System Level Black-Box Models for DC-DC Converters

Luis Arnedo

Dissertation submitted to the faculty of the Virginia Polytechnic Institute and State
University in partial fulfillment of the requirements for the degree of

Doctor of Philosophy
In
Electrical Engineering

Dr. Dushan Boroyevich, Chairman
Dr. Rolando Burgos
Dr. Fred Wang
Dr. Robert D. Lorenz
Dr. Virgilio Centeno
Dr. John Burns


September 05, 2008
Blacksburg, Virginia

Keywords: Black-box, dc-dc, system level, interaction, two-port, system identification,
polytopic, frequency response function

System Level Black-Box Models for DC-DC Converters

Luis Arnedo

ABSTRACT

The aim of this work is to develop a two-port black-box dc-dc converter modeling
methodology for system level simulation and analysis. The models do not require any
information about the components, structure, or control parameters of the converter.
Instead, all the information needed to build the models is collected from unterminated
experimental frequency response function (FRF) measurements performed at the
converter power terminals. These transfer funtions are known as audiosuceptibility, back
current gain, output impedance, and input admittance. The measurements are called
unterminated because they do not contain any information about the source and/or the
load dynamics. This work provides insights into how the source and the load affect FRF
measurements and how to decouple those effects from the measurements. The actual
linear time invariant model is obtained from the experimental FRFs via system
identification.
Because the the two-port model obtained from a set of FRFs is linear, it will be valid in a
specific operating region defined by the converter operating conditions. Therefore, to
satisfy the need for models valid in a wide operating region, a model structure that
combines a family of linear two-port models is proposed. One structure, known as the
Wiener structure, is especially useful when the converter nonlinearities are reflected
mainly in the steady state currents and voltage values. The other structure is known as a
polytopic structure, and it is able to capture nonlinearities that affect the transient and
steady state converter behavior.
iii

The models are used for prediction of steady state and transient behavior of voltages and
currents at the converter terminals. In addition, the models are useful for subsystem
interaction and small signal stability assesment of interconnected dc distribution systems
comprising commericially available converters. This work presents for first time
simulation and stability analysis results of a system that combines dc-dc converters from
two different manufucturers. All simulation results are compared against experimental
results to verify the usefulness of the approach.

iv

To Yovana and Alexandra

ACKNOWLEDGMENTS

I want to express my gratitude to Dr. Dushan Borojevich for giving me the
opportunity to initiate my PhD studies under his supervision, for his guidance during the
course of this investigation, and for all the things I learned from him in the classroom, the
project meetings, and one-on-one meetings. I am also thankful to Dr. Rolando Burgos for
all the time he spent with me giving me week-to-week guidance, and for his suggestions
about the organization of our technical papers and this document. With his help, I was
able to complete this work much faster than if I were working alone.
I want to express my admiration for Dr. Robert Lorenz from the University of
Wisconsin-Madison. From him, I learned a series of skills on frequency response
functions measurements, parameter identification, and controls that were very useful in
this work. I also thank Dr. Fred Wang, Dr. Virgilio Centeno, and Dr. John Burns for
their input during the preliminary examination. Their questions and vision gave me clear
direction to continue and finalize this work.
I thank CPES staff members Marianne Hawthorne, Teresa Shaw, Beth Tranter,
Linda Long, and Trish Rose. They were always helpful and caring. And I dont have the
words to express my admiration and gratefulness to all CPES students from all five
campuses. I feel really proud to have had the opportunity to share this experience with
them.
My wife, my daughter, and I want to express our gratitude to our friends in
Blacksburg. We were fortunate to share with them very important things in our lives that
we will never forget, and we received from them the warm family feeling we missed
from our home country.
To my father, my mother, brothers and sisters, I am always thankful for the love,
education, and advice I receive from all of you.

vi

Table of Content

1. Chapter 1 Introduction ................................................................................................................ 1
1.1 Motivation and Objectives ...................................................................................................... 1
1.2 Literature Review .................................................................................................................... 4
1.2.1 Component Level Models ............................................................................................. 5
1.2.2 Converter Level Models ................................................................................................ 6
1.2.3 System Level Model .................................................................................................... 11
1.3 Dissertation Outline and Summary of contributions ............................................................. 12
1.4 References ............................................................................................................................. 14
2 Chapter 2 Two-port Black-Box Terminal Characterization ..................................................... 20
2.1 Introduction ........................................................................................................................... 20
2.2 Black-box Two-Port Terminal Characterization ................................................................... 22
2.3 Frequency Response Measurements ..................................................................................... 23
2.4 System Identification ............................................................................................................. 29
2.5 Black-box Two-port Model of an Unregulated Buck Converter ........................................... 34
2.6 Model Order Reduction by Means of Balanced Realization and Singular Perturbations ..... 40
2.7 Conclusions ........................................................................................................................... 45
2.8 References ............................................................................................................................. 46
3 Chapter 3 Subsystem Interaction Effects on FRF Measurements ............................................ 47
3.1 Introduction ........................................................................................................................... 47
3.2 Subsystem Interaction Affecting Measurements of Frequency Response Functions ............ 49
3.3 Conclusions ........................................................................................................................... 60
3.4 References ............................................................................................................................. 61
4 Chapter 4 Nonlinear Black-box DC-DC Converter Model ...................................................... 62
4.1 Introduction ............................................................................................................................ 62
4.2 Hammerstein-Wiener Black-box model Structure ................................................................. 64
4.3 Regulated Buck Converter Black-Box model using a Wiener structure ................................ 65
4.4 Polytopic Models .................................................................................................................. 69
4.5 Polytopic Black-Box Model for an Unregulated Bus Converter ........................................... 75
4.6 Polytopic Black-box Model of an Unregulated Flyback Converter ...................................... 77
4.7 Polytopic Black-box Model of a Regulated Flyback Converter ........................................... 81

vii

4.8 Polytopic Black-box Model Analysis .................................................................................... 86

4.9 Conclusions ............................................................................................................................ 93
4.10 References ............................................................................................................................ 94
5 Chapter 5 Analysis and Simulation of Distributed Power Electronic Systems ........................ 96
5.1 Introduction ........................................................................................................................... 96
5.2 Black-box Models of Commercial dc-dc Converters ............................................................ 98
5.3 Modeling of Cascaded Converters ...................................................................................... 103
5.4 Modeling of Parallel Converters ......................................................................................... 105
5.5 Modeling of a Distributed Power System (DPS) ................................................................ 107
5.6 Subsystem Interaction ......................................................................................................... 110
5.7 Conclusions ......................................................................................................................... 124
5.8 References ........................................................................................................................... 125
6 Chapter 6 Conclusions and Future Work ............................................................................... 127
6.1 Conclusions ......................................................................................................................... 127
6.2 Future Work ........................................................................................................................ 129
A. Appendix Linear model from FRF measurments .................................................................. 136
B. Appendix Graph comparison of experimental and simulation results .................................... 138
C. Appendix Obtaining un-terminated FRFs .............................................................................. 140
D. Appendix Flyback polytopic model ....................................................................................... 147

viii

List of Figures

Fig. 1.1 Discrete and modular approach for power electronics converters ........................ 2
Fig. 1.2 Two-port black-box model .................................................................................... 4
Fig. 1.3 Classification of power electronics models ........................................................... 4
Fig. 1.4 Hammerstein model ............................................................................................. 10
Fig. 2.1 Black-box model .................................................................................................. 22
Fig.2.2 Black-box model structure .................................................................................... 23
Fig. 2.3 Series and parallel injection circuits .................................................................... 25
Fig. 2.4 (a) Measurement setup for G
O
and Y
i
.................................................................. 26
Fig. 2.4 (b) Measurement setup for Z
O
and H
i
.................................................................. 26
Fig. 2.5 FRF H
i
from measurement and average model .................................................. 27
Fig. 2.6 FRF Z
o
from measurement and average model .................................................. 28
Fig. 2.7 General linear model structure ............................................................................ 30
Fig. 2.8 AR model structure .............................................................................................. 30
Fig. 2.9 ARX model structure ........................................................................................... 30
Fig. 2.10 ARMAX model structure .................................................................................. 31
Fig. 2.11 Box-Jenkins model structure ............................................................................. 31
Fig. 2.12 Output error model structure ............................................................................. 32
Fig. 2.13 Comparison system identification algorithm ..................................................... 33
Fig. 2.14 Buck converter prototype 25 W 20V to 5V ....................................................... 35
Fig. 2.15 Experimental FRFs ........................................................................................... 36
Fig. 2.16 Open-loop FRFs measured at different output current values ........................... 37
Fig. 2.17 Open loop FRFs from measurements and identified models ........................... 38
Fig. 2.18 Open-loop FRFs from measurements and identified models ............................ 39
Fig. 2.19 Flow chart for black-box modeling ................................................................... 40
Fig. 2.20 Full order and low order FRF comparison ........................................................ 44
Fig. 2.21 Prediction error between full and low order models ......................................... 45
Fig. 3.1 Converter black-box two signal diagram loaded with a current sink .................. 49
Fig. 3.2 Source output impedance impact on H
im
(from average models) ...................... 50
Fig. 3.3 Source output impedance effect H
im
(experimental) ........................................... 51
Fig. 3.4 Source output impedance effect on FRF measurements ..................................... 51
Fig. 3.5 Experimental and simulated transient response using measured FRFs ............... 52
Fig. 3.6 Experimental and simulated transient response using unterminated FRFs ......... 52
Fig. 3.7 Converter black-box signal diagram loaded with a resistor ................................ 53
Fig. 3.8 Removing Source and load dynamics from measured FRFs .............................. 55
Fig. 3.9 Experimental setup for model validation using resistor load .............................. 55
Fig. 3.10 Experimental and simulated transient response during a load step using resistors
........................................................................................................................................... 56

ix

Fig. 3.11 Source and load dynamics ................................................................................. 57

Fig. 3.12 Source and load effect on FRF measurements ................................................. 58
Fig. 3.13 FRFs measured using a resistor as a load and unterminated FRFs .................. 59
Fig. 3.14 Experimental and simulated transient response during three load steps ........... 60
Fig. 4.1. Hammerstein-Wiener model structure. ............................................................... 64
Fig. 4.2. Closed-loop FRFs from measurements and identified models. ......................... 66
Fig. 4.3. Effects of changes of the operating point on H
i
. ............................................... 67
Fig. 4.4. Nonlinear black-box model based on a Wiener structure. ................................. 67
Fig. 4.5. Open loop FRFs from measurements and identified models. ........................... 68
Fig. 4.6. Open-loop FRFs from measurements and identified models. ........................... 69
Fig. 4.7. IBC FRFs measured at different output current values. .................................... 70
Fig. 4.8. IBC linear black-box model prediction between 50% and 100% of the nominal
load. ................................................................................................................................... 71
Fig. 4.9. IBC linear black-box model prediction between 5% and 50% of the nominal
load. ................................................................................................................................... 71
Fig. 4.10. Weighting functions. ....................................................................................... 73
Fig. 4.11. Model interconnection using sigmoid functions. ............................................ 73
Fig. 4.12. General one-dimensional polytopic model. ..................................................... 74
Fig. 4.13. Two-dimensional sigmoid scheduling functions. ................................................ 75
Fig. 4.14. Two-dimensional partition of the operating space. ......................................... 76
Fig. 4.15. Model validation for a two-dimensional IBC black-box model using
a load step from 5% to 25% and 100% of the nominal load. ............................................ 77
Fig. 4.16. Schematic of a flyback prototype. ................................................................... 78
Fig. 4.17. Open-loop flyback FRFs measured at different output current values. ........... 79
Fig. 4.18. Open loop flyback linear model. ...................................................................... 80
Fig. 4.19. Open loop flyback polytopic model. ............................................................... 81
Fig. 4.20. Closed-loop flyback FRFs measured at different output current values. ........ 82
Fig. 4.21. Experimental and predicted input current and output voltage for different load
steps................................................................................................................................... 83
Fig. 4.22. Back current gain (Hi) FRFs measured
at different input voltage and fixed output current of 3 A. ............................................... 84
Fig. 4.23. Flyback polytopic black-box model operating space . .................................... 84
Fig. 4.24. Experimental and simulation results for simultaneous variation of the input
voltage and load current. ................................................................................................... 85
Fig. 4.25. Schematic of a flyback converter referred to the secondary side. ................... 87
Fig. 4.26. Operating space where polytopic models predict the converter nonlinear
behavior............................................................................................................................. 89
Fig. 4.27. Eigenvalues evaluated at the boundaries of the operating space. .................... 89
Fig. 4.28. Eigenvalues crossing into unstable regions. .................................................... 90
Fig. 4.29. FRFs from average models during CCM and DCM........................................ 92

Fig. 4.30. Closed loop flyback converter load step from DCM to CCM. ........................ 92
Fig. 5.1. Black-box terminal behavioral models of commercial converters and filters. ... 99
Fig. 5.2. Interconnected system for FRF prediction. ..................................................... 100
Fig. 5.3. Lamda power supply output impedance. ......................................................... 101
Fig. 5.4. EMI filter FRFs. .............................................................................................. 101
Fig. 5.5. Experimental and predicted output impedance at interface a. ..................... 102
Fig. 5.6. Experimental and predicted output impedance at interface b. ..................... 102
Fig. 5.7. Cascaded system. ............................................................................................. 103
Fig. 5.8. Model prediction of a cascaded system. .......................................................... 104
Fig. 5.9. Semi-regulated bus converters in parallel (passive current sharing) ............... 105
Fig. 5.10. Model prediction of a parallel system. .......................................................... 106
Fig. 5.11. Distributed power electronic system. ............................................................ 107
Fig. 5.12. Model prediction of a distributed power electronic system. ......................... 109
Fig. 5.13. Subsystem interaction. ................................................................................... 110
Fig. 5.14. Equivalent source impedances with different leads lengths. ......................... 111
Fig. 5.15. Low interaction (short wires). ....................................................................... 112
Fig. 5.16. Transient prediction (short wires). ................................................................. 112
Fig. 5.17. Medium interaction (medium-length wires). ................................................. 113
Fig. 5.18. Transient prediction (medium-length wires). ................................................ 113
Fig. 5.19. Strong interaction (long wires). ..................................................................... 114
Fig. 5.20. Transient prediction (long wires). ................................................................. 114
Fig. 5.21. H
im
Pole location as wire inductance increases. ............................................ 115
Fig. 5.22. Subsystem interaction between a regulated bus converter and a POL .......... 116
Fig. 5.23. Regulated bus converter and POL (low interaction ). ................................... 116
Fig. 5.24. Transient prediction regulated bus converter and POL (low interaction ). ... 117
Fig. 5.25. Regulated bus converter and POL (strong interaction ). ............................... 117
Fig. 5.26. Transient prediction regulated bus converter and POL (strong interaction ). 118
Fig. 5.27. Transient prediction regulated bus converter and POL (strong interaction,
system collapse when under large disturbance). ............................................................. 119
Fig. 5.28. System with damping filter. ........................................................................... 120
Fig. 5.29 Bode plot of the minor loop gain Z
s
/Z
in
........................................................... 121
Fig. 5.30 Bode plot of the minor loop gain Z
s
/Z
in
........................................................... 121
Fig. 5.31. Reshaping H
im
to obtain H
id
. .......................................................................... 122
Fig. 5.32. Full order and reduced order damping filter. ................................................. 122
Fig. 5.33. Simulated transient response using full order filter. ...................................... 123
Fig. 5.34. Simulated transient response using reduced order filter. ................................ 123
Fig. 6.1 General Black-box model structure for three-phase converters ........................ 130
Fig. 6.2 Black-box model structure for a open-loop voltage source inverter ................. 131
Fig. 6.3 Measurement setup for three-phase converters ................................................. 132

xi

Fig. 6.4 Voltage source inverter FRFs from linearization of the converter average model
......................................................................................................................................... 133
Fig. 6.5 Comparison between average model and black-box model when under load
disturbance ...................................................................................................................... 134
Fig. 6.6 Comparison between average model and black-box model when under
simultaneous variation of the input voltage and load current ......................................... 135
Fig. A.1 Open loop FRFs from measurements and identified models........................... 136
Fig. B.1 Open-loop FRFs from measurements and identified models ........................... 138
Fig. C.1 Coupled and unterminated FRFs ...................................................................... 140
Fig. D.1 flyback FRF plot 2D and 3D ............................................................................ 147
Fig. D.2 Flyback transient response under large disturbances ........................................ 149

xii

List of Tables

Table 2.1 Network Analyzer settings ............................................................................... 35
Table 2.2 Model Order Reduction .................................................................................... 44

1. Chapter
Introduction
This chapter presents the motivations and objectives for this work. It provides a
review of the most important modeling methods used for analysis and design of power
electronics converters and power electronics distribution systems; it also describes the
state of the art of black-box modeling approaches in power electronics and presents the
outline of the dissertation.
1.1 Motivation and Objectives
It is indisputable that power electronics converters have a strong impact on the
way electrical energy is processed today in a wide range of applications, whereas initially
power electronics converters worked mainly as loads in a electrical distribution systems.
For instance, currently power electronics converters are used in personal digital devices,
computers, and energy efficient home appliances.
However, a growing need for highly reliable electrical distribution systems for
critical infrastructure, such as telecommunications, hospitals, datacenters, high-end
servers, aircraft, satellites, hybrid vehicles, and ships, have promoted the development of
hybrid power systems. Hybrid power systems ensure the availability of energy by
combining renewable with non-renewable energy sources and energy storage elements
such as batteries, flywheels, super-capacitors, etc. Because of the different nature of the
sources, energy storages, and loads, they cannot be directly interconnected to form a
system. Instead, power electronics converters are used as interfaces that allow their
interconnections and ensure a proper performance by controlling the power flow in the
2

system. Thus, in current systems converters are not only loads but also an integral part of
the distribution system by feeding other converters and loads, giving origin to what is
known as electronics power distributions systems (EPDS).
With this new trend, the increasing of the efficiency, reliability, and power
density of converters and a reduction of their cost was necessary. However, these goals
could not have been achieved without a radical change in the way converters were
designed and built. Thus, universities, and later industry, embraced the concept of
modularization In which converters are encapsulated plug and play power processing
units with internal protections and added functionalities to improve system level
performance. Fig. 1.1 shows two commercial products as an example of the
improvements achieved using the modular concept in low power dc-dc converter for
which with nearly half the size of the discrete version, the modularized version can
provide nine times more power.


Input Voltage DC 18V~36V 55W DC-DC
177 x 50.8 x 37 mm
Input Voltage DC 18V~36V 500W DC-DC
116.8 x 61 x 12.7mm
Fig. 1.1 Discrete and modular approach for power electronics converters, photo by author.
For medium and high power applications, there are some physical barriers that
limit but do not eliminate the applicability of the modularization concept. Thus, a similar
building block known as integrated power electronic modules (IPEM) or power
electronics building block (PEBB) have been proposed. For these, the same plug and play
power processor unit concept seen in low power applications has been implemented.
The modularization concept has also changed the way EPDS are constructed
today in that instead of building their own converters, engineers would rather buy highly
efficient low cost power electronics modules, interconnect their power terminals, set the
3

protection settings as well as turn on and turn off sequences, and have a system running
in a very short period of time.
Nevertheless, there is a drawback to this approach that becomes evident during
the design stage of such systems. Because companies need to protect their products, they
limit the amount of information they provide to the user. Therefore, system designers are
not provided with a converter model or the necessary information to build traditional
switching or average models. Hence, the design of these systems relies on engineers
experience and testing of prototypes under multiple conditions and operating modes,
thereby increasing the development cost and time to marketing of a product.
Basically, there is a lack of models for commercial converters that can be used to
study or analyze any subsystem interaction phenomena that could degrade the dynamic
performance of the system during load and/or source transients by making the system
unstable, or, if it remains stable, to know with certain degree of accuracy how far the
system is from an unstable region. It also can be difficult to answer the question of
whether the system meets the transient load requirement.
The aim of this work is to develop a methodology that allows the construction of
unterminated power electronics converter models, without requiring any information
about the converter internal component control or topology. The desired features of the
model are that it should:
Be experimentally obtained.
Capture only the converter dynamic (unterminated).
Predict large signal current and voltage transients when working stand-
alone or when coupled with other converters and loads.
Be used for small signal analysis (bode plots, Nyquist, root locus).
The proposed approach is based entirely on the terminal characteristics of power
converters for which a set of frequency response measurements and the system
identification are used to construct a low-frequency model without requiring any
knowledge either of the converter structure or of any of its internal parameters. Yet, the
model itself intrinsically captures all this information without neglecting parasitic
4

elements or simplifying the converter internal operation as conventional reduced-order

models do. A conceptual representation is given in Fig. 1.2. Where the input signal are
input voltage v
1
and output current i
2
the mode predicts v
2
and i
1
.

Fig. 1.2 Two-port black-box model
In addition, the modular aspect of the approach enables computationally very
efficient analyses of transients and steady state operation of power electronics
distribution systems. These models are suited to system level design, specifically for the
study of dynamic interactions between cascaded series and paralleled converter systems
1.2 Literature Review
The literature on power electronics modeling efforts is focused on three areas of
study: the system itself, the power converter and the single component. In any of these
areas, from the physics based to the behavioral model, there are a variety of intermediate
models as shown in Fig. 1.3. Since the literature on modeling of power electronics is vast,
only the most relevant studies are cited here; however, the cited references can lead the
reader to more references.
3-D Finite element
3-D numerical
1-D physics based
Non-linear quasi-static
Piece-wise linear
Variable resistance
Detailed Switch
Simple switch (w/loss)
Simple switch (on-off)
Average w/loss
Average (no loss)
Simplified average
Transient-switching
Dynamic wtih ripple
Dynamic - control
Stationary multi-freq.
Stat no-lin (load flow)
Stationary linear
P
h
y
s
i
c
s

b
a
s
e
d
B
e
h
a
v
i
o
r
a
l
Component Converter System

Fig. 1.3 Classification of power electronics models
5

1.2.1 Component Level Models

The most detailed models (physics based) are used to study specific phenomena in
switching devices or passive components, e.g., turn-on and turn-off semiconductors [1]
and impedance modeling of electrolytic capacitors [2]. For the case of switching devices,
a complete model based on the semiconductor physics uses a three-dimensional
representation of the device geometry and numerically solves a set of differential
equations of the physical process, such as Poisson and continuity equations or solves by
some approximation method [3]. The complete models are, many times, simplified by
reducing the equations to the two- or one-dimensional case [4]. Simplification of the
device equations can lead to further simplified models [5]. Other approaches do not
reduce the model dimensions; instead, they approximate the spatial behavior by
simplifying the physics equations [6]. One of the techniques is to use transmission line
methods; however, as approximations, these methods do not represent the device as
accurately as the original physics based methods. A lot of effort in device modeling has
been done in obtaining models suitable for circuit simulation.
A non-linear quasi-static approach is a simplified model that calculates the total
charge in the semiconductor, but does not consider its spatial distribution [7]. This
approximation, called the Hammerstein model, combines non-linear static plus linear
dynamic equations. These models are used when simulation speed and flexibility in
model synthesis are required. Piece-wise linear models are usually good enough to
predict the behavior of the power device in the circuit [8], [9].
For passive components, most detailed inductor and capacitor models integrate
Maxwell differential equations using finite elements. In the case of inductors, the goal is
to represent the parasitic capacitances, leakage inductances, and flux distribution [10].
This type of model is used to model high-energy-density capacitors to study the electric
field distribution, corona effect and changes in the dielectric properties [11]. By
simplifying the model but still considering the electromagnetic equations, the model can
become easier to handle than the full order model. Considering the electric model
coupled with a core model of an inductor, saturation and magnetic losses can be predicted
[12]. By simplifying the model, a lumped representation of the passive component is
6

obtained. In the case of an inductor, it includes its parasitic resistance and capacitance;
for a capacitor it includes the parasitic inductance and series resistance. This type of
model is generally used for EMI studies [13], [14]. The simplest description of the
passive element is an ordinary differential equation describing the time domain behavior
of the element.
1.2.2 Converter Level Models
At the converter level, a large number of models are reported in the literature.
They can be classified as white-box or black-box modeling approaches. White-box
models include switching models, average models, discrete models, and sample data
models.
Switching models help designers to gain an understanding of how the circuit
works during the different switching stages and to estimate current and voltage ratings for
semiconductor devices and passive components [15]. Also, switching simulations are
used to study electromagnetic interference (EMI). Accurately modeling EMI noise
generation and propagation in power converters is the first step to predicting and
managing the EMI noise in the system [16 ], [17].
On the other hand, the utilization of averaging in power electronics was reported
for the first time by Wester and Middlebrook [17]. They propose to average the duty
cycle over a switching cycle and justify their approach by noting that in a well-designed
converter the time constant of the circuit is greater than the switching period. The result
of the averaging process is a continuous model that neglects the switching action but
keeps the slow non-linear behavior of the converter. These models can be linearized
around the operating point, allowing designers to use all the linear system theory tools for
control design and to asses the stability of the converter under small perturbations. After
[17], different averaging techniques were developed and can be classified as: state-space
average [19], [20], circuit average [21], switch average [22], [23], and generalized
average [24], which, in essence, will lead to the same average model [25].
7

The next step in the process of developing average models was to include the
effects of variables that affect the real behavior of the converter but were neglected
during the average process. Those models are described in the next paragraphs.
Unified Discontinuous conduction mode (DCM) and continuous conduction mode (CCM)
average model:
Initially average models were differentiated as either CCM or DCM even though
it was known that the DCM model is a general version of CCM models [19],
[20],[22],[23]. Hence, unified representations were proposed in order to model cases like
PFC circuits in which the converter works between these two modes of operation
[27],[28].
Average Model with ripple estimation:
Krein and Bass [29] took the theory developed in [30], [31] and gave a
mathematical justification of averaging in power electronics. Their work confirmed the
theorems presented in [31] and proposed a methodology to recover the ripple information
from the average values. The ripple estimation allows designers to obtain information
about currents and voltage ripples without the need for switching simulation [32].
Frequency dependant Average model: As mentioned above, average models give
a very good approximation of the behavior of the original circuit when a time scale
separation exists between the time constant of the power stage and the switching time
period. However, this dependence is not explicit in the models. When used in cases with
high ripple condition, the regular average model fails to predict the flow of the system.
Hence, average models that depend on the switching frequency were proposed in [33]-
[35].
Average models with current control:
By controlling the current and the voltages, it was possible to improve the
dynamics of the converter, especially for those topologies with a zero in the right half
plane [36], [37]. Also, this approach makes paralleling the power stage possible and is
very well suited for converters which must draw sinusoidal input current from an ac
8

power system. Today, the most commonly used techniques to control the inductor current
are peak current, average current, and charge control.
The first average models for converters under peak current control were accurate
at low frequencies [38], [39]. But they were not able to predict a sub-harmonic instability
for duty cycles above 0.5 [40]. Ridley [41] proposed a model that combines a
simplification of the sample-data models with the three-terminals switch model that is
valid until half of the switching frequency and can predict sub-harmonics instabilities.
Based on the same idea, a state space model was presented in [42]. A unification of the
models was proposed in [43]. Average models were also proposed for average current
mode control and charge control [44]-[46].
Discrete models
These models do not attempt to transform a switched-mode converter into a
continuous one; rather, the system is described in terms of a sequence of samples, one per
switching cycle. Researchers have thought that discrete models were the best modeling
approach because of the discrete nature of the converter [47], [48]. Hence, the converter
is represented by a difference equation that is non-linear and time invariant. Then, after
linearization around the operating point, a small signal discrete models is obtained.
The small signal models obtained using discrete methods are more accurate at
high frequencies than average models and predict the instability of the converter under
current mode control for duty cycles greater than 0.5. Despite the notable accuracy of
these models, they are infrequently used because designers are not accustomed to discrete
models and the dependency on numerical methods does not give any insight on the
behavior of the converter. Hence, simplified discrete models called sample data models
were proposed [49], [50].
The sample data model, which is basically a combination of continuous average
models and discrete models, is the one that has evolved and is used currently for digital
control design of converters [51], [52]. Today, increased attention is given to discrete
models because they are very well suited to studying nonlinear phenomena in power
electronic circuits. For instance, phenomena such as bifurcation, limit cycles, and chaos
can be predicted [53], [54], which is not possible using average models. In addition,
9

because discrete models describe the evolution of the state variables sampling the state
variables, tools like Poincar maps and stroboscopic maps can be used [55], [56].
What is common to all these models is that the designer has knowledge of the
power converter structure. Meanwhile, the most significant parameters, usually parasitic
resistance, inductances of passive components, and PCB traces, are unknown or
neglected in order to simplify the model derivation, even though it is known that these
parasitic resistances are important during transient dynamics and steady state conditions
[57]. Despite their limitations, simplified average models are used for converter level
modeling and analysis because they provides a good tradeoff between accuracy and
simplicity [58 ], [59].
Black-box models:
In general, models that do not require prior knowledge of the converter
parameters are known as black-box models. In the power electronics field there are some
black-box converter models reported in the literature. They are exclusively oriented to
obtain models from the control port to output voltage of the converter for control design
purposes. Because the methodology is experimentally based, the plant model obtained
through this process is claimed to be an accurate representation of the real control to
output converter characteristic. However, like the models listed above, during the
modeling process, the reported black-box models include the load dynamics and ignore
the impact of the source dynamics. Thus, the models would predict the transient response
of the real converter with a certain degree of accuracy as long as the converter is
connected to the same source and the same load used during the modeling process.
Based on the nature of the model structure, black-box models can be classified as
non-linear or linear. Non-linear black-box models are used to represent nonlinear systems
in a wide range of operations. For instance, [60] used the nonlinear autoregressive
moving average with exogenous input, or NARMAX model, to represent the dynamic
behavior from control to output voltage of an open loop boost converter. Time domain
data used to identify this kind of model is generated by perturbing the duty cycle and
measuring the perturbation and the output voltage.
10

The resulting model is compared against conventional average models and

experimental results [61], which also uses a NARMAX model to obtain the control to
output characteristics of a buck converter working in continuous conduction mode.
However, the focus of the paper is on the advantages of using a priori information for the
black-box model structure selection in order to speed up the identification process.
Another example of a nonlinear black-box model of a boost converter for control
design purposes is presented in [62], which describes a Hammerstein model consisting of
a static nonlinearity followed by a linear discrete-time and time-invariant model as shown
in Fig. 1.4. The static nonlinearity describes the steady-state behavior, and the linear part
of the models, together with the static nonlinearity, describes the transient response in the
large signal sense and not only around the neighborhood of an equilibrium point, as linear
black-box models do.

