You are on page 1of 45

Wind turbine control

Control of a 1.5MW wind turbine for load reduction


B.P. Lemmen BSc.
CST 2010.047
Eindhoven, June 22, 2010
2
Contents
1 Linearization 7
1.1 Floquet analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 PID control 11
2.1 Angular velocity high speed shaft . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Tower acceleration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Individual pitch control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Power control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5 Simulations and results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5.1 Wind input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5.2 Sensor noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5.3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 Kalman lter 27
3.1 Linearizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Appending states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4 LQG controller 33
4.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5 Conclusions and recommendations 39
A Coleman transformations 41
3
4
Introduction
In this report the goal is to design dierent ways to control a 1.5 MW wind turbine and compare
the performance of these dierent controllers. All of the controllers will be used to control the
wind turbine in above rated wind conditions. In above rated wind conditions the generator
torque is kept constant while the blades of the rotor are pitched to assure a constant angu-
lar velocity of the rotor. This way the power output of the generator should be constant. In
these conditions increasing the power output further could potentially damage the wind turbine.
A non-linear model of a wind turbine is used for simulations. This wind turbine model is a
3 blade, 1.5 MW wind turbine with a hub height of 85 meters and blades with a length of 35
meters. In above rated wind conditions the desired angular velocity of the rotor is 20.48 rpm
for this turbine. Maintaining this angular velocity is the rst requirement for the controllers.
The second requirement is based on load reduction. Load reduction should, if possible, be fo-
cussed on reducing the loadings on the tower and blades. To do this the blades can be pitched
individually and small changes can be made to the generator torque. For the load reduction
the tower accelerations in both for-aft and side-side directions are measured as well as the blade
edgewise and apwise bending moments at the root of the blade. Here the edgewise direction is
the bending in the rotational plane of the rotor and the apwise direction is the bending out of
this rotational plane.
The controllers that will be designed rst are several PID controllers and secondly a Kalman
lter and LQG controller. For all of these controllers it is necessary to linearize the non-linear
model equations. This will be discussed in the rst chapter where also a structural analysis is
performed. In the second part the design and performance of the PID controllers is discussed.
Controllers will be designed to control the angular velocity of the rotor, to reduce the tower
acceleration and to reduce the apwise bending moment on the blades. Also a multi input
single output controller will be designed to control the power by varying both the torque and
the pitch of the blades. Next the design of a Kalman lter is discussed and the performance of
this lter is assessed. The state estimates from the Kalman lter are than used together with a
LQG controller in the last part of this report. Finally the results from the PID controllers and
the LQG controller are compared.
5
6
Chapter 1
Linearization
Before a controller can be designed a linearized model of the plant is needed. In this chapter
the linearization of the plant model will be discussed. For the simulations the NWTC (National
Wind Technology Center) Design Code FAST [1] is used. FAST has the ability to linearize
the nonlinear model. First a simulation is started and FAST will simulate until a steady state
solution, over one rotation, of the wind turbine is reached. After the steady state has been
reached the model will be linearized at a number of dierent azimuth angles. FAST then
determines system matrices, either for a rst order model, see equation 1.1, or a second order
model, see equation 1.2.
x = Ax +Bu +B
d
u
d
y = Cx +Du +D
d
u
d
(1.1)
M q +C q +Kq = Fu +F
d
u
d
(1.2)
The linearization will be performed for four dierent wind speeds, 12, 16, 20 and 25 meters per
second. Each solution will be determined at 36 dierent azimuth angles. This results in four
sets of 36 linearizations. For these linearizations the following parameters are used:
Inputs:
Torque
Individual pitch of the blades
Degrees of freedom:
First apwise blade mode
Second apwise blade mode
First edgewise blade mode
Generator (angle)
First fore-aft tower bending-mode
Second fore-aft tower bending-mode
First side-to-side tower bending-mode
Second side-to-side tower bending-mode
7
Outputs:
Tower for-aft and side-side acceleration
Low speed shaft angle and speed
High speed shaft speed
Yaw angle
Blade root edgewise bending moments for individual blades
Blade root apwise bending moments for individual blades
Individual pitch angles
In the model that results from these computations some of the variables are in a non-rotating
coordinate frame while others are dened in a rotating coordinate frame. To bring all of the
variables in a non-rotating frame the Coleman transformation is used. These transformations
are discussed in appendix A. After this transformation the found system matrices can be used
for both structural analysis and determination of transfer functions.
1.1 Floquet analysis
A structural analysis can be performed using the results from the linearizations and the sub-
sequent Coleman transformations. This analysis consists of four dierent steps, as described
in [7]:
1. Compute the transition matrix for one rotation of the rotor. The solution after one period
is given by:
x(T) = (T, 0)x(0)
Where is the transition matrix. This matrix can be determined by integration from
t = 0 until t = T with (0, 0) = I, the identity matrix.
2. Determination of the characteristic multipliers
i
by eigenvalue analysis of (T, 0). The
eigenvalues of the matrix (T, 0) are the characteristic multipliers of the system.
3. Using the following two equations the modal damping coecient and the modal frequency
can be computed.

i
=
1
T
ln|
i
|

i
=
1
T

tan
1
Im(
i
)
Re(
i
)
+ 2k

If all the
i
< 0 the system is stable.
4. With the eigenvectors of the eigenanalysis from step 2 the modeshapes of the system can
be computed.
In the gures 1.1 until 1.4 the characteristic exponents
i
=
i
+j
i
, with j =

1, are plotted.
For 20 and 25 meters per second wind speeds there is one characteristic exponent which is larger
than zero. This suggests that for these wind speeds the structure is unstable.
8
12 10 8 6 4 2 0
1.5
1
0.5
0
0.5
1
1.5
Characteristic exponents
Real [rad/s]
I
m
a
g
i
n
a
r
y

