Uncertainty in surface wind speed projections over the Iberian Peninsula: CMIP6 GCMs versus a WRF‐RCM

This study assessed the projected near‐surface wind speed (SWS) changes and variability over the Iberian Peninsula for the 21st century. Here, we compared Coupled Model Intercomparison Project Phase 6 global climate models (GCMs) with a higher spatial resolution regional climate model (RCM; ∼20 km), known as WRF‐CESM2, which was created by a dynamic downscaling of the Community Earth System Model version 2 (CESM2) using the Weather Research and Forecasting (WRF) model. Our analysis found that the GCMs tended to overestimate observed SWS for 1985–2014, while the higher spatial resolution of the WRF‐CESM2 did not improve the accuracy and underestimated the SWS magnitude. GCMs project a decline of SWS under high shared socioeconomic pathways (SSPs) greenhouse concentrations, such as SSP370 and SSP585, while an interdecadal oscillation appears in SSP126 and SSP245 for the end of the century. The WRF‐CESM2 under SSP585 predicts the opposite increasing SWS. Our results suggest that 21st‐century projections of SWS are uncertain even for regionalized products and should be taken with caution.


INTRODUCTION
Near-surface wind speed (SWS, ∼10 m above the ground) is a vital but understudied variable in climate research. 1It plays a significant role in climate system dynamics and can have far-reaching impacts on a variety of socioeconomic and environmental factors, 2 including evapotranspiration and precipitation, 3 wind-related disasters, such as tornadoes and extreme sea level rises, 4 air quality, 5 and the wind energy industry as a motor of decarbonization. 6,7evious studies have reported the observed reduction of global SWS (termed "stilling") before the ∼2010s 2 and a subsequent reversal in the last decade. 8Particularly, the stilling was detected over the Iberian Peninsula (IP), 9,10 while a cessation of this slowdown was recently reported. 1Although the causes behind the changes and variability of SWS have been discussed by the scientific community, 11 decadal-scale variations and long-term trends are probably affected by internal decadal ocean-atmosphere oscillations. 8,12obal climate models (GCMs) have been widely used to investigate climate change at various timescales, including decadal, multi-decadal, and centennial. 13Karnauskas et al. 14 employed 10 models from the Coupled Model Intercomparison Project Phase 5 (CMIP5) to project onshore wind energy and found a decline in future wind energy over the Northern Hemisphere.Carvalho et al. 6 investigated SWS projections over Europe under a multi-model ensemble (MME) of CMIP5 models with good SWS simulation abilities when compared to the European Centre for Medium-Range Weather Forecasts Re-Analysis Interim (ERA-Interim), noticing an increase in SWS over north-central Europe (maximum for Representative Concentration Pathway [RCP] 4.5) and a decrease over the Mediterranean sector (maximum for RCP 8.5) for the end of the century.Carvalho et al. 15 compared the 100 m wind speed projections over Europe using the CMIP5 and the CMIP6, finding discrepancies with the previous study: the CMIP6 models project a general decrease, particularly under SSP585 and at the end of the century (2081-2100) for the whole of Europe.At a global scale, Shen et al. 16 found that generally most of the CMIP6 GCMs could not represent the long-term observed SWS change, and selected the CESM2 as the best performance model.In addition, a decrease in SWS was detected over Europe under all shared socioeconomic pathway (SSP) scenarios for the 21st century.
However, the current spatial resolution of GCMs is not sufficient to accurately simulate the regional-scale processes that influence SWS, particularly in areas with complex topography, such as the IP. 17 A higher spatial resolution is required to study the SWS changes and multidecadal variability over these regions, which can be achieved through the application of a dynamical downscaling by means of regional climate models (RCMs). 18As Jung and Schindler 18 proposed in their review of SWS projections, most studies suggest a general decline of SWS over the IP for the 21st century.For example, Santos et al. 7 used the Coordinated Regional Climate Downscaling Experiment to reproduce SWS over 28 onshore and offshore wind farms over the IP, showing a better performance than the CMIP5 models in simu- Given the crucial role that SWS plays in the climate system, its socioeconomic and environmental impacts, and the questions raised regarding the future SWS, this research aims to assess the ability of CMIP6 GCMs and a higher spatial resolution RCM in reproducing observed SWS across the IP, and compare the SWS projections during the 21st century.The paper is structured as follows: a description of the data and methods; the principal results; a discussion of the main findings; and the conclusions.

