In the docking stage of autonomous aerial refueling, the drogue will have a relatively high frequency swing motion under the combined influence of atmospheric turbulence, wind gust and tanker wake vortex, which poses a great challenge for the receiver to accurately track the drogue. At present, the methods to realize docking are generally to track the drogue directly or to plan the trajectory first and then track the control. However, there are some deficiencies in the research of the two methods. The former is not effective for high frequency sinusoidal maneuvering targets, and the latter does not consider the prediction of drogue motion in trajectory planning. This paper proposes a compound trajectory planning framework based on Minimum-snap method, which takes the predicted drogue trajectory under the LSTM network model as the target point. Firstly, the drogue is modeled as an ideal system with constant length and connected in series through frictionless spherical joints step by step, and a three-degree-of-freedom model is constructed. Secondly, the trajectory of drogue in random wind field is used as the training data of convolutional long-term memory neural network model for training. Finally, the real-time trajectory planning is carried out by Minimum-snap method according to the position of drogue under the prediction model. The simulation results indicate that the trajectory planned by the proposed method is smooth and feasible, satisfies the flight constraints, and has better energy loss. After considering the prediction of drogue motion, the trajectory accuracy planned by this method is higher, which can greatly improve the success rate of autonomous docking.