CFD in Urban Wind Resource Assessments: A Review

Abstract

Urban distributed energy systems play a crucial role in the development of sustainable and low-carbon cities. Evaluating urban wind resources is essential for effective wind energy harvesting, which requires detailed information about the urban flow field. Computational fluid dynamics (CFD) has emerged as a viable and scalable method for assessing urban wind resources. This review paper synthesizes the characteristics of the urban wind environment and resources, outlines the general framework for CFD-aided wind resource assessment, and addresses future challenges and perspectives. It highlights the critical need to optimize wind energy harvesting in complex built environments. The paper discusses the conditions for urban wind resource assessment, particularly the extraction of boundary conditions and the performance of small wind turbines (SWTs). Additionally, it notes that while large eddy simulation (LES) is a high-fidelity model, it is still less commonly used compared to Reynolds-averaged Navier–Stokes (RANS) models. Several challenges remain, including the broader adoption of high-fidelity LES models, the integration of wake models and extreme conditions, and the application of these methods at larger scales in real urban environments. The potential of multi-scale modeling approaches to enhance the feasibility and scalability of these methods is also emphasized. The findings are intended to promote the utilization and further development of CFD methods to accelerate the creation of resilient and energy-efficient cities, as well as to foster interdisciplinary innovation in wind energy systems.

1. Introduction

Wind energy plays an important role in energy decarbonization, with field-installed and offshore wind farms generating 8% of global electricity by 2023 [1]. However, the average physical distance from generation to consumption remains far, causing higher loss during the transmission [2,3]. As a supplementary, small wind turbines (SWTs) could be installed within the built environment and act as a part of an urban-scale distributed energy system [4]. Such distributed wind energy systems are more flexible to adjust according to real-time consumption compared to centralized wind farms, and could be locally operated and maintained on demand, bringing unique advantages to end users.

Urban wind resource evaluation, which acts as the first step of integrating SWTs into the urban built environment, is fundamental for urban wind energy harvesting. Unlike large-scale wind farms, an urban wind energy system is usually distributed and requires detailed information on the urban flow field [5,6]. Although on-site measurements can provide direct data for wind resource assessment, they have notable limitations when applied to urban wind energy studies, due to the limited spatial and temporal resolution of complex urban wind patterns [7] and inability to manipulate the boundary condition [8]. Wind tunnel experiments, though useful for simulating the flow field with a controlled condition, are constrained by their cost, especially when high-resolution data are required [9]. Furthermore, the placement of measurement equipment in urban environments can introduce significant errors, as they interfere with building topography [10]. To address all of these challenges, computational fluid dynamics (CFD), which can reconstruct urban flow with finer resolution both in spatial and temporal variables with high effectiveness and affordability [11], has been utilized as a feasible and scalable approach for urban wind resource assessment [12,13,14]. Meanwhile, the growth of hardware performance and extensive application of commercial CFD codes also improved the efficiency of CFD-aided wind resource assessment [15,16].

Hence, CFD methods in urban wind resource assessment are evolving, with new studies continuously bringing benchmarking and technical correction to the analytical frameworks, models, and boundary conditions. This review article covered the topic of “Computational Fluid Dynamics” or “CFD” and “Urban Wind Energy” or “Urban Wind Resource” in Web of Science. It was found that 39 articles were published from 2020 to 2024, while only 20 were published from 2015 to 2019, as shown in Figure A1 in Appendix A. Although this collection is far from complete, it does show a trend of increasing studies and rising interest in CFD in this field. To promote the standardization of the utilization of the CFD approach in urban wind resource assessment, this review article proposes to summarize the general process and key points in CFD-aided urban wind resource assessment. Specifically, this review focuses on the methodology of urban wind resource assessment; hence, the cited publications are not limited to urban wind assessment, but have been expanded to urban wind engineering, which has a lot in common in its methods.

The main content of this review article includes the following:

(a)

General framework of CFD-aided urban wind resource assessments, in order to produce reliable and mutually recognized results;

(b)

Stringency and permissiveness in case settings, leading to balance between accuracy and computational cost;

(c)

Remarks on current methods, as well as potential challenges and prospects.

2. Urban Wind Environment and Resource

2.1. Characteristics of Urban Wind Environment

Wind resources are fundamentally determined by the meteorological conditions. At the same time, they change significantly with the surface roughness and vary with the height. A vertical profile of the wind flow within the atmospheric boundary layer (ABL) can be described using the logarithmic law [17] or power law profile [18], based on terrain roughness specification. The most used profile set is the logarithmic law [17], defined with velocity, turbulent kinetic energy (k), and turbulence dissipation rate (ε) and based on aerodynamic roughness length (z₀) and friction velocity (u_∗), as Equations (1)–(3) show.

$$U={\displaystyle \frac{u_{*}}{\mathsf{\kappa }}}\ln \left({\displaystyle \frac{z+z_0}{z_0}}\right)$$

(1)

$$k={\displaystyle \frac{u_{*}^2}{\sqrt{{\mathrm{C}}_{\mathsf{\mu }}}}}$$

(2)

$$\epsilon ={\displaystyle \frac{u_{*}^3}{\mathsf{\kappa }(z+z_0)}}$$

(3)

where C_μ is the constant to be determined by the turbulence mode and κ is the von Karman constant. Modified profiles for k and ε were also introduced to consider the decay of k [19] or to more finely adapt different turbulence models [20].

