Wind speeds for a nominal height of 10 m and from the lowest model level (∼70 m above ground level) from the Rossby Center regional climate model (RCM) (RCA3) run at four resolutions between approximately 50 × 50 km and 6 × 6 km are analyzed to assess the effect of model resolution on wind climates. The influence of model resolution in this topographically simple subdomain of northern Europe is more profound in the wind extremes than in the central tendency. The domain-averaged mean wind speed at 10 m increases by 5% as the resolution increases from 50 to 6.25 km, while the 50 year return period wind speed and wind gust at this height increase by over 10% and 24%, respectively. Larger changes are observed in these wind speed metrics at the lowest model level as model resolution increases (∼+10% in the mean and ∼+20% in the 50 year return period wind speed). These differences are of similar magnitude to the climate change signal in extreme wind events derived in prior research and may have implications for climate change risk and vulnerability analyses. Output from the lowest model level indicates some evidence for increased variability at synoptic and meso-α time scales with increased model resolution, but the effect is nonlinear. Furthermore, analysis of power spectra of grid cell average and tile fraction wind speeds at 10 m does not support the assertion that increased model resolution increases model skill at synoptic and meso-α time scales relative to in situ observations.
 Global and regional climate models are being run at increasingly fine horizontal and vertical resolution with the goal of increased fidelity of climate simulations [Leung et al., 2003, 2006]. The implicit assumption is that increased resolution will yield more accurate simulations due to the reduction in the “smearing out” of subgrid-scale phenomena and increasing coverage of wave number space, particularly in regions that are topographically complex and/or exhibit heterogeneous land use. However, increased resolution need not per se be associated with increased model skill or a reduction in model bias [Wang et al., 2004], and it comes at substantial computational cost. Even if some increase in resolution yields increased fidelity, model skill will likely exhibit diminishing returns as resolution is further increased because many models run in climate mode are hydrostatic and all contain parameterizations. Further, simulations with an identical model run at different resolution may lead to different behavior due to nonlinearities in the parameterizations [Christensen et al., 2007], and any change in model behavior and performance may be parameter specific. One study of the effect of increasing horizontal resolution in the HIRHAM4 regional climate model (RCM) [Aðalgeirsdóttir et al., 2009] showed that while precipitation over Denmark shows a weak, but systematic, increase with increasing model resolution from 50 km to 25 km to 12 km, the converse is true for temperature [van Roosmalen et al., 2010].
 Generally, studies focused on the influence of model resolution on thermal and hydrologic regimes have found increased skill with increased resolution [Hohenegger et al., 2008; Kusunoki et al., 2006; Whetton et al., 2001]. For example, 12 km resolution model simulations with HIRHAM4 over Denmark showed greater accord with observed monthly precipitation, particularly during the cold season, than simulations at 25 or 50 km grid spacing [van Roosmalen et al., 2010]. However, the influence of spatial discretization is likely to vary with parameter and spatial scale. Therefore, there is a need to further evaluate the impact of spatial resolution on climate simulations [Leung et al., 2003] and to extend the suite of variables considered beyond temperature and precipitation.
 Modeled wind speeds are naturally a function of model resolution due in part to spectral truncation [Larsén and Mann, 2006]. Flow variations at high frequencies are inherently underrepresented in RCM simulations because of the spatial and temporal averaging and temporal discretization (and disjunct reporting of wind speeds). But energy (variance) manifest at mesoscales and microscales (i.e., at the spectral tail), which is damped or absent in RCM simulations, significantly contributes to intense and extreme wind speeds [Larsén and Mann, 2006].
