In view of the micro-grid random optimization scheduling problem, this paper proposes an online optimization algorithm of micro-grid based on deep reinforcement learning. The deep neural network is used to approximate the state-action value function, and the action of the battery is discretized as the output of the neural network. And then, the nonlinear programming is used to solve the remaining decision variables and calculate the immediate return, and obtain the optimal strategy through the Q-learning algorithm. In order to make the neural network adapt to the randomness of wind, photovoltaic and load power, according to the wind, photovoltaic and load power prediction curves and their prediction errors, Monte Carlo sampling is used to generate multiple sets of training curves to train the neural network. After the training is completed, the weights are saved. According to the real-time input status of the micro-grid, the neural network can output the actions of the battery in real time so as to realize the online optimal dispatching of the micro-grid. Compared with day-ahead optimization results under different fluctuations of wind power, photovoltaic and load power, the effectiveness and superiority of this algorithm in online optimization of micro-grid are verified.