The key parameters, u_∗ and z₀, are determined by the terrain roughness. In urban areas, both parameters change significantly due to the complex geometry of the urban areas, forming a unique urban boundary layer (UBL). Though it adopts the classical ABL structures, the distinctive feature of UBL is that an urban roughness sublayer (RSL) exists, due to the significant drag from the urban elements. The RSL typically extends from the ground up to 1.5–3H with H be the height of urban elements (e.g., buildings and trees) [21]. Within this layer, the mean flow is dominated by the turbulence and ultimately by urban morphology [22]. Within the RSL, urban flow is complex and locally favorable for building-mounted SWTs, as buildings at certain positions may get washed by strong detour flows. Figure 1 gives an example of the varied horizontal pattern (Figure 1a) and vertical profile (Figure 1b) of the urban flow due to the impact from the buildings. As a result, if one SWT were to be installed in such a built environment, it is necessary to simulate the area of interest together with the surrounding urban structures instead of only a few buildings related to the SWT installation, and any inlet profile should be put far enough upstream of the area of interest.

2.2. Description of Urban Wind Resource

As described before, SWTs benefit from the local wind amplification brought by complex urban flow. However, the flow is not fixed, as magnitude and direction always change due to meteorological factors. An urban wind resource assessment, therefore, should consider the changing of weather conditions during a period of time, and as a result, CFD simulation must be performed with multiple and carefully designed conditions. Conditions should be designed with the following factors.

(a)

Wind direction and corresponding frequency;

(b)

Wind velocity and turbulence variables;

(c)

Extreme conditions considering natural windstorm disasters, especially tropical cyclones.

Acquiring the information with observation might be a difficulty, as a rigorous measurement period could be four years or more [23]. A common practice is to fit the short-period observation with a probability distribution and extrapolate the velocity with probability. The Weibull distribution is the most used one to represent the wind characteristics at a fixed location [24], even in the built environment [25]. The cumulative distribution function (CDF) and probability density function (PDF) of a three-parameter Weibull distribution is as in Equations (4) and (5) [26].

$$F\left(x\right)=1-\exp \left[-{\left({\displaystyle \frac{x-\zeta }{\alpha }}\right)}^{\beta }\right]$$

(4)

$$f\left(x\right)={\displaystyle \frac{\beta }{\alpha }}{\left({\displaystyle \frac{x-\zeta }{\alpha }}\right)}^{\beta -1}\exp \left[-{\left({\displaystyle \frac{x-\zeta }{\alpha }}\right)}^{\beta }\right]$$

(5)

where x is the fitted variable, for example, the 10 min mean velocity magnitude of wind, and α, β, and ζ are parameters of the Weibull distribution. However, this fitting method can only shorten but not eliminate the observation process, and a small amount of observation data or other data of the same authenticity is always required. It is proven that the maximum likelihood estimation (MLE) is the most reliable and accurate method to fit the wind characteristics along a fixed period, in both unperturbed [27,28,29] and strongly perturbed urban [30,31] conditions. An MLE algorithm could be easily programmed and executed with any Python or MATLAB libraries accordingly.

With the Weibull distribution, it is also possible to acquire the accurate probability of extreme weather conditions that could be a challenge for the turbines. Typically, an extreme velocity value could be simply calculated by inversing Equation (4) and inputting the threshold probability, which is considered as the extreme.

2.3. Calculation of Energy Yield and Identification of Potential Wind Resource

The ultimate goal for urban wind resource assessment is to obtain the energy yield of SWTs at selected positions in the urban built environment. Therefore, it is crucial to understand the energy extraction process. An aerodynamic way to calculate the power of the wind (P) is described as Equation (6).

$$P={\displaystyle \frac{1}{2}}\rho A{U}^3$$

(6)

where ρ is the density of air, A is the swept area of the wind turbine, and U is the streamwise velocity of wind. However, wind turbines, in practice, can only harvest a small portion of the power of the wind, and the actual output power of the turbine is equal to the power of the wind multiplied by a power coefficient C_P, as Equation (7) describes.

$$P_T=C_P\cdot P$$

(7)

where P_T is the output power of the turbine. The power coefficient C_P is determined by the design and working condition of the turbine. A theoretical maximum of power coefficient C_P is 59.3% and 29.6% for turbines using lifting force and drag force, respectively [32,33]. The power coefficient is typically drawn in a power curve, as a function against wind speed. The cut-in/out speeds define the domain of this function. Once the velocity of wind is beyond the domain, the power coefficient is zero.

