You are on page 1of 10

IEEE Transactions on Power Delivery, Vol. 10, No.

2, April 1995

62 1

REMOVAL OF DC-OFFSET IN CURRENT WAVEFORMS USING


DIGITAL MIMIC FILTERING

Gabriel Benmouyal
Member

mQ

1800montke Ste-Julie,
Varennes, Quebec, Canada, J3X- 1S 1

AbstrPd: This paper presents a theoretical investigation of


the capability of popular digital filters used in relaying
applications to suppress DC offset in current waveforms over a
broad range of time constants as encountered on real power
systems. First, three mathematical performance indices are
inaodUCed to be able to compare filters on the Same basis. Then
the concept of the FIR type of digital mimic filter is described.
It is shown that, for filters exhibiting a null response at
frequency zero on their frequency response, the addition of a
digital mimic circuit, tuned to the middle of the range of time
constants to be suppressed, will almost completely eliminate
the effect of an exponentially decaying component. Moreover,
this is possible without sacrificing the original response speed,
although it increases susceptibility to noisy components.
I<evwor&: digital relaying, relaying filters, suppression of
DC offset effects, protective relays
It is a Well established fXt that lint? relays have a tendency to
overreach in the presence of DC offset components in faultcurrent waveforms. In some relaying applications, like line
relaying or fault location, the decaying component therefore has
to be removed from these waveforms. A number of techniques
have been put forward to deal with such situations [3-6, 9, 111
but So far no theOfetiCd Study has ever been made (0 evaluate
the e f f ~ t i V ~ nOfe sthese techniques. Nor has any mathematical
performanceindex or figure of merit been devised to characterize
this
Of
The purpose of this paper is to Provide the
missing information.
Mimic circuits have proven to be effective in both
electromechanicaland static technology. This paper shows that
by using the exact digital replica of a mimic circuit, the
exponentiallydecaying component of a current waveform can be
removed for a broad class of filters.

Definitio

rf-

Based on the two test waveforms defied in appendix 19 a fist


performance index PI0 with respect to a filter DC offset
removal capability and a PXtkukU time COnStant 7 iS defined

as f 0 k W S :
a) apply a sinewave step function xl(t) at the input of the
filter and record the magnitude of output waveform yl(t)
xl(t) = sin 2x 60 t
(1)
b) apply, at the input of the same filter, a worst-case waveform
of the form:
x2(t) = cos 2x 60 t - e;
and record the magnitude Y2(t) at the filter output.

(2)

yl(t) corresponds to the condition when there is no DC offset


present, whereas y2(t) corresponds to a worst-case condition
when there exists a maximum DC offset.
me fist performance index of the filter will be computed as the
valueof the integral

PM r 1 =

I:'

[yl(t) - yz(01~dt

(3)

where
represents an arbitrary integral number of cycles. In
die case of the present study, this number has' been chosen
arbitrarily to be three. Obviously, if the defined PI0 is close or
equal to zero, it can be concluded that DC offset has no effect
on the filter output. The first performance index integrates
both the transient and the steady state responses of the filter,
which might be seen as a shortcoming because the response
during the transient state could be considered irrelevant. For
this reason, PI0 will not be used in the present study and a
second performance index PI1 will he defined.
It should be home in mind that before y2(t) settles to the
94
431-7 m D A
recommended and approved steady-state value of 1, it will normally oscillate around this
magnitude. The steady state will therefore be considered
by the IEEE Power System Relaying Committee of the
IEEE Power Engineering Society for presentation a t
attained as soon as y2(t) exceeds one at time TO, whereupon
t h e IEEE/PES 1994 Summer Meeting, San Francisco, CA, integration of the difference starts. PI1 is ConSXlUently defined
July 24 - 28, 1994. Manuscript submitted December
as:
15, 1993; made available for printing May 3, 1994.
NT
(4)

While a value of PI0 close to zero indicates that a filter is


completely immune to DC offset both in the transient and the
0885-8977/95/$04.00 0 1994 IEEE

