In response to the high energy consumption of electricity, it has become an important challenge to develop a reasonable grid dispatching strategy in power dispatch to be able to achieve energy saving and emission reduction. Some studies have applied AI technologies to grid dispatching problems, but existing studies do not consider the increasingly complex grid environment that makes it difficult to perform predictive power dispatching. Reinforcement learning has the feature of imitating human learning by interacting with the environment, which is ideal for solving complex grid intelligent dispatching problems. In this paper, we propose a new online grid dispatching method based on deep reinforcement learning, which aims to maximize the proportion of renewable energy supply and grid stability. To efficiently utilize experience in the reinforcement learning training process, we design a grid dispatching strategy with Soft Actor-Critic (SAC) algorithm based on distributed training, which can learn the optimal grid regulation policy through a multi-process approach. The simulation results show that the reward value variance is improved by more than 20% than other methods. It verifies the validity of the method in this paper.