20

Frequency Domain Aspects of Electromagnetic

Transient Analysis of Power Systems.
J. L. Naredo, Senior Member, IEEE, J. Mahseredjian, Senior Member, IEEE, Ilhan Kocar, Member,
IEEE, J. A. GutirrezRobles, Member, IEEE, J. A. MartinezVelasco, Member, IEEE.

Abstract-- Frequency domain (FD) methods have become a x p (t) Periodic signal.
valuable complement to the time domain (TD) ones for the X Vector representation of a periodic signal spectrum.
analysis of electromagnetic transients in power systems. Several X() Input signal, continuous frequency spectrum.
aspects of both Frequency Domain Analysis and Digital Signal H() Frequency response of LTI system.
Processing disciplines, in addition, have become essential for the
analysis of modern power systems. In this chapter, a brief review Y() Output signal, continuous frequency spectrum.
of basic concepts of FD methods is first presented. Then the basic (tt 0 ) Impulse function acting at t=t 0 .
differences between continuoustime and discretetime FD t (t) Train of pulses at intervals t.
Analysis are examined. Next, an overview of transient analysis u(t) Unit step function.
methods based on both, Discrete Fourier Transform (DFT) and N Number of samples.
Numerical Laplace Transform (NLT), are provided along with t Discretization time step.
application examples. Finally, new FD issues related to multirate
Discretization frequency step.
transient analysis are reviewed.
c Damping coefficient.
Index Terms-- Aliasing, Discrete Fourier Transform, X L +(s), X(s) Onesided Laplace Transform of x(t).
Electromagnetic Transients, Fourier Series, Frequency Domain M Cutoff frequency.
analysis, Gibbs Phenomena, Multi rate analysis, Numerical g(t) Ideal interpolator.
Laplace Transform, Phasor Analysis. h R (t) Impulse response of rectangular window.
L () Lanczos window frequency response.
I. NOMENCLATURE
VH () Von Hann window frequency response.
EMT Electromagnetic Transient. rel Relative aliasing error.
LTI Linear TimeInvariant. WN exp(2/N).
TD Time Domain.
FD Frequency Domain. II. INTRODUCTION
DTFT
DFT
FFT
DiscreteTime Fourier Transform.
Discrete Fourier Transform.
Fast Fourier Transform.
T HE electromagnetic transient (EMT) response of a power
system can be determined either by time domain (TD) or
by frequency domain (FD) methods. Common belief, in the
NLT Numerical Laplace Transform. 1980s, was that these two approaches were competing and
LPF Low Pass Filter. that, in the end, only one of these would prevail. Instead,
r.h.s Righthand side of equation. nowadays, TD and FD methods complement each other.
l.h.s. Lefthand side of equation. Devices whose parameters depend on frequency are treated
Angular frequency variable. more conveniently in the frequency domain, whereas those
0 Angular frequency value. elements exhibiting strong nonlinear behavior are better to
x 0 (t) Pure sinusoidal signal. analyze by time domain techniques.
X0 Phasor representation of sinusoidal signal. In practice TDbased methods, like the EMTP, are the
x(t) Time domain signal. most used. These methods are much more intuitive than the
ones based on FD analysis; they also usually require much less
computer resources than the latter ones. On the other hand,
J. L Naredo gratefully acknowledges support from The Mexican Science
and Technology Council (CONACYT) for sabbatical leave, and through however, deep knowledge of FD techniques has become
project 25966. essential for modern power system analysis. Often the
J. L Naredo holds a permanent position at Cinvestav Guadalajara, synthesis of models and of network equivalents is conducted in
Mexico; he currently is a visiting researcher at The Ecole Polytechnique de the frequency domain. In addition, time domain analysis by
Montreal, QC, Canada. (e-mail: jlnaredo@gdl.cinvestav.mx).
Jean Mahseredjian is with the Department of Electrical Engineering, digital computer requires the sampling of all the time
Ecole Polytechnique de Montreal, QC, Canada. (e-mail: jeanm@polymtl.ca) dependent variables. When this sampling is not done properly,
Ilhan Kocar is CYME International, St-Bruno, QC, Canada. (e-mail: it may produce erroneous results. At this respect, FD analysis
ilhan.kocar@polymtl.ca). provides valuable references to check TD results. Sideeffects
J. A. GutirrezRobles is with The Department of Mathematics, CUCEI,
Univ. de Guadalajara, Mexico. (e-mail: alberto.gutierrez@cucei.udg.mx) of sampling processes also are better understood and handled
J. A. Martinez-Velasco is with. Departament d'Enginyeria Elctrica of the in the frequency domain.
Universitat Politcnica de Catalunya, Spain (e-mail: martinez@ee.upc.edu)
 21

This chapter deals with those aspects of Frequency Domain

Analysis and of Digital Signal Processing that have become Linear
essential for the analysis of transients in modern power x(t) Time-Invariant y(t)
systems. The first section of the chapter provides a brief System
review of basic concepts of FD methods. Continuoustime
Fourier Analysis is introduced as an extension of Phasor Fig. 1. Linear timeinvariant (LTI) system
Analysis which is more familiar to power engineers. The
second section of the chapter presents the basic differences x0 (t ) = A
2
e j ( 0t + ) + A2 e j ( 0t + ) (4)
between continuoustime and discretetime Fourier Analysis. Figures 2a and 2b provide illustrations for expressions (3) and
Of special interest here are: 1) The effect of aliasing, 2) The (4), respectively. Whereas (3) leads to phasor representations
Sampling Theorem and 3) The principle of Conservation of of sinusoidal waves, (4) conducts to Fouriertype
Information. The third section provides a brief overview of representations for signals that not necessarily are sinusoidal.
transient analysis methods based on the Discrete Fourier Recall that the phasor representation of x 0 (t) in (1) is by the
(DFT) and the Numerical Laplace (NLT) transforms. The complex exponential function at the r.h.s. of (3) with its factor
fourth section of the chapter deals with issues related to multi
rate transient analysis. e j 0t being removed (see Fig. 2a):
Power systems are increasing substantially in both, size and x0 (t ) X 0 = Ae j (5)
complexity. Even though transient events usually occur
locally, one often needs to analyze their effects on a large The underlaying assumption in Phasor Analysis is that systems
network. Under these circumstances some parts of the network operate in steady state. That is, signal x 0 (t) has always been an
will be subjected to fast dynamics, while others will continue input to Fig. 1 system. Since frequency 0 remains constant,
operating at a slow dynamics, and even close to steady state. It there is no need to refer explicitly the factor e j 0t in phasor
is, therefore, highly attractive to simulate the various parts of a representation (5).
system with different sampling rates [28], each one chosen in Input/output relation (2) can be stated in phasor form as
accordance to its local dynamics. An important issue in multi follows:
rate transient analysis is the interfacing of various simulation Y0 = Ae j ( + ) = e j Ae j , (6a)
processes running at different sampling rates. Interface
where
variables from slower to faster processes should be
interpolated, whereas those from faster to slower processes {
y0 (t ) = e Y0 e j 0t } (6b)
must be decimated [30-33]. These two processes, interpolation Two advantages of phasor representation become apparent
and decimation, introduce aliasing errors which can be treated from the comparison of (6a) and (2). The first is that the
by FrequencyDomain and SignalProcessing techniques [34]. input/output relations are given by the multiplication of two
This is the main topic in the fourth part of this chapter. complex numbers (i.e., phasors). The second is that, at the
particular frequency 0 , the LTI system of Fig. 1 is fully
III. FREQUENCY DOMAIN BASICS characterized by complex number ej and this number can
also be regarded as a phasor.
A. Phasors and FD Representation of Signals
Now consider the representation of x 0 (t) by (4). The first
Figure 1 represents a linear timeinvariant system (LTI). term on the r.h.s. is associated to a phasor A 1 of negative
Consider first that its input is a pure sinusoid with constant frequency 0 , while the second term is to a phasor A +1
amplitude A, frequency 0 and phase : of positive frequency + 0 ; see Fig. 2b:
x0 (t ) = A cos( 0 t + ) (1) Im Im
The input/output relationship for the system is expressed
0 t 0 t
symbolically as follows:
x0 (t ) LTI
= y0 (t ) Ae j
Ae j

A well established fact for LTI systems is that the output y(t) is
Re - Re
always a pure sinusoid with the same frequency 0 [3]:
Aej
y0 (t ) = A cos( 0 t + + ) . (2)
0 t
Only amplitude and phaseangle of an input sinusoid are
Amplitude Amplitude
changed by the system. Sinusoids are therefore said to be 0 t 0 t
characteristic functions (or eigenfunctions) of LTI systems.
Complex exponentials offer a convenient alternative to
sinusoids in the analysis of LTI systems. Consider the time time
following equivalences for x 0 (t) in (1): (a) (b)
{
x0 (t ) = e Ae j ( 0t + ) } (3) Fig. 2. Sinusoidal signal representation. a) Realaxis projection of complex
exponential signal. b) Sum of two complex conjugate exponential signals.
and
 22