Observations
Monthly mean SWS data series from 86 observational stations (1985-2014) were used as a reference to assess the ability of CMIP6 GCMs and RCM in reproducing the historical SWS changes and variability over the IP (including the Balearic Island in the Western Mediterranean basin).The 76 stations located over Spain were supplied by the Spanish State Meteorological Agency (AEMET), and the 10 stations across Portugal came from the Portuguese Institute for Sea and Atmosphere (IPMA).This dataset was examined and calibrated based on a robust quality control and homogenization process in the Climatol package v.3.1.2(https://CRAN.R-project.org/package=climatol;last accessed February 16, 2023). 21For a detailed explanation of the series, the data processing, and the location of the stations, please refer to Utrabo-Carazo et al. 1

CMIP6 GCMs
Monthly SWS data from 25 CMIP6 GCMs' historical simulations (1850-2014) and future projections (2015-2100) under four SSP scenarios were used to evaluate and project the SWS over the IP. 22These SSP scenarios are 126, 245, 370, and 585, corresponding to an end of the century radiative forcing of 2.6, 4.5, 7.0, and 8.5 W m −2 , respectively.The model's name, institution, and spatial resolution are listed in Table S1.Note that only the first realization output for each model was used (r1i1p1f1).In addition, we calculated the MME of all GCMs to remove the internal variability.

RCM: Dynamic downscaling with WRF
To explore the differences of using a higher horizontal resolution product, the Community Earth System Model version 2 (CESM2) has been dynamically downscaled using the WRF model.As Shen et al. 16 proposed, CESM2 has the best performance in reproducing the past terrestrial global SWS. 23The dynamic downscaling method was performed with WRFv3.8.1 using the ARW dynamic core. 24This was forced using the CESM2 model outputs every 6 h.The simulations have 40 vertical layers from the surface to 50 hPa and 480 × 800 grids with a horizontal spacing of 20 km centered on the North Atlantic Ocean.
The parameterizations employed in the WRF-ARW setup are as follows: the WSM6 microphysics scheme, 25 the Yonsei University PBL scheme, 26 the revised MM5 surface layer scheme, 27 the United Noah Land Surface Model, 28 shortwave and longwave RRTMG schemes, 29 and the Kain-Fritsch ensemble cluster scheme. 30Spectral nudging of waves longer than 1000 km was employed to avoid distortion of the large-scale circulation within the regional model domain owing to the interaction between the model solution and lateral boundary conditions. 31For the WRF-ARW simulations, a 1-month spin-up was performed prior to each year to be simulated and the restart mode was used when the WRF-ARW was stopped.We used monthly data for three periods: 1985-2014 for the historical simulation, 2036-2065 for the mid-term, and 2071-2100 for the long-term of the 21st century.The RCM output will be called WRF-CESM2 from now on; for a detailed explanation of the downscaling process, please check Fernández-Alvarez et al. 17 In addition, the SSP used for the WRF-CESM2 was SSP585, 32,33 as the downscaling has high computational demand.The SSP585 is a scenario (worst-case scenario) that represents emissions high enough to produce a radiative forcing of 8.5 W m −2 in 2100 under extreme conditions.

METHODS
Since the CMIP6 GCMs have different spatial resolutions, to be able to compare them we have remapped the data to a common 1 • × 1 • (longitude × latitude) grid using a bilinear interpolation technique. 34,35In addition, we have calculated the MME of the models.
For the evaluation of models in reproducing SWS over the IP, we compared the simulated against the SWS series for each station.We gridded the GCMs, MME, and WRF-CESM2 data to the location of stations by using a bilinear interpolation. 36For the comparison of the three datasets, we used the classic statistics of mean bias (MB; calculated as model minus observations) during the longest historical period (i.e., 1985-2014) over the 86 stations.In addition, we computed the probability density function (PDFs) of each dataset and calculated the overlap percentage (OP) against the observations' PDFs, following Perkins et al. 37 This method has the advantage that if a model can simulate the whole PDF, it can reproduce not only the mean but also the extreme events; previously used for SWS by Carvalho et al. 15 To assess if models can reproduce the multidecadal variability denoted in the observations, we calculated the monthly SWS anomalies series (with respect to the 1985-2014 climatology) for the historical period.
In these evaluations, we focus special attention on MME, CESM2, and WRF-CESM2 for comparison purposes.Finally, we computed the future SWS trends (in m s −1 dec −1 ) by applying linear regression analysis to the averaged SWS.The statistical significance of the spatial trends was given by a nonparametric modified Mann−Kendall's taub test to account for autocorrelation 38 ; with two levels (p<0.05 and p<0.10) to judge if the trend is statistically significant.We particularly emphasized two principal future periods, mid-term (2036-2065) and long-term (2071-2100).All these methods were applied annually and for boreal seasons (see Figures S1-S5 for seasonal results).The OP shown in Figure 1B highlights that models can generally reproduce the observed SWS distribution with most of the OP being higher than 75%.A concordance with the MB appears, with summer presenting higher OP as the bias is weaker.Lastly, the PDFs shown in Figure 1C also denote the strong underestimation of the WRF-CESM2 in reproducing the observed SWS, while the CESM2 and the MME (generally all CMIP6 models) overestimate SWS.In addition, both MME and CESM2 can reproduce the SWS density peak, between 3.5 and 4 m/s, while WRF-CESM2 simulates the SWS density peak at around 1 m/s.Similar results are found seasonally (see Figure S1), with the WRF-CESM2

