We selected the Ensemble Éolien Catalan wind farm, located west of Perpignan in south of France (Fig. ). The wind farm for the study was selected based on three main criteria: its location in France, which allows the use of the AROME model as atmospheric forcing ; the presence of a sufficiently large number of closely spaced turbines to investigate inter-turbine interactions; and its location within vineyards. This wind farm offers several other features that make it particularly suitable for our study: the proximity of a RADOME meteorological station to validate the simulated atmosphere and the combined influence of the Mediterranean Sea and the Pyrenees on meteorological dynamics.

The Catalan wind farm consists of 35 turbines, including 9 Vestas V80 (2 MW) and 26 Vestas V90 (3 MW), resulting in a total installed capacity of 96 MW. The turbines are irregularly arranged in approximately four east-northeast-oriented rows, forming a rectangular layout of about 5 km in length and 2 km in width (farm layout in Fig. c). The orientation of the rows is perpendicular to the prevailing regional winds, which predominantly blow from the north-northwest direction.

The area surrounding the Catalan wind farm is characterized by a complex landscape that combines coastal, mountainous and agricultural features. To the east and southeast we find the proximity of the Mediterranean Sea. To the north and west, the terrain rises gradually toward the Corbières and Pyrénées mountains, which are mainly covered by forested and shrubland areas, creating complex orography. The immediate vicinity of the Catalan wind farm is mostly occupied by vineyards (purple in Fig. ) with smaller patches of cereal crops and orchards. Further away, particularly toward the mountains, the landscape includes forest, Mediterranean shrubland and scattered rural villages. This combination of orography, vegetation types and proximity to the sea makes the region particularly suitable for studying wind farm–atmosphere interactions under realistic, heterogeneous conditions and local wind effects.

As reported in the literature, wind-turbine-induced meteorological variations are more pronounced under strongly stratified and stable atmospheric conditions. These situations also correspond to cases where wind turbine wake extends longer, enhancing interactions between turbines within the wind farm. To investigate such situations, we focus on the night of 7–8 April 2021, during which a strong near-surface cooling event occurred across all of France, leading to nocturnal frost (Fig. ). According to the FNSEA (French National Federation of Farmers' Unions), this episode destroyed up to one-third of the national vineyard production, corresponding to an estimated economic loss of nearly EUR 2 billion. In the Languedoc-Roussillon region, where the Catalan wind farm is located, losses reached approximately 70 % of vineyard areas, with local minima down to -5.5 °C reported by the president of the southern branch of the AOC (Appellation d'Origine Contrôlée; French protected designation of origin) federation. According to the AROME model, on 8 April at 05:00 UTC, the hub-height wind speed at the center of the wind farm was approximately 4.8 ms-1 from the north-northeast (24°), with sunrise occurring at 05:20 UTC.

These extreme temperature conditions resulted from the synoptic situation prevailing at the beginning of April 2021 in Europe, characterized by a persistent anticyclone positioned south of Iceland combined with a low-pressure system over the Iberian Peninsula. This situation generated clear skies and dry air over most of France, creating favorable conditions for intense radiative cooling during nighttime. Additionally, two intrusions of cold and dry polar air affected France during the first half of April, as reported by Météo-France in its April 2021 climate bulletin . These combined conditions resulted in cold nights, with temperatures well below seasonal averages and several new cold records across the country.

Meso-NH is a non-hydrostatic and anelastic research atmospheric model developed by the Centre National de Recherches Météorologiques (CNRM – CNRS/Météo France) and the Laboratoire d'Aérologie (LAERO – UPS/CNRS) . In this work, Meso-NH v5.7.1 is used to simulate the atmosphere from mesoscale to microscale, with the finest domains relying on an LES method. The objective of the numerical setup is to dynamically downscale atmospheric conditions from the French operational numerical weather prediction model AROME to resolutions compatible with the wind turbine actuator parameterization.