 Station observations of gust wind speeds over Germany generally indicate higher agreement with 10 km resolution simulations conducted with the hydrostatic REMO model [Jacob and Podzun, 1997] than two simulations (which differ only in the lateral boundary conditions) at 18 km resolution with the nonhydrostatic COSMO climate local model (CCLM) [Jaeger et al., 2008]. Kunz et al.  found a negative bias (of about 10–30%) in 2 and 10 year return period gusts in all model-derived estimates relative to extreme gust values extrapolated from in situ observations, a high degree of similarity for the two CCLM simulations and that the degree of agreement with the station observations exhibited a dependence on station elevation. A major conclusion of Kunz et al. [2010, p. 919] was that “Higher spatial resolution of the models, such as the 10 km of REMO, permits better representation of the main orographic structures, thus yielding higher spatial variability of the gusts over complex terrain. Reliable representation of the local storm climate, however, requires even higher resolution.” However, the relationship between wind speed metrics and model resolution is not linear. For example, the nonhydrostatic COSMO-CLM model applied at 1.3 and 5 km showed the mean wind speed in southwestern Germany was biased low at the lowest model level (∼50 m above the ground level) in the higher-resolution simulations likely due to the increase in model terrain complexity and friction velocity [Knote et al., 2010].
 Given the importance of synoptic and meso-α scale variability (i.e., motions having temporal scales ∼1 to 10 days and covering spatial scales ∼200 km to a few thousand kilometers [American Meteorological Society, 2000]) to the wind and wave resources in northern Europe [Weisse and von Storch, 2010] and the occurrence of intense and extreme wind speeds over Europe to multiple economic sectors (including the reinsurance and insurance industry [Pinto et al., 2010; Schwierz et al., 2010]), there is a need to assess the ability of RCMs to simulate wind climates. Herein we present analyses of output from a RCM run at four different resolutions – the coarsest typifies the majority of current RCM applications at the continental scale, i.e., ∼50 × 50 km, the finest is ∼6 × 6 km. The objectives of this work are (1) to examine the power spectra for wind speeds and wind gusts to determine which frequencies are affected by changes in model resolution and which resolution is associated with power spectra that exhibits greatest similarity with those derived from in situ observational data and (2) to quantify the impact of model resolution on the magnitude of extreme wind speeds and wind gusts derived from RCM output given the importance of extreme events to the impacts of climate change [Della-Marta et al., 2009; Knote et al., 2010; Schwierz et al., 2010].
 We examine the influence of grid resolution on wind speeds, extreme wind speeds, and wind gusts using output from the Rossby Centre version 3 (RCA3) RCM (a detailed description of the model is given by Samuelsson et al. ). RCA3 uses a mosaic tile approach to characterize subgrid-scale atmosphere-surface exchange. The 10 m wind speeds from RCA3 are the results of weighted averaging over all fractional surface types (or tiles) that the grid box contains. The tile fractions are characterized as open land, forested land, or water (lake or ocean), where the water can be ice covered while the land fractions can also be snow covered [He et al., 2010]. Each fractional type has a specific roughness length, and the surface energy balance calculations are conducted separately before a weighted average is applied according to the fractional areas of the tiles to derive the grid cell average wind speeds. RCA3 also includes a physically based parameterization of wind gusts based on the assumption that wind gusts are a product of deflection of air parcels from high in the planetary boundary layer (PBL) down to the surface [Brasseur, 2001; Brasseur et al., 2002].
 We present RCA3 simulations for 1987–2008 run with grid resolutions of (1) 0.44 × 0.44° (50 km), (2) 0.22 × 0.22° (25 km), (3) 0.11 × 0.11° (12.5 km), and (4) 0.055 × 0.055° (6.25 km), derived by nesting RCA3 in lateral boundaries supplied by ERA-40 for January 1987 to August 2002 and then from the ECMWF operational analysis from September 2002 to the end of the simulation period. In accord with the relatively similar storm climates in ERA-40 and the ECMWF operational analysis, the change of lateral boundary conditions appears to have relatively little impact on the wind climates. For example, annual maximum wind speeds from the first and second periods are not significantly different. All simulations used the same geographical domain, the same RCA3 model version with the convection scheme turned on, and 24 unequally spaced hybrid terrain-following (sigma-pressure coordinate) levels. The lowest model level has a height of ∼70 m above ground level (agl), and all model simulations have a top at ∼10 hPa. The time steps used for the simulations are as follows: 50 km = 30 min, 25 km = 20 min, 12.5 km = 10 min, and 6.25 km = 7.5 min.
