NAVAL

POSTGRADUATE
SCHOOL
MONTEREY, CALIFORNIA

THESIS
ADAPTIVE CONTROL AND PARAMETER
IDENTIFICATION OF A DOUBLY-FED INDUCTION
GENERATOR FOR WIND POWER
by
Stamatios Orfanos - Pepainas
September 2011
Thesis Co-Advisors:
Second Reader:

Roberto Cristi
Alexander L. Julian
Xiaoping Yun

Approved for public Release; distribution is unlimited

THIS PAGE INTENTIONALLY LEFT BLANK

REPORT DOCUMENTATION PAGE

Form Approved OMB No. 0704-0188

Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instruction,
searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of information. Send
comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing this burden, to
Washington headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington, VA
22202-4302, and to the Office of Management and Budget, Paperwork Reduction Project (0704-0188) Washington DC 20503.

1. AGENCY USE ONLY (Leave blank)

2. REPORT DATE
September 2011

3. REPORT TYPE AND DATES COVERED

Masters Thesis
4. TITLE AND SUBTITLE
5. FUNDING NUMBERS
Adaptive Control and Parameter Identification of a Doubly-Fed Induction Generator
for Wind Power
6. AUTHOR(S) Stamatios Orfanos-Pepainas
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES)
8. PERFORMING ORGANIZATION
Naval Postgraduate School
REPORT NUMBER
Monterey, CA 93943-5000
9. SPONSORING /MONITORING AGENCY NAME(S) AND ADDRESS(ES)
10. SPONSORING/MONITORING
N/A
AGENCY REPORT NUMBER
11. SUPPLEMENTARY NOTES The views expressed in this thesis are those of the author and do not reflect the official policy
or position of the Department of Defense or the U.S. Government. IRB Protocol number: N/A.
12a. DISTRIBUTION / AVAILABILITY STATEMENT
1. Approved for public release; distribution unlimited
13. ABSTRACT (maximum 200 words)

12b. DISTRIBUTION CODE

The use of Doubly-Fed Induction Generators (DFIG) for wind energy conversion is addressed in this thesis. It is well
known that when the stator is connected to the electric grid, the rotor voltage can control both mechanical torque and
reactive electric power.
To guarantee efficient wind energy conversion, it is important to research and design more advanced control
schemes. In this thesis, we first review the basic theory behind DFIGs and Adaptive Control. Next we design an
adaptive controller for a wind turbine using a DFIG and model and simulate the system. In order to create a valid
assessment on the results of this method, we compare the systems performance with a standard control scheme based
on proportional integral (PI) controllers as proposed in standard approaches.
14. SUBJECT TERMS
Wind Energy, DFIG, Adaptive Control

17. SECURITY
CLASSIFICATION OF
REPORT
Unclassified

15. NUMBER OF
PAGES
103
16. PRICE CODE

18. SECURITY
CLASSIFICATION OF THIS
PAGE
Unclassified

NSN 7540-01-280-5500

19. SECURITY
CLASSIFICATION OF
ABSTRACT
Unclassified

20. LIMITATION OF
ABSTRACT
UU

Standard Form 298 (Rev. 8-98)

Prescribed by ANSI Std. Z39.18

THIS PAGE INTENTIONALLY LEFT BLANK

ii

Approved for public release; distribution is unlimited

ADAPTIVE CONTROL AND PARAMETER IDENTIFICATION OF A DOUBLYFED INDUCTION GENERATOR FOR WIND POWER

Stamatios Orfanos-Pepainas
Lieutenant, Hellenic Navy
BS, Hellenic Naval Academy, June 2001

Submitted in partial fulfillment of the

requirements for the degrees of

ELECTRICAL ENGINEER
and
MASTER OF SCIENCE IN ELECTRICAL ENGINEERING

from the

NAVAL POSTGRADUATE SCHOOL

September 2011

Author:

Stamatios Orfanos-Pepainas

Approved by:

Roberto Cristi
Thesis Co-Advisor
Alexander L. Julian
Thesis Co-Advisor
Xiaoping Yun
Second Reader
R. Clark Robertson
Chair, Department of Electrical and Computer Engineering
iii

THIS PAGE INTENTIONALLY LEFT BLANK

iv

ABSTRACT
The use of Doubly-Fed Induction Generators (DFIG) for wind energy conversion is
addressed in this thesis. It is well known that when the stator is connected to the electric
grid, the rotor voltage can control both mechanical torque and reactive electric power.
To guarantee efficient wind energy conversion, it is important to research and
design more advanced control schemes. In this thesis, we first review the basic theory
behind DFIGs and Adaptive Control. Next we design an adaptive controller for a wind
turbine using a DFIG and model and simulate the system. In order to create a valid
assessment on the results of this method, we compare the systems performance with a
standard control scheme based on proportional integral (PI) controllers as proposed in
standard approaches.

THIS PAGE INTENTIONALLY LEFT BLANK

vi

TABLE OF CONTENTS

I.

INTRODUCTION........................................................................................................1
A.
BACKGROUND ..............................................................................................1
B.
THESIS OBJECTIVES...................................................................................2
C.
RELATED WORK ..........................................................................................2
D.
APPROACH.....................................................................................................4
E.
THESIS OUTLINE..........................................................................................4
A.
WIND TURBINE THEORY...........................................................................4
1.
Energy in the Wind..............................................................................5
2.
Wind Turbine Aerodynamics Disc Actuator Model......................6
3.
Torque and Power on the Rotor .......................................................10
4.
Regions of Control .............................................................................13
B.
DOUBLY-FED INDUCTION GENERATOR (DFIG) ..............................14
1.
Reference Frames Dynamic Models..............................................15
2.
DFIG Model........................................................................................18
C.
CHAPTER SUMMARY................................................................................19

III.

ONLINE PARAMETER ESTIMATION ................................................................21

A.
INTRODUCTION..........................................................................................21
B.
SYSTEM EQUATIONS ................................................................................22
C.
ESTIMATORS...............................................................................................23
1.
Least Mean Squares (LMS) ..............................................................23
a.
Equations Setup ......................................................................24
b.
Implementation .......................................................................26
2.
Recursive Least Squares (RLS) ........................................................28
a.
Equations Setup ......................................................................28
b.
Implementation .......................................................................29
3.
RLS With Exponential Forgetting (RLS-EF)..................................30
a.
Equations Setup ......................................................................30
b.
Implementation .......................................................................31
D.
SIMULATION OF THE ESTIMATORS....................................................31
1.
Equipment ..........................................................................................31
2.
Simulation Results .............................................................................32
a.
Simulation Using LMS ...........................................................32
b.
Simulation Using RLS ............................................................34
E.
CHAPTER SUMMARY................................................................................36

IV.

ADAPTIVE CONTROL ...........................................................................................39

A.
INTRODUCTION..........................................................................................39
B.
ADAPTIVE CONTROL THEORY .............................................................40
C.
ADAPTIVE CONTROL OF DFIG..............................................................41
1.
First Approach: Adaptive Control for Torque and Reactive
Power...................................................................................................41
a.
Dynamic Model .......................................................................41
vii

D.
V.

b.
Simulation Model....................................................................48
c.
Simulation Results ..................................................................52
2.
Second Approach: Adaptive Control of Torque, PI Control of
the Reactive Power.............................................................................55
a.
Dynamic Model of the Controller...........................................55
b.
Simulation Model....................................................................57
c.
Simulation Results ..................................................................58
3.
Performance Comparison .................................................................62
CHAPTER SUMMARY................................................................................66

CONCLUSIONS AND RECOMMENDATIONS...................................................67

A.
SUMMARY - CONCLUSIONS....................................................................67
B.
RECOMMENDATIONS-FUTURE RESEARCH......................................68

APPENDIX:
MATLAB CODES - SIMULINK MODEL .....................................69
A.
MATLAB CODE FOR INITIALIZATION OF THE SIMULATIONS ..69
B.
SIMULINK BLOCKS ...................................................................................71
1.
First Adaptive Control Approach ....................................................71
a.
Overall System View................................................................71
b.
Aerodynamic Torque Input ....................................................72
c.
General View of the Controller ..............................................72
d.
Preprocess................................................................................73
e.
Parameter Estimation ...............................................................73
f.
Control.....................................................................................74
2.
Second Approach ...............................................................................74
a.
Preprocess................................................................................74
b.
Control.....................................................................................75
LIST OF REFERENCES ......................................................................................................77
INITIAL DISTRIBUTION LIST .........................................................................................81

viii

LIST OF FIGURES
Figure 1.
Figure 2.
Figure 3.
Figure 4.
Figure 5.
Figure 6.
Figure 7.
Figure 8.
Figure 9.
Figure 10.
Figure 11.
Figure 12.
Figure 13.
Figure 14.
Figure 15.
Figure 16.
Figure 17.
Figure 18.
Figure 19.
Figure 20.
Figure 21.
Figure 22.
Figure 23.
Figure 24.
Figure 25.
Figure 26.
Figure 27.
Figure 28.
Figure 29.
Figure 30.
Figure 31.
Figure 32.
Figure 33.
Figure 34.
Figure 35.
Figure 36.
Figure 37.

Global wind power installations, 1996 to 2010. From [1].................................1

A volume of air flowing with speed V. ..............................................................5
Actuator disc. After [14]. ...................................................................................7
Pressure around actuator disk. After [15]. .........................................................8
Turbine disc. ....................................................................................................11
Typical Cp versus speed ratio curve. .............................................................12
Example power curves for a typical commercial WT. From [5]. ....................13
Induction machine cross section. From [20]....................................................15
General view of the DFIG model.....................................................................18
General view of the LMS estimation model. ...................................................26
Detailed block diagram for the LMS estimator model. ...................................27
General view of RLS estimator model.............................................................29
Detailed view of RLS estimator model............................................................30
Convergence of rs and rr' using LMS from simulation....................................33
Convergence of Lm using LMS from simulation..............................................33
Convergence of Lls and Llr' using LMS from simulation. ...............................34
Convergence of rs and rr' using RLS from simulation.....................................35
Convergence of Lm using RLS from simulation. .............................................35
Convergence of Lls and Llr' using RLS from simulation..................................36
Direct MRAC overview. After [31].................................................................40
General view of the first controller in Simulink. .............................................48
Control block for the first approach.................................................................49
Wind speed and torque during simulation of the first approach......................53
Magnified torque plot during simulation of the first approach........................53
Real power (Pe) during simulation of the first approach. ................................54
Reactive power (Qe) during simulation of the first approach. .........................54
Comparison of the rotor currents iqr and idr with their reference values
during simulation of the first approach............................................................55
General view of the second controller model. .................................................57
Control block for the second approach. ...........................................................58
Wind speed and torque during simulation of the second approach. ................59
Magnified torque during simulation of the second approach. .........................60
Real power (Pe) during simulation of the second approach.............................60
Reactive power (Qe) during simulation of the second approach. ....................61
Comparison between real and reactive power Pe and Qe, respectively, for
part of the simulation during the second approach. .........................................61
Comparison of the currents iqs and idr with their reference values during
simulation of the second approach...................................................................62
Wind profile for comparison of the control approaches. .................................63
Comparison of output torque Te and reference torque Tref during the
simulation of a wind gust.................................................................................63
ix

Figure 38.
Figure 39.
Figure 40.
Figure 41.

