IEEE Transactions on Power Delivery, Vol. 13, No.

2, April 1998 605

SIMULATION OF TRANSMISSION LINE TRANSIENTS USING VECTOR FITTING AND MODAL DECOMPOSITION
Bjem Gustavsen* @I) Adam Semlyen (LE")
Department of Electrical and Computer Engineering
University of Toronto
Toronto, Ontario, Canada M5S 3G4

* On leave from the Norwegian Electric Power Research Institute (EFI), Trondheim, Norway.
-
Abstract This paper introduces a fast and robust method for rational propagation and modal characteristic admittance. Unlike
ftting of frequency domain responses, well suited for both scalar and traditional Marti-fitting [3], the resulting rational function
vector transfer functions. Application of the new method results in approximation is not restricted to real poles and zeros, which is
increased computational efficiency for transmission line models using shown to be advantageous when fitting the transformation matrix
modal decomposition with frequency dependent transformation and the modal characteristic admittance.
matrices. This is due to the fact that the method allows the fitted
2 MODAL DECOMPOSITION
elements of each eigenvector to share the same set of poles, and that
accurate ftting can be achieved with a relatively low number of poles. ii 'fir
1 INTRODUCTION
+ t
i - 1 1
Accurate calculation of electromagnetic transients in power +
systems requires the frequency dependent effects of transmission
+ ir
lines to be taken into account. Frequency dependent line models V
can in principle easily be formulated in the time domain via
-
numerical convolutions [l], but the resulting model is
computationally inefficient. The efficiency is greatly improved if
the impulse responses for the line are fitted using rational Fig. 1 Traveling waves at transmission line end
functions in the frequency domain [2-31, z-domain [4-51, or
approximated directly in the time domain [6-71, as this leads to a The frequency domain solution of the traveling wave equation
recursive formulation of the time domain convolution integrals. can at each end of a transmission line be expressed [7] by the well
Further savings can be achieved by introducing modal known matrix-vector expression
decomposition [8], as the transformation matrix is almost constant Y,v-i=2ii =2Hipr (1)
for many overhead lines of practical interest. In such cases, the
number of scalar convolutions to be carried out is reduced from where the propagation function H and the characteristic
4n2 to 4n, where n is the number of conductors. However, in admittance Y, are given by
case of cable systems, multi-circuit overhead lines and strongly H=exp(-mI) (2)
asymmetric overhead lines, the transformation matrix may be
strongly frequency dependent. Taking this frequency dependence Yc=Z-" (3)
into account requires the handling of a frequency dependent Z and Y are the series impedance and shunt admittance per unit
transformation matrix [9]. This increases the number of length of the transmission line. For an n-conductor system these
convolutions to 4n2 +4n , making the modal domain approach are nxn matrices.
slightly more time consuming than the phase domain approach. Equation (1) represents n coupled scalar equations, but may
The propagation function H may in some cases of cable be replaced by n uncoupled equations by introducing modal
systems contain widely different time delays for the individual quantities :
modes. In such cases it is very dinicult to fit H with rational
functions directly in the phase domain, as a very high order fit may i=Tl im (4)
be needed. However, this problem is easily overcome with the v=TVvm (5)
modal domain models, because there is only a single time delay
associated with each mode of propagation. The modal time delays Here, TI and Tv are the right eigenvector matrices of YZ and
can therefore readily be eliminated from the modal components of ZY respectively. Superscript m denotes modal quantities.
H [lo], yielding smooth functions which can be fitted with low Substituting (4) and (5) in (1) gives
order rational polynomials. ycmv -im =2Hm'far
'm
(6)
In this paper we briefly review the methodology of modal
decomposition applied to transmission line modeling. We then The diagonal matrices Hm and Y," are related to their phase
introduce a new powerful fitting technique which allows us to fit domain counterparts by :
the transformation matrix column-by-column. This strategy is
shown to result in substantial savings in computational time for H=TI HmTL1 (7)
the time step loop, as compared to traditional element-by-element Y, =T, Y," TI' (8)
fitting. The method is also used to fit the responses for modal
where we have used [8] the relation :
Tv =TYT (9)
PE-347-PWRD-0-01-1997 A paper recommended and approved
In practical calculations the modal domain transmission line
by the IEEE Transmission and Distribution Committee of the IEEE
model is linked to the phase domain host program by the relations
Power Engineering Society for publication in the IEEE Transactions
on Power Delivery. Manuscript submitted July 26, 1996; made v m =Tf' v=T: v (10)
available for printing January 8, 1997.
i=TI im (11)
Thus, the modal calculations have been manipulated to involve
only the matrices TI , TI' , Y", ,and Hm.