[
r
a
d
/
s
]
Figure 1.1: Characteristic exponents Floquet analysis for 12 mps winds.
12 10 8 6 4 2 0
1.5
1
0.5
0
0.5
1
1.5
Characteristic exponents
Real [rad/s]
I
m
a
g
i
n
a
r
y

[
r
a
d
/
s
]
Figure 1.2: Characteristic exponents Floquet analysis for 16 mps winds.
9
10 8 6 4 2 0 2
1.5
1
0.5
0
0.5
1
1.5
Characteristic exponents
Real [rad/s]
I
m
a
g
i
n
a
r
y

[
r
a
d
/
s
]
Figure 1.3: Characteristic exponents Floquet analysis for 20 mps winds.
12 10 8 6 4 2 0 2
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
Characteristic exponents
Real [rad/s]
I
m
a
g
i
n
a
r
y

[
r
a
d
/
s
]
Figure 1.4: Characteristic exponents Floquet analysis for 25 mps winds.
10
Chapter 2
PID control
System matrices obtained from the Coleman transformations can be used to determine transfer
functions for dierent input output combinations. These can be used in the design of a controller.
In this chapter PID control will be used for two main dierent approaches. First only the angular
velocity of the generator will be used as a control variable. Next both the angular velocity and
the torque of the generator will be controlled, this could potentially reduce the loads on the
wind turbine. As a nal step individual pitch control will be introduced for load reduction on
the blades.
2.1 Angular velocity high speed shaft
In this section the angular velocity of the generator side of the turbine is used as a controlled
variable. With the transformed linearization obtained in the previous chapter the transfer be-
tween the pitch angle as an input and the rotational speed can be determined. The 36 dierent
models for one wind speed are rst averaged. Using equation 2.1 the transfer function can be
determined. So now there are four transfer functions at dierent wind speeds. These transfer
functions now describe the system over one rotation.
G(s) = C(sI A)
1
B +D; With I the identity matrix (2.1)
By extracting dierent columns and rows from B and C respectively dierent transfer functions
can be computed. The derived model can now be veried by means of the step responses. In
gure 2.1 the simulation results for dierent wind speeds are shown. In these simulations the
pitch angle is stepped up by one degree after 100 seconds. Several system properties can be
derived. First is it clear that for each wind speed the transfer is stable. Also when the pitch
angle is stepped up the rotational speed decreases. In the right gure the step responses are
shown zoomed, it can be seen that there are several right half plane zeros for these systems.
The models that where derived using the linearizations show for some wind speeds unstable
behavior, i.e. open loop right half plane poles. These systems are corrected to make sure that
the step response of the model is close to the actual step response from the simulations. The
step response of these corrected models are compared to the simulation results in gure 2.2. All
of the system properties mentioned above have to be considered in the design of a controller.
The bode diagrams of the models are shown in gure 2.3
11
100 150 200
1600
1620
1640
1660
1680
1700
1720
1740
1760
1780
1800
Time [s]
H
i
g
h

s
p
e
e
d

s
h
a
f
t

a
n
g
u
l
a
r

v
e
l
o
c
i
t
y
99 100 101 102 103
1750
1755
1760
1765
1770
1775
1780
1785
1790
1795
1800
Time [s]
H
i
g
h

s
p
e
e
d

s
h
a
f
t

a
n
g
u
l
a
r

v
e
l
o
c
i
t
y


12 mps
16 mps
20 mps
25 mps
Figure 2.1: Step responses from simulations. Right gure is zoomed in on left gure.
100 150 200
1600
1650
1700
1750
1800
Time [s]
H
s
s

a
n
g
u
l
a
r

v
e
l
o
c
i
t
y

[
r
p
m
]


FAST
Model
100 150 200
1600
1650
1700
1750
1800
Time [s]
H
s
s

a
n
g
u
l
a
r

v
e
l
o
c
i
t
y

[
r
p
m
]


FAST
Model
100 150 200
1600
1650
1700
1750
1800
Time [s]
H
s
s

a
n
g
u
l
a
r

v
e
l
o
c
i
t
y

[
r
p
m
]


FAST
Model
100 150 200
1600
1650
1700
1750
1800
Time [s]
H
s
s

a
n
g
u
l
a
r

v
e
l
o
c
i
t
y

[
r
p
m
]


FAST
Model
Figure 2.2: Step responses from simulations and models. Top left: 12 mps. Top right: 16 mps.
Lower left: 20 mps. Lower right: 25 mps.
12
0
20
40
60
80
M
a
g
n
i
t
u
d
e

(
d
B
)
10
5
10
0
900
720
540
360
180
0
P
h
a
s
e

(
d
e
g
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.3: Bode diagram transfer from pitch to HSS angular velocity.
To control this plant rst two integrators are added. Next a lead lter was used to create the
necessary phase lead at the crossover frequencies. In gure 2.3 it can be seen that for low
frequencies the slope of the bode diagram is zero while the phase is 180

. This is because of
the inverse reaction of the angular velocity which was noted in the step responses. To correct for
this the gain of the controller is multiplied by -1. In gure 2.4 the resulting open loop transfer
functions for all four dierent systems are shown. It is clear that for wind speeds near 12 meters
per second the phase and magnitude margins, 16.7

and 10.8 dB respectively, are the smallest


but quickly increase for higher wind speeds. Figure 2.5 shows the sensitivity functions. Here
the sensitivity function for 12 meters per second winds has the highest peak at 11.2 dB, this is
expected because of the low robustness margins. Since all of the systems are open loop stable
the simplied Nyquist stability criterium can be used to assess the stability of the controlled
systems. This is why the Nyquist plots in gure 2.6 only show the positive frequencies. The
Nyquist diagram shows that all of the systems should be stable. This controller will determine a
collective pitch angle for all three blades. To assess the performance of this controller simulations
will be performed. The results of the simulation will be discussed in section 2.5.
13
300
200
100
0
100
200
300
M
a
g
n
i
t
u
d
e

