关键词: 风力发电, 入流风场预测, 多模态混合神经网络, 时空注意力机制, 物理嵌入约束

Abstract:

To address the demand for high-precision inflow wind field prediction in large-scale wind turbines, traditional CFD methods suffer from high computational costs and poor real-time applicability. This paper proposed a multimodal hybrid deep learning-based wind field prediction method. The proposed method took turbine operating parameters and far-range wind field images as inputs and generated short-range wind field images as outputs. By employing a U-Net-Transformer-GAN hybrid architecture, the model achieved multi-scale feature extraction, temporal dependency modeling, and high-resolution wind field image generation. The vorticity transport equation and Kármán-Howarth turbulence statistics were incorporated as weak constraints to enhance physical consistency, while a temporal delay-difference mechanism was introduced to mitigate input and output asynchrony. Experimental results demonstrate that the proposed method outperforms comparative models, with PSNR of 32.51 dB, SSIM of 0.894, and LPIPS of 0.025. It accurately reproduces the Kolmogorov-5/3 power law in the turbulent energy spectrum, achieving a roughly 20-fold increase in predictive efficiency over traditional large eddy simulation.

Key words: wind power generation, prediction of inflow wind field, multimodal hybrid neural network, spatiotemporal attention, physical embedding constraint