Changes in the mean and extreme geostrophic wind speeds in Northern Europe until 2100 based on nine global climate models



This study aims at analyzing the mean and extreme geostrophic wind speeds in Northern Europe. The analyses are based on nine global climate models and the Special Report on Emission Scenarios (SRES) A1B, A2 and B1 scenarios. The time frames studied consist of the baseline 1971–2000 and the future periods 2046–2065 and 2081–2100. The SRES scenarios are considered both separately and combined. The extremes are calculated for the September–April period for various return periods. The analysis is done by applying the program R and the Generalized Extreme Value -methodology.

All projections indicate that both the mean and extreme geostrophic wind speeds will increase in the southern and eastern parts of Northern Europe and decrease over the Norwegian Sea in September–April. The change over the ocean is pronounced already in 2046–2065, over the continents in 2081–2100. For the model mean, the smallest change (2–6%) was projected under the B1 and the largest (4–10%) under the A1B and A2 scenarios. However, spread among the individual global circulation models (GCMs) was fairly large.

The ratios between the return level estimates for various return periods and the annual maximum wind speeds were found nearly homogeneously independent of the time frame studied. For the baseline and future periods, the extreme winds occurring once in 10 or 50 years were 13% ± 2% and 22% ± 5% stronger than the mean annual maxima, respectively. The present findings serve as support for risk assessment such as required when planning the forest management practices. Copyright © 2011 Royal Meteorological Society

1. Introduction

The socio-economic impacts of the violent extreme winds caused by extratropical cyclones have been anomalously large in Europe during the last two decades. For example, in the Decembers of 1990 and 1999, respectively, a total of 100 and 175 Mm3 of timber blew down in winter storms throughout Europe (Ulbrich et al., 2001; Dobbertin, 2002; Schönenberger, 2002; Brüdl and Rickli, 2002). In January 2005, about 70 Mm3 of timber was damaged in Sweden (Alexandersson, 2005; Bengtsson and Nilsson, 2007) and in January 2007, about 45 Mm3 in Central Europe (Fink et al., 2009). In Finland, the most destructive recent wind damages occurred in November 2001 and in July–August 2010. These storms caused a loss of 7.3 and 8 Mm3 of timber, respectively. The economic influence of the damaging winds as well as the secondary damages caused, for instance, by the insect attacks on the remaining trees (Nykänen et al., 1997; Valinger and Fridman, 1997; Schönenberger, 2002; Jönsson et al., 2009; Jönsson and Bärring, 2010) are particularly harmful in the managed forests. This is because the value of harvested timber reduces due to the sudden excess of supply, at the same time the unscheduled and less optimal harvesting procedures increase the costs (Peltola et al., 2010).

Because of the recurrent damage to infrastructure and forests by windstorms (Ulbrich et al., 2001; Dobbertin, 2002; Wernli et al., 2002; Pellikka and Järvenpää, 2003; Alexandersson, 2005; Pinto et al., 2007), the cyclone frequency and intensity in Europe have gained a lot of emphasis in research. For example, Bärring and Fortuniak (2009) found pronounced interdecadal variability in cyclone activity but no significant long-term trend in southern Scandinavian storminess between 1780 and 2005. Correspondingly, Hanna et al. (2008) found no sign of enhancement in storminess associated with climatic change when investigating the Northern European and North Atlantic surface pressure variability measurements since the 1830s. Additionally, according to Bärring and Fortuniak (2009) and Matulla et al. (2007), in northern and central parts of Europe the links of storminess to the North Atlantic Oscillation (NAO) vary depending on the region and the time periods.

In Northern Europe, strong winds are most frequent in the cold season (September–April) because then the spatial variations of temperature are highest (Lau, 1988, Chang et al., 2002) and the conversion of available potential energy to kinetic energy largest. According to Donohoe and Battisti (2009), also of great importance is the global stationary wave pattern that allows the Pacific and Atlantic storm tracks to be connected. If the Pacific and Atlantic cyclone tracks are disconnected, the storms can develop rapidly but the intensities are weaker than if the cyclone tracks are connected. This is independent of the abundance of available potential energy. In line with Donohoe and Battisti (2009), in the investigations of Ulbrich et al. (2008), an increase in the number of the deepest cyclones takes place in the northern North Pacific in tandem with the eastern Northern Atlantic as the global warming proceeds. Also Bengtsson et al. (2009), who investigated winds at the 925 hPa level with the ECHAM5 T213 model (63 km resolution), noticed that the highest wind speeds in winter (December–February) occur predominantly over the mid-latitude areas of the Atlantic and Pacific Oceans.

