Abstractthis paper proposes a MIMO linear quadratic

regulator (LQR) controller designed for a horizontal variable

speed wind turbine with focus on the operating range referring
to the above rated wind speeds. The operating conditions of
wind turbines make them subject to fluctuating loads that
create fatigue and lead to damage. Alleviating these loads would
reduce the needed materials, and increase the lifespan and the
quality of the produced energy. The optimality of the entire
system is defined in relation with the trade-off between the
wind energy conversion maximization and the minimization of
the fatigue in the mechanical structure. The solution of a
control using an LQR regulator is presented. The performances
of the optimal control are assessed and discussed by means of a
set of simulations.
I. INTRODUCTION
Classical control system design is generally a trial and
error process in which various methods of analysis are
iteratively used to determine the design parameters of a
system. Acceptable performance is generally defined in
terms of time and frequency domain criteria such as rise
time, settling time, overshoot, gain and phase margin and
bandwidth.
Radically different performance criteria must be satisfied,
however, by the complex, multiple inputs, and multiple
outputs systems required to meet the demands of modern
technology.
The objective of optimal control theory is to determine the
control signals that will cause a process to satisfy the
physical constraints and at the same time to minimize or
maximize some performance criterion. [1]
The wind industry offers many challenges in designing
effective wind turbines that will harness wind energy and
will transform it into electricity. Wind turbines are large,
complex dynamically flexible structures that operate in
turbulent and unpredictable environmental conditions where
efficiency and reliability are highly dependent upon a well

Manuscript received January 30, 2011.
A. P. Author is with the University Politehnica Of Bucharest,
Romania (corresponding author, e-mail: andreea.pintea@ gmail.com).
H. W. Author is with the MIS Laboratory, University of Picardie Jules
Verne, 7 Rue du Moulin Neuf, 80000, Amiens, France (e-mail:
[email protected]).
N. C. Author is with Universite Lille 1, Lille, France, LAGIS
Department (e-mail: [email protected]).
P. B. Author is with Ecole Centrale de Lille, France, LAGIS Department
(e-mail: [email protected] )
D. P. Author is with the University Politehnica of Bucharest,
Romania, (e-mail: [email protected])
A. B. Author is with the University Politehnica of Bucharest,
Romania, (e-mail: [email protected])

designed control strategy.
From a control point of view, the importance lies not only
on ensuring an optimal operation, but also on load reduction
and grid integration. Another important challenge is to
provide good quality energy delivery from a profoundly
irregular primary source, the wind.
The characteristics of the wind energy source are
important in different aspects regarding wind energy
exploitation. The energy available in the wind varies with the
cube of the wind speed. The wind is variable both in space
and in time [2].
Based on the value of the wind speed, there were two
essential functioning regimes identified for the wind
turbines. The first one corresponds to low wind operation,
and here the main control goal is to maximize the energy
capture.
This region ends when the winds speed reaches the rated
value, above which, the turbine enters the second regime.
This value is usually around 14m/s.
In the above rated region, the pitch angle and the
electromagnetic torque are the control variables that are used
to reduce the structural loads and to maintain the output
power around a constant nominal value, also called the rated
power of the turbine (Fig. 1).

Fig.1 Tipical wind turbine power curve

Therefore, in this regime, the system is multivariable and
multi-objective. Many applications used classical controls to
address more than one control objective, by adding multiple
control loops.
These added complexity to the control design and
systems behavior but, nevertheless, it was difficult to
properly address control-structure interaction issues because
the controller used only a single measured turbine output as
the basis of its control and did not have direct knowledge of
the dynamics of the turbine. Modern control designs using
state space methods, can handle these issues in a better way,
Optimal control of variable speed wind turbines
Andreea Pintea, Haoping Wang, Nicolai Christov, Pierre Borne, Dumitru Popescu, Adrian Badea
19th Mediterranean Conference on Control and Automation
Aquis Corfu Holiday Palace, Corfu, Greece
June 20-23, 2011
WeCT2.1
978-1-4577-0123-8/11/$26.00 2011 IEEE 838

since the controllers in these cases use a model to determine
the systems states. Controllers can be designed not only to
maximize power or to regulate the turbines speed, but also
to add damping to its flexible modes, through state feedback
[3]. In the same context, the LQR regulator, proved to be a
good solution due to the fact that it facilitates multivariable
and multi-objective control design.
The paper is organized as follows: after a short
introduction and the presentation of the context in which the
LQR controller was chosen, one continues with Section II in
which the mathematical model of the turbine is presented in
detail. Section III provides a description of the LQR control
method and Section IV presents the analysis of the results
and the concluding remarks of this study. In the end of the
paper, an APPENDIX with the numerical values of the wind
turbines parameters used is provided.
II. MATHEMATICAL MODEL

