AWFC Cordo

Politecnico di Milano
SCUOLA DI INGEGNERIA INDUSTRIALE E DELL’INFORMAZIONE

Corso di Laurea Magistrale in Ingegneria Meccanica
Wind Farm Optimization redirecting Flow through

Yaw Control: numerical and experimental analysis
Relatore Candidato
Prof. Alberto Zasso Federico Cordò
Matr. 854532
Correlatore
Ing. Paolo Schito
Anno Accademico 2017–2018

Federico Cordò: Wind Farm Optimization redirecting Flow through Yaw
Control
Tesi di Laurea Magistrale in Ingegneria Meccanica, Politecnico di Milano.
c Copyright Luglio 2018.
Politecnico di Milano:
www.polimi.it
Scuola di Ingegneria Industriale e dell’Informazione:
www.ingindinf.polimi.it
Dipartimento di Meccanica:
www.mecc.polimi.it/
Contents
1 Introduction 1
1.1 Wind farm modelling . . . . . . . . . . . . . . . . . . . . . 3
1.2 Active wake control . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Wind turbines performances in yawed operations 13

2.1 Pitt&Peters BEM . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Experimental tests . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 Experimental validation of P&P BEM . . . . . . . 19
2.3 Numerical validation of P&P BEM . . . . . . . . . . . . . 24
3 Wake models 27
3.1 Bastankhah and Porte-Agel gaussian model . . . . . . . . 29
3.1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . 29
3.1.2 Experimental validation . . . . . . . . . . . . . . . 33
3.2 FLORIS: a control-oriented tool . . . . . . . . . . . . . . . 40
3.2.1 Additional code implementations . . . . . . . . . . 41
3.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 WFSim: a 2D RANS solver for control design . . . . . . . 49
3.3.1 Equations and computational framework . . . . . . 50
3.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3.3 New turbine and mixing length model . . . . . . . 56
v
vi CONTENTS
4 From standard controller through wind farm algorithm 61

4.1 Reference torque and pitch controller . . . . . . . . . . . . 63
4.1.1 Baseline torque control design . . . . . . . . . . . . 63
4.1.2 Baseline pitch control design . . . . . . . . . . . . . 66
4.1.3 Simulations . . . . . . . . . . . . . . . . . . . . . . 68
4.2 Observer for estimating wind direction and shear . . . . . 71
4.2.1 Formulation . . . . . . . . . . . . . . . . . . . . . . 72
4.2.2 Simulations . . . . . . . . . . . . . . . . . . . . . . 76
4.3 A proposal for the supercontroller structure . . . . . . . . 81
4.3.1 Torque control in yawed conditions . . . . . . . . . 82
4.3.2 Architecture . . . . . . . . . . . . . . . . . . . . . . 84
5 Wind turbines array optimization 87

5.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Bibliography 97
List of Figures
1.1 Wind farm wakes . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Wake development . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Wake meandering . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Yaw redirection control . . . . . . . . . . . . . . . . . . . . 9
2.1 Blade diagram and velocity decomposition . . . . . . . . . 14

2.2 CL-Windcon wind tunnel test . . . . . . . . . . . . . . . . 16
2.3 G1 model . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 G1 model: CT − λ and CP − λ . . . . . . . . . . . . . . . 20
2.5 July session: thrust coefficient . . . . . . . . . . . . . . . . 21
2.6 July session: power coefficient . . . . . . . . . . . . . . . . 21
2.7 September session: thrust coefficient . . . . . . . . . . . . 23
2.8 September session: power coefficient . . . . . . . . . . . . . 23
2.9 FAST vs P&P BEM: thrust coefficient . . . . . . . . . . . 24
3.1 Wind turbine reference system . . . . . . . . . . . . . . . . 30

3.2 Smoothed experimental wake, low turbulence . . . . . . . . 34
3.3 Smoothed experimental wake, high turbulence . . . . . . . 34
3.4 Self-similarity in wind turbines wake . . . . . . . . . . . . 36
3.5 BPG wake reconstruction, low turbulence . . . . . . . . . . 37
3.6 BPG wake reconstruction, high turbulence . . . . . . . . . 37
3.7 Wake width fit . . . . . . . . . . . . . . . . . . . . . . . . 38
3.8 Wake width vs yaw angle . . . . . . . . . . . . . . . . . . . 39
3.9 CT and CP yaw trend in July session . . . . . . . . . . . . 42
3.10 CT and CP yaw trend, optimal set-point . . . . . . . . . . 43
vii
viii LIST OF FIGURES
3.11 Wake growth rate . . . . . . . . . . . . . . . . . . . . . . . 44

3.12 Wake data: FLORIS and experimental, low TI . . . . . . . 46
3.13 Wake data: FLORIS and experimental, high TI . . . . . . 47
3.14 Wake features, low turbulence . . . . . . . . . . . . . . . . 48
3.15 Wake features, high turbulence . . . . . . . . . . . . . . . 48
3.16 Wake data: WFSim and experimental, high TI . . . . . . . 55
3.17 G1: axial force distribution . . . . . . . . . . . . . . . . . . 56
3.18 Mixing length turbulence new mapping . . . . . . . . . . . 58
3.19 Wake data: new WFSim and experimental, high TI . . . . 60
4.1 Torque control: block diagram . . . . . . . . . . . . . . . . 64

4.2 Torque control: reference trajectory . . . . . . . . . . . . . 65
4.3 Rotor speed response to pitch step input . . . . . . . . . . 67
4.4 Pitch control: block diagram . . . . . . . . . . . . . . . . . 68
4.5 Control performances: step input response . . . . . . . . . 69
4.6 Control performances: disturbances rejection . . . . . . . . 70
4.7 Observer block diagram . . . . . . . . . . . . . . . . . . . 75
4.8 Identication of the observation matrix . . . . . . . . . . . . 76
4.9 Observer performances: ideal ramp input . . . . . . . . . . 77
4.10 Observer performances: low TI . . . . . . . . . . . . . . . 78
4.11 Observer performances: high TI . . . . . . . . . . . . . . . 79
4.12 Observer performances: reference tracking at low TI . . . . 80
4.13 Torque control in yawed condition . . . . . . . . . . . . . . 83
4.14 Wind farm control architecture . . . . . . . . . . . . . . . 85
5.1 Wind turbines array, experimental and numerical results:

test 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
test 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
test 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
test 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
List of Tables
1.1 Wake models scheme . . . . . . . . . . . . . . . . . . . . . 7
2.1 Reynolds over the blade length: λ = 8.15 and V∞ = 5.65 . 19
3.1 WFSim simulation parameters . . . . . . . . . . . . . . . . 54
4.1 Observer performances . . . . . . . . . . . . . . . . . . . . 81

4.2 Torque control in yawed condition . . . . . . . . . . . . . . 84
ix
Abstract
Wind turbines are clustered in wind farms for economic reasons. Aero-
dynamic interaction between the wind and the turbine perturbates the
flow field, generating a flow structure called turbine wake. The wake is
characterized by increased incident turbulence and velocity deficit with
respect to the unperturbed velocity of the flow. Wakes interaction with
downstream turbines results in a decreased turbine lifetime and plant
efficiciency.
Nowadays these effects are neglected by the commercial wind farm
controller on the market. Each turbine is considered at individual level
and not as a player of a global system. Supercontroller capable to consider
aerodynamic interactions are being developed and allows to improve the
wind farm performances considerably.
In this thesis, wind farm control-oriented wake models are analysed.
Validation is carried out through experimental measurements performed at
the Polimi wind tunnel. This work was possible thanks to scaled G1 models
supplied by TU Munchen. The models under investigation are dynamic
as WindFarmSimulator (WFSim), and static as FLOw Redirection and
Induction in Steady-state (FLORIS), developed by TU Delft. Besides of
the validation procedure, model parameters and complexity are evaluated,
with a view to computational cost problems.
A proposal of supercontroller is developed with all components, indi-
vidually tested. Starting from standalone reference wind turbine controller,
supercontroller is introduced comprehensive of wake model and observer
for estimating all input parameters necessary to determine the control
action properly. Technique chosen to optimize the wind farm efficiency is
xi
xii LIST OF TABLES
active yaw control capable to induce significant wake redirection.

At last, control variable optimization is performed for 3x1 wind turbines
array by means of FLORIS model, with specific focus on yaw control. This
configuration was also tested in wind tunnel environment and therefore
experimental data are compared to the predictive numerical model to
optimize the performances of this small wind farm.
Keywords: Wind Farms; Wake Models; Supercontroller; Yaw Control
Sommario
Le turbine eoliche vengono raggruppate in impianti per ragioni eco-

nomiche. L’interazione aerodinamica fra il vento e la turbina genera una
perturbazione all’interno del campo di moto chiamata scia. La scia è
caratterizzata da un aumento della turbolenza incidente e da un deficit
di velocità rispetto alla velocità del flusso imperturbato. L’impatto della
scia sulle turbine retrostanti causa una diminuzione del ciclo di vita della
macchina e una diminuzione in termini di efficienza dell’impianto.
Questi effetti vengono trascurati dai controllori per impianti eolici
attualmente in commercio. Ogni turbina viene considerata a livello in-
dividuale e non come un elemento all’interno di un sistema globale. Su-
percontrollori che considerano le interazioni aerodinamiche sono in fase
di sviluppo e consentono di migliorare le performance dell’impianto in
maniera considerevole.
In questa tesi vengono analizzati i modelli di scia orientati al controllo
di impianti eolici. Una validazione dettagliata viene eseguita attraverso
le misure effettuate presso la galleria del vento del Politecnico di Milano.
Lo studio è stato reso possibile grazie ai modelli in scala G1 messi a
disposizione dalla Technische Universität München. I modelli presi in
considerazioni sono di natura dinamica, WindFarmSimulator (WFSim),
e statica, FLOw Redirection and Induction in Steady-state (FLORIS),
sviluppati dalla Technische Universiteit Delft. Oltre alla validazione, sono
stati analizzati i parametri di modello e la complessità dello stesso con il
derivante costo computazionale.
Un modello di supercontrollore è stato sviluppato con tutti i suoi
componenti, testati singolarmente. Partendo dal controllo di riferimento
xiii
xiv LIST OF TABLES
per singola turbina, si è introdotto il supercontrollore comprensivo del

modello di scia e di un osservatore al fine di stimare tutti i parametri
necessari per determinare l’azione di controllo. La tecnica di controllo per
ottimizzare il rendimento di impianto è il ridirezionamento attivo in yaw
del flusso incidente.
Infine, utilizzando il modello FLORIS sono state ottimizzate le variabili
di controllo per un array di turbine 3x1, con particolare riferimento al
controllo in yaw. Tale configurazione è stata testata all’interno della galleria
del vento e, pertanto, è reso disponibile il confronto fra i dati sperimentali e
il modello numerico predittivo delle performance di questo piccolo impianto
eolico.
Parole Chiave: Impianti Eolici; Modelli di Scia; Supercontrollore; Con-
trollo in Yaw
Chapter 1
Introduction
Nowadays, the importance of renewable energy is not questionable.

Global warming and climate change are perhaps the most difficult human
challenges in 21st century. Fortunately, several renewable energy sources, as
biomass, hydrogen, nuclear, solar and wind are on the rise. In this scenario,
the role of wind energy appears as the most reliable and competitively
price technology. The advantages of wind energy are evident: wind is
easily available, installation area is much lower than the area requested
by other renewable energies, as solar plants, installation and maintenance
costs can be recouped in twenty years. For all these reasons, the growth of
wind energy is impressive: more than 52GW of clean, emissions-free wind
power was added in 2017, bringing total installations to 539 GW globally.
Thanks to this capacity, 3.7% of global electricity is supplied by wind
power in 2015. However, wind energy is not only competitive in the
worldwide markets, but also has a significant impact on environmental
problems. 387.000.000 cubic meters is the amount of water use saved by
wind energy in the EU, equivalent to the average annual household water
use of nearly 7 million EU citizens and 5.497.522 is the amount of avoided
CO2 emissions by wind power each year in Brazil[0].
Wind turbines produced today are operating close to their maximum
possible efficiency, and large machines are used to make them cost effective.
Regarding wind energy, the scientific research is being adressed to two
1
2 Chapter 1. Introduction
challanges: developing offshore wind turbines and implementing wind farm

control. CL-windcon is an EU financed project aiming to investigate real
margin of wind farm control.
Currently, wind turbines are controlled in an individual way (greedy
control), such that each turbine optimize its own performance in terms of
power production and loading, based only on the available information of
its own measurements. However, in wind farms, it is inevitable that wake
interactions take place due to the relative proximity between machines.
Wind turbines in the wake of another turbine experience a wind field with
a lower quality.
In facts, the commercial control logics do not consider these aero-
dynamic interactions and their effects on power production and fatigue
loading. This gets the wind farm to operate in a non-optimum way, since
wind turbines are not controlled as players of a global system. In particular,
for large-scale onshore and offshore farms, there is a great potential with
regard to overall operational optimization which could be exploited by lever-
aging all available data across the farm as well as farm-level interactions
between turbines.
New control strategies are being developed to optimize the overall
output of a wind farm. Open-loop techniques need accurate wake model,
instead of closed-loop which are based on the measurements available
within the wind farm. The most accurate wind farm models are based
on computational fluid dynamic, which is very computationally expensive
and not suitable for control. Therefore, the current control-oriented wake
models are basic static approach, but new dynamic models will be devel-
oped and validated to change standalone controller techniques forward
to comprehensive real-time optimization of the entire wind farm. This
approach will incorporate a balance between energy production, loads,
lifetime and Operations and Maintenance (O&M) costs leading to minimize
the lifetime Levelized Cost of Energy (LCoE).
1.1. Wind farm modelling 3
1.1 Wind farm modelling

Transmission lines costs, operational and maintenance costs, and last,
but not least visual and aeroacustic pollution led wind turbine to be
localized and organized in wind farms. Estimating correctly the wind
power production and the wind loads is a difficult effort since the flow is
very complex inside the wind farm. In fact, downstream turbines could
never work at the designed optimal point, as the flow in the wakes have
decreased velocity and increased turbulence.
Figure 1.1: Clouds forming in the Horns Rev offshore wind plant in Denmark
allows to visualize typical flow structure within the wind farm –
Photo c Christian Steiness
If on one hand power losses affect the plant efficiency, on the other
asymmetric loads affect lifetime of downstream turbines. Currently, greedy
controls operate at an individual level, non-exploiting the overall wind
energy and, therefore, working in a non-optimal way. Wind farm control
algorithm needs wake models to estimate the wind flow and predict the
performances of the overall wind turbines. For open-loop control algorithm,
wake models are essential and the parameters has to be estimated properly
to reconstruct correctly wind flow while the closed-loop ones are based on
the measurements available within the wind farm. If measurements are
available, the wake model parameters may be online estimated to guarantee
good matching between the model suitable for the control design and the
real wind flow.
Figure 1.2: Development of the vertical profile shear in the three regions
Wake Physics According to Crespo[0], the wind turbine wake can be

separated in three regions: a near wake, a far wake and a transition
region. In the near wake pressure and velocity deficit is non-uniform and
the turbine geometry directly affects the flow. In this region velocity
deficit can be considered as induced by the vortices. As the fluid moves
downstream, the cylindrical layer expands and the pressure arises until
the ambient condition is reached. The wake expansion stops at about one
turbine diameter. Turbulence production at the shear layer is enhanced
by the high velocity gradient. Near-wake region ends-up when shear layer
reaches wake centerline at certain distance (approximately 2-5 diameters).
A transition region extending up to 5 rotor diameters downstream leads
to far wake, where the wake is fully developed. Far wake region shows
axisymmetric velocity profile and turbulence intensity and the effect of
the rotor is only seen through more large-scale effects. The centreline
deficit decays monotonically with a rate strongly dependent on the ambient
turbulence. From the prospective of the control design, far wake is the
most interesting region, since in wind farm, generally downstream turbine
operates in far wake region. Far-wake region shows the property of self-
similarity, which is valid with good approximation also in presence of the
1.1. Wind farm modelling 5
ground and the shear of the ambient flow. In this regard, far wake region
is particularly suitable to be modelled as a gaussian profile[0].
Wind tunnel experiments are rather different with respect to full-scale
experiments because of wake meandering[0]. The importance of wake
meandering effect is pointed out by Larsen[0]. The wake meandering
causes continuous change of the lateral and longitudinal position of the
wake centre. It helps the overall efficiency of the plant, misaligning the
turbines most of the time. Since the scale effect and the influence of
terrain are assumed to be responsible for wake meandering, it should
be modelled as stochastic effect. To take into account this effect more
sophisticated numerical models re required, such as Dynamic Dynamic
Wake Meandering (DWM)[0] or Large Eddy Simulation (LES). However
designing a control exploiting this models is computationally expensive
and, nowadays, unfeasible.
Figure 1.3: Ideal sketch of wake meandering
Turbine models Another fundamental part of wind farm model is the

turbine model to predict loads and power production. Actuator disk is
the simplest model based on the 1-D momentum theory. In this case the
rotor disc acts as a drag device slowing the undisturbed wind speed. The
axial momentum equation is coupled to Bernoulli equation to compute
thrust force as function of axial induction only[0]. The classical method for
estimating performances of wind turbines is the Blade Element Momentum.
BEM method starts from 1-D momentum theory discretizing the whole
rotor in several annular infinitesimal elements. This method considers both,
axial and tangential induction. The infinitesimal forces are computed as

function of the inductions, then integrating over the entire rotor disk, thrust
and torque are obtained. BEM shows good agreement with experiments,
but it needs all the constructive parameters of the blades, such as chord,
Drag and Lift static coefficient over the blade length. A popular BEM
based high fidelity wind turbine is Fatigue, Aerodynamics, Structures and
Turbulence (FAST) model, developed at the National Renewable Energy
Laboratory (NREL)[0].
Wake models A detailed analysis of wake models can be found in

literature thanks to Doekemeijer[0]. Simpler models are kinematic models.
In general, these models are steady-state analytical models and describe in
average sense the far wake region. Typically, they require to be coupled to
turbulence models. More complex models are Navier-Stokes based models.
Regarding wind farm modeling, the most accurate simulations are LES
ones, which allow to catch all the effects of wake physics. Wind farm
control design presents two different possible approaches:
• Complex wind farm model, computationally onerous.
• Simple parameter-dependant wind farm model, with online estimation

of tuning variable.
Since many measurements from wind farm are available, at the actual
state of the art, second solution is preferable. Therefore, in this work,
only simple models are considered as FLOw Redirection and Induction
in Steady-state (FLORIS) and Wind Farm Simulation (WFSim). Both
were developed in TUDelft. FLORIS[0] is based on Basthakhan Porte-
Agel Gaussian model[0] and WFSim[0] is a 2-D unsteady RANS solver.
These models are respectively low and medium fidelity, FLORIS is static,
instead of WFSim, which is dynamic. The models are validated through
experimental wind tunnel tests in Chapter 3.
1.2. Active wake control 7
Low fidelity Mid fidelity High fidelity

