A Wind Atlas for Germany and the Effect of Remodeling PDF

B Meteorol. Z. (Contrib. Atm. Sci.), Vol. 31, No. 2, 117–130 (published online February 4, 2022) © 2022 The authors Energy Meteorology A wind atlas for Germany and the eff...

B Meteorol. Z. (Contrib. Atm. Sci.), Vol. 31, No. 2, 117–130 (published online February 4, 2022) © 2022 The authors Energy Meteorology A wind atlas for Germany and the effect of remodeling Martin Schneider1∗ , André Glücksmann1 , Anselm Grötzner2 and Heinz-Theo Mengelkamp1 1 anemos Gesellschaft für Umweltmeteorologie mbH, Reppenstedt, Germany 2 Ramboll Deutschland GmbH, Kassel, Germany (Manuscript received June 28, 2021; in revised form November 19, 2021; accepted November 23, 2021) Abstract Minimizing and quantifying the uncertainty of wind simulations are essential for the wind energy industry during the planning phase of wind farm projects and for financial considerations. Measurements at 118 sites onshore and offshore in Germany are analyzed and used for the verification of wind simulations with the mesoscale model WRF. In order to minimize the difference between simulations and observations a correction of the annual cycle is applied and a remodeling approach is developed which allows for a correction of the simulated wind speed time series. The remodeling methodology is based on a linear regression analysis of simulated and observed wind speed time series accounting for sub-grid variations of orography and roughness. Averaging the regression parameter for 26 measurement sites results in an overall global parameter set which is applied to the wind atlas data. While the “raw” data (without optimization) before any correction showed differences of up to 30 % with respect to the annual mean wind speed the remodeling process reduced the bias to below 5 % for the majority of measurements. When being compared with data from the NEWA wind atlas and the EMD-WRF Europe+ data set an overall bias between 0.6 m/s and 0.8 m/s is found for the NEWA, EMD-WRF Europe+ and anemos “raw” data. This bias is reduced to zero with a small standard deviation when the remodeling process and the site-specific adaptation are applied. Keywords: mesoscale model verification, wind atlas, wind measurements, adaptation of wind simulations 1 Introduction calculations (Serban et al., 2020) and is still widely ap- plied today. However, approximating the energy of the The wind energy industry has developed dramatically wind by statistical means is never totally perfect com- during the last three decades and wind power now has pared to time series. Moreover, wind turbines have to a major share in the transformation from fossil fuels to be shut down temporally i.e. for noise reduction or bat renewable energy. Parallel to this trend wind power me- protection during nighttime and time series analysis is teorology has evolved with focus on three main areas, required to accurately estimate the yield losses. When namely the short-term prediction of electricity produc- information is needed of the market value of the elec- tion, site suitability (turbulence and extreme winds in or- tricity produced a temporal correlation with the variable der to estimate the mechanical stress on wind turbines) stock exchange price for electricity is essential. Highly and resource assessment. Reducing the uncertainty in accurate site-specific time series of the energy yield are wind speed and direction simulations is of utmost impor- appropriate means to meet the requirements for sound tance for wind farm developers, investors and financing financial considerations during the wind farm planning institutions. Even small errors can have a large impact phase. As the life time of wind turbines is assumed more on financial considerations as the electricity production than 20 years the time series should span a climatologi- by wind turbines increases non-linearly with respect to cally relevant time period of a similar length. wind speed. Therefore, it remains a major challenge for Reliable wind information over two or more decades the wind power meteorological community to minimize is rarely available from observations. Except a few tow- the uncertainty in wind resource assessment to the extent ers for research purposes the majority of observations of possible. This study focuses on resource assessment on wind speed and wind direction takes place near the sur- regional to local scales and the quantification of its un- face and is influenced by surface characteristics (orogra- certainty through the comparison of model simulations phy, roughness, obstacles) in the immediate surrounding with observations. (Wieringa, 1980, 1996). Weather station data seem to Since its early stages wind resource assessment has be the first choice but Lindenberg et al. (2012) have been based on statistical methods (Petersen et al., shown that such data are inconsistent in time due to 1981; Mengelkamp, 1999; Mengelkamp et al., 1997; changes in the instrument location or due to changes Badger et al., 2014). Estimating the site-specific wind in surface characteristics of the surrounding area. In potential in terms of the wind speed frequency distri- addition, near surface measurements do not reflect the bution has been the preferred approach for energy yield wind conditions (i.e. the diurnal cycle, low level jets) in ∗ Corresponding author: Martin Schneider, anemos GmbH, Böhmsholzer heights over 100 m which is exceeded by modern wind Weg 3, 21391 Reppenstedt, Germany, e-mail: [email protected] turbine hub heights. Site-specific wind measurements © 2022 The authors DOI 10.1127/metz/2022/1102 Gebrüder Borntraeger Science Publishers, Stuttgart, www.borntraeger-cramer.com 118 M. Schneider et al.: A wind atlas for Germany and the effect of remodeling Meteorol. Z. (Contrib. Atm. Sci.) 31, 2022 during the planning phase of a wind farm commonly data sets almost identical to the one described in this pa- only span a short time period of usually 12 months and per. The New European Wind Atlas (Gottschall et al., require a long-term correlation with a consistent long- 2019; Witha et al., 2019; González-Rouco et al., term wind data set. In addition, hub heights of modern 2019; Dörenkämper et al., 2020) and the EMD-WRF wind turbines will often reach more than 150 m. Mea- Europe+ (ERA5) mesoscale data set (EMD, 2020a, b) surements at these heights are cost intensive and are of- both were simulated with the mesoscale model WRF ten aimed to be avoided. with 3 km horizontal resolution based on ERA5 reanaly- Downscaling reanalysis data by use of a mesoscale sis data. These two data sets cover all of Europe and are model seems an appropriate approach to derive regional described and compared