At present, there are several variable speed wind turbine
configurations that are being widely used. For this study, a
horizontal variable speed wind turbine was chosen. The
variable speed wind turbine is currently the most used
technology and it has proven its advantages over the years
[4]. The major advantage is that by allowing the rotor to
operate at various speeds, one can obtain a more efficient
capture of the wind energy with less stress in the turbine
drive train during wind gusts. The reader can find different
wind turbine modeling techniques in [2] and also detailed
explanations regarding the use of each type of model.
Generally, a model for an entire wind energy conversion
system can be structured as several interconnected subsystem
models: an aerodynamic, a mechanical, electrical and
actuator subsystems. But since the dominant dynamics lie in
the mechanical subsystem, special attention will be paid in
this direction. The mechanical structure that we chose to
study is seen as being arranged into several rigid bodies
linked by flexible joints. The amount of these joints or
degrees of freedom, determines the order of the model.
In [5] [6] and [3] one can observe the way in which the
number of degrees of freedom of the system can increase the
order of the non linear models of the turbine. Therefore, it is
important to consider on the model just those degrees of
freedom that are directly coupled to the control [4].
By this reason, the model presented here, will include just
the first mode of the drive train, the first mode of tower
bending dynamics, and the first mode of the flapping of the
blades. These degrees of freedom will suffice for the
controller design that will be presented (Fig. 2). The drive
train is modeled as a two rigid bodies linked by a flexible
shaft (Fig. 3). Also it was supposed that the two blades move
in unison and support the same forces.
In order to compute the model, we have started from a
theory that states that a mechanical system of arbitrary
complexity can be described by the equation of motion:

Fig. 2 The mechanical structure of the wind turbine

) , , , ( u t q q Q q K q C q M & & & & = == = + ++ + + ++ + (1)
where M, C and K, are the mass, damping and the stiffness
matrices, Q is the vector of forces acting on the system, and
q
i
is the generalized coordinate. For our model, the
generalized coordinates are: ) , , , , (
2 1 T G T
y q = == = , where

T
is the angular speed of the rotor,
G
stands for the angular
speed of the generator,
1
and
2
are the flaps of the blades,
while y
T
represents the horizontal movement of the tower
(Fig. 2).

Fig. 3 The two mass model representation of the drive train

Since the thrust forces acting on the blades are equal, it is
naturally to consider
1
=
2
= and F
aero1
= F
aero2
= F
aero
,
which transforms q into ) , , , (
T G T
y q = == = . In the same
time, one can find Q as being:
) 2 , , , (
aero aero em aero
F F C C Q = == = (2)
The considered forces that are acting on the system are:
C
aero
, the aerodynamic torque, C
em
, the electromagnetic
torque, and F
aero
, representing the thrust. The aerodynamic
torque and the force acting on the entire rotor are expressed
in terms of non-dimensional power coefficient C
P
and thrust
coefficient C
T
respectively, as follows
T
P aero
v
C R C

3
2
) , (
2
1
= == =
2 2
) , (
2
1
v C R F
T aero
= == = (3)
where represents the air density, R is the blade radius, and
978-1-4577-0123-8/11/$26.00 2011 IEEE 839

v is the average speed of the wind.
The power coefficient is one of the most important
parameters of the wind turbine because it offers information
upon the efficiency of the turbine, it helps defining the
control objectives in the below rated regime and also it
characterizes the aerodynamic torque that moves the
turbines rotor. The power and the thrust coefficients can be
expressed in a polynomial form, and depend on two
parameters which are the tip speed ratio and the pitch angle
of the blades.
In order to derive the mathematical model, one has used
the Lagrange equation that offers a systematic procedure to
calculate such models
Q
q
E
q
E
q
E
q
E
dt
d
i
P
i
d
i
c
i
c
= == = + ++ + + ++ +

& &
) ( (4)
Here, E
c
, E
d
, and E
p
denote the kinetic, dissipated and
potential energies. After a few calculations, applied for our
system, one obtains
2 2 2
2 2 2
2 2 2 2
2
) ( ) (
2
2
) ( ) (
2
) (
2 2 2
T
T
P P G T
A
P
T
T
P P G T
A
D
P T P T
T
G
G
T
y
k
r k
k
E
y
d
r d
d
E
r y M y
M J Jt
Ec
+ ++ + + ++ + = == =
+ ++ + + ++ + = == =
+ ++ + + ++ + + ++ + + ++ + = == =