To provide valid information for decision support, wind speed projections based on various models and greenhouse gas (GHG) scenarios should be analyzed in great detail, considering also the associated uncertainties. In this work, we study the modelled responses in the mean and extreme geostrophic wind speeds (hereafter mean and extreme Vg) to GHG forcing. The Vg speeds are considered rather than the true surface wind speeds because the former are less affected by model parameterization than the latter (Rõõm 1998, Zilitinkevich et al. 2002, Schulz, 2008). The analysis concentrates on Northern Europe, especially Finland and the Baltic Sea. The calculations are made using data from simulations performed with nine global circulation models (GCMs) employing the Special Report on Emission Scenarios (SRES, Nakićenović et al., 2000) A1B, A2 and B1 scenarios. The periods to be compared represent the baseline climate 1971–2000 and projections for 2046–2065 and 2081–2100; for the future climate, model data at daily level is only available for these discrete 20-year periods. The surface geostrophic wind speeds are derived from the model-simulated daily mean sea level pressure and temperature fields. The analyses focus on the high-wind season from September to April (SepApr).

The main goal of this paper is to explore the projected changes in the mean and extreme geostrophic wind speeds. The simulated baseline period temporal mean geostrophic wind speeds of nine GCMs are first compared to the observational estimates derived from the ERA-40 dataset (Uppala et al., 2005). Then the methods used for the calculation of the projected changes in the mean, maximum and extreme wind speed are described. A comprehensive analysis of extreme wind speeds for the various return periods is performed for all grid points using the Generalized Extreme Value (GEV) theory and the block maxima approach. For comparison, a simplified method that utilizes the close relationship between the n-year return level estimates and the average annual maximum is developed. This simple method is applicable to daily practices, for instance, in climate centres, insurance companies or other service providers, where service based on climatic statistics is given based on the continuously updated datasets. The uncertainty in the projected changes is explored by finding out the difference among the various GHG scenarios, individual model results as well as the two methods applied in the extreme analyses.

2. Material and methods

2.1. Climate model data

In this work, altogether nine GCMs (Table I) were used to study the changes in the surface geostrophic wind speeds in Northern Europe (50°N, 10°W → 80°N, 40°E). Three periods, i.e., years 1971–2000, 2046–2065 and 2081–2100 were examined. The nine models utilized in this work are a subset of the 23 models used in the Intergovernmental Panel on Climate Change 4th Assessment Report in 2007. These GCMs represent separate institutes, and only models with spatial resolution of about 300 km (T42) or denser were included in the analysis. The model-simulated sea level pressure and the temperature data were downloaded from the Coupled Model Intergovernmental Project 3 (CMIP3) archive (Meehl et al., 2007). For the baseline period 1971–2000, one simulation for each model for the 30-year period was available. For the future, however, only 20-year periods 2046–2065 and 2081–2100 with two to three GHG simulations per model (Table I) were available.

Table I. The outlines for the climate models employed in this work (see details for models from IPCC, 2007, Table VIII)
Model IDInstituteResolutionScenarios
  • a

    The data of BCCR-BCM2.0 available for years 1971–1998, 2046–2065 and 2081–2098.

  • b

    The data of NCAR-CCSM3 analyzed for years 1970–1999, 2045–2064 and 2080–2099.

BCCR-BCM2.0aBjerknes Centre for Climate ResearchT63 (1.9° × 1.9°)A1B,A2
CGCM3.1(T63)Canadian Centre for Climate Modelling and AnalysisT63 (1.9° × 1.9°)A1B,B1
CNRM-CM3Météo-FranceT63 (1.9° × 1.9°)A1B,A2,B1
ECHAM5/MPI-OMMax Planck Institute for MeteorologyT63 (1.9° × 1.9°)A1B,A2,B1
GFDL-CM2.1National Oceanic and Atmospheric Administration (NOAA)2.0° × 2.5°A1B,A2,B1
IPSL-CM4Pierre Laplace Institute2.5° × 3.75°A1B,A2,B1
MIROC3.2(hires)Japan Center for Climate System ResearchT106 (1.1° × 1.1°)A1B,B1
MRI-CGCM2.3.2Japan Meteorological Research InstituteT42 (2.8° × 2.8°)A1B,A2,B1
NCAR-CCSM3bNational Center for Atmospheric ResearchT85 (1.4° × 1.4°)A1B,A2,B1

2.2. Calculation of the geostrophic wind speed

Because the CMIP3 data archive does not provide any projections for the scalar wind velocity, we calculated the daily mean geostrophic wind vector according to Equation (1):

equation image(1)

where ugdaily and vgdaily denote the 24-h means for the zonal and meridional component of the wind, R the gas constant of air, Tdaily the 24-h mean temperature, f the Coriolis parameter and pdaily the 24-h mean surface air pressure (SLP).

