AeroWF: A Geometric Spectral-Temporal Dual-Stream Learning Framework for Aerodrome Weather Forecasting

Accurate and reliable aerodrome weather forecasting is essential for aerodrome safety and efficiency. However, existing approaches often fail to effectively capture the multi-scale dynamics and structural heterogeneity inherent in airport meteorological data, which involves variable numbers of runways, entangled high-frequency sensor streams, and low-frequency environmental contexts. In this paper, we propose AeroWF, a novel geometrically aware spectral-temporal dual-stream learning framework designed to address these challenges. Our approach integrates: (1) a macroscopic context injection module, which explicitly grounds transient micro-scale sensor observations within slow-varying macro-scale meteorological reports (METAR), improving consistency between micro-scale observations and macro-scale environmental states; (2) a spectral-temporal dual-stream encoder that disentangles transient temporal patterns and global spectral periodicities via a transformer-based temporal branch and a complex-valued frequency-domain branch; (3) a structure-adaptive hierarchical aggregator that unifies variable-cardinality runway observations through bidirectional interaction, without relying on predefined topological assumptions; and (4) a geometrically consistent pre-training strategy that combines hybrid masked reconstruction with intrinsic metric alignment, encouraging the learned representations to preserve both local fidelity and the temporal-shape/spectral-distribution structure of weather evolution. Extensive experiments on a new curated real-world multi-airport dataset consisting of 248K samples from four major international hubs (ZBAA, ZBAD, ZSPD, ZSSS) demonstrate that AeroWF achieves state-of-the-art performance across forecasting, imputation, and classification tasks. Furthermore, the model exhibits superior few-shot transferability to unseen airports, establishing a robust foundation for next-generation intelligent aviation meteorology. The source code and datasets are available at: https://github.com/MKC-Lab/AeroWF.