to our data set with the same or local scale wind information. Motivated by the im- forcing data and the same horizontal resolution for the portance of mesoscale model simulations for the wind WRF mesoscale model for Germany in Section 5.4. Sec- industry the Weather Research and Forecast mesoscale tion 6 comprises a summary and a brief outlook. model WRF (WRF, 2017; Skamarock et al., 2008) has become a major tool for investigations into wind condi- tions and sensitivity studies. The latter mainly examine 2 Model set-up and wind atlas the influence of different PBL schemes (Fernández– simulation Gonzales et al., 2018, Yang et al., 2013), or a com- bination of PBL schemes, grid configuration and initial The mesoscale model WRF (Weather Research & Fore- conditions (Siuta et al., 2017). Carvalho et al. (2014) casting Model version 3.7.1) (Skamarock et al., 2008; applied different reanalysis data sets to drive the meso- WRF, 2017) is used to downscale ERA5 reanalysis scale model WRF and compared the simulated hourly data (C3S, 2017) to the region of Germany (Fig. 1). time series of wind speed and -direction with observa- A two-way nesting approach is realized to downscale tions. An overall positive wind speed bias is explained the ERA5 reanalysis data with a horizontal resolution by the smoothing of orography in the model caused by of approx. 30 × 30 km2 to two domains with 9 × 9 km2 its limited spatial resolution. There was no best model for the outer domain and 3 × 3 km2 for the inner nested configuration for all conditions and an ensemble of domain, respectively. 50 vertical levels are prescribed model runs is often suggested (Fernández-Gonzales up to the upper boundary at roughly 15 km height with et al. 2018, Siuta et al., 2017, Deppe et al., 2013) to- 14 levels in the lowest 300 m which are most relevant for gether with some kind of bias correction when using wind energy applications. Initial and boundary condi- model simulations for “real world” applications such as tions were taken from the ERA5 data which are nudged financial considerations (Siuta et al., 2017, Deppe et al., (grid nudging method) into the WRF model every hour. 2013). The specific and the relaxation zone (the first five rows Weiter et al. (2019) follow the idea of using the and columns in each domain) have boundary conditions electricity production from wind turbines directly for from the reanalysis or values blended from the coarser the verification of wind simulations. They transferred domain, respectively. The WRF User Guide provides wind speed time series from simulations with the WRF more information on the model operation. model to production by use of the respective wind tur- Model output is stored every 10 minutes for a tran- bine power curves. The production data from opera- sient simulation starting in 1997 and still continuing. tional wind turbines were analyzed very carefully in Thus, more than 24 years simulation are currently avail- order to reflect the wind conditions without any influ- able. Different long-year simulations were distributed on ence of the wake effects and operation mode of the tur- several CPU with a one-month spin-up time each. bines. For 10-min data from 50 turbines in 12 wind- Orography data are taken from the SRTM (Shuttle farms the relative bias in production ranges between Radar Topography Mission, USGS EROS Data Center, 10 % and 25 % which in terms of wind speed means a Farr et al., 2007) and interpolated from its 90 m resolu- bias of 4 % to 10 % for the annual mean. This seems to tion onto the model grid. Vegetation and roughness in- be an acceptable uncertainty for financial considerations formation is provided by the CORINE data (Keil et al., in the wind industry. In order to reach such small uncer- 2010) of the European Environment Agency and is inter- tainty numbers a bias correction was applied to the wind polated from its 100 m resolution. Soil temperature and speed time series. soil moisture at 4 soil levels and snow cover are taken This paper aims at optimizing a wind atlas for Ger- from the ERA5 data set (C3S, 2017). many by a remodeling process and verifying the opti- The WRF model contains many different schemes mized wind speed time series. It is organized as follows. for the parameterization of physical processes. WRF’s Details of the wind atlas simulation are given in the physics parameterization considered here is the Yon- next section. The observational data used for the adap- sei University (YSU) planetary boundary layer scheme, tation and verification are described in Section 3 and the Monin-Obukhov surface layer description, the Noah the remodeling approach and its expansion compared to land surface model including Mosaic 4, the RRTM the approach described in Weiter et al. (2019) in Sec- scheme for the longwave radiation, and the Dudhia pa- tion 4. Section 5 presents the comparison of model sim- rameterization for the short-wave radiation. No cumulus ulations with observations. There are two further wind parameterization is applied (Skamarock et al., 2008). Meteorol. Z. (Contrib. Atm. Sci.) M. Schneider et al.: A wind atlas for Germany and the effect of remodeling 119 31, 2022 Figure 1: WRF model domain over Germany. Multiple Nesting with domain 1 (9 × 9 km2 ) and domain 2 (3 × 3 km2 ). Despite the many sensitivity studies mentioned the one measurements which, with increasing wind turbine hub optimal set-up for wind energy applications hardly ex- height, become more advantageous by reason of finan- ists. A model performing best in a certain situation over cial and permit issues. A few research towers are oper- a certain region may not do so at other circumstances. In ated by public institutions onshore and offshore within several of the studies cited ensemble simulations were the model domain for periods of several years. In this suggested to be most appropriate. However, limited re- study we used data from onshore and offshore research sources did not allow us to perform ensemble simula- towers and short term measurements for wind farm plan- tions nor to undertake extensive sensitivity studies as ning purposes at met masts and with lidar devices. The done by Olsen et al. (2017) and Hahmann et al. (2020). observational uncertainty is considered very low as all Also, ensemble and sensitivity studies reflect the be- towers and lidar measurements were erected for wind havior of different model set-ups rather than the uncer- energy planning purposes and followed international tainty with respect to observations which only is rel- recommendations (IEC, 2017) or were installed as re- evant for wind energy applications. As the near sur- search towers. Calibration sheets for all anemometers of face