The components equation image and equation image were calculated employing the intermediate points of the native grids of the separate GCMs. This allowed us to use the smallest possible grid size in the calculation. Interpolating the component data linearly onto a common 2.5°× 2.5° grid enabled us to build the multimodel statistics. The observational data applied for evaluation of the model data was represented on the same grid. Time series for the 24-h mean geostrophic wind vector were constructed for all the nine GCMs for the three time spans and the three SRES scenarios. The magnitude of the daily mean geostrophic wind vector can be used as a qualitative measure for the geostrophic wind velocity. This quantity systematically underestimates the daily mean scalar geostrophic wind speed, but the bias occurs both in the baseline period and future GCM-simulated climate. Therefore, by expressing the projected changes in relative terms, despite this bias we can have an idea of how the long-term mean wind conditions will be altered in the future; for the utility of this so-called delta-change approach, see Räisänen (2007) and references therein. For brevity, in this paper the magnitude of the daily mean geostrophic wind vector is simply termed the ‘geostrophic wind speed’.

Using the daily geostrophic wind speed Vgdaily on the 2.5°× 2.5° grid, we calculated the monthly mean Vg(Month). The high-wind season statistics, Vg(SepApr) corresponding to the mean and Vgx(SepApr) to the maximum wind speed of each year, were calculated from 1 September to 30 April and averaged over the number of years available in the model simulations. The calculations were done for each model for the 30-year baseline climate 1971–2000 and for each model and scenario for the two 20-year periods 2046–2065 and 2081–2100. The annual maximum wind speeds of each scenario were combined to form samples. This meant that for the baseline, we had only 30 maxima per GCM. Concerning the future periods, six of the GCMs had all three scenarios (A1B:A2:B1r) and therefore 60 maxima per sample. The other GCMs had only two scenarios either A1B:A2 or A1B:B1 meaning 40 maxima to be used in the extreme analyses. The results based on the model-wise extreme analyses were combined to form the nine-GCM A1B:A2:B1 ensemble. The significance of the projected changes was tested by using the two tailed Student's t-test.

The grid points used in the detailed investigations included one grid point in the southern Baltic Sea and seven land grid points that are located close to the wind measurement stations in Finland (Table II and the Appendix of this paper). For the simplified extreme value analysis, extreme winds on several return periods were calculated for the sea grid point and one of the land grid points. These grid points have been provided with specific names SEA and FOREST.

Table II. Grid points used in the monthly mean and extreme wind speed investigations
Grid point nameLatitude °NLongitude °E
  1. The names of the points refer to synoptic weather stations close to the points. The extreme value analysis was done for the two grid points that have been provided with a specific name (SEA, FOREST).

Baltic Sea/SEA5515

2.3. Assessing the quality of the GCM data

The quality of the model-inferred geostrophic winds was assessed by comparing them with the corresponding observation-based approximation. We employed the ERA-40 dataset (Uppala et al., 2005) in which the pressure analyses are given at 6-h intervals. We first averaged the four consecutive SLP analyses to obtain the daily mean pressure and then calculated the vector (equation image, equation image) from the gradient of this pressure field (Equation (1)).

The systematic error of the modelled geostrophic wind speed for the high-wind season is depicted in the upper panel of Figure 1. The bias is generally fairly small, of the order of magnitude of 0.5 m s−1, the 9-GCM mean tending to slightly underestimate the observed Vg. Larger errors are seen in the vicinity of mountainous areas in Central Europe, and in particular, close to Greenland. In those areas, it is evident that the distribution of Vg is seriously affected by reduction of the surface pressure to standard sea level.

Figure 1.

Map: The difference of the 9-GCM mean geostrophic wind speed for the September to April season from its observation-based counterpart derived from the ERA-40 dataset; an average for the baseline period 1971–2000, with a contour interval of 0.5 m s−1. Small diagrams: The seasonal cycle of Vg at four locations. The thick line shows the multimodel mean and the shading the ± one standard deviation interval, derived from the simulations performed with the 9-GCMs. The corresponding geostrophic wind speeds calculated from the ERA-40 dataset are denoted by dots

The seasonal cycle of the modelled and observation-based geostrophic wind speeds at four grid points are compared in the lower panels of Figure 1. The observation-derived time mean geostrophic wind is weakest in summer, about 7 m s−1, and strongest, 11–13 m s−1, in winter or late autumn. The seasonal cycle is reproduced by the models reasonably well, even though the amplitude tends to be somewhat underestimated. Almost invariably, however, the observed values differ from the multimodel mean by less than one standard deviation of the modelled values.

