Wind farm layout optimization using LES-trained machine learning

Optimizing wind farm layouts remains computationally intensive, especially with high-fidelity methods like large eddy simulation (LES). Although reduced-order models such as the Gaussian-Curl Hybrid (GCH) are efficient, their deviation from LES results limits their reliability for layout optimization. This study extends a machine learning (ML) framework based on a convolutional neural network (CNN) with U-Net skip connections to predict mean flow fields in a utility-scale wind farm with 12 turbines. Trained on LES data and low-fidelity GCH model inputs, the ML model reduced flow field prediction error for an unseen turbine layout from over 11% of the GCH model to under 3%. Integrated into a greedy optimization algorithm, it improved the South Fork Wind Farm layout, yielding a 2.05% increase in annual energy production. The ML predictions closely matched LES results and offered over 99% reduction in computational cost, highlighting the method’s accuracy, scalability, and efficiency for wind farm layout optimization.