See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/227645540

Feedback–feedforward individual pitch control for wind turbine load

reduction

Article in International Journal of Robust and Nonlinear Control · January 2009

DOI: 10.1002/rnc.1324

CITATIONS READS

227 330

5 authors, including:

Stoyan Kanev J. W. van Wingerden

Vattenfall GmbH Delft University of Technology
60 PUBLICATIONS 1,542 CITATIONS 307 PUBLICATIONS 7,016 CITATIONS

SEE PROFILE SEE PROFILE

Michel Verhaegen
Delft University of Technology
670 PUBLICATIONS 16,355 CITATIONS

SEE PROFILE

All content following this page was uploaded by Stoyan Kanev on 17 June 2024.

The user has requested enhancement of the downloaded file.

INTERNATIONAL JOURNAL OF ROBUST AND NONLINEAR CONTROL
Int. J. Robust Nonlinear Control 2009; 19:72–91
Published online 21 April 2008 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/rnc.1324

Feedback–feedforward individual pitch control for wind turbine

load reduction

K. Selvam1, ∗, † , S. Kanev2 , J. W. van Wingerden3 , T. van Engelen2

and M. Verhaegen3
1 Winwind Oy, Itälahdenkatu 15-17, FI-00210 Helsinki, Finland
2 Energy Research Center of the Netherlands, ECN Wind Energy, P.O. Box 1, 1755 ZG Petten, The Netherlands
3 Delft Center for Systems and Control, TU-Delft, Mekelweg 2, 2628 CD Delft, The Netherlands

SUMMARY
This paper focuses on the problem of wind turbine fatigue load reduction by means of individual pitch
control (IPC). The control approach has a two-degree-of-freedom structure, consisting of an optimal
multivariable LQG controller and a feedforward disturbance rejection controller based on estimated wind
speed signals. To make the control design problem time invariant, all signals are transformed to the
non-rotating reference frame using the Coleman transformation. In the Coleman domain, the LQG control
objective is minimization of the rotor tilt and yaw moments, whereas the feedforward controller tries to
achieve even further improvement by rejecting the influence of the low-frequency components of the wind
on the rotor moments. To this end, the tilt- and yaw-oriented components of the blade-effective wind
speeds are approximated using stochastic random walk models, the states of which are then augmented
with the turbine states and estimated using a Kalman filter. The effects of these (estimated) disturbances on
the controlled outputs are then reduced using stable dynamic model inversion. The approach is tested and
compared with the conventional IPC method in simulation studies with models of different complexities.
The results demonstrate very good load reduction at not only low frequencies (1p blade fatigue load
reduction) but also at the 3p frequency, giving rise to fatigue load reduction of the non-rotating turbine
components. Copyright q 2008 John Wiley & Sons, Ltd.

Received 26 July 2007; Revised 7 March 2008; Accepted 10 March 2008

KEY WORDS: wind turbine; fatigue load reduction; individual pitch control; LQG; wind estimation

1. INTRODUCTION

Wind turbine systems are considered as one of the most promising sources of renewable energy
nowadays, with the installed capacity worldwide growing exponentially. Still, in order to reach the
ambitious goals of the western governments, a significant reduction in the kW h price is necessary,

∗ Correspondence to: K. Selvam, Winwind Oy, Itälahdenkatu 15-17, FI-00210 Helsinki, Finland.
†
E-mail: [email protected]

Copyright q 2008 John Wiley & Sons, Ltd.

 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 73

which necessitates even further upscaling of today’s multi-MW wind turbines. These structures
with increasing size are becoming more and more flexible, requiring an intelligent and robust
control system that not only aims at good quality power control at the above-rated wind speeds but
that primarily focuses on the active reduction of fatigue-relevant loads on the turbine components,
as caused by wind shear, tower shadow, oblique inflow and turbulence.
At present, the majority of wind turbines are operating at variable rotor speed and use generator
torque control and collective blade pitch control to produce as much power as possible at wind
speeds below the rated speed, and to limit the captured power and the rotor speed at the above-rated
wind speeds. The above-rated conventional power control strategy consists of two independent
single-input single-output (SISO) control loops: generator torque control is used to achieve the
desired static generator speed–torque relationship, and collective blade pitch angle (or speed)
control consisting of simple PI (or PD) controller for keeping the generator speed at its rated
value [1–3].
Going one step further, this basic power regulation control can be extended by adding additional
feedback loops with the objective of load reduction. For instance, additional generator torque
control acting on the generator speed can be used to damp the drive train torsional vibration [1].
A tower top fore-aft acceleration measurement can be used by an additional collective pitch control
loop to provide increased damping of the first tower fore-aft mode, whereas the sidewards tower
acceleration can be fed back to a generator torque controller to achieve increased damping of the
tower naying motion [4]. As all these (basic and additional) controllers influence each other, it is
important to ‘decouple’ them by constraining their activities at different non-overlapping frequency
ranges by means of proper filtering. To avoid such a disintegrated control design procedure, modern
control design methods can be used for the design of multi-input multi-output (MIMO) controllers
that directly aim at optimizing a trade-off between several design objectives at the same time,
such as basic power regulation, robust stability to high-frequency unmodelled dynamics, improved
resonant mode damping and control bandwidth limitation [5, 6]. Such MIMO controllers, which
are based on linearized turbine models, can be extended even further by making them adaptive to
the constantly changing turbine’s operating point (i.e. wind speed, rotor speed and pitch angle).
This can be achieved by means of a linear parameter-varying (LPV) control design approach, as
proposed in [7–11]. The price to pay for using such advanced optimal MIMO controllers is the
increased computational complexity and the reduced transparency compared with conventional
PID-based controllers.
Modern wind turbines offer the possibility of even more advanced load reductions by means
of individual pitch control (IPC). To take advantage of that, additional measurements are needed,
such as blade root bending moments. Since these are taken in the rotating reference frame, they
give rise to a periodic system. This periodicity can fortunately be circumvented by transforming
all quantities, defined on the rotating reference frame, to the fixed reference frame by using the
so-called Coleman transformation [12], sometimes also referred to as d–q transformation [13].
The blade root flapwise bending moments, for instance, are transformed into fictitious rotor tilt
and yaw moments, which are used by the IPC to compute tilt-oriented and yaw-oriented pitch
signals. These IPC controller outputs are transformed back to the original (rotating) reference
frame, resulting in individual blade pitch signals. Besides the fact that the Coleman transformation
makes the application of well-developed control theory for linear time-invariant (LTI) systems to
the IPC problem possible, it also offers the additional advantage that it allows the collective pitch
control design to be decoupled from the IPC design, since the IPC loop has only a negligible effect
on the collective pitch control loop at low frequencies, where the controllers are active [5, 12].

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
74 K. SELVAM ET AL.