Fig. 1.4 Hammerstein model
The structure of the model is limited to a single-input, single-output model and,
similarly to NARMAX models, it requires time domain data to build the model. In [63],
the Hammerstein model had been also used to model the input voltage to speed
characteristics of a series motor and model the control to output voltage of a buck
converter.
Neural networks were considered in [64] for modeling the control to output
voltage behavior of a step up converter because of the ability of neural networks to
approximate nonlinear functions relating to inputoutput data from a nonlinear system. In
this case a 6-4-1, multilayer feed forward neural network is enough for the
implementation of the model.
On the other hand, for control design purposes, small signal models have been
successfully used for the synthesis of controllers for power converters. Reference [65]
proposed a linear black-box structure to obtain small signal models for converters for
which, because their topology or working principle, it is difficult to obtain the models
11

analytically, such as a series resonant converter and a forward multi-resonant converter.

Reference [66], using a forward converter, shows the resulting control to output voltage
models obtained using some existing system identification methods.
1.2.3 System Level Model
When designing power electronics distribution systems, defining models in terms
of circuit elements is not feasible. Instead, simpler modular behavioral models have been
proposed for system level studies. For instance, [67], [68] proposed to represent each
converter and passive module of a spacecraft power system as an unterminated two-port
network using hybrid G parameters. This methodology was further developed and
expanded in [69], [70] and used for subsystem interaction analysis in [61]-[77] .
However, the models mentioned above did not present strong advantages compared with
existing models because they require first having a small signal converter model to derive
four transfer functions (Gij) that dynamically describe the converter. Hence, theses
modeling approaches have not been widely embraced for system level studies. Instead
average models are preferred.
In addition, there is a need for information about the converter internal topology
and parameters. Average models for system-level simulations have an additional
disadvantage when used for the system level. This is due to the fact that if small values of
inductances and capacitances are included in the model, they will give origin to a system
with multiple time scales in which some states are fast and others are slow. These small
values multiply the left hand sides of the involved differential equations. This fact
produces large right hand sides thus making the model numerically stiff for computer
simulations. As a consequence, small time steps are needed to solve the system of
equations, thereby increasing the simulation time [78] - [81].

12

1.3 Dissertation Outline and Summary of contributions

This dissertation presents a methodology to obtain a system-level black-box
model for dc-dc converters. The models are intended for simulation, analysis, and design
of large-scale distributed power systems. To support the above, model features were
specified up front as being; experimentally obtained, capturing only the converter
dynamic (unterminated), predicting large signal current and voltage transients when
working stand-alone or coupled with other converters and loads, Usable for small signal
analysis (bode plots, Nyquist, root locus).
Each chapter of the dissertation focuses on developing, incorporating and
demonstrating one-by-one the features listed above. For instance, Chapter 2 is devoted to
presenting in detail the black-box internal structure. It describes how to experimentally
obtain the information used to build the model, the procedure to obtain the models from
the experimental data via system identification, the model validation procedure used
throughout this work, and a model order reduction technique. Thus, this chapter presents
the building blocks for the proposed methodology.
Chapter 3 investigates an important issue not addressed in Chapter 2 by
examining how to obtain unterminated frequency response functions (FRFs) given that,
when interactions with the source and the load are significant, the measured FRFs do not
solely capture the internal converter dynamics. Instead, the measured FRFs contain
information of the converter, the source and load dynamics. Two approaches are
presented. One considers the interaction with the source and neglects the interaction with
the load when the converter is loaded with an electronic load working in current sink
mode. The other is the general case because it considers the effect of the source and load
dynamics simultaneously.
Chapter 4 extends the application of the two-port black-box model to not only
represent buck-converter topologies in a reduced operating space but also to represent
other topologies such as buck-boost that present strong nonlinear behavior. The models in
this chapter are valid in a wider operating region compared to the models developed in
Chapter 2. Two nonlinear black-box model structures are considered, the winner model
13

and the polytopic model. The former is used in cases in which nonlinearities are present
in the form of dc gains; the later is used when the states present strong slowly variant
nonlinearities. In addition, a section of this chapter presents the conditions in which this
modeling approach fails.
Chapter 5 shows the application of two-port black-box models to simulating the
dynamic behavior of a commercial converter interconnected with others converters in
parallel and cascade. Also, linear system tools are applied to the models to assess the
stability of interconnected systems. All results are verified with experimental data.
Finally, Chapter 6 presents the conclusions and suggests future work.

14

1.4 References
[1] L. Lu, A. Bryant, E. Santi, J.L Hudgins, P.R Palmer Physics-based model of IGBT
including MOS side two-dimensional effects in proceedings of the 41
st
IEEE Industry
Applications Conference, Volume 3, Oct 8-12. 2006 pp. 1457 - 1464
[2] A. Bajolet,. R. Clerc.; G. Pananakakis, D. Tsamados, E. Picollet, N. Segura, J.-C. Giraudin,
P. Delpech, L. Montes, G. Ghibaudo, Low-frequency series-resistance analytical modeling
of three-dimensional metalinsulatormetal capacitors IEEE Transactions on Electronic
Devices, Volume 54, April 2007 pp.742 - 751
[3] B. Fatemizadeh, P. Lauritzen, D. Silver, Modeling of Power Semiconductor Devices,
Problems, Limitations and Future Trends IEEE Workshop on Computers in Power
Electronics, Aug.1996, pp. 120 - 127
[4] W. Van Dell, W. Kyle, Seminet: further advances in integrated device/circuit simulation
for power electronics, in proceedings of 19
th
IEEE Power Electronics Specialists
Conference, Vol.1, April 1988, pp. 69 - 75
[5] P. Leturcq, A Study of distributed switching processes in IGBTs and other power bipolar
devices, in proceedings of 28
th
Annual IEEE Power Electronics Specialists Conference,
Vol. 1, June 1997, pp.139 - 147
[6] R. Hocine, M. A, Boudghene Stambouli, A. Saidane, A three-dimensional TLM
simulation method for thermal effect in high power IGBTs, in proceedings of 18
th
Annual
IEEE Symposium Semiconductor Thermal Measurement and Management, March 2002,
pp. 99-104
[7] B. Beydoun, H. Tranduc, F. Oms, G. Charitat, P. Rossel, Power MOSFET design and
modeling tool for power electronics, in proceedings of the 5
th
European Conference on
Power Electronics and Applications, Vol.2, Sep 1993, pp. 390 395
[8] C. Batard, State-space models of diode and bipolar transistor for simulation of power
converters, in proceedings of 7
th
IEEE Power Electronics Specialists Conference, Vol.2,
June 1996, pp. 1661 - 1667
[9] P. Bozovic, P. Pejocic, A simple piece-wise linear IGBT model for specialized power
electronics program, Conference Record PCIM98, 1998, pp.475-484
[10] J.M Lopera, M.J. Prieto, A.M. Pernia, F Nuno, A multiwinding modeling method for high
frequency transformers and inductors, IEEE Transactions on Power Electronics, Vol.18,
Issue:3, May 2003 ,pp 896 906
[11] A Wu, M.D, Driga, Mathematical formulation of 2D and 3D finite and boundary element
model for transient electric fields in high performance capacitors, IEEE Transactions on
Magnetics, Vol. 29, Issue: 1, Jan 1993, pp.1088 1092
[12] A.F Witulski, Modeling and design of transformers and coupled inductors, in
proceedings of IEEE Conference Applied Power Electronics, March 1993, pp.589 595
[13] G. Grandi, M.K Kazimierczuk, A. Massarini, U. Reggiani, G. Sancineto, Model of
laminated iron-core inductors for high frequencies, IEEE Transactions on Magnetics, Vol.
40, Issue: 4, July 2004, pp. 1839 1845
15

[14] L. Yang; B. Lu; W. Dong; Z. Lu; M. Xu; F.C. Lee, W.G. Odendaal, Modeling and
characterization of a 1 KW CCM PFC converter for conducted EMI prediction, in
proceedings of the 19
th
Annual IEEE Conference Applied Power Electronics, vol.2, pp.763
- 769
[15] Robert W. Erickson, Dragan Maksimovic Fundamentals of Power Electronics, Second
Edition Springer; 2nd ed. edition (January 2001)
[16] L Arnedo and K. Venkatesan High-Frequency Induction Motor Model for EMI and Over-
Voltage Studies Journal of Electric Power Components and Systems, Volume 31, Number
11, November 2003 , pp. 1047-1061
[17] Andrew C Baisden,.; Dushan Boroyevich,; Jacobus van Wyk Impedance Interaction and
EMI Attenuation in Converters with an Integrated Transmission-Line Filter in
proceedings of the 22
th
IEEE Applied Power Electronics Conference, Feb. 2007, pp.1203-
1208
[18] G. W. Wester and R. D. Middlebrook, Low Frequency characterization of switched dc-dc
converters in proceedings of IEEE Power Processing and Electronics Specialists
Conference, May 1972, pp. 9-20
[19] R. D Middlebrook, S. Cuk, A general Unified Approach to Modeling Switching
Converter Power Stage, in proceedings of IEEE Power Electronic Specialists Conference,
1976, pp 18-34
[20] R. D Middlebrook, S. Cuk Modeling and analysis method for Dc-to Dc Switching
converters, in proceedings of The IEEE international Semiconductor Power Converter
Conference, 1977, pp 90-111
[21] S.R. Sanders, G.C Verghese, Synthesis of averaged circuit models for switched power
converters, IEEE International Symposium on Circuits and Systems, Vol.1, May 1990,
pp.679-683
[22] V. Vorperian, Simplified analysis of PWM converters using model of PWM switch.
Continuous conduction mode, Transactions on Aerospace and Electronic Systems, Vol.26,
Issue:3, May 1990, pp.490496
[23] V. Vorperian, Simplified analysis of PWM converters using model of PWM switch. II.
Discontinuous conduction mode, Transactions on Aerospace and Electronic Systems,
Vol.26, Issue:3, May 1990, pp.490496
[24] S.R. Sanders, J.M. Noworolski, X.Z. Liu, G.C. Verghese, Generalized averaging method
for power conversion circuits IEEE Transactions on Power Electronics, Vol. 6, Issue: 2,
April 1991, pp.251259
[25] E. Van Dijk, J.N. Spruijt, D.M. O'Sullivan, J.B. Klaassens, PWM-switch modeling of DC-
DC converters, IEEE Transactions on Power Electronics, Vol. 10, Issue 6, Nov. 1995,
pp.659-665
[26] R. D Middlebrook, S. Cuk, A general Unified Approach to Modeling Switching Dc-Dc
converters in Discontinuous conduction mode, in proceedings of The IEEE Power
Electronics Specialists Conference,1980, pp.36-57
[27] Y.Amran, F.Huliehel, S Ben-Yaakov, A unified SPICE compatible average model of
PWM converters, IEEE Transactions on Power Electronics, Vol. 6, Issue:4, Oct. 1991
16

[28] V.M Canalli, J.A Cobos, J.A Oliver, J Uceda, Behavioral large signal averaged model for
DC/DC switching power converters, in proceedings of The 27
th
IEEE Power Electronics
Specialists Conference, Vol.2, June 1996, pp.1675-1681
[29] P.T Krein, J. Bentsman, R.M. Bass, On the use of averaging for the analysis of power
electronic systems, in proceedings of The 20
th
IEEE Power Electronics Specialists
Conference, June 1989, Vol.1, pp.463 - 467
[30] N.N. Bogoliubov, Y.A. Mitropolsky, Asymptotic Methods in the Theory of Non-Linear
Oscillations, Hindustan Publishing corporation, 1961
[31] P.R Sethna, M. Balachandra, Some Asymptotic Results for Systems With Multiple Time
Scales, Siam journal Applied Math, Vol 27, No 4, Dec. 1974
[32] J. Sun; R.M Bass, Automated ripple analysis using KBM method in proceedings of The
23
rd
International Conference on Control and Instrumentation, Vol.2, Nov. 1997, pp.621 -
626
[33] B. Lehman, R.M Bass, Extensions of averaging theory for power electronic systems, in
proceedings of 25
th
IEEE Power Electronics Specialists Conference, Vol.2, June 1994, pp.
1053 - 1057
[34] B.Lehman, R.M Bass, Switching frequency dependent averaged models for PWM DC-DC
converters in proceedings of The 26
th
Power Electronics Specialists Conference, Vol.1,
June 1995, pp. 636 - 642
[35] V.A. Caliskan, G.C. Verghese, A.M Stankovic, Multi-frequency averaging of DC/DC
converters. IEEE Workshop on Computers in Power Electronic, Aug. 1996, pp. 113 119
[36] C.W Deisch, Simple switching control method changes change power converter into a
current source, in proceedings of IEEE Power Electronics Specialists Conference,1978, pp
300-306,
[37] A. Capel, G ferrante, D.O Sullivan, A. Weinberg, Application of injected current model
for the dynamics of switching regulators with the new concept of LC3 modulators, in
proceedings of IEEE Power Electronics Specialists Conference , 1978, pp 135-147.
[38] R.D.Middlebrook, Modeling current-programmed buck and boost regulators, IEEE
Transactions on Power Electronics, Vol.4, Issue: 1, Jan. 1989, pp:36 52
[39] R.D Middlebrook, Topics in multiple loop regulators and current mode programming, in
proceedings of IEEE Power Electronics Specialist Conference. 1985, pp 716-732
[40] J.H.B Deane, D.C Hamill, Instability, subharmonics and chaos in power electronic
systems, in proceedings of The 20
th
Annual IEEE Power Electronics Specialists
Conference, Vol.1, June 1989, pp. 34 - 42
[41] R.B. Ridley, A new continuous-time model for current-mode control, IEEE
Transactions on Power Electronics, Vol. 6, Issue: 2, April 1991, pp. 271 280
[42] R. Tymerski, D. Li, State space models for current programmed pulse width modulated
converters, in proceedings of The 23
rd
Annual IEEE Power Electronics Specialists
Conference, Vol.1, June-July 1992, pp. 337 - 344
[43] F.D Tan, R.D Middlebrook, A unified model for current-programmed converters, IEEE
Transactions on Power Electronics, Vol. 10, Issue: 4, July 1995, pp. 397 408
[44] W. Tang, F.C. Lee, R.B. Ridley, Small-signal modeling of average current-mode control,
IEEE Transactions on Power Electronics, Vol. 8, Issue: 2, April 1993, pp, 112119.
17

[45] W.Tang, F.C Lee, R.B Ridley, I. Cohen, Charge control: modeling, analysis, and design
IEEE Transactions on Power Electronics, Vol.8, Issue: 4, Oct. 1993, pp. 396403.
[46] V.Vorperian, The charge-controlled PWM switch in proceedings of The Power
Electronics Specialists Conference, Vol: 1, June 1996, pp. 533 542
[47] F.C Lee, Y. Yu, Computer-Aided analysis and simulation of switched dc-dc, IEEE
Transaction on Industrial Applications, Vol. 15 No 5, Sep/Oct 1979, pp 511-520
[48] D. J. Packard, Discrete modeling and analysis of switching regulators, PhD Thesis
California Institute of Technology, May 1976.
[49] R. Brown, R.D. Middlebrook, Sample-Data modeling of switching regulators, in
proceedings of The IEEE. Power Electronic Specialists Conference, 1981, pp 349-369
[50] D J Shortt, F.C Lee, An improved switching converter model using discrete and averaging
techniques, in proceedings of The IEEE Power Electronic Specialists Conference, 1982,
pp 199-212
[51] J. Sun, B Heck, B Lehman, Continuous approximation and the stability of averaging, in
proceedings of The 7
th
Workshop on Computers in Power Electronics, pp. 139 - 144
[52] M.E Elbuluk, G.C Verghese, J.G Kassakian, Sampled-data modeling and digital control of
resonant converters, IEEE Transactions on Power Electronics, Vol: 3, July 1988, pp.344
354
[53] S. Banerjee, G. C. Verghese, Nonlinear Phenomena in Power Electronics: Bifurcations,
Chaos, Control, and Applications, Wiley-IEEE Press, 1 edition (July 2, 2001)
[54] Chi Kong Tse, Complex Behavior of Switching Power Converters CRC Press (July 28,
2003)
[55] A. H. Nayfeh, B. Balachandran, Applied Nonlinear Dynamics: Analytical, Computational,
and Experimental Methods, Wiley-Interscience (January, 1995
[56] S.K Mazumder, A.H. Nayfeh, D. Boroyevich, An investigation into the fast- and slow-
scale instabilities of a single phase bidirectional boost converter, IEEE Transactions on
Power Electronics, Vol. 18, Issue: 4, July 2003, pp 1063 1069
[57] A. Davoudi, J. Jatskevich, Parasitics Realization in State-Space Average-Value Modeling
of PWM DCDC Converters Using an Equal Area Method. IEEE Transactions on
Circuits and Systems, Volume 54, Sept. 2007 pp. 1960 1967
[58] A. Emadi, Modeling and analysis of multiconverter DC power electronic systems using
the generalized state-space averaging method IEEE Transactions on Industrial
Electronics, Volume 51, June 2004 pp.661 66
[59] M.M. Jalla,; A. Emadi,; G.A. Williamson, B. Fahimi, Modeling of multiconverter more
electric ship power systems using the generalized state space averaging method, in
proceedings of The 30
th
IEEE Conference of Industrial Electronics Society,. Nov 2-6. 2004.
pp. 508 - 513
[60] K.T. Chau, C.C. Chan. Nonlinear identification of power electronic systems in
proceedings of The International Conference on Power Electronics and Drive Systems,
1995, Feb 21-24. 1995, pp329 - 334
[61] L.A Aguirre,P.F Donoso-Garcia, R. Santos-Filho. Use of a priori information in the
identification of global nonlinear models-a case study using a buck converter; IEEE
Transactions on Circuits and Systems I. Volume 47, Issue 7, July 2000, pp.1081 1085
18

[62] F.Alonge, F. D'Ippolito, F. M. Raimondi, S. Tumminaro, Nonlinear Modeling of DC/DC

Converters Using the Hammerstein's Approach IEEE Transactions on Power Electronics,
Volume 22, Issue 4, July 2007, pp. 1210 1221
[63] A. Balestrino, A. Landi, M. Ould-Zmirli, L. Sani, Automatic nonlinear auto-tuning
method for Hammerstein modeling of electrical drives IEEE Transactions on Industrial
Electronics, Volume 48, Issue 3, June 2001, pp. 645 655
[64] R.Leyva, L. Martinez-Salamero, B. Jammes, J.C. Marpinard, F. Guinjoan, Identification
and control of power converters by means of neural networks.; IEEE Transactions on
Circuits and Systems, Volume 44, Issue 8, Aug. 1997 pp.735 742
[65] Choi Ju-Yeop; B.H Cho,.; H.F. VanLandingham,; Mok Hyung-soo; Song Joong-Ho
System identification of power converters based on a black-box approach IEEE
Transactions on Circuits and Systems I, Volume 45,Issue 11, Nov. 1998 pp.1148 1158
[66] B. Johansson, M. Lenells Possibilities of obtaining small-signal models of DC-to-DC
power converters by means of system identification . in proceedings of The Twenty-
second International Telecommunications Energy Conference, Sept. 10-14 2000, pp.65 - 75
[67] B.H Cho, F.C. Lee. Modeling and analysis of spacecraft power systems IEEE
Transactions on Power Electronics Volume 3, Issue 1, Jan. 1988,pp.44 54
[68] J.R Lee, B.H Cho, S.J Kim, F.C Lee, Modeling and simulation of spacecraft power
systems IEEE Transactions on Aerospace and Electronic Systems, Volume 24, Issue 3,
May 1988 pp 295 304
[69] P.G Maranesi, V. Tavazzi,V. Varoli, Two-part characterization of PWM voltage
regulators at low frequencies IEEE Transactions on Industrial Electronics, Volume 35,
Issue 3, Aug. 1988 pp. 444 450
[70] P.Maranesi,; L. Pinola,; V. Varoli,; The incremental voltage control mode for PWM
regulators in proceedings of The 19th Annual IEEE Power Electronics Specialists
Conference, April 11-14 1988 pp. 549 - 554 vol.1
[71] T.Suntio, M. Hankaniemi, M. Karppanen, Analysing the dynamics of regulated
converters. in proceedings of The IEEE Conference on Electric Power Applications,
Volume 153,Issue 6,November 2006, pp.905 910
[72] M.Hankaniemi, M. Karppanen, T. Suntio, Load-imposed instability and performance
degradation in a regulated converter IEE Proceedings on Electric Power Applications,
Volume 153,Issue 6,November 2006 pp 781 786
[73] Kai Zenger; Ali Altowati; Teuvo Suntio. Dynamic Properties of Interconnected Power
Systems A System Theoretic Approach in proceedings of The 1
st
IEEE Conference on
Industrial Electronics and Applications, May 2006, pp.1 6
[74] M. Hankaniemi, M. Karppanen, T. Suntio, EMI-Filter Interactions in a Buck Converter
in proceedings of The 12
th
International Power Electronics and Motion Control Conference
Aug. 2006 pp.54 59
[75] Kai Zenger, Ali Altowati, Teuvo Suntio, Stability and Performance Analysis of Regulated
Converter Systems - in proceedings of The 32
nd
IEEE Conference on Industrial
Electronics, Nov. 2006, pp. 1975 1980
[76] A. Altowati, T. Suntio, K. Zenger, Input filter interactions in multi-module parallel
switching-mode power supplies in proceedings of The IEEE International Conference on
Industrial Technology, Dec 14-17. 2005, pp.851 856
19

[77] T.Suntio, D.Gadoura, Use of unterminated two-port modeling technique in analysis of

input filter interactions in telecom DPS systems ; in proceedings of The 4th IEEE
International Telecommunicationsand Energy Conference, 29 Sept.-3 Oct. 2002 pp.560 -
565
[78] Peponides, G.; Kokotovic, P.; Chow, J.; Singular perturbations and time scales in
nonlinear models of power systems , IEEE Transactions on Circuits and Systems, Volume
29, Issue 11,Nov 1982, pp.758 767
[79] J.A. Ferreira, J.P.E de Oliveira, Modeling multidisciplinary systems with Hybrid State
charts. in proceedings of The IEEE International Symposium on Computer-Aided Control
Systems Design, Oct 6. 2006 pp. 422 427
[80] Jonathan W Kimball, Philip T Krein Singular Perturbation Theory for DC-DC Converters
and Application to PFC Converters in proceedings of The IEEE Power Electronics
Specialists Conference, June 17-21 2007, pp.882 887
[81] Yong Chen Yongqiang Liu Summary of Singular Perturbation Modeling of Multi-time
Scale Power Systems in proceedings of The IEEE Transmission and Distribution
Conference and Exhibition, 2005 pp.1 4

2 Chapter
Two-port Black-Box Terminal Characterization
This chapter presents the development of modular linear black-box converter
models for system-level design and analysis. The model structure is a non-terminated
two-port hybrid parameters network built from experimental frequency response
functions (FRFs) measured at the input and output converter terminals and post-
processed using system identification methods. System identifications tools fit the FRF
data into a state space model in which high order models are typically obtained; thus, a
modeling order reduction method based on a balanced realization and singular
perturbation is implemented for this modeling approach.
2.1 Introduction
To improve the design process of power electronics converters, engineers employ
hierarchical modeling with an increasing level of detail in what is called a top-down
modeling approach. The most detailed models are used to study specific phenomena, e.g.,
turn-on and turn-off semiconductor transients, ripple current or voltages, etc. Simpler
models with ideal switches, for instance, are used for filter design, average models for
behavioral studies, and small-signal models for control system design. What is common
to all these models is that the designer has complete knowledge of the power converter
structure and parameters [1-3].
When designing power electronics conversion systems feeding multiple power
converters, there is a need to employ simpler behavioral modelsfrom the higher levels
of the modeling hierarchyin order to study and assess subsystem interactions of
21

interest, such as power quality, stability, reliability, safety, thermal management, weight,
size and cost of the system [1-3]. However, system engineers will rarely have access to
all the information required to model all the converters, and in fact in most cases they
will have no knowledge at all besides the input and output voltages and power rating of
the system components. This is common in systems built by integrated power converters,
filters, protection devices, and loads designed and manufactured by different companies,
which normally provide little or no information about the design and internal structure of
their products. Consequently, the design and construction of these systems ultimately
relies entirely on the experience of engineers and the testing of the built system under
multiple operating modes.
This chapter presents the development of linear black-box models capturing the
input-output dynamic of power converters with unknown structure, i.e., converters with
only input and output terminals accessible. The model is derived directly from hardware
using a network analyzer to measure input-output transfer functions of interest, which are
then processed and used to build a two-port network equivalent circuit. The models
developed are suitable for system level design, and specifically for the study of dynamic
interactions between cascaded and paralleled converter systems. This differs from
previous black-box converter models that were oriented to obtain plant models for control
design [4]. This chapter includes the complete derivation of a linear black-box model,
providing insight into the experimental measurement and model construction and
validation for an open- and closed-loop buck dc-dc converter power supply (25 W,
20V5V)
A conceptual representation is given in Fig. 2.1, for which the input signals are
input voltage v
1
and output current i
2
,

while the mode predicts v
2
and i
1
. In addition, the
modular aspect of the approach computationally enables very efficient analysis of
transients and steady-state operation of power electronics distribution systems. These
models are suitable for system-level design, specifically for the study of dynamic
interactions between cascaded series and paralleled converter systems.
22


Fig. 2.1 Black-box model
This chapter is divided into seven sections, which include the complete model
derivation, providing insight into the experimental measurement and model construction
and validation. For the sake of simplicity, the entire methodology is presented using an
unregulated non-isolated buck converter. However, it can be used to model isolated buck
converters and EMI filters as well, which are used extensively in intermediate bus
architecture.
2.2 Black-box Two-Port Terminal Characterization
Black-box two-port terminal characterization refers to a hardware-oriented
modeling approach that does not require a priori information about the converter
parameters or the relationship among its internal state variables. Instead, the model is
based on a non-terminated, two-port network such that the bidirectional signal flow of
information that exists in a real converter is captured by the model. The first step in the
modeling process is to define the internal structure of the two-port black-box model,
which can be either linear or non-linear.
A linear network, such as a hybrid parameters (also known as a G parameters
network), is used to model sources, loads, passive modules and converters whose
dynamic behavior is mildly non-linear and may be captured by four transfer functions,
i.e., the hybrid parameters. With this structure, the converter input ports are represented
by a Norton equivalent circuit and the output by a Thevenin equivalent circuit. Similarly,
the source can be characterized by a Thevenin equivalent circuit and the load as a Norton
equivalent circuit. Fig.2.2 shows the linear black-box structure to model sources,
converters, and loads. The converter dynamics are represented by four transfer functions,
namely audiosusceptibility G
o
(s), output impedance Z
o
(s), input admittance Y
i
(s), and
23

back current gain H
i
(s). These transfer functions are considered unterminated because
they reflect only the internal dynamics of the converter. Similarly, the source and load
dynamics are represented by the source output impedance Z
s
(s) and load input admittance
Y
L
(s).

Fig.2.2 Black-box model structure
These transfer functions are obtained from data measured using a network
analyzer at the converter terminals in the form of four frequency response functions
(FRFs). Then, system identification methods are used to obtain a linear time-invariant
model from the measurements [5-7]. The validity region of the model is defined by the
area where the converter behaves nearly as a linear system. The linear operating region is
experimentally obtained by measuring multiple frequency responses at different
operating points.
In contrast, a non-linear structure is required to represent a converter whose
behavior is primarily non-linear and whose transition from one operating point to another
is smooth. The nonlinear structures are obtained by using a physical understanding of the
nonlinearities, then determining a way to model the most important nonlinearities and
incorporate those into the linear model. This case is further explained in a later section in
the context of modeling a buck converter in closed loop.
2.3 Frequency Response Measurements
In the process of obtaining black-box models, there is the option of using either
time domain data or frequency domain data; however, when applied to power electronics,
time domain data have a disadvantage compared with frequency domain data. The origin
of the problem is that transients in power electronics circuits occur in different time
24

scalessome phenomena are fast and others are slow. Thus, in order to capture the fast
dynamics, it would be necessary use many data samples, which makes the identification
process slow. In contrast, frequency domain measurements are able to condense fast and
slow dynamics in a very efficient way [8, 9]. Therefore, frequency response functions are
used in this work.
Two-port black-box terminal characterization models are experimentally based;
therefore, the models will be as good as the data used to build them. Thus, a proper
frequency response measurement is extremely important. In this section, a description of
the measurement setup is provided. Furthermore, we discuss factors that affect FRF
measurement, such as sampling and hold effect, feedback networks, and subsystem
interaction with the source and the load. Finally, the process of obtaining a model from
the data using the MATLAB system identification toolbox is described.
In a setup for FRF measurement, four major components are used: a network
analyzer, such as the Agilent 4395A, an injection circuit, measurement probes, and an
optional linear amplifier. The network analyzer, also known as a frequency response
analyzer, produces a broad band sinusoidal excitation that is injected in the circuit under
test. Two channels equipped with aliasing filters and a configurable bandwidth measure
the circuit response to a frequency sweep. At the same time, magnitude and phase values
of the ratio between the two channels are plotted or stored in a file.
The injection circuit, on the other hand, is in charge of introducing a perturbation
in the voltage or current. Fig. 2.3(a) shows the conceptual schematic for voltage
perturbation at the input converter terminals, in which the injected voltage needs to be
connected in series with the source. The other two schematics, Fig. 2.3 (b) and (c), show
two practical implementations. The former uses an isolation transformer whose
secondary winding is connected in series with the source; the latter employs a MOSFET
working in the active region. Consequently, voltage changes at the gate-to-source
terminals produce a corresponding variation at the drain-to-source voltage. Similarly, a
current injection is often employed to introduce a disturbance in the circuit under test.
Fig. 2.3 (d), (e), and (f) shows a conceptual schematic and two circuits that are basically
the equivalent of those used for series injections [10].
25



Fig. 2.3 Series and parallel injection circuits
Each circuit has its advantages and disadvantages, and the circuit that gives a
better result for a specific situation is usually obtained through experimentation. For this
work, a current injection circuit like the one shown in Fig. 2.3 (e) is used for all FRF
measurements. The main reason for preferring parallel injection over series injection is
that any element that is in series with the source will increase the source output
impedance, which, as will be described in next chapter, plays an important role in FRF
measurements.
On the other hand, measurement probes condition the signals (currents and
voltages) in a way that makes it possible for the network analyzer to receive them. The
measurement probes used should be the least invasive possible; therefore, it is not
recommended to use shunt resistors to sense currents going in or out of the circuit under
test. Also, it is advantageous to use a linear amplifier to increase the power of the
disturbance to be injected in the circuit because network analyzers are manufactured
primarily for radio frequency circuits and microwave applications. Hence, they are
designed to drive low-power circuits, and there are cases in which they do not have
enough power to excite power electronic circuits, for instance when measuring FRFs at
low frequencies.
Fig. 2.4 (a) shows the setup employed to measure G
o
and Y
i
when a current
disturbance is injected at the input using an injection transformer in series with a
capacitor. The measurement probes are connected to the reference channel R to
26

measure the input voltage, and channels A and B measure input current and output
voltage, respectively. Hence, A/R will give Y
i
(j) and B/R will give G
o
(j). An
electronic load working as a current sink sets the operating point. The setup to measure
H
i
(j) and Z
o
(j) is illustrated in Fig. 2.4 (b). In this case, the disturbance is injected at
the output and the reference channel measures the output current of the converters.
Channels A and B measure output voltage and input current. Thus, A/R and B/R give
Z
o
(j) and H
i
(j), respectively.

