This article describes a study in which modellers were challenged to
compute the wind field at a forested site with moderately complex
topography. The task was to model the wind field in stationary conditions
with neutral stratification by using the wind velocity measured at 100 m at
a metmast as the only reference. Detailed maps of terrain elevation and forest
densities were provided as the only inputs, derived from airborne laser scans
(ALSs) with a resolution of 10 m

To respond to the increasing demand for wind power, new areas for wind
turbine sites are being explored. Large offshore farms further away from
the shore are being developed, as are wind farms in more complex onshore
areas, such as terrain with a more varied topography and roughness. This
is the case in northern countries, such as the Scandinavian region, where
large remote forested areas are being explored for wind development.
However, when exploring these complex sites it is evident that new challenges
arise due to comparatively higher turbulence levels and wind shear. While
the magnitude of wind shear and turbulence increase the fatigue load,
uncertainties in the estimation of wind shear and turbulence have shown to be
problematic in forested areas

In addition to the actual difference in wind climate between traditional wind energy sites and complex forested ones, modelling of the wind conditions is challenging. Trees are elevated sources for both momentum absorption and heat transfer, and thus they differ from traditional surfaces since the exchange may be distributed at several model levels. The degree of physical description is a choice by the modeller, from describing plant area densities (PADs) in each grid cell to representing an entire forest with a single roughness length value. The required numerical demand, however, varies with many orders of magnitude when making that choice.

To the knowledge of the authors, no large-scale studies have been published
comparing different micro-scale models over forested terrain with high-quality meteorological data. However,

The progress of forest flow modelling now enables the direct simulation of the
tree densities. Such values as PAD may be derived from airborne laser scans
(ALSs) that are becoming increasingly available from national mapping services

The study started with a call for a benchmarking model validation study to
modellers involved in the European ERANET+ project New European Wind Atlas
(NEWA). The aim of the benchmark is to illustrate how well micro-scale models
are able to simulate winds above a forest in moderately complex topography.
The participating models range from industrial wind models to front-line
research models. The modelled case consists of a typical site located in
Ryningsnäs in southern Sweden, i.e. a patchy forested site with moderately
complex topography

The NEWA project includes several large-scale field campaigns designed for
flow model validation

The paper begins by outlining the following: the benchmark, the validation data, and general modelling. This is followed by a description of first the RANS models and then the LES models. It then continues with the main results and finally concludes with a “Discussion and conclusions” section.

The benchmark task was to model the wind profile at the location

The choice behind having a target wind speed at 100 m of height, rather than having a fixed geostrophic wind speed, was partly due to the lack of measurements of the geostrophic wind speed and partly due to the desire to have as similar a wind speed as possible in the lower part of the boundary layer. This relates to the question of whether or not the models can accurately predict the flow footprint given that the ALS data enable the model to have surface conditions very similar to reality. Setting a fixed geostrophic wind speed would risk the modelled wind speed in the surface layer being lower or higher than the measured, with a subsequent uncertainty of the ability to capture the footprint of the flow. In a strictly neutral boundary layer, the ratio of the turbulence level and the wind speed is expected to be constant, and hence the footprint would be the same for different wind speeds. However, the boundary layer height changes with wind speed, as do the gradients of velocity and turbulence level. Thus, in addition to scaling the wind speed with the friction velocity, one would have to scale the height with the boundary layer height in order to make a fair comparison, which cannot be done since the boundary layer height was not measured.

In order to characterize the forest ALS data from the Swedish map authority,
Lantmäteriet has been utilized

The measurement site is located in Ryningsnäs in south-east Sweden
(

The tower is situated in the north-west corner of a 400 m

Two wind turbines are situated approximately 200 m to the north-east and south of the tower, respectively, but the three sectors used in the validation study exclude directions which these turbines would influence.

The full measurement set-up and the wind climate has been
reported earlier in

The sonic anemometers were sampled at 20 Hz and statistics was evaluated by
30 min block averaging and 3-D rotation of the coordinate system, aligning it
with the local mean wind direction and yielding the wind vector

To select only neutral conditions

a quality check that was passed;

neutral stratification;

stationary flow;

wind direction within the target sector.

Overview of the models used and the model family.

