Journal of Geophysical Research: Atmospheres

Assessing the performance of Intergovernmental Panel on Climate Change AR5 climate models in simulating and projecting wind speeds over China



[1] The ability of nine current generation (Coupled Model Intercomparison Project Phase 5, CMIP-5) coupled atmosphere-ocean general circulation models (AOGCMs) to accurately simulate the near-surface wind climate over China is evaluated by comparing output from the historical period (1971–2005) with an observational data set and reanalysis output. Results suggest the AOGCMs show substantial positive bias in the mean 10 m wind speed relative to observations and the ERA-40, National Centers for Environmental Prediction–Department of Energy, and National Centers for Environmental Prediction–National Center for Atmospheric Research reanalysis. Given that the models generally produce the upper level geopotential height gradients comparatively well, it is postulated that one major reason for the discrepancy between observed and modeled wind fields is the surface characterization used in the AOGCMs. All models exhibit lower interannual variability than reanalysis data and observations, and none of the models reproduce the recent decline in wind speed that is manifest in the near-surface observations. The wind speed of individual model runs during the historical period does not exhibit much influence from the initial atmospheric conditions. The output for the current century from seven of the AOGCMs is examined relative to the historical wind climate. The results indicate that spatial fields of wind speed at the end of the 21st century are very similar to those of the last 35 years with comparatively little response to the precise representative concentration pathway scenario applied.

1. Introduction

[2] Understanding how wind climates have varied in the past and how they might change in the future not only is useful as a diagnostic of the near-surface manifestations of the large-scale circulation but is relevant to possible impacts of climate nonstationarity on ocean ecology [Fitch and Moore, 2007], oceanic carbon storage [Swart and Fyfe, 2012], the agriculture industry [van Hout et al., 2008], forestry [de Langre, 2008; Bang et al., 2010], and the wind energy industry [Pryor et al., 2005; Pryor and Barthelmie, 2011; Rasmussen et al., 2011]. Several recent studies have reported declines in observed in situ near-surface wind speeds during the past 30–50 years over parts of North America [Klink, 1999; Pryor et al., 2009; Tuller, 2004], regions of Europe [Brázdil et al., 2009; Pirazzoli and Tomasin, 2003] and Australia [McVicar et al., 2008]. In one study of near-surface wind speeds over continental areas in the northern midlatitudes from 1979 to 2008,Vautard et al. [2010] reported that wind speeds from 822 surface weather stations have declined by 5–15%.

[3] Prior analyses of in situ daily average wind speed data from China have also indicated recent declines [Fu et al., 2011; Jiang et al., 2010b]. Niu et al. [2010] estimated that spatially averaged surface wind speeds have dropped from 3.7 m s−1 to about 3 m s−1 and the frequency of light wind events has increased significantly over the last three decades. This is consistent with the study of Guo et al. [2011], who estimated the spatially averaged trend in annual mean observed wind speed over China from 1969 to 2005 to be −0.18 m s−1 decade−1. A further study reported that over eastern Asia the annual mean surface wind speed changed by approximately −0.12 m s−1 decade−1 between 1979 and 2008 [Vautard et al., 2010], which may imply a recent reduction in the rate of change. It should further be noted that even in China there are regions that do not exhibit reduced wind speeds, but have actually been characterized by small magnitude increases in wind speeds [Jiang et al., 2010b]. Causes of the declines in recent reductions in observed near-surface daily mean wind speeds over China (and other terrestrial locations) have yet to be fully identified and quantified.Jiang et al. [2010b] proposed that land use change, relocation of stations and changes in anemometry over the period 1956–2004 had only a modest influence on the change in annual mean wind speed and rather the majority of the decrease was linked to changes in the sea level pressure gradients between the Asian continent and Pacific Ocean. However, a sensitivity analysis conducted by Vautard et al. [2010] suggested that increased forest growth over China likely contributed between 40% and 60% of the observed reduction in wind speeds over the past three decades. Based on surface observations, Xu et al. [2006] found “the surface wind speed associated with the eastern Asian monsoon has significantly weakened in both winter and summer in the recent three decades” due to radiative forcing from both enhanced greenhouse gas concentrations and local air pollution.

[4] Atmosphere-ocean general circulation models (AOGCMs) are the primary tools available for investigation of possible future climate scenarios under different radiative forcing, and are used herein to provide projections of the wind climate over China. These models are subject to near-continuous modification with the goal of better describing atmospheric dynamics and reducing uncertainty in climate projections. However, there is a need for detailed evaluation of model performance. Specific to evaluation of AOGCM simulations of past and possible future wind and atmospheric flow climate,Gastineau and Soden [2009]projected that the frequency of extreme near-surface (nominal height of 10 m) wind speed would decrease in the tropics, but increase in the extratropics in response to global warming using a multimodel ensemble of coupled climate model simulations. However, substantial bias in oceanic wind stress was found for several of the Coupled Model Intercomparison Project Phase 3 (CMIP-3) generation AOGCMs [Swart and Fyfe, 2012]. Based on the performance of 19 AOGCMs against reanalysis winds during 1981–2000, McInnes et al. [2011]found the multimodel ensemble exhibit low skill over land areas and spatial correlation coefficient between models and National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP-NCAR) reanalysis is particularly low in the region north of 30°N.

