Content uploaded by Rémi Gandoin
Author content
All content in this area was uploaded by Rémi Gandoin on Jan 13, 2016
Content may be subject to copyright.
Evaluation of two novel wake models in offshore wind farms
Jorge Garza(*)(1), Andreas Blatt(1), Rémi Gandoin(1) and Shiu-Yeung Hui(1)
(1) DONG Energy Renewables. Nesa Allé 1, 2820 Gentofte, Denmark
(*) Corresponding author, jorca@dongenergy.dk
Summary
Two novel CFD based wake models, Fuga
by Risø and WindModeller by ANSYS,
along with a DONG Energy implementation
of the N.O. Jensen model, are tested
against quality filtered measurements from
4 offshore wind farms. The comparison is
done for rows of turbines and overall wake
losses.
WindModeller was found to have the best
agreement with the measurements in
narrow direction sectors, while the accuracy
of Fuga and the N.O. Jensen model
improved as the sector was widened to 30°.
For the only case where all three models
were tested in a 30° sector with 1°
resolution, WindModeller and Fuga reached
the same mean absolute error, which was
lower than that for the N.O. Jensen model.
The assessment of the model accuracy in
the estimation of overall wake losses was
only done for two wind farms. The N.O.
Jensen model was found to perform best
regarding the mean signed error, balancing
out over- and underestimations of the wake
losses. All models resulted in errors lower
than 3%.
1. Introduction
With the increasing size and scale for
offshore wind farms and wind farm clusters,
the accurate assessment of power losses
due to wake effects is becoming an
increasingly vital part of overall project
economics. The traditionally used wake
models are simple and to some extent
unphysical, and their operational frame
could be challenged in large wind farm
clusters.
This paper presents a comparison of two
wake models, Fuga [1] and WindModeller
[2], recently developed by Risø and ANSYS
respectively, as a result of the Carbon Trust
Offshore Wind Accelerator (OWA), along
with the widely used N.O. Jensen model
[3,4]. The benchmarking is performed with
two different approaches. The first is the
assessment of losses in wind direction
cases aligned with rows of turbines, which
has been widely studied in the literature.
The second approach combines results
from all directions similarly to the prevalent
commercially available wind farm simulation
software, resulting in estimates of the
overall park wake losses. The model results
are then validated against measurements
from a vast database comprising 4 offshore
wind farms in different areas. The
combination of the two different approaches
builds a bridge between model accuracy in
very specific settings, and its robustness
and suitability for integration into statistical
estimation tools.
This benchmarking study outlines the
current state-of-the-art of wake modelling
by assessing the capabilities of the two new
models using two types of analysis: single
direction and overall park losses. The
combined understanding of these aspects
is crucial for accurately judging which
model performs best. These results
potentially reduce the uncertainty of layout
design.
2. Methodology
The assessment of wake losses has
traditionally been presented as the turbine
efficiency profile along an aligned row of
turbines for a given wind speed and
direction, normalized by the production of
the turbine in the free stream [5,6]. This
stresses the critical decrease in power
production from the first to the second
turbine in a row for aligned turbines, and is
useful to gain insight into the single wake
interaction. It is, however, not directly
translatable to an indication of the overall
wind farm efficiency. This can be done by
extending the analysis in a way analogous
to the traditional methods used for wind
farm yield estimates. This section will go
through a description of the models, the
measurements that are used to evaluate
them and the method in which this is done.
2.1 Models
The comparison is performed for the full
CFD code WindModeller from ANSYS and
linearised CFD model Fuga from Risø, two
novel wake models whose development
has been accelerated by the OWA
programme. Both models are compared to
an internal DONG Energy implementation
of the N.O. Jensen model which is
prevalent in the industry.
2.1.1 N. O. Jensen
The N. O. Jensen single-wake model was
first presented by Jensen [3] and Kátic [4].
