Multi-objective deep reinforcement learning for emergency scheduling in a water distribution network

In recent years, water contamination incidents have happened frequently, causing serious losses and impacts on society. Therefore, how to quickly respond to emergency pollution incidents is a widespread concern for academic and industry scientists. In this paper, aiming to deal with the uncertain environment and randomness of water demand, we present a method based on a deep reinforcement learning emergency scheduling algorithm combined with Lamarckian local search. This method can effectively schedule water valves and fire hydrants to isolate contaminated water and reduce the residual concentration of contaminants in water distribution networks (WDNs). Furthermore, two optimization objectives are formulated, and then multi-objective optimization and deep reinforcement learning are combined to solve this problem. A real-world WDN is employed and simulation results show that our proposed algorithm can effectively isolate contamination and reduce the risk exclosure to customers.