Toward the full utilization of storage capacity: Urban drainage system operation integrating deep reinforcement learning and green infrastructure.

Green infrastructure (GI) and real-time control (RTC) are jointly employed to enhance flooding and combined sewer overflow (CSO) control in urban drainage systems (UDSs). However, it remains unclear whether RTC fully leverages the additional storage capacity created by GI to effectively mitigate flooding and CSOs, owing to the lack of effective theoretical guidance for the integrated use of GI and RTC. In this study, a new index, the GI cooperative index (GCI), is developed to quantify how effectively an integrated GI-RTC system converts GI-induced inflow reductions into reductions in flooding and CSO. Regression analysis of the relationship underlying GCI further provides an empirical characterization of the controller's baseline control capability and its capacity to exploit the potential operational margins created by GI. GCI is further integrated into the deep reinforcement learning (DRL) training process to explicitly encourage the agent to utilize the favorable hydraulic conditions created by GI, following a knowledge-fusion paradigm. Application to a model UDS located in eastern China demonstrates that training DRL agents in a GI-integrated environment substantially improves coordination between GI and the DRL controller, and incorporating GCI into the reward function yields further enhancements. In addition, a response-function-based interpretation offers a preliminary control-theoretic perspective on RTC in UDSs and provides a basis for further theoretical development.