Magnified plot of output torque Te and reference torque Tref during the
simulation of a wind gust.................................................................................64
Typical DFIG torque versus time curve for different values of rr. ..................65
Comparison of the variation of the output torque curves during simulation
of a variation of rr. ...........................................................................................65
Comparison of the variation of the DFIG speed during the simulation of a
variation of rr. ...................................................................................................66

LIST OF TABLES
Table 1.
Table 2.
Table 3.
Table 4.
Table 5.

Adaptation gains for the LMS simulation........................................................28

Measured DFIG parameters. From [28]...........................................................32
WT and DFIG parameters. From [7]. ..............................................................51
Final simulation parameter values for first approach. .....................................52
Final simulation parameter values for the second approach............................59

xi

THIS PAGE INTENTIONALLY LEFT BLANK

xii

LIST OF ACRONYMS AND ABBREVIATIONS

AC

Alternating Current

COE

Cost of Energy

Cp

Power coefficient of the Wind Turbine

Cq

Torque coefficient of the Wind Turbine

DFIG

Doubly Fed Induction Generator

DMRAC

Direct Model Reference Adaptive Control

Ek

Kinetic Energy

FD

Force that is applied from the air flow on the actuator disc

HAWT

Horizontal Axis Wind Turbine

iqd0r

Rotor currents in the arbitrary dq0 reference frame

iqd0s

Stator currents in the arbitrary dq0 reference frame

Combined rotor and wind turbine inertia coefficient

Llr

Leakage inductance of rotor winding

Lls

Leakage inductance of stator winding

MIMO

Multi Input Multi Output

MMF

Magneto Motive Force

MRAC

Model Reference Adaptive Control

differential operator d/dt

The Number of poles of the generator

PD

Power extracted from the wind by the actuator disc

PDF

Probability Density Function

Per

Rotor Real power output

Pes

Stator Real power output

PI

Proportional Integral

Pv

Power in the Wind

PWM

Pulse Width Modulation

Qer

Rotor Reactive power output

Qes

Stator Reactive power output

Wind Turbine Blade radius

RE

Renewable Energy
xiii

rr

Rotor resistance of the generator per phase.

rs

Stator resistance of the generator per phase.

The Laplacian Operator

SM

Sliding Mode

Te

Electromagnetic Torque output by the Generator

Tr

Aerodynamic Torque produced by the Wind Turbine

Tref

Controller Reference Torque

Wind speed in m/s

Vqd0r

Rotor voltages in the arbitrary dq0 reference frame

Vqd0s

Stator voltages in the arbitrary dq0 reference frame

WE

Wind Energy

WECS

Wind Energy Conversion System

WT

Wind Turbine

Tip speed to wind speed ratio of the Wind Turbine

opt

Tip speed to wind speed ratio of the Wind Turbine for maximum power

qd0r

Rotor flux linkages in dq0 reference frame

qd0s

Stator flux linkages in dq0 reference frame

Rotational speed of rotor

Wind turbine Rotational speed in rads/s

xiv

EXECUTIVE SUMMARY

The possibility of fossil fuel shortages in the near future and environmental concerns
about power generation has lead to an increasing demand for renewable energy (RE)
installations worldwide. Among them, wind energy (WE) is one of the fastest growing.
This form of energy production is suitable for electricity production since wind is
available everywhere. However, the random nature of the wind and the need to provide a
stable source for the grid calls for the design of effective control systems for wind energy
conversion systems (WECS). This would improve both the reliability of the energy
captured and the overall efficiency of the system.
Many large wind power generators today are based on the doubly-fed induction
generator (DFIG). The DFIG is capable of controlling both the active and reactive power
simultaneously and keeping the generator at its maximum efficiency, even in the
presence of wind speed variations. To take advantage of these features, the DFIG requires
proper control in terms of its power electronics and computer algorithms.
Standard controllers, such as proportional integral (PI) controllers for pulse width
modulation (PWM) converters at the rotor and source side of the system are the most
commonly used to provide independent control of active and reactive power output.
A different control approach based on adaptive control theory is investigated in
this thesis. The motivation for this research is the fact that WECS operate in
environments with extreme variations in the operating conditions, and their operation
depends on the wind speed, which, in general is a random process that cannot be
controlled. Adaptive control systems, on the other hand, have the ability to adapt to
variations in the systems parameters, which is a very attractive feature. Among the
various kinds of adaptive control methods, the one examined in this thesis is direct model
reference adaptive control (DMRAC). The basic idea behind this approach is to make the
system behave as closely as possible to a reference model in the presence of uncertain
and/or time-varying dynamics of the system to be controlled. To achieve this, the system
is modeled in terms of a vector of parameters c(t) that may be known or approximately
xv

known. This parameter vector is updated by an adaptive law using input and output data.
The goal of the controller is to continuously estimate the parameters creating a vector

$ c (t ) to drive the error of the output vector z%(t) between the reference model and the
actual plant to zero.
A number of methods for system modeling and parameter identification, in terms
of effectiveness and complexity, are investigated in this thesis. To improve the systems
overall performance and stability under various operating conditions and faults, the
results of the parameter identification are applied to the design of an adaptive controller
for a power grid connected WECS. A basic view of the overall system is illustrated in the
following figure.

General view of the proposed WECS.

xvi

Since there are two independent control channels, one for the active and one for
the reactive power, and they both have different requirements, the effectiveness of using
either two independent adaptive controllers or a combination of one adaptive and one
classical PI controller is considered. One can reason that the reactive power has simple
control requirements since it has to be kept to a constant value. The active power, on the
other hand, has to follow a profile for maximum torque, which changes with wind
velocity and the aerodynamics of the turbine.
To test and assess the effectiveness of the proposed techniques and their ability to
adapt to changes in dynamics, a number of different cases have been simulated, such as
changes in internal resistances and wind gusts.
It was found that DMRAC can be effectively applied in a WECS, which can
significantly improve the performance of the overall system. To that end, a physical
implementation to validate the results of this thesis is recommended.

xvii

THIS PAGE INTENTIONALLY LEFT BLANK

xviii

ACKNOWLEDGMENTS

First, I would like to express my gratitude to my parents, Dimitrio and Eleni for
teaching me the importance of education and for all the sacrifices they made to raise me
and teach me all their principles in life.
I would like to thank my thesis advisor, Professor Roberto Cristi, for his
continuous support and motivation not only throughout my research, but though the
whole period of my studies here in NPS. I would also like to thank professors Xiaoping
Yun and Alexander Julian, for their advice and the expansive knowledge that they passed
on to me during my pursuit of graduate education.
I dedicate this work to my wife, Lina, and my daughter Danae, who have always
been patient and supportive and provide me with the courage and inspiration needed to
achieve my goals in life.

xix

THIS PAGE INTENTIONALLY LEFT BLANK

xx

I.

A.

INTRODUCTION

BACKGROUND
The possibility of a shortage in fossil fuels in the near future in conjunction with

the continuously increasing demand for energy [1], [2] and environmental concerns about
high emissions has contributed to a high interest in renewable energy (RE) worldwide.
Among the various sources of RE, one of the fastest growing and most promising is wind
power.!New wind power capacity added globally during 2010 reached 39 GW, which is
more than any other renewable technology and over three times higher than the 11.5 GW
of wind power added worldwide in 2005. As a result, existing capacity increased more
than 24% relative to 2009, with total global capacity close to 198 GW by the end of 2010
[3]. The growth of the global wind power installations over the past 15 years is
demonstrated in Figure 1.

Figure 1.

Global wind power installations, 1996 to 2010. From [1].

In the U.S., the wind industry started to grow in the 1980s; by 1986, the installed
capacity in California was close to 90% of the global installed wind power. It later hit a
plateau during an electricity restructuring period in the 1990s and regained momentum in
the past decade [4].
The spread of wind energy (WE) installations indicates the need to develop wind
turbines (WT) with higher efficiency. This involves technology advances that reduce the
initial cost of the WT or allow the capture of more energy during operation, which results
in a cost-of-energy (COE) decrease [5]. One approach that can lead to more effective WT
is the design of more sophisticated controllers. The problem of designing an adaptive
controller for a variable speed horizontal axis WT (HAWT) using a doubly-fed induction
generator (DFIG) is addressed in this thesis.
B.

THESIS OBJECTIVES
The objective of this thesis is to apply various methods of system identification to

a DFIG and investigate the design of an adaptive controller for a power grid connected
wind energy conversion system (WECS) to improve the systems overall performance
and stability under various operating conditions and faults. To achieve those goals,
attempts are made to control both mechanical torque and reactive electric power
independently using the rotor voltage components as a control input.
C.

RELATED WORK
Various researchers over the past several years have addressed the problem of

efficient control of WECSs. The complexity of the overall system and the random
conditions under which WTs operate have led scientists to investigate different control
approaches to improve performance and stability under various fault conditions. A novel
control method for smoothing the stator active or reactive power ripple components under
unbalanced grid voltage is proposed in [6]. In [7], the design of a multi-input multi-output
(MIMO) sliding mode (SM) nonlinear control strategy for a WECS is presented. The
resulting robust control law guarantees finite time convergence, whereas smoothing
discontinuities at the basis of sliding mode control achieves chattering reduction. The
2

paper concludes that this type of control is a particularly suitable option to deal with
electronically controlled variable speed operating WECS. Furthermore, SM control has
proven to be robust with respect to system parameter variations and external disturbances
as well as poorly known operating environments [8], [9]. In [10], direct model reference
adaptive control (DMRAC) was used to design an adaptive collective pitch controller for
a HAWT. The objective of the adaptive pitch controller was to regulate generator speed
for wind speeds higher than the rated value for the generator and to reject step
disturbances. This objective was accomplished by collectively pitching the turbine
blades. The results were compared in simulations with a classical proportional integrator
(PI) collective pitch controller. In the simulations, the adaptive pitch controller showed
better speed regulation for high wind speeds.
A different approach of optimal adaptive control is presented in [11] for a
permanent magnet synchronous generator driven by a pulse width modulation (PWM)
inverter. The controller focused on optimization of the system by minimizing Joule losses
in the inverter and generator. To achieve this, the paper proposed the operation of the
system at an operating point with minimum phase current, considering fixed torque.
Other, more classical approaches include the independent control of active and
reactive power using back-to-back PWM voltage source converters in the rotor circuit. A
vector control scheme on the supply side PWM ensures independent control of active and
reactive power drawn from the supply, while vector control of the rotor side PWM results
in a wide speed - range operation. The control is achieved using PI controllers to enable
maximum speed tracking for maximum power extraction from the wind. This approach,
presented in [12], presents two different controls, a current-mode control and a speedmode control. A similar approach was presented in [13], where a Stator-Flux Oriented
vector control approach is deployed for both stator and rotor side converters to provide
independent control of active and reactive power using a decoupled design based on
internal model control.

D.

APPROACH
In the first part of this research, the analytical design of an online identification

algorithm is presented, with the intent of constructing a Simulink model of the physical
system. In the second part, an adaptive control scheme to compensate for dynamic
variations of the generator is proposed and its effectiveness is evaluated by computer
simulations. In the simulations, various fault conditions are implemented, and the results
are discussed subsequently.
E.

THESIS OUTLINE
This thesis is organized as follows: the concept of the basic theory of WTs and

DFIGs is covered in Chapter II, and the model created for the system in the Simulink
environment is described. The theory of online system identification including modeling
and simulation results for different estimation algorithms, is covered in Chapter III. A
basic introduction to adaptive control theory and the design and simulation of two
different adaptive controllers, along with simulation results under different operational
conditions of the system, are presented in Chapter IV. A comparison with simulation
results from other control approaches is also presented. Finally, the conclusions based on
the results from the previous chapters, along with some recommendations for future
research in the area of adaptive control for WTs, are presented in Chapter V. The
Appendix includes Matlab code for initialization of the simulation of the controllers as
well as the basic model blocks in the Simulink environment.II.