A
2
e j ( 0t + ) A1 = A
2
e j and the symbol represents the elementbyelement product
of two vectors. The timedomain output waveform y K (t) is
and
obtained from (13b) as follows:
A e j ( 0t + ) A+1 = A e j . K
y K (t ) = Y e
2 2
jk0t
k . (14)
Sinusoid x 0 (t) is further represented by the following vector k = K
whose elements are phasors: The equivalence between expressions (14) and (11) can be
x0 (t ) = A cos( 0 t + ) X = {A1 ,0, A+1} (7) verified easily. Expressions (12a,b) and (13ac) extend Phasor
Analysis to signals composed by harmonically related
Note here that the frequency associated to each element is sinusoids. Note that for the signal y(t) in (14) to be real
determined by its vector positionindex; that is, (1) 0 for the valued, the following conditions must hold:
first, (0) 0 for the second and (+1) 0 for the third. Note also
k = k and k = k . (15)
the introduction of a zero element as placeholder for a zero
frequency component which, for x 0 (t) in (3), certainly is null. B. Fourier Series
Vector representation (7) is readily extended to signals
A signal x p (t) is said to be periodic when it repeats itself at
composed by a number of harmonically related sinusoids.
constant time intervals T:
Recall that two sinusoids are said to be harmonically related
when their frequencies are multiples of a third one 0 called x p (t) = x p (t+T). (16)
fundamental. Consider the following signal: Figure 3 depicts a periodic signal. The minimum value of
K T > 0 for which property (16) holds is called fundamental
x K (t ) = A0 + A
k =1
k cos(k 0 t + k ) . (8) period.
A periodic waveform x p (t) is further said to be a power
By extension of (7), a phasorvector representation for x K (t) signal when its mean power is finite; that is:
is: 1
x (t )
2
Pxp = dt < .
x K (t ) X K = {X K , , X 1 , X 0 , X 1 , , X K } ,
p
(9a) T
T
where It is straightforward to show that signal x K (t) in (8), and in
X 0 =A 0 (9b) (10), is periodic and has fundamental period T = 2/ 0 . The
and Fourier Theorem establishes that a periodic signal of power
j
1 A e k , k = +1,+2, ,+ K x p (t) can always be approximated by a series x K (t), as in (8) or
X k = 1 2 k j . (9c) in (10), in such way that the power of the difference between
2 A k e k
, k = 1,2, , K
x p (t) and x K (t) tends to zero as the number of seriesterms K
increases towards infinity [2]; i.e.,
The original signal x K (t) (8) is readily recovered from vector
representation (9a) as follows: 1
x (t ) xK (t )
2
+K
lim p dt = 0 . (17)
K T
x K (t ) = X
k = K
ke
jk 0t
(10) T
On the grounds of the Fourier Theorem, the following
A large class of signals in engineering can be represented, equivalence is stated for a periodic signal of power:
+
x p (t ) = X
or at least approximated, by expressions (8), (9a) or (10). If jk0 t
x K (t) of (8) is an input to LTI system of Fig.1, the output can ke (18)
k =
be expressed as follows:
K This expression corresponds to the Fourier Series in its
y K (t ) = 0 A0 +
k =1
k Ak cos(k 0t + k + k ) . (11) complex exponential form. Since x p (t) is assumed realvalued,
the coefficients X k of (18) with negative index should be
For excitations of the form in (8), the LTI system is complexconjugates of their positiveindex counterparts [2]:
characterized by the following vector: X k = X k*
H K = {H K , , H 1 , H 0 , H 1 , , H K } , (12a)
where:
xp(t)
H k = k e j k , k=0,1, 2, , K. (12b)

Input/output relation can thus be expressed in phasorvector

form as follows:
YK = H K X K , (13a)
where
... T 0 T 2T ... t
YK = {Y K , , Y1 , Y0 , Y1 , , YK } , (13b)
Fig. 3. Example of a periodic signal.
Yk = X k H k , k=0,1, 2, , K, (13c)
 23

Fourier coefficients are obtained through the following j 0 2h

ZG = log e ,
expression [1,2,3]: 2 r
x p (t )e jk0t dt

1 Z E is the earth impedance in p.u.l.
Xk = (19)
T j 0 1
T ZE = log e 1 +
2 h j
In the same form as signal x K (t) in (10), x p (t) in (18) can be E
represented in vector form, only that now the dimensions are and Z C is the conductor impedance in p.u.l.
infinite:
x p (t ) X = { , X 2 , X 1 , X 0 , X 1 , X 2 , }
2
j 0 c
Z C = c2 + .
A plot of the magnitudes of elements in X against their r (2 r )2
corresponding frequencies is illustrated in Fig. 4a. A similar The transfer function for the line setup is:
plot for the angles of elements in X is shown in Fig. 4b. Vector Vout / Vin = e ZY length
.
X, as well as its associated graphs in Figs. 4a and 4b, are
referred to as the spectrum of x p (t). Since the components of X Figure 6a shows a plot of the input voltage v in (t) being
are complex, full graphical representation of its spectrum approximated by partial series with K=17. Figure 6b shows a
requires two plots, one for phasor magnitudes and the other for plot of the output voltage v out (t) obtained form the above
phasor angles. An alternate spectrum specification consists in transfer function along with expressions (13), (14) and (19).
providing one plot for the real parts of the phasorelements Note that the use of partial Fourier Series produces oscillation
and a second plot for the corresponding imaginary parts. errors in both figures, 6a and 6b. If better precision is
|Xk | required, the window techniques described in subsection
V.A.1 can be employed here as well.
Xk
+

... 0
... 0 0 0 2 0 3 0 ...
0 0 20 30 ...


(a) (b)
Fig. 4. Periodic signal spectrum. a) Magnitude spectrum. b) Phase angle (a) (b)
spectrum.
vin(t)
In sum, the Fourier (Series) Theorem permits the extension +0.5
of Phasor Analysis to the treatment of linear systems being
excited by periodic signals of power. This is illustrated next by t
means of an application example. 0.5 T=2 ms
1) Example 1; (c)
Fig. 5. Singlephase transmission line excited by periodic signal. a)
A singlephase aerial line is 10 km long and it is excited by Transversal geometry. b) Line layout. c) Input waveform.
a voltage source that produces a square wave with a period of
T=2 ms. The line is terminated in its characteristic impedance. C. Fourier Transform
The voltage waveform is to be determined at the line Fourier Series decomposition of a signal into harmonic
termination assuming that the source has been connected long sinusoids, or into complex exponentials, is extended next to
time enough to consider steady state operation; so, the Fourier non periodic waveforms.
Series method can be used. Figure 5a provides the transversal A signal x(t) is said to be of energy if its total energy is
geometry of the line along with the electrical data required to finite; that is, if
determine the line parameters, Fig. 5b shows a longitudinal

x(t )
2
diagram of the line and its connections and Fig. 5c depicts the Ex = dt < .
input waveform.
Line admittance in per unit length (p.u.l.) is calculated Consider a signal x(t) of finite duration, starting at t = 0 and
through the following expression [21]: ending at t = t 0 , as the input to the LTI system of Fig. 1. A
j2 0 periodic extension for this signal is given by the following
Y= .
log e (2h / r ) expression:
+

x(t nT )
The line impedance parameter in p.u.l. is calculated as follows:
x p (t ) =
Z = ZG + Z E + ZC ,
n =
where Z G is the geometric impedance in p.u.l.
 24

1 Then, consideration of (1/T) = ( 0 /2) is introduced:

0
+ T /2

0.5
x p (t ) = 2
x p (t )e jk0 t dt e jk0 t

Input (p.u.) k = T / 2
0
Next, as the limit of T approaching infinity is taken 0
becomes an infinitesimal and is denoted by d, k 0 becomes
a continuous variable and is denoted by , x p (t) becomes x(t)
-0.5
and the summation becomes an integral [1]:

x(t ) = x(t )e jk t dt e jk t d

-1 1
-1 -0.5 0 0.5 1 (20)
2
Time (ms)
(a) Note that (20) is an identity and that the integral inside braces
1 corresponds to a function of that hereafter is denoted by
X(). Hence:

0.5
X ( ) = x(t )e
j t
Output (p.u.)

dt (21a)

0
and

x(t ) = X ()e
-0.5 1 jt
d (21b)
2

-1 Expression (21a) corresponds to the Fourier Transform (FT)
-1 -0.5 0 0.5 1
Time (ms) and (21b) corresponds to the Inverse Fourier Transform (IFT)
(b) [1,2,3]. The Fourier Transform decomposes non periodic
Fig. 6. a) Fourier Series approximation of square wave input signal. b) signal x(t) into a continuous frequency spectrum X(). Figures
Output signal as obtained by the Fourier Series method.
8a and 8b present typical plots of X(). Since it has been
assumed that x(t) is an energy signal, the existence of its
Figures 7a and 7b provide the respective representations for
Fourier Transform is ensured [1,2]. The relationship between
finite duration signal x(t) and for its periodic extension x p (t).
x(t) and its spectrum X() is stated symbolically as follows:
Note in these figures that, as T > t 0 , x p (t) reproduces x(t)
inside the interval [0,T]. It is clear also that x p (t) becomes x(t ) FT
X ( )
equal to x(t) when T approaches infinity. X()
|X()|
x(t) +

0
0 t0 t
0
(a) (a) (b)
xp(t) Fig. 8. Spectrum of non periodic signal. a) Magnitude spectrum. b) Phase
angle spectrum.