0885-8977/98/$10.00 0 1997 IEEE

606
3 VECTOR FITTING BY OPTIMAL SCALING prescribed stable poles given in the diagonal matrix 2. In
addition, we append a unity to the original vector :
In this chapter we show that substantial savings in the time
step loop can be obtained if the elements of each eigenvector are
fitted using the same set of poles. This is achieved by the
introduction of a new fitting methodology-vector fitting by
optimal scaling. In order to prevent ambiguity of the solution we impose the
condition
3.1 State eguation realization
A linear system can in general be represented by the state equation
realization (SER)
which implies that (J approaches-mity at high frequencies. The
i=Ax+Bu (12) elements of the column vector b are arbitrarily set to unity.
y =Cx+Du (13) Equation (17) is manipulated into an overdetermined set of linear
equations, one equation for each frequency point. This yields the
In transient calculations we need SERs for T I , TF , U,", and sparse system
H"' . Note that in the calculations we need only access to the
variables for input U and output y , whereas the state variables Fx=g (19)
x are used as "internal" variables only. This allows us to use a where the solution vector x contains the elements of as well
diagonal A-matrix, since this corresponds to choosing a particular as of D(S) . Equation (17) is split into itsjeal and imaginary part
set of state variables (via a similarity transformation). The when assembling (19), thereby ensuring C to have real elements
selection of a diagonal A-matrix reduces the computational burden only.
of the time domain calculations. Our experience is that specifying the poles in 2 to be real
Let G(S ) be a matrix transfer function of dimension n x n . G and logarithmically distributed gives a very accurate fit for tI .
is related to its SER by the expression This choice is somewhat arbitrary and is based on the requirement
G(s)= C(sI-A)-' B+D (14) of adequate covering of the whole frequency range considered (see
Appendix A).
The elements of the diagonal A-matrix are seen to be the poles of In order to get a SER for the original vector t j we partition
G . In the case of TI , U," and H ,the transfer matrices have an the SER in (17) into two parts , h and h, , where h, is the last
infiite number of poles Hence, our task is to find a finite order element in the SER :
approximation.
Assume that we fit the transformation matrix T,(s) element-
by-element by a SER of order N . This implies that each element
is represented by a SER of dimension N . For an n-conductor
system the total number of poles for each column of TI equals It is seen from (20) that h, represents a SER for ci . Because the
poles of hl are identical to those of h ,it follows that the poles of
N,, =nN (15) ty become equal to the zeros of h , . The zeros of hl are
calculated directly from its SER, as shown in Appendix C.
We can, however, do better than this. By suitable frequency We next calculate the SER of tj by means of (21), with the
dependent normalization the elements of TI become smooth zeros of hl used as elements (poles) for A :
functions of frequency. This makes it possible to choose the same
poles for all elements of a column of TI . (See Appendix A for t,O=C(sI-A)-'b+d (21)
additional justification). Thus, the total number of poles for a
single column then becomes equal to As in (17), d=t;(s=m), and b is a column vector of ones.
Equation (21) is manipulated into an overdetermined linear set of
Ntot =N ( 16) equations, similarly to (19). However, the solution vector x now
contains only the elements of C as no scaling is involved.
A detailed analysis (Appendix B) shows that this will increase the It should be noted that the new poles used in (21) are both real
computational eficiency for TI in the time step loop by
poles and complex conjugate pairs. This is also the case for the
approximately a factor of 3. It should be noted that an additional zeros for the transfer matrix corresponding to the SER of ty .
requirement of all eigenvectors sharing the same poles would not
The fitting technique by optimal scaling is equally well suited
lead to further savings. This is due to the fact that the SER for for fitting scalar functions as vectors. In (17) the vector t: is then
each vector has to be implemented separately. Similarly, nothing replaced by the scalar function. We use this technique for finding
is achieved by r e q u h g the elements of Y," and H"' (diagonal) the rational function approximation for the elements of the
to have the same poles.
diagonal matrices H m and Y," .
3.2 ODtimal scalinq In case of T f , the SER is calculated from the SER of TI
u s L formulae
~ shown in App-dk D.
In WS section we introduce a new, general fitting methodology
which allows the columns of a matrix to be fitted with the same 3.3 Stability
set of poles. The method is explained for the transformation Because the "new" poles used in (21) are zeros that were
matrix T, ,but will also be used for fitting the diagonal elements
calculated without any restrictions, they may in principle turn out
of Y", and H"'.
As in [9] we normalize the eigenvectors by fixing one element
to be unstable. Our experience is that when the function to be
to unity, giving a transformation matrix T:, . Artificial eigenvector fitted represents a stable system, then the technique by optimal
scaling also results in a stable fit. The reason for this is simply
switchovers are avoided by means of a swtching-back procedure, that if the rational function approximation contained an unstable
similarly as described in [111 pole, then the resulting fit would be poor. An exception is when
In the following we introduce an intermediate scaling the unstable pole is canceled by an unstable zero. This problem is,
(smoothing) of each column tj with the purpose of obtaining however, easily overcome by deleting the unstable poles in (21)
poles suitable for the fitting of the original (unscaled) vector tj . prior to determination of C .
We use for smoothing a scalar (complex) scaling fimction ~ ( s )so
that the scaled column tI = o t j can be fitted with a set of
 607

