Short-term forecasting of 3D wind field based on deep learning and water vapor

Abstract

Short-term wind field forecasting is critical for various applications, including renewable energy and aviation safety, but traditional numerical models often face challenges in computational efficiency and timeliness. This study proposes a Deep Learning Model Framework for Wind Field Prediction using Multi-level Water Vapor (DLMF4-WFPMWV) to improve short-term wind field prediction by utilizing multi-level atmospheric water vapor, surface precipitation and evaporation data. We investigated multiple deep learning architectures, including Artificial Neural Networks (ANN), UNet, Transformer, and a CNN-Transformer fusion model. Our findings indicate that an optimized CNN-Transformer fusion model, enhanced with prior physical knowledge, precipitation, evaporation data, and temporal information from preceding time steps, achieves superior performance in predicting U, V, and W wind components. This is evidenced by a resulting RMSPE below 8% and a correlation coefficient exceeding 0.61. This fusion model demonstrated robust predictive capabilities across different atmospheric levels and weather conditions (clear vs. rainy days) and maintained accuracy even with significantly fewer spatial input points. Furthermore, the framework is highly effective for its primary goal of short-term prediction (1-6 hours). Furthermore, when benchmarked against traditional numerical weather prediction, it also demonstrated competitive, and in some cases superior, performance at longer lead times compared to the ECMWF S2S model, particularly for the vertical wind component, while offering substantial advantages in inference speed and reduced computational resource requirements. This research highlights the feasibility and effectiveness of deep learning for accurate and efficient short-term wind field prediction from water vapor data, providing a valuable foundation for advanced meteorological forecasting.