The reverse is, however, true to a lesser extent as discussed in Section 2.1, so that the IPC would
best be designed for the turbine system with the basic collective pitch control in the loop, as done
in this paper.
There are several IPC design methodologies suggested in the literature recently. In its conven-
tional form, IPC is based on two separate SISO loops for rotor tilt and yaw moment reduction.
Since the 1p (i.e. once per revolution) component of the blade loads is modulated by the Coleman
transform into a static (0p) component in the rotor tilt and yaw moments, the conventional SISO
loops use PI control laws to reduce the low-frequency content of the rotor moments [13, 14]. This
can further be improved by using optimal multivariable controllers, such as LQG [13–15] or H∞
control [5]. Although these methods result in a significant blade fatigue reduction, they offer little
to no reduction on the fatigue loads of the non-rotating parts of the turbine, as those are dominated
by the 3p components of the loads in the fixed reference frame. To improve on that, the basic IPC
control was extended in [12] to 2p and 3p (i.e. twice and three times per revolution) blade load
reduction by means of modification of the Coleman transformation in such a way that, instead of
the usual 1p, now the 2p (respectively, 3p) components in the blade root moments are transformed
into 0p. Similar ideas are pursued in [16].
This paper describes a new IPC design method inspired by the optimal multivariable control in
the Coleman domain [13, 15] on the one hand and by the wind speed estimation methods proposed
in [17–19] on the other hand. In particular, the axial, tilt-oriented and yaw-oriented components of
the blade-effective wind speed signals are modeled by stochastic random walk models, the states
of which are then augmented with the states of the turbine to be estimated all together using a
single Kalman filter. The wind signal estimates are subsequently used in a dynamic disturbance
feedforward controller [20, Section 10.5.2]. The feedforward control action is added to the control
action from an optimal feedback LQG controller. The resulting feedback–feedforward control
structure is similar to the periodic disturbance accommodating control structure in [18], although
the design approach proposed in this paper is fundamentally different in the following three aspects.
First, the complete control design is now performed on an LTI system, as obtained by making
use of the Coleman transform. This offers many possibilities for further improvements such as
extension to LPV control for covering a wider range of operating points of the wind turbine,
similar to the work of Bianchi for collective pitch control [7]. Secondly, the disturbance attenuation
in [18] is based on a simple (static) pseudo-inverse of the system B matrix, whereas the present
approach focuses on a dynamic feedforward controller achieving minimization of the effect of the
disturbance on the tilt and yaw moments. To this end, stable dynamic model inversion (SDMI) is
used [21]. And third, the current method deals with blade-effective wind speeds, which allows for
more accurate approximation of the loads on the individual blades than when only rotor-effective
wind speed is considered, as in [18].
The method is tested in simulation on two models of different complexities. First, a simple
rigid turbine model with just a few degrees of freedom is used, which provides useful insights and
serves as a good basis for the analysis of the presented control strategy. Next, a detailed aero-elastic
linear model, generated by the computer program Turbu [22], is used for more detailed study
of the control strategy. The results are compared with the results obtained with a conventional
PI-based IPC. It is demonstrated that the present method achieves a significant improvement over
the conventional one in that it reduces the rotor tilt and yaw moments over a much larger frequency
band, including the fatigue-relevant 3p components.
The paper is organized as follows. In the following section the models used for the IPC design
are introduced. Section 3 presents the proposed feedback–feedforward method for IPC, which is

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 75

subsequently tested and compared with conventional IPC in simulations in Section 4. The paper
concludes with some final remarks in Section 5.

2. WIND TURBINE MODEL

For IPC synthesis and analysis for load reduction, it is useful to have simple LTI turbine models
that contain enough detail about the system loads. To this end it is necessary to get rid of the
azimuth dependence of some signals, which can be achieved by applying a multi-blade coordinate
transformation (Coleman transformation [23]) on the periodic wind turbine model. One way to
obtain such detailed linearized turbine models is by using TURBU [22], ECN’s linear analysis and
simulation code for horizontal axis wind turbines. However, since the state-space model generated
by TURBU is usually of a rather high-order model and is, hence, less transparent, a simple first-
principles rigid turbine model is first introduced to provide the necessary insight and to outline
the basic principles involved in the generation of a TURBU model.

2.1. Simple wind turbine model

The simple linear model of a three-bladed turbine, taken from [12], has the following features:
(i) individually pitch-controlled rigid blades;
(ii) main rotation and rigid drive train;
(iii) first fore-aft and sidewards tower bending mode;
(iv) controllable generator torque.
The model is excited by realistic blade-effective wind speed signals, accounting for kp-effects on
the rotor blades, k = 1, 2, . . . , due to rotational wind field sampling, tower shadow and wind shear.
Stationary aerodynamic conversion is utilized for the derivation of blade root bending moments.
A schematic layout of the wind turbine model is depicted in Figure 1. How this model is derived
is summarized below.

2.1.1. Linearized aerodynamic conversion. Linear blade element momentum (BEM) theory is
used for computing linearized aerodynamics. To keep the model as simple as possible, unsteady
aerodynamics and wake effects have not been taken into account. The BEM aerodynamics is
based on a linear static map that relates the flapwise wind speeds vfli , i = 1, 2, 3, to flapwise and
leadwise blade root bending moments and forces. The flapwise relative wind speed variation vfli is
defined as the sum of the blade-effective wind speed u i and the upwind motion of the rotor blade.
The latter is caused by fore-aft tower bending only since rigid blades are assumed. The upwind
˙ of the tower top.
structural motion involves both the fore-aft translation x fa and tilt rotation
fa
The latter has an azimuth-dependent effect on the relative wind speed, which varies over the rotor
radius. It is assumed that wind acts at one single point of the blade, located at a distance of 34 of
the blade radius Rb . Hence, denoting H as the tower height, and assuming that the fore-aft tower
motion can be approximated by the motion of a prismatic beam, for which it is known that the
ratio between rotation and displacement is given by 3/2H , the flapwise relative wind speed vfli
can be expressed as follows:
3 3Rb
vfli = u i − ẋfa +sin(i ) ẋfa (1)
2H 4

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
76 K. SELVAM ET AL.

blade 1

blade 3 Fz 1

Mx1
* Fx1
at or
actu Mz1
θ3
xsd θ1
Fs Ω
Ta *
xfa Fa
Mtilt
Myaw
Tg

θ2
blade 2

Figure 1. A schematic layout of the simplified wind turbine model.

