Dimensions of superiority: How deep reinforcement learning excels in urban drainage system real-time control

Reducing combined sewer overflows and flooding is crucial for the efficient operation of urban drainage systems. Traditional real-time control (RTC) methods often fall short in efficiency and performance, which prompts the exploration of innovative approaches. Deep reinforcement learning (DRL) has recently emerged as a promising technique to enhance RTC performance. This study evaluates the effectiveness of RTC using a multi-agent-based DRL approach. We developed a comprehensive evaluation framework incorporating multiple quantitative indicators, including control objectives, decision time, robustness, and adaptability. To validate our framework, we conducted a case study on an urban drainage system in Suzhou, China, analyzing 31 historical rainfall events. Our findings reveal that DRL can reduce flooding and overflow risks by 15.1 % to 43.5 % on average compared to conventional RTC methods. Additionally, DRL demonstrates superior efficiency, robustness, and adaptability. This study not only highlights the potential of DRL in urban drainage management but also provides insights into its broader application in enhancing the resilience of urban infrastructure systems.