[5] Here we focus on evaluation of AOGCM simulations from the CMIP-5 generation of models in terms of near-surface wind and geopotential height fields over China. There have been relatively few previous studies evaluating the ability of climate models to simulate wind speeds in this region.Jiang et al. [2010a] examined the performance of three regional climate models run at spatial resolution of between 20 and 50 km nested within lateral boundary conditions from reanalysis output and AOGCMs, and concluded that these models exhibit some skill in simulating the distribution of mean wind speeds but fail to simulate the recent reductions in wind speeds described above.

[6] Herein, we present analyses of observational data sets, reanalysis products, and nine AOGCMs to do the following. First, we quantify and compare the magnitude, historical trends and temporal variability in 10 m wind speeds derived from direct observations, reanalysis products and output from AOGCMs. Second, we compare the spatial patterns of wind speeds from different data sets. Specifically, we propose possible causes of discrepancies by examining the simulated large-scale atmospheric circulation in different models. Finally, we address whether there is evidence for a long-term climate change signal in the wind regimes over China and investigate the sensitivity of projected wind speeds to radiative forcing and whether any projected change in wind climates is linked to the skill with which wind speeds are simulated during the historical period.

2. Data and Methods

2.1. Data Sets

[7] Our analysis is based on simulations and projections of nine AOGCMs in the World Climate Research Program's (WCRP) Coupled Model Intercomparison Project phase 5 (CMIP-5) for the upcoming Intergovernmental Panel on Climate Change Fifth Assessment Report (AR5) [Taylor et al., 2012]. We analyze monthly near-surface mean wind speeds from historical runs with (i) only greenhouse gas (GHG) forcing, (ii) only natural forcing, and (iii) the combination of runs i and ii, and also examine AOGCM output for a range of future greenhouse gas forcing scenarios. In contrast to the CMIP-3 simulations, the future scenarios in the CMIP-5 exercise employ radiative forcing scenarios derived from representative concentration pathways (RCPs) [Moss et al., 2010; van Vuuren et al., 2011]. The naming convention of the RCP reflects the peak radiative forcing in Wm−2 at or near 2100 (e.g., RCP2.6 indicates the peak forcing is 2.6 Wm−2) [Taylor et al., 2012]. Table 1lists the AOGCMs used in the present work, including model resolution, simulation availability and data record length. It is worthy of note that wind speed data derived from CSIRO-MK3.6 are for a nominal height of 2 m, while for all other AOGCMs wind speeds are output at 10 m. More detailed documentation of CMIP5 models can be found at

Table 1. Summary of the Key Characteristics of the Climate Models Used in This Study and the Number of Simulations Available for the Historical Period and for Each of the Representative Concentration Pathways (RCPs) Used for the Simulations of Future Conditions
AOGCMAtmospheric ResolutionHistorical PeriodNumber of Model RunsKey Reference
BCC-CSM1.1∼2.8° × 2.8°1850–201231111Zhang et al. [2011]
CanESM2∼2.8° × 2.8°1850–2005555 5Arora et al. [2011]
CNRM-CM5∼1.4° × 1.4°1850–20051011 1Voldoire et al. [2012]
CSIRO-MK3.6∼1.9° × 1.9°1850–20051010101010Rotstayn et al. [2010]
GISS-E2-H2.5° × 2°1850–20055    Hansen et al. [2007]
GISS-E2-R2.5° × 2°1850–20055 511 
INM-CM42° × 1.5°1850–20051 1 1Volodin et al. [2010]
IPSL-CM5A-LR∼3.75° × 1.9°1850–20055 313Brient and Bony [2012]
MIROC4h∼0.56° × 0.56°1950–20053    Chikamoto et al. [2012]

[8] Simulations by the nine AOGCMs of the historical period are evaluated relative to NCEP-NCAR reanalysis data [Kalnay et al., 1996], NCEP–Department of Energy (NCEP-DOE) reanalysis data output [Kanamitsu et al., 2002], ERA-40 reanalysis data [Uppala et al., 2005], and a set of homogeneous daily near-surface wind speed observations during 1971–2005 from 540 stations over China [Chen et al., 2012]. Observations of daily mean wind speed at multiple sites across China were obtained from the National Meteorological Center of China Meteorological Administration (CMA). These data meet the World Meteorological Organization's standards and all wind speeds were measured at 10 m above the ground. For every station, a further data quality constraint is applied; that more than 85% observations are valid in each climatological season of each year. The study domain extends from 73° to 136°E, 16° to 54°N, but only grid cells where the centroid lies within China are included in the statistics presented herein (Figure 1).

Figure 1.

Map of the study domain. The triangles indicate the 540 surface observational stations from which data are presented. The dots indicate the three grid cells that are summarized in Table 3. Major geographic features of the study region are also depicted.

