Large-Eddy Simulation of Utility-Scale Wind Farm Sited over Complex Terrain

Abstract

The realm of wind energy is a rapidly expanding renewable energy technology. Wind farm developers need to understand the interaction between wind farms and the atmospheric flow over complex terrain. Large-eddy simulations provide valuable data for gaining further insight into the impact of rough topography on wind farm performance. In this article, we report the influence of spatial heterogeneity on wind turbine performance. We conducted numerical simulations of a $12\times 5$ wind turbine array over various rough topographies. First, we evaluated our large-eddy simulation method through a mesh convergence analysis, using mean vertical profiles, vertical friction velocity, and resolved and subgrid-scale kinetic energy. Next, we analyzed the effects of surface roughness and dispersive stresses on the performance of fully developed large wind farms. Our results show that the ground roughness element’s flow resistance boosts the power production of large wind farms by almost 68% over an aerodynamically rough surface compared with flat terrain. The dispersive stress analysis revealed that the primary degree of spatial heterogeneity in wind farms is in the streamwise direction, which is the “wake-occupied” region, and the relative contribution of dispersive shear stress to the overall drag may be about 45%. Our observation reveals that the power performance of the wind farm in complex terrain surpasses the drag effect. Our study has implications for improving the design of wind turbines and wind farms in complex terrain to increase their efficiency and energy output.

1. Introduction

Large wind farms are growing globally to reduce the energy-related carbon footprint [1,2,3,4]. Optimal onshore locations near coastlines in heavily populated regions are often not allowed for wind farm installation [5]. The other onshore sites are in rural areas, where the meteorological effects of forests and mountains can influence the local wind [6,7,8,9]. For large wind farms in rural onshore locations, the secondary circulation associated with the land–atmosphere interaction contributes to large-scale effects in wind farm layout optimization [1,10,11]. Computational analyses to characterize the wind and atmospheric turbulence are relatively effective because measurements over an entire site are time-consuming. Creating a complete map of the local topography of the whole wind farm is affordable with modern high-performance computing facilities. However, as detailed in several recent articles [12,13,14,15,16], the fine grid causes a computational burden and becomes highly skewed if one resolves a steep mountain, which also deteriorates the numerical error.

This article presents a computational model of the local topography and the surface roughness to simulate the effect of complex terrain. The work of Bhuiyan and Alam [16] led to a scale-adaptive large-eddy simulation (LES) method for the land–atmosphere exchange over complex terrain. The LES of a wind farm in complex terrain can partially resolve the turbulence kinetic energy (TKE) and the Earth’s surface topography, where the subgrid model minimizes the effects of unresolved TKE and other meteorological phenomena [3,13,16,17]. Thus, complex terrain can enhance the dispersive stresses of the secondary motion close to the ground. The yawed flow is the horizontal misalignment of an incoming flow to the rotor axis. Studies have shown that, in wind farms, neglecting the land–atmosphere exchange, yawed flow can reduce power production and increase the load and moments at the rotor blade [1,4,18,19].

This article considers a scale-adaptive LES method to account for the land–atmosphere exchange and focuses on the effects of airflow reduction in complex terrain and forests [20]. To gain better insight into the effects of spatial inhomogeneity of the Earth’s surface, we follow the methodology proposed by Arthur et al. [15] and incorporate the local stress of complex terrain into the momentum equation by adapting resolved stresses at grid points adjacent to the Earth’s surface [21]. We call this approach the near-surface model, which differs from the classical wall models by the way it formulates velocity and stress conditions on the complex terrain [22,23,24,25]. Chow et al. [26] reviewed the existing scale-adaptive LES of atmospheric boundary layer flows (see also [27,28,29]). Among them are blending the characteristic length scales [28], the hybrid RANS-LES in the near-surface region [27], and the canopy stress model [29]. In the current development, we employ Helmholtz’s second vortex theorem and the vortex stretching mechanism to capture the transition of characteristic scales from one flow regime to the other, such as convergence zone, pure shear, or inflection, to dynamically adjust the dissipation rate [16,17,30,31]. We model wind turbines as actuator disks, leading to a drag source in the atmospheric boundary layer [1,4,19]. Readers interested in the scale-adaptive LES method and the actuator disk model may consult the cited reference.

1.1. Surface Roughness

Wind turbines extract energy at heights between 30 m and 150 m and contribute toward the dissipation of energy via surface roughness. The atmospheric sciences community has thoroughly examined the effect of surface roughness caused by mountains, forests, vegetation, etc. Surface roughness is generally represented through the aerodynamic roughness length ($z_0$) and the Monin–Obukhov similarity theory [32,33]. The impact of an isolated hill on the turbine performance depends on the turbine size and its position relative to the mountain [6,8]. Some studies show that wind farm performance can be enhanced over hilly terrain [6,8]. However, other studies have observed that overall surface drag may outweigh the benefit of the local speed-up effect resulting from the disturbances of the surface protrusion (e.g., [11] and the refs in [17]). Past studies of wind farms in complex terrain [5] have either empirically tuned the earth–atmosphere exchange [34,35,36] or ignored it [1].

Wind tunnel experiments verified that a laminar–turbulent transition occurs in the boundary layer due to the surface-mounted roughness elements [34,35,37,38,39]. Cao et al. [40] studied the effects of 2D hills on the turbulent boundary layer. Ye et al. [36] investigated the impact of various roughness elements on flow resistance in mountainous regions. The surface drag and turbulent momentum transport caused by surface roughness primarily constrain the near-surface mean wind speed, dissipate a fraction of the kinetic energy, and alter the atmospheric boundary layer profile [36,40,41,42]. The representation of subgrid-scale effects of mountains within a scale-adaptive LES of atmospheric boundary layer has improved in recent years [15,16,21].

An isolated hill with a smooth surface exerts a blockage effect on the downstream side, which may reduce the power output and enhances the fatigue load on relatively small turbines [6]. The power output of a wind turbine sited on top of a hill may be (about $82\%$) higher than that of flat terrain. Vanderwende and Lundquist [2] observed that the power output of a wind farm over vegetated land increases by 14% when the roughness length ($z_0$) changes from $0.1$ m to $0.25$ m due to the increased crop height. Tobin et al. [43] observed that optimally placed artificial windbreaks may increase power production by at least 10% [8,11,44]. In contrast, the overall power losses due to wake effects could be up to $20\%$ of the total power for wind farms over flat terrain [45].

Each turbine also poses as additional drag and increases near-surface kinetic energy dissipation [20,46]. Therefore, sustaining the high power output of a large wind farm depends on whether another source of kinetic energy can compensate for the extra energy dissipation from a large array of turbines. For example, 3% variation in the hub height velocity caused by a complex terrain could lead to approximately 27% fluctuations in the power output [5]. The irregular topography of complex terrains, such as the Rocky Mountains and Tibetan Plateau, and coastlines, or varying land use are efficient for the vertical kinetic energy transfer from aloft and, thus, can play an essential role in wind farm optimization [2,16,46,47].