622

steady state, a similar value of PI1 denotes a filter free of


oscillation in the steady state. A value of PI1 close to zero does
not allow, however, the speed of the filter to be deduced.
Therefore, PI1, as a figure of merit, is useful when comparing
two filters as it provides an indication of the filter providing the
most efficient DC removal capability.
In order to assess the inherent merit of a filter, a third
performance index PI2 will then be defined. It is equal to the
percentage of maximum overshoot in the output magnitude
y2(t) and corresponds to the next mathematical expression:
PI:! ( ~ ) = ( M ~ [ y 2 ( t ) - l ]100
)*
(5)
There exists a direct relation between this third performance
index and the potential overreach that a filtering system could
yield: a PI2 of 15% for example could results in a maximum
overreach of 115% if the first zone of a distance relay was set
at 100% of the line impedance.
In this study, an anti-aliasing filter is not emulated because its
effect on the response to the testing waveforms, besides a
delayed output, is assumed to be negligible as shown in
appendix 3 of the paper.
Variations in DC offset time constant

Although the frequency response of the filter (see Fig. 11)


shows an infinite attenuation at zero frequency, the filter will
nonetheless show some oscillation on the waveform magnitude
at the output. The longer the DC offset time constant, the
greater the filter immunity, however.

,..y
!.: .. -.'.,
.../

I{

.........

.......

..........

._..
-.__
.......c

:h
S

Fig. 1. Time responseof a full-cycle DFT filter to the test


waveform
Table 1 gives the values of the two performance indices, PI1
and PI2, when the time constant r varies from 0.5 to 5 cycles.

Both PI1 and PI2 can be computed for a particular value of the
DC offset time constant r. Since z depends on the power
system configuration at the moment of a fault and, also, on the
location of that fault on a line, this time constant should be
considered as variable. An important consideration to be
addressed, therefore, is the sensitivity of the filter as z varies.
As a figure of merit of this sensitivity, the average of the
performance indices, whether of type 1 or type 2, is computed
over a range of time constants from 0.5 to 5 cycles as in
PI, (0.5)
AV( PI, ) =

PI,( z

r=1

(i= 1 or 2)

(6)

6
ion to a full-cvcle DFT filter
Full-cycle DIT filters are among the most popular in relaying.
For a current waveform i(t) = A sin (cot+@),the fundamentalfrequencycomponents are provided by [ 1,2]

Performance of tvoe I least-sauares-fittine (LSF)

filter
Assuming that the current equation i(t) consists of an
exponentially decaying component, fundamental, and second to
fifth harmonics, as in
i(t) = A e-:

.x
n=S

and the magnitude of the waveform is


A = V - i 5 7

Figure 1 shows the filter's time responses when the test


functions of Appendix 1 are applied at the input of the fullcycle DFT filter. In this figure, as in the following tests, the
filter has been subjected to three testing waveforms:
a) a sinewave corresponding to Eq. 1.
b) two cosine waveforms with maximum DC offset as in Eq. 2.
Two extreme values of decay rates for the DC components have
been used: 0.5 and 5 cycles at rated frequency .
In this example, as for all filters referred to in this paper, a rate
of 12 sampleskycle is used. In Eys. 7&8, y is therefore equal
to 30'.

Bn sin ( n u t + 8" )

(10)
then the equations for this LSF filter can be derived by
working out the least-squares solution of a set of equations
representing the value of i(t) at different samples of the chosen
data window.
For a data window equal to a full cycle, Fig. 2 [see Ref.1, Fig.
4.10 and Ref. 2, Fig. 3.171 represents the well known
frequency response of the LSF filter.
Figure 3 shows the time responses of the filter when the above
specified three test functions are applied at the input.
Analysis of Fig. 3 indicates that, contrary to common belief,
the DC removal capability of this type of filter is not very
g c d , less performing in fact than the full-cycle Fourier filter.
This could have been predicted by analyzing the filter's
frequency response, which shows non-zero attenuation at zero
frequency.
From a practical point of view, it turns out that this filter does
not remove the exponentially decaying component modeled in
the original curve equation.
n=l

623

Fig. 2. Frequency response of LSF filter.

Fig. 4. Frequency response of type I1 LSF filter.

1.4

1-

.'

-lBYCNJhmtMBmPFuNmoN
- .. ~EIPWSEm "ZSIZSFWXON.

QIOTLE WSET

p e v o N S E ~ c o m ( B S ~ F l u " O N + I C Y ~ O ~ E T

RKWONSE m SINE SKPPUNCnON


..._FSSWNSE TO COSINE SIPP+U"ON

1O.bCYCLE OFRET

........

nEx"e m cosm SIEP.PV"ON

X Y ~ Enmm

I
0 -

SAMPLES
30

Fig. 3. Time responses of type I LSF filter.


