A Deep Learning Method for Predicting Terminal Air Traffic Congestion Levels Based on Multimode Spatiotemporal Propagation Networks

Air traffic congestion and the resulting flight delays have significantly impacted the travel experience of passengers and the development of aviation. Forecasting air traffic congestion is of great importance to intelligent air traffic management. In this study, we propose a novel method named the Three-mode Spatio-Temporal Dynamic Graph Convolutional Network (TSTD-GCN), which employs Graph Convolutional Networks (GCN), Temporal Convolutional Network (TCN), and Transformer to model the spatiotemporal propagation patterns of air traffic congestion in terminal areas. Specifically, we develop a construction method for three-mode time-evolving networks, namely the adjacency network, traffic flow network, and spatiotemporal correlation network. Our approach incorporates serial graph convolution operations across three-mode networks to progressively expand the spatial scope of feature learning. In addition, TSTD-GCN employs TCN’s proficiency in extracting local dependencies and the Transformer’s ability to capture global dependencies, thus improving the precision of prediction. Finally, we validate the proposed model using real-world trajectory data from the Beijing terminal, China. The experimental results indicate that the prediction accuracy reaches 0.92 for a time scale of 10 min. Additionally, prediction accuracy decreases with an increase in time scale or an increase in trajectory complexity.