You are on page 1of 12

IEEE Transactions on Power Delivery. Vol. 8, No.3.

July 1993 1615

A High Frequency Transformer Model for the EMTP


A. Morched (SM) L. M a d (M) J. Ottevangers

Ontario Hydro, Canada

Abstract - A model to simulate the high frequency behaviour of a A substantial number of transformer models have been proposed to
power transformer is presented. This model is based on the date. While it is probably inaccurate to categorize all work done in
frequency characteristics of the transformer admittance matrix high fresuency transformer modelling, it is convenient to identify
between its tcnninals over a given range of frequencies. The two broad trends to describe the model presented in this paper within
transformer admittance characteristics can be obtained from the context of earlier work.
measurements or from detailed internal models based on the physical
layout of the transformer. The elements of the nodal admittance Detailed internal winding models. This type of model consists
matrix are approximated with rational functions consisting of real as of large networks of capacitances and coupled inductances
well as complex conjugate poles and zeroes. These approximations obtained from the discrethtion of distriiuted self and mutual
are nalized in the form of an RLC network in a format suitable for winding inductances and The calculation of
direct use with EMTP. The high frequency transformer model can these parameters involves the solution of complex field problems
be used as a stand-alone linear model or as an add-on module of a and quires informationon the physical layout and construction
more comprehensive model where iron core nonlinearities are details of the transformer. This information is not generally
represented in detail. available as it is considered proprietary by transfonner
manufacturers. These models have the advantage of allowing
Keywords - Transformer, High frequency, Frequency dependence, access to intend points along the winding, making it possible to
Electromagnetictransients, EMTP. assess internal winding stresses. In general, internal winding
models can predict transformer resonances but cannot reproduce
the associated clamping. This makes this class of models suitable
1. INTRODUCTION for the calculation of initial voltage distriiution along a winding
due to impulse excitation, but unsuitable for the calculation of
The transformer is probably one of the most familiar components of transients involving the interaction between the system and the
a power system, but it is also one of the most dficult to model transformer. Furthermore. the size of the matrices involved
accurately. A recent survey comparing EMTP simulations with field (typically 100 x 100 or larger) makes this kind of representation
measurements indica- that studies where transformerbehaviour has impractical for EMTP system studies.
the,greatest influence on the results are those where EMTP
simulations tend to be the least accurate. Terminal models. Models belonging to this class are based on
the simulation of the frequency and/or time domain
To model a transformerin a transient simulation, nonlinearbehaviour characteristics at the terminals of the transformer by means of
as well as frequencydependent effects must be taken into account. complex equivalent circuits or other closed-form
Standard EMTP transformer models such as BCTRAN and representations-. These terminal models have had varying
TRELEG~ can accurately reproduce the response of a transformer at degrees of success in reproducing the frequency behaviour of
the frequency at which the short-circuit and open-circuit Wts are singlephase transformers accurately. The main drawback of the
made; namely, at power frequency. However, theae models do not methods proposed to date appears to be that they are not
account for the frequency dependence of copper and iron losses, or sufficiently general to be applicable to threephase transformers.
the effect of stray capacitances.

The behaviour of the transformer at higher frequencies can be The high-frequency transformer model described here belongs to the
approximated, to some extent, by modelling the distributed stray class of models where the frequency dependent response at the
capacitancesalong the windings with lumped capacitances ~ o ~ e c t e d terminals of the transformer is reproduced by means of equivalent
across the terminals of the transformer. This type of representation networks. Unlike earlier frequencydependent transformer models,
cannot reproduce the behaviour of the transformer beyond the first the new model can simulate any type of multi-phase, multi-winding
resonance frequencies. The calculation of the capacitances is not transformer as long as its frequency characteristics are known either
straightforward, and it is difficult to obtain accurate values, except from measurementsor h m calculationsbased on the physical layout
for simple transformer designs~~. of the transformer. The generation of the parameters for the model
is automatic, and it does not require special skills on the part of the
user. This model has been developed and implemented at Ontario
92 SM 359-0 PWRD A paper recommended and approved Hydro as part of a new and comprehensive transformer model
by the IEEE Transformers Committee of the IEEE Power sponsored by the EMTP Development Coordination Group - DCG
Engineering Society for presentation at the IEEE/PES
1992 Summer Meeting, Seattle, UA, July 12-16, 1992. for the DCG/EPRI version of the EMTP. Although originally
Kanuscript submitted February 3, 1992; made available developedas a high-frequencyrepresentation, thii model can also be
for printing May 1, 1992. used as a stand-alonelinear model, if the frequency characteristicsof
the transformer are known over a sufficiently broad frequency range.
0885-8977/93$03.00 Q 1992 IEEE
1616

2. OVERVIEW The elements of Iyij,J are approximated with rational functions


which contain real as well as complex conjugate poles and zeroes.
Consider a multi-phase, multi-winding transformer. The nodal The rationalfunctions can then be realized with RLC networks which
equations which relate the voltages and currents at the accessible can be combined using (4) and (5) to produce the parameters of the
terminals of the transformer can be expressed as equivalent r-circuit.

where the nodal admittance matrix M is complex, symmetric, and


frequency dependent. In a threephase system, equation (I) can be
expressed as

Fig: 2: Structure of an RLC module.


yu.= yu+ yuw A typical RLC network used in the approximationof the elements of
va = yUb vu# yV& (3) flij,Jis shown in Figure 2. The general structure of these RLC
networks reflects the known frequency characteristics of the
?U- yusb yuw. admittance functions of a transformer:

The basic idea behind the new transfomr model is to produce an


predominantly capacitive behaviour at high frequencies,
equivalentnetwork whose nodal admittancematrix matches the nodal
represented by the single RC branch.
admittance matrix of the original transformer over the frequency
range of interest. Such representation would correctly reproducethe
transient response of the transformer at its terminals. Consider then
3. PRACTICAL CONSIDERATIONS
the multi-phase network shown in Figure 1. This network will be
referred to as a multi-terminal r-equivalent. The parameters of this
The transformer model must be sufficiently robust to produce
circuit can be calculated from its nodal admittance ma& using the
consistentand numerically stable equivalentnetworks even when data
well-known relationships
are obtained from noisy or inconsistent measurements. To this
(4) effect, the following steps are taken:

Fitting the elements ofwd and calculating VE,J by adding the


fitted functions US% (4). instead of fitting WkT1directly.