It is obvious that a slight increase in wind speed can bring a significant increase in energy harvest, as the energy of the wind is linear to the cube of the velocity of wind. A good choice of installation position overshadows any technical advance in SWT itself.

Heterogeneous buildings in the urban area can create local amplifications of wind speed that are regarded as favorable to SWTs [34], and the typical flow patterns were revealed by different studies [35,36,37]. This review further categorizes those flow patterns into two types: detour flow and artificial venturi. Detour flows are the most common patterns, where the wind accelerates at the sides and top of a building. Artificial venturi usually appears on tall buildings that are specially designed for wind resource harnessing. These positions in an urban building cluster should be identified and considered as a sampling point in the related CFD simulation. Currently, the rooftops are the most investigated position for installing SWTs, where the amplification of wind due to detour can be fully utilized, and the spaces for installing SWTs are relatively larger. It was suggested to be feasible to gain more energy by installing both HAWTs [38] and VAWTs [39] on rooftops.

However, a larger wind speed often brings larger turbulence. Turbulence is usually described by turbulence intensity (I) or turbulent kinetic energy (k), and has a negative effect on SWTs’ output and lifespan [40]. Sometimes the overly high wind speed may lead to a cut-off as well. Therefore, a comprehensive wind resource assessment should consider the all of the performance parameters of an SWT.

3. CFD-Aided Assessment of Wind Resource

3.1. Handling of Turbulent Terms

CFD simulation is a numeric and approximate solution of the governing equations of the fluid dynamic based on finite volume method. These governing equations are the Navier–Stokes (N-S) equations, which satisfy the mass conservation and Newton’s second law. There are five major equations for a three-dimensional, confined and incompressible Newtonian flow, containing one continuity equation, three momentum equations, and one energy equation. For urban wind resource studies, the energy equation could be omitted because the Richardson number (Ri) is too low in the studied area, and the buoyancy effect cannot significantly affect the convective flow pattern. The N-S equations, however, are not solvable analytically, for there are partial-differential equation terms which describe the turbulent flow. Therefore, two major approaches were invented to turn convert these terms and solve the N-S equations numerically.

In the Reynolds-averaged Navier–Stokes (RANS) approach, the transient and turbulent flow is averaged in a temporal way, where the flow is regarded as steady and turbulence terms are modeled with steady parameters. Based on this approach, two different types of models were developed, namely the Boussinesq hypothesis-based models, including the Spalart–Allmaras (SA) model [41], k-ε series [42,43,44] and k-ω series [45] models, and the Reynolds stress transportation-based model, including the Reynolds Stress Model (RSM) [46]. For example, in the widely used k-ε and k-ω series models, the turbulent flow is averaged over flow time and modeled by turbulent kinetic energy (k) and its dissipation rate (ε) or specific dissipation rate (ω).

In the other widely adopted approach, large eddy simulation (LES), the flow is averaged spatially, where only large eddies are considered and the small eddies within the discretized cells are computed using sub-grid scale (SGS) models. For urban wind simulation, the Smagorinsky–Lily, wall-adapting local eddy-viscosity (WALE), and algebraic wall-modeled LES (WMLES) models are mostly used.

A figure from a previous review stunningly illustrated the difference between these two models [47], as Figure 2 shows. It should be noted that a direct Navier–Stokes (DNS) approach is not discussed here due to its infeasibility in application.

A list of turbulence model utilization in relevant studies is shown in

Table 1. Recent studies’ choices of turbulence models.

Ref.	Year	Turbulence Model
[48]	2015	RANS, SST k-ω
[49]	2015	RANS, standard k-ε
[50]	2016	RANS, standard k-ε
[51]	2016	LES, dynamic Smagorinsky
[52]	2016	RANS, realizable k-ε
[53]	2016	RANS, realizable k-ε
[54]	2017	RANS, standard k-ε
[55]	2017	RANS, realizable k-ε
[56]	2017	RANS, standard k-ε
[57]	2017	RANS, SST k-ω, and SST-SAS
[58]	2017	RANS, RNG k-ε
[59]	2018	RANS, realizable k-ε
[60]	2020	RANS, realizable k-ε
[61]	2020	RANS, standard k-ε
[62]	2021	RSM
[63]	2021	LES, dynamic Smagorinsky
[64]	2021	RSM
[9]	2022	LES, WALE
[11]	2022	RSM
[65]	2022	RANS, SST k-ω
[66]	2022	RANS, SST k-ω
[67]	2022	RANS, standard k-ε
[68]	2022	RANS, SST k-ω
[38]	2022	LES, dynamic Smagorinsky
[69]	2023	RANS, realizable k-ε
[70]	2023	RSM
[71]	2024	RSM

. It is clear that RANS models are being widely used to reconstruct the urban flow, while only a few studies selected LES models. The critical factor for selecting a turbulence model is the trade-off between simulation accuracy and computational cost, which are distinctly different between RANS and LES models.

