98 IEEE Transactions on Power Delivery, Vol. 14, No.

1, January 1999

Frequency Responses of Phasor-based Microprocessor Relaying Algorithms

Liancheng Wang, Senior Member, IEEE

ABB Power Automation & Protection Division
Coral Springs, FL 33065

&STFACT - The magnitude and phase estimate algorithms in 2. FREQUENCY RESPONSE OF LINEARTIME-INVARIANTSYSTEMS
the phasor-based microprocessor relaying are nonlinear. The The fiequency response of a LTI system represents the
characteristics of the nonlinear relay algorithms w m t be analyzed characteristics of the system under sinusoidal inputs The steady-
using the frequency response concept defied for linear state output of a LTI system to a sinusoid is a sinusoid of the same
theminvariant (Ln)Systems. The term ''frequency response" frequency as its input but with a different amplitude and phase.
needs to be clarified when it is used to characterize nodinear relay ne of or attenuation of the sinusoid
algorithms. The Paper clarifies and then investigates the by the LTI system is defined as magnitude
magnitude responses of deterministic relay algorithms and l&lman characteristics of phase shift of the sinusoid by the LTI system is
filter relay algorithms. The deterministic defied as phase response, and both magnitude response and phase
studled in the paper Mann-Momson$ prOdar, cyc1e response are referred to as frequency response of the system [4].
DFT, half cyc1e DFT, Cosine and Least square. In the The frequency response of a LTI system is completely determined
Kalman filter relay algorithms, two-state and three-state Kalman by its transfer function.
filter models are considered. Finally, the effect of the signal Figure 1 shows a LTI system whose transfer function is
statistic property on the frequency responses of the fGdrnan filter represented by G(z), The input and output of the system is
relay algorithms is studied.
represented by x(k) and Hk), respectively.
KEYWORDS. Microprocessor protective relays, DFT, Least Square,
Kalman Filtering, frequency response, W filters, FIR filters. x(k)
___)

1. INTRODUCTION
Figure 1 Linear time-invariant system
The frequency responses of microprocessor relay algorithms
have been studied by many researchers. The frequency responses We consider a sinusoidal input described by
of some well-known relaying algorithms have been adopted in
many publications and widely used as the fundamental for the x ( k ) = Acos(wkT+@ (1)
understanding of microprocessor relaying algorithms [ 1-31,
However, a careful examination of the frequency responses reveals where A represents its magnitude, w represents its angular
that they are heuristically derived based on the frequency response frequency, (6 represents its initial phase angle, T represents the
concept of linear time-invariant (LTI) systems. Thus they do not sampling period, and k represents the kth sample.
accurately reflect the frequency domain characteristics of the output of the system is determined by
nonlinear magnitude estimate algorithms of the microprocessor
relays Therefore, an appropriate method needs to be developed to y ( k )= A l G ( e ' O ~ ) ( c o s ( ~ ~ + B + L G ( e ' " ) )( 2 )
study the characteristics of the nonlinear relaying algorithms in the
frequency domain. Obviously, the linear system amplified or attenuated the input
In Paper7 a generic model for the Phasor-based signal by a factor of lG(eJoT)land shifted the phase by an angle of
microprocessor relaying algorithms is proposed Based on the
model, the frequency response concept of nonlinear relaying LG(e'OT). The variations of IG(e'")l and LG(eJoT) with
lS The paper is Organizedas
Section respect to are the magnitude response and phase response,
2 briefly summarizes the frequency response of linear time-
invariant systems. Section 3 clarifies the concept of frequency respectively
response in analyzing the nonlinear relaying algorithms Section 4 We note that the frequency response is a concept introduced
studies the magnitude responses of the deterministic relaying for LTI systems. When dealing with the frequency domain
algorithms including Ma-Morrison, Prodar, full cycle DFT, half characteristics of nonlinear systems as described by nonlinear
cycle DFT, Cosine filter, and Least Square (LS). Section 3 studies
relaying algorithms, we must be very careful to avoid any
the magnitude responses of Kalman filter (KF) relaying confusions which may be caused by using the same the terminology
algorithms. Finally, section 6 presents conclusions. but with different meanings.