Table 2 gives the values of the two performance indices as T
varies from 0.5 to 5 cycles.

.........................
10

30

SAMPLES

1595 x 1W6
362 x

1.0032

Performance of tvDe 11 LSF filter

If the exponential in Eq. 10 is expanded using the two first


terms of a Taylor series, a new filter [41 can be derived which
corresponds to the fitting of the next curve:

In this equation, the previously exponentially decaying DC


component becomes the sum of a constant and a linearly
decayingramp.
The frequency response of this new filter is shown in Fig. 4.
This time, the response at DC frequency is zero but a general
amplification of the higher-frequency components has taken
Place.
Analysis of Fig. 5, sbowing the filter's time responses to test
functions, indicates that this filter practically eliminates the DC
offset component over the range of time constants tested.
Table 3 gives the values of the two performance indices for z
varying from 0.5 to 5 cycles.

..

60

2.80
1.81

0.74

115 x

0.39

46 x loe6

0.24

22 x 10-6
32.86

Table 3
2800 x

Fig. 5. Time responses to test functions of modified LSF filter.

Table 2. PI of LSF filter of t

40

0.16

Performance of half-cvcle window DFT filter


The half-cycle Fourier filters introduced by Phadke et al. [5]
offer the advantage that they reduce the time response of the
full-cycle Fourier filter by a factor of two. The main
shortcoming of this filter is its lack of immunity to DC offset
components. This is illustrated in Fig. 6, which shows the
filter's time responses to the test functions, and also in Table 4
representing the performance indices.
Performance of cosine filters
Cosine filters are derived from full-cycle Fourier filters [l I].
The components of the fundamental are obtained by processing
the waveform twice using Eq. 6. First, the waveform is
processed normally; then it is delayed by one quarter of a cycle
to obtain an orthogonal wavefonn and again processed. The
frequency response of this filter is shown in Fig. 7.
Figure 8 shows the filter's time responses to the test
waveforms. From a practical point of view, cosine filters
clearly have an outstanding capability to remove DC offset
influence. Furthermore, contrary to the type 2 LSF filters, this

624
is not at the expense of the frequency response, as may be seen
also from the performance indices for this filter in Table 5 .

1.2

........ ...... - ..-........

....

RESPONSE TO SINE STEP-FIINcI1ON


RESPONSE mCOSME SlCRI"ON
+ 0 . S - C Y a E OFFSET
RESWNSE m COSINE STEF-RNCTION + SCYCLE OFFSET

.......

" " k " ' ' b " " ~ " ' '

SAMPLES

. . . . . . . . . , . . . . , . . . . . . . . , . . . .I

0 ~ :
0

10

SO

40

SAMPLES
30

Fig. 8. Time responses of cosine filter to test waveforms.

M)

Fig. 6. Time responses of half-cycle DFT filter to test


waveforms.
Table 4. PI of half-cycle DFT filter.

'5 (in

cycles)

PI^

1
1

0.9824
3.3994

I
I

I
1

7.7096

14.3624

49.7
79.0

100.0

1
I

108.1
112.6
115.3

10.7462
12.8496

PI^ (%o)

I
1
1

Figure 9 shows the time responses of the Kalman filter phasor


magnitude estimator. It is clear from this figure that this type
of filter will remove the effect of the DC offset almost entirely
when its time constant corresponds to the one modeled in the
state transition matrix (1 cycle). If this is not the case,
however, the filter will be sensitivite to variations in the offset
tune constant.

1.2

..... P E S Y O N S E c
~ os1~~sm.mc1
+ ICYCZEOFR~
10~

........ KESfONSETOC O S N STEP-FUNCIIONt J-CYCZEOFF3FT

Fig. 7. Frequency response of cosine filter.


ce of
Kalman filters were introduced by Girgis as candidates for
voltage and current phasor estimation [6,71. In third-order
Kalman filters, the exponentially decreasing DC offset is
modeled for a particular time constant. IJsing the notations
given in [lo], the test functions were applied at the input of a
Kahnan filter ch,uacterized by 9 s . 12-14

0.866 -0.5 0
0.5 0.866 0
0

1 3

H = [ 1.0 0 1.01.
0.0001

Q=[

. . .

X$:']q

0.92

0
0.0001

0.0001

X3

%{i]

R=[0.0001]

,Pi{:

:]

0 0 1

(12)

(13)
(14)

. . . .Ib. . . . . . . . . . . . . . . . . . . .
-S

Fig. 9. Time responses to test waveforms of Kalman filter.


