Abstract: In order to improve the dynamic response capability and control accuracy of thermal power units in AGC frequency regulation process, reinforcement learning method is adopted to construct optimized control strategy, design reward function, define state variables and action space, and analyze the training convergence and performance indicators in combination with simulation platform. This method can effectively reduce frequency steady-state error and adjustment time, suppress power tracking overshoot, reduce energy consumption of the actuator, and provide an implementable technical path for AGC frequency regulation optimization of thermal power units.