PE-318-PWRD-0-12-1997 A paper recommended and approved by

3. PHASOR-BASED MICROPROCESSOR RELAYINGALGORITHMS
the IEEE Power System Relaying Committee of the IEEE Power All phasor-based microprocessor relaying algorithms may be
Engineering Society for publication in the IEEE Transactions on Power represented by the generic model as shown in Figure 2 The figure
Delivery. Manuscript submitted July 14, 1997, made available for includes two linear filters G,(z) and G,(z) for eliminating
printing December 12, 1997
undesirable harmonics and providing necessary phase shift at the
fundamental frequency. The outputs of these two filters are further
processed to extract the magnitude and phase angle of the
fundamental frequency signal. Figure 2a is the block diagram for
-
0885-8977/99/$10.00 0 1997 IEEE
 99

calculating the magnitude of the input and Figure 2b is the block It can be seen from (4) that, generally, the magnitude estimate is a
diagram for calculating the phase angle. The symbols in Figure 2 function of the sample k. Thus, even for a pure sinusoidal input
are explained as the follows. signal which contains no harmonics, the output of the algorithm is
x(k) represents the input signal at the kth sample, not a constant unless at frequencies where G,(eJmT) and G,(e'")
''NI represents taking the square of,
satisfy a specific relationship. The variation range of the
11 -/, represents taking the square root of, magnitude estimate may be found by letting the sinusoidal
" t I' represents division,
function in (4) take its maximum and minimum values,
"atan" represents taking the inverse tangent of, respectively. Letting cos(2wTk + 0) in (4) be equal to +1, we
Xk and & represent magnitude estimate and phase estimate. have the maximum estimate of the input signal amplitude
According to Figure 2, both magnitude and phase estimate - 1
1

algorithms include nonlinear operations on system states; so Figure - e2 ,( )e) ~~

m a x ( A k ) = - ( ~ l z + ~ , f +JG: + ~ : + 2 ~ : ~ : ~ ~ ~ .~(6)
2a and Figure 2b represent nonlinear systems. Though the Jz
frequency responses of the linear filters in a nonlinear relay Letting cos(2wTk + e) in (4) be equal to -1, we have the minimum
algorithm can be analyzed based on the method described in estimate of the input signal amplitude
Section 2 , they cannot tell the behavior of the relaying algorithm's
- 1
min(A,)=-(G12+G,Z-,/G: + G , 4 + 2 G : G : c o ~ 2 ( 8 ~ - B ~ ) ) ~(7)
.A
response to sinusoids exactly. To study the so-called "frequency
responses" of the nonlinear relaying algorithms, we start our Jz
analysis by considering the sinusoidal signal described by (1) Therefore, the magnitude estimate from the nonlinear algorithm
x ( k ) = Acos(wTk+() represented by Figure 2 is within a band whose upper and lower
boundaries are defined by ( 6 ) and (7), respectively. The variation
of this band with respect to the input signal frequencies reflect the
x(k)+ppq-&t+ characteristics of the relaying algorithm in the frequency domain

*-
We will use the normalized variation band to study the frequency
responses of the relaying algorithms. The upper boundary of the
normalized band is
G2(d
gk) A +
(a) magnitude estimate
The lower bound of the normalized band is

<-qTI
$4 k'

