A Spatial–Temporal Graph Large Language Model Based Air Traffic Flow Prediction in Multi-Airport System

The multi-airport Air Traffic Flow Prediction (ATFP) is a vital area for achieving efficient air traffic operations in regions with multiple airports. ATFP faces challenges while dealing with complex spatiotemporal dependencies and irregular airspace structures. Existing methods struggle from problems like over-smoothing, limiting global pattern recognition while trying to address spatial and temporal patterns. Recently, Large Language Models (LLMs) have shown promise in various areas. Their excellent performance on long term correlation modeling and large-scale cross-domain pre-training based on Transformer suits well in ATFP. However, the extraction of spatial correlations in existing LLM-based methods is insufficient to represent complete topological structures of terminal maneuvering area associated with multi-airport system. To address the challenges mentioned above, we propose Spatial–Temporal Graph Large Language Model (STG-LLM). It extracts spatial–temporal features through a specially designed feature extractor. The extracted features are then fused in an adaptive fusion layer and are subsequently enhanced by a Graph Attention Network (GAT) before being input into a partially frozen LLM for final flow prediction. Experiments on real-world air traffic dataset of Guangzhou-Shenzhen-Zhuhai Airport Group have proved that STG-LLM is able to accurately predict air traffic flow and exceed existing models on cases of airports, fix points and air routes.