where i is the azimuth angle of blade i, measured with respect to the y-axis of the fixed reference
frame. The rotor azimuth angle  is defined as the azimuth angle of the first blade, i.e.  = 1 .
With this notation, for the ith blade, the variations of these aerodynamic moments and forces
are given by
⎡ ⎤ ⎡ ⎤
Mz,i h Mz k Mz
⎢ ⎥ ⎢ ⎥ 
⎢ Fx,i ⎥ ⎢ h Fx k Fx ⎥ vfli
⎢ ⎥=⎢ ⎥ (2)
⎢ M ⎥ ⎢h ⎥ 
⎣ x,i ⎦ ⎣ x M k M x⎦ i
Fz,i h Fz k Fz
where i are the pitch angle variations, Mz,i and Mx,i are the variations in the flapwise and the
leadwise blade root bending moments and Fx,i and Fz,i are the flapwise and the leadwise forces.
The gains h Mz · · · k Fz are derived from the power and thrust coefficient data in a chosen working
point, characterized by wind speed, rotor speed and pitch angle. The derivation is constrained by
the assumption of equal aerodynamic efficiency along the blade radius, which implies a linear
increasing flapwise force per unit span over the rotor radius and constant leadwise force per unit
span, see [4].
The variations in aerodynamic torque Ta , axial force Fa , rotor tilt moment Mtilt and sidewards
force at the tower top Fs are then given by
⎡ ⎤ ⎡ ⎤
Ta Mx,i
⎢ ⎥ ⎢ ⎥
⎢ Fa ⎥ 3 ⎢ Fx,i ⎥
⎢ ⎥= ⎢ ⎥ (3)
⎢M ⎥ ⎢ −M sin  ⎥
⎣ tilt ⎦ i=1 ⎣ z,i i⎦
Fs −Fz,i sin i

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 77

2.1.2. Periodic linear model equations. The drive train is accelerated by the aerodynamic driving
torque Ta and decelerated by the generator torque Tg . Under the assumption of stiff drive train,
the following equation holds for the rotor speed
(generator torque is transformed to slow shaft
equivalent):
˙ = Ta − Tg
J
(4)
3 3
where J is the total inertia (rotor and generator). By noting that i=1 sin i = 0 and i=1 sin2 i =
3
2 , the following expression for the linearized torques and moments can be derived from equations
(1)–(3):
⎡ ⎤
⎡ ⎤ h Mx k Mx −h Mx
Ta ⎢ ⎥⎡ ⎤
⎢ h Fx k Fx −h Fx ⎥
⎢ ⎥ 3 ⎢ ⎥ ui
⎢ Fa ⎥ ⎢ ⎥⎢ ⎥
⎢ ⎥= ⎢ ⎥
⎢ − sin i h Mz − sin i k Mz − b h Mz ⎥ ⎣ i ⎦
9R (5)
⎢ ⎥
⎣ Mtilt ⎦ i=1 ⎢ 16H ⎥
⎢ ⎥ ẋfa
Fs ⎣ 9Rb ⎦
− sin i h Fz − sin i k Fz − h Fz
16H
The variables of the included tower model are the fore-aft x fa and the sidewards xsd displace-
ments, respectively. The fore-aft tower motion is driven by the aerodynamic thrust force Fa and
the tilt moment Mtilt , whereas the sidewards motion is driven by the generator torque Tg and
the sidewards aerodynamic force Fs . Only the first bending modes of the tower in the fore-aft
and sidewards directions are considered. These are approximated by the following second-order
differential equations:
3
m tw ẍfa +stw xfa +dtw ẋfa = Fa − Mtilt
2H
(6)
3
m tw ẍsd +stw xsd +dtw ẋsd = Tg + Fs
2H
where m tw , stw and dtw are tower top equivalent tower mass, stiffness coefficient and damping
coefficient, respectively. Their numerical values can be derived from the following structural data:
• the axial tower displacement at unity force;
• damping rate of the first tower bending mode;
• average of the first fore-aft and sidewards frequencies.
Again, the multiplication factor 3/2H in Equation (6) is derived based on the prismatic beam
approximation of the tower, as discussed above.
Finally, substitution of the expressions in Equation (5) into (4) and (6) results in
3 3
˙ = hM
J
u i +k Mx i −3h Mx ẋfa − Tg
z
i=1 i=1

3 3 3 3
m tw ẍfa = h Fx + sin i h Mz u i + k Fx + sin i k Mz i (7)
i=1 2H i=1 2H
81Rb
+ h M −3h Fx −dtw ẋfa −stw xfa
32H 2 z

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
78 K. SELVAM ET AL.

3 3 27Rb 3
m tw ẍsd = −h Fz sin i u i −k Fz sin i i − h Fz ẋfa + Tg −stw xsd −dtw ẋsd
i=1 i=1 16H 2H

It is assumed in this paper that besides the generator speed, the three blade flapwise bending
moments are measured. Hence, these are included as output equations in the model
9Rb
Mzi = h Mz sin i −1 ẋfa +h Mz u i +k Mz i , i = 1, 2, 3 (8)
8H

2.1.3. The Coleman transformation. Equation (7) contains coefficients that depend on the blade
azimuth angles, giving rise to a time-varying system. However, these equations can be transformed
to a purely time-invariant linear system by means of formulating the structural and aerodynamic
degrees of freedom for the rotor in multi-blade coordinates using the Coleman transform [23].
The Coleman transformation, also called multi-blade transformation, maps the individual blade
coordinates present in the periodic frame of reference into a fixed frame of reference. In other
words, the Coleman transformation maps the blade coordinates to the tower coordinates. In due
process the periodic terms of the aero-elastic equations will be eliminated, as the number of blades
is three, meaning the rotor is isotropic (identical and symmetrically mounted blades) and the inflow
to the rotor is uniform.
The periodic terms are eliminated by multi-blade transformation, because the coordinates of the
model are defined in the same frame of reference as explained in [24]. For example, the physical
coordinate qi for blade i of a three-bladed wind turbine in the rotating frame is represented as
follows:
2 2
qi (t) = q1cm (t)+q2cm (t) sin t + (i −1) +q3cm (t) cos t + (i −1)
3 3
The three multi-blade coordinates qicm , i = 1, 2, 3, replace the blade coordinates qi , i = 1, 2, 3. In
this manner, the blade pitch angles and the blade-effective wind speeds can be represented as
⎡ ⎤ ⎡ cm ⎤ ⎡ ⎤ ⎡ cm ⎤
1 (t) 1 (t) u 1 (t) u 1 (t)
⎢ ⎥ ⎢ cm ⎥ ⎢ ⎥ ⎢ cm ⎥
⎣ 2 (t)⎦ = P ⎣ 2 (t)⎦ , ⎣ u 2 (t)⎦ = P ⎣ u 2 (t)⎦ (9)
3 (t) cm
3 (t) u 3 (t) 3 (t)
u cm
where the matrix
⎡ ⎤
1 sin 1 (t) cos 1 (t)
⎢ ⎥
P = ⎣ 1 sin 2 (t) cos 2 (t)⎦
1 sin 3 (t) cos 3 (t)
is referred to as the Coleman transformation matrix. Its inverse
⎡ 1 1 1 ⎤
3 3 3
⎢2 ⎥
P−1 = ⎢
⎣ 3 sin 1 (t) 3 sin 2 (t)
2 ⎥
3 sin 3 (t) ⎦
2

