Applied Numerical Mathematics 79 (2014) 3–17

Contents lists available at SciVerse ScienceDirect

Applied Numerical Mathematics

www.elsevier.com/locate/apnum

Finite element solution of nonlinear eddy current problems

with periodic excitation and its industrial applications
Oszkár Bíró ∗ , Gergely Koczka, Kurt Preis
Institute for Fundamentals and Theory in Electrical Engineering (IGTE), Graz University of Technology, Inffeldgasse 18, 8010 Graz, Austria

a r t i c l e i n f o a b s t r a c t

Article history: An efficient finite element method to take account of the nonlinearity of the magnetic
Available online 6 June 2013 materials when analyzing three-dimensional eddy current problems is presented in this
paper. The problem is formulated in terms of vector and scalar potentials approximated by
Keywords:
edge and node based finite element basis functions. The application of Galerkin techniques
Finite element method
Fixed-point technique
leads to a large, nonlinear system of ordinary differential equations in the time domain.
Harmonic balance method The excitations are assumed to be time-periodic and the steady-state periodic solution is of
Discrete Fourier transform interest only. This is represented either in the frequency domain as a finite Fourier series
Nonlinearity or in the time domain as a set of discrete time values within one period for each finite
Parallel computation element degree of freedom. The former approach is the (continuous) harmonic balance
method and, in the latter one, discrete Fourier transformation will be shown to lead to a
discrete harmonic balance method. Due to the nonlinearity, all harmonics, both continuous
and discrete, are coupled to each other.
The harmonics would be decoupled if the problem were linear, therefore, a special
nonlinear iteration technique, the fixed-point method is used to linearize the equations
by selecting a time-independent permeability distribution, the so-called fixed-point
permeability in each nonlinear iteration step. This leads to uncoupled harmonics within
these steps.
As industrial applications, analyses of large power transformers are presented. The first
example is the computation of the electromagnetic field of a single-phase transformer
in the time domain with the results compared to those obtained by traditional time-
stepping techniques. In the second application, an advanced model of the same transformer
is analyzed in the frequency domain by the harmonic balance method with the effect of
the presence of higher harmonics on the losses investigated. Finally a third example tackles
the case of direct current (DC) bias in the coils of a single-phase transformer.
© 2013 The Authors. Published by Elsevier B.V. on behalf of IMACS.
Open access under CC BY-NC-ND license.

1. Introduction

The most straightforward method of solving nonlinear electromagnetic field problems in the time domain by the method
of finite elements (FEM) is using time-stepping techniques. This requires the solution of a large nonlinear equation system
at each time step and is, therefore, very time consuming, especially if a three-dimensional problem is being treated. If the
excitations are non-periodic or if, in case of periodic excitations, the transient solution is required, one cannot avoid time
stepping. In many cases however, the excitations of the problem are periodic, and it is only the steady-state periodic solution

* Corresponding author.
E-mail address: biro@tugraz.at (O. Bíró).

0168-9274 © 2013 The Authors. Published by Elsevier B.V. on behalf of IMACS. Open access under CC BY-NC-ND license.
http://dx.doi.org/10.1016/j.apnum.2013.04.007
4 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17

which is needed. Then, it is wasteful to step through several periods to achieve this by the “brute force” method [1] of time
stepping.
A successful method to avoid stepping through several periods in such a case is the time-periodic finite element method
introduced in [12]. To accelerate the originally slow convergence of the method a singular-decomposition technique has
been introduced in [18] and it has even been parallelized in [19].
A new time domain technique using the fixed-point method to decouple the time steps has been introduced in [9] and
applied to two-dimensional eddy current problems described by a single component vector potential. The optimal choice
of the fixed-point permeability for such problems has been presented in [13] both in the time domain and using harmonic
balance principles. The method has been applied to three-dimensional problems in terms of a magnetic vector potential and
an electric scalar potential ( A , V – A formulation) in [14] and, employing a current vector potential and a magnetic scalar
potential (T , Φ –Φ formulation), in [15] and [8]. In contrast to the time-periodic finite element method, the periodicity
condition is directly present in the formulation instead of being satisfied iteratively.
The aim of this work is to present a detailed review of the fixed-point based method and to show its application to
industrial problems arising in the design of large power transformers.
The paper is structured as follows: In the following two sub-sections of the Introduction, two FEM potential formulations
of eddy current problems are briefly reviewed and the continuous and discrete harmonic balance methods to obtain their
steady-state periodic solution are introduced. In Section 2, a method is developed to decouple the harmonics from each
other and hence to solve for each harmonic separately. This is trivial for linear problems, but a special fixed-point iteration
technique is introduced to treat nonlinearity with the harmonics decoupled. Section 3 is devoted to numerical examples
involving large power transformers. The results of the paper are concluded in Section 4.

1.1. Finite element potential formulations

The geometry of an eddy current problem can be naturally split in two: an eddy current domain with unknown current
density distribution and an eddy current free region in which the current density is given [3].
The electromagnetic field problem to be solved in the eddy current domain Ωc consisting of conducting media is de-
scribed by Maxwell’s equations in the quasi-static limit:

curl H = J + curl T 0 , (1)

∂B
curl E = − , (2)
∂t
div B = 0, (3)
div J = 0 (4)
where H is the magnetic field intensity, J is the eddy current density, T 0 is an impressed current vector potential whose
curl is the given current density in coils external to Ωc , E is the electric field intensity, B is the flux density and t is time.
In the eddy current free region Ωn (such as domains containing non-conducting media as well as coils with known current
density) it is sufficient to solve (1) with J = 0 in addition to (3) for the magnetic field quantities. The material relationships
are
   