WIND TURBINES

DFIG
A.

WIND TURBINE THEORY

Wind turbines are designed to capture a portion of the wind's kinetic energy and

convert it to usable energy [14] in the form of electricity. This section presents a brief
explanation of the basic operating principles of WTs.

1.

Energy in the Wind

Wind speed is a random process that depends on several factors. The main
parameter that describes the ability to extract power is the mean wind speed for the
specific area of interest. The probability distribution function (PDF) of the mean wind
speed Vm over an area is modeled as a Weibull distribution [14] of the form:

k V (Vm )
P(Vm ) = m e c
c c

k 1

(0.0)

where the parameters k and c are called the shape and scale factor, respectively, and are
both positive real numbers.

Figure 2.

A volume of air flowing with speed V.

If one takes a volume of air Ax flowing with speed V as shown in Figure 2, the
kinetic energy stored in the wind is
1
E k = mV 2 ,
2

(0.0)

where m is the mass of the air.

If we replace the mass in terms of the air density and the volume as m = x ,
Equation (2.2) can be written as
5

1
E k = (Ax)V 2 ,
2

(0.0)

and from Equation (2.3) one can determine the wind energy per unit volume x = 1 as
1
Ek = V 2 .
2

(0.0)

Since the power is defined as the derivative of the energy with respect to time,
one can derive an expression for the power of the wind:
1
(Ax)V 2 1
2
=
= AV 3 .
PV =
t
t
2

2.

(0.0)

Wind Turbine Aerodynamics Disc Actuator Model

Turbine aerodynamics describes the forces developed on a wind turbine by an

airflow passing through it. The approach used to derive a model for the wind turbine is
actuator disc theory [14]. In this theory, the WT is regarded as an actuator disc, a device
that is designed to extract part of the kinetic energy of the wind. This theory also gives a
value for the maximum efficiency limit. If one assumes the disc is immersed in an airflow
as shown in Figure 3, then the upstream wind speed V has to be greater than the
downstream wind speed V-.

Figure 3.

Actuator disc. After [14].

Furthermore, for the stream tube just enclosing the disc, the upstream cross
sectional area A is smaller than the disc area AD, which in turn is smaller than the
downstream cross-sectional area A-. This happens because, by definition, due to the law
of mass conservation, the mass flow must remain constant within the stream tube [14].
The above statement implies:

= cons tant ,
t

(0.0)

where M is the air mass going through a surface during time t. Now one can derive the
following expression:

x
= A
= AV = cons tant.
t
t

(0.0)

Therefore, if one takes ideal points away from the turbine ( and + infinity) and
on the turbine, one obtains:

V = DVD = V .

(0.0)

Since the air that passes through the disc undergoes a speed reduction equal to
V V , this implies a reduction in its kinetic energy. This results in a force FD developed
by the actuator disc on the incident airflow, which is given by Newton's second law of
motion as the product of mass M and acceleration a:

FD = AD x

(V V )
= (V V ) ADVD .
t

(0.0)

The force FD originated by the pressure drop introduced by the actuator disc can
also be expressed as:

FD = (pD+ pD )AD ,

(0.0)

where, pD+ and pD are the values of the air pressure immediately before and after the disc,
as shown in Figure 4.

Figure 4.

Pressure around actuator disk. After [15].

Since VD is smaller than V, one can write the speed at the disc as
VD = (1 )V ,

(0.0)

where one defines , with 0 1 , which is referred to as the axial flow interference
factor. Substituting Equation (2.11) in Equation (2.9), one gets the force FD as:
FD = (V V ) ADV (1 ) .
8

(0.0)

Equating Equations (2.10) and (2.12), one obtains:

(V V )V (1 ) = (pD+ pD ) .

(0.0)

For an airflow that is stationary, incompressible, and frictionless and with no

external forces applied either upstream or downstream, the Bernoulli equation states that
the total energy of the airflow remains constant. Given that the above conditions are
satisfied, one can apply the Bernoulli equation from far upstream to just in front of the
rotor and from just behind the rotor to far downstream [16]:
1
2
(pD+ pD ) = (V 2 V
).
2

(0.0)

Substituting Equation (2.14) in Equation (2.13), one finally gets

V = (1 2 )V ,

(0.0)

and, since V has to be non-negative, Equation (2.15) gives us the upper bound of ,
which is 0.5.
Substituting Equation (2.15) in Equation (2.12), one gets an expression for the
force that is applied from the airflow on the disc:

FD = 2 ADV 2 (1 ) .

(0.0)

The power that the actuator disc extracts from the wind is given by

PD = FDVD = 2 ADV 3 (1 )2 .

(0.0)

This gives an idea of the maximum power available to the turbine from the flow
of the wind.
A conventional way of characterizing the ability of a wind turbine to capture wind
energy is the power coefficient Cp [14], which is defined as the ratio of extracted power
divided by the total wind power:
Cp =

PD
.
PV

(0.0)

Substituting Equations (2.2) and (2.11) in (2.12), one gets the following
expression for Cp:
Cp =

2 ADV 3 (1 )2
= 4 (1 )2 .
0.5 ADV 3

(0.0)

The power coefficient Cp is a function of the axial flow interference factor . The
maximum theoretical value of the power coefficient can be found from Equation (2.19)
by taking its first and second derivatives. The result, which is called the Betz limit, is

Cp _max = 0.593 for equal t 1/3 [17]. This limit is valid for all kinds of wind turbines, but
the actual efficiency is usually lower, and for most commercial wind turbines, a good
value is between 0.40-0.50.
In practice three effects that lead to a decrease in the maximum achievable
power coefficient [18]:

Rotation of the wake behind the rotor.

Finite number of blades and associated tip losses.

Non zero aerodynamic drag.

3.

Torque and Power on the Rotor

The relation between the energy of the wind flowing through the windmill and the
mechanical power available at the rotor is discussed in this section. The parameter is
the pitch angle of the turbine blades, and the tip speed to wind speed ratio :

r R
.
V

(0.0)

A graphical representation of the parameters involved in Equation (2.20) is shown

in Figure 5.

10

Figure 5.

Turbine disc.

Substituting the expression for the power coefficient Cp from Equation (2.19) into
Equation (2.17), one gets the following expression for the aerodynamic wind power:
PD = 2 ADV 3

C p ( , )
4

1
= ADC p ( , )V 3 .
2

(0.0)

Substituting the expression for the disc area AD = R2 into Equation (2.21), one
gets:
1
PD = R 2C p ( , )V 3 .
2

(0.0)

From Equation (2.22), one can easily derive an expression for the aerodynamic
torque produced by a turbine of radius R rotating at an angular rate r with wind speed V:
2
3
Pr 1 R C p ( , )V
.
Tr =
=
r 2
r

(0.0)

In Equation (2.23), the substitution of r by its value extracted from Equation

(2.20) provides the following expression:
1
Tr = R 3CQ ( , )V 2
2

where the CQ is named the torque coefficient and is related to Cp by

11

(0.0)

CQ =

CP

(0.0)

In Equations (2.21) and (2.22), it can be seen that CQ and Cp are written as
functions of and . However, it ia assumed in this thesis that the pitch angle =0, and
the above coefficients are used as functions of the ratio only.
A parametric approximation of the torque coefficient CQ is given by the following
expression in terms of [7]
CQ ( ) =

ab
c/
1 e ,

(0.0)

where , b, c are dimensionless parameters that characterize the rotor of the wind turbine.
If one substitutes Equation (2.26) in Equation (2.25), then
b

Cp ( ) = a 1 e c / .

(0.0)

The plot of a typical power coefficient curve versus is shown in Figure 6.

Figure 6.

Typical Cp versus speed ratio curve.

The speed ratio is a very important parameter in the dynamic model of the WT.
Since one wants to operate as closely as possible to maximum efficiency to have
12

maximum power capture and given that the wind speed cannot be controlled and the
radius of the blades is fixed, one needs to vary the rotational speed of the wind turbine to
keep close to its optimal value opt. This is one of the advantages of variable speed WTs
over fixed speed WTs, since they operate close to optimal efficiency only for a small
range of wind speed values.
For a typical wind speed profile with a Weibull distribution having a shape
parameter k = 2 and scale parameter c = 8.5, the variable-speed turbine captures 2.3%
more energy per year than the constant-speed turbine. This is considered a significant
difference in the wind industry [18].
4.

Regions of Control

There are three different operational regions for the variable speed WTs, which
are shown in Figure 7.

Figure 7.

Example power curves for a typical commercial WT. From [5].

13

In region 1, the section of the solid curve, usually located to the left of 4 to 5 m/s
wind speed, the power available in the wind is low compared to the losses in the turbine
system; therefore, the turbine is usually not operating in this region. The use of modern
control strategies is not usually critical in region 1 [5].
Region 2, illustrated by the cubic section of the red curve, is generally for winds
speeds between 4 to 5 m/s up to the WT's rated wind speed, usually between 12 to 15
m/s. In this area, the WT is in operational mode, and it is desirable for it to capture as
much power from the wind as possible. The goal in this region is to approach the Betz
Limit curve as closely as possible.
Region 3 is for operation above rated wind speed of the WT, i.e., the wind speed
above which maximum peak power is produced. The goal of control strategies applied in
this region is to limit the wind power captured so as not to exceed the designed electrical
and mechanical load limits of the system.
All research in this thesis centers on Region 2.
B.

DOUBLY-FED INDUCTION GENERATOR (DFIG)

Induction machines are used in a wide variety of applications and are, without
doubt, the workhorse of the electric power industry [19].
An induction motor is an electric motor that is driven by alternating current (AC).
It consists of two basic parts: the stator, which is on the outside and has windings
supplied with AC to produce a rotating magnetic field, and the rotor, which is attached to
the output shaft on the inside. When balanced multi-phase currents flow through the
stator windings, a rotating air-gap magnetomotive force (MMF) is produced, and, if the
rotor rotates at different speed than the speed of the rotating MMF, balanced multi-phase
currents are induced in the rotor windings. There are two different types of rotor
windings: wound and squirrel cage. In Figure 8 the cross section of a typical three-phase
induction machine with a squirrel cage rotor is illustrated.

14

Figure 8.

Induction machine cross section. From [20].

In single-fed induction generators, the rotor windings are short-circuited, while

DFIGs take full advantage of both stator and rotor excitations. The rotor currents and
voltages of the DFIG are at the basis of the control techniques developed in this thesis.
1.

Reference Frames Dynamic Models

Three phase signals (currents or voltages) can be represented either by the

respective time domain components (called abc) or, more conveniently, by the
expansion with respect to a rotating reference frame (called qd0).
The equations converting time domain "abc" voltages into "qd0" components and
vice versa are very standard and can be found in a number of sources [12], [19].
In the "qd0" reference frame for a symmetrical DFIG, we have the following
equations for the flux linkages from [19]:

15

pqd 0 s = rs iqd 0 s jdq 0 s + vqd 0 s

= rriqd
0 r + j( r )qd 0 r + vqd
0r
pqdor

(0.0)
,

where v, i, refer to voltages, current, and flux linkages in the qd0 reference frame of the
stator (subscript s) and the rotor (subscript r). The parameter is the electrical angular
rate of the reference frame, determined by the line frequency (say 60 Hz in the USA); r
is the electrical frequency of the rotor circuit; rs ,rr are the stator and rotor resistances;
and p stands for the differential operator d/dt.
Any qd0 variable, say f, can be expressed in a complex form as:

fqd 0 = fq jfd + 0 .

