Heterogeneous warming impacts of desert wind farms on land surface temperature and their potential drivers in Northern China

To address rapid climate change, wind energy has been widely developed in China in the last two decades. However, wind farm (WF) turbulence effects can change the local climate by redistributing temperature, humidity, and heat fluxes. Previous studies indicate that WFs can significantly increase nighttime land surface temperature (LST); however, their conclusions are mainly derived from individual WFs and ignore heterogeneous impacts among multi-WFs in China. Another large source of uncertainty is that the WFs used in previous studies are mainly located in croplands or grasslands, which may obscure direct WF impacts because of the interactions between vegetation and the atmosphere. In this study, we detect impacts with MODIS LST products during 2001–2018 at sixteen WFs in the desert of northern China. The results suggest that the averaged warming impacts of WFs on LST are similar between nighttime (0.237 °C) and daytime (0.250 °C). However, the uncertainty is much greater for daytime (SD = 0.519 °C) than for nighttime (SD = 0.146 °C) due to spatially heterogeneous impacts of desert WFs on LST. Optimal structural equation models suggest that wind speed, precipitation, and distribution patterns of wind turbines mainly explain the spatial heterogeneity of the desert WF impacts on nighttime LST. Given the rapid development of WFs globally, the local warming impacts of WFs and their corresponding mechanisms should be highlighted as a high priority in the fields of energy and climate.


Introduction
Wind energy plays a major role in renewable energy exploitation, which provides a massive amount of clean energy and reduces greenhouse gas emissions by fossil fuels (Veers et al 2019). The World Wind Energy Association reported that the sum of the global wind turbine capacity has been rapidly growing in the past two decades. It reached approximately 600 Gigawatts by the end of 2018, which could cover 6% of the total human electricity demand (World Wind Energy Association 2019). Moreover, there has been an exploding increase in wind turbine installation in China since the beginning of the 21st century, accounting for approximately onethird (210 Gigawatts) of the global wind energy by the end of 2018 (Chinese Wind Energy Association 2019).
In spite of the fact that the main purpose of rapid wind energy development is to reduce fossil fuel emissions and mitigate global warming, environmental side effects have also appeared as the number of wind turbines have largely expanded (Dai et al 2015, Tabassum et al 2014. Wind turbines usually generate wake turbulence by the rotating blades, which can redistribute the surrounding temperature, humidity, and heat fluxes (Roy and Traiteur 2010, Zhou et al 2012, Armstrong et al 2014. Therefore, the large-scale construction of wind farms (WFs) may change the local climate, which depends on the stability of the atmospheric boundary layer (ABL) (Armstrong et al 2014, Wu andArcher 2021). When the ABL is stable with cooler air near the ground and warmer air in the upper layer, which is more probable to happen in the nighttime, the rotations of wind turbine blades mix the warmer air layer and the air cooler near the surface and heat the land surface. However, when the ABL is unstable, with warmer air near the surface and cooler air in the upper layer, the WF impact on daytime LST is more complicated. Rotations of the wind turbine blades mix the cooler upper air and the warmer near-surface air and cool down the land surface (Miller and Keith 2018, Zhou et al 2020, Qin et al 2022, solar radiation heats the surface and creates upward convection. Besides, when the ABL is neutral, the temperatures of upper and lower air layers are approximate and the heat convection is near zero. The rotation has little impact on the surface. In addition to the ABL process, the conversion of the kinetic energy of the wind into the electric power of the wind turbines also produces massive heat (Corten 2000, Nematollahi et al 2019, which may also contribute to the warming effects. Based on remote sensing time series, previous studies indicated that WFs can significantly increase nighttime land surface temperature (LST) (Zhou et  Although there are some preliminary conclusions in terms of the WF impacts on LST, two main aspects should be further considered to obtain a clear picture of the driving processes. First, previous conclusions were mainly derived from individual WFs in China, which ignored background environmental effects. For example, the magnitude of the local wind speed may affect WF impacts by altering the speed of their wind blade rotation (Abo-Khalil et al 2019, Tahir et al 2019). The other uncertainty source is that the WFs used in previous works were mainly located in croplands and grasslands, which may obscure the direct WF impacts because of the interactions between vegetation and the atmosphere (Feng et al 2016, Grossiord et al 2020. For example, human irrigation in croplands can cool the air temperature by evapotranspiration (Payero et al 2008, Kurylyk et al 2014), which will weaken the WF impacts on LST observed by remote sensing. To minimize the effects of vegetationatmosphere interactions, desert WFs with no vegetation cover are the optimal study areas for detecting the direct WF impacts on LST.
Considering the two main uncertainty sources, in this study, 16 desert WFs (with100 wind turbines in each WF) were selected to evaluate the WF impacts on local LST in northern China. Based on a remote sensing time series, we analyzed the WF impact on the local daytime and nighttime LST by comparing the WF areas and their surrounding control areas (buffer). Then, the spatial distribution of the impact was evaluated at a grid scale in both the WFs and buffers. Furthermore, based on the structural equation model, we ultimately detected the possible environmental drivers of the spatial heterogeneity of desert WF impacts on daytime and nighttime LSTs.