B = μ |H | H or H = ν |B| B, (5)
J =σE or E =ρ J (6)
where μ is the permeability, ν is its reciprocal, the reluctivity and σ is the conductivity with ρ denoting its reciprocal, the
resistivity. In magnetic materials (steel), the relationships (5) are nonlinear, i.e. the permeability and the reluctivity depend
on the magnetic field intensity or the magnetic flux density as indicated.
The numerical solution of the problem is carried out by the method of finite elements. The application of FEM is straight-
forward if potential functions are introduced. Basically, two options are open: the field quantities can either be represented
by a magnetic vector potential A and an electric scalar potential V ( A , V – A formulation) as
∂
B = curl A in Ωc ∪ Ωn , E =− ( A + grad V ) in Ωc , (7)
∂t
or by a current vector potential T and a magnetic scalar potential Φ (T , Φ –Φ formulation) as
H = T 0 + T − grad Φ in Ωc ∪ Ωn , J = curl T in Ωc (8)
with T = 0 in Ωn . The definitions (7) satisfy (2) and (3), whereas those in (8) ensure that (1) and (4) hold. Therefore, the
differential equations (1) and (4) are to be solved in the A , V – A formulation:
∂ 
curl(ν curl A ) +σ ( A + gradV ) = curl T 0 , (9)
∂t
 
∂
− div σ ( A + grad V ) = 0, (10)
∂t
 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17 5

and the Maxwell equations (2) and (3)

∂  ∂
curl(ρ curl T ) + μ( T − grad Φ) = − (μ T 0 ), (11)
∂t ∂t
 
div μ( T − grad Φ) = − div(μ T 0 ) (12)

remain to be solved in the T , Φ –Φ formulation.

Introducing the edge based vector basis functions N i (r ) (i = 1, 2, . . . , ne ) and the node based scalar basis functions N i (r )
(i = 1, 2, . . . , nn ) in the finite elements (ne is the number of edges and nn the number of nodes in the finite element mesh,
r denotes the space coordinates), the potentials are approximated as


ne 
ne
A (r , t ) ≈ A h (r , t ) = ak (t ) N k (r ), T (r , t ) ≈ T h (r , t ) = tk (t ) N k (r ), (13)
k =1 k =1


nn 
nn
V (r , t ) ≈ V h (r , t ) = v k (t ) N k (r ), Φ(r , t ) ≈ Φh (r , t ) = φk (t ) N k (r ). (14)
k =1 k =1

The vector T 0 is represented by edge basis functions similarly to T in (13). The coefficients for T 0 are easily computed as
its line integrals along the edges of the finite element mesh.
Applying Galerkin techniques to (9) and (10) leads to the following ordinary differential equations for the A , V – A for-
mulation:

d
curl N i · ν curl A h dΩ + σ N i · ( A h + grad V h ) dΩ = curl N i · T 0 dΩ, i = 1, 2, . . . , n e , (15)
dt
Ωn ∪Ωc Ωc Ωn ∪Ωc

d
σ grad N i · ( A h + grad V h ) dΩ = 0, i = 1, 2, . . . , nn . (16)
dt
Ωc

In some applications, the voltage U (t ) of the coils is given rather than their current. Assuming there is one single coil
with unknown current I (t ) present, the impressed current vector potential can be written as T 0 = It 0 with t 0 corresponding
to a unit current. Neglecting the resistance of the coil, the given voltage can be written as the integral of −t 0 · ∂ B /∂ t over
the problem domain (see [10]), therefore the Galerkin equations have the form

d
curl N i · ν curl A h dΩ + σ N i · ( A h + grad V h ) dΩ − I curl N i · t 0 dΩ = 0, i = 1, 2, . . . , ne , (15a)
dt
Ωn ∪Ωc Ωc Ωn ∪Ωc

− t 0 · curl A h dΩ = U dt , (15b)
Ωn ∪Ωc

d
σ grad N i · ( A h + grad V h ) dΩ = 0, i = 1, 2, . . . , nn . (16a)
dt
Ωc

Note the symmetry of system with respect to the unknown I (t ) as well.

Gathering the unknown time functions ak (t ) (k = 1, 2, . . . , ne ) and v k (t ) (k = 1, 2, . . . , nn ) in (13) and (14) and, possibly,
the unknown current I (t ) in a vector x(t ), the matrix form of (15), (16) is the system of ordinary differential equations

   dx(t )
S ν x(t ) x(t ) + M(σ ) = f(t ) (17)
dt
where the dependence of the stiffness matrix S on ν and of the mass matrix M on σ is explicitly shown. Since the reluctivity
depends on the field, ν depends on x and hence on t as indicated. The right hand side vector is denoted by f.
In a similar manner, Galerkin’s method applied to (11) and to the time derivative of (12) results in the ordinary differ-
ential equations

d d
curl N i · ρ curl T h dΩ + μ N i · ( T h − grad Φh ) dΩ = − μ N i · T 0 dΩ, i = 1, 2, . . . , ne , (18)
dt dt
Ωc Ωc Ωc

d d
− μ grad N i · ( T h − grad Φh ) dΩ = μ grad N i · T 0 dΩ, i = 1, 2, . . . , nn (19)
dt dt
Ωn +Ωc Ωn +Ωc
6 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17

for the T , Φ –Φ formulation. If a coil with voltage excitation is present, these equations have the symmetric form

d dI
curl N i · ρ curl T h dΩ + μ N i · ( T h − grad Φh ) dΩ + μ N i · t 0 d Ω = 0, i = 1 , 2 , . . . , n e , (18a)
dt dt
Ωc Ωc Ωc

