An aircraft has an obvious state division in a complete flight mission, the data of the trajectory points also have specific characteristics in each flight state, so the trajectory data with the similar characteristics can improve the performance of neural networks. Therefore, a combined online-learning model with K-means clustering and gated recurrent unit (GRU) neural networks for trajectory prediction is proposed in this paper. In the new model, the K-means clustering algorithm is used to adaptively cluster the trajectory points of the aircraft, and the trajectory points with higher similarity are grouped into a same cluster. Then, the online-learning prediction model based on a GRU neural networks is used to learn from the trajectory points of each cluster separately. Finally, the performance superiority of the new model proposed in this paper is tested and verified with the fused flight data of secondary radar and automatic dependent surveillance-broadcast (ADS-B).