[9] To address the question of what may cause the differences in wind speed patterns between the AOGCMs and the reanalysis products, we also use geopotential height data to examine the pressure-gradient force which is the large-scale driver of wind speeds. In this analysis because of the high elevation of the Qinghai-Tibet plateau, the height gradient at 500 hPa is used as the diagnostic parameter over western China, while the height gradient at 700 hPa is used for eastern China (the dividing line between the two regions is set at 105°E; seeFigure 1).

2.2. Data Processing and Methods

[10] Because of slight differences in the AOGCMs model resolution and location of grid centroids and the reanalysis products for the comparison with the reanalysis data set all output was interpolated to 2.5° × 2.5° using inverse distance weighting. When comparing simulations with a specific model under different RCPs, we maintain the original model resolution. Differences between the spatial fields of wind speeds from different models or model runs are evaluated in terms of the ratio of standard deviation (SD), the spatial correlation (r) and root-mean-square difference (RMSD), and are summarized using Taylor diagrams [Taylor, 2001].

[11] Regression analysis is applied to time series of annual mean wind speeds to quantify temporal trends. A bootstrapping technique (using 1000 samples) is used to determine robust confidence intervals on the slope of the regression equation and thus determine the statistical significance of trends at a significance level (α) of 0.05 [Pryor et al., 2009]. This approach provides robust estimates of the trend magnitude, but in the case of high temporal autocorrelation tends to overestimate the significance of any tendency [Pryor and Ledolter, 2010].

3. Results

3.1. Historical Period

[12] Prior research over Europe has indicated that initial conditions exert a strong impact on downscaled projected wind climates (particularly intense and extreme wind speeds) throughout the 21st century [Pryor et al., 2012a]. Thus, output for eight of the nine AOGCMs (BCC-CSM1.1, CanESM2, CNRM-CM5, CSIRO-MK3.6, GISS-E2-H, GISS-E2-R, IPSL-CM5A-LR and MIROC4h, seeTable 1), from multiple simulations which differed only in terms of the initial conditions was examined to assess the degree to which wind climates in the historical period (1971–2005) differ. Differences in the spatial patterns of mean wind speeds over China from these ensemble members during historical period are extremely modest (see the example given in Figure 2 and the time series shown in Figure 3), indicating that the precise initialization period has little influence on the simulated wind speed during 1971–2005 (and indeed all >30 year periods during the historical simulations for each AOGCM). For this reason, in some of the remaining analyses the ensemble mean of simulations with each individual AOGCM is presented.

Figure 2.

Taylor diagram of simulated wind speeds from CanESM2 during 1971–2005. The colors indicate the seasons, and numbers represent realizations with different initial conditions. One member of CanESM2 simulations is arbitrarily selected as the reference (r) and is used to compare each of the four model simulations (m). Taylor diagrams depict three components of the degree to which patterns in the AOGCM simulations are similar: the correlation (r) of the spatial patterns (shown by the azimuthal angle), the ratio of the standard deviation in the fields (SD RATIO (σm/σr)), shown by the radial distance from the origin on the xaxis at a ratio of 1, and the root-mean-square difference (RMSD) in the fields (shown by the distance from the origin on thex axis at (SD RATIO = 1)).

Figure 3.

(a) Annual mean wind speeds over China from individual simulations with the nine AOGCMs (1850 to 2005). Also shown are time series of mean wind speeds from the NCEP-NCAR, NCEP-DOE, and ERA-40 reanalysis output and the average of the observations from the 540 surface stations. Annual mean wind speeds over China from individual simulations with six of the AOGCMs (b) with greenhouse gases forcing only and (c) with natural forcing only.

[13] Domain averaged annual mean wind speeds (1850 to 2005) from individual simulations with a given AOGCM are shown in Figure 3aalong with data from the three reanalysis data sets and the average of the 540 observing stations. Except CSIRO-MK 3.6 (for which the wind speeds are for a nominal height of 2 m), all the other simulated historical 10 m wind speeds display positive bias relative to reanalysis output and observations. This tendency is particularly marked for two models: CanESM2 and BCC-CSM1.1. The spatially averaged wind speeds calculated from these two models are ≥4.5 m s−1, which is nearly double the value from the NCEP-NCAR and NCEP-DOE reanalysis and direct observations, and is over 3 times the value from ERA-40. Because the NCEP-NCAR reanalysis product exhibits best agreement with the spatially averaged observed wind speeds (Figure 3a; and other comparison metrics [Chen et al., 2012]), it is used as the primary evaluation data set against which the AOGCMs are compared.

