IEEE TRANSACTIONS ON MAGNETICS, VOL. 51, NO.

3, MARCH 2015 6300304

Numerical Study of the Core Saturation Influence on the

Winding Losses of Traction Transformers
Jasmin Smajic1 , Giorgio Di Pino1 , Christoph Stemmler2 , Wolfgang Mönig3 , and Martin Carlen4
1 University of Applied Sciences of Eastern Switzerland, Rapperswil CH-8640, Switzerland
2 ABB Schweiz AG, Turgi CH-5300, Switzerland
3 ABB AG, Brilon D-59929, Germany
4 ABB Management Services Ltd., Zurich CH-8050, Switzerland

The winding system of traction transformers is subjected to non-sinusoidal voltages and currents of the adjacent traction converter.
Unlike its fundamental component, the higher harmonics of the load current have a completely non-uniform distribution over
the conductor’s cross section causing hazardous hot spots. Due to dynamic operating conditions in railway grids, the railway
voltage can temporarily increase as much as 20% saturating magnetic cores of traction transformers. This paper presents the
analysis of the core saturation influence on the distribution of the transformer’s winding losses along with suitable numerical
methodologies for their simulation. Thus, a fast 2-D and 3-D frequency-domain methodology for computing winding losses of traction
transformers by considering all the relevant higher current harmonics and core saturation effects caused by the fundamental
voltage component is described in detail. The obtained results are verified by comparison against the available corresponding
measurements and against the results of a rigorous and from the CPU time perspective much more demanding time-domain simulation
approach.
Index Terms— Eddy currents, traction converters, traction transformers, winding losses.

I. I NTRODUCTION

A MODERN traction converter for double-deck electrical

trains consists of a four-quadrant (4Q) line con-
verter, propulsion converter, dc-link circuit, and auxiliary
components [1], as shown in Fig. 1(a). The role of the single
phase traction transformer is either to step down the voltage
of the incoming power from the catenary (15 or 25 kV) in
case of the train’s acceleration or to step up the voltage of
the incoming power from the converter in case of the train’s
deceleration (recuperation).
Due to the switching of the 4Q line converter, the
rectangular voltage is applied to the transformer secondary
windings [1]. The transformer stray inductance is usually used
to transform this rectangular voltage into its harmonic coun-
terpart to avoid a filter between the converter and transformer.
Thus, the resulting secondary current is also a non-sinusoidal
function [Fig. 1(b)].
As already reported in [2], it is possible to accurately
simulate winding losses of converter transformers, including
higher current harmonics, by making a very detailed winding
model. The modeling and simulations presented in [2] did not
consider core saturation effects.
According to the IEC standard [3], the railway voltage can
temporarily increase as much as 20%. At this voltage level,
the transformer’s core goes into saturation causing a different
distribution of the stray magnetic flux between the transformer
coils, especially in the regions near the winding ends, i.e., near Fig. 1. (a) Main components of a modern traction converter attached to the
the saturated transformer’s yokes. The eddy currents induced considered transformer [1]. (b) Time dependence of the transformer voltage
(pulse width modulation) and current.
Manuscript received May 25, 2014; revised September 11, 2014; accepted
September 26, 2014. Date of current version April 22, 2015. Corresponding in the windings by the stray magnetic flux are consequently
author: J. Smajic (e-mail: [email protected]).
Color versions of one or more of the figures in this paper are available
affected by this saturation effect.
online at http://ieeexplore.ieee.org. To the best of our knowledge and according to the extensive
Digital Object Identifier 10.1109/TMAG.2014.2360918 literature review related to the analysis of winding losses
0018-9464 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 09,2021 at 16:45:53 UTC from IEEE Xplore. Restrictions apply.
 6300304 IEEE TRANSACTIONS ON MAGNETICS, VOL. 51, NO. 3, MARCH 2015

