Abstract: Large-scale wind power connected to the gird brings new uncertainties which affects the frequency response characteristics of the system. From the data-driven perspective, an online reinforcement learning method based on adaptive dynamic model for load frequency control is proposed. A low rank autoencoder feature extraction network is established to discover hidden features from the measured low-dimensional data. Based on the feature network, a sparse identification learning model for nonlinear dynamic system is established to detect the potential physical state of the dynamic model of the system and improve the online learning efficiency. Combined with model predictive control, real-time decision control is implemented. The proposed method can effectively solve the problem that traditional model predictive control depends on the accuracy of the system global model, enhance the adaptability of controller to the system dynamic model and effectively track the random fluctuation of wind power. Finally, the validity of the proposed control method is illustrated by load frequency control model integrated with type Ⅵwind turbine.