MAST-LLM: A large language model-driven multimodal adaptive spatio-temporal framework for air traffic complexity prediction

Abstract

Airspace complexity prediction remains challenging due to intertwined spatial, temporal, and modeling factors. Spatially, airspace sectors exhibit heterogeneous dependencies from both geographical adjacency and dynamic traffic flows. Temporally, complexity evolves through multi-scale dynamics with long-range dependencies. Modeling-wise, existing spatio-temporal frameworks rely on static graphs and shallow temporal alignment, limiting their capacity for joint spatio-temporal reasoning, particularly under high-complexity and long-horizon scenarios. To overcome these challenges, this study proposes MAST-LLM, a Multimodal Adaptive Spatio-Temporal framework enhanced by Large Language Models (LLMs). First, a Mixture-of-Experts (MoE)-based temporal alignment phase is introduced to capture heterogeneous temporal patterns and align domain-specific air traffic sequences with pre-trained LLM representations. Second, a full spatio-temporal fine-tuning phase is developed, integrating adaptive multimodal spatial representations, multi-scale temporal features, and a Bidirectional Spatio-Temporal Attention (BSTA) mechanism for joint spatial–temporal interaction modeling. Finally, by leveraging the global reasoning and long-range dependency modeling abilities of LLMs, MAST-LLM establishes a unified representation that captures both local dynamics and global contextual dependencies. Extensive experiments demonstrate that MAST-LLM achieves superior performance at short- and medium-term forecasting horizons and remains competitive at longer horizons, with improvements of up to 9.5 percentage points in accuracy and 12.9 percentage points in macro F1-score. Furthermore, comprehensive ablation, factor importance, sensitivity and visualization analyses verify the effectiveness and robustness of each core component.