winds are most sensitive to the choice of the PBL the wind energy towers and for all lidar systems were scheme (Hahmann et al., 2020) an investigation into the available. The instrumentation of the research towers is most appropriate PBL scheme for wind simulations has considered to be maintained on a regular basis. All data been undertaken comparing the wind conditions above were measured as 10-min averages and aggregated to 100 m height at 10 sites as simulated with the YSU hourly values for the verification while the remodeling and MYJ PBL schemes for August and December 2012. approach is based on the 10-min measurements. These schemes are widely applied in WRF simulations Two data sets are used for this study. Data set 1 and seem to be appropriate for wind simulations in the comprises data from 48 onshore met masts and lidars boundary layer (Hu et al., 2010, 2013), A best scheme of which 26 are used for the remodeling process and for all situations and sites cannot be proposed (Gian- 22 for the independent verification. Measurements at nakopoulou et al., 2014). The YSU scheme performed the four offshore masts FINO1 (North Sea), FINO2 reasonably when wind speed profiles and the diurnal cy- (Baltic), FINO3 (North Sea), and NordseeOst (North cle were compared to measurements at our 10 sites. Sea) are used for the verification and separate offshore remodeling. All these data are at the hands of anemos. In order to support the verification results with a second 3 Observational data independent data set wind speed time series of the wind atlas simulation were provided to the consulting com- It is with the growing wind industry during the last pany Ramboll for 66 locations at which they had tall two decades that a reasonable number of meteorologi- mast and lidar measurements available from their con- cal towers has been installed with heights of more than sulting activities. It is ensured that those data are of sim- 100 m meanwhile. These towers usually are operated ilar very high quality as the first data set and that they only for a period of one year and are privately owned. are analyzed in a similar way. The analysis of the mea- Given access to their data they form an excellent data surements covers a plausibility check, identification of base for model verification. This also holds for lidar icing periods and missing values for any reason and a 120 M. Schneider et al.: A wind atlas for Germany and the effect of remodeling Meteorol. Z. (Contrib. Atm. Sci.) 31, 2022 ucell WRF rem is the wind speed at the grid cell center after the remodeling described in step 3. usite WRF rem is the wind speed after the remodeling pro- cess made comparable to wind measurement sites fol- lowing step 4 resp. step 2. A verification is performed twice, before and af- ter the remodeling process in order to demonstrate the effect of the remodeling on the wind atlas. For both, the verification and the remodeling process observations and model simulations have to be made comparable. While the model grid cell height is an average for the 3 × 3 km2 area a met mast for wind energy purposes commonly is placed at an exposed location advanta- geous for wind farm operation. An elevation and rough- ness correction for wind speed is applied on the model data that accounts for speed-up effects over unresolved crests (Howard and Clark, 2007). This correction has been developed with the computational fluid dynamics (CFD) code Meteodyn WT (2015). The difference be- tween the site-specific wind speed and the wind speed at the 3 km grid cells was compared at 10 sites to the Figure 2: Geographical distribution of measurement sites. height difference Δh = hsite − hcell and the roughness difference Δz0 = z0site − z0cell. It results in the empiri- cally derived speed-up factors α and β. This correction correction for mast shadow effects in case of only one accounts for the height difference Δh between the ele- anemometer per height was installed. If two anemome- vation asl of a particular measurement site and the ele- ters were installed at the respective height pointing into vation of the model grid cell and the difference Δz0 in opposite directions the undisturbed data were used. A roughness length. Because the roughness tables for Me- correlation among all wind measurements at the same teodyn and WRF differ but are both based on the Corine mast helped to identify any implausible data. In summer data set, the roughness table from WRF was transferred 2009 east of FINO1 the offshore wind farm “alpha ven- into the roughness table for Meteodyn. The correction tus” was erected with influence on the FINO1 measure- changes the original model wind speed output ucell raw WRF to ments for easterly winds. Only FINO1 data before this the wind speed usite raw adapted at the site of the mea- WRF date are used in this study. Similarly, if disturbances oc- surements but still with systematic errors of the simula- curred during the measuring period e.g. due to extension tion. This step only makes the simulated data compara- of a wind farm, this particular met mast was excluded ble to the measurements at a particular site from this study or data from times before the disturbance WRF = uWRF · (1 + Δh · α + Δz0 · β) usite raw cell raw affected them were used only. For each measurement (4.1) station a 12-month period was selected, but not neces- sarily the same period for all stations, because the simu- with α and β being constants in units of m−1 empiri- lation covers more than 24 years. All measurement sta- cally derived at 10 sites different from the ones used in tions should reach data availability of at least 80 %. Pe- this study. α and β were developed from speed-up fac- riods of missing observations were also eliminated from tors found by CFD model simulations at the respective the respective simulated time series. The measurements 10 sites. were distributed all over Germany with the majority of them in complex terrain in the southwestern parts, two 4.1 Remodeling in the western part of Poland and one in the northern part An optimization approach is applied to the wind speed of Switzerland. (Fig. 2). time series of the lowest levels up to 300 m of the site- adapted raw WRF simulation usite raw WRF. 4 Verification and remodeling The remodeling process is based on a comparison of simulated wind speed time series usite raw WRF with observa- We denote the wind speed at different stages of the tions at 26 met masts onshore in a height range of 80 remodeling process as follows. to 140 m. A separate correction function was derived for ucell WRF raw is the raw wind speed as simulated in the the offshore met masts. Because there were only 4 off- model grid cell center without