Renewable Energy 81 (2015) 173e181

Contents lists available at ScienceDirect

Renewable Energy
journal homepage: www.elsevier.com/locate/renene

Wind farm power optimization including flow variability

Jürgen Herp a, *, Uffe V. Poulsen b, Martin Greiner c, d
a
The Mærsk Mc-Kinney Møller Institute, University of Southern Denmark, Campusvej 55, DK-5230 Odense M, Denmark
b
Center for Applied Research and Development in Building, Energy and Environment, VIA University College, Chr. M. Oestergaards Vej 4, DK-8700 Horsens,
Denmark
c
Department of Engineering, Aarhus University, Inge Lehmanns Gade 10, DK-8000 Aarhus C, Denmark
d
Department of Mathematics, Aarhus University, Ny Munkegade 118, DK-8000 Aarhus C, Denmark

a r t i c l e i n f o a b s t r a c t

Article history: A model-based optimisation approach is used to investigate the potential gain of wind-farm power with
Received 8 January 2014 a cooperative control strategy between the wind turbines. Based on the Jensen wake model with the
Accepted 12 March 2015 Katic wake superposition rule, the potential gain for the Nysted offshore wind farm is calculated to be 1.4
Available online 30 March 2015
e5.4% for standard choices 0.4
k
0.25 of the wake expansion parameter. Wake model fits based on
short time intervals of length 15sec  T  10 min within three months of data reveal a strong wake flow
Keywords:
variability, resulting in rather broad distributions for the wake expansion parameter. When an optimized
Wind farm
wind-farm control strategy, derived from a fixed wake parameter, is facing this flow variability, the
Wakes
Flow variability
potential gain reduces to 0.3e0.5%. An omnipotent control strategy, which has real-time knowledge of
Model-based data analysis the actual wake flow, would be able to increase the gain in wind-farm power to 4.9%.
Operational wake-flow modelling © 2015 Elsevier Ltd. All rights reserved.
Optimization

1. Introduction The purpose of this paper is to use a model-based optimisation

approach to investigate the potential gain of wind-farm power with
Wind energy is becoming increasingly important in power a cooperative control strategy between the wind turbines. Based on
systems all over the globe. As the technology is maturing, the a simple, but realistic wind farm model and a calibration with real
current focus is on bringing down the cost-of-energy by continuing wind farm data, previous studies are extended by investigating the
to increase the efficiency of the manufacturing, the installation and sensitivity of the gain with respect to the uncertainty and vari-
the operation of a wind turbine [1]. A part of this story is also the ability in the wake flow conditions.
emergence of wind farms, where up to several hundred turbines The paper is organized as follows: Section 2 describes a simple,
are placed in close proximity. The current design and operation of standard wind farm model, which includes a single control variable
wind farms still leave room for improvements. This makes wind for each turbine. The model itself is based on momentum conser-
farms also an interesting object for more fundamental research. vation and empirical rules producing reasonable results [19e21].
From a physics point of view, wind farms are interesting because Based on this model, Section 3 evaluates the potential for
of the interaction between the turbines. The most upwind turbines increasing the farm production by cooperative rather than selfish
extract some of the kinetic energy from the wind and produce operation. This is achieved by sequentially updating the control
wakes for the downwind turbines, which then experience reduced parameters of the individual turbines. Section 4 introduces
power generation and increased fatigue loads. These wake in- ensemble models which include the variability of flow conditions.
teractions between the turbines turn the optimal layout design of The impact of such a variability on the optimization potential is
the wind farm into a non-trivial problem [2e5]. The wake in- presented in Section 5. A conclusion and outlook is provided in
teractions also challenge the design of an optimal operational Section 6.
control, which maximises the overall wind farm power generation
or minimizes the overall fatigue loading [6e18].
2. Standard wind farm models

We consider a wind farm consisting of N identical wind turbines,

* Corresponding author. Tel.: þ45 65509232. indexed by i 2 {1,2,…,N}. The utilized wind farm model consists of
E-mail address: [email protected] (J. Herp). three components: (i) a model to characterize the control of a wind

http://dx.doi.org/10.1016/j.renene.2015.03.034
0960-1481/© 2015 Elsevier Ltd. All rights reserved.
174 J. Herp et al. / Renewable Energy 81 (2015) 173e181

