With the increased awareness of environmental protection and the establishment of the spot market of electric power, renewable energy plays a more and more important role in energy structure adjustment. But the intermittence and instability of renewable energy often result in energy waste. Micro-grid energy storage can effectively improve the utilization rate, economy and reliability of renewable energy. To ensure the economical operation of the micro-grid system and improve the load curve, a two-stage optimal dispatch model was established based on non-dominated sorting genetic algorithm (NSGA-II) and deep Q-value reinforcement learning algorithm (DQN). In the first stage, to get the appropriate battery capacity, the minimum comprehensive operation cost and the minimum capacity fading of battery in the cycle life were selected as the optimization objectives. The optimal solution was calculated by using the improved non-dominated sorting genetic algorithm (NSGA-II). In the second stage, the DQN Algorithm was used to propose a scheduling strategy. Compared with the GA method, the DQN method was able to reduce the operating cost from 0.5829 % to 1.2294 %. The results showed that the method was feasible and effective.