(0.0)

Therefore, the flux linkages are defined as:

(dq 0 s )T = [ds

qs

0]

0 r )T = [dr
(dq

qr

0]

(0.0)

where the flux components are related to the currents by standard equations [19]

qs = Lls iqs + LM (iqs + iqr' )

ds = Lls ids + LM (ids + idr' )
0 s = Lls i0 s
qr = L'lr iqr' + LM (iqs + iqr' )

(0.0)

dr = L'lr idr' + LM (ids + idr' )

0 r = L'lr i0' r
Expanding Equations (2.28) and combining them with Equations (2.31), one gets
the following set of equations to characterize the DFIG [19]:

16

rs + X ss
b

X ss
vqs
b

ds

0
v0 s
=
vqr
p
vdr
X

b
v0 r
r

X
b

X
b ss
rs +

X
b
p
X

where

rs +

X
b

X
b

X ss

rr' +

X ls

0
0
0

0
X 'rr

r '

X rr
b
0

b
p
rr' +

b
0

0
'
X rr

X 'rr
rr' +

0
0
p

iqs

ids
i
0s
iqr
i
dr
i0 r

X lr

. (0.0)

X ss = X ls + XM
X rr' = X lr' + X M
X ls = bLls

(0.0)

X lr' = bL'lr
X M = bLM
express all the impedances. In the above equations, b is the base or rated frequency (say
the standard 60 Hz); r is the electrical rotational frequency of the rotor; and e is the
electrical rotational frequency of the stator. The variables Lls and LM are the leakage and
magnetizing inductances of the stator windings, and L'lr is the leakage inductance for the
rotor windings. Also, as mentioned, rs and r'r are the stator and the rotor resistances.
The electromagnetic torque of the generator is given by the following expressions
[19], which are all equivalent:
3 P
Te = LM (iqs idr' ids iqr' ) ,
2 2

(0.0)

3 P
Te = LM (qr' idr' dr' iqr' )
2 2
.
3 P
Te = LM (ds iqs qs ids )
2 2

(0.0)

In the above expressions, P is the number of magnetic poles of the generator, and
Te is positive for motor action and negative for generator action.

17

The total real and reactive power of a three-phase balanced DFIG in steady state
conditions is given by [19]:

3
Pe = (VqsIqs + VdsIds )
2
.
3
Qe = (VqsIds VdsIqs )
2
2.

(0.0)

DFIG Model

Equations (2.28) clearly define a nonlinear dynamic system where the fluxes are
the states and the voltages are the control signals. Equivalently, from [19], one can
redefine the states in terms of the qd0 current components of the stator and the rotor.
The whole system is a nonlinear dynamic system and was implemented by
Professor A. Julian in Simulink [21], as shown in Figure 9. The time domain voltage
components va, vb, vc of the stator and the rotor are explicitly shown, together with the
state variables ia, ib, ic of stator and rotor, in the time domain. In addition, other physical
quantities, like the torque on the load TL, are shown, along with all relevant power
components Pe (active) and Qe (reactive). The model of the DFIG is described in more
detail in [21].

Figure 9.

General view of the DFIG model.

18

This model is used extensively in the next chapters, both as a benchmark for
simulating the dynamics of the DFIG and as the basis of the analytical design of the
proposed algorithms.
C.

CHAPTER SUMMARY
The basic theory behind WTs and DFIGs was presented in this chapter, and a

description of the simulation model for a DFIG was provided. The quality of the model is
of great importance, because the effectiveness of the controller depends on it. In the next
chapter, three different algorithms of online parameter identification will be presented.

19

THIS PAGE INTENTIONALLY LEFT BLANK

20

III.

A.

ONLINE PARAMETER ESTIMATION

INTRODUCTION
The mathematical model described in the Chapter II, with its Simulink

implementation, is representative of the DFIG physical system. In general, the various

parameters defined (resistances and inductances) have to be estimated based on voltage
and the current measurements of the physical system itself. The mathematical model can
be used for control design and validation by computer simulation.
The purpose of this chapter is to describe the methods used to achieve on-line
estimation of the DFIG parameters. The essential idea behind on-line estimation is the
comparison of the observed system response y(t) with the output of a parameterized
model whose structure is the same as that of the plant model [22]. In this case, the output
vector y(t) consists of all the measured currents in the qd0 reference frame under different
conditions of excitation through the input voltages.
The parameter vector (t) contains estimates of all the parameters in the state
space

model

in

Chapter

II,

and

it

is

adjusted

continuously

so

that

y( ,t) approaches y(t) as t increases. Under certain input conditions, y( ,t) being close
to y(t) implies that (t ) is close to the unknown parameter vector (t) of the plant model.
The online estimation procedure, therefore, involves three steps [22]:

Create an appropriate parameterization of the plant model;

Design an algorithm for generating and updating the parameter vector

estimate; and

Properly excite the plant such that the parameter estimates approach the
actual expected values as t .

In general, the plant model of a system to be studied for identification should be

expressed in the following form:

21

y(t ) = (t ) (t )

(0.0)

which can be explicitly written as:

y1 (t ) 11 (t ) 12 (t )
y (t ) (t )
.
2 21
. = .
.

.
. .
yn (t ) n1 (t )
.

(t )
. . . 1m (t ) 1
(t )
. . .
. 2
.
. . .
.
,
.
. . .
.
.
. . . nm (t )

m (t )

(0.0)

where the nx1 vector y(t) consists of the observed variables; the mx1 vector (t) consists
of the unknown parameters to be determined; and the nxm matrix (t) consists of known
functions that describe the plant model, which are called regression variables[23]. In this
thesis, as one can see later on, n=6 and m=5.
It is shown in [22], [23] that under conditions of persistency of excitation, the
estimated parameters converge to the true parameters. Loosely speaking, this is satisfied
when the signal is sufficiently excited so that the input-output behavior can be modeled
by a unique parameter vector.
B.

SYSTEM EQUATIONS
Based on what was discussed in the previous section, one needs to rewrite the

dynamic equations in terms of parameters and regression vectors. In particular, the five
quantities that form the parameter vector to be estimated are the following:

= rs

rr'

Lls

L'lr

LM .

(0.0)

The vector y(t) of observations is defined by the stator and rotor voltages:
y (t ) = v (t ) = v qs

vds

v0 s

vqr

vdr

v 0 r .

Substituting Equations (3.3) and (3.4) in Equation (2.32), one obtains:

22

(0.0)

 iqs

ids
i
v(t ) = 0 s
0
0

0
0

piqs + ids
pids iqs

0
0

p(iqs + iqr' ) + (ids + idr' )

p(ids + idr' ) (iqs + iqr' )

0
iqr'

pi0 s
0

pi + ( r )idr'

idr'
i0' r

0
0

pi ( r )iqr'
pi0' r

0
'
qr
'
dr

0
(t ) , (0.0)
p(iqs + iqr' ) + ( r )(ids + idr' )
p(ids + idr' ) ( r )(iqs + iqr' )

where the matrix on the right of Equation (3.5) becomes (t ) in Equation(3.2):

v(t) = (t) (t) ,

(0.0)

and is the angular velocity of the arbitrary reference frame that was chosen.

C.

ESTIMATORS
In order to estimate the machine parameters vector (t), three different

identification approaches were used:

least mean squares (LMS),

recursive least Squares (RLS), and

recursive least squares with exponential forgetting (RLS-EF).

The next section describes the three different estimators.

1.

Least Mean Squares (LMS)

The first estimation algorithm is based on the Least Mean Squares (LMS) method.
This was initially formulated at the end of the 18th century by the German mathematician
and scientist K. F. Gauss to calculate the orbits of planets and asteroids. From [23], using
the above method, the parameters of the model should be chosen in such a way that the
sum of squares of the differences between the observed and computed values, multiplied
by numbers that measure the degree of precision, is a minimum.

23

a.

Equations Setup

Based on the model described in Equation (3.1), one can define the
prediction or estimation error e(t) [23], as the difference between y$ (t ) and y (t ) :

e(t) = y(t) y$(t ) = y(t ) T (t )$(t ) .

(0.0)

In addition, one can define the parameter error %(t ) as follows:

%(t) = (t) $(t) .

(0.0)

Thus, from Equations (3.1), (3.7) and (3.8), one can derive the following
expression for e(t):

e(t) =(t)%( ) .

(0.0)

From [24], the general form of an identification algorithm is defined by a

differential equation, called the update law, which is of the form:

&%(t ) = F {y(t ), e(t ),%(t ), (t )} .

(0.0)

By the standard steepest descent algorithm, one can obtain the following
expression for the derivative of %(t ) with respect to time:

&%(t ) = (t )e(t ) ,

(0.0)

where is an arbitrarily chosen strictly positive parameter called the adaptation gain and
allows for a variable rate of adaptation of the parameters. The greater the value of , the
faster the convergence, but the response is more sensitive to noise and might induce
oscillations in the system.
To analyze the behavior of this system, one can define a Lyapunov
function:
V ( (t )) =

1 % 2
(t ) .
2

24

(0.0)

According to [25], for %(t ) to converge, the Lyapunov function described

in Equation (3.12) has to be positive definite, and its total derivative with respect to time
must be negative semi-definite. By definition, it is positive definite, since in this case the
parameters are positive real numbers and this case takes the norm of the parameter vector.
The derivative with respect to time of V((t)) is given by:

dV ( (t )) % T %
= (t ) (t ) .
V& ( (t )) =
dt
.

(0.0)

Substituting the expression of %(t ) from Equation (3.11) into Equation

(3.13), one can obtain:
T
V& ( ,t ) = % (t )(t )e(t ) .

(0.0)

Since the prediction error e ( t ) is related to the parameter error by

Equation (3.9), one can obtain:
T
V& ( ,t ) = % (t )(t )T (t )%(t ) = |T (t )%(t )|2 0 .

(0.0)

The outcome of Equation (3.15) is that V& ( ,t ) is negative semi-definite,

and as a result, %(t ) decreases monotonically with time. Under conditions of persistency
of excitation, this implies that the parameter error actually converges to zero as:

(t) $(t) 0 .

(0.0)

Since the system parameter vector (t) is assumed constant or slowly

changing with time, one can assume that &(t ) 0 . Therefore, by differentiating Equation
(3.8) and substituting Equation (3.11), one gets the following expression for the
derivative of the parameter estimates:
.

$(t) = %(t) = (t)e(t) .

(0.0)

Integrating Equation (3.17), one gets the expression needed for the
estimates:

25

$(t ) = $(0) + ( )[y( ) T ( )$( )]d ,

(0.0)

where the initial condition can be arbitrarily chosen based on any a priori knowledge of
the plant.
b.

Implementation

The simulation of this estimation method required the creation of a

Simulink subsystem which was added to the model described in Chapter II for the DFIG.
A general view of the estimator is shown in Figure 10.

Figure 10.

General view of the LMS estimation model.

The Simulink blocks that provide the vectors (t) and V(t) are common
for all the different methods and were implemented using Equations (3.6). These are the
equations for reference frame transformation described in [19]. The estimator block for
the LMS is shown in more detail in Figure 11.
26

Figure 11.

Detailed block diagram for the LMS estimator model.