turbine and the power generated by it, (ii) a single-wake model to have been put forward [34e36], but we will exclusively deal with
describe the velocity deficit behind a wind turbine, and (iii) a wake the Jensen model [37]. Slight variations of this model are widely
interaction model to describe how wakes from multiple upwind used and by appropriate tuning to data they give quite reasonable
turbines combine at a given downwind turbine. results [19e21]. Within the Jensen model the wake is assumed to be
axially symmetric, with a radius r(x) ¼ R þ kx increasing linearly as
a function of the downwind distance x from the turbine. The
2.1. Turbine model
adjustable wake parameter k is typically chosen to be about 0.07 for
onshore and 0.03e0.05 for offshore wind farms [19,20]. The
A modern wind turbine extracts kinetic energy from the wind,
different values are explained by the lower turbulence intensity
primarily via the lift force created on the fast moving, slender
offshore. Outside the wake, the wind field is assumed to be un-
blades equipped with high-quality airfoils [22]. The detailed
disturbed and the air to move with the free stream wind speed u.
description of this non-steady, high Reynolds number fluid dy-
Inside the wake, the wind speed is reduced, but will recover to-
namic process taking place in the atmospheric boundary layer is
wards u as x increases downstream:
complicated. Using Computational Fluid Dynamics (CFD) with a
realistic representation of the blades is just barely possible and !
then only for the flow field close to the turbine [23e27]. Simplifi- 1q
uw ðx; qÞ ¼ u 1 
2 : (4)
cations of the rotor modelling puts CFD calculations for whole wind 1 þ k Rx
farms within reach [28e32], but such simulations are still not
suited for iterative optimization of e.g. control strategies. Depending on the wind direction q a downwind turbine can be
Fortunately, the turbines power output and influence on the influenced by a wake. If the wake has an overlap Aoverlap with the
average flow field can be modelled much simpler. Consider a tur- downwind turbine's rotor disc, the turbine is ascribed an effective
bine facing uniform inflow with wind speed u. The turbine is pro- relative wind speed deficit given by
ducing energy at a rate P and extracting axial momentum at a rate T.
These key quantities are conveniently parametrized by dimen- 1q Aoverlap ðqÞ
sionless power and thrust coefficients CP and CT: dðq; qÞ ¼  2 ; (5)
A
1þ k xðqÞ
1 þ q
 R
1
P¼ rAu3 CP ; CP ðqÞ ¼ 1  q2 ; (1)
2 2
where x is now the downwind distance between the turbines. This
1 distance depends on the wind direction and the farm layout. Given
T¼ rAu2 CT ; CT ðqÞ ¼ 1  q2 : (2) a fixed layout, the relative velocity deficit at the downwind turbine
2
is only a function of the wind direction q and control parameter q
r is the air density, u the wind speed, and A ¼ pR2 the rotor disc with for the upwind turbine.
radius R.
Both power and thrust are parametrized by the single control
parameter q ¼ u'/u, which describes the reduced wind speed u' 2.3. Wake superposition
behind the rotor disc. When q is decreased, the thrust is increased
and the air passing through the disc looses a larger fraction of its The individual wakes from the turbines of a wind farm expand
kinetic energy. However, the increase in thrust will also cause more downstream and eventually they will start to overlap. A wind farm
air to avoid the disc and hence the total extraction of kinetic energy model must therefore include a recipe for the superposition of
will eventually start to decrease with decreasing q. The optimal wakes. We use a simple empirical rule frequently used in wake
value of q turns out to be qBetz ¼ 1/3. A rotor operated in this way assessment [20,38,39]. To determine the velocity deficit at turbine i,
can in principle have CPBetz ¼ 16=27, the so called Betz-Joukowsky the so-called Katic procedure sums over the squared velocity def-
limit [33]. A real rotor can be reasonably approximated by the icits resulting from all upstream turbines j producing a wake for i,
actuator disc model when aerodynamic drag can be ignored and

X
when operated at a high tip-speed ratio [22]. d2i ðq; qÞ ¼ d2ij q; qj : (6)
isj

2.2. Single wake model Here dij is the velocity deficit at the position of turbine i calcu-
lated as if only the turbine j existed. The effective velocity deficit di
The region directly behind the turbine, where the wind speed depends on the upwind control parameters, but for notational
has been reduced, marks the upstream end of the turbine wake. simplicity we write it as a function of the whole set q ¼ (q1,…,qN).
Further downstream the wake becomes wider and the wake wind
speed uw recovers towards the surrounding free stream wind speed
u due to momentum diffusion. Depending on the degree of tur-
2.4. Example results: Betz optimal turbines
bulent mixing and the downstream distance, a turbine standing in
the wake experiences a deficit
Let us now show a typical result of such a wind-farm model. The
u  uw turbines are modelled as actuator discs with qi ¼ qBetz ¼ 1/3 and
d¼ (3) hence CP ¼ CPBetz ¼ 16=27. The Jensen model with k ¼ 0.04 and the
u
Katic wake superposition is used, and applied to the layout of the
and is not able to produce the same power as a turbine in the free Danish offshore wind farm Nysted; see left panel of Fig. 1. Because
stream. of the regularity of the layout, the farm has a number of ‘‘symme-
In order to describe these wake effects in a wind farm, a wake try’’ axes, along which the turbines are lined up and where wake
model is needed in addition to the turbine model. Again CFD can be effects are particularly strong.
employed, but for our purposes much simpler, semi-empirical The total production of the farm is a function of wind speed and
models are more useful. Several such engineering wake models direction:
 J. Herp et al. / Renewable Energy 81 (2015) 173e181 175

Fig. 1. On the left the Nysted wind farm layout is shown. Some of the major and minor symmetry directions are indicated. On the right the calculated farm efficiency is plotted as a
function of wind direction. Betz optimal turbines, the Jensen wake model with k ¼ 0.04 and the Katic superposition have been used.