d dI
− μ grad N i · ( T h − grad Φh ) dΩ = μ grad N i · t 0 dΩ, i = 1, 2, . . . , nn , (19a)
dt dt
Ωn +Ωc Ωn +Ωc

d dI
t 0 · μ( T h − grad Φ)h dΩ + t 0 · μt 0 dΩ = −U . (19b)
dt dt
Ωn ∪Ωc Ωn ∪Ωc

The vector x(t ) now consists of the unknown time-dependent coefficients tk (t ) (k = 1, 2, . . . , ne ) and φk (t ) (k = 1, 2, . . . , nn )
in (13) and (14) and, if a voltage fed coil is present, of the unknown current I (t ). The matrix form of the Galerkin equations
is the system of ordinary differential equations

d     d    
S(ρ )x(t ) + M μ x(t ) x(t ) = g μ x(t ) , t (20)
dt dt
where the stiffness matrix S is now independent of x and hence of time, but the mass matrix M depends on the perme-
ability which is itself field- and time-dependent. The product of the mass matrix and the unknown vector is differentiated
with respect to time. The excitation vector g depends on x and t, and its time derivative appears on the right hand side.
In the following it is assumed that the excitation of the problem, i.e. the current vector potential T 0 , or the voltage
U (t ) is time-periodic with a frequency f and that we are only looking for the steady-state, time-periodic solution of the
problem. This means that the right hand side vectors of the systems of ordinary differential equations (17) and (20) are
time-periodic, i.e. f(t ) = f(t + T ) and g(μ, t ) = g(μ, t + T ) where T = 1/ f is the period determined by the frequency f .

1.2. Harmonic balance method

Since we are only interested in the steady-state periodic solution satisfying the periodicity condition x(t ) = x(t + T ),
under the assumption that the time average over a period is zero, the solution is approximated by a complex Fourier series
with N harmonics as



N
jkωt
x(t ) ≈ x N (t ) = Re Xk e (21)
k =1

where j is the imaginary unit, ω = 2π f is the angular frequency of the excitation and Xk is the complex Fourier coefficient
of the k-th harmonic at the angular frequency kω . It can be computed as

T
1
Xk = Fk (x) = x(t )e − jkωt dt . (22)
T
0

Setting the approximation (21) into (17) and (20), respectively, and computing the N Fourier coefficients of both sides,
a system of equations with N times as many unknowns is obtained as there are unknown time functions, i.e. degrees of
freedom, in x(t ):

  
Fm S ν (x N ) x N + jmωM(σ )Xm = Fm (f), m = 1, 2, . . . , N , (23)
   
d    d  
S(ρ )Xm + Fm M μ(x N ) x N = Fm g μ(x N ), t , m = 1, 2, . . . , N . (24)
dt dt

In the linear terms in (23) and (24), the Fourier coefficients of the m-th harmonic appear only. The time derivative in (17)
corresponds to a multiplication by jmω in (23). The right hand side of (23) can be computed directly from f as shown in
(22). On the other hand, the nonlinear terms containing the permeability μ(x N ) or the reluctivity ν (x N ) depending on the
unknown solution (21) couple all Fourier coefficients to each other. Therefore, due to the nonlinearity, one cannot solve for
each harmonic alone, a fact which significantly increases the complexity of the problem.
 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17 7

1.3. Discrete harmonic balance method

As an alternative, the periodic time function x(t ) can be represented by a sequence of N equidistant time values within
a period as xk = x(kt ), k = 1, 2, . . . , N with t denoting the time step t = T / N. This sequence is cyclic, since due to the
periodicity of x, we have x0 = x N . Discretizing (17) and (20) by a backward Euler difference scheme, one obtains
  xm − xm−1
S ν (xm ) xm + M(σ ) = fm , m = 1, 2, . . . , N , (25)
t
1       1     
S(ρ )xm + M μ(xm ) xm − M μ(xm−1 ) xm−1 = gm μ(xm ) − gm−1 μ(xm−1 ) , m = 1, 2, . . . , N (26)
t t
where the subscripts indicate time values.
Let us introduce the notations
 T  T
x[1] = x1 x2 . . . xN . . . x N −1
, x[0] = x0 x1 ,
 T
g[1](μ) = g1 (μ(x1 )) g2 (μ(x2 )) . . . g N (μ(x N )) ,
 T
g[0](μ) = g0 (μ(x1 )) g1 (μ(x2 )) . . . g N −1 (μ(x N )) (27)

for the hyper-vectors formed by the cyclic sequences ( T denotes transpose) as well as
⎡ ⎤ ⎡ ⎤
S(ν (x1 )) 0 ... 0 M(σ )
... 0 0
  ⎢ 0 S(ν (x2 )) . . . ⎥0   ⎢ 0
M(σ ) . . . 0 ⎥
S(ν ) = ⎣ ⎢ ⎥, M(σ ) = ⎣ ⎢ ⎥,
... ... ... ... ⎦ ... ... ... ... ⎦
0 0 . . . S(ν (x N )) 0 0 . . . M(σ )
⎡ ⎤ ⎡ ⎤
S(ρ ) 0 ... 0 M(μ(x1 )) 0 ... 0
  ⎢ 0 S (ρ ) ... 0 ⎥   ⎢ 0 M(μ(x2 )) . . . 0 ⎥
S(ρ ) = ⎢
⎣ ...
⎥,
⎦ M(μ) = ⎢ ⎣
⎥
⎦ (28)
... ... ... ... ... ... ...
0 0 ... S(ρ ) 0 0 ... M(μ(x N ))
for the block-diagonal matrices. Hence (25) and (26) can be written as
 
