Global and Local Interattribute Relationships-Based Graph Convolutional Network for Flight Trajectory Prediction

The rapid development of the aviation industry urgently requires efficient airspace traffic management, in which flight trajectory prediction is a core component. Existing trajectory prediction methods mainly capture the relationships between trajectory points. However, there are complex and implicit relationships between the attributes of trajectory points. Furthermore, there is room for improvement when applying previous multivariate time series prediction methods on flight track forecasting, since both local and global attribute features and interattribute relationships have an important impact on the flight track features. In this article, we propose a global and local interattribute relationships-based graph convolutional network (GLAR-GCN) to solve the flight trajectory prediction problem. First, we fuse the local embedded attribute feature and global attribute pattern to obtain the accumulated local embedded features at each point. Meanwhile, we synthesize the local and global attribute correlations to obtain the augmented attribute correlations at each point. Second, we construct an attribute graph at each point, where the nodes are the accumulated local embedded features and the edges are the augmented attribute correlations. Third, we extract the integrated attribute features from each attribute graph with a graph convolutional network. Finally, we perform trajectory prediction with the integrated attribute features as the input of a long short-term memory (LSTM) network. We evaluate the proposed model GLAR-GCN on real short-haul, medium-haul, and long-haul flight datasets for single-step and multistep prediction. Experimental results demonstrate that GLAR-GCN effectively improves the prediction performance on all datasets compared with the classical and the state-of-the-art methods.