The output of LTI system in Fig. 1, when it is excited by

non periodic input x(t), is expressed in the frequency domain
T 0 t0 T t as follows by the product of two complex functions:
(b) Y() = H()X(), (22)
Fig. 7. a) Signal of finite duration x(t). b) Periodic extension of x(t).
As before, X() is the spectrum (or Fourier Transform) of x(t).
Back to finite values of T, when x(t) is an energy signal, H() is a function characterizing the LTI system and is
x p (t) is a signal of power with Fourier Series representation as referred to as its frequency response. The time domain output
in (18) and coefficients given by (19). To extend Fourier (or system response) is obtained by applying the inverse
Analysis to non periodic x(t), first (18) and (19) are applied to Fourier Transform (21b) to Y() in (22):

x p (t) and combined as follows:
y (t ) = X ()H ()e
1 jk t
d . (23)
1 2
+ T /2
x p (t ) =
T
x p (t )e jk0 t dt e jk0 t .


It can be shown that (23) is equivalent to [1]:
k = T / 2
 25

r(tt0)
y (t ) = x( )h(t )d , (24a)
(t t0)

where h(t) is the inverse Fourier Transform of H().
Expression (24a) defines the convolution operation between 1/
two functions, x(t) and h(t). This operation also is represented
symbolically as follows: 0 t0 t 0 t0 t
y(t) = x(t) h(t) (24b) (a) (b)
Fig. 9. a) Impulse function. b) Rectangular pulse.
The Convolution Theorem states that the convolution of
two time domain functions is equivalent to the product of their A. Aliasing Effect
Fourier transforms, or spectra [1,2]. It can be shown as well
The sampling of x(t) at a regular intervals t can be
that the product of two TD signals is equivalent to the
convolution of their Fourier transforms. represented mathematically by its product with t (t):
The Convolution Theorem is a convenient property of the x S (t) = x(t) t (t) (25a)
Fourier Transform. Nevertheless, there are two major or
difficulties for its direct application to practical transient +
problems. The first one comes from the fact that Fourier
Transforms are guaranteed only for signals of energy and this
xS (t ) = x(kt ) (t kt ) ;
k =
(25b)

excludes several cases of practical interest, such as periodic

waves. The second difficulty is due to the Fourier Transform this is illustrated by Fig. 10c.
being an analytical method, and analytical functions that x(t)
represent practical signals usually are very difficult to obtain x(t0)
and handle. Nevertheless, the Fourier Transform provides the
basis to other more practical FD methodologies; among these (t t0)
are the Discrete Fourier Transform (DFT), the Numerical
Laplace Transform (NLT) and the ZTransform. 0 t0 t
(a)
IV. DISCRETETIME FREQUENCY ANALYSIS
In the analysis of systems by digital computer continuous
t (t t0)
time signals must be sampled usually at regular intervals and 1
must be represented by ordered sequences of their samples. A
convenient way to analyze the sampling process is by through ...
impulse functions. Recall that an impulse (tt 0 ) is a
generalized function which is zero all over t, except at t = t 0 , 0 t 2t 3t kt t
where it takes a very large and undetermined value (see Fig. (b)
9a). Figure 9b provides the plot of a rectangular function x S (t)
r(tt 0 ) of width , height 1/ and centered at time t = t 0 :

1 / , t t0 / 2
r (t t 0 ) =
0, t t0 > / 2
0 t 2t 3t ...
The impulse function (tt 0 ) in Fig. 9a is seen as the limit of
(c)
r(tt 0 ) when approaches zero. Fig. 10. Sampling a signal by a train of pulses. a) Continuoustime signal. b)
Consider now a continuoustime function x(t) as the one Train of pulses. c) Sampled signal.
shown in Fig. 10a. The sifting (or sampling) property of the
impulse function states the following result [1,2]: Figure 11a shows the spectrum of x(t) being denoted by
X(). Figure 11b depicts the spectrum of t (t) that also is a
x(t ) (t t )dt = x(t ) .
0 0 train of pulses along the axis [3]:
t (t ) FT
S s ( ) ,

(26a)
Another important (generalized) function is the train of
pulses denoted by t (t) and consisting in an infinite where
sequence of pulses occurring at time intervals of size t.
+ s ( ) = ( k S ) (26b)
t (t ) =
k =
(t kt )
and
k =

Figure 10b provides a plot for t (t). S = 2/t. (26c)

 26

Note that S is the frequency interval between pulses. In t / M (28a)

agreement with the Convolution Theorem, the spectrum of
The inverse of t is the sampling frequency or rate. Its units
sampled signal x S (t) is obtained by the convolution of X()
are samples per second:
with S s () whose result is:
+ F S = 1/t (28b)
X S ( ) = S
k =
X ( kS ) (27) The equality option in (28a) (i.e., t = / M ) corresponds to
the Nyquist sampling interval t Nyq , and its inverse F Nyq =
Figure 11c illustrates the plot of X S () according to (27). 1/t Nyq is known as the Nyquist frequency [1,2,3].
Note from this figure that time domain sampling causes the
repetition of shifted (or frequency modulated) replicas of the |X()|
original spectrum X(). This is the effect of aliasing in
frequency domain. The overlapping of frequency components
from replicas provokes the sampling or aliasing errors. In
extreme cases, poor choice of a sampling rate results in the
original signal not being recoverable from its samples. M 0 M
|X() (a)
X S ( )

0
(a)
S s( ) S 0 S 2 S
M M
S (b)
... ... Fig. 12. a) Spectrum of bandlimited signal. b) Spectrum of sampled band
limited signal.
S 0 S 2S

Consider now that a signal with bandlimited spectrum
(b) X S () has been sampled with an interval complying with
XS ( ) (28a). The original signal is readily recovered by passing its
samples through a low pass filter with the following frequency
response:
1 / S S / 2
... G ( ) = (29)
0 > S / 2
S 0 S 2 S ...
(c) This filter response is plotted in Fig.13a. From (27) and (29):
X ( ) = X S ( ) G ( )
Fig. 11. Effect of sampling on the spectrum of a signal. a) Spectrum of a
continuoustime signal. b) Spectrum of a train of pulses. c) Spectrum of (30)
sampled signal.
The inverse Fourier Transform of G() is obtained as follows
and its plot is shown in Fig. 13b:
In Signal Analysis, timetofrequency relations usually are
sin (t/t )
g (t ) =
symmetric. This has been already observed with the
(31)
Convolution Theorem and it also is the case with the aliasing (t/t )
effect; that is, the sampling of a signal spectrum creates
superposition of timeshifted replicas of the signal, or TD On applying the Convolution Theorem to (30):
aliasing.
x(t ) = x S ( )g (t )d ,
B. Sampling Theorem

A signal x(t) is said to be bandlimited if there is a replacing x S (t) from (25b) and performing the integration:
maximum frequency M above of which its spectrum X() is +
zero (see Fig. 12a): x(t ) = x(kt )g (t kt ) (32)
X ( ) = 0, > M . k =
Figure 14 provides a plot of x(t) in accordance with (32).
For this type of signals one can select a sampling interval that This figure shows that the reconstruction of x(t) is by
avoids the overlapping of frequency replicas. This is illustrated superposing replicas of g(t), each one scaled by a sample value
in Fig. 12b and it follows from (26b) that the required and shifted by an amount of time that is multiple of the
sampling interval is:
 27

|X()| C. Conservation of Information and the DFT

In addition to (27), the spectrum of sampled signal x S (t) can
be obtained as follows:
1/ S +
X S ( ) = x(kt )e
k =
jkt
(33)
S/2 0 S/2 This expression is obtained applying the Fourier Transform to
(a) (25a). It can be shown, either through (27) or (33), that X S ()
is periodic with a repetition interval S = 2/t. The discrete
representation of X S () can be accomplished by sampling only
one period. First suppose that this is done with N samples:
= S / N = 2 / (Nt )
and continuous variable in (33) is replaced by m :
+
X S (m ) = x(kt )e
k =
j 2km / N
.