2.4. Calculation of the extreme geostrophic wind speeds by two alternatives

2.4.1. Primary method: block maxima with the gridded dataset

The datasets of the modelled Vgx(SepApr) were analyzed using the GEV theory (Coles, 2001; Castillo et al., 2004; Wehner et al., 2010). We used the block maxima method by employing the program R, a free software developed by the National Center for Atmospheric Research (NCAR) (Katz et al., 2005) and downloadable from the internet to be used for instance for calculating extreme values of climate parameters.

The estimated probabilities of the extreme geostrophic wind speeds are expressed in terms of the return periods and the corresponding return level estimates. The return level estimates are defined as the threshold that is exceeded at any given period with the probability of p = 1/T, T being referred to as the return period (RP) (or waiting time) in years.

First, the GEV distributions were fitted to the sampled annual maxima using the cumulative distribution function (Equation (2)):

equation image(2)

where µ, σ and ξ represent the location, scale and shape parameters of the distribution function, respectively. The sign of the shape parameter ξ defines the type of the GEV distribution. When ξ = 0, the distribution is of Gumbel type (light-tailed), when ξ< 0, the distribution is of the bounded Weibull type and in the case of ξ> 0, the distribution represents the Frechet distribution with a heavy tail on the right.

After the parameters µ, σ and ξ were estimated from the sample of annual maxima, the return level estimates Vgx(RP) could be determined for various return periods RP (e.g., 10 and 50 years) using Equation (3):

equation image(3)

The software toolkit ‘extRemes’ of the program R uses the Broyden–Fletcher–Goldfarb–Shanno (BFGS) method (Broyden, 1970, Fletcher, 1970, Goldfarb, 1970, Shanno, 1970) for optimization. With the software toolkit version that was suitable for the gridded datasets, the Vgx(RP) could be determined systematically for the chosen return periods for all individual GCMs. After that, it was possible to form the 9-GCM ensemble for the combined A1B:A2:B1 scenario and the 6-GCM ensemble utilizing the various scenarios.

2.4.2. Simplified method: Vgx statistics and block maxima at two grid points

Because the ratio of the Vgx(RP) to the average annual maximum Vgx(SepApr) proved to be rather homogeneous over the domain (not shown) another way of estimating the return level estimates was developed by using a new simplified approach. In this simplified approach, the detailed extreme analyses were conducted by using data at two grid points: 55°N, 15°E (SEA) and 62.5°N, 30°E (FOREST). The ratio of the extreme wind speeds Vgx (RP) to the average September–April maximum Vgx (SepApr) was then calculated for each model according to Equation (4):

equation image(4)

The grid point coefficients CFequation image obtained from Equation (4) were averaged over the nine GCMs for each return period to calculate coefficients CF71 for the period 1971–2000, CF46 for 2046–2065 and CF81 for 2081–2100. The values of the coefficients for the different periods proved to be nearly identical (Table III). Therefore, a time-independent coefficient, CFALL, was derived by averaging the coefficients determined for the three periods 1971–2000, 2046–2065 and 2081–2100. Using the model-wise time-dependent coefficients CF71, CF46 and CF81, we also assessed the spread of the changes (cf Section 3.3.2) between the future and the baseline periods.

Table III. Coefficients determining the ratio of the extreme geostrophic wind speeds Vg(RP) to the mean annual maximum Vgx(SepApr) (for definition, see text) as a function of the return level; an average over the two grid points (SEA and FOREST) and as a function of the return period and the time span
RP (years)51050100
  1. CF71 (RP) corresponds to the baseline period 1971–2000, CF46 (RP) to 2046–2065 and CF81 (RP) to 2081–2100. The average of the three coefficients CFALL is shown on the bottom row.