Averaging the diagonal and off-diagonal elements of wd SO that


Wd become bazancad matrices.
The explicit approximation of the elements flE,J, would result in
models of lower order than those obtained by first appmximating
v$ and then adding the results. However. it has been found that
models obrained by fitting directly. may be numerically
unstable.

Fig. 1: Single-line diagram of a multi-terminal r-equivalent.


If the sub-matrices of M arc balanced, they can be diagonalizad by
a constant transformation [Ql such that 0.80-

m
*
0
d
0.40-

(7)
-
I

M
Y

t
0 . ...
\
0 a . r

Subscriptss and m in (6) stand for "ser and "mutual", respactively; d


subscripts "0" and "1" in (7)stand for the familiar zero and positive Frequency Hz
sequence components. Matrix [Ql could be any of a numbex of
transformation matrices which diagonalize a balanced matrix. Note 1-00
that even though each sub-matrix in M can be diagonalized, M I
itself will not be diagonal. For example, for a thrcuphase, two-
winding transformer

Frequency Hz

To validate the transformermodel, frequency domain as well as time


domain measurementswere conducted on a 125 MVA, 215/44 kV,
threelimbed coretype transformer. The transformer is YY-
Introducing into equation (2) we finally obtain connected with grounded neutrals at the high and low voltage sides.
The transformer has a deltaconnected tertiary winding with no
accessible terminals.

The effect of averaging the elements of Cyd for the transformer


indicated above, is illustrated in Figure 3, where the solid trace
[a-'
v, = VH& ; [a-'I,
= I,& corresponds to the averaged functions, and the dashed tractx
correspond to the raw measurements.
[a-'
'V = -V ; [QpZ' = -z

The admittancesto be approximated with rational functions are now 4. FITTING PROCESS
the elements of Cy,,,,,& instead of that the elements of Cy$. Since
the parameters of the positive and negative sequence networks are The elements of Cy& in (8) are approximated with rational

admittance elements, rather than 3m.(3m+1)/2.


+
identical, the problem reducte to the fitting of only m-(m 1) distinct functions given by
YGr) - Y*(s) = Y=(.s) + Y&) + Y l S ) (10)

Averaging the elements of Cyd to produce balanced matrices has


obvious merits from the point of view of computationalsped. For
example, for a two-winding, threephase transformer, 6 rather than
21 distinct function8 would have to be approximated with RLC
networks. Also, the timastep loop calculationsin the EMTP are also
substantially reduced. Averaging the elements of Cya also adds
some robustness and consistency to the raw measurements, and
contributes further to the numerical stability of the model. While
averaging may, in some instan-, mask the effect of legitimate where k,,, hj,bcand L hCtue
are real constants; ~ R and
asymmetries in the transformer, the differences observed in the real poles; "/o is a real zero; pi and q are complex poles and
transformers studied appear to be relatively small (see Figure 3). ze-, and p; and z'~are their respective complex conjugates. For
1618

the practical example given. the number of real poles NR and the The step8 followed in the approximation of Y(s) are:
number of complex conjupb poles NC m (11) and (13) are typically
6 and 15, rrapoctively. AU polcs arc confined to the left hand side 1) Numerical noise in Y(s) is removed. whose magnitude fall
of the compltx ptaae aad s=jo. YJs) can be described with the below a user-controledpercentageof the largest peak in Y(s) are
equivalent circuit shown m F i i 2. where YRLcomspon&to the dismissed.
RL branches. YRc correspondr, to the RC branch and :Y,
corresponds to the RLC brancher. The single resistive branch comes 2) Initialize YRc. The reaponse of an RC branch is
from & in equation (11). and its conductance is normally very small.

Let us now define fa as the frequency where the first parallel


resonance of Y(8) occurs (seeFigure 4). At frequencies below f-,
the admittance function8 behave as combmation8 of RL branches
without resonances. At frequencies above fm, stray capacitances
come into play and a number of resonancesarc present. Therefore,
for an initial estimate of P&), it b asaumed that the region b e e n The RC branch represents the asymptotic behaviour of the
the first measured data point and fa contains real poles only, m f o n n c r at very high m y . To oalcuku R and C, it k
while the region from f- to the last measured point & contains assumed that for the frequency range of intcmst oRC < < 1 and
complex conjugate pairs only. that the imaginary prut of YRc=oC. The value of C is found by

lo6 U

e 0 . 2 1 ,',(
0.00
102 4 Id2 4
'!k,
l lj/
\ \ I'
,,.,I
I d 2 4 10'2 4
Frequency Hz
I ,,.,;
Id2 4
I
1U'

Frequency Hz

Fi.5: Approximation of^^,^. Solid trace:jirredjuncrion


Dashed trace: raw data

2
1U'