(b) Then, the sum at the r.h.s. is carried out in groups of N terms.
Fig. 13. a) Frequency response G() of low pass ideal filter. b) Time domain
image of G(). This is done by expressing summation variable k as k=n+lN:
+ N 1

sampling interval t. Note that, for instance, g(tkt) is zero at

X S (m ) = x((n + lN )t )ee
l = n=0
j 2 ml j 2mn / N

1
.
all sampling instants, except at the kth one. The implication
of this is that the value of x(t) at t = kt is determined only by Next, the order of summations is interchanged:
N 1 +
the corresponding sample x(kt), whereas a value of x(t)
between sampling points is given by a combination of all the
X S (m ) =
n =0

l =
x((n + lN )t ) e j 2nm / N .

samples, each one weighted by its corresponding shifted
function g(t). Afterwards, a new discretetime signal x(nt) is defined:
+
x(t)
xk1 xk
x' (nt ) = x((n + lN )t ) ;
l =
xk+1
hence:
N 1
X S (m ) = x' (nt )e
n =0
j 2nm / N

Note that x(nt) is a periodized version of x(nt) with

t aliasing. Finally, by assuming that x(nt) is a finite sequence
of N or less terms, x(nt) equals x(nt) for samples between
Fig. 14. Reconstruction of a signal from its samples.
n=0 to n=N1:
N 1
It follows from (32) and from Fig. 14 that the role of g(t) is
the one of an interpolating function. This is in fact known as
X S (m ) = x(nt )e
n =0
j 2nm / N
(34)
the Ideal Interpolator [1,2]. Function g(t) is essentially a
Recall that derivation of (34) started with the assumption of
theoretical tool. Its practical realization as a filter is impossible
a discrete spectrum consisting of N samples and it ended up
since, as it can be observed from Fig. 13, it would have to start
establishing the correspondence with N TD samples at the
acting at time t=. Nevertheless, practical signal most. This is in agreement with the principle of Information
recuperation usually is achieved satisfactorily with a well Conservation.
designed nonideal lowpass filter. Expression (34) also is readily identified as the Discrete
The results expressed by (28a) and (32) conform the Fourier Transform (DFT). Its inverse, the IDFT, is as follows
Sampling Theorem that can be worded as follows: [1,2]:
A bandlimited signal x(t) with maximum frequency F M = N 1
x(nt ) = X (m)e j 2nm / N
1
M /(2) can be fully recovered from its samples, provided it N
S (35)
has been sampled at a frequency F S that is equal at least to m =0

the double of maximum frequency F M ; i.e., sampling
frequency F S must be at least equal to Nyquist frequency F Nyq .
The original signal can be fully reconstructed from its samples
through the ideal interpolator function g(t) defined by (31).
Expressions (34) and (35) establish a unique relation
between one finite sequence of N samples, say in time domain,
and another one lengthN sequence of spectral samples.
Note that sequences X S (m) and x(nt) can be extended
beyond their original lengths N through (34) and (35);
nevertheless, these extensions are mere periodic repetitions.
 28

D. Fast Fourier Transform evaluated by two (N/4)sample DFTs and the required number
Numerical approaches to spectral or frequency domain of multiplications for this is (N2/4) + N/2. Since N is a power
analysis usually end up with DFT (34) and IDFT (35) of 2, the subdivisions can continue until one ends up with N/2
evaluations. It is customary for (34) and (35) to omit the term DFTs, each one with 2 samples and this requires N/2
t in the argument of x(nt) and to denote this variable simply multiplications. The FFT algorithm evaluates DFTs and IDFTs
as x(n), or as x n . Similarly for X S (m), and subindex by continued subdivisions until ending up with N/2 two
s are omitted and this variable is written as X(m), or as X m . It sample transforms. The number of multiplications is thus:
N
is also customary to denote the complex exponentials as log 2 (N )
follows: 2
e 2jmn / N = WNmn From this expression, it can be observed that the number of
operations required by the FFT algorithm increases almost in
Expressions (34) for the DFT and (35) for the IDFT take the linear proportion to the number of samples N, whereas in the
following respective forms: direct evaluation of the DFT by (36), or of the IDFT by (37),
N 1 the number of multiplications increase in quadratic proportion
X (m ) = x(n)W
n =0
mn
N , m = 0, 1, 2, , N1 (36) to N. Table I provides a comparison between the number of
multiplications required by the FFT and the one by direct
and evaluation [23].
N 1
x(n ) = X (m)W
1 mn
N , n = 0, 1, 2, , N1 (37) TABLE I
N m =0 COMPARING NUMBER OF MULTIPLICATIONS REQUIRED BY THE
Clearly form (36) and (37), the evaluation of the DFT and of DIRECT DFT AND THE FFT ALGORITHMS
the IDFT is essentially through the same procedure. It is clear
also that direct evaluation of (36) or (37) takes N2 complex Number of DFT FFT Ratio,
multiplications and N(N1) complex sums. The Fast Fourier Samples (N/2) DFT
Transform (FFT) is an algorithm for evaluating the DFT and N N2 log 2 (N) FFT
the IDFT with very high computational efficiency. Its working 4 16 4 4
principle is outlined as follows. 8 64 12 5.33
As N, the number of samples, is decomposed in its prime 16 256 32 8
factors, the DFT or the IDFT can be evaluated in partial
32 1024 80 12.8
groups of sizes determined by these factors. The evaluation by
64 4096 192 21.3
partial groups requires less operations than direct calculations
128 16384 448 36.57
by (36) or (37). The highest numerical efficiency is obtained
256 65536 1024 64
when N is a power of 2; i.e.:
N = 2 i. 512 262144 2304 113.77
1024 1048576 5120 204.8
Assuming that this is the case, (36) is organized in two groups
of sums. One is for evenindexed samples and the other is for V. FREQUENCY DOMAIN TRANSIENT ANALYSIS
the oddindexed ones. An auxiliary integer variable k
The time domain description of power systems for the
running from 0 to (N/2)1 is introduced; so, n = 2k is for
analysis of transients is through relations involving integrals,
evenindexed samples and n = 2k+1 is for the oddindexed
differentials and convolutions. In the frequency domain these
ones. With these changes (36) yields:
( N / 2 )1 ( N / 2 )1 relations take an algebraic form. FD transient analysis is
X (m ) =
k =0
x(2k )WNmk/ 2 + WNm x(2k + 1)W
k =0
mk
N /2 , performed first by building a system model in the Fourier [15],
Laplace [27] or Z domain [16,18,24]; then, FD transient
responses are obtained solving the corresponding algebraic
m=0, 1, 2, , (N/2)1 (38a)
relations; finally, the TDresponse waveforms are derived
( N / 2 )1 ( N / 2 )1 from their FD counterparts by applying the corresponding
N
X + m =
2
x(2k )W
k =0
mk
N /2 WNm x(2k + 1)W
k =0
mk
N /2 , inverse transform. An additional advantage of FD methods is
that often system elements are synthesized in the frequency
m=0, 1, 2, , (N/2)1 (38b) domain; their incorporation into FD system models is thus
direct.
The original DFT with N samples can thus be evaluated by two
The FD technique described next is referred to as the
DFTs with (N/2) samples. The number of multiplications
Numerical Laplace Transform (NLT) [11, 14]. It is very robust
involved in (38a) and (38b) is
and offers unprecedented numerical accuracy. First, the
(
2 (N / 2 )2 + N / 2 = N 2 / 2 + N / 2 ) problems associated with the numerical inversion of the
Fourier Transform are addressed. Then, the processes
This number is approximately one half of the multiplications developed in the solution of these problems lead in a natural
required by the direct evaluation of the Nsample DFT. Each way to the NLT technique. Finally, the usefulness of the NLT
(N/2)samples DFT in (38a) and (38b) can be further is demonstrated with two application examples.
 29