The models that participated in the benchmark were all CFD models using a
Reynolds-averaged Navier–Stokes (RANS) or large-eddy simulation (LES)
methodology. Table

Computational modelling of the fluid flow employs a filtered version of the
Navier–Stokes equations due to the impracticality of resolving every
temporal and spatial scale. The Reynolds-averaged Navier–Stokes (RANS)
equations make use of the Reynolds decomposition to divide the velocity field
into the time-averaged velocity and the velocity fluctuation around the mean,

RANS modelling supposes that the effect of all ranges of fluctuations on the
mean flow can be accounted for by the models. Conversely, in the LES approach
the energy-containing flow structures are fully resolved, whereas only the
effect of the smaller fluctuations is modelled. This is achieved through the
decomposition of the velocity field into filtered (or resolved) and residual
(or subgrid-scale, SGS) components,

The participating RANS models use one- or two-equation turbulence models,
presented in general form in this section. The specific set-up of each model
is presented in the following sections. The two-equation turbulence-closure
model corresponds to the classical

The length scale in the standard

The two-equation methodology explained above is used by Ellipsys3D, CFDWind,
and Alya. In the case of Meteodyn, a one-equation RANS turbulence model

Furthermore, assuming the canopy elements exert a drag force on the flow,
the effects of the plant drag inside the canopy on the main flow are
parameterized, presuming the form drag dominance in the momentum
Eq. (

Meteodyn WT is a commercial site-assessment software that models the surface
boundary layer (no Coriolis force included) using RANS, in particular a
one-equation

The forest model in Meteodyn is based on a mean flow model, which treats the
forest as a porous media

Summary of the model constants used.

In Meteodyn the canopy height

Two versions of the forest model are available in Meteodyn, which differ in the computation of the mixing length in the one-equation turbulence model, called robust and dissipative in the Meteodyn documentation. The dissipative forest model is used in the Meteodyn simulations presented in this investigation, in which a 15 m extra high dissipation zone is used above the forest.

The computations are performed employing a Cartesian structured mesh on a
square domain with the dimensions

Monin–Obukhov inlet profiles for velocity are defined at the inlet of the
domain, as is a constant turbulent kinetic energy

EllipSys3D is a CFD solver designed for various wind engineering applications – e.g. atmospheric boundary layer flows, turbine rotor computations – it is a multi-block finite-volume solver of the incompressible Navier–Stokes equations in the general curvilinear coordinates. It uses collocated variable arrangement, employing the revised Rhie–Chow interpolation technique in order to avoid the odd–even pressure coupling. The pressure–velocity coupling in the present study was based on the SIMPLE algorithm. Furthermore, the code is designed based on a non-overlapping domain decomposition technique, which combined with its MPI parallelization enables it to highly efficiently run on distributed and/or shared-memory high-performance computation (HPC) systems.

The standard and modified model constants according to
Table

To model the effects of surface roughness on the mean flow and
avoid resolving the laminar sub-layer, wall functions as boundary conditions
at wall surface boundaries are typically applied. In EllipSys3D, the wall
boundary is placed on the top of the roughness elements; this allows large
near-surface velocity gradients to be resolved using shallow (high-aspect-ratio)
computational cells. The wall shear stress is accordingly used to
specify the wall boundary conditions for momentum and

The computational domain is a circular grid with a radius of 17 km, centred
at the Ryningsnäs metmast location. The inner zone surrounding the site
has a quadratic form. It is based on equally spaced grid points and covers a
zone of 5 km

Alya is an HPC code developed at the Barcelona Supercomputing Centre (BSC) to
run large-scale applications. The code was recently tested on 100 000
processors and showed a parallel efficiency above 90 %

The

The Navier–Stokes equation (Eq.

A robust finite-element scheme written in block-triangular form

Once the algebraical system of equations is obtained, a deflated conjugate
gradient

The Ryningsnäs problem was solved using a cylindrical mesh with a radius
of 20 km. The mesh is centred on the metmast. Surrounding the metmast the
mesh resolution is

The inflow boundary conditions are defined from a precursor simulation over flat and homogeneous terrain (i.e. single-column model 1-D). The obtained fields are used also as initial conditions. Zero traction is imposed over the outflow boundaries. No velocity penetration and zero tangential stress are imposed over the top boundary.

Three different geostrophic velocities were set to the three different wind
directions to match the desired velocity at mast. The geostrophic velocities
were set to

CFDWind is a modelling framework developed at CENER on top of the open-source
CFD platform OpenFOAM

As only neutral atmospheric stability was considered, the flow is assumed stationary so the SIMPLE algorithm is employed to solve the pressure–velocity coupling, while second-order upwind schemes are used for the discretization of both velocity and turbulence convective terms.

The Coriolis apparent force is added explicitly to the momentum equation together with the horizontal pressure gradient that drives the system, which is derived from the hydrostatic relation for stationary cases.

