The OWA Anholt Array Efficiency Benchmark

The OWA Anholt Array Efficiency BenchmarkThis initial benchmark of the “OWA Wake Modelling Challenge” will allow participants to test wake models in the assessment of array efficiency for one of the largest offshore wind farms under the influence of coastal wind speed gradients.

Javier Sanz RodrigoBlockedUnblockFollowFollowingFeb 27StatusThe benchmark is open for participation through April 2019.

Please contact the benchmark manager, Javier Sanz Rodrigo, if you would like to take part.

Registered participants will receive and identification code (userID) which they will use to submit their data and identify their results in anonymous model intercomparison.

They will also be required to fill a questionnaire providing details about their simulations that facilitate the assessment of the results.

BackgroundThe “OWA Wake Modeling Challenge” is an Offshore Wind Accelerator (OWA) project that aims to improve confidence in wake models in the prediction of array efficiency.

A benchmarking process comprising several wind farms will allow model developers and end-users test their array efficiency prediction methodologies over a wide range of wind climate and wind farm layout conditions.

You can read more about the scope of this project and the benchmarking process in this document.

The Anholt wind farm has been studied by Peña et al.

(2018), van der Laan et al.

(2017) and Nygaard (2014).

This benchmark follows some of the methodologies described in Peña et al.

(2018) that aimed at developing ways to incorporate mesoscale data in engineering wake models to predict array efficiency under strong horizontal wind speed gradients.

Scope and ObjectivesThe Anholt benchmark is a pilot to define, together with the participants, an open-source model evaluation methodology for array efficiency prediction.

The benchmark is set up as a blind test so you won’t be able to access observational data.

Instead, mesoscale simulation data is available for the modeller to decide on the best interpretation of the input data for the specific needs of the wake model.

Participants will be able to test the evaluation scripts, based on Jupyter notebooks, on own simulation data and submit their best prediction.

The results, together with the first release of the model evaluation methodology will be published at WindEurope Resource Assessment Workshop (27–28 June, Brussels).

This means that you are expected to submit your data by end of April.

Test Case: Anholt Wind FarmAnholt is one of the largest offshore wind farms in the world with 111 turbines totalling 399,6 MW.

Siemens SWT-3,6–120 turbine is installed in all positions with a hub height of 81.

6 m and a rotor diameter of 120 m.

The layout spans 22 km from North to South and the smallest distance between turbines is 4.

9D, being D the rotor diameter.

The large size of the layout and the presence of the coast, at distances from 20 to 120 km, make the wind farm experience significant horizontal wind speed gradients (Figure 1).

Figure 1: Anholt wind farm situation, windrose and distance to the coast in km.

Validation dataSupervisory control and data acquisition (SCADA) operational data will be used to perform the validation.

A period of 2,5 years is available from 1 January 2013 to 30 June 2015.

A quality control process has been carried out to produce a “clean” dataset that only includes situations where a turbine is working in nominal conditions, i.


whose power output is close to the value predicted by the manufacturer’s power curve and, therefore, corresponds to the operational conditions simulated by wake models in the pre-construction phase.

A machine learning technique is used for gap filling to recover time instances when only a few turbines are working in non-nominal conditions.

In these situations, a regression algorithm trained on clean data from neighbour turbines predicts corrected data for the missing turbines to obtain a complete dataset.

The effect on the overall array efficiency assessment is minor compared to the benefits of obtaining a validation dataset that is more statistically significant.

As a result, quality-control corrected data consist on hourly timestamps of power output and nacelle wind direction with all turbines working in nominal conditions.

The nacelle wind direction is calibrated following the method of Hansen (2015).

Then, validation data is defined in terms of sector-wise and stability-wise ensemble averages for a 9±1 m/s velocity bin, when the thrust coefficient is at its maximum resulting in stronger wake effects.

These reference wind conditions are based on mesoscale data as described in the next section.

Input DataUnfortunately there are no meteorological measurements available that could be used to define inflow conditions for wake models.

Peña et al.

(2018) use the equivalent wind speed and direction inferred from the SCADA data as a proxy.

The method depends on selecting a number of “free-stream” turbines to define the reference inflow conditions for each wind direction sector.

This is a bit arbitrary and difficult to generalize consistently to other layouts.

Besides, we would expect the equivalent wind speed and direction to be affected by site effects due to the presence of neighbor turbines, wind farm blockage or local accelerations near corners or through gaps inside the array.

An alternative, which can be generally applied to any wind farm, is to use mesoscale simulations to generate background wind conditions for wake models that are completely free of (microscale) site effects.

Recent validation activities in the New European Wind Atlas (NEWA) have demonstrated high accuracy in offshore conditions (Hahmann et al.

, 2019).

Indeed, a mesoscale simulation has been produced with the Weather Research and Forecasting (WRF) model following the NEWA production run settings.

In this case, three one-way nested domains of 27, 9 and 3 km resolution are configured centred at the wind farm centroid (56.

6ºN; 11.


The vertical grid has 61 terrain-following (sigma) levels, with 10 levels covering the first 200 meters, more specifically at: 6, 22, 40, 56, 73, 90, 113, 140, 179 and 205 meters.

High resolution topography (SRTM 90m) and updated land use categories (Corine Land Cover 2018), together with the Noah land-surface model are used to define the boundary conditions at the surface.

The physical parameterizations are: Mellor–Yamada–Nakanishi Niino 2.

