2017 American Control Conference

Sheraton Seattle Hotel

May 24–26, 2017, Seattle, USA

A Tutorial on Control-Oriented Modeling and Control of Wind Farms

S. Boersma1 , B.M. Doekemeijer1 , P.M.O. Gebraad2 , P.A. Fleming3 , J. Annoni3 ,
A.K. Scholbrock3 , J.A. Frederik1 , and J-W. van Wingerden1

Abstract— Wind turbines are often sited together in wind

farms as it is economically advantageous. However, the wake
inevitably created by every turbine will lead to a time-varying
interaction between the individual turbines. Common practice
in industry has been to control turbines individually and
ignore this interaction while optimizing the power and loads
of the individual turbines. However, turbines that are in a
wake experience reduced wind speed and increased turbulence,
leading to a reduced energy extraction and increased dynamic
mechanical loads on the turbine, respectively. Neglecting the
dynamic interaction between turbines in control will therefore
lead to suboptimal behaviour of the total wind farm. Therefore,
wind farm control has been receiving an increasing amount
of attention over the past years, with the focus on increasing
the total power production and reducing the dynamic loading Fig. 1. Part of the Gemini offshore wind farm located in the Netherlands.
Picture taken from http://geminiwindpark.nl/foto-s.html.
on the turbines. In this paper, wind farm control-oriented
modeling and control concepts are explained. In addition, recent
developments and literature are discussed and categorized. This
paper can serve as a source of background information and in maintenance costs, and furthermore it may exclude more
provides many references regarding control-oriented modeling cost-efficient lightweight turbine designs. Both increased
and control of wind farms. power losses and structural degradation increase the overall
I. INTRODUCTION cost of wind energy. Currently, wind farm control typically
relies on greedy control, in which the turbine’s structural
Ever since the industrial revolution at the beginning of the loading and power production are optimized individually.
19th century, the worldwide energy demand has been grow- However, this often appears to be suboptimal for the overall
ing. The main energy production techniques, using fossil- performance of the wind farm [10]. Current research on
fuel reserves (i.e., coal, oil, and gas), have consequences wind farm control aims to provide plant-wide minimization
on the environment and are finite. The International Energy of the cost of energy while taking wake interactions into
Agency’s annual report on energy, the World Energy Outlook account. The goal of this paper is to provide an overview of
2014 [1], predicts fossil fuels to still account for 55% of the current research on control-oriented modeling and control of
total global energy production by 2040, quantifying the sig- wind farms. This paper will discuss background theory of a
nificance of this problem for current and future generations. wind turbine, wakes, control-oriented modeling and control
Fortunately, renewable energy methods are on the rise, and of wind farms and recent developments therein.
their share in energy generation is expected to grow from This paper is, with respect to the literature overview pre-
21% in 2012 to 33% in 2040 [1]. Wind energy is a significant sented in [11], focused on the corresponding flow control
source of renewable energy in several countries (e.g., in problem and discusses recent wind farm research devel-
Denmark, wind power produced the equivalent of 42.1% of opments and field test experiments in more detail. It is
Denmark’s total electricity consumption in 2015 [2]). organized as follows. In Section II, a brief introduction
Significant advances have been made in the area of wind to wind and wind turbines will be given. At the end of
turbine control for many years (see, e.g., [3], [4], [5], [6]). this section, the concept of a wake will be introduced. In
Placing wind turbines in farms, as in Fig. 1, resulted in Section III, wind farm control objectives in terms of per-
new control challenges researched in, e.g., [7], [8], [9]. Still, formance indicators will be presented. Typically, controllers
many challenges remain in the field of wind turbines and are designed and evaluated according to these indicators. In
wind farms. A number of challenges arise from the complex Section IV, control-oriented wind farm modeling will be
interactions between turbines in a wind farm. These interac- discussed. These models can be used for designing and/or
tions often result in decreased power capture and accelerated testing a controller. In Section V, control of wind farms will
structural degradation of the turbine structures. The conse- be introduced and typical wind farm sensors and actuators
quence of accelerated structural degradation is the increase will be discussed. In Section VI, a categorization of wind
1 Delft University of Technology, The Netherlands.
farm control strategies will be presented. In Section VII,
2 Siemens Wind Power, Denmark. a number of field tests for model validation are briefly
3 National Renewable Energy Laboratory, United States. discussed. In Section VIII, conclusions and an outlook will

978-1-5090-5992-8/$31.00 ©2017 AACC 1

Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
be provided. B. WIND TURBINE

II. WIND AND WIND TURBINES There are different types of vertical-axis and horizontal-axis
wind turbines. The most commonly produced and used wind
This section briefly introduces wind energy and single wind turbine is the upwind horizontal-axis wind turbine. One of its
turbine control as it pertains to the challenge of larger wind advantages can be explained by the fact that the blades are
farm control. A more complete and detailed description can always facing fully into the wind, because incoming wind
be found in [12], [13], [14], [15]. This section will end by does not have to pass the turbine tower first (in contrast to
introducing the concept of a wake and its essential character- downwind turbines) or other blades (in contrast to vertical-
istics relevant for wind farm control-oriented modeling and axis turbines). A horizontal-axis wind turbine consists of a
control. rotor, most often with three rotor blades, that is attached
A. WIND to the generator through a drivetrain. The generator and
drivetrain are housed in the nacelle, which is supported by a
Wind is the source of energy exploited by a wind turbine. tower. See Fig. 3 for a schematic representation of the main
Wind flows are mainly caused by the Earth’s rotation and wind turbine components.
thermal heating of the Earth’s surface by the sun, hence
wind is ubiquitous. However, its force is not everywhere
nacelle
equivalent. The behavior of wind at a specific location and gearbox
for a certain time instant can be characterized by a direction generator torque

and magnitude. The process of energy extraction by turbine

rotors can better be understood by looking at the energy
extracted from the wind flowing through a thin disk (see
Fig. 2), with this disk being equivalent to the rotor swept pitch

area. yaw
generator

blade

tower

Fig. 3. Horizontal-axis wind turbine with labeled main components and

control variables. Figure adapted from [13].

Fig. 2. Flow with velocity U [m/s] through a rotor disk with rotor swept
area A [m2 ]. Figure adapted from [12]. The rotor blades convert the momentum of a wind field
passing the rotor plane into aerodynamic forces that drive the
rotor. The drivetrain transfers the aerodynamic torque from
From the continuity equation of fluid mechanics, the mass the rotor to the generator shaft, either directly (direct drive)
flow of air is a function of air density ρ [kg/m3 ], surface or through a transmission (gearbox). The generator converts
area A [m2 ], and flow velocity U [m/s]. Assuming the latter rotational kinetic power into electrical power by generating a
is uniform across the rotor swept area, A, the mass flow of reactive torque on the shaft. To control the power production
air dm
dt through a rotor disk is defined as and forces (torques) on the wind turbine, a number of degrees
dm of freedom (control variables) are typically available:
= ρAU. (1)
dt • Blade pitch (θ) - The rotor blades can rotate, with their

The instantaneous kinetic power of the wind available at axis of rotation aligned with the blades, using hydraulic
surface A, Pw [W], is calculated by actuators or servo pitch motors. Pitch control can be used
to influence the power capture (see, e.g., [17]) and the
1 dm 2 1 loads (see, e.g., [18], [19], [20], [21]) experienced by the
Pw = U = ρAU 3 . (2)
2 dt 2 wind turbine.
Note that the power expression depends linearly on the rotor • Generator torque (τg ) - The generator converts mechanical
disk area, A, (and thus rotor radius squared), and on the wind power into electricity. Torque control is used to control the
velocity, U, cubed. This implies that relatively higher gains power capture.
in power generation can be achieved by placing turbines at • Yaw (γ) - The nacelle can rotate, with the axis of rotation
locations with high wind velocities. aligned with the tower, using a yaw motor. The yaw angle
However, a wind turbine cannot extract all this available is defined as the angle between the axial rotor axis and
power from the wind, as the flow is required to still have the incoming wind direction. In single turbine control, yaw
velocity behind the rotor. The theoretical limit for energy control is often used to set the rotor plane perpendicular
extraction by a rotor is determined by the Betz limit [16]. to the incoming wind direction to increase the turbine’s
This limit will be, i.a., discussed in the following section. power capture.

2
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
The control variables are shown in Fig. 3 with a number that, by using momentum theory for an ideal rotor, the thrust
of basic components of a wind turbine. With these control coefficient, CT , and power coefficient, CP , can be written as:
variables we can optimize the performance of a single wind
CT (a,γ) = 4a(cos(γ) − a), CP (a,γ) = 4a(cos(γ) − a) 2 , (6)
turbine, such as produced power, P, and turbine loading.
An uncommon, and for now more scientifically interesting, for 0 ≤ a ≤ 21 and the yaw angle, γ. The parameter, a, is
control variable is the tilt angle of a turbine. This is defined called the axial induction factor of a wind turbine. It is the
as the difference of the wind angle of attack and the nacelle ratio of the difference between U∞ and the wind velocity at
angle, with respect to the horizontal plane. In current wind the rotor Ur to U∞ , and is defined as:
turbines, this tilt angle is fixed. U∞ − Ur
A wind turbine exerts a force on the wind flowing through a= . (7)
U∞
the rotor. This thrust force represents the amount of energy
extracted from the flow and can be described by The axial induction factor is thus a measure of the decrease in
wind velocity behind a wind turbine and provides a relatively
1 2 simple expression for coordinated control of wind turbines.
F = CT (θ,λ,γ) ρAU∞ , (3)
2 Note that this factor, or more precisely, Ur , can be controlled
with U∞ [m/s] as the free-stream wind velocity and using the generator torque and blade pitch angle, but is also
CT (θ,λ,γ) as the dimensionless thrust force coefficient, influenced by the yaw angle.
which is a function of the tip-speed ratio, λ, blade pitch, It was already stated that even a perfect wind turbine
θ, and yaw angle, γ. The tip-speed ratio is defined as the cannot fully capture all of the available power in the wind.
ratio of the tangential speed at the blade tip to free-stream There is a theoretical maximum that can be extracted by
wind velocity: a turbine. This maximum can be obtained by calculating
ωR the supremum of CP (a,γ), given in (6), among the axial
λ= , (4) induction factor. It can be found that for any wind turbine,
U∞
the induction factor that results in the maximum power
with R the rotor radius and ω the rotor rotational speed. extraction is a? = cos(γ)/3, which translates to a theoretical
The tip-speed ratio is directly influenced by the rotor speed, limit of CPmax = 16/27 cos3 (γ), which is approximately 0.6
which is influenced by the generator torque or by changing if γ = 0. This theoretical maximum is called the Betz limit.
the pitch angle to change the lift forces on the rotor blades. In a practical sense, the maximum power coefficient for
The generator torque control loop is relatively fast because horizontal-axis wind turbines lies around 0.45 according
the system is manipulated at the electrical level, though to [13]. The maximum force can be found in a similar
changes in the rotor speed itself are not that fast due to way: for a = 1/2, the wind exerts the maximum force on
inertia, especially for large rotors. Although the blade pitch the wind turbine. Note that empirical data published in [22]
control loop is slower than the torque loop, it is still relatively revealed that the thrust coefficient expression given in (6)
fast because of powerful motors that typically can achieve is not accurate when a > 1/2. A possible correction based
up to a 10 [deg/s] blade pitch rate for a utility-scale wind on empirical data has been proposed in that paper. This
turbine. correction is based on the Glauert empirical relation between
The power in the wind across a rotor was given in (2). the thrust coefficient and axial induction.
Although power production can be improved using control, A more detailed representation of the rotor than the ADM is
not all the power in the wind can be extracted by a wind the actuator line model (ALM), which represents each blade
turbine. The wind power available for extraction by a turbine individually in the flow, as a distribution of forces along a
is given by: rotating line.
1 3
P = CP (θ,λ,γ) ρAU∞ , (5) Operating regions:
2
For single wind turbines, different operating regions can be
where CP (θ,λ,γ) < 1 is the dimensionless power coefficient distinguished. Each region has its own control strategy and
and the ratio of generated power by the wind turbine to the is typically determined based on a generator speed feedback
available power in the wind (see (2) and (5)). There are many signal. The ideal power curve for a variable pitch/speed wind
models in literature that provide expressions for the thrust turbine is shown in Fig. 4. In addition, a wind power curve is
and power coefficient. One popular way to get an expression depicted and the ratio between this curve and the power curve
for the force and power coefficients is by exploiting the is defined by the power coefficient. The ideal power curve
momentum theory developed in the 19th century by W. J. exhibits three main regions with distinct control objectives.
M. Rankine, A. G. Greenhill, and R. E. Froude. R. E. Froude, In Region II, the control problem can be seen as a tracking
D. W. Taylor, and S. Drzewiecki combined momentum problem, whereas in Region III, the control problem can be
theory with blade element theory, which resulted in the blade seen as a disturbance rejection problem.
element model (BEM) for calculating the forces that a blade
exerts on a flow. When these forces are then converted into a C. WAKE
disk of distributed forces that model the rotor, this is referred As a wind turbine extracts energy from the wind, it causes a
to as the actuator disk model (ADM). In [12], it is explained change in the wind flow downstream from the wind turbine.