&
&
&
& &
(5)
These energies were calculated under the supposition that
the generalized force that acts on the rotor is applied on a
point situated at the distance r
P
on each blade from the hub
of the rotor (Fig. 2). In the above equations, J
T
and J
G

represent the rotor and the generator moments of inertia, M
T

and M
P
are the masses of the tower and of the blade,

d
P
, d
A

and d
T
represent the damping coefficients for the blade, drive
shaft and tower. Similarly, k
P
, k
A
and k
T
stand for the spring
coefficients of the blade, drive shaft and tower.
T
and
G

are the angular positions of the rotor and generator.
The interconnection of the models of different plant
subsystems, leads to a global highly non linear system,
mainly because of the expressions of the aerodynamic torque
and of the thrust force, both given in (3).
For control design purposes, we linearized the model
around an operating point S
op

v D D D F
v D D D C
fv f T f aero
cv c T c aero
+ ++ + + ++ + = == =
+ ++ + + ++ + = == =




(7)
Here
Sop
aero
cv
Sop
aero
c
Sop
T
aero
c
v
C
D
C
D
C
D










= == =










= == =










= == = , ,


,and
Sop
T
aero
f
F
D










= == =

,
Sop
aero
f
F
D










= == =

,
Sop
aero
fv
v
F
D










= == = respectively.
Besides these equations, in order to interconnect the
models of the individual subsystems, one must include into
the model, the pitch controller. This was modeled here as a
first degree order system [7]:
s T
ref
+
=

1
1
, where
ref
is the desired pitch angle
and is the actual pitch angle of the blades.
We have taken into consideration the fact that the pitch
servomotor has some physical limitations, and we have
modeled them by including into our model one saturation in
the position and one in the speed. For this study we have
supposed that the saturation values in position are -45 and
45, and that the servomotor does not exceed the speed of
10 /s. In Fig. 4 one can observe the way the pitch
servomotors dynamics were modeled.

Fig. 4 The pitch servomotor dynamics modeling
After combining all these equations, one can put (4) into
the into the classical state-space representation
v
m E t u B t x A t x + ++ + + ++ + = == = ) ( ) ( ) ( & (8)
) ( ) ( ) ( t u D t x C t y + ++ + = == =
in which m
v
represents a perturbation acting on the system,
and from a physical point of view it models the eventual
wind gusts that appear.





































= == =
14 . 0 0 0 0 0 0 0 0 0
0 55 . 5 0 0 0 0 0 0 0
07 . 0 21 . 0 42 . 1 57 . 4 0 21 . 0 85 . 242 14 . 457 0
04 . 0 13 . 0 17 . 0 9 . 3 0 15 . 0 35 . 30 47 . 390 0
0 0 0 0 46 . 1 46 . 1 0 0 29 . 268
04 . 0 2 0 0 28 . 0 69 . 0 0 0 4 . 51
0 0 1 0 0 0 0 0 0
0 0 0 1 0 0 0 0 0
0 0 0 0 1 1 0 0 0
A

) 1 0 0 0 0 0 0 0 0 (
0 0 0
0 0 0
0 0 0
10 0 0
,
0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 1 0 0
0 0 0 0 25 . 0 0 0 0 0
0 0 0 0 0 1 0 0 0
0 55 . 5 0 0 0 0 0 0 0
0 0 0 0 10 44 . 2 0 0 0 0
1 0 0 0 0 0 0 0 0
5
5
= == =














= == =














= == =










= == =


E
D C
B
T

The system is multivariable; there have been identified
three inputs and four outputs (Fig. 5). As shown in this
control scheme, the input variables of the system are
considered: v
m
the average value of the wind speed, and the
two control variables: the pitch angle, , and the
electromagnetic torque C
em
.
Here, we have considered the state vector
T
T G T T G T
T
v y y x ) , , , , , , , , (
&
= == = , the output of
the system y = (P_el,
T
, , y
T
), and the command signal u =
(, C
em
). The first component of the output vector represents
the electrical power generated by the turbine. It can be
978-1-4577-0123-8/11/$26.00 2011 IEEE 840

computed as
em G
C P_el = == = but in this paper, its
normalized value was used.

Fig.5 The block scheme of the controlled system
The other output variables that we are interested in are
T

because the goal is to try to maintain it constant to its
nominal value, no matter the changes that appear in the
environment, the flap mode of the blades and of the tower
y
T
respectively, because, it is desired that these variables be
as much as possible.
The two available control variables are the pitch angle and
the electromagnetic torque. The numeric values of the wind
turbines parameters can be found in the APPENDIX, at the
end of the paper.