1.2. Form-Induced Dispersive Stress

Complex terrain acts as a large-scale obstruction to the flow, creating dispersive stress, of which the subgrid-scale effects are essential in scale-adaptive LESs of wind farms. Past studies have addressed drag reduction; empirical formulations of roughness effects in the secondary turbulent motions due to rough walls [48], forests, canopies, vegetation, coral reefs, etc. [49]; and arrays of wind turbines [10,47]. The double-averaging method, introduced by Raupach and Shaw [50] for canopy turbulence [51], provides the dispersive stress of turbulent flow over wind turbines [10]. For dense canopies, the form-induced drag could significantly affect the lower canopy layer [52]. Poggi et al. [53] suggested that dispersive stresses could reach up to 30% of the Reynolds stresses [49,54]. Moltchanov and Shavit [49] observed that normal dispersive stresses are relatively more sensitive to the wake than the recirculation zone. Dispersive stress can originate from the vorticity generated by surface irregularities [55]. In a wind tunnel experiment of urban-like roughness elements, Cheng and Castro [56] observed that dispersive stresses are negligible compared with Reynolds stresses. In turbulent flow over an array of cubes, Santiago et al. [57] observed that Reynolds and dispersive stresses contribute equally [58,59]. However, recent studies suggest that dispersive stresses account for 30% to 40% of the Reynolds stresses [60,61].

For wind farms in complex terrain, the secondary motion of the sparse canopy formed by wind turbines leads to enhanced dispersive stresses (e.g., [50]). However, it remains unclear whether a more significant downward transport of kinetic energy may be sustained when dispersive stresses are due to the simultaneous effects of terrain complexity and wind turbines. In particular, no estimate for a sustained rate of the downward transport of kinetic energy is available, which may compensate for the loss within large wind farms. This work provides the contributions of the dispersive and Reynolds stresses to the total drag generated by wind farms in complex terrain.

1.3. Outline

This article considers a scale-adaptive LES of a wind farm consisting of $12\times 5$ turbines (with a rotor diameter of 100 m) sited over a complex terrain. Instead of explicitly resolving the terrain roughness elements, we employ the geostrophic drag law within a near-surface model to simulate the effects of the terrain. We set the vertical inflow profile of the streamwise velocity with a fixed wind speed at the hub height and employ a stochastic forcing method to inject turbulence intensity into the spanwise and wall-normal components of the inflow velocity. This approach does not recirculate nonlinear eddy dynamics on embedded inflow domains.

We assess the resolution requirements for utility-scale wind farm simulations considering standard criteria. We evaluate the influence of spatial heterogeneity and drag coefficient on the performance of wind farms. Based on the utility-scale wind farm flow fields over complex terrain, we assess the relative contributions of the dispersive and Reynolds stress to the total drag in wind farms. Then, we compare the effects of spatial heterogeneity on the atmospheric turbulence around wind farms. Here, we show that the power production of a wind farm in complex terrain outweighs the associated drag of the roughness elements. Finally, we report the relative contribution of dispersive stresses to the overall drag, where dispersive stresses are due to both the roughness elements and the turbines.

The article is organized as follows. First, Section 2 provides the mathematical description of numerical methods and the theory and implementation of the near-surface model used for the present study. Section 3 presents the validation and the analysis of the result. Finally, Section 4 summarizes the main results indicating potential future research directions.

2. Numerical Methods

2.1. Large-Eddy Simulation (LES)

This study employs an in-house LES code [16,31,62] to simulate the high Reynolds number neutral atmospheric boundary layer flow over an array of wind turbines. The LES method solves the filtered continuity equation:

$$\frac{\partial {\tilde{u}}_i}{\partial x_i}=0$$

(1)

and the momentum equations

$$\frac{\partial {\tilde{u}}_i}{\partial t}+\widetilde{u_j}\frac{\partial {\tilde{u}}_i}{\partial x_j}=-\frac{\partial \tilde{p}}{\partial x_i}-\frac{\partial {\tau }_{ij}}{\partial x_j}+f_i\chi (\mathit{x},t),$$

(2)

where ${\tilde{u}}_i(\mathit{x},t)$ denotes the spatially filtered velocity characterized by a length scale of ${\Delta }_{\mathrm{les}}=2\sqrt[3]{{\Delta }_x{\Delta }_y{\Delta }_z}$. The subscripts $i=1,2,3$ represent streamwise, spanwise, and surface-normal directions, respectively. Also, $\mathit{x}\in {\mathbb{R}}^3$ denotes either $(x_1,x_2,x_3)$ or $(x,y,z)$.

In Equation (2), a subgrid model should correctly represent the second last term so that $-\overline{{\tau }_{ij}{\mathcal{S}}_{ij}}$ accounts for the production of small-scale turbulence kinetic energy, where $\overline{(·)}$ represents the ensemble average, ${\mathcal{S}}_{ij}=(1/2)(\partial {\tilde{u}}_i/\partial x_j+\partial {\tilde{u}}_j/\partial x_i)$ denotes the filtered strain rates, and ${\tau }_{ij}={\widetilde{u_{i}u}}_j-{\tilde{u}}_i{\tilde{u}}_j$ denotes the subfilter-scale stresses. Borue and Orszag [63] indicated that a model of the subgrid stress ${\tau }_{ij}$ would account for the energy cascade through the process of vortex stretching. Several recent investigations have clearly outlined that vortex stretching and strain self-amplification formulate a cornerstone principle for the energy cascade in an LES [30,64]. In this article, we have considered the scale-adaptive LES model [31] to represent the residual stress tensor ${\tau }_{ij}$. The scale-adaptive LES is based on the idea of an energy cascade via vortex stretching [16,31], and its extension to wind farm simulation has been detailed by Alam [17].

2.2. The Actuator Line and the Actuator Disk

The actuator line and the actuator disk are two models appropriate for simulating large arrays of giant wind turbines with rotor diameters of more than 100 m [65]. Wakes of such a large wind turbine are embedded with the wind shear and coherent vortical structures of the atmospheric boundary layer [1]. To accurately capture the tip vortices created by horizontal axis turbines, the actuator line model represents each of the three blades as a rotating line with distributed forces computed using the blade-element theory [18] (e.g., see [66] for the actuator line model of vertical axis turbine). Note that utility-scale wind turbines interact with coherent vortical structures with a length scale of $\mathcal{O}\left(100\right)$ meters, much larger than the diameter of tip vortices [19]. There is thus a need to characterize atmospheric coherent structures and their role, e.g., in the wake recovery for large arrays of utility-scale wind turbines. This work considers the scale gap and employs the actuator disk model to represent each turbine with a stream tube around the rotor disk [7,10].

In this work, the actuator disk model uses the following thrust force:

$$f_1=-\rho \frac{1}{2}{C}_t{\overline{u}}_d^2\frac{\pi }{4}{D}^2,$$

(3)

where ${\overline{u}}_d$ is the local wind speed averaged over the disk region. In a wind farm, the average wind speed ${\overline{u}}_d$ varies from one turbine to the other, which accounts for the interaction of wakes resolved at the cutoff length scale that attenuates all wavenumbers smaller than $\pi /{\Delta }_{\mathrm{les}}$. Here, we adjust the original thrust coefficient $c_t$ used in the standard actuator disk model to a modified thrust coefficient $C_t=c_t{U}_{\infty }^2/{\overline{u}}_d^2$ [65]. Note that $U_{\infty }$ denotes a free-stream wind speed approaching a turbine. Finally, we consider a normal distribution of the thrust force so that the mean drag over the rotor region acts at the center of the rotor (e.g., see [18,65,67]).