The perturbations induced by forests are modelled by adding drag and
source–sink terms in the momentum and turbulence-closure equations,
respectively, as proposed by

Despite the fact that it is expected that wind flow will be dominated by the effects
of forest features near the surface,

Similar to EllipSys3D, the wall functions consider that the computational
grid is placed on top of the roughness elements so that restrictions related
to the height of the cells adjacent to the ground and

The numerical grid was created with the meshing software WindMesh. The tool
has been developed jointly by BSC and CENER for the automatic and fast
generation of grids over terrain. There are currently two different versions
further developed by each institution: BSC-WindMesh

CENER-WindMesh creates structured terrain-following grids optimizing parameters such as orthogonality and skewness by applying filters to the 2-D (ground) mesh and elliptic smoothing techniques for the final 3-D mesh. The mesh is designed so that terrain is smoothed far from the area of interest, whereas towards the central zone the cells are refined to the maximum resolution established. Only real topography is considered for the grid generation in the centre. The “transition” zone between boundaries and the central zone is a progressive blend between the real terrain and flat boundaries.

Similar to previous approaches, a precursor run is conducted prior to the
full-terrain simulation (successor) in order to create the equilibrium
profiles that serve as inlet conditions. Precursor simulations are conducted
on flat domains with periodic boundary conditions on the sides with the top
and wall treatment mentioned above. The PAD is set to a constant value of
19 m

The value of the geostrophic wind is chosen so that the velocity magnitude
obtained in the simulations is approximately

The computational domain is square-shaped and covers an extension of
18 km

The computations by UUCG were carried out with a solver implementation based
on the OpenFOAM platform, version 3.0.1. A neutrally stable wind flow is
computed with LES coupled with an SGS model

It is assumed that the forest acts as a porous surface exerting a drag on the
flow. This is represented in the simulation with the introduction of a source
term in the LES momentum equation (Eq.

A wall model is also used to account for the roughness of the ground,
although it is expected that its influence on the wind flow will be much
smaller in comparison to the forest. For this, the wall model implementation
available in the OpenFOAM libraries of SOWFA

The computational domain consists of a box with the dimensions
32 km

The longitudinal axis of the domain is aligned with the wind direction for
each case, so the inlet is perpendicular to the inflow. All lateral
boundaries are set to periodic boundary conditions. Hence, the inlet flow is
recycled from the outlet. The flow is driven by a uniform pressure gradient
following the procedure described by

PALM is a massively parallelized LES solver designed for studies of the
atmospheric and oceanic boundary layer. It is an open-source code

The forest effect is modelled by adding a sink term to the momentum equation
following

Wind speed profiles.The blue dashed line shows the average from the
cups, and the red full line shows the average from the sonics. Error bars
indicate the 95 % confidence level for the mean value. The various other
markers indicate simulated wind speeds. Please see
Table

The benchmark simulations use a model domain of
2304 m

To summarize, four different RANS codes and two different LES codes are
included in the study. Forest modelling is basically done in the same way in
all codes apart from Meteodyn. All models apart from PALM use heterogeneous
forest, but Meteodyn is based on a different surface data set. Domain sizes
are similar apart from PALM, which uses a significantly smaller domain, but
since PALM has homogeneous forest with recirculation the domain size is
directly comparable. A summary of some key modelling properties is found in
Table

Numerical set-up. Cell size refers to the horizontal grid size in the inner domain.

One main purpose of RANS and LES modelling within the wind energy community is to extrapolate tower measurements vertically and spatially. In the next section, the vertical extrapolation (vertical profiles) is reported first, followed by the horizontal extrapolation (planes).

Wind speed, wind veer, and turbulence are crucial to power production. These
three quantities, in the form of mean wind speed

Modelled and measured profiles are shown for the three different wind
directions described in Sect.

Modelled and observed wind shear. The lower three plots are zoomed
in on the instrument heights to increase readability. The blue dashed line
shows the average from the cups, and the red full line shows the average from the
sonics. Error bars indicate the 95 % confidence level for the mean value.
The various other markers indicate simulated wind speeds. Please see
Table

Wind direction profiles. The red full line shows the average from
the sonics. Error bars indicate the 95 % confidence level for the
difference between the wind direction at each height and direction at 40 m.
The various other markers indicate simulated wind directions. Please see
Table

TKE profiles. The red full line shows the average from the sonics.
Error bars indicate the 95 % confidence level for the mean value. The
various other markers indicate simulated TKE levels. Please see
Table

Most models overestimate the wind speed gradient, as reported in
Fig.

In an earlier publication

Of the three directions, 290

The purple lines in Fig.

In order to evaluate spatial differences the modellers were instructed to
submit horizontal planes surrounding the measurement tower. Planes are shown
here for 40 m above the local elevation in Fig.