Study area
In this study, we extract 7077 wind turbines distributed in 16 desert wind farms (WFs) through the deep learning algorithm You Only Look Once (YOLO) (Zhang et al 2020) in 2018, which are the world's largest desert WF group. YOLO is a fast, high-efficient, and high-precision object detection approach based on a single neural network (Redmon et al 2016). The number of wind turbines installed at the WFs vary from 100 to 1965 (figure 1(a) and table S1). Shuttle Radar Topography Mission (SRTM) datasets (Jarvis et al 2008) are used to represent the elevation (figure S1). The mean annual Moderate Resolution Imaging Spectroradiometer (MODIS) normalized difference vegetation index (NDVI) of the 16 WFs is lower than 0.1 (table S1). Based on the SoilGrids datasets (Hengl et al 2017), there are 4 soil types among the 16 WFs, and 12 WFs are covered by gypsisols. The following analyses are based on 16 WFs, and further analyses have been made on 12 gypsisols WFs to restrict the potential thermal property differences brought by soil components.
To detect the WF impact, a comparison strategy is widely used between WF pixels and their surrounding control region (buffer) (Zhou et al 2012, Slawsky et al 2015, Tang et al 2017. The WF areas are extracted based on 1 km * 1 km pixels to ensure that every pixel contains at least one wind turbine. The buffer is built as 1 km * 1 km pixels 5 to 9 km outside the WF to avoid any air turbulence influence caused by wind turbines ( figure 1(b)). The wake effect of wind turbines generates turbulence that spreads for kilometers downwind, the buffer should be outside the wake range. Meanwhile, the buffer should not be too far from the WF to share a similar climate background (Zhou et al 2012, Tang et al 2017, Qin et al 2022. The desert pixels are finally filtered by MCD12Q1 IGBP land cover data (figure S2), the filtered pixels are defined as wind farm pixels (WFPs) and buffer pixels (BUPs) in the following text.

Land surface temperature
To explore the spatial heterogeneity of WF impacts on desert LST, we use the Moderate Resolution Imaging Spectroradiometer (MODIS) MOD11A2 Land Surface Temperature (LST) time series between 2001 and 2018. The temporal resolution of MODIS LST is 8 days, and the spatial resolution is 1 km. Furthermore, MODIS provides both daytime (10:30 AM) and nighttime (10:30 PM) LST products (Wan, Hook, and Hulley 2015), which can help us better understand the WF impacts on the local climate. Further, we use MODIS Aqua LST time series of MYD11A2 as a supplementary test, the spatial and temporal resolutions are the same as MOD11A2, and the overpass time is 1:30 AM in the nighttime and 1:30 PM in the daytime.