To achieve this, a numerical setup with four nested domains is used (illustrated in Fig. ). Their characteristics are provided in Table . The configuration is divided into two successive offline coupled simulations. The first simulation, referred to as the large-scale, includes domains D1 and D2, which are two-way-nested and forced hourly at the lateral boundaries of domain D1 by AROME analyses. These analyses have a horizontal resolution of 1.3 km and 90 vertical levels. The large-scale simulation is run for 33 h and allows the flow to be downscaled to a horizontal resolution of 125 m. Then, 5 min outputs from this simulation are used to initialize and force the second simulation, referred to as the small-scale, which includes domains D3 and D4.

This separation between the large-scale and small-scale simulation is motivated both by computational cost considerations and by the fact that Meso-NH does not allow vertical grid refinement between directly nested domains. This approach makes it possible to refine the vertical resolution only in the small-scale simulation, where the wind turbines are located. The domain D4 with a horizontal resolution of 5 m and a vertical resolution of 4 m encompasses the Catalan wind farm area. Its horizontal size was chosen to ensure that turbulence is fully developed before the flow reaches the wind turbines and the meteorological tower locations.

The complex topography and the large resolution jump prevent domain D4 from being directly initialized and run from the output of domain D2. Therefore, the small-scale simulation is first run using domain D3 alone for a 45 min spin-up period. The resulting atmosphere is then used to initialize domain D4. Finally, domains D3 and D4 are run in a one-way-nested configuration for a total duration of 55 min. To allow the small-scale atmospheric flow and turbine wakes to develop, the simulation includes a 25 min spin-up period, followed by a 30 min averaging period.

The Meso-NH model solves the non-hydrostatic system of conservation equations for momentum, mass, humidity and other scalar variables, as well as a thermodynamic equation derived from entropy conservation . A detailed description of these governing equations and the model's formulation is available in . In this work, the choice is made to solve the pseudo-incompressible system of equations as defined by to filter out acoustic modes and allow for an explicit time-stepping integration.

The prognostic variables include the three velocity components, the dry potential temperature and the mixing ratios of the water phases. The governing equations are formulated in flux form using the dry air density of a hydrostatic reference state at rest. Enforcing the corresponding continuity constraint requires solving an elliptic equation for the pressure perturbation.

These equations are formulated on a conformal projection plane to account for the Earth's sphericity and discretized horizontally and vertically on an Arakawa C grid using a height-based terrain-following coordinate system. Momentum transport is integrated in time with an explicit Runge–Kutta-centered fourth-order scheme paired with a centered fourth-order spatial discretization. Advection of meteorological scalars and water substances is handled via the piecewise parabolic method.

Subgrid-scale turbulence is parameterized using a three-dimensional, 1.5-order closure scheme based on a prognostic equation for the subgrid TKE , closed with a mixing length as defined by and . Convective-scale motions are explicitly resolved across the finest nested grids; however, for domain D1, shallow convection is parameterized using the eddy-diffusivity–Kain–Fritsch parameterization . Radiation forcing terms are represented by the ecRad modular radiation scheme and the ICE3 mixed-phase single-moment microphysics scheme, which predicts mass mixing ratios for the six water categories .

An online coupling with the SURFEX v9.0 surface transfer model is employed to finely model surface exchange with the atmosphere. Soil and vegetation effects are captured through the explicit multilayer model ISBA-DF , which solves the heat and mass transfer equations in the soil with 14 vertical soil layers. Heat transfer is computed using Fourier's law, with soil thermal properties depending on soil texture, moisture and ice content. Water transfer is described using the mass-conservative mixed form of the Richards equation, accounting for both liquid water transport and water vapor diffusion. The surface heat flux into the soil–vegetation composite is determined by the surface energy balance, which partitions net radiation into sensible and latent heat fluxes . Urban effects are modeled using the Town Energy Balance (TEB) model . The detailed formulations of these physical processes are provided in the SURFEX scientific documentation .

Usually, the SURFEX model is used with the global 1 km resolution ECOCLIMAP-II surface database . However, in our case, its level of detail is insufficient for the resolution of domains D2, D3 and D4, as well as for accurate representation of the Catalan wind farm area. Therefore, a high-resolution surface database was created by extracting surface data from OpenStreetMap using the GeoClimate tool . The surface classes provided in OpenStreetMap were then manually converted into the cover classes defined in ECOCLIMAP-II and given as input to SURFEX (Fig. ). While this method depends on somewhat arbitrary reclassification choices, it nevertheless provides a surface database that captures the fine-scale specification of the wind farm environment with a level of detail that does not exist in standard datasets, particularly at resolutions of 5 m.