Fig. 2.4 (a) Measurement setup for G
O
and Y
i


Fig. 2.4(b) Measurement setup for H
i
and Z
O
plots
Among the many factors that affect the accuracy and range of validity of FRF
measurements in power electronic circuits, the factors that require special attention are
the sampling and hold effect created by the action of the switches, feedback control, and
interaction with the source and the load.
When measuring G
o
(j) and H
i
(j), the disturbance signal is shopped by the
switch and reconstructed at the converter output filter. This can look like a sampling and
hold effect. Thus, based on the sampling theorem, the FRFs can be measured at up to half
27

the switching frequency. The lower limit of 10 Hz is imposed by the network analyzer
and the injection transformer.
As an example, two FRFs for H
i
(j) of an open-loop buck converter with known
parameters are presented in Fig. 2.5. One FRF is measured using a network analyzer, and
the other is obtained using average models. The two FRFs are similar at low frequencies,
but as the frequency increases, the average model fails to predict a degradation of the
phase that occurs before reaching half the switching frequency [11]. After reaching half
the switching frequency, it is difficult to ensure a good measurement because aliasing
phenomena become an important issue [12]. However, this is not a problem for black-box
terminal characterization models, whose purpose is to capture the slow dynamic of the
converter; thus, measurements up to half the switching frequency would be sufficient.
Back current gain Hi

Frequency (rad/sec)
Fig. 2.5 FRF H
i
from measurement and average model
For the FRFs of Y
i
(j) and Z
o
(j), measurements beyond half the switching
frequency are possible because the dynamics of these FRFs in this frequency range are
governed by the input and output converter filters.
Fig. 2.6 shows FRF of Z
o
(j) from measurement and from simulation using
average models, And some differences can be noticed. For instance, at low frequencies,
average models do not account for the effect of PCB and interconnection parasitic
resistances. Something similar occurs at high frequencies; however, in this case the
discrepancy is attributed to un-modeled parasitic inductances and/or capacitances. Thus,
average models are as good as the information available to build the model.

10
3
10
4
10
5
10
0
A
m
p
l
i
t
u
d
e
FRF G22
10
3
10
4
10
5
-400
-200
0
P
h
a
s
e

(
d
e
g
r
e
e
s
)


Measured
Average Models
Half Switching Frequency
28


Output impedance Z
o


Frequency (rad/sec)
Fig. 2.6 FRF Z
o
from measurement and average model
Another important case to consider is the measurement of FRFs on closed-loop
converters. Since closed-loop converters are designed to have good disturbance rejection
at low frequencies, the feedback network would compensate for any low-frequency
disturbance injected into the system. Hence, the network analyzer would not have any
information at the output to compute the FRF, or the response would be so small that
noise in the system would overpower it. To deal with this limitation, is better to use a
higher disturbance magnitude at low frequencies, being careful to not saturate the control
signals and to narrow the network analyzer bandwidth (2 to 10 Hz bandwidth). By doing
this, measurements starting as low as 100 Hz give reliable results for the converter used
in these experiments.
On the other hand, the measurement of un-terminated FRFs can be a difficult task
in practice, since the source and the load dynamics are intrinsically coupled in a running
converter [13]. Consequently, the measurements taken will not reflect the true internal
dynamic of the converter, which naturally limits the applicability of this modeling
approach. To address the situation above, a measurement setup that is designed to have a
weak interaction with the source and the load is initially used.
A weakly-coupled condition is achieved up to a certain frequency by feeding the
converter from a low-output impedance voltage source and an electronic load working as
a constant current sink. With this setup, accurate models are obtained based on FRF
measurements conducted up to 40 kHz. However, for cases requiring frequency sweeps
beyond 40 kHz, this approach is not sufficient, as the source and electronic load
10
3
10
4
10
5
10
0
A
m
p
l
i
t
u
d
e
10
3
10
4
10
5
-100
-50
0
50
P
h
a
s
e

(
d
e
g
r
e
e
s
)
F ( d/ )


Measured
Average Model
29

impedances will become prominent in the higher frequency range, affecting the FRF
measurements considerably and, in consequence, also the accuracy of the models built.
To cope with this problem, an enhanced FRF measurement procedure is presented in
Chapter 3.
2.4 System Identification
The problem of system identification is to infer relationships between past and
future data of unknown dynamic systems. The ultimate goal is to provide a model
approximating the behavior of the process generating the data. In classical approaches,
the search for the optimal approximation model is carried out within a linear
parameterized identification family [14]. System identification is a mature field in which
identification tools have been delivered to the user in the form of various software
packages that are widely used in the industry. In this work, the MATLAB system
identification toolbox is used to obtain the necessary models to build the two-port black-
box model. One of the most important aspect to understand when working with system
identification tools is that there are a variety of model structures available to assist in
modeling a system, and the choice of model structure is based upon an understanding of
the system identification method and insight and understanding into the system
undergoing identification. It is especially important to understand that characteristics of
both system and disturbance dynamics play a role in the proper model selection.
Given the importance of the structure selection for the system identification
process, a brief description of the most often used structures is presented together with
suggestions for appropriate use of a specific structure. The general linear structure, shown
in Fig. 2.7, provides flexibility for both the system dynamics and stochastic dynamics.
However, a nonlinear optimization method computes the estimation of the general linear
model. This method requires intensive computation with no guarantee of global
convergence.
30


Fig. 2.7 General linear model structure
Simpler models that are a subset of the general linear model structure are possible.
By setting one or more of A(q), B(q), C(q) or D(q) polynomials equal to 1, simpler
models such as AR, ARX, ARMAX, Box-Jenkins, and output-error structures can be
created.
The AR model structure, shown in Fig. 2.8 is a process model used in the
generation of models for which outputs are only dependent on previous outputs. No
system inputs or disturbances are used in the modeling. This is a very simple model that
is limited in the class of problems it can solve. This means that the AR model structure is
the model for a signal, not a system.

Fig. 2.8 AR model structure
The ARX model, shown in Fig. 2.9, is the simplest model incorporating the
stimulus signal. The estimation of the ARX model is the most efficient of the polynomial
estimation methods because it is the result of solving linear regression equations in
analytic form. The ARX model, therefore, is preferable, especially when the model order
is high.

Fig. 2.9 ARX model structure
31

The disadvantage of the ARX model is that disturbances are part of the system
dynamics. This coupling can be unrealistic, because the system dynamics and stochastic
dynamics of the system do not share the same set of poles all the time.
Unlike the ARX model, the ARMAX model structure includes disturbance
dynamics as depicted in Fig. 2.10. ARMAX models are useful when you have
dominating disturbances that enter early in the process, such as at the input. The ARMAX
model has more flexibility in the handling of disturbance modeling than the ARX model.

Fig. 2.10 ARMAX model structure
The Box-Jenkins (BJ) structure provides a complete model with disturbance
properties modeled separately from system dynamics as illustrated in Fig. 2.11

Fig. 2.11 Box-Jenkins model structure
The Box-Jenkins model is useful when there are disturbances that enter late in the
process. For example, measurement noise on the output is a disturbance late in the
process.
The Output-Error (OE) model structure in Fig. 2.12 describes the system
dynamics separately and no parameters are used for modeling the disturbance
characteristics.
32


Fig. 2.12 Output error model structure
The previous classical parametric system identification methods minimize a
performance function that is based on the sum of squared errors. These methods work
well in many cases. However, for complex systems characterized by being of high order--
that is, having many parameters, with several inputs and outputs, and having a large
number of measurements--the classical methods can suffer from several problems. They
can experience many local minima in the performance function and, therefore, a lack of
convergence to global minima. The user would need to specify complicated
parameterization of system orders and delays. These methods also may suffer potential
problems with numerical instability and excessive computation time to execute the
iterative numerical minimization methods needed. For cases such as these, the state-space
identification method is the appropriate model structure.
The following equations describe a state-space model.
( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
x t Ax t Bu t Ke t
y t Cx t Du t e t

= + +
= + +

(1)
x(t) is the state vector, y(t) is the system output, u(t) the system input and e(t) is
the stochastic error. A, B, C, D, and K are the system matrices. The dimension of the
state vector x(t) is the only setting needed to provide for the state-space model. In
general, the state-space model provides a more complete representation of the system,
especially for MIMO systems, than polynomial models because state-space models are
similar to first principle models. The identification procedure does not involve nonlinear
optimization so the estimation reaches a solution regardless of the initial guess.
Moreover, the parameter settings for the state-space model are simpler. The user needs
only to select the order, or the number of states, of the model. The order can come from
prior knowledge of the system. The user also can determine the order by analyzing the
singular values of the information matrix.
33

Fig. 2.13 plots the results when different system identification model structures
are used to obtain a linear model from frequency response data. All models are a third
order system and the options for each model are set to their default value. With these
conditions, a state space model structure gives a good result with one iteration. A similar
result can be obtained with the output error structure; however, it will require more
iterations and modification of the order of the model using some options the
identification algorithm provides. The ARX model, on the other hand, will give a good fit
at high frequencies, but at low frequencies the model generally will have some offset.
M
a
g
n
i
t
u
d
e



Frequency (Hz)
Fig. 2.13 Comparison system identification algorithm
The algorithm used to identify models from the measured FRFs can be described
in terms of three steps: pre-processing data, selecting the model structure, and setting
algorithm options. Data pre-processing refers to the creation of a data object or variable
that stores all frequencies in a column vector and stores magnitude and phase in a
complex variable. The object is formed by
Frfdata=idfrd(Mag,freq,Ts).
where the variable Mag is a column vector that stores the magnitude and phase in
complex form, the column vector freq stores all frequencies, and Ts is the sampling time.
For a continuous system, Ts is set to zero.
Once the data are stored in an idfrd object, other aspects of data processing, such
as filtering the data before the estimation, can be implemented. The data are filtered and
compressed using a local smoothing technique
10
3
10
4
10
5
10
-2
10
-1


State Space
ARX
Experimenatal data
Output error
34

Sfrfdata= spafdr(frfdata,Resol).
The argument Resol defines the frequency resolution over the frequency interval
of the data. The wider the resolution is, the smoother the estimate will be. The estimation
using a state-space model structure is obtained in MATLAB by
pem(Sfrfdata,order).
The last step during the identification process is to set the algorithm properties.
These properties serve to find the models that best fit the data. For instance, focus is one
of the most used properties in this work. It is used to apply a weighting to the fit between
the model and the data. For example,
G
o
=pem(Sfrfdata,4,'Foc',[10
3
,10
5
])
will fit the data into a four-order state-space model, concentrating the effort on the
data between 10
3
and 10
5
rad/s. A matlab M file with comments is given in the appendix.

2.5 Black-box Two-port Model of an Unregulated Buck Converter
In this section, a linear two-port black-box model for a buck converter with fixed
duty cycle and working in continuous conduction mode (CCM) is presented step by step.
The converter specifications are as follows: the nominal input and output voltages are 20
V and 5 V respectively, with 25 W output power and 38 kHz switching frequency. Fig.
2.14(a) and Fig. 2.14(b) show a picture and a schematic of the converter built out of
discrete components. If any white-box approach were employed to model the converter, it
would require taking apart the converter and modeling it carefully, component by
component, in order to obtain agreement when comparing the model with experimental
results. With two-port black-box models, on the other hand, all information to build the
model is obtained from FRFs measured at the converter terminals.
35


(a) Buck converter prototype (b) Buck converter prototype schematic
Fig. 2.14 Buck converter prototype 25 W 20V to 5V, photo by author.
The first step is to measure the four FRFs at the converter terminals,
audiosusceptibility G
o
(j), output impedance Z
o
(j), input admittance Y
i
(j), and back
current gain H
i
(j), using the setups shown in Figs. 2.5(a) and 2.5(b). Notice that, in this
case, the setup employed exhibits a weak interaction with the source and the load in the
frequency range of interest. The case in which strong interactions with the load and the
source are present is covered in Chapter 3.
The operating point for these measurements is chosen such that the converter is
working away from the borderline between continuous conduction mode (CCM) and
discontinuous conduction mode (DCM). For instance, when the input voltage of this
converter is 20 V, it will be working in CCM when the output current is between 1.7 A
and 5 A. The distance from the boundary between CCM and DCM is determined by the
amplitude of the disturbance injected in the circuit to perform the measurements. The
network analyzer setting for the measurements shown in Fig. 2.15 are listed in Table 2.1.
Table 2.1 Network Analyzer settings
Setting Description Value for the experiment
Start Start FRF measurement in Hz 50 Hz
Stop Stop FRF measurement in Hz 18 kHz
Num. of points Number of points stored logarithmic list 400 points
If bandwidth Set to 1/5 start frequency 10 Hz
Averaging Active 6
Power Disturbance magnitude (tuned with the linear amplifier) -3mdB
Display Dual (shows magnitude and phase in the same screen) dual
Input voltage Converter input voltage 20 V
Output current Converter output current 2 A



36


G
o
Z
o

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)

Y
i
H
i

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)

Frequency(rad/sec) Frequency(rad/sec)
Measurement

Fig. 2.15 Experimental FRFs
Two assumptions are made to model the converter using a simple linear two-port
network. First, the converter dynamic is mostly linear in the designated operating region,
which is determined by the converter input voltage and output current protections
settings. Second, the converter operates in continuous conduction mode operation. The
above assumptions need to be verified by performing a set of FRFs measurements at
different operating conditions covering the operating space in which the converter model
will be used.
Fig. 2.16 shows a set of FRFs measured at 20 V input voltage and changing the
output current from 2 A to 5 A in steps of 1A. The graphs clearly show that the FRFs are
invariant to changes in the output current. The measurements are repeated at 18 V and 22
V input voltage giving similar results. Thus, the assumption that the converter behaves
closely to a linear system in the designated operating region is verified, and a linear
network is appropriate for modeling the converter.


10
-2
10
0
10
3
10
4
10
5
-300
-200
-100
0
10
0
10
3
10
4
10
5
0
50
100
10
-1
10
0
10
3
10
4
10
5
-100
-50
0
50
10
-1
10
0
10
3
10
4
10
5
-200
-100
0
37

G
o
Z
o


Y
i
H
i


Fig. 2.16 Open-loop FRFs measured at different output current values
The second step in building a two-port black-box model is to select a set of four
measurements from the previous step, filter the data, and obtain a linear model. The
experimental FRFs G
o
(j), Z
o
(j), Y
i
(j)

and H
i
(j) only provide magnitude and phase
information. If the FRFs still contains noise that could affect the accuracy of the model
obtained via system identification, the data maybe filtered using a local smoothing
technique. The linear models are obtained from the experimental FRFs using a state space
model structure from the system identification toolbox in Matlab for which the order of
the model is the control parameter for this process. In this case, it is evident that a second
order model is a good initial choice. However, in general high order models are
preferable at this stage to obtain a good fit in all frequency ranges. Afterwards, a model
order reduction technique presented in section 2.6 can be used to remove the states that
are not dominant. Fig. 2.17 compares the models obtained through system identification
against the measured FRFs. As evinced, the resulting models overlap the measured FRFs
in the frequency range of interest. An example of how to obtain the linear models from
the FRFs is given in Appendix A.

0
5
10
2
10
3
10
4
-60
-50
-40
-30
-20
-10
0
Frequency (Hz)
Output
Current
M
a
g
n
i
t
u
d
e

(
d
B
)
0
5
10
2
10
3
10
4
-40
-30
-20
-10
0
Frequency (Hz)
Output
Current
M
a
g
n
i
t
u
d
e

(
d
B
)
0
5
10
2
10
3
10
4
-60
-40
-20
0
Frequency (Hz)
Output
Current
M
a
g
n
i
t
u
d
e

(
d
B
)
38


G
o
Z
o

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)

Y
i
H
i

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)

Frequency(rad/sec) Frequency(rad/sec)
Measurement Model

Fig. 2.17 Open loop FRFs from measurements and identified models
The final step is model validation. The most obvious and pragmatic way to
validate the model is to investigate how capable the model is of reproducing the system
behavior using a new set of data (the cross-validation data) that was not used to fit the
model. That is, both the model and the system are supplied with an input that is different
from the input used to obtain the estimation data set.
Since the model is built using frequency domain data, the model validation is
carried out in the time domain. In the time domain, measured input-output variables are
compared against variables predicted by the model when applying a load disturbance.
The load disturbance is implemented with an electronic load working in current sink
mode. The current step goes from 2.5 A to 4.5 A and back to 2.5 A, representing a step
load from 50% to 90%. To have a fair comparison between the model and the hardware,
the current step produced by the electronic load is recorded in an electronic file and then
is used in the simulation as the step load for the model.
The output voltage and input current transient response are compared with those
obtained from the black-box terminal characterization model in Fig. 2.18. The figure
shows a good match during transients and steady state. Thus, the models obtained using
10
-1
10
0
10
3
10
4
10
5
-200
-100
0
10
-1
10
0
10
3
10
4
10
5
-100
-50
0
50
10
0
10
3
10
4
10
5
0
50
100
10
-2
10
0
10
3
10
4
10
5
-300
-200
-100
0
39

the proposed methodology are able to predict the converter dynamic behavior from the
input and output power terminals with a degree of accuracy not observed using any other
existing behavioral modeling method.
Graphical comparison for model validation is good enough in this case since
uncertainties on the experimental measurements of the voltages and currents were found
to be small, in the order of 1 and 2% for the voltage and current, respectively. Model
parameter uncertainties are also expected to be small since the model is obtained from
measurements at the converter terminals. However, a formal verification and validation
procedure needs to be developed for this model.





Input Voltage v
1
Output Voltage v
2

V
o
l
t

Input Current i
1
Output Current i
2

A
m
p

Time(msec) Time(msec)
Experimental , Experiemental Average , Model
Fig. 2.18 Open-loop FRFs from measurements and identified models
8 8.5 9 9.5 10 10.5 11 11.5
3
20
20.2
20.4
20.6
20.8
8 8.5 9 9.5 10 10.5 11 11.5
4.4
4.6
4.8
5
5.2
8 8.5 9 9.5 10 10.5 11 11.5
3
0.5
1
1.5
2
2.5
8 8.5 9 9.5 10 10.5 11 11.5
3
2.5
3
3.5
4
4.5
40

The methodology to obtain a linear black-box model is summarized by the
flowchart illustrated in Fig. 2.19.

Fig. 2.19 Flow chart for black-box modeling
2.6 Model Order Reduction by Means of Balanced Realization and
Singular Perturbations
After obtaining a set of frequency responses, system identification is employed to
obtain the actual transfer functions used in the model. During the identification process, it
is very difficult to determine which modes could be neglected to reduce the order of the
system, which means that high-order models achieving a close fit to the experimental
data are preferred at this stage. A second step may then be applied to reduce the order of
the models obtained.
Existing order reduction methods like singular perturbations [9] are well suited
for white box modelswhere all parameters are known. In this case, it is known that
small valued parameters like parasitic inductances and capacitances are related to the fast
states, while energy storage elements are related to the slow states. For black-box models
this method cannot be directly applied, since in this case sometimes there is not an
obvious physical relationship between the identified model and the converter parameters.
To counter this, a balanced state-space realization may be used to represent the
identified model of the converter [15], which enables the use of singular
41

perturbationseliminating fast state dynamicsby reordering the system states
according to their intrinsic dynamics. Let us consider for instance a linear time invariant
system defined by the following equations
x(t) Ax(t) Bu(t) = + y(t) Cx(t) = . (2)
The controllability and observability gramian are defined as
T
At T A t
c
0
W e BB e dt

=
(18)
T
A t T At
o
0
W e C Ce dt

=

.
(3)
The System is considered balanced if the matrices are equal and diagonal, thus
c o 1 2 n
W W diag( , .... ) = = . (4)
The grammians are an unique solution to the Lyapunov equations
T T
c c
AW WA BB 0 + + = (5)
T T
o o
A W WA C C 0 + + = . (6)
Thus, from (5) and (6), if the controllability and observability grammians for an
linear time invariant model (LTI) system exist, we can say the grammians are positive
definite and the eigenvalues of A have a negative real part. If unbalanced, the system
may be balanced by using a similarity transformation where :

1
A' T AT

=
1
B' T B

= C' CT = (7)
~ ~
x(t) A' x(t) TB' u(t) = +
~
1
y(t) C' T x(t)

=
.
(8)
With the similarity transformation given by (7), the Lyapunov equations in the
new coordinates are given by
1 1
1 T 1 T T T
c c
A' T WT T WT A' B' B' 0


+ + = (9)

T T T T
o o
A' T W T T W TA' C' C' + + . (10)
Therefore, the controllability and the observability grammians in the new coordinated
system are given by
42

1
1 T
c c
W ' T WT

= (11)
T
o o
W ' T W T = (12)
And the eigenvalues of
c o
W ' W ' are invariant to similarity transformations

1
1 T T 1 T 2
c o c o c o
W ' W ' T WT T WT T WWT

= = = . (13)
Thus the problem is reduced to find a matrix T such that

1
c o 2
n
0 0
W ' W ' 0 0
0 0


= =



(14)
And having the relations
1
1 T T 2
c o c o
W ' W ' T WT T WT

= = (15)
1 2
c o c o
W ' W ' T WW T

= =
.
(16)
Since
c
W is positive definite, it can be factored using cholesky factorization as

T
c
W RR = (17)
writing the equation (16) as
1 T 2
c o o
W ' W ' T RR WT

= =
,
(18)
which is equivalent to

1 T 2
o c o
W ' W ' T RW R T

= = (19)
and
T 2 T
o
RW R T T =
.
(20)
Then (20) is similar to
T 2 T
o
RW R U U =
,
(21)
and solving for

1 1
T T
2 2
o
U RWR U = (22)
43

defining the nonsingular matrix as

1
T
2
T R U

= and
1
T T
2
T U R

= , (23)
finally, it is shown that
T T
1 1 1 1
1 1 T 1 1
2 2 2 2
c c c
W ' T WT U R WR U

= = = = (24)
1 1 1 1
T T T 2
2 2 2 2
o o o
W ' T WT U RWR U

= = = = . (25)
After which, the system may be partitioned as
11 12
21 22
A A
A'
A A

=



1
2
B
B'
B

=



[ ]
1 2
C' C C =
1
c o
2
0
W ' W '
0

= = =

(26)
where, if a time scale separation between the state variables is present, the values
for
2
in (26) are much smaller than
1
, indicating that the states related to
2
are less
controllable and observable than the states associated to
1
, and thus would have a small
impact on the overall system dynamic behavior [16].
From a singular perturbation perspective, these correspond to the fast, negligible
states. The states associated with
1
in turn are called the slow manifold or dominant
states of the system [16]. The corresponding reduced order model is thus given by:
r r r r
r r r
x (t) A x (t) B u(t)
y (t) C x (t)
= +
=

(27)
where A
r
B
r
and Cr are defined as follows
1
r 11 12 22 21
1
r 11 12 22 21
A A A A A
B B B B B

=
=

1
r 11 12 22 21
C C C C C

= . (28)
This model order reduction method can be implemented with the algorithm
presented in [17], or directly by the functions balreal and modred that are part of the
control system toolbox of Matlab. An example of this methodology is given using the
black-box models developed for the bus converter E48SB12020NRFA. Table 2.1
summarizes the model order reduction obtained in this case.
44

Table 2.2 Model Order Reduction
Full Order Model Low Order Model
G
o
8 2
Z
o
4 2
Y
i
8 8
H
i
6 2
Z
s
11 1

Many of the additional poles and zeroes of the full order models are used to fit the
data at high frequencies that denotes the effect of small parasitic capacitances and
inductances that may not play an important role on the converter dynamics. Therefore, as
observed in Fig. 2.20, in the frequency domain the full order and reduced order models
differ at high frequencies while presenting a perfect match at low frequencies.
Z
o
Full order vs. Z
o
Low order H
i
Full order vs. H
i
Low order
M
a
g

(
A
b
s
)


P
h
a
s
e

Frequency (rad/sec) Frequency (rad/sec)
Full order model, Reduced order model
Fig. 2.20 Full order and low order FRF comparison
Both the high and the low order models are used to predict the transient response
shown when two load steps are applied at the converter terminals using an electronic
load, then the input current and output voltage predicted by the models are subtracted
from the each other in order to show the prediction error of the low order model as
depicted in Fig. 2.21.






10
0
10
4
10
5
10
6
-200
-100
0
10
-2
10
-1
10
0
10
4
10
5
10
6
-100
-50
0
50
45

Input Current Low Order Model Prediction Error Output Voltage Low Order Models Prediction Error
A
m
p
s

V
o
l
t
s

Time (msec) Time (msec)
Fig. 2.21 Prediction error between full and low order models

These results show that in steady state there is a good agreement between the
models and during transients as expected they present small differences. When deciding
whether to use low order models or higher order models to simulate a big system, one
must address a trade off between simulation speed and accuracy.

2.7 Conclusions
This chapter has presented the derivation of a two-port black-box model for dc-dc
converters using hybrid parameters and system identification methods. The results
obtained from the model showed very good agreement with experimental results in both
frequency and time domains for open-loop and closed-loop cases, thereby validating the
proposed methodology. The advantage of this modeling approach is that it requires no
knowledge of the internal structure of the converter or any of its parameters; nonetheless,
the transfer functions that are measured and used to build the model include accurate
values of the parasitic inductances, capacitances, and resistances in the circuit, all of
which are generally neglected when building average models. Furthermore, the two-
terminal-port nature of this model makes it appropriate for building cascaded and
paralleled converter structures, which should allow simple and straightforward simulation
of distributed power electronics systems consisting the converters sold by various
vendors, each converter having different functionalities.

4 4.1 4.2 4.3 4.4 4.5 4.6
-0.02
-0.01
0
0.01
0.02
4 4.1 4.2 4.3 4.4 4.5 4.6
-0.1
-0.05
0
0.05
0.1
46

2.8 References
[1] A. Emadi, "Modeling and analysis of multiconverter DC power electronic systems using the
generalized state-space averaging method," IEEE Transactions on Industrial Electronics vol.
51, pp. 661-668, 2004.
[2] A. Davoudi and J. Jatskevich, "Realization of parasitics in state-space average-value
modeling of PWM dc-to-dc converters," IEEE Transactions on Power Electronics, vol. 21,
pp. 1142-1147, 2006.
[3] H. Mikko, K. Matti, and T. Suntio, "EMI-filter interactions in a buck converter," in
Proceedings of the 12th International Power Electronics and Motion Control Conference
2006, pp. 54-59.
[4] R. Jayabalan, B. Fahimi, A. Koenig, and S. Pekarek, "Applications of power electronics-
based systems in vehicular technology: state-of-the-art and future trends," in Power
Electronics Specialists Conference, 2004. PESC 04. 2004 IEEE 35th Annual, 2004, pp.
1887-1894 Vol.3.
[5] C. Ju-Yeop, B. H. Cho, H. F. VanLandingham, M. Hyung-soo, and S. Joong-Ho, "System
identification of power converters based on a black-box approach," IEEE Transactions on
Circuits and Systems, vol. 45, pp. 1148-1158, 1998.
[6] F. Alonge, F. D'Ippolito, F. M. Raimondi, and S. A. T. S. Tumminaro, "Nonlinear modeling
of dc-to-dc converters using the hammerstein's approach," IEEE Transactions on Power
Electronics, vol. 22, pp. 1210-1221, 2007.
[7] K. T. Chau and C. C. Chan, "Nonlinear identification of power electronic systems," in IEEE
Power Electronics and Drive Systems Conference, 1995, pp. 329-334 vol.1.
[8] L. Ljung, "State of the art in linear system identification: time and frequency domain
methods," in American Control Conference, 2004. Proceedings of the 2004, 2004, pp. 650-
660 vol.1.
[9] L. Ljung, System identification (2nd ed.): theory for the user: Prentice Hall PTR, 1999.
[10] L. Ljung, "Integrated frequency-time domain tools for system identification," in Proceedings
of the 2004 American Control Conference, 2004, pp. 5594-5599
[11] B. Swaminathan and V. Ramanarayanan, "Application of network analyzer in measuring the
performance functions of power supply," Journal of the Indian Institute of Science, vol. 86,
p. 315, 2006.
[12] Y. Qiu, M. Xu, K. Yao, J. Sun, and F. C. Lee, "Multifrequency Small-Signal Model for
Buck and Multiphase Buck Converters," IEEE Transactions on Power Electronics vol. 21,
pp. 1185-1192, 2006.
[13] G. C. Verghese and V. J. Thottuvelil, "Aliasing effects in PWM power converters," in The
IEEE 30
th
Power Electronics Specialists Conference, 1999, pp. 1043-1049 vol.2.
[14] K. Zenger, A. Altowati, and T. Suntio, "Dynamic properties of interconnected power
systems a system theoretic approach," in The 1
ST
IEEE Conference on the Industrial
Electronics and Applications, 2006, pp. 1-6.
[15] P. Kokotovic and P. Sannuti, "Singular perturbation method for reducing the model order in
optimal control design," IEEE Transactions on Automatic Control, vol. 13, pp. 377-384,
1968.
[16] B. Moore, "Principal component analysis in linear systems: Controllability, observability,
and model reduction," IEEE Transactions on Automatic Control vol. 26, pp. 17-32, 1981.
[17] S. K. Singh, S. K. Nagar, and J. Pal, "Balanced Realized Reduced Model of Non-Minimal
System with DC Gain Preservation," in the IEEE International Conference on Industrial
Technology, 2006, pp. 1522-1527.