III. GENERAL PROCEDURE OF THE LINEAR
QUADRATIC CONTROLLER DESIGN

As previously said, there is a large variety of control
techniques that were applied to wind turbines in a permanent
attempt to improve their functioning and to benefit as much
as possible from the energy that they can produce. In
literature, one can find proposed solutions for mono-variable
systems as well as for multi-variable ones.
In [8], for instance, one can find a compared study made
upon the simulation results obtained with three controllers: a
classical PID regulator, a full state feedback and a fuzzy
controller. The authors conclusion is that the PID controller
ensures good performances with power regulation but not
with reducing the structures mechanical loads. In the same
time, the full-state feedback controller manages to reduce
these loads even under turbulent conditions.
The idea of conveniently sizing a trade-off between energy
efficiency and increasing the lifetime of the wind turbines by
alleviating fatigue loads is continuously being paid special
attention, even when employing controllers like PI or PID.
However, these approaches do not allow a rigorous control
design in order to perform a fine tuning of the trade-off
between the energy performance and the reliability demands
[9].
These aspects, together with its design simplicity and the
advantages it could bring, lead us to the idea of choosing a
state feedback linear quadratic controller (LQR) for this
study.
For its design, one imposes a quadratic cost function
defined as
( (( ( ) )) ) dt u R u y Q y J
T T
+ ++ + = == =


0
(9)
The feedback control law that minimizes the value of this
cost is given by:
r K x K u
r
+ ++ + = == = ,
where K is given by
c
T
P B R K = == =
1
, P
c
is given by the
solution to the equation:
0
1
= == = + ++ + + ++ +

Q P B R B P P A A P
c
T
c c
T
c
(10)
while K
r
is being defined by:
Q C K B A B R K
T T T
r
= == =
1 1
1
) ) (( ) ( (11)
This matrix ensures the reference input is scaled in order
to become equal to the feedback signal provided by the LQR
regulator. This algorithm guaranties that no matter, any two
symmetric and positive definite matrixes Q and R that we
chose in order to minimize the quadratic criteria, there is
always a matrix P
c
, also symmetric and positive definite, that
represents the solution of the Ricatti equation (10).
Through this criterion, by replacing the variables y and u
by the corresponding vectors presented in Section 2, one
tries to minimize the flap mode of the blades and the tower
oscillation respectively, maintain the electrical power level
and the angular speed of the rotor at the desired levels while
computing the appropriate command.
The typical rule for choosing the weighting matrixes R
and Q is the Brysons rule, which states that these matrixes
should be selected as diagonal with the non-zero elements
scaled so that the variables that appear in the optimization
criterion have a maximum value of one [10] [11] [12].
This is important especially for the situations when the
units used for the different components of the command and
state vectors are numerically very different from each other.
This is also our case, in the command vector, for instance,
the pitch angle and the electromagnetic torque have different
order of degree units.
Although Brysons rule gives good results, often it is just a
starting point of a trial and error procedure of choosing these
matrixes, in order to obtain the desirable properties for the
closed loop system. Weights reflect the relative importance
given to the state with respect to the control effort.
Therefore, for our system, if one chooses large values for
Q compared to the values in R, one gives a higher
importance on the minimization of the mechanical weights
and a lower importance to the command effort [13] [14].
IV. RESULTS
The simulations were done using MATLAB/SIMULINK
software and the results proved good performances. The
chosen operating point for the linearization of the system
corresponds to the average value of the wind speed of 18m/s.
In Fig. 6 one can see the scheme that was used for the
simulation.
The two reference variables, for the normalized electrical
power P_el_ref and for the angular rotor speed
T_ref
respectively, were chosen as constants with the appropriate
values because the goal is to minimize the variations of the
electrical power extracted around the nominal value of the
978-1-4577-0123-8/11/$26.00 2011 IEEE 841

generator and we also want to keep the rotor speed constant.


Fig. 6 The simulation of the system with LQR regulator
The weighting matrices mentioned in (9) and used for
these simulations are
R = I,














= == =
1 0 0 0
0 1 0 0
0 0 400 0
0 0 0 3 . 0
Q .
These values were chosen using the methods mentioned
above and also based on the fact that they provided very
good performance of the system in terms of achieving good
responses and not very strong control actions.
The cost function was written in the following form:


+ ++ + + ++ + = == =
0
1 1
) 2 ( u S x u R u x Q x J
T T T
, where
1 1 1
C Q C Q
T
= == = ,
1 1 1
D Q D R R
T
+ ++ + = == = ,
1 1
D Q C S
T
= == = ,
and the matrices C
1
and D
1
being the truncated blocks from
the system matrices C and D. These matrices contain the
lines and columns from C and D corresponding to the control
variables C
em
and .
The system is controllable and it does not contain
unobservable modes. One important property of LQ
regulators is that provided these conditions, they guarantee
nominally stable closed loop systems.
In Fig. 7-10, one can see the results obtained in
simulation.