A. Fourier Transforms and Transients HR( ) hR(t)

Consider that the spectrum of a transient signal is available
as N samples of the form Y(m), with m= N/2, , 1, 0, 1, t=/ M
N/21. To obtain the corresponding TD waveform y(t), the 1.0 1/t
inverse Fourier integral (21b) is approximated numerically as
follows:
( N / 2 )1
M 0 M t

y (t ) y1 (t ) =
2 m= N / 2
Y (m )e j (m ) t . (39)
(a)
2t
(b)
This approximation involves two steps. The first one is the Fig. 15. Rectangular (truncating) window a) Frequency response H R (). b)
Time domain image h R (t).
truncation of the integrationrange of in (21b), from the
infinite range [,] to the finite one [ M , M ]:
M
0, t < 0
y (t ) y 2 (t ) = Y ()e
1
u (t ) =
j t
d , (40a) . (45)
2 1, t 0
M
with Figure 16a provides a plot of u(t), while Figs. 16b and 16c
M = (N/2). (40b) illustrate its convolution with h R (t) in (43). The latter figure
depicts the approximation to u(t) obtained by truncating its
The second step is the discretization of the integrand, both in spectrum. Four important features of this approximation
(21b) and in (40a). Continuous variable is replaced by the should be pointed out. The first one is that, as a filter, the
discrete one m. In addition, Y() and ejt are rectangular window is non causal. It is clear from the
represented by their samples at these discrete values. Then, the comparison of Figs. 16a and 16c that the window output y 2 (t)
application of rectangular integration to (40a) yields (39). starts responding before t = 0; that is, before the input y(t) =
The two steps of truncation and discretization are analyzed u(t) starts acting. The second feature is that the discontinuity at
as follows with more detail. t = 0 is approximated by a continuous segment with a non zero
1) Frequency range truncation; risetime amounting to 0.42t. The third one is the presence of
oscillations that are most pronounced near the instant of the
Expression (40a) is equivalent to [4] discontinuity. These oscillations are referred to as Gibbs

phenomena. The fourth feature is the overshoot after the
y 2 (t ) = Y ()H ( )e j t d
1
(41)
2 discontinuity which reaches a peak value in the order of 9.0 %.
R
The step function approximation of Fig. 16c is of special
with interest for transient analysis. Every signal with an isolated
1, M discontinuity is equivalent to a continuous one with a
H R ( ) = . (42) superimposed step function. A major concern in FD transient
0, > M
analysis is the 9.0 % level of overshoot that cannot be
H R () is a rectangular (truncation) window and is plotted in decreased by making the truncation frequency M larger [3].
Fig. 15a. Its time domain image is given by the following In practice, this is decreased and a better approximation to y(t)
expression and is plotted in Fig. 15b: is obtained by applying a smoothing filter. By observing that
the Gibbs errors have an oscillation interval t Gibbs = / M ,
sin (M t )
hR (t ) = M . (43) one can realize that an effective smoother is a sliding window
2 (M t ) of duration / M (see Fig. 16c). For an improved estimate of
y(t), first the sliding window is centered at each point t x of the
Figures 15a and 15b should be compared with Figs. 13a and time range, then a weighted average is performed with all the
13b. Note from Fig. 15b that the zeros of h R (t) occur at regular values of y 2 (t) inside the window, next the result of this
intervals of size t 0 = /(2 M ). If the truncation frequency M average is assigned to the new estimate of y(t) at t x , finally, the
is made larger (i.e., wider bandwidth), the main lobe of h R (t) process is applied to all values of t in range.
becomes taller and, at the same time, narrower. In the limit, as The Lanczos window is a smoothingfilter that applies pure
M , h R (t) approaches the impulse function (t) in much averaging (or, constant weight) along its apperture. Its
the same manner as with R(t) in Fig. 9b. frequency response is [4]:
The timedomain relation between y(t) and its
approximation y 2 (t) is obtained by applying the Convolution sin (/M )
Theorem to (41): (/ ) , M
M
L ( ) =
y 2 (t ) = y (t ) hR (t )
(46)
(44) 0, > M

The effect of truncating the frequency range observed at its
best on signals with discontinuities. Suppose that the original Figure 17a shows a plot of L () that should be compared
signal y(t) is the unit step function u(t) defined as: with the rectangular window in Fig. 15a. This comparison
 30

shows the timedomain averaging is equivalent to a L()

continuous and gradual truncation of the signal spectrum. The
truncation of the spectrum of unit step (43) by the Lanczos
window results in the waveform plotted in Fig. 17b. It can be 1.0
seen that Lanczos window reduces the overshoot to 1.2 %.
This reduction comes at the expense of introducing a slightly
larger delay in the estimated signal; that is, risetime amounts
M 0 M
now to 0.73t. (a)
u(t) u(t)L(t)
1.2 %

t=0 t t=0 t
(a) (b)
u(t)hR(t) VH ()

1.0

t=0 t
M 0 M
(b) (c)
Tw
u(t)hR(t) Sliding Window Fig. 17. Lanczos Window. a) Frequency response. b) Time response when
Width: applied to a step function. c) Von Hann or Hanning window.
9.5 % Tw=/M

v in (t) through partial series (10) with its coefficients further

t0
t=0 t output. Note the practical absence of Gibbs errors and
(c)
Fig. 16. Effects of frequency truncation. a) Unit step function u(t). b)
Convolving signals u(t) and h R (t). c) Convolution result and sliding window.
A highly recommended window for transient analysis is the
one by Von Hann (or Hanning). Apparently, its use for this
purpose was first proposed in [17]. The frequency response of
the Hanning window is:
modified by the Hanning window. Compare this plot with the
one in Fig. 6a. Figure 18b shows a plot of the new system
compare this last plot with the one in Fig. 6b.
Finite Fourier series approximate periodic waveforms by
minimizing the power of the error signal (i.e., least mean
square error of the difference). On the other hand, however, a
series approximation with minimum overshoot error is
preferable in power transient analysis. The reason is that the
main objective here usually is to determine overvoltage and
overcurrent levels.

1 cos(/M ) 2) Discrete frequency range;

2 + , M
2 Consider now that a transient signal y(t) is to be synthesized
VH ( ) = , (47)
from samples of its spectrum Y(), and that the frequency
0, > M
range is not truncated. By applying rectangular integration in
(21b), the following approximation is obtained:


Y (m )e j (m ) t .
and the corresponding plot is shown in Fig. 17c. In addition to
y (t ) y3 (t ) =
applying a continuous and gradual truncation, VH () presents 2 m=
a continuous first derivative at the cutoff frequencies M .
Hanning window reduces further the overshoot to 0.63 %; the This expression is also obtained multiplying Y() by
risetime, however, is increased to 0.87t. () and applying the Inverse Fourier transform to the
product. Recall from (26b) that () is a train of unit pulses
a) Example 2. placed at regular intervals of length along the frequency
Consider again the transmission system described in Figs. 5a axis:

and 5b, as well as the input signal in Fig.5c. Again v in (t) is
y3 (t ) = Y ( )

( k ) e jt d
1
approximated by the finite Fourier series (10) with K=17. This 2 k =
time, however, the series coefficients are multiplied by VH ()

given by (47). Figure 18a shows a plot of the reconstruction of

 31

y3 (t ) = u(t kT )e c (t kT )
1
(49)
k =
0.5 Figure 19 depicts y 3 (t) as in (49). Notice that if attention is
Input (p.u.)
restricted to interval [0,T], y 3 (t) is composed only of y(t) and
0 all its past replicas:
0

-0.5 y3 (t ) = u(t )e
k =
c (t kT )
;

after reordering and factoring terms:

-1

(e )
-1 -0.5 0 0.5 1
Time (ms) y3 (t ) = u (t )e c t cT k
.
(a) k =0
1 This expression is a geometric series and 0<ecT<1; hence:

y3 (t ) = u (t )e c t
1
. (50)
0.5 1 e cT
Output (p.u.)

0
y3(t)
y(t)
-0.5 y(t+3T) y(t+2T) y(t+T)

-1
-1 -0.5 0 0.5 1 t=0 t
Time (ms)
(b) Fig. 19. Aliasing effect on an exponentially decaying step function.

Fig. 18. a) Approximating square wave signal of Fig. 5c by a partial Fourier
Series with coefficients weighted by the Hanning window. b) Response of
system in Fig. 5b obtained by the modified Fourier Series.
From (26a) and from the Convolution Theorem:
y3 (t ) =
( kT ) y (t )d ,
k =
where T = 2/. By exchanging the order between the
integral and the summation and by further performing the
integral the following relation between y 3 (t) and y(t) is
obtained:
y3 (t ) =
k =
y (t kT )
(48)
Expression (48) shows that the discretization of Y()
produces time domain aliasing. As transient signals generally
are not timelimited, the question here is as to what are the
conditions to obtain good approximations to y(t) by y 3 (t) given
in (48). Clearly, since y 3 (t) is periodic the useful range of the
approximation has to be confined to the interval [0,T].
To address the previous question consider first the case of
y(t) being a unit step u(t). According to (48), direct sampling
of its spectrum of u(t) results in an aliasing error in excess; that
is, the value for approximation y 3 (t) turns out to be infinite.
Let now the unit step be multiplied by a damping exponential
[5]:
It follows from (50) that the aliasing error is given by the
factor 1/(1ecT). In addition, if ecT<<1, then:
(
y3 (t ) u (t )e c t 1 + e cT , )
and the relative aliasing error is:
rel =
y3 (t ) y (t )
e cT (51)
y (t )
Finally, the original signal u(t) is recuperated with some
aliasing error after multiplying y 3 (t) by undamping
exponential e+ct:
u (t ) y3 (t )e ct = u (t ) + u (t )e cT
This example with the unit step illustrates the technique for
controlling aliasing errors by introducing exponential
damping. Despite its simplicity, the case of a step function is
highly relevant for transient analysis. Power systems are
composed by passive elements; consequently, their natural
responses are bounded, and mostly decaying; one can
therefore assume that the step function is good representative
of the worst case of natural responses and of excitation signals.
A highly convenient form to introduce the damping coefficient
c in FD transient studies is by working directly in the
Laplace domain. The result in (51) is useful to fix this
coefficient that is incorporated in the Laplace variable as
s = c +j.