3 Chapter
Subsystem Interaction Effects on FRF
Measurements
This chapter analyzes how the subsystem interactions among the converter, the
source, and the load dynamics affect the measurement of frequency response functions
and also investigates how these phenomena can be synthesized into mathematical
expressions in terms of the measured FRFs and the desired unterminated FRFs. This
analysis provides an approach to obtain unterminated FRFs, effectively canceling out
non-converter dynamics from the measurements.
3.1 Introduction
In the previous chapter, it was demonstrated that a converter can be modeled in a
specific operating region by using four transfer functions, namely audiosusceptibility,
output impedance, input admittance, and back current gain, denoted respectively by G
o
,
Z
o
, Y
i
, and H
i
. These transfer functions are considered unterminated because they
reflect only the internal dynamics of the associated converter [1, 2]. However, in practice
FRF measurements are coupled with the source and the load dynamics [3]. Consequently,
the measurements will not reflect the true internal dynamic of the converter. This not
only adds a limitation to this modeling approach but also affects the measurement of any
FRF, including loop gains [4].
To address the problem above, the use of a measurement setup designed to have a
weak interaction with the source and the load has been proposed in [5]. A weakly coupled
condition is achieved up to certain frequency by feeding the converter from a low output
impedance voltage source and an electronic load working as a constant current sink. With
48

this setup, accurate models were obtained from FRF measurements conducted up to 40
kHz. However, the problem of cases in which the frequency sweeps beyond 40 kHz were
needed has not yet been addressed; in these cases the source output impedance and
electronic load input impedance become important. These factors affect the
measurements considerably.
This chapter presents an enhanced FRF measurement procedure in which
unterminated frequency responses can be obtained. Two cases are considered; in the first
case, only the source dynamic is coupled with the converter dynamics, and the load is
considered ideal in the frequency range of interest. The second case is general, where
both the source and load dynamics are coupled with the converter dynamics. Hence, the
source and load dynamics need to be removed from the measurements. The general
approach taken to obtain the unterminated frequency responses from the coupled
measurements is to obtain mathematical expressions of the coupled FRFs and then
calculate the decoupled FRFs.
Once the set of decoupled FRFs is calculated, system identification is used to
obtain actual transfer functions. The order of the resulting transfer functions depends on
how closely they fit the experimental data. Therefore, high-order transfer functions are
not unusual. This is a problem since the models are intended for system-level analysis,
and they are expected to have a large number of interconnected converters. Therefore,
they need to be computationally efficient. To cope with this need, this chapter also
presents a model order reduction methodology for a black-box converter model that is
obtained from the full-order model using a balanced state-space realization and singular
perturbations. The reduced-order models obtained in this way allow for fast simulation of
large interconnected systems with minimal loss of information.


49

3.2 Subsystem Interaction Affecting Measurements of Frequency

Response Functions
The analysis is carried out first by considering only the effect of the source and
neglecting the load by assuming it is a constant current sinkthis assumption is valid
depending on the bandwidth of the electronic load used. The second step is to address the
general case where both the source and the load dynamics are taken into consideration.
For the first analysis, a unregulated buck converter prototype rated at 25 W, 40 kHz, and
20 V input is used.
Fig. 3.1 shows the same black-box parameter model structure but in the form of a
block signal diagram. This is more convenient for visualizing the feedback paths
generated when considering the source and load effect.
G
o
Z
o
Y
i
H
i
Z
s
+
+
+
-
io
i
2
v
2
v
1
V
s
+
-
i
1

Fig. 3.1 Converter black-box two signal diagram loaded with a current sink
Specifically, when perturbing v
1
, i
2
is considered constant in the frequency range
of interest; therefore, G
o
and Y
i
are not affected by the presence of the source output
impedance. Thus
G
om
= G
o
and Y
im
= Y
i
, (1)
where the sub index
m
indicates the measured quantity. On the other hand, when
perturbing i
2
there is a contribution from v
1
and i
2
on v
2
through G
o
and Z
o
respectively,
where v
1
= i
1
Z
s
. Hence, what the network analyzer measures instead of H
i
and Z
o
at the
converter terminals is (2) and (3).
1 i
im
2 s i
i H
H
i 1 Z Y
= =
+

(2)
2 0 i s
om o
2 s i
v
G H Z
Z Z
i 1 Z Y

= = +
+
(3)
50

As seen in (2), (3), the FRFs measurements at the converter terminals are coupled
not only coupled with the source but also with its other internal transfer functions. This
phenomenon is illustrated in Fig. 3.2, where an average model of the buck converter
prototype used in Chapter Two has been linearized repetitively at a specific operating
point for different source output impedance (Z
s
) values ranging from 0 to 50 H, showing
how the frequency responses of H
i
differ in each case in agreement with (2).

M
a
g
(
A
b
s
)



P
h
a
s
e

(
D
e
g
)

Fig. 3.2 Source output impedance impact on H
im
(from average models)
The presence of the inductor in series with the ideal source creates a double pole
that moves towards the origin as the inductor value increases. However, since Y
i
and Z
s

can be obtained from direct measurements as shown in (1), H
i
and Z
o
can be calculated
out from H
im
and Z
om
, as shown in (4) and (5) using (2) and (3).
( )
i im s i
H H 1 Z Y = +

(4)
0 i s
o om
s i
G H Z
Z Z
1 Z Y

=
+
(5)
The simulation result presented above agrees with the experimental measurement
H
im
plotted against the calculated unterminated FRF H
i
in Fig. 3.3. In this experiment, an
8.2 H inductor was placed in series with the source. The double pole predicted by the
average model appears in the measurement as well, evincing the occurrence of the
interaction between the source output impedance and the converter input admittance.



Frequency (rad/sec)
10
0


10
3
10
4
10
5
-360
-270
-180
-90
0
1uH
50uH 40uH
30uH
20uH
10uH
51

H
im
v.s H
i

M
a
g
(
A
b
s
)


P
h
a
s
e

(
D
e
g
)

Frequency(rad/sec)
H
im
(measured FRF) H
i
(unterminated FRF)
Fig. 3.3 Source output impedance effect H
im
(experimental)
The effect of Z
s
on the converter output impedance measurement Z
om
is presented
in Fig. 3.4, where the experimental measurement of Z
om
is plotted against the calculated
unterminated output impedance Z
o
. As predicted by (3), Z
om
is affected mostly at low
frequencies by the source output impedance reflected at the converter output terminals, as
at higher frequencies the output impedance dynamic is governed solely by the converter
output filter.
Z
om
v.s Z
o

M
a
g
(
A
b
s
)


P
h
a
s
e

(
D
e
g
)

Frequency(rad/sec)
Measured FRF Unterminated FRF
Fig. 3.4 Source output impedance effect on FRF measurements
The effect of Z
s
in this case looks small in the frequency domain; however, even
these small differences can affect the prediction accuracy of the black-box models. For
instance Fig. 3.5 shows the experimental and simulated output voltage and input current
10
0
10
-2
10
-1
10
33
3
10
4
10
-100
-50
0
50
0
10
10
-2
10
3
10
4
10
5
-400
-200
0
52

transient response produced by a load step. In this case, the black-box converter model is
built using the terminal measurements H
im
and Z
om
. The differences are noticeable as
observed. For instance, the natural frequency response of the input current predicted by
the models differs from the experimental data. Also, the output voltage presents a steady
state offset when compared with the measured output voltage.
Converter Input Current Converter Output Voltage
A
m
p
s

(
A
)


V
o
l
t
s


(
V
)


Time(msec) Time(msec)
Experimental Experimental (average) Model
Fig. 3.5 Experimental and simulated transient response using measured FRFs
When, on the contrary, the model is built using H
i
and Z
o
calculated from the
measurements, the predicted input current and output voltage perfectly match the
experimental transient waveforms as shown in Fig. 3.6.
Converter Input Current Converter Output Voltage
A
m
p
s

(
A
)

V
o
l
t
s


(
V
)

Time(msec) Time(msec)
Experimental Experimental (average) Model
Fig. 3.6 Experimental and simulated transient response using unterminated FRFs
The general case considers the effect of the source output impedance Z
s
and the
load admittance Y
L
simultaneously. Fig. 3.7 shows the signal diagram where a current
disturbance i
d
is injected at the converter output.
8 9 10 11
4.4
4.6
4.8
5
5.2
8 9 10 11
0.6
0.8
1
1.2
1.4
1.6
8 9 10 11
4.4
4.6
4.8
5
5.2
8 9 10 11
0.6
0.8
1
1.2
1.4
1.6
53

G
o
Z
o
Y
i
H
i
Z
s
+
+
+
-
i
d
i
2
v
2
v
1
V
s
+
-
i
1
Y
L
+
+
+
i
d

Fig. 3.7 Converter black-box signal diagram loaded with a resistor
A mathematical expression for the back current gain H
im
is obtained from the
signal diagram shown in Fig. 11 where the current i
1
is defined in (6)
1 i i 2 i
i v Y i H = +
. (6)
Replacing expression v
1
= -i
1
Z
s
and i
2
= v
2
Y
L
+ i
d
on (6) where Y
L
is the load
admittance and i
d
is the injected current disturbance
( )
1 1 s i 2 L d i
i i Z Y v Y i H = + +
.
(7)
An expression for v
2
in terms of i
1
and i
d
is obtained as in (8)
( )
2 1 s o d 2 L o
v i Z G i v Y Z = +
.
(8)
Solving for v
2
,
( ) ( )
1 s o d o
2
L o L o
i Z G i Z
v
1 Y Z 1 Y Z
=
+ +
(9)
and replacing (9) into (7) and solving for
1
d
i
i
yields,
( )
( )
i
L o
1
im
d
s o L i
s i
L o
H
1 Y Z
i
H
i
Z G Y H
1 Z Y
1 Y Z



+

= =

+ +


+

.
(10)
Similarly, expressions for G
om
, Y
im
and Z
om
can be derived and are given below.
( )
o
om
L o
G
G
1 Y Z
=
+
(11)
( ) 1
= +
+
o L i
im i
L o
G Y H
Y Y
Y Z
(12)
( )
oeq
2
om
d L oeq
Z
v
Z
i 1 Y Z
= =
+

(13)

54

With

( )
i s o
oeq o
s i
H Z G
Z Z
1 Z Y
= +
+
(14)
The resulting expressions are highly coupled; however, expressions for the
unterminated transfer functions in terms of the measured quantities can be obtained as
follows: H
i
and G
o
are obtained from (10) and (11) and replaced in (14).
( ) ( )
( )
2
s om im L o s im
oeq o
s i
Z G H 1 Y Z 1 Z Y
Z Z
1 Z Y
+ +
= +
+
(15)
Then solving (15) for Z
o
, the quadratic equation (16) is obtained. In this equation,
all the coefficients are in terms of the measured quantities H
im
, Y
im
, Z
om
, and G
om
, except
for Y
i
, which is an unterminated FRF.
( )
( )
( )
( )
( )
( )
2 2 s o m i m s o m i m
L s i m o L s i m o
s i s i
s o m i m
s i m o e q
s i
Z G H Z G H
Y 1 Z Y Z 1 2 Y 1 Z Y Z
1 Z Y 1 Z Y
Z G H
1 Z Y Z 0
1 Z Y

+ + + +


+ +


+ + =


+


(16)
To solve equation (16), Y
i
needs to be obtained. This is not a problem, since it can
be determined from the previous case; that is, when the converter is loaded with an
electronic load and Y
im
= Y
i.
Thus, to solve the equation an additional measurement is
performed beforehand using an electronic load in current sink mode.
To show the effectiveness of this approach, the same buck converter prototype
used above and in the first chapter is used here to develop an unterminated black-box
model for the case in which the FRFs measured at converter terminals are coupled with
source and load dynamics. In this case the load is a 2 resistor.
Fig. 3.8 shows the coupled FRFs in red and the unterminated FRFs calculated
using equations (16) , (11), (12) and (13) in blue. It can be noticed there is not much
difference between the two. This implies that, in this case, there exists a low interaction
among the source, the converter, and the load.
G
o
Z
o

55

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)

Y
i
H
i

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)

Frequency(rad/sec) Frequency(rad/sec)
Unterminated FRFs Coupled FRFs

Fig. 3.8 Removing Source and load dynamics from measured FRFs
Following the procedure outlined on the previous chapter, system identification
tools are used to obtain a linear time invariant model from each measurement. Then a
linear hybrid two-port network is implemented in Simulink Matlab. Simulation and
experimental results are compared to explore the prediction capabilities of the model.
In previous cases, the experimental output current was used as the load
disturbance for the model. In this case, the test is more challenging given that all the
simulation data is generated by the models. To achieve this goal instead of using a
constant current sink as a load as in previous cases, the converter is loaded with two 1
resistors connected in series. At a certain time t the switch is activated to produce a
load step from 50 % to 100% of the converter nominal load as illustrated in Fig. 3.9.



Fig. 3.9 Experimental setup for model validation using resistor load, photo by author.

10
3
10
4
10
5
10
-2
10
0
p
om o
10
3
10
4
10
5
-200
-100
0
10
3
10
4
10
5
10
-2
10
-1
10
0
om o
10
3
10
4
10
5
100
-50
0
50
10
3
10
4
10
5
10
0
10
2
im i
10
3
10
4
10
5
0
50
100
(
g
)
10
3
10
4
10
5
10
0
A
m
p
l
i
t
u
d
e
im i
10
3
10
4
10
5
-200
-100
0
(
g
)
F ( d/ )
56

To accurately predict the behavior of the system, it is necessary not only to have a
good converter model but also good models for the source output impedance, and the
load. Therefore, black-box models of the source and load are also obtained. Fig. 3.10
compares experimental data against simulation results during the transient.
The developed black-box models are able to predict with high level of accuracy
the steady state and transient response of the system. But, more importantly this
experiment demonstrates that the transient response at converter terminals does not
depend only on the converter dynamics but on the subsystem interactions among the
source, the converter, and the load dynamics. Such interaction is also present when
unterminated black-box models are interconnected as demonstrated in equations (7) and
(8). Therefore, even though the models of each subsystem are obtained individually, the
two-port nature of the models provides a path for the creation of the coupling terms that
appear in real life when two subsystems are interconnected.

Input Voltage v1 Output Voltage v2
V
o
l
t

(
V
)


Input Current i1 Output Current i2
A
m
p

(
A
)

Time(msec) Time(msec)
Experimental , Experiemental Average , Model
Fig. 3.10 Experimental and simulated transient response during a load step using resistors
6.5 7 7.5 8 8.5 9 9.5 10
19.9
20
20.1
20.2
20.3
20.4
g
(
)
6.5 7 7.5 8 8.5 9 9.5 10
4.4
4.6
4.8
5
5.2
g
(
)


6.5 7 7.5 8 8.5 9 9.5 10
3
0.6
0.8
1
1.2
1.4
1.6
Ti (S )
M
a
g
(
A
m
p
s
)
p
6.5 7 7.5 8 8.5 9 9.5 10
2.5
3
3.5
4
4.5
g
(
p
)
57

The previous experiment and simulation demonstrated that it is possible to

remove the source and load dynamics from FRFs measured at converter terminals.
However, it fails to help the reader to visualize how important this procedure is when
trying to model commercially available converters for which strong interaction among
the system components are expected mainly because these converters work with high
switching frequencies. Thus, FRF measurements up to 300 kHz are common, and the
assumptions made in chapter 2 regarding having a nearly perfect current sink load and
low interaction with the source are no longer valid.
To show the effectiveness of this approach, a commercially available
unregulated bus converter E48SB12020NRFA is used in the following example. This
converter is rated at 48V input, 12 V output, and 240 W, with a switching frequency of
400 kHz i. This converter is commonly found in intermediate bus architecture. In this
case, the FRF measurements need to be performed until 200 kHz. At this frequency, the
assumptions made in Chapter Two, where the source and the load are ideal, are no longer
valid as evinced in Fig. 3.11, where Fig. 11(a) is the source output impedance and Fig.
11(b) is the electronic load input admittance. In the output impedance plot, a resonant
point is observed around 4x10
3
rad/sec. After this point, the output impedance grows
linearly with the frequency. The electronic load input admittance on the other hand is
small at low frequency, accordingly with the assumptions made in Chapter Two.
However, the electronic load input admittance becomes a significant factor at around
1x10
5
rad/sec and presents a resonant point with a magnitude of 20dB at 5.5x10
5
rad/sec.
A
m
p
l
i
t
u
d
e

(
d
B
)

A
m
p
l
i
t
u
d
e

(
d
B
)

P
h
a
s
e

(
D
e
g
)

P
h
a
s
e

(
D
e
g
)


a) Source Output Impedance Zs b) Load input admmitance
Fig. 3.11 Source and load dynamics
40
-40
-20
0
20
1
3
1
4
1
5
1
6
-90
0
90
10 10
6
-45
-30
-20
-10
10
3
10
4 5
0
45
90
58

To observe how the source output impedance and the electronic load input
admittance affect the measurements, both the coupled and unterminated frequency
responses of the converter output impedance are plotted side by side in Fig. 3.12


A
m
p
l
i
t
u
d
e

(
d
B
)

A
m
p
l
i
t
u
d
e

(
d
B
)

P
h
a
s
e

(
D
e
g
)

P
h
a
s
e

(
D
e
g
)

a) Measured converter output impedance b) Unterminated converter output impedance
Fig. 3.12 Source and load effect on FRF measurements
In Fig. 3.12(a), the effect of the source is clearly identified at low frequencies, as
is the effect of the load admittance at high frequencies. To the right, Fig. 3.12(b) shows
the unterminated FRF measurements that, in essence, only capture the converter
dynamics.
The complete converter model is developed using 1 resistor load instead of
using the electronic load. There is not a special technical reason for this; the same model
could be obtained using the electronic load and removing the load and source dynamics
as shown above. However, using a resistor as a load is a more general and less expensive
approach that renders the same results as above.
Once the unterminated measurements are calculated using the described
procedure, system identification tools are used to obtain the corresponding transfer
functions of these indirect measurements. Fig. 3.13 shows the measured FRFs and
unterminated FRFs. With these, a two-port black-box model is finally built using the
calculated G
o
, Z
o
, Y
i
, and H
i
transfer functions.




0
-40
-20
-90
-45
10
3
10
4
10
5
10
6
0
45
0
-40
-20
10
3
10
4
10
5
10
6
-90
-45
0
45
59

Gom vs. Go Zom vs. Zo
M
a
g
(
A
b
s
)

P
h
a
s
e

Yim vs. Yi Him vs. Hi
M
a
g
(
A
b
s
)

P
h
a
s
e

Frequency(rad/sec) Frequency(rad/sec)

Unterminated FRFs Coupled FRFs

Fig. 3.13 FRFs measured using a resistor as a load and unterminated FRFs

The model validation is performed in the time domain in the region where the
converter behaves closely to a linear system. For this, three load steps from 10 A to 20 A,
then down to 15 A, and back to 10 A are applied at the output terminals using an
electronic load working in constant current sink mode. Fig. 14 shows the experimental
and simulated transient response. As seen, the match between waveforms is nearly
perfect, where the blue trace is the measured waveform, the light green trace is the
moving average of the measured waveform, and the red trace is the simulated one.





1 1 1 1
6
10
0
3 4 5
-300
-100
0
2
10 10 10 10
6
10
0
10
3 4 5
0
50
100
-100
10
0
10
-2
10
3
10
4
10
5
10
6
0
100
10 10 10 10
6
10
0


3 4 5
-200
-100
0
-200
60

Input Voltage Output Voltage

Input Current Output Current

Time (msec) Time (msec)

Experimental , Experiemental Average , Model
Fig. 3.14 Experimental and simulated transient response during three load steps


3.3 Conclusions
This chapter has presented a methodology to obtain unterminated frequency
response measurements used for the construction of a two-port converter black-box
model. Two approaches were analyzed. The first approach is simpler because it take into
account the effect of the source and neglects the effect of the load by assuming the load
as a perfect current sink. This assumption is valid and renders good unterminated FRFs
up to 80 kHz. When higher frequency sweeps are needed, the second method is
advantageous. In this case, the converter is loaded with a passive load, then the source
and load dynamics are removed from the measurements. With this approach,
unterminated FRF measurements up to 200 kHz have been demonstrated to produce good
models.
4 4.1 4.2 4.3 4.4 4.5 4.6
x 10
-3
46
47
48
49
50
Time(Sec)
V
o
l
t
s
4 4.1 4.2 4.3 4.4 4.5 4.6
x 10
-3
9
10
11
12
13
14
Time(Sec)
V
o
l
t
4 4.1 4.2 4.3 4.4 4.5 4.6
x 10
-3
2
2.5
3
3.5
4
4.5
5
5.5
Time(Sec)
A
m
p
4 4.1 4.2 4.3 4.4 4.5 4.6
x 10
-3
10
12
14
16
18
20
Time(Sec)
A
m
p
61

3.4 References
[1] C. Byungcho, B. H. Cho, and H. Sung-Soo, "Dynamics and control of dc-to-dc
converters driving other converters downstream," IEEE Transactions on Circuits and
Systems I, vol. 46, pp. 1240-1248, 1999.
[2] C. Byungcho, K. Jaeyeol, B. H. Cho, C. Seungwon, and C. M. Wildrick, "Designing
control loop for dc-to-dc converters loaded with unknown AC dynamics," IEEE
Transactions on Industrial Electronics vol. 49, pp. 925-932, 2002.
[3] L. Arnedo, D. Boroyevich, R. Burgos, and F. Wang, "Unterminated frequency
responce measurements and model Order reduction for black-box terminal
characterization models," in The 23
th
IEEE Applied Power Electronics Conference
and Exposition Austin, Texas, 2008, pp. 1054-1060.
[4] A. Altowati, T. Suntio, and K. Zenger, "Input filter interactions in multi-module
parallel switching-mode power supplies," in IEEE Industrial TechnologyConference,
2005, pp. 851-856.
[5] L. Arnedo, R. Burgos, F. Wang, and D. Boroyevich, "Black-box terminal
characterization modeling of dc-to-dc converters," in The 22
th
IEEE Applied Power
Electronics Conference, Anaheim, California, 2007, pp. 457-463.

4 Chapter
Nonlinear Black-box DC-DC Converter Model

4.1 Introduction
Experimental and simulation results have shown that a linear model can be used
to represent the behavior of an open-loop buck dc-dc converter with a high degree of
accuracy in a wide operating region. Nevertheless, it is known that dc-dc converters are
nonlinear in nature; therefore, linear black-box models are useful as long as the
nonlinearities are mild. For converters with stronger nonlinearities, such as boost and
buck-boost topologies, a linear two-port network is definitely not the best structure for
the black-box model. Another source of nonlinearities comes from the fact that the vast
majority of the applications need a regulated input voltage; therefore, many dc-dc
converters regulate their output voltage using a PI (proportional-integral) or PID
(proportional-integral-derivative) feedback controller. The controller completely changes
the converter dynamic behavior, and, if examined from the input terminals, the converter
behaves as a constant power load. Again, a linear structure will fail in this case.
To cope with the problem above, this work proposes two different nonlinear
structures for the two-port black-box model. The first structure is based on Hammerstein-
Weiner model composed by the series connection of a linear dynamic subsystem with a
nonlinear function. If the non-linearity follows the linear subsystem, the system is called
a Wiener type; otherwise, the system is called a Hammerstein type. This structure is well
suited for cases in which static nonlinearities are dominant, such as the case of a
regulated buck converter or unregulated boost and buck-boost dc-dc converters, but fails

63

in the case of converters with stronger nonlinearities. For instance, when used to model a
regulated boost or buck-boost converters, the resultant model prediction is not accurate.
The second black-box structure is known as a polytopic model. Polytopic models
assume that a nonlinear system can be modeled by a family of local models that are valid
at different operating conditions, which are then combined by means of nonlinear smooth
interpolation to predict the modeled system behavior. Depending on the instantaneous
operating conditions, the model transits smoothlyusing nonlinear interpolationfrom
one linear model to an adjacent one corresponding to the new operating point.
Numerous applications of the Hammerstein-Wiener structure exist for nonlinear
electro-mechanical systems [1], nonlinear biological systems [2], electrochemical
systems [3, 4], and other applications. Hammerstein-Wiener models for dc-dc
converters have been reported in [5, 6] in which a duty cycle to output voltage nonlinear
plant model was developed using a Hammerstein model structure. Then, based on this
model, a robust controller was designed assuming the nonlinear static function as a model
of parameter uncertainty. It is important to emphasize that no attempt has been made
previously to model a converter from its power terminals using a Hammerstein-Wiener
structure.
The polytopic multi-model approach also has been studied extensively as evinced
by the relevant literature; although they all have similar mathematical structure, they have
been given different names. For instance, in the fuzzy logic community, the approach is
called a Takagi-Sugeno model [7], [8] in which each local model is represented as a
fuzzy set and the model is constructed by logical if-then rules. The consequent part of the
rule is described by a linear interpolation between models provided by the membership
function. Similar ideas were developed in the neural network community [9], [10]. In
this approach, local model networks are learning systems able to model nonlinear
processes from observed input and output data. The resulting structure is based on linear
local models, as a generalization of neural networks, which learn and map the inputs to
the outputs at specific operating conditions using localized receptive fields and radial
basis functions.

64

This modeling approach has been successfully used to model a nonlinear

mechanical system, such as in [11], that proposed a polytopic structure to model the
longitudinal motion of an aircraft. This modeling structure also has been used to predict
the dynamic behavior of electromechanical systems that exhibit highly nonlinear
dynamics that vary with the operating conditions, such as a turbogenerator [12]. Then,
based on this model, a nonlinear control of the turbogenerator was obtained by online
blending of multiple PID controllers designed for each linear sub-model.
Application of polytopic modeling to model power electronics converters was
first presented in [13] and [14]. The models were constructed from average models
linearized at different operating conditions. Then, polytopic models, together with linear
matrix inequalities, were used to search for Lyapunov function candidates numerically in
order to assess the stability of the system.
In this chapter, a Hammerstein-Wiener structure and a polytopic structure are
investigated as the means to extend the application of two-port black-box models to
include the nonlinear behavior of a regulated buck converter as well as an unregulated
buck-boost (Flyback) converter. To demonstrate the capabilities of the polytopic black-
box modeling approach, nonlinear black-box models are obtained for a converter that
exhibits strong nonlinearities, specifically for an unregulated bus dc-dc converter
working under light load and a regulated flyback converter.
4.2 Hammerstein-Wiener Black-box model Structure
This model structure comprises a linear system connected in series with one or
two scalar nonlinear functions. Fig. 4.1 shows a block diagram representation.

Fig. 4.1. Hammerstein-Wiener model structure.

65

Where u(t) and y(t) are the inputs and outputs for the system. F and H are
nonlinear functions that correspond to the input and output nonlinearities, respectively.
w(t) and x(t) are internal variables that define the input and output of the linear block.
The Hammerstein-Wiener model has two main limitations. Namely, it assumes
that the dynamics are linear, which is not necessarily true, and that the model structure is
defined for single input single output systems only. Despite its limitations, its rather
simple structure allows an easy way to extend linear models to represent a large number
of nonlinear systems. In this chapter, a Wiener model structure is used to characterize the
constant power load behavior of a regulated buck converter
4.3 Regulated Buck Converter Black-Box model using a Wiener
structure
Having a closed control-loop design is an integral part of many converters.
Therefore, it is important to extend the proposed methodology to converters that have any
form of linear feedback network. In this section, a two-port black-box model was
obtained for the buck converter prototype used in Chapter Two. The converter is operated
in closed-loop mode using a three-pole, two-zero error amplifier to regulate the output
voltage. The process of obtaining a black-box model in this case is very similar to that
used for open-loop converters. The main difference is a modification of the black-box
internal structure to cope with the constant-power load behavior of the converter,
resulting in a nonlinear black-box model.
The FRF measurements are performed in the same way for closed-loop converters
as they are for open-loop converters. However, better measurements, especially for G
om
and Y
im
, are obtained by loading the converter with a resistor and using a source with
some output impedance. Then, the source and load dynamics are removed from the
measurements using the methodology presented in chapter three. If a source with low
output impedance is used for the measurement setup, most of the current disturbance
injected into the system will flow through the source and not the converter.
System identification of a closed-loop system is not an issue as long as a linear
feedback network is used to implement the control [15, 16]. The identification process

66

can be done in two ways. One approach is indirect identification, meaning that
knowledge of the control or the plant is needed to build the model [17]. The other
approach is direct identification, meaning that the data are treated as if they were
obtained from an open-loop system. In this work, direct identification was used because it
is more appropriate, given that no information about the control structure of the converter
is not available. Fig. 4.2 shows the FRFs measured from the hardware and the FRFs
obtained from the models using system identification.
G
o
Z
o

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)

Y
i
H
i

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)

Frequency (rad/sec) Frequency (rad/sec)
Measured FRF , Unterminated FRF , Model

Fig. 4.2. Closed-loop FRFs from measurements and identified models.
The black-box model structure used for the open-loop case is linear; therefore, the
model will fail to predict the inherent constant-power load behavior of regulated
converters. To incorporate this behavior, an analysis of how the FRFs changed when the
input voltage varied from plus or minus 25% the nominal voltage was performed. The
results indicated that the FRFs exhibit slight changes when voltage varies, except for H
i
,
which shows considerable changes on the FRF dc gain. Thus, H
i
cannot be represented
by one FRF. Rather, a family of FRFs is needed, as shown in Fig. 4.3.

67

Because the changes of FRF H
i
were mostly on the dc gain, a Wiener structure is
suitable for this case. For this, a linear model of H
i
is multiplied by a nonlinear functions
implemented by means of a look-up table that, depending on the input voltage, will pull
out a multiplier factor. This simple structure can be used to model this family of FRFs,
avoiding for the moment the use of more elaborate approaches like polytopic models
[36]. In addition to the look-up table, the model for FRF G
o
assumes perfect disturbance
rejection at low frequencies. Therefore, there was no information on the dc output voltage
operating point. This information was introduced in the model using the variable Vo. For
the open-loop case, this is not necessary because G
o
dc gain is the transfer ratio from the
input to the output voltage. The black-box model structure for closed-loop converters is
illustrated in Fig. 4.4.
Fig. 4.4. Nonlinear black-box model based on a Wiener structure.
Time domain validation of the models is performed by applying a load
disturbance from 50% to 100% of the nominal load. The load step is implemented using
resistors going from 2 ohms to 1 ohm. To compare the experimental and simulated
H
i

A
m
p
l
i
t
u
d
e
(
d
B
)


P
h
a
s
e

(
D
e
g
)

Frequency(Hz)

Fig. 4.3. Effects of changes of the operating point on H
i
.
10
2
10
3
10
4
-50
-40
-30
-20
-10
0
Frequency(Hz)
A
m
p
l
i
t
u
d
e

(
d
B
)
10
2
10
3
10
4
-200
-150
-100
-50
0
P
h
a
s
e
(
d
e
g
r
e
e
s
)


FRF for 15 V input Voltage
FRF for 20 V Input Voltage
FRF for 25 V Input Voltage

68

transients, black-box models of the resistors are obtained for the frequency range of
interest. An experimental transient response is plotted against simulation results. Fig. 4.5
shows that the models agreed with the experimental results, confirming that the
measurements performed at converter terminals not only capture the power stage
dynamics but also the control dynamics. Notice that, in this case, all simulation data are
generated by the models. No external disturbance signal is recorded and applied to the
model.
Input Voltage v
1
Output Voltage v
2

V
o
l
t
s

Input Current i
1
Output Current i
2

A
m
p
s


Time (sec) Time (sec)
Experimental , Experimental Average , Model

Fig. 4.5. Open loop FRFs from measurements and identified models.
Since the previous experiment does not show a large variation in the input
voltage, the constant power behavior of the converter is not observed. Hence, to test the
model with a more challenging condition, simultaneous variation of the input voltage and
output current is produced by adding a 2 ohm resistor in series with the source, and an
electronic load is used to produce the load step. Thus, when the load current start rising,
the converter input current will also increase, producing a drop in voltage across the two
ohm resistor. As observed in Fig. 4.6, the drop in the input voltage will trigger the
converter constant power load behavior. This example also shows how the model would
predict the converter transient response when the dynamics of the source and the load are
different than those used to obtain the models.
0.0095 0.01 0.0105 0.011
19.6
19.8
20
20.2
g
(
)
0.0095 0.01 0.0105 0.011
4.8
4.9
5
5.1
g
(
)
0.0095 0.01 0.0105 0.011
0.6
0.8
1
1.2
1.4
1.6
1.8
g
(
p
)
0.0095 0.01 0.0105 0.011
2.5
3
3.5
4
4.5
5
Time(Sec)

69

Input Voltage v
1
Output Voltage v
2

V
o
l
t
s

Input Current i
1
Output Current i
2

A
m
p
s


Time (sec) Time (sec)
Experimental , Experimental Average , Model

Fig. 4.6. Open-loop FRFs from measurements and identified models.
The resistor produces a 1.5 V drop in voltage at the converter input terminals
when the load current is increased from 2.5A to 4.5A. Also, the resistor changes the
transient dynamics considerably compared to the previous case. However, since the
converter model is un-terminated and captures the bidirectional flow of information that
occurs in real converters, the model is able to predict the new transient behavior. This
result confirms that BBTC models can be used as a building block to simulate an
interconnected power electronics system.
4.4 Polytopic Models
As demonstrated above, a Wiener black-box structure successfully captures the
nonlinear behavior of a regulated buck converter basically because static nonlinearities
are dominant. However, this is not always the case. In fact, the general case includes the
presence of strong nonlinearities that shape the dynamic performance of the converter.
For instance, Fig. 4.7 shows the FRFs of a unregulated bus converters with fixed duty
cycles measured at different steady state operating conditions.