3.1. Optimization method

X
N
h i 
Betz
Pfarm ðu; qÞ ¼ P ui u; q; qBetz ; qBetz To improve the control we apply a sequential optimization
i¼1
scheme. Given the ambient wind speed, the wind direction, and a
N
h i3
1 3 Betz X guess on the initial control parameters the turbines are first sorted
¼ ru ACP 1  di q; qBetz : (7)
2 according to their position with respect to the wind direction.
i¼1
Starting with the most upstream one, we then change the operative
When normalized to N times the production of an isolated mode, i.e. the control parameter, of one turbine after the other such
turbine the total production of the farm turns into the direction- that the overall power is maximized. Note that only a single qi is
dependent farm efficiency1: changed at each step, while the others are kept constant. We then
iterate the scheme until no further improvement is made. Algo-
Betz ðu; qÞ
Pfarm rithm 1 summarizes the sequential scheme. Other update schemes,
hfarm ðqÞ ¼ including update steps with multiple q parameters, and other
NP Betz ðuÞ
N
h i3 (8) optimization schemes, like gradient based and particle-swam
1 X
¼ 1  di q; qBetz : methods, have been considered. They produced the same results,
N i¼1 but were all far more time consuming. They did, however, give us
reasonable confidence that the maximum we find is a global one.
The symmetry directions are clearly visible in the polar effi-
ciency plot of Fig. 1: While some wind directions lead to almost no
wake effects, the farm efficiency is less than 40% for the worst wind
direction. A uniform average over wind directions gives a mean
efficiency of 83%.

3. Control optimization in standard models

Now we focus on the cooperative control problem for a wind

farm to improve its efficiency. The farm layout is kept fixed, but the
control parameters q ¼ (q1,…,qN) are allowed to vary:
3.2. Two turbines example

X
N
Pfarm ðu; q; qÞ ¼ Pi ðui ½u; q; q; qi Þ: (9) Let us start by considering just two turbines in a simple geom-
i¼1 etry where the downwind turbine is fully in the wake of the up-
wind turbine. Then Aoverlap ¼ A. Obviously, the downwind turbine
The optimal control design leads to a set qopt, which maximizes should behave selfishly and operate at the Betz optimum with
the total generated power. This is done by increasing qi for the q ¼ qBetz ¼ 1/3. For the upwind turbine we show in the upper panel
upwind turbines, causing them to pass more wind, and therefore in Fig. 2 the optimal q value as a function of kx/R: For small values,
allowing downwind turbines to produce more. the interaction is strong and the optimal q becomes as high as 0.62.
For large values, the wake is weaker and q becomes lower,
approaching 1/3 for vanishing wake interaction. In the lower panel
1
Note that the farm efficiency does not depend on u in our model. This is only
the actual benefit is shown. The optimization gain x is defined by:
realistic for wind speeds where the turbines are producing less than their rated
power and are therefore operated close to the Betz optimum. For high wind speeds,
opt Betz
q and hence CT is lowered significantly in order to limit the extracted power to the Pfarm ¼ ð1 þ xÞPfarm : (10)
rated capacity of the turbine's generator.
176 J. Herp et al. / Renewable Energy 81 (2015) 173e181

Fig. 2. Optimal q value (upper panel) for the upwind turbine and the fractional gain
(lower panel) for a simple two-turbine wind farm. The downwind turbine is assumed
to be completely in the wake of the upwind turbine. Fig. 4. Gain potential as a function of wind direction for three different values of the
wake expansion parameter k.

Strong interaction values obviously lead to a high optimization

potential: 7.5% for kx/R ¼ 0.2. The potential to increase the generated power is largest for wind
It is important to note that if the wind direction is not known directions along the symmetry axes. The magnitude of the increase
exactly, the optimization potential is strongly affected. If the in the production is also observed to be highly dependent on the
downwind turbine is not completely covered by the wake, the distance between successive turbines as larger intra-farm distances
overlap factor from Eq. (5) not only reduces the gain potential at the permit the wind field to recover more of its original energy. As a
downwind turbine, but also alters qopt. In Fig. 3 the gain is plotted as broader wake regains its energy faster, the gain potential is higher
a function of misalignment of the turbines with respect to the wind for more narrow wakes with a smaller wake expansion parameter:
direction for k ¼ 0.04 and two different turbine separations. The For an average over wind directions weighted by the actual wind
fractional gain is calculated using the q that would be optimal for rose for a particular 3 month period (see Fig. 6) we get x ¼ 5.4%,
perfect alignment, i.e. full overlap. As can be seen, when the 1.4%, and 0.39% for k ¼ 0.025, 0.04, and 0.05, respectively.
misalignment is more than 6+, the gain quickly drops and becomes In Fig. 5 we show an example of the gain and the optimal q for
in fact negative. This is because the upwind turbine is still loosing each of the turbines in the farm. The Jensen wake model with
production by increasing its q, but without interaction, the down- k ¼ 0.04 has been used and the wind is coming from the south-east.
wind turbine is not gaining anything. In the remaining part of this We see how the frontline turbines loose production while down-
paper we will always optimize control parameters assuming that wind turbines gain. The qopt values are quite similar in the inner
the wind direction is perfectly known, but it should be noted that part of the farm, but turbines on the downstream edge of the farm
this is in fact a very strong assumption. are of course behaving selfishly (q ¼ qBetz ¼ 1/3) as their wakes are
not affecting any further turbines.