2.3. Near-Surface Model

In the vicinity of the Earth’s surface, simultaneously resolving complex terrain and energy-containing eddies with an isotropic grid incurs significant computational expenses [14,16,29]. For an LES of atmospheric boundary layers, a finer grid resolution is typically applied in the vertical direction compared with the horizontal direction [21,68]. Atmospheric LES codes treat horizontal and vertical dissipation separately, often using a damping function to adjust the eddy viscosity [15,69]. Such an LES method also utilizes the conventional Monin–Obukhov similarity theory to model the effects of the viscous sublayer [70].

The Monin–Obukhov similarity theory states that the vertical distribution of the wind shear is $\partial \overline{u}/\partial z\propto u_{*}/\left(kz\right)$, which leads to the wind profile

$$\overline{u}\left(z\right)=\frac{u_{*}}{k}ln\left(\frac{z}{z_0}\right),$$

(4)

where $z_0$ is the constant of integration and is known as the aerodynamic roughness length and $u_{*}$ is the characteristic velocity scale, also known as friction velocity. The logarithmic law (4) is more appropriate for wind farm simulations than using the power-law profile for the effects of surface roughness (and atmospheric stability) [71]. In the LES of the land–atmosphere exchange over complex terrain [15,22], we consider the commonly used representation of the turbulent stresses on the Earth’s surface (e.g., Equation (1) of Arthur et al. [15]):

$${\tau }_{i3}=-c_f{〈\sqrt{{\overline{u}}_1^2+{\overline{u}}_2^2)}〉}_h{\overline{u}}_i,i=1,2,$$

(5)

where ${\langle ·\rangle }_h$ denotes the average over a horizontal plane. Based on the Monin–Obukov similarity theory, we estimate the frictional coefficient $c_f$ from the velocity $\overline{u}\left(z_1\right)$ given by Equation (4) at some reference height $z=z_1$ [15,21,32]. Typically, one takes a reference height $z_1=\Delta z$, where $\Delta z$ is the height of grid cells adjacent to the boundary at $z=0$.

The roughness length $z_0$ is commonly about one-tenth of the average height of the local surface elevations of complex terrain [10,32,70,72]. However, the upper limit of the roughness sublayer ($z_{*}/z_0$) varies approximately between 35 and 150 [70,73]. For instance, if $z_{*}/z_0=100$ with $z_0=1$ m (e.g., for forests), the depth of the roughness sublayer ($z_{*}$) may be up to 100 m. To capture the rotor region with 5–10 grid points [18], the LES of wind farms in complex terrain will also resolve the roughness sublayer. Thus, the evaluation of the ground shear stress (${\tau }_{i3}$) via Equations (4) and (5) will have $z_1=\Delta z<100$ m, which is not consistent with the boundary layer scaling law [70].

Sullivan et al. [69] proposed the mean-field eddy viscosity to address the above issue. The LES assumes that the subgrid-scale model’s dissipation accounts for the ensemble average of the subgrid-scale turbulence production, $-\overline{{\tau }_{ij}{\mathcal{S}}_{ij}}$ [74,75,76]. For the resolved-scale motion, the filtered energy $(1/2){\overline{u}}_i^2$ evolves according to the following equation [74]:

$$\frac{D}{Dt}\left(\frac{1}{2}{\overline{u}}_i^2\right)=\frac{\partial }{\partial x_j}\left[\mathrm{transport}\mathrm{terms}\right]+\underset{\mathrm{actuation}\mathrm{term}}{\underbrace{{\overline{u}}_i{f}_i\chi (\mathit{x},t)}}-{\tau }_{ij}{\mathcal{S}}_{ij}-2\nu {\mathcal{S}}_{ij}{\mathcal{S}}_{ij}.$$

(6)

The term $-{\tau }_{ij}{\mathcal{S}}_{ij}$ represents energy transfer from large to small scales, and $\mathsf{\Pi }={\tau }_{ij}{\mathcal{S}}_{ij}$ accounts for the subfilter-scale production of turbulence kinetic energy (note the same term with an opposite sign). The term $-2\nu {\mathcal{S}}_{ij}{\mathcal{S}}_{ij}$ represents the resolved scale viscous dissipation. Let us denote the exact viscous dissipation by $-2\nu S_{ij}^e{S}_{ij}^e$, where superscript ${(·)}^e$ refers to the strain rate of the exact flow field, $S_{ij}^e=(1/2)(\partial u_i/\partial x_j+\partial u_j/\partial x_i)$. The exact viscous dissipation $-2\nu S_{ij}^e{S}_{ij}^e$ must equal the sum of the subfilter-scale production $\mathsf{\Pi }={\tau }_{ij}{\mathcal{S}}_{ij}$ and the resolved-scale viscous dissipation $-2\nu {\mathcal{S}}_{ij}{\mathcal{S}}_{ij}$ such that [77]

$$\mathsf{\Pi }-2\nu {\mathcal{S}}_{ij}{\mathcal{S}}_{ij}=-2\nu {\mathit{S}}_{\mathit{ij}}^{\mathit{e}}{\mathit{S}}_{\mathit{ij}}^{\mathit{e}}.$$

(7)

A classical subgrid model based on the filtered strain rate, yielding

$$\mathsf{\Pi }=c_s(\mathit{x},t){\Delta }_{\mathrm{les}}^2{\left(2{\mathcal{S}}_{ij}{\mathcal{S}}_{ij}\right)}^{3/2}$$

(8)

may violate the local isotropy hypothesis in the LES of wind farms. The filtered energy flux $-\mathsf{\Pi }$ will not vanish in local regions (e.g., in a viscous sublayer), where all turbulence eddies may have been locally resolved [77]. To address this problem for the atmospheric boundary layer flows, Sullivan et al. [69] suggested modifying the eddy viscosity ${\nu }_{\tau }(\mathit{x},t)$ inside the roughness sublayer [29] to account for the roughness effects in the filtered momentum equation while evaluating Equation (4) at $z=z_1<z_{*}$ to obtain frictional coefficient $c_f$. This model considers the subgrid-scale stress averaged over a horizontal plane in the following form [69]:

$${\langle {\tau }_{i3}\rangle }_h\approx 2[\gamma \left(z\right){\nu }_{\tau }+{\nu }_T]\frac{\partial {\langle u_i\rangle }_h}{\partial z},i=1,2,$$

(9)

where $\gamma \left(z\right)$ is the damping factor that accounts for the shear-driven nature of the near-surface dynamics [29] and ${\nu }_T$ is the mean-field eddy viscosity needed to force the vertical velocity gradient to match that from the similarity theory at $z=z_1$ [69]. The mean-field eddy viscosity model accounts for the Gabor–Heisenberg uncertainty, where the subgrid turbulence is local in physical space and nonlocal in wave number space.

In the scale-adaptive LES of wind farms, we assume that (i) vortex stretching drives the energy cascade and maintains the Kolmogorov energy spectrum and (ii) that turbulence production, kinetic energy cascade, and viscous dissipation are in local equilibrium [16,17,67]. More specifically, we avoid the horizontal averaging and the damping factor in the mean-field eddy viscosity model (9) proposed by Sullivan et al. [69]. Let us rewrite Equation (5) as

$${[{\tau }_{13}^2+{\tau }_{23}^2]}^{1/4}=C_D^{1/2}\overline{u}(x,y,z_1),$$

(10)

where the drag coefficient is

$$C_D={\left[k/ln(z_1/z_0)-{\psi }_M\left(\xi \right)\right]}^2$$

(11)

and for neutral atmospheric boundary layer flow ${\psi }_M\left(\xi \right)=0$. As we want to account for the impact of heterogeneous surface elevations, we modify the drag coefficient given by Equation (11) such that [15]

$$C_D\left[{\int }_0^{z_1(x,y)}a\left(z\right)dz\right]={\left[k/ln(z_1/z_0)-{\psi }_M\left(\xi \right)\right]}^2,$$

(12)

where $a\left(z\right)$ is a user-defined weight function to switch off from complex terrain to the atmosphere [15,16,62,78]. Note the variability in the reference height $z_1(x,y)$, which follows the irregularity of the terrain. Bhuiyan and Alam [16] considered $z_1(x,y)$ matching with the actual surface of Askervein Hill and observed that the LES data obtained using this method agreed with the wind measurements over Askervein Hill.