RANS models are known to acquire “a decent level of accuracy” with moderate computational cost. The most used k-ε series and k-ω series models solve only two more turbulence equations, while the RSM models must solve seven more equations in three-dimensional flow but give a more accurate solution. The one-equation SA model can of course acquire the solution with even less computational cost but with compromised accuracy. The common point of all RANS models is that their solution is acquired from a convergence process, which will consume a certain number of iterative time steps, and their result is steady and simple to process, with time-averaged velocity representing the momentum distribution within the domain, and turbulent variables representing the fluctuation of this momentum distribution. The LES and LES-based models can provide a transient solution that is closer to wind tunnel tests and on-site measurements, especially in predicting turbulence variables [72], but it requires a greater computational cost, at least one order higher than RANS models. LES only solves the flow within one time step in every iteration. Therefore, apart from its stringent requirement for a computational grid, which will be discussed in Section 3.2, the flow data will have to be collected through a long enough flow time, which further increases the computational cost.

In addition, the requirements of spatial discretization and temporal discretization are in fact coupled with a Courant–Friedrichs–Lewy (CFL) number requirement. This CFL number is defined as the flow characteristic velocity multiplied by the ratio of temporal discretization size to spatial discretization size, and is required to be less than 1.0 in non-iterative transient schemes [73]. Namely, the smaller the computational grid, a smaller time step size and a larger total flow time should be set. Here, a parameter denoting computational cost (CC) for a non-iterative transient scheme is defined in Equation (8).

$$\mathrm{CC}=c_{compute}\cdot {\displaystyle \frac{V}{\Delta V}}\cdot {\displaystyle \frac{T_{init}+T_{sample}}{\Delta T}}=c_{compute}\cdot {\displaystyle \frac{V}{l^3}}\cdot {\displaystyle \frac{u_{\mathrm{CFL}}\left(T_{init}+T_{sample}\right)}{\mathrm{CFL}\cdot l}}$$

(8)

where c_compute is a parameter depending on the researcher’s computing capacity, V and ΔV denote the volume of the whole computational domain and average grid cell, respectively, T and ΔT denote the flow time and average time step size, respectively, l denotes the average cell size, and u_CFL is the characteristic velocity in CFL number. It is recommended that T_init and T_sample are sufficient for the flow to have one and two complete flow-overs, respectively, in order to provide sampling-time-insensitive results. Equation (8) also suggests that, with a fixed CFL number and total flow time, computational cost is directly proportional to l⁻⁴ and will increase drastically with a refined computational grid. Consequently, LES models and LES-based hybrid models (including detached eddy simulation (DES) [74] and embedded LES (ELES)) are chosen less often to reconstruct urban flow, especially high-Re flows, as shown in Table 1.

The LES-based hybrid models, though rarely documented, could be a potential remedy for the high computational cost in urban wind engineering. Both DES and ELES models feature the reduction of the computation of resolved turbulence while keeping a good accuracy as pure LES modes. DES models treat the wall-bounded flow as modeled and the rest as resolved [75], while the ELES model treats the flow according to a manual partitioning of the computational domain. Studies have shown that DES and ELES models can both reach the same level of accuracy when predicting turbulent urban flows when tackling pedestrian-level wind comfort and wind hazard problems [76,77,78], building wind load and vibration problems [79,80], pollutant dispersion problems [81], and regular urban flow problems [82]. With all of these successful implementations in relevant studies, there is no reason that LES-based hybrid models should be excluded from the wind resource assessment toolbox. In practice, however, the accuracy of DES is dependent on finely tuned model constants and wall-bounded grids as well as the carefully chosen boundary layer RANS model to ensure a proper detaching position, and the grid count could be significantly reduced to save computational cost [77,82,83], up to 20% [76]. ELES in contrast, though accurate as well [78,80], can only mitigate the computational burden in the peripheral domain and may not be able to reduce the grid count to the same level as DES. Anyway, LES-based hybrid models are transient models, and will still suffer from the transient sampling process for a long enough period, guaranteeing the same order of magnitude for their computational cost as pure LES models.

Nevertheless, LES-based models provide a high-fidelity numerical method for computation wind engineering [84]. In particular, for a complex urban area, in most cases, the RANS simulation can hardly reconstruct the turbulence variables precisely and may underestimate the turbulent kinetic energy k or turbulence intensity (I) [47,85,86]. This drawback will adversely affect the estimation of lifespan [40] which is usually calculated from k or I. Thus, before choosing a turbulence model, researchers are recommended to fully consider their specific research demand and computing capacity. For instance, LES and LES-based hybrid models can provide detailed and dynamic results (e.g., in verification processes of an SWT’s installation), and RANS could be utilized on a large scale and with large design margins (e.g., in early-stages of wind resource assessment).