However, transferring soil initial conditions from the 2-layer force-restore ISBA-FR model of AROME to a 14-layer soil representation is physically challenging, especially for deep soil temperatures. A strong sensitivity of the initial soil condition to the time of initialization was also identified. Technically, a multi-day surface spin-up would have been required to properly equilibrate the ISBA-DF soil profile, but this approach is computationally expensive. Instead, a simplified initialization was adopted, in which the 14 soil layers were initialized using the mean air temperature of the first AROME level over the previous 24 h. This method produced more realistic deep soil temperatures for April conditions and led to a nocturnal cooling closer to experimental observations (see Sect. ).

The 5 min averaged time series of some meteorological variables from the large- and small-scale simulations are validated using measurements from the meteorological station located at the Perpignan airport (Fig. ). This station is part of the operational surface observation network of Météo-France, known as RADOME (see position in domain D3 in Fig. ). Consequently, only a limited set of meteorological variables is available, measured at 2 and 10 m above ground level. Evaluation is performed for the set of meteorological variables available from the RADOME observation network at this site. Vertical atmospheric profiles derived from domain D2 are also shown in Fig. .

To quantitatively evaluate the performance of the simulation against observations, statistical validation metrics including mean bias error (BIAS), root mean square error (RMSE), systematic (RMSEs) and unsystematic (RMSEu) errors, and Willmott's index of agreement (d) were computed in accordance with the recommendations (Table ). The simulation demonstrates overall good agreement with experimental observations across most evaluated parameters. High values of Willmott's index of agreement (d) are obtained for almost all variables, reflecting strong model skill in capturing atmospheric dynamics. As shown in Fig. a–b, the 10 m wind speed and direction closely match the observed time series, successfully reproducing both the daytime convective regime and the post-sunset wind weakening. This is supported by high Willmott's index values for domain D2 (d=0.944 for wind speed and d=0.971 for wind direction). Furthermore, for 10 m wind speed in D2, the unsystematic error exceeds the systematic error (RMSEu=0.95ms-1>RMSEs=0.66ms-1), indicating that model discrepancies are predominantly driven by unresolvable random fluctuations rather than structural model deficiencies. The model also accurately captures the temporal variations in relative humidity (d=0.967) and shortwave downward radiation (d=1.000).

The main discrepancy between measurements and simulation occurs during nocturnal cooling. A warm bias of approximately 2–3 °C (RMSE of 2.60 °C) is observed in both the 2 m air temperature and surface temperature (solid and dotted lines in Fig. e). This bias is consistent with the driving AROME forecast (black dots in Fig. e). This nocturnal temperature discrepancy is directly reflected in the statistical metrics for temperature: Willmott's index drops to d=0.857, the lowest value among all evaluated variables, and the systematic error clearly dominates the unsystematic error (RMSEs=2.15 °C > RMSEu=1.47 °C). Several sensitivity experiments were conducted, including variations in the mixing-length formulation, soil model equations, soil temperature initialization and atmosphere–surface coupling coefficients. The results show a sensitivity of nocturnal cooling to these parameters; however, no combination was able to reproduce the observed cooling at the measurement site (not shown). Overall, the warm bias appears to be primarily linked to soil temperature overestimation, itself resulting from the initialization of subsurface temperatures in the ISBA-DF model. The revised methodology used to compute soil temperature, described in Sect. , improved nocturnal cooling but did not fully remove the bias. This highlights the complexity of modeling surface–atmosphere interactions and reproducing specific meteorological situations, even at high spatial resolutions .

This validation remains limited due to the availability of only a single near-surface measurement point and the lack of validation at rotor height in the absence of vertical-profile observations. Consequently, while near-surface measurements confirm the occurrence of nocturnal cooling, the exact vertical extent, inversion strength and thermal stratification of the real-world boundary layer cannot be quantitatively verified. The 2–3 °C surface warm bias introduces additional uncertainty regarding the exact magnitude of the simulated atmospheric stability compared to actual conditions. Nevertheless, the atmospheric conditions simulated here are consistent with the primary objective of this study: to simulate a physically consistent stable and stratified boundary layer to investigate its interaction with wind turbine and surface properties.