 Although RCA3 simulations used herein cover the entirety of Europe [Samuelsson et al., 2011], we focus here on a subdomain centered over Denmark (Figure 1). This domain was selected to avoid regions with strong thermotypographic forcing and to focus instead on a wind climate that is strongly linked to synoptic and meso-α scale phenomena and that thus should be well described by a hydrostatic model applied at the grid resolutions used herein. Further, this domain is focused on a subregion within Europe that was projected to exhibit a disproportionately large increase in insurance losses associated with extreme wind storms in a prior analysis such as that using an operational insurance model and RCM simulations from the PRUDENCE project [Schwierz et al., 2010].
 To assess the impact of model resolution on extreme wind speeds, we compute 50 year return period extreme sustained (U50yr) and gust (Gust50yr) wind speeds. Estimates of these metrics are derived by fitting of annual maxima from each simulation and grid cell to a double exponential cumulative probability distribution. Then the method of moments is applied to obtain the Gumbel parameters [Abild et al., 1992], which are used to extrapolate the 50 year return period values for each model grid cell in each simulation. Thus, the 50 year return period wind speeds and gusts are derived from
where UT is the wind speed for a given return period (T = 50 years) and the distribution parameters (α and β) are derived by the method of moments wherein they are determined from the mean and variance of the time series of annual maximum values. Assuming a Gaussian distribution of UT, then 95% of all realizations will lie within ±1.96σ of the mean, and thus, σ(UT) can be used to provide 95% confidence intervals on the estimates of extreme winds with any return period:
and γ is Euler's constant.
 For the estimates derived herein, the domain-averaged 95% confidence interval computed using (2) applied to each grid cell specific U50yr value at 10 m is approximately ±14%. To examine whether the estimates of long-return period extreme wind speeds and gusts derived from the higher-resolution runs exhibit features not present in the 50 km simulations (i.e., that the fields differed beyond simply exhibiting greater or lesser pixelation), the spatial fields of U50yr and Gust50yr from the three higher-resolution simulations are regridded back onto the 50 km grid. This regridding is necessary because these four simulations do not share the same 50 km grid, and regridding was performed as follows: minimizing spherical distance to the 50 × 50 km grid centroids, grid cells within the higher-resolution simulations are allocated to the 50 × 50 km grid, and the values within each 50 km grid are averaged to generate spatial fields that are compared in terms of the correlation, spatial variability, and root-mean-square difference (RMSD) with the extreme value estimates derived from the simulation at 50 km. The resulting metrics are then presented using Taylor diagrams [Taylor, 2001].
 We also present power spectra of the horizontal wind speeds and daily maximum wind gust derived from the four RCA3 simulations to examine which (if any) frequencies are affected by changes in model resolution and whether they differ in the wind speed metrics (3-hourly disjunct mean wind speeds and daily maximum wind gusts at 10 m and the 6-hourly disjunct wind speeds at the lowest model level). The wind speed power spectra from the RCA3 simulations at 10 m are compared with those derived from five long time series of in situ observations to determine which resolution exhibits greatest similarity with station data. These observational data are derived from the National Climate Data Center (NCDC) DS3505 hourly surface data set. The stations used are (listed from west to east) Stavanger/Sola (USAF site 014150), Flyvestation Aalborg (060300), Copenhagen/Kastrup (061800), Jönköping/Axamo (025500), and Hammer Odde (on the northern tip of Børnholm, 061930) (see Figure 1). Data records from all five stations have <2% missing data in the period 1987–2008, making them highly amenable to the data analysis performed herein, but it should be acknowledged that these stations are in complex land cover locations (e.g., four of the five are coastal). The observational time series were resampled at a 3-hourly interval (to match the model output), and the missing data values, which were randomly distributed throughout the records, were replaced by the mean value prior to analysis. This replacement has negligible impact on the moments of the distribution. In the spectral analysis all time series (observational and model derived) are detrended prior to application of fast Fourier transform to minimize aliasing effects, and the results are plotted in semilog form so that the area under any portion of the curve is proportional to the variance. Estimated extreme wind speeds from the RCA3 output are also evaluated relative to values for three stations in Germany: Heligoland, Rostock, and Teterow (see Figure 1) as reported by Kunz et al. .