On the basis of the average and the standard deviation of the coefficients CF for the nine GCMs and the three periods at the two grid points SEA and FOREST, we obtained upper and lower estimates for the values of CFALL for each return period. The regression curves fitted to the minima, average and maxima of CFALL are shown in Figure 2. These, plotted as a function of the return period, give the coefficients by which we can approximate the extreme Vgx(RP) with uncertainty intervals on arbitrary return periods if Vgx(SepApr) is known. According to Figure 2, the wind speeds that occur once in 5-, 10-, 50- and 100-years are approximately 8% ± 1%, 13% ± 2%, 23% ± 5% and 26% ± 6%, respectively, stronger than the average Vgx(SepApr); i.e. the longer the return period, the larger the uncertainty in the coefficient.

Figure 2.

Coefficient CFALL (for definition see text) as a function of the return period RP. Grey boxes show the best estimate values of CFALL for RP of 5, 10, 50 and 100 years The deviation among the model-wise estimates of the three time spans at the two grid points SEA and FOREST per return period is marked with the bars. Regression curves corresponding to the low, best and high estimate are also given

2.5. Sensitivity analysis

As the main part of this work focused on the combined A1B:A2:B1 scenario using an ensemble of nine GCMs, it was also of interest to find out how important the differences among the various GHG scenarios are. For that purpose we repeated all calculations for the mean and extreme wind speeds under the individual scenarios for a set of six GCMs that all had the three SRES scenarios available. Additionally, as the extreme analyses were performed both by using the whole gridded dataset and the simplified method, a comparison of the results of these two methods for the projected changes in the 10- and 50-year return level estimates was made.

3. Results

3.1. Mean geostrophic wind speeds

3.1.1. Projected changes

On the basis of the multimodel mean projections, the surface geostrophic wind speeds increase slightly in the southern and southeastern parts of Northern Europe and decrease over the Norwegian Sea (Figure 3). The geographical pattern of the change is qualitatively very similar for the A1B, A2 and B1 scenarios and the combined A1B:A2:B1 ensemble scenario. This is so, although the responses to the separate scenarios are based on six rather than nine GCMs. The largest increase (>4% by 2081–2100) occurs over southern Baltic Sea and the adjacent land areas under the A1B and A2 scenarios. The smallest and statistically least significant changes are seen under the B1 forcing. The combined A1B:A2:B1 scenario yields fairly similar changes regardless of whether a mean of six (not shown) or nine GCMs is examined. The mean wind responses for each calendar month for a number of individual locations under the A1B-scenario are reported in the Appendix. During the windiest time of the year, the monthly mean wind speeds will start to increase in the Baltic Sea already in 2046–2065 (Table AI). In Finland, increases are largest (5–7%) in November and January by 2081–2100. In November–February 2081–2100, a positive shift of 5–10% is projected to materialize in the Baltic Sea (Table AII).

Figure 3.

The multimodel mean percent changes in the September–April average geostrophic wind speeds from 1971–2000 to 2046–2065 (left) and from 1971–2000 to 2081–2100 (right). From the top (a) and (b) 9-GCM mean A1B:A2:B1; (c) and (d) 6-GCM A1B; (e) and (f) 6-GCM A2; and (g) and (h) 6-GCM B1. Shading indicates the areas where the change is statistically significant at the 95% level. Positive changes are marked with long dashed line and negative with dots. The zero contour is omitted

3.1.2. Uncertainty

The simulated change in the monthly mean geostrophic wind speed varies from one model to another. To assess the resulting uncertainty, we fitted a normal distribution to the set of projections of changes in Vg of the individual models at the SEA and FOREST grid points, and determined the percentage points from this distribution. In winter and late autumn, the multimodel mean change in the mean Vg manifests a positive trend, up to 5–10% in a century for some months (Figure 4). In summer, the multimodel mean response is negligibly small. However, the scatter among the model simulations is quite large. At the SEA grid point, the probability for a positive change is more than 75% from October to February, whereas at the FOREST grid point this limit is only exceeded in November and January. According to the upper-estimate change (the 95th percentage point of the distribution), monthly mean wind speeds would increase by up to 15%. These increases, roughly speaking, occur in the months with the largest climatological wind speeds under the control period (Figure 1). However, as the ensemble consists of only nine GCMs, this uncertainty analysis should be considered as suggestive.

Figure 4.

The seasonal cycle of the projected change in the monthly mean geostrophic wind from 1971–2000 to 2081–2100 under the A1B scenario at the two grid points SEA and FOREST: a probability distribution. The bars represent the 25–75 and whiskers the 5–95 probability interval. Numbers 1–12 refer to the calendar months. In addition to individual months, the corresponding probability intervals are depicted for the entire high- and low-wind seasons (H and L)

The projections also fluctuate between consecutive months (Figure 4). This holds both for the multimodel mean change and the width of the uncertainty interval. Evidently this variability is largely stochastical by its nature rather than caused by any physical grounds. For the mean of the high-wind season, the model-based probability for more intense average winds in the future is quite large, about 88% for both grid points examined in Figure 4. According to the t-test, this indicates that the deviation of the multimodel mean change from zero is confident at more than 95% level (Figure 3(b)). Conversely, the multimodel mean response is insignificantly small at both grid points (Figure 4) in the low-wind season (May–August).

