With the increase in the number of electric vehicles(EVs)

developing rational strategies for charging and discharging scheduling of EV clusters in a distribution network environment with integrated photovoltaic(PV) systems and multiple distributed charging stations has become an important challenge in grid management. In this paper

a multi-agent deep reinforcement learning method based on dynamic grouping for EV cluster charging scheduling is proposed to address this problem. Firstly

physical modeling is carried out in the distribution network environment containing photovoltaic power generation

and the reward function is constructed by considering the charging cost

grid security and other factors. Then

a multi-agent deep reinforcement learning algorithm for dynamic grouping value conversion is proposed to dynamically group EV clusters based on real-time status and charging demand

and the commanders in the group uniformly summarize the information and make decisions in order to reduce the decision coupling and decrease the communication burden. Subsequently

the decision quality of each sub-cluster is evaluated by centralized Critic

and the global Q-value fusion framework based on consistency loss integrates the local value estimation under the inter-group rivalry and competition relationship to achieve efficient and stable global charging scheduling strategy. Finally

the day-ahead scheduling test verifies that the proposed method can reduce the grid load fluctuation and contribute to the stable operation of the grid while satisfying the individual charging demand of EVs.