DISS. ETH NO.

20217

MODELING AND MULTI-OBJECTIVE

OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

A dissertation submitted to
ETH ZURICH

for the degree of

Doctor of Sciences

presented by
JONAS MÜHLETHALER
M.Sc. ETH Zürich
born 16. September 1982
citizen of Lucerne, Switzerland

accepted on the recommendation of

Prof. Dr. Johann W. Kolar, examiner
Prof. Dr. Charles R. Sullivan, co-examiner

2012
Acknowledgments

First of all, I would like to thank Prof. Johann W. Kolar for giving me
the opportunity to do my Ph.D. thesis in his group. The environment
he established in his laboratory is impressive, and I felt lucky to work
there. He always trusted me and gave me plenty of freedom, which was
perfect for me. I also would like to thank Prof. Charles R. Sullivan for
being the co-examiner of my Ph.D. thesis. Also, I would like to thank
him for creating the foundation of my research, which finally resulted
in the i2 GSE.
ABB Corporate Research sponsored the work behind this thesis. I
would like to thank my supervisor Dr. Andreas Ecklebe for the numer-
ous scientific discussions and his exceptional interest in the outcome
of the work. Also, I would like to thank Henri Kinnunen from ABB
Finland for his interests in my work.
I would like to thank the entire PES staff including Ph.D. students,
post-doctoral researchers, secretaries, administrators, and the electron-
ics laboratory for providing a great research atmosphere.
Furthermore, a big thanks goes to the semester and master students
who have supported me in my research: Sascha O. Schneider, Robert
Blattmann, and Marko Tanasković.
This Ph.D. thesis would not have been successful and, more impor-
tantly, would not have been fun without interactions with people in
and outside the PES. Discussions led many (weird) theories in mag-
netism, which (sometimes) were discarded almost as quickly as they
were formed. Also the glorious, unforgettable performances in many
sport events (such as the SOLA-Stafette, the ASVZ soccer champi-
onship, or the ITET Summer Fight) made the time at PES clearly a
good one. Many people were involved in these events, and the bound-
aries between work and pleasure became blurry. Therefore, here is

iii
an alphabetical list of the people with whom I am grateful to have
shared this important phase of my life: Uwe Badstübner, Jürgen Biela,
David Boillat, Daniel Christen, Bernardo Cougo, Thomas Friedli, Ivana
Kovačević, Fabio Magagna, Christoph Marxgut, Gabriel Ortiz, Mario
Schweizer, Thiago Soeiro, Arda Tüysüz, and Benjamin Wrzecionko.
Last but not least, I would like to thank my family and all my friends
for the great support they have given me anytime and anywhere.

iv
Abstract

This thesis deals with the modeling and multi-objective optimization of

inductive power components, in order to improve the efficiency and/or
power density of power electronic systems.
The first part of the thesis introduces how to model magnetic cir-
cuits, i.e. how to set up an accurate reluctance model of an inductive
component. A novel approach to accurately determine the reluctances
of air gaps is introduced. The approach is easy to handle as it is based
on a modular concept where a simple basic geometry is used as a build-
ing block to describe different three dimensional air gap shapes.
The second part of the thesis deals with core loss modeling. The
applied core loss approach can be seen as a hybrid of an improved ver-
sion of the empirical Steinmetz equation and an approach based on a
material loss database (loss map). In order to build the material loss
database, core loss measurements must be made. Therefore, special
focus is placed on how core losses can be measured and what measure-
ments are necessary for an accurate core loss modeling.
Relaxation effects in magnetic materials are discussed. In mod-
ern power electronic systems, voltages across inductors or transformers
generally show rectangular shapes, including periods of zero voltage. In
most core loss models, the phase where the voltage across the magnetic
component is zero (i.e. the flux remains constant) is not considered. It is
implicitly assumed that no losses occur when the flux remains constant.
However, as measurements show, this is not a valid simplification. In
phases of constant flux, losses still occur in the material. This is due
to relaxation processes. A new core loss modeling approach that takes
such relaxation effects into consideration is given.
Another aspect to be considered is the fact that core losses are in-
fluenced by a DC premagnetization. The Steinmetz Premagnetization

v
Graph (SPG) that shows the dependency of the Steinmetz parameters
(α, β and k) on premagnetization is proposed. This permits the calcu-
lation of core losses under DC bias conditions.
Power electronic engineers often work with circuit simulators in or-
der to validate their designs before building costly prototypes. It is
shown, how to calculate core losses from a simulated flux waveform. In
order to do this, the simulated flux waveform is divided into its funda-
mental flux waveform and into piecewise linear flux waveform segments.
The loss energy is then calculated for the fundamental and all piecewise
linear segments, summed and divided by the fundamental period length
in order to determine the average core loss. Another aspect to be con-
sidered in core loss calculation is the effect of the core shape and size.
By introducing a reluctance model of the core, and with it, calculating
the flux density in every core section of (approximately) homogenous
flux density, one can calculate the losses of each core section. The core
losses of each section are then summed to obtain the total core losses.
This generally leads to a high accuracy. However, under certain cir-
cumstances, in tape wound cores a flux orthogonal to the tape layers
can lead to high eddy currents and thus to high core losses.
The second source of losses in inductive components is the ohmic
losses in the windings. The resistance of a conductor increases with
increasing frequency due to eddy currents. Self-induced eddy currents
inside a conductor lead to the skin-effect. Eddy currents due to an
external alternating magnetic field, e.g. the air gap fringing field or
the magnetic field from other conductors, lead to the proximity-effect.
The skin-effect and proximity-effect losses can be calculated for round,
litz, or foil windings; provided that the external field and the current is
known exactly. However, the calculation of the external magnetic field
strength, which has to be known when calculating the proximity losses,
is challenging. In the case of an un-gapped core and windings that are
fully-enclosed by core material, 1D approximations to determine the
magnetic field exist. However, in the case of gapped cores, such 1D
approximations are not applicable as the fringing field of the air gap
cannot be described in a 1D manner. The approach presented in the
thesis is a 2D approach in which the magnetic field at any position can
be calculated as the superposition of the fields of each of the conductors.
The impact of a magnetic conducting material can be modeled with
the method of images. The presence of an air gap can be modeled as
a fictitious conductor carrying a current equal to the Magneto-Motive

vi
Force (MMF) across the air gap.
Another important aspect in modeling inductive components is their
thermal behavior. This is not only important to avoid overheating;
it also has importance in modeling the losses correctly, as they are
influenced by the temperature. Formulae that allow heat conduction,
convection and radiation to be calculated are given.
The last part of the thesis is about the multi-objective optimization
of inductive power components. The optimization of inductive compo-
nents is illustrated using the example of LCL filters for three-phase PFC
rectifiers. The optimization procedure leads to different filter designs
depending on whether the aim of the optimization is more on reducing
the volume V or more on reducing the losses P . Furthermore, an overall
system optimization, i.e. an optimization of the complete three-phase
PFC rectifier, is given.

vii
Kurzfassung

Diese Doktorarbeit beschäftigt sich mit der umfassenden Modellierung

und Optimierung von induktiven Komponenten, um die Effizienz bzw.
Leistungsdichte von leistungselektronischen Systemen zu erhöhen.
Im ersten Teil der Arbeit wird gezeigt wie man Reluktanzmodelle
induktiver Komponenten aufstellt, wobei der Fokus auf der Bestimmung
von Luftspaltreluktanzen liegt. Zur genauen Bestimmung der Luftspalt-
reluktanz wird ein neuer Rechenansatz eingeführt. Der Ansatz basiert
auf einem modularen Konzept bei dem eine einfache grundlegende Geo-
metrie als Baustein verwendet wird, um verschiedene dreidimensionale
Luftspalttypen zu beschreiben.
Der zweite Teil der Arbeit befasst sich mit der Berechnung von
Kernverlusten. Der angewandte Ansatz zur Bestimmung der Kern-
verluste kann als eine Kombination aus einer erweiterten Version der
empirischen Steinmetzgleichung und einem Ansatz basierend auf einer
Kernmaterial-Verlustdatenbank gesehen werden. Für das Aufstellen
dieser Datenbank müssen Kernverluste von verschiedenen Materialien
gemessen werden. Deshalb ist ein besonderer Schwerpunkt darauf
gelegt, wie Kernverluste gemessen werden und welche Messungen für
eine genaue Modellierung der Kernverluste notwendig sind.
Des Weiteren ist eine ausführliche Diskussion über Relaxationsef-
fekte in magnetischen Materialien gegeben. In modernen leistungselek-
tronischen Systemen liegen typischerweise rechteckförmige Spannungen
an den induktiven Komponenten, einschliesslich Zeiten mit Nullspan-
nung. In den meisten Kernverlustmodellen ist die Phase, wo die Span-
nung über den magnetischen Komponenten Null ist (d.h. wo der mag-
netische Fluss konstant bleibt) nicht berücksichtigt. Es wird implizit
davon ausgegangen, dass keine Verluste auftreten, wenn der Fluss kon-
stant bleibt. In der vorliegenden Arbeit werden Messungen vorgestellt,

ix
welche zeigen, dass zu Beginn eines Intervalls mit konstantem Fluss
noch immer Verluste im Material auftreten. Dies ist auf Relaxations-
prozesse im Kernmaterial zurückzuführen. Ein neuer Kernverlust-
-Modellierungsansatz wird vorgestellt, mit welchem diese Relaxations-
Effekte mitberücksichtigt werden.
Ein weiterer Aspekt, den es zu berücksichtigen gilt ist, dass Kern-
verluste durch eine DC Vormagnetisierung beeinflusst werden. Leider
wird dieses Verhalten in gängigen Datenblättern zu Kernmaterialien
nicht weiter spezifiziert. In dieser Doktorarbeit wird der Steinmetz
Premagnetization Graph (SPG) bzw. Steinmetz Vormagnetisierungs-
Graph eingeführt, welcher die Abhängigkeit der Steinmetzparameter
(α, β und k) bezüglich einer Vormagnetisierung zeigt. Der SPG er-
möglicht die Berechnung der Kernverluste in einem Arbeitspunkt mit
DC Vormagnetisierung.
Entwickler von induktiven Komponenten arbeiten oftmals mit Schal-
tungssimulatoren, um ihre Entwürfe vor dem Bau teurer Prototypen
zu validieren. Es wird gezeigt, wie von einem simulierten Flussver-
lauf Kernverluste berechnet werden können. Dazu wird die Verlusten-
ergie für die Grundschwingung und für alle stückweise linearen Seg-
mente einzeln berechnet. Ein weiterer wichtiger Aspekt in der Kern-
verlustberechnung ist der Einfluss der Form und Grösse des Kernma-
terials auf die Verluste. Durch die Einführung eines Reluktanzmodells
können die Verluste der einzelnen Abschnitte mit (ungefähr) homo-
gener Flussdichte berechnet werden. Die Kernverluste der einzelnen
Abschnitte werden dann aufsummiert. Dieser Ansatz führt zu einer ho-
hen Genauigkeit. Allerdings gibt es Situationen in welchen mit diesem
Vorgehen die Kernverluste unterschätzt werden. In Schnittbandkerne
kann sich unter gewissen Umständen ein Fluss ausbilden, welcher or-
thogonal zu den Bändern steht. In dieser Situation bilden sich starke
Wirbelströme aus. Diese Wirbelströme führen zu überhöhten Kernver-
lusten.
Die zweite Quelle von Verlusten in induktiven Bauelementen sind die
ohmschen Verluste in den Wicklungen. Der Widerstand eines Leiters
steigt mit steigender Frequenz aufgrund von selbst-induzierten Wirbel-
strömen. Dieser Effekt nennt sich Skin-Effekt. Die Wirbelströme in
einem Leiter, welche von einem externen magnetischen Wechselfeld
(z.B. dem Luftspaltstreufeld oder dem magnetischen Feld von Nach-
barleitern) induziert werden führen zum Proximity-Effekt. Die Verluste
aufgrund des Skin-Effekts und Proximity-Effekts können für Rundleiter,

x
für Hochfrequenz-Litze und Folienleiter berechnet werden, vorausge-
setzt, dass das äussere Feld genau bekannt ist. Allerdings ist die Be-
stimmung dieses äusseren Feldes zur Berechnung des Proximity-Effekts
nicht ganz trivial. Für den Fall eines Kerns ohne Luftspalt und mit
Leitern, die vollständig von Kernmaterial umgeben sind, existieren 1D
Ansätze für die Bestimmung des äusseren Feldes. Doch im Fall von Ker-
nen mit Luftspalten sind solche 1D Ansätze nicht anwendbar, da für die
Beschreibung des Luftspalt-Streufeldes mindestens eine 2D Beschrei-
bung notwendig ist. Der Ansatz in der vorliegenden Arbeit ist ein
2D Ansatz, bei welchem das Magnetfeld an jeder beliebigen Stelle als
Überlagerung der Felder der einzelnen Leitern abgeleitet wird. Die
Auswirkungen eines magnetischen leitenden Materials lassen sich mit
dem Spiegelungsverfahren beschreiben. Ein Luftspalt kann mittels fik-
tivem Leiter, welcher einen Strom gleich dem magnetischen Spannungs-
abfalle über dem Luftspalt führt, modelliert werden.
Ein weiterer wichtiger Aspekt bei der Modellierung induktiver Bau-
elemente ist ihr thermisches Verhalten. Dies ist nicht nur wichtig, um
eine thermische Zerstörung zu vermeiden, es ist auch wichtig, um die
Verluste korrekt zu modellieren, da diese durch die Temperatur beein-
flusst werden. Formeln zur Bestimmung der Kern und Wicklungstem-
peratur sind gegeben, wobei die Wärmeleitung, Wärmekonvektion, und
Wärmestrahlung gerechnet wird.
Im letzten Teil der Arbeit geht es um die Optimierung induktiver
Komponenten. Die Optimierung von induktiven Bauelementen wird am
Beispiel eines LCL-Filters für dreiphasige PFC-Pulsgleichrichter illustri-
ert. Das vorgestellte Optimierungsverfahren führt zu unterschiedlichen
Filterdesigns mit unterschiedlichen Volumina V und Verlusten P . Des
Weiteren wird eine Optimierung des gesamten Systems, also eine Opti-
mierung des gesamten dreiphasigen PFC-Pulsgleichrichters vorgestellt.

xi
Contents

1 Introduction 1
1.1 Modeling of Inductive Power Components . . . . . . . . 2
1.2 Multi-Objective Optimization of Inductive Power Com-
ponents . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Outline of the Thesis . . . . . . . . . . . . . . . . . . . . 8
1.4 List of Publications . . . . . . . . . . . . . . . . . . . . . 10

2 Magnetic Circuit Modeling 13

2.1 Existing Approaches for Air Gap Reluctance Calculation 15
2.1.1 Method of the Conformal Schwarz-Christoffel Trans-
formation . . . . . . . . . . . . . . . . . . . . . . 15
2.1.2 Increase of the Air Gap Cross-Sectional Area . . 16
2.1.3 FEM Tuned Equations . . . . . . . . . . . . . . . 17
2.1.4 Desired Approach for Air Gap Calculations . . . 17
2.2 Air Gap Reluctance . . . . . . . . . . . . . . . . . . . . 18
2.2.1 Reluctance of Air Gaps with Rectangular Cross-
Section . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.2 Reluctance of Air Gap with Round Cross-Section 24
2.2.3 Adaptation of the Air Gap Reluctance Calcula-
tion Approach to Different Problems . . . . . . . 24
2.3 Core Reluctance . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Experimental Results . . . . . . . . . . . . . . . . . . . . 27
2.4.1 Inductance . . . . . . . . . . . . . . . . . . . . . 27
2.4.2 Saturation . . . . . . . . . . . . . . . . . . . . . . 27

3 Core Loss Modeling 31

3.1 Physical Origin of Core Losses . . . . . . . . . . . . . . 32
3.2 Existing Approaches for Core Loss Calculation . . . . . 35

xiii
CONTENTS

3.3 Outline of Novel Core Loss Calculation Models and Ap-

proaches . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.4 Test Setup to Measure Core Losses . . . . . . . . . . . . 39
3.5 Accuracy of the Measurement System . . . . . . . . . . 41
3.5.1 Phase Shift Error of Voltage and Current Mea-
surement . . . . . . . . . . . . . . . . . . . . . . 43
3.5.2 Equipment Accuracy . . . . . . . . . . . . . . . . 45
3.5.3 Capacitive Coupling . . . . . . . . . . . . . . . . 45
3.5.4 Temperature . . . . . . . . . . . . . . . . . . . . 46
3.5.5 Comparative Measurement and Conclusion . . . 46
3.6 Relaxation Effects in Magnetic Materials . . . . . . . . . 46
3.6.1 Measurement Results . . . . . . . . . . . . . . . 47
3.6.2 Theory of Relaxation Effects . . . . . . . . . . . 49
3.6.3 Shape of B-H Loop for Trapezoidal Flux Waveforms 50
3.6.4 Model Derivation . . . . . . . . . . . . . . . . . . 52
3.6.5 Measurements on Different Materials . . . . . . . 63
3.6.6 Conclusion and Future Work . . . . . . . . . . . 65
3.7 Core Losses Under DC Bias Conditions . . . . . . . . . 66
3.7.1 Measurement Results and the Steinmetz Premag-
netization Graph (SPG) . . . . . . . . . . . . . . 67
3.7.2 Influence of Temperature . . . . . . . . . . . . . 72
3.7.3 Example How to Use the SPG . . . . . . . . . . 74
3.7.4 Core Losses under DC Bias Conditions of Differ-
ent Materials . . . . . . . . . . . . . . . . . . . . 76
3.7.5 Conclusion and Future Work . . . . . . . . . . . 77
3.8 Minor and Major B-H Loops . . . . . . . . . . . . . . . 77
3.9 Hybrid Core Loss Calculation Approach . . . . . . . . . 79
3.9.1 Use of Loss Map . . . . . . . . . . . . . . . . . . 82
3.10 Influence of Core Shape on Core Losses . . . . . . . . . 84
3.10.1 Effective Dimensions of Toroidal Cores . . . . . . 85
3.10.2 General Core Shape . . . . . . . . . . . . . . . . 86
3.10.3 Impact of Core Shape on Eddy Current Losses . 88
3.10.4 Dimensional Resonance . . . . . . . . . . . . . . 91
3.10.5 Losses in Gapped Tape Wound Cores . . . . . . 93

4 Winding Loss Modeling 97

4.1 Skin Effect . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.2 Proximity Effect . . . . . . . . . . . . . . . . . . . . . . 100
4.3 Round Conductor . . . . . . . . . . . . . . . . . . . . . . 102

xiv
 CONTENTS

4.3.1 Skin Effect . . . . . . . . . . . . . . . . . . . . . 103

4.3.2 Proximity Effect . . . . . . . . . . . . . . . . . . 104
4.3.3 Multi-Layer Windings Without Air Gap . . . . . 105
4.3.4 Multi-Layer Windings With Air Gap . . . . . . . 107
4.3.5 Litz-Wire Windings . . . . . . . . . . . . . . . . 110
4.3.6 Accuracy Analysis . . . . . . . . . . . . . . . . . 113
4.4 Foil Conductor . . . . . . . . . . . . . . . . . . . . . . . 115
4.4.1 Skin Effect . . . . . . . . . . . . . . . . . . . . . 117
4.4.2 Proximity Effect . . . . . . . . . . . . . . . . . . 118
4.4.3 Foil Multi-Layer Without Air Gap . . . . . . . . 119
4.4.4 Short Foil Conductors . . . . . . . . . . . . . . . 120
4.4.5 Foil Multi-Layer With Air Gap . . . . . . . . . . 121
4.4.6 Accuracy Analysis . . . . . . . . . . . . . . . . . 123

5 Thermal Modeling 127

5.1 Overview of Thermal Models . . . . . . . . . . . . . . . 127
5.1.1 Total Thermal Resistance . . . . . . . . . . . . . 128
5.1.2 Thermal Resistor Networks . . . . . . . . . . . . 128
5.2 Heat Transfer Mechanisms . . . . . . . . . . . . . . . . . 130
5.2.1 Thermal Conduction . . . . . . . . . . . . . . . . 130
5.2.2 Thermal Natural Convection . . . . . . . . . . . 131
5.2.3 Thermal Radiation . . . . . . . . . . . . . . . . . 136
5.3 Practical Implementation Issues . . . . . . . . . . . . . . 137

6 Magnetic Design Environment and Experimental Re-

sults 139
6.1 Overview of Implemented Loss Modeling Environment . 140
6.2 Experiment I . . . . . . . . . . . . . . . . . . . . . . . . 142
6.3 Experiment II . . . . . . . . . . . . . . . . . . . . . . . . 144

7 Multi-Objective Optimization of Inductive Power Com-

ponents 147
7.1 Three-Phase PFC Rectifier with Input Filter . . . . . . 148
7.2 Modeling of Input Filter Components . . . . . . . . . . 150
7.2.1 Calculation of the Inductance . . . . . . . . . . . 150
7.2.2 Core Losses . . . . . . . . . . . . . . . . . . . . . 151
7.2.3 Winding Losses . . . . . . . . . . . . . . . . . . . 151
7.2.4 Thermal Modeling . . . . . . . . . . . . . . . . . 152
7.2.5 Capacitor Modeling . . . . . . . . . . . . . . . . 152
7.2.6 Damping Branch . . . . . . . . . . . . . . . . . . 153

xv
CONTENTS

7.3 Optimization of the Input Filter . . . . . . . . . . . . . 153

7.3.1 Optimization Constraints and Conditions . . . . 154
7.3.2 Calculation of L2,min . . . . . . . . . . . . . . . . 154
7.3.3 Loss Calculation of Filter Components . . . . . . 155
7.3.4 Optimization Procedure . . . . . . . . . . . . . . 156
7.4 Optimization Outcomes . . . . . . . . . . . . . . . . . . 159
7.5 Experimental Results . . . . . . . . . . . . . . . . . . . . 160
7.6 Overall Rectifier Optimization . . . . . . . . . . . . . . . 162
7.6.1 Overall Optimized Designs . . . . . . . . . . . . 164
7.7 Conclusion and Future Work . . . . . . . . . . . . . . . 169

8 Summary & Outlook 171

8.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

A Appendix 175
A.1 Derivation of the Basic Reluctance . . . . . . . . . . . . 175
A.2 iGSE and Sinusoidal Flux Waveforms . . . . . . . . . . 180
A.3 SPGs of Other Materials . . . . . . . . . . . . . . . . . . 181
A.4 Derivation of the Steinmetz Premagnetization Graph . . 181
A.5 Classic Steinmetz Parameter k . . . . . . . . . . . . . . 185
A.6 Derivation of Effective Dimensions for Toroidal Cores . 185
A.7 Derivation of Winding Losses in Foil Conductors . . . . 187
A.7.1 Skin Effect . . . . . . . . . . . . . . . . . . . . . 188
A.7.2 Proximity Effect . . . . . . . . . . . . . . . . . . 191
A.8 Derivation of Winding Losses in Round Conductors . . . 192
A.8.1 Skin Effect . . . . . . . . . . . . . . . . . . . . . 192
A.8.2 Proximity Effect . . . . . . . . . . . . . . . . . . 194
A.9 Orthogonality of Winding Losses . . . . . . . . . . . . . 196

xvi
Chapter 1

Introduction

The three major reasons why the world’s energy policy has to be rethought
are [1]:
I global warming / climate change that has to be prevented,

I dependency on fossil energy sources as oil, gas, coal, etc. that are
often located in unstable world regions,

I and increasing energy costs due to shortage of resources, e.g. dis-

cussion about the peak-oil problem.

Bose [2] predicted in 2000 that "it appears that cheap and abundant
energy supply which we are now enjoying will be over in future and our
society will be forced to move in an altered direction". One step towards
a more sustainable energy policy and improved stewardship of available
resources could be achieved by promoting electric energy. Taking elec-
tric vehicles as an example, even when electricity is generated mainly
in coal-fired power plants, the fuel chain efficiency of electrical vehicles
compared to gasoline-fulled vehicles increases [2].
Power electronics is clearly a key technology in helping to solve
any of these upcoming energy issues. Highly efficient power electronic
converters allow energy saving through the efficient use of electricity.
Power electronics will also play a key role in providing new solutions
for transportation systems (e.g. electric vehicles), transmission sys-
tems (e.g. HVDC), distribution systems, and integration of new power
sources into the grid [3].

1
CHAPTER 1. INTRODUCTION

The evolution of power electronics closely follows the evolution of

power semiconductor devices. Power semiconductor devices have im-
proved in terms of higher current / voltage ratings, and lower conduc-
tion and switching losses allowing higher switching frequencies. New
power semiconductor devices permit higher system efficiencies and lower
system volumes, i.e. higher power densities. Generally, one can see
a trend in power electronics research towards higher efficiencies and
higher power densities. This trend is driven by cost considerations
(e.g. material economies), space limitations (e.g. in the automotive en-
vironment), and increasing efficiency requirements (e.g. for telecom ap-
plications). The increase of the power density often affects the efficiency,
i.e. a trade-off between these two quality indices exists [4].
A higher switching frequency allows a reduction in the volume of
passive components, such as inductive components. Inductive compo-
nents occupy a significant amount of space in today’s power electronic
systems, and furthermore, considerable losses occur in these compo-
nents. Particularly, today’s increasing switching frequency leads to
High Frequency (HF) losses that are difficult to determine. In order
to increase the power density and/or efficiency of power electronic sys-
tems, losses in inductive components must be reduced, and/or new
cooling concepts need to be investigated. For it, accurate loss and ther-
mal models are crucial. There have been several publications focusing
on modeling inductive components [5, 6, 7, 8, 9, 10]. The aim of this
thesis is to improve the model accuracy. This is done by combining the
best state-of-the-art approaches with newly-developed approaches.

1.1 Modeling of Inductive Power Compo-

nents
Inductive components are widely used in power electronic applications.
Three typical examples that employ inductive power components are
illustrated in Figure 1.1. Figure 1.1(a) shows a buck converter with
a typical inductor current/flux waveform in Continuous Conduction
Mode (CCM). The current is, in case of a linear inductance L, pro-
portional to the magnetic flux. The flux waveform has a triangular
shape with a DC offset. The core losses are influenced by this DC pre-
magnetization; this has been discussed in many previous publications
[8, 11, 12, 13, 14, 15, 16]. Since data that shows the influence of a DC

2
 1.1. MODELING OF INDUCTIVE POWER COMPONENTS

Buck Converter
L Current / Flux Waveform
iL
iL / φL

V1 V2 IL t

(a)
Dual Active Bridge

+ + Current / Flux Waveform

iL L 1:1 iL / φL
V1 v1 v2 V2
t

- -

(b)
Three-Phase PFC Rectifier
IL
Current / Flux Waveform
iL / φL
iL L

VDC IL t
Vmains

(c)

Figure 1.1: Applications of inductive power components: (a) in-

ductor in buck converter, (b) inductor in Dual Active Bridge (DAB)
converter, (c) boost inductors in three-phase PFC rectifier.

premagnetization to the core losses is normally not provided by core

manufacturers, it is difficult to determine core losses under a DC bias
condition.
In Figure 1.1(b) a Dual Active Bridge (DAB) with a typical inductor
current/flux waveform for CCM is given. This flux waveform is typical
for power electronic applications in general. In modern power electronic
systems voltages across inductors or transformers generally show rect-
angular shapes, including periods of zero voltage. In most core loss
models, the phase where the voltage across the magnetic component is
zero (i.e. the flux remains constant) is not considered. It is implicitly

3
CHAPTER 1. INTRODUCTION

Figure 1.2: Overview of different typical flux waveforms.

assumed that no losses occur when the flux remains constant. However,
as measurements show, this is not a valid simplification. At the begin-
ning of a period of constant flux, losses still occur in the material. This
is due to relaxation processes inside the magnetic material.
The third application is a three-phase PFC rectifier with three boost
inductors (cf. Figure 1.1(c)). The flux waveform can be seen as a
fundamental (sinusoidal) waveform and superimposed HF ripple. When
plotting the traversed B-H-curve, one sees a large loop and many small
loops, i.e. one large major loop and many small minor loops. The large
loop originates in the fundamental waveform, while the small loops
originate in the HF ripples.
The three flux waveforms discussed point out the main issues related
to the impact of different flux waveform shapes on core losses. These
three flux waveforms and a standard sinusoidal waveform are summa-

4
 1.1. MODELING OF INDUCTIVE POWER COMPONENTS

rized in Figure 1.2. In this thesis, different means of how to calculate

core losses while addressing the above mentioned difficulties are derived.
The Steinmetz Premagnetization Graph (SPG) that shows the depen-
dency of the Steinmetz parameters (α, β and k) on premagnetization
is proposed. This permits the calculation of core losses under DC bias
conditions. In order to model flux waveforms with periods of constant
flux, a new core loss model that takes relaxation effects into considera-
tion is derived. Furthermore, it is experimentally verified that the loss
energy for each (minor or major) loop can be calculated independently
and totaled. The flux waveform can be divided into its fundamental
flux waveform and into piecewise linear flux waveform segments. The
loss energy is then calculated for the fundamental and all piecewise lin-
ear segments, totaled and divided by the fundamental period length in
order to determine the average core loss. Actually, when doing this,
one does not consider how the minor loop closes: each piecewise linear
segment is modeled as having half the loss energy of its corresponding
closed loop.
In order to improve the model accuracy a loss database has been
built up. The applied core loss approach can be seen as a hybrid of an
improved version of the empirical Steinmetz equation and an approach
based on a material loss database (loss map). In order to build the
material loss database, core loss measurements have to be conducted. A
particular focus is therefore placed on how core losses can be measured,
and what measurements are necessary for accurate core loss modeling.
The situation concerning non-linear current waveforms is much sim-
pler for winding losses. Due to the orthogonality of the cosine-function,
it is acceptable to perform a Fourier expansion of the current, calculate
the losses for each frequency component, and then total up the losses.
This is not valid for core losses, since there is a non-linear relation
between losses and (peak) flux density.
In addition to the impact of the current/flux waveform on core and
winding losses, there are other issues, which make the modeling of in-
ductive components challenging. The procedure to model inductive
components can be structured as:

1. Set up a magnetic circuit model, i.e. a reluctance model.

2. Determine core losses.

3. Determine winding losses.

5
CHAPTER 1. INTRODUCTION

4. Model the thermal behavior.

A magnetic circuit model is set up in order to calculate the induc-

tance, predict the flux density in each section of the core, and model
the effect of the air gap fringing field on winding losses. The flux den-
sity in each core section is important to calculate the core losses and
to avoid saturation of the core. Despite its importance, a satisfying air
gap reluctance model has not been found in literature. A new means
to determine accurately the reluctances of air gaps is accordingly sug-
gested. The approach is easy to handle as it is based on a modular
concept where a simple basic geometry is used as a building block to
describe different three dimensional air gap shapes.
In step two, the core losses are modeled. In addition to the issues
of the impact of the flux waveform on core losses (see above), another
aspect to be considered in core loss calculation is the effect of the core
shape and size. By introducing a reluctance model of the core, and
with it, calculating the flux density in every core section of (approxi-
mately) homogenous flux density, one can calculate the losses of each
core section. The core losses of each section are then totaled to obtain
the total core losses. This generally leads to a high accuracy. However,
under certain circumstances, in tape wound cores a flux orthogonal to
the tape layers can lead to high eddy currents and therewith to high
core losses.
The second source of losses in inductive components is the ohmic
losses in the windings. The resistance of a conductor increases with
increasing frequency due to eddy currents. Self-induced eddy currents
inside a conductor lead to the skin effect. Eddy currents due to an
external alternating magnetic field, e.g. the air gap fringing field or
the magnetic field from other conductors, lead to the proximity-effect.
The skin-effect and proximity-effect losses can be calculated for round,
litz, or foil windings; provided that the external field and the current
are exactly known. However, the calculation of the external magnetic
field strength, which has to be known when calculating the proximity
losses, is challenging. In the case of an un-gapped core and windings
that are fully-enclosed by core material, 1D approximations to calculate
the magnetic field exist. However, in the case of gapped cores, such
1D approximations are not applicable as the fringing field of the air
gap cannot be described in an 1D manner. This thesis applies a 2D
approach in which the magnetic field at any position can be determined
as superposition of the fields of each of the conductors. The impact

6
 1.2. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE POWER
COMPONENTS

L1 L2

VDC IL

Vmains
C

LCL Filter

Figure 1.3: Three-phase PFC rectifier with LCL input filter.

of a magnetic conducting material can further be modeled with the

method of images. The presence of an air gap can be modeled as a
fictitious conductor carrying a current equal to the Magneto-Motive
Force (MMF) across the air gap.
Finally, in step four, the thermal behavior of the inductive compo-
nents must be determined. This is not only important to avoid over-
heating; it also has importance in modeling the losses correctly, as they
are influenced by the temperature. As the temperature and losses are
depending on each other, one has to iteratively perform the four steps
until the problem converges.

1.2 Multi-Objective Optimization of Induc-

tive Power Components
The loss models described in this thesis will form the basis for the
optimization of inductive components. The focus of the last part of
the thesis is placed on how to conduct a multi-objective optimization of
inductive power components employed in power electronic applications.
Using the example of an LCL input filter of a three-phase Power Factor
Correction (PFC) rectifier, an inductive component design procedure
based on a generic optimization approach, guaranteeing low volume
and/or low losses, is proposed. The system considered is illustrated in
Figure 1.3. Limiting factors for the filter design are the filter’s maximum
temperature and its maximum volume. The optimization procedure
leads to different filter designs depending on whether the aim of the
optimization is more on reducing the filter volume or more on reducing

7
CHAPTER 1. INTRODUCTION

the filter losses. In the selected example of an LCL input filter, a

higher switching frequency leads to lower volume, or, when keeping the
volume constant, to lower losses of the inductive components. However,
higher semiconductor switching losses are expected in the case of higher
switching frequencies. Therefore, an overall system optimization, i.e. an
optimization of the complete three-phase PFC rectifier (not only the
filter), is conducted. In so doing it is taken into account that in the
design phase of any power electronics system, it is important to consider
the system as a whole, as there are parameters that bring advantages
for one subsystem but bring disadvantages for another.

1.3 Outline of the Thesis

In Chapter 2 it is shown how to model inductive components us-
ing magnetic circuits. A magnetic circuit model enables a fast and
straightforward inductance calculation. It also allows one to predict
the flux density in each section of the core, thereby making it possible
to calculate core losses and avoid saturation of the core when designing
inductive components. Furthermore, when calculating winding losses
the effect of the air gap fringing flux can be accurately modeled as a
function of the Magneto-Motive Force (MMF) across the air gap. The
air gap reluctance in a magnetic circuit model (or reluctance model) is
by far the most difficult to calculate. An approach to accurately cal-
culate the reluctance of an air gap has been developed in the course of
this thesis.
In Chapter 3 it is shown how core losses can be modeled accurately;
thereby considering all different aspects of core loss modeling. The
impact of peak-to-peak flux density, frequency, DC premagnetization,
temperature, core shape, minor and major loops, flux waveform, and
material on the core loss calculation are all considered. A high level of
accuracy is achieved by combining the best state-of-the-art approaches
and by embedding newly developed approaches into a new hybrid core
loss calculation approach (Section 3.9). A special discussion is given
about relaxation effects in magnetic materials. Additional losses may
occur due to relaxation processes. A new core loss modeling approach
that takes such relaxation effects into consideration has been developed
in the course of this thesis and named the improved-improved General-
ized Steinmetz Equation, i2 GSE (Section 3.6). Another fact that makes
the prediction of core losses difficult is that core losses are influenced by

8
 1.3. OUTLINE OF THE THESIS

a DC premagnetization. The Steinmetz Premagnetization Graph (SPG)

that shows the dependency of the Steinmetz parameters (α, β and k)
on premagnetization is introduced (Section 3.7). This permits the cal-
culation of core losses under DC bias conditions. Other aspects that are
discussed are how core losses are measured (Section 3.4 and Section 3.5)
and how the core shape influences core losses (Section 3.10).

In Chapter 4 the second source of losses in inductive components,

the ohmic losses in the windings, are discussed. The skin effect and
proximity effect losses can be calculated for foil (Section 4.4), round
(Section 4.3), or litz wire (Section 4.3.5) windings; provided that the
external field and the current are exactly known. An approach that
also considers the air gap stray field in the calculation of the external
field for the case of gapped cores is given.

In Chapter 5 another important aspect in modeling inductive com-

ponents, their thermal behavior, is introduced. This is not only impor-
tant to avoid overheating, but also for correctly modeling the losses, as
they are influenced by the temperature. It is shown how to calculate
heat conduction, convection, and radiation in inductive components.

In Chapter 6 experimental tests that confirm the overall accu-

racy are shown. It is shown that a high level of accuracy is achieved
by combining all loss and thermal models introduced in this thesis.
Furthermore, in order to handle the models and enable others to de-
termine losses accurately, a Magnetic Design Environment has been
implemented in the course of this thesis. A short overview is given in
Section 6.1.

The last part of the thesis (Chapter 7) is about the multi-objective

optimization of inductive power components. The design of inductive
components will be illustrated using the example of LCL filters for
three-phase PFC rectifiers. In the design phase all the loss and thermal
models introduced will be used. The optimization procedure demon-
strated leads to different filter designs depending whether the aim of
the optimization is more on reducing the volume or more on reducing
the losses. Furthermore, an overall system optimization, i.e. an opti-
mization of the complete three-phase PFC rectifier including the filter,
has been performed.

9
CHAPTER 1. INTRODUCTION

1.4 List of Publications

Different parts of this work have already been published or are being
published in international journals or conference proceedings. These
publications are listed below.

Conference Papers
1. J. Mühlethaler, J. Biela, J. W. Kolar, and A. Ecklebe, Core Losses
under DC Bias Condition Based on Steinmetz Parameters, Proc.
of the IEEE/IEEJ International Power Electronics Conference
(ECCE Asia), Sapporo, Japan, June 21-24, 2010.

2. J. Mühlethaler, J. Biela, J. W. Kolar, and A. Ecklebe, Improved

Core Loss Calculation for Magnetic Components Employed in
Power Electronic Systems, Proc. of the Applied Power Electron-
ics Conference and Exposition (APEC), Fort Worth, TX, March
6-10, 2011.

3. J. Mühlethaler, J. W. Kolar, and A. Ecklebe, A Novel Approach

for 3D Air Gap Reluctance Calculations, Proc. of the 8th Interna-
tional Conference on Power Electronics - ECCE Asia, The Shilla
Jeju, Korea, May 30-June 3, 2011.

4. J. Mühlethaler, J. W. Kolar, and A. Ecklebe, Loss Modeling of

Inductive Components Employed in Power Electronic Systems,
Proc. of the 8th International Conference on Power Electronics -
ECCE Asia, The Shilla Jeju, Korea, May 30-June 3, 2011.

5. C. Marxgut, J. Mühlethaler, F. Krismer, and J. W. Kolar, Multi-

Objective Optimization of Ultra-Flat Magnetic Components with
a PCB-Integrated Core, Proc. of the 8th International Conference
on Power Electronics - ECCE Asia, The Shilla Jeju, Korea, May
30-June 3, 2011.

