Abstract: As distributed energy storage devices, electric vehicles（EVs） have huge potential to curb power fluctuations.Considering the randomness of EV integration, the uncertainty of renewable energy generation and load, a real-time dispatch model with objectives of the minimum power fluctuations and minimum charging and discharging cost is established by using a model-free deep reinforcement learning method. In order to satisfy the user’s demand, a charging and discharging energy boundary model is used to characterize the charging and discharging behaviors of EVs. After day-ahead training and parameter preservation of the proposed model, the corresponding charging and discharging dispatch strategy is generated in real time for the intraday realtime states of the system operation at each moment. Finally, taking a microgrid as an example, it is verified that the proposed deep reinforcement learning based dispatch method can effectively reduce power fluctuations in the microgrid and reduce EV charging and discharging cost on the premise of satisfying users’ charging demands, while satisfying the demands of real-time dispatch without the need of iterative calculation for intraday dispatch.