Model Kinematic Flow field Flow field Flow field
type models models models models
Costitutive
Heuristic RANS 2D uRANS 3D uLES
equation
Flow
2D/3D 2D 2D 3D
dimension
WFSim
Model Floris Ainslie SOWFA
DWM
Dyn/Static Static Static Dynamic Dynamic
Rotor Actuator Actuator Actuator Actuator
model Disk Disk Disk Line
Turbine
Postproc. Postproc. Postproc. FAST
model
Comput. 101 days
101 s on 102 s on 103 s on
effort on a
desktop PC desktop PC desktop PC
order cluster
Model Low or Low or Low or High or
accuracy Medium Medium Medium very High
Table 1.1: Wake models scheme
1.2 Active wake control
As said, greedy control strategy is a wind farm suboptimal control

solution. First possibility is represented by axial-induction-based control.
Reducing the power capture of the upstream wind turbine, it is possible
to decrease the wake deficit within the flow. Performing this strategy,
following Knudsen[0], it is possible to improve the overall power production
up to 4 − 6% in ideal cases. Alternatively, active control wake is a farm
operation strategy allowing to redirect the wake away from downstream
turbines and increase the wind energy flowing in downstream turbines. Due
to wake redirection, active control wake appears more advantageous than
axial-induction-based control. This control strategy aims at improving

plant efficiency and increase lifecycle of wind turbine reducing the blades
loads. Wake redirection can be achieved in three different ways:
• Yaw control rotating the rotor and nacelle around the tower. In
this way it is possible to misalign the rotor with wind direction and
deflected the wake in the opposite direction to the yaw angle (γ).
• Tilt control rotating the rotor and nacelle in the tilt direction. In
this case the flow is redirected upward and downward in the opposite
direction to the tilt angle (τ ).
• Independent Pitch Control (IPC) rotating the blades and creating

an imbalance of the forces of the blades, causing redirection. IPC
is also used to reduce blades fatigue due to wind shear and gravity
effects[0] at individual wind turbine level, but latest development
shows the unexplored potential of this new technique.
Yaw control Tilt is not a control variable in present-day wind turbine, so

for this investigation tilt control is not considered. Wind turbine wakes and
performances working in yawed condition are investigated by Adaramola[0]
and Bastankhah[0] through numerical simulations and wind tunnel experi-
ments. Yawing the wind turbine, the overall thrust force is decomposed
into the two directions longitudinal and transversal to the flow direction.
Transversal component account for deflecting the wake. Redirection effec-
tively depends on turbulence intensity and wind turbine thrust coefficient.
According to Bastankhah[0] the most interesting application in wind farm
control might be in offshore wind plant where turbulence intensity is very
low and wake redirection very effective. Wind farm control, in general, are
borderline technique, since uncertainty in model parameters and real inflow
condition might affect severely control performances. In fact, upstream
turbines decrease their own power productions, which in special conditions
might not be compensated by the increasing of the downstream ones. Sim-
ple relations allow to compute thrust and power coefficient, which close to
1.2. Active wake control 9
optimal TSR are: CT ≈ cos2 (γ) and CP ≈ cos3 (γ). Therefore, optimal yaw
configuration is a result of a compromise between partializing upstream
wind turbine and optimizing wake flow of the downstream wind turbine.
High fidelity simulations have shown that yaw control can increase power
of 4 − 7%[0], even if the topic is still actively reasearched.
Figure 1.4: Yaw redirection control
Individual pitch control IPC is used to reduce periodic loads on the

blades and support structure avoiding dangerous fatigue phenomena. In
this case, IPC can be used in unconventional manner to redirect the flow
inducing force imbalance on the rotor. Although the thrust force is still
directed parallel to the flow, uneven distribution on the rotor can cause
differences in velocities in the wake, but not significant redirection of the
flow. Moreover, IPC causes the blade torque to be uneven over the rota-
tion period, skewing the wake. This application of IPC is actual under
investigation.
Concerning this work, the proposed control architecture is based on yaw

control. This choice is due to two main reasons: on one hand yaw control
was more investigated, therefore it is readier than the other techniques,
on the other yaw control represents an additional degree of freedom with

respect pitch and generator speed. Advantages and drawbacks of yaw
control are well-known, flow redirection is almost effective, increasing
downstream turbines performances, but decreasing flow quality due to
asymmetric velocity distribution. Partial overlap of the flow might induce
fatigue problems, but thanks to IPC this phenomenon can be mitigated[0].
Yaw control can be coupled to greedy control (composed by torque and
pitch control) independently and avoiding creating conflicts between the
two logics. In practise, this fact is not always true, since optimal TSR
may depends on yaw angle, such that torque control parameters may be
function of yaw angle. Cp − λ curves allows to determine which is the
optimal TSR as function of the yaw angle. Therefore, in the hierarchy
of the logic, yaw control constitutes the highest layer. In Chapter 4, the
control architecture comprehensive of greedy control and yaw control is
explained in detail.
1.3 Thesis outline

The object of this MScThesis is to validate current static and dynamic
wind farm models through experimental tests thanks to Polimi Wind
Tunnel and to investigate the real margin of wind farm optimization by
wake redirection through yaw control. In this Chapter detailed state of the
art of the problem is presented with reference to wind farm modelling and
wake redirection control. In Chapter 2, BEM with Pitt&Peters correction
is compared with FAST and experimental data for evaluating wind turbine
performances in yawed conditions. Then, in Chapter 3, validation of static
(FLORIS) and dynamic (WFSim) wake models is performed through wind
tunnel wake measurements; advantages and drawbacks are deeply analyzed.
In Chapter 4, after a brief overview of the reference standalone wind
turbine controller, an inexpensive observer is developed for estimating
wind direction and shear, then a proposal of supercontroller algorithm
based on yaw control is illustrated. In Chapter 5, FLORIS model is used to
predict the power of 3x1 turbines array tested in Polimi Wind Tunnel, then
1.3. Thesis outline 11
the optimal configuration is compared to experimental one. Conclusions,

recommendations and future outlooks for this research will also be touched
upon in this chapter.
Chapter 2
Wind turbines performances in

yawed operations
Extra code part was implemented in FLORIS and WFSim to reproduce

turbine behaviour. Previously, the lack of aeroelastic code was supplied by
actuator disk model. Both worked coupled to SOWFA, then postprocessed
thrust and power were used as input parameter to reconstruct wake. This
choice is consistent with maintaining low computational cost, however it
limits the potentiality of the tool since it needs to work online. The two
software are designed to work in real practical cases, where there are many
measurements available coming from the real wind turbine. Therefore, this
choice is entirely acceptable, a fortiori in the CL-windcon context, where the
purpose is to use wind turbines as a communicating network. Nevertheless,
in wind tunnel testing and for predicting optimal set-point, working in
online configuration is not possible and it becomes clear that there is ample
need to develop a code for predicting wind turbines performances.
Point of support was found in BEM, recognised by the entire scientific
community. Simple BEM is not able to catch the performances of the wind
turbines working in yawed operations. Pitt&Peters model is used to treat
skew wake. Review and assessment of the model are provided thanks to
several methodologies. This model is not the best model to capture all
the aspect of the aeroelasticity of wind turbine. Probably, induction is
13
14 Chapter 2. Wind turbines performances in yawed operations
treated correctly, but skew flow and dynamic stall are neglected. From
the control point of view, the model appears appropriate and code is
run quickly. Compromise between accuracy and low computational cost
is reached thanks to this model. Robust solution algorithm for BEM
theory is developed following AeroDyn procedure[0]. Incoming flow to
wind turbine are reconstructed according to Atmospheric Boundary Layer
(ABL) measured in wind tunnel test.
2.1 Pitt&Peters BEM
Figure 2.1: Blade diagram and velocity decomposition
Main steps of the procedure are shown in this section. Referring to

figure 2.1, considering the extension of yawed case, the velocities over the
two directions are:
Vx = V∞ (sin(γ)sin(α)sin(ϕ) + cos(γ)cos(ϕ)) (2.1)

Vy = −V∞ sin(γ)sin(α) + ωrcos(ϕ) (2.2)
where:
• α is azimuthal angle
• γ is yaw angle
• a is the axial induction

2.1. Pitt&Peters BEM 15
• a’ is the tangential induction
• V∞ is the total incoming flow velocity
To compute axial induction and tangential induction standard BEM

algorithm is performed. Hub/tip loss corrections are considered and
three different regions are distinguished: momentum region, empirical and
propeller-break region. While the zero-yaw axial induction coefficient, to
compute the residual from standard procedure, Pitt&Peters correction is
applied for yawed case in agreement with experimental data.

15π ξ r
ayaw =a 1+ tan sin(ϕ) (2.3)
64 2 R
The wake-skew angle can be estimated approximately using the relation
from Burton[0]:
ξ = γ(0.6a + 1) (2.4)
In this formulation tilt is neglected. Comparison studies by Snel and

Schepers[0] explored several different yaw models and found a Pitt/Peters
model with a coefficient of 15π
64
to fit best. This topic is still under investiga-
tion, therefore this coefficient is estimated to fit best experimental results.
The model is assumed to be rigid, local blades velocities due to vibration
are neglected.
Several advantages derive from this model. Axial and tangential in-
duction coefficients are computed with great accurate rate. From these
coefficients it is possible to get thrust, bending moments and power. With
respect actuator disk, the trend of these quantities is reconstructed over the
whole blade length. The incoming flow velocity is taken locally considering
the height of the analysed blade section. In this way, turbulence is included
in computations. Especially in WFSim, having a precise trend of the thrust
might improve the accuracy and deflection of the model. In fact, while
floris is heuristic model, with deflection computed by plug into RANS
self-similar velocity profile, WFSim is a CFD tool with all problems related
to mesh. Establishing the correct distribution of thrust over the turbine
mesh is not trivial, but in this way the distribution is assumed to be more
consistent with real wake induction. Drawbacks are represented by the
computational costs which increases and by the data of blades. Chord
and aerodynamic coefficients are needed to compute axial and tangential
coefficient. Aerodynamic coefficient are estimated from static wind tunnel
tests and are function of CD = CD (α, Re, r), CL = CL (α, Re, r). Therefore
a preliminary study of the airfoils has to be performed.
2.2 Experimental tests

The most accurate method for determining wind loads is to perform
measurements on the full scale structure. This is often non feasible for
evident reasons. Galleria del Vento of Politecnico di Milano (GVPM) is a
special closed-circuit wind tunnel, where several tests in controlled wind
scenario are performed. With its own dimension(14mx4m), GVPM is
the biggest wind tunnel in Europe. Fourteen axial fans (total installed
power 1.5M W ) combined with turning vanes at corners and honeycombs
provide high flow quality and closed-circuit feature. ABL is reproduced
in the low turbulence test section at the upper part of the wind tunnel.
Due to this unique feature, GVPM offers the widest possible range of test
arrangements and alternatives.
Figure 2.2: First CL-Windcon campaign: out- and in-wake flow characteriza-
tion allowing for different yaw misalignment angles and different
turbine power set-points.
2.2. Experimental tests 17
2.2.1 Setup
Regarding CL-windcon project, two specific configurations were tested.
The atmospheric boundary layer profile is generated using bricks and
spires[0]. Onshore configuration provides for turbulence around to 12%,
while offshore for 6%. Two configuration are fitted through power law
with α = 0.2 and α = 0.079 respectively. Two different models are tested,
but for this work, tests of G1 only are reported. G1 stays for generic
1-meter diameter rotor. G1 is conceived to satisfy several specific design
requirements:
• Ensure consistency with commercial wind turbines in terms of aero-

dynamic and control.
• Guarantee reliable measurements in wind tunnel environment thanks

to advanced onboard instrumentation, avoiding problems due to
miniaturization.
Since many turbines are tested in wind tunnel to reproduce wind farm,
the turbines have to be very small because of blockage effect. Control is
also crucial for the testing of advanced control strategies. It is structured
in three distinct levels, thanks to Bachmann M1 system (modular real-time
controller) coupled to a CPU module for running control algorithm. Three
standard controls are implemented:
• Torque control and collective pitch control.
• Indvidual and cyclic pitch control (IPC ans CyPC).
• Yaw control.
For BEM code validation in yawed operations, reference torque and collec-
tive pitch control adjust rotor speed depending on inflow condition. The
comparisons are carried out in comparable configurations, considering that
the TSR is not constant yawing the wind turbine.
With reference to Figure 2.3, measurements available for post-processing
are shown. A torque-meter, located after the two shaft bearings, allows for
Figure 2.3: G1 model: dimensions and measurements for post-processing
the measurement of the torque provided by a brushless motor equipped

with a gearhead and a tachometer. The tower stiffness was designed so
that the first fore-aft and side-side natural frequencies of the nacelle-tower
group are properly placed with respect to the harmonic per-rev excitations.
The tower is softened at its base by machining four small bridges, on
which strain gages are glued. An optical encoder, located between the slip
ring and the rear shaft bearing, allows for the measurement of the rotor
azimuth. The entire nacelle can be yawed by means of a brushed motor,
housed within the hollow tower, equipped with a gearhead. Thanks to this
equipment thrust and power can be computed, through moment and force
balance equations.
In addition, Cobra probes and traversing system provide for wake

measurements. Cobra probes are capable to measure dynamic, 3-component
velocity and local static pressure up to 2KHz. These probes require a
complicated calibration procedure. Therefore, preliminary tests are carried
out to determine the calibration matrix. Traversing system enable lateral
motion of supporting structures where wind flow sensors are mounted. In
this way, both in-wake and out-of-wake measurements can be performed.
2.2.2 Experimental validation of P&P BEM

G1 airfoils are identified following Bottasso[0] procedure. Reynolds
effects can not be neglected in computation, due to small dimension of the
blades. This is a typical problem reported in wind turbine miniaturiza-
tion. Therefore, polars are estimated for a set of four different Reynolds.
Reynolds dependency is introduced in P&P BEM code through these
experimental polars. Lift and Drag depend on working conditions:
W∞ (r)c(r)
Re(r) = (2.5)
q ν p p
W∞ (r) = Wx2 + Wy2 = V∞2 + (ωr)2 = V∞2 + (λr V∞ )2
p
= V∞ 1 + λ2r ≈ V∞ λr (2.6)
The approximation is valid only for λr 1 i.e. for r becoming com-

prable to R. Optimal working point of the blade is at 75% of the total
point, therefore the condition is respected. Moreover, blade chord shape is
design to keep constant Reynolds number for almost the total length of the
blade as reported in Table 2.1. λ and V∞ are the two unique parameters
accounting for Reynolds.
BlSpan[m] BlTwist[deg] BlChord[m] Airfoil ID Reynolds

0.0064 23.9008 0.042 1 14260.18822
0.0318 20.7992 0.053 3 25465.97093
0.0445 18.8828 0.078 4 43299.41335
0.0572 16.9839 0.101 5 63071.76467
0.0699 15.3131 0.106 6 73618.22167
0.0827 13.764 0.098 7 75306.70036
... ... ... ... ...
0.3879 1.7609 0.033 2 81228.93944
0.4006 1.5919 0.032 2 81263.51061
0.4261 1.2539 0.030 2 80694.12616
0.4642 0.5814 0.025 2 72824.32133
0.4769 0.2786 0.022 2 65357.56581
0.496 -0.2974 0.015 2 46685.14
Table 2.1: Reynolds over the blade length: λ = 8.15 and V∞ = 5.65
From previous tests, CP − λ and CT − λ curves were estimated for G1

model. Experimental data titting through numerical BEM is carried out
to adjust the results in zero-yaw condition. In this way, error in zero-yaw
operation are minimized. This process produces a ratio dependant of
TSR, used as corrective factor. This factor multiplied CT and CP also in
yawed-condition. Non-corrected numerical CP − λ and CT − λ are reported
to show the differences between numerical and experimental based fit
2.4. This factors account for numerical model error, measurements error,
uncertainties in estimation of static lift and drag coefficient. Moreover,
these miniatures of wind turbine are very complex to be manufactured.
Therefore, design requirements are not respected perfectly. The corrective
term is necessary to achieve correct estimation, which is a strictly condition
in control design. CT ratio is constant and equal to 0.95, while CP ratio
has more complex trend. For larger TSR, measured CP is greater than
the numerical one with differences up to 8%, on the contrary for lower
TSR the two curves are very similar. This factor is computed for zero-yaw
configuration only. Since the coefficient does not depend on yaw angle,
yaw trend is not distorted by the correction.
Figure 2.4: CT − λ and CP − λ are measured in some points and the curves
over the whole TSR domain are estimated through BEM-based fit.
First session In the first campaign, the working condition are set to
reference values: λ = 8.15, V∞ = 5.65m/s, blade pitch β = 1.4259◦ and yaw
angle over the range γ = {0◦ , 10◦ , 20◦ , 30◦ , 40◦ }. The tests are performed
Figure 2.5: July Session: CT in yawed condition for both onshore and offshore
configurations; pitch angle, β = 1.4259◦ .
Figure 2.6: July Session: CP in yawed condition for both onshore and offshore
configurations; pitch angle, β = 1.4259◦ .
at below rated with a non-optimal set of {λ; β} regarding power capture.

