Design of State-Space-Based

Control Algorithms for Wind

Alan D. Wright
National Renewable Energy Laboratory
1617 Cole Blvd.
Golden, CO 80401
e-mail: Alan – wright@nrel.gov
Mark J. Balas
AIAA Fellow
Department of Aerospace Engineering Science
University of Colorado at Boulder
Boulder, CO 30309-0429
e-mail: Mark.balas@colorado.edu
Turbine Speed Regulation
Control can improve the performance of wind turbines by enhancing energy capture and
reducing dynamic loads. At the National Renewable Energy Laboratory, we are beginning
to design control algorithms for regulation of turbine speed and power using state-space
control designs. In this paper, we describe the design of such a control algorithm for
regulation of rotor speed in full-load operation (Region 3) for a two-bladed wind turbine.
We base our control design on simple linear models of a turbine, which contain rotor and
generator rotation, drive train torsion, rotor flap (first mode only), and tower fore-aft
degrees of freedom (DOFs). Wind-speed fluctuations are accounted for using Disturbance
Accommodating Control (DAC). We show the capability of these control schemes to
stabilize the modeled turbine modes via pole placement, while using state estimation to
reduce the number of turbine measurements that are needed for these algorithms. These
controllers are incorporated into a simulation code and simulated for various conditions.
Finally, conclusions to this work and future studies are outlined.
关DOI: 10.1115/1.1621673兴

Introduction several states of the turbine requiring numerous turbine measure-

One of the main goals of wind turbine control is to increase
power production and reduce loads using a minimum number of
control inputs and required turbine measurements. Often, controls
can be designed to simultaneously satisfy more than one objec-
tive, i.e., regulate power and reduce loads. In the 1970s and
1980s, classical control design methods, such as proportional in-
tegral 共PI兲, were used to design controllers to regulate power
while also adding damping to the first drive train torsional mode
of the turbine 关1兴. In Barton 关2兴, a power-system stabilizer was
included to add damping to the drive train mode.
In these various applications, the use of classical controls to
address more than one control objective is not straightforward.
Often, multiple control loops must be used, which adds complex-
ity to the control design. It is difficult to properly address the
control structure interaction issues using classical-control meth-
ods, because the controller uses only a single measured turbine
output as the basis of its control and does not have direct knowl-
edge of the system dynamics of the turbine. Modern control de-
signs using state-space methods more adequately address these
issues since the controller uses a linear model of the turbine to
determine system states.
Most modern control design methods use a full-state feedback
law, in which a linear combination of the turbine states are fed
back to the turbine. The gains in the full-state feedback law are
calculated so that stability of the closed-loop system is enhanced.
For wind turbines, this means that the controllers can be designed
to meet simultaneous control objectives, such as maximizing
power or regulating speed and enhancing damping in several flex-
ible turbine modes. In pole placement control design techniques,
the gains in the full-state feedback law are calculated to place the
plant poles at desired locations in the complex plane to enhance
closed-loop stability. In other control design methods, the full-
state feedback law gains are calculated by minimizing a function
of the states and control inputs, such as in Linear Quadratic Regu-
lator 共LQR兲 design.
In full-state feedback control, the controller must have access to
Contributed by the Solar Energy Division of THE AMERICAN SOCIETY OF ME-
CHANICAL ENGINEERS for publication in the ASME JOURNAL OF SOLAR ENERGY
ENGINEERING. Manuscript received by the ASME Solar Energy Division July 26,
2002; final revision, June 26, 2003. Associate Editor: D. Berg.
ments. To make actual implementation of such controllers more
practical, state estimation is used to estimate states, thereby reduc-
ing the required number of measurements.
Work has been done in Europe using state-space methods for
wind turbine control design. Mattson 关3兴 designed a controller for
regulation of power below rated wind speed for a fixed-speed
machine using blade pitch. In this work, rotor rotation, drive train
torsion, and tower fore-aft DOF were modeled for use in the con-
trol system design. Liebst 关4兴 described the use of individual blade
periodic pitch control to reduce the loads on the Mod 0-A turbine
due to tower shadow, wind shear, and gravity. In Ref. 关4兴, only
blade DOFs were modeled in the dynamics using rigid blade/
hinge models to represent the blade flap, lag, and pitch DOF. In
Refs. 关3–5兴 Linear Quadratic Gaussian 共LQG兲 control-design
techniques were used, which is similar to LQR, except that state
estimation is used to reduce the number of turbine measurements.
Pole placement and LQG control design methods do not pro-
duce robust controllers, which means that they can be sensitive to
errors in the turbine model or measurements. A similar approach
共which is also based on a feedback control law兲 is the H⬁ control-
ler, in which uncertainties in the turbine and wind models can be
explicitly taken into account. This method has been applied to
wind turbine control in Ref. 关6兴.
Another model-based control method is Disturbance Accommo-
dating Control 共DAC兲, which explicitly accounts for input distur-
bances. We know that wind turbines must operate in the presence
of a highly turbulent wind environment. Turbulent winds cause
fluctuations in the blade aerodynamic forces and thus influence
the power, torque, and cyclic loads of the machine. These wind
inputs can be considered to be disturbance inputs, and we must
design controls to account for these effects. We need an approach
to counteract or accommodate these disturbances while also per-
mitting full-state feedback and state estimation. An approach to
the reduction or counteraction of persistent disturbances was de-
veloped by Johnson in Ref. 关7兴 for lumped parameter systems and
extended by Balas in Refs. 关8兴, 关9兴 for large-scale and distributed-
parameter systems via model reduction and residual mode filters.
The basic idea of DAC is the augmentation of the usual state-
estimator-based controller to recreate disturbance states via an as-

