Abstract: Giving full play to the regulatory characteristics of the controllable resource group can greatly improve the dynamic regulation capacity of the regional power grid. Therefore, a collaborative optimal scheduling method for controllable resource groups is proposed, and multi-agent deep reinforcement learning technology is used to solve multi-group complex collaboration problems. Firstly, the regional power grid optimization and dispatching problem considering multiple controllable resource groups is modeled, and the power grid optimization goals and system safety operation constraints are set. Secondly, the basic principle of multi-agent deep deterministic strategy gradient algorithm is expounded. Then, the policy gradient update algorithm is used to seek the optimal scheduling strategy of controllable resource group collaboration, and the corresponding evaluation indicators are defined to test the offline training effect and online application effect of the agent respectively. Finally, based on the improved IEEE test system, the effectiveness of the proposed method is verified.