Greater λ was required to match the real wake with greater accuracy in
wind tunnel tests. In fact, for this specific case, vortexes generated at the
tip persists much more in wind tunnel tests than in real case. Higher λ
balances this effect. In this way, more realistic scaled performance and wake
can be achieved. Still, during tests, greedy control was turned on. Since
aerodynamic torque decreases with increasing the yaw angle, to match the
required torque at the generator side, the rotor speed decreases. Therefore,
the curve identified does not allow to estimate a clear trend with the
yaw angle, because also λ changes. Experimental data are time averaged
to compare CT and CP with the numerical ones. Snel and Schepers[0]
coefficient is best fitted through least square method, giving final value
of 15π
54
. Numerical results show good agreement with experimental tests.
Concerning low turbulence configuration, the wind profile input is modelled
as power law ABL, with an exponent α = 0.079. In this case, the maximum
error is lower than 3%. The model is capable to predict the performances of
the wind turbine, therefore suitable for control design. Same procedure is
followed for high turbulence configuration, fitting with exponential α = 0.2.
Even if the tests are performed at below rated, with increasing turbulence,
the λ decreases too. This effect is caused by torque control, which is
non-linear, since it is related to dynamic pressure and therefore to the
velocity square. Greater velocity contribution are amplified, resuting in
lower λ in high turbulence configuration. In this case, numerical model
shows error up to 9% for high yaw angle γ. The performance estimated by
P&P BEM are always better than the experimental. This fact might be
due to a worsening of the blades aerodynamic efficiency in high turbulence
wind profile. Nevertheless, for low angle yawed operation, the error is
below 5%. Therefore, in high turbulence boundary layer, finding optimal
set-point with a view to wind farm might be difficult.
Second session In the second campaign, the working condition are set to
λ = 8.15, V∞ = 5.75m/s, blade pitch β = 0.4259◦ . Lower pitch guarantees
to improve the match to theoretical scaled thrust force. In fact, both,
thrust force and torque at the rotor are enhanced by this modification. In
this test session, two turbines are aligned to estimate thrust and power
production of the downstream turbines also. Hereafter the results of the
upstream turbine are presented with a two-case-variability of the yaw angle
γ. Same considerations regarding control are still valid and processing is
carried out at the same way as July session. For both configurations errors
are lower than 5%.
The model is tested over wide range of pitch and yaw angle. In general,
good match with experimental results is achieved, paving the way of
an open-loop model-based control design. With increasing turbulence,
P&P BEM results get worse, even if the model seems to be capable to
Figure 2.7: September Session: CT in yawed condition for both onshore and
offshore configurations; pitch angle, β = 0.4259◦ .
Figure 2.8: September Session: CP in yawed condition for both onshore and
offshore configurations; pitch angle, β = 0.4259◦ .
reproduce the performance of the wind turbines. Therefore, P&P BEM is

implemented inside of FLORIS and WFSim code substituting the standard
actuator disk code. In this way optimal set-point can be computed offline.
In Chapter 5 the effectiveness of the entire wind farm model is checked by
means of experimental test. Indeed, offline optimal set-point is calculated
and compared with wind tunnel experiments.
2.3 Numerical validation of P&P BEM

At the same time numerical validation is performed thanks to high
fidelity model. FAST by NREL is a comprehensive aeroelastic simulator,
which allows to predict both the extreme and fatigue loads of two- and three-
bladed horizontal-axis wind turbines (HAWTs)[0]. FAST is recognised
by the scientific community as extremely accurate tool. Moreover, the
model is compared to FAST since it is widely used in Chapter 4, for
the control simulations. Of course, the P&P BEM code implementation
follows accurately aerodyn[0] procedure, which is one of the main module
within FAST. Therefore, results are expected to be comparable, but this
comparison qualifies as further check of the code goodness.
For the sake of simplicity, in this analysis lift and drag dependency
on Reynolds is neglected. Therefore, CL and CD are not function of the
control variable λ and of the inflow condition. Input lift and drag curves
are selected at Re = 75000, which is the closest to the Reynolds in wind
tunnel test configuration among the available ones, as reported in Table
2.1. Inflow condition are uniform and, therefore, ABL is not reproduced.
Figure 2.9: Numerical analysis of CT − λ and CP − λ curves in yawed condition
Results are similar and errors are always below 5%. In particular,
CP and power production is well estimated, while CT and thrust force
are slightly different. Wrong estimation of thrust has an impact in wake
reconstruction in FLORIS and WFSim. In fact, as shown in Chapter
2.3. Numerical validation of P&P BEM 25
3, thrust plays key role for estimating wake potential core and deficit.
Important fact to notice is that optimal λ is function of yaw angle γ. For
this reason, λ and γ are not two independant degree of freedom suitable
for control design, but λ depends on γ. Deeper analysis is presented in
Chapter 4, where it is shown that not adjusting λ in yawed operation,
up to 1% of the upstream wind turbine might be lost in terms of power
production.
Chapter 3
Wake models
The core of wind farm modelling is the wake model. The wake is
a flow structure that is characterized by a reduced wind speed and an
increased turbulence because of the turbine interactions with the incoming
flow. The problem depends on many variables both deterministic as the
morphology of the terrain and stochastic as time-varying atmospheric
conditions. Operating point of the wind turbine should adapt to these
conditions to work always in optimal way. Wake interaction effects have
been studied widely by Vermeer[0]. To catch all these aspects of the complex
physics behind wind turbine interaction, the problem investigation requires
sophisticated tools as CFD simulations. The accuracy of models, like
SOWFA, is very high, while they are very computationally expensive.
Therefore, nowadays, these models are not suitable for control design, but
allows to assess simple wake model. Several wake models are reported in
Table 1.1. In this chapter FLORIS and WFSim are analysed in detail in
terms of accuracy and computational cost.
In this work, active wake control is performed by means of yaw control.
Turbines working in yawed conditions have recently received considerable
attention both as: an unfavourable practical issue in the operation of wind
turbines and more importantly, as a favourable method to increase the
power production of whole wind farm. Indeed, wake deflection combined
to faster wake recovery allows to improve efficiency of the downstream
27
28 Chapter 3. Wake models
wind turbines[0] thanks to lateral mean momentum flux, which boosts the
flow entrainment in only one side of the wake due to yaw misalignment.
Therefore, the wake has non-symmetric distribution, with higher energy.
Models are expected to catch this fact and validation in yawed condition
is carried out by means of experimental wind tunnel data.
P&P BEM supplies the performances of the wind turbines to the two
models considered. FLORIS is a heuristic model and lacks a complete
near-wake model, while WFSim performs bidimensional RANS simulations.
Near-wake region is much more complex than far-wake region and it is
influenced highly by several factors such as tip speed ratio and wake rota-
tion. Describing near-wake region is useless and expensive effort for wind
farm control, since the gap among the turbines of the plant is greater than
the near-wake region. Therefore, downstream turbines always work in the
far-wake region of the upstream turbine. ABL and wind turbines working
conditions affect the wake severely. On one hand these aspects are properly
reproduced by experimental test, on other hand the models should be ca-
pable to consider all these parameters. To simplify the problem and reduce
computational costs, tuning variables are introduced. Further development
will allow to compute online these parameters to guarantee good matching
with real-time wind conditions. FLORIS is based on Bastankhah and
Porte-Agel gaussian (BPG) wake model[0]. With respect other models, like
Jensen[0], Ishiara[0] or Frandsen[0], BPG seems to be more solid and capa-
ble to reconstruct the wind turbine wake very effectively, even if keeping
high parsimony and low computational cost. As shown by Annoni[0], BPG
provides the best representation of the wake characteristics under different
atmospheric conditions, specifically accounting for turbulence intensity,
and different turbine operating conditions. The procedure is carried out as
suggested by Bastankhah[0] in the next section to assess the BPG model,
using G1 wind tunnel measurements in ABL. Model parameters are deeply
analysed and some modifications are proposed to improve the model. On
the other side, some WFSim simulations are performed and compared to
wind tunnel data, however from this analysis FLORIS seems much easier
to manage and provides better results tham WFSim.
3.1. Bastankhah and Porte-Agel gaussian model 29
3.1 Bastankhah and Porte-Agel gaussian model

BPG model is the result of detailed analysis by Bastankhah and Porte-
Agel[0]. Wind tunnel measurements were performed by the two researchers
thanks to a high-resolution stereoscopic Particle Image Velocimetry (PIV)
system. The wake characterization led to some notable features, such
that analytical model was formulated starting from the continuity and
Reynolds-averaged Navier-Stokes (RANS) equations for the wake of a yawed
turbine. Self-similarity is the girder of the model ensuring general validity.
Annoni performs lidar field measurements, demonstrating that BPG is
able to accurately capture different scenarios, even if model parameters
are not properly tuned[0]. Annoni real-scale lidar data covers a range up
to 2.8D downstream. This is a significant difference with the presented
study in GVPM, where wake measurements are carried out in far-wake
region at 5D, 7.5D and 10D. The results show good agreements with what
illustrated by Bastankhah and Annoni and still confirm that the model
is very accurate and thanks to low expensive suitable for control design,
identifying optimal set-point.
3.1.1 Overview
The PIV wake characterization shows that self-similar gaussian distri-
bution represents good solution to fit the wake satisfactorily. However, for
yaw angle approaching to 30◦ , gaussian profile distorts the wake shape.
Fortunately, yaw angle equal or greater than 30◦ are rather impractical for
yaw-angle control methods, main purpose of this work. In fact, as already
said, for very high yaw angle, the upstream turbine undergoes to a too
high power decrement, such that the downstream one is unable to compen-
sate these losses. Moreover, self-similarity provides simplicity and general
validity to the model. Bluff-body wakes achieve self-similarity at a certain
downwind distance, i.e. every velocity deficit collapse onto a single curve,
changing the reference frame to dimensionless one through Buckingham
theorem. At the same time, these considerations are acceptable for wind
turbine wakes also, despite of wake deflection. In fact, concerning wind
turbine wakes, some additional parameters should be introduced to obtain

self-similarity in the horizontal plane at hub height.
Figure 3.1: Wind turbine reference system
Referring to Figure 3.1, the equation, which describe velocity and skew
angle can be written as:
−(y−δ) 2 2
−(z−zh )
ū(x, y, z) 2
= 1 − Ce 2σy e 2σz2 (3.1)
ū∞
−(y−δ+σy )2 −(z−z )2
ϑ(x, y, z) 2
h
= e 2σy e 2σz2 (3.2)
ϑm
where C is the velocity deficit at the wake centre normalized with the
incoming velocity and δ is the wake-centre deflection at each downwind
location. The h¯· i notation stays for time-averaged quantities. The wake
centre is assumed to remain at hub height zh as the vertical displacement
of the wake centre is rather small for lower yaw angles. C, σy , σz and δ
have to be determined to close the model.
The integral form of the RANS equation in the streamwise direction
has been extensively used in classical studies of bluff-body wakes[0]. Under
some assumptions as:
• negligible pressure contribution

• negligible shear effect due to presence of the ground ( dūdz∞ ≈ 0)

(questionable)
2
• convective term is predominant with respect to ū0 contribution in
streamwise direction
manipulating RANS equation in the streamwise direction in conservative

form and integrating leads to
Z ∞ Z ∞
d
ū(ū∞ − ū)dydz ≈ 0 (3.3)
dx −∞ −∞
Plug ū (Equation 3.1) into Equation 3.3 is possible to write down C as

function of other parameters. σy and σz are the so-called wake widths.
Near-wake and far-wake are distinguished. The near-wake has an impact
of the far-wake as an offset. Instead in far-wake region, wake widths vary
linearly with distance:
σy x − x0 σy0
= ky + (3.4)
d d d
σz x − x0 σ z 0
= kz + (3.5)
d d d
Where ky and kz are the wake growth rates, x0 is the onset distance of
the far-wake region and d is the turbine diameter. ky and kz are model
parameters and might be determined thanks to experimental tests or high
accurate numerical simulations. All model parameters are discussed in
the next section. Coupling actuator disk theory, Burton relation(Equation
2.4) for the skew angle and the streamwise momentum deficit allows to
compute the onset:
r
σ z0 1 σy0 σz
= , = 0 cos(γ) (3.6)
d 8 d d
While, x0 can be found reviewing a model proposed by Lee[0]. For a

uniform and laminar incoming flow, Lee suggested that a change in width
of a shear layer is only proportional to the velocity difference between the
potential core and unperturbed surroundings. In real situations, however,
the incoming turbulence also has an impact, enhancing the flow entrainment
and consequently the growth of the shear layer. Therefore, shear layer
width variation may consider both turbulence velocity gradient:
1 ds us ds ue
= = αI + β (3.7)
u∞ dt u∞ dx u∞
Where us is the characteristic velocity of the shear layer and it is assumed

to be equal to 12 (u∞ + u0 ), ue is the characteristic relative velocity in the
shear layer and it is assumed to be equal to 12 (u∞ − u0 ) and α and β are
model parameters. Integrating and exploiting actuator disk theory, useful
relation for onset region can be found:
√
x0 cos(γ)(1 + 1 − CT )
=√ √ (3.8)
d 2[4αI + 2β(1 − 1 − CT )]
According to this equation, it can be readily shown that the length of the
potential core decreases with increasing thrust coefficient of the turbine,
incoming turbulence and yaw angle. These facts provide flow entrainment,
but while last two decrease wake deficit, first one has the opposite effect.
Therefore, increasing the thrust, the overall effect is greater subtraction of
energy from the flow, with shorter near-wake region characterized by high
velocity difference between potential core and unperturbed flow. Since x0
is strictly connected to the beginning of the far-wake region, tip-speed-ratio
and wake rotation have no contribution[0].
Last unknown is δ, the wake-centre deflection. At the same way, as

for wake deficit, RANS equation still helps to provide condition for δ
calculation. However, this time, RANS equation in the spanwise direction
is of particular interest. Following the assumption that shear stress term
is generally smaller than the advection term for a yawed turbine, flow rate
of spanwise momentum conservation is derived:
Z ∞ Z ∞
d
ū2 ϑdydz ≈ 0 (3.9)
dx −∞ −∞
Once again, plug ū (Equation 3.1) and ϑ (Equation 3.2) into Equation
3.6, complex expression for δ is derived. Also in this case, the relation
presents a term due to onset of the far-wake region plus a upper bounded
logarithmic term, as x approaches to infinity.
x0
δ0 = ϑ0 (3.10)
d
ϑ0 from Burton. Wake deflection can be computed at each downwind

location; therefore, all wake characteristics are determined. Of course,
yawing effect are included in these calculations, but the model is limited
by Burton model. Deflection caused by interaction of the rotating wake
and, in addition, incoming shear flow are neglected. Tuning variables of
the model are essentially three ky , α and β, since kz can be assumed to be
equal to ky . Bastankhah proposes wake growth rates linear dependant on
turbulence intensity. This solution is almost simple and effective, however
some measurements show dependence on yaw angle also. The background
of the model is a solid structure based on budget study of the continuity
equation and RANS, experimental data, Lee&Chu model and Burton
theory. Therefore, even if the model parameters are not properly tuned,
the results are almost accurate. Moreover, computationally inexpensiveness
has to be highlighted. According to BPG, CT and turbulence intensity are
the two dominant parameter concerning wake deflection and wake recovery.
Wake deflection is very effective when incoming streamwise turbulence is
very low, while wake recovery is enhanced by high turbulence intensity.
Preferable configuration for yaw control are offshore wind farm, where
turbulence intensity is very low.
3.1.2 Experimental validation