For a given rotor diameter and a free
stream wind speed, it predicts the wind
speed deficit as a function of the distance
downstream of the turbine. The wake
expansion is controlled by the wake decay
constant k. For offshore cases, k is
commonly set to 0.04. The magnitude of
the wind speed deficit depends on the
thrust coefficient. The total wake deficit for
a given turbine location is calculated as the
square root of the sum of squares of the
wake deficit induced by all upstream
turbines.
The N. O. Jensen model is implemented as
an in-house tool, which allows the user to
create results for the entire wind rose, with
a resolution of 1° and at a fixed wind speed.
The typical computation time is around 10
seconds.
2.1.2 WindModeller
WindModeller (release 28/06/2011) is a
series of pre- and post- processing scripts
for Ansys CFX. It uses an actuator disc to
model the wake with a momentum sink
(constant thrust per volume at the rotor disk
location), see [2]. It does not take
atmospheric stability into account.
WindModeller can also be used with
varying roughness and orography for
onshore cases.
The computational domain is a cylinder
centred on the wind farm. It has a diameter
of the order of 20km, and a height of 1km.
The grid is refined around the wind farm
area, to a minimal resolution of 30m
horizontally, and 2m vertically. The mesh
around the rotor is adapted during the
solver stage with the actuator disk
parameters. The mesh generation in
WindModeller implies the advantage of
having only one initial mesh, regardless of
the wind direction simulated.
A steady state RANS approach using k-ε
turbulence model is chosen. The constant
Cµ (ratio of surface shear stress and
turbulence kinetic energy) is set to 0.03 in
order to match the upstream turbulence
conditions (upstream turbulence intensity
set to 0.06). The turbulence model setup is
in accordance with the setup used in [2],
considering a turbulence decay exponent of
0.6. A constant roughness length of z0 =
0.0001m is used.
WindModeller operates with CFX (Ansys
13) on a 48GB RAM machine with two
CPUs of 2.3GHz 6-cores each. The
maximum number of iterations is set to 200,
and the residual target to 1e-5. Indicative
computational time and mesh sizes are
given in Table 1.
Wind
Farm
Initial mesh
nodes [106]
Final mesh
nodes [106]
CPU
time (*)
[min]
HR1
1.4
3.43
56
Burbo Bank
0.98
2.12
26
Barrow
1.42
3.09
40
GFS
1.48
3.38
57
Table 1: Initial and final mesh sizes and
indicative computational time for the four
wind farms. (*) for 8m/s.
2.1.3 Fuga
The linearized CFD code Fuga uses the
wind farm layout and wind turbine
parameters in WAsP formats as inputs. It is
designed for sites with homogeneous
terrain and roughness.
The model solves the time independent
RANS equations with a simple closure
scheme for turbulence and a mixed spectral
solver using pre-calculated look-up tables
[1]. The model neglects the Coriolis force
and buoyancy effects and, due to the latter,
it can only model neutral atmospheric
stratification. An actuator disk is used to
model the wake and the flow is lid-driven,
with the height of the upper boundary zi,
where the velocity is fixed, being one of the
user inputs.
Preliminary-look-up tables (pre-LUTs) are
created once during installation of the
software, and LUTs are created for each
case (turbine + z0 + zi combination). This
results in a very fast computation of the
results, in the order of 1 minute per case.
Fuga version 1.51 is used for this paper. All
simulations were run with a value of
zi=500m and a surface roughness length of
z0=0.0001m. The extents of the maps used
to create the wind farm files were similar to
those used for WindModeller simulations.
2.2 Measurements
SCADA data from 4 offshore wind farms
are available for the analysis, in the form of
10-minute averaged values. The wind
farms under consideration are Horns Rev 1,
Gunfleet Sands, Burbo Bank and Barrow.
Their locations are shown in Figure 1.
The vast database offers a wide span of
test cases, both regarding the turbine type,
turbine spacing and thrust coefficient
values, which are the main drivers behind
wake loss calculation. A big variety
regarding the symmetry of the layouts can
also be found. While all layouts are regular
arrays, two are symmetric and two are
asymmetric. This adds another interesting
dimension by investigating the effect of
layout symmetry on model accuracy. All this
gives a large set of results for turbine rows,
from which only selected cases are shown
here.