3.3. Full wind farm: the Nysted example

4. Ensemble models
We will now show results of the optimization applied to the 72
turbines of the Nysted offshore wind farm in Denmark. Although
Engineering wind farm wake models like the one described in
we use the realistic layout of Fig. 1, the turbines are still represented
Sec. 2 are designed for the prediction of highly averaged quantities
by the simplified model of Sec.2.1.
like the Annual Energy Production (AEP). With standard values of
When changing the wind farm to an optimized cooperative
the wake expansion parameter the model compares well to the
control, a strong dependence on not only the wind direction, but
data. The latter are at first temporally averaged over 10 min, then
also the wake expansion parameter is observed. Fig. 4 shows this
binned into wind speed and direction sectors, and finally averaged
dependency for q 2 [0+,360+] and k ¼ 0.025, 0.04, 0.05.
within each sector. Before the sector averaging is performed the
data contains a lot of variability. Wake losses in any given 10 min
period are not a simple function of wind speed and direction. This is
because real wakes meander rather than move straight down-
stream [40e47] and because atmospheric conditions such as
thermal stratification influence their behaviour [19,48,49]. After
the averaging, the ‘‘wakes’’ are necessarily to be understood as
statistical mean patterns of the wind speed. They will appear
broader and weaker than the typical instantaneous wakes and are
described by relatively high values of k in the Jensen model.
An alternative approach to analyse the data is to focus on single
‘‘events’’: An event describes the production of all turbines in a
wind farm, averaged over just a very short period of time. Still each
of these events can be fitted with a simple wind farm wake model.
Then the wind speed and direction, and also the wake expansion
parameter will vary between events. We will use such an ensemble
Fig. 3. Fractional gain as a function of the angle between wind direction and the line
connecting the two turbines. We have taken a wake expansion parameter of k ¼ 0.04
model to assess the impact of fluctuations on the optimization
and two different values of x. For the Nysted wind farm x ¼ 12.2 R and x ¼ 21.2 R potential for a wind farm. In this Section we describe how the
correspond to wind along 358 and 98 respectively. ensemble model is extracted from data.
 J. Herp et al. / Renewable Energy 81 (2015) 173e181 177

Fig. 5. Optimization results for a complete wind farm using the Jensen wake model with k ¼ 0.04. The wind is coming from the south-east as illustrated by the grey arrow. Left
panel: Gain factor for each turbine. Right panel: The optimal q values.

4.1. The data shown in Fig. 1 of Ref. [19]. Note that periods where some of the
turbines are out of operation do not have to be excluded. Since each
Three months (MarcheMay 2009) of data from the 72 Siemens- event is analysed separately, non-active turbines can be included.
Bonus 2.3 MW turbines comprising the Nysted off-shore wind farm This is one advantage of a single-event method compared to
have been used. The temporal sampling frequency of the recordings standard approaches.
is 1 Hz and they include power and yaw readings from each turbine. It is standard to average the data in each 10 min period. Since we
Since data from nearby meteorology masts were not available, the have access to a finer temporal resolution we can also choose to use
analysis is based on turbine sensor readings only. only the first e.g. 30 s of each period to define an event. This will
The 3 month time span of the data is divided into disjoint 10 min allow us to investigate the impact of temporal averaging with
periods. Only data where the front-line turbines produce more than different averaging times T ¼ 15, 30, 60, 120, 300, and 600 s.
150 kW and less than 2150 kW are considered. This filter excludes
about 20% of the data because of low wind speed and another 20% 4.2. Parametrization of the free wind
because of high wind speed. The filter is applied since power is
translated to wind speed via the power curve of the Nysted tur- Apart from the farm layout and the CT curve of the turbines, the
bines, and this can only be done comfortably between cut-in and input variables of the Jensen model are the free wind speed u, the
rated power. The power and CT curves of the Nysted turbines are wind direction q, and the wake parameter k. We choose to extract u
directly from the data in the following way: First an approximate
wind direction is determined by a simple average of the yaw
readings of all turbines. We apply a whole-farm correction as
described in Ref. [50] to take into account systematic errors in the
yaw sensors. This allows to find the turbines on the farm perimeter
which are not experiencing wakes from other turbines. Because of
the farm geometry there are always (barring turbine outages) at
least 8 such turbines: 8 when the wind is near the South-North
direction, 9 when the wind is near the West-East symmetry axis,
and 16 for the directions in between that are more aligned with the
farm diagonals. For each event, we base our estimate of the free
wind on the 9 most upwind of these free wind turbines, both in
order to work with approximately the same number of turbines for
all wind directions and to extract the free wind near the ‘‘front end’’
of the farm. This approach is similar to the one used in Ref. [19]. The
inverse power curve is used to estimate the wind speed experi-
enced by these upwind turbines. Fig. 6 illustrates the wind rose
derived from the data. Compared to the long term wind rose for
Nysted (see e.g. Ref. [19]), easterly winds are clearly over-
represented. This, however, is of little consequence for our in-
vestigations here.

4.3. Fit cost function

For a given free-stream wind speed, the modelled turbine power

productions depend on the value of the wake expansion parameter
k and the wind direction. For each single event we determine these
Fig. 6. The wind rose derived from the used data. The three blocks for each 15 sector two remaining parameters in a fitting procedure that aims at
indicate on a linear scale the percentage of the data, for which the front-line turbine
generation is below 150 kW (blue), between 150 and 2150 kW (green), and above
adjusting the model to reproduce the observed production pattern
2150 kW (red). (For interpretation of the references to colour in this figure legend, the as well as possible in that particular period of time. The fit cost
reader is referred to the web version of this article.) function is chosen as
178 J. Herp et al. / Renewable Energy 81 (2015) 173e181

P72  
fit 
i¼1 Pi  Pi 
ε¼ P ; (11)
Pi
i

which is the normalized sum of absolute differences between

actual power production and model prediction for each turbine.
Alternatively we could have used the deviation between the
modeled and actual total production for a given event, but as we are
interested in capturing the production pattern it is more relevant to
use Eq. (11) where over- and under-estimations are not allowed to
cancel.