Environmental factors
Four kinds of related environmental factors are used to explain the spatial heterogeneity of the WF impact on desert LST. First, the WorldClim precipitation and Tropical Rainfall Measuring Mission (TRMM) 3B43 monthly precipitation datasets are used. The WorldClim precipitation datasets are downscaled from the Climatic Research Unit (CRU), the spatial resolution is 2.5 arc minutes, and the temporal resolution is monthly (Fick and Hijmans 2017). The TRMM precipitation product algorithmically merges microwave data from multiple satellites. The spatial range is 50°S-50°N globally with a spatial resolution of 0.25°*0.25° (Kummerow et al 1998). Second, wind speed (50 m above ground) is derived from the Modern-Era Retrospective analysis for Research and Applications Version 2 (MERRA-2) monthly datasets (Gelaro et al 2017), with a spatial resolution of 0.625°*0.5°. The height of the wind speed is close to the blades of the wind turbines. Third, topography factors, including elevation and surface roughness, are used. The elevation is derived from the SRTM DEM dataset with a spatial resolution of 90 m (Jarvis et al 2008). Surface roughness is provided by the Sentinel-1 Synthetic Aperture Radar (SAR) Ground Range Detected (GRD) dataset, which is updated daily at a spatial resolution of 10 m and the preprocessing is already radiometric and terrain-corrected (Torres et al 2012). Fourth, soil properties, including soil type and sand content, are obtained from the SoilGrids datasets (Hengl et al 2017), which are modeled and fitted from more than 230000 soil profile observations at a spatial resolution of 250 m.

Wind farm shape factors
In addition to environmental drivers, WF shape factors are also used to explain the different WF impacts on the desert LST. The WF shape factors include the shape index, patch density, landscape division index, and mean Euclidean nearest-neighbor distance, which can be calculated via the Fragstates platform (version 4.2.1) (McGarigal 1995). These factors describe the distribution of wind turbines within WF in different dimensions. First, the shape index represents the ratio of the perimeter and area of patches in a WF, which can be calculated as: where SI i is the shape index of the WF; i is the serial number of the WF; p i is the perimeter of the WF (m); and a i is the area of the WF (m 2 ). The shape index usually increases as the patches in a WF become more irregular and fragmented. Second, the patch density indicates the number of patches within 100 hectares, which can be calculated as: where PD i is the patch density of the WF; i is the serial number of the WF; N i is the number of patches in the WF; and a i is the area of the WF (m 2 ). The patch density increases when there are more patches in a WF. A patch is an independent group of WFs pixels using the 8-neighbor rule in our study. Third, the landscape division index is the divisive level of the WF, which can be calculated as: where LDI i is the landscape division index of the WF; i is the serial number of the WF; n is the number of patches of WFs; a ij is the area of patch j in the WF (m 2 ); and a i is the total area of the WF (m 2 ). The landscape division index is closer to 1 when the WF is more fragmented.
Fourth, the mean Euclidean nearest-neighbor distance is calculated as: where MED i is the mean Euclidean nearest-neighbor distance of the WF; i is the serial number of the WF; n is the number of patch pairs of WFs; and h ij is the distance to the nearest neighboring patch (m). The mean Euclidean nearest-neighbor distance increases when the nearest patches within the WF get further.

Soil thermal admittance
To explore the relationship between surface thermal properties and WFs LST impacts, we calculate thermal admittance of 16 WFs as follows: where m is thermal admittance by J m −2 s − 1/2 K −1 , C is heat capacity by J m −3 K −1 , k is thermal conductivity by J s −1 m −1 K −1 (Oke et al 1991, Runnalls andOke 2000), heat capacity could be estimated by: where f x represent the volume fraction of solid, water, and gas components, and C x are heat capacity of each component (Meyers 2002 where k sat and k dry are saturated and dry thermal conductivities of soil. S r is the degree of saturation, which is set to 0.7 in our study. k is the empirical fabric factor set to 3.55 (Côté and Konrad 2005). k sat and k dry could be calculated by: where k s and k w are thermal conductivities of solid and water components (Barry-Macaulay et al 2015). The thermal conductivities of clay, silt, sand, and water are 2.9, 2.9, 5.5, and 0.6 J s −1 m −1 K −1 . c and η are empirical soil type parameters that account for particle shape effect, which are set to 0.75 and 1.2. n is porosity and set to 0.43 in this study (Côté and Konrad 2005).