3 cos 1 (t) 3 cos 2 (t) 3 cos 3 (t)

2 2 2

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 79

is used to convert the multi-blade coordinates back to the rotating reference frame. For instance,
given the multi-blade coordinates of the blade flapwise root bending moments Mzcmi
, i = 1, 2, 3, the
following can be expressed:
⎡ cm ⎤ ⎡ ⎤
Mz 1 (t) Mz 1 (t)
⎢ cm ⎥
⎢ Mz (t)⎥ = P−1 ⎢ ⎣ Mz 2 (t)⎦
⎥
(10)
⎣ 2 ⎦
Mzcm3
(t) Mz 3 (t)

The second and the third multi-blade flap moment coordinates Mzcm 3
and Mzcm
3
have tilt and yaw
orientation. In fact, neglecting the effects of shearing and tensile forces, and pitchwise moments
by the blades on the hub center, these two multi-blade moments are proportional to the tilt (Mtilt )
and yaw (Myaw ) moments at the hub center [4]:
   cm 
Mtilt 3 −Mz 2
= (11)
Myaw 2 Mzcm
3

The first multi-blade flapwise moment M1cm only has a physical interpretation in terms of hub
loading, and is not commonly used by IPC algorithms.

2.1.4. The LTI model. Although the equations of motion depend on the azimuth angle (t), they
do not include any state variables related to the rotating rotor blade and rotor shaft except for the
rotational speed
. Since
has a co-axial orientation, no state variable needs to be transformed.
The flapwise blade root bending moments, the pitch angles and the blade-effective wind speeds,
on the other hand, are variables that do need to be transformed to arrive at an LTI model. Indeed,
it can easily be verified that using Equations (9)–(11), the time-varying model (7)–(8) takes the
following form:
˙ = −3h M ẋfa +3k M cm +3h M u cm − Tg
J
(12)
x x 1 x 1

81Rb
m tw ẍfa = −stw xfa − dtw +3h Fx − h M ẋfa +3k Fx cm
1 (13)
32H 2 z
9 9
+ k Mz cm
2 +3h Fx u 1 +
cm
h M u cm (14)
4H 4H z 2
3 27Rb 3 3
m tw ẍsd = Tg − h F ẋfa − k Fz cm
2 − h Fz u 2 −stw x sd −dtw ẋ sd
cm
(15)
2H 16H z 2 2
Mzcm
1
= −h Mz ẋfa +k Mz cm
1 +h Mz u 1
cm
(16)
27Rb 3 3
Mtilt = − h Mz ẋfa − k Mz cm
2 − h Mz u 2
cm
(17)
16H 2 2
Myaw = 32 k Mz cm
3 + 2 h Mz u 3
3 cm
(18)
3
where in the derivation of the last equation the equality i=1 sin i cos i = 0 has been used.
Clearly, these equations represent an LTI system. Figure 2 gives a block-schematic representation
of the resulting LTI system. Note that apart from the rotor speed
and the generator torque Tg ,

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
80 K. SELVAM ET AL.

Inverse
Coleman Turbine
Coleman

Figure 2. Conversion of the periodic model into a linear model using the Coleman transformation.

the inputs and outputs of the model are in multi-blade coordinates, and are hence only fictitious.
Therefore, if an LTI controller is designed based on this system representation, then when imple-
menting it with the true turbine it is necessary to convert its inputs and outputs back to the original
rotating reference frame. In other words, the blocks ‘Coleman’ and ‘inverse Coleman’ in Figure 2
will at the end be included into the controller, which will actually make the final controller time
varying.
The conventional IPC methods consist of two individual SISO (usually of PI structure) pitch
controllers, one from Mtilt to cm cm
2 , and the other one from Myaw to 3 . The output Mz 1 is not
cm

used. These IPC controllers are designed separately from the collective pitch controller that has
the rotor speed
as input and cm 1 as output. Clearly, the yaw-oriented channel (18) is (in this
simple model) completely decoupled from the other channels. This, however, is not true for the
tilt-oriented moment Mtilt as it is influenced by the collective control loop via the tower fore-aft
speed ẋfa . Therefore, the design of the tilt moment control loop should be performed keeping in
mind the presence of the collective pitch controller. The latter, in turn, can be designed separately
from both IPCs as the rotor speed is not influenced by the individual pitch actions cm 2 and 3 .
cm

Note also that the sidewards tower motion does not have any affect on the control loops as
it does not influence either the rotor speed or the tilt and yaw moments, i.e. these dynamics are
unobservable. Hence, Equation (15) can be removed from this simple model.

2.2. Detailed TURBU model

The ECN computer code TURBU [22] generates elaborate linearized aero-elastic models of three-
bladed horizontal axis wind turbines. These models include considerable features that are necessary
for control design and aero-elastic stability analysis, such as bending and torsion deformation,
(unsteady) aerodynamic and hydrodynamic conversions and wake dynamics. All model inputs for
the drive train and rotor blades are transformed into multi-blade coordinates before they enter the
LTI model, and the model outputs from the drive train and rotor blades are transformed back to
rotating coordinates. A linear model is computed for a given aerodynamic equilibrium state. The
latter is derived via the BEM theory. The average deformation state is matched to the aerodynamic
equilibrium. This is based on non-linear propagation of the deformation of the individual elements
caused by the average loading. The average deformation per element is based on slender beam
bending theory. A multi-body approach is used to model the structural dynamics, as illustrated in
[22]. The multi-body wind turbine model has N elements per blade ({Di , E i , Fi }, i = 1, 2, . . . , N )

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 81

and M elements for the tower (Si , i = 1, 2, . . . , M), with each element having 5 degrees of freedom.
There are 6 degrees of freedom in the rotor shaft Rr .
A typical TURBU model (with N = 14 and M = 15) has around 600 state variables, and therefore
needs simplification when used for control design. For reducing the model order, TURBU provides
model reduction based on the elimination of high-frequency modes. This allows significant model
order reduction without any loss of accuracy in the dynamic behavior of the lower-frequency
modes that are within the bandwidth of the controllers. This yields a model order of about 150.
This reduced-order model is here referred to as the multi-body model. In addition to that, TURBU
allows for modeling only a distinct number of degrees of freedom in the blade roots and tower
bottom. The model obtained in this manner is referred to as the lumped model. It has only 28 states
and still models dynamic pitch servo actuation. Figure 3 shows the frequency responses from the
multi-blade pitch angle signals icm , i = 1, 2, 3, to the rotor speed
, tilt moment Mtilt and yaw
moment Myaw for both the full multi-body (solid line) and reduced lumped (dashed line) models.
Sufficient accuracy is observed at frequencies below 1 Hz.
From the frequency plots it is clear that at low frequencies (below about 0.3 Hz) there is little
coupling between the channel cm cm
1 →
on the one hand and the IPC channels 2 → Mtilt and
cm
3 → Myaw on the other hand. However, above this frequency there is a clear interaction between
these loops. Therefore, when the collective controller is separately designed from the IPC it is very
important that these loops be further ‘decoupled’ using suitable filters. It is especially important that
structural resonant frequencies that are well inside the bandwidth of the collective controller (e.g.