386 Õ Vol. 125, NOVEMBER 2003 Copyright © 2003 by ASME Transactions of the ASME

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 sumed waveform model; these disturbance states are used as part
of the feedback control to reduce 共‘‘accommodate’’兲 or counteract
any persistent disturbance effects.
Work has been reported in Ref. 关10兴 showing further use of
DAC for wind turbine control design. They developed a linear
model of a turbine using the rigid blade/tower/hinge approach to
model blade and tower flexibility. They developed DAC from a
linear model containing only rotor rotation as the DOF. They then
showed that this DAC adequately controlled a turbine as modeled
in their nonlinear simulator—SymDyn—with just the rotor rota-
tion DOF. This system became unstable when more DOFs were
turned on in SymDyn than were included in the linear model for
control design.
These investigations show that consideration must be given to
unmodeled structural DOF when designing a controller using any
of the modern state-space control design methods. Less-
aggressive control gains could probably be chosen for the lowest-
order controllers, resulting in stable behavior of the complete
wind turbine. However, there will always be components in a
wind turbine that are difficult to model or have uncertain proper-
ties. It is important to begin to assess the importance of these
unmodeled effects in the design of controllers for wind turbines. A
balance must be attained between the controller complexity 共the
controller will be more complex if many turbine DOFs are ac-
counted for in its design兲 and ease of actual implementation. An
appropriate question is: Which turbine DOFs are most critical for
inclusion in the models used for controller design?
In this paper, we will show the design of a control system using
DAC for regulation of turbine speed at full load 共above-rated
power for a variable-speed turbine, i.e., Region 3兲, while also
adding damping to flexible modes of the turbine. The only control
input that we assume in this paper is blade-collective pitch. We
use pole placement to calculate the gains in the full-state feedback
law. We also use state estimation to estimate plant states, which
reduce the number of required turbine measurements. We apply a
constant torque to the generator 共we do not perform any generator
torque control in this paper兲 and use blade pitch to regulate tur-
bine speed. The only measured variables are generator speed and
tower-top deflection 共or acceleration兲.
An objective of this paper is to extend the work performed in
Ref. 关10兴 using linear models and simulation codes, which model
turbine flexibility with an assumed modes approach, rather than
the rigid blade/hinge approach. Kane’s method 关11兴 is used to
develop nonlinear equations of motion for the turbine, as used in
the Fatigue, Aerodynamics, Structures, and Turbulence 共FAST兲
code 关12兴. Linear models are then extracted from FAST for use in
control design. FAST is also used to simulate the nonlinear-
controlled turbine after incorporating the designed control system
into the code for simulation purposes 共Fig. 1兲.
There may be times when we want the controller to add damp-
Fig. 1 Diagram of control system
Journal of Solar Energy Engineering
Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use
ing to several low-damped flexible modes of the turbine. These
modes need to be modeled with more precision than with the rigid
blade/hinge 共or rigid tower/hinge兲 approach. Examples are the ro-
tor first symmetric flap mode, the first drive train torsion mode,
and the tower first fore-aft mode. We also want to establish which
flexible modes are the most likely to become unstable when regu-
lating turbine speed in Region 3 using rotor collective pitch. This
has an impact on the complexity of the linear models used for
control design. Our approach is to begin with the simplest models
possible, then design controls using these simple models and per-
form testing through simulation. Complexity is added to these
linear models only when the need has been firmly established—
when modes neglected in these linear models become unstable
during closed-loop operation. We begin by designing a controller
with just one state 共generator speed兲 and then study the effect of
simulating the controlled turbine with additional DOF switched on
during nonlinear simulation in FAST. If the simulations show un-
stable behavior with additional DOF switched on, then extra states
are added to the linear model to include these modes. We then
show that redesigning the control system from these more com-
plex linear models results in stable closed-loop behavior.
A second goal of this paper is to demonstrate control design
based on a small number of turbine measurements making actual
implementation in a real machine simpler. We begin by using just
generator speed and add measurements only when necessary.
Now, the model linearization is described.
Model Linearization
In order to develop linear models for use in control design, we
developed methods to extract a linear model from FAST using
both a symbolic method and a numerical perturbation method.
The numerical perturbation method allowed us to extract linear
models, which include aerodynamic damping. The symbolic
method was only used as a cross-check to verify results from the
numerical perturbation method while neglecting aerodynamics.
We began by extracting simple models first and then adding com-
plexity in steps.
Our first linear model contained just the generator speed state.
We designed a controller from this model and tested it in FAST.
We then progressed to a model with generator speed, rotor speed,
drive train torsion, rotor first symmetric flap, and tower first fore-
aft bending in steps, designing a controller, and testing it in FAST
at each step. We neglected the effects of gravity, tower shadow,
and wind shear in the control design but accounted for uniform
wind disturbances over the rotor disk. We also neglected the teeter
DOF in the linear models used for control design, because we
were only accounting for uniform wind inputs, which do not ex-
cite this DOF. In this study, we designed controls that accounted
for turbine modes excited by uniform wind disturbances over the
rotor disk such as drive train torsion and rotor symmetric flap.
NOVEMBER 2003, Vol. 125 Õ 387
 This gave us equations of motion with constant coefficients. Fur-
ther information on periodic controls for wind turbines is pre-
sented in Stol and Balas 关13兴.
Linear Models. Each of the linear models that we developed
were expressed in state-space form as:
Mẋគ ⫽Axគ ⫹Buគ ⫹⌫uគ D
yគ ⫽Cxគ
where xគ was the state vector, uគ was the control input, u D was the
disturbance input, yគ was the measured output, A was the state
matrix, M was the mass matrix, B was the control input gain
matrix, ⌫ was the disturbance input gain matrix, and C related the
measured signal yគ to the turbine states. The vectors and matrices
varied in size depending on the number of states in the linear
model.
For the most complex linear model used in this work, states
(1) were included to model the rotor first symmetric flap mode, first
drive train torsional mode, generator speed, and tower first fore-
aft bending mode. This model can be written as:

冤 冥冤 冥 冤 冥冤 冥
1 0 0 0 0 0 0 0 1 0 0 0 0 0
ẋ1 x1
0 M11 M14 0 0 0 M17 ⫺K11 ⫺C11 ⫺C14 0 0 ⫺K17 ⫺C17
ẋ2 x2
0 2M14 Irot 0 0 0 0 ẋ3 0 ⫺2C41 ␥ ⫺Cd ⫺1 Cd ⫺K47 ⫺C47 x3
0 0 0 1 0 0 0 ẋ4 ⫽ 0 0 Kd 0 ⫺Kd 0 0 x4
0 0 0 0 Igen 0 0 ẋ5 0 0 Cd 1 ⫺Cd 0 0 x5
ẋ6 x6
0 0 0 0 0 1 0 0 0 0 0 0 0 1
ẋ7 x7
0 2M71 0 0 0 0 M77 ⫺2K71 ⫺2C71 ⫺C74 0 0 ⫺K77 ⫺C77

冤冥 冤冥
0 0
␨b ␣b
␨ ␣
⫹ 0 ␦␤⫹ 0 ␦w
0 0
0 0
␨t ␣t

冤冥
x1
x2

冋册冋 册
0 0 0 0 1 0 0 x3
y1
⫽ x4 . (2)
y2 0 0 0 0 0 1 0 x5
x6
x7

where, tive pitch. Because rotor collective pitch is the only control input
x1 ⫽ perturbed rotor symmetric flap mode displacement, in this study, the vector uគ has dimension 1⫻1, represented by
x2 ⫽ perturbed rotor symmetric flap mode velocity, perturbations in rotor collective pitch ␦␤.
x3 ⫽ perturbed rotor speed, The dimension of u D is 1⫻1 and is represented by perturba-
x4 ⫽ perturbed drive train torsional spring force, tions in windspeed ␦w. This disturbance is considered to contain
only one component of wind speed, the component normal to and
x5 ⫽ perturbed generator speed,
uniform across the rotor disk. The only nonzero elements in ⌫ are
x6 ⫽ perturbed tower first fore-aft mode displacement, and the second, third, and seventh rows 共for the seven-state model兲.
x7 ⫽ perturbed tower first fore-aft mode velocity. Those elements are the partial derivatives of the blade flap normal
Please note that the mass matrix in Eq. 共2兲 is not symmetric due force, rotor aerodynamic torque, and rotor thrust force with re-
to the selection of the states in this system. For example, x2 , spect to wind speed. These values reflect the influence of uniform
which is associated with rotor first symmetric flap 共which is a wind-speed fluctuations on the rotor symmetric flap mode, rotor
aerodynamic torque, and rotor thrust.
linear combination of blades one and two flap兲 共row two兲, couples
Models with fewer states can be obtained from this equation by
to the tower fore-aft motion (x7 ) through M17 while x7 共row deleting appropriate rows and columns. For the models
seven兲 couples to x2 through 2 M71 . without tower motion, only generator speed was measured, so y
Elements in A consist of various combinations of damping and was 1⫻1.
stiffness terms for the turbine. The elements of B show how the
control input enters the system. For the seven-state model shown Control Design
above, the only nonzero elements in B are the second, third, and
DAC allowed us to regulate turbine rotor speed in the presence
seventh rows. These quantities represent the partial derivatives of of wind-speed disturbances, while placing plant poles through
the blade flap normal force, rotor aerodynamic torque, and rotor full-state feedback 关14兴. It also allowed us to use state estimation
thrust force with respect to pitch angle. These values reflect the to provide the controller with values for those states that were not
capability to control the rotor symmetric flap mode, the rotor aero- measured. This is important when limited turbine measurements
dynamic torque, and tower first fore-aft mode using rotor collec- are available. A basic idea of DAC is the augmentation of the state

