Model Order Reduction for COMSOL - A Compact Model of a Wireless

Power Transfer System

J. Pico, D. Hohlfeld, and T. Bechtold
Institute of Electronic Appliances and Circuits, Department of Computer Science and Electrical Engineering,
University of Rostock,
Albert-Einstein Str. 2, 18059 Rostock, Germany, tamara.bechtold@uni-rostock.de

Abstract: This work presents the application of

mathematical methods of model order reduction
(MOR) for automatic generation of highly
accurate, compact models for wireless power
transfer (WPT) systems. A block two-sided
Second Order Arnoldi (SOAR) algorithm was
implemented to automatically compute a compact
model, which is highly accurate, but demands
several orders of magnitude smaller CPU time.
Thus, it can be used for the co-simulation of the
WPT setup with electrical circuitry Figure 1. Resonant Inductive Coupling
Keywords: Wireless power transfer, Model order
reduction, Block two-sided second order Arnoldi, In 2007, an initial report [5] demonstrated
System-level simulation. efficient WPT over distances up to eight times the
radius of the coil. An efficiency of approximately
1. Introduction 40% was reached over distances of 2 m (see
Figure 2).
Nowadays, a new technological trend towards
Wireless Power Transfer is driven by modern
applications, such as medical implants, mobile
devices, electrical vehicles, drones, and home
appliances, which all profit of cable-free energy
supply [1].

Available charging systems for electrical vehicles

(EVs), which use cables to connect the vehicle to
the grid, feature many disadvantages, such as
dirtiness, broken cables and necessity to establish
an electrical connecting also during rainy
weather. The latter also impose safety risks. To
overcome those issues several WPT charging Figure 2. Comparison of experimental and theoretical
solutions have been implemented as reported in efficiencies as functions of the WPT distance in [5].
[2] and [3].
Another experiment [6] considers aluminum
Inductive power transfer [4] is one of the most plates located next to the resonant coils. The
widely used methods for WPT. It employs the maximum efficiency larger than 94% was
principle of magnetic induction as used in a achieved over a distance of 60 cm, for a coil
transformer, by taking advantage of the resonant radius of 30 cm at a resonant frequency of
coupling effect among coils of two resonant 8.8 MHz (see Figure 3).
inductor-capacitor circuits (see Figure 1).

Excerpt from the Proceedings of the 2016 COMSOL Conference in Munich

 In Figure 5 a WPT an equivalent circuit is shown,
with each coil being serially connected to a
capacitor and a resistor. The latter represents the
coil’s inner resistance. Mutual inductance is
marked with M.

Figure 3. (a) WPT system in close proximity of an

aluminum plate. (b) Transfer efficiency of WPT system
with only a single metal plate (c) WPT system with two
aluminum plates. (d) Transfer efficiency of WPT
system with aluminum plates near both coils. Source:
[6].

In 2014, a full recharging system was reported for Figure 5. WPT equivalent circuit.
EVs [7]. The system consisted of a pair of The resonance frequencies of both RLC-circuits,
rectangular coils each attached to a ferrite layer, can be formulated as:
which magnetically shields the car (see Figure 4).
Aluminum plates have been used as well, and it 1 1
reaches a power efficiency of over 90%. , (1)

A further equivalent circuit, with the mutual

inductance represented as an inductor between
primary and secondary resonator, is shown in
Figure 6.
Figure 4. Magnetic designs for rectangular coils.
Source: [7].

Our work aims at providing a system-level model

of a WPT system for optimization of power
transfer efficiency and tolerance analysis. In
contrast to existing work [8] our contribution uses
Figure 6. Equivalent circuit of coupled resonators.
an accurate numerical model for the inductive
component and enables time-efficient modeling
The Kirchhoff’s voltage law equations for
by mathematical model order reduction
primary and secondary circuits can be formulated
as follows:
2. WPT Circuit Analysis

The WPT system consists of two inductively

0
coupled coils. One inductor is connected to a
power source, which drives a sinusoidal current. (2)
The voltage induced in the second inductor is 1⁄
supplied to a resistive load, which dissipates the
transferred power. Matching capacitors are 1⁄
connected in series with the inductors to
compensate for their impedance at the operating From (2), the circuit input impedance can be
frequency In order to model and analyze such defined as:
systems, traditionally equivalent circuits are used.