[14] In contrast to in situ observations [Chen et al., 2012; Guo et al., 2011; Jiang et al., 2010b; Fu et al., 2011], trend analysis on the ensemble average of simulations from each AOGCM indicate that output from only two models (CNRM-CM5 and CSIRO-MK3.6) exhibit significant declines in 10 m wind speeds over the period 1971–2005 (Table 2). BCC-CSM1.1 also exhibits a weak downward trend, but it is only −0.001 m s−1/decade, and is not significantly different from zero. The other AOGCM ensemble average time series exhibit increasing annual mean wind speeds over the period 1971–2005. The magnitude of these tendencies is statistically significant in output from CanESM2, GISS-E2-H, IPSL-CM5A-LR, and MIROC4h. Comparing the trends during the entire historical period (1850 to 2005) and the period from 1971 to 2005, we found that the wind speed changes in the AOGCM simulations are more marked over the most recent decades (Table 2). To further investigate the origin of the temporal trends and wind speed biases, we analyzed output from six of the AOGCMs for which simulations were conducted with conditional application of radiative forcing. The results indicate that spatially averaged annual mean wind speeds are very similar in simulations with changing greenhouse gas concentrations only and with only natural forcings (e.g., volcanoes and solar variability; Figures 3b and 3c). Thus, simulations conducted with either application of forcing from greenhouse gas concentrations or natural forcing exhibit similar overall magnitudes of wind speed, and interannual variability therein. The majority of simulations with either only natural forcing or only GHG forcing do not exhibit significant wind speed trends (Table 2), even for models that exhibited significant tendencies in simulations conducted with the full suite of climate forcing. Thus, it appears that the AOGCM simulations of historical wind climates over China are not very sensitive to the climate forcing applied, or to the initial conditions, but rather are a stronger function of the specific model used.

Table 2. Trends in Time Series of the Annual Mean Wind Speed From Each AOGCM for Each of the Different Experiments (expressed in m s−1 decade−1 )a
AOGCMHistoricalGreenhouse Gases OnlyNatural Forcing OnlyFuture (2006–2100)
  • a

    Bold values indicate that trends are significant at the 95% confidence level. The last three rows show the temporal trends computed from the average of the in situ observations and the two reanalysis data sets.

CanESM20.0080.029−0.0016−0.0040.0009−0.0070.0060.001 −0.017
CNRM-CM5−0.002−0.0080.0001−0.00050.00060.00160.0040.002 −0.003
CSIRO-MK3.6−0.00002−0.012    0.0030.002−0.0006−0.002
GISS-E2-R0.00050.0020.0004−0.004−0.000340.0004 0.0020.0060.004
INMCM40.0020.001     −0.0004 0.001
IPSL-CM5A-LR0.0080.0280.0080.03−0.0010.005 −0.006−0.003−0.014
Observations −0.18        
NCEP-NCAR −0.14        
NCEP-DOE −0.06        
ERA40 −0.01        

[15] It is worthy of note that models from CMIP-3 also failed to reproduce the recent trend in observed near-surface wind speeds [Jiang, 2009], possibly indicating that (i) the observed downward trend is a manifestation of changes in surface roughness that are not included in the surface boundary conditions used in the climate models [Vautard et al., 2010], (ii) the current AOGCMs have only a relatively weak capability of representing some aspects of the atmospheric flow, or (iii) the observed tendencies are, at least partly, the product of non-climate-related factors (such as inhomogeneities in the station setting or instrumentation). The temporal trends in near-surface wind speed observations are not uniform in space [Chen et al., 2012; Jiang et al., 2010b], thus to examine the performance of the AOGCMs at the subdomain scale, we compared data from three representative grid cells in the NCEP-NCAR reanalysis data set with output from the closest corresponding cells of the AOGCMs (seeFigure 1for the approximate location of these grid cells). While output from the NCEP-NCAR data set exhibits declining trends in all three grid cells (all be it of substantially varying magnitude), output from the AOGCMs show both smaller magnitude upward or downward trends and greater spatial variability (Table 3).

Table 3. Trends in Annual Mean Wind Speed From Different AOGCMs and NCEP-NCAR Output in Three Different Regions in China During 1971–2005a
ModelTrend (m s−1 decade−1)
  • a

    Bold values indicate that the trends are significant at the 95% confidence level.

NCEP-NCAR−0.23 (95°E, 40°N)−0.035 (125°E, 47.5°N)−0.063 (112.5°E, 27.5°N)
BCC-CSM1.10.017 (95.63°E, 40.46°N)0.001 (126.56°E, 48.83°N)0.006 (112.5°E, 26.51°N)
CanESM2−0.005 (95.63°E, 40.46°N)0.038 (126.56°E, 48.83°N)0.076 (112.5°E, 26.51°N)
CNRM-CM50.002 (95.63°E, 39.3°N)−0.046 (125.16°E, 46.9°N)−0.019 (112.5°E, 27.3°N)
CSIRO-MK3.6−0.01 (95.3°E, 40.1°N)−0.001 (125.63°E, 47.56°N)−0.008 (112.5°E, 27.04°N)
GISS-E2-H0.016 (96.25°E, 41°N)−0.002 (123.75°E, 47°N)−0.009 (113.5°E, 27°N)
GISS-E2-R0.004 (96.25°E, 41°N)−0.012 (123.75°E, 47°N)−0.006 (113.5°E, 27°N)
INM-CM4−0.0002 (96°E, 39.75°N)0.017 (124°E, 47.25°N)−0.003 (112.0°E, 27.75°N)
IPSL-CM5A-LR−0.025 (93.75°E, 40.74°N)0.068 (123.75°E, 48.32°N)0.01 (112.5°E, 27.47°N)
MIROC4h0.016 (95.06°E, 40.16°N)−0.024 (124.88°E, 47.46°N)0.04 (112.5°E, 27.24°N)