4.4. Results

First we would like to investigate how the averaging time T

influences the fit quality of the wake model. Fig. 7 shows the mean
fit error for the ensembles based on different values of T. The errors
Fig. 8. Distribution of the fitted wake expansion parameter k for the Jensen model
are in generally rather large. They decrease from 〈ε〉 ¼ 0.249 for the with Katic wake superposition. Results for three different values of the averaging time
T¼ ¼ 15 second ensemble to 〈ε〉 ¼ 0.184 for the T ¼ 10 minutes T are shown. The dashed line is a fitted beta-distribution pbeta(k;a,b) ¼ C ka1(1  k)b1
ensemble. Although the best fits are obtained for the longest with parameters a ¼ 2.186, b ¼ 80.17, and C ¼ 1.382  104.
averaging times considered, the convex shape of the curve indicates
that not much more will be gained by considering even longer T's.
For comparison, a no-wake model, where all wakes are simply In the left panel of Fig. 9 the dependence of the fitted wake
ignored and only the free wind modelling is applied, yields an parameters on the estimated upstream wind speed is illustrated.
average error of 0.32 for T ¼ 10 minutes. We observe a clear, but not dramatic decrease of k with u. We do
The distributions of the fitted wake parameter k for three en- not show plots of fit error as a function of wind speed, but let us just
sembles are plotted in Fig. 8. The distributions show little depen- mention a tendency to better fits (in terms of the chosen cost
dence on T. This is somewhat surprising as wake meandering is function) as the wind speed becomes larger.
expected to be averaged out for T ¼ 10 minutes compared to T ¼ 15 The dependence of the fitted wake parameters on the fitted
seconds. Note however, that via the fit procedure a single wind wind direction is plotted in the right panel of Fig. 9. The wake
direction is enforced upon the whole wind farm, which does not parameter k turns out to be larger along the symmetry axes of the
allow to account for separate wake directions. Although the dis- wind farm. The main symmetry axes of the Nysted wind farm are at
tribution for k does not significantly depend on T, a longer T results q ¼ 98 (278 ) and 178 (358 ), which represent the East-West and
in better fits of the production pattern, as is clearly seen in Fig. 7. the South-North directions. Other symmetry directions are 44
From now on we will use only results for the T ¼ 10 minutes (224 ), 67 (247 ), 124 (304+) and 140 (320 ), see Fig. 1. This
ensemble. finding is in line with Fig. 9 of Ref. [50].
The mean value of the wake expansion parameter is k ¼ 0.025
and turns out to be significantly smaller than the k ¼ 0.04 most 5. Control optimization in ensemble models
often used for offshore wind turbines when working with event-
averaged data [19,20]. This result is similar to what was found in In the previous section we learnt that the wake expansion
a previous study [50] using a single-event analysis on a different parameter varies quite a lot in ensemble models. We will now
Nysted data set. investigate the consequences of this variability when trying to
optimize the operation of the farm.

5.1. Facing the wrong wake flow event

Let us assume that the wind farm operator is using the control
parameters qopt, obtained from a model-based optimization with a
standard wake expansion parameter, say k ¼ 0.04. For wake flow
events described by k ¼ 0.04 this is the optimal choice, but for other
events with ks0.004 this is not optimal. Fig. 10 shows the gain
factor when qopt(k ¼ 0.04) is facing events with k ¼ 0.04 and
k ¼ 0.025. The full line is the expected gain for a k ¼ 0.04 event and
it shows x ~ 14% in a window of about 5+ around the east-west
symmetry axis. The dashed line is obtained by using the same
qopt(k ¼ 0.04) for a k ¼ 0.025 event. The dashed curve is always
below the solid curve, representing a smaller gain. For some wind
directions, the dashed curve even becomes negative.
When the calculation with qopt(k ¼ 0.04) is extended to all wind
directions and weighted with the wind rose of Fig. 6 we get an
Fig. 7. Mean fit error as a function of the averaging time T ¼ 15, 30, 60, 120, 300, and
average gain of x ¼ 2.4% for the k ¼ 0.025 events. This is significantly
600 s. The solid curve represents an exponential ε∞ þ bexp(T/T0) with an asymptotic less than the 5.4% which can be gained by using qopt(k ¼ 0.025).
residual error ε∞ ¼ 0.183; the other parameters are b ¼ 0.0666, and T0 ¼ 194 seconds. Switching the roles of the two k values and using qopt(k ¼ 0.025) in
 J. Herp et al. / Renewable Energy 81 (2015) 173e181 179

Fig. 9. Ensemble wake parameters for T ¼ 10 minutes as a function of the estimated up-wind speed (left panel) and the fitted wind direction (right panel). The dots represent the
ensemble of events while the lines are running averages based on a 0.1 m/s and 10 window width, respectively. Faded vertical lines in the right column indicate the park symmetry
axes, see Fig. 1.