Table 2. Performance metrics of currently proven turbulence models.

Performance	RANS	LES	ELES	DES
Handling of turbulence	Modeled	Resolved	Modeled (peripheral domain)/Resolved (core domain)	Modeled (boundary layer)/Resolved (rest of domain)
Accuracy of turbulence reconstructing	Low (k-ε and k-ω series)/Medium (RSM)	High	High	High
Computational cost	Low (k-ε and k-ω series)/Medium (RSM)	High	Medium-high	Medium-high

shows the performance of currently proven turbulence models for urban wind engineering.

Once the turbulence model is chosen, research should refer to the best practice guidelines (BPGs). There are many general BPGs on urban wind simulation and specific BPGs for pedestrian wind or building wind loads could also be referred to when conducting a computational wind engineering simulation. The COST BPG [87] published in 2004, which was based on the summarization of early CFD and wind-tunnel studies in Europe, is now recognized as the first conventional BPG on general wind engineering. This document was later updated in 2007 [88]. The Architectural Institute of Japan (AIJ) published their BPG on pedestrian wind as well as building wind load in 2008, concentrating on the CFD and wind-tunnel studies [18,89].

3.2. Computational Grid

A high-quality computational grid is fundamental to the computation convergence and accuracy of the results. The selection of grid type is determined by turbulence model, solution scheme, and grid independent analysis. When constructing the computational grid, it is recommended to roughly consider the following three points [18].

(a)

Distinguishing the core region from the peripheral region, and ensuring the cell sizes in the core region are one order smaller than local building size.

(b)

Refining the boundary layer region with cells parallel to the wall, and ensuring the expansion ratio is smaller than 1.5.

(c)

Generating three computational grids for validation purposes.

Once the grid is generated, it should be validated. The grid should be put into computation and results should be evaluated. The major indicator is the non-dimensional wall distance (y⁺), calculated using Equation (9).

$$y^{+}={\displaystyle \frac{u_{\tau }y}{\nu }}=\sqrt{{\displaystyle \frac{{\tau }_w}{\rho }}}\cdot {\displaystyle \frac{y}{\nu }}$$

(9)

where u_τ is the friction velocity, y is the wall distance, ν is the kinematic viscosity, τ_w is the wall shear stress, and ρ is the density of the fluid. The local y⁺ value of a grid cell indicates its perpendicular position in the boundary layer.

Most RANS models have the wall-bounded flow modeled by wall functions (WFs). As a result, the y⁺ requirement is determined by the specific WF. In most cases, first-layer grid cells should fall in the logarithmic sublayer of the boundary layer. For example, when using a standard wall function for k-ε series models [90], the y⁺ of first layer grid cells should be between 30 and 100.

For LES studies, WFs are not applicable, and the wall-bounded flow should be resolved by the grid. In such cases, it is recognized that the y⁺ at all grid cells should be lower than or around 1.0 [91]; as a result, the first layer grid cells fall in the linear sublayer. In practice, however, some studies have shown the possibility of more flexible grid requirements, especially with LES models. For example, the velocity data sampled around the buildings is of high uniformity with different building y⁺ values, which range from 1.5 to 25 [92,93]. Some LES studies accepted even higher y⁺ values [11]. It should be emphasized that a y⁺ value beyond the model-specific range will definitely deteriorate the accuracy of the solution at the near-wall area, and whether this is acceptable in the context of the condition and interested area should be identified, as fully resolved wall-bounded flow is sometimes unnecessary and unpractical, especially for high-Re flows.

There is an additional requirement for LES grids, where they should be fine enough to resolve at least 80% of the total turbulent kinetic energy [94]. An LES_IQ_k is defined to signify this percentage as Equation (10) shows [95].

$${\mathrm{LES}\_\mathrm{IQ}}_k={\displaystyle \frac{k_{resolved}}{k_{total}}}={\displaystyle \frac{k_{resolved}}{k_{resolved}+a_k{h}^p}}$$

(10)

where k is the turbulent kinetic energy, h is the grid size, p is the order of accuracy of the numerical scheme, and a_k is a coefficient to be determined by comparing the performance of two sets of computation grids. For urban wind resource assessment studies, it is critical to accurately reconstruct the velocity and turbulent variables at near-wall areas of urban buildings where the roof-mounted SWTs are installed. Therefore, a strict compliance of y⁺ and resolved turbulent kinetic energy requirement are strongly recommended for LES studies.

