Lifelines 2022

Water distribution networks (WDNs) are critical infrastructure that provides essential support for the community life. Its service can be interrupted by natural hazards such as earthquakes. The resilience of a WDN is related to its capability to resistant damages as well as its speed to recover the service function. This paper introduces the development of a machine learning (ML) model to ensure fast restoration of WDN subjected to earthquake damages, and therefore improves the system resilience. The reinforcement learning (RL) model is trained to optimize WDN restoration sequence considering consumer needs and evolving water demands. The results show that compared with the repair sequence by the commonly used method of static importance ranking, repair sequence by RL model substantially improved the recovery speed of WDN. Besides, under different damage scenarios, the repair sequence by the RL agent consistently achieved efficient recovery (and therefore enhanced resilience), demonstrating the robustness of the RL model. This study demonstrates the potentials of artificial intelligence (AI) techniques to support optimal decision in managing the WDN under emergency conditions.