3.4 Removal of time delavs from the modal propagation functions 4 TIME DOMAIN IMPLEMENTATION

Many years experience with Marti-fitting has shown that the The transmission line model as defined by the traveling wave
scalar transfer function HT can be accurately approximated by a equation ( 1 ) and ( 6 ) is transformed from the frequency domain
minimumphase shift function plus a time delay : into the time domain by application of the convolution theorem.
We start by noting that a convolution between an arbitrary matrix
H r (a)=hmh(o)exp(-jm) (22) G and an input vector U may be expressed :
In the case of Marti-fitting, only the magnitude function was y=G*u=Gou+h (27)
considered in the fitting process. This allowed the time delay to be
calculated by comparing the phase angle of the propagation where * denotes the convolution operator. Gou is an
function with that of the rational approximation. However, in case instantaneous term and h a history term. Considering one end of
of fitting by optimal scaling, both the real and imaginary part of the line, the modal domain traveling wave equation ( 6 )becomes in
H? are used in the fitting process. This makes it necessary to the time domain
remove the time delay before the fitting is carried out, which is
done by multiplying H r with the factor exp(jw7) . As shown in Ycm*vm-im=2Hm*igr (28)
Appendix E, the time delay for each mode can be calculated by the Applying (27)to (28) gives
exmession
(Y$vm+hl)-im =2hz (29)
( H r is equal to 0 due to the time delay between the line ends.)
where 0 is the highest frequency point of interest, v is the The modal domain currents and voltages in (29) are transformed
modal velocity and 1 the length of the transmission line. Because into the phase domain using (lo), ( 1 1 ) . and (27) :
accurate representation of the "toe portion" of H r ( t ) would
require a very high order fitting, we choose R to be the frequency vm=T&v+h3 , i=TIOim+h4 (30)
point where IHr(o)I=O.l,as illustrated in figure 2. This gives an Substituting (30) in (29) and premultiplying with TIo gives the
accurate fit for o <R ,and a reasonable representation for w >Q . final result :
We ensure that the rational approximation approaches zero as
o+oo by specifying the D-matrix of the SER to be zero when
(TIoY$T&)v-i=T1~[2h~ -hl -Y$h,]-h, (31)
doing the fitting. which is conveniently expressed by the Norton equivalent in
figure. 3. This representation is similar to the one used in [9].

R 0
Fig. 3 Network representation of line end
Fig. 2 Upperfrequency limit R
The history terms h h z , h , , h4 and the instantaneous terms
Lh,;, (Q) in (23) is calculated by a formula derived by Bode Y $ , TIo and T i 'are calculated from the respective SERs
[14], which relates the phase angle for a minimum phase shiR assuming trapezoidal integration.
function to its magnitude function. The formula is shown in (24),
in a slightly rewritten form : 5 CALCULATED RESULTS FOR CABLE SYSTEM
In the following we show results for the 66 kV cable system in
figure 4 . The cable series impedance Z is calculated using
simplified formulae given in [12].

and w
U=h- (26)
0 1
The first term alone gives a good estimate for the phase angle.
However, an improved result is achieved by including the second
term A ( u ) . We limit the integration in (25) to the interval \ E,, =2.85
o~[O.lQ,loQ]. This can be done since the shape of the &*g E,~ =2.51
magmtude function more than one decade of frequency away has
only a very small effect on the phase angle at w = Q . The Fig. 4 Cable system data
derivatives in (24)-(26) are evaluated by simple numerical
differsntiutivn, as iii s+lsv the F ~ Q Ffvr the integration. we typ;dly 5.1 Rational function approximation
use 10 frequency points in the integration, which results in the The elements of TI , YF and H were fitted using the optimal
evaluation of A h ; , (Q) to be very fast. scaling technique in the frequency range O.l€&-lM€k, with 5
frequency points per decade of frequency. The cable length was
1O h .
608
Figure 5 shows the magnitude of the six elements of the third
column of TI , and the magnitude of the complex deviation
(fitting error). The fit is seen to be very accurate.

B I 1

1oo 1o2 1o4 lo6

Frequency [Hzl
deviation
0 Fig. 7 Modal characteristicadmittance (7th order approximation)
1oo 1o2 1o4 1o6
Frequency [Hz]
Figure 8 shows the fitted elements of the modal propagation,
H m,after removing time delays by (23). The elements were fitted
Fig. 5. Elements of third column of TI (5th order approximation) separately, using 8 real poles as "starting poles", logarithmically
distributed between O.l& and 1MHz. The fit is seen to be quite
As "starting" poles in (17) we used 5 real poles, good, perhaps with the exception of the toe portion of element #6.
logarithmically distributed between 0.1Hz and 1MHz. The
calculated "new" set of poles used in (2 1) is shown in table 1. It is 1
seen that the fitting technique resulted in a complex conjugate
pair, in addition to three real poles.
Table 1 Poles of rationalfunction approximation
-1.7714E6 -2.6904E3 -944.96kj473.82 -2.0957
-Accurate
Figure 6 shows the phase angles corresponding to the
magnitudes in figure 5. The angles of the rational function
approximation are seen to be in close agreement with the original
angles. We also note that the phase angles of elements (2,3), (6,3)
and (4,3) have a net phase shift of -180", whereas the
corresponding magnitude functions (figure 5) are constant at both
very low and very high fiequencies. This implies that these 0
elements contain a zero in the right half plane. This shows that in 1oo 1o2 io4 1o6
case of cable systems, one cannot assume all the zeros to lie in the Frequency [Hz]
left half plane, as is traditionally done in Marti-fitting [ 3 ] . Fig. 8 Modal propagation (61th order approximation)
5.2 Time domain simulation

-
0)
0 -Accurate
The accurate representation of the fi-equency dependence of the
transformation matrix is particularly important when calculating
25 induced currents [13] and induced voltages [7] in cable sheaths. In
-
al
0)
what follows we show calculated results for the cable system of
2 0 figure 4, (cable length=lOkm). The responses were fitted in the
a interval IHz-lMHz using 5-6 poles for each column of T I , 7
fn
a poles for each element of Y," and 8 poles for each element of
L
a H m . The SER for T; was calculated directly from the SER of
TI ,as previously mentioned