Evaluation of models
showing a good performance in winter.
Comparing the statistics for the different models, no one outperforms clearly.Zooming in on the results for MME, CESM2, and WRF-CESM2, the ensemble mean has a slightly better performance.WRF-CESM2, even though it has higher spatial resolution than CESM2, does not improve the performance of the GCMs.
Figure 2 shows that the long-term evolution of observed SWS clearly experienced a stilling from 1985 to ∼2010 and that it was followed by a recovery.However, none of the MME, CESM2 (generally all CMIP6 GCM models), and WRF-CEMS2 captured the decadal changes and trends shown in the observations (see also historical trends for observations, MME, CESM2, and WRF-CESM2 in Figure 4).In addition, the magnitude of the anomalies varies among datasets.The SWS anomalies for CESM2 exceed the 25th to 75th percentiles (denoting higher anomalies than the CMIP6 models) and are in the same order of magnitude as WRF-CESM2.This contrasts with the high variability of observed anomalies, which vary around ± 0.4 m s −1 .Similar results are found seasonally (see Figure S2).

SWS projections
The projections shown in Figure 3  and especially SSP126, scenarios with a more optimistic vision of carbon emissions, projections present a slowdown until ∼2070 and a cessation or stabilization of the SWS decline until the end of the century, suggesting an inter-decadal oscillation.CESM2 compared with MME denotes some mismatches, especially for SPP245, presenting positive trends for the mid-and long-term.Seasonal results (see Figure S3) that are similar appear, with higher variability in winter compared with summer.

F I G U R E 3
Annual surface wind speed (SWS) anomalies (m s −1 ) of the MME (thick line), inter-CMIP6 models spread (shading the range between the 25th and 75th percentiles), and the WRF-CESM2 (under SSP585).The vertical dashed line divides the past (1985-2014) and future (2015-2100) periods under SSP126, SSP245, SSP370, and SSP585 scenarios.Abbreviations: CESM2, community earth system model version 2; CMIP, Coupled Model Intercomparison Project Phase; MME, multi-model ensemble; SSP, shared socioeconomic pathway; WRF-CESM2, Weather Research and Forcasting-CESM2.WRF-CESM2 under SSP585 keeps exhibiting higher SWS anomalies.Moreover, the historical and projected trends of this downscaled product present discrepancies compared to the CMIP6 models, which can be noticed in The spatial SWS trends of the WRF-CESM2 in Figure 5 show a weak or no significant decrease in most of the IP for the mid-term (Figure 5A), except for a small positive and significant trend area over the Strait of Gibraltar.The opposite occurs for the long-term (Figure 5B), with a dominance of positive and significant SWS trends.Seasonal maps (see Figures S4 and S5) show a marked seasonality, with increases in winter for both future periods and generally no trends in summer.

DISCUSSION
In this study, we assessed the projected SWS changes and variability over the IP, comparing for the first time the CMIP6 GCMs and a dynamical downscaled CESM2 model with WRF.We particularly focused on the MME, CESM2 (GCM), and the downscaled WRF-CESM2 (RCM).
First, we evaluated the abilities of the different datasets in reproducing the historical observed SWS magnitudes, PDF, and trends for the common 1985-2014 period.Although most models can reproduce the PDF, our results show that the CMIP6 GCMs consistently overestimate observed SWS changes, while the WRF-CESM2 underestimates SWS changes.It is challenging to compare our results against previous model evaluations as they strongly depend on each model, the data used as reference, and the location (e.g., Refs.7, 15, and 16).
Overall, there is no consensus among the scientific literature if models overestimate or underestimate SWS.
The products we used show a poor ability to simulate the long-term observed SWS changes and variability.In agreement with Shen et al., 16 CMIP6 models simulated SWS poorly.In fact, the GCMs reproduce a reduction in SWS for the historical period which is one order of magnitude weaker than the observed for the IP, 1 and the recent cessation (or even weak reversal) of SWS after decades of declining is not captured. 9nversely, the WRF-CESM2 does not better represent the past temporal variability than GCMs, and it simulates a nonsignificant increase of SWS for the historical period.Although Shen et al. 16 denoted that the CESM2 has the best abilities among all the CMIP6 models globally, we found that neither CESM2 nor WRF-CESM2 outperforms other models over the IP.