6. J. Mühlethaler, M. Schweizer, R. Blattmann, J. W. Kolar, and

A. Ecklebe, Optimal Design of LCL Harmonic Filters for Three-
Phase PFC Rectifiers, in Proc. of the IECON, Melbourne, Novem-
ber 7-10, 2011

10
 1.4. LIST OF PUBLICATIONS

7. B. Cougo, A. Tüysüz, J. Mühlethaler, and J.W. Kolar, Increase

of Tape Wound Core Losses Due to Interlamination Short Cir-
cuits and Orthogonal Flux Components, in Proc. of the IECON,
Melbourne, November 7-10, 2011.

Journal Papers
8. J. Mühlethaler, J. Biela, J.W. Kolar, and A. Ecklebe, Core Losses
Under the DC Bias Condition Based on Steinmetz Parameters,
IEEE Transactions on Power Electronics, Vol. 27, No. 2, Febru-
ary 2012.
9. J. Mühlethaler, J. Biela, J.W. Kolar, and A. Ecklebe, Improved
Core-Loss Calculation for Magnetic Components Employed in Power
Electronic Systems, IEEE Transactions on Power Electronics, Vol.
27, No. 2, February 2012.

11
Chapter 2

Magnetic Circuit
Modeling

In an analogous manner to Kirchhoff’s voltage law, according to Am-

pere’s law the sum of the magnetomotive force (MMF) in a magnetic
circuit around a closed loop is zero. Gauss’s law for magnetic circuits
yields to the fact that the sum of magnetic fluxes into a node is zero,
similar to Kirchhoff’s current law. Furthermore, the reluctance Rm is
defined as Rm = MMF/Φ, where Φ is the magnetic flux through the re-
luctance Rm . These analogies lead to the concept of magnetic circuits.
The theory of magnetic circuits is, for instance, introduced in [5, 17].
A magnetic circuit model, also named simply reluctance model, of
an inductive component enables a fast and straightforward inductance
calculation. It also allows one to predict the flux density in each sec-
tion of the core, thereby making it possible to avoid saturation of the
core. Furthermore, when winding losses are calculated, the effect of
the air gap fringing flux can be modeled accurately as a function of
the Magneto-Motive Force (MMF) across the air gap [5]. The air gap
reluctance must be known in advance for an accurate MMF calculation.
The inductance of an inductive component with N winding turns
and a total magnetic reluctance Rm,tot is calculated as

N2
L= . (2.1)
Rm,tot
The reluctance of each section of the flux path has to be calculated first

13
CHAPTER 2. MAGNETIC CIRCUIT MODELING

Rc2

Rc1 Rw1 Rw2 Rc4

Rg

Rc3

Figure 2.1: Illustration of a magnetic reluctance model for the ex-

ample of an E-core.

in order to calculate Rm,tot . In Figure 2.1 such a reluctance model for

the example of an E-core is illustrated. The calculation of the reluctance
Rci of the core sections is simple: for a core section of length lc , cross-
section Ac , and permeability µr µ0 it is

lc
Rci = . (2.2)
µr µ0 Ac

Difficulties occur only in the corner sections. However, simple and rea-
sonably accurate approximations for handling these sections exist. The
reluctances Rw1 and Rw2 represent the flux paths that are not closed
over the core and consequently reduce the total magnetic reluctance
Rm,tot . These winding reluctances Rw are rather high and can be ne-
glected in most practical cases. Dominant for the inductance value is
the air gap reluctance Rg . Unfortunately, it is by far the most difficult
to calculate; therefore, the focus of this chapter is on its calculation.
Existing means of calculating the reluctance of air gaps are given in
the following section, and, in Section 2.2, a new model to calculate the
air gap reluctance is derived. Later, in Section 2.3, it is discussed how
core reluctances can be calculated and in Section 2.4, to support the
theoretical analysis, experimental results are given.

14
 2.1. EXISTING APPROACHES FOR AIR GAP RELUCTANCE
CALCULATION

2.1 Existing Approaches for Air Gap Re-

luctance Calculation
Different means of calculating the reluctance of air gaps exist. Under
the assumption of a homogenous flux density distribution in the air gap
and no fringing flux, the air gap reluctance can be calculated as
lg
Rg = , (2.3)
µ0 Ag
where lg and Ag are the air gap length and air gap cross-section re-
spectively, and µ0 is the permeability of free space. Equation (2.3) is
only accurate when the fringing flux is small compared to the total flux,
i.e. when the air gap length is very small compared to the dimensions
of the air gap cross-section.
Different approaches of how to take the fringing flux into considera-
tion for calculating the air gap reluctance have been derived in the past.
The approaches can be basically classified into three categories, which
are discussed within Section 2.1.1, Section 2.1.2, and Section 2.1.3.

2.1.1 Method of the Conformal Schwarz-Christoffel

Transformation
One approach to calculate analytically the reluctance of an air gap
geometry is the method of the conformal Schwarz-Christoffel transfor-
mation [18, 19, 20]. The solutions described in [9, 21, 22] are based on
this transformation. The approach of [21] is based on the capacitance-
to-reluctance analogy for calculating the reluctance of air gaps: if air is
the dielectric, the capacitance C can be expressed as
C = 0 F (g), (2.4)
where F (g) represents the geometry between plates of infinite conduc-
tivity and 0 is the permittivity of free space. The reluctance Rm,airgap
of an air gap between surfaces of infinite permeability and with the
same geometry F (g) is then
1
Rg = . (2.5)
µ0 F (g)
Hence, the calculation of the capacitance of the air gap geometry with
help of the Schwarz-Christoffel transformation leads directly to the air

15
CHAPTER 2. MAGNETIC CIRCUIT MODELING

gap reluctance. However, in [21] only 2D considerations have been

made, which considerably limits the practicability of this approach for
real 3D air gaps.
Also in [9] the magnetic field is calculated via the Schwarz-Christoffel
transformation, though without the capacitance-to-reluctance analogy.
The three-dimensionality of air gaps has been taken into consideration.
The solution of approach [9] leads to a relatively complex, difficult to
handle, but accurate formula for the air gap reluctance.

2.1.2 Increase of the Air Gap Cross-Sectional Area

The fringing flux can be taken into consideration by increasing the
actual air gap cross-sectional area (by increasing the air gap cross-
sectional area the reluctance is reduced to the actual value) [22, 23, 24].
According to [23], 10 % is a typical value by which the cross-sectional
area has to be increased; however, it is admitted that the cross-sectional
area should be increased as a function of the air gap length lg and that
a determination of the exact value by which the cross-sectional area has
to be increased is difficult. In [24], the reluctance of an air gap with
length lg , width c, and depth t is calculated as

lg
Rg = . (2.6)
µ0 (c + lg )(t + lg )

No explanation or derivation of the equation is given. As will be shown

later, the results deviate from comparison with FEM simulations. In
[22] the effective semi-width of the gap, e.g. the radius in case of a round
cross-section, is increased by the amount of
 
1 ba
0.241 + ln lg , (2.7)
π lg

where ba is the total inside length of the leg containing the air gap,
e.g. for a pot core ba equals the width of the winding space. No deriva-
tion for the formula (2.7) is given. However, (2.7) can be derived from
(2.8)1 ((2.8) has not been explained yet but will be introduced in the
next section). The approach of (2.7) is based on the method of the
1µ
w 2 πh


0 2l + π 1 + ln 4l with l = lg /2 and h = ba /2 becomes
h  i h   i
2µ0 w lg πba 2µ0 w 1 ba
lg 2
+ π
1 + ln 4lg
= lg 2
+ 0.241 + π
ln lg
lg .

16
 2.1. EXISTING APPROACHES FOR AIR GAP RELUCTANCE
CALCULATION

conformal Schwarz-Christoffel transformation and is therefore a hybrid

of the two categories. This approach is only capable of calculating the
reluctances of air gaps with shapes as illustrated in Figure 2.2(a); for
this particular air gap shape (air gap type 1/1; cf. next section) the
solution of approach (2.7) and the solution obtained with the model
that will be derived next are the same.

2.1.3 FEM Tuned Equations

A third class of approaches is based on equations with FEM tuned pa-
rameters for different air gap cases. For instance, an approach to derive
the 3D reluctance from the 2D results of [21] has been made in [25]. For
it, the corner reluctance of one air gap shape is described by an equation,
in which a parameter has to be tuned by a 3D FEM simulation. The
structure of the equation for the corner reluctance was found in [26].
The approach in [26] is a rough approximation based on geometrically
constructed flux and equipotential lines of different given standard ge-
ometries. This rough approximation had then been improved in [25] by
a FEM tuned parameter. Another approach to solve the problem with
FEM tuned equations is given in [5, 27], where different two dimensional
cases are modeled with FEM tuned equations. A corner reluctance is
added as well in order to consider the 3D reluctance. Generally, a high
accuracy can be achieved with FEM tuned equations as given in [25] or
[5, 27]. However, in both approaches the added corner reluctance has
more an approximative character and a different approach to consider
the three-dimensionality would be interesting to have.

2.1.4 Desired Approach for Air Gap Calculations

An approach for an air gap calculation that is capable of calculating
the reluctance and

I considers the three-dimensionality,

I is reasonably easy-to-handle,

I is capable of handling different shapes of air gaps, as e.g. the ones

illustrated in Figure 2.2,

I and maintains a high accuracy,

17
CHAPTER 2. MAGNETIC CIRCUIT MODELING

(a) (b)

Figure 2.2: Illustration of different core shapes that lead to different

air gap shapes.

is important to have. In the following, a new approach to calculate the

air gap reluctance is derived. The approach is accurate because it is
based on analytical field solutions.

2.2 Air Gap Reluctance

The novel approach for the 3D air gap reluctance calculation is de-
rived from the example of air gaps with rectangular cross-section in
Section 2.2.1. Later, in Section 2.2.2, it is described how to apply the
approach to air gaps with round cross-sections.

2.2.1 Reluctance of Air Gaps with Rectangular Cross-

Section
The reluctance of the simple basic geometry of Figure 2.3 is taken as a
basis to calculate more complex air gap structures. This basic geometry
is used as a building block to describe different three dimensional air
gap shapes.
The geometry in Figure 2.3 has the 2D reluctance

0 1
Rbasic = w
 , (2.8)
µ0 2l + 2
π 1 + ln πh
4l

where the parameters are as illustrated in Figure 2.3. The 2D reluc-

tance has the unit m/H and corresponds to the permeance per-unit-
length. The derivation of (2.8) has been found in [21] and is given in
Appendix A.1.

18
 2.2. AIR GAP RELUCTANCE

h
µ→ ∞

l w/2

µ→ ∞

Figure 2.3: Basic geometry for air gap calculation.

Basically, one can think of three types of air gap shapes, which are
illustrated in Figure 2.4. All air gap types can be seen as an assembly
of several basic geometries of Figure 2.3. Figure 2.4 shows how to put
together basic geometries to achieve the designated air gap type. In
addition it shows how to set the values for w and l of (2.8). The value h
represents the distance from the air gap edge to the next core corner, as
e.g. illustrated in Figure 2.11. To clarify the idea, another illustration
is given in Figure 2.5, where it is shown how to assemble several basic
geometries in order to build an air gap type 2.
In order to consider the three dimensionality of air gaps, a fring-
ing factor is introduced which describes by which factor the air gap
reluctance decreases due to fringing flux comparing to the idealized re-
luctance of (2.3). This approach is introduced using the example of the
air gap illustrated in Figure 2.7. The 3D air gap of Figure 2.7 is named
a type 1/2 air gap, as it is based on an air gap type 1 (xz-plane) and
an air gap type 2 (yz-plane). Under the assumption that the air gap is
infinitely long in the x-direction (t → ∞), the fringing effects at the air
gap boundaries in the x-direction can be neglected. This air gap has a
cross-section as illustrated in Figure 2.4(b) (air gap type 2). The fring-
ing factor that considers fringing effects in y-direction is determined by
0
calculating the corresponding 2D air gap reluctance Ryz (as illustrated
in Figure 2.4(b)) and dividing it by the 2D reluctance that neglects any
fringing effects:
0
Ryz
σy = a . (2.9)
µ0 b

The fringing factor σy now describes by which factor the air gap re-

19
CHAPTER 2. MAGNETIC CIRCUIT MODELING

R’basic
b a (l = a/2) → R’
(w = b)

(a) Air gap type 1

R’basic
(l = a/2)
b a (w = b) → R’

R’basic
(l = a)
(w = b)

(b) Air gap type 2

b a

→ R’
R’basic
(l = a)
(w = b)
(c) Air gap type 3

Figure 2.4: Different types of air gaps.

luctance decreases due to fringing flux in y direction comparing to the

idealized reluctance of (2.3).
In a similar manner the fringing factor considering fringing effects in
the x-direction can also be determined. Under the assumption that the
air gap is infinitely long in the y-direction (b → ∞), the fringing effects

20
 2.2. AIR GAP RELUCTANCE

l
w/2

Figure 2.5: Illustration of how to assemble several basis geometries

in order to build an air gap type 2.

at the air gap boundaries in the y-direction can be neglected. This

air gap has a cross-section as illustrated in Figure 2.4(a) (air gap type
1). The fringing factor considering fringing effects in the x-direction is
0
determined by calculating the corresponding 2D air gap reluctance Rxz
(as illustrated in Figure 2.4(a)) and dividing it by the 2D reluctance
that neglects any fringing effects:
0
Rxz
σx = a . (2.10)
µ0 t

The fringing factor σx now describes by which factor the air gap re-
luctance decreases due to fringing flux in x direction comparing to the
idealized reluctance of (2.3).
Now, after the two 2D fringing factors have been derived, it should
be discussed how a 3D fringing factor can be derived from them. How
this is done is illustrated in Figure 2.6. In a first step, the idealized
equation of (2.3) is multiplied by the fringing factor σx . This can be
interpreted as if the air gap cross-sectional area is multiplied by 1/σx
and a new cross-sectional area is calculated that allows the reluctance
calculation with (2.3) and the consideration of the fringing flux in x-
direction. An air gap with this new cross-sectional area and no fringing
flux in x-direction has the same properties as the air gap with reduced
size but fringing flux in x-direction. When this new cross-sectional
area is now multiplied with 1/σy , the fringing flux in y-direction is,

21
CHAPTER 2. MAGNETIC CIRCUIT MODELING

1/σx 1/σy
t Ag

Figure 2.6: Illustration of how the 3D fringing factor is derived.

furthermore, taken into consideration. This is illustrated in the second

transition of Figure 2.6. With the above steps, all fringing flux is taken
into consideration.
As discussed above, the product of the fringing factors σx and σy
shows by which factor the air gap reluctance decreases due to the total
fringing flux comparing to the idealized reluctance of (2.3), i.e. the 3D
fringing factor is derived as

σ = σx σy . (2.11)

The reluctance of the air gap can then easily be calculated with (2.3)
and (2.11)
a
Rm,airgap = σ . (2.12)
µ0 · t · b
The novel approach for reluctance calculations (2.12) has been com-
pared to the approach of (2.6), and to the FEM tuned equation of [5, 27],
and to 3D FEM simulations2 . The results are given in Figure 2.8. The
results are given normalized to the "classical" approach of (2.3), i.e. nor-
malized to Rclassic . The comparisons have been made for a type 1/1
air gap with fixed parameters (leg width and depth: w = 40 mm, and
distance between air gap and next core corner h = 60 mm). With the
FEM tuned equation of [5, 27] and the new model a high accuracy is
achieved. At smaller air gaps the FEM tuned equation has a slightly
higher accuracy (presumably this is the range at which the parame-
ters have been tuned). The new model has, in return, a slightly higher
accuracy at larger air gaps.
2 FEM Software: Ansoft Maxwell 14.0

22
 2.2. AIR GAP RELUCTANCE

a
b
z b

y
x

Figure 2.7: 3D air gap type 1/2, based on air gap type 1 (xz-plane)
and air gap type 2 (yz-plane) of Figure 2.4.

1
increase of Ag [24]
(new)
0.8
FEM
Rg / Rclassic

0.6
FEM tuned [5,27]

0.4

0.2 -3 -2 -1
10 10 10
lg / w

Figure 2.8: Comparison of air gap calculation approaches.

Type 1/2 and type 1/3 air gaps have been compared to FEM sim-
ulations as well: the maximum deviation of the new model has always
been smaller than 6 %.

23
CHAPTER 2. MAGNETIC CIRCUIT MODELING

2.2.2 Reluctance of Air Gap with Round Cross-Section

The fringing factor considering fringing effects in the r-direction (polar
coordinate system) is
R0
σr = a , (2.13)
µ0 r

where a is the air gap length, r the radius of the air gap, and R0 rep-
resents the 2D air gap reluctance (this corresponds to the permeance
per-unit-angle) of half of the core leg (including fringing effects). The
accurate reluctance can then be calculated as
a
Rg = σr2 . (2.14)
µ0 r2 π

2.2.3 Adaptation of the Air Gap Reluctance Calcu-

lation Approach to Different Problems
Sometimes the geometry for which the air (gap) reluctance has to be cal-
culated doesn’t allow directly applying the novel calculation approach;
however, in such cases the introduced approach can often be adapted
to meet the problem. In the following this will be illustrated on one
example, where the reluctance model of a PCB-integrated flyback trans-
former for a 1 mm thin PFC rectifier has been calculated.
In [28] the design and implementation of an PCB-integrated flyback
transformer for a 1 mm thin PFC rectifier has been in investigated. The
ultra-flat core of the flyback transformer results in a considerable fring-
ing flux. The setup investigated and the according reluctance model are
shown in Figure 2.9. The air gap reluctances have been calculated as
presented above. However, for b  a (cf. Figure 2.9) the window reluc-
tances Rm,σ are not negligible anymore. Therefore, the flux through the
window Φσ has been calculated via window reluctances as illustrated
in Figure 2.9. These window reluctances have been calculated with a
fringing factor σx in order to consider the fringing flux in x-direction.
The structure is an "air gap" type 1, as can be seen in the xz-plane. No
other dimension for fringing calculation has to be considered. A high
accuracy has been achieved. The interested reader is referred to [28] for
more information.

24
 2.3. CORE RELUCTANCE

b dcore
N N N

Φcore Φσ

a lair Φσ
N N
Φair
la

z lb z
x y x y
(a)
Reasy/nw Rm,easy/nw Φ2 Rm,easy/nw
Φ1
Rm,hard N·I Φσ N·I N·I Rm,hard

N·I (nw-1)∙Rm,σ (nw-1)∙Rm,σ (nw-1)∙Rm,σ N·I

Rm,air N·I N·I N·I Rm,air

(b) Rm,easy/nw Rm,easy/nw Φ2 Rm,easy/nw Φnw

Figure 2.9: (a) Setup of a PCB-integrated inductor; nw winding

packages, each with N turns, are placed around the core in order to
facilitate an interleaving with a secondary winding. The reluctances
Rm between the long magnetic rods have to be considered in the
inductance calculation as they have considerable impact on the in-
ductance for long cores lengths b. (b) Reluctance model of the setup
presented in (a). Figure taken from [28].

2.3 Core Reluctance

In addition to the dominating air gap reluctance, the core reluctances
have to be determined in order to achieve the complete reluctance
model. The reluctance of a core section i can be calculated as
li
Rci = , (2.15)
µ0 µr Ai
where li and Ai are the length and cross-section of section i. Hence,
for every section the magnetic path length and the cross-sectional area
have to be calculated first. This is difficult for corner sections. Practical

25
CHAPTER 2. MAGNETIC CIRCUIT MODELING

d II

c III I

e u e t

Section li Ai
I u c·t
II d e·t
2π (e+c) π t(e+c)
III 4 · 4 = 8 (e + c) 2

Figure 2.10: Approximations of core dimensions for core reluctance

calculation [22].

cores have usually sharp corners, where the flux tends to concentrate
on the inside the bend, so shortening the mean magnetic path.

Sophisticated ways of describing Ai and li of corner sections exist,

e.g. introduced in [9]. However, it is unclear whether the difficult for-
mulae of [9] lead to better results as they neglect the fact that the flux
density concentrates mainly at the inner bend. In any case, the corners
represent only a minor part of the core. Therefore, for this work the
simple approximations from [22] have been taken. They are given in
Figure 2.10.

The accuracy can be further improved by taking the non-linearity

of the core material into consideration, i.e. µr = f (H). This is not
straightforward: since the flux depends on the nonlinear B-H character-
istics, not on a linearization of it represented by reluctance. This prob-
lem must be solved iteratively by using a numerical solving method,
e.g. Newton’s method.

26
 2.4. EXPERIMENTAL RESULTS

h
w lg

Figure 2.11: Built E core for inductance calculation and measure-

ment comparison.

2.4 Experimental Results

This section presents some experimental results that illustrate well the
accuracy that can be achieved.

2.4.1 Inductance
In Table 2.1 measurements and calculated inductance values L of an in-
ductor built of two E-Cores (EPCOS ferrite N27; core E55/28/21 [29];
winding turns N = 80) are given. The core has three air gaps of length
lg as illustrated in Figure 2.11. The inductances for different air gap
lengths have been measured. The calculation has been performed with
the idealized "classic" approach (2.3) and the newly derived approach
(2.12). As can be seen, it is not appropriate to neglect the fringing
flux; the classic approach leads to an underestimation of the induc-
tance value. The calculation with the newly derived approach leads to
accurate results.

2.4.2 Saturation
The saturation current Isat is another very important design parame-
ter. A current that is higher than the saturation current would result
in a flux density above the saturation flux density Bsat . This would
result in a substantial decrease of the relative permeability µr , hence

27
CHAPTER 2. MAGNETIC CIRCUIT MODELING

Table 2.1: Measurement Results of E-Core

Air Gap Length Calculated Calculated with Measured

lg classically (2.3) new approach (2.12)
1.0 mm 1.42 mH 1.97 mH 2.07 mH
1.5 mm 0.96 mH 1.47 mH 1.58 mH
2.0 mm 0.72 mH 1.22 mH 1.26 mH

Table 2.2: Measurement Results of E-Core

Calculated Calculated with

classically (2.3) new approach (2.12)
L 2.75 mH 3.55 mH
Isat 4.6 A 3.6 A

Figure 2.12: Measurement of saturation current.

the inductance would drop. The saturation current can be calculated

as
Bsat Ae N
Isat = . (2.16)
L
An inductor built of two E-cores with an air gap in the center leg
(EPCOS ferrite N27; core E55/28/21 [29]; air gap length lg = 1 mm;
winding turns N = 80) has been built. In Table 2.2, inductance values
and saturation currents are given. The saturation flux density is taken
from the data sheet B-H-curve [29] and is approximately Bsat = 0.45 T
at room temperature (25 ◦ C). The results can be compared to a mea-

28
 2.4. EXPERIMENTAL RESULTS

surement for which the measured current waveform is given in Fig-

ure 2.12. A rectangular voltage waveform has been applied across the
inductor to achieve this current waveform. The measured saturation
current Isat is approximately 3.7 A, which corresponds well to the cal-
culated value.

29
Chapter 3

Core Loss Modeling

There are basically three physical core loss mechanisms: (static) hys-
teresis losses, eddy-current losses, and a third loss component which
is often referred to as residual losses. Hysteresis losses are linear with
the frequency f (rate-independent B-H loop). Eddy-current core losses
occur because of an induced current due to the changing magnetic field
and are strongly dependent on the material conductivity and the core
geometry. The residual losses are, according to [30], due to relaxation
processes: if the thermal equilibrium of a magnetic system changes,
the system progressively moves towards the new thermal equilibrium
condition. When the magnetization changes rapidly, as for example is
the case in high-frequency or pulsed field applications, such relaxation
processes become very important.
Core losses have to be described as a function of the peak-to-peak
flux density ∆B, frequency f , DC premagnetization HDC , temperature,
core shape, flux waveform, and material. This is not a simple task;
predicting core losses is challenging. This chapter shows ways how a
high accuracy in core loss modeling can be achieved.
In Section 3.1 physical origins of the core losses are summarized.
In Section 3.2 state-of-the-art means of core loss modeling are shown.
Later, in Section 3.3, needs for improvement are pointed out and novel
models and approaches for core loss calculation that have been derived
within the course of this thesis are outlined. A test system has been
built to perform studies on core losses and to be able to fully char-
acterize the loss behavior of different materials. The test system and

31
CHAPTER 3. CORE LOSS MODELING

an accuracy analysis are given in Section 3.4 and Section 3.5. A loss
model that allows one to consider relaxation effects in magnetic mate-
rials is introduced in Section 3.6. In Section 3.7 a graph is introduced
that allows one to consider that core losses may vary under a DC bias
condition when modeling core losses. Another important aspect of core
loss modeling is how to consider the presence of major and minor loops;
this is discussed in Section 3.8. In order to improve the model accu-
racy a loss database has been built up. The applied core loss approach
can then be seen as a hybrid of an improved version of the empirical
Steinmetz equation and an approach based on a material loss database
(loss map). This hybrid core loss calculation approach is introduced in
Section 3.9. Another aspect to be considered in core loss calculation is
the effect of the core shape and size, which is discussed in Section 3.10.

3.1 Physical Origin of Core Losses

Diamagnetism, paramagnetism, ferromagnetism, anti-ferromagnetism
and ferrimagnetism constitute five general groups into which materials
can be classified according to their magnetic properties. Diamagnetic
materials have a relative permeability µr less than unity; paramagnetic
materials have a relative permeability µr greater than unity; however,
both material groups have a relative permeability µr close to unity.
Antiferromagnetic materials behave similarly to paramagnetic mate-
rials; however, their underlying magnetic structure differs a lot from
that of paramagnetic materials; hence they have a separate classifica-
tion. Ferromagnetic and ferrimagnetic materials have a relative per-
meability µr much higher than unity and therefore, these materials be-
come very interesting for various technical applications. Ferromagnetic
and ferrimagnetic materials are often named simply magnetic materi-
als. Magnetic materials can be further divided into the two subgroups
soft magnetic materials and hard magnetic materials, depending on the
coercivity of the magnetic material, i.e. depending on the shape of the
B-H loop. The coercivity Hc is defined as the magnetic field required
to bring the magnetization of the material back to zero after the ma-
terial has been saturated. A material with low coercivity is called a
soft magnetic material, while a material with high coercivity is called
a hard magnetic material. In this thesis, only soft magnetic materials
are considered. Other materials, such as diamagnetic, paramagnetic,
antiferromagnetic or hard magnetic materials are not discussed further.

32
 3.1. PHYSICAL ORIGIN OF CORE LOSSES

The interested reader is referred to [31, 32] for more information about
the different material groups.
The rotation of the electron around its nucleus as well as its spin
result in a magnetic moment. The magnetic moment per unit volume is
called the intensity of magnetization. In ferro- and ferrimagnetic mate-
rials, the atoms interact in a way that the materials show spontaneous
magnetization at room temperature, i.e. the material sample is spon-
taneously magnetized even when no external field is applied. However,
to keep the system in a minimum energy state, the sample is divided
into (Weiss) domains of different magnetizing directions, that, together
have a net magnetization of zero. In other words, although the ma-
terial consists of magnetically saturated domains, the material sample
has on a macroscopic scale a net magnetization of zero in the case that
no external magnetic field is applied. These domains are separated by
domain walls (or Bloch walls). In the case that an external magnetic
field is applied to the material sample, the domain walls are shifted
or the magnetic moment within domains change their direction; hence,
the macroscopic net magnetization becomes greater than zero. The sit-
uations without an external field and the magnetizing process by the
movement of the domain walls and the rotation of the magnetic mo-
ments due to an external applied field are illustrated in the Figure 3.1
(a) and (b) respectively. A very important material parameter of ferro-
and ferrimagnetic materials is the Curie temperature TCurie above which
the material becomes paramagnetic (no spontaneous magnetization oc-
curs anymore).
The magnetization process leads to a magnetization curve such as
shown in Figure 3.2, where the magnetic flux density B versus the
external applied magnetic field H is plotted. The flux change in the
magnetic material is partly irreversible, i.e. energy is dissipated as heat.
Even when the loop is traversed very slowly, i.e. in a quasistatic manner,
losses occur. These losses originate in rapid jumps of the domain walls,
the so called Barkhausen jumps. The large local flux changes due to
Barkhausen jumps result in eddy currents that are located in the region
of the jumps and, consequently, in losses. The energy per unit volume
of a magnetic material with intensity of magnetization M and external
field H is −H · M. Consequently, the energy needed to change the
intensity of magnetization from M1 to M2 is
Z M2
W = H dM. (3.1)
M1

33
CHAPTER 3. CORE LOSS MODELING

Domain walls
Domains

Hext = 0 Hext

Magnetization direction
(a) (b)

Figure 3.1: The magnetization process of a ferro- or ferrimagnetic

sample: (a) illustration of domain walls in case no external field is
applied; (b) illustration of the magnetizing process by the movement
of the domain walls and the rotation of the magnetic moments due
to an external applied field.

B [T]

Bsat µ0

Br

-Hc Hc H [A/m]

-Br

Figure 3.2: Typical magnetization curve (hysteresis loop) of a soft

magnetic material.

If the process would be fully reversible, going from M1 to M2 and

back would store potential energy in the magnetic material that is later
released. In a plot similar to Figure 3.2, but with M instead of B for the
ordinate scale, the area of the enclosed loop would be zero. However,
as the process is partly irreversible, in reality a hysteresis loop
H with an
area representing the energy loss per cycle appears. Since H dH = 0,

34
 3.2. EXISTING APPROACHES FOR CORE LOSS CALCULATION

the loss energy per closed loop can be written as

I
W = H dB, (3.2)

i.e. the area of the closed loop in Figure 3.2 represents the energy per
unit volume that is dissipated as heat when going around the loop.
It is important to note that the losses due to eddy currents that are
induced around Barkhausen jumps are not the ones commonly named
eddy-current core losses. The term "eddy-current core losses" has a dif-
ferent meaning. Eddy-current core losses are referred to the losses that
originate from Maxwell’s equation when the presence of magnetic do-
mains is ignored, i.e. only the macroscopic net magnetization is consid-
ered. A discussion about different issues to be considered about (classic)
eddy-current core losses is given in Section 3.10.3. Residual losses are
another loss phenomenon. These losses are, according to [30], due to
relaxation processes inside the magnetic material; state-of-the art mod-
els normally neglect effects related to relaxation phenomena. These
losses will be discussed in Section 3.6 and a new core loss modeling
approach that takes relaxation effects into consideration is introduced.
The discussion above about physical origins of core losses builds a brief
summary from [5, 30, 31, 32, 33]. The interested reader is referred to
these references for more information about the physics of core losses.

3.2 Existing Approaches for Core Loss Cal-

culation
The most used equation for characterizing core losses is the empirical
power equation [22]
Pv = kf α B̂ β (3.3)
where B̂ is the peak induction of a sinusoidal excitation with frequency
f , Pv is the time-average power loss per unit volume, and k, α, β are
material parameters that have to be empirically determined. The equa-
tion is called the Steinmetz Equation (SE). The material parameters k,
α, and β are accordingly referred to as the Steinmetz parameters. They
are valid for a limited frequency and flux density range. Core manu-
facturers often provide data of losses per volume (or per weight) as a
function of frequency f , flux density B̂, and temperature. The Stein-
metz parameters can be extracted out of this data. Sometimes, the

35
CHAPTER 3. CORE LOSS MODELING

Steinmetz parameters are quoted. However, the SE is, in many circum-

stances, not capable of accurately calculating core losses because
I the material is usually exposed to non-sinusoidal flux waveforms
in power electronic applications,
I and a potential DC premagnetization HDC is not considered.

Different approaches have been developed to overcome this limita-

tion and determine losses for a wider variety of waveforms. The ap-
proaches can be classified into the following categories:
1. Improvements of the Steinmetz equation (3.3): for instance, the
analysis in [11] is motivated by the fact that the loss due to do-
main wall motion has a direct dependency of dB/dt. As a result,
a modified Steinmetz equation is proposed. In [34] the approach
is further improved and in [35] a method how to deal with mi-
nor hysteresis loops is presented and some minor changes to the
equation are made. The approach of [11], [34], and [35] leads to
the improved Generalized Steinmetz Equation (iGSE)

1 T 
Z
 dB α

Pv = ki   (∆B)β−α dt (3.4)
T 0 dt
where ∆B is the peak-to-peak flux density and
k
ki = R 2π . (3.5)
(2π)α−1 0
| cos θ|α 2β−α dθ
The parameters k, α, and β are the same parameters as used in
the Steinmetz equation (3.3). By use of the iGSE losses of any
flux waveform can be calculated, without requiring extra charac-
terization of material parameters beyond those for the Steinmetz
equation. This approach is widely applied [6, 36]. If one inserts
a sinusoidal flux density waveform into the iGSE, (3.4) trans-
forms back to the Steinmetz equation (3.3). This is shown in
Appendix A.2.
2. Calculation of the losses with a loss map that is based on mea-
surements. This loss map stores the loss information for different
operating points, each described by the flux density ripple ∆B,
the frequency f , the temperature, and a DC bias HDC (e.g. in
[37, 38, 39]).

36
 3.2. EXISTING APPROACHES FOR CORE LOSS CALCULATION

vL ,iL

vL iL

T
t
T/2

I II III

Figure 3.3: Typical voltage/current waveform of magnetic compo-

nents employed in power electronic systems. Phase I: positive voltage;
phase II: zero voltage; phase III: negative voltage.

3. Methods to determine core losses based on breaking up the total

loss into loss components, i.e. static hysteresis losses, classical
eddy current losses, and residual losses [30, 33, 40].

4. Hysteresis models such as Preisach and Jiles-Atherton used for

calculating core losses.

The approaches of loss separation (category 3) and hysteresis models

(category 4) have a practical disadvantage: such models are based on
parameters which are not always available and are difficult to extract.
However, the approach of loss separation has its relevance, as it gives
a deeper understanding about physical core loss mechanisms. In the
categories 1 and 2 an energy loss is assigned to each section of the
voltage / current waveform as illustrated in Figure 3.3 (e.g. via an
equation as (3.4) or via a loss map), and these losses are summed up
to calculate the power loss occurring in the core. In the course of this
thesis, an approach that can be seen as a hybrid of an improved version
of the empirical Steinmetz equation (category 1) and an approach based
on a loss map (category 2) has been developed; hence, the main focus
is placed on these categories. However, to physically justify the models,
the approach of loss separation (category 3) will be often referred to.

37
CHAPTER 3. CORE LOSS MODELING

3.3 Outline of Novel Core Loss Calculation

Models and Approaches
Existing core loss models have been summarized in the previous sec-
tion. This section now points out unconsidered effects related to core
losses modeling; and an outline of novel core loss calculation models
and approaches, that have been derived in the course of this thesis, is
given.
In modern power electronic systems voltages across inductors or
transformers generally show rectangular shapes, including periods of
zero voltage. In most core loss models, the phase where the voltage
across the magnetic component is zero (i.e. the flux remains constant)
is not discussed. It has been implicitly assumed that no losses occur
when the flux remains constant. However, as measurements show, this
is not a valid simplification. In phases of constant flux, losses still oc-
cur in the material. In the publication [41] about core loss modeling,
a loss increase during zero voltage periods has been observed; but no
explanation or modeling approach is given. Losses still occur a short
period after switching the winding voltage to zero due to magnetic re-
laxation. A further improvement of the iGSE that takes this effect into
consideration is suggested in Section 3.6. The new model is named the
improved-improved Generalized Steinmetz Equation, or simply i2 GSE.
The model equation is
Z T  dB α n
1 X
Pv = ki   (∆B)β−α dt + Qrl Prl , (3.6)
 
T 0 dt
l=1

where Prl is calculated for each stepped voltage change according to

1  αr
d
 t1

Prl = kr  B(t) (∆B)βr 1 − e− τ , (3.7)

T dt
and Qrl is a function that further describes the voltage change and is
 
 dB(t )/dt 
−qr  dB(t+ )/dt 
−
Qrl = e , (3.8)

and α, β, ki , αr , βr , kr , τ , and qr are material parameters.

The approach of (3.6) attempts to solve the problem of different flux
shapes. However, (3.6) does not consider a potential DC premagneti-
zation HDC . For many materials, the impact of a DC bias cannot be

38
 3.4. TEST SETUP TO MEASURE CORE LOSSES

neglected as it may increase the losses by a factor of more than two. An

approach to describe core losses under DC bias condition is introduced
in Section 3.7: a graph that shows the dependency of the Steinmetz
parameters (α, β and ki ) on premagnetization is proposed. The graph
is named the Steinmetz Premagnetization Graph, SPG. This graph en-
ables the calculation of losses via the Steinmetz equation (3.3), the iGSE
(3.4), or the i2 GSE (3.6) using appropriate Steinmetz parameters.
A core loss calculation with the models i2 GSE and SPG described
above has one major drawback: the models are based on parameters
that cannot be extracted from the data provided by the core manu-
facturers. Until core manufacturers provide data to extract the depen-
dency of a DC bias, e.g. an SPG, or provide data with which to predict
relaxation effects, e.g. relaxation parameters αr , βr , kr , τ , and qr of
(3.7) and (3.8), core loss measurements must be performed for an accu-
rate core loss calculation. The measurement results could be stored in
a core material database, which enables one to consider all the effects
described above when designing inductive components. This is one aim
of this thesis: to determine a structure for a core material database
(later named loss map) to permit accurate core loss calculations. The
novel structure is given in Section 3.9.
However, for the purpose of the above introduced studies a core
loss measurement test setup has been built and is described next in
Section 3.4.

3.4 Test Setup to Measure Core Losses

To perform measurements, the best measurement technique has to be
selected first. In [42] different methods are compared. The B-H Loop
Measurement has been evaluated as the most suitable. Amongst other
advantages, this technique offers rapid measurement (compared to other
methods, e.g. calorimetric measurement), copper losses are not mea-
sured, and a good accuracy. In Section 3.5 the accuracy is analyzed in
detail. The principles are as follows: two windings are placed around
the Core Under Test (CUT). The sense winding (secondary winding)
voltage v is integrated to sense the core flux density B
Z t
1
B(t) = v(τ )dτ (3.9)
N2 · Ae 0

39
CHAPTER 3. CORE LOSS MODELING

Power Supply

Oscilloscope 325V

Power Stage
Current
Probe

Voltage CUT
Probe

Heating Chamber

Figure 3.4: Overview of the test system.