Figure 3. Frequency response from multi-blade pitch angles to rotor speed, tilt moment and yaw moment,
for both the multi-body (solid) and the lumped models (dashed).

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
82 K. SELVAM ET AL.

the tower frequency at about 0.47 Hz) are filtered out from cm1 so that it does not get influenced
by the IPC control actions.
Similar conclusions can be made for the IPC loops, which are only decoupled from the collective
loop at low frequencies, and if designed independently, filtering should be used to reduce high-
frequency effects from the collective pitch action. However, in this paper a different approach is
used for the IPC design, which does not need additional filter design. The starting point is the
design of the collective pitch controller and the generator torque controller using conventional
methods, including filters, as discussed above. These controllers are then interconnected with the
turbine, so that the IPC design is performed on the closed-loop turbine system with the generator
torque control and the collective pitch control in the loop. Then, a multivariable robust control
design approach is used to synthesize one MIMO IPC loop that stabilizes the turbine and minimizes
a suitably defined performance criterion based on the rotor moments. Of course, one could choose
to include also the collective pitch and generator torque control loops into the MIMO controller;
this is not done here as the present approach assumes that the speed and power loops are given,
and it focuses on the inclusion of additional actuation through IPC for the purpose of fatigue load
reduction. Note that this MIMO approach also takes into account the coupling between the tilt-
and the yaw-oriented multi-blade coordinates, whose coupling is clearly much stronger than the
coupling with the collective pitch loop (see Figure 3). This coupling was absent in the simple
model from the previous subsection.

3. CONTROLLER DESIGN

In this section the proposed IPC design approach is explained. Since it is compared in the following
section with the conventional SISO approach to IPC, it is briefly described next.

3.1. Conventional SISO control approach

The conventional IPC is a scalar approach based on the assumption that the multi-blade coordinate
transformation, described in Section 2.1.3, results in three independent time-invariant control loops:
from the collective pitch angle to the rotor speed, from the yaw-oriented pitch angle to the yaw
moment and from the tilt-oriented pitch angle to the tilt angle. It is assumed that the generator
speed is measured.
The speed regulation loop is usually based on a PI compensator that has as input the difference
between the filtered generator speed and its rated value. The filter, used in the simulations, includes
• A low-pass filter (inverse Chebychev [type II], fourth order, 20 dB reduction) with a cut-off
frequency of (3p−0.8) rad/s.
• A notch filter (elliptic filter, fourth order, 30 dB reduction) with stop band [0.8tsd , 1.05tsd ],
where tsd is the first tower naying frequency.
• A notch filter (elliptic filter, second order, 30 dB reduction) with stop band [0.8cll , 1.05cll ],
where cll is the collective lead-lag frequency of the blades.
The PI compensator is designed to achieve a gain margin of 2 and a phase margin of 45◦ .
The torque controller is designed for achieving constant power production by keeping the
generator torque equal to the rated power Pr divided by the rotor speed. After linearization around
the rated generator speed
g,r , the generator controller has the form of a P-compensator with gain

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 83

(−Pr /n gb
2g,r ), where n gb is the gearbox ratio. It should be noted that this generator controller
has a slightly destabilizing effect on the rotor speed as the generator torque decreases when the
generator speed increases.
The load regulation loops aim at blade fatigue load reduction by reducing the 1p load component
in the blade root bending moments. It is demonstrated in [12] that (3m ±1) components, m =
1, 2, . . . , in the load spectrum of the blade-effective wind speed signals u i , i = 1, 2, 3, contribute
to the 3m components in the multi-blade wind coordinates u icm . In other words, 1p loads on the
blades are modulated into 0p loads on the tilt and yaw moments, meaning that the former can
be reduced by reducing the static loads on the rotor moments. This can be achieved by means of
a simple I-compensator. In order to prevent the influence of higher harmonics it is necessary to
apply low-pass filtering around and beyond 3p.

3.2. Feedback–feedforward multivariable control approach to IPC

In this section the multivariable control approach to IPC, proposed in this paper, will be introduced.
It has a two-degree-of-freedom structure, containing an optimal feedback LQG controller from the
rotor tilt and yaw moments to the multi-blade pitch angles icm , and a feedforward disturbance
rejection controller acting on the estimated multi-blade wind signals and producing additional
multi-blade pitch angles.
As discussed above, this paper assumes that the collective pitch controller Ccol (s) and generator
torque controller Cgen (s) (including the filters) are available and interconnected with the wind
turbine to regulate the produced power by controlling the generator torque and the rotor speed, as
explained in the previous subsection. Below it is assumed that the controllers Ccol (s) and Cgen (s)
are interconnected to the linear turbine model, be it the simple rigid model (12)–(15) or the multi-
body TURBU model (by substituting cm 1 = C col (s)
and Tg = C gen
in the turbine model), after
which discretization is performed, resulting in

xwt (k +1) = Awt xwt (k)+ Bwt, cm

23 (k)+ Bwt,u u (k)
cm
(19)
y(k) = Cwt xwt (k)+ Dwt, cm
23 (k)+ Dwt,u u (k)+v(k)
cm

where the state xwt contains the state of the wind turbine model and the states of the torque
and pitch controllers, cm cm cm T
23 = [2 , 3 ] is the input, u
cm = [u cm , u cm , u cm ]T is the disturbance
1 2 3
input, y = [Mtilt , Myaw ] is the measured output and v ∈ R2 is a zero-mean white noise process
with covariance matrix Q v = Q Tv >0. Note that the rotor tilt and yaw moments are assumed to be
measured. In practice, conventional wire strain gauges can be used to measure the flapwise bending
moments at the blade roots, which can then be converted into Mtilt and Myaw using the Coleman
transformation. Although strain gauges are not very reliable devices as such for this application,
due to the potential danger caused by lighting, the recent developments of optical strain gages are
likely to overcome this disadvantage.