Excerpt from the Proceedings of the 2016 COMSOL Conference in Munich

 By using Ampere’s Law, the magnetic vector
(3) potential A is calculated, and consequently the
electrical potential V can be derived.
The reactance of the circuit becomes zero at
1
resonance, and (3) yields a fourth degree
μ
polynomial with the following solutions (with (6)
0):

, , , Spatial discretization of (6) via the finite element

method (FEM) resulted in a model with 130.000
4 1 (4) degrees of freedom (DOFs), i. e second order
2 1 ordinary differential equations of the form:

(7)
If the circuit is designed such that primary and
secondary side are matched to ,
only two solutions exist for (4):
where , , ∈ , ∈ , , ∈
, (5) , ∈ , ∈ is the input vector, ∈
√1
is the state vector, and ∈ is the output
The power transfer efficiency yields a maximum vector.
between and and since the coupling
coefficient ≫ 0, and are shifted away The system matrices in (7) can be extracted by
from . This phenomena is well known for using Livelink™ for Matlab.
coupled oscillators and is known as frequency
splitting in the field of WPT. 4. Model Order Reduction

In the following we will explain how this The enormous computational effort for the
phenomena can be analyzed based on the more solution of a full-scale model (7), prevents its co-
accurate numerical model of the coils. simulation with an electrical circuitry and thus the
efficient optimization of power transfer
3. Numerical Modelling efficiency. Luckily, it turns out that model order
reduction can yield a system of the same form as
A 3D geometry (Figure 7) was built in COMSOL (7) but with only 20 DOFs, which is still highly
Multiphysics ® following the specifications accurate. It focuses on reducing the number of
given in [9]. The design goal was to maximize the equations, while preserving the input-output
mutual inductance, while avoiding its sensitivity behavior of the dynamical system. In the
to winding misalignment. following, we explain how MOR works.

The model (7) can be rewritten as:

0 0 0
0
(8)

where the new matrices constitute a new

transformed first-order system of the form:

(9)
Figure 7. 3D model of a pair of disc coils, showing
ferrite and aluminum discs next to the windings.

Excerpt from the Proceedings of the 2016 COMSOL Conference in Munich

By means of Laplace transform, the transfer On the other hand side, one can use only one basis
function of (9) reads: (of or of ) and set for the
projection. Then the method is known as the
(10) input-one-sided or the output-one-sided Krylov
method.
can be developed into Taylor series and its
coefficients are known as moments of the 4.2 Arnoldi Process
transfer function and are defined as:
The Arnoldi process is one of the most popular
, 0,1 … (11) algorithms to find an orthonormal basis for the
input or output Krylov subspace and compute the
These moments are vectors or matrices, which matrices or respectively.
belong to the Krylov subspaces, defined as:
It is based on modified Gram-Schmidt
,.., orthogonalization as follows. Consider for
,.., (12) example the Krylov subspace of order . The
Arnoldi algorithm finds a set of vectors
is called the right/input-sided Krylov subspace , ,…, with length one, which are
and the left/output-sided Krylov subspace of orthogonal to each other and form the basis for the
order n. given Krylov subspace:

MOR aims on finding a reduced system, whose

(16)
first moments will match the first moments of , ,…,
the original system. This can be achieved by
forming one or both Krylov subspaces of order . It is an iterative procedure, as in each step one
more vector, orthogonal to all other previous
4.1. Reduction by Projection vectors, is constructed and normalized.

By using matrices , ∈ , a reduced model The algorithm further generates an upper

of order ≪ can be obtained for the original Hessenberg matrix ∈ , which is related
dynamical system of order , as schematically to the original system matrix as:
represented in Figure 8.
∙ ∙ (17)

and can be interpreted as the matrix of the reduced

system (e. g. the in Figure 8). It can be easily
proven [10] that such reduced system will match
the first moments of the original system (9).