388 Õ Vol. 125, NOVEMBER 2003 Transactions of the ASME

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 estimator for the turbine with additional states for estimation of
the wind-speed disturbances. This means that the original A ma-
trix for the turbine was modified to Ā and contained terms related
to the wind-speed disturbance. In addition, the original C matrix
was modified to C̄ and now accounted for the output of the wind-
speed disturbance estimator. The control law in DAC was as-
sumed in the form of the usual feedback of the plant states as well
as feedback of the wind-disturbance states 关14兴. A requirement for
successful state estimation using DAC was that the pair (Ā,C̄)
must be observable 关14兴.
Another basic idea in DAC is that the wind disturbance gain is
chosen to cancel or minimize the affect of wind-speed distur-
bances. For some of the cases shown in this paper, exact cancel-
lation was possible 共as will be later shown兲, while for other cases,
the wind disturbance gain had to be appropriately chosen to mini-
mize the norm of a vector quantity 关14兴, which resulted in mini-
mization of disturbance effects instead of exact cancellation. For
all of the controllers designed here, we used just one extra state to
estimate the wind-speed disturbance. We assumed that the wind
fluctuations were in the form of step functions, as shown
in Stol 关10兴.
For the results shown in this paper, we designed controls for the
two-bladed, AWT27 CR machine described in Buhl 关15兴. Even
though this turbine is not representative of most commercial ma-
chines, we thought this configuration would represent a good ex-
ample case to illustrate critical issues in wind turbine control de-
sign. In addition, the machine properties and models used in this
work are in the public domain, available for use by other research-
ers who may want to verify or extend these results. We used the
physical and operating parameters of the AWT27 CR to compute
values for the state matrices. To compute elements in the B and ⌫
matrices, we generated tables of machine aerodynamic torque as a
function of blade pitch and wind speed using the FAST code.
These curves are shown in Figs. 2 and 3 and were generated for
the rotor rotating at 53.33 revolutions per minute 共rpm兲. We also
generated tables of blade flap normal force versus wind speed and
pitch using FAST.
We were able to determine the open-loop natural frequencies of
the system described by Eq. 共2兲 by determining the eigenvalues of
the A matrix. This gave the following open-loop poles for this
system: tower first fore-aft bending: ⫺0.07⫾6.0j radians/second
共r/s兲, first drive train torsion: ⫺0.30⫾20.96j r/s, rotor first sym-
metric flap: ⫺2.53⫾16.86j r/s, and generator rotation rate:
⫺0.01 r/s. The tower first fore-aft bending mode had a natural
frequency of 6.0 r/s, while the first symmetric flap mode had a
natural frequency of 16.86 r/s. The first drive train torsional mode
had a frequency of 20.96 r/s. Note that all of the flexible modes,
except the rotor first symmetric flap mode, had a very small
amount of damping, as seen by the real parts of the various poles
corresponding to these modes. A goal of control design was to add
Fig. 2 Plot of rotor aerodynamic torque versus blade pitch for
various wind speeds
Journal of Solar Energy Engineering
Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use
damping to these modes by placing their poles so that the real
parts are farther to the left of the corresponding open-loop pole
values in the complex plane.
We designed our first controller based on a linear model having
just one state: perturbed generator speed. We chose an operating
point for control design 共the control design point兲 at a wind speed
of 16 meters per second 共m/s兲, generator speed of 53.33 rpm, and
a blade pitch angle of 9.5 deg. At this operating point 共16 m/s兲, the
input gain B, which is the partial derivative of rotor torque with
respect to pitch, is nonzero. Control becomes very difficult for
pitch angles close to 5 deg, where the control gain becomes zero
共at a wind speed of 16 m/s兲 and the system becomes uncontrol-
lable as seen in Fig. 2.
In order to use state estimation, the modeled system must be
observable. We confirmed observability of the augmented pair
(Ā,C̄), by examining the rank of the controllability matrix. This
allowed us to perform state estimation by only measuring genera-
tor speed 共using models without the tower first fore-aft mode兲. A
tower-top deflection or acceleration measurement was needed
when designing controls using the model with the tower first fore-
aft mode, shown in Eq. 共2兲. This was due to weak observability of
the tower first fore-aft mode in the generator speed signal.
We usually placed plant poles 共through full-state feedback兲 to
increase the damping in the flexible modes and states modeled in
this study 共compared to their open-loop damping values兲. We did
not place poles to give critical damping ratios 共the classical value
of 0.707兲 because of the resulting high-pitch rates. Through pole
placement, we wanted to improve the speed regulation of the tur-
bine, thus we designed the controller so that the closed-loop pole
for rotor rotation was chosen at ⫺2 r/s 共compared to its open-loop
value of ⫺0.01). Typically, state-estimator poles were placed to
have real parts at ⫺15 r/s, while the corresponding plant poles
were placed to have real parts at ⫺2 r/s. This assured that errors
between the estimated and actual plant states would quickly decay
to zero, providing good estimates of the plant states. We show
later in the paper the effect of pole location on blade pitch rates.
After implementing the controller designed from the one-state
model into FAST, we simulated the turbine operation with a uni-
form wind 共normal to the rotor disk兲, exciting the system with step
changes in wind speed. Figure 4 shows the simulated generator
speed and the blade pitch angle with control beginning at 10 s.
Step changes in wind speed occurred every 5 s after start of con-
trol. The figure shows that generator speed was tightly regulated
to the 53.33 rpm set point. Blade pitch varied from 9.5 to 14.4 deg
during this control effort.
One of the advantages of DAC is that it provides an estimate of
the modeled disturbance state. In this DAC design, we included a
state to model the uniform component of wind speed over the
rotor disk, which gave us an estimate of this wind speed compo-
nent. Figure 5 shows a plot of the actual wind speed input to
FAST and the corresponding estimated wind speed showing good
agreement between actual and estimated wind speed. We used
Fig. 3 Plot of rotor aerodynamic torque versus wind speed for
various pitch angles
NOVEMBER 2003, Vol. 125 Õ 389
 constant gains in the design of this controller, even though the true
gains vary with wind speed and pitch due to aerodynamic nonlin-
earity. As the turbine’s operating point deviated significantly from
the control design point, the controller estimated the wind speed
less accurately.
We then tested the effects of unmodeled modes. The most dra-
matic effect occurred when we simulated this same case with the
drive train torsion mode switched on in FAST 共we neglected this
mode in the controller design兲. As can be seen in Fig. 6, the
generator speed became unstable, with the frequency of oscilla-
tion matching the drive train torsion frequency. We added a large