-200 5.2.1 Short circuit response

1oo 1o2 1o4 1o6 In this test a 1 p.u. step voltage was applied to the core of the
Frequency tHz1 leffmost ,,ble (S-ndLlg -d), -h& thq vther ovrea a d JhEathci
were grounded at this end. At the opposite end (receivmg end) all
Fig. 6 Phase angle of elements of third column of TI cores and sheaths were grounded.
Figure 7 shows the fitted elements of the modal characteristic Figure 9 shows the calculated current flowing in the core and
admittance matrix, Y," . Each element was fitted separately, using sheath of the energized cable at the sending end (solid line). In the
7 real poles as "starting" poles. These were logarithmically same figure are also shown the exact values as calculated by a
distributed between 0.1% and 1MHz. Also is shown the Fourier transform (dotted line). The currents by the simulation
magnitude of the complex deviation. Again, the fit is very model are seen to be on top of those by the Fourier method,
accurate. implying both that the rational function approximation is very
good, and that the model has been correctly implemented in the
time step loop.
 609

Diameter, core conductor : 21.3 mm

-Simulation Thickness, sheath conductor: 0.35 mm
Thickness, outer insulation : 2.2 mm
__---Fourier method pc =4.179E - 8r.Q.m]

-0.2
sheath current
J ps =1.724E-8[Q.m]
E ~ ~ = E ,=2.3
- ~
0 1 2 3 4 5
Time [ms] Fig. 11 Single core coaxial cable.
Fig. 9 Short circuit current in core and sheath of energized The characteristic admittance for the coaxial mode is shown in
cable (step response) figure 12 as the thickness of the main isulation is varied between
5.2.2 Open circuit resuonse 3 mm and 7 mm. We note that a sharp peak occurs as we decrease
the insulation thickness. It should be noted that such peaks are
In this test a 1 p.u. step voltage was applied to the core of the accurately reproduced using our fitting method, as complex
leftmost cable (sending end), while the other cores and sheaths poledzeros are automatically introduced.
were grounded at this end. At the opposite end (receiving end), all
cores and sheaths were open circuited. Figure 10 shows the
calculated receiving end sheath voltage of the energized cable. The
calculated voltage is seen to be in good agreement with the
theoretically accurate solution by the Fourier method.

- Simulation
-
?
0.02 _---- Fourier method

23 0.01
Q)
ol
m I
= O 0'
s 1oo 1o2
Frequency [I-IZ]
1o4 1o6
-0.01
Fig. 12 Influence of insulation thickness on Y," (coaxial mode)
-0.02 6 APPLICATION TO OVERHEAD LINES
0 1 2 3 4 5
Time [ms] The modal domain approach described in the paper has also been
applied to a number of diffaent overhead lines, including both
Fig. 10 Open circuit voltage on sheath of energized cable single and multi-circuit lines. It turned out that the success
(step response) obtained with cables could not be automatically achieved with
overhead lines. Our fitting technique has no difficulties in fitting
5.3 Amlication to other cable aeometries the transformation matrix and the modal quantities, but the
In this study we have only considered systems of single core resulting fit will in general give unstable poles in TI and Y r .
coaxial cables. It appears that in these cases the transformation This problem is due to the fact that unstable poles are needed in
matrix and the modal quantities for characteristic admittance and order to achieve an accurate fit.
propagation are in general smooth functions that can easily be As an example, consider the 132kV untransposed single
fitted with rational functions. An interesting observation is that in circuit line in figure 13.
the case of cables with low system voltage (e.g. 12,24 kV), there
may appear "peaks" in the transformation matrix and in particular
in the modal characteristic admittance. This appears to be a result
of that the thickness of the sheath-ground insulation is no longer
d=2cm
O O4Sm O t
much smaller than that of the main insulation, so that the
capacitances of these two insulations become comparable.
As an example, consider the single core coaxial cable in
figure 11, which has a thickness of 2.2 mm for the sheath-ground
I
I& \L
llm

Qm
insulation.
Fig. 13 Single circuit overhead line
Figure 14 shows the column of the transformation matrix
corresponding to the ground mode (magnitude functions). Two of
the elements are equal to unity,but one element (2,l) is frequency
dependent, particularly between 1Hz and 1OOHz.
610
Figures 16 and 17 show the magnitude functions and phase
I I
angles for the contributions Yck to element (1,l) of & . The
fitting technique was modified so as to fit to magnitude only with
stable poles and zeros. We note that although the magnitude
hctions have been accurately fitted, the phase angles are in error
for contributions #1 and #3. However, when all three elements
were added together to form the phase domain element Yc(l,l),
-Accurate the resulting element could be accurately fitted with stable poles.
_----Fitted to magnitude
z with stable poles
Accurate
----- Fitted to magnitude
n
"
with stable poles
1o-2 1oo 1o2 1o4 1o6 2
0)
Frequency [Hz] U
3
r l
Fig. 14 First column of TI (magnitudefunctions) c

The phase angle for element (2,l) is shown separately in / #3

figure 15. Also is shown the phase angle that results when fitting
this element to magnitude only with stable poles and zeros. We
i /7
note that in this fiequency interval the fitted response has a
0
negative phase shift whereas the original response has a positive 1oo 1o2 1o4 1o6
phase shifl. In this case it is not possible to obtain an accurate fit
using stable poles only. Frequency [Hz]
Fig. 16 Magnitude of modal contributions Y,k(l,l) to Y,(l,l).
10 I I \ I

-
a
Q)
40

3
-a
Q)
Fitted to magnitude
$ 1 I t
\
/'
m I
m 20 . with stable poles
'\
al
a ', Fitted to magnitude m
r
v)

\/
with stable poles a
n
V

1om2 1oo 1u2 1o4 1o6

Frequency [Hz]

Fig. 17 Phase angle of modal contributions&(1,1) toY,[l,l)