4.3 Two-sided Arnoldi Algorithm

If the matrices and are computed as in (16)

to be the orthonormal bases for the input,
Figure 8. Order reduction of the first-order ordinary
differential equation system (9).
respectively output Krylov subspaces and are both
used for the projection of the original system (as
schematically represented in Figure 8), the
If the columns of matrix form the basis of the
method is called two-sided Arnoldi [11].
input-sided Krylov subspace and the columns of
matrix the basis of the output-sided Krylov
It yields better accuracy than the one-sided
subspace, the method is known as two-sided
Arnoldi process, due to the fact that it matches 2
Krylov method.
moments. Simultaneously the algorithm is

Excerpt from the Proceedings of the 2016 COMSOL Conference in Munich

simpler and more stable than the alternative two- The following results presented in section 5 are
sided Krylov methods [12]. gained by using the one-sided and two-sided
block SOAR methods respectively.
4.4 Block Arnoldi Algorithm
5. Simulation Results
Block Arnoldi method is the extension of Arnoldi
process for Multiple-Input-Multiple-Output Simulation experiments were conducted to
(MIMO) systems. The algorithm implements the evaluate the reference model in [9]. The reduced
so called deflation, which is removing of linearly models stemming from the different SOAR
dependent vectors from . methods (one-sided and two-sided) were
compared to the full model. Finally a system-level
The deflation is controlled by checking whether simulation was performed with the purpose of
the norm of each newly constructed vector is characterizing a complete WPT system.
different from zero, after its orthogonalization.
5.1 Sensitivity Analysis: 3D and 2D Comsol
4.5 Block Second Order Arnoldi Algorithm Models versus Reference Model
(SOAR)
A model that was implemented in Ansys ®
The block SOAR method is basically an extension Maxwell [10], was selected as a reference. It was
of the block Arnoldi method to second order used to validate the results of the 3D model (from
systems. It employs the second order Krylov Figure 7), as well as two additional 2D axis-
subspaces (corresponding to system (7)), defined symmetric COMSOL models.
in [12]:
The first 2D model implements disc coils,
, , considered as a bundle of stranded wires and
(18) modelled as multi-turn coils (see Figure 9).
, ,

where is the right/input-sided and is the

left/output-sided second order Krylov subspace of
order .

If the columns of form a basis for the input

second order Krylov subspace, then the first
moments of the original and reduced models Figure 9. 2D axis-symmetric pair of disc coils. Left:
spatial discretization, right: geometry
match [12], provided the reduced model is gained
from (7) by projection analog to those
The second 2D model implements windings turns
schematically represented in Figure 8 and by
represented as concentric rings (see Figure 10).
setting e. g. . The method is called one-
sided block SOAR method.

If the columns of V and W form bases for the

second order input and output Krylov subspaces,
respectively, 2 moments of the original and
reduced order systems match [12], provided the Figure 10. 2D axis-symmetric concentric windings.
Left: spatial discretization, right: geometry
reduced model is gained from (7) by projection
analog to those schematically represented in
In each model, the vertical distance between coils
Figure 8. The reduction procedure is called two-
was varied between 25 mm and 250 mm. This
sided block SOAR method.
analysis was automated from Matlab™ by using
a script. The operational frequency was set to 150
kHz.

Excerpt from the Proceedings of the 2016 COMSOL Conference in Munich

As observed in Figure 11, the relative error for 5.2.1. Input-sided Block SOAR Reduced
self-inductance does not exceed the threshold of Models
1.25 % for all models with regard to the reference
model. From the analysis of the coil’s resistance in Figure
13, it can be observed that all reduced models with
more than 4 moments are accurate. The maximal
relative error was lower than 2.5 % for the
reduced model with 7 moments.
Note, that the relative error for each reduced
model is minimal in the vicinity of 1kHz, which
was chosen the expansion point for the Taylor
series of transfer functions.

Figure 11. Self-inductance and relative error for

different models, as vertical distance between coils was
changed in the range of 25-250 mm.