amount of structural damping to this mode in the FAST input file
and reran the code but found that the response was still not stable.
We proceeded to develop a linear model with additional states
to account for this mode. This resulted in a model with these three
states:
• x1 -perturbed rotor speed
• x2 -perturbed drive train torsional spring force
• x3 -the perturbed generator speed
Fig. 4 Plot of simulated generator speed and blade pitch us-
ing DAC designed from one-state model
Fig. 5 Plot of actual and estimated wind speeds using DAC
designed from one-state model
Fig. 6 Plot of simulated generator speed with and without
drive train shaft torsion using DAC designed from one-state
model
390 Õ Vol. 125, NOVEMBER 2003
Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use
We checked controllability of 共A,B兲 and observability of (Ā,C̄)
and found that these matrix pairs were controllable and observ-
able. Through pole placement we wanted to improve speed regu-
lation and add significant damping to the first drive train torsional
mode. We designed the controller to place the plant poles at ⫺2
⫾20.5j r/s,⫺2 r/s. The first pair of poles corresponded to the
drive train torsion mode, while the third pole corresponded to the
generator speed. The state estimator poles were placed at ⫺14 r/s,
⫺15 r/s, and ⫺16 r/s. The extra estimator pole for measuring
wind speed was placed at ⫺17 r/s.
We incorporated this revised controller into FAST and reran the
same simulation. Figure 7 shows the simulated generator speed
and the blade pitch confirming that the system was stable. The
results were almost identical to the previous results shown in Fig.
4, except that the drive train torsion oscillation 共at 20.5 r/s兲 is
evident in these responses. These oscillations die out about 2 s
after application of the step changes in wind because of the place-
ment of the real part of the poles at ⫺2 r/s. Figure 8 shows the
actual and estimated wind speeds for this case, which were almost
identical to the results for the one-state controller model. Drive
train torsion oscillations are also evident in these results.
We simulated this same case with the blade first flap mode
switched on in FAST using the controller based on the three-state
model. Undesirable fluctuations in generator rotational speed can
be seen in Fig. 9. We added a large amount of structural damping
in the first flap mode, but the results with increased damping
looked similar to these results. We then turned the first flap off and
turned on other modes. Simulation using this controller with the
blade-first edge mode showed some instability, but these results
were dramatically improved with the addition of a small amount
of structural damping in this mode. Simulation with other modes
switched on showed little difference from the results just pre-
sented in Fig. 7. We thus decided to extend our linear model to
include the first flap DOF for blades one and two.
After deriving linear equations of motion with the first flap
mode for blades one and two we determined that this system was
Fig. 7 Plot of generator speed and blade pitch during control
using DAC designed from three-state model
Fig. 8 Plot of actual and estimated wind speeds using DAC
designed from three-state model
Transactions of the ASME
 uncontrollable using rotor collective pitch. This was because we
are trying to control the first flap mode of each blade with just one
control input, rotor collective pitch. In order to obtain a control-
lable system, we made a transformation of coordinates in these
equations to rotor first symmetric and first asymmetric flap. These
transformed states were just linear combinations of blades one and
two first flap. Because we were only using rotor collective pitch
共pitch is assumed identical for both blades兲, it was only possible to
control the rotor symmetric mode, not the rotor asymmetric mode.
We thus deleted states corresponding to rotor-first asymmetric flap
from these equations of motion resulting in a controllable system.
The resulting states are:
• x1 -perturbed rotor first symmetric flap tip displacement,
• x2 -perturbed rotor first symmetric flap tip velocity,
• x3 -perturbed rotor speed,
• x4 -perturbed drive train torsional spring force,
• x5 -perturbed generator speed.
We assessed controllability of this system by examining the rank
of the controllability matrix. We found that nonzero values for the
control gains ensured that the system was controllable. If the gain
corresponding to the partial derivative of blade flap normal force
with respect to pitch was zero, there was still controllability pro-
vided that certain coupling terms in the A matrix were present.
These terms coupled the rotor first symmetric flap to the drive
train torsion mode and other DOF. We did not perform control
designs for turbine operating points in which this gain was close
to zero. At the design points that we chose, elements in rows two
and three of the B matrix had nonzero values, which ensured
controllability.
The other issue was observability of the augmented pair (Ā,C̄).
We assessed observability by calculating the rank of the observ-
ability matrix. In this case, the presence of various coupling terms
in the A matrix was more critical because we were using generator
speed as the assumed measurement. This allowed us to use state
estimation to estimate the unmeasured states of the linear model.
Fig. 9 Plot of simulated generator speed with and without
blade first flap using DAC designed from three-state model
Fig. 10 Plot of generator speed and blade pitch during control
using DAC designed from five-state model
Journal of Solar Energy Engineering
Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use
We found that if certain coupling terms were zero, then observ-
ability of this system was lost. Typical terms in the A matrix that
were important involved coupling terms that were present for non-
zero blade-pitch angles. This resulted in coupling between the
rotor first symmetric flap mode and the other DOF being modeled
in this system: rotor rotation, drive train torsion, and generator
rotation. We chose cases having significant blade-pitch angles 共of
at least 5 deg兲. We did not test the controller for operating points
having small pitch angles, but one would expect that it would take
much larger control efforts to meet the stated control objectives
than for cases with larger pitch angles, due to reduced controlla-
bility and observability.
Having resolved these controllability and observability issues,
we proceeded to design a controller from the five-state model
having poles at ⫺2.5⫾16.4j r/s, ⫺2⫾20.5j r/s, and ⫺2 r/s.
Placement of closed-loop poles at these locations gave improved
speed regulation and improved damping in the first drive train
torsion mode. The first pole pair corresponded to the rotor first
symmetric flap mode, while the second pole pair corresponded to
the first drive train torsion mode. The fifth pole corresponded to
generator speed. We placed the state estimator poles at ⫺14 r/s,
⫺15 r/s, ⫺16 r/s, ⫺17 r/s, and ⫺18 r/s, and the wind speed es-
timator pole at ⫺19 r/s. Placement of the poles at these locations
gave good damping in the flexible modes and generator speed, as
well as quick decay of state estimator error transients.
We simulated the turbine with this controller for the same case