Wake measurements are performed in the first test session of CL-
windcon project, following the procedure illustrated in Chapter 2. Once
again, λ is not the optimal one, to guarantee good matching with real wake
measurements. In figure 3.2 and 3.3, experimental mean streamwise wake
velocity in far-wake region (5D) are presented. Smoothing procedure is
carried out on wake measurements, filtering the signals and considering the
wind tunnel mapping, periodically executed in GVPM. Wake data used for
the subsequent analysis are smoothed. Smoothing has minimum impact
on low turbulence measurements, while high turbulence wake is deeply
changed by this process. However, in this way, some notable features are
guaranteed like symmetry and regular wake shape, aspects that, otherwise,
BPG model would be not able to properly reconstruct. As a first impression,
from these figures it has to be noticed that deflection is strongly non-linear
and how much turbulence intensity enhance flow entrainment and therefore,
wake recovery.
Figure 3.2: Experimental wake measurements: mean streamwise velocity, pre

& post smoothing, TI approx 5%
Figure 3.3: Experimental wake measurements: mean streamwise velocity, pre

& post smoothing, TI approx 10%
Starting from these data, the potential of this model is shown. Pre-
liminary analysis considers as given the model parameters. Optimal wake

reconstruction is performed and best result is assessed to fix the target,
independently on which might be the values of model parameters. Then,
according to Basthankhah, model parameters are investigated in function
of turbulence intensity and yaw angle.
Self-similarity is the girder of the model, σy of the Equation 3.1 has
to be defined. In this sense, from literature, many works investigated
wake characteristics and wake width. Pope[0] proposes a definition of
the wake half-width as the region where the velocity deficit is half of its
maximum value, but this definition has the problem to be rather sensitive
on measurement uncertainty. More robust definition for the wake width is
provided by Bastankhah[0]:
y
1
Z
σy = √ lim (ū∞ − ū)dŷ (3.11)
2π(ū∞ − ūc ) y→∞ −y
where σy is the wake width in y direction, ūc is the velocity of the wake
centre and ŷ is the integration variable. The integral can be calculated
numerically to find the wake width at each downstream position. Note that
for a pure Gaussian profile reduces to the standard deviation of the profile.
As reported in Figure 3.4 self-similarity can be reasonably assumed for
lateral velocity proiles in the far wake. All wake measurements plotted with
respect wake width, maximum wake deficit and wake centre collapse onto a
single curve. This data sample is composed of thirty different configurations.
Even if many cases are considered, the coefficient of determination is close
to 1. High turbulent and high yaw angle data are the most difficult to
handle. In this regard, post-processing allows to smooth data and increase
quality for comparison. Much greater differences appears at the lateral
sides of wake. Fortunately, lateral sides are less important for active wake
control, since wake recovery is higher in these regions. In this case, it
is much more difficult to show precise differences between the two sides,
as reported by Bastankhah[0] in his studies. Indeed, theorically left side
should be influenced by strong spanwise velocity, resulting in less degree
of similarity.
Figure 3.4: Self-similarity: wind turbines wake measurements plotted in self-

similarity variables. Several downwind location 5D, 7.5D and 10D;
yaw angle range{0◦ , 10◦ , 20◦ , 30◦ , 40◦ }; low and high turbulence
configurations
Figure 3.4 takes general overview and demonstrates the goodness of

the model. To better understand which the real degree of self-similarity
is, data sample is reduced to more restrictive cases. In this way, the
worst cases can be identified and analysis on the model parameters can
be performed to evaluate real room for improvement of the model. Figure
3.5 and 3.6 show self-similarity of the wake at 5D-downwind location for
low and high turbulence configurations. In addition, wake is reconstruct
using the gaussian fit curve. The reconstruction is defined optimal for two
different reasons. On one hand gaussian fit minimizes the root mean square
error, on the other hand wake width, centre wake and maximum deficit
are computed by experimental data, therefore they are exact values. In
practical cases, where predicting wind turbines wake velocity is the main
object, these data are not available. Nonetheless, BPG model allows to
compute with good approximation all these parameters, given the wake
growth rate and the shear layer variation rate with respect to turbulence
intensity and velocity difference.
Figure 3.5: BPG vs experimental data: mean streamwise velocity at 5D for

different yaw angle, TI approx 5%
Figure 3.6: BPG vs experimental data: mean streamwise velocity at 5D for

different yaw angle, TI approx 10%
Self-similarity is perfectly verified in low turbulence case (Figure 3.5).

Coefficient of determination approaches to 1. Even if in central region
the wake is reconstructed with excellent results, at lateral side the error
increases, but always maintaining below 5%. When turbulence increases
the model has some trouble (Figure 3.6). Coefficient of determination is
still high, but also for non-yawed case low errors becomes visible. As usual
in aerodynamics, whenever turbulence increases, problems become much
more complex. However, the model shows reasonably good result. Using
dedicated gaussian fit improves low turbulence results, but essentially is not
effective in high turbulence configuration. In this sense, this solution is not

only impractical, but also leads to few benefits and increases complexity
and computational cost.
Assessed the self-similarity, overview of the model parameters is pre-
sented as follows. The shear layer variation rate with respect to turbulence
and velocity difference intensity (α and β in Equation 3.7) are supposed to
be constant by Lee[0]. Therefore, the analysis refers to wake growth rate
only. Many studies on this topic are available from literature. Turbulence
intensity is recognised by the scientific community as the main responsible
of wake growth[0] [0]. Eames proposes linear variation of the wake growth
with x. Figure 3.7 shows good agreement with this assumption, as con-
firmed also by the high values of the coefficients of determination. In this
figure wake width is plotted as function of downwind locations for different
values of turbulence intensity and yaw angle. GVPM wake measurements
on one hand confirm the effect of turbulence effect, on the other lead to
quite different result concerning yaw angle effect.
Figure 3.7: Wake width

Linear variation of wake growth rate with the streamwise turbulence

intensity provides velocity deficit to be comparable or smaller than the
standard deviation of the incoming flow and wake radius to be smaller
than the integral length scale turbulence. These conditions are expected to
be fulfilled for a wind turbine far-wake in ABL. Therefore, Bastankhah pro-
poses wake growth rate as function of turbulence intensity only, neglecting
yaw angle effect. As shown by Figure 3.8, when turbulence increases, yaw
angle effects become negligible. The slope of the curves is slightly affected
by the yaw angle and the trend is not clear changing the yaw angle. While,
for low turbulence intensity, increasing yaw angle, the wake growth rate
decreases, as illustrated by the curves slope, which decreases as well. This
is not only in contrast with Bastankhah, but also with what stated by
Jimenez[0], according to whom increasing yaw angle, wake growth rate
increases. Wake growth rate linear dependant with turbulence intensity is
generally valid, but since, in this case, many measurements are available,
specific formulation valid for this case only is derived and introduced in
FLORIS code. CT has supposed to have an impact as onset of the far-wake
region. Indeed, in this region the wake is fully developed and turbine
characteristics can be seen in large-scale effect only.
Figure 3.8: Wake width as function of the yaw angle in the two turbulent
configurations
3.2 FLORIS: a control-oriented tool
Computational framework within which BPG model is implemented, is

FLORIS. FLORIS was developed by NREL and TUDelft. It is defined as
control-oriented tool for wind farm optimization. The code is developed
in MATLAB . R Model is comprised of several modules. BPG is the core
of FLORIS, but turbine model and wake interactions are mandatory for
wind farm simulations. Simple merging wake is required in FLORIS. Wind
turbine and BPG models have been already analyzed in detail, while wake
interactions have not been treated yet. Many wake interaction models are
available from literature[0]. Among engineering models, quadratic rule
from Katic[0] was validated and gave better agreement then simple linear
superposition. Local wakes are superposed to estimate the overall speed
deficit at the n-th downwind turbines, using a quadratic sum of square of
the local speed deficits:
v
u n
uX
∆un+1 =t ∆u2 |n+1i (3.12)
i=1
Katic model is strictly related to dynamic pressure, however wake physics is

unexplained. Square velocity deficits are related to the thrust force. Since
the thrust force goes ≈ u2 and deficits are direct consequence of the thrust
force, this forumlation is preferable, instead of linear summation. Summing
up the square of the deficits, small deficits are penalized with respect to the
larger ones. For this reason, Katic formulation captures the combination of
many wakes close to each other, showing good agreement with experimental
data. In this regard, it is preferable than other techniques in small wind
farm. On the contrary if the effect of neighbouring wake merging is not
negligible, as in the case of large wind farms, Katic is not able to reproduce
the wake reasonably, since the square summation penalizes these low
deficits.
As said in Chapter 1, referring to wake physics effect, increase in
turbulence can be observed. Niayfair proposes a model to incorporate this
3.2. FLORIS: a control-oriented tool 41
effect[0]:
−0.32
0.73 0.35 x
Iadded = Aoverlap 0.8ai I0 (3.13)
Di
where I0 is the ambient turbulence intensity and a is the axial induction
factor of the turbine. Then also turbulence intensities are summed up
thanks to Katic formula. This formulation was derived experimentally in
case of one single turbine. In wind farm scenario properly formulation
might be rather different from this one. Studies have shown that the added
turbulence intensity reaches an equilibrium point between two and three
turbines downstream[0] and decreases upon a certain distance, which in
FLORIS is fixed to 15D.
At last, optimization module is provided to compute optimal working
conditions for the wind farm. Control parameters are axial-induction and
yaw. Interior point algorithm is used to solve this non-linear optimization
problem. Typically, the cost function is defined as the power production,
however with implementation of aeroelastic code, different cost functions
can be chosen. For instance, weighted cost function between maximizing
power and minimizing flapwise and edgewise bending moment is preferable
to increase the wind turbines lifecycle.
3.2.1 Additional code implementations
Standard FLORIS code consists of a simple wind turbine model. Table

of CT and CP must be derived by the users by means of experimental data
or high fidelity numerical simulation. CT and CP are corrected in yawed
operation, thanks to Gebraad formulation:
CT = CT cospt γ (3.14)
CP = CP cospp γ (3.15)
pt ans pp are two tuning parameters that matches the power losses due to
yaw misalignment. As reported by Fleming[0], pp should be between 1.4
and 2.2. Jonkman found a pp of 1.88 for the NREL 5MW, performing LES
simulations. Instead, pt is reasonably set to 2. These two parameters play

a key role in wind turbine performance estimations. The problem of this
formulation consists of estimating pp and pt , which in general are function
of the operating conditions. Parameters, like TSR or β have strong impact
on the two coefficients. However, it is very compact formulation and paves
the way to derive a sort of efficiency for yaw control. Indeed, the ratio
between p1p / p1t represents a rate of energy capture weighted with respect
to the kinetic energy deficit. The lower pp is, the higher is the power
capture, while the higher pt is, the lower is the thrust force and therefore,
the higher is the downstream wake velocity. As already said, in this model
the impact of the thrust force in far-wake region has less importance than
other parameters like yaw and turbulence, since it affects the onset of the
far-wake region only.
Figure 3.9: CT and CP yaw trend Gebraad formulation in July session
In figure 3.9 experimental and numerical are shown and pp and pt

are computed. In this configuration the turbine is controlled. Greedy
control regulates rotor speed by means of torque control, therefore TSR
is not constant increasing the yaw angle. Numerical optimal set-point is
numerically computed and reported in Figure 3.10. It is interesting to
notice that adjusting pitch angle and TSR, pp and pt change from 2.35
to 2.02 and from 1.65 to 1.68, respectively. In terms of the so-called yaw
control efficiency, the ratio increases from 0.702 to 0.831. Performing yaw
control at this operating condition advantages can be obtained in terms of
both, performances and flow quality.
Figure 3.10: CT and CP yaw trend Gebraad formulation in optimal condition

numerically estimated
P&P BEM provides aeroelastic calculations uniquely. Polars and blades

parameters as twist and chord should be identified and supplied to this
additional model, implemented in FLORIS. In this way, not only power
and thrust can be computed, but also blades load as flapwise and edgewise
bending moment, guaranteeing wider choice of cost function. Moreover,
generator speed and blade pitch can be used as control variable of the
system to optimize power and lifecycle of the plant.
In addition, more accurate computations are included in the code,
regarding wake model. In standard BPG model, Glauert[0] formulation is
used to compute axial induction and, then, flow angle:
1 p
a≈ (1 − 1 − CT0 cosγ) (3.16)
2cosγ
Even though the model suggested by Glauert[0] can estimate the thrust
coefficient correctly, it is unable to properly evaluate the skew angle of
the flow through a yawed rotor. Reason accounts for this problem can
be found in the fact that this model only considers the component of the
induced velocity normal to the rotor, and not the tangential one. As a
result, the model is expected to overestimate the flow skew angle at the
rotor. For these reasons, Burton[0] formulation is preferable to estimate
skew angle:
ϑ = 0.6a (3.17)
0.3γ p
ϑ≈ (1 − 1 − CT0 cosγ) (3.18)
cosγ
In standard Floris code Equation 3.16 is plug into Equation 3.17, but
thanks to P&P BEM code, axial induction coefficient is directly computed.
Theorically, this method allows a more precise prediction of flow angle and
therefore, of wake deflection (Equation 3.10).
Jimenez[0] conjecture states that the wake growth rate increases with
increase in yaw angle. In incoming turbulent flow, the wake quickly
asymptotes to a turbulent free shear flow in the far-wake region. Therefore,
Bastankhah proposes that wake growth rate is not affected by the yaw angle,
as turbulence is the main mechanism responsible for flow entrainment. As
said, measurements of wake growth rate for the G1 model are in contrast
with both Bastankhah and Jimenez. Figure 3.11 shows precise trend with
the yaw angle, with low dispersion especially concerning low turbulence
case. As the turbulence increase, considerations by Bastankhah become
more relevant and yaw angle effect becomes negligible.
Figure 3.11: Wake growth rate as function of the yaw angle in the two turbulent
configuration
Linear trend can reasonably fit the yaw angle trend as shown by coeffi-
cient of determinaton. Therefore, different formulation for wake growth
rate is introduced in the code to improve agreement with experimental
data:
I−Iref
1 Iref
k = ka I + kb γ + kc (3.19)
10
ka = .124, kb = .00088, kc = .0236, Iref = .05
This expression is not aimed to explain physics, but it should be interpreted

as corrective term accounts for wake growth rate trend in yawed conditions.
Last factor is introduced to decrease yaw angle effect, as the turbulence
increase. To achieve good matching with experimental data, this yaw angle
term is essential. Performing optimization and open-loop control needs
accurate model, therefore wake growth rate should be properly estimated.
Since the ultimate goal of this work, it is optimal set-up estimation of the
control variables, the corrective term is defined as acceptable. In future,
online estimation is desirable, to adjust tuning variable with respect to
operating and environmental conditions.
3.2.2 Results
All tested configurations in GVPM are simulated by FLORIS. Wake
growth rate is set up as explained in previous section. The remaining
tuning variables, α and β (Equation 3.8), are chosen to minimize error
between measured and numerical wakes. Proper tuning allows to achieve
good prediction of the wake features. For the sake of completeness, two
sets of measurements should be performed: the first one useful to identify
model parameter and the second to validate the model. In this case,
identifying set only is available. In this phase, a preliminary analysis of the
model is performed, since the ultimate purpose of this work is predicting
optimal operating condition. Thrust and Power are computed by P&P
BEM code. Turbine parameters, like TSR, pitch and rotor speed are
consistent with the time averaged experimental data. Both loads and wake
measurements were performed in July session. Therefore, measurement
set-up and control logic are the same explained in Chapter 2.2. Probes are
equally spaced by 100mm one from each others, providing up to twelve
wake flow measurement points within turbine diameter. Since downstream

turbines operate in far-wake region, measurements are carried out at
5D, 7.5D and 10D from the turbine.
(a) γ = 0◦ (b) γ = 10◦
(c) γ = 20◦ (d) γ = 30◦
Figure 3.12: Wake data: FLORIS and experimental at low TI for different
yaw angle configurations {0◦ , 10◦ , 20◦ , 30◦ }
In Figure 3.12, results in low turbulence shear flow are presented. For
low angle misalignment, results show very good agreement with experi-
mental tests. Nonetheless, concerning all configurations, error is always
below 5%. At larger downstream locations, the model matches better the
experimental data, since wake deficit is lower and flow entrainment has
already reintroduced and distributed energy to the flow within turbine
diameter. Maximum wake deficit is properly reconstructed while wake
centre is slightly different with respect to experimental tests (Figure 3.14).
To notice that, when yaw angle increases, CT decreases and according to