After the computational grid is validated with y⁺ and LES_IQ_k indicators, a grid independence analysis will be conducted. Velocity and turbulent variables are recommended for grid independence analysis, in line with other computational wind engineering studies. The power coefficient of the wind turbine could also be considered [70], as the error from different grids is not essential and will not hide the error with flow. There are two typical ways to report the grid independence. Most studies consider only a set of scatter or line plots in which points of solution from one grid against another are compared. The other way is to compute the grid convergence index (GCI), which is deduced from Richardson Extrapolation [96]. The GCI value is calculated using Equation (11).

$$\mathrm{GCI}=F_S\left|{\displaystyle \frac{{\epsilon }_r}{1-r^p}}\right|$$

(11)

where ε_r is the relative error of interested variables sampled at interested points, r is the size ratio of grids that exceeds 1.0, p is the discretization precision, and F_S is the safety factor. A conservative value for F_S is 3.0, but for practical application a moderate value of 1.25 could also be accepted [96]. The GCI values reported by previous studies were mostly under 5% [11,59,61,62,71].

3.3. Computational Domain and Boundary Conditions

The computational domain for a wind resource assessment study can follow the settings of existing computational wind engineering studies. The dimension of the computational domain should be large enough to prevent the venturi effect and contain the stagnation flow and wake [97]. The BPGs have defined the smallest domain size based on the size of the interested urban geometry [87,89]. It should be at least five times the height of the tallest building (5H) in the urban geometry for the upstream and side margins, 15H for the downstream margin, and 6H for the domain height. Studies from another perspective suggested that the domain should have a blockage ratio (a projected area of building façades against the inlet boundary area) smaller than 3% to contain all of the detour flows [18]. Following the BPGs [18,87], a CFD urban wind resource simulation should generally consider four types of boundaries as

Table 3. Boundary conditions.

Boundary	Condition
Inlet	Velocity inlet with ABL profile
Outlet	Pressure outlet
Sides	Zero-gradient wall
Top	Zero-gradient wall
Ground	Rough wall
Building walls	Smooth wall

shows.

Boundary conditions in Table 3 are based on the recommended size of the computational domain. Special attention should be paid to the pairing of inlet and ground condition. The inlet ABL profile is determined by the terrain, and highly related to the ground roughness [98]. An ABL horizontal homogeneity check should always be performed in an empty domain, with same inlet and ground condition as the actual case, and the profile at the inlet plane and incident plane should be compared to confirm that the urban area is exposed to the correct inflow condition. There is no known study that deliberately implemented poorly-paired profiles to test their impact on building detour flow, which has a direct impact on SWT performance, but a relevant study on building wind load showed that a long upstream terrain (15H) without proper wall function will lead to error in the wind pressure coefficient up to 40% [99]. It is inferred that poorly-paired profiles will lead to errors in the simulated flow pattern and uncertainty in the final results of wind resource assessments.

Figure 3 shows the changes in the inlet profile with different ground roughness specifications [100]. The U and I denote the streamwise velocity and turbulence intensity, respectively, y₀ denotes the aerodynamic roughness length, and K_S denotes the physical or geometrical roughness height. However, in some cases where the upstream margin of the terrain is too large, even a correctly paired ground roughness specification is not capable of maintaining the ABL profile, signifying that the problem is with the model itself. For LES models, the only feasible remedy is to shorten the upstream margin as the boundary layer is fully resolved without a WF. For RANS models, the problem could be solved by replacing the standard WF with a modified one. For example, a modified, k-ε based WF is introduced to keep a perfect ABL homogeneity with less strict y⁺ requirements [101], and could be implanted in the solver with a user-defined function [99].

In addition, the inlet condition for LES and LES-based hybrid models should be equipped with fully developed turbulence flow to ensure the incident profile is with fully developed turbulence. Numerous wind engineering studies have confirmed the superiority of the vortex method (VM) over other models [102], which stochastically generate the velocity fluctuation based on the input turbulence variable [93,103]. If an ELES model is used, this velocity fluctuation should be applied on the RANS-to-LES interface.

3.4. Validation of Authenticity

The validation process is primarily benchmarking the continuous and global data from CFD against the intermittent and local data from on-site measurement or the wind tunnel test. There are two types of validation, direct and indirect [104]. Direct validation is comparing the CFD results directly with observation results or experimental data. This type of validation is of high fidelity and can be applied only to the area of interest. However, direct validation for urban wind resources would require experimental data or observation results in the exact area of interest, which could be unpractical, as the desired condition in the real world cannot be reproduced. Indirect validation is essentially “model validation”, which validates only the CFD model under the same or a similar condition, and assumes the final results have the same fidelity as the ones in model validation. The former would always be preferred and recommended. However, for practical and specific studies, the latter would be used as measurement data are not always available.

For a wind resource assessment study, the validated variable should at least contain two core variables, namely the time-averaged velocity and turbulence intensity (or turbulent kinetic energy), in order to evaluate both the quantity and quality of wind energy. The final output of SWTs could also be considered if on-site measurement is available. For on-site measurement, light detection and ranging (LiDAR) [59,105,106,107] and an anemometer [108] are both capable of producing authentic data, while for the wind tunnel test, the most used is the hot-wire anemometer, which is in fact an invasive method and can only produce data with a lower level of reliability.