3.2.1. Design of the optimal LQG controller. The optimal LQG controller consists of a linear
quadratic regulator (LQR) and a Kalman filter. However, the conventional assumption in the Kalman
filter design that the external input is a random white Gaussian process is clearly not satisfied
for model (19), as the multi-blade wind signals u cm (k) have no flat spectrum. To circumvent this
problem, one can identify a stochastic linear model Mwind (z) that has (approximately) the same
spectrum as the wind signals u cm (k). This would allow modeling of u cm as the output of a filtered

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
84 K. SELVAM ET AL.

white noise process, u cm = Mwind (z)w. Substituting this into the turbine model (19), and augmenting
the states x wt with the filter model states, will yield a new model that has the white noise process w
as external input, so that standard Kalman filter design can be performed. This approach, however,
requires (a) that the spectrum of the multi-blade wind signals is given and (b) that it is accurately
represented by a linear model of low order (in order to keep the order of the final controller low).
To avoid this approach, one might instead use a much simpler wind modeling, by noting that
• the energy of u cm is concentrated at low frequencies (below 0.1 Hz);
• the signal u cm is stationary under mild assumptions, as proved in Lemma A1.
This suggests that a random walk model could be sufficient to represent the relevant low-
frequency behavior of u cm

u cm (k +1) = u cm (k)+w(k) (20)

where w is a random white Gaussian process with zero-mean and covariance matrix Q w . Usually,
the covariance matrix Q w is viewed as a design parameter that provides a trade-off between tracking
speed and smoothness of the estimates. For simplicity, it is often selected as a diagonal matrix.
Faster tracking of the true signals can be obtained by appropriately increasing the elements of Q w ,
which however results in less smooth (i.e. more noisy) estimates and vice versa. A value of 0.1I2
is selected for the simulations in this paper. The random walk model is particularly suitable, and
often used, for the estimation of an unknown time-varying bias on the state and output equations
[25], which has already proved to be an accurate and robust approach to rotor-effective wind speed
estimation [17].
Interconnecting the random walk model (20) with the turbine model (19) results in
        
xwt (k +1) Awt Bwt,u xwt (k) Bwt, cm 0
= + 23 (k)+ w(k)
u (k +1)
cm
0 I u (k)
cm
0 I
    (21)
xwt (k) Dwt, cm
y(k) = [Cwt 0] cm + 23 (k)+v(k)
u (k) 0

A Kalman filter is used to estimate the state of this augmented system. The state estimate x̂ wt (k)
will then be used by the LQR controller, discussed next, while the wind signal estimate û cm (k)
will be used by the feedforward controller, discussed later on.
The conditions under which system (21) is observable are discussed in [17]. A sufficient condition
for the observability is that the pair (Awt , Cwt ) is observable, and that the turbine system (19) has
no poles and zeros at z = 1, which holds for the models considered in this paper. Note, however,
that the augmented system is not controllable, due to the fact that part of the states belong to the
wind model. It is assumed that the remaining part of the augmented states (i.e. the pair (Awt , Bwt ))
is controllable.
Given the Kalman filter turbine state estimate, x̂ wt (k), the LQR control action has the form
(K lqr x̂wt (k)), where the gain K lqr is chosen to optimize the following standard quadratic criterion:
 T   
∞ xwt (k) Q xwt (k)
Jlqr =
k=0 cm R cm

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 85

3.2.2. Feedforward estimated wind disturbance rejection. An additional IPC action is added to the
optimal LQG control action, discussed above. It uses a feedforward disturbance rejection control
23 (k). To this end, define the
algorithm based on the estimated multi-blade wind speed signal û cm
following transfer functions:

G(z) = Cwt (z I − Awt )−1 Bwt, + Dwt,

(22)
H (z) = Cwt (z I − Awt )−1 Bwt,u

so that (with slight abuse of notation) it can be expressed y(k) = G(z)cm

23 (k)+ H (z)u (k)+v(k).
cm

Under the assumption of unbiasedness of the Kalman filter estimates, it follows that

y(k) = G(z)cm
23 (k)+ H (z)û (k)+ ṽ(k)
cm

where ṽ(k) is a zero-mean random process. The feedforward controller Cff (z) should hence be
designed in such a manner that the control action cm
23 (k) = C ff G(z)û 23 (k) minimizes the influence
cm

of û (k) on y(k). The optimal controller will then be given by

cm

Cff (z) = −H (z)G −1 (z)

However, it can happen (as is the case with the numerical models considered in this paper) that
the transfer function G(z) is a non-minimum phase, resulting in an unstable optimal feedforward
controller. To avoid that the inverse of G(z) will be substituted by a stable inverse, as obtained
using the SDMI method in [21].
The basic idea behind the SDMI method is, similar to the wind estimation method above, the use
of a random walk model. To summarize the method, suppose that n(k) is some (unknown) signal
and let q(k) = G(z)n(k). Then finding a stable inverse means computing a stable transfer function
G inv (z) such that n̂(k) = G inv (z)q(k) ≈ n(k). To this end, the signal n(k) can be viewed as an
unknown bias that can be estimated using the same idea as the random walk model in combination
with a Kalman filter. Expressing n(k +1) = n(k)+(k), with E{k } = 0 and E{(k)(k)T } = Q  ,
adding this to the state of G(z) and including (if necessary for numerical reasons) a small additional
process and/or measurement noise terms results in an augmented model G aug (z), for which a
Kalman filter can be designed. The Kalman filter transfer function from q(z) to n̂(k) represents the
inverse of G(z). Hence, for the considered model (22), the SDMI method results in the following
stable inverse of G(z):
  −1
Awt Bwt,
G inv = [0, I ] z I − + K [Cwt , 0] K
0 I

The feedforward controller then takes the form Cff (z) = −H (z)G inv (z), so that the complete
feedback–feedforward control action is formed as shown in Figure 4. This IPC loop, together
with the basic power control loop, including the generator torque P-controller and the collective
pitch PI-controller, is depicted in Figure 5. The figure shows both the conventional collective pitch
control and generator torque control loops, acting on the rotor speed
, as well as the IPC algo-
rithm, acting on the measured flapwise blade root bending moments Mzi and computing additional
actions to the collective pitch angles. Note the required modulation and demodulation of the signals
discussed above.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
86 K. SELVAM ET AL.

Controller

Kalman
filter

Figure 4. The complete feedback–feedforward control scheme.

Figure 5. Layout of the complete control structure.

4. FREQUENCY DOMAIN ANALYSIS

This section presents results that demonstrate a significant improvement in the load reduction
of the non-rotating turbine components, obtained with the presented feedback–feedforward IPC,
compared with the conventional PI-based IPC. The controllers are designed and compared on both
the simple rigid model and the detailed multi-body TURBU model, as described in Section 2. The
parameters of the simple model (12)–(18) for a fictitious 2.5 MW wind turbine are listed in Table I.
This model, as well as the TURBU model, is derived for wind speed of 16 m/s, pitch angle of 10◦
and rotor speed of 1.806 rad/s.
The basic speed and power control loops are the same for both IPCs. These are designed as
discussed in Section 3.1. The conventional IPC consists of two integrators, one for the tilt-oriented

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 87

Table I. Numerical values of the model parameters in the simple model (12)–(18).
Parameter Value Parameter Value