Oscilloscope LeCroy WaveSurfer 24MXs-A

Current Probe LeCroy AP015
Heating Chamber Binder ED53
Power Supply Xantrex XTR 600-1.4
Power Stage 0 − 450 V
0 − 25 A
0 − 200 kHz

Table 3.1: Measurement Equipment

where N2 is the number of sense winding turns and Ae the effective

core cross section of the CUT. The current in the excitation winding
(primary winding) is proportional to the magnetic field strength H

N1 · i(t)
H(t) = (3.10)
le

where N1 is the number of excitation winding turns and le the effective

magnetic path length of the CUT. The loss per unit volume is then the

40
 3.5. ACCURACY OF THE MEASUREMENT SYSTEM

enclosed area of the B-H loop, multiplied by the frequency f 1

I
P
=f HdB. (3.11)
V

The selected approach is widely used [35, 38, 43, 44]. The test system
consists of an oscilloscope, a power supply, a heating chamber, and a
power stage, as illustrated in Figure 3.4. It is controlled by a MAT-
LAB program running on the oscilloscope under Microsoft Windows.
In Table 3.1 the used equipment is listed. In Figure 3.5 a photograph
(a) and the simplified schematic (b) of the power stage is shown. The
power stage has been designed and built in the course of this thesis.
In Table 3.2 the most important components employed in the power
stage are listed. The power stage is capable of a maximal input voltage
of 450 V, output current of 25 A and a switching frequency of up to
200 kHz. With the power stage, it is possible to achieve rectangular
voltage shapes (including phases of zero voltage) across the CUT that
leads to triangular or trapezoidal current shapes including DC bias (if
desired). This behavior is illustrated in Figure 3.6. To control the
DC current, the current is sensed by a DC current transducer. A low
frequency sinusoidal excitation is also possible; for this an output fil-
ter has been designed to achieve a sinusoidal current/voltage shape for
frequencies up to 1 kHz.

3.5 Accuracy of the Measurement System

The different aspects that influence the accuracy of the measurements
are given in the following.

1 The core loss per unit volume is

RT RT dB(t)
P f i1 (t) N
N
1
v2 (t)dt f H(t)le Ae dt
dt
0 2 0
= =
V Ae le Ae le
Z B(T ) I
=f H(B) dB = f HdB,
B(0)

N1
where v (t)
N2 2
is the sense winding voltage transformed to the primary side.

41
CHAPTER 3. CORE LOSS MODELING

(a)

+
S1 S3

CUT
V1 A

V
S2 S4
-
(b)

Figure 3.5: Power stage (a) photograph, (b) simplified schematic.

v, i
iCUT
vCUT
iDC

T
t
T/ 2

Figure 3.6: Current and voltage waveforms of the CUT.

42
 3.5. ACCURACY OF THE MEASUREMENT SYSTEM

Power MOSFETs IXYS IXFB82N60P

Gate Driver IXYS IXDD414SI
Capacitors Electrolytic: 2.75 mF
Foil: 360 µF
Ceramic: 3.86 µF
DSP TI TMS320F2808
Current Sensor LEM LTS 25-NP
Fan San Ace 40 GE

Table 3.2: Power stage components.

3.5.1 Phase Shift Error of Voltage and Current Mea-

surement
The error due to an inaccurate measurement of the voltage and current
phase displacement can be quantified as [45]
cos(ζ + φ) − cos ζ
E = 100 · , (3.12)
cos ζ
where E is the relative error in % of the measured core losses, ζ is the
actual phase shift between the sense winding output voltage and the
excitation winding current, and φ is the error in the measurement of ζ.
Measurements have shown that φ (over some frequency range) depends
linearly on the frequency. In other words, φ originates from a delay
time Td that is independent of the frequency. This delay time Td can
be measured with a rectangular current shape through a low inductance
shunt, and with it the delay time can be compensated. The main cause
of the delay time Td is the current probe.
Different measurements with the material ferrite N87 from EPCOS
(core part number: B64290L22X87) are presented in this thesis; there-
fore a short discussion about phase shift accuracy is given using the
example of this core. This accuracy discussion is similar to the dis-
cussion presented in [12]. In Figure 3.7 a simplified equivalent circuit
of the Core Under Test (CUT) is given. Winding losses and leakage
inductance are assumed to be negligible. The reactance Xm can be
calculated as
Xm = ωAL N12 , (3.13)

43
CHAPTER 3. CORE LOSS MODELING

iCUT,prim

RFE Xm vCUT,sec

Figure 3.7: Equivalent circuit of the CUT.

where AL is the inductance factor, N1 is the number of primary winding

turns, and ω = 2πf is the angular frequency. Hence, for the CUT (AL =
2560 nH [29], N1 = 10) and a frequency f = 100 kHz the reactance is
Xm = 160.8 Ω. At the operating point ∆B = 100 mT (peak-to-peak),
f = 100 kHz, and T = 40 ◦ C losses of PLoss = 0.2 W are expected (from
material data sheet [29]). With this information, the equivalent resistor
RFE that represents the core losses can be calculated
 2
∆B
V 2 N 1 A e ω √
2 2
RFE = rms = , (3.14)
PLoss PLoss
where Ae is the effective core cross section. For the CUT and operating
point, the resistor RFE is 2.26 kΩ. Now, the angle ζ can be calculated
as
RFE
ζ = arctan = 85.9 ◦ . (3.15)
Xm
An uncompensated delay time Td would result in a phase shift error
of voltage and current measurement of

φ = f · Td · 360 ◦ . (3.16)

When (3.16) is inserted in (3.12) and then solved for Td , a tolerable

uncompensated delay time for a desired accuracy is derived; e.g. for an
accuracy of ±3 %, an uncompensated delay time of ±3.5 ns at 100 kHz
and ζ = 85.9 ◦ would be tolerable. Measurements have shown that the
delay time compensation leads to lower residual delay times; although
a quantification is difficult. With a realistic delay time compensation
to an accuracy of ±1.5 ns, and with an expected system accuracy (only
phase shift consideration) of ±4 %, measurements of materials up to
an angle of ζ = 88.7 ◦ (at f = 100 kHz) can be performed. For lower
frequency measurements the permitted angle ζ increases for the same

44
 3.5. ACCURACY OF THE MEASUREMENT SYSTEM

accuracy constraint; e.g. at 20 kHz, for an accuracy of ±4 %, measure-

ments up to an angle of ζ = 89.7 ◦ are permitted. All measurements
presented in the next sections are within this range.
The system has one drawback related to the phase shift: the mea-
surement of gapped cores (or low permeability cores) is difficult because
the angle ζ substantially increases in this situation. A detailed analysis
together with a new method of how gapped cores could be measured is
introduced in [13]. Two measurement methods that improve core loss
measurement for very high frequencies (up to 70 MHz) are proposed in
[46] and [47]. Although the focus of [46] and [47] are on measurements
at very high frequencies, the method could be used to improve the loss
measurement of gapped cores.

3.5.2 Equipment Accuracy

A typical magnitude/frequency characteristic of the current probe has
been provided by the current probe manufacturer LeCroy, from which
an AC accuracy of 3 % could be extracted. Together with the accuracy
of the passive probe (attenuation accuracy of 1 %) and the accuracy
of the oscilloscope itself (1.5 % that originates amongst others from the
limited vertical resolution of 8 bit), an equipment accuracy of ≤ |±5.6 %|
is calculated.

3.5.3 Capacitive Coupling

Capacitive currents may result in errors and must therefore be avoided.
The typical capacitances that are present in windings are
I capacitance between the primary and secondary winding (inter
capacitance),
I self capacitance between turns of a winding (intra capacitance),

I and capacitance between the windings and the magnetic core.

Generally, the inter and intra capacitances increase with increasing area
between the windings and decrease with distance between the windings.
To decrease the inter capacitance, a separation of the primary and sec-
ondary windings is favorable; although a separation of the windings
avoids an absolutely uniform winding distribution around the core (ide-
ally, the primary winding should be distributed uniformly around the

45
CHAPTER 3. CORE LOSS MODELING

core to achieve a homogenous flux density distribution). Another im-

portant aspect of the winding arrangement is the chosen number of
turns of the primary winding. Even with the use of favorable winding
layout, some ringing in current and voltage is inevitable. Fewer turns
are more favorable for two reasons: this additionally decreases parasitic
capacitances, and, because the current for the same magnetic operat-
ing point is higher, capacitive currents are relatively lower compared to
(desired) inductive currents.

3.5.4 Temperature
An important aspect is that the temperature of the CUT is defined and
constant. To keep the temperature constant, the test system performs
the measurement automatically (starts excitation, controls current, reg-
ulates flux (∆B), triggers the oscilloscope, reads values). With such an
automated measurement system, a working point is rapidly measured
and the losses do not heat the core in the short measurement period.

3.5.5 Comparative Measurement and Conclusion

Comparative measurements with the power analyzer Norma D6100
have been performed to confirm the accuracy. The power analyzer
is connected to measure the excitation winding current and the sense
winding voltage to obtain the core losses [42]. The results in the per-
formed working points matched very well. The deviation between the
results of the test system and of the power analyzer was always ≤ |±4 %|
(measured up to 100 kHz).
From the equipment accuracy (≤ |±5.6 %|) and the phase shift ac-
curacy (≤ |±4 %|), a system accuracy of ≤ |±9.8 %| is calculated. How-
ever, based on the results of the comparative measurements, it can be
said that the accuracy achieved is higher.
As a conclusion, a test system has been built up that performs the
measurements quickly and leads to sufficiently accurate results.

3.6 Relaxation Effects in Magnetic Mate-

rials
In modern power electronic systems, voltages across inductors or trans-
formers generally show rectangular shapes, including periods of zero

46
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

BL,vL

∆B t1 t2
t
dB(t)/dt vL
BL

Figure 3.8: Voltage and flux density waveforms.

voltage. In most core loss models, the phase where the voltage across
the magnetic component is zero (i.e. the flux remains constant) is not
discussed. It has been implicitly assumed that no losses occur when the
flux remains constant. However, as measurements show, this is not a
valid simplification. As already mentioned in Section 3.3, during phases
of constant flux (i.e. where the voltage across the magnetic component
is zero) losses still occur in the core material. A literature survey led to
the hypothesis that this is due to relaxation processes in the magnetic
core material. In this section, first, measurements are presented that
illustrate magnetic relaxation. Further, an attempt to theoretically ex-
plain the effect is given, and, with it, the resulting shape of a B-H loop
for a trapezoidal flux waveform is analyzed.

3.6.1 Measurement Results

According to (3.4), the energy loss would only depend on the magni-
tude and the slope of the flux, and consequently, there should be no
loss during periods of constant flux (zero voltage). Measurements on
waveforms as illustrated in Figure 3.8 have been performed to inves-
tigate this. Figure 3.9 shows the corresponding measurement results.
The CUT is made of ferrite EPCOS N87 (size R42). According to (3.4),
the duration of t1 should not influence the energy loss per cycle, but,
as can be seen, increasing t1 has a substantial influence on the energy
loss per cycle. In particular, a change in t1 at low values of t1 influences
the dissipated loss. For larger values of t1 , the core material has time
to reach its equilibrium state and no increase in losses can be observed
when t1 is increased further.
Different experiments have been conducted to confirm that this ef-

47
CHAPTER 3. CORE LOSS MODELING

6
2.∆E
5

4 ∆B = 50mT, t = 10µs
2
E [µJ]

∆B = 100mT, t = 10µs
3 2
∆B = 100mT, t2 = 5µs
2

0
0 τ 10 20 30 40 50 60
t [µs]
1

Figure 3.9: Measurement results for ferrite EPCOS N87 (R42,

B64290L22X87 [29]); temperature = 25 ◦ C. It is further illustrated
how τ according to (3.28) can be extracted.

fect is not due to an imperfection of the measurement setup. One could

mainly think of two sources of error:

1. The effect of a small exponential change of current due to a resid-

ual voltage across the inductor in the "zero" voltage time intervals,
i.e. an effect related to the CUT excitation.

2. An error in measurements due to limited measurement capabilities

of the probes, i.e. an effect related to the measurement equipment
used.

The following experiments have been conducted to make sure that none
of these error sources led to the observed loss increase:

1. A resistor has been connected in series with the primary winding,

i.e. the effect of an exponential change of the current due to a volt-
age drop across the inductor in the "zero" voltage time intervals
has been deliberately increased. A resistor of 10 Ω has been cho-
sen as this value is certainly higher than the residual resistance of
the setup (for instance, the on-resistance of one MOSFET (IXYS
IXFB82N60P) is only RDS(on) = 75 mΩ at Tj = 25 ◦ C; ID = 41 A).
The excitation voltage has been accordingly adjusted to have the
same magnetic operating point. The same core loss increase in
the "zero" voltage time interval has been observed as without an

48
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

additional resistor, which indicates that the effect is not coming

from an improper CUT excitation.
2. The current probe LeCroy AP015 has, with 50 MHz, the lowest
bandwidth of the measurement equipment used. This is enough
to measure the effect observed in Figure 3.9. Generally, according
to the accuracy analysis in Section 3.5, accurate measurement
results are expected for operating points with frequencies / flux
densities such as presented in Figure 3.9. However, to confirm
this, a simple comparative measurement has been performed to
verify that the effect is not originating in the limitations of the
voltage and current probes. For a limited temperature range it can
be approximated that the relative change of the core temperature
is proportional to the losses occurring in the core. According
to this, the core losses can be observed by measuring the core
temperature. The same loss increase in the zero voltage time
intervals as illustrated in Figure 3.9 could be observed by this
simple measurement. This comparative measurement indicates
that the effect is not coming from the measurement equipment
used.
Last but not least, the fact that this effect has been also observed by
another research group [41] greatly increases the credibility of the result.
Concluding, during phases of constant flux, i.e. where the voltage
across the magnetic component is zero, losses still occur. Based on a
literature survey, it is hypothesized that this is because of relaxation
processes in the magnetic core material. Next, a brief introduction
about magnetic relaxation is given.

3.6.2 Theory of Relaxation Effects

There are basically three physical loss sources: static hysteresis losses,
eddy-current losses, and a third loss component which is often referred
to as residual losses. The residual losses are, according to [30], due
to relaxation processes: if the thermal equilibrium of a magnetic sys-
tem changes, the system progressively moves towards the new thermal
equilibrium condition. When the magnetization changes rapidly, as for
example is the case in high-frequency or pulsed field applications, such
relaxation processes become very important.
The Landau-Lifshitz equation describes qualitatively the dynamics
of the magnetic relaxation processes. This is a phenomenological equa-

49
CHAPTER 3. CORE LOSS MODELING

tion that combines all processes that are involved in magnetic relax-
ation. The equation follows directly from equating the rate of change
of the angular momentum L to the torque M×H reduced by a frictional
term that is directed opposite to the direction of motion [30]:

dM
= γM × H − ΛM × (M × H)/M 2 , (3.17)
dt
where γ = ge/2mc is the magnetomechanical ratio M/L, M is the
magnetization vector, H the magnetic field vector, and Λ is called the
relaxation frequency. It describes how the system progressively moves
towards the new thermal equilibrium. The equilibrium is achieved by
rearranging the magnetic domain structures to reach states of lower
energy. The relaxation process limits the speed of flux change, hence
the B-H loops become rate-dependent. Several physical processes con-
tribute simultaneously to magnetic relaxation. The interested reader is
referred to [30, 32, 33] for more information.
Due to magnetic relaxation, the magnetization may change even
when the applied field is constant (the magnetization is delayed). Con-
sequently, a residual energy loss still occurs in the period of a constant
applied field. Furthermore, the shape of the hysteresis loop is changed
depending on the rate of change of the applied field (rate-dependent
loop). An analysis of the impact of magnetic relaxation to a trape-
zoidal flux shape now follows.

3.6.3 Shape of B-H Loop for Trapezoidal Flux Wave-

forms
A B-H loop under trapezoidal flux waveform condition has been mea-
sured to gain a better comprehension of why the losses increase when
the duration of the zero voltage period is increased. The CUT is a
toroid core R42 made of ferrite EPCOS N87. In Figure 3.10(a) the
flux waveform, in Figure 3.10(b) the corresponding B-H loop, and in
Figure 3.10(c) the corresponding current waveform are plotted. Fig-
ure 3.10(b) and (c) are measured figures. The B-H loop always traverses
counterclockwise. The different instants (cf. numbers in Figure 3.10(a),
(b) and (c)) are now discussed step-by-step:

1. A constant voltage at the CUT primary winding results in a time

linear flux increase.

50
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

Figure 3.10: (a) Voltage and flux density waveforms. (b) B-H loop
to illustrate magnetic relaxation under trapezoidal flux shape condi-
tion. (c) Current waveform.

51
CHAPTER 3. CORE LOSS MODELING

2. The CUT primary voltage is set to zero; as a consequence the

flux is frozen (dB/dt = 0). However, the material has not yet
reached its thermal equilibrium. The magnetic field strength H
decreases in order to move towards the new thermal equilibrium
and therewith reaches a state of lower energy. This can also be
observed in the current (the current declines accordingly between
point 2 and 3).

3. This point is reached approximately 24 µs after point 2. It is the

point of the new thermal equilibrium.

4. This point is reached approximately 200 µs after point 3. The

demagnetization in the zero voltage period is due to the small
voltage drop over the on-resistance of the MOSFETs and copper
resistance of the inductor primary winding. This demagnetization
follows a different time constant than the demagnetization due to
relaxation losses (cf. the approximately same distance 2-3 and
3-4, but the different time scale). This demagnetization can be
observed in Figure 3.10(c) between point 3 and 4. In a measured
flux waveform, this demagnetization could be observed too; how-
ever, Figure 3.10(a) is an illustrative figure which neglects this
demagnetization, therefore, it cannot be seen there. At point 4
a negative voltage is applied to the CUT. The small buckle in
the B-H loop is due to small capacitive currents at the switching
instant.

The period between point 2 and 3 obviously increases the area of

the B-H loop, and therewith increases the core losses. The loop area
increases as a function of the duration t1 . After the thermal equilibrium
is reached (in the above example after approximately 24 µs) the loss
increase becomes (almost) zero. In the next section more measurements
are presented to find a method to include this effect into an existing
core loss model.

3.6.4 Model Derivation

Model Derivation 1: Trapezoidal Flux Waveform
Losses can be calculated with the iGSE (3.4), without requiring extra
characterization of material parameters beyond the parameters for the
Steinmetz equation. The Steinmetz parameters are often given by core

52
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

∆B = 100mT, t1 = 0
0
10
∆B = 100mT, t = 60µs
1
∆B = 200mT, t1 = 0
∆B = 200mT, t = 60µs
1
P [W]

-1
10

-2
10
4000 8000 12000 16000 20000
dB/dt [T/s]

Figure 3.11: Core Loss (ferrite N87; measured on R42 core); tem-
perature = 25 ◦ C.

manufacturers, hence core loss modeling is possible without performing

extensive measurements. However, the approach has some drawbacks.
First, it neglects the fact that core losses may vary under a DC bias con-
dition. This will be discussed in Section 3.7, where a graph showing the
dependency of the Steinmetz parameters (α, β and k) on premagnetiza-
tion is introduced. With it, losses can be calculated via the Steinmetz
equation (3.3) or the iGSE (3.4) using appropriate Steinmetz param-
eters. Another source of inaccuracy is that relaxation effects are not
taken into consideration. As approach (3.4) is very often discussed in
literature and often applied for designing magnetic components, im-
proving this method would have the most practical use. Furthermore,
in [6] it has been evaluated as the most accurate state-of-the-art loss
model based on Steinmetz parameters. For this two reasons, in the
following discussion the iGSE will be extended to consider relaxation
losses as well.
When plotting the losses with logarithmic axes, where the x-axis
represents the frequency and the y-axis represents the power loss, an
approximately straight line is drawn. This is because the losses follow a
power function as e.g. the Steinmetz equation (3.3) is. The parameter
α of (3.3) represents the slope of the curve in the plot. In Figure 3.11

53
CHAPTER 3. CORE LOSS MODELING

such plots are given for few operating points. Instead of the frequency
f , dB/dt has been used as x-axis, which, for symmetric triangular or
trapezoidal flux waveforms, is directly proportional to the frequency f .
The time t1 is defined as in Figure 3.10 (t1 = 0 leads to a triangular
flux waveform). As can be seen in Figure 3.11, when a long zero voltage
phase is added between two voltage pulses (having a flux waveform as
given in Figure 3.10) the loss still follows a power function with variable
dB/dt (the losses are still represented by an approximately straight
line). The same conclusion can be made when keeping dB/dt constant
and varying ∆B, hence, the use of a power function with variable ∆B
is justified as well.
It should be pointed out that when a (long) zero voltage interval
(t1 6= 0) is present the average power loss decreases (cf. Figure 3.11).
There is no discrepancy with the observation in Figure 3.9, where an
energy loss per cycle increase has been observed. When having a zero
voltage interval the energy loss per cycle increases, but the period in-
creases as well and leads to a lower average power loss.
The approach of (3.4) will now be extended by taking relaxation
effects into consideration. This is done by adding a new term that
represents the relaxation effect of a transition to zero voltage. As can
be seen in Figure 3.9, the energy loss increase due to the zero voltage
interval can be modeled with the exponential equation
 t1

E = ∆E 1 − e− τ , (3.18)

where ∆E is the maximum energy loss increase (which occurs, when

the magnetic material has enough time to reach the new thermal equi-
librium), τ is the relaxation time that has to be further determined, and
t1 is the duration of the constant flux (zero applied voltage) phase. The
exponential behavior is typical for relaxation processes. Measurements
have shown that τ can be considered to be a constant parameter for a
given core material that does not change for different operating points.
The increase of energy loss per cycle in measurements on waveforms
illustrated in Figure 3.10(a) (or Figure 3.8) leads to twice ∆E, since
there are two transitions to zero voltage. Consequently, in Figure 3.9
the loss increase is labeled as 2 · ∆E. Different measurements on wave-
forms as illustrated in Figure 3.10(a) have been conducted to determine
a formula to describe ∆E. The corresponding results are given in Fig-
ure 3.12, where measured values of 2 · ∆E for different operating points
are plotted. In Figure 3.12(a) dB/dt has been used as x-axis and in

54
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

3.0 ∆B = 50mT
∆B = 75mT
2 . ∆E [µJ] 2.0 ∆B = 100mT

1.0
0.8
0.6
0.4

0.2
1'000 5'000 10'000
(a) dB/dt [T/s]

4.0 dB/dt = 5000 T/s

dB/dt = 10000 T/s
2.0
2 . ∆E [µJ]

1.0
0.8
0.6
0.4

0.2
0.05 0.06 0.07 0.08 0.09 0.1 0.15
(b) ∆B [T]

Figure 3.12: Measured values of 2 · ∆E (ferrite N87; measured on

R42 core); temperature = 25 ◦ C.

Figure 3.12(b) ∆B has been chosen for the x-axis. In both cases ap-
proximately parallel straight lines are drawn, i.e. 2·∆E (approximately)
follows a power function with variables dB/dt and ∆B. Hence, ∆E of
one transition to zero voltage can be described by a power function with
variables ∆B and dB(t− )/dt, where ∆B and dB(t− )/dt define the flux
density waveform before this transition to zero voltage as illustrated
in Figure 3.13. As a consequence, the following power function can be
defined for ∆E:
d αr
∆E = kr  B(t− ) (∆B)βr , (3.19)
 
dt
where αr , βr , and kr are new model parameters which have to be deter-
mined empirically. With (3.19), the relaxation losses of a transition to

55
CHAPTER 3. CORE LOSS MODELING

∆B
dB(t_)/dt

t
Figure 3.13: Definition of dB(t− )/dt and ∆B.

zero voltage can be determined according to the antecedent flux den-

sity slope dB(t− )/dt and the antecedent flux density peak-to-peak value
∆B. Accordingly, when the flux density reaches and remains at zero
as occurs e.g. in a buck converter that is operating in discontinuous
conduction mode, relaxation losses have to be taken into consideration
as well. However, the losses may (slightly) differ in this situation be-
cause the antecedent flux density is DC biased. This DC level of the
antecedent flux density has not been part of investigation of the present
work and could be investigated as part of future work.
Concluding, (3.4) has been extended by an additional term that
describes the loss behavior for a transient to constant flux. This leads
to a new model to calculate the time-average power loss density
Z T  dB α n
1 β−α
X
Pv = ki  (∆B) dt + Prl , (3.20)
 
T dt

0 l=1

where Prl represents the time-average power loss density due to the lth
of n transients to zero voltage. This power loss of each transient to zero
voltage is calculated according to

1  αr
d
 t1

Prl = kr  B(t− ) (∆B)βr 1 − e− τ . (3.21)

T dt

For the sake of completeness, a limitation of the given model should

be pointed out. The curves in Figure 3.12 do not have the shape of
exact straight lines. This illustrates the fact that the newly introduced
parameters αr , βr are only valid for a limited dB(t− )/dt and ∆B range.
The limited parameter validity is a general problem of the Steinmetz
approach.

56
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

Figure 3.14: Triangular flux density waveform.

Model Derivation 2: Triangular Flux Waveform

Often in power electronics, one has a period of zero voltage applied to a

magnetic component winding, e.g. in the transformer of a bidirectional
isolated DC-DC converter with Dual Active full Bridges (DAB). A DAB
will be presented in Section 3.6.4 as an example to illustrate the model.
In this case, (3.20) can directly be used to improve the loss model.
However, another frequently occurring waveform is a triangular flux
waveform in which the flux slope changes to another nonzero value. This
case is illustrated in Figure 3.14. When a duty cycle of 50 % (D = 0.5) is
assumed, directly after switching to the opposite voltage the flux slope
reverses, the material has hardly time to move towards the new thermal
equilibrium. As a consequence, no notable loss increase is expected and
thus this case is well described by the iGSE (3.4). However, when the
duty cycle goes to smaller values, once each period, a high flux slope is
followed by a comparatively very slow flux change. Assuming D to be
infinitely small, it is like a switch to a constant flux. Consequently, in
this case the iGSE (3.4) is not accurate and the relaxation term has to
be added. In all operating points where D > 0 and D < 0.5 (or D > 0.5
and D < 1), a behavior that is in-between these two cases is expected.
In other words, only part of the relaxation term has to be added.
In Figure 3.15 the calculated and measured core losses as a function
of the duty cycle are plotted. One calculation has been performed based
on the iGSE (3.4) which, according to the above discussion, represents
the lower limit of possible losses (as no relaxation effects are taken into
account). It should be noted that two sets of Steinmetz parameters
have been used for the calculation of the iGSE. The reason is that the
Steinmetz parameters are only valid in a limited dB/dt range, and the

57
CHAPTER 3. CORE LOSS MODELING

i2GSE
Upper Loss Limits
Measured Values
P [W]

iGSE

Figure 3.15: Core loss duty cycle dependency. ∆B = 0.1 T, f =

20 kHz.

dB/dt in this experiment is varying in a wide range. This explains

the sharp bend of the iGSE curve at D = 0.15 (change of Steinmetz
parameter). Another calculation has been made always including the
full relaxation loss term and which represents the upper loss loss limit.
In other words, it can be said that losses are expected to have values
between the line representing the upper loss limit and the line represent-
ing the lower loss limit (iGSE). According to the previous discussion,
the real losses are closer to the lower loss limit for D close to 0.5, and
losses closer to the upper loss limit for D close to zero. Measurements
seem to confirm this hypothesis as can be seen in Figure 3.15. Other
operating points showed the same behavior.
Based on the above discussion, the new approach can be further
improved to be also valid for triangular flux waveforms. Basically, (3.20)
can be rewritten as
n
1 T 
Z
 dB α
 X
Pv = ki   (∆B)β−α dt + Qrl Prl , (3.22)
T 0 dt
l=1

where Qrl has to be further defined. In the case of a switch to zero

voltage, Qrl needs to have the value 1. Furthermore, it has to have a
structure such that (3.22) fits the measurement points of a duty cycle
measurement, such as illustrated in Figure 3.15. The following function

58
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

has been chosen:  

 dB(t )/dt
−qr  dB(t+ )/dt 

−
Qrl = e , (3.23)
where dB(t− )/dt represents the flux density before the switching, dB(t+ )/dt
the flux density after the switching, and qr is a new material parameter.
For a triangular waveform as illustrated in Figure 3.14, (3.23) can be
rewritten (for D ≤ 0.5)
∆B
(1−D)T
−qr D
= e−qr 1−D .
∆B
Qrl = e DT (3.24)
In the case of the material Epcos N87 qr = 16 has been found, the re-
sulting loss curve is plotted in Figure 3.15. Before giving an illustrative
example in Section 3.6.4, the new model will be summarized and the
steps to extract the model parameters will be given.

New Core Loss Model: The i2 GSE

A new loss model that substantially increases the expected accuracy
when core losses are modeled has been introduced. This new model is
named the improved-improved Generalized Steinmetz Equation, i2 GSE.
The name has been chosen because it is an improved version of the
iGSE [35]. The time-average power loss density can be calculated with
n
1 T 
Z
 dB α
 X
Pv = ki   (∆B)β−α dt + Qrl Prl , (3.25)
T 0 dt
l=1

where Prl is calculated for each voltage change according to

1  αr
d
 t1

Prl = kr  B(t− ) (∆B)βr 1 − e− τ , (3.26)

T dt
Qrl is a function that further describes the voltage change and is
 
 dB(t )/dt 
−qr  dB(t+ )/dt 
−
Qrl = e , (3.27)
and α, β, ki , αr , βr , kr , τ , and qr are material parameters.
Now, the steps to extract the model parameters are given:
1. First, the parameters ki , α, and β are extracted. The core is
excited with a rectangular voltage waveform that leads to a sym-
metric triangular flux waveform. Measurements at three oper-
ating points are performed, then (3.25) is solved for the three

59
CHAPTER 3. CORE LOSS MODELING

parameters. For symmetric

Pn triangular flux waveforms (with duty
cycle D = 0.5) it is l=1 Qrl Prl = 0. In Table 3.3 the measure-
ment results and the corresponding parameters are given. These
parameters could be extracted directly from the data sheet as
well, as explained in [35].

2. The parameter τ can be read from Figure 3.9 with

∆E dE
= , (3.28)
τ dt
where dE/dt represents the slope of the energy increase directly
after switching to zero voltage. This is illustrated in Figure 3.9.
τ = 6 µs has been extracted for the material N87.

3. The parameters kr , αr , and βr are extracted by performing mea-

surements at three operating points with t1 large enough to let
the material reach the thermal equilibrium. Then, (3.19) is solved
for the three parameters. In Table 3.3 the measurement results
and the corresponding parameters are given.

4. The parameter qr has to be selected such that (3.25) fits the mea-
surement points of a duty cycle measurement, as illustrated in
Figure 3.15.

All model parameters are summarized in Table 3.3. Extracting the

parameters is sometimes difficult and measurements have to be per-
formed very carefully. One error source is a possible current decrease
due to a voltage drop over the inductor winding during "zero" voltage
phase. This can be avoided by choosing a high amount of primary
turns. This increases the inductance value and the current is kept more
constant (by choosing a high amount of primary turns the winding cop-
per resistance increases as well; however, the inductance value increases
quadratically while the resistance value increases linearly).

Example of How to Use the New Model

In the previous section, a new core loss modeling approach was in-
troduced. This section shows now an easy-to-follow example that il-
lustrates how to calculate core losses of a transformer employed in a
bidirectional isolated DC-DC converter with Dual Active full Bridges

60
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

Operating Point Loss Density Model Parameters

(∆B; f ) [kW/m3 ]
(0.1 T; 20 kHz) 5.98 ki = 8.41
(0.1 T; 50 kHz) 16.2 α = 1.09
(0.2 T; 50 kHz) 72.8 β = 2.16
(∆B; dB(t− )/dt) [J/m3 ]
(0.1 T; 4 kT/s) 0.068 kr = 0.0574
(0.1 T; 20 kT/s) 0.13 αr = 0.39
(0.2 T; 20 kT/s) 0.32 βr = 1.31
τ = 6 µs
qr = 16

Table 3.3: Measurement results and model parameters of material

EPCOS N87.

(DAB) [6, 48]. In Figure 3.16(a) the simplified schematic and in Ta-
ble 3.4 the specifications of the transformer are given. The shape of
the core influences the core losses, however, this is not the scope of the
present work, hence a simple toroid is considered as the transformer
core. Phase-shift modulation has been chosen as modulation method:
primary and secondary full bridge are switched with 50 % duty cycle
to achieve a rectangular voltage v1 and v2 across the primary and sec-
ondary transformer side, respectively. The waveforms are illustrated
in Figure 3.16(b), including the magnetic flux density Bµ of the trans-
former core. A phase shift γ between v1 and v2 results in a power
transfer. When the voltages v1 and v2 are opposed (which is the case
in phase tγ ), the full voltage drop is across the transformer leakage
inductance and the magnetic flux density Bµ remains unchanged.
Only the magnetic flux density Bµ time behavior has been con-
sidered for designing the transformer, i.e. no winding losses or leakage
inductance have been calculated. The value of the leakage inductance is
very important for the functionality, however, it is not discussed here.
Therefore, no statement about feasibility is made, the circuit should
only represents solely a simple and easy-to-follow illustrative magnetic
example.
The losses are calculated according to the i2 GSE (3.25). The results
are then compared with measurement results. The peak flux density in

61
CHAPTER 3. CORE LOSS MODELING

Figure 3.16: DAB schematic (a) and waveforms (b) with specifica-
tions given in Table 3.4.

VDC = V1 = V2 42 V
f 50 kHz
N =N1 =N2 20
Effective Magnetic Length le 103 mm
Effective Magnetic Cross Section Ae 95.75 mm2
Core EPCOS N87, R42
(B64290L22X87) [29]

Table 3.4: Specifications of DAB Transformer

the core can be calculated with [6]

 
1 VDC T
B̂ = − tγ (3.29)
2 N Ae 2

62
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

and its time derivative with

 VDC
for t ≥ 0 and t < T2 − tγ ,
 N Ae


dB 0 for t ≥ T2 − tγ and t < T2 ,
= VDC (3.30)
dt − for t ≥ T2 and t < T − tγ ,
 N Ae


0 for t ≥ T − tγ and t < T.
Calculating the losses according to (3.25) leads to the following ex-
pression as a function of tγ
 
T − 2tγ   VDC α  VDC T
 
β−α
P = ki  − t γ  Ae l e
T N Ae N Ae 2
 
2
X
+ Ae l e Qrl Prl , (3.31)
l=1
P2
where l=1 Qrl Prl represents the two transients to zero voltage. There
are two switching instants to zero voltage, each with Qrl = 1. The
values for Prl then have to be determined: it is for each transient
  
1   VDC αr  VDC T βr tγ
  
Prl = kr  − tγ  1 − e− τ . (3.32)
T N Ae N Ae 2
 

The losses have been calculated according to the new approach, and
have been compared to a calculation using the classic iGSE (3.4) and
with measurement results. Open-circuit (no load) measurements have
been performed to validate the new model: the primary winding is
excited to achieve a flux density as illustrated in Figure 3.16(b). Mea-
surements for different values of tγ have been performed, at constant
frequency f and voltage VDC . The new model and measurement results
match very well as shown in Figure 3.17.
In [6] different state-of-the-art core loss calculation approaches are
compared using a very similar example. The iGSE (3.4) showed the best
agreement with measurements, but for increasing zero voltage periods
tγ , the calculated core losses start deviating from the measured core
losses. The reason becomes clear with the new approach i2 GSE and
the calculation can be improved.

3.6.5 Measurements on Different Materials

The approach has been confirmed on different materials, including on
VITROPERM 500F from VAC (measured on W452 core). Measure-

63
CHAPTER 3. CORE LOSS MODELING

i2GSE

Measured Values

iGSE

Figure 3.17: Loss calculation and loss measurement comparison of

the DAB example.

12
∆B = 400mT, t2 = 5µs
10
∆B = 400mT, t2 = 10µs
8
∆B = 200mT, t2 = 5µs
E [µJ]

0
0 5 10 15 20 25 30 35
τ t [µs]
1

Figure 3.18: Measurement results measured on VITROPERM 500F

from VAC (measured on W452 core); temperature = 25 ◦ C. It is
further illustrated how τ according to (3.28) can be extracted.

ments on waveforms as illustrated in Figure 3.10(a) have been per-

formed. Figure 3.18 shows the corresponding measurement results. The
model parameters are given in Table 3.5. Measurements show promise
that the approach is applicable for all material types; however, this
remains to be confirmed as part of future work.

64
 3.6. RELAXATION EFFECTS IN MAGNETIC MATERIALS

Table 3.5: Model parameters of material VITROPERM 500F

(VAC).

α 1.88
β 2.02
ki 137 · 10−6
αr 0.76
βr 1.70
kr 139 · 10−6
τ 4.5 µs
qr 4

3.6.6 Conclusion and Future Work

As experimentally verified, core losses are not necessarily zero when
zero voltage is applied across a transformer or inductor winding after
an interval of changing flux density. A short period after switching the
winding voltage to zero, losses still occur in the material. This work
hypothesizes that this is due to magnetic relaxation. A new loss mod-
eling approach has been introduced and named the improved-improved
Generalized Steinmetz Equation, i2 GSE. The i2 GSE needs five new pa-
rameters to calculate new core loss components. Hence, in total eight
parameters are necessary to accurately determine core losses.
The tested measurement range is given in the following to identify
the range in which the model validity has been confirmed. Two types
of waveforms have been analyzed: trapezoidal as illustrated in Fig-
ure 3.10(a) and triangular waveforms as illustrated in Figure 3.14. For
trapezoidal waveforms, measurements with t1 = 0 . . . 500µs and t2 =
5 . . . 100µs have been conducted. For the triangular waveforms mea-
surements in the range between 20 kHz . . . 100 kHz and D = 0.02 . . . 0.5
have been conducted. No measurements for very low values of t2 have
been conducted; however, the triangular operating point with D = 0.02
and f = 20 kHz, for instance, has a flux rise time of 1 µs, which indicates
that the model is also applicable for very short voltage pulses.
The Steinmetz parameters are valid only for a limited frequency and
flux density range. This is also the case for the additional parameters
of the i2 GSE as has been illustrated in Figure 3.12; this parameter

65
CHAPTER 3. CORE LOSS MODELING

dependency has not been further investigated. Furthermore, a DC level

of the antecedent flux density has not been part of this investigation
and could be considered in the course of future work.

3.7 Core Losses Under DC Bias Conditions

In many power electronic applications magnetic components are biased
with a DC or low-frequency premagnetization, e.g. in Switched-Mode
Power Supplies (SMPS). Within SMPS circuits, magnetic components
that are operating under DC bias conditions are commonly used and
are often among the largest components. Many publications have shown
that the influence of DC bias on the material properties can not be ne-
glected [8, 11, 12, 13, 14, 15, 16]. An approach how to handle DC bias
losses is described in [37, 38, 39]. There, losses are calculated with a
loss map that is based on measurements. This loss map stores the loss
information for many operating points, each described by the flux rip-
ple ∆B, the frequency f , and a DC bias HDC . It is explained how this
loss map can be used to calculate core losses of inductors employed in
power electronic systems. One parameter in the loss map is the DC
premagnetization, thus the loss increase due to DC bias is considered
in this approach. However, extensive measurements are necessary to
build the loss map. Another approach how to consider DC bias losses
is introduced in [14]: the effect of a DC bias is modeled by the given
ratio between losses with and without DC bias for different DC bias
levels HDC and different AC flux densities. This ratio is called the dis-
placement factor DPF. In [14] a graph that shows the DPF is given for
the material ferrite N87 from EPCOS. In [11] an empirical formula that
describes the DPF is given (though it is not named DPF). According
to [11, 14] the DPF does not depend on the frequency f and can be
described as a function of the AC flux density and the DC bias HDC .
A similar approach is suggested in [15], but according to [15] the DPF
does not depend on the AC flux density. The influence of the frequency
on the DPF has not been discussed. The approaches of [11, 14, 15] have
in common that a factor is introduced by which the calculated losses
have to be multiplied to take a premagnetization into consideration. In
other words, the parameter k of (3.3) (or ki of (3.4)) is multiplied by
the DPF and therewith becomes dependent on B̂ (or ∆B).
This thesis proposes a new approach how to describe core losses
under DC bias condition. A graph that shows the dependency of the

66
 3.7. CORE LOSSES UNDER DC BIAS CONDITIONS

Steinmetz parameters (α, β and k) on premagnetization is introduced

in Section 3.7.1. This enables the calculation of losses via the Steinmetz
equation (3.3) or the iGSE (3.4) using appropriate Steinmetz parame-
ters.

