Deep Reinforcement Learning for Adaptive Control of Urban Drainage and Green Infrastructure Under Climate Change

Abstract

Climate-driven intensification of short-duration rainfall is eroding the design margins of urban drainage networks, and static operating rules for storage basins, gates, pumps and green infrastructure (GI) can no longer cope with non-stationary storm regimes. This study develops a deep reinforcement learning (DRL) framework for the coordinated, real-time control of conventional drainage assets and distributed GI under projected climate change. The control problem is formulated as a Markov decision process in which an agent observes water levels, rainfall nowcasts and GI storage states and issues continuous gate, pump and orifice commands. A proximal policy optimization (PPO) agent with a multi-objective reward that penalizes flooding, outfall exceedance and control effort was trained on randomized synthetic storms in a hydraulic environment representing a 4.2 km² urban catchment, and was evaluated on held-out events against passive, rule-based, model predictive control (MPC) and deep Q-network (DQN) alternatives under three scenarios of rainfall intensification (0%, +15% and +30%). The proposed controller reduced total flood volume by 42.6–45.7% relative to passive operation and by 11.4–14.3% relative to MPC, while keeping control actions smooth and decision latency below 5 ms. Performance degraded gracefully as rainfall intensified, and the learned policy exploited GI retention ahead of peak inflow. The results indicate that DRL is a credible option for climate-adaptive operation of hybrid grey–green systems, provided that safety constraints, uncertainty handling and field validation are addressed.