A Lightweight Terrain‐Constraint Model for Wind Spatial Downscaling

• Published in: Journal of geophysical research. Machine learning and computation Vol. 2; no. 2
• Main Authors: Guo, Anboyu, Liu, Qiyang, Wang, Jiuke, Gao, Yuanyong, Wei, Lixin, Song, Xiaojiang, Zhang, Chi, Liu, Kai, Ma, Shuping, Jiang, Jiu
• Format: Journal Article
• Language: English
• Published: Wiley 01.06.2025
• Subjects: complex terrain; lightweight model; terrain‐constraint; wind field downscaling
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000147

High‐resolution wind fields has always been the goal of refined meteorological forecasting. Using advanced deep learning algorithms for wind downscaling is an effective approach to achieve this goal. However, the lack of physical process understanding in deep learning algorithms results in the inability to accurately reconstruct fine‐scale structures after downscaling. In this study, we propose a Terrain‐Constraint Wind Downscaling Model (TCWDM), a lightweight deep learning model consisting of a downscaling module and a terrain‐constraint module. By combining low‐resolution wind with high‐resolution terrain data, the model achieves a tenfold downscaling of spatial wind fields and reconstructs the detailed structure of the wind. Due to the incorporation of an attention mechanism, multi‐feature inputs, and the terrain‐constraint module, TCWDM demonstrates superior downscaling performance. Compared to traditional interpolation methods and other deep learning models, the mean absolute error is reduced by up to 49%. The terrain‐constraint module, in particular, contributes most significantly to the model's performance, especially in complex terrains, where it enables greater optimization of downscaling results. Furthermore, due to the lightweight model structure and a specific fine‐tuning strategy, TCWDM can deliver significantly better downscaling results at a lower cost across different regions, offering potential for broader applications. Plain Language Summary Accurate wind predictions are essential for improving weather forecasts, especially in regions with complex terrains. Traditional methods struggle to capture fine details in wind fields at high resolutions. In this study, we introduce a new machine learning model that improves wind fields predictions by using both low‐resolution wind data and high‐resolution terrain information. Our model, called the Terrain‐Constraint Wind Downscaling Model (TCWDM), enhances the resolution of wind by 10 times and reconstructs detailed wind patterns. The model performs better than previous methods, reducing errors by up to 49%. It also works efficiently in different areas, especially in regions with complex landscapes, by using a lightweight structure and an easy‐to‐apply fine‐tuning process. This approach could lead to more accurate and cost‐effective wind fields predictions in the future. Key Points A lightweight deep learning model is proposed that uses high‐resolution terrain data to enhance the resolution of wind by a factor of 10 The terrain‐constraint can significantly optimize the model's downscaling performance, reconstructing the detailed structure of the wind By using a specific fine‐tuning strategy, the model can achieve performance advantages in different regions at a lower cost