as shown previously. Figure 10 shows the generator speed and
blade pitch for this case. Evident in those responses is a damped
vibration because of drive train torsion coupled with rotor sym-
metric flap motion. That coupling resulted from the cross-coupling
terms in the A and M matrices.
This case also presented some differences compared to results
using the lower-order controllers. We were unable to regulate ro-
tor speed to the 53.33 rpm set point as accurately as with the
lower-order controllers. In this case, the wind-speed disturbances
were not exactly cancelled as they were for the previous models.
Here, we had to choose the wind-speed disturbance gain to mini-
mize the effects of wind-speed disturbances, instead of exact can-
cellation. This was due to increased complexity in the linear
model used for control design because we were modeling several
states of the turbine. We can conclude from those results that as
we attempt to stabilize additional flexible modes of the turbine,
the ability to regulate turbine speed to the desired set point is
degraded.
Figure 11 shows the estimated wind speeds for this case. This
was different from the other cases because the errors between
estimated and actual wind speeds were greater than for the previ-
ous two cases due to the turbine aerodynamic nonlinearity, with
control gains varying with wind speed and pitch angle. It was also
due to the fact that wind speed disturbances can no longer be
exactly cancelled.
Figure 12 shows the rotor symmetric flap tip displacement. This
Fig. 11 Plot of actual and estimated wind speeds using DAC
designed from five-state model
NOVEMBER 2003, Vol. 125 Õ 391
 response is well behaved because the system had been stabilized
with this control system 共because damping had been enhanced
through pole placement兲. We noted that the FAST simulation con-
tained both the rotor first symmetric and first asymmetric flap,
since the first flap mode of each blade was switched on during
nonlinear simulation. This controller was designed from a model
with just the rotor first symmetric flap mode, while ignoring the
rotor first asymmetric flap mode. Those results show that in simu-
lation the rotor first asymmetric flap mode remained stable during
closed-loop operation, even though that mode is neglected in the
linear model used for control design.
Figure 13 shows simulated tower-top fore-aft displacement us-
ing the controller designed from the model with five states. This
motion was dominated by contribution from the tower first fore-
aft bending mode, which is very lightly damped during normal
operation. There was some tendency for the tower motion to be
excited with this controller, although the motion remains stable. It
would have been very desirable, however, to add damping to this
mode with the use of rotor collective pitch control.
We thus designed a DAC controller using the full seven-state
model shown in Eq. 共2兲, which includes the tower first fore-aft
mode. Closed-loop poles were placed at ⫺2⫾6.0j r/s, ⫺2.5
⫾16.4j r/s, ⫺2⫾20.5j r/s, and ⫺2 r/s. The first pole pair corre-
sponded to the tower first fore-aft mode; the second pole pair
corresponded to the rotor first symmetric flap mode; the third pole
pair corresponded to the first drive train torsion mode; and the
seventh pole corresponded to the generator-speed.
We found that when we measured only generator speed for this
control, the tower first fore-aft mode was only weakly observable,
causing numerical problems during closed-loop simulation. When
the tower-top fore-aft deflection or acceleration was measured, the
observability of this mode was greatly increased, resulting in im-
proved simulation results. The expression in Eq. 共2兲 for yគ reflects
measurement of tower-top fore-aft deflection. This equation can
be modified to express tower-top fore-aft acceleration, with only
Fig. 12 Plot of blade tip flap displacement during control us-
ing DAC designed from five-state model
Fig. 13 Plot of tower-top fore-aft deflection during control us-
ing DAC designed from five- and seven-State models
392 Õ Vol. 125, NOVEMBER 2003
Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use
modest additional complexity. This controller was designed as-
suming that tower-top fore-aft acceleration was measured. Figure
13 shows simulated tower-top fore-aft deflections using this con-
troller 共seven-state model兲, with a dramatic improvement in
damping.
This was one of several benefits of using state-space methods to
design control systems, since damping can be added through pole
placement. A disadvantage with this control design is that tower-
top displacement 共or acceleration兲 must be measured in addition
to generator speed.
It was also important to test these control designs using more
realistic wind inputs. Step winds are good for testing the system
dynamics during closed-loop control. We now show results when
using turbulent wind inflow.
Other Simulation Results
We proceeded to test the controller in simulations with the ef-
fects of turbulent wind inflow. We generated turbulent winds using
the SNL WIND3D code 关16兴. We first used a controller designed
from the five-state model for a wind speed of 18 m/s. We designed
two controllers at the 18 m/s wind speed, one by placing the poles
so that the modes had high damping and the other one so that the
modes had low damping. Generator speed from both controllers is
shown in Fig. 14.
Figure 15 shows the actual and estimated wind speed for the
case with high damping. In general, the estimator does a fair job
of estimating the wind speed. When the wind speed is above 18
m/s, the model estimated higher wind speeds than actual values,
while estimating lower than actual values for wind speeds below
18 m/s. Large differences in the actual and estimated wind speeds
were evident when the wind speed drops below 16 m/s. This was
because the turbine operating point deviated significantly from the
control design point.
Fig. 14 Plot of generator rotational speed excited by turbulent
inflow, using DAC designed from five-state model at 18 mÕs
Fig. 15 Plot of estimated wind speed for case with turbulent
inflow using DAC designed from five-state model at 18 mÕs
Transactions of the ASME
 Figure 16 shows the rotor first symmetric flap displacement. As
can be seen from these plots, this turbine state remained stable
even while simulating with turbulent winds, although significant
oscillations due to excitation of this mode were evident.
These figures illustrate the effect of adding different amounts of
damping 共from the control system兲 on system response. For the
low-damped case 共low damping兲, the real parts of all of the poles
were placed at ⫺0.5 r/s, while for the high-damped case, they
were placed at ⫺2 r/s. As can be seen in Fig. 17, the shaft torque
variations were smaller for the high-damped case. In Fig. 14, we
also see that variations in generator speed were lower for the
high-damped case. Figure 16 illustrates the effect of damping on
the flap response of the turbine. It seems as if designing the con-
troller to add significantly to the damping in these modes de-
creases the system response and loads. Fig. 19 Plot of blade pitch rate simulated with turbulent inflow
Next, we simulated with the DAC designed from the model for the baseline case and reduced tower damping case
with seven states, which included the tower first fore-aft mode.