(
d
B
)
10
5
10
0
900
720
540
360
180
0
P
h
a
s
e

(
d
e
g
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.4: Open loop transfer functions for controlled systems.
10
2
10
1
40
30
20
10
0
10
20
M
a
g
n
i
t
u
d
e

(
d
B
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.5: Sensitivity functions for controlled systems.
14
6 5 4 3 2 1 0 1
6
5
4
3
2
1
0
1


Nyquist Diagram
Real Axis
I
m
a
g
i
n
a
r
y

A
x
i
s
12mps
16mps
20mps
25mps
Figure 2.6: Nyquist diagrams for controlled systems.
2.2 Tower acceleration
Now that a controller has been designed to control the angular velocity additional controllers can
be added in order to reduce the loadings on the blades and tower of the wind turbine. The rst
additional controller for this purpose will be a controller to reduce the tower for-aft acceleration.
For the design of this controller the same strategy is used as for the angular velocity controller.
Starting again with the open loop bode diagram from gure 2.7.
100
50
0
50
M
a
g
n
i
t
u
d
e

(
d
B
)
10
6
10
4
10
2
10
0
10
2
900
720
540
360
180
P
h
a
s
e

(
d
e
g
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.7: Open loop transfer functions.
15
To control this system the a weak and a normal integrator is added. Next a lead lter around
a frequency of 1 10
4
Hz. Last a second order low-pass lter with the poles at 0.032 Hz. This
results in the bode diagram in gure 2.8. With this controller the minimal stability margins,
for a wind speed of 20 mps, are a gain margin of 9.15 dB, a phase margin of 15.3 degrees and
a modulus margin of 0.25. For dierent wind speeds these margins increase. In gures 2.9
and 2.10 the sensitivity plots and Nyquist diagrams for the controlled systems are shown.
400
200
0
200
M
a
g
n
i
t
u
d
e

(
d
B
)
10
8
10
6
10
4
10
2
10
0
10
2
540
450
360
270
180
90
0
P
h
a
s
e

(
d
e
g
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.8: Open loop transfer functions for controlled systems.
10
6
10
4
10
2
40
30
20
10
0
10
M
a
g
n
i
t
u
d
e

(
d
B
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.9: Sensitivity functions for controlled systems.
16
5 4 3 2 1 0 1
6
5
4
3
2
1
0
1


Nyquist Diagram
Real Axis
I
m
a
g
i
n
a
r
y

A
x
i
s
12mps
16mps
20mps
25mps
Figure 2.10: Nyquist diagrams for controlled systems.
2.3 Individual pitch control
For the purpose of load reduction individual pitch control can be used, as described in the four
papers [3]-[6]. In [5] the use of a PI-based controller is discussed to reduce the apwise loadings
on the blades. For this controller the measured apwise bending moments, at the root of the
blades, are transformed using the inverse of the transformation matrix as dened in equation 2.2.
This signal is than used for the controller. For this loop a simple PI controller is sucient. The
tuning of this PI controller is done using time simulations and tuning the P-action and I-action
to reduce bending moments as much as possible. After the controller the resulting two control
signals have to be transformed again using the transformation matrix dened in equation 2.3.
The total control scheme is depicted in gure 2.11. Results for the control without individual
pitch control (IPC) and with are also discussed in the simulation section 2.5.
T =
2
3

sin(
1
) sin(
2
) sin(
3
)
cos(
1
) cos(
2
) cos(
3
)

(2.2)
T =

sin(
1
) cos(
1
)
sin(
2
) cos(
2
)
sin(
3
) cos(
3
)

(2.3)
Blade root bending
moments
Inverse Coleman
transformations
PI
PI
Coleman trans-
formations
Individual pitch
angles
Figure 2.11: Control setup for individual pitch control.
17
2.4 Power control
Until now only the pitch was used to control the rotational speed. But if also the torque is used
as a control input this could help to further reduce the loads on the structure. When both pitch
and torque are used as control inputs the power is now chosen as the controlled output. The
power is simply the product between rotational speed and torque.
Again to determine transfer function linearizations have to be performed. This will result in a
MISO (Multi Input Single Output) system. The bode diagrams are shown in gure 2.12.
30
25
20
15
10
5
M
a
g
n
i
t
u
d
e

(
d
B
)
10
5
10
0
135
90
45
0
45
P
h
a
s
e

(
d
e
g
)
Bode Diagram
Frequency (Hz)
20
40
60
80
100
M
a
g
n
i
t
u
d
e

(
d
B
)
10
5
10
0
900
720
540
360
180
0
P
h
a
s
e

(
d
e
g
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.12: Bode diagram for MISO system, torque to power (left), pitch to power (right).
An extra consideration for the tuning of this controller is that preferably the torque has to be
kept constant. To ensure this requirement the gain in the torque controller will be kept low.
The torque controller consists of a double integrator and a lead lter. In gures 2.13 until 2.15
the bode diagrams, sensitivity plots and Nyquist diagrams for the controlled system are shown.
200
100
0
100
200
M
a
g
n
i
t
u
d
e

(
d
B
)
10
6
10
4
10
2
10
0
10
2
315
270
225
180
135
90
P
h
a
s
e

(
d
e
g
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.13: Open loop transfer functions for controlled systems.
18
10
4
10
3
10
2
10
1
50
40
30
20
10
0
M
a
g
n
i
t
u
d
e

(
d
B
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.14: Sensitivity functions for controlled systems.
6 5 4 3 2 1 0 1
5
4
3
2
1
0
1


Nyquist Diagram
Real Axis
I
m
a
g
i
n
a
r
y

A
x
i
s
12mps
16mps
20mps
25mps
Figure 2.15: Nyquist diagrams for controlled systems.
It can be seen that with this controller the stability and robustness margins for this system are
about the same for all the wind speeds. The controller for control of the pitch also contains a
double integrator and lead lter. But as can be seen in the open loop bode diagram in gure 2.12
the phase is 180 degrees at low frequencies where the slope is 0. To correct this a negative gain
is used in the controller. The resulting bode diagram, sensitivity plots and Nyquist diagrams
for the controlled system are shown in gure 2.16 until 2.18.
19
200
0
200
400
M
a
g
n
i
t
u
d
e