H 55.953 m h Mx 8.3806×104 N s
Rb 40 m h Fz 4.0683×103 N s/m
J 11.2553×106 kg m2 h Mz −1.8948×105 N s
m tw 1.5657×105 kg k Fx −6.1478×103 N
stw 1.235×106 N/m k Mx −3.7711×104 N m
dtw 2.7995×103 N s/m k Fz −1.8306×103 N
h Fx 7.2019×103 N s/m k Mz 1.6174×105 N m

Table II. Parameters used for the design of the feedback–feedforward IPC controller.
Parameter Qv Qw Q R Q
Value 0.01I 0.1I2 I 0.8I 0.01

and one for the yaw-oriented channels. The new feedback–feedforward IPC controller is designed
as discussed in Section 3.2. The parameters used in the design process are summarized in Table II.
The comparison between the two control designs is performed in the frequency domain. The
frequency response plot for the simplified model is depicted in Figure 6. The figure shows the
transfer functions from the multi-blade wind signals u cm 23 to the rotor tilt and yaw moment Mtilt
(left plot) and Myaw (right plot). The solid lines in the plots represent the case of no IPC, the
dashed dotted lines correspond to the conventional IPC method, whereas the dashed lines are for
the new feedback–feedforward method. Only the diagonal channels of the transfer function from
u cm to y are given, since the off-diagonal ones are zero (even with the new MIMO controller,
which preserves the intrinsic diagonal structure of the simple model). It can be clearly seen from
the figure that the conventional IPC approach (dashed line) has good load reduction only at very
low frequencies, whereas at 1p and higher frequencies in the fixed reference frame there is no
reduction, and even a slight increase. The low-frequency reduction is due to the integrator structure
of this control method, making the method suitable for blade load reduction (as 0p reduction in Mtilt
and Myaw corresponds to 1p reduction in the flapwise blade root bending moments), but cannot
achieve fatigue-relevant load reduction on the non-rotating components of the wind turbine. On
the other hand, by trading off low-frequency load reduction, the proposed feedback–feedforward
method achieves reduction over a much wider frequency range, including the 3p frequency, which
is very relevant for fatigue on the non-rotating components, such as the nacelle, yaw bearing and
tower [16]. Improved reduction at low frequencies can be obtained by including integral action in
the controller.
Similar conclusions can be made based on the results with the TURBU model, as depicted
in Figure 7. Now there is coupling between tilt- and yaw-oriented moments, so that the off-
diagonal channels are also plotted. Note that although the IPC controller design has been performed
based on the reduced lumped TURBU model, the results in the figure represent the closed-loop
system with the detailed multi-body TURBU model. Note also that the advanced feedback–
feedforward controller achieves improved load reduction over a much wider frequency band than
the conventional IPC method, which only leads to improvement at very low frequencies, whereas
it actually results in load amplification at frequencies of 1p (here 0.3 Hz) and higher in the fixed

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
88 K. SELVAM ET AL.

x 10 TF from ucm2 to Myt TF from ucm3 to Mzt

x 10
4 4

3 3

2 2

1 1

0 0
10 10 10 10 10 10 10 10

Figure 6. Frequency response plot of tilt moment Mtilt (left) and yaw moment Myaw (right) due to the
multi-blade wind inputs u cm cm
2 and u 3 , respectively, for the simple rigid turbine model with no IPC (solid),
conventional IPC (dashed dotted) and new feedback–feedforward IPC (dashed).

x 10 Ucm2 x 10 Ucm3
4 4

3.5 3.5

3 3

2.5 2.5
Myt

2 2

1.5 1.5

1 1

0.5 0.5

0 0
10 10 10 10 10 10 10 10

x 10 x 10
4 4

3.5 3.5

3 3

2.5 2.5
Mzt

2 2

1.5 1.5

1 1

0.5 0.5

0 0
10 10 10 10 10 10 10 10
f [Hz] f [Hz]

Figure 7. Frequency response plot of tilt moment Mtilt (first row) and yaw moment Myaw (second row)
due to the multi-blade wind inputs u cm cm
2 (left column) and u 3 (right column) for the TURBU model with
no IPC (solid), conventional IPC (dashed dotted) and new feedback–feedforward IPC (dashed).

reference frame (observe the high peak on the top-left plot in Figure 7, which is at about 1p).
The performance of the present method at low frequencies can easily be improved by including
integral action in the controller. Finally, it needs to be mentioned that the improved high-frequency
reduction inevitably requires pitch control activity at these frequencies, which might in practice
be undesirable. This can be circumvented by introducing additional penalty on the control signal
at high frequencies in the optimal control optimization.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 89

5. CONCLUSION

This paper presented a new approach to the design of IPC controllers for load reduction. The
method is based on a two-degree-of-freedom control structure, consisting of an optimal feedback
LQG controller and an additional feedforward disturbance rejection controller based on estimated
wind signals. The controller design is performed based on the wind turbine with fictitious inputs
and outputs that are uniquely defined from blade pitch angles and the measured blade flapwise root
bending moments. These fictitious multi-blade inputs and outputs are computed using the Coleman
transformation and lead to a purely time-invariant wind turbine model. This approach, contrary to
the one followed in [18] where the design is performed directly on the periodic system, makes it
possible to exploit the very well-developed control design theory for linear systems, and allows for
extension to the case of varying operating conditions (wind speed, rotor speed and pitch angle),
as suggested for collective pitch control in the work of Bianchi et al. [7]. The wind estimation is
based on a simple but effective random walk model, avoiding the need for an accurate wind model.
The new IPC design method is tested on two linearized turbine models of different complexities,
and the results are compared with those obtained with the conventional IPC control.

APPENDIX A

Lemma A1 (Stationarity of multi-blade wind speeds)

Under the assumption of homogeneous turbulence, constant rotational speed and non-oblique
oriented wind flow, the multi-blade components u icm , i = 1, 2, 3, of the blade-effective wind speeds
are stationary processes.

Proof
For homogeneous turbulence and purely axial wind direction, the blade-effective wind speed u i
on a fixed point on a rotating blade can be expressed as a time-varying Fourier expansion
∞
 2
j pi (t) 1
u i (t, i ) = e û p (t), û p (t) = e j p
u(t,
) d

p=−∞ 2 0

where û p (t) are time-dependent rotational modes. It has been shown in [12] that the following
expression holds for the multi-blade coordinates of the blade-effective wind speeds:
⎡ ⎤ ⎡ ⎤
1 (t)
u cm û 3m (t)
∞
⎢ cm ⎥ ⎢ ⎥
⎣ u 2 (t)⎦ = e j3m  ⎣ j(û 3m+1 (t)− û 3m−1 (t))⎦
m=−∞
3 (t)
u cm j(û 3m+1 (t)+ û 3m−1 (t))

Then, with a ∗ denoting the conjugate of a, it can easily be shown that

⎡ cm ⎤∗ ⎡ ⎤
u 1 (t +) û 3n (t +)
⎢ cm ⎥ ∞ ⎢ ⎥
⎣ u 2 (t +)⎦ = e−j3n  ⎢
⎣ −j(û ∗3n+1 (t +)− û ∗3n−1 (t +))⎥ ⎦
n=−∞
∗ ∗
u cm
3 (t +) −j( û 3n+1 (t +)+ û 3n−1 (t +))

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
90 K. SELVAM ET AL.