It can be seen in Figure 11 that instead of using the same adaptation gain
for all the parameters, different values of for the different channels were employed. The
adaptation matrix in Figure 11 is a diagonal matrix of the form:

11
0

_ atrix = 0

0
0

22

0
0

33

44

0
0
0

0
55

The reason behind this is to compensate for different scaling factors and
bring all the signals to a similar order of magnitude. This seems to improve the
convergence rate and prevent instability.
The values of the non-zero terms in the matrix were chosen to create a
relatively fast convergence without affecting the stability of the system. Large values of
the parameters ii lead to fast convergence, but the estimates are oscillating. Very small
values lead to very slow but stable convergence. To illustrate, the values chosen for the
above parameters that gave a stable output are shown in Table 1.

27

Table 1.

2.

Adaptation gains for the LMS simulation.

ii

Value

11

130

22

820

33

0.007

44

0.38

55

0.0135

Recursive Least Squares (RLS)

The second method was recursive least squares (RLS).

a.

Equations Setup

This method is implemented in discrete time and recursively computes the

parameter estimates $(n) by minimizing the quadratic error [23]:
V ( , n) =

1 n
[y (iTs ) (iTs )$ ]2 ,

2 i =1

(0.0)

with Ts as the sampling interval.

From [26], the recursive estimate can be interpreted as a Kalman Filter for
the process:

$(n + 1) = $(n)
y(n) = (n)$(n) + e(n)

(0.0)

where y(n) (n) are given by Equation (3.4) and Equation (3.6), respectively, by
substituting t with n.
The recursive equations for the RLS can then be derived from the Kalman
filter equations used in [27]:
28

$(n + 1) = $(n) + K (n)[y(n + 1) (n + 1)$(n)]

K (n) = P(n)(n + 1)[ (n + 1)P(n)(n + 1) + I]1

(0.0)

P(n + 1) = P(n) K (n)[ (n + 1)P(n)(n + 1) + I]1 K T (n)

>
n

where the matrix (n) has to be full rank, which means that it has to be nonsingular for
all

0 . The initial condition for Equations (3.21) is:

P(0) = 2 mm

(0.0)

where is a positive number, much higher than the parameters, and m is the dimension of
y(t), which in this case is six.
b.

Implementation

The general outlook of the RLS estimator is shown in Figure 12.

Figure 12.

General view of RLS estimator model.

A more detailed view of the inside of the RLS block is shown in Figure
13. The inputs to the block are (n) and v(n), and the output is the parameter vector
estimates $ ( n + 1) .
29

Figure 13.

3.

Detailed view of RLS estimator model.

RLS With Exponential Forgetting (RLS-EF)

In the above estimations, the parameters are assumed constant or slowly changing.
Now consider the case of varying parameters from either environmental or operational
conditions or a change in the parameters due to a fault in the system. The case that is of
more interest is when the parameters are slowly changing in time. In this case, one needs
to design an algorithm with a forgetting factor which discounts data exponentially
further back in time.
a.

Equations Setup

According to [23], the RLS-EF algorithm is obtained by changing the cost

function of Equation (3.19) from the one that was used for the RLS algorithm to the
following:

V ( ,t ) =

1 t t i
[y(i) (i)$(i)]2 ,
2 i =1
30

(0.0)

where is a parameter such that 0 < 1 and is called the forgetting factor. The meaning
of Equation (3.23) is that one weights the last data by 1, with a factor of n for data that
are n time units old. This means that one reduces exponentially the weight of each set of
measurements. Thus, the equation for the RLS with exponential forgetting becomes:

$(t + 1) = $(t ) + K (t )[y(t + 1) (t + 1)$(t )]

b.

K (t ) =

P(t )

P(t + 1) =

P(t )

(t + 1)[ (t + 1)
K (t )[ (t + 1)

P(t )

P(t )

(t + 1) + I]1

(0.0)

(t + 1) + I]1 K T (t )

Implementation

The general outlook of the RLS-EF estimator is similar to that shown in

Figure 12.
D.

SIMULATION OF THE ESTIMATORS

1.

Equipment

The three estimators described in the previous pages were simulated using
Simulink software. A Matlab code was created to initialize the models. The code used for
initialization of the simulation models in Simulink is presented in the Appendix. To
determine how well the estimators perform, one has to choose a DFIG with known
parameters for comparison. The DFIG that was used is the Lab-Volt model 8231, which
is a 175 W, 120 V, 60 Hz, 4-pole machine. The system parameters were used to
effectively model the system in Simulink. Both the stator and rotor of that machine
consist of 3-phase wye-connected windings and a stator to rotor turns ratio Ns/Nr = 10.
To have an input to determine the values of the estimated parameters required use
of the values that were measured in [28] after a DC test, a no load test, and a blocked
rotor test on the generator mentioned above. The results are given in Table 2.

31

Table 2.

2.

Measured DFIG parameters. From [28].

Parameters (t)

Values

rs

12[]

rr '

15[]

Lls

0.0241[]

L'lr

0.0241[]

LM

0.3342[]

Simulation Results

After conducting a large number of simulations, varying the conditions and some
parameters of the system, it turned out that only the first two methods gave satisfactory
results. The simulation using RLS-EF method was leading the estimator to an unstable
output after a period. The next section describes the results of the first two methods.
a.

Simulation Using LMS

The results for the parameter convergence using LMS are shown in
Figures 14, 15 and 16, where the initial condition for the estimator was set to zero. The
sampling time for all the simulations was set equal to 0.1 ms.

32

Figure 14.

Convergence of rs and rr' using LMS from simulation.

Figure 15.

Convergence of Lm using LMS from simulation.

33

Figure 16.

Convergence of Lls and Llr' using LMS from simulation.

One can see that the parameter estimates converge to the expected values,
but it takes a while for them to settle down. This is a result that depends on the non zero
values of the adaptation matrix given in Table 1. If one sets the values of initial condition
vector closer to the measured values, then the convergence is much faster.
b.

Simulation Using RLS

The results of the simulation using RLS are shown in Figures 17, 18 and
19.

34

Figure 17.

Convergence of rs and rr' using RLS from simulation.

Figure 18.

Convergence of Lm using RLS from simulation.

35

Figure 19.

Convergence of Lls and Llr' using RLS from simulation.

From Figures 17 to 19, one can conclude that the parameter convergence
using RLS is almost instantaneous, but for the inductances, the estimates oscillate around
the expected values. On the other hand, the estimates for the resistances are extremely
close to the experimentally measured parameter values.
E.

CHAPTER SUMMARY
In this chapter, different algorithms for system identification were presented

analytically and simulated. Based on the results and given the fact that the nature of the
system is relatively slow, one can say that the LMS is more suitable in this application
since it performs well and provides relatively fast convergence. In the next chapter, a
basic introduction to adaptive control theory and the design and simulation of two
36

different adaptive controller approaches, along with simulation results under different
operational conditions of the system, will be presented. A comparison with simulation
results obtained from classical control approaches will also be presented.

37

THIS PAGE INTENTIONALLY LEFT BLANK

38

IV.

A.

ADAPTIVE CONTROL

INTRODUCTION
According to Webster's dictionary, the definition of adapt is "to change (oneself)

so that one's behavior will conform to new or changed circumstances" [22]. According to
another definition, an adaptive system is any physical system that has been designed with
an adaptive viewpoint [23]. From the above expressions, one can say that an adaptive
controller is a controller that has the ability to change its behavior to adapt to changes in
the systems dynamics. That means that to be adaptive, the controller has to have
adjustable parameters.
Research on adaptive control systems was introduced during the 1950s, mostly for
the design of high performance flight control systems. For many years, there has been
skepticism among scientists about the use of adaptive control, since, in spite of the fact
that the adaptation feature sounds very attractive, its performance and reliability may be
questionable. In particular, when one applies it to a dynamic system, the designer should
be aware that adaptation adds another loop, which might make the overall system
unstable. Some important issues are the following [29]:

Because of internal modulation effects, fast adaptation produces a high

frequency input. This input excites unmodeled dynamics easily and causes
systematic instability. Typical examples from [30] have shown that under
very mild conditions of unmodeled dynamics, an adaptive control system
can easily become unstable.

There is no guarantee of the performance of the closed loop system during

the transient response.

In the last few years, new applications of adaptive control have started to appear,
and some of these issues have been addressed with a degree of success [31].

39

An adaptive control system can be thought as having two loops. One loop is a
normal feedback loop, with the process and the controller. The second is the parameter
adjustment loop.
The system is modeled in terms of a vector of parameters that may be known or
approximately known. The goal of the controller is to continuously produce a vector of
estimates for the above parameters % and control the output vector.
B.

ADAPTIVE CONTROL THEORY

There are many different approaches to adaptive control. The model reference

adaptive control (MRAC) and, more specifically, the direct MRAC (DMRAC), is used in
this thesis. In this control scheme, the system is modeled in terms of a vector of
parameters c that may be known or approximately known. This parameter vector is
updated by an adaptive law using input and output data. A general view of the direct
MRAC is shown in Figure 20.

Figure 20.

Direct MRAC overview. After [31].

As shown in Figure 20, the output of the system z(t) is compared to the output of a
reference model zref(t). The reference model is chosen to generate the desired trajectory
for the plant output to follow [22]. The controller parameters are updated continuously,
depending on the output error:
40

z%(t ) = z(t) zref (t) .

(0.0)

The goal of the controller is to continuously estimate the parameters giving a

vector $(t ) to drive the error of the output vector z%(t) to zero. The rest of this chapter
examines the application of this to the dynamic model of the DFIG.
C.

ADAPTIVE CONTROL OF DFIG

Designing a control scheme required the evaluation of two different control

approaches according to which part of the system should be adaptive:

First approach: adaptive control of torque and reactive power using as

output the rotor currents iqr and idr.

Second approach: adaptive control of torque and PI control of the reactive

power using as output the rotor currents iqr and idr.

These approaches are described in the following sections and the results obtained
from simulation are discussed.
1.

First Approach: Adaptive Control for Torque and Reactive Power

In this approach, the controller was designed using adaptive control for both rotor
currents components, iqr and idr, which form the z(t) vector. The overall design is
described in the following section.
a.

Dynamic Model

The voltage equations of a DFIG expressed in an arbitrary chosen qd0

reference frame were described in Chapter II.
For the purposes of this design, one should assume that the reference
frame is determined by the line voltage of the stator. In this frame, the stator voltage is a
constant with value:

vs (t) = vqs (t) jvds (t) = jVS

41

(0.0)

In addition, one can assume the stator resistance to be negligible; i.e.,

rs = 0 .