3
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
 Rated Speed • Vertical wind shear, which is the change of wake proper-
100 ties with height, typically an increase of wind speed with
height because of ground friction.
• A kidney-shaped wake, as a result of a yawed turbine [29].
Power Output
(% of Rated Output) Wind Power Note that the external atmospheric properties also have a
critical impact on wakes and their propagation, and thus,
i.e., land-based and offshore wind turbines develop different
wakes. Fig. 5 illustrates a horizontal slice of the wake at
0 I II III turbine hub height with γ = 30◦ . The contour plot with
Wind Speed (m/s)
normalized velocities is obtained from wind tunnel data.

Cut -in Speed Cut -out Speed

Fig. 4. Typical wind turbine power curve. Figure adapted from [15].

The altered flow is called the wake of a turbine. The wake

characteristics are space-, time-, and parameter-dependent.
Fig. 5. A time-averaged, stream-wise wind velocity contour plot at hub
A wake is space-dependent because, e.g., far downstream of height obtained from wind tunnel data. The center of the wake is shown in
a turbine, it is different from the wake closer downstream filled magenta circles. Figure taken from [29].
of the turbine. The wake is also time-dependent because the
operation of a wind turbine changes over time as well as the
surrounding flow. Finally, a wake is parameter-dependent, Using momentum theory and assuming γ = 0, a lower bound
as the external variables (such as temperature) influence the on the wind velocity, U_, and a wind velocity at the rotor,
behavior of the wake. It should be clear that studying and Ur , can be estimated as
modeling a wake is a broad research topic by itself and
ongoing [23], [24]. Models range from low to high fidelity, U_ = U∞ (1 − 2a), Ur = U∞ (1 − a). (8)
where the latter describes the wake in more detail and tries to
capture more of its characteristics than the former. However, As stated before, it is through the wake that an upwind
this increase in precision will result in higher computational turbine can influence the performance of downwind turbines.
costs. A more complete discussion on different wake models The key objective of wind farm modeling and control is to
is presented in Section IV. take these interactions into account and use control variables
Typical characteristics of a wake and its main causes are: to ensure a specific level of performance. One option is to
• Wind velocity deficit, as a result of the turbine’s energy capture the nonlinear stochastic behavior in a mathematical
extraction. model and then use this model to design a controller that
• Increased turbulence intensity, as a result of i.a., the guarantees a performance. The assumption is that, when
turbine blade’s rotation. applying this controller to the real wind farm, equivalent
• Wake recovery, which is the phenomenon that downwind performance will be achieved as predicted by the model used
a wind turbine, i.a., the wind velocity recovers to the free- for controller design. This assumption is based on the validity
stream velocity due to mixing. of the used model. In this approach lies one of the main
• Wake meandering, which is a large-scale stochastic phe- challenges in wind farm modeling and control: understanding
nomenon of a wake in which the entire wake structure will wake behavior and capturing the important dynamics of wake
show horizontal and vertical oscillations over time, rather interactions. An open question is: which wake dynamics
than maintaining a certain fixed position and shape [25], are important for a control-oriented wind farm model. A
[26]. subsequent challenge is controller design for the identified
• Wake expansion, which occurs with distance from the model. Most standard controller synthesis methods known
turbine and can be explained using the law of mass in literature are based on linear state-space models. When
conservation and the assumption of flow incompressibility. dealing with nonlinear and stochastic systems, control design
It can be shown that a decrease in velocity means a techniques are less available, and optimal performance can
proportional increase in the wake’s cross-sectional area (in these cases) not be ensured because of possible local
(see e.g., [14]). minima.
• Wake deflection, which is the phenomenon that the com- Another option is to find an optimal control policy following
plete wake is diverging in the latitudinal direction from a model-free approach. Both model-based and model-free ap-
the rotor center because of blade rotations [27] or the fact proaches will be discussed later in this paper. In Section III,
that the rotor is not oriented perpendicular with respect to an introduction to wind farms will be given. This paper will
the incoming wind, i.e., a yawed or tilted turbine. then discuss control-oriented modeling and control of a wind
• Wake skewing, as a result of veer [28]. farm in Section IV and Section V, respectively.

4
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
 III. WIND FARM: MOTIVATION AND pass in and out of the wake [35], [36]. This imbalance
CHALLENGES can contribute to an accelerated structural degradation of
waked turbines.
The previous section gave a brief introduction on wind
energy and single turbine control and ended by introducing Wind farm control consists of finding control inputs using
the concept of a wake. This was defined as the changed measurements to increase the performance of a wind farm,
downstream flow caused by a wind turbine (see Fig. 6) and thus improving quality or minimizing the cost of wind en-
can result in interactions between wind turbines. It was stated ergy. The latter can of course be carried out by increasing the
that wind farm control aims to take these interactions into spacing between turbines, though this may have a negative
account while ensuring wind farm performance. This section impact on the aforementioned advantages, such as reduced
follows by discussing reasons why it is interesting to study deployment costs of the electricity grid. Also, obtaining the
wind farms, and also the related challenges. required spacing is an increasing challenge as rotor sizes
grow with the newer turbines [37]. Next, the objectives
and corresponding challenges in wind farm control will be
discussed.

A. OBJECTIVES OF WIND FARM CONTROL

In this section, the two most common wind farm performance
indicators will be discussed. In general, the goal of wind
farm control is to minimize the cost of wind energy. This
can be translated into a number of technical objectives,
namely maximizing power production, minimizing structural
degradation, and active power control (APC). APC provides
grid services, such as frequency control and power reference
tracking, and its objective is to improve the quality of
wind energy. It will not be discussed in this paper, though
interested readers can find related information in [38], [39],
Fig. 6. Horns Rev offshore wind farm with, normally invisible, wakes. [40], [41], [42]. The power production and load performance
Source: Christian Steiness. See also [30] for more information on the picture. indicators will be discussed in this section.
Power production maximization:
Placing turbines together has a number of benefits, which Wind turbines extract momentum from the flow, which
are, i.a.: results in the previously explained velocity deficit in the
• Reduced deployment costs of the turbines wake. The amount of this deficit limits the power production
• Reduced deployment costs of the electricity grid of downwind turbines, but can be controlled using the wind
• Reduced operation and maintenance costs farm control variables that will be discussed in Section V-
• Reduced land use and impact. A. In [43], the authors show that, when considering a
Especially in densely populated countries such as the Nether- perfectly aligned two-turbine case, the power loss of the
lands, deploying turbines individually is unfeasible, with downwind turbine scales approximately linearly with the
governments often investing in both land-based and offshore spacing. Losses range from around 25% for radially aligned
wind farms. However, grouping turbines together in farms turbines spaced 16 rotor diameters apart to 80% when the
also introduces a number of complications that often sig- aligned turbines are placed 4 rotor diameters apart. The study
nificantly affect their performance. These complications can in [44] reports a power production loss of 12%, averaged
impact downstream turbines as follows: over different wind directions, in an offshore wind farm as
• Because of the wind velocity deficit in the wakes of up-
a result of wake effects.
stream turbines, the downstream turbine will capture less It is important to note that results like these are in general
power than when operated in free-stream conditions [31], obtained using a specific mathematical model trying to
[32]. capture the wake dynamics for specific atmospheric con-
• As a result of increased turbulence in the wake, fa-
ditions. Outcomes can differ according to the model and
tigue loads on the downstream turbine can increase (see, method used. However, wake loss predictions have also been
e.g., [33], [34]), thereby shortening its lifetime in the measured in real wind farms. Wind farm control can mitigate
absence of control algorithms that take this turbulence part of the wake losses, although given the variable nature of
increase into account. a wake, it is still a point of research to quantify how much
• In most cases, the center of the wake will not coincide with
wind farm control can reduce wake losses exactly.
the center of a downstream rotor. This can be caused by Load minimization:
wake meandering, deflection, and wind direction (mostly). A wind turbine structure has been designed to withstand
Because of this, there is more thrust on one side of steady loads several times larger than nominal loads [45],
the rotor, leading to large cyclic variations as the blades and so it is necessary to study fatigue loading with respect

5
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
to the lifetime of a wind turbine. In [46], it is stated that IV. CONTROL-ORIENTED WIND FARM
modern wind turbines are fatigue-critical machines, i.e., MODELING
the design of many of their components is dictated by The advancements in wind farm control have gone hand in
fatigue considerations. The authors in [47] also conclude that hand with advancements in wind farm modeling, as typically
mostly dynamic loads are responsible for fatigue and reduced modern control algorithms rely on an internal model. These
lifetime of wind turbines in wind farms. In these papers, models are often simple and relatively computationally in-
different loading models were used, hence it is important to expensive. We refer to these types of models as low fidelity
first investigate which type of loading occurs. The three most (possibly parametric) models. High fidelity simulation mod-
important sources for the loading of an upwind horizontal- els are typically used to assess a controller’s performance as
axis wind turbine are [14]: the last step before being put to the test on an actual wind
farm. These models are more accurate, but also significantly
• Gravitational loading
more computationally time consuming, and can therefore
• Inertial loading
not be employed for real-time control. Although wind farm
• Aerodynamical loading.
models are different, two main components can always be
The first type of loading is caused by the gravitational field distinguished:
of the Earth and rotation of the blades. It is clear that a • Turbine model: These models predict the interaction be-
blade rotating downward experiences different forces than a tween the flow and the turbine structure. Additionally,
blade rotating upward. It causes a sinusoidal loading on the structural loads on the turbine given the incoming flow
blades with a frequency corresponding to the rotor rotation field may be predicted, which can include extreme loading,
of once per revolution (1P). Inertial loading occurs when vibrational modes, and fatigue.
the wind turbine changes the rotation speed. Certain parts • Flow model: A model that predicts the flow properties in
on the blades experience different changes that will result in a wake or of the total flow field in a wind farm.
inertial loading. Another source is the centrifugal force acting A turbine model gets a flow field from a flow model as an
on the blades. Aerodynamical loading is caused by the flow input, whereas the turbine loadings are inputs to a flow model
passing the wind turbine and varies in space and time. For that indicates the unavoidable interconnection between the
example, a wind field contains a velocity profile with a bigger two submodels. The two types are described next.
magnitude that is relatively high from the ground because of
A. TURBINE MODEL
shear effects, whereas the turbulent effects introduce time-
varying behaviour in a wind field. Also, according to [14], Wind turbine models describe the flow effect on the turbine
the yaw (and tilt) angle of a wind turbine causes additional structure, including loading and vibrations. A flow field
aerodynamical periodical forces on a wind turbine. In a serves as an input with which the turbine model evaluates
wind farm, a wind field will also be perturbed by wind the resulting loading. Two models traditionally used for
turbines causing changes in a wind field as highlighted at estimating aerodynamic loading are the ADM and ALM,
the end of Section II. Downwind turbines in a wind farm both introduced in Section II. These models can predict
can then experience a changing wind field over the rotor turbine flow interactions and provide estimations of the
that can introduce additional aerodynamical loading. Loading turbine’s power capture and forces exerted on the flow. A
can, in the end, lead to fatigue damage and breakdowns. more elaborate turbine model is FAST [49], developed by the
There are different measures of fatigue loading, such as National Renewable Energy Laboratory (NREL). It contains,
the rainflow counting, spectral, stochastic, and hysteresis i.a., the ALM and takes into account, given an incoming flow
operator method. This paper does not cover these methods, field, all of the three types of loading discussed in Section III-
but the interested reader is referred to [48]. A. DEL values can be determined and the lifetime of a
turbine can be assessed. Other turbine models exist, such as
The purpose of single turbine control is to mitigate the effects
HAWC [50], but will not be further discussed in this paper.
of gravitational, inertial, and aerodynamical loading. On a
By using models such as these, accurate predictions can be
wind farm control level, it is more important to focus on
made on the (extreme and fatigue) forces, moments, and
the effects of the changed aerodynamical loading caused by
vibrations of a turbine structure for given wind conditions.
the upwind wind turbines in the farm. Damage equivalent
Also, these models provide accurate predictions of power
load (DEL) is a measure that is commonly used in literature
capture of the turbine at given inflow conditions. It should
to quantify loading, and allows for direct quantitative com-
be clear that more advanced turbine models require relatively
parisons of different loading types on the turbine structure.
more computation time. An overview of the components
DEL defines the equivalent fatigue damage caused by a load,
generally present in such turbine models can be found
taking into account the fatigue properties of the material.
in [51].
In this section, two wind farm performance indicators were
introduced. Wind farm control aims to optimize these indica- B. FLOW MODEL
tors. For synthesis and evaluation of controllers, wind farm It was previously stated that the dynamical behaviour of a
models are typically used. This will therefore be the topic of wake (or more general, a flow) is governed by the three-
the following section. dimensional (3-D) unsteady Navier-Stokes equations. These

