Quantify uncertainty in historical simulation and future projection of surface wind speed over global land and ocean

Quantifying uncertainty in simulations of surface wind speed (SWS) has significant implications for its applications. Here, we examine the SWS changes from the 6th coupled model intercomparison project (CMIP6) model outputs, and analyzed the simulation uncertainties in CMIP6 both in the historical period and future projections. The results show that the both trend and interannual variability of SWS are underestimated in the CMIP6. The SWS over most of the Northern Hemisphere will reduce by 4%–6% under the high emission scenario in the last 21st Century, whereas it will increase by 6%–10% over South America and Southeastern Pacific. Over land, the majority of projection uncertainties is dominated by model uncertainty, followed by the internal variability and scenario uncertainty. Over ocean, the simulation uncertainty is greatly influenced by model uncertainty and internal variability, with the scenario uncertainty accounting for around 20% of total simulation uncertainty in the late 21st century.


Introduction
The changes in surface wind speed (SWS) have significant impacts on the ecological environment, social life, industrial and agricultural development.Strong winds are easily to trigger grass land fires and dust events over arid and semi-arid regions (Kurosaki et al 2011, Abram et al 2021, Liu et al 2022, Wang et al 2023).Low SWS can readily lead to air pollution events in places with high emissions, compromising human health (Wang et al 2018, Fan et al 2020, Han et al 2020).Furthermore, with the aggravation of energy shortage and environmental problems in social development, wind energy resources have gradually attracted global attention.Rational development and utilization of wind energy is an important way to solve the energy shortage and achieve sustainable development (Pryor and Barthelmie 2011, 2021, Porte-Agel et al 2020).Because the estimation of wind energy is mostly interpolated by SWS, the accurate projection of SWS is also crucial for wind energy research (Jacobson andArcher 2012, Zeng et al 2019).
Based on the state-of-the-art global climate models from the 6th coupled model intercomparison project (CMIP6), many studies have shown that the global land SWS will decrease in the future, and the decrease of SWS would be higher in shared socioeconomic pathways (SSP) SSP5-8.5 than SSP2-4.5 (Wu et al 2020, Deng et al 2022, Shen et al 2022, Miao et al 2023b), which indicates that there will be less wind energy available under the high-emissions scenarios in the future.Although the ensemble average results based on multi models give the long-term trend change of SWS, it is worth noting that there are substantial numerical and geographical discrepancies for different models (Karnauskas et al 2017, Zha et al 2021a, 2021b), and that the multi model ensemble projections still have large uncertainty (Bichet et al 2012, Olson et al 2019, Deng et al 2021, Chen et al 2022).Therefore, in order to gain a deeper understanding of CMIP6 projections, it is necessary to conduct in-depth research on the simulation uncertainty.
For the historical simulations of CMIP6, the simulation uncertainty can be obtained by comparing it with observation directly.Previous studies have found considerable disparities in SWS long-term changes across land and ocean during historical periods.Specifically, the observed SWS has shown a significant decreasing trend in the past 50 years over land (Wu et al 2018, Zhang et al 2019, Minola et al 2024), and this decreasing has slowed down or even it started to increase in the past 20 years (Zhang and Wang 2020, Minola et al 2022), which is known as the SWS 'stilling' and 'recovery' phenomenon (Roderick et al 2007, Zeng et al 2019).However, recent studies show an increasing trend over ocean in the last 40 years (Zheng et al 2016, Young andRibal 2019).Although many studies have evaluated the performance of CMIP6 in simulating historical SWS, most of the studies focus on land SWS (Hueging et al 2013, Tobin et al 2016, Guo et al 2019, Tian et al 2019, Zha et al 2020), while most of the area on earth is covered by oceans (71%), so in order to maximize wind energy usage in the future, it is also important to investigate its changes over the oceans.Noting that the observed SWS trend differences over land and ocean, it is significant to explore the simulated SWS differences between land and ocean in CMIP6.
For the future projection, there are no observational data to verify the simulation uncertainty of CMIP6.As a consequence, the uncertainties are mainly explored from the following three sources, which are internal variability, model uncertainty, and scenario uncertainty (Hawkins and Sutton 2009).The El Niño-Southern Oscillation (Timmermann et al 2018), the Pacific Decadal Oscillation (Newman et al 2016), and other natural variations are examples of internal variability that is inherent to the climate system (Song et al 2014, Ghil andLucarini 2020).The internal variability includes not only interannual variability, but also decadal and multi-decadal variabilities in the climate system.Since various models react to the same radiative forcing in different ways, which is primarily caused by variations in the mathematical representation and parameterization techniques used by different models, and thus inducing model uncertainty (Asseng et al 2013).Scenario uncertainty refers to model differences responses to various anthropogenic forcing (Hawkins and Sutton 2011).Based on the above classifications, many studies have analyzed the uncertainty sources of various state-of-the-art model projections in precipitation (Zhou et al 2020), runoff (Miao et al 2023a), soil moisture droughts (Chen and Yuan 2022), sea level (Little et al 2015, Jin et al 2024), and so on.However, there have never been quantitative studies of uncertainty sources of SWS in CMIP6 projections.At the same time, there are few studies explore the consistency of CMIP6 multi model projection on SWS over globe, and we will fill the above gapes in this study.