reported in [2], there are no scientific papers describing the

analysis of the correlation between the core saturation level
and winding losses caused by non-sinusoidal currents of the
power converter attached to the traction transformer. Thus, the
original scientific contribution of this paper is to present
the details and results of the said analysis and to demonstrate
the most promising numerical algorithms for simulating this
phenomenon in 2-D and 3-D.
The remainder of this paper is organized as follows.
Section II describes the field formulations and simulation Fig. 2. Non-sinusoidal current source used in the simulation models
methods used for obtaining the presented results. Section III (the red curve) compared with the ideal harmonic current (the blue curve) in
time-domain (left) is presented. The harmonic content of the source current
presents the obtained numerical result and their verification. (the blue bars) and its most dominant harmonics considered in the frequency-
Finally, the conclusion is drawn in Section IV. domain simulations (the red bars) (right) are also depicted. The relevance
threshold (0.02 · In , where In is the nominal transformer’s current) of the
harmonic components is also shown (the green horizontal dashed line).
II. F IELD F ORMULATION AND S IMULATION M ETHODS
For the axisymmetric 2-D and 3-D magneto-quasistatic
analysis (i.e., the eddy-current analysis) in time and frequency frequency-domain analysis is possible, according to [7], pro-
domain, a field solver with the following implemented time- vided that the effective magnetization curve from the original
domain field formulation is used [4]–[6]: measured B–H curve of the material is constructed. Several
different methods for constructing the effective magnetization
1 ∂ H curve were suggested and evaluated in [7]. The best perfor-
∇× ∇ × H + μr = 0 in C (1)
σ · μ0 ∂t mance, however, with the field formulation implemented in the
∇ × H = JS in C (2) used field solver shows the following rms method applicable
to current sources [7]:
∇ · μr H = 0 in 0 (3) 

H × n = 0 in ∂ D  (4) T
T
1
B 2 (t) · dt

μr H · n = 0 in ∂ N  (5) Beff 0
μeff = = √ (6)
where C is the conducting regions, 0 is the non-conducting
Heff Hamp / 2
region,  = C ∪ 0 is the computational domain, ∂ D  is where Hamp is the amplitude of the magnetic field, B(t) is the
the Dirichlet’s boundary of the computational domain, magnetic flux density obtained from the measured magnetizing
∂ N  is the Neumann’s boundary of the computational domain, curve, and T is the source period.
σ is the electric conductivity, μ0 is the magnetic permeability It is also worth mentioning that the magnetic field of
of vacuum, μr is the relative magnetic permeability of the the BVP (1)–(5) is in the magnet solver represented in the
material, n is the normal unit vector to the boundary, JS is the following way [5]:
known source current density, and H is the magnetic field. H = H S − ∇φ in 0 (7)
The boundary value problem (BVP) (1)–(5) is written
in time-domain because the source current is not harmonic H = H S + H  in C (8)
(the effect of the traction converter) and non-linear magnetic where φ is the scalar magnetic potential, H S is the source
materials (the transformer’s core) are involved. magnetic field produced by the known source current
The field formulation (1)–(5) implemented in the within density JS , and H  is the magnetic field that accounts for
this paper used field solver enables efficient eddy-current the difference between the total current density (including the
simulations of the current-driven problems, which is essential induced eddy currents) and the source current density.
for the analysis of winding losses [2], [5]. At the interface between conducting and non-conducting
Considering the non-sinusoidal time dependence of the material, the following condition is valid [5]:
current shown in Fig. 1(b) (the blue curve) and considering
n × ( H − H S ) = n × (−∇φ). (9)
the non-linear character of the magnetic core, only the time-
domain solution of the BVP (1)–(5) yields accurate results. As reported in [5], it is very simple to impose this condition
Due to its extremely long CPU time even in its axisym- over the conducting to non-conduction interface to eliminate
metric 2-D version, the time-domain method is not suitable the edge variables in favor of the nodal variables.
for daily design involving numerous simulations of slightly According to (7) and (8), it is visible that the magnet
different geometries. Therefore, it would be very beneficial solver computes the unknown scalar magnetic potential in
if the frequency-domain solution of the BVP (1)–(5) for the non-conductive regions using scalar tetrahedral elements.
dominant harmonic components could deliver the winding On the other hand, in conductive regions, the magnetic field
losses of the acceptable accuracy for daily design. is computed and vector tetrahedral elements are used [5], [6].
Considering the theory of the frequency-domain solution
of the BVP (1)–(5), its fundamental assumption that all III. N UMERICAL R ESULTS AND T HEIR V ERIFICATION
the involved vector and scalar fields are harmonic time- All the presented winding loss simulations are current-
dependent functions is violated, if non-linear magnetic mate- driven eddy-current problems described by (1)–(9). The cur-
rials are involved. However, even in this case, an approximate rent source used in the simulation models is shown in Fig. 2.