Wind farm impacts on land surface temperature
The spatial distribution of the WFPs and BUPs is close (∼5 km distance) for each WF ( figure 1(b)). Because they usually have similar regional background climate conditions, signals derived from the LST differences between WFPs and BUPs mainly represent the impacts of the WF (Zhou et al 2020). In this study, we used three different methods to determine the desert WF impacts on LST. The methods and the workflows of this study are given in figure 2. The first method is the trend in the LST differences time series on regional scale, which can be expressed as: LST WFPs i j is the mean annual LST in the WFPs in year j; i is the serial number of the WF; The second method is calculating the ordinary least square slopes and then the ΔLST between 2001 and 2018 at a grid scale. Then, we compare the pixel trends of LST between the WFPs and BUPs using a two-sample t-test Figure 2. Workflows of the set up of study areas, regional and grid scale analysis, and driving factors analysis in this study. at a significance level of 0.05. If the ΔLST in the WFPs are significantly larger than those in BUPs, they could be considered significant warming impacts of WF.
The third method is to compare the preconstruction and postconstruction periods of the WFs at a grid scale. We define the postconstruction period as 2016-2018 and the preconstruction period as [2001][2002][2003]. Then, we compare the Δ LST between the WFPs and BUPs using a two-sample t-test at a significance level of 0.05.

Driving factor analyses with structural equation modeling
To determine the possible driving processes of the spatial heterogeneity of the desert WFs on LST, we first build linear relationships between the trends in the ΔLST (daytime and nighttime) and 15 environmental factors (i.e., climate, terrain, soil, and shape factors) among the 16 WFs. The coefficients of determination (R 2 ) are used to evaluate the robustness of the relationships. Then, based on a priori knowledge, we use structural equation modeling (SEM) to find the main driving factors of the spatial heterogeneity of the desert WFs on daytime and nighttime LST. The SEM is a multivariate statistical model used for complex relationships between directly and indirectly observed variables by multivariate statistical techniques of factor analysis and path analysis (Maruyama 1997). During the model optimization, the individual path coefficients with p>0.05 are removed to obtain a minimum chi-square value (χ 2 ), Akaike information criterion (AIC), and maximum coefficient of determination (R 2 ). To eliminate the potential effects of soil properties, we also establish an SEM with the 12 WFs covered by gypsisols.

Results
3.1. Wind farm impacts on land surface temperature at a regional scale Based on the MODIS LST time series from 2001 to 2018 in the 16 desert WFs, the results suggest that average ΔLST between the WFPs and BUPs is significantly increasing for both daytime and nighttime, which indicates that WFs increase the local temperature. The averaged warming impacts of WFs on LST are comparable between nighttime (0.237°C) and daytime (0.250°C) (figures 3(a), (d)); however, the uncertainty is much larger for daytime (SD=0.519°C) than for nighttime (SD =0.146°C). This suggests that the WF impacts on nighttime LST are more consistent than those on daytime LST.
For daytime LST ( figure 3(b), figure S3), 10 of the 16 WFs show warming impacts, and 5 of those are significant (p<0.05). The range of the WF warming impacts is between 0.368°C at WF No. 3 and 1.456°C at WF No. 13. However, 6 WFs suggest cooling impacts on LST, which range from −0.016°C to −0.729°C. In contrast, the WF impacts on nighttime LST are more robust (figure 3(e), figure S4). Specifically, 14 of the 16 WFs show warming impacts, and 10 WFs are significant (p<0.05). The range of the WF warming impacts at nighttime is between 0.181°C at WF No. 3 and 0.543°C at WF No. 11. Only two WFs (No. 2 and No. 4) suggested nonsignificant cooling impacts on LST. Interestingly, the WFs with significant impacts on daytime and nighttime LST are mainly located at the center of our study area .
In addition to the WF impacts at an annual scale, we also analyzed the WF impacts on daytime and nighttime LST in different seasons (i.e., spring, summer, autumn, and winter). Similar to those at the annual scale, the WF impacts on daytime LST are more divergent than those on nighttime LST (figure S5). For daytime, the average ΔLST range from 0.063°C in spring to 0.244°C in winter. In contrast, for nighttime, the maximum impacts of the WFs are in summer (0.285°C) and the minimum impacts of the WFs are in autumn (0.188°C).