0.008 0.01 0.012 0.014 0.016
0.6
0.8
1
1.2
1.4
1.6
0.008 0.01 0.012 0.014 0.016
18
18.5
19
19.5
20
20.5
g
(
)
0.008 0.01 0.012 0.014 0.016
18
18.5
19
19.5
20
20.5
Ti (S )
0.008 0.01 0.012 0.014 0.016
2.5
3
3.5
4
4.5

70

G
o
Audiosusceptibility Z
o
Output Impedance
Y
i
Input Admittance H
i
Back Current Gain
Fig. 4.7. IBC FRFs measured at different output current values.
When the converter output current is between 10 A and 20 A, the FRFs do not
change with the output current. This indicates that the converter behaves as a linear
system in this region. Therefore, a linear black-box model will be enough to model the
converter. However, when it operates below 10 A, the converter starts approximating the
boundary between continuous conduction mode (CCM) and discontinuous conduction
mode (DCM). As a consequence, the FRFs start exhibiting some nonlinearities as at this
boundary, and the order of the converter dynamics is reduced.
To show the impact of the phenomena described above, experimental data is
compared with simulation results obtained using a linear black-box model developed
from the FRFs measured in the region where the converter behaves as a linear system.
First, a load step will test the predictive capabilities of the model in the region where the
converter behaves as a linear system. This is from 50% to 100% and 75% of the nominal
load. The match between waveforms is nearly perfect, as illustrated in Fig. 4.8 in which
the blue trace is the experimental measurement, the green is the moving average of the
measured waveform, and the red trace corresponds to the simulation.

71

Input Voltage (V) Output Voltage (V)

Input current (A) Output current (A)
Experimental Experimental Average Model
Fig. 4.8. IBC linear black-box model prediction between 50% and 100% of the nominal load.
If the linear model is used to predict the converter behavior outside of the
specified operating region, the prediction becomes less accurate because the nonlinear
behavior at light loads is not included in the model (Fig. 4.7). Fig. 4.9 shows, as an
example, the output voltage predicted by the linear model for a load step from 1 A to 5 A,
then to 10 A and back to 1 A, clearly illustrating this shortcoming. A more complex
black-box model structure is thus required if the model is expected to capture the
nonlinearities observed in the FRFs depicted in Fig. 4.7.
Output Voltage (V) Output current (A)
Experimental Experimental Average Model
Fig. 4.9. IBC linear black-box model prediction between 5% and 50% of the nominal load.
To cope with the above problem, this paper proposes to use a family of local
black-box models that are valid at different operating points instead of a single model as
4 4.1 4.2 4.3 4.4 4.5 4.6
x 10
-3
46
47
48
49
50
Time(Sec)
4 4.1 4.2 4.3 4.4 4.5 4.6
x 10
-3
9
10
11
12
13
14
Time(Sec)
4 4.1 4.2 4.3 4.4 4.5 4.6
x 10
-3
2
2.5
3
3.5
4
4.5
5
5.5
Time(Sec)
4 4.1 4.2 4.3 4.4 4.5 4.6
x 10
-3
10
12
14
16
18
20
Time(Sec)
p
4 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8
x 10
-3
10.5
11
11.5
12
12.5
Time(Sec)
4 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8
x 10
-3
0
1
2
3
4
5
6
7
8
9
10
11
Time(Sec)
p

72

in the linear open-loop converter case. This family of local FRFs is then employed to
approximate the converter behavior across the converter operating space.
In this framework, a nonlinear system described by
x = f(x, u)

(2)
can always be written as
x = A(x, u)x + B(x, u)u


(3)

This system can be approximated by a set of linear local models that are an
adequate description of the system in these different local conditions. The parameters of
the local models are obtained from linearization of a nonlinear model or from measured
data using system identification methods. The transition from one operating regime to
another is performed by means of weighting functions w
i
(x,u)also known as
scheduling function, such that
( )( )
( )( )
n
i i i
i 1
n
i i i
i 1
x w x, u A x Bu
y w x, u C x Du
=
=
= +
= +

i
(4)
These scheduling functions are chosen as normalized radial basis functions that
are placed in the center d(u
i
) of each operating regime u
i
, such that w
i
(u) 1 if u u
i
,
and w
i
(u) 0 if u u
i
. There are many different types of weighting functions. For
instance, Takagi-Sugeno uses a triangular weighting function [15]; however, bell-shaped
functions, double sigmoid functions, and trapezoidal functions are also extensively used
[18-20]. The double sigmoid function is defined by
( )
( ) ( )
1 1 2 2
a x c a x c
1 1
f x, a, b, c
1 e 1 e


=

+ +
(7)
where the parameters a
1
and c
1
define the slope and the center for the left sigmoid,
respectively. Likewise, a
2
and c
2
define the slope and the center for the right sigmoid.
Examples of the different weighting functions are depicted in Fig. 4.10.

73

Triangular Weighting Function Sigmoid Weighting Function

Trapezoidal Weighting Function Gaussian Weighting Function

Fig. 4.10. Weighting functions.
For the modeling task, the sum of the scheduling functions at any point in the
modeling space should be equal to unity. This is a necessary requirement for the set of
local models to be able to approximate the nonlinear system. Therefore, in many cases,
these functions need to be normalized. However, normalization introduces undesirable
side effects when the centers are not equally spaced or if adjacent scheduling functions
have different widths. These side effects are known as shifting of the equilibrium points
and reactivation [21]. These effects are more pronounced as the input dimension
increases.
To avoid the effect of normalization, the models developed in this paper use a
sigmoid scheduling function since it allows the interconnection of linear models with
different width and center locations. The sum of the scheduling functions is equal to one
as long as the adjacent sigmoid functions intersect in the middle point of their rising and
falling edges with the same slopes. This guarantees that the sum of the scheduling
functions is equal to unity as depicted in Fig. 4.11.
Weighting Functions Summation of Weighting Functions

Fig. 4.11. Model interconnection using sigmoid functions.
To illustrate how a polytopic structure works, a one-dimensional smooth
nonlinear function Y is approximated using set of linear functions as depicted in Fig.
4.12. This structure can be implemented in Matlab/Simulink using the fuzzy logic
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
0
0.2
0.4
0.6
0.8
1
1.5 2 2.5 3 3.5 4
0
0.2
0.4
0.6
0.8
1
1.5 2 2.5 3 3.5 4 4.5 5
0
0.2
0.4
0.6
0.8
1

74

toolbox, although, in principle, polytopic models can be implemented in any circuit

simulation environment.


Fig. 4.12. General one-dimensional polytopic model.
The multidimensional polytopic model occurs when the local sub-models depend
on more than one operating condition variable. For the two-dimensional case, for
instance, the model is constructed by simply blending together a set of one-dimensional
polytopic models. Similarly, a three-dimensional model is obtained by blending a set of
two-dimensional models. However, it should be noted that the number of local models
required grows exponentially with the dimension. This is known as the curse of
dimensionality.
This work focuses on two-dimensional models, assuming that the operating point
of a dc-dc converter is determined by its input voltage and output current. Fig. 4.13
shows a representation of a two-dimensional polytopic model where the highlighted blue
lines indicate the operating space, i.e., the minimum and maximum input voltage and
output current values in which the converter was designed to work. For the sake of
simplicity, the operating space in the figure is divided into four operating regions
indicated by the dotted lines. In practice, the partitioning of the space is one of the most
difficult decisions to make and several publications have been devoted to this issue, in
particular [22-24].
The center point of each function indicates the operating conditions in which the
local models were obtained, corresponding as well to the weighting functions midpoint

75

location. The red surface that intersects the operating regions defines the area where the
output of a local model is equal to the output of the global model. The light green area of
the plane defines where the output of the global model is the result of the interpolation of
two or more adjacent local models. The size of the area defined by the intersection of the
red and light green areas depends on the value assigned to the sigmoid weighting function
slope. Thus, the slope is an additional variable to be determined during the modeling
process.

Fig. 4.13. Two-dimensional sigmoid scheduling functions.

4.5 Polytopic Black-Box Model for an Unregulated Bus Converter

In this section, a model of the unregulated bus converter E48SB12020NRFA that
is valid in a wide operating region is derived. The model is expected to predict the
converter behavior when used between an output current of 0.5 A to 20 A and between an
input voltage of 38 V to 52 V. Notice that these two variables, namely input voltage and
output current, define the operating conditions of the converter; therefore, a two-
dimensional partition of the operating space is used to build the model as shown in Fig.
4.14. Each of the rectangular regions in this figure represents the set of FRFs measured at
the corresponding input voltage and output current operating point. For instance, the
model whose center is located at 15 A and 40 V is used to model the operating regime
from 10 A to 20 A and 38 V to 44 V.

76

Fig. 4.14. Two-dimensional partition of the operating

space.
The partitioning of the operating space is determined by how the FRFs change
with the operating conditions. For instance, in the area where the converter behaves as a
linear system, it is logical to use a single model. In contrast, at light loads, the FRFs
experience large changes, which requires smaller partitions in this area. The number of
local models is determined on a trial and error basis in which preliminary local models
obtained every 2 A load steps are simulated and compared with experimental data. The
preliminary partitioning is then further refined as necessary by adding, for instance, more
local models obtained over 1 A steps up to a load of 9 A as shown in Fig. 4.14. This is
repeated until satisfactory results of the model are obtained when compared with
experimental data.
For variations on the input voltage, the converter FRFs change slightly with the
operating conditions; therefore, two one-dimensional models characterized at 48V and
40V, respectively, were enough to represent the whole operating space. Once each sub-
model is obtained and validated, a polytopic structure is built using each of the sub-
models. After this, the polytopic model is validated in a wide operating region.
To validate the model, an experiment is designed to produce the simultaneous
variation of the input voltage and output current. For this, a 2 resistor is placed in
series with the source and a load step is applied at the output terminals using an electronic
load working as a constant current sink. The conditions of the experiment were later
reproduced in simulation, using the actual electronic load current, measured
experimentally, as input to the simulated system.
Fig. 4.15 shows the simulated and experimental results after applying the series of
load steps from 1 A to 5 A, from 5 A to 20 A, and finally back to 1 A. As observed, the
black-box model accurately predicts the transient and steady state behavior that occurs
during each of the load steps, demonstrating the predictive capabilities of the model in

77

both linear and nonlinear operating regions of the converter. This test verifies that the
developed polytopic model contains the necessary information to accurately represent the
converter dynamic in a wide operating region.
Input Voltage (V) Output Voltage (V)
Input current (A) Output current (A)
Time (ms)
Experimental Experimental Average Model
Fig. 4.15. Model validation for a two-dimensional IBC black-box model using
a load step from 5% to 25% and 100% of the nominal load.

4.6 Polytopic Black-box Model of an Unregulated Flyback

Converter
Based on small-signal average model analysis, a flyback converter, as shown in
Fig. 4.16, with fixed duty cycle and working in CCM could be represented by a linear
network where G
o
(s), Z
o
(s),Y
i
(s) and H
i
(s) are expressed in terms of the converter
parameters as in (6), (7), (8), and (9).
2 2.5 3 3.5
38
40
42
44
46
48
50
p g
2 2.5 3 3.5
8
9
10
11
12
p g
2 2.5 3 3.5
0
1
2
3
4
5
A
m
p
p
2 2.5 3 3.5
0
5
10
15
20
A
m
p

78

Fig. 4.16. Schematic of a flyback prototype.

u
o
(s) =
DD

(CR
c
s+1)
LCs
2
+C(D
2
R
c
+R
I
)s+D
2
(6)
Z
o
(s) =
(CR
c
s+1)(Ls+R
I
)
LCs
2
+C(D
2
R
c
+R
I
)s+D
2
(7)
Y
I
(s) =
D
2
Cs
LCs
2
+C(D
2
R
c
+R
I
)s+D
2
+
C
I
s
(C
I
R
cI
s+1)
(8)
B
I
(s) =
DD

(CR
c
s+1)
LCs
2
+C(D
2
R
c
+R
I
)s+D
2
(9)
D and D` are the duty cycle and its complement, C is the output capacitor, L is the
equivalent transformer inductor reflected to the secondary side, and C
f
is the input filter
capacitor reflected to the secondary side. Also, the model includes the parasitic resistance
of all energy storage elements denoted by R
l
, R
c
, and R
cf
.
However, in practice there are some factors, such as the switch parasitic
resistance, the transformer flux leakage, the diode drop voltage, proximity to the
boundary between CCM and DCM, and others that, although for the sake of simplicity
are not included in the average model, do affect the real terminal characteristic of the
converters significantly.
These effects, neglected during a conventional averaging process, are
incorporated in the measurements shown in Fig. 4.17. For instance, the output impedance
Z
o
slightly increases as the current is changed from 5 A to 2 A; however, below 2 A a
sharp increase on Z
o
is observed as the converter approaches the boundary between CCM
and DCM.

79

G
o
Audiosuceptibility

Z
o
Output Impedance
Y
i
Input Admittance H
i
Back Current Gain
Fig. 4.17. Open-loop flyback FRFs measured at different output current values.
The back current gain H
i
also experiences big changes as the converter
approaches this boundary condition. Specifically, it can be observed that H
i
becomes
more damped as the current decreases. The input admittance Y
i
on the contrary remains
unchanged, mainly because this measurement depends on the converter input filter. For
G
o
the changes are not visible because they are small and do not have a strong influence
on this converter due to the low current ratings handled.
Based on the FRFs in Fig. 4.17, a linear black-box model and a polytopic black-
box model are constructed and validated against experimental data. Fig. 4.18 shows the
experimental and simulated results using a linear black-box model when a load step from
3A to 4 A, and then to 5 A is applied using an electronic load working as a constant
current sink.

80

Experimental Experimental Average Model

Fig. 4.18. Open loop flyback linear model.
The linear model accurately predicts the input current, but the voltage prediction
shows a small dc offset when the converter is running at full load. This result
demonstrates that a linear model is able to represent the converter behavior in this
operating regime. However, other experimental results not shown here indicate that the
prediction error of the linear model increases as the model is used away from the
specified operating regime, just as for the buck bus converter analyzed in the previous
section. For instance, when it was used to predict the converter behavior at light loads.
Even though the prediction error is small when using a simple linear network, a
poly-topic model structure can substantially eliminate this problem as depicted in Fig.
4.19, where now the model prediction accurately matches the converter behavior in the
specified operating regime.



0.012 0.014 0.016 0.018 0.02
0
0.5
1
1.5
2
2.5
Time(Sec)
A
m
p
Input Current
0.012 0.014 0.016 0.018 0.02
4.6
4.8
5
5.2
Time(Sec)
V
o
l
t
Output Voltage

81

Experimental Experimental Average Model

Fig. 4.19. Open loop flyback polytopic model.

4.7 Polytopic Black-box Model of a Regulated Flyback Converter
When a flyback converter works in closed-loop mode, its nonlinearities become
accentuated, especially regarding its output impedance and back current gain. Fig. 4.20
shows the frequency responses of the converter transfer functions G
o
, Z
o
, Y
i
, and H
i
when
the input voltage is kept constant at 18 V and the output current changes from 2 A to 5 A.
As seen, the FRFs of the audiosusceptibility, output impedance, and back current gain
change significantly with the operating conditions, reflecting the much less damped
behavior of the converter as the output current increases. On the contrary, the input
admittance of the converter remains virtually invariant throughout the complete operating
region. For this reason, its transfer characteristic is modeled using a single local model
for the sake of simplicity in the final model structure.



0.012 0.014 0.016 0.018 0.02
0
0.5
1
1.5
2
2.5
Time(Sec)
A
m
p
Input Current
0.012 0.014 0.016 0.018 0.02
4.6
4.8
5
5.2
Time(Sec)
V
o
l
t
Output Voltage

82

G
o
Audiosusceptibility Z
o
Output Impedance
Y
i
Input Admittance H
i
Back Current Gain
Fig. 4.20. Closed-loop flyback FRFs measured at different output current values.
Accordingly, and due to their higher operating point dependance, the remaining
FRFs are modeled using a one-dimensional polytopic black-box structure and initially
partitioned every 1 A. The model was then builtmeasuring the corresponding set of
FRFsand verified by comparing the simulated and experimental results in order to
ensure that the number of local models chosen could indeed approximate the converter
transient behavior. Otherwise, a finer partition would have been required.
Fig. 4.21 shows specifically the experimental and simulation results when a load
step from 3 A to 4 A, from 4 A to 5 A, and back to 3 A, is applied to the converter using
an electronic load working in current sink mode. Here, it is clearly observed that the
model accurately predicts not only the steady state value of the input current and output
voltage but also the overshoot and natural frequency response. Notice that, as expected,
per the converter frequency response shown in Fig. 9, the transient response for the two
load steps applied is different, with the second load step presenting a much larger current
overshoot and settling time. All three transients were nonetheless predicted accurately by
the polytopic black-box model constructed.

83

Input current (A) Output Voltage (V)
Load step from 3 A to 4 A
Load step from 4 A to 5 A
Load step from 5 A to 3 A
Time (sec) Time (sec)
Experimental Experimental Average Model
Fig. 4.21. Experimental and predicted input current and output voltage for different load steps.
To further test the capabilities of the proposed black-box modeling approach, its
predictive ability under simultaneous variations of the input voltage and output current
was evaluated. To do so, a two-dimensional model similar to the one presented for the
bus converter was constructed, blending several one-dimensional models at different
input voltages. The partition resolution in this dimension would naturally depend on how
the converter FRFs change with the input voltage. Fig. 4.22 shows specifically how H
i

changes when the input voltage varies between 20 V and 16 V with a load current fixed
at 3 A. As observed, the transient dynamics are dominated by a double pole initially
located at 7 kHz that moves down to 5 kHz when the input voltage changes to 16 V.
Also, an increase on the dc gain of H
i
is observed, evincing that there is an increase on
0.0495 0.05 0.0505 0.051 0.0515 0.052 0.0525 0.053 0.0535
1
1.5
2
Ti (S )
p
0.0495 0.05 0.0505 0.051 0.0515 0.052 0.0525 0.053 0.0535
5.05
5.1
5.15
5.2
0.057 0.058 0.059 0.06 0.061
1
1.5
2
2.5
3
0.057 0.058 0.059 0.06 0.061
5.05
5.1
5.15
5.2
0.063 0.0635 0.064 0.0645 0.065 0.0655 0.066 0.0665
0
0.5
1
1.5
2
0.063 0.0635 0.064 0.0645 0.065 0.0655 0.066 0.0665
5.05
5.1
5.15
5.2
5.25
V
o
l
t

84

the transfer ratio from output current to input current. This is expected, as it is a direct
reflection of the constant power behavior of a regulated converter. From this figure, it is
also possible to infer that just a few one-dimensional models would be required to
approximate the converter nonlinear behavior when subjected to input voltage variations.
Back Current Gain (Hi)


M
a
g

(
d
B
)


Vin = 20 V Vin= 18 V Vin = 16 V
Fig. 4.22. Back current gain (Hi) FRFs measured
at different input voltage and fixed output current of 3 A.
The final partition used for the operating space is shown in Fig. 4.23

Fig. 4.23. Flyback polytopic black-box model operating space .
It should be noticed that, in the end, a greater number of models were used to
cope with the accentuated constant power load behavior that the converter presents at
lower voltages, implying that large variations on the dc operating point were to be
expected. However, the number of local models is not limited to the ones presented here.
In fact, if a higher number of local models were used, the prediction of the polytopic
black-box model would be increasingly more accurate [25]. The operation of the
converter under discontinuous conduction mode (DCM) was not included in the model
due to the intrinsic constraints it presents regarding the measurement of FRFs. This is
under current investigation.
Using the partition presented above, the resultant two-dimensional polytopic
black-box model was constructed and validated against experimental data. For this, an
10
2
10
3
10
4
-25
-20
-15
-10
-5
0
5
10
Frequency (Hz)

85

experiment similar to the one used to validate the bus converter model was implemented
for the flyback converter. In this case, a 1.6 resistor was placed in series with the
source. Fig. 4.24 shows the simulation and experimental results obtained when two
consecutive load steps are applied. As seen, the flyback black-box polytopic model is
able to predict not only the steady state but also the transient behavior of the converter
when subjected to large simultaneous disturbances at its input and output terminals.
As discussed, the model prediction could be further improved by increasing the
number of local models; however, it would be difficult to justify the effort knowing that
any new added partition in the voltage or current dimensions implies adding 12 additional
local models.
Input voltage (V) Input current (A)
Output voltage (V) Ouptut current (A)
Experimental Experimental Average Model
Fig. 4.24. Experimental and simulation results for simultaneous variation of the input voltage
and load current.



0.05 0.052 0.054 0.056 0.058 0.06 0.062
16
17
18
19
20
21
0.05 0.052 0.054 0.056 0.058 0.06 0.062
0.8
1
1.2
1.4
1.6
1.8
2
1
0.05 0.052 0.054 0.056 0.058 0.06 0.062
5
5.05
5.1
5.15
5.2
5.25
Ti (S )
2
0.05 0.052 0.054 0.056 0.058 0.06 0.062
2
2.5
3
3.5
4
4.5
5
Ti (S )
p
2

86

4.8 Polytopic Black-box Model Analysis

Clearly a polytopic model structure is able to capture most important nonlinear
dynamics present in a dc-dc converter. However, it is important to emphasize that this is
an approximation as stated in (3) and (4), therefore, there will be cases in which a
polytopic model will fail to approximate the behavior of a nonlinear system.
Consequently, it is very important to understand the conditions in which it is possible to
confidently use a polytopic model and when it will fail.
The justification for polytopic models in dynamics system theory is that the local
behavior around an equilibrium point or trajectory approximates the behavior of the
nonlinear system for initial conditions that deviate slightly from the operating point in
which the linearization was performed. Moreover, it is required that contiguous local
model parameters are similar, meaning that the system dynamics change smoothly when
moving from one operating regime to another [25, 26].
The conditions above are explained and rigorously mathematically treated in the
singular perturbation theory framework in which, if a system changes slowly with the
operating conditions, it is called a slowly varying system. In essence, singular
perturbation theory examines linear, linear time variant, and nonlinear systems by
decomposing the system in time scales in which a fast and a slow subsystem are
identified.
For linear time invariant systems, the only source of multiple time-scales is the
separation of the system eigenvalues into two or more groups of eigenvalues with
different orders of magnitude. In the case of linear time variant systems and nonlinear
systems, an additional source of time-scales exists based on the variation of model
parameters. For instance, a slowly varying system defined by
( ) ( ) ( ) ( ) z t A x t z t =
i

(10)


where x(t) satisfies
( ) ( ) ( )
x t f x t =
i
.
(11)

87

Then, it is demonstrated in [9, 27] that, given the following conditions,

the function f(x(t)) is bounded and continuous in C
1

4
f (x(t)) c
x

,
(12)

the eigenvalues of A(x) have real negative parts
( )
1
Re A(x(t)) c 0 < , (13)
and the norm of A(x(t)) and its first derivative are bounded for all x(t)

2
A(x(t)) c
and 3
A(x(t)) c

.
(14)
For a sufficiently small , the dynamic behavior of the system could be predicted
by studying a family of time-invariant systems with parameters frozen at specific
operating conditions.
Thus, for the regulated output voltage flyback converter modeled in section IV
using a black-box approach, it is important to show that in the operating space defined in
Fig. 4.23, the system meets the conditions stated in (12), (13), and (14). To achieve this,
the converter governing equations are expressed in the state space form and averaged
over a switching period. A schematic of the converter, including the PID controller, is
shown in Fig. 4.25.

Fig. 4.25. Schematic of a flyback converter referred to the secondary side.

88

The averaged converter governing equations are written in a form similar to (10)
and (11). Where the inductor L is the small parameter

( )
( )
L
L
o L
g C
C
R 1 d(t) 0 d(t)
i i i
L
1 L 1 d(t)
v v 0
0
v
C
C

= +







,
(15)

the PID controller state space equation is given by

( )
C1
1 2 1 2
C1
C2
C2 ref C
2 3 2 3
C3
C3
3 2 3 3 3 2
3 1 3
1 1
0
C R C R 0 v
v
1 1
v 0 0 v v v
C R C R
v
v
1 1 1
1 1 1
C R C R C R
R R C











= +










+




i
i
i
,
(16)
and f(x(t)) in this case correspond to the duty cycle d(t), which is finally defined
by (17) being the modulator gain Fm equal to 0.28.
C3
d(t) Fm v = 0 d(t) 1 (17)
As observed, (15) depends on d(t) that is proportional to v
C3
,

which depends on t
(time). Furthermore, d(t) is bounded and continuous in C
1
in compliance with (12). If,
from (15), d(t) is replaced by (16) and (17) gives the nonlinear average governing
equations with slowly variant coefficients shown in (18).
( )
( )
C3 m L
C3
L
C3 m
L
C
C
C1 C1
1 2 1 2
C2
C2
C3
2 3 2 3
C3
3 1 3 3 2 3 3 3 2
1 v F R
0 0 0 v F
L L
i
1 v F
0 0 0 0 i
C
v
v
1 1
0 0 0
v
v
C R C R
v
1 1
v
0 0 0
v
C R C R
v
1 1 1 1 1 1
0
R R C C R C R C R

















= +

















+

i
i
i
i
i
m
g
o
ref
2 3
3 1 3
0 0
L
1
0 0
C
V
0 0 0
i
1
V
0 0
C R
1 1 1
0 0
R R C

+



(18)
To calculate the eigenvalues, a linearization of (18) is performed at the boundaries
of the operating space where the polytopic structure is used to model the converter. The
linearized state space equation is given in (19).

89

( ) ( )
( )
c g
L
L
L
L
C
C
C1 C1
1 2 1 2
C2
C2
C3
2 3 2 3
C3
3 1 3 3 2 3 3 3 2
V V Fm
1 D R
0 0
L L L
i
1 D I Fm
i 0 0 0
C C
v
v
1 1
0 0 0
v
v
C R C R
v
1 1
v
0 0 0
v
C R C R
v
1 1 1 1 1 1
0
R R C C R C R C R
+

= +












+



i
i
i
i
i
g
o
ref
2 3
3 1 3
D
0 0
L
1
0 0
C
V
0 0 0
i
1
V
0 0
C R
1 1 1
0 0
R R C

+




(19)
The exact steady state conditions where the eigenvalues are calculated are marked
with red dots in Fig. 4.26

Fig. 4.26. Operating space where polytopic models predict the converter nonlinear behavior.
Fig. 4.27 shows that all eigenvalues are located in the left half plane, away from
the imaginary axis. However, there exist two eigenvalues that move towards the
imaginary axis as the current increases and toward the real axis as the input voltage
decreases. All others eigenvalues just move slightly from their original position. This
explains why the converter transient dynamic becomes less damped as the output current
increases as observed in Fig. 4.20 and 21. Despite of having two eigenvalues wandering
in the left half plane, they are sufficiently far away from the imaginary axis. Therefore,
the condition stated in (13) is met.

Fig. 4.27. Eigenvalues evaluated at the boundaries of the operating space.

90

If the input voltage is kept constant, for instance at 10V, and the output current is
allowed to go beyond 5A, the moving eigenvalues will cross the imaginary axis and the
system will become unstable as illustrated in Fig. 4.28

Fig. 4.28. Eigenvalues crossing into unstable regions.
From the Hartman-Grobman theorem, it is well known that quadratic and cubic
nonlinear terms neglected by local models become important, and, in fact, they will
determine the dynamics of the system when a set of eigenvalues approaches the
imaginary axis [28]. Because polytopic models are built from a family of local models,
they do not contain any information about quadratic or any high order nonlinear terms.
Therefore, polytopic models will be less and less accurate as the eigenvalues approach
the imaginary axis
The conditions outlined in (14) discard the use of polytopic models to
approximate the dynamics of non-smooth or discontinuous dynamical systems. To better
explain this limitation, it is important to mention that three types of non-smooth
dynamical systems exist:
Systems described by differential equations with a discontinuous right hand side;
Non-smooth continuous system with a discontinuous jacobian; and
Systems with discontinuities or jumps in the state, such as impacting systems.
Not surprisingly, dc-dc converters may exhibit all three types of non-smoothness.
Therefore, by showing the conditions in which this occurs while at the same time
showing the boundaries where polytopic models are, a valid approximation of the
nonlinear system is achieved.