 There is considerable uncertainty in RCM-derived 10 m wind speeds due to issues associated with extrapolating winds to 10 m from the lowest model level. Thus, in addition to analysis of winds at a nominal height of 10 m and comparison with observations from that height, we also present analyses of wind speeds derived from the lowest model level.
 Although the timing and occurrence of modeled intense/extreme wind speeds at a height of 10 m at each of the five observational stations are consistent between the simulations at the four resolutions, the grid cell specific annual maximum wind speeds (used to derive the U50yr estimates) from each of the RCA3 simulations do not exhibit a clear signal of increased values with increased resolution (Figure 2). This feature is due in part to the complex land cover in the coastal locations for which data are available. This heterogeneity in the land surface means that as the resolution changes, the tile fraction of the grid cell containing the station that is characterized as open water, open land, and forested also changes. The associated changes in roughness length and other parameters that dictate atmosphere-surface exchange have a substantial impact on the 10 m wind speeds.
 Consistent with the spectral truncation discussed in further detail below and the caveats and limitations of extreme value extrapolation described elsewhere [Cook, 1986, 2004], it should be emphasized that the U50yr estimates for a nominal height of 10 m derived from RCA3 output and shown in Figure 1 are not representative of the “true” extreme wind speeds (or the extreme values that derived from data time series as measured at observational stations). The 3-hourly disjunct time series from the five stations (Stavanger, Aalborg, Copenhagen, Jönköping, and Hammer Odde) yield U50yr estimates for a nominal height of 10 m agl (listed from west to east) of 28.6, 28.8, 28.7, 19.5, and 37.3 m s−1, respectively, which bound similar estimates from the Westermarkelsdorf site on the north German coast (U50yr = 26.2 m s−1) [Pryor et al., 2012]. Spatial averaging (and other processes) causes the RCA3 grid cell derived estimates to deviate by up to ±10 m s−1 from values obtained from the station time series. Conversely, 50 year return period gust estimates for three observing stations in Germany [Kunz et al., 2010] exhibit relatively good accord with estimates from RCA3 applied at 6.25 km resolution. The values for the Heligoland, Rostock, and Tetrow stations derived from observations are approximately 43, 40 and 38 m s−1, respectively [Kunz et al., 2010], while the RCA3-derived estimates for grid cells containing those observing stations are 51.9, 39.5, and 36.8 m s−1.
 It should be acknowledged that there is some evidence of nonlinearity in the Gumbel plots both of annual maximum sustained and gust 10 m wind speeds (Figure 2). At some locations and some model resolutions the extreme wind speed distributions thus exhibit evidence for a mixed wind climate [Cook et al., 2003] (as indicated by curvature in the Gumbel plots), and hence extrapolation of wind speeds with long return periods (e.g., 50 years) is subject to considerable uncertainty.
 Consistent with the findings of earlier studies [Kunz et al., 2010], spatial patterns of annual mean wind speed (〈U〉), U50yr and Gust50yr generally indicate increasing complexity of the spatial patterns and increasing values with model resolution (Figures 1 and 3 and Table 1). Accordingly, the domain-averaged U50yr for a height of 10 m increases from 17.33 to 19.15 m s−1 as the model resolution increases from 50 to 6.25 km. Equivalent comparisons for Gust50yr at the two extremes of the four resolutions are 33.96 to 42.18 m s−1, respectively (Table 1). Thus the response of the three wind climate metrics to changes in model resolution is larger for the descriptors of the tail of the probability distribution. The domain- averaged 10 m wind speed increases by 5% as the resolution increases from 50 to 6.25 km, U50yr increases by over 10%, and the 50 year return period wind gust increases by 24%. Some fraction of these differences is attributable either to the caveats given above or to differences in the simulation time step (and thus implicit) temporal averaging period. On the basis of prior analyses of observed wind speeds in this region, the difference in extreme winds computed from 30 min averages versus 7.5 min averages is approximately 5% [Larsén and Mann, 2006]. The residual differences (i.e., changes beyond 5%) are likely to reflect differences in the simulated flow fields that are not merely a product of temporal sampling.