A terrain-following plan of horizontal wind speed Uh at hub height (80 m above ground level) in domains D1 and D2 is shown in Fig.  (a video is also available at 10.5446/73018; ). During the morning and early afternoon of 7 April, strong winds blow from the Languedoc coast toward the Mediterranean Sea (Fig. a). After sunset, the associated nocturnal cooling leads to a marked decrease in wind speed in the Perpignan area (Fig. b). Over the Mediterranean Sea, a mesoscale frontal zone develops as a result of the convergence of three air masses, further reinforcing the weak-wind conditions over Perpignan (Fig. c). Under these conditions, the local flow becomes primarily controlled by orography, with slope and valley winds dominating the atmospheric circulation (Fig. f). Between 20:00 and 04:00, wind speeds remain around 3–4 m s⁻¹, which is close to the cut-in wind speed of the turbines.

After 04:00, the frontal system progressively shifts over the Mediterranean. A strong marine flow then penetrates inland, guided by the Albères massif south of Perpignan. A recirculation pattern develops, leading to the formation of a mesoscale vortex over the Perpignan area (Fig. j–k). As a consequence, toward the end of the night, the wind farm is exposed to easterly to northeasterly winds, which enhance wake interactions along the turbine rows (see wind farm layout in Fig. ). This late-night period, characterized by a strong nocturnal cooling, a highly stratified atmosphere and wind speeds sufficient for turbine operation, represents a particularly interesting case for the present study.

Vertical cross-sections of domain D3 at 04:50 are shown in Fig. . This cross-section is aligned with the wind direction at hub height and focuses on the eastern and southern parts of the wind farm (see segment AB in the horizontal cross-sections), which are subsequently analyzed in Sect. . The wind turbine actuator disk model is not activated at this step; however, turbine locations are indicated to provide a clearer view of the spatial scales and extent of the wind farm. A pronounced horizontal and vertical heterogeneity is visible in the different meteorological fields. In the vertical cross-sections, an air mass originating from the Mediterranean Sea enters the domain from the right (Fig. b). This air mass is characterized by higher wind speeds, warmer temperatures and greater moisture content compared to the continental air. Because the land surface is cooler, the warmer marine air mass tends to overrun the cold and dry air near the surface (see bottom left of Fig. h and j). This configuration leads to global positive vertical velocities (Fig. f) and the development of strong wind shear at the interface between the two air masses. Such shear conditions are favorable to Kelvin–Helmholtz instabilities, which are noticeable at low altitude on the left side of the vertical cross-section by the presence of vortical structures in the wind and temperature fields (Fig. b and h), as well as by enhanced turbulence and alternating ascending and descending air motions (Fig. d and f).

At the surface, temperature also exhibits strong spatial heterogeneity: hilltops are relatively warmer, while valleys remain cooler, a pattern typical of thermally driven slope and valley circulations under stable nighttime conditions (Fig. g). Near-surface winds follow the slope orientation and are partially decoupled from the flow at hub height, indicating the presence of a stratified boundary layer regime. Surface relative humidity (Fig. i) further highlights the contrast between the two air masses: a moist marine air mass advected from the sea and a drier air mass descending from the mountains.

To characterize the vertical structure of the atmosphere inside the wind farm before turbine activation, 5 min averaged vertical profiles between 05:15 and 05:45 at the center of domain D4 are presented in Fig. . Across the rotor layer, wind speeds exceed the turbine cut-in value, ranging from approximately 4.5 to 6.0 m s⁻¹ (Fig. a), with a progressive speed increase observed over the 30 min study period. A strong directional shear (veering) is evident with altitude (Fig. b), alongside a gradual rotation toward easterly directions over time. The thermal structure (Fig. c) exhibits a sharp surface-based temperature inversion within the lowest 40 m (rising from 5.0 °C at the ground slightly above 7 °C). Relative humidity reaches its maximum near the surface (∼65 %) and stabilizes around 55 %–60 % across the rotor area (Fig. d). Finally, TKE values remain low (Fig. e), and the turbulent heat flux (THF) is strictly negative throughout the lower levels within the surface inversion layer (Fig. f), characteristic of downward sensible heat transport under strongly stratified nocturnal conditions.