(
d
B
)
10
5
10
0
900
720
540
360
180
0
P
h
a
s
e

(
d
e
g
)


Bode Diagram
Frequency (Hz)
12mps
16mps
20mps
25mps
Figure 2.16: Open loop transfer functions for controlled systems.
Bode Diagram
Frequency (Hz)
10
4
10
3
10
2
60
50
40
30
20
10
0
10
M
a
g
n
i
t
u
d
e

(
d
B
)


12mps
16mps
20mps
25mps
Figure 2.17: Sensitivity functions for controlled systems.
20
Nyquist Diagram
Real Axis
I
m
a
g
i
n
a
r
y

A
x
i
s
6 5 4 3 2 1 0 1
4
3.5
3
2.5
2
1.5
1
0.5
0
0.5


12mps
16mps
20mps
25mps
Figure 2.18: Nyquist diagrams for controlled systems.
In this case the stability and robustness margins are smallest for wind speeds of 12 mps. The
magnitude margin in this case is 42 dB, the phase margin is 34.3 degrees and the maximum of
the sensitivity function is 5.44. So all of these margins are within the standard guidelines.
Now that all of these dierent controller are designed they can be used in a simulation. Af-
ter the simulations the results can be compared.
2.5 Simulations and results
In this section simulations will be performed using the various PID controllers from the previous
sections. For these and later simulations it is important that simulation conditions are identical
for a objective comparison between dierent controllers and control strategies. So before actually
stating the simulations and discussing the results rst the wind input will be discussed. Secondly
the sensor noise which is used in all of the simulations is presented.
2.5.1 Wind input
For the wind used in the simulations a two dimensional prole in the wind is desired. This means
that wind speeds will only vary according to the height and not in the horizontal direction. Now
a wind input le for Aerodyn will be computed. In Aerodyn there are various parameters that
determine the wind prole. A full description can be found in the Aerodyn manual [8]. Here
only the parameters that are used will be discussed.
The prole of the wind in these simulations will be a combination between linear vertical shear
and vertical power law shear. These components are described by the functions 2.4 and 2.5.
V
lin
=

h H
2R

(2.4)
V
power
=

h
H

V
mean
(2.5)
21
Where is the linear vertical shear parameter, the vertical power law parameter, H the rotor
hub height, R the rotor radius (blade length), V
mean
the mean hub height wind speed and h the
height. This height only dened over the rotorplane. So from the hub height minus the blade
length till the hub height plus the blade length. By choosing = 0.14 and varying the this
results in the wind proles shown in the left side of gure 2.19. It can be seen that all of these
proles cross at the hub height. To get the required proles a correction is added to the mean
wind speed. The equation for this correction velocity is in equation 2.6. Now the proles are
like the ones depicted on the right of gure 2.19.
10 12 14 16 18 20
40
50
60
70
80
90
100
110
120
Wind speed [m/s]
H
e
i
g
h
t

[
m
]
10 15 20 25
40
50
60
70
80
90
100
110
120
Wind speed [m/s]
H
e
i
g
h
t

[
m
]
Figure 2.19: Wind proles, (left) uncorrected, (right) corrected.
V
corr
= 10 (2.6)
Now that the wind prole is as desired it needs to be put into a format suitable for simulation.
This would mean adding noise, since the wind basically is a disturbance on the system. But
since wind is not white noise the noise needs to be colored. This is done by ltering white noise
with a second order lter. This lter is based on a wind turbulence model [9]. The transfer
function for this lter is in equation 2.7.
H
f
(s) = K
v
sT
v
a
1
+ 1
(sT
v
+ 1)(sT
v
a
2
+ 1)
(2.7)
K
v
=

2T
v
(1 a
2
2
)

a
2
1
a
2
2
a
2
+ 1 a
2
1

1
(2.8)
Where a
1
= 0.4 and a
2
= 0.25. The ltered noise is now used to determine the linear vertical
shear parameter at each simulation step. After this the mean wind speed is adjusted using the
correction velocity is using equation 2.6. This data can now be written into a simple data le
and be used in a simulation.
2.5.2 Sensor noise
For an accurate model description sensor noise is added to the measured outputs. The amount
of noise for each of these measurements is based on [10]. The noise added is Gaussian white
noise with a zero mean and a standard deviation as dened in table 2.1.
22
Measured variable Standard deviation
Generator speed 0.0158 rad/s
Generator torque 45 Nm
Pitch rotor angle 0.2

Rotor speed 0.025 rad/s
Tower accelerations 0.5 m/s
2
Table 2.1: Table with standard deviations for sensor noise.
2.5.3 Simulations
Now the simulations can be performed. Each of the simulations will be 600 seconds long. In
total four dierent control setups will be compared. First only the PID controller for the angular
velocity of the rotor. Second this PID controller but now together with the controller for the
tower acceleration and the individual pitch controllers. Third is the MISO controller. And
nally the MISO with again the controller for the tower acceleration and the individual pitch
controllers. Comparison will be based on two main criteria. In the above rated wind conditions
considered the rotational speed of the rotor should be kept constant. The generator used in these
simulations should have a rotational speed of 20.48 rpm. Next to this the controller should try
to reduce loads on both the tower and the blades. For this the root mean square values of the
tower accelerations in the two directions and the average and standard deviation of the bending
moments in the two directions are compared. Also the frequency content of the tower for-aft,
edgewise bending moment and apwise bending moment are compared.
In gure 2.20 the angular velocity of the rotor over time is shown for the four dierent control
setups. From this gure it can be concluded that all of the controllers are able to keep the
angular velocity around the desired value of 20.48 rpm. So if any of the controllers are able to
reduce loadings this should not have a signicant eect on the generated power.
0 100 200 300 400 500 600
19
19.5
20
20.5
21
21.5
22
22.5
23
23.5
24
Time [s]
R
o
t
o
r

a
n
g
u
l
a
r

v
e
l
o
c
i
t
y

[
r
p
m
]