7 OVERVIEW OF RESULTS
The frequency dependence of the modal transformation matrix
encountered in transmission line transient calculations can be
handled by a modal domain approach involving an additional
convolution for the transformation matrix For an n-conductor
system this gives a total number of convolutions equal to
4 n 2 + 4 n , as compared to 4n when a constant transformation
matrix is used. A significant computational burden obviously lies
in the 4n2 convolutions for the iransfomation matrix.
In this paper we reduce the computational burden for these
convolutions to approximately one third by using vectorjtting for
the transformation matrix. Vector fitting means that all elements
in a vector share the same set of poles. This has been possible by
where Y$ is element (k,k) of Y,". When the sum in (32) is the introduction of a new jitting technique--rational fitting by
taken to form a phase domain element, the poles of that element optimal scaling. The method produces a state equation reaIization
becomes equal to the poles of the contributing terms The resulting (SER) for the entire vector by solving a well-conditioned
zeros of the phase domain element will, however, be different overdetermined set of linear equations. The fitting routine takes as
from the zeros of the contributing terms. This permits unstable input the vector as function of frequency plus a set of starting
poles in Yck to be canceled by unstable zeros. Thus, some Yck poles. The key idea of the method is to multiply the vector by an
could be unstable while the resultant is stable In case of "optimal scalar", o(w), so that the starting poles give a good fit
overhead lines this problem appears to frequently take place as for the vector. By appending a unity to the original vector, the
non-perfect cancellations between unstable poles and unstable zeros for ~ ( w )are then readily obtained. These zeros are now
zeros. This makes it generally impossible to obtain a perfectly used as new starting poles, and the original vector is in the end
accurate fit for the modal components using stable poles only. fitted by solving an overdetermined set of linear equations. One
strength of the method lies in that the accuracy of the resulting fit
 611
is only slightly dependent on the location of the starting poles. Our 9 ACKNOWLEDGMENTS
experience is that the startingpoles can be assumed to be real and
logarithmically distributed. The fitting routine produces a SER Financial assistance by the Natural Sciences and Engineering
with both real and complex poles. Research Council of Canada is gratefully acknowledged. The first
In the paper we implemented in time domain a modal domain author wishes to express his gratitude to the Norwegian Electric
model with fTequency-dependent transformation matrix, similarly Power Research Institute (EFI), Trondheim, Norway, for granting
as in the L. Marti model [9]. The new fitting methodology was and financing his leave at the University of Toronto.
used to find SERs for the transformation matrix, the modal 10 REFERENCES
characteristic admittance matrix and the modal propagation
matrix. Application to a 66kV single core cable system showed [l] H. Nakanishi and A Ametani, "Transient calculation of a transmission
that highly accurate results could be achieved with a relatively h e using superposition law", IEE Proc., vol. 133, pt. C, no. 5, July 1986,
low number of poles : 5-6 for each column of the transformation pp. 253-269.
matrix and 7-8 for each element of the modal characteristic and [2] A. Semlyen, "Contributions to the Theory of Calculation of
propagation matrix. Of particular interest was the observation that Electromagnetic Transients on Transmission Lines With Frequency
zeros in the right half plane are intrinsically necessary for some Dependent Parameters", IEEE Trans. PAS, vol. 100, no. 2, February
elements of the transformation matrix, while traditional Marti- 1981, pp. 848-856.
[3] J. R. Marti, "Accurate Modelling of Frequency-Dependent Transmission
fitting [3] assumes all zeros to lie in the left half plane.
Application of the new fitting method, however, automatically Lines in Electromagnetic Transient Simulations", IEEE Trans. PAS, vol.
produces a SER which has the correct zeros. Another observation 101, no. 1, January 1982, pp. 147-157.
was that in case of cables where the outer insulation is not much [4] G. Angelidis and A. Semlyen, "Direct Phase Domain Calculation of
thinner than that of the main insulation (low system voltage), the Transmission Line Transients Using Two-sided Recursions", IEEE
transformation matrix and in particular the characteristic Trans. PWRD, vol. 10, no. 2, April 1995, pp. 941-949.
admittance matrix may have strong peaks. These can, however, be [SI T. N o h N. Nagaoka and A. Ametani, "Phase Domain Modeling of
easily fitted using optimal scaling, as the needed complex poles Frequency-Dependent Transmission Lines by Means of an ARMA
will automatically be introduced. Model", IEEE Trans. PWRD, vol. 11, no. 1, January 1996, pp. 401-411.
In case ofoverhead lines the method ofmodal decomposition [6] A Ametani, "A Highly Efficient Method for Calculating Transmission
with a frequency dependent T seems to be less successful than for Line Transients", IEEE Trans. PAS, vol. 95, Sept./Od. 1976, pp. 1545-
cables. Application to a single circuit line showed that T could 1551.
not be fitted accurately with stable poles only. It was concluded [7] B. Gustavsen, J. Sletbak and T. Henriksen, "Calculation of
that in case of the characteristic admittance, we may get an Electromagnetic Transients in Transmission Cables and Lines Taking
unstable modal decomposition. Similar results have also been Frequency Dependent Effects Accurately Into Account", IEEE Trans.
obtained for other line geometries,' including multi-circuit lines. PWRD, vol. 10, no. 2, April 1995.
This fact suggests that a directphase domain approach may be the [SI L. M. Wedepohl, "Application of matrix methods to the solution of
right choicefor such cases. travelling-wave phenomena in polyphase systems", Proc. IEE, vol. 110,
The method of optimal fitting has been implemented as a no. 12, December 1963, pp. 2200-2212.
subroutine where the number of poles is given in the input. This [9] L. Marti, "Simulation of Transients in Underground Cables With
allows the vector and scalar functions to be repeatedly fitted with Frequency-Dependent Modal Transformation Matrices", JEEE Trans.
an increasing number of poles, until an error criterion is met. PWRD,vol. 3, no. 3, July 1988, pp. 1099-1110.
Consequently, in the case of overhead lines, columns which turn [lo] A. Semlyen and A. Roth, "Calculation of Exponential Propagation Step
out to be frequency independent are fitted with a constant vector -
Responses Accurately for Three Base Frequencies", IEEE Trans. PAS,
(order 0). Thus, in the case of a constant transformation matrix, vol. 96, no. 2, MarcWApril 1977, pp. 667-672.
the proposed simulation model becomes similar to the J. Marti [l11L. M. Wedepohl. H. V. Nguyen and G. D. Irwin,"Frequency-Dependent
model [31. Transformation Matrices for Untransposed Transmission Lines Using
Newton-Raphson Method", paper 95 SM 602-3 PWRS, presented at
8 CONCLUSIONS IEEE Summer Meeting, Portland, 1995.
A new method has been introduced which allows vectors and [12]L. M. Wedepohl and D. J. Wilco& "Transient analysis of underground
scalars to be fitted with rational functions containing complex power-transmission systems. Systemmodel and wave-propagation
poles and zeros, in addition to real ones. The fitting process is charaderistics", Proc. IEE, vol. 120, no. 2, February 1973, pp. 253-260.
very robust and there is no problem in specifLing a very high order [13] L. Marti, "Simulation of electromagnetictransients in underground cables
for the fitting if needed. using the EMTP", Proc. AF'SCOM 93, Hong Kong, December 1993.
The new fitting method has been applied with success to cable [14]H. W. Bode, "Network analysis and feedback amplifier design", D. Van
systems using modal decomposition and frequency dependent Nostrand, New York, 1945 (book).
transformation matrices. In this model, we need to fit the 11 APPENDICES
transformation matrix and the diagonal matrices for propagation
and characteristic admittance. The resulting cable model is A - Fitting With Simple Fractions
computationallyefficient for the following reasons : Simple fractions
Accurate fitting can be achieved with a relatively low number
of poles. An order of 5-10 is usually sufficient.
The new fitting method allows each vector of the with real poles p , constitute a basis in the function space for
transformation matrix to be fitted with the same set of poles. fitting smooth functions h ( o ) . Indeed such functions are smooth,
This gives approximately a 3-fold increase in computational as illustrated in figure A.l for pi=-2.5,-5,-7.5,-10.
efficiency in the time step loop for all convolutions involving However, the functions g i ( o ) do not span the whole
the transformation matrix. function space and therefore they are useful only for relatively
In the case of overhead lines it is not always possible to obtain an smooth functions. We can see this if w e represent
accurate fit with stable poles when relying on modal g i ( o k ) , w k = o I , w z...,O N , as vectors forming the columns of a
decomposition with a frequency dependent transformation matrix. matrix G . Figure A.2 shows the singular values obtained by
A stable model will in such cases be an approximation. Singular Value Decomposition (SVD) of a 100x100 G- matrix
(for 100 points for p i and KJ uniformly distributed from 0 to
612