Therefore, for the variance of u cm

2 one has
∗
2 (t)(u 2 (t +)) }
E{u cm cm

 
∞
=E e j3(m−n) (û 3m+1 (t)− û 3m−1 (t))(û ∗3n+1 (t +)− û ∗3n−1 (t +))
m,n=−∞

∞
= e j3(m−n) E{(û 3m+1 (t)− û 3m−1 (t))(û ∗3n+1 (t +)− û ∗3n−1 (t +))}
m,n=−∞

since under the assumption of constant rotational speed one has  =
t +(0), so that e j3(m−n)
are purely deterministic signals. Furthermore, in [26] it is proved that, under the considered
assumptions, the rotational modes are orthogonal and stationary, i.e.

{û p (t)û q∗ (t +)} =  p,q û p ()

where  p,q denotes the Kronecker delta function and û p () is the covariance function of û p .
∗
2 (t)(u 2 (t +)) } all terms for n  = m drop, giving
Therefore, in the above expression for E{u cm cm

∞
∗
2 (t)(u 2 (t +)) } =
E{u cm ( û 3m−1 ()+ û 3m+1 ())
cm
m=−∞

2 (t) is not a function of the time t. The same lines can be

Clearly, the correlation function of u cm
followed for the first and the third multi-blade components u cm1 (t), and u 3 (t) to arrive at the same
cm

conclusion. Therefore, u i (t), i = 1, 2, 3, are stationary processes.

cm

REFERENCES
1. Bossanyi E. The design of closed loop controllers for wind turbines. Wind Energy 2000; 3:149–163.
2. Burton T, Sharpe D, Jenkins N, Bossanyi E. Handbook of Wind Energy. Wiley: New York, 2001.
3. van der Hooft E, van Engelen T, Schaak P. Development of wind turbine control algorithm for industrial use.
Proceedings of the European Wind Energy Conference, Copenhagen, Denmark, 2001.
4. van Engelen T, Markou H, Buhl T, Marrant B. Morphological study for aeroelastic control concepts for wind
turbines. Technical Report ECN-E-06-056, Energy Research Center for the Netherlands (ECN), 2006 (available
from http://www.ecn.nl/publicaties/default.aspx?nr=ECN-E-06-056).
5. Geyler M, Caselitz P. Individual blade pitch control design for load reduction on large wind turbines. Proceedings
of the European Wind Energy Conference, Milan, Italy, 2007; 82–86.
6. Cutululis N, Ceanga E, Hansen A, Sørensen P. Robust multi-model control of an autonomous wind power system.
Wind Energy 2006; 9:399–419.
7. Bianchi F, De Battista H, Mantz R. Wind Turbine Control Systems: Principles, Modelling and Gain-scheduling
Design (Advances in Industrial Control). Springer: London, 2006.
8. Lescher F, Zhao J, Borne P. Robust gain scheduling controller for pitch regulated variable speed wind turbine.
Studies in Informatics and Control 2005; 14(4):299–315.
9. Bianchi F, Mantz R, Christiansen C. Control of variable-speed wind turbines by LPV gain scheduling. Wind
Energy 2004; 7:1–8.
10. Bianchi F, Mantz R, Christiansen C. Gain scheduling control of variable-speed wind energy conversion systems
using quasi-LPV models. Control Engineering Practice 2005; 13:247–255.
11. Lescher F, Camblong H, Curea O, Briand R. LPV control of wind turbines for fatigue loads reduction using
intelligent micro sensors. Proceedings of the 2007 American Control Conference, New York City, U.S.A., 2007;
6061–6066.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc
 FEEDBACK–FEEDFORWARD INDIVIDUAL PITCH CONTROL 91

12. van Engelen T. Design model and load reduction assessment for multi-rotational mode individual pitch control
(higher harmonics control). Proceedings of the European Wind Energy Conference, Athens, Greece, 2006 (available
from: http://www.ecn.nl/publicaties/default.aspx?nr=ECN-RX-06-068).
13. Bossanyi E. Individual blade pitch control for load reduction. Wind Energy 2003; 6:1919–128.
14. Bossanyi E. Wind turbine control for load reduction. Wind Energy 2003; 6:229–244.
15. Kallesœ BS. A low-order model for analysing effects of blade fatigue load control. Wind Energy 2006; 9:421–436.
16. Bossanyi E. Further load reductions with individual pitch control. Wind Energy 2005; 8:481–485.
17. Mesic S, Verdult V, Verhaegen M, Kanev S. Estimation and robustness analysis of actuator faults based on
Kalman filtering. Proceedings of the 5th Symposium on Fault Detection, Supervision and Safety for Technical
Processes, Washington, DC, U.S.A., 2003.
18. Stol K, Balas M. Periodic disturbance accommodating control for blade load mitigation in wind turbines. Journal
of Solar Energy Engineering 2003; 125(4):379–385.
19. Kendall L, Balas M, Lee Y, Fingersh L. Application of proportional–integral and disturbance accommodating
control to variable speed variable pitch horizontal axis wind turbines. Wind Engineering 1997; 21:21–38.
20. Goodwin G, Graebe S, Salgado M. Control System Design. Prentice-Hall: Englewood Cliffs, NJ, 2001.
21. George K. Verhaegen M, Scherpen J. A systematic and numerically efficient procedure for stable dynamic model
inversion of LTI systems. Proceedings of the 38th Conference on Decision and Control, Pheonix, AZ, U.S.A.,
1999; 1881–1886.
22. van Engelen T. Control design based on aero-hydro-servo-elastic linear models from TURBU (ECN). Proceedings
of the European Wind Energy Conference, Milan, Italy, 2007.
23. Coleman R, Feingold A. Theory of Self-excited Mechanical Oscillations of Helicopter Rotors with Hinged Blades.
USGPO: Washington, DC, 1958.
24. Hansen M, Thomsen K, Fuglsang P, Knudsen T. Two methods for estimating aeroelastic damping of operational
wind turbine modes from experiments. Wind Energy 2006; 9:179–191.
25. Kanev S, Verhaegen M. Two-stage Kalman filtering via structured square-root. Journal of Communications in
Informations and Systems 2005; 5(1):143–168.
26. van Engelen T, Schaak P. Oblique inflow model for assessing wind turbine controllers. Proceedings of the 2nd
Conference on the Science of Making Torque from Wind, Denmark, 2007.

Copyright q 2008 John Wiley & Sons, Ltd. Int. J. Robust Nonlinear Control 2009; 19:72–91
DOI: 10.1002/rnc

View publication stats