The same phenomenon is reflected by the values of the
performance indices presented in Table 6.

625

.. . .

DC

offset " ~ v a l bv "lc clrclut


A new digital " i c filter will now be defined which allows
the removal of harmful oscillations that occur when DC offset
is present at some ftlter inputs.
If a current waveform has to pass through a mimic circuit
consisting of a resistance in series with an inductance or
impedanceof the form

K (1 + ~ 1 1 )
(15)
then the exponentially decaying component at the output will
vanish, provided its time constant is equal to 11 (as proven in
Appendix 2). In this equation, 71 is expressed in number of
samples.
The sum of a gain and a differentiator circuit is represented by
Eq. 15. The differenriator circuit, represented in the Laplace
transform by s. can be emuhted digitally by the following FIR
filter:
(16)
1 2-1
Introducing Eq. 16 into Eq. 15, we eventually obtain
K [(l +TI) - 71 i' ]
(13

Here the gain K has to be set in such a way that, at rated


frequency, the filter gain will be 1. At 60 Hz. the gain is given
by the following expression
Gain (60Hz)=

I K [(l +TI) -71 e-jCoT1 I = 1

(18)

Solving this equation for K, we obtain


K2=
1
[(l+.rl) - T~*COS/2*x*60_)lz+[~1*sin1)-(2
F-P
F-P
(19)
where Fsamp is the sampling frequency.
Figure 10 represents the frequency response of the mimic filter
with a time constant of 2 cycles, which is in the middle of the
range studied in this paper, 0.5 to 5 cycles. This is obviously a
high-pass filter and its use in combination with any other type
of filter should have the same impact as the one observed with
the type I1 LSF filter, i.e. amplification of the higher-frequency
components.
Let us now analyze the influence of this mimic filter on the
performance indices of two filters that show sensitivity to DC
offser full- and half-cycle DFT filters.
t t o a full-cvcle DFT

film
Prior to passing through the Fourier filter, the current signal is
processed by the mimic filter represented by Eq.17. The overall
frequency response of the filtering system is shown in Fig. 11.
It should be bome in mind that an anti-aliasing filter exists in
all digital relays. This filter is an ideal low-pass filter with cutoff frequency at half the sampling frequency. Therefore the
overall frequency response as shown in Fig. 11, should be
considered, for practical purposes, as not affected by the antialiasing filter
Analysis of Fig. 12 shows that the effect of the mimic filter is
to substantially reduce, if not entirely eliminate, the effect of
the decaying DC component over the entire range of time
constants tested. This is reflected in Table 7, where the
performance index figures indicate a substantial, if not
outstanding, improvement over the full range of time constants.

This improved performance is therefore obrained at tbe expense


of a slight increase in the sensitivity to off-nominal higherfrequency components because of the amplification in the
overall frequency response.

Fig. 10. Frequency response of mimic filter with time constant


of 2 cycles.
I2

I
0.8

0.6

0.4
0.2
0

Fig. 11. Frequency response of a combined mimic and fullcycle DFF filter.

..I ,

I
.I

_ . . .i b . . .

.zb'

3b

'

SAMPLES

'

4b

*'.'.

Fig. 12. Time response of a full-cycle DFT and 2-cycle mimic


filter combination .

626

Since the cosine filter showed a very good natural response to

1.2

DC offset, it is interesting to compare it with the others, as in


Fig. 13. From this comparison of the time responses of the
Fourier, Fourier+mimic, and cosine filters to a sine step
function, it can be seen that the Fourier+mimic combination is
slightly slower than a Fourier filter alone hut somewhat faster
than the cosine filter.

.-

RUiPDNSE

To COElNE P r r P N N C C n O N

N m W

RESPONSE TO COSINE

Q.5CYCLE 0
"

ICICLE O F K C

RFSPONSE TO COSWE ETEP-NHCIION 1 I C I C L E OFFSET

-_

DFT + MIMIC FILTER

.......

Fig. 15. Time responses of half-cycle DFT filter to test


waveforms.