Discrepancies between the GCMs and RCMs and observations
could hypothetically be related to: (i) the coarse horizontal resolution, which causes models to oversimplify the complex terrain, leading to a poor reproduction of surface roughness and local-scale process that controls SWS, 39 particularly in areas with complex topography, such as the IP; (ii) instrumental errors associated with loss of cup anemometer performance or encoding issues that create biases in the observations; 40,41 and (iii) assimilation processes that do not get enough SWS observations, 36 which could be behind models' problems to simulate the observed multidecadal variability.More efforts are strongly needed to improve the physics of the models, the parametrizations, the spatial resolution, and the assimilation of observations.Particularly, our downscaled WRF-CESM2, although having a higher spatial resolution (∼20 km) than GCMs (100-250 km), does not outperform the CESM2 GCM.Other factors that can influence the poor performance of the WRF-CESM2 are the quality of the large-scale forcing data (i.e., CESM2) or that the domain is not centered over the IP.Shen et al. 16 proved that, globally, the CESM2 has the better performance in reproducing SWS, but other variables necessary for the downscaling with WRF (e.g., air temperature, near-surfacespecific humidity, or surface altitude) could not be well simulated, leading to biases in the RCM.Sub-grid orographic drag schemes, especially turbulent orographic form drag schemes, play a key role in model performance to reproduce near-SWS. 42The implementation of a turbulent orographic form drag scheme in WRF could significantly improve the model performance in capturing the SWS changes in areas of complex terrain. 42,43Different studies show that to avoid large biases, a correction should be made to the GCM bias with respect to the ERA5 reanalysis.This procedure has the potential to significantly improve dynamic downscaling simulations (e.g., Refs.44 and 45).
Our results show that the CMIP6 models project a consistent decline of SWS under high anthropogenic forcing scenarios (i.e., SSP370 and SSP585) for the whole 21st century over the IP.In contrast, the SSP245, and especially the "green path" of SSP126, project an SWS decline until ∼2070 followed by an interruption of the slowdown until the end of the century.These results are in agreement with previous studies over the Northern Hemisphere regions. 16,46To summarize, the higher the greenhouse gas emission scenarios, the higher the projected stilling for the 21st century.In contrast, positive SWS trends are predicted for mid-(2036-2061) and end-(2071-2100) of the century for the WRF-CESM2 under SSP585, the highest external forcing scenario.
8][49] The pole is warming faster than mid-latitudes so a reduction in this meridional difference will take place, 50 resulting in a weakening of the Hadley, Ferrel, and Polar cells that explain the projected stilling in SWS. 51 actual observed stilling versus reversal will have a stronger effect, 8,52 giving a reason for the projected interruption of SWS decline under optimistic 21st-century SSPs. 47e results of this study provide detailed information about SWS projections over the IP that might be useful for stakeholders.Nonetheless, we want to clarify the limitations of our results.SWS is a very complex and difficult atmospheric parameter to study, as it is dynamic and heavily dependent on the topography and surface roughness.Our SWS projections are subject to the difficulties of both GCM and RCM to reproduce observations, so they should be taken with caution.4][55] The use of a range of CMIP6 GCMs, MME, and multiple SSPs can reduce the uncertainties in models and internal variability. 18For the WRF-CESM2 (RCM), only the SSP585 was used, as the process is computationally demanding.Therefore, further research will focus on using a downscaling centered over the study region, using boundary conditions for an ensemble of CMIP6 models, and examining different anthropogenic forcing scenarios to show wider possible SWS realities.In future research, the representation of the wind field will be evaluated focusing on different planetary boundary layer (PBL) parameterization schemes, which play a dominant role in wind simulations.In addition, other key parameters tuning within the PBL schemes, such as the turbulent kinetic energy (TKE) dissipation rate, the TKE diffusion factor, and the turbulent length scale coefficients, will be considered.

