Recent Advances in Automatic Control, Information and Communications

Wind Turbine Control with Multiple Model Predictive Control

JAKUB NOVAK, PETR CHALUPA
Faculty of Applied Informatics
Tomas Bata University in Zlin
Nam T.G.Masaryka 5555, Zlin
CZECH REPUBLIC
jnovak@fai.utb.cz
Abstract: - In this work, model predictive control of a wind-turbine system is studied. The nonlinear process is
represented by a set of local linear models which were linearized for different wind speeds. Since not all the
states are measurable, operating point and active local model are determined by effective wind speed estimation
using an extended Kalman filter (EKF). The estimated states and active model parameters are then used within
predictive control strategy for computation of control signals. Due to different control demands in different
operating regimes of the wind turbine the weighting matrices are also scheduled for different wind speeds.
Simulations show that the wind-turbine system can be successfully controlled by this methodology.
Key-Words: Extended Kalman filter, Predictive control, Wind turbine
constraints on the model of the 5MW wind turbine.
In Section 2, nonlinear mathematical model of the
wind turbine is addressed and the linearization of
this model for different steady state is given. State
and wind estimation method based on Extended
Kalman filter is presented in Section 3. The concept
of Predictive control is introduced in Section 4.
Finally, in Section 5, simulation results for model of
the 5MW wind turbine are shown with brief
conclusion.

1 Introduction
Wind energy is by far the fastest-growing renewable
energy resource. In 2012, the European Union
installed capacity of wind turbines reached
105 000MW. The energy available in the wind is
obtained through the wind conversion process,
which is strongly nonlinear and very challenging
from the control system viewpoint as the power
obtained from the wind is proportional to the third
power of the effective wind speed. The control
system have to be able to optimize the wind energy
conversion by application of appropriate generator
torque in weak wind and also manipulate the pitch
blade angle to keep the output power at nominal
value in strong winds. The effective control strategy
also reduces structural fatigue and load affecting the
drive-train and tower structure, leading to
potentially longer lifetime of the wind turbine [1].
Several methods have been developed to control the
wind turbines. Conventional PI controllers have
been tested in [2]. In [1] multivariable approach to
the control of wind turbines in the context of gain
scheduling techniques is applied. Soliman et al.
developed a Multiple Model Predictive controller to
control variable speed variable-pitch wind turbine
over its full operating range [3]. Robust control
techniques of the same type of wind turbine have
been investigated in [4], [5]. To cope with nonlinear
behavior of the wind turbine system, adaptive
control approaches have been proposed in [6].
Review of the control techniques for horizontal axis
wind turbines is presented in [7]. This work presents
application of the MIMO predictive control with

ISBN: 978-960-474-316-2

2 Mathematical Model of a Wind

Turbine
The wind turbine characteristics can be divided into
4 parts:
Aerodynamics
Turbine mechanics
Generator dynamics
Actuator dynamics
The rotor blades convert the kinetic energy of the
wind into mechanical energy, providing the torque
on the rotor:

(1)
where the is the rotational speed of the rotor and
the rotor power is given by the relation:
,
(2)
is the blade radius and is
is the air density,
the effective wind speed. The efficiency coefficient
is a function of tip speed ratio and blade pitch
angle .

97

Recent Advances in Automatic Control, Information and Communications

(9)
where variable describes the twist of the shaft,
and
are the rotor and generator inertia,
respectively. The effective wind velocity is modeled
superimposed by
as a mean wind velocity
turbulent wind velocity .
(10)

0.5
0.4
cp

0.3
0.2
0.1
0
30

The turbulent wind velocity is modeled as a second

order filter driven by a white noise process:
(11)
(12)
, ,
are functions of the mean wind
where
speed and 0,1 is a white noise process. The
, , are found by second order
parameters
approximation of the wind power spectrum
density[9].
The Equations 5-12 are combined to yield the
nonlinear model of the wind turbine:

100

20

50

10

0 0

Fig. 1 The efficiency coefficient

as a function of
tip speed ratio and blade pitch angle (negative
values were zeroed)
The tip speed ratio is given as a ratio between the
blade tip speed and effective wind velocity:
(3)

, ,

The values of power coefficient are usually

provided in the form of look-up table [8]. A three
dimensional plot of curve is depicted on Fig. 1.
The generated power is given by:
(4)
where
is rotational speed of the generator. No
.
losses due to bearings etc. are assumed
The generator torque can be controlled and is
modeled as a first order system with time constant
:
(5)
where
is the generator torque setpoint. The pitch
of the blades is also changed by actuator with first
order dynamics with time constant :
(6)

0
0 0
0 0

where
is the blade pitch angle setpoint. The
mechanistic part of the turbine can be divided into
rotor and generator side connected by a shaft. The
dynamic nature of the shaft is described by the
and spring constant
. The torques
damping
acting on each side of the transmission are related
by the gear ratio
. Using the Newtons second
law for rotating bodies the following equations can
be formulated to describe the dynamics of the
mechanical part of the wind turbine:

0
0
0
0
0
0

(13)
The slow varying element of the wind velocity
is
assumed to be known. The fuzzy global model of
the process is obtained through piecewise
linearization. The nonlinear system:
x (t ) f ( x(t ), u (t ))
(14)

y (t ) f ( x(t ), u (t ))

where x(t ) R n , u (t ) R p , y (t ) R q represent the

states, input value and output value, respectively.
The system is linearized around the operating point
using the Taylors series approximation. This results
in a series of M local linear models.
As the controller will be working in all operating
modes it is beneficial to convert the incremental

(7)
(8)

ISBN: 978-960-474-316-2

0 0
0 0
0 0
0

98

Recent Advances in Automatic Control, Information and Communications

The linearization of the functions

, ,
is
accomplished off-line during the modeling process
and the Jacobians can be computed using the
linearized local models matrices as:
(26)
|

model into affine and discretize to yield model in

the form:
(15)
1
(16)
with
1,2, , .
In order to provide off-set free control in case of
unmeasured disturbances and modeling errors the
model is augmented with the model of unmeasured
input disturbances with integral character:
1
(17)
1
(18)
The augmented system is then given as:
1
1
1

0
0

1 0
0 1

0
0

4 Model-based Predictive Control

Model-based predictive control (MPC) has been
successfully used in many industrial applications
due to its ability to handle MIMO control problems
with constraints on the system variables [11]. The
objective for control of a wind turbine is to
maximize power production and minimize the
mechanical stress on the components of the wind
turbine. In order to control the wind turbine in the
whole spectrum of wind speeds the parameters of
model are scheduled based on the current mean
wind speed. The state-space model based predictive
control is based on linear time varying model that is
obtained at every sampling interval and its
parameters are used for the entire prediction horizon
:
x(k 1) Ax(k ) Bu (k ) E
y ( k ) Cx(k ) Du (k ) F

(19)

0 0

(20)

The augmented model is used for state estimation in

extended Kalman filter.

3 State Estimation with Extended

Kalman Filter
Not all the states of the variable speed wind turbine
system can be measured and furthermore the
measurement of available outputs is bound to be
corrupted with measurement noise. In order to
estimate the states of the nonlinear model the
extended Kalman Filter (KF) is designed. The
extended KF is the nonlinear extension of the linear
Kalman filter where the model is nonlinear [10]:
1
,
(21)
(22)
where and are process and measurement noises,
which are assumed to have Gaussian distribution
with covariances
and
. Since the system is
nonlinear the Jacobians are used instead of
nonlinear expressions to compute covariance and
consequently Kalman gain. The update equations

i 1

i 1

i 1

i 1

i 1

i 1

(28)

where are the weights of each local model and are

a function of the current wind speed. The
computation of a control law of MPC is based on
minimization of the following criterion

T
J MPC Y W Q Y W URU +
(29)

U
U
S
U
U

s
s
with input constraints:
(30)

(23)

(31)

1,2, ,
1,
is the control horizon,
where

Y is a vector future prediction of the system, W

is a vector of reference trajectory and Q, R, S are
positive definite weighting matrices. The last term
represents the deviation from inputs that
led to linearization point of the model.