91

In essence, high fidelity dc-dc converter models, such as switching models, are in
fact differential equations with discontinuous right hand sides. What allows the use
polytopic structure for modeling of dc-dc converters is the time scale separation between
the switching dynamics and the dynamics created by the converter energy storage
elements such as capacitors and inductors. Therefore, as conventional average models,
polytopic models ignore the switching dynamics and only predict average currents and
voltages values. Also, as average models, polytopic models will correctly predict the
converter average behavior as long as the switching frequency is several orders of
magnitude higher than the inverse of the smallest time constant defined by any of the
energy storage elements.
It is well known that flyback and, in general, all dc-dc converters have two
operation modes named continuous conduction mode (CCM) and discontinuous
conduction mode (DCM). The discontinuous conduction mode occurs when the current
through the inductor takes a zero value for a finite sub-interval of the switching period.
For converters designed to work in CCM at nominal load, DCM occurs at light load. The
transition between DCM to CCM and vice versa occurs naturally when the output current
is below certain value. However, during the transition, there is a significant change in the
converter dynamics as observed in Fig. 4.29, where H
i
and Z
o
FRFs are plotted from an
average model of the flyback converter, revealing the differences between CCM and
DCM.
For instance, H
i
and Z
o
exhibit a double pole when working in CCM. Once the
converter enters in DCM, the double pole abruptly disappears. In addition to that, Z
o

experience a considerable increase in its magnitude when going from CCM to DCM. This
is well known in power electronics and there are many reference devoted to explain this
phenomenon. What had not been clear until now were the implications this phenomenon
has when modeling the converter using a polytopic black-box modeling approach.
Based on the conditions stated in (14), it is concluded that polytopic model can be
used to model the converter when working either in CCM or DCM, but, if these models
are combined using a smooth interpolation functions, it will not predict the converter
behavior accurately when moving across the boundary between CCM and DCM since

92

there is not a smooth transition between these two operating modes. In fact, at the
boundary between CCM and DCM, the conditions stated in (14) are violated since the
jacobian is not defined at that point. Thus, in these conditions, the converter is basically a
non-smooth continuous system with a discontinuous jacobian and at the same time the
system dynamics show a jump in the state related with the inductor current.
Back current gain (H
i
)
Output impedance (Z
o
)
Fig. 4.29. FRFs from average models during CCM and DCM.
This is analysis is confirmed in Fig. 4.30 where a load step is applied in the
borderline between DCM and CCM. As evinced the model is not able to describe the
transient behavior during the transition
Input Voltage (V) Output Voltage (V)
Time (sec) Time (sec)
Experimental Experimental Average Model
Fig. 4.30. Closed loop flyback converter load step from DCM to CCM.

0.05 0.051 0.052 0.053 0.054
0.25
0.3
0.35
0.4
0.45
0.5
0.55
0.6
A
m
p
Input Current
0.05 0.052 0.054 0.056 0.058 0.06
5.02
5.04
5.06
5.08
5.1
5.12
5.14
5.16
V
o
l
t
Output Voltage

93

Also, it is important to emphasize that jumps in the state can also occur during
abnormal conditions such as short circuits in which a state could either increase or go to
zero in a very short period of time.

4.9 Conclusions
This chapter has presented a two-port black-box model using one- and two-
dimensional polytopic structures. With this new structure, black-box two-port models are
able to capture changes in the dynamic behavior of the converter when its operating
conditions change. In consequence, the proposed models are valid in a wide operating
region, and thus are capable of representing converter topologies with known nonlinear
behavior, such as the flyback converter studied throughout the chapter. Simulation results
validated with experimental data showed this, demonstrating that the proposed modeling
approach indeed captures the nonlinear dynamics of these converters. Furthermore, the
proposed structure for black-box two-port models was shown to be able to accurately
predict the dynamic behavior of converters when subject to simultaneous variations on
both input and output terminals (input voltage and output currents).

94

4.10 References
[1] F. Jurado, M. Valverde, and M. Ortega, "A method for the identification of micro-
turbines using a Hammerstein model," in Canadian Conference on Electrical and
Computer Engineering, 2005, pp. 1970-1973.
[2] W. I. Hunter and M. J. Korenberg, "The identification of nonlinear biological
systems: Wiener and Hammerstein cascade models," Biol. Cybern., vol. 55, pp.
135-144, 1986.
[3] J. Francisco, V. Manuel, and G. Manuel, "Identification of Hammerstein Model
for Solid Oxide Fuel Cells," in The IEEE Electrical and Computer Engineering
Conference, 2006, pp. 442-445.
[4] J. S. Lai and D. J. Nelson, "Energy Management Power Converters in Hybrid
Electric and Fuel Cell Vehicles," Proceedings of the IEEE, vol. 95, pp. 766-777,
2007.
[5] F. Alonge, F. D'Ippolito, F. M. Raimondi, and S. A. T. S. Tumminaro, "Nonlinear
modeling of dc-to-dc converters using the hammerstein's approach," IEEE
Transactions on Power Electronics, vol. 22, pp. 1210-1221, 2007.
[6] F. Alonge, F. D'Ippolito, and T. Cangemi, "Hammerstein Model-Based Robust
Control of DC/DC Converters," in 7
th
IEEE Power Electronics and Drive Systems
Conference 2007, pp. 754-762.
[7] T. Takagi and M. Sugeno, "Fuzzy identification of systems and its applications to
modeling and control," IEEE Transactions on Systems, Man,Cybernetics, vol. 1,
pp. 116-132, 1985.
[8] Z. Huaguang and L. Derong, Fuzzy modeling and fuzzy control 1ed. Boston:
Birkhuser 2006.
[9] R. Shorten, R. Murray-Smith, R. Bjorgan, and H. Gollee, "On the interpretation of
local models in blended multiple model structures," International Journal of
Control, vol. 72, pp. 620 628, May 1999 2000.
[10] R. Murray-Smith and H. Kenneth, " Local model architectures for nonlinear
modelling and control " in Neural network engineering in dynamic control
systems Berlin, 1995, pp. 61-82.
[11] A. Fujimori and L. Ljung, "A Polytopic modeling of aircraft by using system
identification," in The International Conference on Control and Automation,
2005, pp. 107-112 Vol. 1.
[12] R. Lixin, G. W. Irwin, and D. Flynn, "Nonlinear identification and control of a
turbogenerator an on-line scheduled multiple model controller approach," IEEE
Transaction on Energy Conversion,, vol. 20, pp. 237-245, 2005.
[13] S. D. Sudhoff, S. F. Glover, S. H. ak, E. J. Zivi, J. D. Sauer, and D. E. Delisle,
"Stability analysis of a DC power electronics based distribution system," in
proceedings of Society of Automotive Engineers 2002 Power Systems Conference,
Coral Springs, Florida, 2002.
[14] S. Glover, "Modeling and stability analysis of power electronics based systems,"
in Electrical engineering. vol. Ph.D United States - Indiana: Purdue University,
2003, p. 179.

95

[15] L. Ljung, System identification (2nd ed.): theory for the user: Prentice Hall PTR,
1999.
[16] L. Ljung and U. Forssell, "Variance results for closed-loop identification
methods," in Decision and Control, 1997., Proceedings of the 36th IEEE
Conference on, 1997, pp. 2435-2440 vol.3.
[17] L. Ljung and U. Forssell, "An alternative motivation for the indirect approach to
closed-loop identification," IEEE Transactions on Automatic Control, vol. 44, pp.
2206-2209, 1999.
[18] N. B. Karayiannis and M. M. Randolph-Gips, "On the construction and training of
reformulated radial basis function neural networks," IEEE Transactions on
Neural Networks, vol. 14, pp. 835-846, 2003.
[19] G. Gregorcic and G. Lightbody, "Local Model Network Identification With
Gaussian Processes," IEEE Transactions on Neural Networks, vol. 18, pp. 1404-
1423, 2007.
[20] S. G. Cao, N. W. Rees, and G. Feng, "Quadratic stability analysis and design of
continuous-time fuzzy control systems," International Journal of Systems
Science, vol. 27, pp. 193 - 203, 1996.
[21] R. Shorten and R. Murray-Smith, "On Normalizing basis function networks," in
proceedings of 4th Irish Neural Networks Conference, Univ. College Dublin,
1994, pp. 5594 - 5599
[22] C. Hametner and S. Jakubek, "Neuro-Fuzzy Modelling Using a Logistic
Discriminant Tree," in American Control Conference, 2007, pp. 864-869.
[23] O. Nelles, "Axes-Oblique Partitioning Strategies for Local Model Networks," in
IEEE International Symposium on Intelligent Control, 2006, pp. 2378-2383.
[24] K. Gasso, G. Mourot, and J. Ragot, "Structure identification in multiple model
representation: elimination and merging of local models," in Proceedings of the
40th IEEE Conference on Decision and Control, 2001, pp. 2992-2997 vol.3.
[25] H. O. Wang, J. Li, D. Niemann, and K. A. T. K. Tanaka, "T-S fuzzy model with
linear rule consequence and PDC controller: a universal framework for nonlinear
control systems," in The Ninth IEEE International Conference on Fuzzy Systems,
2000, pp. 549-554 vol.2.
[26] R. Murray-Smith and T. A. Johansen, Multiple model approaches to modeling
and control. Bristol,PA: Taylor&Francis, 1997.
[27] P. Kokotovic , H. Khalil , and J. O'Reilly Singular perturbation methods in
control: analysis and design Orlando, Florida: Academic Press, 1986.
[28] L. Perko, Differential equations adn dynamical systems. New York: Springer,
2001.

5 Chapter
Analysis and Simulation of Distributed Power
Electronic Systems
This chapter presents the application of black-box terminal characterization models
(BBTC) for the simulation an analysis of distributed power electronics systems compose
by commercial power electronics modules and filter modules. The models predict the
transient and steady state response of a cascade, parallel and distributed power system.
They are also used to predict the small signal stability between source and load
converters. All simulation results are validated with experimental data.
5.1 Introduction
Direct current power electronics distribution systems are present in many
applications such as telecommunications [1], spacecrafts [2], electric vehicles [3], and
industrial applications, with power ratings ranging from a few watts up to kilowatts.
With the observed trend toward using multiple renewable energy sources and energy
storage elements to improve the availability of energy, it is expected that dc distribution
systems will become predominant in many other applications that today are mostly ac
distribution systems, such as self-sustainable homes, data centers, and intelligent
buildings.
Many of todays dc distribution systems are built using commercial dc-dc
converters mainly because they have proven to have high reliability, high efficiency, and
high power densities and their use contributes to reducing the time to market of a product.
97

However, there is a drawback to this approach that becomes evident during the design
process of a system. For instance, it may be difficult to try to answer questions about the
details of the dynamic response of the system for a set of possible architectures. Simply
changing a converter from one manufacturer to another or changing the load type is a
daunting task because the transient dynamics depend not only on the converter itself but
also on the characteristic of the source and the load [4, 5]. Todays models for system
level simulation are not an answer to this problem because they are based on a white-box
approach for which detailed models for each converter component are needed to build a
converter model. Generally, that information is not available. Therefore, such questions
are answered today by building multiples prototypes and testing the system in multiple
conditions.
The main goal of developing black-box dc-dc converter models is to fill this gap,
thereby giving engineers in charge of building dc power electronics distribution system
an appropriate tool to design, analyze, and test the system before is actually built. Thus,
architecture studies, sub-system interaction issues, protection coordination issues, and
transient specifications issues are addressed during the design process.
Previous chapters have been devoted to explaining and testing the modeling
methodology for an individual converter. However, they did not address questions about
how closely black-box models would predict the terminal behavior of commercial
converters nor consider how the model would predict the dynamic behavior of complete
systems. This chapter shows how BBTC models perform in the cases mentioned above.
For this, a cascade, parallel and a distributed power system with intermediate bus
architecture is constructed out of commercial power modules. Each module is modeled
using BBTC models, then the models are interconnected in a circuit simulation
environment to predict current and voltage transients of the real system when under load
disturbances. Also, BBTC models are able to predict the stability of a system by
analyzing the input and output impedance of the converters models at their interface. All
simulation results are validated with experimental data.
98

5.2 Black-box Models of Commercial dc-dc Converters

Dc distribution systems comprise several converters with specific characteristics
and functions in the systems; for instance, in the case of intermediate bus architecture,
there are three conversion stages and a filter stage before reaching the load. These
conversion stages are known as the front end converters, bus converters, and point of load
converters [6]. A front end converter consists of an active rectifier stage, also known as
power factor correction (PFC) circuit cascaded with a dc-dc converter in charge of
providing a regulated 48 V output voltage. This converter is the interface between the ac
and dc systems. A filter is usually required to meet the standards for conducted
electromagnetic emissions, and even though the filters are designed to cut off high
frequency currents, they introduce significant low frequency dynamics that need to be
taken into account during the design of a distribution system. The bus converter connects
to the 48 V regulated node and, depending on the design, it will step the voltage down to
12 V, 8V, or 5V, regulated or unregulated. The bus converter also provides electrical
isolation from the 48 V node to the converters connected downstream. The final
conversion stages are the point of load converters (POLs) that will take the intermediate
distribution voltage and convert it to the actual voltage level required by the load.
Before discussing the modeling of interconnected dc distributions systems, it is
important to show all the modeling effort being done to build a set of black-box models
for individual converters manufactured by Delta and Artesyn. It is equally important to
show whether interconnected black-box converter models resemble the characteristics of
the real interconnected converters.
First, for different kinds of converters used in intermediate bus architecture, Fig.
5.1 shows the experimental and simulated transient response to a load step from 50% to
100% of the converter nominal load. Despite the fact that this is a small sample compared
to what is available in the market, the converters chosen have very different
characteristics, depending on whether they are isolated, non-isolated, regulated, semi-
regulated, or unregulated. The black-box models developed using the methodology
presented in previous chapters is able to capture their internal dynamics seamlessly.
99

Output voltage Input current

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)

Regulated bus converter
E48SR12007, 84W, 48V in ,
12V/7A out

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)

Semi-regulated bus converter,
IBC30AQS4812, 360W
42Vto 53V in ,12V/30A out

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)

Unregulated bus converter
E48SB, 200W , 48V in,
12V/16.6A out

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)

Non-isolated point of load
PTV12020 8.3-14 Vin, 0.75-
5V/16A out

V
o
l
t
a
g
e

(
V
)


C
u
r
r
e
n
t

(
A
)

FL75L20 Filter module 75 V
Vin, 20A out
Time (ms) Time (ms)
Fig. 5.1. Black-box terminal behavioral models of commercial converters and filters, photo by author.


4.5 5 5.5 6 6.5 7
11.4
11.6
11.8
12
12.2
12.4
(S )


Experimental
BBTCModel
5 5.5 6 6.5 7
0.5
1
1.5
2
2.5
4 4.2 4.4 4.6 4.8
3
10
11
12
13
14
Ti (S )


Measured
BBTC Model
4 4.2 4.4 4.6 4.8
3
0
2
4
6
8
p
Ti (S )


Measured
BBTC Model
4 4.5 5 5.5
10.8
11
11.2
11.4
11.6
11.8
12

Experimental
BBTC Model
4 4.5 5 5.5
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
p


Experimental
BBTC Model
4 4.2 4.4 4.6 4.8 5 5.2
3
4.7
4.8
4.9
5
5.1
5.2
Ti (S )
g
(
)


Measured
BBTCModel
4 4.2 4.4 4.6 4.8 5 5.2
3
3
4
5
6
7
8
Time(Sec)
(
p
)


Measured
BBTCModel
4 4.2 4.4 4.6 4.8
3
47.2
47.4
47.6
47.8
48
Ti (S )
g
(
)


4 4.2 4.4 4.6 4.8
2
3
4
5
6
Ti (S )

conve
predi
FRFs
meas
what
conve
amon
are c
match
stabil
outpu
FL75
const
meas
calcu
only
BBTC
impe
exper
impe
Although
erters and fi
ict the transi
s calculated
ured at the s
was observ
erter termina
ng the source
cascaded, th
h the outpu
lity assessm
ut impedance
The syst
5L20, a bus
tant current
ured un-term
ulated at the
the output
C models we
Fig
To devel
dance were
rimental and
dance chang
black-box m
filters, there
ient dynamic
at the inte
same interfa
ved in Cha
als depends n
e, the conve
he equivalen
ut impedanc
ment since th
e at each inte
tem consisti
s converter
sink mode a
minated imp
same locat
impedance
ere develope
g. 5.2. Interco
lop a black
e measured
d its corresp
ges with the
models are
is no guara
c of a real in
erface betwe
ace of the rea
apter Three,
not only on
erter, and the
nt output im
e measured
he analysis i
erface.
ing of a La
E48SB120
are interconn
pedances Z
ion using bl
characterist
ed.
nnected system
k-box mode
at differen
onding blac
operating c
able to pred
antee that in
nterconnecte
een two con
al system. T
which dem
the converte
e load. Ther
mpedance me
in a real s
is based pur
amda dc po
10NRFA, a
nected as sh
sa
and Z
sb

lack-box mo
tic is model
m for FRF pr
el of the s
nt output cu
ck-box mode
onditions th
dict transien
nterconnecte
ed system un
nverters mo
This requirem
monstrated t
er dynamic b
refore, when
easured at a
system. Th
rely on com
ower supply
and an elec
hown in show
are compare
odels. For th
led. For the
ediction, phot
ource, seve
urrent valu
el FRFs. It i
herefore, a si
nt dynamics
d black-box
nless it is sh
odels resemb
ment comes
that the tran
but also on th
n two or mo
all the inter
his is also i
mparison of t
y LT804, a
ctronic load
wn in Fig. 5
ed with the
he Lamda p
e other com

to by author.
eral FRFs o
es. Fig. 5.3
s evinced th
ingle transfe
10
of individua
x models wi
hown that th
ble the FRF
directly from
nsient at th
he interactio
ore converter
rfaces shoul
important fo
the input an
an EMI filte
d working i
5.2. Then, th
e impedance
power supply
mponents, fu
of its outpu
3 shows th
hat the outpu
er function i
00
al
ill
he
Fs
m
he
on
rs
ld
or
nd
er
in
he
es
y,
ull
ut
he
ut
is
101

used to represent the dynamic behavior of the source. In the case of the EMI filter, a
complete two port black-box model was developed. Since the filter is a passive network,
the un-terminated Z
o
and H
i
are measured with the input terminals short circuited. G
o
and
Y
i
are measured with all terminals open. Fig. 5.4 shows the experimental FRFs and the
FRFs from the black-box model. The bus converter model is already available from
Chapter Three.

Z
s


A
m
p
l
i
t
u
d
e

(
d
B
)

P
h
a
s
e

(
D
e
g
)

Frequency (Hz) Frequency (Hz)
3D view Measured FRF Model FRF
Fig. 5.3. Lamda power supply output impedance.
G
o
Z
o

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)


Y
i
H
i

A
m
p
l
i
t
u
d
e

P
h
a
s
e

(
D
e
g
)

Frequency (rad/sec) Frequency (rad/sec)
Measured FRF Model FRF

Fig. 5.4. EMI filter FRFs.
10
0
10
3
10
4
10
5
10
6
-100
-50
0
10
-2
10
-1
10
3
10
4
10
5
10
6
-20
0
20
40
60
10
-1
10
0
10
1
10
3
10
4
10
5
10
6
-50
0
50
100
10
0
10
3
10
4
10
5
10
6
-150
-100
-50
0
-40
-30
-20
-10
0
10

10
2
10
3
10
4
10
5
-45
0
45
90
102

Fig. 5.5 shows the experimental Z

sa
FRF measured at the interface a and the
FRF obtained with black-box models at the same location, the FRFs present very good
matching in the frequency range of interest. It is important to highlight that the resulting
output impedance is not obtained by simply adding the two output impedances but is
from the result of the interaction between the Lamda dc source output impedance and the
four transfer functions used to model the filter as expressed in (1).
Z
sa
= Z
o
+
H

Z
s
G
o
1+Z
s
Y

(1)
For the next case, a bus converter is cascaded with the filters and the source, the
electronic load is set to 20 A to match the operating conditions of the previous case then
the total output impedance Z
sb
is measured at the interface b. Fig. 5.6 shows the
experimental FRF and the findings obtained through simulations.

Z
sa

A
m
p
l
i
t
u
d
e

(
d
B
)


P
h
a
s
e

(
D
e
g
)

Frequency (Hz)
Measured FRF, Model FRF

Fig. 5.5. Experimental and predicted output impedance at interface a.

Z
sb

A
m
p
l
i
t
u
d
e

(
d
B
)


P
h
a
s
e

(
D
e
g
)

Frequency (Hz)
Measured FRF, Model FRF

Fig. 5.6. Experimental and predicted output impedance at interface b.
-30
-20
-10
0

10
2
10
3
10
4
10
5
-45
0
45
0

-40
-20
10
2
10
3
10
4
10
5
-90
-45
0
45
103

Again, the models provide a very good prediction of the resulting output
impedance in almost all the frequency ranges of interest. Notice that three resonant points
are easily identified in the measurement and predicted FRFs; the first resonant point is
attributed to the source and occurs at 700 Hz, a second resonant point at 7 kHz is added
by the EMI filter, and the last resonant point is located at 60 kHz is the largest and
dominates the output impedance dynamics. A noticeable difference in the magnitude and
phase is observed beyond 90 kHz. However, from the model order reduction technique in
Chapter Two, it was concluded that these differences at high frequencies do not play an
important role on the predictive capabilities of the model.
5.3 Modeling of Cascaded Converters
This section shows how BBTC models would predict transient responses when
two converters connected in cascade are fed by a front end converter. The experimental
setup consists of a front end converter DPS-1001AB-1, an unregulated bus converter
E48SB12020, and one POL DNL10SIP16. The system is simulated and compared against
experimental measurements. Fig. 5.7 shows the interconnecting diagram.

Fig. 5.7. Cascaded system.

Fig. 5.8 illustrates the regulating action of the POL keeping the voltage at 4.97 V
(v
3
). The models predict the same behavior, demonstrating that BBTC models capture the
power stage dynamic together with control dynamics. On the other hand, voltage v
2
drops
during the load step, which produces an increase of the POL input current (i
2
) due to its
constant power behavior. Similarly, the nonlinear BBTC model senses the drop in the
voltage and adjusts the transfer current ratio to successfully predict the new input current
value. When looking at the dc source output current (i
1
), it is observed that the models
capture the slow transient dynamics including the disturbances speed and pattern as it
propagates from the load to the source. For instance, the i
3
step up is at 3.92 ms, but the
104

step is seen by the dc source 0.3 ms later because of the presence of energy storage
elements such as inductors and capacitors in the system.
v
1
i
1

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)

v
2
i
2

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)

v
3
i
3

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)

Time (ms) Time (ms)
Experimental, Model

Fig. 5.8. Model prediction of a cascaded system.

3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4

48.15
48.2
48.25
48.3
48.35


Measured
BBTC Model
3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4
x10
-3
11.2
11.4
11.6
11.8
12
12.2
12.4
V2
Time(Sec)
V
o
l
t
a
g
e

(
V
o
l
t
s
)


Measured
BBTCModel
3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4
x10
-3
3
4
5
6
7
8
i2
Time(Sec)
C
u
r
r
e
n
t

(
A
m
p
s
)


Measured
BBTCModel
3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4
x10
-3
4.8
4.85
4.9
4.95
5
5.05
V3
Time(Sec)
V
o
l
t
a
g
e

(
V
o
l
t
s
)


Measured
BBTCModel
3.6 3.8 4 4.2 4.4 4.6 4.8 5 5.2 5.4
x10
-3
8
9
10
11
12
13
14
15
16
i3
Time(Sec)
C
u
r
r
e
n
t

(
A
m
p
s
)


Measured
105

5.4 Modeling of Parallel Converters

Parallel configuration of converters is used in distributed power electronic
systems as a means to improve the reliability of N+1 power module configuration by
reducing the stress on each converter in the system. Therefore, it is important to show
that the proposed modular behavioral modeling approach can be used for parallel
converter configurations. For this example, a passive current sharing method is
implemented using two semi-regulated intermediate bus converters IBC30AQS4812. In
addition to the converters, two shottky diodes are used to decouple the outputs. The
interconnecting diagram is shown in Fig. 5.9.

Fig. 5.9. Semi-regulated bus converters in parallel (passive current sharing)

Since passive current sharing is used, no resistors were added in series with the
converters output. Only mismatches of the converters internal impedances will
determine how the current is shared between the two. Therefore, a black-box model of
each converter is obtained separately.
Fig. 5.10 shows experimental measurements plotted against a simulation using
BBTC models. From the experiments, it is shown that the total output current i
2
is shared
unevenly by the converters where converter A is drawing more current than converter
B. Also, it is observed that their transient dynamics are slightly different. Simulations
show that BBTC models are able to capture those small differences with only four FRF
measurements. If the same level of detail were to be achieved with a white box approach,
each passive and active component would need to be characterized with a high level of
detail.


106

Output current
A
m
p
s

(
A
)

4 4.2 4.4 4.6
10
15
20



Time (ms)

i
2a
experiemental i
2b
experimental
i
2a
simulation i
2b
simulation
Input current
A
m
p
s

(
A
)

4 4.2 4.4 4.6
0
1
2
3
4
5
6
7



Time (ms)

i
1a
experiemental i
1b
experimental
i
1a
simulation i
1b
simulation
Output Voltage v
2b

V
o
l
t
s

(
V
)

4 4.2 4.4 4.6
10
11
12
13

14



Time (ms)

v
2b
experiemental v
2b
simulation
Fig. 5.10. Model prediction of a parallel system.

5.5
how
conve
regul
transf
to a l
conve
an E
PTV1
electr
black
is pro
extern
meas
agree
Modelin
A DPS sy
BBTC mod
ersion stage
lation of the
formation an
ower and tig
The expe
erters. The s
MI filter FL
12020W. Tw
ronic load w
k-boxes. Fig.
F
The trans
oduced by a
nal input to
urements an
ement betwe
ng of a D
ystem with in
els can be u
s. The first
48 V dc dis
nd isolation,
ghtly regulat
rimental set
setup compr
L75L20, a b
wo POLs are
working as a
. 5.11 shows
Fig. 5.11. Dist
ient respons
load disturb
the model is
nd simulation
en the exper
Distribute
ntermediate
used to mod
stage, as me
stribution vo
, and the thi
ted voltage.
tup used for
rises a front
bus convert
e loaded wit
a current sin
s a one-line i
tributed powe
se at the inpu
bance at one
s the current
n results are
rimental data
d Power
bus architec
el a complet
entioned bef
oltage, the se
ird stage tran
this section
end convert
ter E48SB,
th a 0.5 ohm
nk. The pow
interconnect
er electronic sy
ut and outpu
e of the POL
t i
4
measured
compared in
a and the pre
System (
cture (IBA) i
te system. T
fore, is in ch
econd stage
nsforms the
n is build en
ter DPS-100
and three p
ms resistor, a
wer resistors
ion diagram
ystem, photo b
ut terminals
Ls from 8 am
d at the POL
n Fig. 5.12.
ediction usin
(DPS)
is used as a s
This system
harge of rec
is in charge
intermediate
ntirely out o
01AB-1 used
points of loa
and one is lo
are also cha
m of the syste


by author.
of the modu
mps to 16 am
L terminals.
The results
ng BBTC mo
10
setup to show
include thre
tification an
e of providin
e bus voltag
f commercia
d as a source
ad converter
oaded with a
aracterized a
em.
ules upstream
mps. The onl
Experimenta
show a goo
odels.
07
w
ee
nd
ng
ge
al
e,
rs
an
as
m
ly
al
od
108

For instance, the load disturbance produced by the electronic load travels through
the POL and interacts with the bus converter output impedance. A transient and a voltage
drop in the intermediate distribution voltage v
3
are direct consequences of the interaction.
However, the interaction also affects the other two POLs because the drop in the
intermediate distribution voltage generates input current transient at the other two POLs.
These transients are also added to the total current i
3
. The good correlation between the
experimental and simulated current i
3
indicates that the models capture the behavior
described above, which is commonly known as an interaction between loads.
Looking at the bus converter side, although the model used in this simulation is
the same model used for the cascade case, the transients voltages v
3
and v
2
and currents i
3

and i
2
at converter terminals are modified by all impedances upstream and downstream.
This is a very important result because it shows that BBTC models capture the internal
converter dynamics using frequency domain methods. However, the time domain
response would depend on the subsystem interaction of the power modules that compose
the system.
The BBTC model obtained from the EMI filter might not be valid for EMI
studies but is good for subsystem interaction where the phenomenon under study occurs
at much lower frequencies and should not be ignored. The predicted transient and steady
state response of v
2
and i
1
in Fig. 5.12 demonstrate the importance of the EMI filter
model.








109

v
1
i
1

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)


v
2
i
2

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)


v
3
i
3

V
o
l
t
a
g
e

(
V
)

C
u
r
r
e
n
t

(
A
)


v
4
i
4


Time (ms) Time (ms)
Experimental, Experimental Average, Model

Fig. 5.12. Model prediction of a distributed power electronic system.

4 4.2 4.4 4.6 4.8
48
8.1
8.2
8.3
8.4
4 4.2 4.4 4.6 4.8
2.8
3
3.2
3.4
3.6
3.8
4
4.2
4.4
4 4.2 4.4 4.6 4.8
47.8
47.9
48
48.1
48.2
48.3
48.4
48.5
48.6
4 4.2 4.4 4.6 4.8
2.5
3
3.5
4
4.5
5
4 4.2 4.4 4.6 4.8
11.2
11.4
11.6
11.8
12
12.2
4 4.2 4.4 4.6 4.8
11
12
13
14
15
16
17
18
4 4.2 4.4 4.6 4.8
4.9
5
5.1
5.2
5.3
4 4.2 4.4 4.6 4.8
8
10
12
14
16

5.6
on th
been
contr
persp
regul
and E
stabil
equiv
Midd
distri
interc
apply
conve
stand
not a
some
could
the co
testin
and d
Subsys
The incre
he dynamic p
pushing con
rol bandwid
pective, this
lated conver
EMI filters
lity analysis
valent sourc
dlebrook imp
The imp
ibuted powe
connected, a
ying the Ny
erter output
dard test that
available. T
e papers are c
The objec
d reduce the
onstruction
ng multiple
design of eff
stem Inte
eased expect
performance
nverter desig
dth in order
trend direct
rters are pro
[7-9]. Thu
s of interco
ce output im
pedance crite
edance crit
er systems.
as shown in
yquist criteri
impedance a
t needs to be
There exists
cited here [1
ctive in this
cost of dev
of time-cons
distributions
ficient, reliab
eraction
tations for h
e requested b
gners to wo
to meet n
ly affects th
one to stron
us, system l
onnected sy
mpedance, a
erion [10].
terion is va
It states th
n Fig. 5.13,
ion to the m
and Z
in
is th
e performed
an extensiv
11-14].
Fig. 5.13. Su
s section is
veloping pow
suming prot
s configurati
ble, and stab
higher power
by very sens
ork with high
new load re
he stability o
ng subsystem
level design
ystems that
and the loa
alid for sm
hat when tw
, the stabili
minor-loop g
e load conve
on a system
ve literature
ubsystem inte
to provide
wer electron
totypes, ther
ions and pro
ble dc distri
r demands a
itive loads t
her switchin
quirements.
of the system
m interaction
ners rely on
are based
ad input im
mall-signal s
wo stable su
ity of the s
gain Z
s
/Z
in
erter input im
m prototype if
devoted to
eraction.
insight into
nics distribut
eby allowing
oviding a po
bution syste
and tighter s
o voltage va
ng frequenci
From the
m in the sens
n with sourc
n methods d
on the ana
mpedance als
stability an
ubsystems M
system is de
where Z
s
i
mpedance. T
f high fidelit
this topic, a

o how black
tion systems
g engineers
owerful tool
ems. No less
11
specification
ariations hav
ies and large
system leve
se that tightl
ce converter
developed fo
alysis of th
so known a
alysis of d
M and N ar
etermined b
is the sourc
This is a ver
ty models ar
although onl
k-box model
s by avoidin
the option o
l for analysi
s importantly
10
ns
ve
er
el
ly
rs
or
he
as
dc
re
by
ce
ry
re
ly
ls
ng
of
is
y,
111

this section also shows a limitation of the models when in the presence of strong
interactions. For this, three examples are presented.
The first example involves a regulated dc power supply, a POL DN16, and three
pairs of interconnecting leads AWG #14 with different lengths. The objective is to
emulate the case in which there is a considerable distance between the source converter
and the POL and to compare experimental results against simulations using nonlinear
black-box models. Fig. 5.14 shows the measured output impedance Zs 1, Zs 2, and Zs 3
that correspond to the output impedance of the dc power supply with short medium and
large leads, respectively.