1 1  
S[1](ν ) + M(σ ) x[1] − M(σ ) x[0] = f[1], (29)
t t
 
1 1   1  
S(ρ ) + M[1](μ) x[1] − M[0](μ) x[0] = g[1](μ) − g[0](μ) . (30)
t t t
Note that the matrices depending on the permeability μ or reluctivity ν vary in time. This is reflected in (29) and (30) by
the symbols [1] and [0] following these matrices indicating the sampling operations defined in (27).
The discrete Fourier transform of the sequence x[1] is defined as [11]:

   T   
N
k
x̂ = D x[1] = x̂1 x̂2 . . . x̂ N , x̂m = Dm x[1] = xk e − j2π ·m N , m = 1, 2, . . . , N . (31)
k =1

This has the advantage that, according to the shift theorem [11], the discrete Fourier transform of x[0] can simply be
obtained as
⎡ 1
⎤
Ie − j2π N 0 ... 0
  ⎢ 2 ⎥
⎢ 0 Ie − j2π N ... 0 ⎥
D x[0] = Px̂, P=⎢ ⎥ (32)
⎣ ... ... ... .... ⎦
N
0 0 . . . Ie− j2π N
where I is the unit matrix.
Applying the discrete Fourier transformation to (29) and (30), a system of equations with N times as many unknowns is
obtained as there are elements, i.e. degrees of freedom, in xk :
  
1  1   
D S[1](ν ) + M(σ ) x[1] −
M(σ ) Px̂ = D f[1] , (33)
t t
  1     1 
S(ρ ) x̂ + D M[1](μ) x[1] − M[0](μ) x[0] = D g[1](μ) − g[0](μ) . (34)
t t
8 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17

In the linear terms in (33) and (34), the elements of x̂, i.e. the discrete harmonics, are decoupled. The shift in time cor-
responds to a multiplication by the block-diagonal matrix P in (33). The right hand side of (33) can be computed directly
from f[1] as shown in (31). On the other hand, the nonlinear terms containing the permeability μ or the reluctivity ν de-
pending on the unknown solution couple all elements of the discrete Fourier transform, i.e. the discrete harmonics to each
other. Therefore, due to the nonlinearity, one cannot solve for each discrete harmonic alone, a fact which, again, significantly
increases the complexity of the problem.

2. Decoupling of harmonics

It is highly desirable that the harmonics be decoupled and hence be determined independent of each other. This would
lead to N systems of equations, each with as many unknowns as there are degrees of freedom in the FEM approximation.
As shown below, the decoupling is trivial in the linear case but, for nonlinear problems, special techniques are needed.

2.1. Linear problems

If the permeability and the reluctivity are independent of the magnetic field, the systems of ordinary differential equa-
tions (17) and (20) become linear, since S in (17) and M in (20) do not depend on x(t ).
Hence, on the one hand, the Fourier coefficients indicated by Fm in (23) and (24) become
   
 d  d
Fm S(ν )x N = S(ν )Xm , Fm M(μ)x N = jmωM(μ)Xm , Fm g(μ, t ) = jmωFm (g). (35)
dt dt

Consequently, (23) and (24) indeed become decoupled, each harmonic can be determined independently:

 
S(ν ) + jmωM(σ ) Xm = Fm (f),
m = 1, 2, . . . , N , (36)
 
S(ρ ) + jmωM(μ) Xm = jmωFm (g), m = 1, 2, . . . , N . (37)

The right hand side vectors in (36) and (37) can be easily computed by traditional Fourier decomposition as in (22). Once
(36) and (37) are solved, the time functions can be obtained via (21).
On the other hand, the discrete Fourier transforms in (33) and (34) simplify to
    
1 1
D S(ν ) + M(σ ) x[1] = S(ν ) + M(σ ) x̂,
t t
       
D M(μ) x[1] − M(μ) x[0] = M(μ) (I − P)x̂, D g[1](μ) − g[0](μ) = (I − P)D g[1] . (38)

Therefore, the discrete harmonics in (33) and (34) are decoupled:

 
  1    
S(ν ) + M(σ ) (I − P) x̂ = D f[1] , (39)
t
 
  1    
S(ρ ) + M(μ) (I − P) x̂ = (I − P)D g[1] . (40)
t
Indeed, these can be written as
 
1  − j2π m
  
S(ν ) + M(σ ) 1 − e N x̂m = Dm f[1] , m = 1, 2, . . . , N , (41)
t
 
1  m   m   
S(ρ ) + M(μ) 1 − e − j2π N x̂m = 1 − e − j2π N Dm g[1] , m = 1, 2, . . . , N . (42)
t
Since the matrix M and the vector g are real, the m-th and the (m + N /2)-th equations in (41) and (42) are complex
conjugate to each other assuming N to be even, i.e. only N /2 linear systems have to be solved. The right hand side vectors
in (41) and (42) can be easily computed by discrete Fourier transformation as shown in (31). Having solved (41) or (42),
the time values can be obtained by inverse discrete Fourier transformation:

1 
N
 T k
D −1 (x̂) = x[1] = x1 x2 . . . xN , xm = x̂k e j2π ·m N . (43)
N
k =1
 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17 9

2.2. Fixed-point iteration technique for nonlinear problems

The fixed-point iteration method for the solution of nonlinear equations reduces the problem to finding the fixed point
of a nonlinear function. The fixed point xFP of the function G(x) is defined as

xFP = G(xFP ). (44)

The fixed point can be determined as the limit of the sequence

 
x( s +1 ) = G x( s ) , s = 0, 1 , 2 , . . . , (45)

provided G(x) is a contraction, i.e. there exists a contraction number −1 < q < 1 so that for any x and y
 