The wind regime during the end of the night over the Perpignan wind farm is dominated by a mesoscale vortex, due to sea and mountain flows interactions. At the hub height, the wind blows from the north-northwest and exhibits a counterclockwise deflection between the upstream and downstream parts of the farm. As a result, the velocity deficit reveals wake directions that can differ by up to 45° within the wind farm, despite its relatively compact layout (Fig. a). Horizontally, wake expansion is enhanced by the gradual turning of the wind direction across the farm. Velocity deficits generated by upstream turbines impact downstream rows. Figure  shows the ratio between the mean power produced by the ADR and the maximum theoretical power output of the wind turbine. Power ratio is maximal for turbines located along the outer edges of the farm, unaffected by upstream wakes, whereas turbines within the interior produce around half of the power. This contrast is further reinforced by the fact that the background wind speed at southern wind turbine locations is weaker even in the absence of upstream wake effects (Fig. a).

A localized acceleration zone is visible near the ground beneath each rotor (Fig. b). Such sub-rotor accelerations have been previously reported in idealized LES studies , particularly under stable atmospheric conditions. This effect is attributed to turbine-induced pressure field adjustments that channel the flow beneath the rotor, producing a Venturi acceleration. The phenomenon is enhanced under stable stratification, where vertical motions that would otherwise allow the flow to pass the obstacle are strongly suppressed . At the wind farm scale, flow acceleration can also be observed above the farm (at around 600 m, not shown on the plot Fig. e) and, to a lesser extent, along its eastern lateral edges (Fig. a). These wind speed anomalies arise from the same mechanism as the sub-rotor acceleration but acting at the scale of the farm. This phenomenon, commonly referred to as wind farm blockage , occurs because the wind farm behaves as an obstacle that locally modifies the atmospheric circulation, forcing the flow to decelerate in front of the wind farm and accelerate around and above the farm to satisfy mass conservation. Similar large-scale acceleration patterns have been reported in offshore wind farm simulations using wind farm parameterization .

We now focus on the near-surface meteorological variations induced by the wind turbines. To simplify the analysis, we restrict the study to the eastern part of the wind farm, where the turbine density is highest and where the largest perturbations are observed. The study domain is therefore limited to the red rectangle in Fig. . The subdomain origin is centered and aligned with the wind direction at hub height of the most upstream turbine in the farm: wind turbine number 32.

Figure  shows the wind speed deficit induced by the turbines within the subdomain at 80 and 2 m above ground level (3D iso-surface of ΔUh is plotted in Fig. ). A strong difference in wind direction is observed between hub height and the near-surface layer, with deviations approaching 30–50°. At 80 m, the wake structures exhibit the typical downstream spreading, while an upstream induction zone is visible in front of each turbine as well as a larger-scale induction zone associated with each turbine row. Near the ground, wind speed variations are also present. Although their magnitude is smaller than at hub height, they represent relatively large perturbations when compared to the weak background wind speeds close to the surface.

Near the ground a reduction in wind speed is observed upstream of turbine rows, while beneath the rotor localized acceleration occurs, consistent with the sub-rotor acceleration mechanism described in Sect. . In addition, downstream of the last turbine row, a region characterized by alternating positive and negative wind anomalies is visible. This region coincides with a convergence zone between two air masses with different flow directions (see converging arrows in Fig. b) and distinct temperatures (see the next paragraph and Fig. a). These wind anomalies therefore appear to result primarily from a wind-farm-induced modification of the position of the katabatic flow and the convergence zone, rather than from the individual influence of the turbines.

The 2 m air temperature T and temperature difference ΔT fields are illustrated in Fig. . As with the wind speed deficit analysis, the study is restricted to the subdomain (red rectangle in Fig. ). Subdomain temperature distributions for the Ref and ADR simulations are evaluated in Fig. , with the corresponding ΔT field shown in Fig. a. Finally, the 3D iso-surface of ΔT is plotted in Figs.  and .