where Pfarm(u,q,k,q) is described by Eq. (9). As the objective is to

maximize Pfarm, the choice of the control vector q becomes crucial.
Let us assume that the wind-farm operator is able to determine the
upwind speed u and direction for a single wake-flow event in a
unique manner, but has no knowledge about the actual wake
expansion parameter k. The operator will then pick a control vector
q ¼ q(u,q,k0), which is optimal for a typical reference wake
parameter k0. When compared to the Betz case qBetz with qi ¼ 1/3
for all turbines, the overall gain in average wind-farm power gen-
eration (12) is 0.35% and 0.42% for the choice k0 ¼ 0.025 and 0.04,
respectively. These gain factors are significantly smaller than the
previously obtained values in Table 1, when the wake flow
ensemble was assumed to have only a single wake parameter.
Actually, the wind farm operator could do better with
Fig. 10. Gain factor as a function of wind direction when optimization is done based on
a wrong value of the wake expansion parameter. The full line is for a Jensen wake
k0 s 0.025 or 0.04 for q(u,q,k). For each u,q an optimal k0 ¼ kopt(u,q)
model with k ¼ 0.04 and correspondingly optimized q values. The dashed line is the can be found:
result of using the same q's but assuming k ¼ 0.025. Z
kopt ðu; qÞ ¼ arg max dkpðu; q; kÞPfarm ðu; q; qðu; q; k0 ÞÞ: (13)
a k ¼ 0.04 event gives a gain of 1.1%, again less than the 1.4% pro- k0
duced with qopt(k ¼ 0.04). These results are summarized in the
upper left part of Table 1. The use of the control vector q(u,q,kopt) increases the overall
gain factor to 0.52%, which is about 20% better than for the control
vectors with k0 ¼ 0.025 and 0.04.
5.2. Facing the wake-flow ensemble So far we have assumed that the wind-farm operator has no
knowledge about the wake expansion coefficient of the actual flow
The wind-farm is facing a large number of different wake-flow field. If the wind farm controller has some way of obtaining k in
situations. All of them together define the wake-flow ensemble, real-time, we could always employ the appropriate qopt(u,q,k). We
which is characterized by the multivariate probability density denote such a controller as omnipotent since within our model we
p(u,q,k) for wind speed, wind direction and wake expansion have full information. This case should be considered as a bench-
parameter. The probability density p(u,q,k) is extracted from the mark which gives an upper limit to the possible gain. For the data
three month of data, described in Section 4.1. It includes the cor- ensemble we find that the omnipotent controller can reach
relations between k and u, and between k and q, which have been x ¼ 4.9%. This number is comparable again to the gain from a
shown in Fig. 9. k0 ¼ 0.025 controller with k ¼ k0 wake flow events. Consult again
The average wind-farm power generation can now be expressed Table 1.
as:
Z Z Z 6. Conclusion
Pfarm ¼ du dq dkpðu; q; kÞPfarm ðu; q; k; qÞ; (12)
We have studied the optimization potential for the total power
production of a wind farm by employing cooperative control
schemes where each turbine is not necessarily maximizing its own
output. Based on an engineering wind farm modelling framework a
Table 1
Gain factor x for different combinations of qopt and wake flow events. simple sequential optimization scheme has been applied. The
resulting gain was found to strongly depend on the chosen wake
qopt Wake flow events
expansion parameter. Narrow wakes lead to much higher gains
k ¼ 0.025 k ¼ 0.04 Data ens. than broader wakes. For the Nysted offshore wind farm, the stan-
K ¼ 0.025 5.4% 1.1% 0.35% dard value k ¼ 0.04 for the wake expansion parameter leads to a
k ¼ 0.04 2.4% 1.4% 0.42% potential gain of 1.4%, while k ¼ 0.025 yields 5.4%. The importance
data ens. e e 0.52% of knowing the exact wind direction has also been pointed out in
data ens., omni. e e 4.9%
order to avoid unnecessary downregulation. The most important
180 J. Herp et al. / Renewable Energy 81 (2015) 173e181