6
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
equations are mathematically defined as a nonlinear infinite one-to-one quantitative comparisons after, e.g., changing a
dimensional system with equality constraint. Under boundary control policy.
conditions (inflow conditions) and forcing terms (the wind The authors in [56], [57], [58] present more control-oriented
turbines) typically used in a wind farm model, and without and relatively less computationally expensive wind farm
making significant assumptions, no analytic solution has models based on the unsteady 2-D Navier-Stokes equations
been found yet for these equations. Hence, in such a case, following a LES approach. It is attempted to solve the set of
it is impossible to solve the governing equations directly. discretized equations governing the wake and wind turbines
Computational fluid dynamics (CFD) is a branch of fluid directly, without model reduction nor any assumptions other
mechanics that uses numerical analysis and algorithms to than incompressibility. The number of states in these models
solve and analyze this type of problem. can easily be 103 or more, which makes it challenging to
Spatial discretization is a method that is applied to obtain a use them for controller design. A second challenge using
set of solvable equations. Because turbulence exists on many this approach is the choice of a (relatively simple) turbu-
different temporal and spatial scales in a wind farm, the most lence model, which should be included to account for wake
accurate way to simulate turbulent flows is to directly solve recovery. In [58], the authors include a simplified mixing-
the obtained set of equations on a very dense grid, capturing length turbulence model to create wake recovery behind a
all eddy scales. This method is referred to as direct numerical turbine, whereas in [56], no turbulence model is included. In
simulation (DNS). It is computationally expensive because, these dynamic wind farm models, the turbines are modeled
after spatially discretizing, the dimensionality of the obtained using the ADM. The cost of solving these wind farm models
set of equations is huge as a result of the fact that every is relatively low because of the exploitation of sparsity and
cell in the wind farm has its own Navier-Stokes equations. structure in the system’s matrices.
Large-eddy simulations (LES), on the other hand, resolve the Another approach is using simplified versions of the gov-
governing equations (after spatially or temporally filtering erning equations. For example, in the 2-D Ainslie [59] and
the Navier-Stokes equations) on a coarser mesh (capturing 2-D dynamic wake meandering (DWM) model (also called
only the large-scale eddies), but can approximate the smaller- the Larsen model) [60], assumptions are made such that the
scale eddies with subgrid models. Small-scale turbulence is Navier-Stokes equations can be approximated with a thin
then calculated within each coarse cell using this subgrid shear layer approximation that is less computationally ex-
model. Most wind farm flow solvers that are considered as pensive. Currently, NREL is developing FAST.Farm, which
high fidelity models employ this method. extends the DWM model to include more control-relevant
Less computationally expensive models are also present dynamics [61]. WakeFarm (also referred to as Farmflow),
in literature. Most of these models consider a two- developed at Energy research Centre of the Netherlands
dimensional (2-D) space to reduce the model complexity (ECN), simulates the wind turbine wakes by solving the
and assume incompressibility of the flow, and only have steady parabolized Navier-Stokes equations in perturbation
a simplified turbulence model to induce wake recovery. In form in three dimensions [62], [63]. When applying time
addition, parametric models exist that only estimate specific averaging on the Navier-Stokes equations, the Reynolds
characteristics of a wake, such as velocity deficit and wake Averaged Navier-Stokes (RANS) equations can be obtained.
deflection. This paper will continue giving a brief overview With this approach used in, e.g., [64], a time-averaged
of some wind farm models that exist. (mean) flow is computed and the effects of turbulence
are implemented using the mixing-length hypothesis. The
computational cost for using RANS equations in a wind
C. EXAMPLES
farm model will also be computationally less expensive than
Wind farm models that use LES flow models include Simu- for high fidelity flow solvers. A combination is presented
lator fOr Wind Farm Applications (SOWFA) [52] and UTD in [65], in which the authors present a RANS wind farm
Wind Farm (UTDWF) [53], a wind farm model developed model for which model parameters are updated online using
at UT Dallas, and SP-Wind (Leuven) [54], and PArallelized the high fidelity flow solver UTDWF. The authors in [66]
LES Model (PALM) [55]. These 3-D, high fidelity flow present a, with wind tunnel experiment data validated, wind
solvers contain, in general, sophisticated wind turbine models farm model based on simplified RANS equations. The sim-
and 106 or more states. The resulting computation time plification results in the approximate governing equations
can be on the order of days or weeks using distributed upon which an inexpensive analytical model is built. A
computation. It should be clear that these types of models completely different dynamic wind farm model is presented
are not useful for online control, wherein measurements in [67] where the Navier-Stokes equations are solved using a
are fed into a controller that calculates optimal actuator semi-Lagrangian approach. The interested reader is referred
settings based on an internal model in real time. However, to [68], [69] for more background information on the Navier-
these models can serve as analysis tools. The cost of doing Stokes equations and its varieties.
simulation experiments using these solvers is significantly One way to circumvent the complexity of wake modeling
less than the cost of doing experiments on a real wind is by using 2-D parametric models. The idea is to capture
farm. Moreover, simulation experiments can be done in only the most dominant wake characteristics. Most of these
controlled atmospheric conditions, which is important for parametric wake models estimate a steady-state situation for,

7
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
i.a., a given inflow direction. If the wind farm is large,
A CLASSIFICATION AND PROPERTIES OF DIFFERENT MODELS .
this inflow direction should then hold for the whole farm, Table 1. Wind farm models overview.

which can be an unrealistic assumption. Examples are the Low fidelity Medium fidelity High fidelity

Frandsen model [70], the model presented in [71], and Model

type
Kinematic models Flow field models Flow field models

the Jensen Park model [72], [73], which predict a linearly Funda- Parametric 2D NS 3D NS
expanding wake with a velocity deficit that only depends mentals

on the distance behind the rotor. Extending Jensen’s model Models

Jensen,
FLORIS,
FLORI- DWM, WFSim, SOWFA, WakeFarm,
resulted in the parametric model called FLOw Redirection Frandsen, ...
Dyn,... Ainslie... UTDWF, SP-Wind,...

and Induction in Steady-state (FLORIS) [74]. A dynamical Flow 2D 2D 3D

dimension
version named FLOw Redirection and Induction Dynamics
Dy-
(FLORIDyn) of this is presented in [75] and a similar model namic/Static
Static Dynamic Dynamic

is SimWindFarm [76], wherein relatively simple dynamical Turbine

ADM
ADM/ALM and/or an aerodynamic package (e.g.,
model FAST)
equations are used to estimate the velocity deficit in a
Comp. Order of seconds on a desktop Order of minutes on a Order of days on a
wake. Interestingly, recent findings have shown that these effort PC desktop PC cluster of 102 CPUs
simple parametric models, such as the Jensen model and Model Low – medium Medium – high High – very high
models based on the Jensen model like FLORIS, can in
accuracy

some cases predict wake losses accurately, if uncertainty is

included in the calculation [77], [78], [79], [80]. Inclusion of
uncertainty in wake models, and evaluating controllers based
applied to a wind farm model are [88], [89]. In [88] and [90],
on uncertain wake models, is an active field of research.
the authors illustrate that by using data from a high fidelity
Note that never all flow behaviour will be captured when flow solver and dynamic mode decomposition, a low-order,
simulating a wind farm using a model, especially when two-turbine wind farm model can be obtained in which states
employing a 2-D model. For example, in the latter case, retain a physical interpretation.
the inflow from above and below is not taken into ac- Note that model order reduction techniques rely on specific
count, even though it influences wake properties. In addition, operating conditions, i.e., they provide linear models for
the underlying assumption of infinitely tall turbines in 2- a specific operating point and are only valid within small
D models that are based on the unsteady Navier-Stokes deviation from this point. These reduced-order linear models
equations results in flow speedup effects on the right and left can be defined at, e.g., specific wind speeds and directions.
downwind the turbines. Interestingly, the model presented However, techniques in parameter-varying control exist that
in [81] includes information on the third dimension in the can help link these models together [91]. Also, reduced-order
2-D unsteady Navier-Stokes equations, effectively reducing models could be data-driven in a more system identification
this undesired effect. However, for some specific cases, 2-D approach (see e.g., [92]) or model-based considering, e.g.,
wind farm models have been validated with high fidelity 3-D the Navier-Stokes equations. Examples of the latter are
models, which hints to the fact that the assumption could be balanced proper orthogonal decomposition and the Galerkin
reasonable. It is the time-reducing property that makes 2-D projection (see e.g., [93]). This paper will not discuss these
models attractive for wind farm control. different techniques.
In addition to the above, the authors in [23], [82], [83], [84], A note for this section is that a model’s accuracy and appli-
[85], [86] provide overviews of rotor blade models and wake cability are highly dependent on the atmospheric conditions
models. In Table I, a classification of previously described of the relevant wind site. It is shown in, e.g., [94] that wake
models is given, noting that in this table the term "fidelity" is characteristics change a lot relatively when the atmosphere is
mainly used to describe the amount of detail described by the stable or not. In addition, in unstable atmospheric conditions,
model. This does not automatically imply that more detailed often wakes are extremely hard to control, and no significant
models are more suitable for (online) control purposes, as improvements can be yielded by active wind farm control. In
discussed before. In addition, it has been shown in several stable conditions, wakes are less difficult to control; however,
wind farm simulation cases that medium and low fidelity the problem is still challenging.
models are able to estimate wind velocity and power data In this section, two main components of a wind farm model
from a high fidelity model. The acronym NS stands for the were discussed and examples of wind farm models were
Navier-Stokes equations. given. These models can be used to conduct wind farm
Reduced order models: Performing model order reduction analysis or control. The latter will be the topic of the
techniques, such as proper orthogonal decomposition and following section.
dynamic mode decomposition on high fidelity flow solver
V. WIND FARM CONTROL
data, is another method to obtain a model. The authors
in [87] illustrate that it is possible to apply proper orthogonal It was stated earlier that wind farm control is aimed at
decomposition and compare the flow fields obtained with optimizing the previously presented objectives: minimization
the low-order model with that of a high fidelity flow solver. of power losses and structural loading. More precisely, wind
Other articles that deal with proper orthogonal decomposition farm control aims to find control actions that increase the

