Author(s): Wenxuan Jia; Zhenju Chuang; Nan Gao

Linked Author(s):

Keywords: Wind turbine blade icing; Cold region; Data augmentation; Conditional instance normalization

Abstract: Wind turbine blade icing in cold climates severely threatens the operational safety and efficiency of wind farms. Traditional numerical simulation methods struggle to yield a substantial volume of high-quality ice shape data under extreme conditions due to their high computational cost, creating a major bottleneck for the development of ice shape prediction models. To address this, focusing on the FFA-W3-211 airfoil at the IEA 15MW wind turbine blade tip, this paper proposes a high-quality ice shape data augmentation method based on an improved Generative Adversarial Network (CIN-SAM-GAN). First, a Conditional Instance Normalization (CIN) module is integrated into the bottleneck and upsampling stages of a U-Net to achieve adaptive modulation of global feature representations in the channel dimension based on icing condition parameters. Subsequently, a Spatial Attention Module (SAM) is introduced at the decoder's end to effectively suppress redundant interference from background and non-salient features in the spatial dimension. Ablation studies demonstrate that the synergistic effect of CIN and SAM significantly improves the generated ice shape quality. Furthermore, comparative experiments prove that, compared to baseline models trained solely on the limited original dataset (OD), the prediction model trained on the CIN-SAM-GAN augmented dataset achieves a steady improvement in F1-score.

DOI:

Year: 2026