DfTFlLIER

Ta

.... COSINEFILTER

10
SAMPLES

I5

tion

20

Fig. 13. Comparison of DFT and cosine filter responses to step


functions.
Application of a mimic circuit to a half-cvcle DFT

filter
The frequency response for the same mimic filter used in
combination with the half-cycle Fourier filter is shown in Fig.
14.
When subjected to test waveforms, the half-cycle+mimic filter
combination proves to be very effective in reducing the effects
of DC offset components. This may be seen in Fig. 15, which
shows that DC offset removal is complete when the time
constant is equal to that implemented in the mimic filter.
Table 8 gives the performance indices of the new combination,
from which the same conclusion can be drawn.

Modification of the state transition matrix proved to be a more


effective way of improving this filter. The state transition
matrix was therefore expanded to include an additional constant
term and the previous time constant was changed to 0.5 cycles,
as seen in Eqs. 20-22.
0.866 -0.5

0.846 0
0

x3

x4

H = [ 1.0 0 1.0 1.01.

1
Fig. 14. Frequency response of a half-cycle D l T filter
combined with a mimic filter.
It should be borne in mind that further reduction of DC offset
effects could be achieved by cascading more than one mimic
circuit tuned to at least two time constrmts. However, this
would increase the high-pass filtering effect, which is not
desirable.
With further improvement of the performance indices of the
type I1 LSF and cosine filters, a mimic filter could equally be
applied to them but both are already performing their task of
removing the J
X offset quite adequately.
Improvement of Kalman filter response
After a few tests, application of a mimic circuit prior to
entering a Kalman filter did not improve the performance.

o
o
o

10.~

o
o

lo3

10.3

R=[0.0001]

(21)

0 0 1 0

Another means of improvement is to further modify the process


noise covariance matrix by increasing the terms related to the
DC offset, as in Eq. 22. The effects on the time responses are
illustrated in Fig. 16.

Fig. 16. Time response of Kalman filter to test waveforms.

627

On the contrary, in filters where the decaying current was


modeled by a constant plus a decaying ramp, the performance is
significantly improved.
c) Third-order Kalman filters prove to be
sensitive to
variations of the DC offset time constant. It has been shown
that the addition of a constant term, corresponding to the
modeling of a DC component in the signal significantly
improves the performance.
AcknowledPmentz

Summarv of results
Table 10 summarizes the results presented above and gives the
average values of the two performance indices for the filters
studied in this paper.
Considering only the DC offset removal capability, the
mimic/full-cycle DFT filter combination offers the type 2
performance index closest to zero over the defined range of time
constants. The type I1 LSF and improved Kalman perform
almost as well.
It should be borne in mind that adding a mimic filter to the
half-cycle filter substantially improves the original.

The author thanks Lesley Kelley-Regnier of Hydro-Quebec for


careful editing of the text.
ADDendix 1
..
or t e s u wave-

Consider the network of Fig. A1 representing a voltage source


supplying a power system modeled by an equivalent resistance
in series with an inductance. In this model, R and L represent
the impedance of the source together with the one of the line.
Since the impedance of the source varies with the network
topology, the L/R ratio or the impedance time constant has to
be considered as a variable.

Table 10. PI of filters.

Fig. A l . Equivalent circuit of a power system during a fault.


If the voltage source is of the form
v (t) = Vpeak sin ( at + 8 )
(AI)
the current in the circuit, when the switch is closed, will be
i (t) = Ipeak [ sin ( at + 8 - 8,)

- sin (e - ez)exp (-

X/R

)1

with

8, = atan ( X / R ) = atan ( a L / R )
Conclusion
This paper has presented an analysis of the performance of some
popular relaying digital filters with respect to their DC removal
capability. In light of this analysis, the following points can be

made:
a) The addition of a mimic digital filter when processing a
current waveform prior to entering a full-cycle D I T filter
completely suppresses the effect of an exponentially decaying
component at the output over a broad range of time constants
(0.5 to S cycles and more). The perform'ance of the DFT filter
will then surpass that of most known filters. The price to be
paid is reduced immunity to off-nominal frequencies or noise
because of the amplification of high-frequency components.
AS illustrated by the case of the half-cycle Fourier filter,
application of the same mimic filter turns out to be very
effective for filters displaying non-zero attenuation on their
zero-frequencyresponses .
b) In least-squares-fitting filters, where the DC offset was
modeled as a decreasing exponential, the performance is very
poor.