8
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
wind farm performance by taking measurements (and pos- turbine is in operation, these measurements are disturbed
sibly an internal model) into account. By (partially) relying because of the interactions between the flow and tur-
on measurements, a controller can cope with changing en- bine rotor, especially at small distances around the rotor
vironments. In this section, wind farm actuators and sensors plane [10].
will first be discussed, followed by the discussion of two • Power sensors.
wind farm actuation methods. The categorization of different • Strain sensors, which measure structural deformations.
control strategies will be covered in Section VI. • Accelerometers, which measure the turbine’s acceleration.
• Generator shaft-speed sensors.
A. ACTUATORS AND SENSORS • Torque sensors.
In wind farm control, measurements from sensors and/or • Temperature sensors, which are used for anti- and de-icing
possibly an internal model are used to compute control techniques [99].
settings. These control settings are assigned to the turbine’s
These turbine sensors are also useful for wind farm control.
actuators, which can be considered the degrees of freedom in
Examples of sensors on a wind farm level are:
the wind farm control problem. In this section, typical wind
farm actuators and sensors will be discussed. • Separate meteorological measurement masts, which are
located in the farm, and provide information on the flow
Actuators: For a single turbine, actuators were defined as velocity for their respective positions.
turbine yaw, γ, generator torque, τg , and blade pitch angles, • Remote-sensing (RS) technologies, which measure the
θ. Tilting the turbine’s rotor provides an additional actuator flow field at different positions upstream or downstream
for control, though this approach has only been used in of turbines, without the need for repositioning the sensor.
simulations until now [27], [95], [96], [97]. In a real wind RS can use sodar, lidar, or radar technology, or satellite
farm and some wind farm models, the control variables are scatterometry.
(γi ,τg i ,θ i ) for i = 1,2,. . . , N, with N the number of turbines
in the farm. However, it is common in wind farm modeling According to [100], sodar systems use sound waves and
to define the axial induction (see (7)) or similarly the thrust are deemed too slow and of too low accuracy for wind
force coefficient (see (6)), and the yaw angles as actuators. farm applications, although they are capable of wind field
Although this approach neglects the dynamics between the monitoring [101], [102]. More recently, lidar technology
physical turbine actuators τg i ,θ i and the axial induction or has been applied, which relies on the same principle as
thrust force coefficient, it simplifies the modeling and control sodar but using laser instead of sound waves [103]. The
problem. Some studies (see e.g., [98]) include a first-order authors in [104] show that sodar and lidar can achieve
time filter to circumvent sudden unrealistic axial induction similar accuracy in field tests on one of the Vattenfall wind
changes in simulations. The following wind farm actuators farms. However, theoretically, lidar is able to achieve higher
can be defined for models employing the ADM: measurement accuracy because of the nature of light [100].
Furthermore, both [103], [105] show that lidar has real
• γi for i = 1,2,. . . , N,
potential to improve the accuracy of current wind speed
• ai (or CTi ) for i = 1,2,. . . , N.
measurements above the resolution of a mast. Also, a lidar
Note that by changing ai , the thrust force, i.e., the amount of device can be placed on top of a wind turbine to measure
energy the turbine extracts from the flow, will change. These upwind or downwind. Because a lidar device is relatively
two variables are illustrated for one turbine in Fig. 7. expensive, it is interesting to investigate how to use it in
an optimal way such that expenses can be minimized. The
authors in [106] present such a study. Interestingly, lidar
technology was initially applied for single-turbine control,
incorporating feed-forward control (see, e.g., [105], [107],
[108], [109], [110], [111]). A project in which the use of
radar in a wind farm is investigated can be found in [112].
At this moment, radar devices are relatively expensive and
large regarding dimension.
The main challenges in RS technology are data outliers
Fig. 7. A schematic representation of wind farm actuators typically used in because of hard targets and interference with the turbine
simulations. The thrust force, F (a i ), is determined by the axial induction blades, and problematic wind field reconstruction due to
a i of turbine i.
the cyclops dilemma. For example, a single lidar system
measures the wind from only one angle of view. Thus, with
Sensors: Proper placement and choice of sensors is key to a single lidar system, it is not possible to reconstruct the full
the success of wind farm control. Examples of wind turbine 3-D wind field without making any assumptions [113]. An
sensors include: example of this can be found in [114], which shows that it
• Anemometers and wind vanes. These devices are mounted is possible to estimate a 3-D wind field using lidar.
on the nacelle to locally determine the wind speed and Given the actuators and sensors, the next question is which
direction at the rotor plane. However, note that when a actuation methods can be used to optimize performance

9
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
within the wind farm. This will be the topic of the following diameter. Hence, the power that is purposely not captured by
section. the upwind turbine will not be captured completely by the
downwind turbines for the case in Fig. 8. It is the deviation
B. ACTUATION METHODS FOR WAKE CONTROL
of the upwind turbine wake which, according to the authors
Currently, most wind farms are operated using individually in [121], is, i.a., determined by:
optimal wind turbine control settings referred to as greedy
control. As stated before, wind farm control consists of • Wind direction
finding control inputs using measurements (and possibly an • Relative position of turbines
internal model) to increase the performance of a wind farm, • Wake meandering
thus minimizing the cost of wind energy. It has been an active • Atmospheric conditions
research topic since the 1990s and relies on the assumption • Wake expansion.
that the performance of a wind farm can be increased by Wind direction is important in two ways. First, changes in
operating turbines in the farm at configurations different the wind direction in the farm contribute to deflection and
from their individual optimal settings. Two general control skewing of the wake, which causes the wake to overlap
methods exist for this purpose: axial induction control (AIC) less with the downwind rotor. Second, the wind direction
and wake redirection control (WRC). Simulation studies is never exactly perpendicular to the rotor, and hence there
such as [115], [116], illustrated that both methods have a will always be a deviation of the upwind turbine wake from
potential to increase the power production and can influence the downwind rotor. Note that especially the latter could
structural loading. Another possible future method is to be captured using uncertainties in the model. Importance of
actively reconfigure the wind turbines in a wind farm with wake expansion is emphasized because the authors in [121]
floating turbines. Wind farm layout optimization can be show that, when using pitch offsets, most power passed
considered as initial work towards such a strategy. This will, by the upwind turbine is located in the outer ring of the
however, not be discussed further in this paper though the wake. This complicates AIC, as it becomes more difficult
interested reader is referred to [117], [118], [119], [120]. AIC for the downstream to capture this energy because of wake
and WRC will be topics of the remainder of this section. expansion. Note that this specific spatial distribution of the
Axial induction control: power cannot be modeled using the standard ADM, but can
The idea of AIC is to reduce the power production of upwind be captured using ALM.
turbines by changing the axial induction so that downwind Another interesting point regarding AIC is that, when the
turbines can generate more. The axial induction is changed thrust force is reduced, the turbulent wake mixing and
by adjusting the blade pitch angles and generator torque thereby wake recovery will be reduced. There are thus two
away from individually optimal settings. AIC is worthwhile counteracting effects: increased velocity in the near wake,
if the reduced power production of the upwind turbines but reduced recovery downstream (effectively decreasing
can be compensated for by the downwind turbines, and if velocity of the far wake). In [121], it is shown that including
performance of a turbine is significantly impacted by an this effect in an engineering model reduces the expected
upstream turbine through its wake, e.g., in situations with power production increase from AIC.
little wake recovery, dense turbine spacing, and relatively Research of AIC is done quite extensively, showing inconclu-
high wake-rotor overlap. Fig. 8 illustrates an aligned two- sive results on its feasibility. Most work in recent literature
turbine situation in which this is not completely the case. only takes power production into account, whereas loading
is neglected. An example is the LES simulation results
cut−through at 3D downstream of turbine
500
cut−through at 4D downstream of turbine
500
cut−through at 5D downstream of turbine
500 presented in [122], wherein power production is increased
400 400 400
using AIC by enforcing quick variations in the thrust force.
altitude z [m]

altitude z [m]

altitude z [m]

300

200
300

200
300

200
These variations will increase turbulence in the wake and
100 100 100 mixing with the upper boundary layer containing a higher
0
−200 0 200
0
−200 0 200
0
−200 0 200
flow velocity, which is beneficial for the power production.
lateral distance to hub [m] lateral distance to hub [m] lateral distance to hub [m]
Subsequent work [123] shows that by constraining thrust
force variations, the power gain will again be reduced. The-
cut−through at 6D downstream of turbine cut−through at 7D downstream of turbine cut−through at 8D downstream of turbine
500 500 500

400 400 400

oretically, it can be possible to increase power, but of course
altitude z [m]

altitude z [m]

altitude z [m]

300 300 300

200 200 200

quick variation of thrust force will have implications on loads
100 100 100 and these should be taken into account. Differentiation of the
0
−200 0 200
0
−200 0 200
0
−200 0 200 results can be made with respect to the used models: steady-
lateral distance to hub [m] lateral distance to hub [m] lateral distance to hub [m]
state [115], [124], [125], [126] or dynamical [122], [127],
[128]. In general, early results based on relatively simple
Fig. 8. A cut-through at different distances of a wake from SOWFA data
(red colored area contains higher wind velocity than the blue colored area).
steady-state parametric models illustrate increases in wind
Figure taken from [9]. farm power production. The simplified models in mentioned
studies might not represent the relevant wake phenomena
Typically, in a wind farm, the distance between up and down- in AIC, and thus it is questionable if the optimized control
wind turbines is around 7D, meaning seven times the rotor settings would work for the atmospheric conditions under