y(t) = u(t)ect, B. Fourier and Laplace Transform

While the Fourier Transform is appropriate for steady state
and the spectrum of the resulting function y(t) be sampled.
analysis, the Laplace Transform is far better suited for
From (48):
transient studies. It is thus convenient to establish the
relationship between these two transforms. A large class of
signals of practical interest are not of energy and their Fourier
 32

Transforms cannot be assured. Often, however, when these ( N / 2 )1

signals are damped by a decaying exponential factor as it has x(nt ) =
X (c + jm )e cnt e jmnt ,
2 m = ( N / 2 )
been shown above, the Fourier Transform becomes applicable.
Consider a signal x(t), along with the following modification: n = 0, 1, 2, , N1. (57)
x MOD (t) =x(t)u(t)e ct
(52) In agreement with the principle of Conservation of
Information, the number of TD samples in (57) has been made
Assume that x MOD (t) is an energy signal and obtain its Fourier
equal to N; that is, the number of samples in FD. In addition, N
Transform as follows:
determines the following observation time for x(t):

T = Nt.
X MOD ( ) = x(t )u (t )e ct e jt dt = x(t )e (c + j )t dt

(58)
(53)
0 Recall that a maximum value for the observation time has been
already established as follows by virtue of T being the
By introducing the Laplace variable s=c+j: repetition (or aliasing) period in (48):

T = 2/, (59)
X MOD ( j (c s )) = x(t )e st dt

0 The combination of (58) and (59) yields the following relation:
On the grounds of this last result, the one-sided Laplace t = 2/N (60)
Transform is introduced as follows [2]:



The introduction of (60) in (57) yields:
X L+ (s )= x(t )e st dt = X MOD ( )
(54) ( N / 2 )1
e cnt 1
0
Note that the lower bound of the integral is taken as 0. This
x(nt ) = X (c + jm )e 2jmn / N ,
t N m = ( N / 2 )

choice is convenient for resolving ambiguities that can arise
n = 0, 1, 2, , N 1. (61)
from signals with a discontinuity at t=0 [2]. Such
discontinuities occur commonly in transient analysis. Note in (61) that the term inside the braces is an IDFT. It is
The corresponding Laplace inversion integral is obtained thus convenient to modify the summation index as follows:
now. First, the inverse Fourier Transform is applied to (53):
e cnt 1 N 1
x(t )u (t )e ct =
1
2
X MOD ( )e jt d .
(55) x(nt ) =
t

N
X (c + jm)e
m =0
,
2 jmn / N

ct n = 0, 1, 2, , N 1, (62)
Then, both sides of (55) are multiplied by e . Next, the
Laplace variable s=c+j is introduced and the assumption is where, for m>N/2:
made for x(t) being zero as t<0. Finally, all these changes lead X(c+jm) = X*(c+j(Nm))
to the Inverse (one sided) Laplace Transform:
c + j and X*( ) is denoting the complex conjugate of X( ).
x(t ) =
1
X L+ ( s )e st ds (56) To minimize Gibbs (frequency truncation) errors in (62),
2j the discrete FD samples X(c+jm) are multiplied by a data
c j
For ease of notation reasons, here as in most texts on the window. The Von Hann window is recommended here [17],
subject the symbol X( ) is hereafter used indistinctly to denote and:
the Laplace or the Fourier transform of x(t). To avoid the
confusion that this may bring, strong recommendation is made
x(nt ) =
e cnt
t
{
ifft [X (c + jm ) VH (m ) ]mN=10 }
(63)
here as to always keep in mind that X(s) is a shorthand notation This expression is the Inverse Numerical Laplace Transform.
for X L+ ( ) in (54) and in (56). Note that the summation of (62) is performed in (63) through
the FFT algorithm. Aliasing error minimization is attained
C. The Numerical Laplace Transform there by a proper choice of damping coefficient c.
One of the major advantages of analyzing transients in It follows from (51) that the relation between the overall
frequency domain is that signal relations involving integrals, relative aliasing error and the damping coefficient is:
c = [log e ( rel )] T
derivatives and convolutions become algebraic expressions.
(64)
For practical analysis, Fourier and Laplace transforms must be
applied in discrete form and it has been shown already that FD Ideally, one would like to specify an arbitrary small value for
discretization produces TD aliasing errors. These errors can be rel ; however, there are practical limits for this. For the
controlled by introducing a damping coefficient and this is frequency sampling used in (57), Wedepohl reports in [14] the
most conveniently done with the Laplace transform. following rule that has been found by experience:
Let now X(s) denote the (one sided) Laplace transform of rel = 1/N (65)
transient signal x(t). A first approximation to the numerical
solution of the inverse Laplace transform (56) is: In sum, Numerical Laplace Transform inversion is attained
through (63). In practice one must choose two or at most three
parameters to apply this expression. By setting the maximum
 33

observation time T, frequency resolution is fixed 20 s. Figure 21c shows also the results obtained with the
automatically by (59). Choice of maximum or cutoff frequency EMTP using the FDLine model with two integration steps:
M automatically determines t as the Nyquist sampling t = 20 s and t = 2 s. It can be observed in this figure that
interval in (28a). In power system analysis the bandwidth for when the resolution of t = 20 s is used, the NLT and EMTP
the different types of transient events is generally well result differ substantially. As the resolution is increased ten
established. As T and t, or as M and , are given, the times in EMTP (t = 2 s), its result becomes closer to the
number of samples N becomes determined. When N is not of one with the NLT. Stress is made here that essentially the same
the form 2i, it is recommended here to choose the next larger results are obtained by using the Universal Line Model (ULM)
value that is an integer power of 2. To select a suitable value [35]with the EMTP.
for damping coefficient c, recommendation is made here for For the time being, the NLT cannot produce sequential real
the use of Wedepohls relation (65). As an observation time T time and offline simulations. In these cases one must rely on
is being set for a specific analysis, one has to take in advanced TDEMTP models. Nevertheless, this last example
consideration that between 3 % and 5 % of the last samples illustrates the form in which Frequency Domain methods assist
obtained by (58) are useless due to amplification of runaway in the development and finetuning of EMTP study cases.
aliasing and Gibbs errors.
D. Application Examples with the NLT
1) Example 3.
Consider again the transmission line example 1 in
subsection III.B. The line data are given in Fig. 5a, and the
longitudinal layout is provided by Fig. 5b. This time, however,
the line is terminated in a threephase open circuit and the
input signal is a step function starting at t=0. As opposed to (a)
example 1, the system response includes now a transient
component. Figure 20a depicts the input step function, while
Fig. 20b shows the farend response obtained with the NLT
using N = 2048 samples. Observe in Fig. 20b the response
delay due to the travel time of the line. Observe also the
transient oscillations caused by the reflection at the line open
end. Note that the oscillation period is four times the line
travel time.
2
Input (p.u.)

1
0
-0.5 0 0.5 1 1.5
Time (ms) (b)
(a) NLT
Recovery Voltage (p.u.)