In this work, we focus on modelling near-surface turbulence and assume that Equation (4) provides local values of the shear stress on each point of the surface layer (particularly for $z=z_1$). Then, Equations (4), (10) and (11) provide local values of the friction velocity $U_{*}={[{\tau }_{13}^2+{\tau }_{23}^2]}^{1/4}$ at each (grid) point $(x,y,z_1)$, which we use for each component of the ground shear stress, such as $U_{*}^2={\tau }_{i3}$ for $i=1,2$. Finally, we estimate the mean-field eddy viscosity ${\nu }_T$ from the following equation (without the horizontal average)

$${\tau }_{i3}=2[{\nu }_{\tau }(x,y,z_1)+{\nu }_T]\frac{\partial {\overline{u}}_i}{\partial z},i=1,2.$$

(13)

2.4. Remark

In the preceding discussion, we adopt the approach of Sullivan et al. [69] and Basu and Lacser [70] to address atmospheric turbulence and near-surface effects in a scale-adaptive LES applied to wind farms in complex terrain. When the grid resolution allows for capturing some large-scale terrain features, we employ the immersed boundary method, as demonstrated by Bao et al. [21]. Notably, recent studies, including references [16,78], have shown the efficacy of this approach, referred to as the canopy stress method, in an LES of atmospheric boundary layer flow over hilly terrain.

For our investigation concerning utility-scale wind farms in complex terrain, where the shape of the terrain remains entirely unresolved, it becomes essential to assess the methodology thoroughly. As actual wake measurements from wind farms in complex terrain are limited, we rely on knowledge of turbulent flow to validate the findings of our numerical investigation.

3. Results and Discussions

The following analysis employs an in-house LES code to study the interaction between a wind farm and complex terrain [16,17,62,67]. The present study does not explicitly resolve irregular topography, such as mountains, coastlines, or urban variations in land use [20,46,79]. Wind energy applications may characterize complex terrain based on the roughness class associated with a roughness length scale (see Elgendi et al. [79] and the refs therein). This work focuses on wind farms’ flow structure and performance, considering each wind turbine a Gaussian actuator disk [67]. First, we discuss the impact of grid resolution on ABL profiles, wake profiles, turbulence, and wind farm performance. Second, we study the effect of surface roughness on the performance of wind farms. We investigate various factors contributing to the increased power production in wind farms. Third, we examine the relative contribution of dispersive stresses and Reynolds stress for wind farms in complex terrain.

3.1. Simulation Setup

This work simulates a $12\times 5$ array of wind turbines in the neutrally stratified atmospheric boundary layer over a complex terrain. Each turbine has a hub height of 100 m and a rotor diameter of $D=100$ m. Turbines are separated by a distance of $S_x=7$D in the x-direction and $S_y=5$D in the y-direction. The wind farm is schematically shown in Figure 1. The presence of the wind turbine array affects the upwind flow conditions known as the blockage effect [80,81]. Strickland and Stevens [82] observed that the induction region is more apparent and may affect up to 13D on upwind flow conditions if the inter-spacing between the wind turbine is relatively less. Therefore, we kept the first row at a distance of 14D downstream from the inlet boundary at $x=0$. We have the initial condition ${\overline{u}}_i(\mathit{x},0)=\langle U\left(z\right),0,0\rangle$. At the inlet boundary, $x=0$, we assign the streamwise component of the velocity to $U\left(z\right)=(u_{*}/k)ln(z/z_0)$, where $u_{*}$ is obtained for a fixed value of $z_0$ according to a desired Reynolds number ($\mathcal{R}e=6.7\times {10}^7$) based on the undisturbed velocity at hub height. We solve the stochastically forced linearized Navier–Stokes equation at the inlet boundary $x=0$ to provide transient perturbations $\langle 0,v^{\prime }(0,y,z,t),w^{\prime }(0,y,z,t)\rangle$ into the spanwise and surface-normal velocities and, thus, consider $\langle U\left(z\right),v^{\prime }(0,y,z,t),w^{\prime }(0,y,z,t)\rangle$ for the boundary condition at $x=0$. This method assumes an ensemble of (longitudinal) eddies with centers randomly distributed in space at $n_{\mathrm{spot}}$ locations while their axis of rotation aligns in the x direction. For a relatively small number of such eddies, e.g., $n_{\mathrm{spot}}=100$, the method generates a 3D synthetic turbulence with the spectrum of each velocity component exhibiting a logarithmic slope of $-5/3$ [17,76]. Meneveau [76] explained that such eddies maintain an enhanced level of variance (in a corresponding time series) by extracting energy from the mean flow and passing it to the perturbation fields (see also [17,83]). The stochastic force is dynamically adjusted to match the desired turbulence intensity level.

To demonstrate the transitional flow development in a scale-adaptive LES for surface roughness of $z_0=1$ m, we collected velocity time series data at a specific point $(x,y,z)=(12D,y,1D)$ in front of three turbines, namely T2, T3, and T4, as indicated in Figure 1. The time series of velocity would represent the phenomena that atmospheric turbulence produces chaotic motions in the atmospheric boundary layer, which inhibit turbulence mixing. The streamwise component of each of the three velocity time series is shown in Figure 2. This result indicates that episodes of turbulent bursts persist in the streamwise velocity and pass through the first row of wind turbines.

A moving average with a window size of 512 s was applied to the streamwise velocity in front of turbine T2. The bottom plot of Figure 2 displays the results, revealing a range of energy-containing length scales at the measurement location T2. This observation highlights the complex and intricate inflow turbulence simulated using our stochastic forcing method.

In the following section, we try to understand the grid resolution and time step necessary to capture wind turbine wakes.

3.2. Effect of Mesh Resolution

The present study considers a range of resolutions, including isotropic and non-isotropic refinements, to ensure that we had both poorly resolved and well-resolved simulations. Following the criteria outlined by Pope [75], we discuss the effects of mesh resolution on the wake profiles, turbulence statistics, and the performance of a wind farm [68,84].

Table 1. Cases selected for mesh convergence analysis. $N{x}_2$ and $N{x}_3$ are the grid points across the rotor region in the spanwise and surface-normal directions, respectively. $z_1$ is the first cell height from the Earth’s surface. The computational domain for the simulated cases is ${\mathcal{L}}_x\times {\mathcal{L}}_y\times {\mathcal{L}}_z=[12\times 3\times 1]{\mathrm{km}}^3$.

