A Wind Speed Prediction Model Based on the Dual-Branch 3DUT-Net Architecture

To improve the accuracy and temporal coherence of wind speed forecasting under complex wind field conditions, this paper proposes a multi-module hybrid model named 3DUT-Net, which integrates a three-dimensional convolutional neural network (3D CNN), U-Net architecture, and Transformer attention mechanism. The model combines the advantages of local spatiotemporal feature extraction, multi-scale spatial representation, and global dependency modeling to establish an end-to-end framework for multi-step wind speed prediction. Experiments were conducted using lidar-based wind field imagery collected from Urumqi Diwopu International Airport. Results demonstrate that 3DUT-Net outperforms conventional models in terms of MAE, RMSE, SSIM, and PSNR metrics, achieving higher prediction accuracy and structural fidelity. This study offers a new technical solution for short-term wind forecasting in airport scenarios and lays a foundation for intelligent meteorological services.