G(x) − G(y)  qx − y (46)

where  ·  is a suitable norm. Furthermore, the sequence (45) converges to the same fixed point independent of the choice
of the initial guess x(0) .
A general nonlinear equation F(x) = 0 can be transformed to a fixed-point problem by selecting a suitable linear operator
A and defining G as

G(x) = x + A −1 F(x). (47)

The fixed-point iterations (45) then become

 
A (s) x(s+1) = A(s) x(s) + F x(s) , s = 0, 1 , 2 , . . . (48)

where the superscript s of A (s) indicates that the linear operator A can be changed at each iteration step to accelerate
convergence.
In case of the ordinary differential equations (17) and (20) obtained by Galerkin FEM techniques, the selection of a linear
operator is straightforward: the permeability or reluctivity has to be set to a value independent of the magnetic field. This
value, μFP or νFP , is not necessarily independent of the space coordinates r, i.e. generally μFP = μFP (r ) or νFP = νFP (r ) are
permeability or reluctivity distributions varying in the problem domain but independent of the field and hence of time. By
the same argument as the one used for the linear operator A above, μFP or νFP can also change at each iteration step. This
(s) (s)
fixed-point permeability or reluctivity function will be denoted by μFP or νFP below.
Once a suitable fixed-point permeability or reluctivity has been selected, (17) and (20) can be iteratively solved by
obtaining x(s+1) (t ) from the equations

 
( s ) ( s +1 ) dx(s+1) (t )  (s) 
S νFP x (t ) + M(σ ) = S νFP − ν (s) x(s) (t ) + f(t ), s = 0, 1 , 2 , . . . , (49)
dt
d   ( s )  ( s +1 )  d   (s)   d  
S(ρ )x(s+1) (t ) + M μFP x (t ) = M μFP − μ(s) x(s) (t ) + g μ(s) , t , s = 0, 1 , 2 , . . . (50)
dt dt dt
at each step. The permeability or reluctivity distributions μ(s) or ν (s) are determined from the solution x(s) (t ), i.e., in
(s) (s)
contrast to μFP or νFP , they are time-dependent. The stiffness matrix S on the right hand side of (49) is obtained with ν
(s) (s)
replaced by νFP − ν (s) and the mass matrix M on the right hand side of (50) is computed with μFP − μ(s) written instead
of μ. Indeed, these matrices depend linearly on ν and μ, respectively.
Since (49) and (50) are linear ordinary differential equation systems, they can be solved by the continuous harmonic
balance method with decoupled harmonics. Indeed, the continuous harmonic balance method yields equations similar to
(36) and (37) to be solved for s = 0, 1, 2, . . . :
  (s) ( s +1 ) (s)    
S νFP + jmωM(σ ) Xm = Fm S νFP − ν (s) x(s) (t ) + f(t ) , m = 1, 2, . . . , N , (51)
  (s)  (s+1)   (s) (s)
 (s)
 
(s)
S(ρ ) + jmωM μFP Xm = jmωFm M μFP − μ x (t ) + g μ , t , m = 1, 2, . . . , N (52)

where x(s) (t ) is obtained from the harmonics similarly to (21) as



N
(s) jkωt
(s)
x (t ) = Re Xk e . (53)
k =1

On the other hand, the time discretized forms of (49) and (50) are:
10 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17

 
 (s)  1 1     (s) 
S νFP + M(σ ) x(s+1) [1] − M(σ ) x(s+1) [0] = S[1] νFP − ν (s) x(s) [1] + f[1], s = 0, 1, 2, . . . , (54)
t t
 
1  ( s )  ( s +1 ) 1   (s)  (s+1)
S(ρ ) + M μFP x [1] − M μFP x [0]
t t
1   (s)      (s)   
= M[1] μFP − μ(s) x(s) [1] + g[1] μ(s) − M[0] μFP − μ(s) x(s) [0] − g[0] μ(s) , s = 0, 1, 2, . . . . (55)
t
Applying the discrete harmonic balance method leads to equations similar to (41) and (42) for s = 0, 1, 2, . . . :
 
 (s)  1  − j2π m
 ( s +1 )   (s)  
S νFP + M(σ ) 1 − e N x̂m = Dm S[1] νFP − ν (s) x(s) [1] + f[1] , m = 1, 2, . . . , N , (56)
t
 
1  (s)  m ( s +1 )
S(ρ ) + M μFP 1 − e − j2π N x̂m
t
1   (s)      (s)   
= Dm M[1] μFP − μ(s) x(s) [1] + g[1] μ(s) − M[0] μFP − μ(s) x(s) [0] − g[0] μ(s) ,
t
m = 1, 2, . . . , N (57)
where x(s) [1] is obtained from the discrete harmonics by inverse discrete Fourier transformation as shown in (43):

 T 1  (s) j2π ·m k
N
(s)
x(s) [1] = x(1s) (s)
x2
(s)
. . . xN , xm = x̂k e N . (58)
N
k =1

A time shift back yields x(s) [0] according to the definition in (27).
The nonlinear iterations of solving the linear systems in (51), (52), (56) or (57) are terminated once the change of μ(s)
or ν (s) between two iteration steps becomes less than a suitable threshold.
The most computational effort is needed for the solution of the N linear equation systems in (51), (52), and N /2 ones in
(56) and (57), respectively. Since these are independent of each other, they can be solved parallel with each core responsible
(s+1) (s+1)
for the solution for one harmonic Xm or x̂m . Once these parallel computations are ready, the right hand side for
the next iteration can be determined by first computing the time function of the solution as in (53) or (58) and then
carrying out the Fourier decompositions indicated in (51) and (52) or (56) and (57). This is the part of the process when no
parallelization is possible, but since the computational effort necessary for it is negligible in comparison to the solution of
the large linear algebraic systems, the method is massively parallel.
One of the most important factors influencing the rate of the convergence of the fixed-point technique is the choice of
the fixed-point permeability or reluctivity. As pointed out above, this is not necessarily constant with respect to the space
coordinates, i.e. it can be selected to be different at each Gaussian integration point of the finite element mesh. The analysis
(s)
of the optimal choice has been carried out in [13], the result for μFP below is taken from there:
 T 
(s) 0
[μ(s) ]2 dt mint ∈[0,T ] (μ(s) ) + maxt ∈[0,T ] (μ(s) )
μFP = max T , . (59)
μ(s) dt 2
0

