A Novel 3D Physics-Integrated Swin-Transformer Model for Precise High-Resolution Urban Boundary Layer Wind Speed Estimation

Abstract

Accurately estimating low-altitude wind speed (WS) is a critical and challenging task, with significant implications for urban meteorology and pollution dispersion modeling. This study developed a novel three-dimensional Physics-Integrated Swin-Transformer (3D-PST) deep learning model to estimate high-resolution WS in the urban boundary layer. Comprehensive evaluations demonstrate that the 3D-PST model outperforms existing methods across all key metrics, achieving state-of-the-art results. Notably, the Root Mean Square Error reaches 1.10 m/s with a spatial resolution of 130 layers, while the correlation coefficient (R) is as high as 0.93, indicating a strong predictive capability. Furthermore, ablation studies reveal that key components like dynamic physical variables improve accuracy by 10%, and the convolution patch merging yields 12% improvement by more effectively capturing multi-scale turbulence features. These results highlight the adaptability and robustness of the 3D-PST model, showing its potential as a powerful tool for urban meteorological monitoring and supporting low-altitude economic development.