Abstract: In the context of multi regional interconnection in the main network, existing centralized optimization methods face difficulties in data processing, delayed dynamic response, and cannot effectively solve the problem of voltage exceeding the limit of the main network. Therefore, research on dynamic coordination optimization of multi regional power flow in the main network based on deep deterministic policy gradient is carried out. By constructing a decoupled multi region optimal power flow model and a distributed coordination framework, a foundation is laid for intelligent agent learning. By designing a deep reinforcement learning environment with system state, coordinated actions, and safety economic rewards as its core, high-dimensional dynamic feature extraction and intelligent decision-making can be achieved. By introducing the semi positive definite relaxation technique to handle non convex flow constraints within the region, the reliability and efficiency of solving subproblems are ensured. Through comparative experiments, it has been proven that the proposed method is significantly superior to existing methods in ensuring system security, improving economy, accelerating computing speed, and optimizing resource utilization, providing an effective solution for achieving real-time intelligent coordinated operation of cross regional power grids.