3.2. Extreme geostrophic wind speeds determined from the gridded dataset with the primary method

3.2.1. Projected changes based on the combined A1B:A2:B1 scenario

The projected changes in the extreme geostrophic wind speeds are shown in Figure 5. By the period 2046–2065, the extreme wind speeds change very little and only a small fraction of these changes are systematically significant. The pattern of change is also rather noisy. For the latter 20-year period the changes become more pronounced. The 10- and 50-year return level estimates increase by 2–4% and 2–8%, respectively, in the eastern part of the study area (20°E–40°E). The changes are statistically more significant for the 10-year than for the 50-year return level estimates. The decreases over the Norwegian Sea are of the order of 2–4% but only the changes in the 10-year return level estimates are statistically significant.

Figure 5.

Changes in the return level estimates in percentages derived from the gridded dataset when comparing the periods 2046–2065 (left) and 2081–2100 (right) to 1971–2000 climate using the 9-GCM mean and the combined A1B:A2:B1 scenario. In (a) and (b) depicted are the changes for the 10-year and in (c) and (d) for the 50-year return level estimates. Shading indicates the areas where the change is statistically significant at the 95% level. The zero contour is omitted

As far as the individual GCMs are concerned, the weakest changes by 2046–2065 were projected by IPSL and the strongest ones by CNRM (not shown). For the 20-year period 2081–2100, the weakest changes for the whole domain were produced by GFDL-CM2.1 and the largest ones by CNRM. The 10-year return level estimates produced by the individual models seemed to be rather reasonable. With the longer return periods, by contrast, anomalously high values of change appeared in the analyses around some grid points. This phenomenon occurred in the 50-year return periods in the extreme value analyses of BCCR-BCM2.0, GFDL-CM2.1 MRI-CGCM2.3.2, MIROC3.2 (hires), ECHAM5/MPI-OM and CGCM3.1 (T63). Simulated by these individual GCMs, local increases up to 20% and decreases of 16% occurred at some grid points.

3.2.2. Difference due to the GHG emission scenarios

Changes in the 10-year return level estimates of Vg as responses to the individual A1B, A2 and B1 scenarios are shown in Figure 6. As in the case of the mean Vg speed (Figure 3), the B1 forcing gives smallest and the A1B and A2 forcing largest decreases and increases. Additionally, the A1B- and A2-scenarios project fairly similar changes both in time and location (Figure 6). Positive changes of 2–6% are projected to materialize in Finland and the northwestern part of Russia. Negative changes of similar magnitude appear in the analyses over the Norwegian Sea.

Figure 6.

The changes in the 10-year return level estimates from 1971–2000 to 2046–2065 (on the left) and from 1971–2000 to 2081–2100 (on the right) for individual SRES scenarios as an average of six GCM simulations. From the top the 6-GCM mean (a) and (b) A1B; (c) and (d) A2; (e) and (f) B1. Shading indicates the areas where the change is statistically significant at the 95% level

3.3. Extreme wind speeds using the simplified method

3.3.1. Comparison of the two methods used for analyzing the extreme winds

The spatial distributions of the projected changes in the return level estimates of the extreme geostrophic wind speeds proved to be were qualitatively similar regardless of whether the primary (Figure 5) or the simplified method (Figure 7) was used. As for the gridded dataset, the simplified method gives weaker changes for 2046–2065 than for 2081–2100. Similarly, the main pattern of changes followed that described already when using complete method: an increase of 2–4% in Finland and the northwestern part of Russia and a decrease over the Norwegian Sea. One clear advantage in the simplified method is that it filters the anomalously large, evidently physically unrealistic grid point estimates that were present in six of the GCMs (see Section 3.2.1) when using the primary method.

Figure 7.

