AGC is the main means to maintain the active power balance of the power system and ensure the system frequency quality. In this paper, deep learning techniques are used to optimize AGC active coordinated control. The linearized model simplifies the dynamic system at the operating point, and in the AGC coordinated control system, the model parameters are estimated using the recursive least squares method. The discrete reinforcement learning DQN algorithm is used to construct the AGC optimization model, and the continuous reinforcement learning PPO algorithm is used to propose the optimization strategy for the AGC active coordinated control, which solves the problems of discretization error as well as dimensional catastrophe. The AGC active coordinated control test is being carried out at power station A in the northwest region of China as the study site. After the optimization, the average regulation rate, average response time, and average regulation accuracy of Generation Station A are improved to 0.4995, 0.7835, and 0.8082, respectively, and the comprehensive FM performance index is improved by 22.08% compared with that before the optimization. The economic benefits indexes such as FM capacity and FM mileage compensation revenue fee are improved, and the average response time qualification rate and regulation rate in October-December after optimization are improved by 5.87% and 6.38%, respectively, compared with the pre-optimization period.