Abstract: To address the challenge of balancing accurate simulation with efficient control in the real-time control of urban drainage systems, this study proposed an intelligent real-time control framework integrating a graph neural network (GNN) surrogate model with reinforcement learning (RL), aiming to enhance the effectiveness of overflow risk management. First, a GNN-based surrogate model, capable of efficiently simulating overflow in complex systems, was developed. Then, the surrogate model served as a virtual environment in which RL was used to achieve dynamic regulation of multiple actuator gates, with the objective of minimizing overflow. Using a drainage district in Yueyang City, China as a case study, the surrogate model achieved a Nash-Sutcliffe efficiency coefficient of 0.969 during the testing period, indicating high simulation accuracy. Under heavy rainfall conditions, the RL strategy based on the surrogate model reduced overflow volumes by up to 18.1%, with a single-step decision time of only 0.19 seconds. This was substantially faster than conventional methods, such as genetic algorithms, that required 16.2 minutes per step. Moreover, RL exhibited greater robustness under uncertainty scenarios. This study provides technical support for intelligent regulation of urban drainage systems.