be easily replicable to multiple DWNs. The proposed methodology has been validated in different Agbar water utilities: 8 DMAs operated by Aigües de Barcelona and one DWN operated by Hidralia.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221924

Year: 2022