10
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
consideration. High fidelity studies such as [121] and wind the axial induction of the first turbine, it is actuated with
tunnel experiments such as [129] show that it is not always a yaw angle of 25 degrees. The authors conclude that the
possible to increase power by AIC, and this can be explained induced velocity increase caused by yaw actuation is better
by phenomena mentioned earlier. Interestingly, the authors concentrated within the rotor area of a downwind turbine
in [130] show that, although it seems that the power pro- placed more than 3D behind the upwind turbine. Wake
duction cannot be increased, it can be interesting to employ behavior as a result of yaw actuation is an actively researched
AIC to reduce turbine loading while maintaining equivalent topic [29], [133], [66]. The developments in WRC as an
power production. In addition, AIC can possibly be used in actuation method go hand in hand with these studies and
APC. more details regarding this method need to be investigated
It is still difficult to make conclusive statements on AIC. using simulation and field studies.
Perhaps a solution lies in adjusting the structural design of An interesting but rarely seen approach is to use both AIC
wind turbines in a farm according to previously described and WRC [134]. Here, a relatively simple engineering model
phenomena, but this is beyond the scope of this paper. In to capture wake dynamics is used, but its parameters are
conclusion, although the concept of AIC is promising, recent calibrated using one data set from a high fidelity flow solver.
advances in wind farm modeling and wind tunnel and field The AIC and WRC analysis is done for different wind
tests have shown that possible production gains may be directions. It is shown that a power increase is achieved for
smaller and more difficult to harvest than initially expected all studied wind directions using the proposed approach. The
based on static control strategies and more simplified models. obtained control settings are not tested on a real wind farm
Further research is needed to conclude whether it is possible nor a high fidelity flow solver.
to find a wind farm controller that will use AIC to reduce In this section, we discussed the most common wind farm
loads and/or increase production by dynamically adjusting actuation methods. Section VI will discuss possible control
pitch and torque settings to atmospheric conditions. strategies.
Wake redirection control: VI. WIND FARM CONTROL STRATEGIES
In this approach, the rotor of the upstream turbine is pur-
posely misaligned with the incoming flow to deflect the wake In wind farm control, a supervisory controller determines a
downstream so that it will not at all or partially overlap a collective control policy using measurements (and possibly
downwind turbine. The deflection can be done using: an internal model) so that performance (as defined in Sec-
tion III-A) is achieved. According to this control policy, the
• Tilt actuation
supervisory controller assigns individual control settings as
• Individual pitch control (IPC)
defined in Section V-A to each turbine in the farm. Then,
• Yaw actuation.
relatively simple internal controllers enforce the tracking of
Tilt actuation will not be further discussed in this paper, but this assigned turbine setting. In this closed-loop approach,
the reader is referred to [27], [95], [96], [97]. In simulation not only the atmospheric conditions but also quantities such
studies, IPC is shown to be effective at inducing wake redi- as power production and a turbine’s structural loading can
rection, though this results in a large increase in loads [27]. be defined as measurements. Hence, control actions can
Fig. 9 depicts a schematic illustration of yaw actuation. adapt to the changing wind farm and atmospheric properties,
which has the potential to lead to robust control solutions.
wind 8 7 6 5 4 3 m/s Controllers are evaluated using an internal model, which can
speed be dynamic or static. For a dynamical model, the model states
φ can have a physical meaning, such as wind flow velocity, but
it can also be a nonvariable. No system states are present with
(parametric) steady-state models.
γopt A distinction between closed-loop controllers can be made
with respect to the measurements used. In closed-loop state-
feedback, all the states of the model (e.g., flow velocity
vectors or power signals from the turbines) are assumed
y to be measured and fed back to the controller. This as-
sD sumption can be unrealistic, because measuring each system
x
state can be impractical and often impossible depending on
Fig. 9. An illustration of wake redirection control with inflow angle φ
and a second turbine placed s rotor diameters D downstream of the first the used model. In closed-loop output feedback, only the
turbine. Figure taken from [36]. measurements, e.g., a subset of the states, are fed back to
the controller and used to evaluate control actions. State
Wake redirection promises significant improvements in sim- estimators (observers) can be used to estimate the system
ulation with power production increases on the order of 4%- states using only measurements. For example, the state of a
7% [11] and an annual energy production increase on the model can contain all flow velocities (or a linear combination
order of 3%-4% [131]. In [132], a similar simulation is done of these velocities) in a wind farm, whereas the output may
as previously presented in Fig. 8, but instead of changing be only the flow velocity at hub height of the rotors. An

11
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
observer (discussed in Section VI-C) can estimate all flow sented. In [125], AIC using gradient-based ESC (therein
velocities using only these few measured flow velocities at defined as maximum power-point tracking (MPPT)) while
the rotors. Different closed-loop control strategies and their having information only from neighboring turbines is applied
applications to wind farms will be discussed next. to maximize the power output of a wind farm for different
atmospheric conditions. An extended Jensen Park wind farm
A. OPTIMIZATION-BASED CLOSED-LOOP CONTROL model is used and a benchmark power production is obtained
In this strategy, wind farm measurements are fed into a using GT. The results illustrate that, by using gradient-based
controller. Here, an optimization procedure evaluates, using ESC, the power production can be improved with respect
an internal model, optimal control inputs such as yaw angles, to the benchmark results. The optimal control inputs are
pitch angles, and generator torque (or axial induction) values not tested on a wind farm nor a high fidelity model, hence
for the turbines in the farm. In addition, the model parameters results depend on the validity of the model used. In this
can be updated using the wind farm measurements. Then, case study, the information for the individual wind turbines
optimal control inputs are sent to the turbines in the farm is also limited, hence a global optimum cannot be guaranteed,
and new measurements are taken. but the computation time is reduced. In [141], AIC and
An algorithm that can be used for finding optimal inputs WRC using a similar gradient-based ESC algorithm as [125]
is game theory (GT). Here, favorable actions lead to high is applied in a wind tunnel with power maximization as
rewards and unfavorable actions to low rewards. The algo- an objective. The case study includes three turbines with
rithm tries to find the most rewarding action according to the limited information for the individual turbines. Hence, again,
used model. The reward can, e.g., be the amount of power it is observed that a global optimum cannot be guaranteed
or the experienced loading. Because of the random search though a power production increase with respect to a baseline
actions, the algorithm needs time to converge to optimal controller is presented. In [142], AIC and WRC using a
control settings. The duration depends on the complexity Bayesian Ascent method is presented. Simulation and wind
of the internal model, but even if the model is a simple tunnel test results are shown for a four-turbine case. Dynamic
parametric steady-state model and consequences of certain programming is another algorithm also applied to wind farm
control actions can be evaluated quickly, GT needs many models (see, e.g., [143], [144], [145]). The latter aims at
iterations to converge to an optimal solution. If atmospheric optimizing the power production among the yaw angles
conditions in a wind farm change during the search for employing an extended Jensen Park model. These results will
optimal control settings, the algorithm has to start again not be discussed further in this paper.
finding optimal settings for these new atmospheric condi- Note that the optimization-based closed-loop control results
tions. Literature such as [126], [135] illustrate AIC using presented so far, except for [74], [140], [141], [142], are
GT. For specific conditions, power production improvements obtained using a relatively simple model. The control actions
are shown with respect to a baseline controller. However, are not tested in high fidelity simulations nor a real or scaled
relatively simple engineering wake models are used, and the wind farm, and the question is if similar results will be
found optimal inputs are not applied on a wind farm nor obtained when doing so. It is also important to note that
a high fidelity model. It is therefore not clear how these GT and ESC are, in essence, model-free approaches, hence
results would apply to real wind farms. The authors in [136] they could be applied directly on a wind farm. However, this
illustrate AIC and WRC using GT to optimize the power is due to, i.a., wake traveling delays, unpractical hence these
production. Using their approach on an engineering model methods are applied on relatively simple (fast) models. With
results in improvements, though again the control settings are MPPT, only information from neighboring turbines is used,
not tested on a more realistic situation. The authors in [74] which decreases the necessary wake traveling time. This is
apply WRC using GT with FLORIS, a steady-state model the reason why MPPT can be applied in a wind tunnel,
introduced in Section IV. The optimal inputs are then applied and possibly a wind farm, but not GT or ESC. GT, ESC,
to a high fidelity model SOWFA. An increase in power with and MPPT are typically applied using parametric steady-
respect to a baseline controller is presented. state models and optimal control settings are evaluated for
Another approach is extremum seeking control (ESC), an specific atmospheric conditions and in steady state. Because
optimization approach that can work for nonlinear, time- a wind farm most likely never reaches the steady state, it
varying systems. ESC algorithms estimate the gradient of is questionable if the found results are really optimal. In
the cost function (e.g., the total power of a wind farm) using addition, the computation time these optimization algorithms
measurements. In literature such as [137], [138], [139], AIC need to converge remains a critical issue because of the time-
using ESC and a greedy controller are applied on a relatively varying conditions in a wind farm. We therefore encourage
simple wind farm model and the results are compared. The researching methods that can increase the convergence rate
found optimal values are not sent to a high fidelity model of these optimization algorithms.
or real wind farm to validate the results. The authors show Closed-loop control based on a dynamic model has potential
that, for different cases, power production can increase with to find a temporally optimal solution. An example of this is
respect to greedy control. In [140], AIC using ESC is applied presented in [122], [123]. Here, model-predictive control is
on the high fidelity model UTDWF and power production applied using the high fidelity model SP-Wind. Knowledge
improvements with respect to a baseline controller are pre- of all the flow velocities and wind turbine power signals

12
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
 wk
is assumed and the algorithm maximizes the total power
production among axial induction factors for a given time Controller Plant vk

horizon. Computationally, this is a heavy task, but the results

give insights into the possibilities of AIC. The authors uk + + yk

in [45], [47], [128] also present AIC using MPC via a

medium fidelity flow model to reduce the computational
effort. Power increase (and load reduction in [45], [47]) with
respect to a baseline controller are presented, though the ŷk
+
rk −
controller is not tested in a high fidelity model. In all the x̂ k
MPC examples, full state knowledge is assumed. As stated
before, this is in general not realistic. error

B. LINEAR DYNAMIC CLOSED-LOOP CONTROL Fig. 10. A closed-loop wind farm control scheme with an observer.
Examples of these approaches are PID, H2 , and H∞ con-
trollers. These controllers are defined as dynamic controllers
and can be designed using (mostly) linear models. Tracking • Deal with noise in measurements and may act as a low-
behavior and disturbance rejection are time-domain specifi- pass filter in estimating the system states
cations that can be imposed relatively easily on closed-loop • Enrich the state estimation with small-scale flow behavior
systems. in which a control-oriented model is able to estimate the
The authors in [146] did implement a H2 controller using a large-scale flow behavior.
medium fidelity wind farm model that neglects turbulence. The latter property is especially interesting because model
The controller is tested on a nonlinear model and the authors mismatches are most likely to occur as a result of the
conclude that the controller provides a distribution of power dynamic complexity in a wind farm. In [151] and [152], the
references between wind turbines so that demanded wind authors implement a type of observer called the Ensemble
farm power is ensured and structural loading is minimized. Kalman filter using a medium fidelity flow model. Although
The authors claim that their method can also be used to the initial results are promising in simulation with LES data,
evaluate a H∞ controller. Unfortunately, the controller is not no real closed-loop simulations have yet been performed with
evaluated on a high fidelity model. a controller and state observer. The authors show that the
In [147], the authors designed a PID controller for wake flow estimations can be improved using an observer, and
tracking. The controller is applied in SimWindFarm, a model flow fields can better approximate high fidelity flow data
discussed in Section IV. In [148] and [149], a H∞ and a when applying an observer.
robust H∞ controller are designed, respectively, to steer the In [153], the authors use a relatively simple dynamic wake
wake while employing a dynamic wind farm model based on model in an observer while taking measurements from the
the 2-D Navier-Stokes equations. Perfect knowledge of the SP-Wind flow solver. The objective of the control framework
center of the wake using lidar is assumed in both papers. The is power reference tracking, and AIC using a MPC controller
concept of steering the wake to a certain position makes the is applied. The results look promising. Another example of
work in these papers unique. However, the question remains applying an observer in a wind farm simulation can be found
as to which position the wake should be steered to increase in [154].
wind farm performance as discussed in Section III-A. Open questions regarding the application of an observer in
wind farms are 1) what are the optimal sensor locations and
C. OBSERVER
2) how many sensors should be used such that state recon-
An observer is able to estimate the full state (and possibly struction is still possible and qualitatively acceptable. The
update model parameters) based on specific measurements. first question relates to increasing the information density
A closed-loop control scheme using an observer is depicted from each sensor. Minimizing the number of sensors is from
in Fig. 10. In this figure, r k is a reference signal, x̂ k is the an economical perspective important. These questions are
system state estimated by the observer, yk are measurements not easy to be answered due to the time-varying behaviour
taken from the plant, uk is the control signal, and wk and a wind farm exhibits.
vk are process and measurement noise, respectively. For In Section VI-A, VI-B and VI-C, a summary of wind farm
example, given only rotor velocities, an observer can, when control strategies has been given. It can be concluded that
containing a proper model, estimate the flow velocity vectors most of these strategies are optimization-based and evaluate
in the whole farm assuming observability. The latter holds optimal control settings by optimizing a cost function. How-
true if initial conditions can be inferred from measurements ever, most controllers in literature are not implemented in a
(see, e.g., [150] for more information on this topic). Ob- wind farm or a high fidelity flow solver to validate their
servers (also called estimators) contain a dynamical model true performance. Less research has been done regarding
and can be used in combination with, e.g., a model-predictive the application of modern control strategies in wind farms,
controller. Observer properties include the ability to: thereby making this a relatively undiscovered research area.
• Estimate states from specific measurements Applying observers in wind farms shows promising results,