Data
The observed SWS in the historical period mainly comes from two parts: ocean data and land surface data.The SWS over ocean are obtained from the International Comprehensive Ocean-Atmosphere Data Set (ICOADS), which contains observations from many different observing systems.There have been studies shown that the SWS over ocean are in excellent agreement between observations and satellites (Zheng et al 2016, Young and Ribal 2019).The ICOADS records observations of ocean SWS from 1800 to now and is updated near-real-time, while the ocean SWS from satellites are mainly started from the 1990s.Therefore, the SWS from ICOADS are adopted here to ensure the consistency of the study time with the land SWS.The SWS in ICOADS are gridded at 1 • × 1 • or 2 • × 2 • grids, in order to match the resolution of the different models in CMIP6, a resolution of 2 • × 2 • grids has been used here.The global land SWS are collected from two main sources: the SWS over China are obtained after homogenization based on geostrophic wind (Zhang and Wang 2023), and the SWS in other lands except China is collected by global surface summary of the day (GSOD).Because most of the stations in the GSOD were recorded starting in 1973, and CMIP6 historical period simulation ended in 2014, the study period of the historical simulation is 1973-2014.We collect all monthly observations and exclude stations with greater than 20% missing data, and there are approximately 4500 stations left finally.
The monthly SWS from all historical forcing experiments  and projections under the SSP1-2.6,SSP2-4.5 and SSP5-8.5 emission scenarios (2015-2099) from CMIP6 are examined, and there are 30 models left which provide both of historical and future simulations, detailed information of the selected models are shown in table S1 in the supplementary information.To ensure consistency, all the models are assigned the same code of 'r1i1p1f1' .This code signifies that all these models share the same realization (ensemble number) and are initialized using identical initial conditions.Additionally, the parameterizations of physical processes and the external forcing inputs are the same across these models.This allows for more reliable and comparable results when analyzing the SWS from these models.The model data are regridded to a standard grid size of 2 • × 2 • by bilinear interpolation function on two-dimensional grids to facilitate analysis.The period from 1973 to 2003 sever as the baseline of our analysis, when present future projections, we define 2020-2049 and 2070-2099 as the near term and long term period, respectively.We investigated the SWS over land in North America, South America, Europe, Africa, Asia, and Australia, as well as over the oceans including the Pacific Ocean, Atlantic Ocean, and Indian Ocean.The areas of these zones are shown in figure S1 in the supplementary information.

Methods
The interannual variability is calculated by the following steps.First, the original time series is detrended, which removes the long-term trend component.Next, a 9 year moving average is applied to the detrended data, capturing the decadal variations.Then, the detrended sequence is subtracted from the decadal variation sequence, resulting in the interannual variations.Finally, the standard deviation of the interannual variations is calculated, representing the magnitude of interannual variability.The linear trend is obtained by the ordinary least square regression method.
To examine the models' agreement in simulating SWS, the signal to noise ratio (SNR) and simulation spread are also examined.The SNR is estimated by dividing the projected mean of the model ensemble by its standard deviation, and a larger SNR indicates relatively low uncertainty.The simulation spread refers to the range between the highest and lowest simulated values, taking into account all models.The simulation spread of SWS at the 10th, 50th, and 90th percentiles are inspected in this study.
To quantify the uncertainty sources of SWS projection in CMIP6, we adopt the method developed by Hawkins and Sutton (2009).The specific details are described below: Each individual simulation was fitted with a fourth-order polynomial over 1973-2099 using ordinary least squares.The formula for the raw simulations X m, s, t is written as where m, s, t represents model, scenario, and year, respectively.x m, s, t stands for the smooth fit, i m, s for the reference SWS, and ε m,s,t for the residual, which denotes internal variability.The averaged variance of the residuals across multiple models serves as the definition of internal variability, where N m is the model numbers and var s,t (ε m,s,t ) signifies the variance for the model m over scenarios and time.Also, the internal variability can be obtained by calculating the standard deviation of single model initial-condition with large ensemble.Study has shown that the internal variability derived by the two methods are very close at large scales, such as the three oceans in this study (Jin et al 2024), and we adopt the method introduced above.The variance in the various model simulation fits is used to estimate the model uncertainty for each scenario, and the estimation of the model uncertainty component is based on the multi-scenario mean where N s is the number of scenarios.The variance of the multi model means for the three scenarios is used to compute the scenario uncertainty: The total variance is calculated as: The average of three scenarios and multiple models is used to obtain the mean change of all simulations in relation to the reference period Based on the above method, we can analyze the sources of simulation uncertainty of CMIP6 for SWS and obtain their respective contributions.