any adaptation. shore masts no independent data were available for the usite raw WRF is the raw wind speed before the remodeling verification. process but made comparable to the site of a wind mea- The remodeling or optimization of the simulated surement according to step 2. wind speed ucell raw WRF basically consists of four steps: Meteorol. Z. (Contrib. Atm. Sci.) M. Schneider et al.: A wind atlas for Germany and the effect of remodeling 121 31, 2022 1. The first step is a correction of the annual cycle of the of global slope and offset parameter c0 to c4 “raw” wind speed data of the WRF simulation on the 3 × 3 km2 grid (ucell raw m = c0 + c1 · x1 + c2 · x2 + c3 · x3 + c4 · x4 (4.4) WRF ). Since ERA5 reanalysis data and consequently the raw data of the wind atlas from and similarly for b. The goal of this step is the calcu- the WRF simulation driven by ERA5 show a bias lation of global slope- and offset-parameter (c0–4 ) for of the annual cycle, a correction of this was imple- the investigated type of sub-grid information (x1–4 ) mented in the optimization approach. The monthly from the training data. Finally, the c1 till c4 parame- bias of 48 onshore stations was analyzed. The mean ters are in a range between zero and one. bias of each month is used to determine a correc- tion function for the annual cycle. Finally, a peri- With the global scaling parameter derived from the odic function with amplitude and phase derived from 26 training met masts scaling factors are applied the mean bias values generates time dependent and for the wind speed at each model grid cell taking periodic scaling factors for each 10-min time step. into account the respective sub-grid information. The This scaling factor time series is multiplied with the scaling factors are applied for each wind direction wind speed time series. This approach successfully sector and result in a corrected simulated wind data WRF for the 3 × 3 km grid cells. set ucell rem 2 minimized the monthly bias and did not impair other statistical parameters. The annual cycle correction is based on the bias of all 48 onshore met masts while 4. Site-specific time series usite WRF rem are calculated from the following remodeling steps use only 26 data sets the 3 × 3 km2 data ucell rem WRF by applying step (2) again for training. A separate annual cycle correction was but now on the remodeled data. derived for the offshore locations. The remodeling process changed the wind speed fre- 2. At this stage the simulated wind speed time series on quency distribution and the vertical wind speed profile at the 3 × 3 km2 grid ucell WRF raw are made comparable to some sites in a non-acceptable way. Constraints for the the observations with a correction based on elevation frequency distribution were applied in the remodeling and roughness according to equation (4.1) resulting approach to minimize the error. A 0.1 % bias threshold in usite WRF. raw was implemented for both tails of the frequency distribu- tion, so that the mi, j and bi, j were slightly rescaled in the 3. In a third step, both modeled usite WRF raw and observed remodeling process until this threshold has reached. The uobs wind speed data with 10-min resolution are par- analysis of different heights in the remodeling approach titioned into eight wind direction sectors to account resulted in a vertical correction of the global parame- for various surface characteristics depending on wind ters to reduce the bias at most of the heights. This verti- direction sector δ. A linear regression analysis cal correction is handled by a height dependent scaling uobs (δ) = m(δ) · usite factor derived as a mean from the training set. The raw WRF + b(δ) raw (4.2) and corrected vertical profiles for station 19 are shown follows for simulated and observed wind speed pro- in Fig. 4 exemplary and the wind speed frequency dis- viding regression coefficients for each measurement tribution in Fig. 5. site and 8 wind direction sectors, respectively. The results of the remodeling process were verified There are now 16 regression coefficients for each of with data from 48 onshore met masts. The raw data the 26 measurement sites (offset and slope of the re- ucell WRF raw show a clear positive bias of 0.7 m/s while a gression line for each of 8 wind direction sectors). slight overcorrection of the remodeling on the 3 × 3 km2 WRF ) results in a negative bias of −0.4 m/s. Site- grid (ucell Following Staffell and Pfenninger (2016) and rem Thøgersen et al. (2007) a multiple linear regression specific time series usite rem WRF show no mean bias and also analysis is performed separately for slope and off- the lowest bias variation (Fig. 6). set parameter taking into account sub-grid informa- tion. The sub-grid information is taken from the re- spective variables on a 1 × 1 km2 grid within each 4.2 Verification metrics 3 × 3 km2 cell for the height (x1 ), the height dif- The primary objective of the wind atlas is its use for ference between the respective 1 × 1 km2 grid and wind energy purposes. This basically requires frequency 3 × 3 km2 cell (x2 ), the latitude (x3 ) and the surface distributions of wind speed and -direction to character- roughness (x4 ). For each of the 26 sites (i) and for ize the local or areal wind climate. Additional informa- each of 8 direction sectors (j) this results in the equa- tion is provided by time series which reflect the temporal tion for m variability of the wind speed. The mean wind speed is denoted as ū and calculated mi, j = c0i, j + c1i, j · x1i, j + c2i, j · x2i, j as + c3i, j · x3i, j + c4i, j · x4i, j (4.3) 1 n and for bi, j accordingly. A multiple linear regression ū = ui (4.5) n i=1 is applied on these sets of equations resulting in a set 122 M. Schneider et al.: A wind atlas for Germany and the effect of remodeling Meteorol. Z. (Contrib. Atm. Sci.) 31, 2022 Figure 3: Bias in simulated monthly mean wind speed at 48 measurement locations. Right: Raw data, Left: after correction (line: mean bias, dots: individual measurements) And the mean bias is: 1 m Bias = Bias j (4.8) m j=1 with m being the number of verification sites. The stan- dard deviation of the bias is     m 1 σBias = (Bias j − Bias)2 (4.9) m j=1 Model data are instantaneous data every 10 minutes and observations are averages over a 10-min time interval. The remodeling procedure is based on these 10-min data while the verification is based on 1 h averages which should minimize the differences and is considered ad- equate for the verification of the diurnal cycle. A small shift in timing of changes in wind speed may lead to lower correlations, which, however, are insignificant for wind energy planning purposes (This paper does not Figure 4: Exemplary vertical profiles of wind speed before (red) and cover short-term forecasts). after (blue) the remodeling process for measurement site 19. 