For transect analysis, in order to isolate
wake effects from other possible power
losses, measurements are filtered for ideal
operating conditions in a manner inspired
by [7]. This is achieved by removing
records where turbine availability within the
Figure 1: Location of the wind farms used in
the analysis.
10-minute period was not perfect, along
with records in which the turbine controller
imposed a limitation to the turbine power
output. Due to the rounding of the data in
the database, perfect availability is defined
as 95% or more of the 10-minute period
operating correctly, and not exclusively
100%. Furthermore, for every individual
case, only records with the specific wind
speed and direction are used.
For overall wake losses calculations, the
data quality filter greatly reduces the
amount of available data, since all turbines
are now required to be operating correctly
at the same time. In order to maximise the
amount of data available, the filter criteria is
modified and a record is deemed
acceptable if more than 95% of the turbines
were operating correctly. This means 77 out
of 80 turbines in Horns Rev 1 and 46 out of
48 in Gunfleet Sands. This was found to
increase the amount of available data for
the analysis by a factor of 6 to 8, and its
effect on the general trend of the results
was small. This relaxation of the filter
criteria can add uncertainties to the
analysis, since it is assumed that the
missing turbines are distributed uniformly,
while it might not necessarily be the case.
This uncertainty was accepted due to the
great benefit of a drastic increase in the
data sample size.
Since the models produce one result per
wind direction and the measurements can
provide many data points for each direction,
with population distributions that are not
necessarily uniform, the data averaging
applied to the measurements is not trivial. A
simple average would assign the same
weight to every single data point, which
could in practice mean that the observed
wake losses at X° wind direction are given
a higher weight than those at Y°, depending
on the number of observations at each
direction; while averaging of model results
gives the same weight to all directions. In
order to avoid this unequal weight, a
smoothing method is applied by first
averaging all measurements at every 1°
wind direction and then averaging the
smoothed 1° results falling in the sector in
turn.
2.3 Method
The SCADA data is quality-filtered by
removing the effects of availability and
power curtailment, in order to isolate wake
losses. The wind direction is obtained from
the yaw alignment signals of all available
turbines, a procedure that is verified with
met mast data when available. A set of
reference, free stream turbines is defined
according to the wind direction. Finally, the
free stream wind speed is derived from the
power production at the reference along
with the density corrected power curve. The
method for correcting power curves for air
density is verified with density-specific
power curves from several manufacturers.
For the analysis of a row of turbines, the
power from the turbines in the row is
normalized by that of the reference turbine
to create efficiency profiles. For wind farm
efficiency, the power from all the wind farm
is divided by the product of the number of
turbines and the reference power. Then the
efficiency of the wind farm at a given wind
speed is calculated for all directions,
grouped in 30° sectors or other angular
resolution. The contributions from each
sector are then scaled according to a wind
rose and their sum results in the wind farm
efficiency at that wind speed.
For this process to be accurate, the sample
size for all wind speed - direction
combinations should be sufficiently large.
This data should be representative of the
general climatology of the site including
atmospheric stability, a factor that is not
explicitly included in the calculations but is
known to have an effect on wake losses
[8,9].
The models are used as recommended by
the software developers, and part of the
analysis focuses on the reproducibility of
previously published results, in order to
assess how dependent are the models to
the user.
Results from Fuga and N.O. Jensen
models are produced for 1° bins and then
averaged linearly in the different 30° sector
sizes. This places WindModeller on a
different ground from the others, since it is
computationally expensive and impractical
to run simulations for the whole wind rose
at 1° resolution. Therefore, in the cases of
rows of turbines, WindModeller results are
referred to the average of results every 1°
in a certain directional sector, while in the
overall wake losses only the results for the
sector centre angle are taken into account.
This work does not aim at testing the
models on all speed – direction
combinations, therefore the total wind farm
wake losses are not considered.
3. Results
3.1 Rows of turbines
3.1.1 Horns Rev 1
Horns Rev 1 provides here the longest
datasets. It has been extensively used for
previous studies [5,7,8,9], and is
investigated here as a starting point.