Although the models all show similar wind speed patterns there is a difference in the amount of wind speed streaks present and in the strength of the streaks. All the models show more intense streaks at higher height. The LES models show more tendency for streaks than the RANS models. EllipSys3D shows almost no streaks, whereas Alya and OpenFoam CFDWind have similar streak patterns as LES OpenFoam UUCG. The streaks correlate with topographical features, but there are also clear streaks in the PALM LES, which ran without topography.

The main aim of the study has been to investigate the state-of-the-art abilities of modelling groups to replicate measurements in neutral conditions over a forested site. Six modelling groups in total completed the whole process and submitted results. The RANS modellers using research codes used a fairly homogeneous approach to the model task, while the LES modellers took quite different approaches. Overall, a variety of options were used and in this section we will try to discuss some of the implications of these different choices.

All models except Meteodyn and PALM use ALS input for topography and forest
data. The fact that a variety of models (including LES and RANS) were able to
use the ALS input was considered a success. The use of PAD data from ALS
removes the uncertainty of having to guess the PAD or the roughness length
and displacement height, which in practice can be a large source of
uncertainty when estimating the wind resource at a potential site. The only
model not to use the ALS was Meteodyn (instead using a constant PAD of roughly 0.025 m

An interesting observation noted by several of the participants was that the roughness was totally dominated by the drag of the forest and that the value of the ground roughness did not make a visible impact on the results. That could, on the other hand, be expected, since even though the forest characteristic is heterogeneous, the landscape as a whole can be considered forest covered.

An initial question at the start of the study was whether the differences in the measured profiles among the three directions would be captured by the models, given the detailed surface data. In summary, the differences between the directions turned out to be small for the RANS models, but the LES model that used the detailed surface data produced profiles that resembled the measured profiles.

From Fig.

The LES version of OpenFoam furthermore showed a much more pronounced difference between the inflow angles both in terms of shear and TKE; a possible explanation may be that the RANS models are over-diffusive, something also indicated by the fact that RANS models show fewer streaks in the horizontal planes.

One of the most striking outcomes of the study is that the

As seen in Table

Following the reasoning in

Another computational expense in LES modelling is the integration time. This particular study was aimed at simulating a stationary case, and since the Coriolis force may introduce inertial oscillations it is important to make sure that the influence of oscillations does not impact the results. Another important conclusion of the study is that stationary, neutral conditions are very rare in the atmosphere, and hence future studies should investigate naturally occurring transient conditions such as diurnal cycles, evening transitions, and developing unstable turbulence.

The orders of magnitude difference in numerical challenge, both in the set-up and in the computational time used, should be considered in terms of the accuracy of the modelling results.

Many modellers expressed the difficulties involved in trying to determine the
correct value of the

Based on the problem of adjusting forcing in order to match a target wind speed, measurement campaigns designed for flow model validation should attempt to measure the boundary conditions and forcing, such as the boundary layer height, vertical and horizontal fluxes, radiation, ground temperature, and geostrophic wind speed. Another option is to use a mesoscale model to compute the boundary forcing for the micro-scale models, but then care has to be taken that no additional uncertainty is introduced due to bias between the mesoscale model results and reality.

Future micro-scale model comparisons at complex forested sites should focus on the modelling of thermal stability effects. The radiative cooling during cold nights strongly affects the wind profile, introducing strong variation of heat flux in the forest. Heat transfer models over forested sites have been already implemented over flat terrain. However, new thermal models need to be developed and validated that account for complex terrain. It is also clear that the given thermally stratified conditions are more difficult to compare to stationary conditions, both due to model drift and to the fact that the atmosphere itself is mostly in transition. Comparing natural cycles would also remove the uncertainty of integration time since the physical time modelled would be the same for all participants. The representativeness of the results would also increase significantly if thermally stratified conditions were included since the strictly neutral conditions used in this study are rare in the atmosphere. Such a study would of course necessitate the use and development of unsteady RANS models.

The data used to validate the models (selected as described
in Sect.

SI coordinated the project. JA selected the data, was responsible for the ALS to PAD conversion, gathered the results, and made the plots. MA ran Alya, DC ran EllipSys3D, RAC-A ran CFDWind, HO-E ran UUCGWind, CP and JA ran Meteodyn, and BW ran PALM. All authors contributed to writing the text.

The authors declare that they have no conflict of interest.

The work is mainly performed within the ERANET+ project NEWA. This work was partly conducted within StandUp for Wind, a part of the StandUp for Energy strategic research framework in Sweden. The simulations were partly performed on resources provided by the Swedish National Infrastructure for Computing (SNIC) within the project SNIC 2017/11-10. Ebba Dellwik is greatly acknowledged for contributions regarding benchmark design, converting the ALS to PAD, and measurement data selection. Edited by: Luciano Castillo Reviewed by: two anonymous referees