AGST-WindFormer: A Forecasting Model for Multi-Station Near-Surface Winds

Abstract

Accurate forecasts of near-surface winds are important for low-altitude aviation, unmanned aerial vehicle operations, and local weather services. Compared with single-station forecasting, multi-station forecasting can exploit complementary spatial information to improve wind-field representation. However, pronounced spatial heterogeneity and complex inter-station dependencies make this task inherently a spatially structured multivariate time-series forecasting problem. To address these issues, we propose AGST-WindFormer, an Adaptive-Graph Spatiotemporal Gated Wind Transformer for directly forecasting the horizontal wind components U and V. A shared Transformer captures station-wise temporal dependencies, while node embeddings generate an adaptive static graph shared across samples. A feature-wise gate then integrates the temporal and graph-aggregated representations, and a multi-output head predicts future U and V values. Experiments used hourly Meteostat observations from 854 stations during 2022–2025. The model achieved a mean wind-speed RMSE of 1.507 ± 0.009 m/s over the 10-h forecast horizon and a wind-direction MAE of 23.5 ± 0.1° for observations with wind speed ≥ 2 m/s. Relative to the strongest baseline for each metric, errors decreased by 1.1–3.2%. Ablations identified cross-station aggregation and nonuniform learned edges as the principal contributors. These results indicate that AGST-WindFormer can effectively exploit the spatiotemporal information contained in multi-station near-surface wind observations and improve forecasting accuracy across multiple consecutive future time steps.