Cases	${\Delta }_{\mathit{x}}\times {\Delta }_{\mathit{y}}\times {\Delta }_{\mathit{z}}\left[{\mathbf{m}}^3\right]$	${\mathit{z}}_1$	${\mathit{Nx}}_2$	${\mathit{Nx}}_3$
M10	$10\times 10\times 10$	10	10	10
M15	$15\times 15\times 15$	15	6	6
M20-1	$20\times 20\times 20$	20	5	5
M20-2	$20\times 20\times 10$	10	5	10
M20-3	$20\times 20\times 8$	8	5	12
M30	$30\times 30\times 30$	30	3	3
M40	$40\times 40\times 40$	40	2	2

summarizes seven representative cases of LES for a mesh resolution study. The aerodynamic roughness length ($z_0$) of the inlet wind profile was set to 1 m. In the LES Equation (2), one assumes that ${\tilde{u}}_i(\mathit{x},t)$ is a spatially filtered velocity. However, the numerical solution ${\tilde{u}}_i(\mathit{x},t)$ of Equation (2) is, in principle, a resolved part of the filtered velocity [75]. Thus, to extract the statistics of the resolved velocity ${\tilde{u}}_i(\mathit{x},t)$, Pope [75] suggests averaging the flow field within a time interval, which is $45T_{*}$ in the present study, where $T_{*}=D/u_{*}$ is the large-eddy turnover time unit. Based on a global error tolerance of ${10}^{-6}$, the flow field was sampled at a dynamically adjusted time step such that CFL = 1, and CFL stands for Courant–Friedrichs–Lewy condition.

3.2.1. Mean Profiles and Turbulence

Figure 3 shows the vertical profile of the mean streamwise velocity on the vertical mid-plane at 2D upstream from the first row of the wind farm. The mean was obtained within a time interval of size $45T_{*}$, and the result has been compared with the log-law wind profile given by Equation (4). The vertical profile of the mean streamwise velocity follows the logarithmic profile given by Equation (4) for all resolutions listed in Table 1. The atmospheric boundary layer flow is poorly resolved when the grid spacing is 40 m in all three directions [68]. Wurps et al. [68] suggested that a grid spacing of about 10 m is necessary for a well-resolved LES to capture the vertical profile of mean streamwise velocity.

Figure 4 shows vertical profiles of the mean streamwise velocity on the vertical mid-plain at wind turbine rows 1–4. For each row, we compare the profiles at locations 2D, 3D, 4D, and 5D behind the turbine. Similarly, Figure 5 shows vertical profiles for rows 5, 7, 10, and 12. Figure 4 and Figure 5 compare the wake profiles among seven resolutions. The results indicate that the solution converges to the wake profiles of case M10. The profiles of cases M30 and M40 show deviations from that of case M10. A grid spacing of 20 m in all directions is necessary to capture the wakes using the scale-adaptive LES.

Turbulence statistics of a wind farm provide insight into how kinetic energy is entrained from aloft. For this analysis, we consider the Reynolds decomposition:

$${\tilde{u}}_i(\mathit{x},t)={\overline{u}}_i+u_i^{\prime }(\mathit{x},t),$$

(14)

where ${\tilde{u}}_i(\mathit{x},t)$ is the numerical solution of Equation (2) and ${\overline{u}}_i$ is the ensemble average within a time interval of length $45T_{*}$. The Reynolds stress resolved via the LES method is ${\tau }_{ij}^R=\overline{u_i^{\prime }{u}_j^{\prime }}$. The fraction of the turbulence kinetic energy modelled using the scale-adaptive LES method is $k_{sgs}=(1/2){\tau }_{ii}$, where the resolved turbulence kinetic energy is $k_{res}=(1/2){\tau }_{ii}^R$. The ratio of resolved to the total turbulence kinetic energy is the effective resolution $\gamma =k_{res}/(k_{res}+k_{sgs})$ [74]. A value of $\gamma =0.8$ indicates that the LES method has captured about 80% of the total turbulence energy. In this context, we have analyzed the shear stress $u_{*}^2$, resolved kinetic energy $k_{res}$, subgrid-scale kinetic energy $k_{sgs}$, and effective resolution $\gamma$.

Figure 6 shows vertical profiles of $u_{*}^2$, $k_{res}$, $k_{sgs}$, and $\gamma$, where $u_{*}^2={\left({\overline{u^{\prime }{w}^{\prime }}}^2+{\overline{v^{\prime }{w}^{\prime }}}^2\right)}^{1/2}$ represents the shear stress. Vertical shear stress profiles exhibit negligible dependence on the grid resolution except for cases M30 and M40, (see Figure 6a). Figure 6b shows that, in most parts of the boundary layer, the resolved TKE for coarser grid cases (M30 and M40) were higher than that for finer grid cases. Wurps et al. [68] also noted a similar scenario in atmospheric boundary layer simulations without wind turbines. Celik et al. [85] observed that, in wall-bounded flows, resolved strain ${\overline{\mathcal{S}}}_{ij}$ tends to be small, thereby resulting in relatively less resolved dissipation $ϵ=2{\nu }_{sgs}{\overline{\mathcal{S}}}_{ij}{\overline{\mathcal{S}}}_{ij}$, which leads to a higher resolved TKE.

Figure 6d illustrates the ratio ($\gamma$) of the resolved TKE to the total TKE in the wind farm, which shows a clear dependence on the resolution. Notably, finer grids consistently exhibit larger $\gamma$ values than coarser grids across the entire boundary layer. As the vertical resolution increases, such as in cases M20-2 and M20-3, the value of $\gamma$ approaches that of the finest resolution case M10. Furthermore, the value of $\gamma$ is not constant within the boundary layer. As discussed by Pope [74] and Celik et al. [85], $\gamma >80\%$ may lead to a well-resolved condition for LES. Figure 6d indicates that $\gamma$ exceeds $90\%$, even for the coarsest grid. These findings for scale-adaptive LES of wind farms are consistent with the results of Wurps et al. [68] for the atmospheric boundary layer flows without wind turbines.

Figure 7a–d demonstrate colour-filled contour plots of $Q=-(1/2){\mathcal{G}}_{ij}{\mathcal{G}}_{ij}$ for $Q>0$, where the colour represents the vertical vorticity ${\omega }_z$. The second invariant Q of the velocity gradient tensor ${\mathcal{G}}_{ij}$ represents the relative magnitude of vorticity over the strain; thus, $Q>0$ identifies vortex-dominated regions. Figure 7a,c show the coherent structures and wind turbine wakes for cases M10 and M20-3, respectively. Figure 7b,d show the coherent structures in cases M20-1 and M40, respectively. For a grid spacing of 20 m and 40 m, coherent structures are poorly resolved (Figure 7b,d). When we refine the vertical grid keeping horizontal grid spacing of 20 m (Figure 7c), the scale-adaptive LES method captures the coherent structures similar to case M10.

Figure 8a–f present contour plots of the streamwise and surface-normal components of the resolved velocity at $z=100$ m and $t=45T_{*}$ for cases M10, M20-1, and M40, respectively. These contour plots compare the results between one well-resolved case (M10) and two poorly resolved cases (M20-1 and M40). The results indicate that the horizontal flow structures have been resolved at hub height $z=100$ m.

3.2.2. Effects of Resolution on the Power Output of the Wind Farm

Although Churchfield et al. [1] coupled the actuator line model with the aerodynamics structural response model, there is still a need for a computational framework capable of capturing atmosphere–turbine interactions in a realistic atmospheric environment [18,65,86]. For example, a generalized actuator disk model represents wind turbine wake effects by applying instantaneous forces to each component of the Navier–Stokes equations. Here, we employ the actuator disk theory; however, as discussed in Section 2.2, we calculate the thrust force based on the local disk-averaged velocity using a modified thrust coefficient (see [65,67]).