3.7.1 Measurement Results and the Steinmetz Pre-

magnetization Graph (SPG)
In this section, measurement results are presented and a new approach
to describe core losses under DC bias conditions is introduced that is
based on a graph that shows the dependency of the Steinmetz parame-
ters (α, β and k) on premagnetization. This is done on the example of
the material ferrite N87 from EPCOS (core part number B64290L22X87
[29]). In Figure 3.19 the core losses and in Figure 3.20 the core losses
normalized to the losses P0 at zero premagnetization are shown for dif-
ferent DC bias values. In Figure 3.21 the losses are plotted as a function
of the frequency f and in Figure 3.22 the losses are plotted as a function
of the peak-to-peak flux density ∆B, with and without DC bias. To
describe the losses via the Steinmetz equation (3.3) or the iGSE (3.4) is
the most common method, hence improvements of this method would
be most beneficial for design engineers. As the iGSE (3.4) is more suit-
able for the description of core losses in power electronic applications,
in all following considerations the three discussed parameters are α, β,
and ki of the iGSE (α, β are the same as in (3.3), while ki is described in
(3.5)). For the applied waveform as illustrated in Figure 3.6 (symmetric
triangular current/flux shape) (3.4) leads to

Pv = ki (2f )α ∆B β . (3.33)

When core losses are plotted with logarithmic axes, where the x-
axis represents the frequency and the y-axis represents the power loss,
approximately straight lines are drawn (cf. Figure 3.21). This is because
the losses follow a power function as e.g. the laws stated in (3.3) and the
iGSE (3.4) are. The parameter α represents the slope of the curve in
this plot. The same can be said when the frequency f is kept constant
and ∆B is varied; hence, the use of a power function with variable
∆B is justified as well (cf. Figure 3.22). The parameter β represents
the slope of the curve in this plot. When a core is under DC bias
condition, the losses over a wide range of HDC still can be described
with the Steinmetz equation (3.3) or the iGSE (3.4), i.e. the losses

67
CHAPTER 3. CORE LOSS MODELING

1.8
∆B = 50mT, T = 40°C, f = 100kHz
1.6 ∆B = 100mT, T = 40°C, f = 100kHz
∆B = 150mT, T = 40°C, f = 100kHz
1.4

1.2
P [W]

0.8

0.6

0.4

0.2

0
0 10 20 30 40 50 60
HDC [A/m]

Figure 3.19: Core losses under DC bias conditions (ferrite N87;

measured on R42 core), f = 100 kHz, T = 40 ◦ C.

still follow the power equation stated by Steinmetz (cf. Figure 3.21
and Figure 3.22). However, for very high values of HDC and high flux
densities ∆B the use of a power function is not appropriate anymore
(cf. Figure 3.22, curve for HDC = 80 A/m). This is due to saturation
effects. The curve for HDC = 50 A/m has been determined as the last
one that can be considered as an approximate straight line over a wide
flux density range. For most applications it is not desired to operate
at higher DC bias levels; hence, the majority of inductive components
are operated in a range where the losses still follow the power equation
stated by Steinmetz.
As described above, the Steinmetz parameters must be adjusted
according to the DC bias present. As will be shown in the following, a
DC bias causes changes in the Steinmetz parameters β and ki , but not
in the parameter α.
I The losses change when ∆B and frequency f are kept constant
and the DC bias HDC is varied (cf. Figure 3.19). Thus, the Stein-
metz parameter ki depends on the DC bias HDC (ki = f (HDC )).
I When the frequency f is kept constant, the factor by which the
losses increase due to a premagnetization HDC differs for different
∆B (cf. Figure 3.20). Thus, the Steinmetz parameter β depends
on the premagnetization HDC as well (β = f (HDC )). The slopes

68
 3.7. CORE LOSSES UNDER DC BIAS CONDITIONS

4
∆B = 50mT, T = 40°C, f = 100kHz
∆B = 100mT, T = 40°C, f = 100kHz
3.5
∆B = 150mT, T = 40°C, f = 100kHz

3
P/P0

2.5

1.5

1
0 10 20 30 40 50 60 70
HDC [A/m]

Figure 3.20: Core losses under DC bias conditions, normalized to

losses P0 at zero premagnetization (ferrite N87; measured on R42
core), f = 100 kHz, T = 40 ◦ C.

of the curves in Figure 3.22 represent the parameter β. As can

be seen the curve for HDC = 20 A/m is slightly steeper compared
to the curve of HDC = 0 A/m, though the difference is very little.
However, a small change in β already considerably influences the
core losses, as one can see when comparing with Figure 3.20. It
should be again pointed out that it is only valid to define a β
within the range of HDC where the logarithmically plotted losses
lead to an approximate straight line (cf. Figure 3.22).
I According to [14], the influence of a DC bias does not depend
on the measurement frequency f . This has been confirmed for
frequencies up to 100 kHz. As can be seen in Figure 3.21, at
a constant ∆B, the factor by which the losses increase due to
a premagnetization HDC is the same for different frequencies f
(the slopes of the curves remain the same). Hence, the Steinmetz
parameter α is in this frequency range independent of the premag-
netization HDC (α = const.). The fact that α is constant has been
confirmed to frequencies up to 100 kHz; no measurements above
this frequency have been performed, hence no information can be
given as to whether and up to which frequency α is constant.
Next, at each tested DC bias level the Steinmetz parameters have

69
CHAPTER 3. CORE LOSS MODELING

1
10
∆B = 100mT, T = 40°C, H =0
DC
∆B = 100mT, T = 40°C, HDC =50A/m
∆B = 200mT, T = 40°C, HDC = 0
0 ∆B = 200mT, T = 40°C, H =50A/m
10 DC
P [W]

-1
10

-2
10 1 2
10 10
f [kHz]

Figure 3.21: Core losses vs. frequency (ferrite N87; measured on

R42 core), T = 40 ◦ C.

1
10

0
10
P [W]

-1
10 f = 100kHz, T = 40°C, HDC = 0
f = 100kHz, T = 40°C, H =20A/m
DC
f = 100kHz, T = 40°C, H =50A/m
DC
f = 100kHz, T = 40°C, HDC =80A/m
-2
10
50 100 150 200
∆B [mT]

Figure 3.22: Core losses vs. flux density (ferrite N87; measured on
R42 core), T = 40 ◦ C.

been extracted. A least square algorithm has been implemented that

fits measured losses with calculated data by minimizing the relative
error at 3 different values of ∆B, each measured at two frequencies.
The markers on top of the curves in Figure 3.23 represent these values.
As not only the Steinmetz parameters at discrete operating points are of

70
 3.7. CORE LOSSES UNDER DC BIAS CONDITIONS

interest, a curve fitting algorithm has been implemented to extract the

dependencies β = f (HDC ) and ki = f (HDC ). Its derivation is discussed
in Appendix A.4.
For the material N87 from EPCOS the dependencies β = f (HDC )
and ki = f (HDC ) are given in Figure 3.23, and normalized to β0 and ki0
in Figure 3.24. β0 and ki0 are the Steinmetz parameters at zero premag-
netization. We call the graph illustrated in Figure 3.24 the Steinmetz
Premagnetization Graph (SPG). The SPG is very useful and it would
be valuable to have such a graph in the data sheet of a magnetic ma-
terial as it would then be possible to calculate core losses under a DC
bias condition. Figure 3.25 shows how the measured and, based on the
SPG, calculated curves compare. For the considered working points the
accuracy obtained has always been ≤ ±15 %.
In Appendix A.3 SPGs of other materials (Ferroxcube 3F3 (ferrite),
EPCOS N27 (ferrite), and VAC VITROPERM 500F (nanocrystalline
material)) are given. Furthermore, a discussion how to extract the
Steinmetz parameter value k from the SPG is given in Appendix A.5.
The markers on top of the curves in the SPG represent the Steinmetz
parameter values that are directly supported by measurement data.
The SPG could be improved by an increase of the HDC resolution to
minimize interpolation errors. All given SPGs consider only the pre-
magnetization range where it is still appropriate to use the Steinmetz
approach, i.e. the losses still follow a power equation.
In the SPG, the Steinmetz parameters are plotted as a function of
HDC . For an ideal toroid HDC can be calculated according to (3.10) as
IDC N1
HDC = , (3.34)
le
where IDC is the DC current, N1 is the number of excitation winding
turns and le the effective magnetic path length of the CUT. It would
also be possible to use BDC instead of HDC . For cores without air gaps,
HDC has the advantage that it is directly calculable from the current
(as it is done in this work). For gapped cores, one would need to set up
an accurate reluctance model (cf. Chapter 2) to calculate HDC inside
the core. The relationship BDC (HDC ) is customarily assumed to be the
initial magnetization curve [14].
For the derivation of the SPG the losses are calculated according
to (3.33). For the frequency f the unit Hertz (Hz) has been used and
for the peak-to-peak flux density ∆B the unit Tesla (T) has been used.
Consequently, the SPG is only valid when this set of units is used.

71
CHAPTER 3. CORE LOSS MODELING

12

10

8
ki
α, β, ki

4
β
2 α

0
0 10 20 30 40 50
HDC [A/m]

Figure 3.23: Steinmetz parameters as a function of premagnetiza-

tion HDC (ferrite N87), T = 40 ◦ C.

3 1.2
2.8 ki / ki0; T = 40°C 1.18
2.6 1.16
2.4 1.14
2.2 1.12
2 β / β0; T = 40°C 1.1
1.8 1.08
1.6 1.06
i0

β / β0
k /k

1.4 1.04
i

1.2 β / β0; T = 80°C 1.02

1 1
0.8 0.98
ki / ki0; T = 80°C
0.6 0.96
0.4 0.94
0.2 0.92
0 0.9
0 10 20 30 40 50
HDC [A/m]

Figure 3.24: SPG of the material ferrite N87 (EPCOS).

3.7.2 Influence of Temperature

For an accurate core calculation, the temperature is another important
parameter that considerably influences core losses. In Figure 3.26 the
losses normalized to losses P0 at zero premagnetization are given for
different temperatures. As can be seen for the material ferrite N87, at

72
 3.7. CORE LOSSES UNDER DC BIAS CONDITIONS

0
10
∆B/f 100mT/100kHz (cal.)
100mT/100kHz (meas.)
150mT/20kHz (cal.)
150mT/20kHz (meas.)
-1
10
150mT/10kHz (cal.)
P [W]

150mT/10kHz (meas.)
50mT/20kHz (cal.)

-2
10
50mT/20kHz (meas.)

-3
10
0 10 20 30 40 50
HDC [A/m]

Figure 3.25: Core losses under DC bias conditions: measured

(meas.) and calculated (cal.) curves (ferrite N87), T = 40 ◦ C.

2.6
∆B = 100mT, T = 40°C, f = 100kHz
2.4 ∆B = 100mT, T = 60°C, f = 100kHz
∆B = 100mT, T = 80°C, f = 100kHz
2.2

1.8
P/P0

1.6

1.4

1.2

0.8
0 10 20 30 40 50
H [A/m]
DC

Figure 3.26: Core losses under DC bias conditions: measured at

different operating temperatures. Normalized to losses P0 at zero
premagnetization. Material N87.

higher temperatures the influence of a premagnetization on core losses

reduces. The temperature influence is described by extending the SPG
to curves of different operating temperatures, as shown in Figure 3.24.

73
CHAPTER 3. CORE LOSS MODELING

S
IL L Iload

+
Vin D C Rload Vout

(a)

iL

I load ∆i Lpp
V –V –V
-----in----------out
---- ------out
---
L L
t
0 DT T
(b)

Figure 3.27: Buck converter schematic (a) and current waveform

(b) with specifications given in Table 3.6.

Vin / Vout 12 V / 6 V
f 100 kHz
P 2W
Iload 0.33 A
L 150 µH (EPCOS N87; R25; N =8; no air gap)
(core part number: B64290L618X87 [29])

Table 3.6: Buck converter specifications.

3.7.3 Example How to Use the SPG

In the previous sections the SPG has been introduced. This section
presents now an easy-to-follow example that illustrates how to calcu-
late core losses of the inductor of a power electronics converter with help
of the SPG. In Figure 3.27 the schematic and the inductor current wave-
form of a buck converter, and in Table 3.6 the corresponding specifica-
tions are given. For the inductor L a DC bias of HDC = 44 A/m (accord-
ing to (3.34)), and a peak-to-peak flux density ripple of ∆B = 73 mT
is calculated. The following steps lead to the core losses that occur in
the inductor:

74
 3.7. CORE LOSSES UNDER DC BIAS CONDITIONS

I For the material used, the corresponding Steinmetz parameters

are extracted from the datasheet. This is done by solving (3.3) at
three operating points for α, β, and k: α = 1.25, β = 2.46, k =
15.9 (values for temperature = 40 ◦ C, at zero premagnetization).

I Next, ki is calculated according to (3.5): ki = 1.17.

I ki and β are now adjusted according to the SPG of the material

N87 (cf. Figure 3.24) for an operating point with HDC = 44 A/m:
ki0 = 2.8 · ki = 3.28 and β 0 = 1.04 · β = 2.56.

I Now, the losses are calculated according to (3.4). For piecewise

linear waveforms, as is the case in the presented example, the
integral of (3.4) may be split into one piece for each linear segment,
such that a complicated numerical integration is avoided [35]. The
losses follow as
0
k 0 (∆B)β −α
P = Ve i
 T
 ∆B α  ∆B α
   
· 
  DT + 
  (1 − D)T
DT  (1 − D)T 
0
k 0 (∆B)β −α
= Ve i
T 
 Vin − Vout α  −Vout α
   
· 
  DT + 
  (1 − D)T
N Ae  N Ae 
= 52.8 mW, (3.35)

where Ve = 3079 mm3 is the effective core volume, Ae = 51.26 mm2

is the effective core cross section, T = 1/f is the period length,
and D = 0.5 is the duty cycle.

Under the assumption that the Steinmetz parameters had not been
adjusted according to the SPG in the example above, the losses would
have been calculated as P = 24.5 mW, which is an underestimation by
a factor of more than two.
In case of a load change one has to redo the core loss calculation as a
load change leads to a change in the premagnetization and, accordingly,
to a change of the core losses. This fact is rarely considered when
modeling magnetic components.

75
CHAPTER 3. CORE LOSS MODELING

1.8 ∆B = 100mT, T = 40°C, f = 100kHz

∆B = 200mT, T = 40°C, f = 100kHz
1.6 ∆B = 300mT, T = 40°C, f = 100kHz

1.4

1.2
P [W]

0.8

0.6

0.4

0.2

0
0 5 10 15 20 25 30 35
HDC [A/m]

Figure 3.28: Core losses under DC bias conditions; material VIT-

ROPERM 500F (VAC); core: W452; f = 100 kHz, T = 40 ◦ C.

3.7.4 Core Losses under DC Bias Conditions of Dif-

ferent Materials
Different materials have been tested to gain information how core losses
are influenced by a premagnetization. Measurements on the nanocrys-
talline material VITROPERM 500F and on a molypermalloy powder
core (Magnetics MPP 300u) and cores of silicon steel (tested: M165-35S
grain-oriented steel with lamination thickness 0.35 mm, M470-50A non-
oriented steel with lamination thickness 0.5 mm) have been conducted.
The measurements on the silicon steel cores have been performed up
to a DC magnetic flux density of 1 T, which is before the core starts to
saturate. A loss increase of approximately 50 % has been observed. The
tested powder core (Magnetics MPP 300u; part number: C055433A2)
has been tested up to a DC magnetic field strengths of 1200 A/m; up
to that operating point the loss change is negligibly small.
Losses in the nanocrystalline material VITROPERM 500F from
Vacuumschmelze increase under DC bias condition, as can be seen in
Figure 3.28. The SPGs of the material VITROPERM 500F and of some
more ferrites are given in Appendix A.3. In Table 3.7 an overview of
the tested materials is given. The reason for the distinctive behavior
of each material class hasn’t been studied for this work and could be
investigated as part of future work. Tests have been performed only on

76
 3.8. MINOR AND MAJOR B-H LOOPS

Material Class Measured Material(s) Impact on Losses

Soft Ferrites EPCOS N87, N27, T35 very high
Ferroxcube 3F3
Nanocrystalline VITROPERM 500F (VAC) yes
Silicon Steel M470-50A non-oriented steel yes
M165-35S grain-oriented steel
Molypermalloy Magnetics MPP300 negligible
Powder

Table 3.7: Impact of DC bias to core losses, an overview of different

material classes.

the above listed components; hence, a general declaration of the whole

material class cannot be made with 100 % certainty.

3.7.5 Conclusion and Future Work

A graph that shows the dependency of the Steinmetz parameters (α, β
and k) on premagnetization, i.e. the Steinmetz Premagnetization Graph
(SPG) has been introduced. Based on the SPG, the calculation of core
losses under DC bias condition becomes possible. For the considered
frequency range it is shown that the graph is independent of the fre-
quency f . This new approach how to describe losses under DC bias
condition is promising due to its simplicity. Graphs are given for dif-
ferent materials and different operating temperatures.
Furthermore, different material classes have been tested to gain in-
formation how core losses are influenced by a premagnetization. Mea-
surements on molypermalloy powder, silicon steel, nanocrystalline ma-
terial, and ferrite cores have been performed.

3.8 Minor and Major B-H Loops

In the previous sections, different aspects of core loss modeling have
been discussed; it became clear how to model core losses of a single B-H
loop of different shape. However, in many power electronic applications,
e.g. in PFC rectifiers, a high frequency flux ripple is superimposed on a

77
CHAPTER 3. CORE LOSS MODELING

B(t) B(t) B(t)

t = t + t

B B B

H = H + H

Figure 3.29: Illustration of minor and major loop separation.

low frequency flux waveform. When plotting the traversed B-H-curve,

one sees a large loop and many small loops, i.e. one large major loop and
many small minor loops. The large loop originates in the fundamental
current waveform, while the small loops originate in the HF ripple.
According to [10, 35], the total loop can be separated into its major
and minor loops and then the loss energy of each loop can be calculated
independently and summed. The concept is illustrated in Figure 3.29.
This has been experimentally verified in [10, 35]. In the following,
measurement results are presented to reconfirm the concept.
This loop separation has been confirmed on the amorphous alloy
2605SA1 from Metglas (core PowerLite AMCC8) as core material. The
measured current waveform consists of a sinusoidal low frequency part
and, superimposed to it, piecewise-linear high frequency segments, as
plotted in Figure 3.30. The flux waveform can be divided into its funda-
mental flux waveform and into piecewise linear flux waveform segments.
The loss energy is then calculated for the fundamental and all piecewise
linear segments, summed and divided by the fundamental period length
in order to determine the average core loss. Actually, when doing this,
one does not consider how the minor loop closes: each piecewise linear
segment is modeled as having half the loss energy of its corresponding
closed loop. This leads to a high accuracy, as measurements confirm.
For the loss energy calculation of each piecewise linear flux segment
and for the fundamental waveform, a loss map has been built up in

78
 3.9. HYBRID CORE LOSS CALCULATION APPROACH

0.5
i [A]

-0.5

-1
-2 -1 0 1 2
t [ms]

Figure 3.30: Measured (and low-pass filtered) current waveform of

superposition experiment; fLF = 200 Hz, ∆BLF = 1.2 T (peak-to-
peak), fHF = 10 kHz, and ∆BHF = 0.2 T (peak-to-peak).

Operating Loss Loss

Point cal. [W] meas. [W]
fLF = 200 Hz, ∆BLF = 1.2 T 0.82 0.76
fHF = 10 kHz, ∆BHF = 0.2 T
fLF = 50 Hz, ∆BLF = 2 T 2.27 2.14
fHF = 10 kHz, ∆BHF = 0.4 T

Table 3.8: Results of minor major loop measurements.

advance. For it, sinusoidal low frequency measurements and triangular

high frequency measurements at different operating points have been
performed first. How exactly such a loss map is built up and used will
be discussed in the next section.
The calculated losses for different operating points agree well with
measurements, as can bee seen in Table 3.8.

3.9 Hybrid Core Loss Calculation Approach

In the previous sections, different aspects of core loss modeling have
been discussed. With the derived knowledge it is now possible to ac-

79
CHAPTER 3. CORE LOSS MODELING

curately determine core losses in inductive components employed in

power electronic systems. In order to improve the model accuracy a
loss database has been built up. The applied core loss approach can
be seen as a hybrid of the improved version of the empirical Steinmetz
equation (i2 GSE) and an approach based on a material loss database,
i.e. a loss map.
As mentioned in Section 3.3, until core manufacturers provide data
for loss calculation based on i2 GSE, SPG, or DPF, core losses can be
calculated with a loss map that is based on measurements. This loss
map stores the loss information for different operating points, each de-
scribed by the flux density ripple ∆B, the frequency f , the temperature
Tc , and a DC bias HDC . This has been implemented e.g. in [37, 38, 39].
This approach is independent of a parameter set, e.g. Steinmetz param-
eter set, hence a higher accuracy over a wide frequency and flux density
range can be expected since the problem that Steinmetz parameters
are only valid for a limited frequency and flux density range does not
appear. A novel loss map structure has been developed in the course
of this thesis and will be presented within this section.
The loss map stores loss densities (in our case losses per volume)
as the data should be applicable for all different type of core shapes.
A core with homogenous flux density distribution is basically needed
for measuring loss densities. This would be the case when having an
"ideal" toroid with very small radial thickness. If the radial thickness
is not small, the flux density is not distributed homogenously over the
radius and the situation is more complicated. In this case "effective"
dimensions (effective length le and effective cross-sectional area Ae ) are
needed to permit calculation as if it were an ideal toroid. The core losses
per volume can then be extracted by dividing the measured losses by
the effective volume of the core Ve = Ae le . These effective dimensions
for a toroid can be calculated as (cf. Section 3.10)

h ln2 r2 /r1
Ae = , (3.36)
1/r1 − 1/r2
and
2π ln r2 /r1
le = . (3.37)
1/r1 − 1/r2
After it has been clarified on what cores measurements for setting
up a loss map can be performed, it has to be further discussed what
type of waveforms are to be measured. In some applications, e.g. PFC

80
 3.9. HYBRID CORE LOSS CALCULATION APPROACH

Relaxation B-H-Relation
αr βr
kr τ
qr

LF HF
ΔB, f, T, HDC ΔB, f, T, HDC
ΔB, f, T, HDC ΔB, f, T, HDC
ΔB,f,f,T,T,HHDCDC
ΔB, ΔB, f, T, HDC
ΔB, f, T, HDC ΔB,f,f,T,T,HHDCDC
ΔB,
ΔB, f, T, HDC ΔB, f, T, HDC

E E
E E
EE EE
E E
E E

Figure 3.31: Content of loss map.

rectifiers, a high frequency (piecewise-linear) flux ripple is superimposed

to a low frequency (sinusoidal) flux waveform. This well illustrates the
typical situation in modern power electronic systems: the low frequency
fundamental waveform usually has a sinusoidal shape, while the high
frequency ripple consists of piecewise-linear segments. The B-H loop
of the low frequency fundamental flux part is called the major loop,
while the high frequency loops are called minor loops. According to
the discussion in Section 3.8, the flux waveform can be divided into
its fundamental flux waveform and into piecewise linear flux waveform
segments. The loss energy is then calculated for the fundamental and
all piecewise linear segments, summed and divided by the fundamental
period length in order to determine the average core loss. As mentioned
before, when doing this, one does not consider how the minor loop
closes: each piecewise linear segment is modeled as having half the loss
energy of its corresponding closed loop. This led to a high accuracy, as
measurement have confirmed.
Loss information for different operating points is stored in the loss
map, each described by the flux density ripple ∆B, the frequency f , the
temperature Tc , and a DC bias HDC . On the basis of the above discus-
sion, sinusoidal waveform measurement results are stored for frequencies
below 1 kHz and triangular, i.e. piecewise-linear, waveform measure-
ment results are stored for frequencies above 1 kHz. The classification
limit of 1 kHz has been selected as 1 kHz is slightly above the funda-
mental frequency of the mains in modern aircraft (up to 800 Hz [49]).
In addition, the loss database contains the initial B-H curve and one

81
CHAPTER 3. CORE LOSS MODELING

set of relaxation parameters αr , βr , kr , τ , and qr . The extraction of the

relaxation parameters has been carried out as described in Section 3.6.
The initial B-H curve enables introducing a reluctance model that takes
the non-linearity of the core material into consideration (cf. Chapter 2).
The relaxation parameters permit the taking of relaxation effects into
consideration by using the i2 GSE (3.25). In Figure 3.31 an overview
of the loss map content is illustrated. The B-H relation as well as the
loss map operating points have to be measured and stored at different
temperatures in order to consider the temperature behavior of the core
material.

3.9.1 Use of Loss Map

As mentioned before, the applied core loss approach in this work can
be described as a hybrid of the Steinmetz approach and a loss map
approach: a loss map is taken in order to provide accurate parameters
for the i2 GSE or SE. This principle is illustrated in Figure 3.32. The
fundamental principle of the implemented hybrid approach can also be
interpreted differently: a loss map is taken to calculate core losses, while
the interpolation and extrapolation between operating points is made
with the SE or the i2 GSE.
The flux density waveform for which the losses have to be calculated
could, for instance, be simulated in a circuit simulator. This simulated
waveform is then broken up into its fundamental (mostly sinusoidal)
flux waveform and into piecewise linear flux waveform segments, as
illustrated in Figure 3.32. The loss energy is then calculated for all
segments, summed and divided by the fundamental period length. The
piecewise-linear waveforms are translated into a symmetric triangular
flux waveform with the same peak-to-peak flux density ∆B, the same
flux density slope dB/dt and the same DC premagnetization HDC , re-
sulting in a symmetric triangular operating point that is defined as
(∆B ∗ , f ∗ , HDC
∗
, Tc∗ ). The translation is necessary as the loss map
stores operating points of symmetric triangular waveforms.
The losses to the (sinusoidal or triangular flux waveform) operating
point (∆B ∗ , f ∗ , HDC
∗
, Tc∗ ) are calculated by the following steps:

1. The close-by operating points of the loss map have to be found.

Nine operating points are necessary for the interpolation: three
for the interpolation of f and ∆B, multiplied by three for the in-
terpolation of the temperature Tc and DC magnetic field strength

82
 3.9. HYBRID CORE LOSS CALCULATION APPROACH

Loss Map
Relaxation B-H-Relation
αr βr
kr τ
qr

LF HF
ΔB, f, T, HDC ΔB, f, T, HDC
ΔB, f, T, HDC ΔB, f, T, HDC
ΔB,f,f,T,T,HHDCDC
ΔB,
ΔB,f,f,T,T,HHDCDC
ΔB,
ΔB, f, T, HDC ΔB, f, T, HDC
ΔB, f, T, HDC ΔB, f, T, HDC

EE EE
EE E
E EE
E E

k , ki , α , β , α r , β r , qr ,τ

T α
1 dB n
P
(∆B )β −α dt + ∑ Qrl Prl
T ∫0 dt
Pv = ki
l =1 V

Figure 3.32: Illustration of hybrid approach: a loss map is built in

order to provide accurate parameters for the i2 GSE or SE.

HDC . This is illustrated in Figure 3.33 and will be discussed be-

low.

2. In a second step, a linear interpolation for the temperature Tc

and the DC magnetic field strength HDC is performed. This is
illustrated in Figure 3.33(a) and carried out for each pair ∆B/f .
∗
Therefore the following points are derived: (∆B1 , f1 , HDC , Tc∗ ),
∗ ∗ ∗ ∗
(∆B2 , f1 , HDC , Tc ), (∆B1 , f2 , HDC , Tc ), i.e. three operating
points remain with inter-/extrapolated temperature Tc and mag-
netic field strength HDC .

3. The peak-to-peak flux density ∆B and the frequency f are in-

ter/extrapolated by extracting the Steinmetz parameters from the
three remaining operating points and calculating the losses by the
SE (in the case of a sinusoidal flux waveform) or the i2 GSE (in
the case of a piecewise linear waveform). This is illustrated in
Figure 3.33(b). The relaxation term of the i2 GSE can directly be
evaluated with the relaxation parameters from the loss map; for
this, information about the antecedent piecewise-linear flux phase

83
CHAPTER 3. CORE LOSS MODELING

(∆B1, f1, H*DC,T*c)

(∆B1, f1, HDC,2,Tc,1) (∆B1, f1, HDC,1,Tc,2)

Tc
Tc,2
HDC,2
(∆B1, f1, HDC,1,Tc,1)
(a) HDC

PV
∆B 3
(∆B2, f1, H*DC,T*c)
∆B 2
(∆B*, f*, H*DC,T*c)
∆B 1

f
f1 f2
(b)

Figure 3.33: Illustration of interpolation.

is required.
This hybrid approach already enables a high degree of accuracy for a
limited number of pre-measured operating points in the loss map as
the inter/extrapolation by the SE or i2 GSE takes the basic core loss
behavior into account.
As the frequency increases, the loss map operating points change
from sinusoidal to triangular waveforms; hence, for the Steinmetz ex-
traction the curve fitting function changes from the SE to the i2 GSE. In
one case the Steinmetz parameter k, and in the other case the Steinmetz
parameter ki is extracted. However, with (3.5) they can be translated
into each other.

3.10 Influence of Core Shape on Core Losses

The Steinmetz equation and its improvements, such as the i2 GSE, give
information about losses per volume (or per weight). But, simply mul-
tiplying the loss density by the volume may lead to substantially wrong
results. Within this section it is discussed how to model core losses of
cores of general shape.

84
 3.10. INFLUENCE OF CORE SHAPE ON CORE LOSSES

In a first step, in Section 3.10.1, it is given how to calculate the

effective dimensions of a toroid in order to perform calculations as if the
flux density distribution were homogenous. These effective dimensions
are important to characterize core materials, as they allow determining
the loss densities (loss per volume or loss per weight) of a material from
the measured losses. Later, in Section 3.10.2, it is analyzed how the core
losses for cores of general shapes can be calculated. By introducing a
reluctance model of the core, and with it, calculating the flux density
in every core section of (approximately) homogenous flux density, one
can calculate the losses of each core section. The core losses of each
section are then summed to obtain the total core losses. This leads to
a generally high accuracy. However, under certain circumstances, in
tape wound cores a flux orthogonal to the tape layers can lead to high
eddy currents and therewith to high core losses. This will be discussed
in Section 3.10.5. Additionally, discussions about eddy-current losses
in general and dimensional resonance are given in Section 3.10.3 and
Section 3.10.4 respectively.

3.10.1 Effective Dimensions of Toroidal Cores

When considering an ideal toroid with a very small radial thickness, it
is possible to speak of its magnetic lengths l and cross section A. In case
the radial thickness is not small, the flux density is not homogenously
distributed over the radius, and the core property is more complicated.
However, it is possible to find effective dimensions le and Ae , which allow
to proceed calculations as if it were an ideal toroid. When measuring
core losses of a toroid, one can approximate the core losses per volume
with these effective dimensions. The flux in the core divided by the
effective cross-sectional area Ae gives an average flux density. This
allows approximating the core losses per volume to this flux density.
The effective dimensions for toroidal cores are

h ln2 r2 /r1
Ae = , (3.38)
1/r1 − 1/r2

and
2π ln r2 /r1
le = . (3.39)
1/r1 − 1/r2
The derivation of (3.38) and (3.39) is given in Appendix A.6.

85
CHAPTER 3. CORE LOSS MODELING

These effective dimensions are important to characterize core ma-

terials, as they allow determining the loss densities (loss per volume or
loss per weight) of a material from the measured losses.

3.10.2 General Core Shape

[22] proposes a method to find the dimensions le and Ae for general
core shapes, which allows to calculate the inductance value as if it
were an ideal toroid. The variation of permeability with field strengths
is neglected, and the core is divided into sections with approximately
constant flux densities. Core factors are introduced as
X li
C1 = , (3.40)
i
Ai

and
X li
C2 = , (3.41)
i
A2i

where li and Ai are the length and cross section of section i. The
effective dimensions are then
C12
le = (3.42)
C2
and
C1
Ae = . (3.43)
C2
These effective dimensions are only suitable for an inductance cal-
culation and may lead to substantial errors in loss estimation. For
instance in case of an E-core, a simple multiplication of the loss density
with Ae and le leads to wrong loss values. This is because the flux
density is not uniform at every position of the core. The calculation of
losses with effective dimensions is only valid in case of an uniform flux
density distribution.
The core losses have to be calculated differently. By setting up a
reluctance model of the core, and with it, calculating the flux density in
every core section of (approximately) homogenous flux density, one can
calculate the losses of each core section. The core losses of each section
are then summed to obtain the total core losses. This also allows to
consider an air gap. How to set up a reluctance model has been shown

86
 3.10. INFLUENCE OF CORE SHAPE ON CORE LOSSES

Working Point Loss Density Steinmetz

(∆B; f ; HDC ) [kW/m3 ] Parameters
(0.05 T; 50 kHz; 0 A/m) 3.09 ki = 6.84
(0.05 T; 100 kHz; 0 A/m) 6.89 α = 1.16
(0.1 T; 100 kHz; 0 A/m) 36.5 β = 2.41

Table 3.9: Steinmetz parameters of material EPCOS N87.

D C B

E A

(a) (b)

Figure 3.34: Photo and the section segmentation of the core under
test EPCOS B66319G0X187.

in Chapter 2. The fact that the flux density in core parts very close to
the air gap is (slightly) reduced as part of the flux already left the core
has been neglected. In the following, some experimental results, which
illustrate and confirm this approach, are given.
In a first step, the selected material is characterized, i.e. the core loss
densities of the selected material are measured on a toroid. The material
N87 from EPCOS has been selected, and the characterizing measure-
ments have been conducted on the toroidal core EPCOS B64290L22X87
[29]. The effective magnetic length le and effective core cross section
Ae are 103 mm and 95.75 mm2 respectively. The Steinmetz parameters
have been extracted from three operating points. They are given in
Table 3.9, where ki is the parameter of (3.4).
The core for which the core losses are calculated is the E-core EP-
COS B66319G0X187. In Figure 3.34 a photo and the section segmen-
tation of the core are given. For every core section the length li and
cross section in Ai is calculated as proposed in Figure 2.10. The results

87
CHAPTER 3. CORE LOSS MODELING

Section li [mm] Ai [mm2 ]

A 9.7 26.3
B 3.6 33.2
C 6.2 40.2
D 4.2 39.2
E 9.7 38.3

Table 3.10: Dimensions of core EPCOS B66319G0X187.

are summarized in Table 3.10. The peak-to-peak flux ∆φ in one outer

leg is calculated as
∆φ = ∆B · AA , (3.44)
where AA is the cross section of core section A. Note that section A
of the core represents only half of the middle leg (symmetry). ∆B
represents the peak-peak flux density in the middle leg. With it, in
every section the flux density can be calculated as

∆φ
∆Bi = . (3.45)
Ai
The losses per core section for a triangular current shape can be calcu-
lated with (3.33) as

Pi = li · Ai · ki (2f )α ∆Biβ . (3.46)

The total core losses, including the right leg, are then

P = 2 · (2PA + 2PB + 2PC + 2PD + 2PE ). (3.47)

The calculated losses for different operating points agree well with mea-
surements, as can bee seen in Table 3.11.

3.10.3 Impact of Core Shape on Eddy Current Losses

An alternating magnetic flux inside the core material induces an electric
field which leads to eddy currents and, consequently, to eddy current
core losses. Eddy current core losses are referred to the losses that orig-
inate from Maxwell’s equation when the presence of magnetic domains
is ignored, i.e. only the macroscopic net magnetization is considered.

88
 3.10. INFLUENCE OF CORE SHAPE ON CORE LOSSES

Working Point Loss Meas. Loss Cal.

(∆B; f ) HDC = 0 [mW] [mW]
(0.1 T; 50 kHz) 42.7 40.9
(0.1 T; 100 kHz) 91.6 91.2
(0.15 T; 50 kHz) 117 108
(0.15 T; 100 kHz) 257 242
(0.2 T; 50 kHz) 233 217
(0.2 T; 100 kHz) 509 483

Table 3.11: Results of loss calculations and measurements of core

EPCOS B66319G0X187; Steinmetz parameters from Table 3.9.

Eddy currents and their corresponding losses depend a lot on the elec-
trical conductivity and geometry of the core. In case of a core material
with high electrical conductivity, eddy currents of high magnitude are
induced; these eddy currents lead to high core losses. The material
can be divided into electrically insulated segments, e.g. laminations or
grains, in order to reduce eddy current losses. The eddy current losses
per unit volume depend then not on the shape of the bulk material,
but on the size and geometry of the insulated regions [22]. Particularly,
iron-based material (e.g. silicon steel) has to be laminated in order to
limit eddy current losses.
The eddy currents are such that the magnetic field generated by
them is opposed to the applied magnetic field. In other words, the eddy
currents have a shielding effect: the resulting magnetic field, which is
the sum of the applied field and the field originating from eddy currents,
decreases exponentially towards the inside of the core. The distance
from the outer boundary to where the resulting magnetic field has fallen
to 1/e of the outer boundary field value, is called the skin depth (or
penetration depth) δ, and can be calculated as [5]

1
δ=√ , (3.48)
πµσf

where µ is the permeability of the core material, σ is the conductivity

of the core material, and f is the frequency of the sinusoidal applied
magnetic field. At lower frequencies, where the applied field penetrates
(almost) the entire core, the eddy current losses per unit volume for

89
CHAPTER 3. CORE LOSS MODELING

Geometry kec
laminations of thickness d 6
cylinder of diameter d 16
sphere of diameter d 20

Table 3.12: The eddy current loss coefficient kec for some geometries.

y
x
z
B

Figure 3.35: Geometry of considered lamination layer.

sinusoidal excitation can be calculated with [22]

(π B̂f d)2
Pce = , (3.49)
kec ρ

where B̂ is the peak value of the flux density, which is perpendicular

to the plane with cross-sectional dimension d, ρ is the resistivity of the
core material, and kec is a dimensionless eddy current loss coefficient.
The eddy current loss coefficients for some geometries are given in Ta-
ble 3.12. The power loss density of the lamination geometry shown in
Figure 3.35 is then calculated as

(π B̂f d)2 π2 2 2 2
Pce = = B̂ f d σ, (3.50)
6ρ 6
where σ is the conductivity of the core material. The same equation
has been derived in [33]. In case of a solid material or at very high
frequencies, where eddy currents shield the applied magnetic field to
entirely penetrate the core material, other equations than (3.49) or
(3.50) have to be taken. Equations for these situations are given e.g. in
[5, 33].
One important conclusion of the above discussion is the fact that,
in case of laminated iron cores or tape wound cores such as amorphous

90
 3.10. INFLUENCE OF CORE SHAPE ON CORE LOSSES

or nanocrystalline iron material, it is still appropriate to calculate with

core loss densities that have been measured on a sample core with a
geometrically different bulk material, but with the same lamination or
tape thickness. This is because the eddy current losses per unit volume
depend on the size and geometry of the insulated regions and not on
the geometry of the core itself. In other words, it is appropriate to set
up a reluctance model, calculate the flux density in each core section
of approximately homogenous flux density and calculate the core losses
based on a previously measured loss density.
In iron based core materials, the skin depth δ is small and the core
has to be laminated in order to prevent from high core losses. Ferrites,
on the other hand, are usually considered as homogeneous materials
with respect to eddy current losses [22]. However, the conductivity σ,
in return, is much smaller and eddy current core losses can be neglected
in most practical cases. Typical resistivity values for Manganese Zinc
(MnZn) ferrites are in the range of 2 . . . 100 Ωm [31]. The dominating
loss effects in ferrites are the static hysteresis and residual (relaxation)
losses. However, there might be situations in which eddy current losses
are not negligible; particularly at very high frequencies, in cores with
a relative high conductivity and with large cross sectional area, eddy
currents are observed. For instance, in [50] measurements on different
ferrite materials on different toroidal core sizes have been conducted.
On the material 3E6 from Ferroxcube with a comparably low resistivity
of 0.1 Ωm [51] the loss density increased about 50 % at 100 kHz when
the toroid is changed from size R16 to size R36. In such situations, it
is important to extract the loss density for the loss map on small core
samples where the eddy currents are still low. In the modeling process
of larger cores, eddy current losses, calculated with (3.49), are added
to the core losses subsequently. However, in most practical cases, eddy
currents can be neglected in ferrites.