(24)
is the estimation covariance error matrix,
is the measured output vector,
is the
Kalman gain. The covariance error matrix is
obtained as a solution of algebraic Riccati equation:

where

(25)

ISBN: 978-960-474-316-2

C = pi Ci , D = pi Di , F = pi Fi

,
|

A = pi Ai , B = pi Bi , E = pi Ei

of the EKF are given as:

1|

(27)

99

Recent Advances in Automatic Control, Information and Communications

Th
he wind turb
bine can be ooperated in four differennt
reg
gimes (Fig. 3).
3 In the reggion I the rottational speeed
of the rotor is at its low
west allowab
ble level. Thhe
con
ntrol strategy
y is to keep the pitch off the blades at
a
thee optimum value and control thee system by
b
maanipulating the
t generatorr torque . In the regioon
II the rotationaal speed of the rotor
is within thhe
mits ,
and the blade pitch is
i
lim
,
kep
pt at the optimum value prroviding thhe
maaximization of . The m
maximal valu
ue of is foor
blaade pitch
0 and tip sp
speed ratio
7.5.

5 Implem
mentation
The
T proposedd predictive control schem
me is presennted
in
i Fig. 2.

Fig.
F 2 Predictive control scheme

1.5

Based
B
on thee measured outputs an
nd inputs the
states
s
of the system arre estimated using multiiple
Kalman
K
filterrs whose outtputs are weiighted using the
fuzzy
f
weightts scheduled with mean wind speed
which
w
is asssumed know
wn and estim
mated turbullent
speed
s
. Thhe actual paarameters off the linearizzed
model
m
are obbtained usingg (30). The setpoints
s
aand
steady-state
s
inputs
foor computatiion of criterrion
(31)
(
are alsoo computed for actual effective w
wind
speed.
s
The
T wind tuurbine that is
i being con
nsidered in the
paper
p
is a w
wind turbinee model wid
dely known as
NREL

offshhore 5-MW baseline win

nd turbine tthat
was
w developped by Naational Reneewable Enerrgy
Laboratory
L
of the Uniteed States off America. T
This
wind
w
turbinee is a convenntional three--bladed upw
wind
variable-spee
v
ed
variablle
blade-p
pitch-to-feathhercontrolled
c
tuurbine. The parameters
p
of the model are
presented
p
in Table 1 [12].

II.

[rad/s]

I.

IV.

1
III.
0.5

P [[MW]

10

0
6
x 10

10

15

2
20

25

10

15

2
20

25

0
4
x 10

10

15

2
20

25

10

15

2
20

25

[deg]

40

20

0
5

T [N.m/s]
/ ]

4
3
2
1
0

v [m
m/s]

Fig
g. 3 Operatio
on modes of tthe NREL wind
w
turbine

Table
T
1 NRE
EL wind turbine parameteers
parameter
nominal pow
wer ,
rated rotor sppeed ,
drive-train sppring const.
drive-train daamp. const.
generator ineertia
rotor inertia
blade radius
gear ratio
max blade piitch
min blade pittch
max bl. pitchh rate
min bl. pitch rate
max gen. torqque ,
min gen. torqque ,
max
geen.
torquue
rate ,
min
genn.
torquue
rate ,
pitch actuatoor const.
gen. actuatorr const.