Frequency Hz Frequency Hz

Fig.4: Elrmenr Y ~ J Solid


. trace: raw data; Fig. 6 @proxi" of yu,o. Solid trace:fittedficncrio
Dashed trace: low frequency mudel Dashed trace: raw data.
1619

fitting the imaginary part of Y(s) with OC in the least squares large, no causal rational function will be able to approximate the
sense. With C known, R is found by matching YRc to a point on transition region and still produce a negligible contribution at 60 Hz.
the lower envelope of the d part of Y(s). This technique is In these cases, the approximation in the neighbourhood of f,, will
very simple but surprisingly effective: optimization seldom be somewhat degraded.
changes this initial estimate by more than five percent.
When the response of the low fresuency model is subtracted from the
3) InitializeYw: by identifyingthe local maxima and minima of the measureddata, the region between 60 Hz and f- defines a transition
magnitude of the real part of Y(s). Each local maximum or peak area where the magnitude of Y(s)=Y,,-Y,,,,,, is determined by the
comsponds to a complex conjugate pole pi=q+j&, and each consistencybetween measured data and the low frequency model and
local minimum or valley comeponds to a complex conjugate where the magnitude of Y(s) at 60 Hz should be zero. If the low
zero q=vi+j6p The angular frequency at which a maximum and frequency model were accurate from 60 Hz to & and the
minimum occurs determines Bi and 4 (&=2rLk, 4=2rLby). measurements were error free,then the magnitude of F(s) would be
The real parts are arbitrarily initialized to 2.5% of their very small at L. The smaller the difference the better the fit
compondmg imaginary parts. The real zero v0 is initialid to around f- If the difference is very large, then is not possible to
y0=21-f,. The number of poles and zeroes assigned is approximatethis transition region accurately and some compromises
determined by the shape of Y(s) and by the tolerance that are necessary, namely, a larger error between & and fa and a
determines which peah are considered meaningful. relatively large Y,(s).

3) Optimize the initial guess of Yl(s)=YRc+Ym, assuming A high firesuency model was developed for the measured transformer
YRL=O, over the fresuency range from to LX using a as an add-on module to a TRELEG model of the same transformer.
modified Marquardt algorithm'2. The error function is defined Figures 5 and 6 show the approximation of the elements Ym,o, and
as the magnitude of the difference function Y(s)-Y,(s). YU,,produced by the combined model. These illustrate the best and
the worst fits, respectively, obtained for this particular transformer.
4) Initialize YRL(s)from 60 Hz to f, using the asymptotic fitting
procedure described in the Appendix. This initialization
algorithm does not require that the function to be approximated 6. TRANSIENT RESPONSE
be a minimum p b s h i f t function in order to produce an
accurate initial fit. The default number of poles is four, but this With the elements of [y ] available in a closed form, inclusion of the
number is under user's control. high frequency model in the EMTP is conceptually straightforward.
Each branch in the equivalent network shown in Figure 1 is
5) Optimize the entire function Y,(s) from 60 Hz to f, using the represented by a constant conductance matrix in parallel with a past
same optimization algorithm indicated in item 3 above. history current source. During a transient simulation, modal
voltages and currents are calculated from terminal voltages and the
During the optimization process, all poles are confined to the left uncoupled sequence networks are solved to produce an updated set
hand side of the complex plane. Zeroes are not so constrained. This of modal history current sources. These current sources are
often leads to the realization of branches with negative values of R, tmnsformed into phase quantities and used by the EMTP for the
L and C. Nevertheless, because of the constraints indicated above, solution of the system in the next time step.
these branches still have a positively damped response.
The approximationsgenerated with the techniques described above,
The entire fitting process is fully automatic and no user input or are ultimately combined using equations (1)and (2)to produce the
special skills are necessary to initialize it. Some of the fitting branches of the equivalent sequence representation of the
parameters can be overridden by the user to control the number of transformer. These networks can be simulated in the EMTP by
poles and zeroes of Y,(s) and to control the desirable error levels in means of the new FDB (Frequency Dependent Branch) model. This
the approximations. model was designed as a general-purposetool to simulate multi-phase
coupled RLC networks in the EMTP. The type of networks which
can be modelled with the FDB model are more general than those
5. USAGE AS AN ADD-ON MODULE required by the high frequency transformer model. In fact, even the
EMTP implementation of the FDNE (Frequency dependentNetwork
When the high frequency model is used as an add-on module of a Equivalent) is a sub-set of the FDB model.
more complex representation with linear and nonlinear components,
the nsponse of the low frequency components Yb must be It is not within the scope of this paper to describethe implementation
subtracted from the measured response YmW before it is of the FDB model in the EMTP or Ontario Hydro's experience with
approximated. In a typical application, the frequencyresponseof the the new high frequency transformer model. Due to space limitations
low frequency nameplate model (e.g., BCTRAN or TRELEG), these topics will have to be part of a separate paper.
including iron core losses, is subtracted.
To verify the EMTP implementation and to validate the developed
The subtraction of the effect of the low frequency models introduces model, a comparison between simulated versus measured transients
some complications if the difference between their low frequency on the same 125 MVA, 215/44kV unit used earlier was conducted.
iesponse and the measurements is not negligiile in the transition Figure 7 shows the measured response of the transformer measured
region around the lowest measured frequency f& (see Figure 4). on phase 1 of the high voltage terminals when a step voltage is
Ideally, an add-on high frequency model should have no effect on applied on phase 3 of the high voltage terminals. All other terminals
the response at power frequency. If the difference at f- is too are grounded. Figure 8 shows the results of the corresponding EMTP
1620
~ ~ ~~

Cherryrood 125 HVR 215/4'1 k V . Voltage a t H3


0.90, 1 v o l tages: Scale IO** l * J I
4
8.00--

-b.OO

0 -8.00
0 . 0 0 0.02 0.04 0.06 0.08.O.IO 0 . 1 2 0 . l r 0 . 1 6 0 . 1 8 0 . 2 0
Tine 1 0 n w (-2)
Fig. 7: Step response. Aeld test. Fi. 8: Step response. EMTP simulation.