5 Results with n being the total number of hourly data. The Pear- son correlation coefficient R 5.1 Bias and correlation of hourly wind speed n Based on the 26 data sets involved in the remodeling (uobsi − ūobs )(uWRFi − ūWRF ) R = i=1 (4.6) process a general correction function was derived by a σuobs σuWRF multiple linear regression model for the slope and off- set parameter. This general correction function was then is a measure for the temporal interrelation between mea- applied to each of the raw data sets. These data sets are sured “obs” and simulated “WRF” data. considered “semi-independent” as their specific correc- The bias as the difference between mean values tion function was used to derived the overall mean cor- ūWRF − ūobs or their ratio ūWRF /ūobs points out a sys- rection function but the latter was applied for the final tematic deviation. correction. The same argument holds for the offshore met masts. The remaining 22 onshore data sets not used Biasū = ūWRF − ūobs (4.7) for the remodeling process are considered independent. Meteorol. Z. (Contrib. Atm. Sci.) M. Schneider et al.: A wind atlas for Germany and the effect of remodeling 123 31, 2022 Figure 5: Exemplary wind speed frequency distribution of the raw data (left) and after the remodeling process (right) for measurement site 19. The color shifts from blue to green for higher wind speed bins. than 5 percent for offshore conditions. The mean bias for all data sets before and after the remodeling process is shown in Fig. 7. Obviously, the bias is low for the off- shore met masts as there were only four data sets and the deviation of the respective scaling factor to the mean fac- tor is small in any case due to the similar surface charac- teristics. For the onshore sites the effect of the remodel- ing is remarkable. Most of the onshore data show a mean bias well within the range of ±5 % with some exceptions for very complex sites (Black Forest and Swiss Alps). It seems that even with a site-specific correction mesoscale data with 3 km resolution find their limit of reasonabil- ity at those sites. An error of 5 % in mean wind speed would result in an error for the wind turbine electric- ity production of approximately 10 % to 15 %, depend- ing on the wind speed frequency distribution and turbine power curve characteristics (a factor between 2 and 2.5 is a reasonable choice). From our long year consultancy experience an overall uncertainty of up to 15 % for the mean annual power production appears to be acceptable for the wind industry. Figure 6: Three box plots showing the biases of the comparison of raw wind atlas data, remodeled data (“cell”) and site-specific data The mean bias over all measurements (Fig. 8) is close (“site”) for the 48 onshore measurement sites for 100 m height. The to zero for all three data sets (onshore semi-independent, crosses mark the average of all stations (mean bias). The horizon- onshore independent, offshore) while the standard devi- tal bars in the box indicates the median and the box borders the ation and outliers (extremes of the bias) show a wider quartiles, respectively. Minimum and maximum are indicated by the spreading for the independent data set. whiskers. For a subset of the measurements the analysis was also performed for 60 m, 80 m, and 140 m height (Ta- ble 1). As expected, the bias and its standard deviation As a result of the remodeling process the bias of the decrease with increasing height as the influence of a ma- mean wind speed is reduced at all met masts. Before jor uncertainty factor, namely the parameterization of the remodeling process the model showed a positive bias the surface characteristic in the mesoscale model, is re- (model winds were too strong) for almost all onshore duced with increasing height. The bias reduction is high- met masts of up to 27 percent and a negative bias of less est from 60 m to 80 m and reduced also for the 80 m 124 M. Schneider et al.: A wind atlas for Germany and the effect of remodeling Meteorol. Z. (Contrib. Atm. Sci.) 31, 2022 Figure 7: Bias in mean wind speed for “raw data” (grey) and “remodeled site-specific data” divided into “semi-independent” (green) and “independent” (red) onshore data and the offshore met masts (blue) for 100 m height. Table 1: Mean bias and correlation coefficient in wind speed for measurements at different heights after the remodeling approach. Height [m] No. of Bias [%] Bias [m/s] R [%] sites 60 38 3.82 ± 7.18 0.15 ± 0.31 84.3 ± 4.8 80 45 0.50 ± 5.02 0.01 ± 0.26 85.9 ± 3.9 100 52 0.08 ± 4.38 0.00 ± 0.24 86.9 ± 3.6 140 17 0.10 ± 3.45 0.01 ± 0.21 87.4 ± 2.9 Ramboll 85–164 66 0.40 ± 4.5 — 87.3 ± 2.3 only slightly higher than our results as is the correlation coefficient with 87.3 %. Poulos and Stoelinga (2020) compared mesoscale model simulations with measurements at 105 sites with 2 to 10 meteorological towers each. No detailed infor- mation about model settings and measurement heights is available but for different model resolutions the bias Figure 8: Mean bias in wind speed, standard deviation and extreme for wind speed was found in the range of 1.9 % for 200 m values for “semi-independent” (green) and “independent” (red) on- resolution to 3.6 % for 600 m resolution. This data are in shore data and the offshore met masts (blue) for 100 m height. line with our statement that with downscaled mesoscale model data a bias for the mean wind speed below 5 % is reasonable. This complies with the uncertainty require- ments for the wind energy industry. to 100 m step. The correlation of the hourly time se- ries increases with height from 84.3 % at 60 m to 87.4 % at 140 m. In order to demonstrate the quality of the remodeled 5.2 Correction of the annual cycle mesoscale data set by a fully independent process, time series of the wind atlas were provided to an institution The effect of correcting the annual cycle was also inves- not linked to anemos GmbH for a given set of 66 sites tigated for the Ramboll data set. For 39 met masts the (Fig. 9). The analysis was performed on the whole by the mean monthly deviation, its standard deviation, and the market competitor Ramboll Deutschland GmbH. Only extremes are shown for the raw wind atlas data (no re- the site coordinates were known by the anemos team. modeling) and the time series after the remodeling pro- Ramboll’s analyses are in remarkable agreement with cess in Fig. 10. Like our analysis (Fig. 3) the correction the results shown before. With a mean bias of 0.4 % reduced the positive deviation in winter and the negative and a root mean square error of 4.5 % for the measure- deviation in summer and lead to a more realistic annual ments between 85 m and 164 m height the results are cycle of the wind speed. Meteorol. Z. (Contrib. Atm. Sci.) M. Schneider et al.: A wind atlas for Germany and the effect of remodeling 125 31, 2022 Figure 9: Bias of 66 measurements for “raw data” (red) and “remodeled site-specific data” (blue) based on Ramboll analysis. Figure 10: Monthly bias of 39 measurements for the Ramboll data, raw data (left) and corrected data (right). The boxes represent the standard deviation and the thin bars show extremes. 