Equation 3.8, the onset of the far-wake region shifts forward. In fact, since
in FLORIS far-wake region model only is implemented, in case of low CT
and low turbulence, the model is not able to reproduce wake correctly. Flat
zone of the curve at 5D of the Figure 3.12 (d) is due to this lack in the code.
However, this situation might occur only under particular circumstances
and, therefore, the problem is not relevant from the control prospective.
In this regard, the most important parameter is the spaced-averaged wake
velocity, which is estimated with error below 2%.
(a) γ = 0◦ (b) γ = 10◦
(c) γ = 20◦ (d) γ = 30◦
Figure 3.13: Wake data: FLORIS and experimental at high TI for different
yaw angle configuration {0◦ , 10◦ , 20◦ , 30◦ }
In Figure 3.13, high turbulence results are shown. Also in this case, as
reported in Figure 3.15, wake deficit is reproduced with good approximation,
but wake centre position is not properly caught by the model. Increasing
yaw misalignment, the spaced-averaged wake velocity is slightly different
from the experimental one and the error approaches to 5%, which is
reasonable estimation error.
Figure 3.14: Wake features: deficit and centre, TI approx 5%
Figure 3.15: Wake features: deficit and centre, TI approx 10%
As said BPG model is particularly suitable to reproduce wake in envi-

ronmental conditions, as ambient turbulence and shear inflow conditions,
which is strength point of the model. However, the tests carried out in
GVPM have several differences with respect to real cases. The most im-
portant one is wake meandering. Of course, time-averaging the flow this
effect is damped, but it has to be remembered that FLORIS does not
consider wake meandering and this effect in wind tunnel testing is not
present. Therefore, if good matching is achieved in wind tunnel tests, this
3.3. WFSim: a 2D RANS solver for control design 49
model quality is not guaranteed in real cases.
3.3 WFSim: a 2D RANS solver for control

design
Three-dimensional high fidelity simulations may have 106 or more states.
The resulting computation time may be in order of days or weeks using
parallel computing through powerful CPU. In other words, the computation
time needed for LES is in general more than the total time that is simulated.
Of course, for a control-oriented model, this is a significant issue.
To reduce the high complexity of wake modelling, two-dimensional
tool has been developed. Supposing axis-symmetric conditions, the flow is
estimated in the horizontal plane at hub height. Generally, medium fidelity
models are based on simplified version of the Navier-Stokes equations.
Coupling bidimensional flow and assuming thin shear layer, Navier-Stokes
equations can be simplified and, therefore, computational effort is reduced
significantly. In this regard, WFSim[0] is a control-oriented tool, which
exploits all these factors, with the aim to work in real-time. In this scenario,
model parameters can be estimated online by means of flow measurements
and blades loads. Therefore, the great advantage of this model is to
use simple turbine and turbulence model, dependant of model parameters
which are tuned continously. In this way, the model self-adapts to real-time
situation.
Doekemeijer illustrates that Kalman filter[0] might have important
consequences in online parameters estimation, although these operations
would increase the computational cost. Currently, main problem is related
to wake measurements, which are not available extensively. Nowadays,
researchers are focusing on estimating wake characteristics using LIDAR
devices, which are not mounted on commercial wind turbines. Therefore,
practical issues are still present in estimating turbulence model parameters,
which have significant impact on wake recovery and, therefore, on wind
farm control. Since WFSim is more recent than FLORIS, model sensitivity
to these parameters is not deeply analysed.

From the study performed in this work, WFSim seems to be much
more sensitive to the model parameters, while FLORIS guarantees good
accuracy, even if properly estimation is not performed. For this reason,
in the last Chapter where numerical optimal set-point is compared to the
experimental one, the analysis is carried out with the FLORIS model only.
However, an overview of the WFSim model is presented hereafter and one
single turbine wake simulations are performed consistently with the July
test session.
3.3.1 Equations and computational framework
As any wind farm model, wake, turbines and turbulence model are
needed to define the problems completely. Each turbine is modelled
through actuator disk model (ADM) and 2D Navier-Stokes are used to
derive a model with a heuristic turbulence feature. In this way, the
resulting matrices from discretization process are sparse, leading to low
computational effort.
As starting point, 3D incompressible filtered Navier-Stokes equation
are used, as in LES case:
∂ṽ 1
+ ṽ · ∇ṽ + ∇ · τm + ∇p̃ − f = 0 (3.20)
∂t %
∇ · ṽ = 0 (3.21)
Where ṽ represent the 3-component filtered velocity vector and p̃ the

filtered pressure field. Instead, f represents the effect of the turbines on
the flow. The filter is supposed to be a top-hat filter, with width D, as the
turbine diameter. Of course, this choice is in contrast with LES, where the
filtered length scale is related to subgrid scale model. Fixing z at the hub
height, it is possible to pass from 3D to a 2D system. Moreover, assuming
w velocity along the z-component negligible, the following equations can
be derived:
∂u ∂uw + τM,13 ∂vw + τM,23

+u·∇u+∇·τH +∇p−f = − e1 −− e2 (3.22)
∂t ∂z ∂z
∂w
∇·u=− (3.23)
∂z
With u = (ṽ1 (x, y, zh ), ṽ2 (x, y, zh ))T , τH is the 2D stress tensor and
e1 , e2 are the two unit vector along x and y. The right-hand side is
roughly neglected in the Equation 3.22 and to adjust continuity equation
it is supposed that for a quasi axis-symmetric case: ∂w ∂z
∂v
≈ ∂y . This last
additional term allows to avoid senseless speed-up in the flow. Therefore,
the set of equation which has to be discretized is rather similar to RANS
equation, even if model is derived starting from LES:
∂u
+ u · ∇u + ∇ · τH + ∇p − f = 0 (3.24)
∂t
∂v
∇·u=− (3.25)
∂y
The set of equations are spatial discretized over a staggered grid follow-
ing Versteeg[0]. It is carried out by employing the Finite Volume Method
and the Hybrid Differencing scheme. Temporal discretization is performed
using the implicit method that is unconditionally stable. Dirichlet inflow
boundary condition are imposed to one boundary, while on the others
Neumeann (zero stress) conditions are prescribed. Initial pressure field is
set to zero.
The most controversial issue concerning the model is the mixing length
turbulence definition. Of course, mixing length relates eddy viscosity to
∂u
∂y
, following Prandtl[0].

2 ∂u

τh = −νt S; νt = lu (x, y) (3.26)
∂y
Mapping mixing length over the entire domain is a difficult effort.
However, the terms related to mixing length enter in x-momentum and y-
momentum equations, playing a key role concerning wake recovery. Iungo[0]

suggests that mixing length parameter is roughly invariant in near-wake
region and he proposes linear variation in the far-wake region. This
formulation lead to very simple parametrization of the mixing length
parameter depending essentially on one single model parameter:

(x0 − n0 )l , if x0n ∈ Xn0 and if yn0 ∈ Yn0
n s
lun (x0n , yn0 ) = (3.27)
0, otherwise
With the notation h ·0n i, n-th turbine local system is indicated. Xn0 =
{x0n m0 ≤ x0n ≥ n0 } and. Yn0 = {yn0 |yn0 |0 ≤ D}, where m’ represents the
near wake region limit, while n’ defines the region affected by the wind
turbine interaction. Providing this formulation only three parameters have
to be estimated, but it might occurr that the formulation is not be able
to capture all the complexity of the wake physics in this way. Moreover,
m’ can be estimated a fortiori, leading to very simple formulation suitable
for online estimation. ls is the main tuning variable and expresses the
amount of wake recovery. According to Lee[0], as said in previous section,
ls should be influenced by ambient turbulence for sure, but also by turbine
characteristics as thrust coefficient and yaw angle. This formulation might
be not valid in case of two superimposed wake. Indeed, in FLORIS the two
wakes, and, therefore, the wake growth rates are considered separately, then
summed up following Katic formula, while in this case the two contributions
are considered together in the x-momentum and y-momentum. In case of
small wind farm, where several wakes interact each others, this fact might
lead to estimation error.
At last, turbine model is provided to estimate wind turbine perfor-
mances thanks to classical non-rotating actuator disk. Thrust force is
uniformly distributed on the wind turbine mesh cells. This is a critical
point, since this distribution is not consistent with the problem physics.
The distribution of thrust force has greater impact in near-wake region
than in the far-wake one. Therefore, estimation errors are expected in
near-wake region. Forces decomposition along x and y direction accounts

for wake deflection. CT0 is used instead of CT , according to the disk-based
thrust by Meyers[0].
2
1 2 1 Urp
T = %ACT U∞ = %ACT
2 2 1−a
1 CT 1
= %A 2
2
Urp = %ACT0 Urp
2
(3.28)
2 (1 − a) 2
where Urp is the flow velocity at the rotor plane computed following
actuator disk theory. Of course, induction is not considered directly, but it
derives from the numerical computations. Furthermore, a tuning parameter
is introduced to match the thrust force and flow velocity at the rotor plane,
which can be assumed to be independent of mixing length turbulence.
At the same way, the power is calculated from the resolved flow velocity
components. As already said, both CP and CT are input parameters. For
these reasons, WFSim performs at its own best, if coupled to high fidelity
software, like SOWFA. In this way, as reported by Doekemeijer, tuning
variables and performances coefficients can be estimated thanks to Kalman
filter.
3.3.2 Results
WFSim simulations are compared to experimental wake data. The

analysis is carried out, as in FLORIS case, in terms of wake features, as wake
deficit and wake deflection. Of course, WFSim is more computationally
expensive than FLORIS, but theorically, it is more accurate. Furthermore,
WFSim is dynamic tool, considering wake evolution and transition in case
of changes to environmental condition. Simulation parameters are shown in
Table 3.1, the mesh was defined consistently with the condition illustrated
by Boersma[0]. The mesh is not well refined, that aims to keep the code
lean and to allow real-time simulation, mandatory in closed-loop control.
In this way, using Kalman filter, it is possible to estimate online model
parameters, guaranteeing good matching in real-time situation. Turbine

parameters are set up as measured in experimental tests. Mixing length
turbulence is chosen to minimize RMS in zero-yaw case and then kept
constant for other yawed condition.
Mesh Time
L[m] N Points per D Aspect Ratio Tend [s] dT [s] Courant
x 34 50 1.617 1:4.87 10 0.1 0.822
y 14 100 7.857
Table 3.1: WFSim simulation parameters
Results are provided in Figure 3.16. In terms of spaced-average velocity

over turbine diameters, the model is capable to reproduce wake experi-
mental data. Maximum deficit is well caught by the model, however the
gradient in y direction is much higher in WFSim models. The resulting
wake shape shows speed up at the lateral side. This effect was corrected in
the latter version of WFSim, where the bidimensional continuity equation
is corrected to avoid lateral side speed-up. Concerning this analysis, the
problem is still evident and it might be due to bidimensional flow, which
accelerates according to continuity equation. To solve this problem different
formulation of mixing length is introduced within WFSim code, increasing
the domain of lmu along the y direction.
Another problem shown by the results is wake deflection. Following
WFSim simulations, wake centre is shifted laterally not enough. Moreover,
the trend of the wake centre along the downwind position is physically
unreasonable. In facts, according to the budget study of RANS by Bas-
tankhah, wake deflection increases over axial direction in logarithmic way.
WFSim is deeply in contrast with this fact, since wake deflection decrease
over the x direction. In addition, this effect has an important impact on
active yaw redirection control, which appears to be low effective. To solve
this problem, P&P BEM supplies different distribution of the thrust force
over the turbine diameter. The distribution is skewed in yawed condition
and allows to improve wake deflection. Nonetheless, the problem might not
be fixed changing the thrust distribution only. Results are very sensitive to
mesh grid and, therefore, the problem might be detected in cells dimension.
However, restriction on computational time limits the size of the cells.
Of course, changing the distribution of the thrust force, mixing length is
changed also over the x direction. From this preliminary analysis, WFSim
is reported to be not suitable for yaw control design in this configuration.
Since main problem of WFSim are related to wake deflection, it might
be useful for axial induction control. However, concerning yaw control, it
needs to be coupled to a high fidelity model or wind farm real measurements
to estimate online model parameters.
(a) γ = 0◦ (b) γ = 10◦
(c) γ = 20◦ (d) γ = 30◦
Figure 3.16: Wake data: WFSim and experimental at high TI for different
yaw angle configuration {0◦ , 10◦ , 20◦ , 30◦ }
3.3.3 New turbine and mixing length model

P&P BEM supplies different distribution of the axial force on the
rotor diameter, as reported in Figure 3.17. Three different effects might
affect WFSim simulations. Since in the centre of the rotor axial force is
minimum, typical near wake region is properly reproduced, even if it is not
important for control design. The efficiency of the blades is maximum at
about the 75% of the radial coordinate, then decreases. Lower axial forces
at the lateral sides might avoid the speed-up effect. At last, yawing the
rotor, the distribution skewness increases, and, therefore wake deflection
is improved. Since results are deeply changed by this new distribution,
thoerically, to balance this result, new distribution of the mixing length
should be provided along y direction. However, from literature review,
Figure 3.17: Axial force distribution and blade factor, implemented in WFSim.
Turbine parameters are the same as in September session λ = 8.15,
β = 0.4259, γ = 0◦
Iungo[0] and Madsen[0] state that mixing length can be assumed constant
along the y direction. In this case, the wake is distorted in the centre, due
to impractical zero axial thrust force in that zone. New parameter, Blade
factor is introduced and reasonably fixed at 30% of the blade. Within
the blade factor, axial force is kept constant to the value of the axial
force evaluated at the blade factor. To maintain overall axial force, the
distribution is scaled. In this way, the axial force integrated over the rotor
diameter is not changed.
According to these significant additional changes to turbine-flow inter-
actions, new mixing length is introduced. Adding some degrees of freedom
to mixing length accounts for better matching with experimental results.
On one hand the new formulation should be effective in reproducing ex-
perimental results, on other hand mixing length has still physical meaning
and it cannot be turned upside down with respect to the original one. In
this regard, a good proposal is formulated by Madsen[0]. As said, wake
recover is affected mainly by ambient turbulence and slightly by turbines
parameter. Original mixing length in WFSim is responsible for ambient
turbulence only and it is assumed to vary linear over the x direction.
According to eddy viscosity concept, in this case slightly different with
respect to Equation 3.26:
νt = lm Umi (3.29)
Where lm and Umi are suitable turbulence length scale and velocity. νt
provided by Madsen is chosen to be dependant of ambient turbulence and
wake shear layer:
νt = F2 k2 b(U0 − Udef,min ) + F1 νT A (3.30)
With F2 and F1 filter functions, depending on downstream x position. U0

is the unperturbed velocity, Udef,min is the minimum wake velocity, k2 is
an empirical constant for the wake flow field and νT A is the viscosity term
directly linked to the ambient turbulence level. In WFSim eddy viscosity
is expressed by the Equation 3.26, but these two terms participate to eddy
viscosity through the filter functions. Following Madsen, at the end of the
dissertation, both terms are supposed to be dependent of model parameters.
Therefore, this formulation really matches to wind farm closed-loop control
paradigm. All the complexity is discharged on model parameters, online
tuned. In this way mixing length instead of eddy viscosity is formulated
by means of filter functions, which allow to decouple and weight the two
effects separately. F1 filter function varies linearly from 0 to 1 over the
distance 0 to 2D, while F2 varies quadratically up to 10D.
Furthermore, in y direction smooth trend is chosen from the tip of rotor
diameter to one additional external diameter. This solution has physical
meaning and decreases speed up at the lateral side. New resulting mixing
length mapping is shown in Figure 3.18.
Figure 3.18: Mixing length turbulence new mapping, according to P&P BEM
new axial force distribution
Results achieved with new code modifications are presented in Figure