3.10.4 Dimensional Resonance

In MnZn ferrites at very high frequencies dimensional resonances may
occur in which standing electromagnetic waves are established [22, 30,
52]. A greater amount of energy is then dissipated. In [52] it was
discovered that the observed permeability of brick-shaped core samples
decrease rapidly to a very low value at about 2 MHz. The reason is
that the very high permittivity  and the high permeability µ leads to

91
CHAPTER 3. CORE LOSS MODELING

Figure 3.36: Dimensional resonance in MnZn ferrite cores, calcu-

lated from typical properties and three different permeabilities. If the
smallest cross-sectional dimension of the core is half of the wavelength
λ then the fundamental mode standing wave will be established. Fig-
ure copied from [22].

relatively low wavelengths of electromagnetic waves.

The wavelength λ in a loss-free medium is

1
λ= √ , (3.51)
f µ

where µ is the permeability of the medium,  is the permittivity of

the medium, and f is the frequency of the electromagnetic wave. In a
typical MnZn ferrite the permeability is µ = 103 µ0 and the permittivity
is  = 105 0 [22]. This leads to a wavelength of λ = 30 mm at a
frequency of 1 MHz. If the smallest cross-sectional dimension of the
core is half of the wavelength then the fundamental mode standing
wave will be established [22]. In case of a dimensional resonance, the
observed permeability is decreasing to a very low value. The frequencies
at which dimensional resonances occur differ if the medium is not loss-
free. In Figure 3.36 the according wavelengths of a typical MnZn ferrite
as a function of the frequency are given.
This topic is somewhat outside the scope of this thesis, as the effect
occurs mainly at very high frequencies in the megahertz range. How-
ever, in case of very large core samples, the frequencies where dimen-
sional resonance occurs drops and therefore, for the sake of complete-
ness, one should be aware of the existence of dimensional resonances.

92
 3.10. INFLUENCE OF CORE SHAPE ON CORE LOSSES

3.10.5 Losses in Gapped Tape Wound Cores

The core losses can be calculated based on an accurate reluctance model,
which, if needed, consists of an accurate air gap reluctance. However,
it has been found that gapped cores of amorphous alloys or nanocrys-
talline materials show a behavior that cannot be modeled by a reluc-
tance model: gapped tape wound cores have higher core losses compared
to un-gapped cores [53, 54, 55, 56, 57]. There are two causes assumed
to be dominant for this distinctive behavior: (1) in the cutting process,
interlamination short circuits may be introduced (insulation layer dam-
age) and (2) flux lines that are orthogonal to the lamination layers due
to fringing flux. Both effects would result in higher eddy-current core
losses.
In [53] a doubling of the core losses of an iron-based amorphous
alloy after the cutting process has been observed. However, additional
core losses could be partly reduced by polishing the cut surface after
the cutting process. In [57], the cut core has been put in an 40 %
ferric chloride solution, in order to remove short circuits introduced by
the cutting process. The losses could be substantially (more than 50 %)
decreased. Both cited papers confirmed the existence of interlamination
short circuits and showed a way how to (partly) get rid of them.
The second cause for this distinctive behavior has been observed, for
instance, in [55, 56, 57]. In [55, 56] it has been observed that the core
losses in amorphous cut cores with air gaps increase with an increase
of the air gap length. This increase has been attributed to the in-
plane eddy-current losses generated by the fringing flux perpendicular
to the layers of the tape wound cores. Losses originating in flux lines
orthogonal to the layers of the tape wound core have also been observed
in [57], where their existence has been confirmed with an experiment
which is illustrated in Figure 3.37. Two core halves of nanocrystalline
core material are taken, one of the halves is shifted, first in horizontal,
later in vertical direction. In case of a shift in horizontal direction
a flux orthogonal to the layers of the tape wound core close to the
region of contact occurs, while in the case of a shift in vertical direction
no orthogonal flux is generated. As can be seen in Figure 3.37(c), in
case of an orthogonal flux, the core losses substantially increase. These
experiments have also been conducted with ferrites, where no difference
between the two shift directions has been observed. The results with
the tape wound cores show clearly that an orthogonal flux substantially
increases core losses in these types of core. In [57] it is furthermore

93
CHAPTER 3. CORE LOSS MODELING

(a) (b)

10
Horizontal displacement
8 Vertical displacement
Core Losses [W]

0
0 10 20 30 40 50
Misalignment [% of the corresponding dimension]
(c)

Figure 3.37: Displacement of the core halves in (a) horizontal di-

rection or (b) vertical direction. (c) Results. Figures from [57].

pointed out that the leakage flux in transformers built of tape wound
cores may generate significant core losses.
In [54] the loss increase in cut tape wound cores has been observed
as well. However, an increase in the air gap length didn’t lead to a
further core loss increase. A possible explanation can be given by the
fact that there is a trade-off between the two causes: an increase of the
air gap length results in more orthogonal flux; however, on the other
hand, the flux through the cut surface, where the short circuits are
located, is reduced. It seems that in [54] the two effects balance each
other.
For the sake of completeness, it should be briefly discussed why in
stacked laminated iron cores, e.g. silicon steel cores, the effect of a loss
increase in gapped cores is much smaller. It is obvious that interlam-
ination short circuits are not present in these type of cores (different
machining process, mechanically less fragile). Furthermore, the layer

94
 3.10. INFLUENCE OF CORE SHAPE ON CORE LOSSES

orientation is different; the laminations are normally stacked, hence,

the flux through the core window doesn’t lead to a flux orthogonal to
the layer.
In this thesis only a brief overview about this important topic is
given. The interested reader is referred to the cited articles. This topic
gives opportunities for further research, since, to the author’s knowl-
edge, there exists no approach which allows analytically describing the
presented effects.

95
Chapter 4

Winding Loss Modeling

The second source of losses in inductive components is the ohmic losses

in the windings. The resistance of a conductor increases with an increase
of frequency due to eddy currents. Self-induced eddy currents inside
a conductor lead to the skin-effect. Eddy currents due to an external
alternating magnetic field, e.g. the air gap fringing field or the magnetic
field from other conductors, lead to the proximity effect. In order to
limit losses related to skin- and proximity effects, a careful winding
design is important.
Most windings are realized with solid round conductors since they
are low priced. However, in return, they have high skin- and proximity-
effect losses. In order to reduce these high frequency (HF) losses, differ-
ent other types of windings are available, such as foil and litz wire wind-
ings. However, there are situations in which foil or litz wires are even
lossier than solid round windings. The winding arrangement is another
degree of freedom that allows controlling winding losses. For instance,
interleaved transformer windings substantially reduce proximity-effect
losses.
In Section 4.1 and Section 4.2 the skin- and proximity-effects are
qualitatively explained. Later, the different winding types are discussed,
including a discussion about the determination of the external magnetic
field which is needed in order to calculate the proximity-effect losses.
In Section 4.3 round conductors (including litz wires in Section 4.3.5),
and in Section 4.4 foil conductors are discussed. Analytical equations
are given which allow calculating winding losses.

97
CHAPTER 4. WINDING LOSS MODELING

H-field
y
x
z
r
d d/2
Jz
(a) (b)

Figure 4.1: (a) Cross section of a round conductor that is infinitely

long in the z-direction with homogenous current density Jz . (b) Mag-
nitude of the corresponding H-field.

Hϕ

E / Jeddy
Jz
Hϕ

Figure 4.2: Self-induction of eddy currents due to Faraday’s law.

4.1 Skin Effect

The magnetic field of a current through a conductor can be determined
with Ampere’s law. Ampere’s law is given as
I ZZ
Hdl = J dA, (4.1)

where H is the magnetic field strength vector and J is the current

density vector. The magnitude of the resulting H-field of a current
through a round conductor is qualitatively illustrated in Figure 4.1.
In case the current through the conductor is an alternating current
(AC), the associated alternating magnetic field self-induces an electric
field inside the conductor. Faraday’s law of induction is the basic law

98
 4.1. SKIN EFFECT

350

300
f = 100kHz
Current density [kA/m ]
2

250

200

150 f = 20kHz
f = 50Hz f = 5kHz
100

50

0
0 0.5 1 1.5 2
Distance to the conductor center [mm]

Figure 4.3: Current distribution for different frequencies in a round

conductor with a radius of 2 mm. The current magnitude is 1 A.

that describes this effect; it is given as

I ZZ
d
Edl = − BdA, (4.2)
dt
where E is the electric field vector and B is the magnetic flux den-
sity vector. The effect in case of a round conductor is illustrated in
Figure 4.2. The induced electric field results in a current, which coun-
teracts the excitation current in the center of the conductor, hence,
the major part of the current is flowing in an outer layer (skin) of the
conductor. This change in the current distribution is illustrated in Fig-
ure 4.3 for different frequencies using the example of a round conductor.
The current density is decreasing from the outer conductor boundary
to the conductor center. The distance from the outer boundary to where
the current density falls to 1/e of the maximum is called the skin depth
(or penetration depth) δ, and, for µr = 1, can be approximated as [5]
1
δ=√ , (4.3)
πµ0 σf
where µ0 is the magnetic constant, σ is the conductivity of the con-
ductor material, and f is the frequency of the sinusoidal current. The
assumption that µr is one is valid for conductor materials such as cop-
per, aluminum, etc. As a simplification, the current can be considered

99
CHAPTER 4. WINDING LOSS MODELING

as concentrated in a skin layer with width δ and with a constant current

density.
The skin-effect losses (including DC losses) per unit length can be
calculated as
PS = FR/S (f ) · RDC · Iˆ2 , (4.4)

where RDC is the DC resistance, Iˆ is the peak current, and FR/S =

RAC /RDC is a factor that describes the increase of the conductor resis-
tance due to the skin effect. Different formulae for FR/S (f ) are taken
for different conductor geometries, whereas the subscribed letter stands
for the conductor type (e.g. FR (f ) for round conductors).
The term "skin-effect losses" PS describes the losses due to the cur-
rent through the conductor, including the losses due to self-induced
eddy currents. Therewith, the DC losses are included in the skin-effect
losses.

4.2 Proximity Effect

The proximity effect is illustrated in Figure 4.4 using the example of
a round conductor. A current in conductor A results in a magnetic
field that induces eddy currents in conductor B. On the facing side, the
induced current in conductor B has a direction opposite to the current
in conductor A. In other words, the current distribution of a conductor
that is penetrated by an external magnetic field changes and additional
winding losses occur. The induced current distribution of a round con-
ductor is illustrated in Figure 4.5 for different frequencies. In Figure 4.6
the current density of two neighboring conductors is illustrated; in Fig-
ure 4.6(a) the currents have opposite directions, while in Figure 4.6(b)
both conductors have the same current direction. As can be seen, the
current concentrates there where the magnetic field H concentrates.
The losses due to the proximity effect are described as

PP = RDC · GR/S (f ) · Ĥe2 , (4.5)

where RDC is the DC resistance, Ĥe is the peak external magnetic field,
and GR/S (f ) is a factor that describes the amount of winding losses due
to the proximity effect. Different formulae for GR/S (f ) are taken for
different conductor geometries, whereas the subscribed letter stands for
the conductor type (e.g. GR (f ) for round conductors).

100
 4.2. PROXIMITY EFFECT

Hϕ

Hϕ

Jz E / Jeddy
Hϕ

A B

Figure 4.4: Induction of eddy currents due to Faraday’s law.

Current density [kA/m2]

f = 100kHz

f = 15kHz
f = 50Hz

f = 5kHz

Distance to the conductor center [mm]

Figure 4.5: Induced current distributions of a conductor with a

radius of 2 mm. The conductor is penetrated by an external alternat-
ing H-field of amplitude Ĥe = 50 A/m and of different frequencies.
Figure from [58].

The term "proximity-effect losses" PP describes the losses due to

eddy currents induced by an external magnetic field He .

101
CHAPTER 4. WINDING LOSS MODELING

Figure 4.6: Current density of two neighboring conductors; The

currents in the two conductors have in (a) opposite directions, and in
(b) same directions.

Figure 4.7: Skin- and proximity-effect losses per unit length of a

round conductor. f = 100 kHz, Iˆ = 1 A, and the magnitude of the
external magnetic field Ĥ = 1000A/m.

4.3 Round Conductor

Most windings are realized of solid round conductors since they are
low priced. However, in return, they have high skin- and proximity-
effect losses. In order to keep these losses low, the diameter must be
selected well. The skin effect losses in solid round conductors decrease
for increasing conductor diameters, while the proximity effect losses
increase with increasing diameters; accordingly, there is an optimum
conductor diameter to a given problem. This is illustrated in Figure 4.7.

102
 4.3. ROUND CONDUCTOR

y
Jz
x
z d

Figure 4.8: Cross section of the considered round conductor with

a current density in z-direction. The conductor is infinitely long in
z-direction.

In Section 4.3.1 and Section 4.3.2 the expressions to determine ana-

lytically skin- and proximity-effect losses are given. The external mag-
netic field strength He of every conductor has to be known in order to
calculate the proximity-effect losses. In Section 4.3.3 a 1D calculation
of the external field for un-gapped transformers is given. In the case
of gapped cores, such 1D approximations are not applicable since the
fringing field of the air gap cannot be described in a 1D manner. A 2D
approach is introduced in Section 4.3.4. Winding losses can be reduced
by using litz-wire windings, which are discussed in Section 4.3.5. In
Section 4.3.6 calculations are compared to FEM simulations.

4.3.1 Skin Effect

The geometry considered to calculate the skin effect in round conductors
is illustrated in Figure 4.8. The round conductor has the diameter d,
and the length l, whereas it is assumed that d  l, thus the round
conductor is considered as infinitely long in z-direction. The skin-effect
losses (including DC losses) per unit length can be calculated as

PS = RDC · FR (f ) · Iˆ2 (4.6)

with
1
δ=√ ,
πµ0 σf
d
ξ=√ ,
2δ
4
RDC = ,
σπd2

103
CHAPTER 4. WINDING LOSS MODELING

y Jz He
x
z d

Figure 4.9: Cross section of a round conductor that is influenced by

an external magnetic field in x-direction. The conductor is infinitely
long in z-direction.

and1 
ξ ber0 (ξ)bei1 (ξ) − ber0 (ξ)ber1 (ξ)
FR = √
4 2 ber1 (ξ)2 + bei1 (ξ)2
 (4.7)
bei0 (ξ)ber1 (ξ) + bei0 (ξ)bei1 (ξ)
− .
ber1 (ξ)2 + bei1 (ξ)2
The derivation of (4.6) is from [58, 60] and is given in Appendix A.8.

4.3.2 Proximity Effect

A round conductor with diameter d that is positioned parallel to the
z-axis, is penetrated by an alternating magnetic field with magnitude
Ĥe . The considered situation is illustrated in Figure 4.9. The resulting
proximity-effect losses per unit length can be calculated as

PP = RDC · GR (f ) · Ĥe2 (4.8)

with
1
δ=√ ,
πµ0 σf
d
ξ=√ ,
2δ
4
RDC = ,
σπd2
1 The solution is based on a Bessel differential equation that has the form x2 y 00 +

xy 0 + (k2 x2 − v 2 )y = 0. With the general solution y = C1 Jv (kx) + C2 Yv (kx),

whereas Jv (kx) is known as Bessel function of the first kind and order v and Yv (kx)
is known as Bessel function of the second kind and order v [59]. Furthermore, to
resolve Jv (kx) into its real- and imaginary part, the Kelvin functions can be used:
3
Jv (j 2 x) = berv x + j beiv x.

104
 4.3. ROUND CONDUCTOR

and
ξπ 2 d2 ber2 (ξ)ber1 (ξ) + ber2 (ξ)bei1 (ξ)

GR = − √
2 2 ber0 (ξ)2 + bei0 (ξ)2
 (4.9)
bei2 (ξ)bei1 (ξ) − bei2 (ξ)ber1 (ξ)
+ .
ber0 (ξ)2 + bei0 (ξ)2
The derivation of (4.8) is from [58, 60] and is given in Appendix A.8.

4.3.3 Multi-Layer Windings Without Air Gap

With (4.6) and (4.8) the losses in a single solid round conductor can be
calculated. For it, the current Iˆ through the conductor and the external
magnetic field Ĥe , mostly present due to neighboring windings, has to
be known. In this section it is shown how the winding losses of an un-
gapped transformer with many round conductor turns are calculated.
The applied method to calculate the external magnetic field has been
derived in [61].
The case considered is illustrated in Figure 4.10(a). The conductors
are assumed to be enclosed by an ideal magnetic conductor (µ → ∞).
The magnetic field is assumed to have only a direction as illustrated.
The following considerations are made for only one winding shown in
Figure 4.10 (i.e. winding 1); the solution for winding 2 is the same due to
symmetry. The magnetic field strength Hz parallel to every conductor
has to be known to calculate the proximity-effect losses. This magnetic
field can be determined with Ampere’s law. For instance, for the field
calculation between the first and second layer, Ampere’s law can be
written as
Ĥz,1 bF = NL Iˆ1 , (4.10)
where NL is the number of conductors per layer and bF is the core
window width (cf. Figure 4.10). Hence, the magnetic field between the
first and second conductor is

NL Iˆ
Ĥz,1 = , (4.11)
bF
between the second and third conductor

2NL Iˆ
Ĥz,2 = , (4.12)
bF
etc.

105
CHAPTER 4. WINDING LOSS MODELING

Round Conductor Core

x
µ→∞
y Hz Hz
z

bF

(a) C Winding 1 Winding 2

Symmetry
Hz

(b)
y

Figure 4.10: (a) Cross section of the left core-winding window of

an E-core transformer that is built of solid round conductors. (b)
Corresponding H-field distribution.

The distribution of the magnetic field is illustrated in Figure 4.10(b).

The magnetic field is constant between two conductors, and ascending
within a conductor. The same magnetic field strength has to be assumed
on both sides of the conductor in order to use (4.8). The average value
of the two magnetic fields of the two sides is taken (Havg = 12 (Hleft +
Hright )).
Hence, for a setting as in Figure 4.10 with NL ML turns (ML : the
number layers), the average magnetic field for layer m is

2m − 1 NL Iˆ
Ĥavg = , (4.13)
2 bF

with m = 1 . . . ML .
The total losses for a winding with an average winding length lm

106
 4.3. ROUND CONDUCTOR

are then
ML
!
X
P = RDC ˆ2
FR I NL ML + NL GR 2
Ĥavg lm
m=1 (4.14)
4ML2 − 1
 
ˆ2 3
= RDC I NL ML FR + NL ML GR lm
12b2F

with
4
RDC = ,
σπd2
NL the number of conductors per layer, and
ML the number layers.

4.3.4 Multi-Layer Windings With Air Gap

The external magnetic field strength He of every conductor has to be
known when calculating the proximity-effect losses. In the case of an
un-gapped transformer core and windings that are fully-enclosed by
core material, a 1D approximation to calculate the magnetic field exist,
as shown in the preceding section. However, in the case of gapped
cores, this 1D approximation is not applicable since the fringing field
of the air gap cannot be described in a 1D manner. An approach has
to be selected to overcome this limitation. The approach selected for
this thesis can be used for inductive components with and without air
gaps and is based on the work presented in [5, 62]. The approach is
presented in the following.
A winding arrangement as illustrated in Figure 4.11(a) leads to an
external field vector Ĥ across conductor qxi ,yk due to the current îxu ,yl
of conductor qxu ,yl of

îxu ,yl
Ĥ = − j p
2π (xu − xi )2 + (yl − yk )2
(xu − xi ) + j(yl − yk )
·p (4.15)
(xu − xi )2 + (yl − yk )2
îxu ,yl ((yl − yk ) − j(xu − xi ))
= .
2π ((xu − xi )2 + (yl − yk )2 )
Complex numbers are used to identify the conductor position as this
simplifies the calculation. The magnitude of the total external field Ĥe

107
CHAPTER 4. WINDING LOSS MODELING

jy
∆x
qxm ,yn
jyn
ix m,yn
∆y
jy2
jy1

x1 x2 xm x

(a)

µ→ ∞

(b)

Figure 4.11: (a) Illustration of winding arrangement. (b) Illustra-

tion of mirroring.

across conductor qxi ,yk can be calculated with

 
m Xn
X îxu ,yl ((yl − yk ) − j(xu − xi )) 
Ĥe =  (u, l) , (4.16)


u=1 l=1
2π ((xu − xi )2 + (yl − yk )2 ) 

where (u, l) = 0 for u = i and l = k, and (u, l) = 1 for u 6= i or l 6= k.

The proximity-effect losses can be calculated based on (4.8) with the
determined external magnetic field Ĥe .
The impact of a magnetic conducting material can be modeled with
the method of images (also known as mirroring). There, the shape of
the magnetic field can be modeled by removing the magnetic material
and injecting currents in a way such that the magnetic field remains

108
 4.3. ROUND CONDUCTOR

µ→ ∞ µ→ ∞

Figure 4.12: Illustration of modeling an air gap as a fictitious con-

ductor.

Section 2
Section 1

Core

Winding

Figure 4.13: Illustration of different sections in the winding path of

an E-core.

unchanged. The additional currents are the mirrored version of the

original currents, as illustrated in Figure 4.11(b). In case of windings
that are fully-enclosed by magnetic material (i.e. in the core window),
a new wall is created at each mirroring step as the walls have to be
mirrored as well. The mirroring can be continued to push the walls
away. This is illustrated in Figure 4.11(b).
The presence of an air gap can be modeled as a fictitious conductor
without eddy currents equal to the magneto-motive force (MMF) across
the air gap [5] as illustrated in Figure 4.12.
In the case of an E-core, for instance, the winding losses have to be
calculated differently for each section illustrated in Figure 4.13 as the
mirroring leads to different results.
In the case of a non-sinusoidal current it is permissible to express
the current as a Fourier series, calculate the losses for each frequency
component independently, and then total them up. It is also permissible
to total up the independently calculated skin effect and proximity-effect
losses [63]. Both are shown in Appendix A.9.
It is possible to accurately model winding losses of solid conduc-

109
CHAPTER 4. WINDING LOSS MODELING

Hi
r0
y He da di

x
z

Figure 4.14: Cross-sectional area of litz-wire winding.

tors based on the above discussion. In Section 4.3.6 a discussion about

the expected accuracy and comparisons with FEM simulations are pre-
sented. Based on the above derivations, a litz wire calculation is also
possible, which will be shown in the following Section 4.3.5.

4.3.5 Litz-Wire Windings

Winding losses can be reduced by using litz-wire windings. Litz-wire
windings are made up of multiple individually-insulated strands as il-
lustrated in Figure 4.14. In litz-wire windings, skin- and proximity
effects can be further divided into strand-level and bundle-level effects
[64]. The effect related to eddy currents circulating in paths involving
multiple strands is called the bundle-level effect, whereas strand-level
effects take place within individual strands. The strands are twisted
such that, ideally, each strand equally occupies each position in the
bundle. Therewith, bundle-level effects can be strongly reduced and
are neglected in the following formulae. However, bundle-level effects
may have an impact on winding losses in case the litz wire is not ideally
twisted, as have been reported in [65].
The skin-effect losses, including DC losses, of a litz-wire winding
that consists of n strands, each with strand diameter di are calculated
as [58, 62]
!2
Iˆ
PS,L = n · RDC · FR (f ) · , (4.17)
n
4
where RDC = σπd 2 , and FR (f ) is calculated as (4.7) with the strand
i
diameter d = di .
The magnetic field that leads to proximity-effect losses is the sum of
the external magnetic field He and the internal magnetic field Hi . The
external magnetic field He originates from the neighboring conductors

110
 4.3. ROUND CONDUCTOR

250

200

150
P [mW]

Internal Prox. Effect

100 External Prox. Effect
Skin Effect
50

0
0 20 40 60 80 100
f [kHz]

Figure 4.15: Skin, internal prox. and external prox. effect losses
per unit length in a litz wire as a function of the frequency f . With
25 × (di = 0.5 mm), Iˆ = 5 A, and the magnitude of the external
magnetic field Ĥ = 300 A/m.

or the air gap fringing field and can be calculated in the same manner as
described for the case of solid round conductors. The internal magnetic
field Hi across one strand originates from its neighboring strands. It
is assumed that the current is equally distributed over the litz-wire
cross-sectional area for the calculation of the internal magnetic field.
Each strand is, furthermore, assumed to be penetrated by the average
internal magnetic field.
The proximity-effect losses in litz-wire windings can then be calcu-
lated as [58, 62]

PP,L = PP,L,e + PP,L,i

!
Iˆ2 (4.18)
= n · RDC · GR (f ) Ĥe2 + 2 2
2π da
4
where RDC = σπd 2 , and GR (f ) is calculated as (4.9) with the strand
i
diameter d = di . The losses PP,L,e are named the external proximity-
effect losses, since they originate from the external magnetic field. The
losses PP,L,i are named the internal proximity-effect losses, since they
originate from the internal magnetic field (the field from the neighboring
strands). In Figure 4.15 the skin-, internal proximity- and external
proximity-effect losses of a litz wire are plotted as a function of the
frequency f .

111
CHAPTER 4. WINDING LOSS MODELING

250

200 Litz Wire (25 x (di = 0.5 mm))

150
P [mW]

Solid Wire (da = 2.5 mm) Litz Wire (50 x (di = 0.35 mm))
100
Litz Wire (100 x (di = 0.25 mm))
50

0
0 10 20 30 40 50 60 70 80
f [kHz]
(a)

-4
10

Litz Wire (25 x (di = 0.5 mm))

-6
10 Solid Wire (da = 2.5 mm)
P / He2

-8
10

-10
10 3 4 5 6
10 10 10 10
f [Hz]
(b)

Figure 4.16: Losses (per unit length) of litz wires and a solid round
wire with all the same cross-sectional area, as a function of the fre-
quency f ; the current per wire is Iˆ = 5 A. (a) Skin and internal
proximity effects (no external magnetic field); (b) external proximity
effect.

An important question is whether litz wires are always better than

solid round conductors. In Figure 4.16 losses of litz wires and a solid
round wire are compared to each other. All wires have the same cross-
sectional area. As can be seen in Figure 4.16(a), above a certain fre-
quency the litz wire losses exceed the solid wire losses; the reason is
that the internal proximity-effect losses become dominant and deterio-
rate the performance of litz wires. But, not only because of the internal
proximity-effect losses, but also because of the external proximity-effect
losses, litz wire become worse above a certain frequency. This can be
seen in Figure 4.16(b), where the external proximity-effect losses of a

112
 4.4. FOIL CONDUCTOR

litz wire above some frequency exceed the (external) proximity-effect

losses of a solid round conductor. Generally, it can be said that the
higher the number of strands the better the litz wire (in case the cop-
per cross-sectional area is kept constant). A high number of strands is
favorable in case the copper cross-sectional area is kept constant; how-
ever, a higher number of strands brings the disadvantage that the space
occupied by insulation is high. In [64] this issue is addressed and it is
shown how to optimally select the number and diameter of strands.

4.3.6 Accuracy Analysis

The formulae above are based on analytical equations. The major sim-
plification that has been made is to neglect the magnetic field of the
induced eddy currents. The calculated field He is assumed to penetrate
completely thorough the conductors. According to [5] this approxima-
tion is valid if the largest dimension of the conductor, i.e. the diameter
for a round conductor/strand, is less than 1.6 times the skin depth δ.
Consequently, the frequency fmax , up to which the calculation for a
given conductor diameter d is accurate, can be calculated as

2.56
fmax = . (4.19)
πµ0 σd2

Another simplification that has been made is that it was neglected

that the external field might be bent within the conductor, i.e. the
field might not be homogeneous inside the conductor. However, for the
considered cases, the approach at hand showed to have a high accuracy.
Solutions to overcome the two mentioned problems, i.e. high frequency
effects or bent flux lines, can be found in literature, e.g. in [5].
Three winding arrangements have been calculated and compared to
2D FEM simulations. The FEM simulations have been carried out with
the software FEMM2 . The results are given in Figure 4.17. Simulations
for frequencies below fmax lead to deviations of always less than 5 %,
above fmax more than 5 %, but, in all illustrated cases, never more than
25 %.

113
CHAPTER 4. WINDING LOSS MODELING

80 MA/m2 3 MA/m2 3 MA/m2

40 MA/m2 1.5 MA/m2 1.5 MA/m2

0 MA/m2 0 MA/m2 0 MA/m2

Arrangement 1 Arrangement 2 Arrangement 3

f = 20 kHz f = 10 kHz f = 100 kHz

cal.
FEM
fmax cal.
FEM

Arrangement 1
fmax
P [W/m]

cal.
Arrangement 2

FEM
fmax

Arrangement 3

f [Hz]

Figure 4.17: Comparison of winding loss calculation to FEM sim-

ulation. Size of core window: 37 mm x 10.15 mm. Air gap length
1 mm. Winding arrangement 1: number of turns N = 423; Iˆ = 1 A;
winding diameter d = 0.5 mm. Winding arrangement 2: number of
turns N = 108; Iˆ = 1 A; winding diameter d = 1 mm. Winding ar-
rangement 3: number of turns N = 5; number of strands n = 37;
Iˆ = 1 A; strand diameter di = 0.4 mm.

114
 4.4. FOIL CONDUCTOR

12

11.5
d = 1.95 mm
11
P [mW]

10.5
µ→∞

10
10 mm x 0.3 mm

9.5

8.5
0 1 2 3 4 5 6 7 8
(a) f [kHz]
3.8

3.6
d = 1.95 mm
3.4
P [mW]

3.2 10 mm x 0.3 mm

2.8
0 2 4 6 8 10 12 14 16
(b) f [kHz]

Figure 4.18: Comparison between the losses of solid round and foil
conductors. Size of foil conductors: 10 mm x 0.3 mm; diameter of
round conductor d = 1.95 mm; current per conductor Iˆ = 1 A. All
losses per unit length. (a) Windings enclosed by magnetic material;
(b) single conductors not enclosed by magnetic material.

4.4 Foil Conductor

Foil windings are another winding type that allow reducing high fre-
quency losses. Furthermore, they show advantages compared to litz
wires such as a higher filling factor or lower price. However, they have
drawbacks such as increased winding capacitances and the risk of an
orthogonal flux. A flux orthogonal to the foil conductor leads to high
eddy current losses.
In Figure 4.18(a) foil and solid round windings are compared to
2 FEMM 4.2, freeware from www.femm.info (date of download: February 2011).

115
CHAPTER 4. WINDING LOSS MODELING

11.5
0.7 MA/m2
8 kHz
10 mm x 0.3 mm
11 0.35 MA/m2

0.0 MA/m2

10.5
P [mW]

10
d = 1.95 mm

9.5

8.5
0 1 2 3 4 5 6 7 8
(a) f [kHz]

13.5
0.4 MA/m2
8 kHz d = 1.95 mm

13
0.2 MA/m2

0.0 MA/m2

12.5
P [mW]

12
10 mm x 0.3 mm

11.5

11
0 1 2 3 4 5 6 7 8
f [kHz]
(b)

Figure 4.19: Comparison between the losses of solid round and foil
conductors. Size of foil conductors: 10 mm x 0.3 mm; diameter of
round conductor d = 1.95 mm; current per conductor Iˆ = 1 A. All
losses per unit length. (a) Three windings not enclosed by magnetic
material, all conductors with current in one direction. (b) Windings
with return conductors not enclosed by magnetic material.

each other. As can be seen the foil winding has substantially lower
losses compared to a solid round winding. The reason is that the "skin"
of a foil conductor with the same cross-sectional area is larger com-
pared to the "skin" of a solid round conductor, hence the skin-effect
losses are substantially reduced. In Figure 4.18(a) the foil winding is
enclosed by magnetic material which guarantees that the magnetic field
is always parallel to the conductor, i.e. no orthogonal flux is present.
In Figure 4.18(b) and Figure 4.19(a) the situation is different. The foil
winding is not enclosed by a magnetic material, i.e. the parallel field

116
 4.4. FOIL CONDUCTOR

HS1
y
x Jx h
z HS2
b

Figure 4.20: Cross section of a foil conductor with a current density

in x-direction. The conductor is infinitely long in x-direction.

lines are not guaranteed. In this situation, the current density is not
constant along the width, i.e. the current is concentrated at the ends
of the windings. This increases the winding losses in the foil windings
substantially. The losses can be higher than in solid round conductors.
The current is concentrated on the ends of the winding, since there is a
flux orthogonal to the foil conductors, as can be seen in Figure 4.19(a).
In the situation where return conductors are placed anti-parallel to the
main conductors, as illustrated in Figure 4.19(b), the field is parallelized
and the losses in foil conductors are reduced. It can be concluded that
HF losses can be substantially reduced by the use of foil windings. How-
ever, a careful design, in which orthogonal flux is avoided, is crucial.
In Section 4.4.1 and Section 4.4.2 the analytical expressions for skin
and proximity-effect losses in foil windings are given. The external
magnetic field strength He of every conductor has to be known when
calculating the proximity-effect losses. In Section 4.4.3 a 1D derivation
of the external field for un-gapped transformers is given. However, in
the case of gapped inductors, such 1D approximations are not applicable
as the fringing field of the air gap cannot be described in a 1D manner.
In Section 4.4.5 an approach how to calculate losses in gapped inductors
is given. In Section 4.4.6 calculations are compared to FEM simulations.

4.4.1 Skin Effect

The geometry considered to calculate the skin-effect losses in foil wind-
ings is illustrated in Figure 4.20. The conductor has a width b and a
height h, whereas h  b. It is assumed that the current is flowing in
x-direction, with frequency f and magnitude I. ˆ The skin-effect losses
(including DC losses) per unit length of this geometry can be calculated

117
CHAPTER 4. WINDING LOSS MODELING

as
PS = FF (f ) · RDC · Iˆ2 (4.20)
with
1
δ=√ ,
πµ0 σf
h
ν= ,
δ
1
RDC = ,
σbh
and
ν sinh ν + sin ν
FF = .
4 cosh ν − cos ν
The derivation of (4.20) is from [58, 60] and is given in Appendix A.7.

4.4.2 Proximity Effect

y
x h He

z
b

Figure 4.21: Cross section of a foil conductor that is influenced by

an external magnetic field in z-direction. The conductor is infinitely
long in x-direction.

The geometry considered to calculate the proximity-effect losses in

foil windings is illustrated in Figure 4.21. On both conductor sides
the magnetic field strength in z-direction has the magnitude Ĥe . The
proximity-effect losses per unit length of this geometry can be calculated
as
PP = RDC · GF (f ) · Ĥe2 (4.21)
with
1
δ=√ ,
πµ0 σf
h
ν= ,
δ

118
 4.4. FOIL CONDUCTOR

1
RDC = ,
σbh
and
sinh ν − sin ν
GF = b2 ν .
cosh ν + cos ν
The derivation of (4.21) is from [58, 60] and is given in Appendix A.7.

4.4.3 Foil Multi-Layer Without Air Gap

With (4.20) and (4.21) the losses in a single foil conductor can be cal-
culated. For it, the current Iˆ through the conductor and the external
magnetic field Ĥe , mostly present due to neighboring windings, has to
be known.
In the following, the winding losses of an un-gapped transformer
with many foil conductor turns are calculated. The case considered
is illustrated in Figure 4.22(a). The conductors are assumed to be
enclosed by a magnetic ideal conductor (µ → ∞). The magnetic field
between two layers is then calculated with Ampere’s law similar as in
the approach for round conductors (cf. Section 4.3.3). For instance, the
magnetic field between the first and second conductor is

Iˆ
Ĥz,1 = . (4.22)
bF

The distribution of the magnetic field is illustrated in Figure 4.22(b).

The average value of the two magnetic fields of the two sides is taken
to calculate the proximity-effect losses for each conductor.
Hence, the losses in winding 1 of an un-gapped transformer with foil
windings with average winding length lm (cf. Figure 4.22) is
N
X
P = RDC (FF Iˆ2 N + GF 2
Ĥavg )lm (4.23)
m=1
2
 
ˆ2 4N − 1
= RDC I FF + GF N lm (4.24)
12b2F

with
1
RDC = ,
σbh
and N the number of turns per winding.

119
CHAPTER 4. WINDING LOSS MODELING

Foil winding Core

x
µ→∞
y Hz Hz
z

bF

(a) C Winding 1 Winding 2

Symmetry
Hz

(b)
y

Figure 4.22: (a) Cross section of the left core-winding window of an

E-core transformer that is built of foil conductors. (b) Corresponding
H-field distribution. Figure from [58].

4.4.4 Short Foil Conductors

The width bL of foil conductors are in practical implementations smaller

then the width of the core window bF . There are also situations in which
more then one foil conductor per layer is present. Short foil conductors
can be transformed into an equivalent foil conductor with width b0L = bF
in order to be able to use the equations derived in the previous section.
The transformation is illustrated in Figure 4.23. Dowell introduced
this transformation in [61] such that the DC resistances of the original
and the transformed foils are the same. For it, Dowell introduced the
"porosity factor" η, which is calculated as

NL bL
η= , (4.25)
bF

120
 4.4. FOIL CONDUCTOR

bL

bF

bL

µ→∞

Figure 4.23: Transformation of short foil conductors into foils of

window size height.

with NL the number of foil conductors per layer. The conductivity σ,

skin depth δ, and variable ν are accordingly redefined as

σ 0 = ησ, (4.26)
1
δ0 = √ , (4.27)
πf σ 0 µ0
and
h
ν0 = . (4.28)
δ0
Equations (4.26), (4.27), and (4.28), together with the adapted width
b0L = bF , can directly be used in the equations for the skin- and proximity-
effect losses, (4.20) and (4.21). In case more than one foil conductor per
layer is present, (4.24) has to be (slightly) adapted in order to consider
that the field is now generated by the current NL Iˆ per layer (and not
ˆ
only I).