5.3 Frequency distribution of wind speed and The bias and correlation of the wind direction for the direction 52 measurement sites is shown in Fig. 12. The realistic representation of the wind direction is of paramount im- The wind speed frequency distribution may be approx- portance for the siting of wind turbines in a wind farm imated by a Weibull distribution with scale parame- as it governs the wind farm wake effect. Although wind ter A [m/s] and shape parameter k. The scale parameter vanes are usually aligned carefully small mounting er- is linked to the position of the maximum of the distribu- rors cannot be excluded. The overall bias in direction tion and is slightly larger than the mean wind speed. The shows a positive mean value of 1.5° with a standard de- shape parameter k is a measure of the width of the distri- viation of 3.0°. These deviations are covered in the un- bution i.e. the variance of wind speed. Fig. 11a shows the certainty of the wind vanes. No correction was made for absolute k values for the measurement and remodeled wind direction. The correlation reaches 95 % for more simulation results. Most of the k values of the 52 mea- than half of the stations. An example of the wind direc- surements are in a small box with a range of ±0.2. The tion distribution is given exemplary for 100 m height at simulated mean and 25/75 quantile are slightly shifted site 19 (Fig. 13). The excellent agreement between sim- to higher values, but the outliers (extremes of k) are in ulated and observed wind direction is considered an ef- a similar range. Fig. 11b displays the bias in k (simu- fect of the reduced roughness influence at 100 m height lation minus measurement) with a mean value of 0.06 and the dominating pressure gradient. and standard deviation of 0.1. The Weibull A parameter shows a zero biased distribution with a standard devia- 5.4 Comparison with NEWA and EMD data tion of 4.1 % (Fig. 11c). The Weibull distribution com- sets monly is considered to reasonably well represent the wind speed distribution for Northern Europe long year Two other data sets from model simulations are quite wind conditions. similar to the one discussed so far. There is the 126 M. Schneider et al.: A wind atlas for Germany and the effect of remodeling Meteorol. Z. (Contrib. Atm. Sci.) 31, 2022 (a) (b) (c) Figure 11: Consistency of the wind speed frequency distribution of 52 measurements at 100 m height represented by Weibull parame- ter k (a)–(b) and A (c). New European Wind Atlas NEWA, (NEWA, 2020; SST with 6-hourly assimilation), and simulation runs Gottschall et al., 2019; Witha et al., 2019; Gon- (anemos: multi-year simulations with 1 month spin-up, zález-Rouco et al., 2019, Dörenkämper et al., 2020) NEWA: 8 day simulations with 24 h spin-up) exist. No and the EMD-WRF EUROPE+ mesoscale data set such information is available for the EMD-WRF data (EMD, 2020a). The official wind atlas data sets have set. All three simulations, however, use the ERA5 re- been downloaded from the respective data base and no analysis as forcing data, a validated version of the WRF correction has been applied by us. Differences in model mesoscale model and a 3 km horizontal grid resolution. version (anemos: WRF 3.7.1, NEWA: WRF 3.8.1) and Use of the data for the wind energy industry is the main set-up (anemos: 50 vertical layers, NEWA: 61), parame- purpose of all data sets. It is not our intention to discuss terization schemes (anemos: Yonsei University (YSU) potential discrepancies between these data sets in terms planetary boundary layer scheme, NEWA: modified of model configuration or simulation approach but rather Mellor–Yamada–Nakanishi–Niino (MYNN) planetary highlight the effect of the remodeling process. We used boundary-layer scheme), boundary conditions (anemos: those 22 wind measurements which were not part of the ERA5 SST with hourly assimilation, NEWA: OSTIA determination of the remodeling parameters and calcu- Meteorol. Z. (Contrib. Atm. Sci.) M. Schneider et al.: A wind atlas for Germany and the effect of remodeling 127 31, 2022 Figure 12: Consistency of the wind direction of 52 measurements at a mean height of 98 m. Figure 13: The wind roses simulated (left) and observed (right) at site 19 at 100 m height. lated the overall bias and correlation for the 100 m height Data sets a)–d) are is the model output on the data from the different simulations. Figs. 14a and b show 3 km grid while for data set e) the respective 3 km data the bias and correlation in wind speed for the following have been itemized for the met mast site to be compared data sets: with. This process uses orography and land-use infor- mation to adapt the time series on the 3 km grid to the a) NEWA data surface characteristics in the immediate vicinity of the met mast. The adaptation is based on findings from the b) EMD-WRF Europe+ data remodeling process. At this point it must be emphasized that the general findings from the remodeling process c) anemos raw data without any remodeling or adapta- using 26 met masts are applied here to an independent tion process data set consisting of the remaining 22 measurements. d) anemos wind atlas after remodeling (= cell) With a mean coefficient of 81.2 % the correlation of hourly data is lowest for the NEWA data followed by e) anemos wind atlas after remodeling site specific the EMD-WRF Europe+ data with 85.2 % and 86.8 % (= site). for the three anemos data sets. Because the ERA5 re- 128 M. Schneider et al.: A wind atlas for Germany and the effect of remodeling Meteorol. Z. (Contrib. Atm. Sci.) 31, 2022 Figure 14: Boxplots of the bias (left) and correlation (right) in wind speed for 22 independent sites. Compared are three versions of the anemos wind atlas, namely the raw data (black) and the corrected data on a 3 km grid (orange) and the site specific data (green) with the NEWA (blue) and EMD-WRF Europe+ wind (red) atlas. analysis is the forcing for all simulations the difference of wind simulations to the extent possible is of particular in the correlation might be due to various WRF model importance for applications in the wind energy sector. configurations