message to take home from the present work is that the wake [21] Hassan Garrad. WindFarmer theory manual. Bristol, England. v4.2 Edition
2011.
expansion parameter is observed to be highly variable and that this
[22] Burton T, Sharpe D, Jenkins N, Bossanyi E. Wind energy handbook. John Wiley
variability has a negative effect on the wind-farm power optimi- & Sons, Ltd; 2002.
zation potential. The optimization gain is dramatically reduced [23] Hansen M, Sørensen J, Voutsinas S, Sørensen N, Madsen H. State of the art in
below 0.5% once the wind farm controller does not account for the wind turbine aerodynamics and aeroelasticity. Prog Aerosp Sci 2006;42(4):
285e330.
generic variability of the wake parameter. Only if the omnipotent [24] Hsu M.-C, Akkerman I., Bazilevs Y., Finite element simulation of wind
knowledge of the instantaneous k value is available as online input turbine aerodynamics: validation study using Nrel phase vi experiment.
to the running controller, the optimization gain can increase again Wind Energy.
[25] Lynch C, Smith M. Unstructured overset incompressible computational fluid
to 4.9%. dynamics for unsteady wind turbine simulations. Wind Energy 2013;16(7):
It will be an interesting challenge to develop more dynamic 1033e48.
wind-farm control schemes, potentially approaching the perfor- [26] Zahle F, Sørensen NN, Johansen J. Wind turbine rotor-tower interaction
using an incompressible overset grid method. Wind Energy 2009;12(6):
mance of an omnipotent controller, and taking also fatigue opti- 594e619.
mization into account [16]. The necessary real-time wake [27] Li Y, Paik K-J, Xing T, Carrica PM. Dynamic overset CFD simulations of wind
information could be provided by e.g. lidar measurements, in a turbine aerodynamics. Renew Energy 2012;37(1):285e98.
[28] Calaf M, Meneveau C, Meyers J. Large eddy simulation study of fully developed
manner similar to load reduction and power capture optimization wind-turbine array boundary layers. Phys Fluids 2010;22(1):015110.
schemes for a single turbine [41,42,51e54]. [29] Lee S, Churchfield MJ, Moriarty PJ, Jonkman J, Michalakes J. A numerical study
of atmospheric and wake turbulence impacts on wind turbine fatigue load-
ings. J Sol Energy EngTrans ASME 2012;135(3).
Acknowledgements [30] Churchfield MJ, Lee S, Michalakes J, Moriarty PJ. A numerical study of the
effects of atmospheric and wake turbulence on wind turbine dynamics.
We are grateful to DONG Energy A/S and Siemens Wind Power J Turbul 2012;13(N14):1e32.
[31] Wu Y-T, Porte-Agel F. Simulation of turbulent flow inside and above wind
for giving us access to the Nysted data and to Emil Hedevang for farms: model validation and layout effects. Boundary Layer Meteorol
useful discussions. 2013;146(2):181e205.
[32] Andersen SJ, Sorensen JN, Mikkelsen R. Simulation of the inherent turbulence
and wake interaction inside an infinitely long row of wind turbines. J Turbul
References 2013;14(4):1e24.
[33] Okulov VL, van Kuik GA. The betzejoukowsky limit: on the contribution to
[1] Kusiak A, Verma A, Wei X. Wind turbine frontier from scada. Wind Syst Mag rotor aerodynamics by the british, german and russian scientific schools.
2012;3(9):36e9. Wind Energy 2012;15(2):335e44.
[2] Larsen GC, Buhl T, Madsen HA, Troldborg N, Larsen TJ, Ott S, et al. TOPFARM - [34] Rathmann O, Barthelmie R, Frandsen S. Turbine wake model for wind
next generation design tool for optimisation of wind farm topology and resource software. In: Proceedings (online): 2006 european wind energy
operation. 2011. conference and exhibition, Athens (GR), 27 Feb e 2 Mar 2006. Brussels: Eu-
[3] Gonza
lez JS, Rodriguez AGG, Mora JC, Santos JR, Payan MB. Optimization of ropean Wind Energy Association; 2006.
wind farm turbines layout using an evolutive algorithm. Renew Energy [35] Frandsen S, Barthelmie R, Pryor S, Rathmann O, Larsen S, Højstrup J, et al.
2010;35(8):1671e81. Analytical modelling of wind speed deficit in large offshore wind farms. Wind
[4] Kusiak A, Song Z. Design of wind farm layout for maximum wind energy Energy 2006;9:39.
capture. Renew Energy 2010;35(3):685e94. [36] Nygaard NG, Downey R, Me
chali M, Jensen LE. An offshore wake model built
[5] Meyers J, Meneveau C. Optimal turbine spacing in fully developed wind farm from extensive data sets. In: Proceedings (online) of EWEA annual event 2013.
boundary layers. Wind Energy 2012;15(2):305e17. Vienna: European Wind Energy Association; 2013.
[6] Machielse L, Barth S, Bot E, Hendriks H, Schepers G. Ecn-ee07e105: evalua- [37] Jensen NO. Risø-m-2411: a note on wind turbine interaction. Roskilde: Risø
tion of “heat and flux” farm control e final report. Petten: ECN; 2007. National Laboratory; 1983.
[7] Spudi
c V, Baoti