Figure 2 shows the results for a row of 10
turbines with 7 rotor diameter (RD) spacing
and 10m/s winds from 270°. They are
shown for specific direction, 10° and 30°
wide sectors (centre angle ± 5° and ± 15°,
respectively). WindModeller is the closest to
the measurements in the 1° and 10° sector
cases. The accuracy of Fuga and N.O.
Jensen improves as the direction sector is
widened to 30°, where both Fuga and
WindModeller have excellent predictions.
3.1.2 Other wind farms
The same analysis was performed for rows
of turbines in other wind farms which have
not been previously used for this kind of
studies. Namely Gunfleet Sands in the
North Sea; Burbo Bank and Barrow in the
Irish sea.
Figure 2: Results for wake losses along a row of turbines in Horns Rev 1 for winds from 270° ± a)
1°, b) 5° and c) 15°. 10 WTG/ and 7D spacing
Figure 3: Results for wake losses along a row of turbines in Gunfleet Sands for winds from 239° ±
a) 5° and b) 15°. 9WTG and 8.4D spacing
T02 T12 T22 T32 T42 T52 T62 T72 T82 T92
0
0.2
0.4
0.6
0.8
1
POW [-]
270deg - 10m/s
Measurements
FUGA
N.O. Jensen
WindModeller
01deg bin
T02 T12 T22 T32 T42 T52 T62 T72 T82 T92
0
0.2
0.4
0.6
0.8
1
POW [-]
270deg - 10m/s
Measurements
FUGA
N.O. Jensen
WindModeller
10deg bin
T02 T12 T22 T32 T42 T52 T62 T72 T82 T92
0
0.2
0.4
0.6
0.8
1
POW [-]
270deg - 10m/s
Measurements
FUGA
N.O. Jensen
WindModeller
30deg bin
G01 G02 G03 G04 G05 G06 G07 G08 G09
0
0.2
0.4
0.6
0.8
1
POW [-]
239deg - 08m/s
Measurements
FUGA
N.O. Jensen
WindModeller
10deg bin
G01 G02 G03 G04 G05 G06 G07 G08 G09
0
0.2
0.4
0.6
0.8
1
POW [-]
239deg - 08m/s
Measurements
FUGA
N.O. Jensen
30deg bin
Figure 4: Results for wake losses along a row of turbines in Burbo Bank for winds from 313° ± a) 5°
and b) 15°. 8WTG, 5.3D spacing
Figure 5: Results for wake losses along a row of turbines in Barrow for winds from 312° ± a) 5° and
b) 15°. 8WTG, 5.5D spacing
T38 T37 T36 T35 T34 T33 T32 T31
0
0.2
0.4
0.6
0.8
1
POW [-]
313deg - 08m/s
Measurements
FUGA
N.O. Jensen
30deg bin
T38 T37 T36 T35 T34 T33 T32 T31
0
0.2
0.4
0.6
0.8
1
POW [-]
313deg - 08m/s
Measurements
FUGA
N.O. Jensen
WindModeller
10deg bin
B8 B7 B6 B5 B4 B3 B2 B1
0
0.2
0.4
0.6
0.8
1
POW [-]
312deg - 08m/s
Measurements
FUGA
N.O. Jensen
WindModeller
10deg bin
B8 B7 B6 B5 B4 B3 B2 B1
0
0.2
0.4
0.6
0.8
1
POW [-]
312deg - 08m/s
Measurements
FUGA
N.O. Jensen
30deg bin
Results for a test case at 8m/s along the
longest row in each wind farm are shown in
Figure 3 for Gunfleet Sands (9 turbines with
8.4RD spacing), Figure 4 for Burbo Bank (8
turbines with 5.3RD spacing) and Figure 5
for Barrow (8 turbines with 5.5RD spacing).
The result for the specific 1° directional bins
has been omitted, since its practical value
is limited when considering the
uncertainties in measured wind direction
within a 10 minute period. It was not
possible to produce runs every 1° for
WindModeller in all cases due to limited
time and computational resources; and
these plots therefore do not show results for
this model in 30° sectors.