(0.0)

Then, defining the rotational frequency of the frame to be the same as

the electrical rotational frequency of the stator e, one can write Equation (2.28) as:

ps = jdqs + jVs .
The solution of Equation (4.4), if the initial condition is s (0) =

(0.0)
Vs

, is a

constant given by

s (t ) =

Vs

(0.0)

Since the flux is complex s (t ) = qs (t ) jds (t ) , this yields

qs (t ) =

Vs

, cons tant

(0.0)

ds (t ) = 0
The fact that the flux is a constant in this reference frame leads to the
definition of a constant magnetizing current ims. The goal of the rest of this section is to
rewrite the dynamic model of the DFIG in terms of the rotor current components iqr and
idr and the magnetizing current ims only.
Therefore, using Equations (2.31) from Chapter II relating the fluxes with
the currents, one can obtain:

LM is
s Lls + LM
.
= L
Llr + LM ir
r M

Substituting Equations (4.6) in (4.7), one gets:

42

(0.0)

(Lls + LM )iqs + LM iqr =

Vs

(0.0)

(Lls + LM )ids + LM idr = 0

Solving Equation (4.8b) for ids , one gets:
ids =

LM
idr .
Lls + LM

(0.0)

Now defining the constant magnetizing current ims as

LM ims = qs ims =

qs
LM

ims =

Vs
,
LM

(0.0)

and substituting Equation (4.10) in Equation (4.8a) and solving for iqs, one obtains:

LM ims =

Vs

(Lls + LM )iqs + LM iqr

L
L
iqs = M ims M iqr
Lls + LM
Lls + LM

(0.0)

Now one can easily rewrite all the equations in terms of ims and i'r.
Substituting Equations (4.9), (4.10) and (4.11) in Equations (4.7) and expanding the
expressions for qr and dr , one obtains:

qr = LM iqs + (Llr + LM )iqr

qr = LM (
qr =

LM
L
ims M iqr ) + (Llr + LM )iqr
Lls + LM
Lls + LM

(0.0)

LM 2
LM 2
ims + (1
)(Llr + LM )iqr
Lls + LM
(Lls + LM )(Llr + LM )

and

dr = LM ids + (Llr + LM )idr

dr = LM (
dr = (1

LM
)idr + (Llr + LM )idr
Lls + LM

LM 2
)(Llr + LM )idr
(Lls + LM )(Llr + LM )

43

(0.0)

The flux linkages equations can then be summarized from Equations (4.6),
(4.12) and (4.13) as:

qs =

Vs

ds = 0

L 2
qr = M ims + (Llr + LM )iqr
Lls + LM
dr = (Llr + LM )idr

(0.0)

where is defined as a real parameter given by the following expression:

=1

LM 2
,0 < < 1 .
(Lls + LM )(Llr + LM )

(0.0)

In most cases, LM is significantly larger than Lls and Llr, which means that
is very small.
From Equations (2.28) in Chapter II, after ignoring the 0 components,
which are negligible and defining

r = % ,

(0.0)

one gets:

% dr + vqr
pqr = rr iqr
% qr + vdr
pdr = rr iqr +

(0.0)

Substituting Equations (4.14) in Equations (4.17), one finally obtains the

desired dynamic model of DFIG in terms of magnetizing and the rotor currents:
piqr =

rr
1
iqr % idr +
v
(Lr + LM )
(Lr + LM ) qr

rr
LM 2
1
pidr =
idr + % iqr + %
ims +
v
(Lr + LM )
(Lr + LM )(Ls + LM )
(Lr + LM ) dr

Now, define 0,1,b the following parameters:

44

(0.0)

a0 =

rr
(Lr + LM )

LM 2
i .
(Lr + LM )(Ls + LM ) ms
1
b=
(Lr + LM )

a1 =

(0.0)

Substituting Equations (4.19) in Equations (4.18), one ends up with two

simple first order models:

piqr = a0 iqr % idr + bvqr

pidr = a0idr + % iqr + % a1 + bvdr

(0.0)

iqr
vqr
By defining the state vector z = and the control vector v = and
vdr
idr
choosing two arbitrary positive constants m, m, one can rewrite Equations (4.20) in
terms of the state vector and a desired dynamic model:

idr
z&(t ) = am z(t ) + m v(t ) + (t ) + % ,
iqr

(0.0)

with
i
(t ) = qr
idr

0 vqr
% vdr

(0.0)

0 + m
= 1
b m

The significance of Equation (4.22) is that by defining the control signal as

follows

idr

(t) (t ) + W ref (t)

i
qr

m v(t) = % (t)

45

(0.0)

with (t ) an estimate of the system parameter vector (t) , the closed loop system
becomes

z&(t ) = m z(t ) + (t )%(t ) + W ref (t )

(0.0)

where

%(t) = $(t) .

(0.0)

In the ideal case of perfect knowledge of the DFIG dynamics, the

parameter error vector %(t ) in Equation (4.25) is zero, and the state vector z ( t ) tracks the
reference model:

z& ref (t ) = m z ref (t) + W ref (t)

(0.0)

The tracking error z%(t) is defined as:

z% (t ) = z(t ) zref (t ) .

(0.0)

Then, combining Equations (4.24) and (4.26) and substituting in Equation

(4.27), one ends up with the following expression:
.

z%(t ) = m z% (t ) + (t )%(t ) .

(0.0)

Now, estimating the controller parameters requires the development of a

differential equation, called the update law. That differential equation is of the form:
.

%(t ) = F {y(t ), e(t ),%(t ), (t )} .

(0.0)

Using the steepest descent algorithm, we ensure that the parameter error
.

decreases monotonically with time. In particular the definition of (t ) is as follows:

&

(t ) = (t )z% (t),

(0.0)

where >0 is the adaptation gain of the controller. The definition of the following
Lyapunov function is:
46

V ( ,t ) =

1 % 2 1 % 2
(t ) .
z(t ) +
2
2

(0.0)

As also stated in Chapter III, for the system to be stable, the Lyapunov
function described in Equation (4.31) has to be positive definite, and its derivative with
respect to time must be negative semi-definite. By definition, it is positive definite, since
in this case the parameters are positive real numbers and the norm of the parameter
vectors are used. The derivative with respect to time of V ( (t )) is
.

1 T .
V& ( (t )) = z% (t ) z% (t ) + % (t )%(t ) ,

(0.0)

where, since &(t ) 0 , one can use the fact that:

.

&(t ) = %(t ) .

(0.0)

In Equation (4.32), if one substitutes Equations (4.28) and (4.30), one gets:
2

V& ( ,t ) = z% (t )( m z% (t ) + (t )%(t )) ((t )z% (t ))T %(t ) = m z% (t ) .

(0.0)

From Equation (4.34), V& ( ,t ) is negative semi-definite for every value of

time t.
Equation (4.34) implies that:

V (t ) V (0)t
1 % 2 1 % 2 1 % 2 1 % 2
(t ) z(0) +
(0)
z(t ) +
2
2
2
2

(0.0)

which means, that if one chooses

z(0) = z(0) z%(0) = 0 ,

And from Equation (4.35)

47

(0.0)

z% (t )

%(0)

(0.0)

Equation (4.37) means that z%(t) is bounded.

b.

Simulation Model

A Simulink model for the controller was created using the above equations
for the system. The inputs to the controller are the following quantities, all expressed in
the abc reference frame: stator voltages Vabcs, stator and rotor currents iabcs and iabcr
respectively, r and the reference currents iqr_ref and idr_ref. The outputs of the controller
are the rotor voltages that go into the DFIG model. A general block diagram of the
controller is shown in Figure 21.

Figure 21.

General view of the first controller in Simulink.

48

This model was created using Equations (4.18) to (4.37). A detailed view
of the controller block, which gives the values of the control vector vqr

vdr according

to Equation (4.23), is shown in Figure 22. The output is then transformed back to the abc
reference frame to supply the rotor with the proper control voltage.

Figure 22.

Control block for the first approach.

The reference model shown in Figure 21 has a first order transfer function
of the form

T (s) =

am
.
s + am

(0.0)

The control goal is to independently control the rotor currents to minimize

the reactive power of the DFIG and to produce the maximum real power for every
different value of the wind speed respectively.
From Equation (2.36), after substituting the values of Vqs and Vds from
Equation (4.2), one can obtain the following expression for the reactive power:
3
Qe = Vs iqs ,
2

(0.0)

which means that Qe is proportional to iqs, so to have Qe = 0 , iqs has to be equal to zero;
i.e.,

49

i q s _ re f = 0 .

(0.0)

From equation (4.11), substituting the above value for iqs _ ref from
Equation (4.40), one gets:

Iqr _ ref = ims ,

(0.0)

which means that iqr should be equal to the magnetizing current, which is constant after
the system reaches steady-state, and is given by Equation (4.10).
On the other hand, keeping the power output maximum requires that the
electric torque of the DFIG be equal to the aerodynamic torque that gives the maximum
power output of the WT for each wind speed, which is given by [12]
Te _ ref = Taero _ optimal = K opt r 2 ,

(0.0)

C ( )
1
K opt = R 5 p 3 opt
opt
2

(0.0)

where

and r is the rotational speed of the WT in revolutions per minute (RPM). This is given
by

r =

1
2
r ,
Gear _ Ratio P

(0.0)

where Gear_Ratio is the gear box ratio, P is the number of poles of the DFIG, and r is
the electrical rotational speed of the rotor in radians per second.
The expression of the electric torque output of the DFIG is given by
Equation (2.35). Substituting the values of the flux linkages from Equation (4.6), one gets
3 P V
Te = LM ( s ids ) ,
2 2 b

and by substituting ids from Equation (4.9) into Equation (4.45), one obtains
50

(0.0)

L
3 P V
Te = LM s ( M idr ) .
Lls + LM
2 2 b

(0.0)

The goal is to track a reference torque, which requires Te = Te _ ref . That

means that from Equations (4.44) and (4.46) one should set the reference current idr that
is driving the reference model to be

idr _ ref =

Te _ ref
LM Vs
3 P
LM
2 2 Lls + LM b

(0.0)

Running the model required the inclusion of a block that takes the current
wind speed as input and gives the instant aerodynamic torque as output, which is the
excitation of the DFIG. The WT turbine speed is then given by

& = T T ,
J
r
aero
e

(0.0)

where J is the combined inertia of the WT and the DFIG rotor. In this thesis, the rotor
inertia was neglected due to its very small value compared to the inertia of the WT.
The comparison of the control approaches presented here with other
controllers required the use of the parameters of the DFIG and the WT described in [7],
which are shown in Table 3.

Table 3.

WT and DFIG parameters. From [7].

DFIG
Pmax=50 hp=37.285 kW
P=4
Rs=0.082
Rr=0.228
Ls=0.0355 Hy
Lr=0.0355 Hy
LM=0.0347 Hy
b=_=260 rad/s
Vs =380 V

Wind Turbine
R =3.8 m
J =3.362 kgm2
a=19.346
b=9.4117
c=20
Gear_Ratio 16:1
opt=6.4

51

c.

Simulation Results

After a large number of simulations, the system parameters that gave the
best results are shown in Table 4. The sampling time for the simulation was chosen equal
to 0.8 ms. For these values, the system was stable and tracked the reference model for
wind speeds from 3.2 m/s to 10.8 m/s, which is a good range. This covered almost all the
operational range of the WT.

Table 4.

Final simulation parameter values for first approach.

Value
100
250
0.75
50
0.4

Parameter
am
bm
11
22
33

In Table 4, it can be seen that the adaptation gain for the parameter
estimation was chosen to be a diagonal matrix as seen in Chapter III to improve the
convergence of the estimates. The initial condition for the estimates was chosen equal to
70% of the expected values calculated using Equation (4.22).
The wind speed profile used is shown in Figure 23 along with a plot of the
absolute values of the actual output torque Te and the reference torque Tref. The wind
speed profile was chosen in order to prove the performance of the controller even for
large rapid changes in wind speed, greater than 5 m/s, which can be considered wind
gusts according to [33].
The output torque is tracking the reference value for a range of wind
speeds from 3.2 m/s to 10.7 m/s as shown in Figure 23.

52

Figure 23.

Wind speed and torque during simulation of the first approach.

Zooming in on the torque plot from time 90 s to 120 s in Figure 24, one
can see that even for wind changes of 6.8 m/s, the difference between reference and
actual output is less than 1.65 %. For lower wind speeds, the difference is always less
than 0.7 %, which is a very satisfactory result.