Once the data are obtained, a benchmarking should be conducted to compare the two datasets. Apart from benchmarking point by point [66], a comprehensive set of validation metrics is also utilized. For example, factor of two observations (FAC2) is defined to compare the data sampled at all points, as Equation (12) shows [109].

$$FAC2={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{N}_{FAC2i}}$$

(12)

where

$$N_{FAC2i}=\left\{\begin{array}{cc}\begin{array}{c}1\\ 1\\ 0\end{array}& \begin{array}{c}0.5\le {\displaystyle \frac{P_i}{Q_i}}\le 2\\ \left(\left|P_i\right|\le W\right)\&\left(\left|Q_i\right|\le W\right)\\ \mathrm{otherwise}\end{array}\end{array}\right.$$

(13)

where P and Q are normalized CFD results and observations, and W is the threshold depending on the variable [88]. This validation indicator equals to one when CFD results are in good agreement with observations. Stricter validation indicators including FAC1.05 [110] and FAC1.3 [111] could be conducted by replacing 0.5 and 2 in Equation (13). Other validation indicators include hit rate (q), fractional bias (FB), and normalized mean square error (NMSE). Researchers are recommended to follow the exact quality acceptance criteria for each of the validation indicators [109] and compare the validation results with other relevant studies. For example, wind tunnel test data of a nine-block case could be a basis for benchmarking [112].

4. Challenges and Prospects

4.1. Limited Fidelity Brought by Turbulence Model

Through this review of CFD methods, several challenges are identified. These challenges limited the development of CFD-aided urban wind resource assessment.

As validated by multiple studies, LES and LES-based hybrid models are regarded as high-fidelity CFD models for urban flow [113,114]. However, as listed in Section 3.1, there is no sign of a significant increase in studies using LES and LES-based hybrid models. Witnessing the leap in computational capacity over the last few years, this review does not regard the high computational demand as the only reason for this lack of growth in the use of LES and LES-based hybrid models. As another review article pointed out, most BPGs in wind engineering are for RANS and there is a lack of BPGs for LES [104]. The lack of BPGs surely contributed to the limited use of LES models, which lowers the overall fidelity of this field of research. It should also be pointed out that this is even more serious for LES-based hybrid models. With validated accuracy and improved efficiency, these hybrid models should be utilized more widely, but they are not. As a result, recognized reviewers are appealed to for supplementing current BPGs and including these high-fidelity models.

Still, the choice of the turbulence model always depends on the scale and purpose of the study. High-fidelity models should be considered when assessing the wind resources on a building and community scale. Correspondingly, for assessing larger scales such as the overall wind resource of a whole city or town when larger computational resources are not available, RANS models should be considered as the feasible alternative.

4.2. Limited Fidelity Brought by Geometry Simplification

The fidelity of studies is also subject to the level of detail in building geometries. As Figure 4 depicts, in simplified geometries, the whole block is integrated into one object; in approximated geometries, the buildings are separately modeled; and in detailed geometries, the details including sloped rooftops are modeled. It is argued that an approximated building geometry provides a satisfactory degree of accuracy in the context of a wind profile above average building height [115]. However, the fine details on the rooftops cannot be neglected in wind resource assessment for SWTs. It is found that rooftop shapes can create a significant difference in detour flow above the building, with turbulent kinetic energy deviating up to 50% [9,116]. It is also found that even the slightest rounding of sharp edges of the rooftops can significantly change the shape of power density curves within the detour flow [48,117]. However, refining the geometry is only feasible in single-building studies or method studies. For large-scale simulation, using detailed geometry will cause a series of problems concerning the turbulence model and computational grid, as discussed in Section 3. Even if the turbulence model is specially tuned and validated to adapt those details, there is no source for such detailed building geometry, as the building geometry is normally extracted from GIS databases using a Shapefile format based on vectors [118]. A prospect for future studies is to quantitatively evaluate the deviation brought by geometry approximation, so the results for SWTs could be calculated from approximated building geometry and then calibrated by qualitative information on rooftops and façades.

4.3. Wake Problems, Limitation of Benchmarking, and Extreme Conditions

An SWT can have 7.5 to 75 m² of swept area, equivalent to a rotor diameter up to 10 m [33], which is close to the characteristic length scale of urban buildings. The wake of the turbine can create significant disturbance to urban flow pattern and generate strong turbulence for downstream SWTs, lowering the power generation and shortening the lifespan with increased fatigue [119]. Since a combined CFD simulation of SWTs and the building cluster is unpractical, turbines are never resolved in CFD simulations but modeled using wake models [120,121,122]. Studies have demonstrated the feasibility of applied wake models for HAWTs [38] and VAWTs [39,123] in the urban built environment. However, a lot of studies regarded the turbine wake as negligible and assessed the urban wind resource without an SWT wake model, leading to an overestimation in energy yield and underestimation of turbulence level.