The temperature field exhibits strong spatial variability as cold air accumulates on valley floor, while terrain at higher elevations remains comparatively warmer. A katabatic flow is visible in the valley located downstream of the wind farm (blue/yellow air mass in Figs. a and ), producing a front between the different local flows.

In the ADR simulation (Fig. c), localized warming zones appear downstream of the turbines and are advected by the near-surface flow. The temperature averaged along the x axis (Fig. b) shows a difference of nearly 2 °C between the upstream and downstream parts of the farm in both simulations. The net contribution of the wind farm has to be evaluated by comparing the two curves. The difference is limited to the immediate vicinity of the farm and shows a mean warming anomaly of up to 0.4 °C.

The maximum convergence of the 2 m wind field is indicated by the dotted line in Fig. . This line marks the position of the katabatic front separating the two interacting air masses. The comparison of its position in simulations Ref and ADR (Fig. a) shows that the presence of the wind farm shifts this convergence zone southward by approximately 100–400 m downstream of the farm. As a result, strong temperature variations are observed along this frontal zone, reflecting the different thermodynamic properties of the interacting air masses. Temperature variations ΔT (Fig. a) confirm the localized warming patterns downstream of wind turbines. However, they also highlight the presence of cooling anomalies of around 0.2 °C in some areas surrounding the wind farm.

By comparing the terrain contours with the temperature variations induced by the wind turbines, we can conclude that the strongest warming beneath the rotors occurs for turbines located within valleys, where cold air tends to accumulate (e.g., turbines 9, 10 and 32). Figure  helps illustrate this phenomenon by showing average vertical cross-sections of domain D4 passing through turbines 32, 26 and 21. Turbine 32 is located within a small valley a few tens of meters deep, whereas turbines 26 and 21 are located over relatively flat terrain. Within the valley, turbine-induced mixing interacts with stronger temperature inversions (Fig. e), bringing warmer air from aloft down towards the surface and thereby generating stronger temperature anomalies. For example, a temperature increase of up to 1.6 ^∘C is observed downstream of turbine 32. Another region where strong meteorological variations are observed is along the katabatic flow. In this area, a pronounced warming trail reaching nearly 2 °C is visible (Fig. b). These temperature anomalies are therefore primarily associated with the displacement of the convergence zone induced by the wind farm rather than with the direct contribution of the third turbine row.

At night, surface radiative cooling makes the ground colder than the air above, establishing a stable air temperature inversion. Under these conditions, heat is transported downward by background turbulence from the warmer air aloft toward the colder surface, resulting in a negative turbulent heat flux (Fig. f). At and below rotor height, a negative flux anomaly develops downstream of each turbine, making the turbulent heat flux even more negative (Fig. j, k and l). This indicates an enhanced downward transport of warm air or an upward transport of cold air caused by wake-generated turbulence. This increased vertical mixing within the stable stratification drives the positive temperature anomalies below the rotor and the negative anomalies at and above rotor level (Fig. h and i).

The histogram of temperature variations (Fig. c) indicates that, although both warming and cooling occur within the subdomain, the distribution is positively skewed, with warming anomalies dominating in both frequency and amplitude. Specifically, 50 % of the subdomain pixels experience a temperature increase greater than +0.1 °C (reaching a maximum local warming of +1.96 °C), whereas 30 % remain unaffected, and only 20 % show minor cooling. The spatial distribution shows that under stable atmospheric conditions turbine-induced mixing tends to increase near-surface temperatures downstream of the wind farm, while modifications of the flow field around the farm tend to slightly cool the surrounding regions. Across the subdomain the mean temperature change is +0.17 °C with an RMS of 0.38 °C, reflecting significant spatial heterogeneity. These spatial statistics should, however, be considered indicative rather than absolute due to their dependence on the subdomain boundaries. Nevertheless, they provide a meaningful quantitative basis confirming the dominance of 2 m air temperature warming across the study area.

Figure a shows the ground temperature variation induced by the wind turbines at the first level (-0.1 mm) of the SURFEX/ISBA surface model (rivers and lakes are excluded). Warming patterns are visible starting from the rotor location of each turbine, corresponding to the near-surface air temperature increases previously identified.