The same general trend as for Horns Rev 1
is observed, with WindModeller having the
best accuracy for smaller sector widths.
Fuga and N.O. Jensen improve their
accuracies as the sector widens to 30°.
For Gunfleet Sands, WindModeller is very
accurate at 10°, slightly overpredicting
wake losses; while the other two models
overpredict the losses by a wider margin.
When switching to a 30° sector, N.O.
Jensen captures the behaviour at the first
three waked turbines and then
progressively underpredicts losses, while
Fuga underpredicts all along but gets the
last turbine correct.
For the cases of Burbo Bank and Barrow,
all models were inaccurate in the 10° case
and all overpredict the wake losses.
WindModeller could capture the losses at
the first waked turbine, but not the
development of the loses along the row.
Fuga and N.O. Jensen improved when
averaging over 30°; with N.O. Jensen
performing best in the Burbo Bank test case
and both performing equally well in Barrow.
Fuga was found to underpredict the losses
in both cases.
3.1.3 Rows of turbines – General
It was observed that in general terms, Fuga
and N.O. Jensen accuracy increases
notably with increasing directional sector
bin size; while WindModeller accuracy
improves as well, but in a much less
dramatic way.
In the Horns Rev 1 case, where the full
span of 0° to 30° averaging sector width
could be tested, both new models were
found to reach comparable levels of
accuracy for 30° sectors, achieving mean
absolute error values lower than 1%. The
relationship between mean absolute error
and averaging sector width for the Horns
Rev 1 case is shown in Figure 6.
Figure 6: Evolution of mean absolute error
for all three models as a function of the width
of the direction sector in Horns Rev 1 with
10m/s winds from 270°.
3.2 Overall wake losses
Wake losses along a perfectly aligned row
of turbines has been studied extensively in
wake modelling research. For its application
in energy yield estimates, the most
important is the accuracy of the wake
model when integrated with wind statistics
binned in (typically) 30° sectors and
weighted with the corresponding wind
speed and wind direction frequency. The
estimation of wake model accuracy when
predicting overall wind farm wake losses is
part of the bridge between the in-row wake
loss analysis and its application in industry
standard procedures.
The smoothing technique of averaging the
measurements in every 1° and then
averaging all 1° results in the sector was
not applied in this case, since it would
drastically reduce the amount of data
available for the analysis as well as adding
uncertainties as a result. The metric for
comparing the models is the mean signed
error. This is considered the most sensitive
approach since in the case of a direct
application to yield assessment, the fact
0 5 10 15 20 25 30
0
0.05
0.1
0.15
0.2
0.25
Average Absolute Errror [-]
Sector Width [deg]
FUGA
N.O. Jensen
WindModeller
that positive errors can cancel out negative
errors is relevant.
Figure 7 shows the results for wind farm
efficiency as a function of wind direction
every 10°, predicted by all three models
and from measurements at Horns Rev 1
and 10m/s. It shows that all models capture
the general trend, including the directions
with the largest losses and those with peak
efficiency. It is also seen that, for some
particular directions, all models disagree
with measurements. It was verified that
model error does not correlate with the
amount of data in each sector, which
indicates that there was enough data to
perform the calculation and the error stems
from other factors.
Figure 7: Wind farm efficiency as a function
of wind direction every 10°, predicted by all
three models and from measured data from
Horns Rev 1 at 10m/s.
It is worth stressing that the overall wake
loss results presented are obtained with the
wind rose of the final filtered dataset. It is
not representative of a full year. Moreover,
these results are given for one wind speed
value and are therefore not representative
of the total wind farm wake losses.
Table 2 shows the resulting mean signed
error for all models in two test cases where
enough measurements were available and
WindModeller simulations with a 30°
resolution were performed: Horns Rev 1
with 10m/s wind speed and Gunfleet Sands
with 8m/s wind speed. Furthermore, 10°
WindModeller simulations were available
for the Horns Rev 1 case, and an additional
field is included in the table for this case. It
is found that all models perform well, within
3% error.