transient simulation. Numerical stability has been verified for this would like to thank A. Narang from Ontario Hydro's Electrical
and other similar tests by allowing the simulation to run for Research Division for providing the measurements for the
extremely long times (several seconds). transformer used in the numerical examples. Thanlrs are also due to
CEA for permission, on behalf of DCG, to publish this work.
Funding for this project was provided by DCG.
7. CONCLUSIONS
REFERENCES
This paper presents a model to simulate the behaviour of a multi-
phase, multi-winding transformer over a wide frequency range. This [l] J. Skliutas, and J. Panek, "Electromagnetic Transients Program
model reproduces the behaviour of the transformer by means of (FMTP) - Field Test Comparisons". EPRI EL-6768, March 1990.
combmations of RLC nemorks that match the frequency response of
[2] H. W. Dommel, Electromagnetic Transients Program Reference
the transformer at its terminals. The fresuency response of the Manual THEORY BOOK), Printed by The University of
~~

transformer is assumed to be known from measurements, or from British Columbia, Vancouver B.C.,Canada, August 1986,pp. 6-
calculations with models based on geometry and construction details. 62 - 6-63.
Its most,important featurea are:
[3] R. C Degeneff, "A Method for Calculating Terminal Models of
Single Phase n-winding transformers". Paper No. A 78 539-9
1) It can be used to model multi-winding, multi-phase transformers
for which the frequency response is known. presented at the IEEE PES Summer meeting in Los Angeles, July
1978.
2) It can be used as an add-on module for a more complex [4] T. Adielson, A. Carlson, H. B. Matgolis, and J. A. Hallady,
transformer repmentation. It can also be used as a stand-alone "Resonant Overvoltages in EHV Transformers - Modelling and
linear model if the frequency response of the transformer is Application",IEEE Transactionson Power ApparatmmdSystmr,
known over a sufficiently wide fresuency range. vol. PAS-100,pp. 3563-3572,July 1981.
[5] P.I. Fergerstadand T. Henriksen, 'Inductancesfor the Calculation
3) The fitting techniquee developed to approximate the admittance of Transient Oscillations in Transformers", ZEEE Transactions on
functions of the transformer produce approximations of Power Apparatus and Systems, vol. PAS-93,No. 2,pp. 500-509,
exceptional quality. MarcWApril 1974.
4) tests performed indicate that the models produced a x L61 R. c*Degeneff, "A Method for m i n i n g Resonances
in Transformer Windings", IEEE TransactionsonPower Apparatus
accurate and numerically stable.
and Systems, vol. PAS-%, No., pp. 423-430,MarcWApril 1977.
5 ) The process to generate parameters for the model is completely [7] R. C. Degeneff, W.S. McNult, W.Neugebauer, J. Pan&, M.E.
automatic: no special skills or experience are required from the McCallum, and C. C. Honey, "Transformer Response to System
user. Switching Voltages" ,IEEE Transactionson Power Apparatus and
Systems, vol. PAS,No. 6,pp. 1457-1470,June 1982.

AcknowIedgements [8] P. T. M.Vaessen, "Transformer Model for High Frequencies',


IEEE Transactions on Power Delivery, vol. 3, No. 4, pp. 1761-
The authors would like to acknowledge the use of the Marquardt 1768,October 1988.
optimization routine from the Harwell Subroutine Library. Also we

- _
1621

[9] Q. Su, R. E. Jamcs, and D. Sutanto, "A ZTransform Model of


Transformersfor the Study of ElccttomagneticTransientsin Power
Systems",IEEE TransactionsonPower Systems. vol. 5 , No. 1, pp.
27-33, Febr~ary1990.
[lo] A. Keyhani, H. Tesai, and A. Abur, "Maximum Lkelyhood
Estiimation of High FrequencyMachine and TransformerWinding
Parameters", IEEE Transactionson Power Systems, vol. 5 , No. 1,
pp. 212-219, January 1990.
1111 A. Keyhani, S. Cbua, and S. &bo, "Maximum Likdyhood
W i i o n of Transformer High Frequency Parameters from Test h3
Data", IEEE Transactionson Power Drlivery, vol. 6, No. 2, pp.
858-865, April 1991. \/;.I.'
1121 D. W.Marquardt, "AnAlgorithm for --Square Estimation of h4
Nonlinear Parameters", J. Soc. Indust. Appl. Math, vol. 11, No.
2, pp. 431-441, June 1963. frequency
[13] J. R. Marti, "Accurate Modelling of Frequency-Dependent 1) Calculate R(s), where
Transmission Lines in Electromagnetic Transient Calculations". Rcs) = 4WHOI - c
IEEE TransactionsonPowerApparatusandSystems,pp. 147-157,
January 1982. where C is an arbitrary constant such that Re{H(s)} > 0Vo 2 0
1141 L. Ma", "Low-order approximation of Transmission Line
Parametersfor Frequency-DependentModels". IEEE Transactions 2) Use the fitting technique described above to approximate R(s) with
on PowerApparatusandSystems, pp. 3584-3589,November1983. R(S).

3) Calculate the partial fraction expansion of the rational function


R'(s)
APPENDIX
R'(s) = ke + *-
4
To approximate a minimum phaseshift function H(s) with a rational
1-1 b - PJ
function P(s) that contains only real poles and zeroes which lie in the 4) The approximation of H(s) is then given by
left hand side of the complex plane, it is sufficient to match the
magnitude functions of H(s) and P(s). This is possible because the H(s) .
I A', + -
kJP,
phaseangle of a minimum phaseshift function is uniquely determined 1-1 (s - PJ
by its magnitude function: if IH(s) I and IP(s) I match, their phase
angles will also match. where kb = k, +C
H(s) P(s) = n-
ko
- PJ
i-1
(s-zJ
(s BIOGRAPHIES
A very effective technique to match the magnitude of a minimum
phaseshift function is suggested in [13] and [14]: Atef S. Morched (M'77SM'90) received a BSc. in Electrical
Engineering from Cairo University in 1964, a Ph.D. and a D.Sc.
1) Subdivide the magnitude function into N equally-spaced from the Norwegian Institute of Technology in Trodheim in 1970 and
segments. These segments define the location of the horizontal 1972. He has been with Ontario Hydro since 1975 where he
asymptotes hi, (i=l, ...,N+1) of IP(s)l. currently holds the position of Section Head - Electromagnetic
Transients in the Power System Planning Division.
2) Place the corresponding vertical asymptotes at the frequency
where IH(s)l equals the geometric mean of two adjacent Luis M a d (M'79) received an undergraduate degree in Electrical
horizontal asymptotes. Engineering from the Central University of Venezuela in 1979,
MA% and PhD degrees in Electrical Engineering in 1983 and 1987,
3) The initial location of poles and zeroes is d e h e d by the respectively, from The University of British Columbia. He did post-
intersection of vertical and horizontal asymptotes. doctoral work in cable modelling in 1987-1988, and joined Ontario
Hydro in 1989, where he is currently working in the Analytical
4) O p t i m i i the initial location of the poles and z e m by Methods & Specialized Studies Department of the Power System
reducing the error function in the least squares sense. Planning Division.