Figure 13. Sensitivity analysis for input-sided block

SOAR reduced models with different number of
moments; coil resistance (above) and relative error
(bellow).

5.2.2. Output-sided Block SOAR Reduced

Models
Figure 12. Mutual-inductance and relative error as
vertical distance between coils was changed in the Models that are created by using output-sided
range of 25-250 mm. SOAR show better accuracy than those
constituted by input-sided SOAR algorithm, as
The impact of misalignment on the mutual shown in Figure 14. The model with 4 moments
inductance was also evaluated. The error between has the maximum relative error of 25 %. The
models was lower than 4 %, as shown in model using 7 moments exhibits negligible error
figure 12. over the whole frequency range.

5.2 Sensitivity Analysis of Reduced Models

with Different Orders

The analysis was focused on comparing the

accuracy of the implemented block SOAR
algorithms versus the original full model.
Reliability and accuracy of the algorithms were of
outmost importance, as the reduced models were
created with only a small number of moments
(DOFs).

Excerpt from the Proceedings of the 2016 COMSOL Conference in Munich

 Figure 16. Reduced model connected as a 2-port
network to an electrical circuit.

Apart from the voltage source E, each coil is

Figure 14. Sensitivity analysis for output-sided block connected to a serial capacitance CS and CL, and
SOAR reduced models with different number of
RL is the load resistance.
moments; coil resistance (above) and relative error
(bellow).
The voltage over the load resistor is shown in
5.2.3. Two-Sided Block SOAR Reduced Figure 17, where frequency splitting around the
Models operating frequency, 150 kHz is observed. The
frequency splitting matches the analytically
As expected, models reduced by two-sided block calculated values listed in Table 1.
SOAR are more accurate than those reduced by
one-sided methods. The maximum relative error Figure 18 depicts the power efficiency with its
of reduction amounts to solely 5·10-3% (see maximum within the frequency splitting range.
Figure 15). As power efficiency is directly dependent on the
coupling coefficient k, the bandwidth is affected
when the vertical distance is increased.

Figure 15. Sensitivity analysis for two-sided block

SOAR reduced models with different number of
Figure 17. Load resistor voltage for several vertical
moments; coil resistance (above) and relative error
distances (left); magnified view in the vicinity of the
(bellow).
operating frequency of 150 kHz (right).
5.3 Power Efficiency Analysis
Table 1. Approximated calculated range of frequency
splitting for horizontal displacements.
As a final step, power efficiency analysis was
conducted on the reduced order models. For this
Horizontal ⁄√1 ⁄√1
displacement [mm] [kHz] [kHz]
purpose, the reduced models were exported to
VHDL format and imported into ANSYS 0 132,34 177,36
50 134,50 172,49
Simplorer® as 2-port networks (see Figure 16).
100 140,02 162,46
150 146,44 153,83