Fig. 20 Plot of tower-top fore-aft deflection simulated with tur-

bulent inflow for baseline case and reduced tower damping
Fig. 16 Plot of blade symmetric flap tip displacement excited case
by turbulent inflow using DAC designed from five-state model
at 18 mÕs

Fig. 21 Plot of pitch rate simulated with turbulent inflow for

baseline case and reduced drive-train torsion damping case
Fig. 17 Plot of shaft torque for case with turbulent inflow, us-
ing DAC designed from five-state model at 18 mÕs

Fig. 22 Plot of generator speed simulated with turbulent in-

Fig. 18 Plot of tower-top fore-aft deflection simulated with tur- flow for baseline case and reduced drive-train Torsion Damping
bulent inflow for two different controllers Case

Journal of Solar Energy Engineering NOVEMBER 2003, Vol. 125 Õ 393

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Figure 18 shows a comparison of the tower-top fore-aft displace- over a range of wind speed and pitch angles. If these controllers
ment when simulating with this controller compared to the case of resulted in stable behavior at some wind speeds and unstable re-
simulating with the DAC designed from the five-state model. The sults at other wind speeds, then gain scheduling would be more
decrease in tower-top fore-aft deflection with the DAC designed critical.
from the seven-state model is very evident. We also tested the effects of turbulent wind inflow on the con-
Another issue was pitch actuator rates using these control algo- trol of the nonlinear turbine, and found that all turbine states re-
rithms. Figure 19 shows pitch rates using the DAC designed from mained stable while excited by this inflow for this test case.
the seven-state model. We wanted to explore the effect of pole
placement on pitch rates. Figure 19 shows pitch rates for a case in
which the DAC was designed to place the poles of the tower first Future Work
fore-aft mode at ⫺0.5⫾6.0j r/s 共reduced tower damping兲 instead Further issues must be resolved before implementing these con-
of ⫺2⫾6.0j r/s 共baseline兲. Placing the poles closer to the imagi- trol algorithms, such as reducing blade-pitch rates. One possible
nary axis resulted in reduced pitch rates because the demand to solution is to use generator torque as a control actuator to add
add damping by the controller was decreased. There was a trade- damping to the drive train torsion mode. This might reduce the
off though, since decreasing the damping results in increased re- demand on the pitch control system to perform this function,
sponse. Figure 20 shows tower-top fore-aft deflection showing thereby reducing blade-pitch rates. We must also assess the effects
that reducing the damping in the tower first fore-aft mode results of uncertainties in the turbine properties and models, as well as
in increased deflections. measurement noise and actuator dynamics, etc. A thorough com-
We also studied the effect of reducing damping in the first drive parison of results from these control algorithms to results from
train torsion mode. Figure 21 shows reduced pitch rates when simple classical-control results must be performed in order to
designing the DAC to place the pole of the first drive train torsion show the advantages of using these modern state-space control
mode at ⫺0.5⫾20.5j r/s 共reduced drive train damping兲 compared designs.
to ⫺2⫾20.5j r/s 共baseline兲. As can be seen from Fig. 22, the Finally, we must implement and thoroughly test these algo-
speed regulation for both cases was about the same showing that rithms on a field test turbine. We have already designated a con-
reducing damping in this mode does not penalize this control ob- trols test machine at the National Wind Technology Center for
jective. However, the drive train shaft torsional loads will increase these field tests, and we are beginning implementation of these
when the damping in the drive train is decreased. control algorithms on this machine.
It is clear from this study that a trade-off must be performed
between allowable pitch rates and the amount of desired damping
in various flexible modes. Placing poles to add higher amounts of Nomenclature
damping 共compared to the baseline case兲 will result in even higher A ⫽ state matrix,
pitch rates than shown here. B ⫽ control input gain matrix
Another solution to reducing pitch rates may be to use genera- C ⫽ relates plant output to states
tor torque as a control input with the objective to add damping to Cd ⫽ drive train torsional damping coefficient
the first drive train torsion mode. This may decrease the demand Cij ⫽ Coefficients of damping matrix for various values of
placed on the pitch control system, thereby reducing blade-pitch i and j
rates. This issue will be studied in future work. Igen ⫽ generator mass moment of inertia
Irot ⫽ rotor mass moment of inertia
Conclusions Kd ⫽ drive train torsional spring stiffness
Kij ⫽ coefficients of stiffness matrix for various values of i
We concluded that it was possible to stabilize the rotor first and j
symmetric flap mode, the first drive train torsional mode, and the M ⫽ mass matrix
tower first fore-aft bending mode using only rotor collective pitch Mij ⫽ coefficients in mass matrix for various values of i
as the control input for this machine. It was possible to measure and j
only generator speed for design of controls using the models that uគ ⫽ control input
neglected tower fore-aft motion. Control designed from the model uD ⫽ disturbance state-space model output
containing the tower first fore-aft bending mode yielded poor xគ ⫽ state vector
simulation results when measuring only generator speed due to ẋគ ⫽ time derivative of xគ
the weak observability of that mode in the generator speed signal. xi ⫽ i th state-in-state vector xគ
The addition of a tower-top deflection or acceleration measure-
ẋ i ⫽ time derivative of i th state
ment resulted in much improved simulation results due to im-
proved observability.
yគ ⫽ control 共or measured兲 signal
For this model, various terms in the A matrix, which coupled yi ⫽ value in i th row of measured output yគ
the rotor first symmetric flap mode with the other DOF, were ⌫ ⫽ disturbance input gain matrix
important and ensured observability of the system. These terms ␥ ⫽ partial derivative of rotor aerodynamic torque with
had nonzero values for cases in which the blade pitch was non- respect to rotor speed
zero. We chose design cases that had significant nonzero blade- ␦␤ ⫽ perturbation in rotor collective pitch
pitch angles 共at least 5 deg兲. This study did not include design ␨ ⫽ partial derivative of rotor aerodynamic torque with
points with small pitch angles 共close to zero兲, where the control- respect to rotor collective pitch angle
lability and observability of the system was degraded. ␨b ⫽ partial derivative of blade root flap normal force
We also saw the effects of the nonlinear behavior of the con- 共sum of both blades兲 with respect to rotor collective
trolled turbine. We found that the nonlinear aerodynamics of the ␨t ⫽ partial derivative of rotor thrust force with respect to
turbine caused variations in the gains with wind speed and pitch rotor collective pitch angle
angle. The controller was designed using one set of gains appro- ␦w ⫽ perturbed wind disturbance 共uniform over rotor disk兲
priate for a particular design wind speed and blade-pitch angle. As ␣ ⫽ partial derivative of rotor aerodynamic torque with
the turbine’s operating point deviated significantly from this de- respect to wind speed
sign point, the wind-speed estimator became less accurate. A ␣b ⫽ partial derivative of blade root flap normal force with
strong need for gain scheduling was not firmly established. It was respect to wind speed
demonstrated in the paper that the controls designed from the ␣t ⫽ partial derivative of rotor thrust force with respect to
higher-order linear models resulted in stable closed-loop behavior wind speed