Table 1. Domain-Averaged Mean Annual Wind Speed (〈U〉) and 50 Year Return Period Extreme Wind Speed (U50yr) and Wind Gust (Gust50yr) From the Four RCA3 Simulations at Different Resolutions Based on Output From 1987 to 2008a
Height Resolution (km)
Lowest Model Level
〈U〉 (m s−1)
U50yr (m s−1)
Gust50yr (m s−1)
〈U〉 (m s−1)
U50yr (m s−1)
Domain is shown in Figure 1. The two sets of results indicate values for a nominal height of 10 m agl and computed at the height of the lowest model layer (∼70 m agl).
 To further examine the degree to which the spatial fields of U50yr and Gust50yr from the 50, 25, 12.5, and 6.25 km simulations differ, estimates of these extreme value metrics for the three higher-resolution simulations were regridded onto the 50 km grid, and the results were compared to extreme value estimates derived from the lowest-resolution simulations. The results indicate that even when the spatial fields are “degraded” to 50 km, the estimates of long return period extreme events derived from the 25 km simulation resolution appear to exhibit spatial variability not present in the 50 km results (i.e., σm/σr values are above 1, Figure 4). However, the regridded fields from the two highest-resolution simulations do not exhibit substantial additional variance. This might imply that the transition to resolutions higher than 25 km is not introducing important new features to the 10 m flow fields during situations associated with extreme wind events.
 Power spectra of 3-hourly model time step “instantaneous” wind speeds at all four resolutions exhibit some features present in the observations particularly up to the Nyquist frequency of the large-scale (synoptic and meso-α scale) forcing (approximately 6–12 h). In all cases, except the 50 km resolution simulation at Stavanger and perhaps the 12 km simulation at Copenhagen, the variance in the synoptic-scale frequencies (i.e., the peak in variance at a period of about 5 days) tends to exhibit a negative bias in the RCA3 simulations. Nevertheless, the model derived spectra show clear peaks in the variance associated with biannual, annual (i.e., the peak near 3 × 10−3 d−1), seasonal and synoptic to meso-α scale frequencies that also dominate the observational time series (Figure 5). However, there is no consistent signal that the variance at these frequencies is related to model resolution nor indeed is there clear evidence for consistency in the resolution for which greatest accord is found with the observations. This may be due in part to the selection of the grid cell which encompasses the observational station. For example, at these coastal stations the amount of land area versus water surface in the selected grid cell differs with model resolution.
 The tile fraction wind speeds within each grid cell were saved for each of the simulations and were analyzed to establish if there was a well-defined resolution signature in the spectral features from the tile fraction that best represents the fetch of the sites from which the in situ observations are available. The results indicate that the spectra for those tile fractions do indeed more closely approximate those derived from the observations but also that there is not a clear consistent signal for increased variance at the synoptic and meso-α temporal (and spatial) scales with increased resolution (Figure 6). For example, in the case of the Copenhagen grid cell, there is higher energy (i.e., variance) in the water tile fraction wind speeds in frequencies peaking between 0.1 and 1 d−1 in simulations conducted at 50 km than at any of the other resolutions, while the spectra for the 25, 12.5, and 6.25 km resolutions are virtually indistinguishable.
 Similar dominant frequencies of variation to those for wind speeds are also evident in the model derived wind speed gust power spectra, though the gust time series exhibit a greater dominance of the synoptic frequencies (peaking at periods of approximately 4–7 days (f ∼ 0.14 to 0.25 d−1), Figure 7). There is stronger evidence for an increase in variance at synoptic and meso-α time scales in wind gusts with model resolution than in the mean wind speeds. In all cases the 6.25 or 12.5 km simulations exhibit greater variance in the synoptic-scale frequencies, though again the tendency toward higher variance is not entirely systematic across the five locations. The variance associated with the diurnal cycle and frequencies above 1 d−1 (i.e., associated with horizontal motions that characterize the meso-α, meso-β, and meso-γ scales, i.e., 2 km to a few hundred kilometers [American Meteorological Society, 2000]) are underestimated in the model (both for wind speeds and wind gusts) but in the case of wind gusts show consistent increases for frequencies above 0.1 d−1 with increasing model resolution.
