Deep Reinforcement Learning for Pressure Optimization in Water Distribution Networks with Multiple Pumping Stations

Pressure control is essential in water distribution networks (WDNs) with multiple pressure sources for ensuring operational efficiency, maintaining infrastructure integrity, and providing reliable service. Excessive pressure leads to unnecessary energy consumption and increases the risk of pipe failure and water loss, whereas minimum pressure levels must be maintained in district metered areas (DMAs) to ensure service standards. Technological advancements have enabled the application of reinforcement learning (RL) to advance pressure management. However, previous work focused on improving pump efficiency and have overlooked pressure control using RL with continuous action spaces. We conducted a case study on a WDN in Abu Dhabi, applying deep RL (DRL) techniques for pressure control optimization. A custom training environment was developed using the Gym framework, integrating a calibrated hydraulic model of the network. The DRL agent was trained to dynamically adjust the pumping station pressure setpoints to control the maximum pressure across the network while maintaining the required minimum pressure within all DMAs. The trained DRL agent demonstrated effective control of the network pressure, resulting in measurable reductions in the energy consumption and risk of pipe damage while maintaining service reliability. The methodology and tools applied in this study are fully adaptable and can be readily implemented for pressure optimization in water networks of varying sizes and configurations. By leveraging open-source resources, this approach offers a practical, transferable, and cost-effective solution for enhancing the operational efficiency, resilience, and sustainability of WDNs.