Fig. 5.14. Equivalent source impedances with different leads lengths.
The Matlab-Simulink simulation environment as well as other simulation packages such
as Saber and Simplorer allow the user to perform small signal ac analysis on the
nonlinear black-box models, extract a linear model at the specific operating condition,
and draw Bode plots and location of poles and zeroes. Since the software used in this
work to build the black-box models is Simulink-Matlab, the linearization tool of
Simulink is used for the analysis. Fig. 5.15. shows a Bode plot of the output impedance of
the source with short wires and the POL input impedance obtained by linearization of the
black-box models. To set the operating conditions, the source is loaded with an ideal
current sink and the POL is fed by and ideal voltage source as suggested in [15, 16].

Source Output Impedance
10
10
10
100
Frequency (rad/s)
10
3
10
4
10
5
10
6
-2
-1
0


10
3
10
4
10
5
10
6
0
50
Source Zs 1
Source Zs 2
Source Zs 3
112

The Bode plot indicates a small interaction between the two impedances around
150 kHz; therefore, the stability of the system is determined by plotting the Nyquist of
the minor-loop gain as shown in Fig. 5.15. The Nyquist plots indicate that the -1+j0 point
is not encircled and, in fact, the system is far away from any unstable region; thus, a large
signal disturbance will not likely drive the system to an unstable operation. The
prediction of the models about the stability of the systems is confirmed by time domain
simulations and experimental data gathered at the POL terminals. The input current and
the output voltage transients are produced by a load step from 8 A to 16 A.
Zin vs Zs 1 Nyquist plot
M
a
g
n
i
t
u
d
e

(
d
B
)

I
m
a
g
i
n
a
r
y

a
x
i
s


Frequency (Hz) Real axis
Zin, Zs 1
Fig. 5.15. Low interaction (short wires).

POL input current POL output voltage
A
m
p
s

(
A
)

V
o
l
t
s

(
V
)

Time(msec) Time(msec)
Experimental, Model
Fig. 5.16. Transient prediction (short wires).
When the wires length is increased, a strong interaction between the source and
the POL is expected. Fig. 5.17 shows that effectively the interaction between the source
output impedance and the load input impedance is stronger and start occurring at around
3.9 4 4.1
2
3
4
5
6
7
8
9
(
p
)


Measured
BBTC Model
3.9 4 4.1
4.85
4.9
4.95
5
5.05


Measured
BBTC Model
-2
-1
-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5
0
1
2

10
-30
-20
-10
0

10
2
10
3
10
4
10
5
-180
-135
-90
-45
0
45
90
113

50 kHz with longer wires. When the Nyquits of the minor-loop gain is plotted, the
contour approaches but does not encircle the -1+j0 point. Therefore, the system is stable
in the small signal sense but is close to an unstable region and, in this case, a large
disturbance may be more likely to drive the system to an unstable operation although this
is not conclusively determined.
To confirm what the black-box models predict, time domain simulations together
with experimental data are used for this purpose. Likewise, a large load disturbance from
8A to 16 A is applied at the converter terminals. The transient predicted by the models
matches very well with the experiments. A higher input current overshoot as well as an
increase of the ringing before the current reaches a steady state value as observed in Fig.
5.18.
Zin vs Zs 2 Nyquist plot
M
a
g
n
i
t
u
d
e

(
d
B
)


I
m
a
g
i
n
a
r
y

a
x
i
s

Frequency (Hz) Real axis
Zin, Zs 2
Fig. 5.17. Medium interaction (medium-length wires).

Converter Input Current Converter Output Voltage
A
m
p
s


V
o
l
t
s

(
V
)

Time(msec) Time(msec)
Experimental Model
Fig. 5.18. Transient prediction (medium-length wires).
3.9 3.95 4 4.05 4.1 4.15 4.2 4.25
3
2
3
4
5
6
7
8
9
3.9 3.95 4 4.05 4.1 4.15 4.2 4.25
3
4.9
4.95
5
5.05
-20
-10
-15 -10 -5 0 5 10 15
0
10
20
10

()
-180
-30
-20
-10
0
10
2
10
3
10
4
10
5
-90
0
90
114

When the longest interconnecting wires are used, the Bode and Nyquist plots shows a
stronger interaction than in the previous two cases, with the contour of the Nyquist plot
nearly encircling the -1+j0 point as evinced in Fig. 5.19.
Time domain simulation predicts a dynamic behavior according to the previous
analysis. It shows that more ringing is expected during transients, but at the end, voltages
and current will settle down to the new operating conditions. However, in this case,
experimental data does not support the models predictions. The system becomes
unstable after a large disturbance from 8A to 16 A is applied at the output terminals of
the POL as shown in Fig. 5.20.
Zin vs Zs 3 Nyquist plot
M
a
g
n
i
t
u
d
e

(
d
B
)


I
m
a
g
i
n
a
r
y

a
x
i
s


Frequency (Hz) Real axis
Zin , Zs 3
Fig. 5.19. Strong interaction (long wires).

Converter Input Current Converter Output Voltage
A
m
p
s

(
A
)

V
o
l
t
s

(
V
)

Time(msec) Time(msec)
Experimental, Model
Fig. 5.20. Transient prediction (long wires).
The reason why black-box model fails to predict the system response in this case is that
the inductance of the interconnecting wires introduces a pair of complex poles. As the
4 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8
2
3
4
5
6
7
8
9
10
4 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9
3
4.88
4.9
4.92
4.94
4.96
4.98
5
5.02
5.04
5.06
5.08
Ti (S )
-20 -10 0 10 20
-40
-30
-20
-10
0
10
20
30
40
10

-180
-30
-20
-10
0
10
2
10
3
10
4
10
5
-90
0
90
115

inductance increases, the complex poles move towards the imaginary axis and become
dominant, as illustrated in Fig. 5.21. From Chapter Four is known that polytopic black-
box models are less accurate when any of the system eigenvalues are too close to the
imaginary axis due to the fact that nonlinear black-box models are built from a family of
local models and during the linearization process some nonlinear terms are neglected.
These nonlinear terms becomes dominant and determine the stability of the systems when
a set of eigenvalues are close to the imaginary axis.
I
m
a
g
i
n
a
r
y

a
x
i
s

Real axis
Fig. 5.21. H
im
Pole location as wire inductance increases.
Under these circumstances, variation on the output current will produce large
variation on the input current. The input current reacts with the source output impedance
producing a drop in the input voltage which will increase the input current magnitude
even more because of the constant power load behavior of the converter creating a kind
of positive feedback. This phenomenon is amplified until the system collapses.
The next example shows another consequence of subsystem interaction with
catastrophic results that could be avoided during the design stage by using black-box
models. In this case, a regulated bus converter DSE48SR12007 is used. The
specifications for this converter are 48 V input voltage, 12 V output voltage, and 7 A
output current. This converter is able to deliver 84 watts with 93% efficiency. The load
converter is a POL with 12 V input voltage, and the output is set to 5 V and a maximum
current of 16 A. The POL may provide up to 80 watts with 93% efficiency.
4
5
Pole-Zero Map
-10 -8 -6 -4 -2 0
x 10
-3
-2
-1
0
1
2
3
x 10

With
oppor
condi
conve
electr
manu
impe
intera
determ
point

M
a
g
n
i
t
u
d
e

(
d
B
)




transi
system
to 16
POL
watts
-6
-4
-2
2

-18
18
power ratin
rtunity to t
itions. Fig. 5
As a benc
erters is fir
rolytic capa
ufacturer. Th
dance obtai
action occur
mine the sta
t, thus, the m
Fig. 5.22
Time dom
ient at the P
m is stable a
6A and back
input curren
s which is ju
0
0
0
0
0
10
1
10
2
0
0
0
ngs so close
test the pre
5.22 shows a
chmark for t
rst investiga
citor mount
he Bode plo
ined by lin
rs at approxi
ability of the
models predic
2. Subsystem
Zs vs Z
in

Frequency (H
Fig. 5.23.
main simul
POL termina
and remains
to 8 A is ap
nt reaches a
st 6 watts ab

10
3
e, it is expec
ediction cap
a diagram of
this example
ated. Low
ted close to
ot of the bu
nearization
imately 50 k
system. In
ct a stable sy
interaction be
Hz)
Regulated bu
ation of th
als that matc
stable even
pplied at the
a peak input
bove the nom
10
4
cted the sys
pabilities of
f the intercon
e, a low subs
interaction
the POL i
us converter
of black-bo
kHz. Conseq
this case the
ystem.
etween a regu

I
m
a
g
i
n
a
r
y

a
x
i
s

Z
s
, Z
us converter an
he interconn
ches very w
when a larg
POL output
t current of
minal load.
10
5

stem will hit
f black-box
nnected syste
system intera
n is achieve
input termin
r output im
ox models,
quently, a N
e contours d
lated bus conv
Z
in

nd POL (low i
nected syste
well with the
ge load pertu
t terminals.
7.5 A, imp
-15
-10
-5 0
-5
0
5
10
15
t some limit
models ne
em.
action betwe
ed by placi
nals as sugg
mpedance an
indicates t
Nyquist plot
do not encirc

verter and a P
Nyquist plot
Real axis
interaction ).
em shows t
experiment
urbation from
During the
posing a pea
5 10 15
11
ts, giving th
ear abnorma
een these tw
ing a 44
gested by th
d POL inpu
that a sma
is needed t
cle the (-1+j0
POL
the predicte
tal data. Th
m 8 A to 12 A
transient, th
ak load of 9
5 20 25
16
he
al
wo
F
he
ut
all
to
0)
ed
he
A
he
90
30
117

Input Voltage v
1
Output Voltage v
2

V
o
l
t
s


Input Current i
1
Output Current i
2

A
m
p
s


Time (sec) Time (sec)
Experimental, Experimental Average, Model

Fig. 5.24. Transient prediction regulated bus converter and POL (low interaction ).
A strong subsystem interaction is created by placing a 4 H inductor in series
with the source converter. A Bode plot of the equivalent source impedance and the POL
input impedance are shown in Fig. 5.25. The added inductor increases the bus converter
output impedance at high frequencies, thereby now the interactions start from 10 kHz.
The Nyquist plot of the minor-loop does not encircle the (-1+j0), but the contour of the
plot passes very close to this point.
Z
s
vs Z
in
Nyquist plot
M
a
g
n
i
t
u
d
e

(
d
B
)


Frequency (Hz) Real axis
Z
s
Z
in

Fig. 5.25. Regulated bus converter and POL (strong interaction ).

4 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8
-3
11
11.5
12
12.5
13
4 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8
3
4.9
5
5.1
5.2
5.3
4 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8
2
3
4
5
6
7
8
9
4 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8
3
6
8
10
12
14
16
18
400
-100 -50 0 50 100 150 200 250 300 350
-400
-200
0
200

-20
0
20

10
1
10
2
10
3
10
4
10
5
-180
-90
0
90
118

Fig. 5.26 shows the experimental and predicted transient response for a load step
from 8 A to 12 A to 15 A. Because of the presence of the inductor, the input current
presents a larger overshot reaching 8 A peak, overloading the bus converter by
approximately 15% of the nominal load, but the system remains stable. In addition to
this, black-box models of the bus converter and the POL are able to accurately predict the
transient response of the system.


Input Voltage v
1
Output Voltage v
2

V
o
l
t
s



Input Current i
1
Output Current i
2

A
m
p
s


Time (sec) Time (sec)
Experimental, Experimental Average , Model

Fig. 5.26. Transient prediction regulated bus converter and POL (strong interaction ).
A very interesting event occurs when a load step from 8 A to 12 A to 16 A is
applied at the POL terminals. As evinced in Fig. 5.27 when the second load step is
applied the POL input current goes above 8.4 A which is the over-current protection
setting of the regulated bus converter. The protection mechanism automatically turns off
the bus converter and stop feeding the POL.
As a demonstration of how black-box models are useful in this case, a simple
over-current protection scheme is implemented in the bus converter black-box model as
follows: the current going out of the model is sensed, and, after the protection trips it, an
ideal switch opens the output terminals of the black-box model. At this point, the model
is not designed to describe the converter dynamics after the protection trips; therefore, the
4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6
3
11
11.5
12
12.5
13
4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6
3
4.85
4.9
4.95
5
5.05
5.1
5.15
5.2
4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6
3
2
3
4
5
6
7
8
4 4.2 4.4 4.6 4.8 5 5.2 5.4 5.6
3
6
8
10
12
14
16
119

differences between the models prediction and the experimental data are expected after
the converter enters into failure mode.
Converter input current Converter input voltage
A
m
p
s

(
A
)

V
o
l
t
s

(
V
)

Time(msec) Time(msec)
Experimental, Experimental (average), Model
Fig. 5.27. Transient prediction regulated bus converter and POL (strong interaction, system
collapse when under large disturbance).
What is important in this example is the highly accurate prediction capabilities of
the models. System designers would be able to simulate the system to test overload
coordination protections schemes, and search for protection trips when the system is
working under large signal disturbances. Also, it would be possible to look for under- or
over-rated fuses. Black-box models, as explained in Chapter Four, are not intended for
short-circuit analysis.
On the other hand, the problem above could be avoided by designing a damping
filter consisting of a capacitor in series with a resistor and placing it close to the input
terminals of the POL. Current design approaches focus on intermediate one-stage and
two-stage damped filters [8, 17, 18], but there are no guidelines for designing simple RC
network filters for the case in which an intermediate filter is not required. In this case,
black-box models are a very interesting solution to this problem.
As evinced in Fig. 5.27, the bus converter trips because of the current overshoot
demanded by the POL. If the current overshoot is prevented during large transients, the
problems could be solved. In Chapter Three, it was shown that the dynamic of the
current when interconnected with other converters depends on the interaction between the
source converter output impedance and the load converter input impedance and is
expressed as
3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8
-2
0
2
4
6
8
8.4
3 3.5 4 4.5
3
4
6
8
10
12
120

B
Im
=
H

_1+
Z
s
Z
n
]
(2)
where H
im
is the resulting back current gain and can be obtained by measuring its FRFs
when the converter is interconnected with other converters. H
i
, Z
s
, and Z
in
are the
unterminated back-current gain, equivalent source output impedance, and the load
converter input impedance, respectively. Since H
i
and Z
in
are known from the POL black-
box model it is possible to modify the equivalent source impedance in order to reshape
the frequency response of H
im
by adding an unknown impedance Z
x
in parallel with Zi.
The resulting interconnected system is shown in Fig. 5.28. The back current gain is
measured at the points a and b.

Fig. 5.28. System with damping filter.
For the schematic above, H
id
is the desired back-current gain characteristic when
connected to a non ideal source and can be calculated as
B
Id
=
H

_1+
Z
s
Z
x
(Z
s
+Z
x
)
Z
n
_
(3)

In the equation above Z
x
and H
id
are unknown; however, H
id
could come from the
analysis of the minor loop gain of the original circuit Z
s
/Z
in
. As evinced in Fig. 5.29 there
is a point where the phase margin drops considerably. The stability of the interconnected
system is improved by boosting the phase margin of the minor loop gain by placing a
zero as indicated in Fig. 5.30. The phase now drops only up to 45 degrees.



121

M
a
g
n
i
t
u
d
e

(
d
B
)

P
h
a
s
e

(
D
e
g
)

Frequency (rad/sec)

Fig. 5.29 Bode plot of the minor loop gain Z
s
/Z
in




M
a
g
n
i
t
u
d
e

(
d
B
)

P
h
a
s
e

(
D
e
g
)

Frequency (rad/sec)

Fig. 5.30 Bode plot of the minor loop gain Z
s
/Z
in







10 10 10 10
-100
0
100
2 4 6 8
45
90
135
180
225
Boostedphasemargin
AddedZero
10 10 10 10
-100
0
100

2 4 6 8
-45
0
45
90
135
180
Lowphasemargin
122

Fig. 5.31 shows the frequency response functions of the original back current gain H
im

and H
id
where the phase of the minor loop gain has been increased to ameliorate the
possibility of unstable behaviors.
Back current gain
M
a
g
n
i
t
u
d
e

(
d
B
)


Frequency (Hz)
H
im
H
id

Fig. 5.31. Reshaping H
im
to obtain H
id
.
Thus, given H
id
the impedance Z
x
can be calculated as follows
Z
x
=
H

Z
s
H
d
(1+Y

Z
s
)-H

(4)
Fig. 5.32 shows two FRFs, the first one corresponding to the full order model of Z
x
and
the second one to the reduced order model of Z
x
. Notice that the reduced order model is
the FRF of a RC series network, which is the damping filter designed to meet the
specifications for the back-current gain.
Z
x

M
a
g
n
i
t
u
d
e

(
d
B
)


Frequency (Hz)
Z
x
full order Z
x
reduced order
Fig. 5.32. Full order and reduced order damping filter.

-20
0
20
10
0
10
2
10
4
-90
-45
0
45
90

-60
-40
-20
0

10
2
10
3
10
4
10
5
-225
-180
-135
-90
-45
0
45
123

Fig. 5.33 compares experimental data when no damping filter is used against the
transient response predicted by black-box models when the full order damping filter is
placed at the POL input terminals. With that, as observed, the overshoot is eliminated as
is any possibility to trip the overload protection of the bus converter.
Converter input current Converter input voltage
A
m
p
s

V
o
l
t
s

Time(msec) Time(msec)
Experimental Experimental (average) Model
Fig. 5.33. Simulated transient response using full order filter.
If, instead of the full order model for Z
x
, a series RC network is placed at the POL
terminals, the current overshoot is eliminated and the input current behaves as a first
order system. This is a very important result because now specifications for H
im
can be
formulated by modifying the minor loop gain such that minimum or some interaction is
allowed, thereby giving system level engineers a powerful tool to design their systems
and meet the specifications.
Converter input current Converter input voltage
A
m
p
s

V
o
l
t
s

Time(msec) Time(msec)
Experimental Experimental (average) Model
Fig. 5.34. Simulated transient response using reduced order filter.
This procedure should be used cautiously because, depending on the back-current
gain (H
id
) specifications, it is very possible to end up with designs for Z
x
that are not
practical. For instance, it is possible to create a need for a large capacitor, the filter may
0.01 0.01020.01040.01060.01080.0110.01120.01140.0116
2
3
4
5
6
7
8
0.01 0.01020.01040.01060.01080.0110.01120.01140.0116
11
11.5
12
12.5
13
0.01 0.01020.01040.01060.01080.0110.01120.01140.0116
2
3
4
5
6
7
8
0.01 0.01020.01040.01060.0108 0.0110.01120.01140.0116
11
11.5
12
12.5
13
124

be difficult to implement because of the passive components tolerances, or the design

may be impossible to implement because of the presence of positive real eigenvalues in
Z
x
. For instance, in the example above H
id
was specified only as an exercise to illustrate
the methodology. Thus, the designed damping filter for the reduced order case consists of
a 3.8 mF capacitor in series with a 0.2 resistor. Such capacitor will may take a volume
larger than the bus converter and the POL together.
5.7 Conclusions
The results reported in this chapter has focused on presenting how black-box models
are used for system level simulation analysis and design. The good agreement between
the experiments and the models increase the confidence on the predictive capabilities of
the model and indicates that the models truly capture the terminal dynamic behavior of
the characterized converters and passive components.
It is shown in the frequency domain and in the time domain that when black-box
models are cascaded or paralleled, they predict the dynamic characteristic of the original
interconnected system. thereby, as demonstrated in this chapter these models allows
subsystem interaction studies using small signal analysis at the interconnecting interfaces,
with extensively used procedures such as the Middlebrook impedance criterion or less
conservatives approaches such as the forbidden regions criterion.
Black-box models are used to design a damping input filter to reshape the transient
dynamics of the load converter input current. Thus, avoiding undesirable current
overshoots that produce dangerous drops and oscillations in the input voltage and
consequently driving the converter to unstable operation.







125

5.8 References
[1] F. C. Lee, X. Ming, and W. Shuo, "High Density Approaches of AC to DC
Converter of Distributed Power Systems (DPS) for Telecom and Computers," in
Power Conversion Conference - Nagoya, 2007. PCC '07, 2007, pp. 1236-1243.
[2] E. W. Gholdston, K. Karimi, F. C. Lee, J. Rajagopalan, Y. Panov, and B.
Manners, "Stability of large DC power systems using switching converters, with
application to the International Space Station," in Energy Conversion Engineering
Conference, 1996. IECEC 96. Proceedings of the 31st Intersociety, 1996, pp. 166-
171 vol.1.
[3] J. S. Lai and D. J. Nelson, "Energy Management Power Converters in Hybrid
Electric and Fuel Cell Vehicles," Proceedings of the IEEE, vol. 95, pp. 766-777,
2007.
[4] M. Hankaniemi, M. Karppanen, and T. Suntio, "Load-imposed instability and
performance degradation in a regulated converter," in Proceedings of the IEE
Electric Power Applications, . vol. 153, 2006, pp. 781-786.
[5] K. Zenger, A. Altowati, and T. Suntio, "Dynamic properties of interconnected
power systems a system theoretic approach," in The 1
ST
IEEE Conference on the
Industrial Electronics and Applications, 2006, pp. 1-6.
[6] S. Abe, H. Nakagawa, M. Hirokawa, T. Zaitsu, and T. Ninomiya, "System
stability of full-regulated bus converter in distributed power system," in Twenty-
Seventh IEEE Telecommunications Conference, 2005, pp. 563-568.
[7] S. Schulz, B. H. Cho, and F. C. Lee, "Design considerations for a distributed
power system," in Power Electronics Specialists Conference, 1990. PESC '90
Record., 21st Annual IEEE, 1990, pp. 611-617.
[8] C. Byungcho and B. H. Cho, "Intermediate line filter design to meet both
impedance compatibility and EMI specifications," IEEE Transactions on Power
Electronics, vol. 10, pp. 583-588, 1995.
[9] G. S. Thandi, R. Zhang, K. Xing, F. C. Lee, and D. Boroyevich, "Modeling,
control and stability analysis of a PEBB based DC DPS," IEEE Transactions on
Power Delivery vol. 14, pp. 497-505, 1999.
[10] R. D. Middlebrook, "Input Filter Considerations in Design and Application of
Switching Regulators," in Proceeding on Industry Application Society 1976, pp.
11 - 14.
[11] L. Jinjun, F. Xiaogang, F. C. Lee, and D. Borojevich, "Stability margin
monitoring for DC distributed power systems via perturbation approaches," IEEE
Transactions on Power Electronics, vol. 18, pp. 1254-1261, 2003.
[12] F. Xiaogang, L. Jinjun, and F. C. Lee, "Impedance specifications for stable DC
distributed power systems," IEEE Transactions on Power Electronics vol. 17, pp.
157-162, 2002.
[13] C. M. Wildrick, F. C. Lee, B. H. Cho, and B. Choi, "A method of defining the
load impedance specification for a stable distributed power system," IEEE
Transactions on Power Electronics, vol. 10, pp. 280-285, 1995.
126

[14] Z. Weidong, S. Pekarek, J. Jatskevich, O. Wasynczuk, and D. Delisle, "A model-

in-the-loop interface to emulate source dynamics in a zonal DC distribution
system," Power Electronics, IEEE Transactions on, vol. 20, pp. 438-445, 2005.
[15] Y. Panov and M. M. Jovanovic, "Small-signal measurement techniques in
switching power supplies," in Applied Power Electronics Conference and
Exposition, 2004. APEC '04. Nineteenth Annual IEEE, 2004, pp. 770-776 vol.2.
[16] Y. Panov and M. Jovanovic, "Practical issues of input/output impedance
measurements in switching power supplies and application of measured data to
stability analysis," in Applied Power Electronics Conference and Exposition,
2005. APEC 2005. Twentieth Annual IEEE, 2005, pp. 1339-1345 Vol. 2.
[17] D. M. Mitchell, "Power line filter design considerations for DC-DC converters,"
Industry Applications Magazine, IEEE, vol. 5, pp. 16-26, 1999.
[18] R. W. Erickson, "Optimal single resistors damping of input filters," in The 14
th

IEEE Power Electronics Conference and Exposition, 1999, pp. 1073-1079 vol.2.

6 Chapter
Conclusions and Future Work
6.1 Conclusions
The research reported in this dissertation has focused on the development and
demonstration of two-port black-box models for dc-dc converters. The basic structure of
the model is a hybrid two-port network constructed from frequency response
measurements performed at the converter terminals.
The advantages of this modeling approach can be enumerated as follows:
The models are experimentally obtained and no previous knowledge of the
converter parameter or topology is needed to construct the models.
The models are able to predict steady state and transient voltages and currents at
the converter input and output terminals when under large signal disturbances.
The models are valid until half the switching frequency, meaning that they will
predict average voltages and current values.
The models can be connected in cascade, parallel, and series to form distribution
systems.
The numerical solution of the model is computed by solving a set of algebraic
equations instead of differential equations, which is especially advantageous
when multiple time scales exist.
After linearization of the models at specific operating conditions, linear tools
such as Bode and Nyquist plots can be used for analysis and assessment of the
128

small signal stability of the system, given the condition that all system
eigenvalues are located in the right half plane.

Since a perfect model does not exist, two-port black-box models exhibit the
following disadvantages:
There is not a systematic way to determine how many models are needed to
describe a converter in a specific operating range and with a specific accuracy
range.
The number of models grows exponentially with the number of variables that
determine the operating conditions. This is known as the curse of dimensionality.
To obtain a two-port black-box model, the converter needs to be dynamically
stable or, at least, it needs to be capable of being stabilized by an external control
with known dynamics.
The models become less accurate as any system eigenvalue of the interconnected
system approaches the imaginary axis.
The models are not able to predict the kind of unstable behavior the system will
experience when any of the system eigenvalues cross the imaginary axis. For
instance, the models cannot predict if the system is undergoing a Hopf
bifurcation, a saddle-node bifurcation, a pitchfork bifurcation, or a transcritical
bifurcation.
The main difficulties experienced during the course of this investigation were
addressed in each chapter. For instance, Chapter 1 deals with FRF measurement issues;
Chapter 2 studies subsystem interaction effects on FRF measurements and how to obtain
unterminated FRFs; Chapter 4 extends the application of the modeling approach to
converters with strong nonlinearities whose parameters are slowly variant with respect to
the operating conditions; and Chapter 5 demonstrates the effectiveness of the approach by
showing that the converters black-box models can be used for prediction and analysis of
interconnected systems.
129

However, there are still unresolved issues such as: how to perform an optimal
model partition for polytopic models; how to ameliorate the problem generated by the
curse of dimensionality, how to increase the predictive capabilities of the model when
any of the system eigenvalues are close to the imaginary axis; how to incorporate
verification and validation activities during the model development process; and how to
extend the proposed approach to model three-phase converters. These unresolved issues
constitute opportunities to continue further research on this topic.
6.2 Future Work
Future research on this topic is needed to find a solution for the unresolved issues
listed in the section above; however, there is a special interest in extending the
application of the black-box modeling approach for three-phase converters because of
their extensive application in power conversion but, specially the models will be useful to
study a system of interconnected three-phase converters in a microgrid. This section
presents some initial ideas for these pursuits.
The General Three-phase Converter Black-box Model in the Synchronous d-q Reference
Frame
The conceptual black-box model for three-phase converter has a similar structure
to the one used for the dc-dc black-box model except that the input and output variables
are now vectors defined in (2) and the transfer functions are actually matrices of transfer
functions defined in (3). The model (1)-(3) in Fig. 13 can be easily reduced to represent
dc-ac or ac-dc converters. For a dc to three-phase ac inverter, the input q-channel could
be set to zero, i.e. v
iq
i
iq
0, which then simplifies the matrices to

o
i
i i
o o
i
o
i
v
H Y
Z G
i
v
) ( ) (
) ( ) (
s s
s s
(1)
where
130

od id id id
oq iq iq iq
v v i i
, , ,
v v i i

= = = =


o i o i
v v i i
(2)
dd dq dd dq
qd qq qd qq
dd dq dd dq
qd qq qd qq
G (s) G (s) Z (s) Z (s)
, ,
G (s) G (s) Z (s) Z (s)
Y (s) Y (s) H (s) H (s)
, .
Y (s) Y (s) H (s) H (s)

= =



= =


o o
i i
G Z
Y H
(3)


Fig. 6.1 General Black-box model structure for three-phase converters
For instance, for a dc to three-phase ac inverter, the input q-channel could be set
to zero, i.e. viq iiq 0, which then simplifies the matrices as

dd qd
dd qq
dq qq
dd
id dc d dc dd
qq
Z (s) Z (s)
G (s) G (s) ,
Z (s) Z (s)
H (s)
v , i , Y (s),
H (s)

= =




= = = =


o o
i i
G Z
v i Y H
. (4)

Rearranging the terms yields

d od dd qd dc
q oq dq qq d
dc in id iq q
v G (s) Z (s) Z (s) v
v G (s) Z (s) Z (s) i
i Y (s) H (s) H (s) i



=






.
(5)
The circuit representation of (5) is given in Fig. 6.2.
v
i
v
Z ii
id
id
dd
Y
iq dq
v Y
od dd
i H
oq dq
i H
iq
iq
i
qq
Y
id qd
v Y
od qd
i H
oq qq
i H
id dd
v G
dd
iq dq
v G
oq dq
i Z
od
od
v
iq qq
v G
qq
Z
id qd
v G
od qd
i Z
oq
i
oq
v
dd
Y v G
oq dq
i Z
od
iq qq
v G
qq
Z
id qd
v G
od qd
i Z
oq
i
oq
v
131

Fig. 6.2 Black-box model structure for a open-loop voltage source inverter
Once the internal structure for the black-box model is defined, then an
experimental setup is needed to measure the unterminated FRF that will capture the
internal converter dynamics. For this, an experimental setup for the measurement of the
unterminated FRF in three-phase converters is proposed. The setup follows the same
principles as the one used for dc-dc converters. The unterminated FRF can be obtained by
feeding the three-phase converter with a near ideal voltage source and loaded with a
current sink. A near ideal voltage source means a source with low output impedance, and
a current sink load refers to a load whose input current is tightly regulated. These two
requirements are needed to minimize the subsystem interactions among the converter, the
source and the load that could affect the obtainment of unterminated measurements as
demonstrated in Chapter 3.
Fig. 6.3 shows a proposed measurement setup for which the converter under test
(CUT) is loaded with a three-phase boost rectified through a transformer that serves four
purposes: to increase the stiff current characteristic of the load, to limit the switching
noise generated by the active rectifier, to provide isolation, and to eliminate any dc signal
injection into the CUT.
The operating conditions are set using the commanded currents id* and iq*.
Current injection circuits in the load side are eliminated by adding the disturbance signals
id~ and iq~ to the commanded current. This imposes a design constraint on the load
current loop in that the current control bandwidth should be at least twice the maximum
frequency of the frequency sweep.
132