Wind farm impacts on land surface temperature at the grid scale
In addition to the regional scale, the ΔLST are also applied at the grid scale to detect the WF impacts on LST  4(m)). In contrast, WF No. 7 shows the most obvious cooling effect of −0.749°C ( figure 4(g)).
For nighttime, the averaged ΔLST is 0.231°C or 0.172°C, which is 30.92% or 21.03% larger in the WFPs than in the BUPs between preconstruction and postconstruction or 18 years periods (figures S6(b) (d)). The largest warming impact occurs at WF No. 11, the LST in the WFPs gets warmer 0.555°C faster than in the BUPs, while in WF No. 2, WFPs get cooler by −0.037°C ( figure 5). Furthermore, the differences between the preconstruction and postconstruction periods of the WFs at the grid scale also suggest similar patterns of the WF impacts on daytime and nighttime LST (figures S7, S8).

Driving factors of wind farm impacts on land surface temperature
To understand the potential mechanisms behind the different WF impacts on annual mean daytime and nighttime LST, linear regressions between the ΔLST and environmental factors are first established at the 16 WFs (figures S9, S10). For nighttime LST, the most robust relationship is found between the ΔLST and annual precipitation (R 2 =0.24). For daytime, the highest R 2 is observed between the ΔLST and the Δ surface roughness (R 2 =0.21). For other factors, the results suggest that the relationships with the ΔLST are weak. When we establish the relationships in the 12 WFs covered by gypsisols, the relationships become more robust, which indicates that soil types might also affect the spatial heterogeneity of the desert WF impacts on daytime and nighttime LST (figures S11, S12).
Based on a priori knowledge of WF affected LST, SEM is used to find the main driving factors for the spatial heterogeneity of desert WFs on daytime and nighttime LST. After model optimization based on χ 2 , AIC, and maximum R 2 , three factors, including the shape index, wind speed, and annual precipitation, explain 60% of the variation in the WF impacts on nighttime LST ( figure 6(a)). The path coefficients of the shape index (0.35, p=0.09) and wind speed (0.43, p<0.05) are positive, while the path coefficient of the annual precipitation is negative (−0.54, p<0.05). However, the three factors only explain 9% of the variation in the WF impacts on daytime LST ( figure 6(b)). Furthermore, none of the three path coefficients are significant (p >0.05), which also implies that the processes of the WF impacts on daytime LST are more complex than those on nighttime LST.
To further remove the possible influences of different properties of soil types, we also use SEM based on the 12 WFs covered by gypsisols. The results suggest that the three factors can explain 76% of the variation in the WF impacts on nighttime LST (figure 6(c)), and the path coefficient of the shape index is also similar to the SEM established with all the WFs. However, the SEM still could not adequately explain the variations in the WF impacts on daytime LST, although the coefficients of determination increased (R 2 =0.16, figure 6(d)).
In addition, we also apply the SEM in different seasons (figure S13). The results suggest that the three factors can well explain the variations in the WF impacts on night LST in spring (R 2 =0.54), autumn (R 2 =0.49), and winter (R 2 =0.54), but the results suggest a low explanation in summer (R 2 =0.13). The sign of the path coefficients is the same as the results on the annual scale. In contrast, this factor is still unable to well explain the variations in the WF impacts on daytime LST across the different seasons.   In comparison, the impacts of desert WFs are more stable (−0.033°C to 0.543°C) in the nighttime and more divergent (−0.729°C to 1.456°C) in the daytime. This phenomenon may be explained by the interactions between different land cover types and the atmosphere. For example, potential evapotranspiration in croplands and grasslands may increase under the warmer environment caused by WFs. The cooling effects of increased evapotranspiration may ease the WF impacts on daytime LST observed by remote sensing. The higher water contents of leaf and soil in grassland or cropland lead to slower land surface cooling than desert soil (Ceccato et al 2001). It should be noted that the previous conclusions of WF impacts on LST are derived from individual WFs with different climatic and environmental backgrounds. Therefore, the background conditions should be considered when comparing the impacts of WFs on different land cover types in the future. The warming impacts are higher in the nighttime in spring, summer, and autumn, but the opposite in winter (figure S5). These might be because of the enhanced likelihood of turbulent induced warming by the increase of ABL stability, which happens more likely in the nighttime. However, in winter, there is a higher probability of inversions by changes in snow cover, and the surface thermal properties are altered by higher albedo and various thermal admittances of snow and ice as well (Oke et al 1991). In this study, there is no obvious correlation between nighttime ΔLST and snow cover in winter, the winter ΔLST might be driven by other factors (figure S14).