Thus, the instantaneous power extracted by a wind turbine is

$$P\left(t\right)=2a/(1-a)\rho A{\overline{u}}_d^3,$$

(15)

where $a=1-{\overline{u}}_d/U_{\infty }$ is the axial induction factor, $\rho$ is the air density, A is the rotor’s swept area, and ${\overline{u}}_d$ is the velocity averaged over the rotor’s swept area [65]. Local velocity ${\overline{u}}_d$ and the axial induction factor account for the wake interaction. Due to additional complexity and challenges in properly considering the effects of all the influencing parameters affecting the wakes, it is not straightforward to evaluate wind farm performance using the classical power calculation (Equation (15)). If the inflow condition is laminar, wake effects are a major source of power production losses in wind farms [76]. Having turbulent inflow conditions in the simulations, vertical transport and turbulence mixing are primary mechanisms for wake recovery for turbines operating in the wake of another turbine [10].

In this analysis, we first calculate the temporal average $P_{wt}=(1/T){\int }_0^{T}P\left(t\right)dt$ of the instantaneous power of each turbine using Equation (15) and consider an average power $P_n$ for every 5 turbines at the n-th row of the $12\times 5$ array of turbines. Figure 9 compares the average power of a turbine at each row with various mesh resolutions. Cases M20-1 and M40 over-predict the power production, indicating the need for a finer mesh resolution; however, as we refine the vertical grid, such as in cases M20-2 and M20-3, the average power converges towards the well-resolved high-resolution case. As indicated by the analytical wake models [87], the turbines on the second and third rows are affected mainly by the effects of the first row. The wakes start to recover from the fourth row.

3.3. Effect of Vertical Mixing on Wind Farms in Complex Terrain

Wind speed over the hills is often higher than those in the areas over flat land [6]. However, it remains unclear whether a more significant downward transport of kinetic energy may be sustained in mountainous and forested terrain [5]. The wake of a single wind turbine often exhibits reduced vertical mixing [65]. In stably stratified atmospheric conditions, temperature rise below the rotor tip suggests that the downward flux may persist in large wind farms (see [17] and the refs therein). In the atmospheric boundary layer, the flux of mean kinetic energy at height z is approximately $\mathsf{\Phi }\left(z\right)=-\overline{u^{\prime }{w}^{\prime }}\overline{u}\left(z\right)$. Using Equation (4) and assuming that $u_{*}^2=-\overline{u^{\prime }{w}^{\prime }}$, the energy flux $\mathsf{\Phi }\left(z\right)=(u_{*}^3/z_0)ln(z/z_0)$ at height z increases as $z_0$ increases.

The geostrophic drag law relates the surface friction velocity $u_{*}$ (i.e., surface shear stress ${\tau }_w=u_{*}^2$) and the aerodynamic roughness length $z_0$ [32]. In meteorological applications, the geostrophic drag law assumes that the Coriolis and the pressure gradient force are in geostrophic balance, and the atmospheric background stratification is neglected. The commonly used wind profile based on geostrophic drag law for wind resource assessment is [87]

$$u\left(z\right)-U_g=\frac{u_{*}}{k}ln\left(\frac{fz}{u_{*}}\right)+\mathcal{C}.$$

(16)

Here, $f=2\mathsf{\Omega }sin\varphi$ represents the Coriolis force. Using Equation (16), we approximate the depth of the atmospheric boundary layer as $H_G\approx 0.16u_{*}/f$[10]. Comparing Equation (16) with the near-surface logarithmic wind profile (Equation (4)), we can relate the geostrophic wind speed $U_g$ to the ground friction velocity $u_{*}$ as

$$u_{*}=\frac{k{U}_g}{\left[ln\left(\frac{U_g}{f{z}_0}\right)-{\mathcal{C}}_{*}\right]},$$

(17)

where ${\mathcal{C}}_{*}\approx 4$ [10,73]. In conditions with high wind above flexible surface protrusions, such as crops or forests, non-dimensional arguments suggest that $z_0$ may depend on $u_{*}$ as

$$z_0={\alpha }_c{u}_{*}^2/g,$$

(18)

where ${\alpha }_c$ is Charnock’s constant and g is the gravitational constant [32]. An important observation from Equations (17) and (18) is that there is a positive wind speed bias in the inertial sublayer whenever the ground roughness elements exert a relatively large roughness length $z_0$.

Here, we simulated ten flow fields in a wind farm considering ten values of the neutral drag coefficient $C_{DN}={\left[\kappa /ln(z/z_0)\right]}^2$ (as discussed in Section 2.3, Equation (11)). For each flow field, we scatter the resolved streamwise velocity $\tilde{u}(x,y,z)$ against the vertical coordinate z, and thus, estimate $u_{*}$ and $z_0$ to obtain the best fit to Equation (4).

Table 2. The table shows the cases used to study the effect of complex terrain on the performance of wind farms. The grid resolution was fixed at ${\Delta }_x={\Delta }_y=20\mathrm{m}$ and ${\Delta }_z=10\mathrm{m}$ after we had identified the most suitable one from the mesh convergence analysis. Classification and corresponding aerodynamic roughness length ($z_0$) are also listed.

Cases	${\mathit{z}}_0$	Classification	${\mathit{u}}_{*}$
A	$0.0005$	sea	$0.3358$
B	$0.01$	beaches, morass	$0.4451$
C	$0.03$	grass prairie	$0.5054$
D	$0.05$	airports, heather	$0.5390$
E	$0.1$	low crops	$0.5935$
F	$0.2$	high crops	$0.6597$
G	$0.3$	scattered obstacles	$0.7058$
H	$0.4$	trees, hedgerows	$0.7426$
I	$0.5$	mixed farm fields	$0.7738$
J	$1.0$	suburban houses, regular coverage of obstacles	$0.8903$

shows estimated values of $z_0$ and $u_{*}$ from these ten velocity fields.

To deal with the overall effect of hills and other roughness elements, the concept of an effective roughness and effective surface force ($u_{*}^2$, per area and normalized by density) is not new, particularly for hills, obstacle arrays, forests, and urban canopies [17,32,50,62]. Also, several past works have focused on “wall modelling" [25] and “near surface modelling” [29]. The present investigation’s near-surface model accounts for the complex terrain’s overall effects on the residual stress ${\tau }_{ij}$ and, hence, the wall shear stress ${\tau }_w\left(t\right)$. We have analyzed 10 LES flow fields corresponding to 10 values of effective roughness length $z_0$, listed in Table 2, to estimate the instantaneous shear stress on the Earth’s surface, ${\tau }_w\left(t\right)$. Here, we compute the shear stress as ${\tau }_w\left(t\right)=\sqrt{\left|\right|{\tau }_{13}{\left|\right|}_{\infty }^2+\left|\right|{\tau }_{23}{\left|\right|}_{\infty }^2}$, where $\left|\right|{\tau }_{i3}{\left|\right|}_{\infty }={max}_{x,y}\left|{\tau }_{i3}(x,y,\Delta z/2,t)\right|$ for $i=1,2$. Figure 10 shows the time evolution of the friction velocity $U_{*}\left(t\right)=\sqrt{{\tau }_w\left(t\right)}$ associated with four values of the roughness height $z_0$. Note the upper case symbol $U_{*}\left(t\right)$ for transient friction velocity. We see that a relatively smooth surface (with $z_0=5\times {10}^{-4}\mathrm{m}$) produces a relatively low friction velocity $U_{*}\left(t\right)$. However, increased roughness enhances fluctuations, impacting turbulent structures and momentum flux aloft. Based on a scaling analysis of the atmospheric boundary layer [32], the influence of complex terrain on the shear stress ${\tau }_w\left(t\right)$ indicates the corresponding impact on the downward flux of energy and the geophysical potential for the wind energy density.