2 EMTP 1
EMTP 2
Output (p.u.)

2 1
1
0
0
0 0.5 1 1.5 2 -1
Time (ms)
(b) -2
Fig. 20. a). Excitation signal. b) Response.
0 0.005 0.01 0.015
2) Example 4. Time (s)
Figure 21a shows the connection diagram for the line in Fig. (c)
Fig 21. a) System Layout. b) Line data. c) Comparing NLT and EMTP
21b [20,22]. The transient recovery voltage of switch t 2 is to results: NLT (t = 20 s), EMTP1 (t = 2 s), EMTP2 (t = 20 s).
be obtained. The simulation starts at t=0 with the simultaneous
phase energizing and with a permanent fault at the line end. E. Brief History of NLT Development
The fault condition is represented by the three 0.1 shunt
In addition to the NLT technique that is described here,
resistances. After 2 ms of energizing, the switches open
various other methods have been developed for applying
simultaneously.
Laplace Transforms to transient system analysis. The one
The transient recovery voltage at switch t 2 is shown in
presented in [27] deserves special attention; although, it still is
Fig. 21c as calculated with the NLT using a resolution of t =

at a very early stage of development. The technique described
here originated in the early 1960s by the works of N.
Mullineux, et al, [57]. It then evolved in the late 1960s
through the works of L. M. Wedepohl, et al, [8,9,12,13]. At
that time the technique was referred to as The Modified
Fourier Transform. In the early 1970s A. Ametani introduced
the use of the FFT algorithm [10]. In the late 1970s D. J.
Wilcox produced a systematic view of the methodology
relating it to the Laplace Transform theory and provided
important criteria for its practical application [11]. In the early
1980s Wedepohl introduced in the technique further
refinements that have permitted to attain very high accuracy
[14]. More recent work on the NLT can be found in [17] to
[23], and in [25] and [26].
VI. MULTIRATE TRANSIENT ANALYSIS
Consider a large network in which two regions can be
distinguished: region 1 with an ongoing fast disturbance and
region 2 with operation close to steady state. The network may
be conveniently subdivided for its analysis in these two regions
as illustrated in Fig. 22. The interface between the regions is
through the exchange variables x(t) and y(t). The first one
conveys the necessary information from the fast to the slow
dynamics region. The second one carries this information from
the slow to the fast region. The network can be simulated
digitally with two different sampling rates. A high rate
F XS = 1/t x is assigned to simulation of region 1 and it is to
be in agreement with its fast dynamics. The other is a slower
rate F YS = 1/t y used for region 2. In practical situations these
two rates can be up to three orders of magnitude apart and the
savings in computational time make highly attractive the
pursuit of two-rate and even multi-rate simulation techniques.
REGION 1
FAST
DYNAMICS
tx
x(t) y(t)
REGION 2
SLOW
DYNAMICS
ty
Fig 22. Subdivision of a large network into two regions operating with
different dynamics..
Signal x(t) is produced by the ongoing processes in region
1. Digital simulation of this region delivers a discrete version
of x(t) as follows:
x d (n) = x(nt x ), n=0, 1, 2,
The spectrum of x d (n), denoted here as X d (), has to be
periodic with repetition interval 2 XM = 2F XS . This is
illustrated by Fig. 23a. The maximum frequency XM must be
chosen sufficiently high so x d (n) provides an accurate
representation of x(t) .
34
Consider now that the samples of x d (n) are produced at the
rate F XS . Region2 simulation cannot accept all these samples,
as it runs at the lower rate F YS . Suppose that F XS is L times
faster than F YS , or that timestep t y is L times larger than t x .
Signal x d (n) must thus be decimated by an Lfactor; that is, for
every L samples of x d (n) one is kept and the other L1 are
discarded. A new signal x d (n) = x d (nL) is produced by this
decimation process that is represented as follows:
( L) x d (n) = x d (n) = x d (nL)
The spectrum of x d (n) is a further periodization of X d (),
now with a repetition interval 2 YM = 2 XM /L. Figure 23b
illustrates this and it can be observed that aliasing errors could
be severe. To avoid these errors, signal x d (n) should be filtered
before the decimation. Figure 23c illustrates the filtering of
x d (n) by an ideal lowpass filter (LPF). Now the maximum
frequency of the filtered signal is YM . Figure 23d shows the
spectrum of the new signal after being decimated by an L
factor. Note the absence of aliasing. Although ideal filters are
non realizable, good results can be obtained in practice from a
real filter with cutoff frequency YM and with sufficient
attenuation in its transition band.
Xd()
XM XM
(a)
Xd()
YM YM
(b)
Xd()
YM YM
(c)
Xd()
YM YM
(d)
Fig 23. a) Spectrum of signal x d (n). b) Spectrum of x d (n) after decimation by
an Lfactor. c) Filtering of by an ideal filter. d) Spectrum of x d (n) after being
filtered and before decimated.
Concerning signal y(t), being produced by the ongoing
processes in region2, suppose that it is reproduced in discrete
form as y d (n) by digital simulation. The following time step is
used:

t y = / YM .
with YM = XM ./L. Hence:
y d (n) = y(nt y ), n=0, 1, 2,
In the same form as with signal x d (n), the spectrum of y d (n)
is periodic with repetition interval 2 YM . This is depicted
graphically in Fig. 24a. The samples of signal y d (n) are
produced at a much lower rate than the one required by
region1 simulation. The rate of y d (n) has to be increased L
times by interpolating L1 samples inbetween every two
consecutive values. The interpolation is performed
conveniently first by inserting L1 zeros between the
consecutive samples, and then by lowpass filtering the new
signal with cutoff frequency YM . The process of
interpolation is represented as follows:
y (n / L ) n = 0, L, 2 L,
( L)y d (n) = y d (n) = d
0 otherwise
The insertion of zeros in y d (n) increases the signal sampling
rate in L times without modifying the information contents of
the signal. This means that now the repetition interval is 2 XM
and the effective bandwidth has been increased to XM . This
new bandwidth is also marked in Fig. 24a. Note that the
spectral components outside the range YM are unwanted
aliasing replicas. Their removal by an ideal LPF with YM cut
off frequency is illustrated in Fig. 24b.
Yd()
YM
YM
2XM
(a)
Yd()
YM
YM
XM
(b)
Fig 24. a) Spectrum of signal y d (n). b) Spectrum of signal y d (n) after the
interpolation and filtering processes.
Finally, Fig. 25 provides a modification to the diagram in
Fig. 22, where the required interfaces have been included to
avoid the aliasing errors that otherwise would be caused by
interconnecting two simulation processes running at different
rates. This interfacing is based on the assumption that the
faster rate is an integer multiple of the slower one.
VII. CONCLUSIONS
Transient analysis of power systems can be done by
methods, either in the time domain or in the frequency domain.
It has been stated here that time domain methods, such as the
EMTP, are the most used in practical applications. Two
important reasons for this are: 1) they are easier to understand
35
and apply to a very wide class of problems, 2) their
requirements of computation resources are moderate. On the
other hand, however, it has been shown here that frequency
domain methods are valuable complements to the time domain
ones. Most times, power system element models and power
system equivalents are synthesized in the frequency domain,
and their frequency domain analysis can attain unmatched
numerical accuracies. Frequency domain methods can thus be
used to verify and finetune time domain models and
procedures.
REGION 1
FastDynamics
tx
x(t)
Low Pass Filter Low Pass Filter
CutOff: YM CutOff: YM
Decimator Interpolator
( L) ( L)
y(t)
REGION 2
SlowDynamics
ty
Fig 25. Interconnection of a tworate process with an interface to avoid
aliasing errors.
The increasing size and complexity of modern power
systems requires every time more powerful tools for analysis,
and power system engineers must be well acquainted now with
the Frequency Domain Analysis (FDA) and the Digital Signal
Processing (DSP) disciplines. In this chapter, first FDA has
been introduced as an extension of Phasor Analysis, as the
latter is more familiar to power engineers. Then, the
differences between continuous time and discrete time FDA
have been examined. A central topic here is the Sampling
Theorem and its implications in the simulation and analysis of
power system transients. Next, the Numerical Laplace
Transform method to analyze power transients has been
explained and its application has been illustrated by two
examples. Finally, the possibility of conducting multirate
transient analysis in time domain has been examined through
frequency domain analysis.
VIII. ACKNOWLEDGMENT
The authors gratefully acknowledge the assistance of
Octavio Ramos Leaos and Efran Cruz Chan at preparing the
examples.
IX. REFERENCES
[1] B. P. Lathi, Linear Systems and Signals, Berkeley Cambridge Press,
1992.
[2] J. G. Proakis and D. G. Manolakis, Digital Signal Processing.
Principles, Algorithms and Applications, Prentice-Hall, 4th Edition,
2007.
[3] H. P. Hsu, Fourier Analysis, Simon &Schuster, Inc., 1970.