4.4.5 Foil Multi-Layer With Air Gap

In case of gapped inductors with a foil winding, it is not appropriate
to use the equations of the previous sections since the air gap leads
to a field orthogonal to the foils of the winding. This field has to
be taken into consideration when modeling winding losses. References

121
CHAPTER 4. WINDING LOSS MODELING

bF dwg

dw

Figure 4.24: Illustration of situation considered for winding loss

modeling of gapped foil winding inductors.

[5, 66] give approaches how to calculate losses in gapped inductors.

The approach in [66] is based on a superposition of the losses from the
1-D field calculation of the previous sections and losses due to eddy
currents caused by the air gap fringing field calculated with a 2-D field
calculation. The derivation of approach [66] leads to relatively complex,
difficult to handle formulae. The approach derived in [5] is much easier
and straight forward to implement, therefore, this approach has been
selected to be briefly discussed in the following.

The situation considered for the winding loss modeling of gapped

foil winding inductors is illustrated in Figure 4.24. It is assumed that
the tip of the foil winding is very close to the magnetic material, the air
gap is very small compared to the distance between air gap and first
foil conductor dwg , and the skin depth δ is considerably larger than the
foil thickness.

The winding loss per unit length for a foil winding structure as

122
 4.4. FOIL CONDUCTOR

illustrated in Figure 4.24 with a single air gap becomes

 2  2 
Iˆ

0 ∞
ρ N 2  F
√ Z b /2 X 1 1
P = 0
    dx − 
δ bF
 (x+kb )π

F
−bF /2 k=−∞ 2dwg cosh
2dwg

(4.29)

with
Nd
ρ0 = ρ ,
dw
r
0 dw
δ =δ ,
Nd
√
where ρ is the resistivity of the foil material and δ = 1/ πµ0 σf is the
skin depth.
For cases where dwg /bF < 0.25, (4.29) can further be simplified to

ˆ 2
 
ρ0 N √I2  dwg

P = 1−π . (4.30)
δ 0 πdwg bF

4.4.6 Accuracy Analysis

Four winding arrangements have been calculated and compared to 2D
FEM simulations. The FEM simulations have been carried out with the
software FEMM3 . The results are given in Figure 4.25. The winding ar-
rangements 1 and 2 are in transformer configuration, i.e. with a primary
and a secondary winding. The winding arrangements 1 and 2 are calcu-
lated based on the discussion given in Section 4.4.3 and Section 4.4.4.
The winding arrangements 3 and 4 are in gapped inductor configura-
tion, i.e. with only a primary winding. The winding arrangements 3
and 4 are calculated based on the discussion given in Section 4.4.5.
It can be seen in Figure 4.25, that the calculation of winding arrange-
ment 1 is very accurate, with a deviation of always less than 3.1 %. The
winding arrangement 1 has windings with the same width as the core
window, i.e. bL = bF . For the winding arrangement 2 the winding width
is smaller than the core window width, i.e. bL = 33 mm < bF = 37 mm,
hence, the windings have to be transformed as discussed in Section 4.4.4.
The accuracy is still good, with a deviation of always less than 6.5 %.
3 FEMM 4.2, freeware from www.femm.info (date of download: February 2011).

123
CHAPTER 4. WINDING LOSS MODELING

1
10

Arrangement 4
0
10 cal.
FEM Arrangement 3
d=δ
P [W/m]

-1
10 cal.
FEM FEM
cal.
Arrangement 2
-2
10 Arrangement 1

-3
10 3 4
10 10
f [Hz]

Figure 4.25: Comparison of winding loss calculation to FEM simu-

lation. Size of core window: 37 mm x 10.15 mm (bF = 37 mm).
Winding arrangement 1: number of turns N = 2 × 7; Iˆ = 1 A;
foil thickness h = 0.4 mm; winding width bL = 37 mm; distance
between foils 0.2 mm; distance between core leg and first winding
dwg = 1.0 mm; no air gap; transformer configuration.
Winding arrangement 2: number of turns N = 2 × 10; Iˆ = 1 A;
foil thickness h = 0.3 mm; winding width bL = 33 mm; distance
between foils 0.1 mm; distance between core leg and first winding
dwg = 1.0 mm; no air gap; transformer configuration.
Winding arrangement 3: number of turns N = 14; Iˆ = 1 A;
foil thickness h = 0.4 mm; winding width bL = 37 mm; distance
between foils 0.2 mm; distance between core leg and first winding
dwg = 1.0 mm; air gap length 0.5 mm; inductor configuration.
Winding arrangement 4: number of turns N = 20; Iˆ = 1 A;
foil thickness h = 0.3 mm; winding width bL = 37 mm; distance
between foils 0.1 mm; distance between core leg and first winding
dwg = 1.0 mm; air gap length 0.5 mm; inductor configuration.

The winding arrangements 3 and 4 have been calculated with (4.30).

The winding arrangements 3 and 4 have windings with the same width

124
 4.4. FOIL CONDUCTOR

as the core window, i.e. bL = bF . The deviation between (4.29) and

(4.30) has been determined to be very small, hence, the losses have been
calculated with (4.30). However, the accuracy is not very high. In both
arrangements, the highest accuracy has been achieved in a range where
the winding thickness d is smaller than the skin depth δ, as can be seen
in Figure 4.25 where the points of d = δ are labeled. However, for very
high and for very low frequencies the accuracy is worse. In conclusion,
the approach introduced in Section 4.4.5 gives, in a limited range, a
reasonable good estimation of the losses. Nevertheless, before building
the inductor, a confirmation with a FEM simulation is indispensable.

125
Chapter 5

Thermal Modeling

In the last two chapters, the losses occurring in magnetic components

have been discussed. Another important aspect in the design phase of
inductive components is the expected temperature of the component.
This is not only important to avoid overheating, it has also importance
to correctly model losses, since the losses are influenced by the tem-
perature. Hence, a thermal model is indispensable for an accurate loss
model. This chapter illustrates ways to model the thermal behavior of
the components. The thermal modeling is based on a simplified two
dimensional resistor network. Each resistor is determined based on cal-
culations of heat conduction, radiation, and natural convection. In this
thesis, only natural convection and no forced convection is considered.
In Section 5.1 a brief overview about thermal modeling is presented,
and in Section 5.2 the heat transfer mechanisms are summarized and
the formulae needed for modeling are given.

5.1 Overview of Thermal Models

Basically, the three physical mechanisms conduction, radiation, and
convection lead to transported heat. These three mechanisms have to
be considered when modeling the thermal behavior of a given geometry.
Generally, in order to avoid a computationally intensive finite element
simulation, a common approach is to calculate the thermal behavior
via thermal resistor networks. In the following the concept of thermal
resistor networks is discussed.

127
CHAPTER 5. THERMAL MODELING

TL Ploss Rth TA

Figure 5.1: Thermal model with only one thermal resistance.

5.1.1 Total Thermal Resistance

The temperature of a magnetic component is often considered as homo-
geneous and the thermal behavior is characterized by a single resistance
Rth . Such a model is illustrated in Fig. 5.1. The resistor is formally de-
termined according to the thermal-electric analogy and the ohmic law
with
∆T
Rth = P , (5.1)
P
where ∆T is the difference
P between the ambient and magnetic compo-
nent temperature and P is the sum of the occurring losses inside the
magnetic component. Major drawback of this approach is that a linear
function as (5.1) is not valid, i.e. Rth changes with changing (absolute)
temperature. However, when one defines Rth as a function of the am-
bient and magnetic component temperatures, the approach improves
substantially. The resistor is then determined based on calculations of
heat conduction, radiation, and convection.

5.1.2 Thermal Resistor Networks

The above introduced model can be improved by taking into considera-
tion that the temperature is unequally distributed inside the magnetic
component. For this, the resolution of nodes of constant temperature
can be chosen arbitrarily. Models of different complexity have been
derived in the past. For example, in [10] for each winding turn a new
temperature node is introduced, that leads to a quite complex resistor
network. Simpler approaches have been implemented e.g. in [7] or [22],
where one node of constant temperature represents the core temper-
ature, and one node of constant temperature represents the winding
temperature. This simplification is motivated by the fact that the core
and winding materials are usually good heat conductors. The accord-
ing resistor networks for inductors and transformers are illustrated in
Figure 5.2. Generally, the goal of any modeling task is to evaluate a
model that is as simple as possible, but still guarantees a reasonably
high accuracy.

128
 5.1. OVERVIEW OF THERMAL MODELS

TA

Rth,CA
PC
TC

Rth,CW
TA PW1
TW1

Rth,CA
Rth,W1
PC
TC

Rth,CW Rth,W1W2
PW PW2
TW
TW2

Rth,W
Rth,W2

Rth,WA
Rth,WA

TA
TA
(a) (b)

Figure 5.2: Thermal model of (a) inductor and (b) transformer

(with two windings).

In a model such as illustrated in Figure 5.2, it is recommended to

add a resistor for the heat conduction from the inside to the outside of
the winding. This is favorable, as otherwise only the winding surface
temperature would be calculated, which is lower compared to the inner
temperature. This resistor is important to model the temperature drop
within the winding. By doing this, the winding losses are assumed to
occur all on the inner winding side, which leads to a presumed higher

129
CHAPTER 5. THERMAL MODELING

temperature. However, the simplification is justified by the reduced

complexity that is achieved with it.
For the sake of completeness, it is mentioned that only the steady
state temperature is calculated. The dynamics of the temperature is
not modeled. This could be done by adding capacitors representing the
heat capacity.
The major difficulty is to calculate the values of the network resis-
tances correctly. This is done based on calculations of heat radiation,
conduction and convection, and will be introduced in the following.

5.2 Heat Transfer Mechanisms

Heat is energy transferred from one body to another due to thermal
contact. Energy can only be transferred by heat between bodies of
different temperatures, where the sink has to be cooler. The physical
mechanisms of heat transfer that occur in magnetic components are
conduction and radiation. The term "convection" is used to describe
the combined effect of conduction and fluid flow, which for this work
is considered as an additional heat transfer mechanism. Each network
resistance is determined based on calculations of heat conduction, con-
vection, and radiation. In the following, the three heat transfer mecha-
nisms are explained one by one. Formulae that allow to determine the
resistor network are given.

5.2.1 Thermal Conduction

Thermal conduction (or heat conduction) is the transfer of heat between
neighboring molecules in a substance due to a temperature gradient

~q = −λ · ∇T, (5.2)

where ~q is the heat flux, λ is the heat conductivity, and ∇T the tem-
perature gradient. The heat conductivity λ is a material parameter and
can be considered as constant for most materials employed in magnetic
components. Material parameters can e.g. be extracted from [67].
The thermal resistance for a cuboid as illustrated in Figure 5.3 with
length l and cross section A is calculated with

∆T l
Rth = = . (5.3)
P Aλ

130
 5.2. HEAT TRANSFER MECHANISMS

q A

l
Figure 5.3: Thermal resistance of a cuboid: Rth = l/(Aλ).

Although (5.3) looks rather simple, the calculation of heat conduction

can be challenging. For instance, one difficulty is to model interfaces
between materials.

5.2.2 Thermal Natural Convection

The term convection is used to describe the combined effect of conduc-
tion and fluid flow. When a body is facing a fluid, the fluid absorbs
and transports heat. Convection can be calculated by solving the dif-
ferential Navier Stokes equations analytically or numerically. However,
empirical approaches that overcome solving this complex mathemati-
cal problem exist. They are widely applicable for different geometries.
Basically, the heat transfer via convection is described with
P = αA(Tb − Tg ), (5.4)
where P is the heat flow, A the surface area, Tb the body surface
temperature, and Tg the fluid temperature. α is a coefficient that is
influenced by [10]
I the absolute temperature,

I the material property of the fluid,

I the flow rate of the fluid,

I the dimensions of the considered surface,

I orientation of the considered surface,

I and the surface texture.

A set of characteristic dimensionless numbers has been introduced to
determine α; they are summarized in Table 5.1. In the following, this
characteristic numbers are described one by one.

131
CHAPTER 5. THERMAL MODELING

Name Symbol Measure of . . .

Nusselt number Nu . . . improvement of heat transfer
compared to the case with hypothetical
static fluid.
Grashof number Gr . . . ratio between buoyancy and
frictional force of fluid.
Prandtl number Pr . . . ratio between viscosity and
heat conductivity of fluid.
Rayleigh number Ra . . . flow condition (laminar or
turbulent) of fluid.

Table 5.1: Characteristic numbers to describe convection empirically

[10].

Characteristic Dimensionless Numbers

I The Nusselt number is a measure of improvement of heat transfer
compared to the case with hypothetical static fluid. The fluid
next to a hot surface heats up and transports heat constantly
away, hence the heat transfer capability is improved. The Nusselt
number is calculated with
αl
Nu = , (5.5)
λ
where λ is the heat conductivity of the fluid, α is the coefficient
of (5.4), and l is the characteristic length that will be discussed
later. Interesting is that the Nusselt number follows a law that
can be described with an empirical equation as
N u = f (Gr, P r), (5.6)
where Gr and P r are other characteristic numbers. Formulae for
(5.6) are given in literature, e.g. in [67], for different heat transfer
problems. Equations (5.5) and (5.6) lead to an equation for α
N u(Gr, P r)λ
α= . (5.7)
l
Equation (5.7) inserted in (5.4) solves the heat transfer problem.

132
 5.2. HEAT TRANSFER MECHANISMS

I The Grashof number approximates the ratio of the buoyancy to

viscous force acting on a fluid. The Grashof number is calculated
with
gl3
Gr = 2 β∆T, (5.8)
v
where g is the acceleration due to Earth’s gravity, β is the volu-
metric thermal expansion coefficient (equal to approximately 1/Tg
for ideal fluids, where Tg is the absolute fluid temperature), v is
the kinematic viscosity1 , l the characteristic length, and ∆T the
temperature difference between fluid and considered surface tem-
perature.

I The Prandtl number is approximating the ratio of kinematic vis-

cosity and thermal diffusivity. It is dependent only on the fluid
and the fluid state. For air at 20 ◦ C . . . 100 ◦ C it is [67]

P r = 0.7081 . . . 0.7004(= 0.7). (5.9)

I The Rayleigh number is defined as

Ra = Gr · P r (5.10)

It is a measure of flow condition (laminar or turbulent) of the

fluid. When the Rayleigh number is below a critical value for
that fluid, laminar flow occurs. Above that critical value the flow
condition is turbulent. There exists a third fluid state, where the
air is basically not moving, hence for very low Rayleigh numbers
the heat transfer is mainly in the form of conduction.

Formulae for Nusselt numbers

In the last section, the characteristic dimensionless numbers have been
introduced, wherewith the thermal problem can be solved. The remain-
ing open question is to find the appropriate function for the Nusselt
number N u = N u(Gr, P r). For this, in literature empirical formulae
are given. The presented equations in this work are from [67]. Formu-
lae for different geometries that are of interest for magnetic components
are introduced in the following.
1 kinematic viscosity of air at 30 ◦ C [67]: v = 162.6 · 10−7 m2 /s.

133
CHAPTER 5. THERMAL MODELING

b
a
h

b
a

(a) Vertical Plane (b) Horizontal Plane: Top (c) Horizontal Plane: Bottom

h
s

(d) Horizontal Closed Gap (e) Vertical Closed Gap

Figure 5.4: Illustration of convection geometries.

I Vertical Plane
The formula for the Nusselt number to determine the mean heat
transfer capability of a vertical plane as illustrated in Figure 5.4(a)
is h i 2
1/6
N u = 0.825 + 0.387 (Ra · f1 (P r)) , (5.11)

with
"  9/16 #−16/9
0.492
f1 = 1 + . (5.12)
Pr

134
 5.2. HEAT TRANSFER MECHANISMS

The formula above is valid for Ra = 10−1 . . . 1012 and P r > 0.001.
For the characteristic length in (5.5), (5.7), and (5.8) the height
h of the plane is chosen.
I Horizontal Plane: Heat Emission on Top Side
The formula for the Nusselt number to determine the mean heat
transfer capability of a horizontal plane with heat emission on top
side (cf. Figure 5.4(b)) is
(
1/5
0.766 (Ra · f2 (P r)) for Ra · f2 (P r) ≤ 7 · 104
Nu = 1/3
0.15 (Ra · f2 (P r)) for Ra · f2 (P r) > 7 · 104
(5.13)
with "  11/20 #−20/11
0.322
f2 = 1 + . (5.14)
Pr

The characteristic length in (5.5), (5.7), and (5.8) is calculated

for rectangular planes of size a × b with

ab
l= , (5.15)
2(a + b)

and for circular disks with diameter d with

d
l= . (5.16)
4

I Horizontal Plane: Heat Emission on Bottom Side

The formula for the Nusselt number to determine the mean heat
transfer capability of a horizontal plane with heat emission on
bottom side (cf. Figure 5.4(c)) is
1/5
N u = 0.6 (Ra · f1 (P r)) (5.17)

with f1 according to (5.12). The characteristic length in (5.5),

(5.7), and (5.8) is calculated according to (5.15) or (5.16), de-
pending on the geometry.
I Horizontal Gap
The air is static in case of a heat transfer from top to bottom in
a horizontal gap as illustrated in Figure 5.4(d), hence N u = 1.

135
CHAPTER 5. THERMAL MODELING

In case the heat transfer is from bottom to top, the formula for
the Nusselt number for the mean heat transfer capability of a
horizontal gap is

 1 for Ra < 1708
Nu = 0.208Ra0.25 for 1708 ≤ Ra < 2.2 · 104 (5.18)
0.092Ra0.33 for Ra ≥ 2.2 · 104


For the characteristic length in (5.5), (5.7), and (5.8) the gap
length s of Figure 5.4(d) is chosen.
I Vertical Gap
The formula for the Nusselt number for the mean heat transfer
capability of a vertical gap as illustrated in Figure 5.4(e) is

−0.25
Nu = 0.42P r0.012 Ra0.25 hs for 104 ≤ Ra < 107
0.33
0.049Ra for 107 ≤ Ra < 109
(5.19)
The dimensions s and h are illustrated in Figure 5.4(e). For the
characteristic length in (5.5), (5.7), and (5.8) the gap length s is
chosen.

Formulae for other geometries and for forced convection are given
e.g. in [67].

5.2.3 Thermal Radiation

Each surface of an object that has a temperature above absolute zero
radiates its thermal energy in the form of electromagnetic waves. The
Stefan-Boltzmann law states the total energy radiated per unit surface
area of a black body in vacuum

q = σT 4 , (5.20)

with
2π 5 k 4
σ= , (5.21)
15c2 h3
where k = 1.38·10−23 J/K is the Boltzmann constant, h = 6.62·10−34 Js
is Planck’s constant, and c = 3 · 108 m/s is the speed of light in vacuum.
(5.20) is an idealized equation for a black body that in reality does
not exist. The radiation property of a real surface deviates from the

136
 5.3. PRACTICAL IMPLEMENTATION ISSUES

Material 
Silver (polished) 0.02
Copper 0.1-0.2
Aluminium (oxidized) 0.2-0.3
Enamel 0.8-0.95
Ceramic 0.9-0.95

Table 5.2: Values for selected emissivities (for an industrial temper-

ature range) [10].

radiation property of a surface of a black body. This is taken into

consideration by an empirical correction factor, the emissivity . The
emissivity  is a function of temperature, emission angle, and wave-
length. The Stefan-Boltzmann law for the heat transfer between two
surfaces with emissivities 1 and 2 becomes then [10]
P = eff A1 σ(T14 − T24 ). (5.22)
with
1

1 A 1

,
+ A1

 1 2 −1
2
eff = (5.23)

1 for A2  A1 .


In Table 5.2 the emissivities of some selected materials are given.

Important to note is that optically transparent materials are not nec-
essarily transparent for heat radiation. For example, for an enamel-
insulated wire, the emissivity of enamel has to be taken, not the one of
copper.
The linear resistance value of a resistor network can be calculated
with (5.22) for one thermal operating point with
∆T T1 − T2
R= = . (5.24)
P eff A1 σ(T14 − T24 )

5.3 Practical Implementation Issues

With above formulae, an arbitrarily magnetic component can be ther-
mally modeled. However, the thermal resistors are nonlinear with tem-

137
CHAPTER 5. THERMAL MODELING

A1 A2 = A1 π/2

Figure 5.5: Illustration of increased winding surface.

perature; therefore, the thermal calculation has to be iteratively per-

formed: first a starting temperature is assumed, then the network is
linearized and the calculation is conducted. The calculation is then re-
done with the new calculated temperature. This is repeated until the
algorithm has converged.
Another important thing to be discussed is the actual surface of the
component to be modeled. The shape of conductors may increase the
surface of the winding-ambient boundary. In case of having round con-
ductors and modeling heat convection, one has to multiply the idealized
flat surface with π/2 in order to consider the conductor shape. This is
illustrated in Figure 5.5. Note that this is not valid for heat radiation,
as the round surface does not improve the heat transfer capability and
simply leads to heat transfer between nearby conductors.

138
Chapter 6

Magnetic Design
Environment and
Experimental Results

It is difficult to conduct a calculation that combines all loss and thermal

aspects discussed in the previous chapters. The impact of peak-to-peak
flux density ∆B, frequency f , DC premagnetization HDC , temperature,
core shape, minor and major loops, flux waveform, and material on core
loss calculation have been considered in Chapter 3. In order to calculate
winding losses, formulae for round, foil and litz wire conductors, each
including skin- and proximity-effects (including the influence of an air-
gap fringing field) have been given in Chapter 4. Furthermore, thermal
models in order to avoid overheating and improve the model accuracy
have been discussed in Chapter 5. In order to handle these models
and enable others to determine losses accurately, in the course of this
thesis a Magnetic Design Environment has been implemented. This
environment is introduced in Section 6.1.
Experimental tests that confirm the overall accuracy have been con-
ducted. The experimental results are presented in Section 6.2 and Sec-
tion 6.3. It will be shown that a high level of accuracy is achieved by
combining all loss and thermal models introduced in the previous chap-
ters. More experimental results will be given in Chapter 7, where a
design procedure for the mains side LCL filter of an active three-phase
rectifier is introduced.

139
CHAPTER 6. MAGNETIC DESIGN ENVIRONMENT AND
EXPERIMENTAL RESULTS

6.1 Overview of Implemented Loss Model-

ing Environment
In order to increase the power density of magnetic components, it is
indispensable to model their losses and their thermal behavior accu-
rately. For it, a Magnetic Design Environment consisting of a core loss
measurement system, a core material database, and magnetic design
software has been built. The system is illustrated in Figure 6.1. It
applies all models that have been introduced in the previous chapters.
The Automated Core Loss Measurement System is being built to
analyze core losses under general flux waveform excitation. The built
test system is performing all measurements automatically (starts exci-
tation, controls current, regulates flux, triggers the oscilloscope, reads
value). Rectangular and sinusoidal voltage excitation is possible when
performing such automated measurements. Furthermore, it is possible
to automatically extract the relaxation parameters. The idea is to au-
tomatically set up the loss map introduced in Section 3.9, which enables
an accurate core loss calculation.
The Core Material Database (Loss Map) is a database linked to the
Automated Core Loss Measurement System. It allows a standardized
storage of measurement results. Loss density data are stored in order
to calculate core losses of cores of general shapes.
The Magnetic Design Software is being built to make magnetic com-
ponent modeling as simple as possible, while still taking all important
loss and thermal effects into consideration. With its graphical user
interface it enables a straightforward design of magnetic components.
It is possible to read data from the database in order to improve the
core loss calculation. Furthermore, it is possible to import voltage and
current waveforms simulated on a circuit simulator, such as e.g. Mat-
lab/Simulink, Simplorer, or Gecko Circuits. With it, it is possible to
model inductive components under actual in-circuit conditions.

140
 Core Material Database Magnetic Design Software Circuit Simulator
(Loss Map)
Relaxation B-H-Relation
αr βr
kr τ
qr

LF HF
(Ansoft Simplorer)
ΔB, f, T, HDC ΔB, f, T, HDC
ΔB,
ΔB, f,f,T,T,HHDC ΔB, f, T, HDC
ΔB, DC ΔB,
ΔB,f,f,T,T,HHDCDC
ΔB,f,f,T,T,HHDCDC ΔB, f, T, HDC
ΔB, f, T, HDC ΔB, f, T, HDC

EE EE
E EE
EE
E
E E

Automated Core Loss Measurement System Prototype Verification

Power Supply (e.g. with calorimeter,
Oscilloscope 325V power meter, ...)

Power Stage
Current
Probe

Voltage CUT
Probe

Heating Chamber

Figure 6.1: Illustration of the implemented Magnetic Design Envi-

ronment.

141
ENVIRONMENT
6.1. OVERVIEW OF IMPLEMENTED LOSS MODELING
CHAPTER 6. MAGNETIC DESIGN ENVIRONMENT AND
EXPERIMENTAL RESULTS

Figure 6.2: E-core used for measurements in Experiment I.

6.2 Experiment I
Loss measurements have been conducted on an inductor to verify the
loss calculations. The inductor has been built with two E-cores having
an air gap in the center leg (EPCOS Ferrite N27; Core E55/28/21; air
gap length lg = 1 mm [29]). The windings are made of a solid copper
wire with diameter 1.7 mm. The number of turns has been N = 18. A
photo of the inductor is given in Figure 6.2. The measurements have
been carried out with a Yokogawa WT3000 Precision Power Analyzer.
Two different types of loss measurements have been performed and
compared to loss calculations. The first type of measurements have
been conducted with symmetric triangular flux waveforms at different
operating points for which the results are given in Table 6.1. The second
type of measurements have been conducted with a low frequency LF
(100 Hz) sinusoidal current with a superimposed high frequency HF
(10 kHz) triangular current ripple. A low modulation index has been
chosen, hence the high frequency peak-to-peak flux density ∆BHF is
(almost) constant over the full low frequency period. The results of
the second measurement are given in Table 6.2. A high level of overall
accuracy with a maximum deviation of 10 % has been achieved.

142
 Operating Points Calculated Losses Measured Losses Comparison
∆B f HDC Core Losses Winding Losses Total Losses Total Losses Rel. Error
[T] [kHz] [A/m] [W] [W] [W] [W] [%]
0.25 1 0 0.06 0.11 0.17 0.18 -5.56
0.5 1 0 0.32 0.42 0.74 0.77 -3.90
0.25 2 0 0.14 0.12 0.26 0.27 -3.70
0.5 2 0 0.72 0.49 1.21 1.21 0.00
0.25 5 0 0.4 0.23 0.63 0.61 3.28
0.5 5 0 2.04 0.92 2.96 2.70 9.63
0.1 10 0 0.11 0.08 0.19 0.20 -3.06
0.2 10 0 0.56 0.33 0.89 0.88 1.14
0.3 10 0 1.36 0.74 2.10 2.01 4.48

Table 6.1: Results to measurements with symmetric triangular flux

waveforms (Experiment I).

Operating Points Calculated Losses Measured Losses Comparison

IˆLF ∆BLF ∆BHF Core Losses Winding Losses Total Losses Total Losses Rel. Error
[A] [T] [T] [W] [W] [W] [W] [%]
5 0.25 0.15 0.44 0.35 0.79 0.76 3.95
10 0.5 0.3 1.83 1.51 3.35 3.6 -6.94

Table 6.2: Results to measurements with sinusoidal current with

a superimposed high frequency triangular current ripple (Experi-
ment I). LF: 100 Hz; HF: 10 kHz.

143
6.2. EXPERIMENT I
CHAPTER 6. MAGNETIC DESIGN ENVIRONMENT AND
EXPERIMENTAL RESULTS

Label Core Air gap N Wire diam.

length [mm] [mm]
L1 E25/13/7 1.05 27 0.8
L2 E32/16/9 0.55 18 0.8
L3 E20/10/6 1 80 0.45

Table 6.3: Different investigated designs in Experiment II for ther-

mal and loss measurements. Cores from EPCOS (material N87).

6.3 Experiment II
In the second experiment, thermal and loss measurements on different
inductors have been conducted. The different designs are listed in Ta-
ble 6.3. The loss measurements have been carried out with a Yokogawa
WT3000 Precision Power Analyzer. The surface core and surface wind-
ing temperatures have been measured with the infrared camera FLIR
ThermaCAM. The results are given in Table 6.4, Table 6.5, and Ta-
ble 6.6. Again, a high level of accuracy with a maximum deviation of
12 % in losses and temperatures has been observed.

144
 Op. Points Calculated Losses and Temperature Meas. Losses and Temperatures
∆B f Core Winding Total Core Winding Total Core Winding
[kHz] Losses Losses Losses Temp. Temp. Loss Temp. Temp.
[W] [W] [W] [◦ C] [◦ C] [W] [◦ C] [◦ C]
0.3 10 0.07 0.40 0.47 35.6 44.2 0.50 40 49
0.3 20 0.14 0.59 0.73 40.6 52.5 0.74 43 58
0.35 20 0.18 0.81 0.99 45 60.9 1.01 49 66

Table 6.4: Results to inductor L1 in Experiment II.

Op. Points Calculated Losses and Temperature Meas. Losses and Temperatures
∆B f Core Winding Total Core Winding Total Core Winding
[kHz] Losses Losses Losses Temp. Temp. Loss Temp. Temp.
[W] [W] [W] [◦ C] [◦ C] [W] [◦ C] [◦ C]
0.3 10 0.14 0.31 0.45 33.3 38.0 0.45 35 43
0.3 20 0.28 0.37 0.65 37.2 41.4 0.60 37 45
0.35 20 0.38 0.52 0.90 40.7 46.5 0.94 44 52

Table 6.5: Results to inductor L2 in Experiment II.

145
6.3. EXPERIMENT II
146
Op. Points Calculated Losses and Temperature Meas. Losses and Temperatures
∆B f Core Winding Total Core Winding Total Core Winding
EXPERIMENTAL RESULTS

[kHz] Losses Losses Losses Temp. Temp. Loss Temp. Temp.

[W] [W] [W] [◦ C] [◦ C] [W] [◦ C] [◦ C]
0.3 10 0.04 0.24 0.28 34.2 40.6 0.27 37 42
0.3 20 0.07 0.33 0.40 37.8 45.7 0.38 39 46
0.35 20 0.10 0.45 0.55 41.5 52.2 0.52 44 52
0.4 50 0.22 1.44 1.66 63.9 96.7 1.60 60 91

Table 6.6: Results to inductor L3 in Experiment II.

CHAPTER 6. MAGNETIC DESIGN ENVIRONMENT AND
Chapter 7

Multi-Objective
Optimization of
Inductive Power
Components

The models introduced in the previous chapters will form the basis for
the optimization of inductive components employed in key power elec-
tronic applications. The aim of this chapter is to use these previously
derived models and to show the optimization procedure on a particular
example. The chosen example is an LCL input filter structure for a
three-phase Power Factor Correction (PFC) rectifier.
LCL input filters are an attractive solution to attenuate switching
frequency current harmonics of active voltage source rectifiers [68, 69].
The design procedure for LCL filters based on a generic optimization
approach is introduced guaranteeing low volume and/or low losses. Dif-
ferent designs are calculated showing the trade-off between filter volume
and filter losses. Furthermore, the converter (consisting of semiconduc-
tor switches, DC link capacitor, and cooling system) is also taken into
consideration in the optimization procedure. This is necessary as, for
instance, a high switching frequency leads to a lower filter volume and
losses, but on the other hand, leads to higher switching losses in the
semiconductors of the converter. To find the overall optimum, such

147
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

Parameter Variable Value

Input Phase Voltage AC Vmains 230 V
Mains Frequency fmains 50 Hz
DC-Voltage VDC 650 V
Load Current IL (nominal) 15.4 A
Switching Frequency fsw 8 kHz

Table 7.1: Specification of the three-phase PFC recitifier.

trade-offs have to be considered. Generally, it is important to consider

the system to be optimized as a whole, since there are parameters that
bring advantages for one subsystem while deteriorating another subsys-
tem.
In Section 7.1 the three-phase PFC rectifier is introduced, in Sec-
tion 7.2 the applied models of the LCL filter components are discussed,
and in Section 7.3 the optimization algorithm for the LCL filter is de-
scribed. Simulation and experimental results are given in Section 7.4
and Section 7.5 respectively. In Section 7.6 the converter volume and
losses are taken into consideration, i.e. an overall system optimization
is performed.

7.1 Three-Phase PFC Rectifier with Input

Filter
The three-phase PFC rectifier investigated in this work is shown in Fig-
ure 7.1. The rectifier comprises in each phase a boost inductor L2 , a
damped LC filter L1 /C/Cd /Rd and a pair of switches with free-wheeling
diodes. The load is assumed to be a DC current source. The considered
operating point of the PFC rectifier is described in Table 7.1. A Space
Vector Modulation (SVM) scheme with loss-optimal clamping has been
implemented and the system is operated with a fundamental displace-
ment factor of cos φ = 1. The functionality and the detailed properties
of the used SVM scheme are described in [70]. The three-phase PFC
rectifier with input filter has been simulated in MATLAB/Simulink,
where the non-linearity of the core material of the inductors, i.e. the
change of the inductance value with changing current, is taken into
account.

148
 a

IL d
Winding

h ww
i2,a w
a i1,a L1,a L2,a
b L1,b L2,b N/2 N/2
VDC IL w
c L1,c L2,c
do
Vmains Rd,a Rd,b Rd,c
Ca Cb Cc B

Cd,a Cd,b Cd,c

t

(a) (b) A

Figure 7.1: (a) Three-phase PFC rectifier with LCL input filter. (b)
Cross-sections of inductors employed in the input filter.

149
7.1. THREE-PHASE PFC RECTIFIER WITH INPUT FILTER
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

The three inductors L1,a , L1,b , and L1,c , and the three capacitors
Ca , Cb , and Cc in star arrangement, together with the three boost
inductors L2,a , L2,b and L2,c , result in a third order LCL low pass fil-
ter between the mains and the switching stage. The capacitor/resistor
branches Cd,a /Rd,a , Cd,b /Rd,b , and Cd,c /Rd,c are necessary to damp
the resonance of the LC input filter. All inductors are assumed to have
the same geometry, which is illustrated in Figure 7.1(b). The cores are
made of grain-oriented steel (M165-35S, lamination thickness 0.35 mm).
Solid copper wire is taken for the conductors. The windings are divided
into two halves arranged on the two legs which leads to a more dis-
tributed winding structure. A more distributed winding structure has
advantages such as better heat dissipation capabilities, lower inductor
volume, etc.

7.2 Modeling of Input Filter Components

The modeling of the inductive components has been made according
to the discussions in Chapters 2 - 5. In the following, these models
are briefly recapitulated. Furthermore, models for capacitors are given,
and it is introduced how the damping branch is designed.

7.2.1 Calculation of the Inductance

The inductance of an inductive component with N winding turns and
a total magnetic reluctance Rm,tot is calculated as

N2
L= . (7.1)
Rm,tot
Accordingly, the reluctance of each section of the flux path has to be
determined first in order to calculate Rm,tot . The total reluctance for a
general inductor is calculated as a function of the core reluctances and
air gap reluctances. The core and air gap reluctances can be determined
applying the methods described in Chapter 2. The reluctances of the
core depend on the relative permeability µr which is extracted from the
nonlinear initial B-H-relation of the core material, hence the reluctance
is described as a function of the flux. Therefore, as the flux depends on
the core reluctance and the reluctance depends on the flux, the system
can only be solved iteratively by using a numerical method. In the case
at hand, the problem has been solved by applying the Newton approach.

150
 7.2. MODELING OF INPUT FILTER COMPONENTS

The reluctance model of the inductor of Figure 7.1(b) consists of

one voltage source (representing the two separated windings), one air
gap reluctance (representing the two air gaps) and one core reluctance.

7.2.2 Core Losses

The applied core loss approach is described in Chapter 3 in detail and
can be seen as a hybrid of the improved-improved Generalized Steinmetz
Equation i2 GSE and a loss map approach: a loss map is experimentally
determined and the interpolation and extrapolation for operating points
in between the measured values is then made with the i2 GSE.
The flux density waveform for which the losses have to be calcu-
lated is e.g. simulated in a circuit simulator, where the magnetic part
is modeled as a reluctance model. This simulated waveform is then di-
vided into its fundamental flux waveform and into piecewise linear flux
waveform segments. The loss energy is then calculated for the funda-
mental and all piecewise linear segments, summed and divided by the
fundamental period length in order to determine the average core loss.
The DC flux level of each piecewise linear flux segment is considered,
as this influences the core losses. Furthermore, the relaxation term of
the i2 GSE is evaluated for each transition from one piecewise linear flux
segment to another.
Another aspect to be considered in the core loss calculation is the
effect of the core shape/size. By introducing a reluctance model of the
core, the flux density can be calculated. Subsequently, for each core
section with (approximately) homogenous flux density, the losses can
be determined. In the case at hand, the core has been divided into four
straight core sections and four corner sections. The core losses of the
sections are then summed to obtain the total core losses.

7.2.3 Winding Losses

The second source of losses in inductive components are the ohmic
losses in the windings. The resistance of a conductor increases with
increasing frequency due to eddy currents. Self-induced eddy currents
inside a conductor lead to the skin effect. Eddy currents due to an
external alternating magnetic field, e.g. the air gap fringing field or the
magnetic field from other conductors, lead to the proximity effect.
The sum of the DC losses and the skin effect losses per unit length
in round conductors can be calculated with (4.6). The proximity effect

151
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

TL Ploss Rth TA

Figure 7.2: Thermal model with only one thermal resistance.

losses in round conductors per unit length can be calculated with (4.8).
The external magnetic field strength He of every conductor has to be
known when calculating the proximity losses. The applied approach is
a 2D approach and is described in detail in Chapter 4.

7.2.4 Thermal Modeling

A thermal model is important when minimizing the filter volume, since
the maximum temperature allowed is the limiting factor when reducing
volume. The model used in this work consists of only one thermal re-
sistance Rth and is illustrated in Figure 7.2. The inductor temperature
TL is assumed to be homogenous; it can be calculated as

TL = TA + Ploss Rth , (7.2)

where Ploss are the total losses occurring in the inductor, consisting
of core and winding losses, and TA is the ambient temperature. The
ambient temperature TA is assumed to be constant at 25 ◦ C.
The heat transfer due to convection is described with

P = αA(TL − TA ), (7.3)

where P is the heat flow, A the surface area, TL the body surface tem-
perature (i.e. inductor temperature), and TA the fluid (i.e. ambient air)
temperature. α is a coefficient that is determined by a set of char-
acteristic dimensionless numbers, the Nusselt, Grashof, Prandtl, and
Rayleigh numbers. Radiation has to be considered as a second impor-
tant heat transfer mechanism and is described by the Stefan-Boltzmann
law. Details about thermal modeling are given in Chapter 5.

