10-min time series

GEO-NET recently offers the WRF-GN_ML wind time series for Germany (other countries on request). These have a temporal resolution of 10 minutes. The spatial resolution of the wind time series is 3 km.

Method

The data basis of the WRF-GN_ML time series is provided by the ERA5 reanalysis data. These are globally available free of charge with a resolution of about 30 km. The temporal resolution of the data, which are updated daily with a time lag of about five days, is one hour. 

The Weather Research and Forecasting model WRF is used to downscale the ERA5 data to a spatial resolution of 3 km and a temporal resolution of 10 minutes. The microphysical processes, surface physics, boundary layer physics, atmospheric radiation physics, and cloud physics can be accounted for in WRF using various parameterization approaches. The optimal settings of these parameterizations for the most realistic wind distribution were determined in a sensitivity study. The WRF wind time series were validated with one-year or longer measurement time series from 50 sites.

In a final step, the wind speeds are optimized using machine learning (ML). For this purpose, an artificial neural network (ANN) is used, which is trained with the wind speeds from WRF, other time-dependent and site-specific variables, such as the time of day (input neurons) and the matching measured data of the wind speed (output neuron). As a result of the training algorithm, a function is obtained that optimizes the WRF results as a function of the input parameters. This procedure is regularly refined by additional measurement data and input parameters.

Validation

To validate the WRF-GN_ML time series, the wind speeds and wind directions from 100m height were compared with the values of 50 measurement sites. To evaluate the quality of the data, statistical quantities considered included the correlation coefficient R², the coefficient of variation CV, and the deviation of wind direction.

Comparison of the 10-min values

The median of the correlation coefficient, which is a measure of the degree of a linear relationship between two data sets, is 0.73 for the WRF-GN_ML time series, which was 0.66 before optimization using machine learning (see WRF-GN). The other models, such as ERA5, EMD-WRF EUROPE+, and ConWx, also have a lower average correlation with the measured data than WRF-GN_ML.

Besides the correlation coefficient, the coefficient of variation CV, which is a measure of the wind distribution, is also taken up for the evaluation of the results. If the CV of the model and measured data are similar, it can be concluded that the Weibull distribution of the two time series is similar. The figure shows the percent deviation between the CV values of the model and measured data. The median of this deviation is 1.20% for the WRF-GN_ML time series. Thus, the deviation of the WRF-GN time series is reduced by 2.80 percentage points by the ML optimization.

For the wind direction, no ML optimization is performed for the WRF-GN_ML time series, so that the wind directions of the WRF-GN_ML time series correspond to those of the WRF-GN. The median deviation of the wind direction from the measured values for these models is 24.39°, which is in the range of the other models considered (EMD-WRF EUROPE+: 22.02°; ERA5: 21.13°; ConWx: 26.78°).

Purpose

• Basis of time series dependent yield calculations in the absence of wind measurements
• Basis for extreme wind assessments
• Data source for long-term correction of reference data
• Calculation of combined electricity production for hybrid power plants (wind-PV)
• Better comparison of load profile and production possible → Calculation of self-sufficiency levels