Fig. 6.3 Measurement setup for three-phase converters
Given that the transfer function for these models are defined in the synchronous
reference frame, a digital signal processor card is needed to convert the disturbances and
responses from the CUT into the d-q rotating reference frame so they can be processed by
the network analyzer to obtain the FRFs that are defined by (6)
~
d
od ~
in
v
G
v
=
~
q
oq ~
in
v
G
v
=

~
d
dd ~
d
v
Z
i
=

~
q
qq ~
q
v
Z
i
=

~
d
dq ~
q
v
Z
i
=
~
q
qd ~
d
v
Z
i
=
(6)
~
in
i ~
in
i
Y
v
=

~
in
dd ~
d
i
H
i
=

~
in
qq ~
q
i
H
i
=

.
The black-box modeling approach developed for three-phase converters is
implemented for a unregulated voltage source inverter (VSI). Use of an unregulated VSI
is an appropriate approach to prove this modeling concept, considering that the dynamic
behavior of this kind of converter is nearly linear in a wide range of operating conditions;
thus, nine linear transfer functions are enough to describe the converter dynamics in a
wide operating region.
The original methodology indicates that the FRFs are obtained using the
experimental setup shown in Fig. 6.3, and then the FRFs are processed using system
identification to obtain the actual transfer function to be used to build the model. Since, in
133

this case, the setup is not available, the transfer functions are obtained by linearization of
the VSI average model.
God Goq
Zdd Zqq
Zdq Zqd
Hid Hiq
Yi

Frequency (rad/sec) Frequency (rad/sec)
Fig. 6.4 Voltage source inverter FRFs from linearization of the converter average model

10
-2
10
0
A
m
p
l
i
t
u
d
e
10
2
10
3
10
4
-200
-100
0
P
h
a
s
e

(
d
e
g
r
e
e
s
)
10
0
A
m
p
l
i
t
u
d
e
10
2
10
3
10
4
-400
-200
0
P
h
a
s
e

(
d
e
g
r
e
e
s
)
10
-2
10
0
A
m
p
l
i
t
u
d
e
10
2
10
3
10
4
-300
-200
-100
0
P
h
a
s
e

(
d
e
g
r
e
e
s
) 10
-2
10
0
A
m
p
l
i
t
u
d
e
10
2
10
3
10
4
-300
-200
-100
0
P
h
a
s
e

(
d
e
g
r
e
e
s
)
10
0
A
m
p
l
i
t
u
d
e
10
2
10
3
10
4
-400
-200
0
P
h
a
s
e

(
d
e
g
r
e
e
s
)
10
0
A
m
p
l
i
t
u
d
e
10
2
10
3
10
4
-200
0
200
P
h
a
s
e

(
d
e
g
r
e
e
s
)
10
0
A
m
p
l
i
t
u
d
e
10
2
10
3
10
4
-200
-100
0
P
h
a
s
e

(
d
e
g
r
e
e
s
)
F ( d/ )
10
0
A
m
p
l
i
t
u
d
e
10
2
10
3
10
4
-300
-200
-100
0
100
P
h
a
s
e

(
d
e
g
r
e
e
s
)
F ( d/ )
10
-2
10
0
A
m
p
l
i
t
u
d
e
10
2
10
3
10
4
-100
0
100
P
h
a
s
e

(
d
e
g
r
e
e
s
)
F ( d/ )
134

The model validation is performed in the time domain in which conventional

averages models are compared to black-box models. For this validation, two cases are
considered. First, only a step load disturbance is considered, and, second, simultaneous
load and input voltage disturbances are considered.
For the first case, the load disturbance is applied in the d channel and the input
voltage is kept constant (ideal source). Fig. 6.5 compares the predicted transient
responses using average models and black-box models. As expected, based on the linear
characteristics of the VSI, the linear black-box model agrees 100% with the nonlinear
average model.
Output voltage Vd Output voltage Vq
V
o
l
t
a
g
e

(
V
)

V
o
l
t
a
g
e

(
V
)

Input current Idc Input voltage Vin
C
u
r
r
e
n
t

(
A
)

V
o
l
t
a
g
e

(
V
)

Time (sec) Time (sec)

Fig. 6.5 Comparison between average model and black-box model when under load disturbance
The second test compares the prediction of the two models for a not-so-evident
case. Basically, in both a step load disturbance together with a input voltage disturbance
are present. This case is recreated by placing a 2 ohm resistor in series with the source,
such that when the load step occurs, it will produce a voltage drop proportional to the
input current. The results are depicted in Fig. 6.6.


0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095
645
650
655
Output Voltage Vd
0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095
-4
-2
0
2
4
p g q
0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095
8
10
12
14
16
p
0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095
799
799.5
800
800.5
801
135

Output voltage Vd Output voltage Vq
V
o
l
t
a
g
e

(
V
)

V
o
l
t
a
g
e

(
V
)

Input current Idc Input voltage Vin
C
u
r
r
e
n
t

(
A
)

V
o
l
t
a
g
e

(
V
)

Time (sec) Time (sec)

Fig. 6.6 Comparison between average model and black-box model when under simultaneous
variation of the input voltage and load current
The presence of the resistor produces not only a change on the operating
conditions but also a change of the transient response making it more damped. Despite
the changes introduced in the circuit, the black-box models still holds while it
incorporates the added element in order to produce the same response predicted by
average models. These result are encouraging and serve as a proof of concept for further
research on this topic.

0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095

626
628
630
632
634
636
638
p g
( )
V
o
l
t
s
0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095
-1.5
-1
-0.5
0
0.5
V
o
l
t
s
g
0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095
774
776
778
780
782
784
V
o
l
t
s
p g
0.06 0.065 0.07 0.075 0.08 0.085 0.09 0.095
8
9
10
11
12
13
Ti ( )
A
m
p
s
136

A. Appendix
Linear model from FRF measurement
G
o

A
m
p
l
i
t
u
d
e


P
h
a
s
e

(
D
e
g
)

Frequency(rad/sec) Frequency(rad/sec)
Measurement Model
-2981.S1(s - 1.9SS 1u
4
)(s + 1.u79 1u
4
)(s
2
+ 9977s - 1.u8S 1u
8
)
(s - S.884 1u
4
)(s
2
+ 2SS6 1u
4
s + 4.u7S 1u
7
)(s
2
+ 1.22S 1u
4
s + 1.u77 1u
8
)


Fig. A.1 Open loop FRFs from measurements and identified models

% reading from a data file created by another m-file that reads the
%data directly from the Network analyzer using a USB-GPIB interface.
%Measured back current gain two identification algorithms are used PEM
and FITFRD more information can be found in Matlab Help

load Hi20V2A.mat
S21 = size (chan1.y); %Number of data points
AMP21= zeros(S21, 1); % preallocate vector
FREQHi= chan1.x*(2*pi); % Frequency (from Hz to rad/seg)

for ix = 1:1:S21(1);
AMP21(ix,1)= 1*10^(chan1.y(ix)/20); %transform magnitude in dB to absolute value
end
PHA21=chan2.y; % store phase information in a variable PHA21

i = sqrt(-1);
zfr21 = AMP21.*exp(i*((PHA21*pi/180))); %Create a complex variable (magnitude and phase)

Hi20V2idfrd= idfrd(zfr21 ,FREQHi,0); % Create and idfFRD object used by PEM

Hi20V2frd=frd(zfr21,FREQHi); % Create and FRD object used by fitfrd

Hi20V2PemModel = pem(Hi20V2idfrd,3,'Foc','stability'); % model using PEM

Hi20V2FitfrdModel= fitfrd(Hi20V2frd,3); %Model using fitfrd

10
-1
10
0
10
3
10
4
10
5
-200
-100
0
137

figure
bode(Hi20V2idfrd,Hi20V2PemModel,Hi20V2FitfrdModel) % bode comparing data vs pem
vs fitfrd

figure
step(Hi20V2PemModel,Hi20V2FitfrdModel) % will tell if the system is stable

% Comments: PEM is slower than FITFRD because uses a different
%identification algorithm. PEM provides more options to the user
%therefore, more opportunities for automatization of the identification
%process. FITFRD is more prone to produce models with positive
eigenvalues especially when the order of the model is increased.


















138

B. Appendix
Graph Comparison of Experimental and
Simulation Results
A Matlab (BuckOpenLoopIdealVali.m) file and a Simulink file (BuckOpenLoopVal.mdl)
are used for this example
% This M-file load data into the Matlab work space, call and run a
% simulink implementation of a linear black-box model

clear all
close all

s=tf('s')

Go=0.998*(-2987.4891 *(s-1.953e004)* (s+1.079e004) *(s^2 + 9977*s +
1.083e008))/((s+5.884e004) *(s^2 + 2366*s + 4.075e007) *(s^2 +
1.225e004*s + 1.077e008))
%
Zo=(62500 *(s+2038) *(s^2 + 1.208e005*s +
9.319e009))/((s+5.506e005)^2* (s^2 + 2393*s + 4.073e007))

Yi=(22851011.0854 *s *(s^2 + 2283*s + 5.303e007))/((s+4.057e005)^2
*(s^2 + 2251*s + 4.12e007))
Input Voltage v
1
Output Voltage v
2

V
o
l
t

Input Current i
1
Output Current i
2

A
m
p

Time(msec) Time(msec)
Experimental , Experiemental Average , Model
Fig. B.1 Open-loop FRFs from measurements and identified models
8 8.5 9 9.5 10 10.5 11 11.5
3
20
20.2
20.4
20.6
20.8
8 8.5 9 9.5 10 10.5 11 11.5
4.4
4.6
4.8
5
5.2
8 8.5 9 9.5 10 10.5 11 11.5
3
0.5
1
1.5
2
2.5
8 8.5 9 9.5 10 10.5 11 11.5
3
2.5
3
3.5
4
4.5
139

Hi=(731.6297 *(s+2.623e004)*(s^2 + 1.256e004*s + 2.148e008)* (s^2 +
7.816e004*s + 6.72e009)* (s^2 - 7526*s + 1.117e010))/((s+6.67e005)*
(s+3.382e004)* (s^2 + 2308*s + 4.023e007) *(s^2 + 1.171e004*s +
2.023e008)* (s^2 + 5.443e004*s + 5.975e009))


%% Source Output Impedance


Zs=(418204.751*(s+7.656e004) *(s+1469)* (s+240.2)* (s^2 + 3.284e004*s
+ 2.898e008) *(s^2 + 2850*s + 1.025e009)*(s^2 + 3286*s + 4.275e009)*
(s^2 + 5048*s + 9.458e009))/((s+9.153e004) *(s+6.85e005)^2* (s+481.8)*
(s^2 + 2184*s + 1.091e007) *(s^2 + 2831*s + 1.026e009) *(s^2 +
3229*s + 4.274e009)*(s^2 + 4948*s + 9.443e009))


%% Load Step From 2.5 to 4.5 Amps Electronic Load

load i2Ideal.txt %Load output current experimental data

[t2]=0.008+i2Ideal(:,1); % shift data by 0.0008 sec
[i2]=100*i2Ideal(:,2); %scaling factor depend on current amplifier
settings

load i1Ideal.txt
[t1]=0.008+i1Ideal(:,1); %Load input current experimental data

[i1]=100*i1Ideal(:,2);
[i1a]=smooth(i1,800,'moving'); %moving average experimental data

load v1Ideal.txt
[v1]=v1Ideal(:,2); %Load input voltage experimental data
[v1a]=smooth(v1Ideal(:,2),800,'moving'); %moving average experimental
data

load v2Ideal.txt
[v2]=v2Ideal(:,2); %Load output voltage experimental data
[v2a]=smooth(v2Ideal(:,2),800,'moving');%moving average experimental
data

%% Run simulink model
sim('BuckOpenLoopVal.mdl',[0 12e-3]);

%% plot results
plotfigures(time1,voltagev1)
plotfigures(time1,voltagev2)
plotfigures(time1,currenti1)
plotfigures(time1,currenti2)

140

C. Appendix
Obtaining un-terminated FRFs
Gom vs. Go Zom vs. Zo
M
a
g
(
A
b
s
)

P
h
a
s
e

Yim vs. Yi Him vs. Hi
M
a
g
(
A
b
s
)

P
h
a
s
e

Frequency(rad/sec) Frequency(rad/sec)

Unterminated FRFs Coupled FRFs

Fig. C.1 Coupled and unterminated FRFs

% Delta Bus Converter Unterminated Measurements

% Experimental FRFS having interaction with the source and the load
% to obtain the un-terminated FRFs the following steps need to be
% followed .
% 1. Obtain un-terminated models of the source and the load
% 2. Measure FRFs at the converter terminal when connected to the
% source and load in step 1
% 3. Obtain and additional FRF when the converter is loaded with and
% electronic load
% 4. Solve for unterminated Zo using the quadratic equation (16) in
% Chapter 3
% 5. Once havin Zo solve for unterminated Hi and Go using equations
% (10) and (11) from Chapter 3
% 6 Use appendix A to obtain a model from the un-terminated FRFs

% Note: When solving equation (16) all FRFs needs to have the same
% number of points and same frequency range. In case is not possible to
% meet this requirement for some FRFs. Find a linear model for those
% measurements first, the quadratic equation will accept the model.
%%
close all
1 1 1 1
6
10
0
3 4 5
-300
-100
0
2
10 10 10 10
6
10
0
10
3 4 5
0
50
100
-100
10
0
10
-2
10
3
10
4
10
5
10
6
0
100
10 10 10 10
6
10
0


3 4 5
-200
-100
0
-200
141

clear all

%% Unterminated Source Model
% The source is loaded with a 25 ohms resistor

% Obtain model for the 25 ohm load resistor
N11='R25FRF.txt' ;
load(N11);
A11=eval(strtok(char(N11),'.'));
clear (strtok(char(N11),'.')) ;
%Number of data points
S11 = size (A11,1);
%Frequency data
FREQ11= A11(1:S11,1)*2*pi;
%create Variables for the Magnitude, frequency and Phase
for ix = 1:S11;
AMP11(ix,1)=A11(ix,2);
end
PHA11=A11(1:S11,3);
i = sqrt(-1);
R25frf = AMP11.*exp(i*PHA11*pi/180);
Ts = 0;

%creates FRD object
R25lti=frd(R25frf,FREQ11, 'Hz');

% Creates and Identification Object
gfrR25 = idfrd(1/R25lti,FREQ11,0);
% obtain the model using PEM
m12 = pem(gfrR25 ,1,'Foc','stability');
rx=minreal(tf(m12));
G12bus=rx(1,1);
s=tf('s')
% Resistor impedance and admmittance model
YR25= 2.156e004/(s + 5.262e005)
ZR25=4.6391e-005 *(s+5.262e005)
bode(gfrR25,G12bus)

%% Source output Impedance Model

%Load measured source output impedance data
N12='ZsLow.txt';
load(N12);
A12=eval(strtok(char(N12),'.'));
clear (strtok(char(N12),'.')) ;
S12 = size (A12,1);
FREQ12= A12(1:S12,1)*(2*pi);

for ix = 1:S12;
AMP12(ix,1)=0.1*10^(A12(ix,2)/20);
end
PHA12=A12(1:S12,3);

i = sqrt(-1);
ZsmedFRF = AMP12.*exp(i*(PHA12*pi/180));
142

Ts = 0;

s=tf('s')
YR25= 2.156e004/(s + 5.262e005);

ZsLowltia=frd(ZsmedFRF,FREQ12, 'Hz');

%decouple the source and 25 resistor load dynamics
ZsLowlti=ZsLowltia/(1-ZsLowltia*YR25);

% find a model for the decoupled source output impedance

%this case is easier to find the input admmittance model invert it
later
gfrZsLow = idfrd(1/ZsLowlti,FREQ12,0);
m12 = pem(gfrZsLow,11,'Foc','stability');
rx=minreal(tf(m12));
G12bus=rx(1,1);
bode(gfrZsLow,G12bus)

% final decoupled source output impedance model
YsLow=(3375916.8772 *(s^2 + 2825*s + 1.847e007)* (s^2 + 8.704e004* s +
4.856e009)* (s^2 + 1.049e005* s + 8.173e010)* (s^2 + 5.446e004* s +
1.133e011) * (s^2 + 7128* s + 1.245e011))/((s+3.24e004) * (s+7514) *
(s+3089)* (s^2 + 2.053e005* s + 1.108e010)* (s^2 + 1.133e005* s +
8.162e010)* (s^2 + 3.313e004* s + 1.127e011)* (s^2 + 6669* s +
1.243e011))


%% Converter Load Model ( 1 ohm resistor )

N11='R1FRF.txt' ;
load(N11);
A11=eval(strtok(char(N11),'.'));
clear (strtok(char(N11),'.')) ;
S11 = size (A11,1);
FREQ11= A11(1:S11,1)*2*pi;
for ix = 1:S11;
AMP11(ix,1)=A11(ix,2);
end
PHA11=A11(1:S11,3);
i = sqrt(-1);
R1frf = AMP11.*exp(i*PHA11*pi/180);
Ts = 0;

R25lti=frd(R1frf,FREQ11, 'Hz');

gfrR1 = idfrd(1/R25lti,FREQ11,0);
m12 = pem(gfrR1 ,1,'Foc','stability');
rx=minreal(tf(m12));
G12bus=rx(1,1);

% Final load model
s=tf('s')
143

YR1= 8.399e004/(s + 8.501e004)

%% Measured Audiosuceptibility
N11='GORLZ1.txt'
load(N11);
A11=eval(strtok(char(N11),'.'));
clear (strtok(char(N11),'.')) ;
S11 = size (A11,1); %Number of data points
FREQ11= A11(1:S11,1)*(2*pi);
for ix = 1:S11;
AMP11(ix,1)=1*10^(A11(ix,2)/20); %create Variables
for the Magnitude, frequency and Phase
end
PHA11=A11(1:S11,3);
i = sqrt(-1);
zfr11 = AMP11.*exp(i*PHA11*pi/180);

%Create FRD object
Gom=frd(zfr11,FREQ11);

%% Measured Output Imprdance Zom

N12='ZORLZ1.txt';
load(N12);
A12=eval(strtok(char(N12),'.'));
clear (strtok(char(N12),'.')) ;
S12 = size (A12,1);
FREQ12= A12(1:S12,1)*(2*pi);
for ix = 1:S12;
AMP12(ix,1)=0.1*10^(A12(ix,2)/20);
end
PHA12=A12(1:S12,3);

i = sqrt(-1);
ZsmedFRF = AMP12.*exp(i*(PHA12*pi/180));

%Create FRD object
Zom=frd(ZsmedFRF,FREQ12, 'Hz');




%% Converter input admmitance when loaded with a 1 ohm resistor

N21='YIRLZ1.txt';
load(N21);
A21=eval(strtok(char(N21),'.'));
clear (strtok(char(N21),'.')) ;
S21 = size (A21,1);
FREQ21= A21(1:S21,1)*(2*pi);
for ix = 1:S21;
AMP21(ix,1)=10*10^(A21(ix,2)/20);
end
144

PHA21=A21(1:S21,3);
i = sqrt(-1);
zfr21 = AMP21.*exp(i*((PHA21*pi/180+pi)));
%Create FRD object
Yimm=frd(zfr21,FREQ21);

G21lti=frd(Yimm,FREQ21, 'Hz');
Yi=frd(zfr21,FREQ21);
Yis=spafdr(Yi,100)
gfrYi= idfrd(Yis);
m12 = pem(gfrYi ,7,'Foc',[10^3 10^7]);
rx=minreal(tf(m12));
G12bus=rx(1,1);

s=tf('s')
Yim=(691069019.5796 *(s-2.72e006)* (s+2276) *(s^2 + 9.645e004*s +
1.23e011)* (s^2 + 2.852e005*s + 1.214e012))/((s+3.238e007)*(s-
1.155e007) *(s+1.131e005)* (s^2 + 3.625e004*s + 1.174e011) *(s^2 +
1.616e005*s + 1.221e012))
bode (Yim,G12bus)
%% Converter input admmitance loaded with an electronic load

N21='YIELZ1.txt';
load(N21);
A21=eval(strtok(char(N21),'.'));
clear (strtok(char(N21),'.')) ;
S21 = size (A21,1);
FREQ21= A21(1:S21,1)*(2*pi);
for ix = 1:S21;
AMP21(ix,1)=10*10^(A21(ix,2)/20);
end
PHA21=A21(1:S21,3);
i = sqrt(-1);
zfr21 = AMP21.*exp(i*((PHA21*pi/180+pi)));

G21lti=frd(zfr21,FREQ21, 'Hz');
Yi=frd(zfr21,FREQ21);
Yis=spafdr(Yi,100)
gfrYi= idfrd(Yis);
m12 = pem(gfrYi ,8,'Foc',[10^3 10^7]);
rx=minreal(tf(m12));
G12bus=rx(1,1);

Yi=(-7654214.084* s* (s-8.356e006)* (s+3478) *(s^2 + 8.802e004*s +
1.081e011) *(s^2 + 2.415e005*s + 1.162e012))/((s+1.154e005)* (s+3479)
*(s^2 + 5.325e004*s + 1.028e011)* (s^2 + 1.468e005*s + 1.18e012)* (s^2
+ 2.447e006*s + 1.215e013));




%% Converter Back Current Gain

N22='HIRLZ1.txt';
load(N22);
145

A22=eval(strtok(char(N22),'.'));
clear (strtok(char(N22),'.')) ;
S22 = size (A22,1);
FREQ22= A22(1:S22,1)*(2*pi);
for ix = 1:S22;
AMP22(ix,1)=10^(A22(ix,2)/20);
end
PHA22=A22(1:S22,3);
i = sqrt(-1);

zfr22=AMP22.*exp(i*PHA22*pi/180);
Ts = 0;

Him=frd(zfr22,FREQ22, 'Hz');

%% Obtaining Unterminated Measurments

s=tf('s');

% Source Output impedance model
Zs=(16.5547*(s+3.24e004)* (s+7514) *(s+3089) *(s^2 + 2.053e005*s +
1.108e010)* (s^2 + 1.133e005*s + 8.162e010)*(s^2 + 3.313e004*s +
1.127e011)* (s^2 + 6669*s + 1.243e011))/((s+5.589e007)* (s^2 + 2825*s +
1.847e007) *(s^2 + 8.704e004*s + 4.856e009)* (s^2 + 1.049e005*s +
8.173e010)*(s^2 + 5.446e004*s + 1.133e011) *(s^2 + 7128*s + 1.245e011))

% Resistor admitance model
YRL=8.399e004/(s + 8.501e004) ;

%load admmittance = 1 ohm resistor admmitance
Yl=YRL;

Yi=(-7654214.084* s* (s-8.356e006)* (s+3478) *(s^2 + 8.802e004*s +
1.081e011) *(s^2 + 2.415e005*s + 1.162e012))/((s+1.154e005)* (s+3479)
*(s^2 + 5.325e004*s + 1.028e011)* (s^2 + 1.468e005*s + 1.18e012)* (s^2
+ 2.447e006*s + 1.215e013));
% Equivalent output impedance
Zoeq=Zom/(1-Zom*Yl);
%coefficients of the quadratic equation
a=Yl^2*(1+Zs*Yim)*Zs*Gom*Him/(1+Zs*Yi);
b=1+2*Yl*(1+Zs*Yim)*Zs*Gom*Him/(1+Zs*Yi);
c=(1+Zs*Yim)*Zs*Gom*Him/(1+Zs*Yi)-Zoeq;
% solve quadratic equation
X=b^2-4*a*c;

[response,freq] = frdata(X);

N=frd(response.^0.5,FREQ11);
%obtain unterminated Zo
Zo=(-b+(N))/(2*a);

Zo2=(-b-N)/(2*a);

%obtain unterminated Hi
146

Hi=Him*(1+Yl*Zo)*(1+Zs*Yim);
%obtain unterminated Go
Go=Gom*(1+Yl*Zo);

Yim2=Yi+(Gom*Yl*Hi);


%% System Identification of the Unterminated FRF's

%Go Model
gfr11 = idfrd(Go,FREQ11,0);
sgfr11= spafdr(gfr11,100)
m11 = pem(gfr11,8,'Foc','stability');
rx=minreal(tf(m11));
G11bus=rx(1,1);
figure
bode(sgfr11 ,G11bus)



%Zo Model

gfr11 = idfrd(Zo,FREQ11,0);
sgfr11= spafdr(gfr11,100)
m11 = pem(sgfr11,6,'Foc',[10^3 10^7]);
rx=minreal(tf(m11));
G11bus=rx(1,1);
figure
bode(sgfr11 ,G11bus)


%Hi Model

gfr11 = idfrd(Hi,FREQ11,0);
sgfr11= spafdr(gfr11,100)
m11 = pem(sgfr11,6,'Foc','stability');
rx=minreal(tf(m11));
G11bus=rx(1,1);
figure
bode(sgfr11 ,G11bus)






147

D. Appendix
Flyback Polytopic Model
Output Impedance Z
o
2D Output Impedance Z
o
3D

Fig. D.1 flyback FRF plot 2D and 3D

%% Plot 2D and 3D view of experimental FRFs

close all
clear all
%% Load FRF Experimental FRFs
load ZomFlyBkCL5A20V
x20=[ 1.7 2 3 4 5 ]
y20=chan1.x;
z5=chan1.y;
load ZomFlyBkCL4A20V
z4=chan1.y;
load ZomFlyBkCL3A20V
z3=chan1.y;
load ZomFlyBkCL2A20V
z2=chan1.y;
load ZomFlyBkCL1o7A20V
z1o7=chan1.y;

Zo20=[z1o7, z2 , z3 , z4 , z5]-20;

%% Plot experimental FRFs in 2D

figure
axes('GridLineStyle','--','FontSize',30)
semilogx(y20,Zo20,'LineWidth',6)
xlim([100 20000])
ylim('auto')
xlabel('Frequency (Hz)','FontSize',30)
ylabel('Magnitude (dB)','FontSize',30)
title('Z_o Output Impedance','FontSize',30)
grid on




10
2
10
3
10
4
-40
-35
-30
-25
-20
-15
-10
Frequency (Hz)
M
a
g
n
i
t
u
d
e

(
d
B
)
148

%% Plot experimental FRFs in 3D

figure1 = figure;
% Create axes
axes1 = axes('Parent',figure1,'YScale','log','YMinorTick','on',...
'YMinorGrid','on',...
'YDir','reverse',...
'FontSize',30);
% Uncomment the following line to preserve the Y-limits of the axes
ylim([50 2e+004]);
% Uncomment the following line to preserve the Z-limits of the axes
zlim([-45 -10]);
view([-116.5 34]);
grid('on');
hold('all');

% Create surf
surf(x20,y20,Zo20,'Parent',axes1,'AmbientStrength',0.5,...
'FaceLighting','gouraud',...
'FaceColor','interp',...
'EdgeColor','none');

% Create light
light('Parent',axes1,'Position',[1 0 0]);

% Create light
light('Parent',axes1,'Style','local','Position',[6.844 -1.811e+005
324.9]);

% Create xlabel
xlabel('Output Current
(Amp)','HorizontalAlignment','right','FontSize',30);

% Create ylabel
ylabel('Frequency (Hz)','FontSize',30,'HorizontalAlignment','left');

% Create zlabel
zlabel('Magnitude (dB)','FontSize',30);







149

%% Load set of linear models for Hi and Zo obtained from FRF

measurments and run a simulink file containing a black-box polytopic
model

Input voltage (V) Input current (A)
Output voltage (V) Ouptut current (A)
Experimental Experimental Average Model
Fig. D.2 Flyback transient response under large disturbances
clear all
close all

load FlyBkModelsWO16V
load FlyBkModelsWO18V
load FlyBkModelsWO20V

% Go and Yi present small variation with the operating conditions.
Thus they are not modeled as a polytopic structure
s=tf('s');
Go=(0.062335 *s *(s+282) *(s^2 + 3.693e004*s + 1.084e009) *(s^2 -
4.027e004*s + 2.798e009))/((s+2.351e005) *(s+5438) *(s+1112)
*(s+150.8)* (s^2 + 1.53e004*s + 2.329e009))

Yi=(18.3371 *s *(s+2.654e006)* (s+2621))/((s+1.909e006) *(s+1.257e005)
*(s+3029))

%% Load source model

Zs3ohms=(8.7023* (s+6.254e005) *(s+5.869e004)* (s+103.5) *(s^2 +
53.66*s + 1.222e006) *(s^2 + 2237*s + 1.693e007) *(s^2 + 4809*s +
7.02e007)* (s^2 + 1.157e004*s + 1.426e008))/((s+1.832e006)
*(s+6.12e004)* (s+114.6)* (s^2 + 52.66*s + 1.22e006)* (s^2 + 2193*s +
1.669e007)* (s^2 + 4858*s + 6.942e007)* (s^2 + 1.134e004*s +
1.387e008))
0.05 0.052 0.054 0.056 0.058 0.06 0.062
16
17
18
19
20
21
0.05 0.052 0.054 0.056 0.058 0.06 0.062
0.8
1
1.2
1.4
1.6
1.8
2
1
0.05 0.052 0.054 0.056 0.058 0.06 0.062
5
5.05
5.1
5.15
5.2
5.25
Ti (S )
2
0.05 0.052 0.054 0.056 0.058 0.06 0.062
2
2.5
3
3.5
4
4.5
5
Ti (S )
p
2
150

%% Load experimental data step electronic load from 2 Amps to 3.5
amps to 4.5 Amps
%
load i2FlyBk23o54o5AWORL.txt

ib=i2FlyBk23o54o5AWORL;
[t2]=0.05+ib(:,1);
[i2]=100*ib(:,2);
[i2a]=smooth(i2,100,'moving');


load i1FlyBk23o54o5AWORL.txt

ia=i1FlyBk23o54o5AWORL;
[t1]=0.05+ia(:,1);
[i1]=100*ia(:,2);
[i1a]=smooth(i1,100,'moving');


load v1FlyBk23o54o5AWORL.txt
va=v1FlyBk23o54o5AWORL;
[v1]=va(:,2);
[v1a]=smooth(v1,100,'moving');


load v2FlyBk23o54o5AWORL.txt
vb=v2FlyBk23o54o5AWORL;
[v2]=vb(:,2);
[v2a]=smooth(v2,100,'moving');


%% Run simulink model
sim('FlyBkValWO.mdl',[0 632e-3]); % for details of the Simulink
%implementation please see the file
%% plot results
plotfiguresFbk(time1,voltagev1)
plotfiguresFbk(time1,voltagev2)
plotfiguresFbk(time1,currenti1)
plotfiguresFbk(time1,currenti2)