Multi-agent Deep Reinforcement Learning Based Multi-objective Charging Control for Electric Vehicle Charging Station

This paper proposes a multi-agent deep reinforcement learning (MADRL) based algorithm for charging control of multiple electric vehicles (EVs) in an electric vehicle charging station (EVCS) with dynamic operations. By taking the uncertainties of EV users’ behaviors and charging prices into consideration, a Markov game with multiple objectives is first established, where the individual objective of reducing EV charging costs and the collective objective of mitigating charging load fluctuations are both studied. Then, a multi-agent actor-double-critic (MAADC) algorithm with a unified advantage actor-critic framework is proposed, where two advantage functions are designed as the local and global critics, and the integrated actor is employed to balance the learning towards the individual and collective objectives. The simulation results obtained in a delicately built EVCS environment show the effectiveness of MAADC.