Assessing engineering wake models against operational data: insights from the Lillgrund wind farm wake steering campaign

Validating engineering wake models under real-world operational conditions is essential for improving wind farm performance predictions. This study uses a unique dataset from the Lillgrund offshore wind farm collected during the Horizon 2020 TotalControl project, combining synchronous SCADA and LiDAR measurements under baseline operation (no intentional yaw offset) and active wake steering. Four analytical wake model combinations are assessed, employing different formulations for velocity deficit, added turbulence, wake superposition, and deflection, implemented in the LongSim software developed by DNV. The analysis focuses on time-averaged wake velocity deficits and turbine- and farm-level power output, with model accuracy quantified using mean absolute error metrics. The models reproduce general wake deficit trends and wake deflection across a range of atmospheric conditions, with normalised velocity deficit errors between 7% and 15%. Power prediction errors increase with farm depth, with turbine-level errors between 3% and 23% and farm-level errors between -13% and +30%. Some analytical models achieve accuracy comparable to reported LES results while requiring substantially lower computational cost. The results highlight the value of field campaigns for benchmarking engineering wake models and inform trade-offs between model fidelity and operational practicality for wake steering applications.

💡 Research Summary

This paper presents a comprehensive field‑based validation of four analytical wake‑model configurations using the unique Lillgrund offshore wind‑farm dataset collected during the Horizon 2020 TotalControl campaign. The dataset combines synchronized SCADA (2 Hz) and long‑range/short‑range LiDAR (0.033 Hz) measurements, providing 10‑minute averaged inflow wind speed, direction, turbulence intensity (TI), and wake velocity fields at 70 m height for periods when all sensors were operational. Both baseline operation (no intentional yaw offset) and active wake‑steering cases (selected turbines yawed clockwise or anticlockwise) are examined, allowing assessment of model capability to reproduce wake deflection as well as power changes.

The analytical models are implemented in DNV’s LongSim tool and differ in four key aspects: (1) velocity‑deficit formulation (standard Gaussian, Gaussian‑Curl‑Hybrid, vortex‑sheet, etc.), (2) added‑turbulence growth model (basic TI‑growth vs. refined turbulence‑intensity‑growth factor), (3) wake‑superposition rule (linear/energy‑based vs. cumulative/energy‑based), and (4) wake‑deflection treatment (simple rotation vs. rotation‑Curl coupling). Two of the four configurations employ cumulative (energy‑based) superposition combined with the refined turbulence scheme, which the authors hypothesize will improve performance for deep‑array conditions.

Model performance is quantified using mean absolute error (MAE) for (i) normalised wake‑velocity deficit (relative to the reference hub‑height speed) and (ii) normalised turbine‑ and farm‑level power output (P/P₀). Across all cases, normalised velocity‑deficit MAE ranges from 7 % to 15 %, indicating that the models capture the general wake centre‑line shift, width expansion, and deflection caused by yaw misalignment. Power‑prediction errors, however, increase with farm depth. Turbine‑level power MAE varies between 3 % (upstream turbines) and 23 % (mid‑farm and downstream turbines), while farm‑wide power errors span –13 % to +30 %. The spread reflects both systematic model bias and compensating errors among turbines.

Key findings include:

• Cumulative superposition + refined turbulence delivers the lowest power‑prediction errors, especially for turbines experiencing multiple overlapping wakes. This configuration reduces the error growth that is evident in linear‑superposition models as wake depth increases.
• Gaussian‑Curl‑Hybrid with cumulative superposition reproduces wake deflection patterns comparable to large‑eddy simulation (LES) results reported in the literature, yet its computational cost is on the order of seconds, far lower than the hours required for LES.
• Vortex‑sheet based models capture yaw‑induced deflection well but tend to over‑predict turbulence recovery, leading to under‑prediction of downstream power.
• All models struggle with flow heterogeneity, wind‑direction variability, and blockage effects that become pronounced in the densely spaced Lillgrund layout (average spacing ≈ 3.5 d). The authors note that the absence of explicit blockage or farm‑scale boundary‑layer representation limits accuracy for deep‑array turbines.
• Measurement uncertainties are carefully discussed. LiDAR’s low sampling rate underestimates TI, while nacelle‑mounted wind speed overestimates it; the authors correct the latter using a TI‑ratio derived from a previous mast‑vs‑nacelle study (Göçmen & Giebel, 2021). Atmospheric stability is inferred from ERA5‑derived Monin‑Obukhov length, but its impact on model performance is only briefly examined.

The study demonstrates that, for operational wake‑steering applications, a well‑chosen analytical model can achieve LES‑level fidelity for wake‑deflection and overall power loss while retaining the computational efficiency required for real‑time control. Nevertheless, the authors stress that further improvements are needed: incorporation of dynamic inflow variability, explicit blockage and secondary‑steering effects, and systematic parameter‑tuning procedures.

In conclusion, the Lillgrund field campaign provides a valuable benchmark for engineering wake models. The paper shows that models employing cumulative wake superposition and refined turbulence treatment strike the best balance between accuracy and speed, making them strong candidates for integration into wind‑farm control strategies. Future work should focus on extending the validation to longer time series, exploring adaptive model calibration, and embedding the validated models within closed‑loop wake‑steering controllers.