[16] Wind climates vary on the time scales of hours, months, years and decades. Thus, a key component of model evaluation is assessment not only of the mean climate state but also the temporal variability. Accordingly, we evaluate the intra-annual and interannual variability of wind speed from the AOGCM simulations, reanalysis products and observed wind speed over China for 1971–2005 (Figure 4). In both cases, the variability is quantified as the standard deviation of mean annual wind speeds between years (interannual variability) and month to month variability (intra-annual variability). In contrast to results for Europe [Pryor et al., 2006], the magnitude of interannual variability is smaller in the AOGCM output and ERA-40 reanalysis than in direct observations or either the NCEP-NCAR and NCEP-DOE reanalysis output (Figure 4). The intra-annual variability (i.e., month to month variability) from the AOGCMs shows more divergent behavior with some AOGCMs exhibiting higher intra-annual variability than is manifest in the observations/reanalyses, while others show lower variability (Figure 4). As in the time series of annual mean wind speeds, simulations with individual AOGCMs that differ only in the initialization conditions exhibit very close accord in terms of intra-annual and interannual variability. The tendency for underestimation of interannual variability of 10 m wind speeds and more variable representation of intra-annual variability is also manifest in Regional Climate Model simulations over North America [Pryor et al., 2012b].

Figure 4.

Comparisons of simulated and observed temporal (interannual and intra-annual or monthly) variability of mean wind speeds during 1971 to 2005. The dashed lines are centered on the observations. The dot color indicates the model or reanalysis data set, and the individual dots depict the members from a given AOGCM.

[17] Wind speed seasonality over China differs between the direct observations and reanalysis data sets. The observations indicate a spring peak while all three analysis products exhibit highest wind speeds during the winter months (Figure 5). The spatially averaged monthly mean wind speeds among different realizations for a given AOGCM are very modest, and the difference between the minimum and maximum value of the mean wind speed in any calendar month across the realizations with a given AOGCM are all <5% of the mean value which further justifies use of an ensemble mean from each AOGCM. The springtime maximum is captured by the ensemble mean from some AOGCMs (e.g., CanESM2, despite the marked overestimation of intra-annual variability;Figure 4) but is not manifest in all models, e.g., CSIRO-MK3.6 exhibits greater consistency with the NCEP-NCAR reanalysis (Figure 5 and Table 4). This comparison illustrates one of the challenges in evaluation of AOGCMs: that discrepancies in the characterizations of the contemporary wind climate in different data sets leads to ambiguities in identifying the “target” against which skill should be assessed. This is discussed further below in the context of the upper level flow.

Figure 5.

Spatially averaged monthly mean wind speeds (1971–2005) over China from nine AOGCMs, NCEP-NCAR, NCEP-DOE, and ERA-40 reanalysis products and in situ observations. Note that in the case of ERA-40 the data period is 1971–2001 because the data product ends in 2002.

Table 4. Correlation Coefficients of Monthly Wind Speed From the AOGCMs Against In Situ Observations and Output From the NCEP-NCAR Reanalysis During 1971–2005a
  • a

    Also shown is the correlation between the multimodel ensemble (MME) and both the in situ observations and NCEP-NCAR. Bold values indicate that the correlations are significant at the 95% confidence level.


[18] The decadal variability of simulations during the historical period was assessed by setting the first decade of each simulation as the reference and comparing each subsequent decade with it. Thus, spatial patterns of wind speeds are computed for each AOGCM for each decade and climatological season. The results show that the AOGCMs generally indicate lower decadal variability (Figures 6a–6i) than that manifest in the NCEP-NCAR reanalysis data set (Figure 6j). This, in conjunction with the underestimation of interannual variability (Figure 4), implies that AOGCMs are staying too close to the mean state and may be suppressing the variability of wind climates associated with internal climate models (e.g., the Arctic Oscillation (AO) and El Niño–Southern Oscillation (ENSO) patterns [Chen et al., 2012] or lower-frequency modes, such as the Pacific Decadal Oscillation (PDO) [Goh and Chan, 2010]).

Figure 6.

(a–i) Taylor diagrams of decadal variability of seasonal mean wind speed over China for each AOGCM. The wind speed during the first decade is set as the reference. Each number indicates the wind speed in the subsequent decade (e.g., 2 means wind speed in the second decade, 3 is wind speed in the third decade, and so on). The colors indicate the seasons (green, spring; red, summer; magenta, autumn; blue, winter). (j) Comparable analyses of the NCEP-NCAR reanalysis.