A null DC offset will occur when


( n= 0,1,2,3 etc...)
0 - eZ = n x
arid the current waveform becomes

i (t) = Ipeak sin ( at )


A maximum value of DC offset will occur when

6 - e, = n

( n= 1, 3 ,S etc...)
2
with the current waveform equal to

i (t) = Ipeak [ cos ( at ) - exp (-

)I

L/R
The test waveforms chosen for this paper were Eq. AS for
current corresponding to a null DC offset and Eq. A7
corresponding to a maximum DC offset.
Aupendix 2
PrinciDle of a mimic circuit

.
References

F l
vo

(t 1
Fig. A2. Principle of a mimic circuit.
Consider the circuit of Fig. A2 where a current source supplies
a resistance in series with an inductance. The voltage across the
combination is provided by
Vo (s) = ( s L + R) I (s)
If the current has a DC offset of the form
I(t) = e x p ( - L

(A8)
(A9)

then the voltage due to the D(:offset term alone will be


L-I [ I (s)] =L-' [ s L + R ]
(A101
s + 1 / T
If r = L/R, then the inverse Laplace transfonn of a constant is a
unit pulse uo (t) at time t = 0 as in
L-I [ I (s)] = 1U0 (1)

(A11)

In practical terms, transformation of the exponentially decaying


current into a unit pulse voltage means that the DC offset in
the measured output voltage has vanished.
Awendix J

E-tg
T
m

.-alia.
. .

filter
The study conducted in this paper has neglected the effect of the
anti-aliasing filter. It could be shown both dieoretically and by
simulation that this effect is negligible with respect to the
results presented here. As an example, Fig. A3 shows the
response of a full-cycle DFT filter when suhjected to a test
waveform corresponding to Eq. 2 with a DC component
decaying rate of one cycle, with and without a simulated antialiasing filter. In this case, the anti-aliasing filter is a fourth
order Butterworth with a cut-off frequency of 360 Hz. It can be
seen that the effect of the anti-aliasing filter is to delay the
output magnitude by one sample without significantly
changing the output.

Fig. A3. Effect of an anti-aliasing filter.

[I] A.G. Phadke, J.S. Thorp, COMPUTER RELAYING FOR


POWER SYSTEMS, Research Study Press Ltd., John Wiley &
Sons Inc., 1988.
[2] M.S. Sachdev, Coordinator, MICROPROCESSOR RELAYS
A N D PROTECTION SYSTEMS, IEEE Tutorial Course,
88EH0269-1-PWR.
[3] M.S. Sachdev, M.A. Baribeau, "A Digital Computer Relay
for Impedance Protection of Transmission Lines," Trans. of the
Engineering and Operating Division, Canadian Electrical
Association, Vol. 18, Part 3, No. 79-SP-158, 1979, pp. 1-5.
[4] M.S. Sachdev, M.A. Baribeau, "A New Algorithm for
Digital Impedance Relays," IEEE PAS-98, Vol. 2, No. 4, Dec.
1979, pp. 253-260.
[SI A.G. Phadke, T. Hlibka, M. Ibrahin, M. G. Admiak, "A
Microprocessor-Based Symmetrical Component Distance
Relay," Proceedings of PICA, May 1979, Cleveland.
[6] A.A. Girgis, R. Grover Brown, "Application of Kalman
Filtering in Computer Relaying," IEEE Trans. PAS-100, No.
7, July 1981, pp. 3387-3397.
[7] A.A. Girgis, R. Grover Brown, "A New Kalman FilteringBased Digital Distance Relay," IEEE Trans. PAS-101, No. 9,
September 1982, pp. 3471-3480.
[8] A.A. Girgis, R. Grover Brown, "Modeling of Fault-induced
Noise Signals for Computer Relaying Applications," IEEE
Trans. PAS-102, No. 9, September 1083, pp. 2841.
[ 9 ] A.G. Phadke, T. Hlibka, M. Ibrahin, "Fundamental Basis
tor Distance Relaying with Symmetrical Components", IEEE
Trans. PAS-96, No. 2, March/April 1977, pp. 635-646.
[ 101. G. Benmouyal, "Frequency-Domain Characterization of
Kalman Filters as Applied to Power System Protection," IEEE
Trans. on Power Delivery, July 1992, Vol. 7, pp. 1129-1138.
[ l l ] E. 0. Schweitzer 111, D. Hou, "Filtering for Protective
Relays", 19th Annual Western Protective Relay Conference,
Spokane, Wa\hington, 1992.
B i o c w
Gahriel Benmauyd received his
B.A.Sc. in Electrical Engineering
and his M.A.Sc. in Control
Engineering from Ecole
Polytechnique, Universitk de
Montreal, Canada in 1968 and
1970 respectively. In 1969, he
joined Hydro-Quebec as an
instrumentation and control
specialist. He worked on different
projects in the field of substation
control systems and dispatching
centres.
Since 1978, he has been working at IREQ where his main
field of activity is the application of microprocessors and digital
techniques to substation and generating-station control and
protection (ystems.
tfe is a registered professional engineer in the Province of
Quebec, is an IEEE member and has served on the Power
System Relaying Committee since May 1989.

