In order to reduce investment cost, the wind farm development trend includes mainly the two points: 1) rotor diameter of wind turbine will be larger; 2) compactness degree of offshore wind turbine will be higher. Therefore, the wake interference effect will be more serious, and predictive controller has become one of the main study for reducing wake effect. However, the online calculation time is usually too long, which is unfavorable to the real-time power control. To reduce the online calculation time, an offline predictive controller based on convolutional neural network-general regression network(CNN-GRNN) is proposed. Firstly, a dynamic wake model considering yaw angle and axial thrust coefficient is established, and a power prediction model is obtained. The particle swarm optimization(PSO) is used to solve the power optimization and get the optimal values of control variable under various wind directions and wind speeds. The mapping between system states and optimal values is established through CNN-GRNN. The implementation of receding horizon is replaced by CNN-GRNN, and the offline predictive controller for wind farm is constructed. The results show that this strategy can effectively increase the total wind farm power with less online calculation and high control accuracy.