Traditional yaw tracking control methods for large-scale wind turbine exhibit inherent lag, which would lead to loss of power generation. In this paper, a new imitation learning-based yaw control (ILYC) framework with better tracking performance is proposed. The framework consists of two modules, the demonstration generator based on a new ternary genetic algorithm (TGA) and imitation learning module based on the time-stamped feature. A discrete kinetics model with three actions is established for the convenience of obtaining optimal control sequences for given wind sequences. TGA is developed from the standard binary coding-based genetic algorithm to cater for this model and generate demonstrations offline with abundant wind samples from all over China. Considering the annual and daily periodicity of wind, we combine the wind with feature with the time stamp to utilize the latent periodic information. Experiments show that the proposed imitation learning framework effectively estimates the change of wind direction and makes the yaw system move in advance, and hence achieves an average increase in power generation revenue by 5% as well as an average decrease in yaw frequency by 30%.