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 09,2021 at 16:45:53 UTC from IEEE Xplore. Restrictions apply.
 SMAJIC et al.: NUMERICAL STUDY OF THE CORE SATURATION INFLUENCE 6300304

TABLE I TABLE II
R ESULTS OF THE 2-D F REQUENCY-D OMAIN A NALYSIS R ESULTS OF THE 2-D T IME -D OMAIN A NALYSIS W ITH
W ITH THE I DEAL C URRENT S OURCE F ROM F IG . 2(L EFT ) THE N ON -S INUSOIDAL C URRENT S OURCE G IVEN
(B LUE L INE ) V ERSUS M EASUREMENTS IN F IG . 2(L EFT ) (R ED S OLID L INE )

TABLE III
R ESULTS OF THE 2-D F REQUENCY-D OMAIN A NALYSIS
W ITH THE 10 D OMINANT H ARMONIC C URRENT
C OMPONENTS [T HE R ED BARS IN F IG . 2(R IGHT )]

Fig. 3. The structure of a detailed axisymmetric transformer model (left)

and the current density distribution for the ideal harmonic nominal current simulations were performed with the original B–H curve, as
shown in Fig. 2(left) (the blue line) is presented.
the construction of the effective magnetizing curve in this case
is not needed.
To test the predictive power of the axisymmetric 2-D trans- Table II shows that the axisymmetric 2-D time-domain
former model, a frequency-domain simulation according to the windings model of the highest modeling resolution (each turn
method presented in [2] with the ideal harmonic source [blue modeled as a separate entity) did not reveal any influence of
line in Fig. 2(a)] at the fundamental frequency (16.7 Hz in the the core saturation level on the total winding losses with the
Swiss traction grid) was perform and its results were compared non-sinusoidal current source shown in the left-hand side of
against measurements. The obtained results are presented in Fig. 2 (red line). This, to some extent, unexpected result is a
Table I. The measured losses corresponding to the 100% value very important information for the subsequent thermal design
are 46.15 (±1%) kW. of the transformer. Such 2-D time-domain simulation models
Evidently, the comparison of the obtained simulation results with detailed winding modeling require a very long CPU
with the ideal harmonic current source with the corresponding time, namely four days on the modern computer (Intel Xeon
measurements has revealed a very high level of accuracy. The CPU E5-2670 0 2.6 GHz, 128 GB RAM). This CPU time is
main reason for such a high accuracy is a fine mesh resolu- evidently not acceptable for the daily design point of view.
tion and detailed modeling of the low voltage (LV) winding As a comparison, the previously mentioned non-linear
(aluminum foils) and high voltage (HV) winding (aluminum frequency-domain analysis with the fundamental current har-
wires) shown in Fig. 3. Fig. 3 shows the eddy currents induced monic component has required the CPU time of only 1 h,
by the stray magnetic flux in the LV-foils at the top of the 28 min, and 5 s. Thus, it would be very beneficial, if a
winding. series of non-linear frequency-domain simulations performed
The simulation result shown in Fig. 3 was obtained by for each relevant harmonic component of the transformer’s
performing a non-linear eddy-current analysis in frequency current [right-hand side of Fig. 2 (red bars)] could be used
domain with the effective B–H curve constructed according to obtain the total distribution of the winding losses as a
to (6). The transformer core was magnetized by an auxiliary sum of the results at each frequency. If possible, this would
magnetizing coil to easily control the core saturation level. lead to a radical reduction of the required CPU time and
In this simulation, the core was magnetized in accordance with to possibility to use those simulations in daily design. From
the nominal transformer’s voltage level (Un ). the theoretical point of view, it is not clear to what extent
After this encouraging result, the same model was is this possible to do due to the non-linearity of the core.
used for the time-domain eddy-currents analysis with the To clarify this question, a series of 2-D non-linear frequency-
non-sinusoidal current source shown in the left-hand side domain simulations was performed and the corresponding
of Fig. 2 (red curve). The transient analysis was performed results are shown in Table III. The results in Table III show
for three different core saturation levels corresponding to the main problems of the frequency domain approach in case
60%, 100%, and 120% of the nominal transformer’s voltage. of the non-sinusoidal current source and non-linear materials
The obtained results are given in Table II. The time-domain involved. If the time dependence of the non-sinusoidal current

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 09,2021 at 16:45:53 UTC from IEEE Xplore. Restrictions apply.
 6300304 IEEE TRANSACTIONS ON MAGNETICS, VOL. 51, NO. 3, MARCH 2015

Fig. 4. Average loss density distributions for the chosen current harmonics by
the saturated (left) and non-saturated core (right) are presented. The induced
eddy currents are at these frequencies visible not only in the LV-foils, but also
in the LV-wires.