The primary source of kinetic energy for a wind farm is the geostrophic wind in the free atmosphere, which is transferred to the wind turbine. A basic understanding of flow phenomena associated with performance loss for downstream turbines in a wind farm may lead to improvements in wind energy harvesting. For instance, consider the wake effects in the Horns Rev wind farm [86], which consists of 80 Vestas V-80 wind turbines covering an area of $20{\mathrm{km}}^2$. The layout of Horns Rev consists of an array of 10 rows and 8 columns. The rows are rotated approximately 7 degrees counterclockwise from the north–south direction. Wu and Porté-Agel [86] demonstrated the wake effects in Horns Rev wind farm using an LES in a domain of approximately $16{\mathrm{km}}^2$. The simulation of Wu and Porté-Agel [86] assumed a fixed hub-height velocity ($z_h=80\mathrm{m}$) of $8\mathrm{m}/\mathrm{s}$ and an aerodynamic roughness length of $z_0=5\times {10}^2\mathrm{m}$. The LES of Horns Rev wind farm [86] provides a reference for computational studies of wind farms.

In order to compare the present method’s accuracy concerning Wu and Porté-Agel [86], we simulate a wind farm where the hub height ($z_h=100\mathrm{m}$) wind speed is $10\mathrm{m}/\mathrm{s}$ and the aerodynamic roughness length is $z_0=0.05$ m. The wind farm consists of an array of $12\times 5$ wind turbines, covering approximately an area of $36{\mathrm{km}}^2$. Thus, the present wind farm is different from the Horns Rev wind farm (and that of Wu and Porté-Agel [86]). However, in Figure 11, we observe that the performance of the present wind farm is very similar to that of Wu and Porté-Agel [86].

Let us briefly discuss the impact of surface roughness on wind farm performance, where we varied the surface roughness between $z_0=5\times {10}^{-4}$ m and $z_0=1.0$ m. As discussed above, increasing the overall roughness height $z_0$ due to hills and forests would also increase the downward energy flux and the wind speed in the inertial sublayer [76]. The relative power $P_n/P_1$ measures the resulting impact on the wind farm performance, where $P_n$ is the average power extracted via a turbine at the n-th row (see Section 3.2.2). Figure 12 compares $P_n/P_1$ for four values of the roughness height $z_0$. For $z_0=5\times {10}^{-4}$ m, each turbine after the third row indicates having the same performance of approximately 40% relative to the first row. There are known reasons for such an observed performance. For instance, the partial wake recovery within the wind farm after a few rows (three in the present study) is a geophysical potential due to the vertical flux of atmospheric turbulence. For $z_0=1.0$ m, the wake is recovered to approximately 80% after the fifth row. Thus, the overall effect of complex terrain is likely to increase the available kinetic energy for wind turbines. An enhanced momentum exchange between the land and atmosphere for a fully rough surface results in more energy entrainment from aloft [1,2].

The present study needs to provide a detailed analysis of the local production of turbulence kinetic energy and the associated load on wind turbines [1,6]. However, Figure 12 suggests an overall reduction in turbulence kinetic energy from complex terrain in wind farms. As we know, turbulence is highly intermittent; local production of turbulence and shear stress may be responsible for fatigue load.

3.4. Relation between the Power and the Drag Coefficient

As the complexity (i.e., slope, distribution, etc.) of the roughness elements increases, the drag coefficient also increases, thereby causing an increase in the roughness length $z_0$ [32]. Evaluating the drag coefficient from observations and identifying the variables that influence it, such as wind speed, is a costly experimental process. For a neutrally stratified atmospheric boundary layer over a fully rough surface, $C_{DN}$ is obtained by using Equations (11) and (18) (see [32]):

$$C_{DN}=k^2/{\left[ln(zg/{\alpha }_c{u}_{*}^2)\right]}^2.$$

(19)

Formulating a relationship between the power production and the drag coefficient for wind farms sited over a complex terrain would be interesting. Wind turbines may also contribute to the roughness length $z_0$. In other words, the wind farm’s density and layout can affect the drag coefficient $C_{DN}$ affecting the power output. The present study unveils a relationship between heterogeneous spatial surfaces and power dependence, which can be instrumental in optimizing power generation and drag in large-scale wind farms. We obtain the drag coefficient ($C_{DN}$) from Equation (19) for each case shown in Table 2, where $z_0$ and $u_{*}$ are computed by fitting the LES data with Equation (4). We took the value of Charnock’s constant, ${\alpha }_c\approx 0.0144$ [32].

Figure 13a shows a correlation between the neutral drag coefficient $C_{DN}$ and the square of the friction velocity $u_{*}$.

The best least-square fit indicates a linear relationship between $C_{DN}$ and wall shear stress (i.e., squared friction velocity). This finding is consistent with the Charnock formulation illustrating the impact of wind speed on the drag coefficient [32]. Notably, as the roughness of the terrain increases, wind speed near the ground decreases while accelerating it near the hub height or aloft and thus, leading to enhanced wind power production [6]. However, increasing the complex terrain’s roughness also increases the overall drag in the wind farm [11]. In the present study, we have six hundred time series corresponding to 10 cases of $12\times 5$ wind turbine arrays. For Case J, the aerodynamic roughness height $z_0=1$ m is an order of magnitude (e.g., about one-tenth) smaller than the average height of the roughness elements of a suburban region [70,73]. We calculate each 600 wind power time series using Equation (15), accounting for the wake effects. We obtain the relative power $P=P_m/P_1$ for each of the cases listed in Table 2, where $P_m$ is the wind power averaged over $12\times 5$ turbines for m-th case (e.g., $m=1$ refers to Case A in Table 2). We obtained the best least-square fit between $P_m/P_1$ and the drag coefficient $C_{DN}$. Figure 13b shows the relative power $P_m/P_1$ as a function of the neutral drag coefficient $C_{DN}$. We observed approximately a 68% increase in the power production for Case J (suburban roughness) relative to Case A (offshore roughness). This result is consistent with the literature. It is worth mentioning that past studies resolved the shape of the roughness element, such as scaled models of isolated hills or ridges (e.g., [6,88]). Such experimental and numerical studies reported an increased normalized power of 50–80%. Resolving the roughness elements’ shape is crucial for optimizing wind farm layouts [1,88]. However, the present work contributes additional insights into the turbulent flow around onshore wind farms in the presence of different classes of complex terrain.

The best-fit linear equation for the wind power as a function of $C_{DN}$ is

$$P=({\beta }_1+{\beta }_2{C}_{DN})\times {10}^{-3},$$

where the parameters ${\beta }_1$ and ${\beta }_2$ are approximately $0.747$ and $0.114$, respectively. This relationship helps advance the analytical model for analyzing the power production of wind farms in complex terrain.