The percent changes in September–April extreme geostrophic wind speeds when using the simplified method with the time-dependent coefficients (Table III). For denotations see Fig. 5

3.3.2. Differences among the model simulations at two grid points

Using the model-wise coefficients CFGCM obtained by the simplified approach and fitting a normal distribution to the set of projections based on these coefficients, we can define the percentage points for the changes of Vgx(RP10) and Vgx(RP50). These distributions are shown in Figure 8. In the extreme analyses considering the 10-year return level and the period 2046–2065, the probability of a positive change at the grid points SEA and FOREST is about 75 and 50%, respectively (Figure 8). In 2081–2100, the extreme wind speeds of the 10-year return level show, at both points, an increase at about 75% probability. For the changes of the 50-year return level, the uncertainty intervals are much broader than for the 10-year level. The interval for the period 2081–2100 ranges from about − 9 to + 13% at both grid points. In other words, even the sign of the change in the 50-year return level estimates of geostrophic wind speeds at these two grid points, SEA and FOREST, is highly uncertain.

Figure 8.

The probability distribution of the projected changes in the high-wind season (Sep–Apr) extreme geostrophic wind speeds from 1971–2000 to 2046–2065 and from 1971–2000 to 2081–2100 under the A1B:A2:B1 scenario on the 10-year (dark shading) and 50-year (light shading) return levels. The bars represent the 25–75 and whiskers the 5–95 probability interval. Horizontal line inside the bars depicts the multimodel median. This figure is available in colour online at

4. Discussion

In this work, the main focus was on analysing the projected changes of the mean and extreme geostrophic winds (Vg) in Northern Europe as well as the uncertainties and sensitivity in the analyses. The analyses of the extreme Vg on the different return periods were done using the GEV theory (Coles, 2001, Castillo et al., 2004) and the block maxima approach with the program R utilizing the extRemes software toolkit package (Katz et al., 2005). As the responses in extreme wind speeds were approximately homogeneously proportional to the average maximum wind speeds changes in September–April, a simplified method for estimating wind extremes was developed. This method used the annual maxima statistics at two grid points (55°N, 15°E and 62.5°N, 30°E). Applying the simplified approach, the unrealistic return level estimates found at some grid points in the study area in some of the GCMs could be filtered.

The projected changes with the mean and extreme geostrophic winds became clearly significant in the latter 20-year period 2081–2100. For the mean wind the changes and their significance covered larger areas than the changes in the extreme wind speeds did (Figure 3(a),(b) vs Figure 5). However, in all the calculations the resulting pattern bared several similarities: there was an increase in southern and eastern Northern Europe and a decrease over the Norwegian Sea. The A1B and A2 forcing showed stronger increases at the end of the century than the B1 forcing.

By comparing the GCM-based average high-wind season maximum Vgx to the corresponding extreme geostrophic wind speeds on the various return periods Vgx(RP), we found that for instance the 10 and 50 year return period extreme wind speeds are, respectively, 13% ± 2% and 22% ± 5% stronger than the average September–April maxima. This finding was true for the three time periods and two grid point locations studied. The extreme wind speeds will increase on average by 2–4% in the southern and eastern parts of Northern Europe, whereas a decrease of 2–6% dominates over the Norwegian Sea. Nikulin et al. (2011), who evaluated the future projections of 20-year return level estimates of gust winds by performing the downscaling of six GCMs, found similarly that the increase in winds is dominant in a zone stretching from northern parts of France over the Baltic Sea towards northeast. In their work the decrease over the Norwegian Sea was somewhat less robust and stretched over Norway.

5. Conclusions and final remarks

The two methods used for analyzing the extreme wind speeds in this work gave qualitatively similar results. The results are more robust for the short return periods. Therefore, we assume that only a small part of the uncertainty in the estimated return level estimates originated from the extreme value analysis itself and that most of the uncertainty related to the extreme wind speed analyses resulted from the spread among the different GCMs. Similarly, Kharin et al. (2007), who studied the changes in temperature and precipitation extremes of a GCM ensemble, concluded that model differences generally dominate the uncertainty. Additionally, as the block maxima approach is somewhat sensitive to the sample size, a better way could be to employ the peak over threshold (POT) and bootstrapping methods (Naess and Clause, 2000) instead. Wehner (2010) who focused on the sources of uncertainty in the extreme temperature analyses found, however, that the return level estimates obtained by employing averaged empirical and parametric methods compare reasonably well even then when the cumulative distribution function is problematic to fit to the sample.