HSS PID
HSS PID+IPC+TA
MISO
MISO+IPC+TA
Figure 2.20: Angular velocity over time for dierent control setups.
As stated above also the frequency content of various signals can be used as basis for the com-
parison. In gure 2.21 the frequency content of three signals is shown. First the tower for-aft
acceleration, second the edgewise bending moment and third the apwise bending moment.
What can be noted in these gures is that, instead of using normal frequencies. periodic compo-
nents are used. These periodic components are based on an N-times per revolution component
23
HSS PID HSS PID + IPC + TA MISO MISO + PID + IPC
RMS Acc. for-aft [m/s
2
] 1e
3
1e
3
1e
3
1e
3
RMS Acc. side-side [m/s
2
] 2e
4
3e
4
3e
4
4e
4
RMS Pitch rate [

/s] 1e
5
0.03 8e
6
0.03
Avg. edgewise bending moment 43 39 43 40
Avg. apwise bending moment 657 664 623 630
Std edgewise bending moment 45 43 45 44
Std apwise bending moment 185 121 183 120
Table 2.2: Results simulations PID control setups.
in the signal. For instance the tower shadow will have an eect on each of the three blades in
a single revolution. This would then show up as a 3p component in the tower acceleration but
as a 1p component on the blade bending moments. For the results the 1p components of the
bending moments are considered and the 3p component of the tower acceleration.
0.9 0.95 1 1.05 1.1
0
10
20
PSD edgewise bending moment
Periodic component
|
Y
(
f
)
|
0.9 0.95 1 1.05 1.1
0
50
PSD flapwise bending moment
Periodic component
|
Y
(
f
)
|


2.9 2.95 3 3.05 3.1
0
0.05
PSD ForAft Tower Acceleration
Periodic component
|
Y
(
f
)
|
HSS PID
HSS PID+IPC+TA
MISO
MISO+IPC+TA
Figure 2.21: Frequency content for dierent signals.
Clearly this result shows that the addition of the tower acceleration and IPC controllers has the
desired eect. In all three signals a reduction in the content of the periodic component can be
seen.
Finally the RMS values, averages and standard deviation of some signals are in table 2.2. In
this table also the RMS values of the pitch rate are added. These show that the more extensive
control schemes do increase the pitch rate. For the rest of the values it can be concluded that the
control setups with the IPC and tower acceleration controller are able to decrease the averages
and standard deviations on the bending moments in both directions of the blades. But the
standard deviation of the side-side tower acceleration is slightly increased as a consequence.
From these results it shows that the designed controllers perform as expected. They decrease
most of the loadings on the tower and blades. But the blade pitch rate is increased as a
consequence. In the next part of the report the focus will be on the second control setup.
24
This will involve a combination of a Kalman lter, for state estimation, together with a LQG
controller. After the design of these elements simulation results will be compared to the results
from this part.
25
26
Chapter 3
Kalman lter
In this chapter the design and performance of a Kalman lter for the wind turbine model will
be discussed. The Kalman lter will be used to estimate the states of the model to be able to
use them with a LQG controller.
3.1 Linearizations
The model used for the Kalman lter will be based on linearizations from FAST. These lin-
earizations will be computed for dierent wind speeds and dierent azimuth positions. In the
simulations a linear combination of the systems matrices resulting from these linearizations will
be used to get a model for the Kalman lter. Since this is a non-linear system it will need to
be veried that there are enough linearizations to accurately describe the system dynamics. If
this is not the case the Kalman lter will not be able to estimate the states properly.
In total there will be linearizations for 11 dierent wind speeds ranging from 12 until 25 mps.
For each wind speed there will be 24 linearizations at equally spaced azimuth positions. Using
24 here will ensure that each blade is linearized at exactly the same position and therefor should
lead to repeating models for each one third azimuth rotation. After the computations the results
can be examined to assess the use of linear combinations for the determination of the system
matrices. In gure 3.1 the system matrix A1 is shown in a mesh plot, this is for a azimuth
angle of 0

. Figure 3.2 shows also the A2 matrix but now for a azimuth angle of 15

. Now the
linear combination of 0.5(A1 + A2) is computed, the result is shown in gure 3.3 Comparing
these three gure is can be concluded that the mesh plots for the three dierent matrices are
very similar. This suggest that using a linear combination as an estimate of the system matrix
is allowed. Similar results can be shown for dierent azimuth positions, wind speeds as well as
the other matrices B, C and D.
27
0
10
20
30
0
10
20
30
6000
4000
2000
0
2000
4000
6000
Figure 3.1: Mesh plot for A matrix at 0

azimuth position.
0
10
20
30
0
10
20
30
6000
4000
2000
0
2000
4000
6000
Figure 3.2: Mesh plot for A matrix at 15

azimuth position.
28
0
10
20
30
0
10
20
30
6000
4000
2000
0
2000
4000
6000
Figure 3.3: Mesh plot for linear combination for A matrix at 7.5

azimuth position.
3.2 Appending states
For a linear plant in equation 3.1 a Kalman lter is able to estimate the states of the system by
the inputs and measurements. In this equation w and v are the process and measurement noise
respectively. A requirement for these signals is that they are white noise with a zero mean. As
discussed in section 2.5.2 the sensor noise is white noise. But for the signal w the mean hub
height wind speed and the linear vertical shear parameter will have to be used. But as discussed
in section 2.5.1 these parameters are colored noise. And a second order lter was used to color
white noise. So the Kalman lter will have to be adjusted to correct for this. This is simply
done by addition of states to the system. These state can be computed from the second order
lter that was used to color the white noise. By converting the transfer function to a second
order state space system in equation 3.2.
x = Ax +Bu +Gw
y = Cx +Du +v (3.1)
x
f
= A
f
x
f
+B
f
m
y
f
= C
f
x
f
+D
f
m (3.2)
Now by simply appending the computed states x
f
to the state vector x the x
A
from equation 3.3
is used in the Kalman lter. The system dynamics can now be written into the form of equa-
tion 3.4, where m is now white noise for the two signals for mean wind speed and linear vertical
shear.
x
A
= [x x
f
]
T
(3.3)
x
A
=