7.2.5 Capacitor Modeling

The filter and damping capacitors have been selected from the EPCOS
X2 MKP film capacitor series; which have a rated voltage of 305 V.
The dissipation factor is specified as tan δ ≤ 1 W/kvar (at 1 kHz) [71],

152
 7.3. OPTIMIZATION OF THE INPUT FILTER

which enables an approximation of the capacitor losses. The capaci-

tance density to calculate the capacitors volume can be approximated
with 0.18 µF/cm3 . The capacitance density has been approximated by
dividing the capacitance value of several components by the according
component volume.

7.2.6 Damping Branch

An LC filter is added between the boost inductor and the mains to meet
a THD constraint. The additional LC filter changes the dynamics of the
converter and may even increase the current ripple at the filter resonant
frequency. Therefore, a Cd /Rd damping branch has been added for
damping. In [72, 73] it is described how to optimally choose Cd and Rd .
Basically, there is a trade-off between the size of damping capacitor Cd
and the damping achieved. For this work, Cd = C has been selected as it
showed to be a good compromise between additional volume needed and
a reasonable damping achieved. The value of the damping resistance
that leads to optimal damping is then [72, 73]
r
L1
Rd = 2.1 . (7.4)
C

The Cd /Rd damping branch increases the reactive power consump-

tion of the PFC rectifier system. Therefore, often other damping struc-
tures, such as the Rf -Lb series damping structure, are selected [73]. For
this work, however, the Cd /Rd damping branch has been favored as its
practical realization is easier and lower losses are expected. Further-
more, as will be seen in Section 7.4, the reactive power consumption
of the PFC rectifier system including the damped LC input filter is
in the case at hand rather small, and, if necessary, could be actively
compensated by the rectifier.
The losses in the damping branch, which occur mainly in the re-
sistors, are calculated and taken into consideration in the optimization
procedure as well.

7.3 Optimization of the Input Filter

The aim is to optimally design a harmonic filter of the introduced three-
phase PFC rectifier. For the evaluation of different filter structures, a

153
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

cost function is defined that weights the filter losses and filter volume
according to the designer needs. In the following, the steps towards an
optimal design are described. All steps are illustrated in Figure 7.4.
The optimization constraints are discussed first.

7.3.1 Optimization Constraints and Conditions

The high-frequency ripple in the current i2,a/b/c is limited to the value
IHF,pp,max , which is important as a too high IHF,pp,max , e.g., impairs
controllability (for instance, an accurate current measurement becomes
more difficult). Furthermore, the THD of the mains current is limited.
In industry a typical value for the THD that is required at the rated
operating point is 5 % [74]. In the design at hand, a THD limit of
4 % is selected in order to have some safety margin. Two other design
constraints are the maximum temperature Tmax and the maximum vol-
ume Vmax the filter is allowed to have. A fixed switching frequency
fsw is assumed. The DC link voltage VDC and the load current IL
of the converter are also assumed to be given and constant. All con-
straints/condition values for the current system are given in Figure 7.4.

7.3.2 Calculation of L2,min

The minimum value of the inductance L2,min can be calculated based
on the constraint IHF,pp,max as

2
√
3 VDC −2Vmains
L2,min = δ(100) · , (7.5)
IHF,pp,max · fsw

with the relative turn-on time √ of the space vector (100) when the current
of phase "a" peaks of δ(100) = 3M 2 cos(π/6). Equation (7.5) is based on
the fact that, in case of a fundamental displacement factor of cos φ = 1,
the maximum current ripple IHF,pp,max occurs when the current reaches
the peak value IˆLF of the fundamental. As a consequence, the minimum
value L2,min √has to be met at the current IˆLF . With the modulation
2 2Vmains
index M = VDC , (7.5) becomes
√ √ 2
√
3 2|Vmains | VDC − 2Vmains
L2,min = cos(π/6) · 3 . (7.6)
VDC IHF,pp,max · fsw

154
 7.3. OPTIMIZATION OF THE INPUT FILTER

L1 L2
i1 i2
C

Figure 7.3: Idealized current waveforms for each filter component.

7.3.3 Loss Calculation of Filter Components

In the foregoing sections (respectively chapters) it has been shown how
an accurate loss modeling based on simulated current and voltage wave-
forms is possible. However, such a calculation based on simulated wave-
forms is time consuming and therefore, for an efficient optimization,
simplifications have to be made. In Figure 7.3, idealized current wave-
forms for each filter component of a phase are illustrated. The current
in L1 is approximated as purely sinusoidal with a peak value of
2 IL VDC
Iˆ = √ , (7.7)
3 2Vmains
where Vmains is the RMS value of the mains-phase voltage. The reactive
current drawn by the filter capacitors is rather small and has been
neglected. With the mains frequency fmains = 50 Hz, losses, volume,
and temperature of L1 can be calculated.
The current in L2 has a fundamental (sinusoidal) component, with
an amplitude as calculated in (7.7) and a fundamental frequency of
fmains = 50 Hz, and a superimposed ripple current. The ripple current
is, for the purpose of simplification, in a first step considered to be sinu-
soidal with constant amplitude IHF,pp,max over the mains period. The
losses for the fundamental and the high-frequency ripple are calculated
independently, and then summed. By doing this, it is neglected that
core losses depend on the DC bias condition and it is neglected that
ripple amplitude varies over the cycle.
The ripple current is assumed to be fully absorbed by the filter
capacitor C, hence, with the given dissipation factor tan δ, the losses in
the capacitor can be calculated as well.
How the above mentioned simplifications affect the modeling accu-
racy will be discussed in Section 7.4.

155
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

7.3.4 Optimization Procedure

After the optimization constraints and the simplifications chosen for the
loss calculation have been described, the optimization procedure itself
will be explained next.
A filter design is defined by
 
aL1 aL2
 wL1 wL2 
 
 NL1 NL2 
 
 doL1 doL2 
X= , (7.8)
 hL1 hL2 
 
 tL1 tL2 
 
 wwL1 wwL2 
dL 1 dL2

where all inductor parameters are defined as in Figure 7.1(b). The

subscripts L1 and L2 describe to which inductor each parameter corre-
sponds. The capacitance value of the filter capacitor is calculated based
on the L1 value to guarantee that the THD constraint is met. The pa-
rameters in X are varied by an optimization algorithm to obtain the
optimal design. The optimization is based on the MATLAB function
fminsearch() that applies the Downhill-Simplex-Approach of Nelder
and Mead [75].
The optimization algorithm determines the optimal parameter val-
ues in X. A design is optimal when the cost function

F = kLoss · qLoss · P + kVolume · qVolume · V, (7.9)

is minimized. kLoss , kVolume are weighting factors, qLoss , qVolume are

proportionality factors, and P , V are the filter losses and filter volume,
respectively. The proportionality factors are chosen such that, for a
"comparable performance", qLoss · P and qVolume · V are in the same
range1 . In the case hand, this was achieved with qLoss = 1/W and
qVolume = 3 · 104 /m3 .
The following steps are conducted to calculate the filter losses P
and the filter volume V of (7.9) (cf. Figure 7.4):
1 The filter losses are in the range of approximately 100 W . . . 250 W, whereas

the filter volumes are in the range of approximately 0.001 m3 . . . 0.01 m3 . With
qLoss = 1/W and qVolume = 3 · 104 /m3 , the volume range is lifted to 30 . . . 300 and
therewith becomes comparable to the losses.

156
 7.3. OPTIMIZATION OF THE INPUT FILTER

Optimization Constraints and Conditions

- max. IHF,pp,max in boost inductors L2 4A
- max. THD of mains current 4%
- max. temperature Tmax 125 °C
- max. volume Vmax 10 dm3
- switching frequency fsw 8 kHz
- DC link voltage VDC 650 V
- load current IL 15.4 A
Calculate L2 L2

L2,N
L2,min
iLF
ILF

Optimization Procedure

Optimization algorithm
(varies parameter in matrix X)

1) Calculate
L2, PL2, VL2, TL2

Constraint Yes
violation?
No

2) Calculate
L1, PL1, VL1, TL1

Constraint Yes
violation?
No

3) Calculate and Select C

Calculate PC, VC

Constraint Yes
violation?
No

4) Calculate
F = f(P, V)

Optimization No
finished?
Yes

Output

Figure 7.4: The design procedure for three-phase LCL filters based
on a generic optimization approach.

157
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

1. Calculate the inductance value L2 as a function of the current.

Furthermore, the losses PL2 , the volume VL2 , and temperature
TL2 of the boost inductors L2 are calculated. In case a constraint
cannot be met, the calculation is aborted and the design is dis-
carded and a new design will be evaluated.
2. Calculate the inductance value L1 as a function of the current.
Furthermore, the losses PL1 , the volume VL1 , and temperature
TL1 of the filter inductors L1 are calculated. In case a constraint
cannot be met, the calculation is aborted and the design is dis-
carded and a new design will be evaluated.
3. For the purpose of simplification, the THD without filter is ap-
proximated with THD = IHF,pp,max /I(1),pp , where I(1),pp is the
fundamental peak-to-peak value of the mains current2 . Further-
more, it is assumed that the dominant harmonic content appears
at fsw (this assumption is motivated by simulation results). The
LC-filter has then to attenuate the ripple current by


A = 20 log10 I(1),pp THDmax /IHF,pp,max (7.10)
(in dB) at a frequency of fsw . Therewith, C is calculated as
1 1
C= = , (7.11)
L1 ω02 A
L1 (2πfsw · 10 40 dB )2
where ω0 is the filter cutoff frequency. The losses and the volume
of the capacitors and damping resistors can then be calculated.
In case a constraint cannot be met, the calculation is aborted and
the design is discarded and a new design will be evaluated.
4. The volume, temperature, and losses are now known and the cost
function (7.9) can be evaluated.
The optimal matrix X is found by varying the matrix parameters, eval-
uating these matrixes by repeating the above steps, and minimizing the
cost function (7.9). After the optimal design is found, the algorithm
quits the loop of Figure 7.4.
p
I22 +I32 +I42 +···+In
2
2 The THD is defined as THD = , where In is the RMS value
I1
of the nth harmonic and I1 is the RMS value of the fundamental current. Under
the assumption that only the dominant harmonic content IHF,pp,max (which, for
the purpose of simplification, is assumed to appear at one frequency) leads to a
harmonic distortion, the THD becomes IHF,pp,max /I(1),pp , consequently.

158
 7.4. OPTIMIZATION OUTCOMES

Investigated Design (cf. Fig. 7.6)

L1 L2
a 5.70 mm a 2.20 mm
250 w 12.0 mm w 21.7 mm
N 56 N 78
Prototype do 45.1 mm do 81.2 mm
h 60.8 mm h 28.4 mm
built
t 8.50 mm t 50.0 mm
200 ww 61.3 mm ww 27.7 mm
d 4.25 mm d 3.30 mm C 3 5.52 µF
P [W]

150

fSW = 4 kHz
fSW = 8 kHz
100
0 2 4 6 8 10
V [dm3]

Figure 7.5: The P -V -Pareto front showing filter volumes V and

filter losses P of different optimal designs.

7.4 Optimization Outcomes

The optimization procedure leads to different filter designs depend-
ing on the chosen weighting factors k1 and k2 in (7.9), i.e. depending
whether the aim of the optimization is more on reducing the volume V
or more on reducing the losses P . Limiting factors are the maximum
temperature Tmax (limits the volume from being too low) and a maxi-
mum volume Vmax (limits the efficiency from being too high). Different
designs are shown by a P -V -plot, i.e. a P -V -Pareto front in Figure 7.5;
the trade-off between losses and volume can be clearly identified.
One design of Figure 7.5 has been selected for further investiga-
tions. Particularly, a comparison between the (for the optimization
procedure) simplified and the more elaborate calculation based on volt-
age / current waveforms from a circuit simulator has been made. The
filter parameters of the selected design are detailed in Figure 7.5. The
circuit of the three-phase PFC rectifier with the selected input filter has
been simulated in MATLAB/Simulink, and the simulated current and
voltage waveforms have been taken to calculate the losses according to
Section 7.2. The results are given in Figure 7.6. The THD constraint
is met and the current IHF,pp,max is only insignificantly higher. The
simplified loss calculation used for the optimization leads to an over-
estimation of the boost inductor losses. This is due to the fact that

159
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

L1 L2

20 20

10 10
i [A]

i2 [A]
0 0
1

-10 -10

-20 -20
15 20 25 30 35 15 20 25 30 35
t [ms] t [ms]

THD = 3.97 % IHF,pp,max = ~ 4.1 A

PLoss = 2.23 W (simplified calculations: 1.65 W) PLoss = 27.3 W (simplified calculations: 34.7 W)
(a) (b)

Figure 7.6: Simulations and calculations to one selected design (cf.

Figure 7.5): (a) filter inductor L1,a , (b) boost inductor L2,a .

the maximum ripple current has been assumed to be constant over the
mains period. The losses in the filter inductors, on the other hand,
have been underestimated as any high frequency ripple in the current
through L1 has been neglected in the simplified calculations. One could
try to improve/change the simplifications made for the optimization and
therewith improve the simplified loss calculation. However, the differ-
ence between the two calculation approaches has been considered as
acceptable for this work.
Another important design criteria is the achieved power factor. The
reactive power consumption of the the PFC rectifier system, includ-
ing the damped LC input filter, is in the case at hand rather small
(power factor = [real power]/[apparent power] = 0.998).
So far, all results are based on simulations and calculations. The
models have to be verified experimentally to prove the validity of the
optimization procedure. In the following section experimental results
are shown.

7.5 Experimental Results

Experimental measurements have been conducted to show that the
above introduced calculations are valid. The filter prototype built in
the course of this work has been assembled of laminated sheets and
coil formers of standard sizes, in order to keep the costs low. This,
however, avoids an exact implementation of an optimum as can be seen

160
 7.5. EXPERIMENTAL RESULTS

Figure 7.7: Photo of the PFC converter.

Simulated Current Waveform Measured Current Waveform

20 20

10 10
i1 [A]

i [A]

0 0
1

-10 -10

-20 -20
20 25 30 35 40 -10 -5 0 5
t [ms] t [ms]
THD = 3.23 % THD = 3.86 %
PLoss = 3.0 W PLoss = 3.4 W
(a) (b)

Figure 7.8: (a) Simulations and (b) measurements on one of the

filter inductors L1 of the implemented design.

in Figure 7.5 (the prototype built is not on the 8 kHz-line). However,

this does not impair the significance of the measurement results. Spec-
ifications, dimensions and photos of the LCL filter built are given in
Figure 7.10. All measurements have been carried out with a Yokogawa
WT3000 Precision Power Analyzer.
The measurements have been conducted with the T-type converter
introduced in [76]. In Figure 7.7 a photo of the converter is given.
The T-type converter is a 3-level converter, however, a 2-level opera-
tion is possible as well. A 2-level operation with the same modulation
scheme (optimal-loss clamping modulation scheme) as in the MATLAB
simulation has been implemented.
The results of the comparative measurements and simulations are
given in Figure 7.8 and Figure 7.9. As can be seen, the calculated and

161
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

Simulated Current Waveform Measured Current Waveform

20 20

10 10
i [A]

i [A]
0 0
2

2
-10 -10

-20 -20

20 25 30 35 40 -10 -5 0 5 10
t [ms] t [ms]
IHF,pp,max = 4.1 A IHF,pp,max = 4.7 A
PLoss = 44.9 W PLoss = 46.8 W
(a) (b)

Figure 7.9: (a) Simulations and (b) measurements on one of the

boost inductors of the implemented design.

measured loss values are very close to each other. The maximum current
ripple in the actual system is (slightly) higher than in the simulation.
This can be explained by the fact that in the simulation the inductance
of the boost inductor is assumed to be constant over the full frequency
range. However, in reality the effective inductance decreases with in-
creasing frequency due to inductor losses and parasitic capacitances.
The higher THD value can also be explained with the same effect since
the filter inductance decreases with increasing frequency as well. The
frequency behavior could be modeled analytically by representing the
inductors as RLC networks.

7.6 Overall Rectifier Optimization

As can be seen in Figure 7.5, an increase in switching frequency leads

to lower filter losses and lower filter volumes. However, in return, an
increase in switching frequency leads to higher switching losses in the
converter semiconductors. In other words, it is important to consider
the system as a whole in order to achieve a truly optimal design. In
the following, first, a model for the converter is derived, which allows
to determine a P -V -Pareto front of the converter. Later, the trade-off
in the switching frequency is illustrated and the optimal frequency for
the overall system is determined.

162
 7.6. OVERALL RECTIFIER OPTIMIZATION

Figure 7.10: Specifications, dimensions and photos of (a) one filter

inductor L1 , (b) all filter and damping capacitors/resistors C, Cd and
Rd , and (c) one boost inductor L2 .

163
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

AT, AD, VCS,

Converter model Tj,T, Tj,D, V, P
I(1),pp, L2, fsw

Figure 7.11: Illustration of converter model.

Parameter Variable Value

◦
Max. junction temp. Tj,max 125 C
Max. cooling system vol. VCS,max 0.8 dm3
Heatsink height 4 cm
Max. area per chip AT,max /AD,max 1 cm2
DC link voltage VDC 650 V
Max. DC link voltage overshoot ∆VDC 50 V
Fund. peak-peak current I(1),pp 20.5 A

Table 7.2: Constraints and conditions for converter optimization.

7.6.1 Overall Optimized Designs

The Infineon Trench and Field Stop 1200 V IGBT4 series has been
selected to determine the P -V -Pareto front of the converter. These
semiconductors are the same as the ones employed in the converter of
Figure 7.7. A converter model has been set up, which determines the
junction temperature of the transistors Tj,T , the junction temperature
of the diodes Tj,D , the volume of the converter V , and the losses of
the converter P . As input variables it needs the transistor chip area
AT , the diode chip area AD , the cooling system volume VCS , and the
operating point defined by the switching frequency fsw and peak-to-
peak fundamental current I(1),pp . The high frequency ripple current
IHF,pp has, as long as IHF,pp  I(1),pp , only a negligible impact on the
losses and is therewith not taken into consideration. The inputs and
outputs of the converter model are illustrated in Figure 7.11 and all
optimization constraints and conditions are listed in Table 7.2. This
optimization based on the semiconductor chip area is motivated by
previously presented works [77, 78].
The thermal resistance of the cooling system, i.e. the heat sink sur-
face to ambient thermal resistance Rth,sa , has been modeled with the

164
 7.6. OVERALL RECTIFIER OPTIMIZATION

Cooling System Performance Index (CSPI) [79]; the Rth,sa is then cal-
culated as
1
Rth,sa = , (7.12)
CSP I · VCS
where the CSPI can be approximated to be constant for a given cooling
concept. The value 15 W/(K · liter) has been calculated for the cooling
concept of the converter shown in Figure 7.7 according to [79]. Accord-
ing to [77], for the selected IGBT series the junction-to-sink surface
thermal resistance Rth,js can be approximated as a function of the chip
area A as  −0.88
K A
Rth,js = 23.94 · . (7.13)
W mm2
The switching and conduction losses per chip area have been ex-
tracted from data sheets. Different IGBTs with different current rat-
ings, but from the same IGBT series have been analyzed. Since the
data sheets doesn’t provide information about the chip area, the IGBT
and diode chip area as a function of the nominal chip current IN has
to be known in order to determine the losses as a function of the chip
area. This chip area-current dependency has been taken from [77] and
is for the transistor

mm2
AT = 0.95 · IN + 3.2 mm2 , (7.14)
A
and for the diode

mm2
AD = 0.47 · IN + 3.6 mm2 . (7.15)
A
The switching losses scaled to the same current do not vary much with
the chip size in the considered range, as a comparison of different data
sheets of the selected IGBT series has shown; therefore, for this work,
the switching loss energies have been considered as independent of the
chip area. The switching loss energies have been extracted at two junc-
tion temperatures (25 ◦ C and 150 ◦ C) and, for intermediate tempera-
tures, a linear interpolation has been made. These, from the data sheet
extracted and interpolated, switching loss energies are then assigned to
each switching instant in order to determine the switching losses.
In return, the conduction losses depend on the chip area. This has
been modeled as in [78], but with taking the impact of the temperature

165
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

into consideration. The following equation describes the conduction

losses
Ron,N (Tj,T/D ) · AN 2
Pcond (A, i, Tj,T/D ) = Vf · i + ·i , (7.16)
A
where Vf is the forward voltage drop, Ron,N (Tj,T/D ) is the nominal on-
resistance, AN is the nominal chip area (for which Ron,N (Tj,T/D ) is
taken from the data sheet), i is the current through the transistor, and
A is the actual chip area. The on-resistance has been been extracted
at two junction temperatures (25 ◦ C and 150 ◦ C) and, for intermediate
temperatures, a linear interpolation has been made. The current i has
been calculated for the chosen modulation scheme. Equation (7.16) is
evaluated with different values for Vf and Ron,N , depending whether the
conduction losses of the transistor or the diode are calculated.
The converter volume is the sum of the cooling system volume, the
DC link capacitor volume, and the volume of the switching devices.
The switching device volume has been calculated by multiplying the
chip areas by a depth of 1 cm in order to approximate the transistor
volume. The cooling system volume is known as it is a model input
parameter. The DC link capacitor CDC has been selected such that the
DC link voltage in case of an abrupt load drop from nominal load to
zero load doesn’t exceed a predefined value. The maximum DC link
voltage increase ∆VDC after the load drop can then be approximated
as
√ !
1 3 2Vmains Iˆ2 1ˆ
∆VDC = · + I2 tmax ,
CDC 2 VDC fsw 2

where Iˆ2 is the peak value of the phase current through the boost in-
ductor L2 and tmax is the time difference between the moment where
the system sampled the load drop (i.e. latest 1/fsw after the load drop
occurred, since the sampling interval is 1/fsw ) and the moment where
the DC link voltage peaks (i.e. VDC + ∆VDC is reached). For the calcu-
lation of tmax , it is assumed that the maximum possible demagnetizing
voltage is applied to the boost inductor from the moment the system
samples the load drop. The chosen value for the maximum voltage over-
shoot is ∆VDC = 50 V. The DC link capacitors have been selected from
the EPCOS MKP DC link film capacitor series; which have a rated
voltage of 800 V. The capacitor losses are low and, therefore, have been

166
 7.6. OVERALL RECTIFIER OPTIMIZATION

110

100

90
fSW = 8 kHz
P [W]

80

70

60 fSW = 4 kHz

50
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
V [dm3]

Figure 7.12: The P -V -Pareto front showing converter volumes V

and converter losses P of different optimal designs.

neglected. The capacitance density to calculate the capacitors volume

can be approximated with 0.6 µF/cm3 .
The losses and volumes of other system parts, such as the DSP,
auxiliary supply, gate driver, etc. have not been considered.
An optimization algorithm has been set up similar to that of the
filter. A design is characterized by the parameters AT , AD , and VCS .
The optimization algorithm varies these parameter in order to find op-
timal designs. The optimization procedure leads to different designs
depending whether the aim of the optimization is more on reducing the
volume V or more on reducing the losses P . Different designs are shown
by a P -V -plot, i.e. a P -V -Pareto front in Figure 7.12; the trade-off be-
tween losses and volume can be clearly identified. It becomes clear that
a higher switching frequency leads to higher losses for a given volume.
Therefore, there must exist an overall optimal switching frequency at
which the system (including filter) losses are minimized.
Basically, one can combine the results from the filter and from the
converter and, therewith, determine the overall system performance.
The losses of loss optimized designs for different frequencies have been
calculated and are shown in Figure 7.13(a). In the case at hand, the
optimal switching frequency is at approximately 5 − 6 kHz. In Fig-
ure 7.13(b) the losses of a volumetric optimized designs are given for
different frequencies; it can be seen that the optimal switching frequency
is in the same range.
In case of loss optimized designs, the major volume part comes from

167
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

250

200 Total Losses

150 Filter Losses

P [W]

100

50
Converter Losses

0
2000 4000 6000 8000 10000 12000
f [Hz]
(a)

400

350 Total Losses

300

250
P [W]

200
Filter Losses
150
Converter Losses
100

50
2000 4000 6000 8000 10000 12000
f [Hz]
(b)

Figure 7.13: (a) The losses of loss optimized designs for different
frequencies; (b) the losses of volumetric optimized designs for different
frequencies.

the filter volume. This becomes clear when Figure 7.5 is compared with
Figure 7.12: the maximum filter volume is much higher than the max-
imum converter volume. This has to do with the selected constraints;
however, a further increase of the converter volume wouldn’t have a big
impact in further decreasing the losses, therefore, it can be concluded
that the constraints have been selected well.
In Figure 7.14 the volume of volumetric optimized designs for dif-

168
 7.7. CONCLUSION AND FUTURE WORK

5
Total Volume
4
V [dm ]
3

3
Filter Volume
2

1
Converter Volume
0
2000 4000 6000 8000 10000 12000
f [Hz]

Figure 7.14: The volume of volumetric optimized designs for differ-

ent frequencies.

ferent frequencies are plotted. As both, the filter size and the converter
size, decrease with increasing frequencies, a high frequency is favorable
with respect to system size. The converter size decreases with increas-
ing frequency because the DC link capacitance reduces with increasing
frequency. The fact that the volume decreases with increasing switching
frequency is only true in a limited frequency range; above this frequency
range the losses become very high and the size of the components in-
creases again so that the heat can be dissipated. This effect is not
visible here.

7.7 Conclusion and Future Work

A design procedure for three-phase LCL filters based on a generic op-
timization approach is introduced guaranteeing low volume and/or low
losses. The cost function, which characterizes a given filter design, al-
lows a weighting of the filter losses and of the filter volume according
to the designer’s need. Different designs have been calculated to show
the trade-off between filter volume and filter losses. Experimental re-
sults have shown that a very high loss accuracy has been achieved. To
improve the THD and current ripple accuracy the frequency behavior
of the inductors could be modeled as well.
As can be seen in Figure 7.5, a higher switching frequency leads

169
CHAPTER 7. MULTI-OBJECTIVE OPTIMIZATION OF INDUCTIVE
POWER COMPONENTS

to lower filter volume, or, when keeping the volume constant, to lower
filter losses. However, higher switching losses are expected in case of
higher switching frequencies. Therefore, an overall system optimization,
i.e. an optimization of the complete three-phase PFC rectifier including
the filter, has been performed. Generally, it is important to consider the
system to be optimized as a whole, since there are parameters that bring
advantages for one subsystem while deteriorating another subsystem.

170
Chapter 8

Summary & Outlook

8.1 Summary
The four steps proposed towards an accurate inductance modeling are:

1. Set up a reluctance model of the inductor that is modeled.

2. Calculate core losses of the inductor.

3. Calculate winding losses of the inductor.

4. Calculate temperatures of the inductor.

The steps 1) to 4) are iteratively repeated until the algorithm has con-
verged.
Step 1 (reluctance model) has been discussed in Chapter 2, where
the main focus was put on an accurate air gap reluctance calculation,
which is essential when designing inductive components. An approach
has been proposed, which is easy-to-handle because of its modular con-
cept (different shapes of air gaps can be built from a simple structure
that is easy to calculate), and which still achieves a high level of accu-
racy as the approach is based on analytical field solutions.
Step 2 (core losses) has been discussed in Chapter 3. For the scien-
tific analysis of core losses as well as for the practical characterization of
core materials, a test setup, which allows a core excitation with a wide
variety of waveforms, has been built. With it, new core loss models
have been derived.

171
CHAPTER 8. SUMMARY & OUTLOOK

A graph that shows the dependency of the Steinmetz parameters

(α, β and k) on premagnetization, i.e. the Steinmetz Premagnetization
Graph (SPG) has been introduced. Based on the SPG, the calculation
of core losses under DC bias condition becomes possible. This new
approach how to describe losses under DC bias condition is promising
due to its simplicity.
A new loss model that considers relaxation effects has been pro-
posed. As experimentally verified, core losses are not necessarily zero
when zero voltage is applied across a transformer or inductor winding
after an interval of changing flux density. A short period after switching
the winding voltage to zero, losses still occur in the material. This is
due to magnetic relaxation. A new loss modeling approach has been
introduced and named the improved-improved Generalized Steinmetz
Equation, i2 GSE. The i2 GSE needs five new parameters to calculate
new core loss components. Hence, in total eight parameters are neces-
sary to accurately determine core losses.
Other issues have been discussed, such as the impact of the core
shape on core losses or how to handle minor and major B-H loops. A
high level of accuracy has been achieved by combining the best state-
of-the-art approaches with the newly-developed approaches.
Step 3 (winding losses) has been discussed in Chapter 4. Formulae
for round conductors, foil conductors and litz wires, each including skin-
and proximity effects (including the effect of an air gap fringing field)
have been given for the calculation of winding losses. The accuracy has
been confirmed by FEM simulations.
Step 4 (temperature) has been discussed in Chapter 5. A thermal
model is not only important to avoid overheating; it has also importance
to model the losses correctly, as they are influenced by the temperature.
Formulae to set up a thermal resistor network have been given. The
heat transfer mechanisms are conduction, convection, and radiation.
The steps 1) to 4) lead to an accurate loss and thermal modeling of
inductive power components. The concepts and models have been con-
firmed using the example of an LCL input filter of a three-phase Power
Factor Correction (PFC) rectifier in Chapter 7. A multi-objective op-
timization procedure for three-phase LCL filters based on a generic
optimization approach is introduced guaranteeing low volume and/or
low losses. The cost function, which characterizes a given filter design,
allows a weighting of the filter losses and of the filter volume accord-
ing to the designer needs. Different designs have been calculated to

172
 8.2. OUTLOOK

show the trade-off between filter volume and filter losses. Experimental
results have shown that a very high loss accuracy has been achieved.
Furthermore, an overall system optimization, i.e. an optimization of
the complete three-phase PFC rectifier including the filter, has been
performed.

8.2 Outlook
Within the thesis inductive components have been thoroughly modeled
and it has been shown how a multi-objective optimization leads to op-
timal designs of inductive components. However, there are still some
(modeling) issues that could be addressed in the future in order to fur-
ther improve the model applicability. In the following some open tasks
are listed, with no claim of completeness.
I Improvement of measurement system. Particularly, measurements
of gapped-cores, cores with low permeability, or measurements at
very high frequency are difficult. There are publications address-
ing this issue, e.g. [13, 46, 47]; however, a solution that is easy
to implement and allows any shape of waveform as an excitation
would be very interesting to have.
I In this thesis, a brief overview about tape wound cores and the
losses that additionally occur in them has been given. This topic
gives opportunities for further research, since, to the author’s
knowledge, there exists no approach which allows to analytically
describe the presented effects.
I Foil windings in (gapped) inductors. The approach introduced in
Section 4.4.5 gives, in a limited range, a reasonably good estima-
tion of the losses. However, its accuracy is much worse compared
to any other model discussed in this thesis. An improved model
would be very interesting to have.
I The thermal model presented in this thesis could further be im-
proved. Particularly, a model that determines the thermal re-
sistance of litz wires has, to the authors knowledge, not been
successfully addressed yet.
I A multi-objective optimization procedure, similar as presented in
Chapter 7, could be applied to any kind of problem. Therewith

173
CHAPTER 8. SUMMARY & OUTLOOK

the existing efficiency or loss density boundaries of existing power

electronic systems could be determined and possibly shifted to
higher levels.

174
Appendix A

Appendix

A.1 Derivation of the Basic Reluctance

The derivation of the basic reluctance (2.8) is given in this section. All
considerations are based on works presented in [9, 18, 21]. A function
of the type
z = f (t) = x(u, v) + jy(u, v) (A.1)
that defines a complex variable z = x + jy as a function of another
complex variable t = u + jv and preserves angles is termed conformal.
Conformal transformations are very helpful to derive solutions of field
problems. A transformation equation must be found relating the given
field to a simpler one to which a solution is known or easily found.
One particular conformal transformation equation that transforms
the real axis of one plane to the boundary of a polygon of another
plane was first given, independently, by Schwarz and Christoffel. The
transformation is in such a way that the upper half of the first plane
transforms into the interior of the polygon.
For the situation illustrated in Figure A.1, the transformation from
the real axis of the t-plane to the polygon boundary in the z-plane is
derived by integrating the equation
dz
= S(t − a)(α/π)−1 (t − b)(β/π)−1 (t − c)(γ/π)−1 . . . , (A.2)
dt
which is named the Schwarz-Christoffel differential equation. S is a con-
stant, a, b, c, . . . are points on the real axis in the t-plane corresponding

175
APPENDIX A. APPENDIX

y
A
α

B β

z-plane
γ
C
(a) x

t-plane

a b c
u
(b)

Figure A.1: Illustration of Schwarz-Christoffel Transformation.

to the points A, B, C, . . . in the z-plane, and α, β, γ, . . . are the interior

angles of the vertices of the polygon in the z-plane.
The Schwarz-Christoffel transformations that are used to calculate
the reluctance of the basic geometry of Figure 2.3 are given in the fol-
lowing. Two transformation equations have to be found: first a trans-
formation equation that relates the basic geometry in the z-plane (cf.
Figure A.2(a)) to the real axis in the t-plane (cf. Figure A.2(b)) is
derived. Second, a transformation equation that relates the real axis of
the t-plane to a geometry of a parallel-type capacitor (cf. Figure A.2(c))
is further derived. The geometry of a parallel-type capacitor permits
easy calculation of the capacitance that, with (2.5), directly leads to
the reluctance.
The dashed line in Figure A.2(a) illustrates the chosen polygon
structure that will be transformed to the real axis in the t-plane of
Figure A.2(b). In Table A.1 the vertex coordinates of the z-plane poly-
gon, the vertex angles, and the corresponding points on the real axis of

176
 A.1. DERIVATION OF THE BASIC RELUCTANCE

z-plane jy

jh

w/2 x
2
l 1

jv
t-plane

-1 0 1
3 1 (w/2) 2 ht u
t

v-plane jb
3

1 jV
a
(w/2)v 2 hv

Figure A.2: The Schwarz-Christoffel transformation of the basic

geometry.

the t-plane are given. This leads to the Schwarz-Christoffel differential

equation
√
dz t−1
= S1 t−1 (t − 1)1/2 = S1 , (A.3)
dt t
hence
√ √

z(t) = −jS1 2 ln (1 + 1 − t) − ln t − 2 1 − t + C1 , (A.4)

where S1 and C1 are constants that have to be further determined. For

z(1) = 0, C1 becomes C1 = 0. When z(t → 0) = ∞ is assumed for the
second boundary condition, some further steps are necessary. Assume

177
APPENDIX A. APPENDIX

1 2
zi ∞ 0
α/π 0 3/2
ti 0 1

Table A.1: Transformation Table 1.

t = ejθ and thus dt = jejθ dθ. For t → 0, (A.3) becomes

j jejθ dθ
dz = S1 dt = jS1 = −S1 dθ. (A.5)
t ejθ
Near t = 0 (point 1 in Figure A.2), z varies from x − jl to x and θ varies
from π to zero. Thus,
Z x Z 0
dz = − S1 dθ, (A.6)
x−jl π

hence, jl = S1 π or S1 = j πl . Thus, the transformation equation is fully

described with
l √ √

z(t) = 2 ln (1 + 1 − t) − ln t − 2 1 − t . (A.7)
π
The transformation equation relating the real axis of the t-plane to a
geometry of a parallel-type capacitor (cf. Figure A.2(c)) will be derived
next. The appropriate transformation table is given in Table A.2, which
leads to the Schwarz-Christoffel differential equation

dv S2
= S2 t−1 = , (A.8)
dt t
hence
v(t) = S2 ln t + C2 , (A.9)
where S2 and C2 are constants that have to be further determined.
With v(1) = 0, C2 becomes C2 = 0. With v(−1) = jV , where V
represents a constant that will cancel out later,

jV = S2 ln(−1) = S2 jπ, (A.10)

178
 A.1. DERIVATION OF THE BASIC RELUCTANCE

1
vi −∞
α/π 0
ti 0

Table A.2: Transformation Table 2.

hence, S2 = V /π. Thus, the transformation equation is

V
v(t) = ln t. (A.11)
π
For the geometry in Figure A.2(c) the capacitance per unit length
can be calculated as
hv − w2 v


0 (A.12)
C = 0 ,
V
where hv and (w/2)v have to be determined based on the above derived
transformation functions. First, (w/2)v will be derived. From (A.7) and
with t ≈ 0
l √ √

z(t) = x + j0 = 2 ln (1 + 1 − t) − ln t − 2 1 − t
π (A.13)
l
= (2 ln 2 − ln t − 2) ,
π
hence,
πx
− ln t = + 2(1 − ln 2). (A.14)
l
With (A.11) and x = w/2, (w/2)v becomes
w V V  πw 
= ln t = − + 2(1 − ln 2) . (A.15)
2 v π π 2l
Next, hv will be derived. From (A.7) and with t → ∞
l √ √

z(t) = 0 + jy = 2 ln (1 + 1 − t) − ln t − 2 1 − t
π (A.16)
l  √
= ln (−1) + 2j t ,
π
hence,
l  √
jy = j π+2 t (A.17)
π

179
APPENDIX A. APPENDIX

or πy π √
= + t. (A.18)
2l 2
For t → ∞, (A.18) can be simplified to
√ πy
t≈ (A.19)
2l
and  πy 2
t= (A.20)
2l
With (A.11) and y = h, hv becomes
V 2V πh
hv = ln t = ln . (A.21)
π π 2l
Therewith, the capacitance of the geometry in Figure A.2 is deter-
mined. (A.12) becomes
2V πh V πw


0 π ln 2l + π 2l + 2(1 − ln 2)
C = 0
  V  (A.22)
w 2 πh
= 0 + 1 + ln .
2l π 2l
(2.5) and (A.22) lead to the basic reluctance of (2.8).

A.2 iGSE and Sinusoidal Flux Waveforms

If one inserts a sinusoidal flux density waveform into the iGSE, its
equation transforms back to the Steinmetz equation (3.3). This is shown
in the following. The iGSE is given as
1 T 
Z
 dB α

Pv = ki   (∆B)β−α dt (A.23)
T 0 dt
where ∆B is the peak-to-peak flux density and
k
ki = R 2π . (A.24)
(2π)α−1 0
| cos θ|α 2β−α dθ
dB(t)
A sinusoidal flux density waveform B(t) = ∆B 2 · sin 2πf t with dt =
∆B
2πf · 2 · cos 2πf t inserted in (A.23) leads to
α
1 T 
Z 
∆B  β−α
Pv = ki 2πf ·
 · cos 2πf t
 (∆B) dt. (A.25)
T 0 2

180
 A.3. SPGS OF OTHER MATERIALS

Inserting (A.24) into (A.25) and replacing 1/T with f , one gets
RT
f k( ∆B α
2 ) (∆B)
β−α
(2πf )α 0 | cos 2πf t|α dt
Pv = R 2π . (A.26)
(2π)α−1 2β−α | cos θ|α dθ
0

With (substitution of variables)

RT
0
| cos 2πf t|α dt 1
R 2π = , (A.27)
α
| cos θ| dθ 2πf
0

(A.26) simplifies to
 β
α ∆B
Pv = kf . (A.28)
2
Equation (A.28) is the Steinmetz equation (3.3).