13
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
though more research is necessary. wind farms is yet to be proven. Still, the use of techniques
in parameter-varying control that can help link multiple
VII. FIELD TESTS linear reduced-order models is promising.
From Section IV, it can be concluded that there are many • Current literature tells us that axial induction control
wind farm models that predict flow fields, power capture, will most likely not result in power production increases
and/or loading in a wind farm. Parametric and medium without increasing structural loading. Open questions are
fidelity models are sometimes validated using flow data from if axial induction control can be used to minimize the
high fidelity wind farm models. However, validation of these turbine’s structural loading while maintaining power pro-
solvers using real wind farm data is still ongoing. Although duction and if it is applicable in active power control.
it is expensive to do field testing, it is essential for further • Wake redirection control is a promising actuation method
development. Field tests are not only used to validate high for wake control. Additional field tests are required to
fidelity flow models, but also to obtain results that show provide more information on the true potential of this
that wind farm control can be worthwhile in general. For actuation method. Furthermore, it could be beneficial
example, fields tests are described in [155], [156], [157], to study the combination of axial-induction and wake
[158], [159]. A less expensive approach is doing wind tunnel redirection control in greater detail.
experiments (see e.g., [29], [34], [66], [129], [141], [160]). • Designed controllers should be tested on real wind farms,
Although wind tunnel tests can provide interesting data, the or at least in a high fidelity wind farm simulator for dif-
experiment environment remains a scaled conditioned one. ferent test cases, to get a better idea of their effectiveness
This prevents a one-to-one comparison to real wind farms. in realistic wind farm scenarios.
In addition, it appears to be challenging to have realistic • Remote-sensing technologies, or other measurement de-
turbines and flow characteristics at a smaller scale. However, vices used in wind farms, should be researched further.
the advantage of this is that a more idealized experiment can These methods are critical for control algorithms to obtain
be performed, which can better be represented in simulation, reliable measurements of wake dynamics used for deter-
and thus provide a better comparison between a simulation mining a certain control policy and to update an internal
and an experiment. model.
• The application of an observer in a wind farm is promis-
VIII. CONCLUSIONS
ing. It provides the ability to take a few measurements and
In this paper, basic wind farm control-oriented modeling thereby estimate the full state space of the model. From
and control concepts have been explained and literature has a practical point of view, this is much more realistic than
been categorized and discussed. The following summarizing assuming full state knowledge. An observer is based on a
conclusions can be drawn: dynamical model and can be used in combination with,
• High fidelity models are suitable for flow and wind farm e.g., a model-predictive controller. However, relatively
controller analysis. They are also suitable for exploring little research has been done regarding this topic, and its
the possibilities of wind farm control. However, more true potential is still a question.
validation of high fidelity models with field test data is • More field experiments should be conducted to further
necessary to improve their quality. Because high fidelity investigate if wind farm control can improve the perfor-
models are computationally complex, they are not suitable mance of a real wind farm and to obtain data to validate
for online control. existing models.
• The use of medium fidelity dynamical models can, e.g., • For long-term research challenges in wind energy,
be employed to predict the available power and/or flow see [161].
fields in a wind farm. In addition, they can deal with
changing atmospheric conditions over space and time. IX. ACKNOWLEDGEMENT
However, current medium fidelity dynamical models based The authors would like to thank Mehdi Vali, Andreas
on the Navier-Stokes equations are still computationally Rott, Steffen Raach, Ben Wilson and Raymond de Rui-
complex, hence studying simple dynamical and parametric jter for their discussions and providing feedback on this
steady-state models could be helpful. The question is if a manuscript. The authors B.M. Doekemeijer, J.A. Frederik
sufficient amount of dynamics can still be captured with and J-W van Wingerden would like to acknowledge the
these models so that they can be used for wind farm CL-Windcon project. This project has received funding
control resulting in realistic results. In some specific cases, from the European Union’s Horizon 2020 research and
medium fidelity dynamical and low fidelity steady-state innovation programme under grant agreement No 727477.
models have shown similar simulation results with respect P.A. Fleming, J. Annoni, A.K. Scholbrock were supported
to high fidelity models, though no conclusive statement by the U.S. Department of Energy under Contract No.
can be made yet. DE-AC36-08GO28308 with the National Renewable En-
• Reduced-order models can provide information on im- ergy Laboratory. Funding for their work was provided
portant wake farm dynamics with limited computational by the DOE Office of Energy Efficiency and Renewable
complexity. However, these models are valid for one Energy, Wind and Water Power Technologies Office. The
specific atmospheric condition, and applicability in real U.S. Government retains, and the publisher, by accepting

14
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
 this article for publication, acknowledges that the U.S. [26] D. Medici and P. H. Alfredsson, “Measurements on a wind turbine
Government retains a nonexclusive, paid-up, irrevocable, wake: 3d effects and bluff body vortex shedding,” Wind Energy, 2006.
[27] P. A. Fleming, P. M. O. Gebraad, S. Lee, J. W. van Wingerden,
worldwide license to publish or reproduce the published K. Johnson, M. Churchfield, J. Michalakes, P. Spalart, and P. Mo-
form of this work, or allow others to do so, for U.S. riarty, “Evaluating techniques for redirecting turbine wakes using
Government purposes. SOWFA,” Renewable Energy, 2014.
[28] P. M. O. Gebraad, M. Churchfield, and P. A. Fleming, “Incorporating
atmospheric stability effects into the FLORIS engineering model of
R EFERENCES wakes in wind farms,” TORQUE, 2016.
[29] M. F. Howland, J. Bossuyt, L. A. Martinez-Tossas, J. Meyers, and
[1] International Energy Agency, “World Energy Outlook,” 2014. C. Meneveau, “Wake structure in actuator disk models of wind tur-
[2] www.euractiv.com/section/climate-environment/news/denmark- bines in yaw under uniform inflow conditions,” Journal of Renewable
breaks-its-own-world-record-in-wind energy/. Sustainable Energy, 2016.
[3] M. H. Hansen, A. Hansen, T. Larsen, O. Stig, P. Sorensen, and
[30] B. Hasager, L. Rasmussen, A. Peña, L. E. Jensen, and P. Réthoré,
P. Fuglsang, “Control design for pitch-regulated, variable-speed wind
“Wind farm wake: The Horns Rev photo case,” Energies, 2013.
turbine,” Tech. Rep. Riso-R-1500(EN), Risø National Laboratory,
2005. [31] M. Steinbuch, W. de Boer, O. Bosgra, S. Peters, and J. Ploeg,
“Optimal control of wind power plants,” Journal of Wind Engineering
[4] V. Rezaei, “Advanced control of wind turbines: Brief survey, catego-
and Industrial Aerodynamics, 1988.
rization, and challenges,” American Control Conference, 2015.
[5] S. T. Navalkar, E. van Solingen, and J. van Wingerden, “Wind tunnel [32] K. Johnson and N. Thomas, “Wind farm control: Addressing the
testing of subspace predictive repetitive control for variable pitch aerodynamic interaction among wind turbines,” American Control
wind turbines,” IEEE Transactions on Automatic Control Systems Conference, 2009.
Technology, 2015. [33] T. Hahm and S. Wußow, “Turbulent wakes in wind farm configura-
[6] J. Laks, L. Y. Pao, and A. D. Wright, “Control of wind turbines: tion,” European Wind Energy conference & exhibition, 2006.
Past, present, and future,” American Control Conference, 2009. [34] J. Bossuyt, M. F. Howland, C. Meneveau, and J. Meyers, “Measure-
[7] Á. Jiménez, A. Crespo, and E. Migoya, “Application of a LES ment of unsteady loading and power output variability in a micro
technique to characterize the wake deflection of a wind turbine in wind farm model in a wind tunnel,” Experiments in Fluids, 2017.
yaw,” Wind Energy, 2010. [35] M. T. van Dijk, J. W. van Wingerden, and T. Ashuri, T. Li,
[8] S. Kanev and F. Savenije, “Active wake control: loads trends,” Wind “Wind farm multi-objective wake redirection for optimizing power
Energy, 2016. production and loads,” Energy, 2017.
[9] P. M. O. Gebraad, Data-Driven Wind Plant Control. PhD thesis, [36] D. S. Zalkind and L. Y. Pao, “The fatigue load effects of yaw control
Delft University of Technology, 2014. for wind plants,” American Control Conference, 2016.
[10] L. Y. Pao and K. E. Johnson, “A tutorial on the dynamics and control [37] www.share-ng.sandia.gov/news/resources/news_releases/big_blades/.
of wind turbines and wind farms,” American Control Conference, [38] J. Aho, A. Buckspan, J. Laks, P. A. Fleming, Y. Jeong, F. Dunne,
2009. M. Churchfield, L. Y. Pao, and K. Johnson, “A tutorial of wind turbine
[11] T. Knudsen, T. Bak, and M. Svenstrup, “Survey of wind farm control- control for supporting grid frequency through active power control,”
power and fatigue optimization,” Wind Energy, 2015. American Control Conference, 2012.
[12] T. Burton, D. Sharpe, N. Jenkins, and E. Bossanyi, Wind energy [39] V. Ela, E. Gevorgian, P. A. Fleming, Y. C. Zhang, M. Singh, E. Mul-
handbook. John Wiley and Sons, 2001. jadi, A. Scholbrock, J. Aho, A. Buckspan, L. Y. Pao, V. Singhvi,
[13] F. Bianchi, H. D. Battista, and R. Mantz, Wind Turbine Control A. Tuohy, P. Pourbeik, D. Brooks, and N. Bhatt, “Active power
Systems; Principles, modelling and gain scheduling design. Springer- controls from wind power: Bridging the gaps,” tech. rep., National
Verlag London, 2007. Renewable Energy Laboratory (NREL), 2014.
[14] M. Hansen, Aerodynamics of wind turbines, third edition. Routledge, [40] P. A. Fleming, J. Aho, P. M. O. Gebraad, L. Y. Pao, and Y. Zhang,
2015. “Computational fluid dynamics simulation study of active power
[15] W. Tong, Wind power generation and wind turbine design. Wit Press, control in wind plants,” American Control Conference, 2016.
2010. [41] T. Göçmen, G. Giebel, J. E. Sørensen, and N. K. Poulsen, Possible
[16] A. Betz, “Das Maximum der theoretisch moglichen Ausnutzung des Power Estimation of Down-Regulated Offshore Wind Power Plants.
Windes durch Windmotoren,” Zeitschrift für das gesamte Turbinen- PhD thesis, Technical University of Denmark, 2016.
wesen, 1920. [42] J. W. van Wingerden, L. Y. Pao, J. Aho, and P. A. Fleming, “Active
[17] M. M. Hand and M. J. Balas, “Systematic controller design method- power control of waked wind farms,” International Federation of
ology for variable-speed wind turbines,” 2002. Automatic Control, 2017.
[18] E. Bossanyi, “Individual blade pitch control for load reduction,” Wind [43] T. M. Fletcher and R. E. Brown, “Simulation of wind turbine wake
Energy, 2003. interaction using the vorticity transport model,” Wind Energy, 2010.
[19] E. Bossanyi, “Further load reductions with individual pitch control,”
[44] R. Barthelmie, S. Frandsen, K. Hansen, J. Schepers, K. Rados,
Wind Energy, 2005.
W. Schlez, A. Neubert, L. Jensen, and S. Neckelmann, “Modelling
[20] K. Selvam, S. Kanev, J. W. van Wingerden, T. van Engelen, and the impact of wakes on power output at Nysted and Horns Rev,”
M. Verhaegen, “Feedback-feedforward individual pitch control for European Wind Energy Conference, 2009.
wind turbine load reduction,” International Journal of Robust and
[45] V. Spudic, M. Jelavic, M. Baotic, and N. Peric, “Hierarchical wind
Nonlinear Control, special issue on Wind turbines: New challenges
farm control for power/load optimization,” TORQUE, 2010.
and advanced control solutions, 2009.
[21] R. Ungurán and M. Kühn, “Combined individual pitch and trailing [46] H. Sutherland and J. Herbert, “On the fatigue analysis of wind
edge flap control for structural load alleviation of wind turbines,” turbines,” tech. rep., Sandia National Laboratories, 1999.
American Control Conference, 2016. [47] M. Soleimanzadeh, R. Wisniewski, and S. Kanev, “An optimization
[22] L. Marshall and J. Buhl, “A new empirical relationship between thrust framework for load and power distribution in wind farms,” Journal
coefficient and induction factor for the turbulent windmill state,” tech. of Wind Engineering and Industrial Aerodynamics, 2012.
rep., National Renewable Energy Laboratory, 2005. [48] B. Berglind and R. Wisniewski, “Fatigue estimation methods com-
[23] B. Sanderse, S. P. van der Pijl, and K. B., “Review of computational parison for wind turbine control,” eprint arXiv:1411.3925, 2014.
fluid dynamics for wind turbine wake aerodynamics,” Wind Energy, [49] J. M. Jonkman and M. L. Buhl, “Fast v6.0 user guide, technical
2011. report,” tech. rep., National Renewable Energy Laboratory (NREL),
[24] J. Bartl and L. Sætran, “Blind test comparison of the performance 2005.
and wake flow between two in-line wind turbines exposed to different [50] T. J. Larsen, H. A. Madsen, G. C. Larsen, and K. S. Hansen,
atmospheric inflow conditions,” Wind Energy Science, 2016. “Validation of the dynamic wake meander model for loads and power
[25] G. España, S. Aubrun, S. Loyer, and P. Devinant, “Spatial study of production in the Egmond aan Zee wind farm,” Wind Energy, 2012.
the wake meandering using modelled wind turbines in a wind tunnel,” [51] P. J. Moriarty and S. B. Butterfield, “Wind turbine modeling overview
Wind Energy, 2011. for control engineers,” American Control Conference, 2009.

