Due to the randomness of traffic in the computing power network and the diversity of paths, congestion is inevitable. When network congestion occurs, it will cause increased queuing delay, low network utilization, resulting in deterioration of application performance. Nowadays, there are many congestion control methods. By continuously adjusting the transmission rate of the end side, the capacity of the incoming network is finally approached as close as possible to the carrying capacity of the network to solve the congestion problem in the network. However, the current mainstream method has the disadvantages of insufficient control precision and is incapable of self-adaptive adjustment. Therefore, it is necessary to propose a new type of intelligent congestion control algorithm based on deep reinforcement learning to achieve more active control capabilities and more precise rate regulation, and to achieve the goals of full bandwidth, low latency, fast convergence, and excellent fairness. Thereby effectively improving the transmission efficiency of the computing power network and ensuring the efficient completion of large-scale distributed computing tasks.