[19] The spatial distributions of seasonally averaged wind speeds during 1971–2005 from all nine AOGCMs were also compared to those from the NCEP-NCAR reanalysis data set (Figure 7a). Although the spatial variability of wind speeds during winter is better simulated than in the other three seasons, model skill in simulating the spatial patterns of wind speeds is comparatively low. With the exception of MIROC4h, the AOGCMs underestimate the spatial variability of wind speeds relative to NCEP-NCAR in all four seasons (i.e., the ratio of standard deviation of the AOGCM to the NCEP-NCAR is <1, seeFigure 7a). The spatial correlation coefficients with NCEP-NCAR are also below 0.75 for all models in all seasons (Figure 7a).

Figure 7.

Taylor diagram of (a) near-surface wind speeds derived from the average of simulations conducted with a specific AOGCM and the multimodel ensemble (MME) against NCEP-NCAR output from 1971 to 2005. Height gradients calculated from the average of simulations conducted with a specific AOGCM against NCEP-NCAR output from 1971 to 2005 over (b) western China (at 500 hPa) and (c) eastern China (at 700 hPa). The colors show the seasons (green, spring; red, summer; magenta, autumn; blue, winter), while the numbers indicate the specific AOGCM (or a hash mark for the multimodel ensemble, MME).

[20] Differences in near-surface wind climates documented above may originate from many sources. The reanalysis output may differ due to different data assimilation approaches and the output may differ from reanalysis products or from AOGCMs due to factors such as different surface characterizations or different descriptions of the large-scale dynamics. With respect to the latter, it is worthy of note that the two versions of the GISS model (GISS-E2-H and GISS-E2-R) differ only in terms of the oceanic model [Hansen et al., 2007], and simulations with these models exhibit a high degree of accord with respect to the temporal trends (Table 2), mean wind speeds (Figure 3), interannual and intra-annual variability (Figure 4) and seasonality (Figure 5) of wind speeds over China. However, as shown in Figure 7a, output from these two model versions differ in terms of the spatial fields of springtime wind speeds over China, which may be causally linked to the differing representation of key teleconnection indices (e.g., ENSO and AO).

[21] Use of multimodel ensemble averages, where several independent models are combined, have generally been shown to exhibit enhancement of skill in seasonal forecasts (see examples given in Tebaldi and Knutti [2007]). However, the multimodel ensemble does not exhibit greater accord with the NCEP-NCAR seasonally averaged spatial patterns of wind speed variability over China than the individual AOGCMs (Figure 7a).

[22] Following work by Troccoli et al. [2012]and others, as a first diagnostic of possible causes of the discrepancies between AOGCM derived wind fields and temporal trends therein relative to the reanalysis output, we examined geopotential height gradients from the AOGCMs relative to the NCEP-NCAR output at 500 hPa over western China (i.e., west of 105°E;Figure 7b) and at 700 hPa over eastern China (Figure 7c). Although the two GISS models and MIROC4h exhibit rather large discrepancies with the geopotential height gradients from NCEP-NCAR, most of the AOGCMs reproduce the seasonally averaged spatial patterns of gradients very well, indicating that these models accurately simulate the large-scale driving flow as depicted in NCEP-NCAR (Figure 7). Figure 8illustrates some of the causes of the discrepancy in the GISS models and MIROC4h relative to NCEP-NCAR. In the western part of China, the geopotential height gradient decreases from north to south in NCEP-NCAR but this feature is not manifest in GISS-E2-H (seeFigures 8a and 8b). Meanwhile over the eastern China, the geopotential height gradient derived from GISS-E2-H exhibit high consistency with NCEP-NCAR except the edge of east Tibetan Plateau (seeFigures 8a and 8b). This suggests that discrepancies in the GISS-E2-H depiction of the large-scale atmospheric flow might be caused by the topography of the Tibetan Plateau. For MIROC4h, the height gradients over western China are well captured but in eastern China, except in summer, the patterns of geopotential height gradients show large discrepancies with those from the NCEP-NCAR data set, particularly in northeast China and the lower reaches of the Yangtze River (seeFigures 8a and 8c). With the exception of MIROC4h and GISS-E2-H, the RMSD between gradients of geopotential height from the AOGCMs and NCEP-NCAR (Figures 7b and 7c) is generally (though not uniformly) smaller than for the seasonally averaged wind speeds (Figure 7a). Thus, the “observed” features of the upper air pressure gradients are better captured by the AOGCMs simulations than the near-surface (10 m) wind speeds. For example, CNRM-CM5 reproduces the seasonal variability and magnitude of the geopotential height gradient over China relative to NCEP-NCAR (Figures 7b and 7c) but fails to capture the spatial patterns of near-surface wind speeds (Figure 7a). This implies that the surface characterization (e.g., the surface roughness) in the AOGCMs [see Masson et al., 2003, Table 2; Dorman and Sellers, 1989, Table 3] may be an important cause of the discrepancies between the near-surface wind speed climates from the AOGCMs and NCEP-NCAR. For example, the roughness length of evergreen broadleaf forests is 0.13 h (h is a typical tree height) in the CNRM-CM5 land scheme but is set to 0.05 m in the Simple Biosphere Model applied in the NCEP-NCAR reanalysis.

