A low-computational physics-guided deep learning model for wind farm flow control under time-varying wind conditions

Wind farm flow control (WFFC) under time-varying wind conditions (TVWC) is essential for enhancing power production. Existing studies typically achieve real-time control through the look-up table (LUT) method. However, it suffers from high computational costs due to individual optimization of diverse TVWC and exhibits limited generalization under unknown TVWC. Consequently, this study proposes a low-computational, physics-guided deep learning model for real-time control under TVWC, which is trained on data precomputed by the low-computational method and demonstrates strong generalization. Initially, a comparative analysis of WFFC under time-averaged wind conditions (TAWC) and TVWC is conducted to underscore the necessity of considering TVWC. Subsequently, a low-computational method is proposed to precompute data essential for WFFC by clustering and selecting representative TVWC, thereby significantly reducing computational demands. Finally, a physics-guided deep learning model is built by incorporating physical features into neural networks to rapidly and accurately acquire control strategies, which can be integrated with wind prediction to enable engineering applications. Simulation results demonstrate that the low-computational method reduces computational demands by 85.5 % while maintaining 84 % of the optimal control performance compared to individually optimizing TVWC. The proposed model outperforms the LUT method by approximately 40 % in control performance, both using the same precomputed data.