Case
Fuga
N.O.J.
WM
HR1-30,
10m/s
1.4%
0.7%
-0.1%
HR1-10,
10m/s
2.1%
1.5%
3.0%
GFS-30,
8m/s
2.5%
1.8%
-1.2%
Table 2: Mean signed error in overall wake
losses for the three models in three test
cases in Horns Rev 1 and Gunfleet Sands.
4. Discussion
Models are easy to use. In the N.O. Jensen
model, the user can only change the wake
decay constant, which makes it very robust
and well suited to the wind analysts
community. Fuga has a limited amount of
parameters that can be tuned, namely the
height of the boundary layer (forcing), the
roughness length and the averaging
method. Considering that the height of the
boundary layer does not drastically
influence the results if it is set high enough,
the influence of user is limited to the
averaging method and the roughness
length. This makes Fuga and the N.O.
Jensen very simple and easily reproducible.
In contrast, like any industrial CFD code,
WindModeller has to be used with caution
due to the multitude of user defined
parameters. The turbulence model setup is
very well documented [2], but the user has
to check grid and numerical convergence.
The results obtained for the analysis of
wake losses along a row of turbines in
Horns Rev 1 confirm the findings of the
OWA and previously published studies, for
instance [10]. The comparison of wake
models with 1° transect results are found to
be irrelevant, because of the significant
scatter in the wind direction measurements.
In addition, unsteady flow features like
wake meandering are beyond the capability
of the models in this study.
Fuga and Windmodeller are the closest to
the measurements for the 10 and 30° bin
width. The absolute error for each of the
models decreases with the bin width
(Figure 6). This is especially true for Fuga
which shows similar errors as WindModeller
when the width exceeds 10°. These values
allow the population of records to be
representative enough with regards to the
030 60 90 120 150 180 210 240 270 300 330
0.4
0.5
0.6
0.7
0.8
0.9
1
1.1
Wind Direction [deg]
POW[-]
Meas.
FUGA
N.O. Jensen
WindModeller
model assumptions (steadiness,
especially).
Similar comments can be made on the
three other wind farms, with different wind
speed, Ct, and spacing values. With a 10°
bin width, WindModeller comes first, Fuga
second, and N.O. Jensen third. One could
then expect similar behaviour with other
regular arrays. The CFD codes are better
suited to capture lateral turbine interaction,
which in the case of regular arrays is
important [6]. This could explain the better
accuracy of Fuga and WindModeller when
looking at 10deg bins. When comparing on
30deg bins, the N.O. Jensen performs well
in comparison to Fuga, and nothing
indicates that one model should be
preferred.
Regarding the propagation of the wake
losses along a row of turbine, one can
observe that the pattern is different for each
wind farm. For Horns Rev 1 and Gunfleet
Sands, the losses keep decreasing,
whereas in Barrow and Burbo the efficiency
slightly increases after the second turbine,
and then decreases. This is especially
visible with 10° averaged bin size. One
could suggest this is due to the low turbine
spacing combined with the high thrust at
8m/s. Fuga seems to be the only method
properly capturing this pattern.
Due to the high computational resources
required, it is unfeasible and undesirable to
run 360-1° simulations for yield estimate
and layout optimisation purposes.
Therefore Ansys WindModeller is not
equally represented as the others because
simulations have not been run for every 1°.
It might be missing a smoothing effect the
others have, but the results from the HR1-
30° resolution case (Figure 6) suggest that
this effect would not be as drastic as with
Fuga.
Regarding overall wake losses (Figure 7),
the three models seem to capture the
general trend in some directions, but not all.
In some sectors, all models are very close
to each other, yet none replicates perfectly
the measurements.
