Enhancing Water Distribution Network Leak Detection with Deep Reinforcement Learning: A Hybrid AI-Hydraulic Approach

<p dir="ltr">Water distribution networks (WDNs) are vital for delivering potable water yet leaks result in significant water loss and financial burdens, affecting service efficiency and sustainability. Traditional detection methods, including manual inspections, often struggle with delays and false alarms owing to water network complexity. While AI and machine learning (ML) offer potential, they typically require extensive labelled datasets, limiting their adaptability. Conversely, Deep Reinforcement Learning (DRL) leverages trial-and-error learning and real-time feedback, enabling adaptive decision-making in complex environments. DRL has proven effective in robotics and control systems, and in the water sector, it has been applied to pump optimisation and pressure control. However, direct DRL applications for comprehensive leak detection in WDNs remain underexplored. Therefore, this research introduces an innovative leak detection approach that integrates DRL with EPANET hydraulic simulation. This adaptive system formulates leak detection as a Markov Decision Process (MDP) which needs to represent the state, action, rewards, etc. States represent pressure and flow data, actions determine which network locations to inspect, and rewards are based on detection speed, accuracy, and efficiency. A simple custom Python-based OpenAI Gym environment has been developed for future DRL agent training using Stable Baselines3 and EPANET integration. Preliminary model components demonstrate the technical feasibility of the proposed approach. Future work will train and evaluate for enhanced leak detection efficiency, reducing water loss, and operational costs.</p><p dir="ltr">This paper was presented at the 21st Computing and Control in the Water Industry Conference (CCWI 2025) at the University of Sheffield (1st - 3rd September 2025).</p>