Accurate and computationally efficient prediction of wind turbine fatigue loads is essential for load-constrained wind farm co-design. This work compares surrogate models trained using a hierarchy of local inflow parameterizations, ranging from rotor-averaged to sector-averaged, proper orthogonal decomposition, and convolutional neural networks, operating on rotor-resolved inflow fields. All surrogate models are trained on a dynamic wind farm aero-servo-elastic simulations database, and evaluated for predictive accuracy, wake-induced load sensitivity, and generalization to different wind farm layouts. Across all cases, rotor averaged based surrogates show the largest errors, while for the other parameterizations the error is within the seed-to-seed uncertainty. The higher-fidelity representations reliably capture partial- and full-wake effects. Coupling with quasi-static wind farm simulators is done via the newly developed wind-farm-loads python package, whose application reveals that a mismatch in wake-added turbulence between dynamic and quasi-static simulators biases load predictions, highlighting the need for improved wake-added turbulence modeling and calibration.
Link to paper: https://zenodo.org/records/22030258