Jan H. Ottevaneers received an MSc. in Electrical Engineering from


Thii technique can be extended to approximate any analytical, non- the Delft Institute of Technology in 1956. He has been with Ontario
minimum phaseshift function. The basic premise is that the Hydro since 1967 where he is currently working in the Analytical
imaginary part of an analytical function is uniquely determined by its Methods and Specialized Studies Department of the Power System
real part. The modified method proceeds as follows: Planning Division.
1622

Discussion different parameters than the other two phases.) Then, as a result of the
balancing, a real, constant, transformation [e]
can be used to obtain
Q. Su (Monash University, Clayton, Australia): The authors are to be the desired modes. After fitting, the modal approximations are
congratulated in having presented a comprehensive high frequency transformed back to the original phase domain. They now correspond
transformer model of equivalent networks. In electrical power systems, to the given matrix [U of frequency domain measurements. My first
the transient overvoltages of high voltage power transformers, either question is whether a final refinement of the fitted results is performed
at the terminals or inside the windings, are of great importance for the or contemplated for obtaining a best match with the original set of
reliability of electricity supply. The model developed by the authors data, in order to compensate for the approximation made by the initial
will be useful for the study of system transients in which the high balancing process?
frequency characteristics of transformer winding are to be considered.
Obviously, the model represented by a number of R, L. C compo- The transformer connection used in the paper is Y-Y, with the
nents can easily fit in EMTP programs. For a detailed internal winding particular feature that the [Tj]block can in fact be balanced by a small
model, several hundred components may be used resulting in a large adjustment. This is so because the connection does not produce an
size of matrix impractical for EMTP system studies, as mentioned in internal phase shift. When this is not the case, for instance in the
the paper. The authors RLC module in Figure 2 consists of at least 1.5 important class of Y-A or Y-Z (zig-zag) connected transformers, the
components and 12 such modules are used to represent a three-phase, offdiagonal block [qj] has cyclic symmetry. For instance, in one
two-winding transformer. It is therefore necessary to simulate each
transformer of interest in power system with a network of 180 or more particular Y-A connection (with an admittance y associated to each
RLC components. Would this be a problem with EMTP system phase of the Y-connected winding), we have
studies?
In my previous papers [1,2], a closed-form transformer high fre-
quency model was presented, as shown in Figure A(a). Extended to
the mode form in Figure A(b), the model has also been used for three
phase transformers. From my experience, the computing time for
system transient studies increases significantly for a transformer repre-
sented by RLC networks rather than closed-form models.
Another question concerns the higher frequency response of a Eigenanalysis of this matrix leads to the symmetrical component
transformer under step voltages. The functions in Figures 5 and 6 fit transformation matrix with three, rather two, decoupled modes.
measured data up to about 200 kHz and the calculated step voltage Balancing would yield the zero matrix. An a-type input gives a &type
response in Figure 8 agrees with the measured in Figure 7. This output and vice versa (as expected, see for instance PI; thus the (real)
confirms the fitting accuracy of the authors method. Could the au-
thors indicate the rise time of the step voltage and the time step Clarke transformation does not result in modal decoupling). Positive or
interval used for the calculations of the step voltage responses? negative sequence voltages result in currents of the same sequence with
the expected phase rotation. In the modal domain there is of course no
References strict symmetry, as reciprocity now implies a rotation in the opposite
direction if the voltages are applied to the secondary rather than to the
Q. Su, R. E. James and D. Sutanto, A Z-transform Model of primary winding.
Transformers for the Study of Electromagnetic Transients in
Power Systems, CO-authored with R. E. James and D. Sutanto, Clearly, transformers with internal phase shifting effects pose more
IEEE Transactions on Power System, No. 1, Vol. 5 , 1990, pp. complex problems. Could the authors please elaborate on their thoughts
27733. regarding the solution of these problems?
Q. Su and T. Blackburn, Application of Z-Transform Method Finally, I wish to reassert my appreciation regarding the merits of this
for Study of Lightning Protection in Electrical Power Systems,
Proceedings of the 7th Intemntional Symposium on High Voltage paper and would like to congratulate the authors for their fine
Engineenng, Dresden, Germany, Aug. 26-30, 1991, pp. 139-142. contribution.
[A]A.S. Morched, J.H. Ottevangers, and L. Marti, Multi-Port
Frequency Dependent Network Equivalents for the E M , IEEE
paper no. 92 SM 4614 PWRD, presented at the 1992 IEEEPES
Summer Meeting, in Seattle, WA.
[B] Edith Clarke, Circuit Analysis of A-C Power Systems, Volume I:
Symmetrical and Related Components, John Wiley & Sons, Inc.,
New York, 1943.

M a n u s c r i p t r e c e i v e d J u l y 2 7 , 1992.
( a ) ( b )
Fig. A A close-formhigh lrequency model for two-winding, (a) single-phaseand
(b) Ihree-phase Iranslormers.