629

Discussion
David G. Hart and Damir Novosel, ABB-TTI,
Raleigh, NC: The author has presented an interesting
paper which attempts to define indices which can be used in
evaluating phasor estimation algorithm. The author explores
the complex topic and makes several conclusion. The
following remarks are made regarding the paper:
1. The author defines yl(t) and y2(t) as the output of the
filter. Since the filter is estimating the frequency components,
how does the author define the output as a function of time ?
How are frequency domain terms integrated in the time
domain in equations 3 and 4 ?

2. When a waveform is sampled, the sampling frequency


must be a minimum of twice the highest frequency in the
waveform. Anti-aliasing filters are required to attenuate the
higher frequency components which may be present in the
waveform. However, if the author is testing the waveform
with, for example, a 1 Hertz signal, for one 60 Hertz cycle,
this will only account for 1/60th of one cycle of the input
waveform. Which part of the input waveform the algorithm
processes is dependent on the phase angle of the input and
will affect the frequency response plots in the paper. Did the
author consider this in the present study ? How would the
inclusion of a 1 Hertz waveform in addition to a 60 Hertz
input impact the 60 Hertz estimate ?

3. The authors compares several estimation techniques,


including the DFT and least squares. For the ideal noiseless
inputs defined, the DFT is attempting to estimate two
unknowns using 12 samples. The least squares filter, LSF 2,
is trying to estimate twelve unknowns ( 5 harmonics +
exponential terms) using twelve samples. The DFT filter can
be considered an over redundant set of equations while the
LSF 2 has one equation for each unknown. Did the author
consider either a higher sampling rate for the LSF 2
algorithm or reducing the number of unknowns (harmonics)
before making a comparison? Is it possible higher sampling
rates could improve one algorithm while not offering much
benefit to another ? other least squares implementations ?
4. The author states the proposed method does not allow, "the
speed of the filter to be deduced". It should be noted that this
is an important parameter of interest for the application of
phasor estimation for relaying and should be considered in an
evaluation.

5 . Last, there are several different means of implementing


particular algorithms such as least squares and Kalman
filtering. How one implements the filter in terms of
unknowns, formulation, and parameters can significantly alter
results. Did the author consider other least squares
implementations ? The author evaluate the Kalman filter in
the paper. Kalman filtering is a recursive optimal estimation
algorithm which requires statistical parameters of the signal
to be processed. As a result, the output of the filter is only
good as the "fit" of the statistical model of the signal being
processed. Will the author comment on the selection of these
values for both Kalman filters evaluated ?
6. Equation 2 is incorrect (exponential term should be
negative). The correct equation is included in the appendix.
In summary, the author has presented an interesting paper,
which tries to establish some means of evaluating the

performance of various phasor estimation algorithm. To the


discussers, it appears difficult to rank the algorithms using
only the indices outlined in the paper. Does the author plan
additional work in this area ?
The author comments on the above remarks would be
appreciated.
Manuscript received August 22, 1994.