10). It is seen that only a fraction of the singular values are not The computational time inferred from (B.5) - (B.6) is then found
very small. Therefore, identifLing simple fractions (A.l) for to be about equal to :
fitting leads to a poorly conditioned problem, especially when the Km(nN+n+2N)(ka+km)wnN(ka+km) 03.7)
number of simple fractions is high. (Only the orthogonal Lfeach element is fitted separately, each element gives a SER with
U - matrix of the SVD (G=USVH) provides a well-conditioned
one input and one output. Since there are n elements we get:
basis.) Conversely, a larger set of simple fractions (A.l) that
covers the whole frequency range of interest can be chosen with Kmn(N+1+2N)(ku+k,)w3nN(ku +k,) (B.8)
great flexibility when fitting any number of smooth functions By compaSing'(B.7) with (B.8) we see that time domain evaluation
NCO). of a SER based on vector fitting gives about 3 times faster
1, I evaluation than that of a SER based on element-by-element fitting.
C - Calculating New Set of Poles by Optimal Scaling
The new set of poles produced by the optimal scaling is equal to
the zeros of the particular element h of the scaled vector. These
zeros are calculated as follows :
The SER of the scaled vector is given by (20). The SER of h has
the same A and b as that of the entire vector, and a row vector cT
and scalar dl equal to the corresponding row of C and d for the
,
vector. From (12) and (13) we get the SER of h in the frequency
domain :
0'
0 2 4 6 8 10 sx=Ax+bu , y=cTx+dlu (C.1)
Frequency [rads] Interchanging input and output gives the SER of the inverse of h :
Fig. A1 Frequency domain behavior of simple fractions sx=(A-bd;' cT)x+bd;' y , u=-d;' cTx+d;' y (C.2)
The zeros of hl are then calculated as the eigenvalues of the
I 1
matrix A'=A-bd;'cT. Note that d;' always exists because h ,
approaches unity at high frequencies.
D - State Equation Realization for &r
Let A ,B ,C andD represent the SER for TI . By transposing
(15), we find a SER for TI' given by A ' = A , B ' = C T , C ' = B T ,
D'=D~.
E - Time Delays For Modal propagation
Each element of the modal propagation matrix H m (diagonal)
can be written as :
0 20 40 GO 80 100
count
Fig. A2 Singular values of G where a , v , and 1 denote attenuation, velocity and line length,
respectively. It is well known [ 3 ] that h m ( a ) can be fitted by a
B - Computational Efficiency of State Equation Realization Based minimumphase shift function h,, (a) plus a time delay, T :
on Vector Fitting
h (a)=h,, (a (E4
Consider the scalar-input / vector-output convolution
From (E.l) and (E.2) it follows that T must be chosen so as to
y=h*u (B.1) satisfjr the relation
whose finite State Equation Realization has been found :
i=Ax+bu (B.2)
y =Cx +du 03.3)
In our implementation, (E.3) is evaluated at a single frequency
point, C2 . Thus, the time delay becomes :
Trapezoidal integration of (A.2) gives :
xn-xn-l = x A+xn-,
L +U,_1
+ b L (E.4)
At 2 2
03.4)
where At is the time step length used in the simulation. Further The phase angle Lh,, (CO),which is uniquely defied by the
manipulation gives * magnitude h c t i o n Ih,; (w )I , is calculated by equations (24)-
(261.
x,=Pn,-, +qlu, +U,-] 1 (B.5) 12 BIOGRAPHIES
y n =CX,+dun (B.6) Bjsrn Gustavsen was bom in 1965 in Harstad, Norway. He received the
In the following we analyze the time needed in the time step loop M.Sc. degree in Electrical Engineering in 1989 and the Dr. Ing. Degree in
for an eigenvector when its elements have been fitted separately, 1993, both &om The Norwegian Institute of Technology. Since 1994 he has
and when they have been forced to share the same set of poles been working at the Norwegian Electric Power Research Institute (EFI),
(vector fitting). It is assumed that N poles are used in both cases, mainly in the field of transient studies. He is currently on leave at the
Department of Electrical and Computer Engineering University of Toronto.
and that the system has n conductors.
Adam Semlyen was born in 1923 in Rumania where he obtained a Dipl. Ing.
Let k, and k , be the computational time for one addition and degree and his Ph.D. He started his career there with an electric power utility
one multiplication, respectively. When vector fitting is used, the and held academic positions at the Polytechnic Institute of Timisoara. In 1969
SER has one input and n outputs, thus he joined the University of Toronto where he is a professor in the Department
of Electrical and Computer Engineering emeritus since 1988. His research
P : ( N x N ) (diagonal) q : ( N x l ) (full) interests include steady state and dynamic analysis as well as computation of
C : ( n x N ) (full) d : ( n x l ) (full) electromagnetictransients in power systems.
 613

