A fully three-dimensional deep learning surrogate model for large-eddy simulation-resolved wake dynamics in wind farm

Wake interactions in large wind farms significantly reduce power output and increase structural fatigue loads, making accurate wake prediction essential for farm design and control. However, large-eddy simulation (LES) is computationally prohibitive for repeated evaluations, and existing low-fidelity models have limited capability to capture three-dimensional unsteady wake dynamics and complex multi-turbine interactions. To address these challenges, a deep learning framework is proposed for long-term prediction of three-dimensional turbulent wake flows in wind farms, employing a Swin Transformer backbone integrated with a U-shaped convolutional neural network architecture. It is trained on high-fidelity LES data that include multiple turbine tilt configurations to capture three-dimensional wake dynamics and vertical deflection. The model accurately predicts instantaneous three-dimensional streamwise velocity fields over long horizons, reproducing large-scale wake structures and turbine-induced flow variations. Detailed physics-based analyses confirm that the framework captures velocity deficit profiles, wake centerline trajectories, and rotor inflow velocities, even for the unseen tilt configuration. Incorporating limited upstream inflow information further improves the inflow velocity predictions for turbines located near the inflow boundary. Overall, the proposed model serves as an efficient surrogate for LES, capturing the key wake dynamics and multi-turbine interactions with substantially lower computational cost.