A G[10C
f
C
f
]
T
0 A
f

x
A
+

B
0

u +

0 0
B
f
B
f

m
y = Cx +Du +v (3.4)
29
The factor 10C
f
in the A
A
matrix comes from the correction velocity used in the computation
of the wind input. This correction velocity was ten times the linear vertical shear parameter.
With these appended states the Kalman lter can now be computed. These computations will
result in a gain L which is used to produce state estimate

x using equation 3.5 The gain L is
determined by solving an algebraic Riccati equation and will ensure stable estimations, i.e. the
estimated state should converge to the real states over time.

x = A x +Bu +L(y C x Du) (3.5)


3.3 Performance
Now that a Kalman lter is designed the quality of the estimations can be assessed in a simu-
lation. In this simulation the measured outputs are compared with the estimated outputs. As
mentioned in the previous section linear combination as a function of both mean wind speed and
azimuth position is used to determine the system matrices which are then used for the estima-
tions. Initial simulations suggested that the Kalman lter did not function. In these simulations
the mean wind speed is a very noisy signal. This also means that the dynamics in the Kalman
lter change rapidly. Because the mean wind speed is used in the computation of a linear combi-
nation to determine the model at each time instance. When adding a weighted moving average
to the mean wind speed signal the changes are less rapidly. When using a weighting over just
0.12 second the results improve signicantly. The results are shown in gures 3.4 and 3.5.
From these results it can be concluded that the Kalman lter is not able to correctly estimate
all of the outputs. For some of the outputs like the bending moments and generator torque
the estimation is suciently close to the measured value. But for the low and high speed shaft
angular velocity a steady state oset is visible. This could be because the linearization algorithm
of FAST only searches for a steady state between certain bounds. When the dierences in the
states in one rotation are within these bounds the algorithm assumes steady state has been
reach. Furthermore as discussed in the PID control design the linearized model errors can even
cause a stable system to seem unstable. As for the tower for-aft acceleration the estimations
dier between good and bad. This could be because of the amount of linearization point which
are used. The total estimation is build up from two distinct parts. First there is the nominal
trajectory which is computed in the linearization. And second there is the estimator part. If the
steady state in the simulation is about the same as in the linearization the estimation part should
be relatively small. This would suggest that the biggest part of the estimation comes from the
nominal trajectory. But what can be noted in the gure for the tower for-aft acceleration is that
in one rotation (about 3 seconds), there are a total of nine oscillations. These oscillations have
to be described for the biggest part by the nominal trajectory. But in one rotation the nominal
trajectory only consists of 24 points. So 24 point are available to describe nine oscillations. This
could also lead to incorrect estimations. Unfortunately because of the amount of linearizations
it is not possible to correct each of these models like in the PID case. The amount of points
of the nominal trajectory could be solved by just performing new linearizations with a more
detailed description of this nominal trajectory.
Still the Kalman lter works suciently and can be used in union with a LQG controller. In
the next chapter the design of this controller will be discussed as well as the simulation of the
system with the Kalman lter and LQG controller.
30
300 301 302 303 304 305 306
1
0.5
0
0.5
1
Time [s]
T
o
w
e
r

f
o
r

a
f
t

a
c
c
e
l
e
r
a
t
i
o
n

[
m
/
s
2
]


Measured output
Estimated output
300 301 302 303 304 305 306
0.25
0.2
0.15
0.1
0.05
0
0.05
0.1
0.15
0.2
0.25
Time [s]
T
o
w
e
r

s
i
d
e

s
i
d
e

a
c
c
e
l
e
r
a
t
i
o
n

[
m
/
s
2
]


Measured output
Estimated output
300 301 302 303 304 305 306
20.1
20.2
20.3
20.4
20.5
20.6
20.7
20.8
20.9
21
Time [s]
L
o
w

s
p
e
e
d

s
h
a
f
t

a
n
g
u
l
a
r

v
e
l
o
c
i
t
y

[
r
p
m
]


Measured output
Estimated output
300 301 302 303 304 305 306
1770
1780
1790
1800
1810
1820
1830
1840
Time [s]
H
i
g
h

s
p
e
e
d

s
h
a
f
t

a
n
g
u
l
a
r

v
e
l
o
c
i
t
y

[
r
p
m
]


Measured output
Estimated output
Figure 3.4: Measured and estimated outputs.
31
300 301 302 303 304 305 306
50
0
50
100
150
200
250
300
350
400
Time [s]
A
z
i
m
u
t
h

a
n
g
l
e

[

]


Measured output
Estimated output
300 301 302 303 304 305 306
6.5
7
7.5
8
8.5
9
9.5
10
Time [s]
G
e
n
e
r
a
t
o
r

t
o
r
q
u
e


Measured output
Estimated output
300 301 302 303 304 305 306
50
0
50
100
150
200
Time [s]
E
d
g
e
w
i
s
e

b
e
n
d
i
n
g

m
o
m
e
n
t

[
k
N
m
]


Measured output
Estimated output
300 301 302 303 304 305 306
200
400
600
800
1000
1200
Time [s]
F
l
a
p
w
i
s
e

b
e
n
d
i
n
g

m
o
m
e
n
t

[
k
N
m
]