X. Chen (Department of Electrical Engineering, Seattle University,


Seattle, WA): This paper is very impressive in scope and in detail. The
Adam Semlyen (University of Toronto): This is an interesting and authors and their organization must be commended for their contribu-
useful paper as it solves the problem of providing a realistic model for tions to the accurate modeling of the high frequency behavior of
multi-phase transformers for the purpose of EMTP simulations. It has multi-winding, multi-phase transformers. The fitting techniques to
approximate the admittance functions of a transformer is both novel
benefited of the authors expertise and experience in fitting stable and practical. This discussor has learned a lot from their paper and
circuit models to frequency domain data [A]. One of its outstanding the authors earlier papers on transformer modeling. I would appreci-
features is the modal decomposition they have used: it has not only ate the authors comments on the following questions:
simplified the problem of fitting but, more importantly, it has reduced
the dynamic size of the model to that of a minimal realization. The (a) I have developed a computer program which can form the induc-
following remarks and questions are mainly related to the problem of tance matrix for a two-winding, three-phase, multi-legged trans-
modal decomposition. former.
The inductance matrix for the primary winding of an unsatur-
I note that the authors assume that the offdiagonal block [qj]can be ated three-phase three-legged transformer computed by BC-
adjusted to become a balanced matrix. This, of course, implies an TRAN (pages XIX-C-15 to 20, ATP Rule Book, 1987-1992,
approximation. (The central phase may, for instance, have somewhat BPA) is shown in Eqn. (A).
1623

: :1 2 I: i
La, L,, Lo,

=
879.72
-438.02
-438.02
-438.02
879.72
-438.02
-483.02
-438.02
879.12 I
The inductance matrix computed by my program is shown in Eqn.
Henry (A)
R. Malewski (Westmount, Quebec, Canada): This study can serve as an
excellent example of successful and realistic approach to modeling of a
large HV power transformer complex internal circuit. The authors
recognize a necessity of taking measurements of the examined trans-
former characteristics in order to develop the EMTP model. As an
alternative, they refer to the transformer design parameters; these
(B). however, are considered proprietary by the manufacturer and not

;:1 2: 2.j i
accessible to the utility engineers.

1
La, Lab Lac 879.80 -584.68 -292.14 The paper title includes the mention of high frequency, and at the
-584.68 1172.06 -584.68 Henry (B) end of paragraph #2 a statement is made on the predominantly
= -292.14 -584.68 879.80 capacitive behavior of the winding at high frequencies. This is correct
if the f,, is set at some 200 kHz, as indicated in Fig. 1. After all, it
It is striking to note that Lo, is two times greater than L,,, and
L,, is 1.33 times greater than La, and Lee. Because of the T r a n e f e r Cunstlon. dlec to dlec. svmry sscond d1.c
asymmetry of the iron core of a three-legged, core-type trans-
former, my work is very possibly correct. If this is the case, then
Eqs. (6) to (8) of the authors' paper might not be valid. It is
common practice to represent a transformer by its sequence
impedances for short circuit analysis. To apply the symmetrical
components method to an unloaded three-phase core-type trans-
former is not always valid, even if there is no saturation involved.
(b) Figures 7 and 8 of the paper showed the comparison between the
computed and measured step response of a 125 MVA, 215/44
kV transformer. The applied step voltage is much lower than the
rated voltage of the high voltage terminals. The main objective of
.
I
fn

5 0.2
developing high frequency transformer models is to study trans- >
0

former overvoltages caused by switching and lightning. Although


many researchers claimed that magnetic saturation of the core
0.1
has minor influence on fast transients, and therefore can be
disregarded, this discussor is interested in knowing if the authors
have compared the results of their model to the field test or
measurements for overvoltages caused by a lightning surge on a 0.0
2 5 . OkHz 5 0 . OhHr 7 5 . OhHz LOO. OkHz
transmission line which is connected to a transformer and operat-
ing at rated voltage. This discussor has a strong opinion that
harmonic analysis is valid for linear and slightly nonlinear sys-
tems. Wherever severe nonlinearity is involved, differential equa-
tions should be used and nothing else.
Again, the authors are to be congratulated for their effort in develop-
ing a comprehensive transformer model.
Manuscript received August 7, 1992.

I
;; 0.2
H. M. Beides and A. P. Sakis Meliopoulos (Georgia c
_I
z
0
Institute of Technology). The authors should be commended
for revisiting the problem of power transformer modeling.
0. 1
As it is widely known, a comprehensive and generally
acceptable transformer model for transient simulation does
not exist. One of the reasons is that transformers come in
different designs and configurations and with different 0.0
1 0 0 . OhHz 2 0 0 . OhHz
parameters of parasitic capacitances, etc. We would FRED

appreciate the authors response to the following comments


and questions:

Has the proposed modeling method been tested using


transformers with tertiary windings? If yes, the authors'
comments on the accuracy and performance of the derived
models will be appreciated. What are the effects of U 0.1
hysteresis losses and skin effect on the accuracy of the .
P

estimated resistive components of the transformer model?

The method requires measuring the frequency response of 0.05

the transformers. Is the frequency response dependent on


the design of the transformer alone (i.e. two transformers of
the same manufacturer and type will have identical 5 0 0 . OhHi 1 . OUHz I . SMHr 2. OYHz 2.5YHz
frequency response)? If this is not the case, it appears to us FRED