From another perspective, a problem remains when wake models are applied on a large scale, where traditional wake models are still too complicated [124]. It is expected that future studies could apply parameterized models into the CFD wind assessment and realize the large-scale and wake-included simulations.

As mentioned in Section 2, extreme conditions considering natural windstorm disasters should be considered when carrying out a wind resource assessment. Studies have found that tropical cyclones, as a source of wind hazard, are increasing in frequency and intensity due to climate change [125,126]. Sadly, few studies discussed specifically the definition and simulation method of extreme conditions in wind resource assessment. Hence, future studies, when simulating the wind resources in the urban area, could also apply the extreme condition to the building cluster and acquire the extreme working condition of the SWTs. In addition, the urbanization around the globe is still ongoing [127], leading to changes in building morphology and corresponding boundary conditions for wind resource assessment.

4.4. Integration of Multi-Scale Modeling

Apart from the method-related challenges, the objects and focuses of those wind resource studies also provided insights for the potential evolution of CFD methods.

The urban environment and climate consist of different scales, including single facet, building, canyon and blocks, neighborhood, and city. The flow is affected by the forces across the range of different urban scales [21]. In most cases, the area of interest in wind resource assessment is surrounded by a dense urban area, which determines the local boundary condition and inlet profiles. The weather research and forecasting (WRF) model combined with the urban canopy model (UCM) is also a useful tool to provide authentic and local-scale (down to 1 × 1 km) conditions for the CFD simulation [128]. The weather research and forecasting (WRF) model could provide mesoscale velocity data with a resolution of 1 km and be converted into the format of a CFD inlet condition [129]. In fact, the inlet condition is of high complexity, and has to be considered together with wall functions (WFs). Usually, z₀ follows the updated Davenport terrain classification [130], and the u_* could be calculated from the given velocity at the given height. Studies have found that the coupled WRF-UCM performs better than conventional WRF in providing local-scale wind data [131,132].

4.5. Optimization Towards Application

The influencing factors of building-installed SWTs are relatively clear. A lot of studies discussed the effects of rooftop shape/angle [116,133], installation positions [53,134], and building layout/urban morphology [135]. Sadly, there is still a gap in practical application. Single building and idealized building arrays are repetitively investigated, the findings of which helped only local improvements such as rooftop shapes. Practical studies on a large scale and for real cities are, in contrast, seldom studied. A possible reason for this is the conflict between universality and particularity, in that many researchers might see a city-specific study as a multi-parameter problem, and its findings are too synthetic and case-specific to provide any insights for other scenarios.

Hence, this review sees two potential approaches towards a deeper understanding of factors affecting urban wind pattern and the distribution of urban wind resources. For idealized studies, multi-parameter optimization studies could be considered. These parameters include building geometry (height, height-width ratio, rooftop shape, etc.), turbine characteristics (turbine type, swept area, power curve, cut-in/out velocity, etc.), and conditions (wind rise, wind velocity distribution, upstream terrain roughness, etc.). The targets of this optimization include power, annual energy yield, and the lifespan of the SWT. In such studies, the computational cost for CFD simulation should be relatively low, and hundreds of cases could be performed with varying parameters, resulting in a quantitative and continuous understanding of what affects the SWT performance.

The other approach is to assess the wind resource of real cities and provide useful insight to the industry or policymakers, but that could raise the concern of high computational demand, as the largest scale for a pure CFD-based wind resource assessment is documented up to only 10 km [50] and cannot cover a modern megacity. Thus, it is promising that the CFD simulation could be combined with parameter analysis to realize the rapid reconstruction of the distribution of urban wind resources, as multiple studies have attempted for urban ventilation and wind comfort [136,137,138]. Data-driven methods could also be utilized to accelerate the reconstruction and optimization of urban flow based on CFD data [139]. Therefore, future studies on large-scale real cities will certainly deepen the understanding on urban flow and the application of urban SWTs.

5. Conclusions

This review systematically summarized the general process of a CFD-aided urban wind resource assessment for urban wind harnessing studies. Firstly, the conditions for urban wind resource assessment were addressed, including the extraction of boundary conditions and the performance of SWTs. Secondly, the commonly used CFD models are analyzed and provided with key points in configuration. It was found that LES, as a high-fidelity model, is still less used than the RANS model, which could be attributed to a lack of BPGs. However, LES-based hybrid models have not yet been used in urban wind resource assessment, and could offer a compromise as they balance the computational cost and accuracy. Thirdly, some challenges and corresponding prospects for future studies are discussed. These challenges include the fidelity problem of turbulence models and geometry simplification, the inclusion of wake models and extreme conditions, and larger scale application for real urban environments. At the same time, recent developments in multi-scale modeling and data-driven approaches show good potential for application, as they can be used in real city case modeling and accelerate the flow reconstruction and optimization. This review provides the technical basis for the CFD-aided urban wind simulation that can help to locate the starting point of such studies, and promote the application of state-of-the-art techniques in urban wind resource assessments.