[4] S. J. Day, N. Mullineux, and J. R. Reed, Developments in Obtaining
Transient Response using Fourier Transforms. Part I: Gibbs
Phenomena and Fourier Integrals, Int. J. Elect. Engng. Educ., Vol. 3,
pp. 501-506, 1965.
[5] S. J. Day, N. Mullineux, and J. R. Reed, Developments in Obtaining
Transient Response using Fourier Transforms. Part II: Use of the
Modified Fourier Transform, Int. J. Elect. Engng. Educ., Vol. 4, pp.
31-40, 1966.
[6] S. J. Day, M. J. Battisson, N. Mullineux, and J. R. Reed,
Developments in Obtaining Transient Response using Fourier
Transforms. Part III: Global Response, Int. J. Elect. Engng. Educ.,
Vol. 6, pp. 259-265, 1968.
[7] N. Mullineux and J.R. Reed, Developments in Obtaining Transient
Response Using Fourier Transforms: Part IV-Survey of the Theory,
Int. J. Elect. Engng. Educ., Vol. 10, pp. 259-265, 1973.
[8] L. M. Wedepohl and S. E. T. Mohamed, Multiconductor transmission
lines. Theory of Natural Modes and Fourier integral applied to
transient analysis, Proc. IEE, Vol. 116, No. 9, pp. 1553-1563,
September 1969.
[9] L. M. Wedepohl and S. E. T. Mohamed, Transient analysis of
multiconductor transmission lines with special reference to nonlinear
problems, Proc. IEE, Vol. 117, No. 5, pp. 979-988, September 1970.
[10] A. Ametani, The Application of the Fast Fourier Transform to
Electrical transient Phenomena, Int. J. Elect. Engng. Educ., Vol. 10,
pp. 277-287, 1973.
[11] D. J. Wilcox, Numerical Laplace Transformation and Inversion, Int.
J. Elect. Enging. Educ., Vol. 15, pp. 247-265, 1978.
[12] L. M. Wedepohl and C. S. I. Indulkar, Switching overvoltages in short
crossbonded cable systems using the Fourier transform, Proc. IEE,
Vol. 122, No. 11, pp. 1217-1221, November 1975.
[13] L. M. Wedepohl and C. S. I. Indulkar, Switching overvoltages in long
crossbonded cable systems using the Fourier transform, IEEE Trans.
on Power Apparatus and Systems, Vol. PAS-98, No. 4, pp. 1476-1480,
July/Aug 1979.
[14] L. M. Wedepohl, Power System Transients: Errors Incurred in the
Numerical Inversion of the Laplace Transform, Proc. of the Twenty-
Sixth Midwest Symposium on Circuits and Systems, pp. 174-178,
August 1983.
[15] B. Gustavsen, Validation of Frequency-Dependent Line Models,
IEEE Trans. on Power Delivery, Vol. 20, No. 2. April 2005.
[16] W. D. Humpage, K. P. Wong, Electromagnetic transient analysis in
ehv power networks, Proc. IEEE, Vol 70, Nr. 4, pp. 379-402, 1982.
[17] J. L. Naredo V., P. Moreno V., J. L. Guardado Z., and J. Alberto
Gutirrez R., "The Numerical Laplace Transform as a Tool for
Research and Development in Electrical Engineering", (in Spanish),
Proceedings of the II International Conference on Electrical and
Electronics Conference, CIIIEE'98, Aguascalientes, Mexico, Sept 14-
18, 1998.
[18] N.R. Watson and G. D. Irwin, Comparison of root-matching
techniques for electromagnetic transient simulation, IEEE Trans. on
Power Delivery, Volume: 15 , Issue: 2, Pages: 629 634, April 2000.
[19] F. A. Uribe, J. L. Naredo, P. Moreno, and J. L. Guardado,
Electromagnetic Transients in Underground Transmission Systems
Through the Numerical Laplace Transform, International Journal of
Electrical Power and Energy Systems, Elsevier Science LTD, Vol.
24/3, pp 215-221, March 2002.
[20] P. Gmez, P. Moreno, J. L. Naredo, and J. L. Guardado, Frequency
Domain Transient Analysis of Transmission Networks Including Non-
linear Conditions, 2003 IEEE Bologna PowerTech Proceedings, paper
BPT03-115, Bologna, Italy, 23-26 de junio de 2003.
[21] A. Ramirez, P. Gomez, P. Moreno, and A. Gutierrez, Frequency
domain analysis of electromagnetic transients through the numerical
Laplace transforms, IEEE Power Engineering Society General
Meeting, 2004, vol., no., pp.1136-1139 Vol.1, 10-10 June 2004
[22] P. Moreno, P. Gmez, J. L. Naredo, and J. L. Guardado, Frequency
Domain Transient Analysis of Transmission Networks Including Non-
linear Conditions, International Journal of Electrical Power & Energy
Systems, Volume 27, issue 2, pp. 139-146, February 2005.
[23] J. L. Naredo, J. A. Gutierrez, F. A. Uribe, J. L. Guardado, and V. H.
Ortiz, Frequency Domain Methods for Electromagnetic Transient
Analysis, IEEE Power Engineering Society General Meeting, 2007,
vol., no., pp.1-7, 24-28 June 2007
[24] T. Noda and A. Ramirez, z -Transform-Based Methods for
Electromagnetic Transient Simulations, IEEE Trans. on Power
Delivery, vol.22, no.3, pp.1799-1805, July 2007
36
[25] P. Moreno and A. Ramirez, Implementation of the Numerical Laplace
Transform: A Review, IEEE Trans. on Power Delivery, Vol. 23, No. 4,
pp. 2599-2609, October 2008.
[26] P. Gmez and F. A. Uribe, The numerical Laplace transform: An
accurate technique for analyzing electromagnetic transients on power
system devices, International Journal of Electrical Power & Energy
Systems, Volume 31, pp. 116-123, 2009.
[27] J. Vlach and K. Singhal, Computer Methods for Circuit Analysis and
Design, Kluwer Academic Publishers, Second Edition, Chapter 10, pp.
349-390, 2003.
[28] A. Semlyen and F. de Len, Computation of Electro-Magnetic
Transients Using Dual or Multiple Time Steps, IEEE Trans. on Power
Systems, Vol. 8, No. 3, August 1993.
[29] G. Strang and T. Nguyen, Wavelets and Filter Banks, Wellesley
Cambridge Press, 1997.
[30] J. Szczupak, Real Time, Oscillation Free Network Digital
Simulation, IPST99 International Conference on Power System
Transients, IPST99_Paper_032, Budapest, Hungary, June 24-29, 1999.
[31] M. Pller and M. Schiemieg, Exploiting Multiple Time Scale
Properties for Fast Simulation Algorithms, Proceedings of the 13th
Power Systems Computation Conference, Trondheim, Norway, 1999.
[32] J. Szczupak, S. T. Facerolli, and K. B. Guedes, Electrical Network
Simulation By Multirate Parallel Digital Filter Structures,
Proceedins of the IEEE Bologna Power Tech Coference, Bologna, Italy,
June 23-26 th, 2003.
[33] H. Zini and G. Ratta, Multirate Modeling Scheme for
Electromagnetic Transients Calculation, IEEE Trans. on Power
Delivery, Vol 19, No. 1, pp. 240-247, January 2004.
[34] J. L. Naredo, J. L. Guardado, J. A. Gutierrez-Robles, P. Moreno, F. A.
Uribe, V. H. Ortiz, and L. A. Snider, Trends in the frequency domain
analysis of electromagnetic transients, IEEE Power & Energy Society
General Meeting, 2009, PES '09, pp.1-5, 26-30 July 2009.
[35] Atef Morched, Bjorn Gustavsen, Manoocher Tartibi, A Universal
Model for Accurate Calculation of Electromagnetic Transients in
Overhead Lines and Underground Cables, IEEE Transactions on
Power Delivery, Vol. 14, No. 3, pp. 1032-1038, July 1999.
X. BIOGRAPHIES
Jos Luis Naredo (SM) graduated from The University of British Columbia
as M. A. Sc. (1987) and as PhD (1992). He conducted R&D work at The
Electrical Research Institute of Mexico (IIE) in the areas of power system
communications, power system transients and power system protections,
(1978-1985, 1992-1994). Since May 1997 to present, he is full professor of
CinvestavGuadalajara, Mexico. Dr. Naredo currently is spending a
sabbatical year at The Ecole Polytechnique of Montreal.
Jean Mahseredjian (SM) graduated from cole Polytechnique de Montral
with M.A.Sc. (1985) and Ph.D. (1991). From 1987 to 2004 he worked at
IREQ (Hydro-Qubec) on research and development activities related to the
simulation and analysis of electromagnetic transients. In December 2004 he
joined the faculty of electrical engineering at cole Polytechnique de
Montral.
Ilhan Kocar (M) received the B.S. and M. Sc. degrees from METU, Ankara,
Turkey, in 1998 and 2003, respectively, both in Electrical and Electronic
Engineering. He provided custom designed power conversion system
solutions to railway industry as a Project Engineer at Aselsan Electronics Inc.
between 1998 and 2004. In 2009 he received Ph.D. degree from the
department of Electrical Engineering at cole Polytechnique de Montral. Dr.
Kocar is with CYME International, St-Bruno, QC, Canada.
Jos A. Gutirrez Robles (M) received his BSEE and M. Sc. degrees from
Universidad de Guadalajara (UdeG), Mexico, in 1993 and 1998, and his PhD
from CinvestavGuadalajara in 2002. He currently is Professor with the
Department of Mathematics of CUCEIUdeG. His research interests are in
the field of Transient Phenomena in Power Systems.
Juan A. Martinez (M) was born in Barcelona (Spain). He is Profesor Titular
at the Departament d'Enginyeria Elctrica of the Universitat Politcnica de
Catalunya. His teaching and research interests include Transmission and
Distribution, Power System Analysis and EMTP applications.