Figure 8.

Spatial patterns of the height gradient in four seasons during 1971–2005 from (a) NCEP-NCAR reanalysis and (b) GISS-E2-H and (c) MIROC4h (units are 10−4 m s−2). Height gradients over western (eastern) China are plotted by cross (dots) and are computed at a geopotential height of 500 (700) hPa.

Figure 8.


Figure 8.


3.2. Projected Wind Speed Scenarios

[23] Projections of future wind climates over China from the seven AOGCMs for which future projection output is available exhibit only very modest differences relative to simulations from 1971 to 2005. The spatial distribution of mean wind speeds in the final 35 years of the current century (2066–2100) versus 1971–2005 exhibit a very high degree of similarity in terms of the correlation coefficient, RMSD and ratio of the standard deviation for all models (Figure 9). Except CSIRO-MK3.6 and GISS-E2-R, all models exhibit similar or smaller wind speed spatial variability by the end of the century. The patterns for the different RCPs are also very similar indicating that based on these simulations, there is only a weak response to the range of climate forcing considered (of approximately 2 to 9 W m−2).

Figure 9.

(a–g) Taylor diagrams of wind speed in each season for each AOGCM simulation. The comparisons are undertaken for the climate change projection period relative to the historical period (i.e., 2066–2100 versus 1971–2005). Numbers represent the different climate scenarios (1, RCP2.6; 2, RCP4.5; 3, RCP6.0; 4, RCP8.5), while the colors show the seasons (green, spring; red, summer; magenta, autumn; blue, winter).

[24] Temporal trends in domain averaged wind speeds (2006–2100) computed for all four RCP scenarios are also very modest (Table 2), indicating that the near-surface (10 m) wind climate will remain relatively stable over the current century. This is consistent with previous research on CMIP-3 output that reported that under the A1B emission scenario the frequency of extreme wind speeds at 850 hPa over China during 2095–2100 do not differ significantly from 1995 to 2000 except in a small area of the Tibetan plateau where the 95th percentile wind speed decreased by 4% per degree of warming [Gastineau and Soden, 2009]. An example of the spatial distribution of the changes in projected mean wind speeds for each AOGCM is given in Figure 10 for the RCP4.5. As shown, there is no consistency between the AOGCMs with respect to the regions of upward and downward tendencies, but all AOGCMs exhibit some grid cells which are characterized by declining values through the projection period, and some grid cells that exhibit increases (Figure 10). The model-to-model variability is demonstrated by the following: Consistent with trends from spatially averaged wind speeds given inTable 2, except over the Tibetan plateau, wind speeds from CSIRO-MK3.6 exhibit significant positive trends. CNRM-CM5 exhibits upward trends for the spatially averaged wind speeds (Table 2), which derive primarily from the center of eastern China (Figure 10). In the case of IPSL-CM5A-LR, the spatial coverage of grid cells that exhibit declines is much greater than that which shows increasing wind speeds (Figure 10), leading to the domain averaged trends toward declining wind speeds (Table 2).

Figure 10.

Spatial patterns of the magnitude and sign of the significant trends from the CMIP-5 AOGCMs during 2006 to 2100 under RCP4.5. The color depicts the AOGCM, while the size and direction of the triangles shows the magnitude and direction of the trend (upward or downward) in units of %/yr.

[25] For most AOGCMs, the increasing (decreasing) tendencies weaken (intensify) as the greenhouse gas forcing is enhanced. However, the tendencies are generally not statistically significant and the sign of response is generally not consistent between AOGCMs or across RCP scenarios (Figures 9 and 11b). The observation that increasing radiative forcing does not give a proportional response in wind climates over China (see Table 2) may be linked to the importance of teleconnection patterns to wind climates over China, and the nonlinear response of those climate modes to radiative forcing. For example, an observed feature of the CMIP-5 AOGCMs is that the intensity of the Central Pacific ENSO “increases steadily from the preindustrial to the historical and the RCP4.5 simulations,” but the intensity of Eastern Pacific ENSO “increases from the preindustrial to the historical simulations and then decreases in the RCP4.5 projections” [Kim and Yu, 2012]. For the RCP2.6 only CanESM2 exhibits a mean wind speed in 2066–2100 that differs significantly from 1979 to 2000 according to the t test applied at a confidence level of 95%. For RCP4.5 the three of the seven AOGCM exhibit a significantly different value (but the direction of change is inconsistent between the models). Simulations under the two higher RCP exhibit a greater fraction of significant differences, with the majority of the AOGCM exhibiting lower mean wind speeds in the future period. The synthesis of these analyses is thus: while the RCP does influence the change in wind speeds over China, the sensitivity of wind speeds from these AOGCMs to the greenhouse gas concentration and radiative forcing is small compared to differences in wind speeds from the different AOGCMs (see Figures 3, 10, and 11).

Figure 11.