CONCLUSION
The main findings of this study can be summarized as follows: lating SWS and a general SWS decrease for the near-(2019-2045), mid-(2046-2072), and long-term (2073-2099) period.Martins et al. 19 also detected an overall reduction of SWS and an increase in extreme events over 26 cities of the IP for 2080-2099 by using RCM forced with the Max Planck Institute Earth System Model-Low Resolution (MPI-ESM-LR) under RCP 8.5.Lastly, Claro et al. 20 have recently employed an ensemble of six CMIP6 models under SSP585 and the Weather Research and Forecasting (WRF) model with the initial conditions of the (Max Planck Institute Earth System Model-High Resolution, MPI-ESM-HR) MPI-ESM1.2-HR,finding an intensification of northern SWS along the western IP coast for 2046-2065, especially during summer.

Figure 1
Figure 1 presents the evaluation of the models' abilities to reproduce the observed SWS magnitude and distribution of values.The positive MB (Figure 1A) exhibited by most of the GCMs indicate an overestimation of SWS against the observations, while WRF-CESM2 shows an opposite, negative MB.A weak seasonality on MB is also denoted: that is, summer presents weaker biases compared to the rest of the seasons and annually, suggesting a better performance during this season.

F I G U R E 2
for the CMIP6 models and the MME under high anthropogenic forcing scenarios (i.e., SSP370 and SSP585) exhibit a clear decline of SWS for the whole 21st century.For SSP245 F I G U R E 1 Performance statistics between observed and simulated surface wind speed (SWS) series: (A) mean bias (m s −1 ).Names of the models are to the left of the figure; (B) overlap percentage (%); and (C) annual probability density functions (PDFs) of SWS (m s −1 ) over the location of the stations.Abbreviations: CESM2, Community Earth System Model version 2; DJF, December, January, and February; JJA, June, July, and August; MAM, March, April, and May; MME, multi-model ensemble; OBS, observed; SON, September, October, and November; WRF-CESM2, Weather Research and Forcasting-CESM2.Annual surface wind speed (SWS) anomalies (m s −1 ) of the MME, CESM2, WRF-CESM2, and OBS over the stations for 1985-2014.The shaded area indicates the uncertainty (25th to 75th percentiles) of the CMIP6 GCMs.Abbreviations: CESM2, Community Earth System Model version 2; CMIP, Coupled Model Intercomparison Project Phase; GCM, global climate model; MME, multi-model ensemble; OBS, observed; WRF-CESM2, Weather Research and Forcasting-CESM2.
a) The CMIP6 GCMs and the WRF-CESM2 RCM poorly perform the simulation of the observed SWS changes and multidecadal variability over the IP(1985-2014).Despite the improvement in spatial resolution, the WRF-CESM2 does not outperform the CESM2 (GCM).b) Under high anthropogenic forcing pathways (i.e., SSP370 and SSP585), the CMIP6 GCMs project a continuous decline in SWS, while under SSP245 and SSP126 show an interdecadal oscillation over the IP at the end of the century.Contrarily, the WRF-CESM2 projects a reinforcement on SWS for mid-and long-term in the 21st century.c)The projections are subject to the abilities of models to reproduce the historical period, and due to the large uncertainty found they should be taken with caution.Further efforts are strongly needed to improve the parametrizations and assimilation in GCMs and RCMs for accurately simulating SWS.AUTHOR CONTRIBUTIONSM.A.-M., C.A.-M., and C.S. conceived the study and designed its implementation.M.A.-M., C.S., and J.C.F.-A.performed the analysis and drafted the figures.M.A.-M., C.S., and J.C.F.-A.wrote the first draft of the manuscript.All authors edited, revised, and approved the final version of the manuscript.ACKNOWLEDGMENTS This research was funded by the 2021 Leonardo Grant for Researchers and Cultural Creators, BBVA Foundation.Our Climatoc-Lab also received funding support from the following projects: IBER-STILLING (RTI2018-095749-A-I00, MCIU/AEI/FEDER, UE) and VENTS (GVA-AICO/2021/023).The study was carried out in the framework of the CSIC Interdisciplinary Thematic Platform (PTI) "PTI+ Clima y Servicios Climáticos" and was also supported by the "Unidad Asociada CSIC-Universidad de Vigo: Grupo de Física de la Atmosfera y del Océano", the National Natural Science Foundation of China (42005023), and the Yunnan Province Basic Research Project (20220AT070403).The