Historical simulation and projection of CMIP6
The SWS time series in observations and CMIP6 during the historical period are shown in figure 1.Compared to observations, the SWS in CMIP6 is overestimated over land while underestimated over the ocean.Generally, except for Australia, the observed SWS over the global land showed a decreasing trend of −0.08 m s −1 decade −1 from 1973 to 2014 (figure 1(a 1 )), which is known as SWS 'stilling' (Roderick et al 2007).Also, the decreasing trend ceased since the 2010 and started to show 'recovery' phenomenon as previous studies revealed (Zeng et al 2019, Yang et al 2021).However, the SWS in CMIP6 neither captured the stilling nor recovery phenomenon (figure 1(b 1 )), the mean trend of multi model ensemble results was −0.0072 m s −1 decade −1 , which indicate that there are large simulation uncertainties in CMIP6.The above defects also occur over the ocean, the observed SWS presented an increasing trend of 0.27 m s −1 decade −1 during the studied period (figure 1(a 2 )), while it was −0.013 m s −1 decade −1 in the CMIP6 (figure 1(b 2 )), and the spatial pattern of SWS trend is show in figure 2. It is worth noting that our previous studies show that the inhomogeneity of the observed SWS leads to the overestimation of its decreasing trend over China (Zhang and Wang 2023), compared to the trend of homogenized SWS (−0.030 m s −1 decade −1 ), the trend in CMIP6 multi-model ensemble results was −0.0066 m s −1 decade −1 , and it still significantly underestimated the SWS decreasing trend.The spatial pattern of SWS interannual variability in observations and CMIP6 are shown in figure 3. Similarly, the CMIP6 projection underestimated the SWS interannual variability, and this phenomenon was more significant over the ocean than over the land.The observed interannual variability were 0.18 and 0.43 m s −1 over land ocean, respectively, while which were 0.10 and 0.22 m s −1 in CMIP6.
The SWS anomalies under three SSPs in the 21st century and their uncertainties are shown in figure 4.
In different development scenarios, both of the future SWS over land and ocean will show a decreasing trend, the decreasing trend over the land will be −0.0024,−0.0017, and −0.0024 m s −1 decade −1 under SSP1-2.6,SSP2-4.5 and SSP5-8.5 scenarios, respectively.However, the decreasing trend over ocean is slightly faster, reaching −0.0016, −0.0037, and −0.0054 m s −1 decade −1 under the above three SSPs, respectively.Furthermore, even though the global SWS will show a downward trend in the future,  there is a high degree of uncertainty in the SWS changes under different scenarios.This uncertainty increases with the extension of projection time, and the estimated uncertainty under SSP5-8.5 is greater than the other two scenarios.
In order to present the changes in future SWS in more detail, we analyze the spatial changes of SWS under different SSPs during different periods compared to historical periods, and the results are shown in figure 5.In near term period, except for South America and small parts of Africa, wind energy resources in most of the land will decrease (figures 5(a 1 )-(c 1 )).However, this downward trend is also relatively weak, and which are mainly concentrated in the range of −2% to 0 over most Northern Hemisphere lands.For the ocean, the decreasing of SWS is mainly located over the Northern Hemisphere, while the Arctic ocean and the ocean over Southern Hemisphere show a slight increase.In the long term period, the SWS decreasing trend in the Northern Hemisphere will be more pronounced than in the near term period, and will become more pronounced as emissions scenarios increase (figures 5(a 2 )-(c 2 )).Under the high emission scenario of SSP5-8.5, its decreasing trend are mostly concentrated in −6% to −4% over Northern Hemisphere.However, the SWS over South America and Southeastern Pacific will increase at a rate of 6%-10%.The above results indicate and the global wind energy will southward shift in the last 21st century (Karnauskas et al 2017).