3.5. Dispersive Stress Analysis

Characterizing secondary turbulent motions is essential to quantify the degree of spatial heterogeneity in large wind farms. However, the turbulent transport of momentum in wind farms is very complicated due to the interaction of wind turbines and complex terrain. Here, we consider the Double-Averaged Navier–Stokes (DANS) equations to evaluate the dispersive (or form-induced) stresses, representing turbulent momentum transport due to spatial heterogeneity.

Let us consider the Reynolds decomposition of the solution of Equation (2):

$${\tilde{u}}_i(\mathit{x},t)={\overline{u}}_i(\mathit{x},t)+u_i^{\prime }(\mathit{x},t),$$

(20)

where

$${\overline{u}}_i(\mathit{x},t)\equiv \frac{1}{\mathcal{N}}\sum _{n=1}^{\mathcal{N}}{\tilde{u}}_i(\mathit{x},t;n),$$

(21)

is the time-average of $\mathcal{N}$ snapshots of the LES resolved velocity ${\tilde{u}}_i(\mathit{x},t;n)$. In the present analysis, each snapshot is the velocity at n-th time step with CFL = 1. The number of snapshot ($\mathcal{N}$) is large enough to capture the flow for a duration of 45 eddy turnover time units at a sampling rate of 1 Hz, yet average velocity varies slowly in time. Next, we consider the spatial averaging of the time-averaged velocity field:

$$\langle {\overline{u}}_i\rangle =\frac{1}{A}\int \int \overline{u}(\mathit{x},t)dxdy,$$

(22)

where A is the area of a horizontal plane parallel to the ground surface at $z=0$. The application of the dual operation of time and spatial averaging in Equations (21) and (22) results in the triple decomposition of the LES resolved field $\widetilde{u_i}(\mathit{x},t)$. Thus, we have

$${\tilde{u}}_i(\mathit{x},t)\equiv \underset{{\overline{u}}_i(\mathit{x},t)}{\underbrace{\langle {\overline{u}}_i\rangle +u_i^{″}(\mathit{x},t)}}+u_i^{\prime }(\mathit{x},t).$$

(23)

According to the above decomposition, the dispersive stress associated with the horizontal spatial averaging is

$$D_{ij}\left(z\right)\equiv \langle ({\overline{u}}_i-\langle {\overline{u}}_i\rangle )({\overline{u}}_j-\langle {\overline{u}}_j\rangle )\rangle =\langle u_i^{″}{u}_j^{″}\rangle .$$

(24)

Note, however, that the Reynolds stress is due to the joint variability of the randomness between the velocity fluctuations:

$$R_{ij}\left(\mathit{x}\right)=\overline{(\widetilde{u_i}-{\overline{u}}_i)(\widetilde{u_j}-{\overline{u}}_j)}=\overline{u_i^{\prime }{u}_j^{\prime }}.$$

(25)

We have computed the dispersive and the Reynolds stresses using Equations (24) and (25), respectively, and compared their relative contributions with the total drag. We obtained the maximum value of the stresses as a function of the drag coefficient $C_{DN}$ (see Equation (19)). Figure 14 shows the variations in dispersive stress components concerning the drag coefficient. The first notable observation from Figure 14a is that the streamwise normal-component $\langle u_1^{\prime \prime }{u}_1^{\prime \prime }\rangle$ of the dispersive stress is relatively insensitive to the drag coefficient. However, the spanwise $\langle u_2^{\prime \prime }{u}_2^{\prime \prime }\rangle$ and surface-normal components $\langle u_3^{\prime \prime }{u}_3^{\prime \prime }\rangle$ vary linearly with the drag coefficient (Figure 14b,c). In contrast, the corresponding Reynolds stresses vary linearly with the drag coefficient, as shown in Figure 15. It is interesting to note that the magnitude of the streamwise component of dispersive stress $\langle u_1^{\prime \prime }{u}_1^{\prime \prime }\rangle$ is higher than its Reynolds stresses $\overline{u_1^{\prime }{u}_1^{\prime }}$ counterpart. This result suggests that the dominant source of spatial heterogeneity within a wind farm is in the streamwise direction, which is the dominant “wake-occupied” region. Additionally, the dispersive stress shows minimum variations because the present study considers an aligned arrangement of 60 wind turbines. Further research could explore this idea by quantifying the degree of spatial heterogeneity with different layouts and numbers of wind turbines and by resolving the complex terrain. Furthermore, the relative contribution of the dispersive shear stress component $\langle u_1^{\prime \prime }{u}_3^{\prime \prime }\rangle$ is almost $45\%$ of the Reynolds shear stress to the drag.

4. Conclusions

This article presents a scale-adaptive LES of wind farms in complex terrain, incorporating a near-surface model. Using the LES data of an idealized wind farm, we have analyzed the impact of mountainous terrain on the performance of wind farms, where the simulation does not explicitly resolve the shape of the mountain or other roughness elements. We observe that the near-surface model helps improve the accuracy of capturing the atmospheric boundary layer turbulence without needing an extremely refined mesh. We investigated the effect of grid resolution on the ABL and wake profiles and observed that the ABL wind profile is not very sensitive to grid resolution. In contrast, the wake profile depends on grid resolutions. When the grid has about ten grid points across the rotor, at least in the vertical direction, the wake profiles converge according to the grid refinement analysis.

Our findings indicate a 68% increase in normalized power production in wind farms situated in mountainous and forested terrain compared with a flat or offshore location. This study treats the subgrid-scale effects of terrain complexity without resolving the shape of roughness elements. We observe a linear correlation between the power production and the drag coefficient and show that drag increases as the surface roughness of the terrain increases. These findings help us understand the consequences of atmospheric turbulence on the performance of wind farms.

We provide a quantitative assessment of the secondary motion due to spatial heterogeneity in onshore sites of wind farms. We analyze the dispersive and Reynolds stresses. We follow the double-averaged Navier–Stokes equations to formulate the dispersive stresses. We show that a dominant source impacting a wind farm through spatial heterogeneity is due to the incoming turbulent flow, where the wind turbines are perpendicular to the streamwise direction. Furthermore, we observed that the dispersive shear stress contributes nearly 45% of the drag of wind farms in complex terrain compared with the Reynolds shear stress.

The development of wind farms in complex terrain poses significant challenges. The site assessment requires assessing performance due to the interaction of wind turbine wakes and atmospheric boundary layer flow. Different from flat terrain, it is challenging to characterize the complex terrain’s wind patterns because measurements at one point may not represent the flow state at points only a few hundred meters away. The present study highlights the importance of accurately modelling the subgrid-scale effects of the secondary motion due to the spatial heterogeneity imposed by the complex terrain. Future research may focus on capturing the grid-scale variation in the topography while modelling the relevant subgrid-scale effects. There are two crucial questions. Can artificial windbreaks be considered to compensate for the blockage of mountains on the reduced performance of turbines? If trees can provide potential windbreaks, it would be interesting to characterize the optimal placement of trees on mountainous terrain. Mountain and internal waves are essential to assess the potential role of atmospheric turbulence on wind farms in complex terrain. Since the middle of the twentieth century, the atmospheric science community has confirmed that complex terrain substantially impacts large-scale circulation and stationary waves, subsequently affecting the regional climate. Suppose we aim for a global net-zero emissions by 2050 in the context of a 1.5 degrees change scenario. In that case, we must incorporate such a large-scale effect of mountains in optimizing the wind farm layout.