ISBN: 978-960-474-316-2

units
[MW]
[rad/s]
[N.m/rad]]
[N.m/rad.s]
[kg.m2]
[kg.m2]
[m]
[-]
[deg]
[deg]
[deg/s]
[deg/s]
[N.m]
[N.m]
[N.m/s]

value
5
1.2671
867.637E
E6
6.215E6
534.116
3.8768E77
63
97
90
0
8
-8
47402.977
0
15000

[N.m/s]

-15000

[s]
[s]

0.12
0.1

he primary ob
bjective is too keep tip speeed ratio at itts
Th
op
ptimal valuee
to
t maximizze
turrbines aerod
dynamic effi
ficiency. In the
t region IIII
thee rotational speed
s
of the rrotor is at its maximum
m
bu
ut generated power
is below its nominal valuee.
In this transition region iss system is controlled
c
thhe
sam
me way as in
i the regionn II. In the top
t region IV
V
bo
oth the rotaational speedd of the rotor
annd
gen
nerated pow
wer
are at their rated values.
v
In thiis
reg
gion the torq
que is kept cconstant and blade pitch
is used
u
to comp
pensate the vvariation in wind
w
power.
Th
he wind turbiine is controolled in all reegions. As thhe
dy
ynamics of the
t system is different for differennt
wiind speeds a local m
model was obtained by
b
lin
nearization of the nonlineear model of the process.
21 local models were deeveloped forr mean winnd
25 . The paarameters of
o the system
m
speeed 4,2
bettween these local modelss are obtaineed using fuzzzy
meembership function
f
as shown on Fig. 4. Thhe
sam
mpling perio
od is choseen to be 0.02
25s.

100

Recent Advances in Automatic Control, Information and Communications

M1

MPC was implemented with constraint on the blade

pitch and generator torque given by:

(39)
(40)
,
,

(41)
,

,
(42)
The weights for the MPC were also scheduled using
fuzzy membership function and their values for
different wind speeds are presented in Table 2.

M21

M2

p [-]

0.8
0.6
0.4
0.2
0
-0.2

10

15

20

25

v [m/s]

Fig. 4 Fuzzy membership functions

It is assumed that only the generator speed , blade
pitch and output power can be measured and
is assumed to be known:
the mean wind speed

Table 2 MPC weights scheduling

v
Q

(32)

The other system states and unmeasured

disturbances are estimated via extended Kalman
filter. Since the wind does not affect the actuators
states the process noise is added to these states in
order to obtain nonzero Kalman gain. The process
noise covariance matrix is then given as:

where

0.001 0
0 0 0
0
0.1
0 0 0
0
0 1
0
0
(34)
0
0 0 0.001 0
0
0 0
0
0.1
0
0
0
0
0
0
0
0
0
0
0
0
0
/
0
0
0
0
/
(35)
0
0
0
0
0
0
0
0
0
0
/
0
0
/
0
0
0
0 0 0 0 0 0
0 0
(36)
The measurement covariance matrix was set to [13]:

v [m/s]

10

50

100

150
T [s]

200

250

300

200

250

300

200

250

300

Fig. 5 Turbulent wind in the simulation

1.3
r,max

1.2
1.1
1
0.9

(37)

0.8
0

0
0
Model predictive controller was employed for
computation of manipulated variables. Prediction
and control horizons were set to:
2 ,
1
(38)
For such values acceptable performance was
obtained and these values does not represent big
computational burden for the processing unit. The

ISBN: 978-960-474-316-2

15

5
0

12-25 m/s
10
0
0
0
10 0
0 0
0 0
0 0

0
10
0
10
0
10

20

r [rad/s]

11m/s
10
0
10
0
10
0

The proposed control strategy was tested through

simulation with turbulent wind that drives the
turbine through different operating modes. With the
MPC weights given in the Table 2 both the output
power and rotor speed are tracked accurately in
region IV. The turbulent wind affecting the wind
turbine during the simulation is presented in Fig. 5.
The control courses for reference tracking test are
presented in Fig. 6 and 7. In the region II the main
focus of the control is to keep the tip speed ratio at
the optimal value so the figure of tip speed ratio is
also presented in Fig. 8.