The optimal fixed-point reluctivity is obtained in a similar way. The permeability μ(s) and the reluctivity ν (s) are functions
of the space coordinates and also of time since they are determined by the magnetic field distribution, itself space- and
time-dependent. According to (59), the fixed-point permeability depends on the space coordinates but not on time. The
computational effort necessary for the evaluation of (59) in each nonlinear iteration step is negligible.

3. Numerical examples

The numerical examples presented here illustrate the industrial applications of the method presented. They have been
taken from recent publications [6–8].

3.1. Time domain analysis of a single-phase transformer

A model of a single-phase power transformer has been analyzed. The model includes a conducting steel tank carrying
eddy currents, a non-conducting core made of high grade laminated steel and two cylindrical coils as shown in Fig. 1. One
fourth of the arrangement has been modeled; the traces of the hexahedral second-order finite elements on the tank and
the core are also shown in the figure. Applying the T , Φ –Φ formulation has resulted in 148,765 degrees of freedom. The
cylindrical coils are not part of the finite element model, they are taken into account by the appropriate choice of the
impressed current vector potential T 0 . The coil currents are described by a sine function with a frequency of 50 Hz.
The nonlinearity of steel in the tank and the core is taken into account, the corresponding B–H curves are shown in
Fig. 2.
 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17 11

Fig. 1. Finite element model of one fourth of a transformer. The core is yellow and the tank is transparent. (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)

Fig. 2. B–H curves of the steel parts. (a) Tank material, (b) core material.

Fig. 3. The flux density in the tank at t = T /4 = 5 ms obtained by the fixed-point method of the paper. The thick line indicates the portion of the cross
section of the tank wall where the time function of the flux is shown in Fig. 5.

The problem has been attempted to be solved by the “brute force” method using N = 40 time steps per period, i.e. with
t = T / N = 0.5 ms. Altogether 10 periods have been stepped through. Simultaneously, the fixed-point method of the paper
has been applied to the discrete harmonic balance equations with the same time step. In this case, since the excitation is
an odd function, the number of different equations (57) is further halved to N /4 = 10.
The magnetic flux density in the tank at t = T /4 and the current density at t = 0 are shown in Figs. 3 and 4 as obtained
by the fixed-point technique. The two time instants have been chosen to approximately give the maxima of the quantities
plotted, since the field of the coils is maximal at t = T /4. Obviously, the tank wall is saturated at this moment.
The time function of the flux through the tank wall along the surface indicated in Fig. 3 is shown in Fig. 5. The first, fifth
and tenth periods obtained from the “brute force” method are plotted along with the result of the fixed-point calculation.
12 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17

Fig. 4. The current density in the tank at t = 0 obtained by the fixed-point method of the paper.

Fig. 5. Time function of the flux through the portion of the tank wall cross section shown in Fig. 3 for the first, fifth and tenth periods of the “brute force”
method as well as the solution obtained by the fixed-point method of the paper.

Obviously, steady state has not been achieved after 10 periods, since the positive and negative maxima of the flux are still
quite different. In view of the tiny change between the periods 5 and 10, the number of necessary periods can be estimated
to be over 100.
In order to show the efficiency of the proposed fixed-point technique, the computation times are compared in the
following. The architecture used is Intel Xeon X5570 at 2.93 GHz with two quad processors. The “brute force” method has
taken about 5000 seconds per period with 5 nonlinear iterations per time step on average. The fixed-point method of the
paper needed 14 nonlinear iterations and altogether 13,000 seconds with 8 processors used. This time would rise to about
70,000 seconds without parallelization, enough to compute 14 periods by the “brute force” technique. As seen in Fig. 5, this
is by far not sufficient to achieve steady state. Indeed, the difference between the fifth and the tenth periods is invisible in
the plot indicating that the “brute force” method requires much more than 14 periods.

3.2. Computation of transformer losses in the frequency domain

The eddy current losses of a transformer can be obtained by integrating the Joule loss density computed from the current
density distribution. The current density can be computed from the potentials as shown in (6) and (7) in case of the A , V – A
formulation and as given in (8) for the T , Φ –Φ formulation. Since the potentials are provided as Fourier series of the form
(21) by the harmonic balance method presented, the current density is obtained as



N
jkωt
J (r , t ) ≈ Re J k (r )e (60)
k =1
 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17 13

Table 1
Name plate data of a single-phase autotransformer.