Measured output
Estimated output
Figure 3.5: Measured and estimated outputs.
32
Chapter 4
LQG controller
In this chapter the design and performance of a linear quadratic gain controller is discussed.
This controller will use the state estimations from the Kalman lter to control each of the blades
individually. For this controller the choice is made to use an output weighting instead of an
state weighting. With this the quadratic cost function that needs to be minimized changes to
equation 4.1. The choice to use an output weighting is because of the fact that all rotating states
and outputs are transformed using the Coleman transformations. After these transformations
the states change and it is much more dicult to nd a proper weighing matrix Q. But for the
outputs a simple simulation where the outputs are transformed is enough to give a good view of
how the weighting have to be adjusted. The same adjustment has to be made to the weighting
on the pitch inputs. Since these inputs are also in a rotating frame and are transformed. Next
to tuning the weights on in- and outputs the reference that the LQG-controller should try to
follow can be tuned for better performance. After this tuning the controller can be used in a
simulation. In the next section results are discussed and compared with the results from the
PID controllers.
J =


0
(y
T
Qy +u
T
Ru)dt (4.1)
4.1 Results
As with the results from the PID the rst priority is the angular velocity of the rotor. In
gure 4.1 the angular velocity is shown over time. It shows that the LQG controller is faster to
get to desired the steady state value of 20.48 rpm.
Now the frequency content can be compared. This is done in gure 4.2. What can be noted
is that here no improvement is shown for the LQG-controller compared to the simplest PID
controller for just the angular velocity of the rotor. Also when looking at the data in table 4.1
the same conclusion can be made. So based on this data it seem like the LQG-controller is not
able to improve on any the load reduction compared to PID controllers. When looking at the
tower accelerations in both for-aft and side-side direction there is an improvement to be seen.
In gures 4.3 and 4.4 it can be seen that the for-aft acceleration is reduced by 0.5 m/s
2
and the
side-side acceleration by about 0.1 m/s
2
. This while not increasing the bending moments which
are plotted in gures 4.5 and 4.6.
33
0 100 200 300 400 500 600
19.5
20
20.5
21
21.5
22
22.5
23
Time [s]
A
n
g
u
l
a
r

v
e
l
o
c
i
t
y

L
S
S

[
r
p
m
]


LQG
MISO + IPC + TA
Figure 4.1: Angular velocity over time for dierent control setups.
0.9 0.95 1 1.05 1.1
0
20
40
PSD edgewise bending moment
Periodic component
|
Y
(
f
)
|
0.9 0.95 1 1.05 1.1
0
50
100
PSD flapwise bending moment
Periodic component
|
Y
(
f
)
|


2.9 2.95 3 3.05 3.1
0
0.05
PSD ForAft Tower Acceleration
Periodic component
|
Y
(
f
)
|
LQG
HSS PID
MISO+IPC+TA
Figure 4.2: Periodic frequency content for dierent control setups.
34
0 100 200 300 400 500 600
1.5
1
0.5
0
0.5
1
1.5


MISO + IPC + TA
LQG
Figure 4.3: Tower for-aft acceleration for PID controller and LQG controller.
0 100 200 300 400 500 600
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4


MISO + IPC + TA
LQG
Figure 4.4: Tower side-side acceleration for PID controller and LQG controller.
35
HSS PID LQG
RMS Acc. for-aft [m/s
2
] 1e
3
1e
3
RMS Acc. side-side [m/s
2
] 2e
4
2e
4
RMS Pitch rate [

/s] 1e
5
0.057
Avg. edgewise bending moment 43 43
Avg. apwise bending moment 657 656
Std edgewise bending moment 45 47
Std apwise bending moment 185 193
Table 4.1: Results simulation LQG control setup.
200 220 240 260 280 300 320 340 360
150
100
50
0
50
100
150
200


MISO + IPC + TA
LQG
Figure 4.5: Edgewise bending moments for PID controller and LQG controller.
200 220 240 260 280 300 320 340 360
200
0
200
400
600
800
1000
1200
1400
Figure 4.6: Flapwise bending moments for PID controller and LQG controller.
36
As was concluded in the section about the Kalman lter design there is a lot of improvement
possible in the state estimations. Because the performance of a LQG-controller is heavily depen-
dent on the quality of the state estimates the best way to improve is to get a better estimator.
Since the FAST model is now a black box no non-linear model information is available. When
a non-linear model would be available an extended Kalman lter could be used for better esti-
mations.
37
38
Chapter 5
Conclusions and recommendations
In this report a comparison was made between PID controllers and a LQG controller for the
control of a 1.5 MW wind turbine. The control was focused on keeping a steady angular velocity
while reducing loads on tower and blades. From the results several conclusions can be made. For
the PID controllers it was shown that a single PID controller could control the non-linear model
for all considered wind speeds. Gain scheduling could potentially improve the performance of
these controllers further and is dependent on the specications for the controlled system. When
comparing results for the dierent PID control setups it was shown that the most elaborate con-
troller was able to achieve the biggest load reduction on the system. This controller used both
individual pitches and the generator torque to control the power while reducing tower for-aft
acceleration and the apwise bending moments on the blades.
Next a Kalman lter was designed. The model in the lter was based on linearizations of
the system for dierent winds speeds and azimuth angles. It was shown that is was allowed to
take linear combinations of the linearizations for the determination of the system dynamics in
the simulations. By using a weighted moving average on the wind speed signal the Kalman lter
was able to estimate the outputs of the system. The performance of the lter is dependent of
the linearizations which, as was already shown in the PID controller part, can contain errors. A
big improvement for the performance of the Kalman lter can be potentially made by using a
non-linear model in the Kalman lter and using an extended Kalman lter.
Finally a LQG controller was designed to use the state estimates from the Kalman lter to
control the system. Comparing results with the PID controllers it could be seen that the per-
formance of the LQG controller was, for the biggest part, comparable with the high speed shaft
angular velocity PID controller. The LQG controller was able to reduce tower acceleration by
a considerable amount. Because the performance of this controller is dependent on the esti-
mations from the Kalman lter improving the Kalman lter could also the performance of the
LQG controller. When comparing the tuning of the LQG controller with the PID controllers
it is much easier to adjust the LQG controller and the weighting lters make it easier to see
the connection between dierent signals. With this the potential of the use of a LQG controller
shows.
39
40
Appendix A
Coleman transformations
Before the linearization can be used for structural analysis and controller design a coordinate
transformation has to be performed. In this section the Coleman transformations are explained.
The Coleman transformation can be used to transform the variables in a rotating frame into
variables in the xed, non-rotating frame. This transformation process is described in [2]. Below
the process will be described briey for a three blade wind turbine.
A transfer matrix

