Calibration and bias prediction of engineering wake models

Biased predictions of wind farm energy production using engineering wake models are a persistent problem in the wind energy sector. Systematic bias in engineering models is non-linear and a primary contributor to the uncertainty of wake model predictions. In this contribution, we present an approach to calibrate and predict bias of wind farm wake models. The methodology is then applied to a validation dataset of 182 CFD simulations. We observe that a geometric feature called blocking distance is a primary driver of model bias. We compare reductions in bias relative to default model parameters through different procedures (calibration only, bias correction only, and calibration followed by bias correction). Our results show that all three approaches outperform a fully data driven power prediction. Calibration with bias correction reduced the median bias by between 84.5% and 97.9% compared to data-driven prediction, which reduced the bias by between 30.6% and 72.7%, for the wake models used. Reductions in the standard deviation of the bias were between 53% and 87.5% for the calibrated and bias-corrected models. For the data-driven model, bias standard deviation varied between an increase of 58% and a reduction of 45% depending on the wake model.