TABLE IV
Fig. 5. 3-D transformer model (top-left), the numerical results of the
R ESULTS OF THE 3-D T IME -D OMAIN A NALYSIS W ITH frequency-domain analysis in the form of current density (bottom) for the
THE N ON -S INUSOIDAL C URRENT S OURCE G IVEN three chosen current harmonics, and the results of the time-domain analysis
IN F IG . 2(L EFT ) (R ED S OLID L INE ) in the form of the time average loss density evaluated along the top edge of
every LV-foil (top-right) (the angle equal to zero represents the position under
the yoke, and the numbers 1, 2,. . .,28 name the LV-foils) are presented.

IV. C ONCLUSION
The details and results of the winding loss analysis of trac-
tion transformers with the non-sinusoidal current of the adja-
is of such a shape that its fast Fourier tranform consists of cent 4Q converter by considering the level of magnetic core
numerous harmonic components not distinctively separated in saturation are presented. The obtained 2-D and 3-D results
the frequency domain a high number of frequency domain using the presented fast frequency-domain methodology are
simulations are needed to accurately obtain the total winding verified by comparison against the available measurements and
losses. This effect is visible by comparing Tables II and III against the results of the accurate but, in terms of the CPU
(102% of the losses by considering the 10 dominant current time, very expensive time-domain approach.
harmonics versus 110% of the losses obtained by the brute The obtained 2-D and 3-D results did not show any mea-
force time-domain approach). surable influence of the core saturation level on the total
In addition to the previous statement, Table III shows us winding losses but have revealed its influence on the local
a measurable influence of the core saturation on the winding loss distribution of the higher current harmonics. In converter
losses not visible in the time-domain results (Table II) that regimes with very pronounced higher current harmonics for
we consider more accurate. This effect is attributed to the longer time, this effect could potentially lead to hazardous hot
construction of the effective magnetizing curve within the spots and lifetime reduction of the transformer.
frequency-domain simulations.
Table III consists of the total loss data. However, the R EFERENCES
influence of the core saturation on the local winding loss [1] Propulsion and Auxiliary Converter for Double-Deck EMU,
distribution is visible in Fig. 4, especially in case of higher BORDERLINE CC1500 AC_15–25 kV, ABB Switzerland Ltd.,
Turgi, Switzerland, 2013.
current harmonics. This could lead in some cases to local hot [2] J. Smajic, J. Hughes, T. Steinmetz, D. Pusch, W. Mönig, and
spots if the higher harmonics are more pronounced. M. Carlen, “Numerical computation of ohmic and eddy-current winding
In the axisymmetric 2-D model, the yoke is modeled as a losses of converter transformers including higher harmonics of
load current,” IEEE Trans. Magn., vol. 48, no. 2, pp. 827–830,
ferromagnetic disk above and below the windings. In reality, Feb. 2012.
only the windings and to some extent the core limb could [3] Railway Application—Electric Equipment for Rolling Stock,
be considered axisymmetric. The yokes possess no axial International Electrotechnical Comission, Geneva, Switzerland,
IEC Standard 60077-1, 1999.
symmetry and therefore could be accurately modeled only in [4] Infolytica Corporation. Infolytica MagNet: Design and Analysis Software
a 3-D model, such as the model shown in Fig. 5. for Electromagnetics. [Online]. Available: http://www.infolytica.com,
Due to the memory constraints, the HV-coil of the accessed 2014.
3-D model was modeled as a stranded winding block without [5] J. P. Webb and B. Forghani, “A scalar-vector method for 3D eddy current
problems using edge elements,” IEEE Trans. Magn., vol. 26, no. 5,
induced eddy currents and only the LV-coil was modeled pp. 2367–2369, Sep. 1990.
in its full complexity. The obtained results are presented in [6] J. P. Webb and B. Forghani, “T- method using hierarchal edge
Table IV. Similar to the 2-D model, also the 3-D model did not elements,” IEE Proc.-Sci., Meas. Technol., vol. 142, no. 2, pp. 133–141,
Mar. 1995.
reveal any measurable change in winding losses. However, the [7] G. Paoli, O. Biro, and G. Buchgraber, “Complex representation in
influence of the yoke is visible in the loss density distribution nonlinear time harmonic eddy current problems,” IEEE Trans. Magn.,
shown in Fig. 5 (top-right). vol. 34, no. 5, pp. 2625–2628, Sep. 1998.

Authorized licensed use limited to: Universiti Teknikal Malaysia Melaka-UTEM. Downloaded on September 09,2021 at 16:45:53 UTC from IEEE Xplore. Restrictions apply.