3.19. Simulations are carried out with same parameters, as in Table 3.1
and model parameters are set up to improve correlation between numerical

results and experimental data in zero-yaw case, then kept constant in
other cases. Radial velocity gradient is more similar to experimental in
this case and at least, in this sense, results are improved. Increasing yaw
angle, new model shows better agreement with experimental results, in
term of average wake deficit. However, other critical points, previously
highlighted, are still evident. Even if wake centre is better estimated, wake
deflection trend is in contrast with experimental data and BPG model,
since it decrease with downstream position. To solve the problem, in offline
configuration, strong dependence over y direction might be implemented,
considering also the skewness of the thrust force distribution. Another
problem is related to near to far-wake transition which is shifted up to 5D
downwind position. Flat zone is visible at 5D in zero-yaw condition and it
might be due to new axal force distribution, which is lower in the central
zone of the rotor diameter. Further investigation concerning mixing length
turbulent are mandatory for this offline configuration of WFSim to balance
new axial distribution.
In this regard, possibilities of yaw control operation in wind farm are
enhanced. Compared to FLORIS, WFSim shows more limits in reproduc-
ing wake flow field in offline configurations, without considering its own
more computational power required. Therefore, in offline configuration to
estimate optimal wind farm set-point, FLORIS is preferable than WFSim.
However, it has to be noticed that FLORIS depends at least on four model
parameters, while the original version of WFSim depends on one single
mixing length turbulence parameter. For this reason, for online closed-loop
control, WFSim is easier to manage, if online tuning required. At last, even
if FLORIS depends on four parameters, it is more stable, in terms of wake
flow field estimation, while WFSim relies very much on model parameters,
which strongly affects final results. This is due to the different nature of
the model, since WFSim is essentially numerical CFD tool, while FLORIS
is a self-similarity-based heuristic model. In last Chapter, starting from
these consideration, optimization is carried out thanks to FLORIS in 3x1
wind turbines array for different inflow condition.
(a) γ = 0◦ (b) γ = 10◦
(c) γ = 20◦ (d) γ = 30◦
Figure 3.19: Wake data: WFSim with new mixing length and turbine model
and experimental at high TI for different yaw angle configuration
{0◦ , 10◦ , 20◦ , 30◦ }
Chapter 4
From standard controller

through wind farm algorithm
In this chapter, a proposal of the algorithm for wind farm operations

is illustrated. The best operation strategy is difficult to be evaluated
since many variables has a considerable influence on the problem. In this
scenario, the selected one among the possible approaches for redirecting
the wake is yaw control. The reason for this choice is the fact that allows
to redirect wakes and to reduce power capture in one fell swoop. In facts,
also IPC shows this feature, however, nowadays, it is more difficult to
be managed due to rotor imbalance. Moreover, estimating downwind
flow field is still being investigated for this technique. This algorithm is
closed-loop algorithm, due to presence of an observer for estimating state
variables, thanks to blade root bending moments measurement. However,
FLORIS model parameters are pre-tuned and, therefore, look-up table
is computed offline. In this regard, the reference trajectory is computed
in open-loop. In this case, the effectively of the control totally relies on
model accuracy. Validation procedure performed in previous chapter is
mandatory whenever the open-loop approaches are used. According to
standard procedure, validation tests are performed at different conditions
with respect to the identification tests suitable for model parameters tuning.
This is a proof of the validity of the model to a certain extent. Since many
61
Chapter 4. From standard controller through wind farm
62 algorithm
information are available within real wind turbines network, it is reasonable

to think about online parameter estimation. Nonetheless, until flow field
measurements will not be easily accessible, giving concrete form to this
possibility would be difficult.
The control architecture is composed of three parts:
• Wind shear and direction are estimated by means of an observer
• FLORIS pre-tuned Look-up table provides reference value for rotor

speed, pitch and yaw angle
• Torque, pitch and yaw control guarantee to maintain reference tra-

jectory
Unfortunately, control architecture has not been tested experimentally

in its entirely. Indeed, G1 models do not have strain gauge at the root
blades for measuring the bending moments. Shear and direction observer
is tested in numerical simulations thanks to FAST, an aeroelastic high
fidelity tool, while control is tested in wind tunnel regularly. Normally,
wind direction is estimated as a system state variable involving certain
uncertainty. While, in wind tunnel environment, the control “knows” the
wind direction exactly and, therefore, the performances are not affected by
this fact.
To design the observer, torque and pitch are required to properly
reproduce the wind turbine behaviour. In first section, briefly overview
of standard control greedy control is introduced. Since the object is
to estimate the performances of the observer, many simplifications are
carried out concerning the model and the control design. Then, physical
considerations leading to the observer are explained and the observer is
tested in terms of reference tracking and disturbances rejection. Turbulence
intensity may affect considerably observer performances. At last, greedy
control is considered in wind farm operations, where wind turbines are
requested to work in yawed condition. In particular, torque control might
be adjusted as function of the yaw angle to correct the TSR to the optimal
one.
4.1. Reference torque and pitch controller 63
4.1 Reference torque and pitch controller

Baseline control design has two main functions. On one hand generator
torque control regulate rotor speed, in below rated, on other hand, pitch
control keeps constant rotor speed in above rated. Torque control aims to
keep constant TSR in below rated, in this way, the power and the thrust
coefficient can be assumed as constant value, if Reynolds dependency can
be neglected. In below rated Pitch control is turned off, while it is triggered
at high velocity to avoid rotor reaching too high velocity. In this way it
regulates rotor speed at certain set point, while constant generator torque
and, therefore, also power are maintained at the design values.
Greedy control is required to properly design the observer. Indeed,
the observer reconstructs wind direction and shear thanks to blades root
flapwise in and out-of-plane bending moments measurements. In this
regard, the control regulates wind turbines state variables and has consid-
erable impact on the blades root bending moments values. The observer
design procedure is the same, however, the resulting calibration matrix is
completely different turning off the control.
4.1.1 Baseline torque control design

Typical large commercial wind turbines are variable-speed machines
due to the presence of an inverter. The standard industry controller is
implemented following the procedure by Wright[0]. Generally, three regions
can be distinguished concerning torque control. Small pitch angle results
in sufficient aerodynamic torque to overcome bearing friction allowing the
rotor to start up from rest. Once the generator speed has accelerated to
sufficiently high values, the generator torque is switched on and power is
produced normally. Now the turbine is operating in Region 2, which is the
first operating region.
In region 2, pitch is held constant to maximize power coefficient and
generator torque is controlled in such a way that, keeping constant TSR,
allows to operate at the optimum power coefficient. The region is defined
64 algorithm
as quadratic function of rotor speed:

1
Popt %πR2 CP,opt U∞3
Topt = = 2 =
ω ω
3
1
%πR2 CP,opt (ωR)
λ3 1 CP,opt
= 2
= %πR5 3 ω 2 (4.1)
ω 2 λ
Therefore, the only parameter to be computed is the torque constant:
1 CP,opt
k = %πR5 3 (4.2)
2 λ
In region 3, generator torque is simply held constant at rated torque.

In some machines region 3 generator torque control is set to maintain
constant power instead of constant torque, with generator torque inversely
proportional to rotor speed. If rotor speed is tightly controlled to rated
speed in region 3, this type of control will be almost identical to setting
generator torque to maintain constant torque in region 3.
Transitional region 2 1/2 is needed if rated torque is reached at higher
speed than the rated one. Thus, a new region is introduced where torque
depends linearly on rotor speed, reaching rated torque at slightly different
speed. In this work, region 2 1/2 is not required, since no restrictive
conditions are imposed on rotor speed.
Figure 4.1: Simulink block diagram for torque control

Resulting block diagram for closed-loop control is shown in Figure 4.1,

translating in Simulink R environment operating wind turbine regions
(Figure 4.2). Fast is run from Simulink by means of Level-2-S Function. In

this way, control design can be simple managed by Simulink. In addition,
reference power is provided to FAST.
Figure 4.2: Variable-speed turbine operating regions
While for pitch control several design methodologies are developed,

torque control always shows this control architecture, since aerodynamic
efficiently is directly linked to TSR. Therefore, degrees of freedom are
minimum to determine control objective for torque regulation. Instead,
for pitch control many choices are available to define control targets,
as mitigate damaging fatigue loads and detect fault conditions. In this
scenario, advanced multi-input multi-output (MIMO) multivariable control
design methods, such as those based on state-space models, can be used
to meet these objectives and use all the available actuators and sensors
in a reduced number of control loops. However, these techniques are very
complex and, of course, they have many application, but in this work,
the main object is to investigate yaw control effectiveness in wind farm
operation. Therefore simple pitch control is chosen.
66 algorithm
4.1.2 Baseline pitch control design
In region 3, pitch control is turned on to maintain constant rotor speed

at designed value. Simple PI control is implemented in closed-loop to
ensure null steady-state error and to improve the dynamic response of the
system:
∆Ω̇ = A∆Ω + B∆ϑ + Bd ∆w (4.3)

Z
∆ϑ = Kp ∆Ω + KI ∆Ωdt (4.4)
Where A represents the state-space matrix, B the input matrix for

the control action and Bd the input matrix for the disturbances. PI is
suitable to regulate rotor speed, even if the state-space matrix is not
defined. However, to design properly integral and proportional gains of
the controller, input-output relations between pitch angle and omega has
to be established. Several simulations are performed at different wind
speed, where the rotor speed at the equilibrium position (reference value)
is perturbed by pitch step. The results are shown in Figure 4.3, where
the time reponses are fitted by first order system. In this way, state-
dependant transfer function suitable for control design can be evaluated.
The simulations are performed at ideal condition of uniform constant inflow.
For each inflow conditions, a precise pitch value guarantees to maintain
the designed reference rotor speed. Supposing that wind measurements
are not available, state-space matrix can be expressed as function of the
pitch angle.
This procedure leads to two advantages: possibility to choose the poles
of the closed-loop system and avoid the dependence on the inflow condi-
tion. Therefore, dynamic performances, as response time and maximum
overshoot, are not affected by the inflow conditions. Of course, the control
gains will depend on the state of the system to achieve this result. This
solution is implemented in Simulink by means of look-up table.
Design requirements are defined depending on the values of natural
frequency and damping. PI zero is fixed by cancellation and the natural
Figure 4.3: Rotor speed response to pitch step input: Equlibrium position is
perturbated at different inflow condition. Transfer function are
estimated by this process
frequency of the closed-loop is reasonably fixed at ω = 0.6rad/s, as

suggested by Wright[0]. Gain scheduling technique provides correct values
to the control system. Actuation pitch is lower and upper bounded, due to
practical issues. Control architecture is shown in Simulink block diagram
4.4, where saturation block is placed consistent with actuator physical
limits. Integral action might lead to delay between disturbances and control
action, in presence of pitch saturation. Typically, in region 2 the blade pitch
is saturated at the lower pitch limit. The negative speed error is fed to the
integrator part of the controller and the integrator continuously integrates
this negative error resulting in a larger and larger negative pitch angle.
Whenever wind flow conditions changes getting operating point to region
3, it takes a long time for this positive speed error contribution to cancel
the effects of the negative pitch angle contribution that has been built up
from integration. Of course, this effect, the so-called windup, might affect
the performances. To avoid this problem, anti-windup implementation for
the PI controller is introduced. Feedback gain multiplies by the difference
68 algorithm
between saturated signal and the original one provides anti-windup term.
By trial and error procedure, the gain is set to 1000rad/s/rad. In this
way, if the pitch is not saturated, anti-windup signal is zero, while, if the
pitch is saturated, the error is not integrated, and, therefore a delay in
control action is avoided. Both coupled torque and collective pitch control
defines the reference control for wind turbines.
Figure 4.4: Simulink block diagram for pitch control
4.1.3 Simulations
Numerical analysis is carried out by means of FAST. This analysis
presents many approximations with respect to experimental tests. Lift and
Static coefficients are constant with Reynolds, the performances of the wind
turbine are reported in Figure 2.9. Elastic transmission is considered as
infinitely rigid. Thus, rotor speed and generator speed are perfectly equal.
Besides of these wind turbine related approximations, inflow conditions are
rather different. Indeed, the simulations are performed at uniform wind
speed. Shear and time history of the inflow wind are consistent with the
wind tunnel data, however the wind is uniform along the y direction.
These tests are not a check for the control design but aim to understand
the behaviour of the wind turbine in ambient turbulence. In Chapter 2, it
has illustrated that with increasing ambient turbulence, the performances
decrease and the control gets trouble to keep constant the TSR. In this
analysis, this fact cannot be seen, due to many approximation and, in
particular, the which one related to elastic transmission. However, if the

operating point of the wind turbine changes, the performances might be
affected significantly. Concerning wind farm operations, this fact needs to
be further investigated.
Typically, in control design performances are evaluated in terms of
reference tracking and disturbances rejection. As said, main topic of this
section is the observer design. Therefore, two simulations only about
control performances are reported for the sake completeness.
Figure 4.5: Ideal case of step input response
In Figure 4.5, step input response is illustrated. This case is totally

ideal, since inflow conditions are steady and uniform. The control works in
region 3, pitch actuation brings back the rotor speed to its reference value.
Since the case is perfectly ideal, the response is an ideal exponential. Pitch
reaches proper value in a short time, however rotor dynamic imposes more
time to rotor speed to achieve its reference value.
In Figure 4.6, disturbances rejection is tested. In this case, inflow
70 algorithm
conditions are taken consistent with the pitot measurements in wind

tunnel experiments and shear is reproduced with experimental power law.
However, the wind is uniform, thus constant along y direction. Transient
region is included to show response time. Since inflow condition is in the
above rated region, pitch control is involved. For this turbulence index,
T I ≈ 5%, the control is capable to maintain rotor speed to reference
value. Disturbances rejection works reasonably well and blade pitch do not
saturate.
Figure 4.6: Disturbances rejection with inflow condition at T I ≈ 5%. The

inflow is chosen consistent with wind tunnel experimental data.
Red curve in rotor speed is the reference in region 3.
Of course, this control is simple and not optimized in terms of perfor-

mances and targets, but it is very reliable. For this reason, it is applied to
commercial wind turbines generally. Thus, this methodology is chosen to
be consistent with practical cases.
4.2. Observer for estimating wind direction and shear 71
4.2 Observer for estimating wind direction and

shear
Final goal of this work is to perform yaw control to optimize wind farm
power production and to reduce fatigue-damaging loads on the blades.
All these benefits have to be weighted against the cost of frequent yaw
actuation. In fact, yawing the nacelle and rotor of a modern large wind
turbine requires moving a very massive structure (for example weighing
in excess of 150 tons for a typical 3.0MW machine) [0]. This means
overcoming the static friction in the yaw bearing when initiating the
maneuver and slowing down the motion once the new desired configuration
is reached. Hence, to reduce cost, complexity, size and maintenance of the
yaw actuation system, its duty-cycle must be carefully limited.
Unfortunately, high-quality wind measurements suitable for wind tur-
bines control are difficult to obtain. In fact, on-board yaw sensors, typically
wind vanes, are affected by various sources of inaccuracy, including dis-
turbances caused by the rotor wake and its turbulence, the presence of
the nacelle, the periodic passing of the blades upstream of the instrument.
Nowadays, accurate measurements are available from lidar. However, these
instruments are very expensive, thus they are not mounted on standard
wind turbines. Commercial existing sensors, even when well compensated
for all sources of error, can only provide point information, usually at
hub-height. For the very large diameters of modern wind turbines this
limitation can provide for an additional source of error, and a more global
view of the wind direction over the rotor disk would be more appropriate.
In this scenario, designing an observer is necessary to support the
operation of wind turbine control systems. Generalized wind observers
allow to estimate several wind states, as yaw and tilt angle or vertical and
horizontal shear. These observers are designed providing an expanded set
of blades dynamic equilibrium equations to Kalman filter. This formulation
is very computationally expensive and complex, thus not convenient in
wind farm control.
Simpler approach is developed hereafter, following the procedure illus-
72 algorithm
trated by Bottasso [0]. Since turbulence intensity level and wind direction
are input variables for both wake models, FLORIS and WFSim, they
are the reasons of interest of the observer. While wind direction can be
directly estimated by the observer, turbulence intensity can be derived by
the wind shear. Basic idea of the observer design is to exploit 1P response
of the measured flapwise bending moment. Starting from the physics of
the problem, general linear formulation is derived, whose parameters are
obtained by system identification, based on the least-square method.
4.2.1 Formulation
To achieve general formulation, physical terms involved into the dy-

namic equilibrium has to be identified. For a flapping hinged rigid blade:
d2 β γ

4 dβ γ
2
+ 1 − cosψ(Ū0 + q̄l) + (K + 2Bcosψ + Ū0 sinψ)β =
dψ 8 3 dψ 6
γ q̄
= (A − sinψ) − 2q̄cosψ−
2 4
γ λ 2ϑp K1 V̄0
− (Ū0 + q̄l) + + cosψ (4.5)
2 2 3 4
Where ψ is the azimuthal angles, ω the angular velocity, K the non-

dimensional flapping frequency, A an aerodynamic term accounting for
non-dimensional inflow, λ the TSR, ϑp the pitch, B the gravity term, γ the
lock number, V̄0 the non-dimensional free-stream, Ū0 the non-dimensional
cross-flow, K1 the vertical wind shear, q is non-dimensional yaw-rate and l
the non-dimensional yaw moment arm(for detail see [0]).
β can be written according to Fourier series arrested to the first order:
β = β0 + β1c cosψ + β1s sinψ (4.6)
Plug the Equation 4.6 into the Equation 4.5 and rearranging the terms, it
is possible to invert the system and compute cross-flow and vertical wind
shear as:
" # " #
Ū0 0
= +
K1 V̄0 −16Bβ0 /γ
" #" #
3/4 −6(K − 1)/γ β1c /β0
+ +
−8(K − 1)β0 /γ − 3A3 −β0 + 24A3 (K − 1)/γ −β1s /β0
" #
−3/(4β0 )
+ q̄ (4.7)
−16/γ + A3 (3/β0 − 4l)
This equation provides linear observation state-space model, based only

on the blade flap angle and the yaw rate during the maneuver. Cross-
flow and shear induce different effects on the blade such that using one
single observer both can be estimated. For considerations regarding the
invertibility of the matrix and, therefore the observability of the system
see Bottasso [0].
Since blade flap angle is typically not available from wind turbine
measurements, this formulation is impractical. However, supposing direct
relation between flapwise bending moment and blade flap angle, general
observation model can be introduced with similar one to the previous
one, which is validated by the physics of the problem. In this case, the
matrices coefficients are not linked to the physical quantities but are
identified using least-square procedure by identification tests set. The
general wind-scheduled observation model can be reformulated:
w̄ = w̄0 (V0 ) + Ā(V0 )m̄ + B̄(V0 )q̄ (4.8)
Where w̄ = (Ū0 , K1 V̄0 )T and m̄ contains the 1P load harmonic amplitudes

of in-plane and out-of-plane bending moments. Both components are
needed, since the flaps is not always orthogonal to the rotor plane. In facts,
as the blade is pitched, the flap motion assumes a component parallel to
the rotor plane.
Identification procedure is performed for different incoming flow velocity,

such that observation matrix is composed of wind-dependent coefficient.
74 algorithm
Incoming flow velocity is supposed to be point measured at hub height.

