Ensuring Reliability of Large-scale Charging Piles: An Efficient DRL-based Scheme

The rapid development of electric vehicles has triggered a surge in the deployment of charging stations, accompanied by an increasing risk of safety incidents. It is necessary to deploy preventive maintenance in order to maintain the reliability and safety of charging pile operations. However, developing effective preventive maintenance strategies face various challenges like limited maintenance budgets, large-scale deployment of charging piles, and the dynamic health status of charging piles. Note that, for a charging pile, the extent of damage varies, and the degree of health recovery after repair also differs. This paper considers a preventive maintenance scheduling system for large-scale charging piles, in which there exist the budget constraint and the dynamics of the charging pile status. We aim to maximize the operational revenue of the charging piles while minimizing maintenance costs. To this end, we first set up the preventive maintenance scheduling problem as a partially observable Markov decision process (POMDP). Then, we apply deep reinforcement learning techniques to design a scheme in order to handle the POMDP. This scheme addresses the issue caused by the high-dimensional action space of the POMDP, by proposing an action space splitting method to reduce the complexity of the action space. We conduct experiments to verify that our proposed scheme performs better in both maximizing the operational revenue and minimizing the maintenance costs than other baselines.