Excerpt from the Proceedings of the 2016 COMSOL Conference in Munich

 9. References

[1] Héctor Vázquez-Leal, Agustín Gallardo-Del-

Angel, Roberto Castañeda-Sheissa and
Francisco Javier González-Martínez (2012).
The Phenomenon of Wireless Energy
Transfer: Experiments and Philosophy,
Wireless Power Transfer - Principles and
Engineering Explorations, Dr. Ki Young Kim
(Ed.), ISBN: 978-953-307-874-8, InTech,
Available from:
Figure 18. Power efficiency for several vertical http://www.intechopen.com/books/wireless-
distances (left); magnified view of the area with power power-transfer-principles-and-engineering-
efficiency larger than 50%
explorations/the-phenomenon-of-wireless-
energy-transfer-experiments-and-philosophy.
6. Conclusions and Outlook [2] Morris Kesler. Highly Resonant Wireless
Power Transfer: Safe, Efficient, and over
Several models of the WPT coil setup were Distance. White Paper [online].WiTricity
created in COMSOL ® based on the geometry Corporation, 2013. URL:
suggested in [10]. The results for self- and http://www.witricity.com/assets/highly-
mutual-inductance matched the reference model resonant-power-transfer-kesler-witricity-
within 1.4 % and 4 %. 2013.pdf
[3] Benedikt Schmuelling, Sarp G Cimen,
For creating reduced order models, the following Thomas Vosshagen, and Faical Turki,
block SOAR reduction algorithms (see chapter "Layout and Operation of a Non-Contact
4.5) were implemented and tested: input-sided, Charging System for Electric Vehicles," in
output-sided and two-sided. As expected, the two- 15th International Power Electronics and
sided algorithm yielded more accurate reduced Motion Control Conference, EPE-PEMC
order models than the one-sided algorithms. 2012 ECCE Europe, Novi Sad, Serbia, 2012,
p. 7.
Furthermore, reduced order models of very small [4] Aviva Brecher and David Arthur, "Review
size (between 4 and 7 second order ordinary and Evaluation of Wireless Power Transfer
differential equations) showed the same response (WPT) for Electric Transit Applications,"
as a full 3D model with more than 130 000 DOFs. Federal Transit Administration (FTA)
Research, Washington DC, USA, FTA
The effect of frequency splitting was analyzed Report 60, 2014.
analytically by calculating approximated values [5] André Kurs et al., "Wireless Power Transfer
for the mutual inductance maxima. Those values via Strongly Coupled Magnetic Resonances,"
were similar to the results obtained in ANSYS Science - AAAS, vol. 317, pp. 83-86, July
Simplorer®. Moreover, a continuous maximum 2007.
in power transfer efficiency over the frequency [6] Xiaofang Yu et al., "Wireless power transfer
splitting range was observed. in the presence of metallic plates:
Experimental results," AIP Advances, 2013.
The next step would be to study the mathematical [7] Faical Turki and Abdelkader Guetif,
methods of parametric model order reduction [13] "Contactless Charging Electric Vehicles with
for their applicability to construct the reduced Renewable Energy," in The fifth
models in which the coil distance would be International Renewable Energy Congress
preserved as parameter [14], which would enable IREC 2014, Hammamet, Tunisia, 2014.
fast design optimization of WPT systems. [8] Alexey Bodrov and Seung-Ki Sul, "Analysis
of Wireless Power Transfer by Coupled
Mode Theory (CMT) and Practical
Considerations to Increase Power Transfer
Efficiency," in Wireless Power Transfer -

Excerpt from the Proceedings of the 2016 COMSOL Conference in Munich

 Principles and Engineering Explorations.
Rijeka, Croatia: InTech, 2012, ch. 2, p. 272.
[9] Z. Tang, M. Christini, T. Koga,“Wireless
Power Transfer using Maxwell and
Simplorer”, Automotive Simulation - World
Congress, 2012.
[10] Roland W Freund, "Model reduction
methods based on Krylov subspace," Acta
Numerica, vol. 12, pp. 267-319, May 2003.
[11] B. Salimbahrami, B. Lohmann, T.
Bechtold, J.G. Korvink, Two-sided
Arnoldi algorithm and its application in
order reduction of MEMS, in: I. Troch, F.
Breitenecker (Eds.), Proc. 4th MATHMOD,
Vienna, 2003, pp. 1021–1028.
[12] Behnam Salimbahrami, Rudy Eid, and
Boris Lohmann, "Model Reduction by
Second Order Krylov Subspaces. Extensions,
Stability and Proportional Damping," in
Conference on Computer Aided Control
Systems Design, Munich, Germany, 2006,
pp. 2997-3002.
[13] Panzer, Heiko; Mohring, Jan; Eid, Rudy;
Lohmann, Boris, “Parametric Model Order
Reduction by Matrix Interpolation”, at –
Automatisierungstechnik 58 (8), 2010.
[14] Klis, D.; Burgard, S.; Farle, O.; Dyczij-
Edlinger, R., “Fast Simulation of Wireless
Power Transfer Systems With Varying Coil
Alignment” IFAC – PapersOnLine 48 (1),
2015, pp. 248-253.

Excerpt from the Proceedings of the 2016 COMSOL Conference in Munich