 To examine the degree to which the inferences drawn above are specific to a nominal height of 10 m and thus may be partly a result of vertical extrapolation of modeled wind speeds, similar analyses were also conducted for 6-hourly disjunct wind speeds derived from the wind components at the lowest model level (∼70 m agl). The results indicate the following:
 1. The domain-averaged mean wind speed and U50yr at the lowest model level increase more with increased resolution than is the case for the 10 m winds (Table 1). The mean wind speed at the lowest model level increases by approximately 9% as resolution increases from 50 to 6.25 km, while domain-averaged U50yr increases by ∼20%. These values are almost double the fractional change in wind speeds at a height of 10 m.
 2. As in the results of wind speeds from 10 m, when the fields of 50 year return period wind speeds are “degraded” to 50 km and compared with those from the simulations conducted at 50 km, there is a slight increase in the overall spatial variability (Figure 4a), but the change is relatively modest indicating that while the absolute values of U50yr are higher in the higher-resolution simulations, the overall spatial patterns are relatively similar (Figure 8).
 3. The Gumbel plots constructed for output from the lowest model layer for grid cells containing the five surface observing stations exhibit more consistent increased values with higher resolution at four of the five stations (all except Stavanger) (Figure 9) than is the case for the 10 m wind speeds. This implies that the higher-resolution simulations are generating higher extreme values, but as discussed above, the effect is mediated in the lowest levels of the model and particularly in the 10 m interpolated wind speeds by variations in grid cell land cover characteristics.
 4. Power spectra of wind speeds from the lowest model level sampled every 6 h indicate that with the exception of the Stavanger location, there is evidence that the variance associated with the synoptic and meso-α scale motions is increased with increased model resolution. However, the response is not linear (Figure 10).
4. Concluding Remarks
 The effects of horizontal resolution on simulation of wind climates over northern Europe is assessed using RCA3 applied at four resolutions (approximately 50, 25, 12.5, and 6.25 km). The results indicate that increased model resolution does not lead to uniform improvement of the characterization of power spectra of 10 m wind speeds relative to observations, and specifically that increased resolution does not uniformly add variance at the synoptic and meso-α scale frequencies. Nevertheless, increased model resolution does increase the domain-averaged mean wind speed and both the 50 year return period sustained wind speed and wind gust for a nominal height of 10 m agl and at the lowest model level. The impact of model resolution appears to be greater in wind climate extremes than the mean wind speed and is larger in wind speeds at the lowest model level than in those interpolated to a nominal height of 10 m agl.
 Presuming current and possible future economic and societal vulnerabilities derive principally from the extreme wind events, risk analyses may be very sensitive to the resolution of the model used to simulate the wind climate. Further, the differences in extreme wind speeds and gusts with model resolution reported herein are of comparable magnitude to those reported previously for analyses focused on simulations from different RCMs and analyses of the climate change signal in extreme wind speeds and gusts [Della-Marta and Pinto, 2009; Pinto et al., 2007, 2010; Pryor et al., 2012]. Thus, it may be desirable that the impact of resolution be incorporated into analyses of uncertainty in evaluating climate change projections of extreme wind phenomena. This analysis has explicitly not considered the impact of model resolution on the climate change signal, and it may be worthy of note that one study of the impact of resolution on precipitation climates over Denmark resolved that “the choice of RCM resolution did not affect the projected climate change in this study” [van Roosmalen et al., 2010, p. 417]. The degree to which this is also true for metrics of the wind climate should be addressed in future research, particularly in light of evidence of large and increasing vulnerability of high value assets to wind extremes in the study region and beyond [Schwierz et al., 2010].
 Financial support was supplied by the National Science Foundation (grant 1019603). The comments of three anonymous reviewers are gratefully acknowledged.