Compared to the 2 m air temperature response, ground temperature exhibits a similar spatial pattern downstream of the turbines, where downward turbulent mixing of warmer air creates distinct surface warming plumes. However, a key difference lies in the higher prevalence of cooling anomalies (30 % of the subdomain for ground temperature versus 20 % for air temperature) and a lower overall mean warming impact (+0.043 vs. +0.173 °C). Comparing ground temperature variations with the near-surface turbulent heat flux (resolved + subgrid) reveals an identical spatial pattern (Fig. b) featuring positive turbulent heat flux anomalies where cooling occurs and negative anomalies where warming is observed. These cooling anomalies are primarily located immediately upstream of the turbine rows, in regions affected by the upstream blockage effect (Fig. b). In these deceleration zones, reduced near-surface wind speeds weaken turbulent exchange, thereby reducing the downward (negative) turbulent heat flux toward the ground. This drop in turbulent exchange leads to a local decoupling between the ground and the near-surface air. Consequently, nocturnal radiative cooling of the topsoil layer prevails, enhancing ground cooling upstream of the turbines. Conversely, because the upstream near-surface air is continuously replaced by the background flow, its temperature remains unaffected by this localized surface cooling. Overall, ground temperature variations reflect a coupled response to both air temperature changes and wind speed changes through modifications of turbulent transfer.

The results highlight the importance of flow heterogeneities when investigating wind farm–atmosphere interactions in complex environments. In the present case, both the complex terrain and the stable, strongly stratified atmospheric conditions play a key role in shaping the flow. These findings align with recent literature emphasizing the necessity of multi-scale modeling chains to capture the full spectrum of interactions across spatial scales.

The simulations reveal a strong heterogeneity of meteorological conditions across the domain, with significant spatial variability in both wind speed and wind direction horizontally as well as vertically. This heterogeneity becomes particularly pronounced near the surface, where local katabatic flows induce the convergence and interaction of air masses with distinct thermal and momentum properties. As recently highlighted by in complex terrain setups, terrain-induced atmospheric structures can strongly interact with or even dominate near-surface microclimatic signals, making realistic multi-scale forcing essential to isolate turbine-induced perturbations from natural background variability.

Within this already complex background flow, wind turbines introduce additional perturbations that further modify local atmospheric circulation. Under stable and strongly stratified conditions, turbines facilitate vertical transport through two main physical mechanisms: (i) organized vertical motion driven by blade rotation and wake divergence, which generates coherent vertical velocity pairs , and (ii) vertical convergence and divergence of turbulent heat fluxes downwind of the rotor . In stably stratified boundary layers, both mechanisms act to transport warmer, higher-momentum air from aloft down toward the surface.

At the single-turbine scale, the wind turbine acts as localized obstacles to the flow, manifesting as upstream induction zones and near-surface Venturi acceleration beneath the rotor. Local sub-rotor wind acceleration under stable stratification is consistent with fine-scale LES studies of isolated and aligned turbines , where buoyancy forces in stable stratification reduce vertical displacement over the obstacle and enhance localized flow acceleration near the ground. At the wind farm scale, cumulative blockage causes broader flow redistribution, including lateral edge acceleration and slight displacement of surrounding air masses, echoing findings from mesoscale and microscale wind farm simulations .

The amplitude of the meteorological variations induced by each turbine appears to be strongly influenced by local topography. In particular, turbines located within valleys generate the largest perturbations, as they interact with stronger near-surface temperature inversions. The strong relationship between near-surface thermal stratification and the magnitude of turbine-induced temperature variations is consistent with the experimental findings of and the numerical results of . In contrast, the amplitude of these variations does not seem to be primarily controlled by the turbine power production (Fig. ) or by its spatial position within the wind farm.

The atmospheric perturbations generated by the turbines extend to near-surface air temperatures before propagating to the land surface through coupled energy exchanges. At 2 m height, the simulations indicate that wind turbines induce a mean near-surface warming of approximately 0.4 °C within the wind farm immediate area, with local maxima reaching up to 1.6 °C. This magnitude aligns well with field observations and numerical modeling under stably stratified nocturnal conditions, which typically report air temperature increases ranging between 0.1 and 2.2 °C depending on inversion strength, turbine hub height and farm layout .