When we examine the individual months in Finland and the Baltic Sea, we can see that wind speeds are projected to increase from October to February. Despite the fact that the extreme winds are not expected to increase greatly in Northern Europe in the future, the wind-induced risks to forests could still be expected to increase under the changing climate. The influence of an increase in wind speed in forests in Finland is accompanied and strengthened by the reduction of the frozen soil period during the windiest seasons (i.e. from late autumn to early spring) (Peltola et al., 1999, Gregow et al., 2011). Moreover, under the warming climate the forest growth is projected to increase (Kellomäki and Leinonen, 2005, Kellomäki et al., 2008), which increases the need to manage forests more often or with higher intensity (i.e., thinning, final fellings) (Peltola et al., 2010). In the future work, similar analyses as presented in this paper would be important to do by examining the wind speeds together with other climate parameters. For instance, the changes in the wind speeds and directions, snow loads and soil frost (Gregow et al., 2008, 2011) could be combined model-wise for assessing the uncertainties in the projected climatological risks to wind- and snow-induced damages in forests. Wind speeds and snow loads in unfrozen and frozen ground have different impacts on the exposure of trees to bending, breakage and uprooting depending also on the tree age, stand characteristics, tree type and the recent management practices (Peltola et al., 1999, Gregow et al., 2011).

Although the GCM simulate the extratropical cyclone tracks and even the storm developments rather well (Geng and Sugi, 2003, Leckebusch and Ulbrich, 2004, Lambert, 2004, Bengtsson et al., 2006, Leckebusch et al., 2008, Ulbrich et al., 2008, Fink et al., 2009), there are still many details that need further research, especially when the focus is on winds. Aspects such as the projected reduction of snow cover (Jylhä et al., 2008, Räisänen and Eklund, 2011), retreat of sea-ice (Gordon et al., 2000, Mueter and Litzow, 2008), uncertainties in the land surface parameterization (Fischer et al., 2010) and the boundary layer processes as well as the modes in the ocean circulation can have essential unknown feedbacks to the storm and wind climate. As many such parameters and feedbacks are currently not yet adequately described in the GCMs, there is clearly still such uncertainty, i.e. in the climate projections that cannot even be estimated. Therefore, the results of this work should be considered mainly suggestive.


We acknowledge the modelling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) and the WCRP's Working Group on Coupled Modelling (WGCM) for making available the WCRP CMIP3 multi-model dataset. Seppo Saku is thanked for technical assistance. This work was financially supported by the Academy of Finland (2008–2010) via funding of the project on Impacts of temporal and spatial variability of critical weather events and forest management on the risk of wind- and snow-induced damage in forest stands, led by Dr Heli Peltola, University of Eastern Finland, School of Forest Sciences. Additional funding was provided by the Finnish Ministry of Agriculture via Finnish Research Programme on Adaptation to Climate Change. Dr Ari Venäläinen from the Finnish Meteorological Institute is thanked for the initial discussions. An anonymous referee is acknowledged for several useful suggestions that have improved the paper significantly.

Appendix: Monthly mean Vg changes at various grid points in Finland and over the Baltic Sea

In 2046–2065, over the southern Baltic Sea, the Vg (Month) increases during October–November and January–February by 3–6% (Table AI). In Finland, such an increase is only seen in November while other months are only showing 1–3% changes of either positive or negative sign. Decrease dominates in Finland in March and April.

For the period 2081–2100, the changes are qualitatively similar to those for 2046–2065 but larger in amplitude. However, a general decrease is now only seen in March (Table AII). In November, the increase is 5–10%, being largest in the Baltic Sea and smallest in northern Finland. In January, the corresponding increase is 4–7%. More intense winds are likewise expected in October in the future although the change is mostly 2–4%. The differences between individual months indicate that an increase in the wind speeds is most likely in September–February.

Table AI.. The 9-GCM average percentage changes in monthly mean wind speeds at the grid points presented in Table I as a response to the A1B scenario for the period 2046–2065 relative to 1971–2000.
Baltic Sea350− 114− 3− 10462
Helsinki23− 2− 1020− 13261
Joensuu02− 2− 1− 1− 1− 1− 23251
Jyväskylä22− 3− 1− 101− 13251
Kauhava22− 2− 1− 11003241
Kajaani11− 2− 2− 10− 2− 11230
Rovaniemi21− 2− 2− 11001130
Sodankylä11− 2− 300− 100230
Table AII.. The 9-GCM average percentage changes in monthly mean wind speeds at the grid points presented in Table I as a response to the A1B scenario for the period 2081–2100 to 1971–2000.
Baltic Sea6613110004105
Helsinki62− 2101112362
Joensuu62− 30− 1− 10− 12372
Jyväskylä72− 3100001451
Kauhava62− 2100001450
Kajaani53− 30− 10− 3− 11251
Rovaniemi53− 3001− 101350
Sodankylä42− 2− 2− 20− 110352