c M, Peri

c MJN. Hierarchical wind farm control for power/load [38] Katic I, Højstrup J, Jensen N. A simple model for cluster efficiency. In: Palz W,
optimization. In: Proceedings of torque 2010, Heraklion, Greece; 2010. Sesto E, editors. European wind energy association conference and exhibition,
[8] Madjidian D, Mårtensson K, Rantzer A. A distributed power coordination Rom, 1986, Vol. 1. Rome: A. Raguzzi; 1987. p. 407.
scheme for fatigue load reduction in wind farms. In: 2011 American control [39] Thøgersen ML. Wake and turbulence models in WindPRO. EMD International
conference, San Francisco, CA, USA; 2011. A/S; 2011. accessed on 10-12-2013.
[9] Bjarnason B, Sveinsson O. € Wind farm controller e optimization of power [40] Larsen GC, Madsen HA, Thomsen K, Larsen TJ. Wake meandering - a pragmatic
production. Master’s Thesis. Lyngby: DTU Electrical Engineering; July 2010. approach. Wind Energy 2008;11:377.
[10] Soleimanzadeh M, Wisniewski R. Controller design for a wind farm, consid- [41] Bingol F, Mann J, Larsen GC. Light detection and ranging measurements of
ering both power and load aspects. Mechatronics 2011;21(4):720e7. wake dynamics part i: one-dimensional scanning. Wind Energy 2010;13(1):
[11] Soleimanzadeh M, Wisniewski R, Kanev S. An optimization framework for 51e61.
load and power distribution in wind farms. J Wind Eng Indus Aerodynamics [42] Trujillo J-J, Bingol F, Larsen GC, Mann J, Kuehn M. Light detection and ranging
2012;107e108:256e62. measurements of wake dynamics. part ii: two-dimensional scanning. Wind
[12] Nilsson K. Numerical computations of wind turbine wakes and wake inter- Energy 2011;14(1):61e75.
action e optimization and control. Ph.D. thesis. Royal Institute of Technology, [43] Espana G, Aubrun S, Devinant P. Is the meandering of a wind turbine wake
KTH Mechanics; Dec. 2012. due to atmospheric length scales? In: Peinke J, Oberlack M, Talamelli A, edi-
[13] Soleimanzadeh M, Wisniewski R, Johnson K. A distributed optimization tors. Progress in turbulence iii, vol. 131 of Springer proceedings in physics.
framework for wind farms. J Wind Eng Indus Aerodynamics 2013;123(Part A): Berlin: Springer-Verlag; 2010. p. 91e4.
88e98. [44] Espan~ a G, Aubrun S, Loyer S, Devinant P. Spatial study of the wake
[14] Lee J, Son E, Hwang B, Lee S. Blade pitch angle control for aerodynamic per- meandering using modelled wind turbines in a wind tunnel. Wind Energy
formance optimization of a wind farm. Renew Energy 2013;54(0):124e30. 2011;14(7):923e37.
[15] Park J, Kwon S, Law KH. Wind farm power maximization based on a coop- [45] Keck R-E, Veldkamp D, Madsen HA, Larsen G. Implementation of a mixing
erative static game approach. In: Society of Photo-optical instrumentation length turbulence formulation into the dynamic wake meandering model.
engineers (SPIE) conference series, vol. 8688 of society of photo-optical J Solar Energy Eng-Trans ASME 2012;134(2).
instrumentation engineers (SPIE) Conference Series; 2013. [46] Yang Z, Li Y, Seem JE. Model predictive control for wind turbine load reduction
[16] Knudsen T, Bak T, Svenstrup M, Survey of wind farm control e power and under wake meandering of upstream wind turbines. In: Proceedings of the
fatigue optimization. Wind Energy. URL http://dx.doi.org/10.1002/we.1760. American control conference; 2012. p. 3008e13.
[17] Marden J, Ruben S, Pao L. A model-free approach to wind farm control using [47] Espan~ a G, Aubrun S, Loyer S, Devinant P. Wind tunnel study of the wake
game theoretic methods, control systems Technol. IEEE Trans 2013;21(4): meandering downstream of a modelled wind turbine as an effect of large
1207e14. http://dx.doi.org/10.1109/TCST.2013.2257780. scale turbulent eddies. J Wind Eng Indus Aerodynamics 2012;101(0):24e33.
[18] Ahmad MA, Azuma S-i, Sugie T. A model-free approach for maximizing power [48] Motta M, Barthelmie RJ, Vølund P. The influence of non-logarithmic wind
production of wind farm using multi-resolution simultaneous perturbation speed profiles on potential power output at danish offshore sites. Wind En-
stochastic approximation. Energies 2014;7(9):5624e46. http://dx.doi.org/ ergy 2005;8(2):219e36.
10.3390/en7095624. URL, http://www.mdpi.com/1996-1073/7/9/5624. [49] Hansen KS, Barthelmie RJ, Jensen LE, Sommer A. The impact of turbulence
[19] Barthelmie RJ, Jensen LE. Evaluation of wind farm efficiency and wind turbine intensity and atmospheric stability on power deficits due to wind turbine
wakes at the Nysted offshore wind farm. Wind Energy 2010;13:573. wakes at horns rev wind farm. Wind Energy 2012;15(1, SI):183e96.
[20] Mortensen N, Heathfield D, Rathmann O, Nielsen M. Wind atlas analysis and [50] Cleve J, Greiner M, Enevoldsen P, Birkemose B, Jensen L. Model-based analysis
application program: WAsP 10 help facility. Roskilde, Denmark: Risø National of wake-flow data in the Nysted offshore wind farm. Wind Energy 2009;12:
Laboratory for Sustainable Energy, Technical University of Denmark; 2009. 125.
 J. Herp et al. / Renewable Energy 81 (2015) 173e181 181

[51] Wang N, Johnson K, Wright A. Comparison of strategies for enhancing energy [53] Mikkelsen T, Angelou N, Hansen K, Sjo € holm M, Harris M, Slinger C, et al.
capture and reducing loads using lidar and feed forward control, control A spinner-integrated wind lidar for enhanced wind turbine control. Wind
systems technology. IEEE Trans 2013;21(4):1129e42. Energy 2013;16(4):625e43.
[52] Harris M, Hand M, Wright A. Lidar for turbine control. Tech. Rep. TP- [54] Florian S, Christophe B, Alain D. A fast atmospheric turbulent parameters
500e39154. Golden, Colorado, USA: National Renewable Energy Laboratory; estimation using particle filtering. application to lidar observations. J Phys
2006. Conf Ser 2011;318(7):072019.