G. Benmouyal : The author would like first to thank the


discussers for their very stimulating remarks and questions.
Each paragraph will be addressed separately.
Answer to auestion 1: A phasor estimation algorithm can be
evaluated from different perspectives:
-immunity to harmonics
-immunity to off-nominal components
-immunity to frequency shifts
-immunity to noise
--immunity to dc-offset
-speed and response time etc.
Because a filter cannot be outstanding with respect to all these
criteria, a filtering system is generally a compromise.
Testing of a filter can be accomplished by using waveforms
obtained from different sources like a real network, EMTP or
a Transient Network Analyzer. The problem with this kind of
testing is that in order for a parameter to be tested it has to be
present in the waveform: the immunity to frequency shifts
cannot be tested with waveforms exhibiting a fixed
frequency, immunity to harmonics cannot be tested with a
linear network. A filter can, however, be tested independently
with respect to each of the listed features. This is precisely
what has been done in the paper: a number of popular filters
have been tested with respect to their immunitv to dc-offset
only.Therefore the performance indices defined in the paper
are only good with respect to dc-offset immunity. Should we
have other frequency components, different performance
indices should be used.
The performance indices defined in the paper are based on the
premise that the testing functions contain only noise-free 60Hertz sine components and exponentiallydecreasing dc-offset.
Should the waveform contain additional
noxious
components, the performance should degrade, not improve.
y2(t) is the magnitude time response when there is a dc-offset
and yl(t) is the response when there is no dc-offset. If the
filter is completely immune to dc-offset, the difference
between the two responses should be zero. It tums out that
these indices are sound because for filters more immune to dcoffset effects, PI0 and PI1 are very small values.
Answer to auestion 2: Refer to the preliminary statement in
answer to question 1. A 1 Hertz component is an off-nominal
frequency signal and the purpose of the study was not to
answer this matter which, nonetheless is important.
Answer to auestion 3: It should be borne in mind that so
called DFT filters are in fact least squares fitting filters where
only fundamental and no exponential are modeled. Higher
sampling rates do not improve the dc-offset removal
capability of an LSF 2 filter. Reducing the number of
harmonics does not change the basic principle stated in the
paper: when the dc-offset is modeled as an exponential, the
capability of the LSF 2 filter to remove the dc-offset effect is
very limited.

630

As an example, a filter (identified LSF A) has been


synthesized using a least squares technique based on a
waveform consisting of a fundamental and an exponential as
in:
i(t) = A e-k + B ( sin cot + e )

(AI)

A second filter (LSF B) has been synthesized based this time


on the dc-offset modeled as a constant and a ramp as in:
i(t)=A( 1 - t ) + B ( s i n o t + O )

af(

In both cases, the number of samples has been increased to 16


sampleskycle and the filter data window is 8 samples (or halfcycle). The time constant is one cycle.
Fig. A1 shows the time responses of both filters to a test
function corresponding to equation 2 of the paper with an
incoming dc-offset of one cycle.
Fig. A2 shows the frequency response of both filters. One can
infer that changing the sampling frequency or the data window
do not alter the conclusions of the paper: the filter where the
dc-offset has been modeled as a constant and a ramp has a
better capability in removing the dc-offset effect. However its
frequency response is less immune to higher frequencies
components.
1.2

Answer to auestion 5: In Kalman filters, two kind of signals


are modeled : deterministic and noise. The exponentially
decaying component is a deterministic component normally
modeled with a fixed time constant in the state transition
matrix. In that perspective, modeling the statistical
parameters of the signal is a matter which is independent
from the immunity of the filter to dc-offset effect. The same
principle apply to harmonics and off-nominal components
that have to be modeled in the state transition matrix.
It turns out that a Kalman filter is sensitive to variations of
the dc-offset time constant. The paper has shown that a
Kalman filter will only be outstanding in removing a dcoffset which time constant is the same as the one modeled
in the state transition matrix. For that reason, the state
transition matrix has to be modified to account for these
variations. A mimic circuit, installed prior to the filter
processing, provides poor results. This matter is fully
deterministicand independent from noise statistics.
Answer to auestion 6: a minus sign is indeed missing in
equation 2.
As an answer to the last comment, Table 10 is definitely not
the one to be used if one is interested to provide a rank to
filters in general. It refers only to the performance of filters
with respect to a specific parameter: immunity to dc-offset.

, , , , l , , , , l , , , . I , , . ,

rl

Answer to auestion 4: The speed of a filter is the paramount


parameter. It was not however the focus of attention in the
present study.

_....__LSF B

This paper is underlining a testing philosophy: use particular


testing waveforms depending upon which filters feature one
is interested in. The overall performance of a filter could be
assessed by summing a number of weighted indices. We
intend to foster this approach.
Furthermore if a manufacturer was providing in its relays
specifications, the value of one of the indices used in this
paper, the user would get a clear picture of the performance
without the need for further time consuming and expensive
testing.
Manuscript received October 24, 1994.

Fig. A1 Time responses of filters LSF A & B.

6 I

\
\

f
5

a
........ LSF B

100

200

300

400

FREQLENCY

Fig. A2 Frequency responses of filters LSF A & B.

500

You might also like