Figure 24.

Magnified torque plot during simulation of the first approach.

The plots of the real power Pe and reactive power Qe during the simulation are
shown in Figures 25 and 26, respectively.

53

Figure 25.

Figure 26.

Real power (Pe) during simulation of the first approach.

Reactive power (Qe) during simulation of the first approach.

In Figures 25 and 26, one can see that the ratio of Qe over Pe is very small
and remains less than 0.2% for most the simulation time. This result indicates that the
second goal of the design is also satisfied. The only time that Qe increases, but still with
values less than 1% of Pe, is during extreme changes in wind speed, and that happens for
a very short period.
The rotor currents during the simulation are shown in Figure 27.

54

Figure 27.

Comparison of the rotor currents iqr and idr with their reference values
during simulation of the first approach.

2.

Second Approach: Adaptive Control of Torque, PI Control of the

Reactive Power

A second approach was attempted to simplify the controller with only one
adaptive loop for torque control. The difference in this approach is that instead of
controlling both iqs and idr adaptively, a PI controller was used for iqs and the adaptive
control for idr was kept.
a.

Dynamic Model of the Controller

Equations (4.2) to (4.18) are valid for this approach. From Equation
(4.11), solving for iqr, one gets

iqr = ims

LM + Lls
iqs .
LM

Substituting Equation (4.49) for the last one of Equations (4.18),

one obtains
55

(0.0)

pidr =

L +L
rr
LM 2
1
idr + % (ims M ls iqs ) + %
i +
v
LM
(Llr + LM )
(Llr + LM )(Lls + LM ) ms (Llr + LM ) dr

. (0.0)
pidr =

L +L
rr
LM 2
1
idr % ( M ls )iqs + % (1
v
)ims +
LM
(Llr + LM )
(Llr + LM )(Lls + LM )
(Llr + LM ) dr

One defines the following parameters:

rr
(Llr + LM )
L +L
a1 = ls M
LM
a0 =

LM 2
a2 = (1
)i
(Llr + LM )(Lls + LM ) ms
1
b=
(Llr + LM )

(0.0)

Substituting Equations (4.51) in Equation (4.50), one gets

pidr = a0 idr a1 (% iqs ) + a2% + bvdr .

Now zr = idr

(0.0)

and v = vdr are called the state and control signal,

respectively, which in this approach are scalars. Then, for any arbitrary constant m, m
one can write Equation (4.52) in the form

z&(t) = am z(t) + m v(t) + (t) .

(0.0)

where

(t ) = idr

% iqs % vdr

(0.0)

and
0 + m
a

1
.
=
a2

b m

In this case, the control signal is

56

(0.0)

 m v(t ) = (t ) + W ref (t ) .

(0.0)

Now substitute Equation (4.56) into Equation (4.53), and one gets:

z&(t ) = m z(t ) + (t )%(t ) + W ref (t ) ,

(0.0)

where %(t ) is given by Equation (4.25).

b.

Simulation Model

The general view of the controller in Simulink is shown in Figure 28.

Figure 28.

General view of the second controller model.

Equations (4.30) to (4.37), which were used in the first approach for the
parameter estimates, were also used for this case.
For iqs, a PI controller was used with the transfer function
TPI (s) =

K1 s + K 0
,
s

57

(0.0)

where K0, K1 are the gains of the PI controller. The input of this controller, the error
between the reference and the actual measured value of iqs and the output is vqr.
Furthermore, Equations (4.42) to (4.48) that give the expression for idr_ref
and r were again used here.
In this approach, the output of the adaptive controller is vdr according to
Equation (4.57); two low pass (LP) filters were added in the output of the adaptive part to
filter higher frequency components that might affect the stability of the system. A
detailed view of the control block diagram is shown in Figure 29.

Figure 29.

Control block for the second approach.

The cutoff frequency ac in the LP filter shown in Figure 29 was 100, and
for the second filter, which is a 12th order Butterworth filter, the cutoff frequency was 10
times higher.
c.

Simulation Results

After a large number of simulations, the system parameters that gave the
best results for this approach are those shown in Table 5. The sampling time for the
58

simulation was equal to 0.8 ms. For these values, the system was stable and tracked the
reference model for wind speeds from 3.2 m/s to 10.8 m/s.

Table 5.

Final simulation parameter values for the second approach.

Value
100
250
100

Parameter
am
bm
11

In Table 5, it can be seen that the adaptation gain for the parameter
estimation in this case was a single value, which is about two orders of magnitude larger
than the adaptation gains for the first approach given in Table 4. This value was chosen
because comparable results could not be achieved between the two control approaches
with smaller values of . The initial condition for the estimates was chosen equal to 70%
of the expected values calculated using Equations (4.56).
The wind speed profile is the same as the one used in the previous
approach, which allows for a comparison between approaches in the next chapter. This
plot, along with a plot of the absolute actual output torque Te and the reference (optimal)
torque Tref , is shown in Figure 30.

Figure 30.

Wind speed and torque during simulation of the second approach.

59

From Figure 30, it can be seen that the output torque tracks the reference
value pretty well for a range of wind speeds from 3.2 m/s to 10.7 m/s, but during the
sudden wind changes, the actual output torque spikes for a few moments. A zoomed in
view of the torque plot from time 97 s to 119 s is illustrated in Figure 31.

Figure 31.

Magnified torque during simulation of the second approach.

It can be seen in Figure 31 that for large and fast wind changes that
simulate wind gusts, the torque tends to lose tracking for few moments. This is especially
true when the wind decreases, e.g. at 100 s or at 117 s of simulation time, where the
difference between the two values even reaches 100% but recovers quickly.
The plots of the real power Pe and reactive power Qe during the simulation
of the second approach are shown in Figures 32 and 33, respectively.

Figure 32.

Real power (Pe) during simulation of the second approach.

60

Figure 33.

Reactive power (Qe) during simulation of the second approach.

It is can be seen in Figure 33 that Qe is quite high compared to Pe. This is

clearer if one zooms in on a part of the simulation, as shown in Figure 34.

Figure 34.

Comparison between real and reactive power Pe and Qe, respectively, for
part of the simulation during the second approach.
Finally, the plots of the currents iqs and idr are shown in Figure 35.

61

Figure 35.

Comparison of the currents iqs and idr with their reference values during
simulation of the second approach.

It can be seen in Figure 35 that idr is tracking its reference value very well,
but iqs seems to lose tracking during sudden wind speed changes.
3.

Performance Comparison

To be able to reach a conclusion on the performance of the controllers designed in

this study, a Simulink model of a classical approach for the control of a WECS was
created. This is described in [12], again using the parameters of the system given in [7].
The model was created for [33].
Two different fault cases were simulated to compare the robustness of the control
algorithms. These two are:

a wind gust, i.e., a sudden increase in the wind speed from 6 to 10 m/sec,

doubling of the value of the rotor resistance rr to demonstrate what

happens during a parameter variation.

62

The wind profile shown in Figure 36 was assumed.

Figure 36.

Wind profile for comparison of the control approaches.

The output torque of the three different control approaches during simulation is
shown in Figure 37.

Figure 37.

Comparison of output torque Te and reference torque Tref during the

simulation of a wind gust.

Magnifying Figure 37 for about 10 s during the top of the wind gust provides
Figure 38.

63

Figure 38.

Magnified plot of output torque Te and reference torque Tref during the
simulation of a wind gust.

From Figure 38, one can see that even though the performance of all the
approaches considered is very good, the output torque of the first approach and the
classical approach with the PI controllers are identical and almost identical, respectively,
to the reference torque throughout the simulation, while the second approach shows some
oscillation during sudden wind speed changes. Furthermore, the torque of the classical
approach is slightly closer to the reference value, but has more overshoot compared to the
first approach.
The resistance of the rotor is a parameter that affects the shape of the torque
output of the DFIG [19] and is illustrated in Figure 39.

64

Figure 39.

Typical DFIG torque versus time curve for different values of rr.

For the purposes of this thesis, in order to investigate the performance of the
controllers that we proposed in previous sections, the rotors resistance rr was increased
to twice its initial value at t=15 s of the simulation, and all the control approaches where
simulated. The output torque plots during the above change in rr from the simulations are
shown in Figure 40.

Figure 40.

Comparison of the variation of the output torque curves during simulation

of a variation of rr.
65

As seen in Figure 40, Te for the second approach has a high overshoot and even
after three seconds does not completely track the reference torque Tref. Between the other
two cases, the first approach appears to have the best performance since it has a small
spike in Te but recovers faster than the PID approach. This can also be seen in Figure 41,
where the speed of the DFIG for the three approaches is presented around the step change
in rr; it is clear that the first adaptive approach gives the smoother speed curve among the
three different cases.

Figure 41.

D.

Comparison of the variation of the DFIG speed during the simulation of a

variation of rr.

CHAPTER SUMMARY
In this chapter, the analytical design, modeling and simulation of two MRAC

approaches for a DFIG-based WECS were presented. The simulation results were
compared with results from simulation of a more classical control approach for this kind
of system, which was presented in [12]. The conclusions from these comparisons along
with recommendations for future research are discussed in Chapter V.

66

V.

A.

CONCLUSIONS AND RECOMMENDATIONS

SUMMARY - CONCLUSIONS
The problem of the control of a doubly-fed induction generator based wind energy

conversion system with fixed pitch horizontal axis wind turbine was addressed in this
thesis. The adaptive control used was direct model reference adaptive control. This
research effort addressed issues relative to the identification of the systems parameters
and the design of adaptive control techniques in order to compensate for uncertain or
time-varying dynamics.
In particular three system identification algorithms were presented based on
standard techniques such as least mean squares (LMS) or recursive least squares (RLS)
with and without exponential forgetting. The significance of these techniques is that a
dynamic system of the DFIG can be properly parameterized on the basis of experimental
results.
Furthermore, two different adaptive control approaches were introduced and
simulated for application in wind power generation. In this case it turns out that the
parameterized dynamic model, based on the magnetic fluxes of stator and rotor, can be
simplified considerably and only a few parameters have to be estimated. This is due to
the stator being connected to the grid so that the stator voltage is constant in magnitude,
frequency and phase. Given this characteristic, the magnetizing current becomes one of
the parameters to be estimated, which results in considerable simplification of the
dynamic model.
These approaches have been tested in numerous simulations and seem to present
an attractive solution to the problem of compensating for uncertain dynamics such as
varying resistances and impedances in the circuitry.

67

B.

RECOMMENDATIONS-FUTURE RESEARCH
Extensive computer simulations demonstrate that the algorithms presented for

adaptive control work as expected. Next step in this research will be the implementation
of the proposed controllers in a laboratory environment using a direct current motor to
simulate the wind turbine and building the controllers using Xilinx software in Simulink
and field programmable gate arrays to establish bidirectional communication between the
software and the equipment.
This effort will involve not only actual real-time implementation but also the
design of an appropriate model for the wind profile so that wind effects can be simulated
by the generation of appropriate torque by the DC motor. The ultimate benefit of this
research will be more efficient wind power generators.

68

APPENDIX:

MATLAB CODES - SIMULINK MODEL

In this Appendix, the code that was used for the initialization of the simulations,
as well as the basic block of the Simulink models for the controllers are presented.
A.