394 Õ Vol. 125, NOVEMBER 2003 Transactions of the ASME

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 References
关1兴 Hinrichsen, E. N., 1984, ‘‘Controls for Variable Pitch Wind Turbine Genera-
tors,’’ IEEE Trans. Power Appar. Syst., 103, pp. 886 – 892.
关2兴 Barton, R. S., Bowler, C. E. J., and Piwko, R. J., 1979, ‘‘Control and Stabili-
zation of the NASA/DOE MOD-1 Two Megawatt Wind Turbine Generator,’’
American Chemical Society, 101, 325–330.
关3兴 Mattson, S. E., 1984, ‘‘Modeling and Control of Large Horizontal Axis Wind
Power Plants,’’ Ph.D. Thesis, Lund Institute of Technology, Department of
Automatic Control, Lund, Sweden.
关4兴 Liebst, B. S., 1983, ‘‘Pitch Control System for Large Scale Wind Turbines,’’ J.
Energy, 7共2兲, pp. 182–192.
关5兴 Steinbuch, M., 1989, ‘‘Dynamic Modeling and Robust Control of a Wind
Energy Conversion System,’’ Ph.D. Thesis, Delft University of Technology,
The Netherlands.
关6兴 Knudsen, T., Andersen, P., and Toffner-Clausen, S., 1997, ‘‘Comparing PI and
Robust Pitch Controllers on a 400 kW Wind Turbine by Full-scale Tests,’’
Proc. European Wind Energy Conference, Dublin, Ireland, October, pp. 546 –
550.
关7兴 Johnson, C. D., 1976, ‘‘Theory of Disturbance Accommodating Controllers,’’
Advances in Control and Dynamic Systems, 12, pp. 387– 489, C.T. Leondes,
ed.
关8兴 Balas, M. J., 1980, ‘‘Disturbance Accommodating Controllers for Distributed
Parameter Systems: An Introduction,’’ J. of Interdisciplinary Modeling and
Simulation, January.
Journal of Solar Energy Engineering
Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 02/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use
关9兴 Balas, M. J., 1990, ‘‘Active Control of Persistent Disturbances in Large Pre-
cision Aerospace Structures,’’ SPIE Cont., Orlando, Florida, April.
关10兴 Stol, K., Rigney, B., and Balas, M., 2000, ‘‘Disturbance Accommodating Con-
trol of a Variable-Speed Turbine using a Symbolic Dynamics Structural
Model,’’ Proceeding of the 2000 ASME Wind Energy Symposium, Reno, Ne-
vada, pp. 84 –90.
关11兴 Kane, T. R. and Levinson, D. A., 1985, Dynamics: Theory and Applications,
McGraw-Hill Book Company, New York, NY, pp. 1–379.
关12兴 Wilson, R. E., Walker, S. N., and Heh, P., 1999, ‘‘Technical and User’s Manual
for the FAST Advanced Dynamics Code,’’ OSU/NREL Report 99-01, Oregon
State University, Corvallis, Oregon.
关13兴 Stol, K., and Balas, M., 2001, ‘‘Full-State Feedback Control of a Variable-
Speed Wind Turbine: A Comparison of Periodic and Constant Gains,’’ J. Eng.
Mech. Div., 123共4兲, pp. 319–326.
关14兴 Balas, M. J., Lee, Y. J., and Kendall, L., 1998, ‘‘Disturbance Tracking Control
Theory with Application to Horizontal Axis Wind Turbines,’’ Proceeding of
the 1998 ASME Wind Energy Symposium, Reno, Nevada, pp. 95–99.
关15兴 Buhl, M. L., Wright, A. D., and Pierce, K. G., 2001, ‘‘FAST Code Verification:
A Comparison to ADAMS,’’ Proceeding of the 2001 ASME Wind Energy Sym-
posium, Reno, Nevada, pp. 368 –377.
关16兴 Kelley, N. D. 1992, ‘‘Full Vector 共3-D兲 Inflow Simulation in Natural and Wind
Farm Environments Using an Expanded Version of the SNLWIND 共Veers’兲
Turbulence Code,’’ NREL/TP-442-5225. National Renewable Energy Labora-
tory, Golden, Colorado.
NOVEMBER 2003, Vol. 125 Õ 395