Model agreement for SWS changes
Model agreement in SWS trend under three SSPs during the near term and long term period are shown in figure 6.In the studied area, if more than twothirds (20) of the studied CMIP6 models have consistent trends, it is considered to have good consistency in simulating local future SWS trends.Overall, the CMIP6 multi model simulations of SWS trends have better consistency over ocean than over land, and which is also more consistent in long term period.In the near term period, the projected SWS in CMIP6 over land mainly have good consistency under three SSPs in North America, South America, and Asia (figures 6(a 1 )-(c 1 )).For long term period, the good consistency increased from the above regions to Eurasia (figures 6(a 2 )-(c 2 )).The specific proportion of trend consistencies in different regions are shown in figure S2.
Furthermore, we examine the SNR and simulation spread in different periods and scenarios among different models.The ratio of SNR in different future periods compared to historical periods is shown in figure 7.In general, the CMIP6 multi model simulations of SWS exhibit a higher SNR in the long term period compared to the near term period, and the SNR over the ocean presents more obvious spatial differences than that over land.The simulation spread of SWS at the 50th percentile in different future periods compared to historical period is shown in figure 8. Generally, the simulation results show stable patterns over land throughout the entire study period.However, the spatial differences over the ocean are more apparent, which is consistent with the findings of the SNR analysis.The simulation spread of SWS at the 10th and 90th percentiles is similar to the results obtained for the 50th percentile (figures S3 and S4), and we will not elaborate on this again.The above findings indicate that even though the multi models in CMIP6 have more consistent trends over ocean than over land, the simulation differences between the models are also greater over the ocean.In section 3.1, it is also found that the SWS underestimation of interannual variability in CMIP6 is more significant over the ocean than over the land.The high simulation    differences of SWS over the ocean may accentuate this phenomenon, and the specific reasons may be complicated and need to be further studied in the future.

Uncertainty decomposition of SWS
Figure 9 illustrates the decomposition of SWS's uncertainty into three different categories: internal variability, model uncertainty, and scenario uncertainty.Similarly, to analyze spatial differences in more detail, we quantify the uncertainty for three partitions across six continents and three oceans.For land, except for Australia, the composition of uncertainty in the other five continents is relatively similar, with model uncertainty accounting for the main source of uncertainty, followed by the impact of internal variability, and scenario uncertainty having the smallest impact.In Australia, the situation is somewhat different from other regions, in which the internal variability is relatively high compared to other continents, with a contribution of up to 40% in the near term period (approximately 20% in other continents).From the long-term time change, all continents show consistent changes, that is, as forecast time increases, the contribution of model uncertainty increases, as does the contribution of scenario uncertainty, while the contribution of internal variability decreases.However, both of scenario uncertainty and internal variability are relatively minor in the long term period, making up just around 5% of the overall contribution lastly.
For the ocean, the projection uncertainties are significantly different from the land: the internal variability and model uncertainty contribute most of the projection uncertainties in the near term period.With the increase of forecast time, the contribution of internal variability decreases, and the contribution of scenario uncertainty increases, which gradually stabilizes and reaching about 20% in the late 21st century.The proportion of model uncertainty shows a nonlinear increasing trend generally, and the spatial pattern of the specific three uncertainties in the near term and long term period is shown in figure 10.The above results show that the uncertainties in CMIP6's simulation of SWS are significantly different between land and ocean.Therefore, it should be cautious when predicting future SWS based on CMIP6 multi models.

