Robust real-time urban drainage control under communication failures: an information-aware multi-policy control framework.

With climate change and urbanization, urban drainage systems face increasing pressure. By dynamically regulating storage and conveyance capacities in response to system states, real-time control offers a cost-effective approach to mitigate combined sewer overflow (CSO) risks without expanding physical infrastructure. This capability relies on online state information obtained from sensor measurements and transmitted through communication networks, making real-time control vulnerable to communication failures. Existing studies have mainly focused on evaluating the passive robustness of control policies under disturbances, while methods for active performance improvement under communication failures remain unexplored. Leveraging the flexibility of neural networks (NNs) in handling different input structures, this study proposes an information-aware multi-policy switching control framework. Specifically, a set of NN-based control policies is trained under different system state configurations, and the appropriate policy is selected according to information availability during communication failures. The proposed framework is evaluated using the Astlingen benchmark network, where scenarios with individual sensor disconnections and random communication failures are simulated. Results show that the proposed multi-policy switching framework outperforms single-policy control across all the simulated failure scenarios, with performance improvements becoming more pronounced as disconnection duration increases. Under random communication failure scenarios, overall CSO volume is reduced by approximately 3.7-9.6% relative to single-policy control. Further analysis indicates that these gains are mainly associated with reduced CSO volumes within low-CSO-rate intervals. This suggests that the framework can better exploit the system's regulation capability during the onset and recession stages of CSO events, whereas single-policy control is more constrained.