5-level planetary boundary-layer scheme (MYNN), WRF Single-Moment 5-class microphysics scheme, the Rapid Radiative Transfer Model for GCMs shortwave and longwave radiation schemes and the Kain-Fritsch cumulus scheme in the outermost domains 1 and 2.

The simulation is driven by input data from ERA-5 in blocks of 5 days with additional spin-up time of 24 hours.

We will use the centroid of the mesoscale simulation to define reference wind conditions in terms of hub-height interpolated wind speed and direction and surface-layer stability, defined by z/L parameter where z = 10 m and L is the Obukhov length computed by WRF.

Mean profiles at the reference site are produced by horizontally averaging data from a 30-km wide 10×10 squared grid around the centroid.

Ensemble averaged conditions are defined in terms of 30º wind direction sectors (centred at 0, 30, 60, etc) and three stability classes (Figure 2):Unstable (u): -0.

2 < z/L < -0.

02Neutral (n): -0.

02 < z/L < 0.

02Stable (s): 0.

02 < z/L <0.

2Figure 2: Wind climate distribution simulated with WRF at the reference site.

The number of hourly samples at the 9±1 m/s velocity bin is shown in Table 1, indicating the statistical representativeness of each class in the validation range.

Table 1: Number of hourly samples in each ensemble-averaged class considered in the validation range.

Mesoscale and SCADA hourly data is synchronized and flagged to filter out registers in non-nominal conditions that will not participate in the validation.

To facilitate the interpretation of mesoscale data by modellers different input datasets are produced, namely:Fields: netCFD files, cropped from the original WRF output files, of 3D fields at 3 km resolution covering the wind farm area (30×30 km) and surface quantities.

3D fields: velocity components (U, V), potential temperature (θ), turbulent kinetic energy (TKE).

Surface quantities: 2-m temperature (T2), surface temperature (TSK), 10-m wind speed (U10, V10), Obukhuv length (L), friction velocity (u*), heat flux (HFX)Reference mean profiles: netCDF file with horizontally-averaged ABL profiles over a 30×30 km area around the layout centroid including momentum and potential temperature tendencies, representative of ABL conditions and mesoscale forcings across the wind farm.

Time series at turbine positions: Hub-height interpolated U, V, θ and TKE and surface quantities.

Additionally, the turbine coordinates and the manufacturer’s power curve are provided.

All the input data is provided to registered participants through a b2drop input folder.

Output DataThe ultimate goal is to analyse ensemble-averaged statistics of array efficiency η, at individual turbine level and for the whole wind farm, defined as:where Pi and Si are the power and “free-stream” (mesoscale) wind speed at turbine position i.

Participants should submit their power predictions for ensemble-averaged quantities per wind direction and stability classes (Table 1), in a .

csv file with the following format.

Turbine #, P1n, P2n, …, P12n, P1u, P2u, …, P12u, P1s, P2s, …, P12swhere P1n means power in MW in sector 1 for neutral conditions and so on.

Hence, for Anholt, a 111×37 matrix would follow the one-line header above.

If your model does not handle stability explicitly you still need to submit results for all classes and explain what kind of proxy you are using to differentiate from neutral conditions.

Alternatively, if your method produces a time series of power data, you can submit directly the time series concurrent with the mesoscale input data and the ensemble averaged table will be processed by the benchmark manager.

Time-series allows extending the analysis to other array efficiency predictors that can be obtained from the mesoscale data (TKE, tendencies, etc).

You should submit your results in a .

csv file with the following formatdatetime, P1, P2, … , P111where P1 means power for turbine 1 and so on and you shall use the same datetime indices of the mesoscale data and format YYYY-mm-dd HH:MM:SS.

As a bonus, you are welcome to provide uncertainty estimates on the power predictions.

Simply add the sigma values next to the mean values in the same files, e.


datetime, P1, P2, … , P111, P1std, P2std, … , P111stdPlease use the following file naming convention: Anholt _modelID_userID.


Output data will be shared with the benchmark manager by a private b2drop output folder, that only allows uploading files.

Schedule6 March 2019: benchmark launch11 April 2019: webinar to discuss on-going results and evaluation script30 April 2019: webinar, results due to be included in WindEurope’s Resource Assessment Workshop17 May 2019: Submission of poster27–28 June 2019: Presentation in BrusselsAcknowledgementsThis benchmark is organized with support from the Offshore Wind Accelerator (OWA).

The Anholt observational data is provided by Ørsted A/S.

ReferencesHahmann et al.

(2018) WRF sensitivity experiments for the mesoscale NEWA wind atlas production run.

EGU 2018, accepted for manuscript.

Hansen K.


(2015) Guideline for qualification of SCADA data for wake efficiency analysis.

In “WAKEBENCH best practice guidelines for wind farm flow models”.

Edited by Sand Rodrigo J.

and Moriary P.

IEA Task 31 Report to the IEA-Wind Executive Committee, IEA Wind, 2015Peña A, et al.

(2018) On wake modeling, wind-farm gradients, and AEP predictions at the Anholt wind farm.

Wind Energ.


3:191–202, https://doi.


5194/wes-3–191–2018Nygaard N.

G (2014) Wakes in very large wind farms and the effect of neighbouring wind farms.



: Conf.


524 012162, https://doi.


1088/1742–6596/524/1/012162van der Laan, et al.

(2017) Challenges in simulating coastal effects on an offshore wind farm, J.




, 854, 012046, https://doi.


1088/1742-6596/854/1/012046, 2017.

. More details

Leave a Reply