t is dened in equation A.1. This matrix describes the relation between three
DOFs in a rotating and a non-rotating frame. To transform three DOFs (q
rot
1
,q
rot
2
,q
rot
3
) in a
rotating frame to a non-rotating frame (q
nr
1
,q
nr
2
,q
nr
3
) the transformation in equation A.2 can be
used. It can be noticed that the transformation is dependent on the azimuth angles (
1
,
2
,
3
)
of the individual blades.

t =

1 cos(
1
) sin(
1
)
1 cos(
2
) sin(
2
)
1 cos(
3
) sin(
3
)

(A.1)

q
nr
1
q
nr
2
q
nr
1

=

t
1

q
rot
1
q
rot
2
q
rot
3

(A.2)
The transformation of the systems matrices, determined by the linearization, is dependent on
the number of non-rotating and rotating variables. For the setup chosen the following division,
given in the table below, can be made between non-rotating and rotating variables.
Non-rotating Rotating
Generator (angle) First apwise blade mode
First fore-aft tower bending-mode Second apwise blade mode
Second fore-aft tower bending-mode First edgewise blade mode
First side-side tower bending-mode Blade root apwise bending moments
Second side-side tower bending-mode Blade root edgewise bending moments
Individual pitch of the blades
Torque
Individual pitch angles
Tower for-aft and side-side acceleration
Low speed shaft angle and speed
High speed shaft speed
Yaw angle
Transformation for the system matrices, both for the rst and second order model, are now
given in equation A.3.
41
M
NR
= MT
1
C
NR
= 2MR
2
+CT
1
K
NR
=
2
MT
3
+

MT
2
+ CT
2
+KT
1
F
NR
= FT
lc
F
d NR
= F
d
C
v NR
= T
1
lo
C
v
T
1
C
d NR
= T
1
lo
(C
v
T
2
+C
d
T
1
)
D
NR
= T
1
lo
DT
lc
(A.3)
D
d NR
= T
1
lo
D
d
A
NR
=

T
1
1
0
0 T
1
1

A

T
1
0
T
2
T
1

T
2
0

2
T
3
+

T
2
2T
2

B
NR
=

T
1
1
0
0 T
1
1

BT
lc
B
d NR
= B
d
C
NR
= T
1
lo
C

T
1
0
T
2
T
1

D
NR
= T
1
lo
DT
lc
D
d NR
= T
1
lo
D
d
Where and

are the angular velocity and acceleration of the rotor shaft. The transformations
matrices: T
1
, T
2
, T
3
, T
lo
and T
lc
are dened in equation A.4.
T
1
=

I
nFnF

T
2
=

0
nFnF

t
2

t
2

t
2

T
3
=

0
nFnF

t
3

t
3

t
3

(A.4)
T
lo
=

I
FoFo

T
lc
=

I
FcFc

Here nF is the number of xed DOF, Fo is the number of xed outputs and Fc is the number
of xed control inputs. The two transformations matrices

t
2
and

t
3
are given in equations A.5
and A.6.
42

t
2
=

0 sin(
1
) cos(
1
)
0 sin(
2
) cos(
2
)
0 sin(
3
) cos(
3
)

(A.5)

t
3
=

0 cos(
1
) sin(
1
)
0 cos(
2
) sin(
2
)
0 cos(
3
) sin(
3
)

(A.6)
43
44
Bibliography
[1] NWTC Design Codes (FAST by Jason Jonkman). http://wind.nrel.gov/designcodes/simulators/fast/.
Last modied 12-August-2005; accessed 12-August-2005.
[2] Bir G., Multiblade Coordinate Transformation and Its Application to Wind Turbine Analysis,
NREL, January 2008
[3] Bossanyi E.A., The Design of Closed Loop Controllers fo Wind Turbines, Garrad Hassan
and Partners Ltd, St Vincents Works, Bristol, UK, Wind Energy 2000; 3(143-163)
[4] Bossanyi E.A., Wind Turbine Control for Load Reduction, Garrad Hassan and Partners Ltd,
St Vincents Works, Bristol, UK, Wind Energy 2003; 6(229-244)
[5] Bossanyi E.A., Individual Blade Pitch Control for Load Reduction, Garrad Hassan and Part-
ners Ltd, St Vincents Works, Bristol, UK, Wind Energy 2003; 6(119-128)
[6] Bossanyi E.A., Further Load Reductions with Individual Pitch Control, Garrad Hassan and
Partners Ltd, St Vincents Works, Bristol, UK, Wind Energy 2005; 8(481-485)
[7] Stol K.A., Dynamics Modeling and Periodic Control of Horizontal-axis Wind Turbines, Uni-
versity of Colorado, December 2001
[8] Laino D.J., Hansen A.C., USERS GUIDE to the Wind Turbine Aerodynamics Computer
Software AeroDyn,Windward Engineering, LC, Salt Lake City, AeroDyn 12.50, 24 December
2002
[9] Fernando D. Bianchi, Ricardo J. Mantz and Hernn De Battista, Wind Turbine Control
Systems, Springer London 2007; 3(29-48)
[10] Thomas Esbensen, Christoer Sloth, Fault Diagnosis and Fault-Tolerant Control of Wind
Turbines, Aalborg University, Department of Electronic Systems, Section for Automation
and Control, June 2009
45

You might also like