MATLAB CODE FOR INITIALIZATION OF THE SIMULATIONS

%% INITIAL CONDITIONS FILE

clc;
clear all;
omega_b = 2*pi*60;
tstep=.0005;
tstop=120;
twopiby3 = 2*pi/3;
poles = 4;
J=3.362;
% kg m^2
polesby2J = poles/2/J;
GearRatio=16;
%Parameters from paper on nonlinear control
rr0 =0.228; % ohms
rs=0.02;
% ohms
Xls=omega_b*(0.0355-0.0347); % ohms
Xm =omega_b*0.0347;
% ohms
Xlr =omega_b*(0.0355-0.0347);
% ohms
Vs=380; % Volts Grid phase voltage
omega_e=omega_b; %electrical frequency
%Parameters - Equivalent circuit
Ls=Xls/omega_b;
Lm=Xm/omega_b;
Lr=Xlr/omega_b;
rsbyXls = rs/Xls;
rrbyXlr = rr0/Xlr;
Xaq = 1/(1/Xm+1/Xls+1/Xlr);
Xad = Xaq;
XaqbyXls = Xaq/Xls;
XaqbyXlr = Xaq/Xlr;
XadbyXls = Xad/Xls;
XadbyXlr = Xad/Xlr;
V_phase = 380*sqrt(2)/sqrt(3);
model

%Peak value used in Simulink for source

69

%Steady state analysis

omegar_ic = omega_b*.85;
slip = 0.001;
%start at synchronous speed
Z1 = rs + j*Xls;
Z2 = j*Xm;
Z3 = rr0/slip + j*Xlr;
Ias = V_phase/sqrt(2)/(Z1 + Z2*Z3/(Z2+Z3));
Iar = -Ias*(Z2*Z3/(Z2+Z3))/Z3;
Te = 3*poles/2*Xm/omega_b*real(j*conj(Ias)*Iar);
Iqse = sqrt(2)*real(Ias);
Idse = -sqrt(2)*imag(Ias);
Iqre = sqrt(2)*real(Iar);
Idre = -sqrt(2)*imag(Iar);
psi_mqe = Xm*(Iqse+Iqre);
psi_mde = Xm*(Idse+Idre);
psi_qse = Iqse*Xls + psi_mqe;
psi_dse = Idse*Xls + psi_mde;
psi_qre = Iqre*Xlr + psi_mqe;
psi_dre = Idre*Xlr + psi_mde;
psi_qsic=psi_qse;
psi_dsic=psi_dse;
psi_qric=psi_qre;
psi_dric=psi_dre;
%WIND
rho=1.225; % air density
lam=0.1:.1:10;
C=(19.346./lam).*(9.4117./lam -1).*exp(-20./lam);
% plot(lam, C)
lam_opt=6.4;
C_opt=C(64);
K_opt=rho*pi*(3.8^5)*C_opt/(2*lam_opt^2);
lamda=0.8;
%MRAC parameters
am=100.0; % this works
bm=250;
% Filter frequency
ac=500;
sigma=1-(Lm^2)/((Ls+Lm)*(Lr+Lm));
b=1/(sigma*(Lr+Lm));
a0=-rr/(sigma*(Lr+Lm));
c=50;
% controller parameters
im0=Vs/(omega_e*Lm);
% estimated magnetizing current
% true values
theta(1)=a0+am;
theta(2)=Lm/(Ls+Lm);
theta(3)=-im0*a0*Lm/(Ls+Lm);
theta(4)=b*Lm/(Ls+Lm) - bm;

70

theta(5)=(Ls+Lm)/Lm;
theta(6)=(1+(Lm^2)/(sigma*(Ls+Lm)*(Lr+Lm)))*im0;
theta(7)=b-bm;
% Estimation bounds
rho_theta=50; % parameter error in "percent"
d_theta=abs(theta)*rho_theta/100;
theta_min=theta-d_theta;
theta_max=theta+d_theta;
% initial estimates
theta0=theta+d_theta.*(2*rand(1,7)-1);

B.

SIMULINK BLOCKS
1.

First Adaptive Control Approach

a.

Overall System View

71

b.

Aerodynamic Torque Input

c.

General View of the Controller

72

d.

Preprocess

e.

Parameter Estimation

73

f.

2.

Control

Second Approach

The above Simulink blocks, in sections

a, b, c, e, are common for both

approaches. Those which are different for the second approach follow.
a.

Preprocess

74

b.

Control

PID Controller

iqs_error

vqdr

Matrix
M ultiply
Analog
Filter Design
theta
Phi_trans(t)theta_ c_ hat(t)

bm'Vqdor

LP FILTERIN G

75

THIS PAGE INTENTIONALLY LEFT BLANK

76

LIST OF REFERENCES
[1]

U.S. Energy Information Administration (EIA),International Energy Outlook

2010 [Online]. Available: http://www.eia.gov/oiaf/ieo/electricity.html (accessed
August 5, 2010).

[2]

F. Birol, Power to the People, The World Outlook for Electricity Investment,
International Atomic Energy Agency Bulletin, Bulletin 46/1, pp. 912, June 2004.

[3]

REN21, Renewables 2011 Global Status Report, Technical Report, REN21

Secretariat, Paris, 2011.

[4]

U.S. Department of Energy, Office of Energy Efficiency and Renewable Energy,

20% Wind Energy by 2030, Technical Report No. DOE/GO-102008-2567, U.S.
DOE, Washington, DC, 2008.

[5]

K. E. Johnson, Adaptive Torque Control of Variable Speed Wind Turbines,

Technical Report No NREL/TP-500-36265, National Renewable Energy
Laboratory, Golden, Colorado, August 2004.

[6]

J. I. Jang, Y. S. Kim, and D. C. Lee, Active and Reactive Power Control of

DFIG for Wind Energy Conversion Under Unbalanced Grid Voltage, Presented
at Power Electronics and Motion Control Conference, 2006. IPEMC 2006.
CES/IEEE 5th International, Shanghai, August 2006.

[7]

F. Valenciaga, P. F. Puleston, and S. K. Spurgeon, A Geometric Approach for

the Design of MIMO Sliding Controllers, Application to a Wind-driven Doubly
Fed Induction Generator, Int. J. of Robust and Nonlinear Control, vol. 19, issue
1, pp. 2239, January 2009.

[8]

V. Utkin, J. Guldner, and Shi J., Sliding Mode Control Theory in

Electromechanical Systems, Taylor & Francis, London, 1999.

[9]

A. Sabanovic, L. Fridman, and S. K. Spurgeo, Variable Structure Systems: from

Principles to Implementation, IEEE Book Series. IEEE, London, 2004.

[10]

S. A. Frost, M. J. Balas, and A. D. Wright, Direct Adaptive Control of a Utilityscale Wind Turbine for Speed Regulation, Int. J. of Robust and Nonlinear
Control, vol.19, issue 1, pp. 5971, January 2009.

[11]

A. Martinez, F. Lescher, and J. Y. Zhao, An Adaptive Control for Wind

Turbine, presented, at the International Conference on Renewable Energies and
Power Quality 05, Zaragoza, Spain, March 2005.

77

[12]

R. Pena, J. C. Clare, and G. M. Asher, Doubly Fed Induction Generator Using

Back-to-Back PWM Converters and its Application to Variable Speed Windenergy Generation, IEE Proc.-Electr. Power Appl., vol. 143, no. 3, pp. 231241,
May 1996.

[13]

S. K Salman and B. Badrzadeh, New Approach for Modeling Doubly-Fed

Induction Generator (DFIG) for Grid-Connection Studies, presented at 2004
European Wind Energy and Exhibition, London, November 2004.

[14]

F. D. Bianchi, H. De Battista, and R. J. Mantz, Wind Turbine Control Systems, pp.

228, Springer, London, 2010.

[15]

A. I. Munteanu, A. I. Bratcu, N-A Cutululis, and E. Ceang, Optimal Control of

Wind Energy Systems Towards a Global Approach, Springer, London, 2008.

[16]

M. O. L. Hansen, Aerodynamics of Wind Turbines, Second Edition, Earthscan,

London, 2008.

[17]

A. Betz, Introduction to the Theory of Flow Machines, Permagon Press, Oxford,

1966.

[18]

J. F. Manwell, J. G. McGowan, and A. L. Rogers, Wind Energy Explained:

Theory, Design, and Application, John Wiley and Sons Ltd., West Sussex,
England, 2002.

[19]

P. C. Krause, O. Wasynczuk, and S. D. Sudhoff, Analysis of Electric Machinery

and Drive Systems, Second Edition, John Wiley & Sons, 2002.

[20]

R. De Doncker, D. W. J. Pulle, and A. Veltman, Advanced Electrical Drives,

Analysis, Modelling, Control, Springer, London, 2011.

[21]

A. L. Julian, Notes for EC4130 (Advanced Electrical Machinery Systems), NPS,

2010 (unpublished).

[22]

P. Ioannou and J. Sun, Robust Adaptive Control, Prentice Hall, 1996.

[23]

K. J. Astrom and B. Wittenmark, Adaptive Control, Second Edition, Dover

Publications, New York, 2008.

[24]

S. Sastry and M. Bodson, Adaptive Control, Stability, Convergence, and

Robustness, Prentice Hall, New Jersey, 1989.

[25]

J. J. E. Slotline and W. Li, Applied Nonlinear Control, Prentice Hall, New Jersey,
1991.

78

[26]

K. J. Astrom and B. Wittenmark, Computer Controlled Systems, Theory and

Design, Third Edition, Prentice Hall, New Jersey, 1997.

[27]

R. G. Brown and P. Y.C. Hwang, Introduction to Random Signals and Applied

Kalman Filtering, Third Edition, Wiley, New York, 1997.

[28]

J. G. Massey, Doubly Fed Induction Machine Control for Wind Energy

Conservation, M.S. Thesis, NPS, Monterey, CA, 2009.

[29]

J. H. Yang, Robust Adaptive Control Using a Filtering Action, PhD

dissertation, NPS, Monterey, CA, 2009.

[30]

C. Rohrs, L. Valavani, and M. Athans, and G. Stein, Robustness of Continuous

Time Adaptive Control Algorithms in the Presence of Unmodeled Dynamics,
IEEE Transactions on Automatic Control, vol. 30, no. 9, pp. 881889, September
1985.

[31]

N. Hovakimyan and C. Cao, L1 Adaptive Control Theory, Society for Industrial

and Applied Mathematics, Philadelphia, 2010.

[32]

National Oceanic and Atmospheric Administration, National Weather Service,

Definition of Wind Gust. [Online]. Available:
http://www.weather.com/glossary/g.html (accessed August 9, 2011).

[33]

R. Cristi, Notes for EC4300 (Advanced Topics in Control Systems: Wind Turbine
Control), NPS, 2010 (unpublished).

79

THIS PAGE INTENTIONALLY LEFT BLANK

80

INITIAL DISTRIBUTION LIST

1.

Defense Technical Information Center

Ft. Belvoir, Virginia

2.

Dudley Knox Library

Naval Postgraduate School
Monterey, California

3.

Dr. Clark Robertson

Electrical Engineering and Computer Department
Code EC/R1
Naval Postgraduate School
Monterey, California

4.

Dr. Roberto Cristi

Code EC/C1
Naval Postgraduate School
Monterey, California

5.

Dr. Alexander Julian

Electrical Engineering and Computer Department
Code EC/J1
Naval Postgraduate School
Monterey, California

6.

Dr. Xiaoping Yun

Electrical Engineering and Computer Department
Code EC/Yx
Naval Postgraduate School
Monterey, California

7.

Embassy of Greece
Office of Naval Attach
Washington, District of Columbia

8.

LT Stamatios Orfanos - Pepainas

Hellenic Navy General Staff
Athens, Greece

81