Driven by these near-surface air anomalies, local modifications of wind and turbulent heat flux, coupled land–atmosphere energy exchanges lead to topsoil ground temperature variations ranging from -0.7 to +1.6 °C. This range and its spatial heterogeneity are consistent with satellite-derived land surface temperature observations over operating wind farms as well as LES with surface energy balance modeling . Upstream blockage and velocity deficits weaken surface turbulent exchanges, leading to localized ground cooling, whereas downwind advection of turbine-mixed air enhances downward heat transfer, resulting in localized ground warming .

From the perspective of vineyard protection, the simulations indicate that wind turbines induce a mean near-surface warming of approximately 0.4 °C within the wind farm immediate area, with local maxima reaching up to 1.6 °C. Although these temperature increases remain relatively small, they are of the same order of magnitude as those produced by frost protection fans commonly used in vineyards . As an example, the Food and Agriculture Organization of the United Nations (FAO) indicates that the 2 m temperature variation corresponds to roughly 30 % of the inversion strength between 2 and 10 m and recommends the use of fans when the inversion exceeds 1.5 °C, and wind speeds near the ground are below 2.5 m s⁻¹. It should be noted, however, that frost protection fans rotate, providing a 360° coverage, and can be operated on demand. Unlike wind turbines, they do not require sufficient wind speed at hub height, as radiative cooling events are typically associated with weak wind conditions .

However, the simulations also reveal ground cooling zones and wind speed reductions in some areas surrounding the farm. According to FAO recommendations, wind plays an important role in limiting frost damage by preventing cold air stagnation near the ground. Therefore, the spatial redistribution of wind and temperature induced by the wind farm may lead to both beneficial and adverse effects depending on the local position within the vineyard landscape.

Our simulations are consistent with reports from winegrowers indicating that frost damage can vary significantly from one location to another during the same frost event. This spatial variability is also present in the simulations, which show strong heterogeneity in near-surface temperature. Consequently, the warming induced by wind turbines is also highly heterogeneous and depends strongly on the local flow configuration.

Several in situ experimental studies (San Gorgonio campaign; see ) estimate turbine-induced meteorological variations by computing the difference between measurements from an upstream and a downstream sensor. However, the present results suggest that such approaches may be insufficient in complex flow conditions. In particular, the wake direction at hub height and the advection direction of near-surface meteorological variations can differ significantly in complex terrain.

Furthermore, even in the absence of a wind farm, meteorological properties may evolve substantially between upstream and downstream locations due to natural atmospheric variability. For these reasons, experimental setups should ideally include multiple near-surface meteorological stations distributed around the wind farm (as for the VERTEX campaign of ). Such measurements would help characterize local flows and better distinguish the background atmospheric evolution from the perturbations directly induced by the turbines.

While the present simulations provide high-resolution insights into turbine–atmosphere–surface interactions in complex terrain, several limitations and uncertainties must be emphasized.

First, model validation is inherently constrained by the availability of observational data. Quantitative evaluation was restricted to a single operational surface station at 2 or 10 m height. In the absence of vertical-profile observations at rotor height, the exact vertical extent, inversion strength and thermal stratification of the actual boundary layer could not be quantitatively verified against field measurements.

Second, the simulation exhibits a nocturnal warm bias in air temperature of approximately 2–3 °C, inherited from the driving AROME forecast and subsurface soil temperature initialization. Although sensitivity experiments were performed on mixing length formulations, soil parameters and atmosphere–surface coupling coefficients, this systematic bias could not be fully removed. Consequently, while the simulated boundary layer is physically consistent and stably stratified, this thermal bias introduces quantitative uncertainty regarding the exact magnitude of the simulated atmospheric stability relative to the study period.

Finally, the results presented here are specific to a single nocturnal event featuring a complex synoptic transition with weak rotor wind speed, local katabatic flows and a late-night mesoscale vortex. While the fundamental physical processes identified in this study are generalizable to stably stratified boundary layers over complex topography, the exact quantitative impacts on near-surface warming and soil temperature variations remain inherently linked to the specific thermal stratification, wind farm geometry and soil properties of this case study.