(33)

4-10 m/s
10
0
0
10
10
0
0
10
10
0
0 10

50

100

150

x 10

Pe [MW]

Pe,norm

2
0
-2
0

50

100

150
Time [s]

and rotor
Fig. 6 Wind turbines power output
in simulation with turbulent wind
speed

101

Recent Advances in Automatic Control, Information and Communications

References:
[1] F.D. Bianchi, H. Batista, R.J. Mantz, Wind
Turbine
Control
Systems:
Principles,
Modelling and Gain Scheduling Design,
Heidelberg: Springer, 2007.
[2] A.D. Hansen, P. Sorensen, F. Iov, F. Blaabjerg,
Control of variable speed wind turbines with
doubly-fed induction generators, Wind Eng.,
Vol. 28, pp. 411443, 2004.
[3] M. Soliman, O.P. Malik, D.T. Westwick,
Multiple Model predictive Control for Wind
Turbines With Doubly Fed Induction
Generators, IEEE Transactions on Sustainable
Energy, Vol. 2 (3), 2011, pp. 215-225.
[4] M. Mirzaei, N.K. Poulsen, H.H. Niemann,
Robust Model Predictive Control of a Wind
Turbine, Proceeding of the American Control
Conference, 2012, pp. 4393-4398.
[5] H. Camblong, Digital robust control of a
variable speed pitch regulated wind turbine for
above rated wind speeds, Control Engineering
Practice, vol. 16, no. 8, 2009, pp. 946-958.
[6] S.A. Frost, M.J. Balas, A.D. Wright, Direct
Adaptive Control of a Utility-Scale Wind
Turbine for Speed Regulation, Int. Journal of
Robust Nonlinear Control, Vol. 19 (1), 2009,
pp. 5971.
[7] J.H. Laks, L.Y. Pao, A. D. Wright, Control of
Wind Turbines: Past, Present, and Future,
Proceedings of the American Control
Conference, 2009, pp. 2096-2103.
[8] J.D. Grunnet, M. Soltani, T. Knudsen, M.
Kragelund, T. Bak, Aeolus Toolbox for
Dynamic Wind Farm Model, Simulation and
Control, Proceedings of the 2010 European
Wind Energy Conference, 2010
[9] S. Thomsen, Nonlinear Control of a Wind
Turbine. Masters thesis, Technical University
of Denmark, 2006.
[10] D. Simon, Optimal State Estimation: Kalman,
H-Infinity, and Nonlinear Approaches, WileyInterscience, 2006.
[11] J. Maciejowski, Predictive Control With
Constraints, Prentice-Hall, 2000.
[12] J.M. Jonkman, S. Butterfield, W. Musial, G.
Scott, Definition of a 5-MW reference wind
turbine for offshore system development,
Technical report, NREL, 2009.
[13] A. Gosk, Model Predictive Control of a Wind
Turbine, Masters thesis, Technical University
of Denmark, 2011.

30
25
[deg]

20
15
10
5
0
0
4
x 10

50

100

150

200

250

300

50

100

150
Time [s]

200

250

300

Tg [N.m/s]

4
3
2
1
0
0

Fig. 7 Blade pitch angle

and generator torque
in simulation with turbulent wind
8

opt

[-]

7
6
5
4
0

50

100

150
T [s]

200

250

300

Fig. 8 Tip speed ration in simulation with turbulent

wind

6 Conclusion
In the paper, fuzzy predictive control concept is
applied to the simulated wind turbine model. It was
assumed that measurement noise was present and
not all the states are measurable to make the control
more realistic. The weight scheduled predictive
controller was implemented for control of the wind
turbine in all operating regimes in the turbulent
wind conditions. The MPC controller enables the
desired tradeoff between power smoothing and
speed regulation while reducing generator torque
fluctuations and pitch actuator activity. The use of
future wind speed measuring device as LIDAR
might improve the performance of the control
system when implemented in the predictive control
structure. The simulation on 5MW wind turbine
model shows that the strategy can be applied to the
wind turbine effectively.

Acknowledgement
The authors kindly appreciate the financial support
which was provided by the European Regional
Development Fund under the project CEBIA-Tech
No. CZ.1.05/2.1.00/03.0089.

ISBN: 978-960-474-316-2

102