Several aeroelastic simulations are run using FAST. In general, for this kind
of observer numerically-calibrated coefficient are not consistent with real
cases. Therefore, whenever the observer is tested on a scaled or full-scale
wind turbine the observation matrix has to be re-calibrated by means
of high-quality field measurements [0]. Standard least-square procedure
is carried out collecting all suitable osservations and associated driving
inputs:
 
h i 1 1 ... 1
W = w̄1 , w̄2 , . . . , w̄n ; M = m̄1 m̄2 ... m̄n  (4.9)
 
q̄1 q̄2 ... q̄n
All unknown coefficients are collected into the observation matrix:

h i
T(V0 ) = w̄0 (V0 ), Ā(V0 ), B̄(V0 ) (4.10)
Then, coefficients are computed by least-square procedure:
T(V0 ) = WMT (MMT )−1 (4.11)
The driving 1P loads harmonics appearing in the input vector can be

computed by demodulation of the projection of the blade root bending
moments or by Coleman transformation[0]. In this work, demodulation
is chosen instead of Coleman transformation, since it provides effective
filtering of the rapid fluctuations and transient at the cost of increasing
the delay. As drawback, demodulation needs the revolution period, which,
of course in practical case is not constant. Therefore, demodulating over a
not precise period can lead to large error estimation. Moreover multi-blade
coordinate transformation of Coleman[0] is easier than demodulation to
be implemented. Demodulation projects the bending moment over the
azimuthal component. The resulting formulation is very close to Fourier
transform:
Z ψ−2N π
1
m0b = mb (t)dψ (4.12)
2N π ψ
Z ψ−2N π
1
m1cb = mb (t)cosψdψ (4.13)
Nπ ψ
Z ψ−2N π
1
m1sb = mb (t)sinψdψ (4.14)
Nπ ψ
Final observer architecture is illustrated in Figure 4.8. Once the sig-

nals are obtained, filtering is preferable to remove fast fluctuations, as
the maneuvering is limited to significant enough misalignment with slow
dynamic.
Figure 4.7: Complete block diagram of the observer, comprehensive of filtering,

gain scheduling of the observation matrix and demodulation for
the projection of the moment component. Input variables for the
observation are blade root bending moment, rotor speed and the
incoming normal flow point measured at the hub height suitable
for gain scheduling
76 algorithm
4.2.2 Simulations
FAST is used to identify observation matrix coefficients. Simulations

approximations include Lift and Drag static coefficient, inflow and trans-
mission, as previously explained. The identification is carried out at several
wind speed in both below and above rated condition. Gain scheduling
is performed in function of incoming wind speed in a range of 3.5 m/s
to 8.5 m/s, with a step of 0.2 m/s, while operating conditions ranging
from −20◦ to 20◦ with a step of 5◦ and shear power law exponent equal
to ν = 0.1, 0.2. Considering in-plane and out-of-plane loads identification
by least-square procedure is capable to fit the observed cases, as shown in
Figure 4.8.
Figure 4.8: Identification set to determine the coefficient of the observation

matrix through least square matrix. Below and Above rated data
.
While offline matrix calibration can be easily performed and gives

reliable results, online tests needs adjustments to reduce the effect of
turbulence. Since the performances of the observer should be evaluated in
terms of steady state error, maximum overshoot, delay in transient and
disturbance rejection, simulations are carried out initially at constant and
uniform inflow conditions, then in ambient turbulent conditions. In both
cases, input noise on measurement is neglected.
In this regard, first simulation illustrated in Figure 4.9, helps to under-
stand the performances of the observer in ideal case. Maximum overshoot
is about 11◦ , but the transient is short and the observer is capable to track
the reference signal in less than 1s, while during the ramp the observer
exhibits a small delay of about 0.4s, due to low-pass filter and moving
average filter. Moreover, the maximum steady state error arises at zero-yaw
condition and it has a magnitude of −0.2◦ , which, of course is a reasonable
quantity.
Figure 4.9: Observer tracking a reference ideal ramp of 2◦ /s.

78 algorithm
Ambient turbulent results are shown in Figure 4.10. Wind speed time
history and the shear power law of 0.079 are consistent with the data
measured by the pitot tube during wind tunnel test in low turbulence
configuration. Therefore, the inflow conditions are similar to real case,
despite of the perfect spatial coherence over the rotor area. Theoretically,
the coherence has a negative impact on the performances of the observer.
Real turbulence varying over the rotor area should balance itself, helping
disturbance rejection. Instead of previous case, this time gain scheduling
is involved.
Figure 4.10: Observer performances with inflow condition at T I ≈ 5%. The

In this case, disturbance rejection capability can be analysed. Small

oscillations are present around the mean value which matches the reference
value. Wind shear estimation is more stable than wind directions. Filtering
the output signal allows to reduce these high frequency oscillations induced
by turbulence, in exchange for delay in signal. Since the principal idea is
to evaluate wind ambient turbulence and direction variation over a large

period, the delay has limited importance in this analysis.
Increasing turbulence up to ≈ 10%, oscillations increase even if the
mean value is properly reproduced, as shown in Figure 4.11. In this case
the effect of filtering is evident in the normal velocity at the hub height.
The performances are deeply affected by the increasing turbulence, but
still for low frequency dynamic, as in this case of yaw control, the observer
is able to reproduce mean value and therefore, the information provided
to FLORIS are almost correct. In addition, it has to be noticed that
some filter based on the rate of variation of yaw and shear might be
implemented, correcting somehow these sudden variations due to ambient
wind turbulence. Of course, decreasing cut-off frequency of the filter allows
to achieve this result, but it leads to too high phase decrement and might
lead to some kind of instability.

80 algorithm
Last case considered in Figure 4.12 is a yaw variation from 0◦ to 20◦

and back again to 0◦ in ambient turbulence of ≈ 5%. To perform additional
analysis the shear coefficient is set to 0.2, instead of 0.079. Observer is able
to catch yaw rate of change also in turbulent environment. From Table
4.1, it should be noticed that RMS depends only on turbulence intensity.
Since shear dynamic is independent of yaw misalignment, changing the
power law factor the results are still comparable. Furthermore, even if in
presence of yaw variation, the two cases with low turbulence has almost
equal RMS. Same considerations are valid for the maximum error.

inflow is chosen consistent with wind tunnel experimental data,
but the wind direction is changed.
Turbulence intensity identifies as the principal responsible for the

performances of the observer. Indeed, maximum error and RMS have
significant increment increasing turbulence. This analysis is carried out
under inflow condition with average wind speed of 5.65 m/s. The control
4.3. A proposal for the supercontroller structure 81
Wind Direction [◦ ] Wind Shear

Mean RMS errmax Mean RMS errmax
LT -0.04 0.698 2.85 0.0783 0.0092 0.0258
HT -0.432 2.87 14.76 0.196 0.0228 0.0613
LT+ramp / 0.846 2.84 0.1904 0.0191 0.0533
Table 4.1: Observer performances
is regulated, as properly explained in previous section. Since the rated

velocity is 5.8 m/s, the control continuously switches its status due to
turbulence. Thanks to gain scheduling, the two regions are described
correctly by the observation matrix. Therefore, it is expected that results
are not affected by the average wind speed velocity with the behaviour of
the observer constant even if changing the operating region of the wind
turbine, as demonstrated also by Figure 4.8.
4.3 A proposal for the supercontroller struc-

ture
As already said, wind farm control is not yet implemented in commercial
wind farms. The topic is still being actively researched and due to its
own complexity, several methodologies are proposed but each one has
advantages and drawbacks with respect each other. Wake interaction and
environmental conditions affect the problem, such that any case has to be
analysed individually. Therefore, at the state of the art, optimal control
methodology does not exist.
In this section, first recall on the effects of yawed operation on the
wind turbines are evaluated considering interaction with torque control.
Suitable degree of freedom for wind farm control are analysed and some
considerations are carried out about necessary information to define the
optimal set-point for a wind turbine. Then, a model of supercontroller
is obtained coupling all the elements mentioned in the previous analysis,
82 algorithm
as wake model, observer and optimizer. The chosen wake redirecting

techniques is yaw control.
4.3.1 Torque control in yawed conditions

Conventional degrees-of-freedom (DOFs) on wind turbines are collective
pitch and generator torque, which accounts for the axial induction of the
rotor and thus the velocity deficit in the wake. This DOFs might be used
in wind farm operation by controlling the energy capture from the flow.
In this case the set-point of TSR and pitch is directly affected by the
supercontroller logic. Also, unconventional DOFs can be used in wind
farm control as yaw, tilt (actually, not available) and independent pitch.
In this section, the effect of yaw on TSR is treated. The definition of
DOF implies itself that each DOF can be moved arbitrarily. However,
moving on optimal path, constraints are added and once the yaw setting
is determined, TSR might be fixed as function of the yaw angle.
In this regard the TSR, which is optimal in zero-yaw condition, might
be non-optimal progressively changing the yaw angle. This fact does not
represent a problem considering individual control, which theoretically
works normal to the incoming flow. However, in wind farm operation where
flow redirection is performed by means of yaw control, wind turbines might
work in yawed condition for a large time. Non-optimal TSR might affect
the performance of the wind turbine in terms of power capture and lifecycle.
Therefore, in wind farm control, classical control might be reconsidered
and should include this fact. In practical case, the torque constant defined
in the Equation 4.2 should be recalculated as function of the yaw angle
since both optimal power coefficient and optimal TSR are function of the
yaw angle. Of course, with a view to wind farm control this effect has
marginal importance, since not all the wind turbines are involved in yawed
operations. Moreover, in theory, also optimal pitch might change as the
yaw angle increases.
To show this fact simple numerical investigation is performed exploiting
the P&P BEM model explained in Chapter 2. As shown by the Figure
4.13 on the left, the model can predict the operating point of the wind
turbine working in yawed condition with reasonable accuracy. Coupling
point is obtained by the superimposition of the aerodynamic torque and
of ideal generator torque fixing the working condition as TSR = 8.15. It
has to be noticed that as the yaw progressively increases, the aerodynamic
torque decreases and TSR decreases as well. The amount of the decrement
in terms of torque and TSR does not depend on this kind of torque control,
while it depends only on the aerodynamic characteristic of the wind turbine.
In practise, this torque control is not able to control the TSR, as the wind
turbine works in yawed condition. The curves are obtained fixing incoming
flow velocity U = 5.65 m/s as in the wind tunnel tests. It should be added
that no relation between TSR and normal flow can be obtained, since the
phenomenon is extremely non linear.
Figure 4.13: On the left, numerical estimation of operating conditions when

torque control works in yawed operations. On the right, standard
torque control and numerical estimated optimal TSR. Standard
torque control does not guarantee optimal working condition in
yawed operation
On the left of the Figure 4.13, optimal TSR is illustrated. The inter-
sections between the black curve and the coloured ones represent the real
working condition with standard torque control. Of course, optimal TSR
is computed by simple numerical approach as BEM and, therefore, it is
affected by estimation error. However, significant differences between the
two points are present, such that it can not be justified by numerical error.
84 algorithm
Yaw misalignment [◦ ]
40 30 20
Operating cond. Opt. 55.9-0.4465 60.98-0.5046 66.06-0.5458
ω[rad/s] - T[Nm] Std. 60-0.41 64.6-0.476 68.05-0.526
Opt. 24.95 30.77 36.05
Power [W]
Std. 24.6 30.74 35.8
Decrement -1.4% -0.1% -0.7%
Table 4.2: Torque control in yawed condition
In Table 4.2, the results in terms of power production are illustrated. Of

course, in this case, the losses are not significant and represent a concept
more than necessity of engineering. However, since standard torque control
is completely ineffective, these losses might increase depending upon the
wind turbine design characteristics.
4.3.2 Architecture
The architecture of wind farm controller is illustrated in Figure 4.14.
The control logic is based on active wake redirection thanks to yaw con-
trol. Each element is assessed by high fidelity numerical simulation or
experimental test. However, in this work, the structure is not tested as a
whole. Further developments might be addressed to supercontroller testing
within the SOWFA environment. In the last Chapter, FLORIS is used to
estimate optimal set-point of 3x1 turbines array. The results are compared
to experimental set-point found by means of trial and error procedure
changing progressively the yaw settings of wind turbines array.
The wind farm supercontroller architecture needs the wake model to
consider the overall wind farm, including mutual interaction among wind
turbine wakes, the observer to estimate state-space variables, which cannot
be directly measured due to practical issue and standard wind turbine con-
trollers. Wind farm can work as an active network, exchanging information
in real-time. In this particular case, wake model needs the data of incoming
flow, as mean wind speed, incoming turbulence and wind direction. The
mean speed is supposed to measured, while other two input variables are
estimated by the observer. Regarding the observer, blade root bending

moment are required to estimate observation variables. Wake model is pre-
tuned making identification set for model parameters. Offline optimization
is run, giving three-dimensional look-up table. In this sense, computational
costs of wake model are inexistent or very low. The major responsible
for the computational cost is the observer, even if all the procedure is
translating in filters and gain scheduling. Nonetheless, reducing even more
the computational cost is possible by providing accurate measurements
of wind characteristics as input of the look-up table for optimal set-point
estimation. At the lower level of the architecture, individual controller
accounts for maintaining reference set-point computed by FLORIS.
Main problem might be related to exploit unconventional DOF for
control, due to legislative issues. However, due to the massive inertia of the
whole system, yaw control is supposed to be used with low frequency during
the day just to catch large variation of wind condition with slow dynamic.
In this sense accurate analysis of wind data might help to estimate real
effectiveness of this control system.
Figure 4.14: Wind farm control architecture composed of wind observer, wind
fam optimizer and greedy controller.
Chapter 5
Wind turbines array

optimization
In this Chapter, optimization process is carried out by numerical models

concerning a 3x1 wind turbines array, then tested experimentally. Several
wind tunnel tests are performed changing the yaw configuration of the
upstream wind turbines. Measuring the power capture of the overall mini
wind farm, it is possible to compute the optimal configuration and evaluate
the possible enhance in terms of wind farm efficiency. Wind direction is
changed to understand the real effectiveness of the yaw control, even if the
wind tubines are not perfectly aligned to the incoming wind flow.
In parallel, FLORIS model, pre-tuned as explained in Chapter 3, is

used to compute a surface of the possible yaw configuration, suitable
to determine the optimal configuration. The performances of the wind
turbine in yawed condition are computed including within the code the
torque control regulation. Predictive capability of the model is compared
to evaluate real possibility of performing open-loop control.
At the end, some conclusions and future outlooks are illustrated. Indeed,
from this study great potential of yaw control is revealed, but further steps
are required to improve FLORIS model and to develop real closed-loop
control architecture for wind farm operations.
87
88 Chapter 5. Wind turbines array optimization
5.1 Results
In the CL-windcon campaign of May 2018, 3x1 wind turbines array
configuration is tested in Polimi wind tunnel. Several tests are performed
varying the wind direction. In this analysis, four different cases are con-
sidered consistent with practical case. The yaw direction is obtained by
misaligning the wind turbines along the lateral direction:
1. Totally misaligned: 1D lateral displacement (equivalent to Wind

direction −11.54◦ );
2. Almost misaligned: 0.8D lateral displacement (equivalent to Wind

direction −9.21◦ );
3. Almost aligned: 0.4D lateral displacement (equivalent to Wind direc-

tion −4.58◦ );
4. Perfectly aligned: 0D lateral displacement (equivalent to Wind direc-

tion 0◦ );
The results are provided below. FLORIS model parameters are tuned
following the wake measurements of the first CL-windcon campaign. There-
fore, the tuning is not properly dedicated to this case. In addition, added
turbulence model parameters are not changed with respect to the original
ones. Although model parameters are unaltered, a scaling factor is intro-
duced to fit the experimental results with the numerical. This scaling factor
is constant and, therefore, the yaw trend is unaffected by this coefficient.
In practise, the optimal point is not changed by this corrective term, which
accounts for model errors, measurements uncertainty and wind tunnel
blockage effects.
The model shows good agreement with experiments for not perfectly
aligned configuration, while progressively changing the wind direction to
aligned wind turbines configuration, the experimental yaw trend becomes
rather different. In the perfectly aligned configuration, the numerical
yaw trend of the second turbine is totally in contrast with experimental
data. Probably, wake superimposition is over simplified in FLORIS model,
5.1. Results 89
neglecting complex mixing phenomena. Moreover, the optimal point is