(a) Comparison of the projected interannual variability of wind speed from 2066 to 2100 and historical mean wind speed from 1971 to 2005. (b) Comparison of the climate change signal (i.e., difference in mean wind speed in 2066–2100 versus 1971–2005) and the historical mean wind speed from 1971 to 2005. (c) Comparison of the projected interannual variability of wind speed from 2066 to 2100 and interannual variability in the 1971–2005 simulation. (d) Comparison of the projected intra-annual variability of wind speed from 2066 to 2100 and intra-annual variability in the 1971–2005 simulation. The colors indicate the specific AOGCM.

[26] We investigated whether the wind speed response to radiative forcing was linked to the magnitude or variability of modeled wind speeds in the historical period (Figure 11). In part because wind speeds are zero bounded, AOGCMs that were characterized by higher mean wind speeds in the historical period also tend to be associated with higher interannual variability in both the historical period and the future. There is also some evidence that enhancement of the radiative forcing in the future is associated, at least in the majority of the AOGCMs, with increased interannual variability (Figure 11a). However, the projected change in mean wind speed 2066–2100 versus 1971–2005 for a given AOGCM does not show any association with the mean wind speed during the historical period (1971–2005) (Figure 11b). In addition, the projected interannual variability in 2066–2100 is only weakly correlated with modeled interannual variability in the historical period (Figure 11c). Conversely, the intra-annual (month-to-month) variability of domain-averaged wind speeds in the future (2066–2100) is highly consistent with the historical simulation (Figure 11d), but appears to be insensitive to the degree of radiative forcing.

4. Summary

[27] There remain relatively large differences between near-surface (10 m) wind climates from different AOGCMs. Analyses of near-surface wind climates over China from AOGCM simulations relative to direct observations and reanalysis output indicates (i) that the AOGCM initialization condition has little influence on the simulated wind speed and temporal variability during 1971 to 2005, (ii) that each of the nine CMIP-5 generation AOGCMs considered display both positive bias in mean near-surface wind speeds and negative bias in the interannual variability relative to reanalysis output and observations, and (iii) that no AOGCM exhibit uniformly best agreement with either the reanalysis products or the in situ observations. In order to investigate the origin of the AOGCM biases, we evaluate simulated wind speeds from historical runs with only GHG forcing and only natural forcing. The results indicate that the overestimation of wind speeds is present in both sets of simulations.

[28] In situ observations and output from the NCEP-NCAR reanalysis (and to a lesser extent NCEP-DOE) indicate declining wind speeds over the period 1971–2005, but this tendency is not reproduced by any of the AOGCMs. However, the AOGCMs reproduce at least some aspects of the seasonality of wind speeds over China and some models, e.g., CNRM-CM5, exhibit good agreement both with the magnitude and spatial variability of wind climates as manifest in the NCEP-NCAR reanalysis product.

[29] Multicentury simulations from the AOGCMs do not indicate any long-term tendency over the historical period or the current century. The spatially averaged and spatial fields of seasonal wind speeds for 2066–2100 exhibit very close accord with simulations from the AOGCM for the historical period (1971–2005). The mean wind speeds from each model computed for 2066 to 2100 do not show a substantial, consistent dependence on the degree of radiative forcing, although there is some evidence that the modeled interannual variability in the future period is somewhat higher under scenarios of stronger radiative forcing. The projected change in mean wind speed 2066–2100 versus 1971–2005 for a given AOGCM does not show any association with the simulated mean wind speed during the historical period. However, AOGCMs that exhibit high interannual and intra-annual variability in the historical period also tend to exhibit higher interannual and intra-annual variability in the future, but both inter and intra-annual variability appear to be relatively insensitive to the degree of radiative forcing. Thus, any climate change signal is considerably smaller than the discrepancies between the different AOGCMs during the historical period. The degree to which these findings are robust is hard to ascertain and they should be contextualized with information regarding the underestimation of historical variability in the AOGCM simulations and the large biases in the historical period.

[30] Results analyzed herein and previous research conducted using the prior generation of AOGCMs (i.e., those in CMIP3) [Jiang, 2009] indicate that in contrast to temperature, AOGCM derived near-surface wind climates exhibit substantial discrepancies with those obtained from direct observations and reanalysis products. The AOGCMs tend to more closely approximate the large-scale pressure gradients over China from the NCEP-NCAR reanalysis product than wind speeds therefrom, which may indicate that the discrepancies in wind climates derive largely from simulation of the atmosphere-surface coupling.


[31] L.C. and D.L. acknowledge financial support from Special Funds for Scientific Research on Public Causes (grant GYHY201006038), Innovation Project of Postgraduates Research in Jiangsu Province (grant CX10B_287Z) and Priority Academic Program Development (PAPD) of Jiangsu Higher Education Institutions. S.C.P. acknowledges financial support from the U.S. National Science Foundation (grant 1019603). We acknowledge the World Climate Research Programme's Working Group on Coupled Modeling, which is responsible for CMIP, and we thank the individual climate modeling groups for producing and making available their model output. For CMIP the U.S. Department of Energy's Program for Climate Model Diagnosis and Intercomparison provides coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals. The authors gratefully acknowledge the insightful comments of three anonymous reviewers.