Rated power 450/450/85 MVA

√ √
System voltage 500/ 3//230 3//13.8 kV
Rated current 1559/3389/6159 A

where J k (r ) is the complex amplitude of the k-th harmonic of the current density. Hence, the eddy current losses are
obtained as
T  
1
N
1 | J (r , t )|2 | J k (r )|2
P eddy = dΩ dt = dΩ. (61)
T σ 2 σ
0 Ωc k=1Ω
c

The iron losses can be computed by integrating the specific losses per unit volume given as a function p (| B |) of the
flux density provided by the manufacturer as described in [5] for the case of sinusoidal time variation. In fact, the specific
losses are customarily given for unit weight but multiplying them by the specific weight yields the losses per unit volume.
Usually, the specific losses are measured for one single frequency f 0 (e.g. f 0 = 50 Hz), this is denoted by p (| B |, f 0 ). In order
to approximately take account of the dependence of the specific losses on frequency, the following algorithm is adopted. It
is assumed that, neglecting excess losses,
     
p | B |, f = p cl | B | f 2 + p hyst | B | f (62)
2 2
where p cl (| B |) = σ π6 d | B |2 (d is the thickness of the laminates, see [2]). Hence, p hyst (| B |) can be obtained as
 
  1   σ π 2 d2
p hyst | B | = p | B |, f 0 − | B |2 f 02 , (63)
f0 6
and, finally,
 
  σ π 2 d2 f   σ π 2 d2
p | B |, f = | B |2 f 2 + p | B |, f 0 − | B |2 f 02 . (64)
6 f0 6
Similarly to the current density, the magnetic flux density is also obtained in the form of a Fourier series when using
the harmonic balance technique:



N
jkωt
B (r , t ) ≈ Re B k (r )e . (65)
k =1

In lack of any better assumption, the specific losses are simply computed for each harmonic from (64) and then added:
N
  
P iron = p | B k |, kω/2π dΩ. (66)
k =1 Ω

As an example, the autotransformer analyzed in [5] is presented here. Its name plate data are given in Table 1.
The FEM model used has been improved in comparison to [5], it consists of 334,110 finite elements. The problem has
been solved using the T , Φ –Φ formulation, resulting in 2,217,625 degrees of freedom for the potentials. The model is shown
in Fig. 6.
Two short circuit computations have been carried out with the winding currents taken to be sinusoidal and the mag-
netization current neglected. In one of them, the method of [5] assuming sinusoidal time variation for all field quantities
has been used and, in the second one, the harmonic balance method of the present paper using N = 9 harmonics has been
employed (only odd harmonics appear in the field quantities). The losses have been computed as described above. The com-
puted losses in the two cases are summarized in Tables 2 and 3, given as a percentage of the total measured short circuit
losses.
These results indicate that in parts of the transformer where significant saturation is present, like in the first laminates
of the core exposed to stray magnetic fields (see Fig. 7), the losses due to the higher harmonics are considerable.

3.3. Analysis of a single-phase transformer with DC bias

Geomagnetically induced currents (GIC) are direct currents that enter and leave the directly earthed neutrals of high-
voltage star connected windings, causing a direct current (DC) bias in the magnetizing current of the transformer [16]. The
frequency of GIC ranges typically from 0.001 Hz to 0.01 Hz, and the peak value was measured to be about 200 A in Finland
in March 1991. In England and Wales, on the National Grid Company (NGC) transmission system, values of 10–15 A are
14 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17

Fig. 6. FEM model of the analyzed single-phase autotransformer. The model comprises one half of the transformer. The tank is shown transparent, the core
is yellow, the clamping plates and the tie bars are shown green. The windings and the tank shieldings are red. (For interpretation of the references to color
in this figure legend, the reader is referred to the web version of this article.)

Table 2
Losses in percentage of total measured losses of autotransformer analyzed. All quan-
tities are sinusoidal.

DC copper losses (measured) 66.92%

AC copper losses (computed from 2D FEM) 21.75%
Tank (computed from model presented) 3.76%
Clamping plates (computed from model presented) 2.80%
Tie bars (computed from model presented) 0.21%
Tank shielding (computed from model presented) 0.51%
Core (computed from model presented) 2.75%
Total 98.70%

Table 3
Losses in percentage of total measured losses of autotransformer analyzed. Harmonics
up to the 9th are taken into account.

DC copper losses (measured) 66.92%

AC copper losses (computed from 2D FEM) 21.75%
Tank (computed from model presented) 2.55%
Clamping plates (computed from model presented) 2.08%
Tie bars (computed from model presented) 0.18%
Tank shielding (computed from model presented) 0.97%
Core (computed from model presented) 6.05%
Total 100.50%

more typical [17]. In recent years, high-voltage direct current (HVDC) transmission is widely used for intercontinental dis-
tribution of electric power. The large DC potential difference between two converting plants also generates a DC that flows
into the windings of the power transformer [20]. The core of the transformer is saturated during the half cycle in which the
bias current is in the same direction as the magnetizing current, causing undesirable effects like increased noise, additional
core losses as well as eddy current losses due to the higher leakage flux.
Here, we focus on solving a nonlinear steady-state power transformer problem under DC bias in the discrete Fourier
domain. The T , Φ –Φ formulation is used with the voltage in the winding directly used as the excitation, i.e. Eqs. (18a),
(18b), (19b) are solved.
The geometry of a single-phase power transformer is shown in Fig. 8, including the core, a winding with the magnetizing
current and a tie bar carrying eddy currents. The model comprises 54,144 second-order hexahedral finite elements.
The tie bar is made of massive steel (same material as the tank in the model of Fig. 1) and the core of laminated steel.
Both ferromagnetic materials are nonlinear, the corresponding B–H curves are shown in Fig. 2. The single winding is driven
by a given sinusoidal voltage of 60 Hz. The current of the winding has a known DC bias.
 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17 15

Fig. 7. Magnetic flux density in the core at the time instant of maximal winding current.

Fig. 8. Model of a single-phase power transformer with a three-limb core.