The pattern of wake losses per sector is
symmetric for the models in HR1, which is a
symmetric layout. The measurements,
however, do not show this perfect
symmetry. The east half of the wind rose
shows higher wake losses than the
western. This could be due to both the fact
that easterly and westerly winds in this
location are climatologically different, or to
the fact that the data subset used to build
these statistics is not symmetric regarding
the content of different turbulence levels or
atmospheric stability conditions. For
example, the sectors between 110°-130°
are clearly less efficient than those between
290°-310°, and the proportion of stable
conditions in these two directions is found
to be very different (75% for the range
110°-130° against 20% for the range 290°-
310°).
The most sensible way to compare the
models is with the mean error in overall
wake losses, rather than the mean absolute
error. This is closer to what would be done
in an energy yield estimate, where the
positive errors in a sector are combined
with the negative errors in another one. It is
complicated to pick a definite best with
these results, but a clear conclusion is that
under this analysis, the N.O. Jensen model
is found to be robust and perform as well as
the novel, more complicated models.
After commenting on the results, it is worth
stressing that comparing wake models with
power measurements involves strong
filtering of the data: for availability, wind
speed, etc. In order to maximise the
amount of data, additional filters like
stability, turbulence, speed up effects
around the wind farm, or unsteady
situations within 10min period have not
been taken into account. These problems
can be counter-balanced by a high number
of representative samples, in each of the
wind direction bins. And indeed, similar
results are obtained on all wind farms,
although they are in different climates and
equipped with different machines. However,
the interaction between stability and layout
efficiency, both varying with the wind
direction, needs to be better understood.
5. Conclusion
The OWA results have been replicated
independently on a similar but not identical
dataset without a concurrent on-site met
mast for the full measurement period. The
relevance of single degree transects can be
questioned, since it exceeds both
measurements (yaw precision) and model
limitations (steadiness).
From the previously unpublished
comparison with three wind farms in the
North Sea and the Irish Sea, WindModeller
and Fuga perform better than N.O. Jensen,
and give comparable results when the wind
direction bin size exceeds 10°. CFD models
are likely to better capture the lateral
interaction of wakes in regular arrays.
When comparing with 30° bins, all models
perform equally well.
A way of assessing model accuracy in
overall wake losses has been proposed.
The N.O. Jensen model is robust: despite
its simplicity it withstands the comparison
and proves to be useful even when
compared to much more complex models.
The data filtering focused on maximising
the amount of measurements. Thus, the
dataset contains various stability conditions,
as well as other complex flow features. This
study focuses on including three new wind
farms to wake model comparisons and
extending these comparisons to overall
wake losses, instead of investigating the
effect of these features on the
measurements. Future work should focus
on better understanding the interaction
betwen stability, turbulence, availability,
park speed-up and inter-park effects.
6. References
[1] Ott, S. Linearized CFD. Risø-I-3093.
January 2011.
[2] Montavon C. et al. Offshore wind
Accelerator: Wake modelling using CFD.
EWEA 2011, Brussels.
[3] Jensen, N.O. A note on Wind Generator
Interaction. Risø-M-2411, Risø National
Laboratory, 1984, Roskilde, Denmark
[4] Katic, I., Højstrup, J., Jensen, N.O.. A
Simple Model for Model for Cluster
Efficiency. EWEC 1986 Proceedings, 7-9
October 1986, Rome.
[5] Méchali M. et al. Wake effects at Horns
Rev and their influence on energy
production. EWEC 2006, Athens.
[6] Barthelmie, R.J. et al. Quantifying the
Impact of Wind Turbine Wakes on Power
Output at Offshore Wind Farms. Journal of
Atmospheric and Oceanic Technology,
2010.
[7] Barthelmie, R.J. et. al. Wp8: Flow.
Deliverable D8.1 Data. Wake
measurements used in the model
evaluation. UpWind Project.
[8] Jensen, L. Array efficiency at Horns Rev
and the effect of atmospheric stability.
EWEA 2007, Milan.
[9] Hansen, K. et al. The impact of
turbulence intensity and atmospheric
stability on power deficits due to wind
turbine wakes at Horns Rev wind farm.
Wind Energy Journal, 2011.
[10] Gribben, B. Wake effect development
in the Offshore Wind Accelerator. EWEA
Workshop, Brussels, 2011.
