A physics-guided deep learning model for real-time wind farm flow control with interpretability analysis

Wind Farm Flow Control (WFFC) is a promising approach for reducing the levelized cost of energy in wind farms. Due to the difficulty of real-time online optimization in satisfying the timeliness requirements of WFFC, current research mainly relies on the look-up table (LUT)-based control methodology. However, the LUT method exhibits limited generalization under unknown wind conditions, which may affect control performance. Consequently, this paper proposes a physics-guided deep learning model for real-time control, utilizing the powerful generalization ability of neural networks to ensure control performance. During model construction, the physical feature characterizing power loss caused by wake effects is integrated into the model to provide a priori knowledge. After model training, the Integrated Gradients (IG) method is employed to elucidate the model's decision-making processes. To validate its effectiveness, thirteen wind turbines (WTs) from an offshore wind farm in China were selected for modeling and analysis. Simulation results demonstrate that the proposed model achieves 95.5 % of the optimal control performance of real-time optimization while satisfying the timeliness requirements of WFFC, and it outperforms the LUT method in control performance with the same computational cost. Furthermore, the model's interpretability is enhanced, and the contribution of the constructed physical feature is validated. • A physics-guided deep learning model is built for real-time wind farm flow control. • Wake-induced power loss is integrated into the model as the physical feature. • The developed model demonstrates good control performance. • The model's interpretability is improved through the Integrated Gradients method.