Conclusion and discussions
In this study, we examine the model capacity of historical simulation and future projection of CMIP6 globally, and also analyze the future SWS changes, as well as the simulation uncertainty and sources of uncertainty, and come to the following conclusions: 1.Both of the trend and interannual variability are underestimated in CMIP6.The observed SWS trend from 1973 to 2014 were −0.08 and 0.27 m s −1 decade −1 over land and ocean respectively, while which are −0.0072 and −0.013 m s −1 decade −1 in CMIP6.The observed interannual variability in the studied period were 0.18 and 0.43 m s −1 over land ocean, respectively, while which were 0.10 and 0.22 m s −1 in CMIP6.2. The rich area of global wind energy resources will gradually migrate to the southern hemisphere in the future.That is the SWS will decline over the northern hemisphere, and the decline will accelerate as the emission scenario rises, which reach −6% to −4% over Northern Hemisphere under SSP5-8.5 in the long term period, whereas it will increase by 6%-10% over South America and Southeastern Pacific.Furthermore, the SWS in CMIP6 multi model estimations have more consistent trends over ocean than over land, the simulation differences between the models are also greater over the ocean.Compared with the near term forecast, the long term forecast results have better consistency.3. The projected SWS uncertainties in CMIP6 are significantly different between land and ocean.
For land, model uncertainty accounts for the majority of uncertainties over land (contributing from 60% to 80% over six continents), followed by the internal variability, and scenario uncertainty.The contribution of model uncertainty and scenario uncertainty grows as projection time increases, whereas the contribution of internal variability decreases.For ocean, the internal variability and model uncertainty contribute most of the projection uncertainties in the near term period, and the scenario uncertainty show a contribution of about 20% uncertainty in the late 21st century.Generally, the proportion of model uncertainty and scenario uncertainty show a nonlinear increasing trend with the increase of forecast time, and the contribution of internal variability will decrease.
By means of multi models in CMIP6, we use the widely used method proposed by Hawkins and Sutton (2009) to decompose the sources of uncertainty in future SWS projections.Because of the importance of wind energy to future energy policy making, its accurate projection and uncertainty analysis are also necessary.In this study, we find that there are some differences of the uncertainty sources contributions between land and ocean.Future work should integrate new techniques like statistical learning (Sippel et al 2019) and dynamical adjustment (Deser et al 2016) to further uncover the causes uncertainty differences over land and ocean.Furthermore, the underestimation of trend and interannual variability of SWS in CMIP6 during historical period may also affect the accuracy of future prediction results.To mitigate the potential uncertainty in future SWS simulations, the prediction results can be constrained by the principle detected in CMIP6 historical simulations and observations (Bai et al 2022, Zhang et al 2022), which requires further studies in the future.Moreover, this study investigates the projection uncertainties among CMIP6 multi models.In order to reduce projection uncertainties within the models, it is recommended to employ large ensemble simulations such as involving varying initial conditions and incorporating multiple physical processes in each model.The availability of a single-forcing large ensemble will allow us to explore rare events and gain a better understanding of the influence of internal variability on forced climate change (Rodgers et al 2021, Lin et al 2022).
In this study, we found that the multi models in CMIP6 cannot reproduce the observed SWS trends.The dynamic reasons for SWS involve differences in the atmospheric pressure gradient forces and frictional forces (Zhang et al 2019).Various factors such as changes in atmospheric pressure (Zhang and Wang 2021), surface roughness (Vautard et al 2010) can affect the simulation capacity of SWS.Therefore, in order to accurately simulate SWS, it is necessary to have a deeper understanding and analysis of the various physical processes within the boundary layer.
This implies that further in-depth research is needed in the future.

Figure 1 .
Figure 1.The surface wind speed (SWS) time series in observations (a1)-(a2) and CMIP6 (b1)-(b2) in the historical period.The first row depicts SWS changes over land, and the second row depicts SWS changes over ocean.

Figure 2 .
Figure 2. The spatial pattern of surface wind speed (SWS) trend in observation (a1-2) and CMIP6 (b1-2) during the historical period (1973-2014).The first column represents the SWS over land, and the second column represents the SWS over ocean.

Figure 3 .
Figure 3.The spatial pattern of surface wind speed (SWS) interannual variability in observation (a1-2) and CMIP6 (b1-2).The first column represents the SWS over land, and the second column represents the SWS over ocean.

Figure 4 .
Figure 4.The surface wind speed (SWS) anomaly (compared to historical period) of CMIP6 under three SSPs in the future.The first row and second row represent SWS over land and ocean, respectively.The green, yellow and red lines represent the SWS anomaly changes under SSP1-2.6,SSP2-4.5 and SSP5-8.5 scenarios, respectively.The shading indicates the standard deviations for all models.

Figure 5 .
Figure 5.The surface wind speed (SWS) changes in different future periods compared to historical periods.(a1)-(c1) represent the changes in SWS under different scenarios in the near term period, (a2)-(c2) represent its changes in the long term period.

Figure 6 .
Figure 6.Agreement in surface wind speed (SWS) trend under three SSPs during the near term period (a1)-(c1) and long term period (a2)-(c2), the colorbar represents the number of models with the same trend among the 30 studied models, and the areas where the model consistency is less than 20 are shown in gray.

Figure 8 .
Figure 8.The ratio of simulation spread of surface wind speed (SWS) at the 50th percentile for three shared socioeconomic pathways (SSPs) during the near term period (a1)-(c1) and long term period (a2)-(c2) compared to historical period.

Figure 10 .
Figure 10.The spatial distribution of the contribution of internal variability (a), model uncertainty (b) and scenario uncertainty (c) to total uncertainty in the near (first column) and long term (second column) period.