Correlating wake-model performance to the atmospheric state: a SCADA-based analysis at two offshore wind farms
Abstract. Engineering wake models are routinely validated against SCADA measurements with discrepancies between predicted and observed power often attributed directly to limitations in the wake model itself or to inadequate calibration. Wind farm power assessment however involves a sequence of processing steps, of which wake modelling is only the final stage. Disagreement between predictions and observations may therefore stem from the wake-model physics itself but also from the software implementation, the reconstructed inflow, or the atmospheric conditions sampled.
This work aims to assess these different contributions independently by evaluating the entire wake modelling chain, from wind-resource assessment to the final power estimation. SCADA data from two offshore wind farms, Westermost Rough and Anholt, are processed using two independently developed filtering workflows and complemented with profiling LiDAR measurements and 1 km-resolution WRF simulations. The resulting inflow conditions are used to drive two wake models, Jensen and TurbOPark, both sensitized to the ambient turbulence intensity. Three wake-modelling software packages are considered (FLORIS, PyWake, and FOXES) while WindIO is used to ensure identical inputs across all implementations.
The two SCADA-filtering workflows yield nearly interchangeable wind speed timeseries (bias ≤ 0.04 m s−1, r ≈ 1), but differ in wind direction by up to 7.27° at Anholt, likely due to differences in northing calibration procedure. For identical inputs, software implementation alone produces bias spreads of up to 4 % for TurbOPark, comparable to the effect of the wake-model choice itself, while Jensen remains consistent across packages. Farm-level bias is further shown to be an insufficient performance indicator that should therefore be complemented by additional metrics such as farm-level RMSE, correlation, and turbine-level statistics. Indeed, bias compensation can lead to near-zero farm-wide biases that hide binned farm-mean errors of ±5 % and farm-MAE exceeding 20 % under very stable or strongly heterogeneous inflow conditions. At the turbine-level, the analysis highlights the need to account for local wind speed gradients when evaluating individual turbine power, establishing clear spatial correlation between wind speedup maps and turbine power error spatial distribution.
Overall, these results support a shift away from AEP-based validation toward time-series or condition-binned approaches. They additionally motivate a better inclusion of physical phenomena such as heterogeneity-aware models. They finally call for better community standards, through, e.g. the development of shared standards such as WindIO.