To validate the fixed-point method, the problem has been first solved by the method in [4]. This method is capable
of predicting the waveform of the magnetizing current based on three-dimensional static finite element analyses of the
transformer. With the aid of a flux–current curve, the waveform of the magnetizing current with the prescribed DC value is
predicted, the computed waveforms have been shown to agree well with measured ones.
The fixed-point method with the time-periodic technique uses N = 40 time steps per period leading to N /2 = 20 equa-
tions in the discrete Fourier domain.
The flux density distribution in the tie bar at t = T /4 is shown in Fig. 9, the tie bar is saturated at this moment.
The waveform of the winding current at a DC component of 45 A obtained by the present method is in good agreement
with the result of [4] as shown in Fig. 10.

4. Conclusion

It has been shown that the use of FEM in conjunction with the continuous or discrete harmonic balance method is
capable of providing the steady-state solution to large, complex real-world eddy current problems with the nonlinearity of
16 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17

Fig. 9. The flux density in the tie bar at t = T /4 obtained by the fixed point-method with a given DC bias of 45 A.

Fig. 10. Comparison of magnetizing current waveforms at DC bias of 45 A.

ferromagnetic media taken into account. The method developed allows for the decoupled computation of the harmonics
and is hence massively parallel.

References

[1] R. Albanese, E. Coccorese, R. Martone, G. Miano, G. Rubinacci, Periodic solutions of nonlinear eddy current problems in three-dimensional geometries,
IEEE Trans. Magn. 28 (1992) 1118–1121.
[2] G. Bertotti, Hysteresis in Magnetism, Academic Press, 1998.
[3] O. Bíró, Edge element formulations of eddy current problems, Comput. Methods Appl. Mech. Engrg. 169 (1999) 391–405.
[4] O. Bíró, S. Ausserhofer, Prediction of magnetising current waveform in a single-phase power transformer under DC bias, IET Sci. Meas. Technol. 1 (2007)
2–5.
[5] O. Bíró, U. Baumgartner, G. Koczka, G. Leber, B. Wagner, Numerical modelling of transformer losses, in: Proceedings of the International Colloquium
Transformer Research and Asset Management, Cavtat, Croatia, November 12–14, 2009, Session I, Numerical modelling, 9 pages.
[6] O. Bíró, U. Baumgartner, K. Preis, G. Leber, Finite element method for nonlinear eddy current problems in power transformers, in: Proceedings of the
International Colloquium Transformer Research and Asset Management, Dubrovnik, Croatia, May 16–18, 2012, Session I, Numerical modelling, 10 pages.
[7] O. Bíró, Y. Chen, G. Koczka, G. Leber, K. Preis, B. Wagner, Steady-state analysis of power transformers under DC bias by the finite element method
with the fixed point technique, in: 18th International Conference on the Computation of Electromagnetic Fields, Sydney, Australia, July 12–15, 2011,
11. Numerical techniques, Paper 429, 2 pages.
[8] O. Bíró, G. Koczka, K. Preis, Fast time-domain finite element analysis of 3D nonlinear time-periodic eddy current problems with T, Φ –Φ formulation,
IEEE Trans. Magn. 47 (2011) 1170–1173.
 O. Bíró et al. / Applied Numerical Mathematics 79 (2014) 3–17 17

[9] O. Bíró, K. Preis, An efficient time domain method for nonlinear periodic eddy current problems, IEEE Trans. Magn. 42 (2006) 695–698.
[10] O. Bíró, K. Preis, G. Buchgraber, I. Ticar, Voltage driven coils in finite element formulations using a current vector and a magnetic scalar potential, IEEE
Trans. Magn. 40 (2004) 1286–1289.
[11] E.O. Brigham, The Fast Fourier Transform and Its Applications, Prentice Hall, Englewood Cliffs, NJ, 1988.
[12] T. Hara, T. Naito, J. Umoto, Time-periodic finite element method for nonlinear diffusion equations, IEEE Trans. Magn. 21 (1985) 2261–2264.
[13] G. Koczka, S. Ausserhofer, O. Bíró, K. Preis, Optimal convergence of the fixed-point method for nonlinear eddy current problems, IEEE Trans. Magn. 45
(2009) 948–951.
[14] G. Koczka, S. Ausserhofer, O. Bíró, K. Preis, Optimal fixed-point method for solving 3D nonlinear periodic eddy current problems, COMPEL 28 (2009)
1059–1067.
[15] G. Koczka, O. Bíró, Fixed-point method for solving nonlinear periodic eddy current problems with T , Φ –Φ formulation, COMPEL 29 (2010) 1444–1452.
[16] S. Lu, Y. Liu, Harmonics generated from a DC biased transformer, IEEE Trans. Power Deliv. 8 (1993) 725–731.
[17] P.R. Price, Geomagnetically induced current effects on transformers, IEEE Trans. Power Deliv. 17 (2002) 1002–1008.
[18] Y. Takahashi, T. Tokumasu, A. Kameari, H. Kaimori, M. Fujita, T. Iwashita, S. Wakao, Convergence acceleration of time-periodic electromagnetic field
analysis by the singularity decomposition-explicit error correction method, IEEE Trans. Magn. 46 (2010) 2947–2950.
[19] Y. Takahashi, T. Iwashita, H. Nakashima, T. Tokumasu, M. Fujita, S. Wakao, K. Fujiwara, Y. Ishihara, Parallel time-periodic finite-element method for
steady-state analysis of rotating machines, IEEE Trans. Magn. 48 (2012) 1019–1022.
[20] X. Zhao, J. Lu, L. Li, Z. Cheng, T. Lu, Analysis of the saturated electromagnetic devices under DC bias condition by the modified harmonic balance finite
element method, in: 14th Biennial IEEE Conference on Electromagnetic Field Computation (CEFC), Chicago, IL, USA, May 9–12, 2010, http://dx.doi.org/
10.1109/CEFC.2010.5480338.