Discussion Taku Noda and Akihiro Ametani (Doshisha University,

Kyoto, Japan): We would like to congratulate the authors for
BrandHo Faria, Senior Member, IEEE (CETME, Instituto Superior developing the new fitting approach - vector fitting - for
TBcnico, AV.Rovisco Pais, 1096 Lisboa Codex, Portugal): modal-domain modeling of a transmission line.
The proposed line model is based on the modal decompo-
This paper, together with a companion one [ 11, presents an impor- sition approach. The modal decomposition at a single
tant and lasting value contribution to the field of transmission line
transients computation. The authors are, thus, to be commended by frequency is always possible and is a perfect theory as shown
the excellent work they have done. The few comments and ques- in [SI. But is it guaranteed that a modal-transformation
tions that next follow have the sole purpose of permitting the clarifi- matrix TI represents a physically realizable system? If a
cation of some minor details. system is not physically realizable, its numerical implemen-
tation would be impossible. For example, a non-causal
I ) On stating that Hmand Ypare diagonal matrices, eqs. (8-9), the system cannot be implemented in computer code, because
authors are tacitly assuming that the product matrix ZY (or YZ) is future input and output cannot be obtained.
always diagonalizable. However, unfortunately, that assumption is a The authors have concluded that the present line model is
false one [2-41. What kind of strategy should one then adopt in or-
suitable for cable modeling but not always for overhead line
der to keep using the method proposed by the authors if a non-
diagonalizable situation IS faced at some frequencies? modeling. But it should be noted that the frequency charac-
2) In Section 3.2 it is referred an eigenvector normalization proce- teristics of a modal-transformation matrix highly depend on
dure consisting in fixing one element to unity. It is well known that, impedance formulas used in the calculations. Impedance
for bilaterally symmetric transmission line configurations, certain formulas used for modeling the single-circuit overhead line
propagation modes may have some elements null in the eigenvector of Fig. 13 should be clarified.
columns. Therefore, some caution must be exercised in the process The comments of the authors would be highly appreciated.
of fixing elements to unity.
3) The accuracy of the vector fitting technique developed by the
authors is shown in a series of illustrations throughout the paper. Manuscript received March 3, 1997.
However, the illustrations refer to the use of different orders of ap-
proximation (i.e. different number of poles). It is not clear which
criterion should one adhere to in deciding on the minimum number
of poles sufficient for achieving a "good" fitting. This question may
be decisive for the application of the proposed technique to new
computation cases where the "accurate" solution is not known ab Bjsrn Gustavsen and Adam Semlyen: We wish to
initio, that is, when a reference is not available for comparison. thank the discussers for their valuable comments and
4) In Section 6 the authors technique is applied to overhead line useful contributions. The following are itemized answers
configurations with much less success than the one they have got to the problems raised.
with cable configurations in Section 5. The mathematical reason for
this relative failure (explained in the paper) lies on the presence of
unstable poles. Can a simple explanation for this be provided based Professor Brandgo Faria :
on physical grounds? Ad 1 Non-diagonalizable matrices ZY (or n) are
extremely rare and we never encountered one. Even if at
The authors reply to these comments will be highly appreciated. certain values of the parameters, notably at some
particular frequency, such a situation would occur, a
small disturbance of the frequency will lead to a regular
References: matrix. In the integration process implicit in the
transition from frequency to time domain, any isolated
[ I ] B. Gustavsen and A. Semlyen, "Combined phase and modal irregular point would be filtered out. Therefore, we
domain calculation of transmission line transients based on believe that, apart from special, contrived or
vector fitting," paper PE-346-PWRD-0-01-1997. pathological situations, there should be no concern for
non-diagonalizability .
[2] J. A, BrandHo Faria, Multiconductor Transmission-Line Stvuc-
tures, New York: Wiley & Sons, 1993. Ad 2 We agree with the cautionary note of the discusser
[3] J. A. BrandHo Faria, "Irregular wave propagation characteristics regarding the selection of an element of the vector to be
in underground transmission cables," IEE Proceedings, Vol. normalized to unity. In our implementation, the elements
135, Pt. C,No. 2, March 1988, pp. 131-136. of T were scanned as function of frequency. The element
[4] J. A. Brandgo Faria, "Overhead three-phase transmission lines: with the largest minimum value was normalized to unity.
Nondiagonalizable situations," IEEE Transactions on P WRD, We wish to add that only in the case of transformation
Vol. 3, NO. 4, Oct. 1988, pp. 1348-1355. matrices is an element equal to unity obtainable by
normalization. In most other applications of vector
fitting an element equal to unity is appended to the given
vector. This permits a transitional scaling and smoothing
of the vector, an important step in the state equation
Manuscript received February 25, 1997. approximation process.
614