Spatial heterogeneity of wind farm impacts on land surface temperature
Based on the optimal structural equation model, annual precipitation, annual mean wind speed, and shape index mainly explain the spatial heterogeneity of the WF impact on nighttime LST. First, the path coefficient of annual precipitation is negative, which means that WFs located in wetter regions with relatively higher evaporation levels may have weaker warming effects. In winter, the thermal properties of snow-and ice-covered surfaces are variable. For instance, the thermal admittance of new snow is low and can support faster warming than desert soil. However, after the snow turns into ice, the thermal admittance becomes higher than that of desert soil, allowing more heat conduction and storage heat and mitigates the observed warming impacts compared to that of the desert soil surface (Oke 2002). In our results, the 4 WFs (Nos. 1, 2, 3, and 4) with higher precipitation (10.38 to 19.91 mm) in winter show weaker warming effects than the remaining 12 WFs (figure S15). Second, the path coefficient of wind speed is positive, which means that WFs located in higher wind speed regions may have stronger warming effects. A higher wind speed can increase the rotation speed of the wind turbine's blades before it meets the rated power (Ragheb and Ragheb 2011); therefore, it may increase the wake turbulence by rotating the blades and then increasing the LST. Third, the path coefficient of the shape index is positive, which means that the WF impacts on LST become more obvious when the WF shape becomes more irregular and fragmentary. Meanwhile, the higher the shape index is, the more dispersed the wind turbines in WF will be, which would help alleviate the wake effects of turbines and mitigate the wind speed loss (De-Prada-Gil et al 2015), which might lead to a faster rotation speed. The path coefficients of the other three shape factors are not as high as the shape index because it is the only factor that can indicate the shape complexity within a single WF patch.
However, annual precipitation, annual mean wind speed and shape index cannot well explain the spatial heterogeneity of WF impacts on daytime LST. The Δ surface roughness between the WFPs and BUPs negatively correlates with the ΔLST, which implies that the WFs on surfaces with a lower surface roughness than buffer regions may cause stronger warming effects. The Δ surface roughness can explain 21% of the variations in all 16 WFs and 31% of the variations in the 12 WFs covered by gypsisols (figures S9(j), S11(j)). The relative surface roughness between WFPs and BUPs might modify the convection of sensible heat between the two regions (Zhao et al 2014, Manoli et al 2019. When the BUPs are relatively rougher, the heat generated by turbulence effects in the WFPs might be restricted to diffuse into BUPs and could lead to more obvious warming effects. Besides, higher surface roughness in the BUPs may lead to stronger verticle sensible heat flux and suppress surface warming (Potter et al 1987), which could make more obvious relative warming in the WFPs. In the nighttime, thermal admittance might affect ΔLST, the lower thermal admittance is in the BUPs, the higher ΔLST is more likely to happen because of low thermal sensitivity match. Moreover, when the Δthermal admittance is higher between WFPs and BUPs, the ΔLST might be higher (Oke et al 1991). The relationships between nighttime ΔLST and thermal admittance and Δthermal admittance are weak in our study (figure S16), which might be the result of two sources. The first source is the uncertainties brought by the parameters of thermal admittance calculation, which are given by literatures without experiments. The second source might be the uncertainty of SoilGrids datasets, which are rasterized by soil samples and machine learning algorithms. The sparse desert soil samples in Northwestern China might lead to higher uncertainty than other soil types.