that it will be necessary to measure the frequency response Fig. 1. Transfer function of five first discs of an interleaved HV transformer
of each transformer to be modeled. winding presented at three frequency scales. These transfer functions were
deconvoluted in frequency domain from transients recorded on the un-
Manuscript received August 11, 1992. tanked winding.
1624
has been known since the time of Wagner [l] that the standing wave practice may not be necessary and the phase data of the measured
type of resonant frequencies is confined to a few hundred kilohertz transformer admittances should be used for the transfer function
interval. estimation, because by including the phase data in the estimation
A distinctly different behavior of typical HV transformer windings process does not increase the number of unknown parameters which
starts some 2 to 5 MHz, where the internal disc resonances come to determines the size of the estimation problem. Furthermore it adds an
play. It may be of interest to inspect a typical winding transfer important constraint on the variation of the estimated parameters.
function spanning all the four frequency intervals. Such a graph was Another important aspect of the high frequency transfer function
obtained from an impulse voltage distribution measured along the estimation is the numerical stiffness problem. This problem becomes
discs on an untanked medium voltage unit. The transfer function was more severe if the resonant points are spreaded in a wide frequency
deconvoluted from the digitally recorded transients and the applied range. In general, this problem can be resolved if a proper scaling
(low voltage) impulse. First five disc characteristic is shown in Fig. 1 at scheme is adopted during the curve fitting. It would be interesting to
threefrequency scales: 125 kHz, 300 kHz and 2.8 MHz. It can be seen know if any frequency scaling was performed or needed for this
that the interval from some 500 kHz to nearly 2 MHz can be modeled particular study.
by a real pole circuit, but beyond that limit a different representation The authors have provided the power industry with a valuable and
is required. practical technique for modeling the transformer high frequency dy-
Clearly, this paper does not address the issue of very high frequency namics for the EMTP. We would appreciate the authors comments
phenomena, although they are of practical importance for transform- concerning the questions and issues raised in this discussion.
ers directly connected to SF6 insulated bus bars [2]. Manuscript received October 16, 1992
Practical implementation of the EMTP model presented in this
paper calls for measurements of the transformer transfer function in
the frequency range of at least 200 kHz.Such measurements can not A. S. Morched, L. Marti, and J. Ottevangers: We would like to thank
be easily taken on a large unit in substation, but the required mea- the discussers for their interest and their many relevant questions
sured characteristics can be obtained from an industrial laboratory presented.
performing the acceptance test of new transformers. At present, many Regarding Dr. Sus questions, we would like to make the following
laboratories use a digital recorder for monitoring the impulse test comments: After the admittance functions are approximated with
[3,4]. The obtained records are usually processed in order to enhance rational functions, they become closed-form representations of the
the efficiency of fault detection. The processing often includes calcula- original functions. Expressing the admittance functions in terms of
tion of the frequency spectrum of the output and input impulses, and
finding the transformer transfer function as quotient of these two RLC modules is a convenient form of visualization and it does not
imply that a number of RLC branches have to be connected explicitly
spectra. in the EMTP. Inside the EM, the fitted functions are modeled with
An analysis of the transfer function required for the dielectric fault
detection, is not pertinent to the study presented by the authors. FDB modules. Each n-phase FDB module consists of a constant
However, at a reduced voltage level, additional records can be taken conductance matrix and a set of past history current sources. Figure I
during the impulse test, if requested by the utility purchasing the illustrates the EMTP representation of a two-winding transformer
transformer. Such additional measurements can be included in the test using FDB modules.
program, on demand of the utility system planning department. An Therefore, the presence of a high frequency transformer (HFT) in
the EMTP does not increase the size of the nodal admittance matrix
incremental cost of the additional measurement is negligible, since the of the system modeled, and it only adds n entries to the EMTP branch
impulse generator and recording system are anyhow prepared for the
acceptance test. tables for each n-phase FDB module. For example, the transformer
The algorithms for measuring the HV to LV transfer function, and shown in Figure I only adds 9 branches to the EMTP branch tables.
for retrieving the parameters required for modeling can be imple- Additional storage is needed to keep track of the updating of the past
mented on existing commercial digital impulse recorders, or a special- history current sources. In broad terms, this additional storage amounts
ized recording and signal processing system can be developed using to 2 cells for each complex conjugate pole in (13) and one cell for each
the accumulated experience in high frequency measurement of trans- real pole in (11) and (12).
former winding characteristics. The computational burden of Dr. Sus transformer model should be
comparable with that of the HFT model. For a single-phase, two-wind-
ing transformer Dr. Sus model requires the approximation of three
References functions and four numerical convolutions per time step of a transient
Wagner, K. Das Eindringen einer elektromagnetischen Welle in solution. The HFT model, also requires the approximation of three
eine Spule mit Windungkapazitat, Elektrotechnik und Maschi- distinct functions, but only three numerical convolutions per time step
are needed. A comparison of the performance of both representations
nenbau, 1915, p. 89. will depend largely on the number of terms needed in the fitting
Muller, W. Fast Transients in Transformers, CIGRE SC12, process.
WG12.11 Report presented at the Transformer Colloquim in The time step used in the EMTP simulation of the step response
Graz, 1990.
Malewski, R., Poulin, B., Impulse Testing of Power Transformers shown in Figures 7 and 8 of the paper, was the sampling rate used in
the field measurement, i.e., 0.5 ys. An EMTP step function was used
using the Transfer Function Method, IEEE Trans. Vol. PWRD-3, in the simulation ( A t rise time). In the field test, the input step
1988, p. 476.
Malewski, R., Gockenbach, E., Maier, R., Fellmann, K. H., Claudi, reached 95% of its peak value in 1 ys.
We agree with Messrs Keyhani and Tsai when they indicate that a
A., Five Years of Monitoring the Impulse Test of Power Trans- non-minimum phase shift function cannot be described uniquely with
formers with Digital Recorders and the Transfer Function the magnitude function alone. However, any causal function is uniquely
Method, CIGRE Paper 12-201, 1992. defined if its real part is known. In the identification/optimization
Manuscript received August 21, 1992 process described in the paper, both real and imaginary parts are used
to compensate for possible inconsistencies in measured data. Fre-

