Simulated fatigue load data set of the Lillgrund wind farm

Data-set creation

This data set was created in association with a conference paper submitted to DeepWind 2023 titled Efficient Mann Turbulence Generation for Offshore Wind Farm Fatigue Load Surrogates.

1000 HAWC2Farm simulations were performed of the Lillgrund wind farm consisting of 48 turbines, each with a nameplate capacity of 2.3MW. Each simulation runs for 1 hour @100Hz with a random inflow wind direction, ambient wind speed (between 4 and 20 m/s), power law shear exponents (between 0.1 and 0.3), and turbulence intensity (between 2.5% and 10%).

The windfields immediately infront of each turbine are recorded every second in a 30x30 grid with dimensions 1.1Dx1.1D (where the rotor diameter is D=92.6m). The loads of the blade roots, tower base, tower top, and main shaft are recorded.

Each 1 hours HAWC2Farm simulation is split into 6x10minute observation. The wind fields are temporally averaged over the 10 minute window, providing averaged wind field profiles (Fig. 1). Additionally, the residual background turbulence is measured over the 10 minute period, providing wind field deviation profiles (Fig. 2).

Damage equivalent load (DEL) quantities are calculated for the corresponding 10 minute periods. DELs for blade components use a wholer exponent of 10, whereas tower and main shaft components use a wholer exponent of 4.

With 6 time intervals and 48 turbines, each HAWC2Farm simulation provides 6x48=288 observation pairs, with the wind field as the input and DELs as output. The total data-set consists of 288000 observations. 230400 (80%) of these observations make up the training data set provided in this data set. The other 57600 (20%) observations have been embargoed as a blind testing set for future investigations.

Data-set description

The data has been split into two groups:

• Training set input (dataset_input.parquet)
• Training set output (dataset_output.parquet)

Data files each contain dataframes in Apache Parquet format, which can easily loaded using Data Frame packages. For example, in Python, parquet files can be loaded with Pandas (using the pandas.read_parquet() function), Polars (using the polars.read_parquet() function), or PyArrow (using the pyarrow.parquet.read_table() function).

Training set inputs

The mean and standard deviation profiles for each observation are 30x30 arrays of floats. These are stored in row-major order as two vectors of 900 elements each. The elements of the mean wind field are stored in the column labeled mean_XXX where XXX is the index (from 0 to 899) of the element. Similarly, the standard deviation field is labelled std_XXX. The 1800 columns of the input data set are therefore:

• mean_000
• mean_001
• ...
• mean_899
• std_000
• std_001
• ...
• std_899

Training set outputs

All output quantities are averaged over a 10 minute period. Blade damage equivalent loads (DELs) are computed using a wholer exponent of 10. All other loads use a wholer exponent of 4.

• pelec - Electrical power [kW]
• MxBr - Blade root edgewise DELs [kNm]
• MyBr - Blade root flapwise DELs [kNm]
• MzBr - Blade root torsion DELs [kNm]
• MxTB - Tower bottom side-side DELs [kNm]
• MyTB - Tower bottom fore-aft DELs [kNm]
• MzTB - Tower bottom torsion DELs [kNm]
• MxTT - Tower top fore-aft DELs [kNm]
• MyTT - Tower top side-side DELs [kNm]
• MzTT - Tower top torsion DELs[kNm]
• MxMB - Main bearing bending DELs [kNm]
• MyMB - Main bearing bending DELs [kNm]
• MzMB - Main bearing torsion DELS [kNm]

Task

Build a predictive model mapping the wind field parameterization at the turbine rotor to the turbine damage equivalent loads and power output.

Performance

Performance of the predictions shall be compared against baseline predictions, where the mean values of the outputs from the training data-set are used as predictions.