A.3 SPGs of Other Materials

In Figure A.3 the SPG is given for the material EPCOS N27, and in
Figure A.4 for the material Ferroxcube 3F3. In Figure A.5 the SPG for
the nanocrystalline material VITROPERM 500F from Vacuumschmelze
(VAC) is depicted. The independency of α with the frequency has been
confirmed for all materials (α = constant). All given SPGs consider
only the premagnetization range where it is still appropriate to use the
Steinmetz approach, i.e. the losses still follow a power equation.

A.4 Derivation of the Steinmetz Premag-

netization Graph
The Steinmetz parameters as a function of HDC are described with a
fourth order series expansion



   1
α α0 0 0 0 0 HDC 
 2 
 β  =  β0 pβ1 pβ2 pβ3 pβ4  · 
HDC

 (A.29)
3 
ki ki0 pki1 pki2 pki3 pki4 HDC
4
HDC

181
APPENDIX A. APPENDIX

4 1.2
3.6 ki / k ; T = 30°C 1.16
i0
3.2 β / β0; T = 30°C 1.12
2.8 1.08
2.4 1.04
ki / ki0

β / β0
2 1
1.6 β / β0; T = 100°C 0.96
1.2 0.92
0.8 ki / ki0; T = 100°C 0.88
0.4 0.84
0 0.8
0 10 20 30 40 50
HDC [A/m]

Figure A.3: SPG of the material ferrite N27 (EPCOS); measured

on R25 core.

4.4 1.12

4 ki / k ; T = 40°C 1.1
i0

3.6 1.08

3.2 1.06

2.8 β / β0; T = 40°C β / β0; T = 100°C

1.04
ki / ki0

β / β0

2.4 1.02

2 1

1.6 0.98

1.2 0.96

0.8 0.94
ki / ki0; T = 100°C
0.4 0.92
0 10 20 30 40 50
HDC [A/m]

Figure A.4: SPG of the material ferrite 3F3 (Ferroxcube); measured

on core type TN25/15/10.

or
S = P · H. (A.30)

To extract the dependency of the Steinmetz parameters on the pre-

magnetization, one has to find the right coefficients of the matrix P.

182
 A.4. DERIVATION OF THE STEINMETZ PREMAGNETIZATION
GRAPH

4 1.3
3.6 1.26
3.2 1.22
2.8 ki / k ; T = 40°C 1.18
i0

2.4 1.14
ki / ki0

β / β0
2 1.1
1.6 ki / ki0; T = 100°C 1.06
1.2 1.02
0.8 0.98
β / β0; T = 40°C
0.4 β / β0; T = 100°C 0.94
0 0.9
0 5 10 15 20 25 30 35
HDC [A/m]

Figure A.5: SPG of the material VITROPERM 500F (VAC); mea-

sured on W452 core.

This is an optimization problem. A least square algorithm has been im-

plemented that fits measured curves with calculated data by minimizing
the relative error at 3 different values of ∆B, each measured at two fre-
quencies, and 6 premagnetization values HDC (including HDC = 0).
The losses are calculated according to (3.33) with Steinmetz parame-
ters from (A.29)/(A.30). In the initial matrix P, all elements p∗ (cf.
(A.29)) are set to zero. The values that represent the Steinmetz values
under no DC bias condition (α0 , β0 , and ki0 ) have reasonable initial
values. As an optimization constraint, it is assumed that α(HDC ) > 1
and β(HDC ) > 2 for all values of HDC . The optimization is based on the
MATLAB function fminsearch() that applies the Downhill-Simplex-
Approach of Nelder and Mead [75]. This optimization procedure leads
to graphs for the Steinmetz parameter dependency as Figure 3.23, or
normalized to β0 respectively ki0 to the SPG as e.g. shown in Fig-
ure 3.24.
For the sake of completeness, a drawback of the chosen straight-
forward fitting procedure is discussed in the following. The above de-
scribed fitting procedure to calculate the SPG may, in some cases, result
in flawed SPGs that lead to partly wrong core loss calculations. This is
illustrated in Figure A.6, where an initial dip in the k/ki0 curve (cf. Fig-
ure A.6(a)) leads to an underestimation of core losses for very low values
of HDC (cf. Figure A.6(b)). This behavior is not supported by any mea-

183
APPENDIX A. APPENDIX

6 1.4
5.6 1.36
5.2 ki / k ; T = 40°C 1.32
i0
4.8 1.28
4.4 1.24
4 1.2
β / β0; T = 40°C
3.6 1.16
3.2 1.12
ki / ki0

β / β0
2.8 1.08
2.4 1.04
2 1
1.6 0.96
1.2 0.92
0.8 0.88
0.4 0.84
0 0.8
0 10 20 30 40 50
(a) HDC [A/m]

0
10
∆B/f

-1
10 150mT/20kHz (cal.)
100mT/20kHz (meas.)
P [W]

150mT/20kHz (meas.)
100mT/20kHz (cal.)
-2
10 50mT/20kHz (cal.)

50mT/20kHz (meas.)
-3
10
0 10 20 30 40 50
HDC [A/m]
(b)

Figure A.6: (a) Illustration of a flawed SPG (material ferrite N27

(EPCOS) at 40 ◦ C; measured on R25 core). The initial dip in the
curve k/ki0 is not supported by measurement data and (b) leads to a
partly wrong core loss calculation.

surement data. Such interpolation errors could e.g. be avoided/limited

by an increase of the HDC resolution. However, all published SPGs (ex-
cept the one in Figure A.6) are tested to be (almost) free from anomalies
like that.

184
 A.5. CLASSIC STEINMETZ PARAMETER K

A.5 Classic Steinmetz Parameter k

A short discussion how to extract the Steinmetz parameter value k (not
ki ) from the SPG is given in the following. According to (3.5), for k we
have
R 2π
k ki (2π)α−1 0 | cos θ|α 2β−α dθ
= R 2π (A.31)
k0 ki0 (2π)α−1 | cos θ|α 2β0 −α dθ
0

that is, under the assumption α = constant,

k ki 2β ki β


−1 β0
= = ·2 β0
, (A.32)
k0 ki0 2β0 ki0

where β/β0 can be extracted from the SPG. Of course, it is conceivable

to write k/k0 in the SPG, instead of ki /ki0 . However, because the built
test system excites the core with a triangular current shape, ki /ki0 has
been chosen for the graph. The iGSE is in any case very broadly used,
hence, to avoid further calculations, to have directly the information
about ki is often desired.

A.6 Derivation of Effective Dimensions for

Toroidal Cores
In the following, the effective dimensions for toroidal cores (introduced
in Section 3.10.1) are derived. The derivations are from [22]. An in-
finitesimal small flux part is

NI
dΦ = µ0 µr HdA = µ0 µr dA, (A.33)
l

where µ0 is the magnetic constant, µr the relative permeability, N the

amount of turns around the core, I the current through the winding,
and l the magnetic path length. With it, the flux is described by
Z
µr dA
Φ = µ0 N I . (A.34)
l

185
APPENDIX A. APPENDIX

With the second order Peterson relation1 for the relative permeability
µr , (A.34) can be written as
 Z Z 
dA dA
Φ = µ0 N I a10 + a11 N I . (A.35)
l l2
The flux in an equivalent ideal toroid (with dimensions le and Ae ) is
 
Ae Ae Ae
Φ = µ0 N Iµr = µ0 N I a10 + a11 N I 2 . (A.36)
le le le
By comparison of coefficients in (A.35) and (A.36), it follows that
Z
Ae dA
= , (A.37)
le l
and Z
Ae dA
= . (A.38)
le2 l2
With dA = hdr and l = 2πr (r is the radius and h is the axial thickness),
(A.37) and (A.38) can be solved
Z r2
Ae hdr h ln r2 /r1
= = , (A.39)
le r1 2πr 2π
and Z r2  
Ae hdr h 1 1
= = − , (A.40)
le2 r1
2
4π r 2 4π 2 r1 r2
where r1 is the inner radius and r2 the outer radius.
Hence, the effective dimensions for a toroidal core are
h ln2 r2 /r1
Ae = , (A.41)
1/r1 − 1/r2
and
2π ln r2 /r1
le = . (A.42)
1/r1 − 1/r2
1 Peterson expressed the flux density as a double power series of the instantaneous

field strengths H. Ignoring the higher powers of H, it is

B = µ0 (a10 H + a11 H2 )
or
µr = a10 + a11 H
[22].

186
 A.7. DERIVATION OF WINDING LOSSES IN FOIL CONDUCTORS

A.7 Derivation of Winding Losses in Foil

Conductors
Within this section, formulae, which allow to quantitatively calculate
losses in foil conductors will be derived. The current, voltage and fields
are assumed to be sinusoidal, hence the derivative with respect to time
d/dt simplifies to a multiplication with jω. In case of current and voltage
shapes that are non-sinusoidal, a Fourier expansion has to be performed
first2 . The presented derivation is from [58, 60].
The Maxwell equations are
ρ
div E = (A.43)

rot E = −jωB (A.44)
div B = 0 (A.45)
rot B = jωµE + µJ (A.46)
With the ohmic law J = σE, (A.46) simplifies to

rot B = (σ + jω)µE (A.47)

The substition of (A.44) in (A.47) leads to

rot rot E = ∇(div E) − ∇2 E = −(σ + jω)jωµE (A.48)

that, with (A.43), can be simplified to

∇ρ
∇2 E = + (σ + jω)jωµE. (A.49)

Substituting (A.46) in (A.44), and performing similar transforma-
tions as above, leads to

rot rot B = ∇(div B) − ∇2 B = −(σ + jω)jωµB, (A.50)

and, with div B = 0 (A.45),

∇2 B = (σ + jω)jωµB. (A.51)
2 Because of the orthogonality of the cosine/sine-functions, it is valid to perform
a Fourier expansion of the current, calculate the loss for each frequency component,
and total the losses up (cf. Appendix A.9).

187
APPENDIX A. APPENDIX

HS1
y
x Jx h
z HS2
b

Figure A.7: Cross section of a foil conductor with a current density

in x-direction. The conductor is infinitely long in x-direction.

Equations (A.49) and (A.51) differ in the term ∇ρ  , that describes the
induced charge distribution perpendicular to the current flow, due to
an external quasi static electric field. Such an electric field exists, for
instance, due to a voltage between the conductors of a winding. In case
the displacement current density is neglected3 and under the assump-
tion of no external quasi static electric field, it is

∇2 E = jωσµE, (A.52)

∇2 B = jωσµB, (A.53)
and with J = σE
∇2 J = jωσµJ . (A.54)
In the following, formulae for calculating losses, and considering the
skin- and proximity-effect, will be derived based on the above derived
equations. It is assumed that the current is flowing in x-direction, with
frequency f and magnitude I. ˆ

A.7.1 Skin Effect

The geometry considered to calculate the skin effect in foil windings is
illustrated in Figure A.7. In case of conductor materials with relative
permeabilities of unity (e.g. copper), we have B = µ0 H. Equation
(A.53) can then be written as

∇2 H = α2 H, (A.55)
3 The displacement current density is described with the term −ωµ in (A.49)

and (A.51).

188
 A.7. DERIVATION OF WINDING LOSSES IN FOIL CONDUCTORS

with
1+j
α=
δ
and
1
δ=√ .
πµ0 σf
In a conductor with width b and height h, whereas h  b, and
whereas the current flows only in x-direction, the magnetic field strength
H can be considered as independent of its z and x position. Hence, the
magnetic field strengths H can be described with only a z-component
and, consequently, (A.55) can be simplified to a one dimensional prob-
lem
d2
Hz = α2 Hz . (A.56)
dy 2
The general solution of (A.56) is

Hz = K1 eαy + K2 e−αy . (A.57)

The magnetic field strengths on the surface (boundary) of the conductor

can be calculated with Ampere’s Law (cf. Figure A.7) as

Iˆ
HS1 = −HS2 = . (A.58)
2b
Out of it, the constants K1 and K2 can be determined

Iˆ
K1 = = −K2 , (A.59)
4b sinh αh
a

and the magnetic field strength Hz becomes

Iˆ sinh αy
Hz = . (A.60)
2b sinh αh
2

With dHz /dy = Jx (cf. (A.46) and displacement current density ne-
glected), the current distribution Jx becomes

αIˆ cosh αy
Jx = . (A.61)
2b sinh αh
2

189
APPENDIX A. APPENDIX

y
x h He

z
b

Figure A.8: Cross section of a foil conductor that is influenced by

an external magnetic field in z-direction. The conductor is infinitely
long in x-direction.

With the current distribution the ohmic losses per unit length can be
calculated as
h
Iˆ2 sinh ν + sin ν
Z
b
PS = |Jx |2 dy = , (A.62)
2σ 0 4bσδ cosh ν − cos ν

where
h
ν= .
δ
Hence, the skin-effect losses (including DC losses) per unit length
can be calculated as

PS = FF (f ) · RDC · Iˆ2 (A.63)

with
1
δ=√ ,
πµ0 σf

h
ν= ,
δ
1
RDC = ,
σbh
and
ν sinh ν + sin ν
FF = .
4 cosh ν − cos ν

190
 A.7. DERIVATION OF WINDING LOSSES IN FOIL CONDUCTORS

A.7.2 Proximity Effect

The geometry considered to calculate the proximity-effect losses in foil
windings is illustrated in Figure A.8. On both conductor sides the mag-
netic field strength in z-direction has the magnitude Ĥe . This boundary
condition inserted in (A.57) leads to the field distribution

cosh αy
Hz = Ĥe (A.64)
cosh αh
2

and with dHz /dy = Jx (cf. (A.46) and displacement current density
neglected), the current distribution Jx becomes

α sinh αy
Jx = Ĥe . (A.65)
cosh αh
2

With the current distribution the ohmic losses per unit length can be
calculated as
Z h
b b sinh ν − sin ν 2
PP = |Jx |2 dy = Ĥ (A.66)
2σ 0 σδ cosh ν + cos ν e

with
h
ν= .
δ
Hence, the proximity losses per unit length can be written as

PP = RDC · GF (f ) · Ĥe2 (A.67)

with
1
δ=√ ,
πµ0 σf
h
ν= ,
δ
1
RDC = ,
σbh
and
sinh ν − sin ν
GF = b2 ν .
cosh ν + cos ν

191
APPENDIX A. APPENDIX

y
Jz
x
z d

Figure A.9: Cross section of the considered round conductor with

a current density in z-direction. The conductor is infinitely long in
z-direction.

A.8 Derivation of Winding Losses in Round

Conductors
In this section, formulae to quantitatively calculate losses in round con-
ductors will be derived. The considered round conductor is illustrated
in Figure A.9. The round conductor has a diameter d, and the length
l, whereas it is assumed that d  l, thus the round conductor is con-
sidered as infinitely long in z-direction. All following calculations are
performed in cylindrical coordinates. Furthermore, the magnetic field
has only a ϕ-component (cylindrical coordinates) and the current only
an axial z-component. The presented derivation is from [58, 60].

A.8.1 Skin Effect

∂
Under the assumption that the problem is cylinder symmetric ( ∂ϕ B=
0) and the displacement current density is neglected, the Maxwell equa-
tion (A.46) can be written as (in cylinder coordinates)
∂Hϕ Hϕ
Jz = + (A.68)
∂r r
The Maxwell equation (A.44) together with the ohmic law (J = σE)
leads to
∂Jz
= jωσµHϕ . (A.69)
∂r
(A.68) and (A.69) lead to the differential equation
d2 1 d
Jz + Jz = jωσµJz (A.70)
dr2 r dr

192
A.8. DERIVATION OF WINDING LOSSES IN ROUND CONDUCTORS

for the current density inside the conductor. This is a Bessel differential
equation 4 . The solution of (A.70) is
3√
Jz = CJ0 (j 2 ωσµr), (A.71)

where Jv (kx) is known as Bessel function of the first kind and order
v. The integration constant C is calculated by integrating the current
density
Z d/2
3√
ZZ
I= Jz dA = 2πC rJ0 (j 2 ωσµr)dr (A.72)
AL 0

xv Jv−1 (x)dx =
R
where AL is the cross section area of the conductor. With
xv Jv (x) + C [59], (A.72) simplifies to
2π d 3√ d
I= 3 √ C J1 (j 2 ωσµ ). (A.73)
j2 ωσµ 2 2

The integration constant C is eliminated by substituting (A.73) into

(A.71), hence
3√ 3√
j 2 ωσµ J0 (j 2 ωσµr)
Jz = I 3√ , (A.74)
2π d2 J1 (j 2 ωσµ d2 )
where Jz ∈ C. On the surface, i.e. r = d2 , the voltage drop arises from
the resistance R and reactance ωL. The voltage drop per unit length is
3√ 3√
j 2 ωσµ J0 (j 2 ωσµ d2 )
(R + jωL)I = I 3 √ . (A.75)
2π d2 σ J1 (j 2 ωσµ d2 )

To resolve the right-hand side of (A.75) into its real- and imaginary
part, the Kelvin functions can be used
3
Jv (j 2 x) = berv x + j beiv x. (A.76)

After some mathematical conversions, the skin-effect losses (includ-

ing DC losses) per unit length can be calculated as

PS = RDC · FR (f ) · Iˆ2 (A.77)

4 The Bessel differential equation has the form x2 y 00 + xy 0 + (k 2 x2 − v 2 )y = 0.

With the general solution y = C1 Jv (kx) + C2 Yv (kx), whereas Jv (kx) is known

as Bessel function of the first kind and order v and Yv (kx) is known as Bessel
function of the second kind and order v [59]. Equation (A.70) can be transformed
to r2 Jz00 + rJz0 + j3 ωσµr2 Jz = 0, which is a Bessel differential equation.

193
APPENDIX A. APPENDIX

y Jz He
x
z d
Figure A.10: Cross section of a round conductor that is influenced
by an external magnetic field in x-direction. The conductor is in-
finitely long in z-direction.

with
1
δ=√ ,
πµ0 σf

d
ξ=√ ,
2δ
4
RDC = ,
σπd2
and 
ξ ber0 (ξ)bei1 (ξ) − ber0 (ξ)ber1 (ξ)
FR = √
4 2 ber1 (ξ)2 + bei1 (ξ)2
 (A.78)
bei0 (ξ)ber1 (ξ) + bei0 (ξ)bei1 (ξ)
− .
ber1 (ξ)2 + bei1 (ξ)2

A.8.2 Proximity Effect

A round conductor with diameter d that is positioned parallel to the z-
axis is influenced by an alternating magnetic field H(r, ϕ) = Hr (r, ϕ)er +
Hϕ (r, ϕ)eϕ with magnitude Ĥe . The considered situation is illustrated
in Figure A.10. The corresponding vector potential has the form H(r, ϕ) =
Az (r, ϕ)ez , where the vector potential is defined by

H = rot A. (A.79)

194
A.8. DERIVATION OF WINDING LOSSES IN ROUND CONDUCTORS

When the displacement current density in the Maxwell equation (A.46)

is neglected, it is
J = rotH = rot rot A, (A.80)
hence
1 ∂Az ∂ 2 Az 1 ∂ 2 Az
Jz = − − 2
− 2 . (A.81)
r ∂r ∂r r ∂ϕ2
With (A.44) (rot E = −jωB) and the ohmic law (J = σE), it is

J = −jσµωA, (A.82)

and the partial differential equation for the vector potential becomes

1 ∂Az ∂ 2 Az 1 ∂ 2 Az
jσµωAz = + 2
+ 2 . (A.83)
r ∂r ∂r r ∂ϕ2

How to solve this differential equation is presented in [60]. The solution

for the magnetic vector potential Az is
3
2µ20 Ĥe J1 (j 2 ξr)
Az = 3 3 sin ϕ, (A.84)
j 2 ξ J0 (j 2 ξ d2 )

with
d
ξ=√ .
2δ
The relative permeability µr of the conductor is considered to be one,
which is valid for e.g. copper, aluminium, etc. Also the space around
the conductor has the relative permeability of one, as e.g. valid for air.
With (A.82) and (A.84), the current density inside the conductor can
be calculated 3 3
2µ20 Ĥe j 2 ξJ1 (j 2 ξr)
Jz = 3 sin ϕ. (A.85)
J0 (j 2 ξ d2 )
Out of it, the resulting proximity losses per unit length can be cal-
culated
Z 2πZ d2
1
PP = |Jz |2 rdrdϕ = RDC · GR (f ) · Ĥe2 (A.86)
2σ 0 0

with
1
δ=√ ,
πµ0 σf

195
APPENDIX A. APPENDIX

d
ξ=√ ,
2δ
4
RDC = ,
σπd2
and
ξπ 2 d2 ber2 (ξ)ber1 (ξ) + ber2 (ξ)bei1 (ξ)

GR = − √
2 2 ber0 (ξ)2 + bei0 (ξ)2
 (A.87)
bei2 (ξ)bei1 (ξ) − bei2 (ξ)ber1 (ξ)
+ .
ber0 (ξ)2 + bei0 (ξ)2

A.9 Orthogonality of Winding Losses

In the derivations in Chapter 4, a current with sinusoidal waveform
has been assumed. However, often a non-sinusoidal current is flowing
through the conductor. This current can be expressed as a Fourier
series with complex fourier coefficients I v

I(t) = I 0 + I 1 cos ωt + I 2 cos 2ωt + I 3 cos 3ωt + . . . (A.88)

and accordingly, the corresponding current density is

J(x, y, t) = J 0 + J 1 cos ωt + J 2 cos 2ωt + J 3 cos 3ωt + . . . (A.89)

The losses per-unit-length are

Z Z T
1
P = |J(x, y, t)|2 dtdA (A.90)
Tσ A 0
with A the conductor cross section, and T the current period. Because
of the orthogonality of the cosine-function,
R 2πthe product of two Fourier
parts with different frequencies is zero ( 0 cos kx · cos lx dx = 0 for
l 6= k), hence
∞ Z
1 X
P = J i J ∗i dA. (A.91)
2σ i=0 A
The part J i can be split into the current density due to skin- J Si and
due to proximity-effect J Pi
∞ Z
1 X
P = (J Si + J Pi )(J ∗Si + J ∗Pi )dA. (A.92)
2σ i=0 A

196
 A.9. ORTHOGONALITY OF WINDING LOSSES

When a conductor has an axis of symmetry and the applied field is

uniform and parallel to the symmetry axis (as in round-, or foil con-
ductors), the current density due to the skin effect JS has an even
symmetry, while the current density due toR the proximity effect JP has
an odd symmetry [63]. Consequently, it is A JP · JS dA = 0 and (A.92)
simplifies to
∞ Z
1 X
P = (J Si J ∗Si + J Pi J ∗Pi )dA (A.93)
2σ i=0 A
∞
X
= (PSi + PPi ). (A.94)
i=0

Above it is shown that the assumption to directly sum the skin- and
proximity losses is valid. Furthermore, because of the orthogonality of
the cosine-function, it is valid to perform a Fourier expansion of the
current, calculate the loss for each frequency component, and total the
losses up.

197
Bibliography

[1] Steven Milunovich and Jose Rasco. The sixth revolution: The
coming of cleantech. Merrill Lynch Report, Nov. 2008.
[2] B. K. Bose. Energy, environment, and advances in power elec-
tronics. IEEE Transactions on Power Electronics, 15(4):688–701,
2000.
[3] G. Andersson. Power electronics solutions for sustainability. In
Proc. of the IEEE Power Engineering Society General Meeting,
pages 2304–2308, 2004.
[4] J. W. Kolar, J. Biela, S. Waffler, T. Friedli, and U. Badstuebner.
Performance trends and limitations of power electronic systems.
In Proc. of the 6th International Conference on Integrated Power
Electronics Systems (CIPS), pages 1–20, 2010.
[5] Alex van den Bossche and Vencislav Cekov Valchev. Inductors and
Transformers for Power Electronics. CRC Press, Taylor & Francis
Group, 2005.
[6] I. Villar, U. Viscarret, I. Etxeberria-Otadui, and A. Rufer. Global
loss evaluation methods for nonsinusoidally fed medium-frequency
power transformers. IEEE Transactions on Industrial Electronics,
56(10):4132–4140, Oct. 2009.
[7] R. Petkov. Optimum design of a high-power, high-frequency trans-
former. IEEE Transactions on Power Electronics, 11(1):33 –42,
Jan. 1996.
[8] Thomas Komma. Allgemein gültiger Entwurfsalgorithmus für ma-
gnetische Komponenten in Schaltnetzteilen mit unterschiedlichen

199
BIBLIOGRAPHY

Topologien und Schaltfrequenzen bis 2 MHz. PhD thesis, Techni-

sche Universität Dresden, 2005.
[9] Peter Wallmeier. Automatisierte Optimierung von induktiven
Bauelementen für Stromrichteranwendungen. PhD thesis, Univer-
sität - Gesamthochschule Paderborn, 2001.
[10] Ansgar Brockmeyer. Dimensionierungswerkzeug für magnetische
Bauelemente in Stromrichteranwendungen. PhD thesis, RWTH
Aachen, 1997.
[11] J. Reinert, A. Brockmeyer, and R. De Doncker. Calculation of
losses in ferro- and ferrimagnetic materials based on the modified
Steinmetz equation. IEEE Transactions on Industry Applications,
37(4):1055–1061, 2001.
[12] C. A. Baguley, B. Carsten, and U. K. Madawala. The effect of
DC bias conditions on ferrite core losses. IEEE Transactions on
Magnetics, 44(2):246–252, Feb. 2008.
[13] C. A. Baguley, U. K. Madawala, and B. Carsten. A new technique
for measuring ferrite core loss under DC bias conditions. IEEE
Transactions on Magnetics, 44(11):4127–4130, 2008.
[14] G. Niedermeier and M. Esguerra. Measurement of power losses
with DC bias - The Displacement Factor. In Proc. of PCIM, pages
169–174, 2000.
[15] A. Brockmeyer. Experimental evaluation of the influence of DC-
premagnetization on the properties of power electronic ferrites. In
Proc. of the 11th Annual Applied Power Electronics Conference
(APEC), volume 1, pages 454–460, 3–7 March 1996.
[16] M. S. Lancarotte, C. Goldemberg, and Ad. A. Penteado. Esti-
mation of FeSi core losses under PWM or DC bias ripple voltage
excitations. IEEE Transactions on Energy Conversion, 20(2):367–
372, 2005.
[17] Manfred Albach. Grundlagen der Elektrotechnik 1. Pearson
Studium, 2008.
[18] K. J. Binns, P. J. Lawrenson, and C. W. Trowbridge. The Analyt-
ical and Numerical Solution of Electric and Magnetic Fields. John
Wiley & Sons, Inc., 1992.

200
 BIBLIOGRAPHY

[19] Miles Walker. The Schwarz-Christoffel transformation and its ap-

plications - A simple exposition. Dover Publications, Inc., 1964.

[20] J. J. Thomson. Notes on recent researches in electricity and mag-

netism. Oxford, 1893.

[21] A. Balakrishnan, W. T. Joines, and T. G. Wilson. Air-gap re-

luctance and inductance calculations for magnetic circuits using a
Schwarz-Christoffel transformation. IEEE Transactions on Power
Electronics, 12(4):654–663, July 1997.

[22] E. C. Snelling. Soft Ferrites, Properties and Applications. 2nd

edition. Butterworths, 1988.

[23] Marian K. Kazimierczuk. High-Frequency Magnetic Components.

John Wiley & Sons, Inc., 2009.

[24] Ned Mohan, Tore M. Undeland, and William P. Robbins. Power

Electronics - Converters, Applications, and Design. John Wiley &
Sons, Inc., 2003.

[25] A. F. Hoke and C. R. Sullivan. An improved two-dimensional

numerical modeling method for E-core transformers. In Proc. of
the Applied Power Electronics Conference and Exposition (APEC),
pages 151 –157, 2002.

[26] Herbert C. Roters. Electromagnetic Devices. John Wiley & Sons,

Inc., 1944.

[27] A. van den Bossche, V. Valchev, and T. Filchev. Improved approx-

imation for fringing permeances in gapped inductors. In Proc. of
the 37th IAS Annual Meeting, volume 2, pages 932–938, 2002.

[28] C. Marxgut, J. Mühlethaler, F. Krismer, and J. W. Kolar. Multi-

objective optimization of ultra-flat magnetic components with a
PCB-integrated core. In Proc. of the 8th International Conference
on Power Electronics - ECCE Asia, pages 460–467, 2011.

[29] Ferrites and Accessories, Edition 2007. EPCOS AG.

[30] J. B. Goodenough. Summary of losses in magnetic materials. IEEE

Transaction on Magnetics, 38(5):3398–3408, Sept. 2002.

201
BIBLIOGRAPHY

[31] B. D. Cullity and C. D. Graham. Introduction to Magnetic Mate-

rials. John Wiley & Sons, Inc., 2009.
[32] Soshin Chikazumi. Physics of Ferromagnetism. Oxford University
Press, 1997.
[33] Giorgio Bertotti. Hysteresis in Magnetism. Academic Press, Inc.,
1998.
[34] J. Li, T. Abdallah, and C. R. Sullivan. Improved calculation of
core loss with nonsinusoidal waveforms. In Proc. of the 36th IEEE
IAS Annual Meeting, volume 4, pages 2203–2210, 2001.
[35] K. Venkatachalam, C. R. Sullivan, T. Abdallah, and H. Tacca. Ac-
curate prediction of ferrite core loss with nonsinusoidal waveforms
using only Steinmetz parameters. In Proc. of IEEE Workshop on
Computers in Power Electronics, pages 36–41, 2002.
[36] J. Biela, U. Badstuebner, and J. W. Kolar. Impact of power density
maximization on efficiency of DC-DC converter systems. IEEE
Transactions on Power Electronics, 24(1):288–300, Jan. 2009.
[37] S. Iyasu, T. Shimizu, and K. Ishii. A novel iron loss calculation
method on power converters based on dynamic minor loop. In Proc.
of European Conference on Power Electronics and Applications,
pages 2016–2022, 2005.
[38] T. Shimizu and K. Ishii. An iron loss calculating method for AC
filter inductors used on PWM inverters. In Proc. of 37th IEEE
Power Electronics Specialists Conference (PESC), pages 1–7, 2006.
[39] K. Terashima, K. Wada, T. Shimizu, T. Nakazawa, K. Ishii, and
Y. Hayashi. Evaluation of the iron loss of an inductor based on
dynamic minor characteristics. In Proc. of European Conference
on Power Electronics and Applications, pages 1–8, 2007.
[40] W. A. Roshen. A practical, accurate and very general core loss
model for nonsinusoidal waveforms. IEEE Transactions on Power
Electronics, 22(1):30–40, 2007.
[41] C. R. Sullivan, J. H. Harris, and E. Herbert. Core loss predictions
for general PWM waveforms from a simplified set of measured data.
In Proc. of Applied Power Electronics Conference and Exposition
(APEC), pages 1048–1055, 2010.

202
 BIBLIOGRAPHY

[42] B. Carsten. Why the magnetics designer should measure core loss;
with a survey of loss measurement techniques and a low cost, high
accuracy alternative. In Proc. of PCIM, pages 163–179, 1995.
[43] F. Dong Tan, J. L. Vollin, and S. M. Cuk. A practical approach for
magnetic core-loss characterization. IEEE Transactions on Power
Electronics, 10(2):124–130, Mar. 1995.
[44] W. Shen, F. Wang, D. Boroyevich, and C. W. Tipton. Loss
characterization and calculation of nanocrystalline cores for high-
frequency magnetics applications. IEEE Transactions on Power
Electronics, 23(1):475–484, Jan. 2008.
[45] V. J. Thottuvelil, T. G. Wilson, and H. A. Owen. High-frequency
measurement techniques for magnetic cores. IEEE Transactions
on Power Electronics, 5(1):41–53, Jan. 1990.
[46] M. Mu, Q. Li, D. Gilham, F. C. Lee, and K. D. T. Ngo. New
core loss measurement method for high frequency magnetic mate-
rials. In Proc. of IEEE Energy Conversion Congress and Exposition
(ECCE), pages 4384–4389, 2010.
[47] Yehui Han, G. Cheung, An Li, C. R. Sullivan, and D. J. Perreault.
Evaluation of magnetic materials for very high frequency power
applications. IEEE Transactions on Power Electronics, 27(1):425–
435, Jan. 2012.
[48] F. Krismer and J. W. Kolar. Accurate power loss model derivation
of a high-current dual active bridge converter for an automotive ap-
plication. IEEE Transactions on Industrial Electronics, 57(3):881
–891, 2010.
[49] M. Hartmann, J. Miniboeck, and J. W. Kolar. A three-phase delta
switch rectifier for more electric aircraft applications employing a
novel PWM current control concept. In Proc. of the Applied Power
Electronics Conference and Exposition (APEC), pages 1633 –1640,
Feb. 2009.
[50] Alexander Stadler. Messtechnische Bestimmung und Simulation
der Kernverluste in weichmagnetischen Materialien. PhD thesis,
Universität Erlangen - Nürnberg, 2009.
[51] Ferroxcube data sheet to material 3E6, Sep. 2008.

203
BIBLIOGRAPHY

[52] F. G. Brockman, P. H. Dowling, and W. G. Steneck. Dimensional

effects resulting from a high dielectric constant found in a ferro-
magnetic ferrite. Physical Review, 77(1):85–93, January 1950.
[53] H. Kakehashi and Y. Ohta. A study of amorphous alloy cutting
core. IEEE Translation Journal on Magnetics in Japan, 1(5):608–
609, 1985.
[54] W. Shen, F. Wang, D. Boroyevich, and C. W. Tipton. High-density
nanocrystalline core transformer for high-power high-frequency
resonant converter. IEEE Transactions on Industry Applications,
44(1):213–222, 2008.
[55] H. Fukunaga, T. Eguchi, Y. Ohta, and H. Kakehashi. Core loss in
amorphous cut cores with air gaps. IEEE Transactions on Mag-
netics, 25(3):2694–2698, 1989.
[56] H. Fukunaga, T. Eguchi, K. Koga, Y. Ohta, and H. Kakehashi.
High performance cut cores prepared from crystallized Fe-based
amorphous ribbon. IEEE Transactions on Magnetics, 26(5):2008–
2010, 1990.
[57] B. Cougo, A. Tüysüz, J. Mühlethaler, and J. W. Kolar. Increase
of tape wound core losses due to interlamination short circuits
and orthogonal flux components. In Proc. of. the 37th Annual
Conference of the IEEE Industrial Electronics Society (IECON),
2011.
[58] Jürgen Biela. Optimierung des elektromagnetisch integrierten
Serien-Parallel-Resonanzkonverters mit eingeprägtem Aus-
gangsstrom. PhD thesis, ETH Zurich, 2005.
[59] Larry C. Andrews. Elementary Partial Differential Equations with
Boundary Value Problems. Academic Press College Division, 1986.
[60] Jiri Lammeraner and Milos Stafl. Eddy Currents. Iliffe Books Ltd.,
1966.
[61] P. L. Dowell. Effects of eddy currents in transformer windings.
Proceedings of the Institution of Electrical Engineers, 113(8):1387–
1394, 1966.
[62] J. A. Ferreira. Electromagnetic Modelling of Power Electronic Con-
verters. Kluwer Academics Publishers, 1989.

204
 BIBLIOGRAPHY

[63] J. A. Ferreira. Improved analytical modeling of conductive losses in

magnetic components. IEEE Transactions on Power Electronics,
9(1):127–131, Jan. 1994.
[64] C. R. Sullivan. Optimal choice for number of strands in a litz-wire
transformer winding. IEEE Transactions on Power Electronics,
14(2):283 –291, March 1999.
[65] H. Rossmanith, M. Doebroenti, M. Albach, and D. Exner. Mea-
surement and characterization of high frequency losses in nonideal
litz wires. IEEE Transactions on Power Electronics, 26(11):3386–
3394, 2011.
[66] P. Wallmeier. Improved analytical modeling of conductive losses in
gapped high-frequency inductors. IEEE Transactions on Industry
Applications, 37(4):1045–1054, 2001.
[67] VDI-Wärmeatlas. Springer-Verlag Berlin Heidelberg, 2006.
[68] M. Liserre, F. Blaabjerg, and S. Hansen. Design and control of an
LCL-filter-based three-phase active rectifier. IEEE Transactions
on Industry Applications, 41:1281–1291, 2005.
[69] S. V. Araujo, A. Engler, B. Sahan, and F. Antunes. LCL filter
design for grid-connected NPC inverters in offshore wind turbines.
In Proc. of the 7th Internatonal Conf. Power Electronics (ICPE),
pages 1133–1138, 2007.
[70] Thomas Friedli. Comparative Evaluation of Three-Phase Si and
SiC AC-AC Converter Systems. PhD thesis, ETH Zurich, 2010.
[71] Film Capacitors, Edition 2009. EPCOS AG.
[72] R. D. Middlebrook. Design techniques for preventing input filter
oscillations in switched-mode regulators. In Proc. of Powercon 5,
pages A3.1–A3.16, 1978.
[73] Robert W. Erickson and Dragan Maksimovic. Fundamentals of
Power Electronics. Springer Science+Business Media, LLC, 2004.
[74] J. W. Kolar and T. Friedli. The essence of three-phase PFC rectifier
systems. In Proc. of. the 33rd IEEE International Telecommuni-
cation Energy Conference (INTELEC), Amsterdam, Netherlands,
2011.

205
BIBLIOGRAPHY

[75] J. A. Nelder and R. Mead. A simplex method for function mini-

mization. The Computer Journal, 4:308–313, 1965.

[76] M. Schweizer and J. W. Kolar. High efficiency drive system with

3-level T-type inverter. In Proceedings of the 14th European Con-
ference on Power Electronics and Applications - ECCE Europe,
Birmingham, UK, 2011.
[77] T. Friedli and J. W. Kolar. A semiconductor area based assessment
of ac motor drive converter topologies. In Proc. of the 24th Annual
IEEE Applied Power Electronics Conf. and Exposition (APEC),
pages 336–342, 2009.
[78] M. Schweizer, I. Lizama, T. Friedli, and J. W. Kolar. Comparison
of the chip area usage of 2-level and 3-level voltage source converter
topologies. In Proc. of the 36th Annual Conf. on IEEE Industrial
Electronics Society (IECON), pages 391–396, 2010.
[79] U. Drofenik, G. Laimer, and J. W. Kolar. Theoretical converter
power density limits for forced convection cooling. In Proc. of
the International PCIM Europe Conference, Nuremberg, Germany,
2005.

206
Curriculum Vitae

Personal Data

Name Jonas Mühlethaler

Date of Birth 16.09.1982
Citizen of Lucerne, Switzerland

Education

2008 - 2012 Ph.D. Studies at the Power Electronic Systems

Laboratory, ETH Zurich
2003 - 2008 M.Sc. Studies in Information Technology and
Electrical Engineering, ETH Zurich
2002 - 2003 Preparation for the Entrance Exam ETH Zurich,
AKAD
1998 - 2002 Apprenticeship as an Electronic Technician
(with Berufsmatura), ABB Lernzentren, Baden
1989 - 1998 Primary School & Bezirksschule, Mellingen

Work Experience

2012 - . . . Researcher (Post-Doc) at the Power Electronic Systems

Laboratory, ETH Zurich
2008 - 2012 Research Assistant at the Power Electronic Systems
Laboratory, ETH Zurich
2002 - 2003 Electronic Technician, Varian Medical Systems,
Baden

207