Uncertainties, implications, and future works
In this work, we use the latest remote sensing products to detect WF impacts on LST, but uncertainties still remain that should be further studied in future research. First, all the wind turbines in each WF were built in the last two decades. However, there are still no datasets on the construction period and operation time for each wind turbine. Because the WF impacts on LST mainly occur when the wind turbines are running, lacking these data obscures the quantification of the WF impacts on LST in our work. Second, wind turbines with higher rated power may generate stronger weak turbulence than turbines with lower rated power (Fan and Zhu 2019). Therefore, providing the rated power for each wind turbine can help us better understand the magnitudes of the WF impacts on LST. Third, there is no obvious averaged difference between MODIS Terra and Aqua ΔLST in the nighttime (0.001°C), while it is higher in the daytime (0.083°C) (figure S17). The overpass time of satellite might affect observed ΔLST due to the lower ABL stability in the daytime. Fourth, although long-term remote sensing series are an optimal technological method for studying WF impacts, we still need in situ measurements and field experiments to evaluate our results and discover the driving mechanisms.
In previous studies, researchers have simulated WFs with Weather Research and Forecasting (WRF) models The terrain within WFs will influence the wind flow and shortwave radiation (Wood 2000), and further affect local ABL stability. However, the coarse spatial resolution of the present MODIS LST datasets is a not good match for analysis of terrain induced micro climate changes. Moreover, the ABL stability in complex terrain in WFs remains uncertain (Kit et al 2017). In the future, the impact of terrain could be further discussed with finer resolution datasets based on remote sensing or in situ measurement. The sky view factor might influence income shortwave and longwave radiation and further affect ΔLST in WFs (Oke 1981). The sky view factor of WF could be changed by terrain, turbine density, and turbine size, the effect of the sky view factor could be extracted in the future with in situ measurements.
The main aim of wind energy is to provide clean electricity and replace fossil fuels to mitigate climate change. However, WFs also increase the local LST based on our results and those of previous studies (Keith et al 2004, Zhou et al 2012, Li et al 2018. Furthermore, large-scale WFs in vegetated regions may also affect ecosystem dynamics (e.g., vegetation growth and soil carbon stability) by changing the local climate (Knapp et al 2002, Armstrong et al 2014 and affect animal diversity (Marques et al 2014, Dai et al 2015, Smallwood and Thelander 2008. The WFs reduce greenhouse gas emissions of fossil fuels and mitigate global warming trends while having localized side effects on the environment. Although there might be some disadvantages of desert WFs compared with those in grasslands or croplands, such as high electricity transport costs and potential mechanical damage caused by sand storms (Fiore andSelig 2016, Li et al 2018), desert WFs can still yield global energy profits by minimizing environmental costs. Therefore, the potential trade-offs in energy and ecosystems of WFs should be further evaluated in the future to estimate the realistic efficiency of wind energy.

Conclusion
Wind farms (WFs) can change the local land surface temperature (LST) through turbulence effects. In this study, we detect the direct impacts on daytime and nighttime LST of 16 desert WFs. Although the averaged impacts of daytime and nighttime LST are comparable, the impacts on nighttime LST are more convergent than those on daytime LST. The spatial heterogeneity of the desert WF impacts on nighttime LST can mainly be explained by environmental factors. However, the divergent impacts of desert WFs on daytime LST cannot be well explained by environmental factors. Our study is the first to provide the spatial patterns of direct WF impacts on LST, which greatly overcomes the potential uncertainties in previous studies induced by interactions between vegetation and the atmosphere. In general, our results provide a benchmark for parameterizing Weather Research and Forecasting (WRF) model processes. Given the rapid development of wind energy, quantifying WF impacts on the local environment besides LST should be considered a high priority in climate change research.