Fig. 7 The normalized electrical output power of the turbine

Fig. 8 The variation of the angular speed of the rotor

Fig. 9 The tower bending movement in the direction of the nacelle

Fig. 10 The variation of the first flap mode of the blades
It can be observed that the electrical output power and the
angular speed of the rotor manage to follow the reference
and to maintain their nominal imposed values.
In the same time, the variables that were meant to be
minimized, namely the first flap mode of the blades and the
bending of the tower, have extremely small values. The
blades have a deviation of about 5mm while the tower has an
insignificant movement on the horizontal direction.
APPENDIX
THE NUMERICAL VALUES OF THE WIND TURBINE
PARAMETERS
Symbol Physical measure Value
J
t
Turbine inertia 214 000 Kg
* m
2
978-1-4577-0123-8/11/$26.00 2011 IEEE 842

J
g
Generator inertia 41 Kg * m
2

M
T
Tower and nacelle mass 35000 kg
M
p
Blade mass 3000 kg
k
P
Blade Stiffness Coefficient 1000 Kg *
m
2
/s
2

k
T
Tower Stiffness Coefficient 8500 Kg *
m/s
2

k
A
Drive Shaft Stiffness
Coefficient
11000 Kg *
m
2
/s
2

d
P
Blade Damping coefficient 10 000 Kg *
m
2
/s
d
T
Tower Damping coefficient 50 000 Kg *
m/s
d
A
Drive shaft damping
coefficient
60 000 Kg *
m
2
/s
r
P
Distance from the rotor hub 8 m
N Number of blades 2
D The rotor diameter 34 m
P
n
Nominal Power 400 kW

nom
Nominal rotor speed 4 rad/s
h Tower height 47 m
REFERENCES
[1] D. O. Kirk, Optimal control theory-An Introduction, Dover
Publications, 2004, pp. 3-5.
[2] N.A Cutululis, I. Munteanu, E. Ceanga, Optimal control structure for
variable speed wind power system, The annals of Dunarea de Jos
University of Galati, Fascicle III, pp 95-102, 2002.
[3] A. D. Wright, Modern control design for flexible wind turbines
Technical Report, NREL/TP-500-35816, July 2004, pp.23-26.
[4] D. F. Bianchi, H. Battista, Wind turbine control systems Principles,
Modeling and gain scheduling design, Springer-Verlag, London,
2002
[5] L. Lupu, B. Boukhezzar, Pitch and torque control strategy for
variable speed wind turbines, Proceedings EWEC, Athens, 2006.
[6] F. A. Vanegas, M. Zamacona, Robust control solution of a wind
turbine - A simulation study, International Masters Thesis in
Information Technologies, Halmstadt University, February, 2008.
[7] F. Lescher, P.Borne, Robust gain scheduling controller for pitch
regulated variable speed wind turbine, Studies in Informatics and
Control, vol 14, No.4, pp 299-315, 2005.
[8] M. Jelavic, I. Petrovic, Design of a wind turbine pitch controller for
loads and fatigue reduction, Electrical Engineering Institute,
Proceedings of the European Wind Energy Conference & Exhibition
EWEC, Milan, Italy, 2007.
[9] I. Munteanu, E. Ceanga, Optimal control of wind energy systems
Towards a global approach, Springer, 2007.
[10] B. D. O. Anderson, J. B. Moore, Optimal control: Linear quadratic
methods, Prentice Hall, Englewood Cliffs, NJ, 1990.
[11] F. C. Callier, J. L. Willems, Criterion for convergence of the solution
of the Riccati differential equations, IEEE Trans. Automat. Contr.
AC-26, pp. 1232-1242, 1981.
[12] R. Sivan, H. Kwakernaak, Linear optimal control systems, Wiley-
Interscience, 1
st
edition, October 1972.
[13] A. E. Bryson, Applied linear optimal control Examples and
Algorithms, Cambridge University Press, 2002.
[14] J. L. Hellerstein, Y. Diao, S. Parekh, D. M Tilbury, Feedback control
of computing systems, John Wiley & Sons, Inc., 2004.

978-1-4577-0123-8/11/$26.00 2011 IEEE 843