15
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
[52] M. Churchfield, S. Lee, J. Michalakes, and P. Moriarty, “A numerical impact on wake modeling at the Horns Rev offshore wind farm,”
study of the effects of atmospheric and wake turbulence on wind Wind Energy, 2014.
turbine dynamics,” Journal of Turbulence, 2012. [79] A. Peña, P. E. Réthoré, and P. M. van der Laan, “On the application
[53] L. A. Martinez-Tossas, M. J. Churchfield, and S. Leonardi, “Large of the Jensen wake model using a turbulence-dependent wake decay
eddy simulations of the flow past wind turbines: actuator line and coefficient: the Sexbierum case,” Wind Energy, 2015.
disk modeling,” Wind Energy, 2014. [80] A. Rott, B. M. Doekemeijer, J. W. van Wingerden, and M. Kühn,
[54] J. Meyers, “Large eddy simulations of large wind-turbine arrays in “Robust active wake control in consideration of wind direction
the atmospheric boundary layer,” Aerospace Sciences Meeting, 2010. variability and uncertainty,” Manuscript in preparation.
[55] B. Witha, S. Gerald, and D. Heinemann, “Advanced turbine pa- [81] S. Boersma, B. M. Doekemeijer, M. Vali, J. Meyers, and J. W.
rameterizations in offshore LES wake simulations,” International van Wingerden, “A control-oriented dynamic wind farm flow model:
Symposium on Computational Wind Engineering, 2014. WFSim,” Manuscript in preperation.
[56] M. Soleimanzadeh, R. Wisniewski, and A. Brand, “State-space rep- [82] B. Sanderse, “Aerodynamics of wind turbine wakes, literature re-
resentation of the wind flow model in wind farms,” Wind Energy, view,” Report 16, ECN, 2009.
2014. [83] A. Crespo, J. Hernandez, and S. Frandsen, “Survey of modelling
[57] S. Boersma, J. W. van Wingerden, M. Vali, and M. Kühn, “Quasi methods for wind turbine wakes and wind farms,” Wind Energy, 1999.
linear parameter varying modeling for wind farm control using the [84] L. J. Vermeer, J. N. Sorensen, and A. Crespo, “Wind turbine wake
2D Navier Stokes equations,” American Control Conference, 2016. aerodynamics,” Progress in aerospace sciences, 2003.
[58] S. Boersma, P. M. O. Gebraad, M. Vali, B. M. Doekemeijer, and [85] T. Göçmen, P. van der Laan, P.-E. Réthoré, A. P. Diaz, and G. C.
J. W. van Wingerden, “A control-oriented dynamic wind farm flow Larsen, “Wind turbine wake models developed at the technical
model: WFSim,” TORQUE, 2016. university of Denmark: A review,” 2016.
[59] J. F. Ainslie, “Calculating the flowfield in the wake of wind turbines,” [86] J. Annoni, P. Seiler, K. Johnson, P. A. Fleming, and P. M. O. Gebraad,
Journal of Wind Engineering and Industrial Aerodynamics, 1988. “Evaluating wake models for wind farm control,” American Control
[60] G. C. Larsen, H. A. Madsen, J. Bingöl, F. Mann, S. Ott, J. N. Conference, 2014.
Sørensen, V. Okulov, N. Troldborg, M. Nielsen, K. Thomsen, T. J. [87] J. Annoni, P. M. O. Gebraad, and P. Seiler, “Wind farm flow
Larsen, and R. Mikkelsen, “Dynamic wake meandering modelling,” modeling using an input-output reduced-order model,” American
tech. rep., Risø National Laboratory, 2007. Control Conference, 2016.
[61] J. M. Jonkman, J. Annoni, G. Hayman, B. Jonkman, and
[88] N. Hamilton, M. Tutkun, and R. B. Cal, “Wind turbine boundary
A. Purkayastha, “Development of FAST.Farm: A new multiphysics
layer arrays for cartesian and staggered configurations: Part II, low-
engineering tool for wind farm design and analysis,” AIAA SciTech,
dimensional representations via the proper orthogonal decomposi-
2017.
tion,” Wind Energy, 2015.
[62] A. Crespo, J. Hernandez, E. Fraga, and C. Andreu, “Experimental
[89] D. Bastine, B. Witha, M. Wächter, and J. Peinke, “Towards a
validation of the UPM computer code to calculate wind turbine wakes
simplified dynamic wake model using POD analysis,” Energies, 2015.
and comparison with other models,” Journal of Wind Engineering and
Industrial Aerodynamics, 1988. [90] G. V. Iungo, C. Santoni-Ortiz, M. Abkar, F. Porté-Agel, M. A. Rotea,
and S. Leonardi, “Data-driven reduced order model for prediction of
[63] H. Özdemir, M. C. Versteeg, and A. J. Brand, “Improvements in ECN
wind turbine wakes,” TORQUE, 2015.
wake model,” ICOWES conference, 2013.
[64] J. Annoni and P. Seiler, “A low-order model for wind farm control,” [91] J. Annoni and P. Seiler, “A method to construct reduced-order
in American Control Conference, 2015. parameter-varying models,” International Journal of Robust and
[65] G. V. Lungo, F. Viola, U. Ciri, M. A. Rotea, and Leo, “Data-driven Nonlinear Control, 2016.
RANS for simulations of large wind farms,” TORQUE, 2015. [92] P. J. Schmid, “Dynamic mode decomposition of numerical and
[66] M. Bastankhah and F. Porté-Agel, “Experimental and theoretical experimental data,” Journal of Fluid Mechanics, 2010.
study of wind turbine wakes in yawed conditions,” Journal of Fluid [93] C. W. Rowley, T. Colonius, and R. M. Murray, “Model reduction for
Mechanics, 2016. compressible flows using POD and Galerkin projection,” Physica D:
[67] A. Rott, S. Boersma, J. W. van Wingerden, and M. Kühn, “Dynamic Nonlinear Phenomena, 2004.
flow model for real-time application in wind farm control,” Wake [94] M. Abkar and F. Porté-Agel, “The effect of atmospheric stability
Conference, 2017. on wind-turbine wakes: A large-eddy simulation study,” TORQUE,
[68] J. M. Mcdonough, “Introductory lectures on turbulence physics, 2016.
mathematics and modeling,” tech. rep., University of Kentucky, 2004. [95] P. A. Fleming, P. M. O. Gebraad, S. Lee, J. W. van Wingerden,
[69] J. Blazek, Computational Fluid Dynamics: Principles and Applica- K. Johnson, M. Churchfield, J. Michalakes, P. Spalart, and P. Mo-
tions. Elsevier, 2001. riarty, “Simulation comparison of wake mitigation control strategies
[70] S. Frandsen, R. Barthelmie, S. Pryor, O. Rathmann, S. Larsen, for a two-turbine case,” Wind Energy, 2014.
J. Højstrup, and M. Thøgersen, “Analytical modelling of wind speed [96] S. Guntur, N. Troldborg, and M. Gaunaa, “Application of engineering
deficit in large offshore wind farms,” Wind Energy, 2006. models to predict wake deflection due to a tilted wind turbine,” in
[71] F. Porté-Agel and A. Niayifar, “Analytical modeling of wind farms: European Wind Energy Conference & Exhibition, 2012.
A new approach for power prediction,” Energies, 2016. [97] J. Annoni, A. Scholbrock, M. Churchfield, and P. A. Fleming, “Eval-
[72] N. Jensen, A note on wind generator interaction. 1983. uating tilt control for wind farms,” American Control Conference,
[73] I. Katic, J. Hojstrup, and N. O. Jensen, “A simple model for cluster 2017.
efficiency,” EWEC, 1986. [98] T. Knudsen and T. Bak, “Simple model for describing and estimating
[74] P. M. O. Gebraad, F. W. Teeuwisse, J. W. van Wingerden, P. A. wind turbine dynamic inflow,” American Control Conference, 2013.
Fleming, S. D. Ruben, J. R. Marden, and L. Y. Pao, “Wind plant [99] O. Parent and A. Ilinca, “Anti-icing and de-icing techniques for wind
power optimization through yaw control using a parametric model turbines: Critical review,” Cold Regions Science and Technology,
for wake effects - a CFD simulation study,” Wind Energy, 2014. 2011.
[75] P. M. O. Gebraad, P. A. Fleming, and J. W. van Wingerden, [100] A. Peña, C. B. Hasager, J. Lange, J. Anger, M. Badger, F. Bingöl,
“Wind turbine wake estimation and control using FLORIDyn, a O. Bischoff, J. P. Cariou, F. Dunne, S. Emeis, M. Harris, M. Hofsäs,
control-oriented dynamic wind plant model,” in American Control I. Karagali, J. Laks, S. E. Larsen, J. Mann, T. Mikkelsen, L. Y.
Conference, 2015. Pao, M. Pitter, A. Rettenmeier, A. Sathe, F. Scanzani, D. Schlipf,
[76] J. D. Grunnet, M. Soltani, T. Knudsen, M. N. Kragelund, and T. Bak, E. Simley, C. Slinger, R. Wagner, and I. Würth, “DTU-Wind-Energy-
“Aeolus toolbox for dynamics wind farm model, simulation and Report-E-0029,” tech. rep., DTU Wind Energy, 2013.
control,” The European Wind Energy Conference & Exhibition, 2010. [101] P. S. Anderson, R. S. Ladkin, and I. A. Renfrew, “An autonomous
[77] V. Rostampour, K. Margellos, M. Vrakopoulou, M. Prandini, G. An- doppler sodar wind profiling system,” Journal of Atmospheric and
dersson, and J. Lygeros, “Reserve requirements in ac power systems Oceanic Technology, 2005.
with uncertain generation,” Innovative Smart Grid Technologies Eu- [102] R. J. Barthelmie, L. Folkerts, F. T. Ormel, P. Sanderhoff, P. J.
rope, 2013. Eecen, O. Stobbe, and N. M. Nielsen, “Offshore wind turbine
[78] M. Gaumond, P.-E. Réthoré, S. Ott, A. Peña, A. Bechmann, and wakes measured by sodar,” Journal of Atmospheric and Oceanic
K. S. Hansen, “Evaluation of the wind direction uncertainty and its Technology, 2003.

