With the increasing computing demand and the diversification of resource types, traditional network architectures can no longer meet the complexity and dynamics of modern computing tasks. Computing Power Network, as a new network architecture for managing distributed and heterogeneous computing resources, can achieve computing power diversification, computing network integration, and global AI, but computing power networks need to manage heterogeneous computing power resources to meet the access needs of massive tasks. This paper proposes a three-layer network architecture for intelligent scheduling of computing power network resources, and designs a CPN controller to manage and schedule resources within a single domain to reduce task latency and meet massive access needs. A resource scheduling algorithm based on deep reinforcement learning for latency and cost is studied. The algorithm is suitable for multi-task scenarios. When making decisions for tasks, it can allocate the best computing power service node for current tasks. Experiments show that the cost is about 31% lower than that of the random algorithm, and about 20% lower than that of the greedy algorithm. In terms of latency, it is about 23% lower than that of the random algorithm and about 17% lower than that of the greedy algorithm.