or simulation realizations (e.g. nudging This paper compares wind simulations and observa- technique and time step, length of simulation period). In tions at 118 met masts and describes a remodeling ap- addition, the anemos and EMD-WRF EUROPE+ hourly proach to reduce the inherent difference between model data represent the average of 10-min instantaneous val- output and observations. The mesoscale model WRF is ues whereas the NEWA data are based on 30-min in- used to simulate the wind conditions over Germany with stantaneous values. An investigation into the sensitivity a horizontal resolution of 3 × 3 km2 and a temporal res- of the correlation coefficient on model settings is not olution of 10-min. The analysis, however, is made with our intent. However, a difference of 5.6 % in the cor- aggregated hourly data. ERA 5 reanalysis data are used relation seems to be rather large keeping in mind that for forcing purposes. e.g. the electricity price is dealt on the Stock Exchange In order to minimize the inherent difference between half-hourly. The picture is quite different for the bias. simulations and observations a remodeling approach is All three raw data sets without any remodeling or adap- applied which corrects simulated wind speed time series tation have a mean bias in a similar range between +0.6 taking into account differences in surface characteristics and +0.8 m/s. After remodeling the bias of the anemos (height and roughness variations) between model grid wind atlas is overcompensated with a negative value cell and observation site. The remodeling process com- of −0.3 m/s. This is explained by the fact that met masts prises 4 main steps followed by a site-specific adapta- for wind energy purposes commonly are positioned at tion: exposed sites compared to the average height of a model grid cell. The site-specific mean wind speed corrects for a) Correction of the annual cycle. this overestimation and finally shows a mean bias of 0 with an extend of 0.3 m/s for the 25 % to the 75 % quar- b) Correction for height difference between model grid tile (50 % of the data within this box). Also, the spread cell and measurement point. The speed-up factors are of the bias is reduced compared to the non-remodeled deduced from CFD model simulations. data. c) Derivation of regression parameter (slope and offset) between simulated and observed wind speed for 8 di- rection sectors at 26 measurement sites. Multiple lin- 6 Conclusion and outlook ear regression analysis separately for slope and off- set to derive a global parameter set. Remodeling the This paper is based on the idea that even small uncer- mesoscale wind speed time series resulting in a re- tainties in wind speed may result in large uncertainties conditioned wind atlas. for the wind potential and wind turbine power produc- tion and may imply an increased risk for investments in d) Site-specific adaptation of mesoscale wind speed wind energy projects. Hence, reducing the uncertainty time series. Meteorol. Z. (Contrib. Atm. Sci.) M. Schneider et al.: A wind atlas for Germany and the effect of remodeling 129 31, 2022 The “raw” wind atlas data (before step a), the re- Carvalho, D., A. Rocha, M. Gómez-Gesteira, C. Silva- conditioned wind atlas data (after step c) and the site- Santos, 2014: WRF wind simulation and wind energy pro- specific data (after step d) were compared to observa- duction estimates forced by different reanalyses: Comparison tions at 100 m height. The “raw” data showed a positive with observed data for Portugal. – Appl. Energy 117, 116–126. C3S – Copernicus Climate Change Service, 2017: ERA5: bias of up to 27 % for the mean wind speed onshore and Fifth generation of ECMWF atmospheric reanalyses of the a negative bias of less than 5 % for the offshore towers. global climate. – Copernicus Climate Change Service Cli- The remodeled data (after step d) showed a mean nega- mate Data Store (CDS), date of access. https://cds.climate. tive bias for onshore sites. This is explained by the fact copernicus.eu/cdsapp#!/home. that measurements for the wind industry tend to be per- Deppe, A.J., W.A. Gallus Jr., E.S. Takle, 2013: A formed at exposed sites. This effect was corrected for WRF ensemble for improved wind speed forecasts at tur- with the site-specific adaptation. bine height. – Wea. Forecast. 28, 212–228, DOI:10.1175/ WAF-D-11-00112.1. Accounting for all 118 met masts and a height range Dörenkämper, M., B.T. Olsen, B. Witha, A.N. Hahmann, between 80 m and 160 m the mean bias in wind speed N.N. Davis, J. Barcons, Y. Ezber, E. García-Bustamante, is roughly within a range of 0.08 % and 0.50 % with J.F. González-Rouco, J. Navarro, M. Sastre-Marugán, a standard deviation below 5 %. A 5 % uncertainty in T. Sile, W. Trei, M. Žagar, J. Badger, J. Gottschall, wind speed is in the range expected today for financial J.S. Rodrigo, J. Mann, 2020: The Making of the New Euro- considerations in the wind energy industry. pean Wind Atlas – Part 2: Production and evaluation. – Geosci. Model Dev. 13, 5079–5102. The wind speed frequency distribution expressed by EMD, 2020a: EMD-WRF Europe+ MesoScale Data Set. – Weibull parameter showed a bias close to zero with a https://www.emd.dk/data-services/mesoscale-time-series/ standard deviation of roughly 4 % for the scale parame- pre-run-time-series/emd-wrf-europe-mesoscale-data-set/. ter A and a mean bias less than 0.1 for the form param- EMD, 2020b: windpro Technical Note: Validation of EMD-WRF eter k. The small mean bias of 1.5 deg in wind direction EUROPE+ (ERA5) mesoscale dataset, EMD Interna- is not corrected for as it might be in the range of the tional A/S. – http://help.emd.dk/knowledgebase/content/ mounting uncertainty. Technotes/TechnicalNote7_EMDWRF_EUROPE_Validation. pdf. Comparison of the NEWA and EMD-WRF Europe+ Farr, T.G., P.A. Rosen, E. Carop, R. Crippen, R. Duren, wind atlas data with observations showed a mean bias S. Hensley, M. Kobrick, M. Paller, E. Rodriguez, between +0.6 m/s and +0.8 m/s similar to our “raw” data. L. Roth, D. Seal, S. Shaffer, J. Shimada, J. Umland, The remodeling and site-specific adaptation process sig- M. Werner, M. Oskin, D. Burbank, D. Alsdorf, 2007: The nificantly reduced the mean bias to 0.03 m/s with a stan- Shuttle Radar Topography Mission. – Rev. Geophys. 45, 2, dard deviation of 0.27 m/s. DOI:10.1029/2005RG000183.; Fernández-Gonzales, S., M.L. Martin, E. Garcia-Ortega, A method was presented to minimize the discrepancy A. Merino, J. Lorenzana, J.L. Sanchez, F. Valero, between wind simulations and measurements by a re- J.S. Rodrigo, 2018: Sensitivity analysis of the WRF model: modeling process and a site-specific adaptation. The re- Wind-resource assessment for complex terrain, J. Appl. Cli- sulting uncertainty is in a the range which is acceptable matol. 