16
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
[103] A. Rettenmeier, D. Schlipf, I. Wurth, and P. W. Cheng, “Power [129] F. Campagnolo, V. Petrović, C. L. Bottasso, and A. Croce, “Wind
performance measurements of the NREL CART-2 wind turbine using tunnel testing of wake control strategies,” in American Control
a nacelle-based lidar scanner,” Journal of Atmospheric and Oceanic Conference, 2016.
Technology, 2014. [130] C. Santoni, U. Ciri, M. Rotea, and S. Leonardi, “Development of a
[104] S. Goossens, “Field-test of nacelle-based lidar to explore its appli- high fidelity CFD code for wind farm control,” in American Control
cations for Vattenfall as wind park operator,” Master’s thesis, Delft Conference, 2015.
University of Technology, 2015. [131] P. M. O. Gebraad, J. J. Thomas, A. Ning, P. A. Fleming, and
[105] D. Schlipf, S. Kapp, J. Anger, O. Bischoff, M. Hofsäß, A. Ret- K. Dykes, “Maximization of the annual energy production of wind
tenmeier, U. Smolka, and M. Kühn, “Field testing of feedforward power plants by optimization of layout and yaw-based wake control,”
collective pitch control on the CART2 using a nacelle-based lidar Wind Energy, 2016.
scanner,” European Wind Energy Association, 2011. [132] P. M. O. Gebraad, P. Fleming, and J. W. van Wingerden, “Comparison
[106] M. Mirzaei and J. Mann, “Lidar configurations for wind turbine of actuation methods for wake control in wind plants,” American
control,” TORQUE, 2016. Control Conference, 2015.
[107] D. Schlipf, D. J. Schlipf, and M. Kühn, “Nonlinear model predictive [133] L. Vollmer, G. Steinfeld, D. Heinemann, and M. Kühn, “Estimating
control of wind turbines using lidar,” Wind Energy, 2013. the wake deflection downstream of a wind turbine in different
[108] D. Schlipf, P. Grau, S. Raach, R. Duraiski, J. Trierweiler, and P. W. atmospheric stabilities: An LES study,” Wind Energy Science, 2016.
Cheng, “Comparison of linear and nonlinear model predictive control [134] J. Park and H. Law, K, “Cooperative wind turbine control for
of wind turbines using lidar,” American Control Conference, 2014. maximizing wind farm power using sequential convex programming,”
[109] D. Schlipf, “Prospects of multivariable feedforward control of wind Energy Conversion and Management, 2015.
turbines using lidar,” American Control Conference, 2016. [135] J. Marden, S. D. Ruben, and L. Y. Pao, “Surveying game theoretic
[110] A. Scholbrock, P. A. Fleming, D. Schlipf, A. D. Wright, K. Johnson, approaches for wind farm optimization,” AIAA aerospace sciences
and N. Wang, “Lidar-enhanced wind turbine control: Past, present, meeting, 2012.
and future,” American Control Conference, 2016. [136] J. Jinkyoo Park, S. Kwon, and K. H. Law, “Wind farm power
[111] M. Mirzaei, M. Soltani, N. K. Poulsen, and H. H. Niemann, “Model maximization based on a cooperative static game approach,”
predictive control of wind turbines using uncertain lidar measure- [137] K. Johnson and G. Fritsch, “Assessment of extremum seeking control
ments,” American Control Conference, 2013. for wind farm energy production,” Wind Engineering, 2012.
[112] www.dongenergy.com/en/media/newsroom/news/articles/first-data [138] Z. Yang, Y. Li, and J. E. Seem, “Optimizing energy capture of
BEACon-radar. cascaded wind turbine array with nested-loop extremum seeking
[113] D. Schlipf, S. Kapp, J. Anger, O. Bischoff, M. Hofsäß, A. Retten- control,” Journal of Dynamic Systems, Measurement, and Control,
meier, U. Smolka, and M. Kühn, “Prospects of optimization of energy 2015.
production by lidar assisted control of wind turbines,” European Wind [139] A. Menon and J. S. Baras, “Collaborative extremum seeking for
Energy Association, 2011. welfare optimization,” Conference on Decision and Control, 2014.
[114] D. Raach, D. Schlipf, F. Haizmann, and P. W. Cheng, “Three [140] U. Ciri, M. Rotea, C. Santoni, and S. Leonardi, “Large Eddy
dimensional dynamic model based wind field reconstruction from Simulation for an array of turbines with Extremum Seeking Control,”
lidar data,” TORQUE, 2014. in American Control Conference, 2016.
[115] T. Horvat, V. Spudić, and M. Baotić, “Quasi-stationary optimal [141] F. Campagnolo, A. Croce, E. M. Nanos, V. Petrovic, J. Schreiber, and
control for wind farm with closely spaced turbines,” Information & C. L. Bottasso, “Wind tunnel testing of a closed-loop wake deflection
Communication Technology Electronics & Microelectronics, 2012. controller for wind farm power maximization,” TORQUE, 2016.
[116] P. A. Fleming, P. M. O. Gebraad, J. W. van Wingerden, S. Lee, [142] J. Park and H. Law, K, “Bayesian ascent: A data-driven optimization
M. Churchfield, A. Scholbrock, J. Michalakes, K. Johnson, and scheme for real-time control with application to wind farm power
P. Moriarty, “The SOWFA super-controller: A high-fidelity tool for maximization,” IEEE Transactions on Control Systems Technology,
evaluating wind plant control approaches,” Proceedings of the EWEA 2016.
Annual Meeting, Vienna, Austria, 2013. [143] Y. Tang, H. He, Z. Ni, J. Wen, and X. Sui, “Reactive power control of
[117] R. Stevens, “Dependence of optimal wind turbine spacing on wind grid-connected wind farm based on adaptive dynamic programming,”
farm length,” Wind Energy, 2015. Neurocomputing, 2014.
[118] P. A. Fleming, A. Ning, P. M. O. Gebraad, and K. Dykes, “Wind [144] M. A. Rotea, “Dynamic programming framework for wind power
plant system engineering through optimization of layout and yaw maximization,” The International Federation of Automatic Control,
control,” Wind Energy, 2015. 2014.
[119] R. N. King, “Adjoint optimization of wind farm layouts for systems [145] Z. Dar, K. Kar, O. Sahni, and J. H. Chow, “Windfarm power
engineering analysis,” Wind Energy Symposium, 2016. optimization using yaw angle control,” Sustainable Energy, 2017.
[120] P. Mittal, K. Mitra, and K. Kulkarni, “Optimizing the number and [146] M. Soleimanzadeh, R. Wisniewski, and K. Johnson, “A distributed
locations of turbines in a wind farm addressing energy-noise trade- optimization framework for wind farms,” Journal of Wind Engineer-
off: A hybrid approach,” Energy Conversion Management, 2017. ing and Industrial Aerodynamics, 2013.
[121] J. Annoni, P. M. O. Gebraad, A. Scholbrock, P. A. Fleming, and [147] S. Raach, D. Schlipf, F. Borisade, and P. W. Cheng, “Wake redirecting
J. W. van Wingerden, “Analysis of axial-induction-based wind plant using feedback control to improve the power output of wind farms,”
control using an engineering and a high-order wind plant model,” American Control Conference, 2016.
Wind Energy, 2016. [148] S. Raach, J. W. van Wingerden, S. Boersma, D. Schlipf, and P. W.
[122] J. Goit and J. Meyers, “Optimal control of energy extraction in wind- Cheng, “Hinf controller design for closed-loop wake redirection,”
farm boundary layers,” Journal of Fluid Mechanics, 2015. American Control Conference, 2017.
[123] W. Munters and J. Meyers, “Effect of wind turbine response time on [149] S. Raach, S. Boersma, J. van Wingerden, D. Schlipf, and P. W.
optimal dynamic induction control of wind farms,” TORQUE, 2016. Cheng, “Robust lidar-based closed-loop wake redirection for wind
[124] M. Mirzaei, T. Göçmen, G. Giebel, P. E. Sørensen, and N. K. Poulsen, farm control,” International Federation of Automatic Control, 2017.
“Turbine control strategies for wind farm power optimization,” Amer- [150] K. J. Aström and R. M. Murray, Feedback Systems. Princeton
ican Control Conference, 2015. University Press, 2008.
[125] P. M. O. Gebraad and J. W. Wingerden, “Maximum power-point [151] B. M. Doekemeijer, J. W. van Wingerden, S. Boersma, and L. Y.
tracking control for wind farms,” Wind Energy, 2015. Pao, “Enhanced kalman filtering for a 2D CFD NS wind farm flow
[126] J. Marden, S. D. Ruben, and L. Y. Pao, “A model-free approach to model,” TORQUE, 2016.
wind farm control using game theoretic methods,” Control Systems [152] B. M. Doekemeijer, S. Boersma, J. W. van Wingerden, and L. Y.
Technology, 2013. Pao, “Ensemble kalman filtering for wind field estimation in wind
[127] J. G. Schepers and S. van der Pijl, “Improved modelling of wake aero- farms,” American Control Conference, 2017.
dynamics and assessment of new farm control strategies,” TORQUE, [153] C. R. Shapiro, J. Meyers, C. Meneveau, and D. F. Gayme, “Wind
2007. farms providing secondary frequency regulation: Evaluating the per-
[128] M. Vali, J. W. van Wingerden, S. Boersma, V. Petrovic, and M. Kühn, formance of model-based receding horizon control,” TORQUE, 2016.
“A predictive control framework for optimal energy extraction of [154] C. R. Shapiro, J. Meyers, C. Meneveau, and D. F. Gayme, “Dynamic
wind farms,” TORQUE, 2016. wake modeling and state estimation for improved model-based reced-

17
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.
 ing horizon control of wind farms,” American Control Conference,
2017.
[155] J. W. Wagenaar and J. G. Schepers, “Wake measurements in ECN’s
scaled wind farm,” TORQUE, 2012.
[156] B. D. Hirth, J. L. Schroeder, W. S. Gunter, and J. G. Guynes,
“Investigating the impact of turbine control on turbine wakes using
advanced doppler radar,” Windpower, 2014.
[157] P. A. Fleming, M. Churchfield, A. Scholbrock, S. S. Clifton, K. John-
son, A. Wright, P. M. O. Gebraad, B. Naughton, J. Berg, T. Herges,
J. White, T. Mikkelsen, M. Sjöholm, and N. Angelou, “Detailed field
test of yaw-based wake steering,” TORQUE, 2016.
[158] Y. Sakagami, P. A. Santos, R. Haas, J. Passos, and F. T. Taves,
“Effects of turbulence, wind shear, wind veer, and atmospheric
stability on power performance: a case study in Brazil,” EWEA
Annual Event, 2015.
[159] P. A. Fleming, J. Annoni, J. J. Shah, L. Wang, S. Ananthan, Z. Zhang,
K. Hutchings, P. Wang, W. Chen, and L. Chen, “Field test of wake
steering at an offshore wind farm,” Wind Energy Science conference,
2017.
[160] www.zephirlidar.com/wake-steering-results-15-power-increase-total-
wind-farm-control demonstration/.
[161] G. van Kuik and J. Peinke, Long-term Research Challenges in Wind
Energy - A Research Agenda by the European Academy of Wind
Energy. Springer, 2016.

18
Authorized licensed use limited to: University of Grenoble Alpes. Downloaded on January 31,2025 at 08:52:43 UTC from IEEE Xplore. Restrictions apply.