Abstract: With the rapid development of power digitalization technology, digital converter stations, as a key component of the power system, are facing increasing computational pressure. A cloud-edge-local three-tier computing offload framework is proposed for the computing service requirements of fusion terminals in digital converter stations. Considering the mobility of terminal devices, this paper aims to minimize the total system delay, constructs a Markov decision model according to business requirements, and adopts a dynamic task offloading optimization method based on deep deterministic policy gradient algorithm to achieve the optimal scheduling of overall computing resources and network resources at the cloud-edge-local layer. By building a deep reinforcement learning environment and simulating experiments, the results show that the proposed scheme reduces the total system delay by 29.6% compared with the edge-local two-layer offloading scheme.