57, 3, 733–753, DOI:10.1175/JAMC-D-17-0121.1 by the wind industry even for financial considerations. Giannakopoulou, E.-M., R. Nhili, 2014: WRF Model Such data sets will form an essential contribution to the Methodology for Offshore Wind Energy Applications. – wind farm development process. The anemos wind at- Adv. Meteor., published online, 319819, DOI:10.1155/ 2014/319819. las data sets are commercially available via the awis González-Rouco, J.F., E. Garcia-Bustamante, A.N. Hah- (anemos wind information system, awis.anemos.de) on- mann, I. Karagili, J. Navarro, B.T. Olsen, T. Sile, line tool. B. Witha, 2019: NEWA. – Report on uncertainty quantifica- tion. – https://zenodo.org/record/3382572#.XnHzAHdFxPY. Gottschall, J., M. Dörenkämper, B. Lange, P. Kühn, Acknowledgements D. Callies, B. Witha, G. Steinfeld, S. Voss M. Kühn, D. Heinemann, 2019: NEWA – New European Wind Atlas This paper is based on data provided by wind farm own- Joint Programme. – Final report, personal communication. ers and developers. The study was partly supported by Hahmann, A.N., T. Sile, B. Witha, N.N. Davis, M. Dören- N-Bank Niedersachsen and the Federal Ministry for the kämper, Y. Ezber, E. García-Bustamante, J.F. González- Environment, Nature Conservation and Nuclear Safety Rouco, J. Navarro, B.T. Olsen, S. Söderberg, 2020: The represented by the German Environment Agency of the Making of the New European Wind Atlas, Part 1: Model Sensitivity. – Geosci Model Dev. 13, 5053–5078, 10.5194/ Federal Republic Germany. The final form of this paper gmd-13-5053-2020. benefitted from the helpful comments and suggestions Howard, T., P. Clark, 2007: Correction and downscaling of of the two anonymous reviewers. NWP wind speed forecasts. – Meteor. Appl. 14, 105–116, DOI:10.1002/met.12. Hu, X-M., J.W. Nielsen-Gammon, F. Zhang, 2010: Evaluation References of Three Planetary Boundary Layer Schemes in the WRF Model. – J. Appl. Meteor. Climatol. 49, 1831–1843, DOI: Badger, J., H. Frank, A.N. Hahmann, G. Giebel, 2014: Wind- 10.1175/2010JAMC2432.1. climate estimation based on mesoscale and microscale model- Hu, X-M., P.M. Klein, M. Xue, 2013: Evaluation of the updated ing: statistical-dynamical downscaling for wind energy appli- YSU planetary boundary layer scheme within WRF for wind cations. – J. Appl. Meteor. Climatol. 53, 8, 1901–1919, DOI: 10.1175/JAMC-D-13-0147.1. 130 M. Schneider et al.: A wind atlas for Germany and the effect of remodeling Meteorol. Z. (Contrib. Atm. Sci.) 31, 2022 resource and air quality assessments – J. Geophy. Res.: At- bution models. – Science Direct, Energy Reports 6, Suppl. 6, mos., 118, 10490–10505, DOI:10.1002/10.1002/jgrd.50823. 250–2676, DOI:10.1016/j.egyr.2020.08.048. IEC, 2017: International Electrotechnical Commission IEC Skamarock, W.C., J.B. Klemp, J. Dudhia, D.O. Gill, 61400-12-1:2017, Wind energy generation systems – D.M. Barker, M.G. Duda, X.-Y. Huang, W. Wang, Part 12-1: Power performance measurements of electricity J.G. Powers, 2008: A description of the advanced re- producing wind turbines. – International Electrotechnical search WRF Version 3. – NCAR Technical Note NCAR/ Commission (IEC). TN-475+STR, June 2008. Keil, M., M. Bock, T. Esch, A. Metz, S. Nieland, Siuta, D., G. West, R. Stull, 2017: WRF hub-height wind A. Pfitzner, 2010: CORINE Land Cover Aktualisierung forecast sensitivity to PBL scheme, grid length, and initial 2006 für Deutschland. Abschlussbericht zu den F+E Vorhaben condition choice in complex terrain. – Wea. Forecast. 32, UBA FKZ 3707 12 200 und FKZ 3708 12 20. – Deutsches 493–509, DOI:10.1175/WAF-D-16-0120.1. Zentrum für Luft- und Raumfahrt e. V., Deutsches Fernerkun- Staffell, I., S. Pfenninger, 2016: Using bias-corrected re- dungsdatenzentrum Oberpfaffenhofen, Januar 2010. analysis to simulate current and future wind power output. – Lindenberg, J., H.-T. Mengelkamp, G. Rosenhagen, 2012: Energy 114, 1224–1239, DOI:10.1016/j.energy.2016.08.068. Representativeness of near surface wind measurements from Thøgersen, M.L., M. Motta, T. Sørensen, P. Nielsen, 2007: coastal stations at the German Bight. – Meteorol. Z. 21, 1, Measure-Correlate-Predict Methods: Case Studies and Soft- 99–106. ware Implementation. – presented at European Wind Energy Mengelkamp, H.-T., 1999: Wind climate simulation over Conference, Milan. complex terrain and wind turbine energy output estima- Weiter, A., M. Schneider, D. Peltret, H.-T. Mengelkamp, tion. – Theor. Appl. Climatol. 63, 129–139, DOI:10.1007/ 2019: Electricity production by wind turbines as a means for s007040050098. the verification of wind simulations. – Meteorol. Z. 28, 1, Mengelkamp, H.-T., H. Kapitza, U. Pflüger, 1997: 69–77, DOI:10.1127/metz/2019/0924. Statistical-dynamical downscaling of wind climatologies. – Wieringa, J. 1980: Representativeness of wind observations at J. Wind Eng. Ind. Aerodyn. 67–68, 449–457. airports. – Bull. Amer. Meteor. Soc. 61, 962–971. METEODYN WT, 2015: Technical note-meteodyn WT. – Tech- Wieringa, J., 1996: Does Representative Wind Information Ex- nical report, https://meteodyn.com/. ist? – J. Wind Eng. Ind. Aerodyn. 65, 1–12. NEWA, 2020: About the New European Wind Atlas, https://map. Witha, B., A.N. Hahmann, T. Sile, M. Dörenkämper, neweuropeanwindatlas.eu/about. Y. Ezber, E. Garcia-Bustamante, J. Fidel Gonzalez- Olsen, B.T., A.N. Hahmann, A.M. Sempreviva, J. Badger, Rouco, G. Leroy, J. Navarro, 2019: NEWA, Report on H.E. Jørgensen, 2017: An intercomparison of mesoscale WRF model sensitivity studies and specifications for the models at simple sites for wind energy applications. – eWind mesoscale wind atlas production runs. – https://zenodo.org/ Energ. Sci. 2, 211–228, www.wind-energ-sci.net/2/211/2017/ record/2682604#.XnH3GohKiHt. DOI:10.5194/wes-2-211-2017. WRF, 2017: User’s Guides for the Advanced Research Petersen, E.L., I. Troen, S. Frandsen, K. Hedegaard, 1981: WRF (ARW) Modeling System, Version 3. – WRF Danish Windatlas: A rational method for wind energy siting, user page, http://www2.mmm.ucar.edu/wrf/users/docs/user_ Risø-R-428. – Risø National Laboratory, Denmark, 229 pp, guide_V3/contents.html. ISSN 0106-2840. Yang, Q., L.K. Berg, M. Pekour, J.D. Fast, R.K. New- Poulos, G.S., M. Stoelinga, 2020: Mesoscale and Microscale som, M. Stoelinga, C. Finley, 2013: Evaluation of Modelling for Wind Energy Applications. – Windtech Inter- WRF-Predicted Near-Hub-Height Winds and Ramp Events national. 16, 6–8. over a Pacific Northwest Site with Complex Terrain. – Serban, A., L.S. Paraschiv, S. Paraschiv, 2020: Assessment J. Appl. Meteor. Climatol. 52, 8, 1753–1763, DOI:10.1175/ of wind energy potential based on Weibull and Rayleigh distri- JAMC-D-12-0267.1.