A. Keyhani and T. Tsai (The Ohio State University, Electrical Engr.,


Columbus, OH): We would like to commend the authors for a well-
written paper and for their efforts to develop a practical high fre-
quency transformer model for the EMTP.
The essential ingredient of high frequency transformer modeling is
to represent the transformer admittances as frequency dependent
nonlinear functions. In general, these transfer functions are non-
minimum phase system. Therefore both magnitude and phase of the
transfer function are needed to uniquely identify the transfer function
model. In this paper, since only the magnitude data were used for the
transfer function identification, the transfer function model had to be Fig. I. Single-line diagram representation of a two-winding transformer with
modified into a minimum-phase plant. It is our belief that such the FDB model.
1625

quency scaling is commonly used in the solution of least squares


optimization problems. However, its use would not be advantageous
within the context of the modified Marquardt optimization algorithm.
Messrs Beides and Meliopoulos ask whether the frequency response
of a transformer depends on its design alone. We have observed that
transformers of the same design and make show essentially the same
frequency behaviour. It is unclear to us, at this point in time, how far
to generalize these observations. In absence of actual measurements, it
might be better to use the HFT model of a similar transformer than to
use no high frequency model at all. While the transformer used in the
paper has a buried delta winding, we have not yet modeled a delta-
connected tertiary winding explicitly. This requires some special con-
siderations which will be explained in more detail in our response to
Prof. Semlyen's questions. Hysteresis and eddy current effects are
normally taken into account by dedicated models (e.g., [il) when the 1 U'
HFT is used as an add-on module. In this case, the HFT model will d
match the difference between the measured data and the frequency Frequency Hz
response of the linear portions of these models. Fig. 11. Magnitude of ylrIh1.
We will now address Dr. Malewski's comments. The choice of 200
kHz as the maximum frequency was based on the simulation needs of
the transient simulations for which the transformer model was re-
quired. Figure I1 shows the magnitude of yhlh from 400 Hz to 1 MHz.
Other than additional poles and zeros and the added computational
burden, we do not feel that the extended frequency range presents a
problem that the HFT model cannot handle. The graphs shown by Dr.
Malewski also suggest to us that frequency range beyond 1 MHz does
not pose any special problems either.
With regard to the techniques used to obtain the frequency re-
sponses, we feel that direct, low voltage frequency domain admittance
measurements are probably simpler, cheaper and more reliable for the
purposes of the HFT model. This type of measurements can be made
with relative ease in the field. It would probably be difficult to
persuade manufacturers to perform all the full scale chopped-wave
tests required to obtain all the data required by the HFT model.
Nevertheless, Dr. Malewski's measurement techniques could provide
an alternative way to obtain data for the H F I model since some of
them are normally done in acceptance tests anyway.
Fig. 111. Off-diagonal elements of [ Y H ] .
Professor Semlyen suggests an adjustment of the final fitted func-
tions to account for phase asymmetries. It is not clear to us how this
adjustment could be made after the fitted functions are obtained. It
should be possible, however, to choose a real constant transformation
matrix other than a , p, o to account for unbalances. This would be differ only by a constant. For example,
roughly the same type of approximation used to model frequency
dependent unbalanced lines. Whether this constant transformation (iii)
matrix would also give acceptable answers at higher frequencies is
probably a subject of further research. Another possibility is to ap-
proximate each element of the Y matrix. This would not rely on any
assumptions of symmetry, but the additional computational burden
would be substantial.
Professor Semlyen correctly points out that in the case of Y-D or
[ Y ~ =~ Y,(~ w~) .iJ5
~ I0
[: 1
I!-
In this case, only one admittance function yo( o)has to be approxi-
(iv)

Y-Z-connections, the off-diagonal [E: ] sub-matrices do not lend them- mated. In the time-step loop of the EMTP the complex algebra does
selves to be approximated by a baianced matrix. Depending on the not present a problem because even if intermediate functions are
type of delta connection and node numbering scheme, variations of a nominally complex, the final phase voltages and currents are always
cyclic matrix can be obtained. For instance, real. In other words, the existing FDB model can easily be modified to
account for cyclic symmetric modules.
Professor Chen correctly points out that a multi-legged transformer
should show some asymmetry, which would degrade the accuracy of
the assumption that the sub-matrices of [Y] are balanced. However,
the measurements available to us do not show the severe asymmetry
indicated in Prof. Chen's calculations. Figure 3 in the paper shows that
the diagonal elements of the high voltage winding block are nearly
identical over a wide frequency range. Figure 111 below, shows the
measured off-diagonal elements of the same sub-matrix.
From this plot it can be seen that while one element is indeed
different, the unbalance ratio at low frequencies is in the order of 1.2
There are several ways in which this situation can be handled: The to 1.3 rather, 2.0 as Professor Chen's calculations suggest. Based on
most obvious one is to fall back on the approximation of every these measurements we are inclined to accept the balancing procedure
element of [Y]. On the other hand, it might be more practical to use a as a reasonable simplification. It is clear, however, that the use of
constant transformation matrix [Q] whose elements qi,kfor a n-phase modal transformation matrix that accounts for center phase asymme-
system are given by tries would be desirable. Strictly speaking, this transformation matrix
would also be frequency dependent. Therefore, further investigation
would be needed to find what constant transformation matrix would
represent an acceptable compromise over the entire frequency range
of interest.
The well-known symmetrical components transformation matrix is just The question of the validity of superimposing linear high frequency
a special case of the matrix defined by equation (ii). The resulting behaviour on the nonlinear response due to saturation does not have a
modal admittance matrix only has two non-zero elements, and these simple answer. Short of solving the nonlinear field problem with
1626
detailed knowledge of core and winding design, accounting for nonlin- requirements. This may contribute to the lack of field measurements
ear and frequency dependent effects will always involve a certain that would validate the assumption of superposition under near-
degree of approximation. If the transformer is unsaturated, high saturation conditions.
frequency excitation cannot drive the transformer into saturation as Reference
the flux produced by a voltage input is inversely proportional to its
frequency. If a transient of sufficient magnitude is impressed on a [i] E. Tarasiewicz, A, S. Morched, A. Narang and E. P. Dick,
transformer which is already saturated or near saturation, then super- Frequency Dependent Eddy Current Models for Nonlinear Iron
position is not strictly valid. On the other hand, situations where a Cores, Paper No. 92 WM 177-6 PWRS, Presented at the IEEE-
transient is impressed on a transformer which is already in saturation PES Winter Meeting, New York, Feb. 1992.
may not be all that common. It is Ontario Hydros practice not no
operate transformers near saturation because of acoustic pollution Manuscript received October 16, 1992.

You might also like