always settled to yaw angles similar for the first and second turbines
differently with respect to experimental cases. Of course, to better predict
experimental data, dedicated tuning can be performed or some model
modification can be implemented.
Nonetheless, FLORIS is capable to predict reasonably the performances
of the 3x1 wind turbines array in wind tunnel environment. Wind tun-
nel tests are different with respect to real scale cases. Big difference is
represented by the blockage effect which can enhance the performances
considerably, due to wind speed-up. Also wake meandering is completely
absent in wind tunnel tests, leading to low efficiency in aligned configura-
tion, not representative of real cases. Moreover, the tested configuration
is the low turbulence one, where wake redirection is very effective. The
control margin illustrated are expected to be much lower in high turbu-
lence configuration. Thus, the most important application can be related
to offshore wind plants, where the turbulence intensity level is very low
allowing powerful wake redirection to yaw operations.
Although these significant differences between real environment and
wind tunnel, the strong effectiveness of yaw control is demonstrated by
these analyses. Both experimental and numerical calculations prove that
yaw control can improve the performances of the wind farms significantly,
depending upon the incoming wind conditions. The maximum increment
reaches +24.8% according to experimental data and +23.5% according to
FLORIS, for intermediate configurations. Of course, in totally misaligned
condition the power increment is about to +1.7% and for perfectly aligned
about +6.7%. These two cases lower bounds the enhancing of the power
capture and shows that there are real margin for almost every kind of
situation. These are very important percentage, resulting into a significant
decrement of the Levelized Cost of Energy (LCoE).
Figure 5.1: Test 1: numerical surface and experimental results
Numerical Experimental
Baseline CP 1.226 1.229
Optimum 6◦ /6◦ 11◦ /6◦
Optimum CP 1.25 1.25
Efficiency +1.9% +1.7%
RMS 0.0082 /
5.1. Results 91
Optimum 13◦ /14◦ 15◦ /10◦
RMS 0.0127 /
Optimum 24◦ /23◦ 27◦ /15◦
RMS 0.0197 /
5.1. Results 93
Optimum 23◦ /29◦ 31◦ /19◦
RMS 0.0236 /
5.2 Conclusions
The basic idea of this work is to develop a simple control architecture
for wind farm to optimize the performances of the overall wind farm. Wind
farm numerical models are validated thanks to wind tunnel measurements.
The control architecture is closed-loop by means of the observer, but the
wake model is pre-tuned and, therefore, the computational cost is almost
non-existent.
First, the control-oriented wake models are investigated by means of
wind tunnel data in low and high turbulence ABL. BPG wake model is
validated and dedicated modifications are implemented according to wake
measurements. From the analyses, FLORIS results to be capable to prop-
erly reproduce the flow field within the wind farm in offline configurations,
while WFSim needs online tuning of model parameters to get adequate
results. For this reason, in wind tunnel testing, FLORIS is preferable to
estimate and predict optimal working condition for wind farm.
Secondly, simple formulation of control architecture for wind farm is
formulated. Starting from an overview of reference control, an observer
for estimating wind conditions, as wind shear and direction is developed.
The observer needs blade root bending moments to estimate the inflow
conditions, which are general available for commercial wind turbines. In
this sense, the control architecture can exploit wind farm information
as communicating network, in agreement with wind farm closed-loop
paradigm. The observer is effective in low turbulence configuration, but it
needs some corrections in the high turbulence atmospheric inflow. In this
way, it is possible to estimate input variable for the FLORIS wake model.
In last chapter, the real performances of FLORIS are tested in wind
tunnel environment for 3x1 wind turbines array in low turbulence ABL. For
array misaligned with wind direction, the FLORIS model is able to estimate
the optimum set-point with good agreement with respect to experimental
results, while for array aligned with wind direction, the model is in contrast
with experimental results.
In conclusion, this work defined as a preliminary analysis in wind farm
5.2. Conclusions 95
control, since measurements are performed in wind tunnel tests, which

present some differences with respect to real full scale. Nonetheless, the
great potential of yaw control is shown and FLORIS might be the right
model to develop a wind farm control algorithm, since it is capable to repro-
duce the flow field and optimize the wind farm with good approximation,
in offline configurations also.
Future work might address to improve FLORIS model and to formulate
universal procedure for the tuning of model parameters. In addition, it
would be interesting to develop a heuristic dynamic model starting from
the BPG wake model. In this way, it would be possible to extend FLORIS
model to unsteady cases. Of course, in this sense, an inexpensive observer
might be developed for online tuning of model parameters, ensuring a
model which self-adapt to changes of the atmospheric conditions in real
time situations.
Bibliography
[0] Global Wind Energy Council–GWEC–Global. Wind Report 2016.

Retrieved on September 15, 2017.
[0] Antonio Crespo, J Hernandez, and Sten Frandsen. “Survey of modelling
methods for wind turbine wakes and wind farms”. In: Wind Energy: An
International Journal for Progress and Applications in Wind Power
Conversion Technology 2.1 (1999), pp. 1–24.
[0] Majid Bastankhah and Fernando Porté-Agel. “Experimental and the-
oretical study of wind turbine wakes in yawed conditions”. In: Journal
of Fluid Mechanics 806 (2016), pp. 506–541.
[0] Gunner C Larsen et al. “Wake meandering: a pragmatic approach”.
In: Wind energy 11.4 (2008), pp. 377–395.
[0] Gunner C Larsen et al. “Dynamic wake meandering modeling”. In:
Risø National Laboratory, Technical University of Denmark, Roskilde,
Denmark, Risø (2007).
[0] Martin OL Hansen. Aerodynamics of wind turbines. Routledge, 2015.
[0] Andrew Ning et al. “Development and Validation of a New Blade
Element Momentum Skewed-Wake Model within AeroDyn”. In: 33rd
Wind Energy Symposium. 2015, p. 0215.
[0] BM Doekemeijer. “Enhanced Kalman filtering for a 2D CFD Navier-
Stokes wind farm model”. In: (2016).
[0] Sjoerd Boersma et al. “A control-oriented dynamic wind farm model:
WFSim”. In: Wind Energy Science 3.1 (2018), p. 75.
97
98 Bibliography
[0] Torben Knudsen, Thomas Bak, and Mikael Svenstrup. “Survey of

wind farm control—power and fatigue optimization”. In: Wind Energy
18.8 (2015), pp. 1333–1351.
[0] EA Bossanyi. “Individual blade pitch control for load reduction”. In:
Wind energy 6.2 (2003), pp. 119–128.
[0] MS Adaramola and P-Å Krogstad. “Experimental investigation of
wake effects on wind turbine performance”. In: Renewable Energy 36.8
(2011), pp. 2078–2086.
[0] M Bastankhah and F Porté-Agel. “Wind tunnel study of the wind
turbine interaction with a boundary-layer flow: Upwind region, turbine
performance, and wake region”. In: Physics of Fluids 29.6 (2017),
p. 065105.
[0] David G Wilson et al. Combined Individual Pitch Control and Active
Aerodynamic Load Controller Investigation for the 5MW UPWIND
Turbine. Tech. rep. Sandia National Laboratories (SNL-NM), Albu-
querque, NM (United States), 2009.
[0] Tony Burton et al. Wind energy handbook. John Wiley & Sons, 2011.
[0] H Snel and JG Schepers. Joint investigation of dynamic inflow effects
and implementation of an engineering method. Netherlands Energy
Research Foundation ECN, 1995.
[0] IS Gartshore and KA De Croos. “Roughness element geometry required
for wind tunnel simulations of the atmospheric wind”. In: Journal of
Fluids Engineering 99.3 (1977), pp. 480–485.
[0] Carlo L Bottasso, Stefano Cacciola, and Xabier Iriarte. “Calibration
of wind turbine lifting line models from rotor loads”. In: Journal of
Wind Engineering and Industrial Aerodynamics 124 (2014), pp. 29–45.
[0] Jason M Jonkman and ML Buhl Jr. “FAST user’s guide, national
renewable energy laboratory”. In: No. NREL/EL-500-38230, Golden,
CO (2005).
Bibliography 99
[0] LJ Vermeer, Jens Nørkær Sørensen, and Antonio Crespo. “Wind

turbine wake aerodynamics”. In: Progress in aerospace sciences 39.6-7
(2003), pp. 467–510.
[0] Majid Bastankhah and Fernando Porté-Agel. “A wind-tunnel investiga-
tion of wind-turbine wakes in yawed conditions”. In: Journal of Physics:
Conference Series. Vol. 625. 1. IOP Publishing. 2015, p. 012014.
[0] Niels Otto Jensen. “A note on wind generator interaction”. In: (1983).
[0] Takeshi Ishihara, Atsushi Yamaguchi, and Yozo Fujino. “Development
of a new wake model based on a wind tunnel experiment”. In: Global
wind power 6 (2004).
[0] Sten Frandsen et al. “Analytical modelling of wind speed deficit in
large offshore wind farms”. In: Wind energy 9.1-2 (2006), pp. 39–53.
[0] Jennifer Annoni et al. Analysis of Control-Oriented Wake Modeling
Tools Using Lidar Field Results. Tech. rep. National Renewable Energy
Lab.(NREL), Golden, CO (United States), 2018.
[0] Hendrik Tennekes and John Leask Lumley. A first course in turbulence.
MIT press, 1972.
[0] Joseph Hun-wei Lee and Vincent Chu. Turbulent jets and plumes: a
Lagrangian approach. Springer Science & Business Media, 2012.
[0] Stephen B Pope. Turbulent flows. 2001.
[0] PB Johnson et al. “On the spread and decay of wind turbine wakes
in ambient turbulence”. In: Journal of Physics: Conference Series.
Vol. 555. 1. IOP Publishing. 2014, p. 012055.
[0] Majid Bastankhah and Fernando Porté-Agel. “A new analytical model
for wind-turbine wakes”. In: Renewable Energy 70 (2014), pp. 116–123.
[0] Ángel Jiménez, Antonio Crespo, and Emilio Migoya. “Application of
a LES technique to characterize the wake deflection of a wind turbine
in yaw”. In: Wind energy 13.6 (2010), pp. 559–572.
100 Bibliography
[0] Ewan Machefaux, Gunner Chr Larsen, and JP Murcia Leon. “Engineer-
ing models for merging wakes in wind farm optimization applications”.
In: Journal of Physics: Conference Series. Vol. 625. 1. IOP Publishing.
2015, p. 012037.
[0] I Katic, J Højstrup, and Niels Otto Jensen. “A simple model for
cluster efficiency”. In: European wind energy association conference
and exhibition. 1986, pp. 407–410.
[0] Amin Niayifar and Fernando Porté-Agel. “A new analytical model
for wind farm power prediction”. In: Journal of Physics: Conference
Series. Vol. 625. 1. IOP Publishing. 2015, p. 012039.
[0] Leonardo P Chamorro and Fernando Porte-Agel. “Turbulent flow
inside and above a wind farm: a wind-tunnel study”. In: Energies 4.11
(2011), pp. 1916–1936.
[0] Paul Fleming et al. “Field test of wake steering at an offshore wind
farm”. In: Wind Energy Science 2.1 (2017), p. 229.
[0] Hermann Glauert. A general theory of the autogyro. Vol. 1111. HM
Stationery Office, 1926.
[0] S Boersma et al. “A control-oriented dynamic wind farm flow model:“WFSim””.
In: Journal of Physics: Conference Series. Vol. 753. 3. IOP Publishing.
2016, p. 032005.
[0] Henk Kaarle Versteeg and Weeratunge Malalasekera. An introduction
to computational fluid dynamics: the finite volume method. Pearson
Education, 2007.
[0] Ludwig Prandtl. “Bericht uber Untersuchungen zur ausgebildeten
Turbulenz”. In: Zs. angew. Math. Mech. 5 (1925), pp. 136–139.
[0] Giacomo V Iungo et al. “Data-driven RANS for simulations of large
wind farms”. In: Journal of Physics: Conference Series. Vol. 625. 1.
IOP Publishing. 2015, p. 012025.
Bibliography 101
[0] Johan Meyers and Charles Meneveau. “Large eddy simulations of

large wind-turbine arrays in the atmospheric boundary layer”. In: 48th
AIAA aerospace sciences meeting including the new horizons forum
and aerospace exposition. 2010, p. 827.
[0] H Aa Madsen et al. “Calibration and validation of the dynamic wake
meandering model for implementation in an aeroelastic code”. In:
Journal of Solar Energy Engineering 132.4 (2010), p. 041014.
[0] Alan Duane Wright and Lee Jay Fingersh. Advanced Control Design
for Wind Turbines: Control Design, Implementation, and Initial Tests.
National Renewable Energy Laboratory, 2008.
[0] CL Bottasso and CED Riboldi. “Estimation of wind misalignment
and vertical shear from blade loads”. In: Renewable Energy 62 (2014),
pp. 293–302.
[0] CL Bottasso and CED Riboldi. “Validation of a wind misalignment ob-
server using field test data”. In: Renewable Energy 74 (2015), pp. 298–
306.
[0] Robert P Coleman and Arnold M Feingold. “Theory of self-excited
mechanical oscillations of helicopter rotors with hinged blades”. In:
(1957).

AWFC Cordo

Uploaded by

Copyright:

Available Formats

AWFC Cordo

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

AWFC Cordo

Uploaded by

Copyright:

Available Formats

Politecnico di Milano

SCUOLA DI INGEGNERIA INDUSTRIALE E DELL’INFORMAZIONE

Wind Farm Optimization redirecting Flow through

Anno Accademico 2017–2018

2 Wind turbines performances in yawed operations 13

4 From standard controller through wind farm algorithm 61

5 Wind turbines array optimization 87

1.1 Wind farm wakes . . . . . . . . . . . . . . . . . . . . . . . 3

2.1 Blade diagram and velocity decomposition . . . . . . . . . 14

3.1 Wind turbine reference system . . . . . . . . . . . . . . . . 30

3.11 Wake growth rate . . . . . . . . . . . . . . . . . . . . . . . 44

4.1 Torque control: block diagram . . . . . . . . . . . . . . . . 64

5.1 Wind turbines array, experimental and numerical results:

1.1 Wake models scheme . . . . . . . . . . . . . . . . . . . . . 7

2.1 Reynolds over the blade length: λ = 8.15 and V∞ = 5.65 . 19

3.1 WFSim simulation parameters . . . . . . . . . . . . . . . . 54

4.1 Observer performances . . . . . . . . . . . . . . . . . . . . 81

active yaw control capable to induce significant wake redirection.

Le turbine eoliche vengono raggruppate in impianti per ragioni eco-

per singola turbina, si è introdotto il supercontrollore comprensivo del

Nowadays, the importance of renewable energy is not questionable.

challanges: developing offshore wind turbines and implementing wind farm

1.1 Wind farm modelling

Wake Physics According to Crespo[0], the wind turbine wake can be

Figure 1.3: Ideal sketch of wake meandering

Turbine models Another fundamental part of wind farm model is the

axial and tangential induction. The infinitesimal forces are computed as

Wake models A detailed analysis of wake models can be found in

• Complex wind farm model, computationally onerous.

• Simple parameter-dependant wind farm model, with online estimation

Low fidelity Mid fidelity High fidelity

Table 1.1: Wake models scheme

1.2 Active wake control

As said, greedy control strategy is a wind farm suboptimal control

axial-induction-based control. This control strategy aims at improving

• Independent Pitch Control (IPC) rotating the blades and creating

Yaw control Tilt is not a control variable in present-day wind turbine, so

Figure 1.4: Yaw redirection control

Individual pitch control IPC is used to reduce periodic loads on the

Concerning this work, the proposed control architecture is based on yaw

on the other yaw control represents an additional degree of freedom with

1.3 Thesis outline

the optimal configuration is compared to experimental one. Conclusions,

Wind turbines performances in

Extra code part was implemented in FLORIS and WFSim to reproduce

2.1 Pitt&Peters BEM

Figure 2.1: Blade diagram and velocity decomposition

Main steps of the procedure are shown in this section. Referring to

Vx = V∞ (sin(γ)sin(α)sin(ϕ) + cos(γ)cos(ϕ)) (2.1)

• a is the axial induction

• a’ is the tangential induction

• V∞ is the total incoming flow velocity

To compute axial induction and tangential induction standard BEM

In this formulation tilt is neglected. Comparison studies by Snel and

2.2 Experimental tests

• Ensure consistency with commercial wind turbines in terms of aero-

• Guarantee reliable measurements in wind tunnel environment thanks

• Torque control and collective pitch control.

• Indvidual and cyclic pitch control (IPC ans CyPC).

The approximation is valid only for λr 1 i.e. for r becoming com-