A d 3 Vector fitting is used to find a rational function Replacing in (a) and (b) the state vector xi with the
approximation of a given function. In deciding the modified vector
minimum number of poles needed for a given accuracy,
a trial and error approach will have to be used. x; = xi --qui (c)

Ad 4 We cannot explain the unstable modal gives

decomposition encountered for overhead lines on
physical grounds. We can only point to the fact that xi’= Px;-l + (Pq + q)uz-l (d)
several eigenvalues for overhead lines lie very close over yi = cx;-l + (Cq $. d)% (4
a wide range of frequencies, while this is usually not the
case for underground cables. It is interesting to note that Equation (d) can be evaluated more efficiently by using a
in figure 12 the increasing “peak” in the characteristic similarity transformation via the matrix
admittance, observed with decreasing thickness for the S = diag(Pq + q ) (9
inner insulation, is due to the two eigenvalues getting
close. Eventually we arrived at a situation where the Since P is diagonal, the final result takes the form
transformation matrix could no longer be fitted with
xi‘= Px;:,+ eui-l (g)
stable poles only.
yi = R$+ Gu, (h)
The robustness, accuracy, and efficiency of vector fitting
makes it attractive for many other applications in where e is a vector of ones and
addition to the one presented in the paper. In a more
general setting, the transfer functions are in general not
R=CS (0
strictly proper but co$d be just proper or even improper G=Cq+d 6)
(order or numerator equal or greater than that of the We next look at the computational effort for updating the
denominator). It turned out to be a simple matter to past history source Rq” in the time step loop. Let
generalize the vector fitting program to be applicable to k,and k, denote the computational effort for one
all these cases. These would often occur in the addition and multiplication, respectively. Assume that
identification of transformers or of external system we have fitted a column of n elements using N poles.
equivalents over a wide band of frequencies.
Dr. Taku Noda and Professor Akihiro Ametani : If (9) and (h) are evaluated separately for each element,
we find the total computational cost to be:
The discussers raise the interesting question whether it is
guaranteed that a modal transformation matrix T will K = 2nN(k, + k,) (k)
represent a physically realizable system. All we can say
is that the successful state equation realizations obtained If (g) and (h) are evaluated taking advantage of all
for T indicate physical realizability. In addition, there is elements having the same state vector we get:
a plausibility argument to support this: since both the
phase and modal domain input-output relations (transfer K = (H + l ) N ( k , -I-k,,,) c: nN(k, + k,) (1)
functions) are causal, the link T between them should
have the same property. The only pertinent difficulty we Comparing (k) with (1) shows that the columnwise
have experienced is that the realization may result realization gives a two-fold increase in computational
unstable in the case of overhead lines. This instability efficiency as compared to element-by-element
cannot be removed by scaling. realization.
For modeling the single circuit overhead line we used
the complex depth method [A] for the ground return [A] A. Deri, G. Tevan, A. Semlyen, and A.
impedance and approximate formulae [11,B] for the Castanheira, “The Complex Ground Return Plane:
internal conductor impedance. A Simplified Model for Homogeneous and Multi-
We would finally take advantage of this opportunity to Layer Earth Return”, IEEE Trans. PAS, vol. 100,
present a more efficient implementation of the time no. 8, August 1981, pp. 3686-93.
domain integration than the one in the paper. The [B] A. Semlyen and A. Deri, “Time Domain Modelling
approach, suggescecl by Dr. Thor Henriksen (EFI), can of Frequency Dependent Three Phase Transmission
be summarized as follows : Line Impedance”, IEEE Trans. PAS, vol. 104, no.
We start with equations (B.5) and (B.6) : 6 , June 1985, pp. 1549-55.

xi = Pxi-l + q(ui + ui-l) (a>

yi = Cxi +dui (b) Manuscript received September 9, 1997.