Thus, the frequency responses of the nonlinear relaying

algorithms may be studied by finding the upper and lower
Xz(4
G2@) boundaries of the normalized variation band. On the other hand,
once the normalized magnitude variation band of a relaying
(b) phase angle estimate algorithm is known, the magnitude estimate offered by the
Figure 2 phasor-based relaying algorithm algorithm is specijkd for any signal frequencies. The closeness of
the upper boundary to zero at harmonic frequencies indicates the
Based on (2), the steady-state outputs of GI@) and G2(z) algorithm's capability in rejecting harmonics. Obviously, the ideal
to x(k) are described as relaying algorithm with respect to harmonics rejection capability
x,(k)=AGl~~(&T+#+6,) would be the one whose upper and lower boundaries are equal to 1
at the fundamental frequency and 0 at all other frequencies. But
x,(k)= A G , ~ ~ ( ~ T + ( + ~ , ) the algorithm with this frequency characteristics would take an
where infinite time to reach its steady state.
G, = IGl(e")l 4 = LG,(eJoT) If the angle difference between the phase shift of linear
filters G,(z)and G 2 ( z ) is 90 degrees, 0, -0, =90", the upper
G2 = IG2(eJmT)l 0, = LG,(e'")
and lower boundaries can be simplified as
Thus, we the have magnitude estimate of the input sinusoidal ma('_ _)-- max(G,,G , )
_
signal as A
= (G,'A2 cos2(uTk ( + + 6, )+ G,'A2 COS' (mTk + 4 + O,))? (3) and

Equation (3) may be rearranged as --

min('k ) - min(G, ,G, )
A
2 --.!-(GIZ+G,' +,,/G: +G~+2G:Gico~2(8,-8~) where max(G,,G,) represents taking the maximum of GI and
"A 1
G, , and min(G,,G,) represents taking the minimum of G, and
. cos( 2wTk + e)) .A (4)
G, . Equations (10) and (1 1) indicate that the magnitude variation
where band is formed by the magnitude responses of Gl(z) and G,(z)
0 = tan-' G: sin 2( (+ 0,) + G: sin 2(( + 0,) (5)
when the phase shift of the two filters are 90 degrees apart.
G: cos2(( + 0, ) + G i cos2((+ 0, ) Furthermore, if the angle difference between the phase shifts of the
two linear filters is 90 degrees and the magnitude GI is equal to
100

the magnitude G I , the magnitude estimates of the relaying Figure 3 . It is found from fi a fundamental
algorithm are the same as the input signal magnitude. Indeed, frequency sinusoidal signal with f 1, the steady-
except the Ma-Momson and Prodar, all the algorithms studied state magnitude estimate of th n algorithm is
in the paper satisfy the condition at the fundamental frequency and between 6/nsin(6/z) and 1; for a third harmonics with the
thus give the accurate magnitude estimate at the fundamental magnitude of 1, the steady-
frequency. and 6/27.
Conventionally, the frequency responses of nonlinear
magnitude estimate algorithms in microprocessor relays were 4.2 Prodar
defined by considering only the first two terms in (8) or (9). The In the Prodar algorithm [ l , 21, the
discussions above show that such a combination of these two terms estimated as
has a practical meaning only at some specific frequencies; thus the
conventional definition is not precise.

4. FREQUENCY RESPONSEOF DETERMINISTIC 1

where x' ( k ) = -(x(k + 1 ) - x( k - 1))
RELAYING ALGORITHMS 2T
As discussed in the previous section, the magnitude 1
x"(k)= -(x(k + 1) - 2 4 k ) + x ( k -
responses of nonlinear relaying algorithms are characterized by the T2
normalized variation band of the signal magnitude estimates x'(k) and x"@) are approximates for the first and
corresponding to different frequencies. This section will study the derivatives of x(k) at sample k. Referri
frequency responses of deterministic relaying algorithms by finding functions for the two filters are
their normalized variation bands. The sampling rate (samples per
fundamental frequency cycle) is assumed to be 12.
4 1 Mann-Morrison: Substituting z by eJmT yields
In the Mann-Morrison algonthm [l, 21, the magnitude of a
sinusoidal signal is estimated as
Since the phase shift of the two filters are 90 degrees, according to
(10) and ( 1 1) the upper and lower boundaries of the magnitude
1
estimates are max(Gl,G2) and min(G,,G,) , respectively. Figure
where x ' ( k )= - ( x ( k + 1 ) - x(k - 1)) 4 shows the frequency response o f the Prodar algorithm wth a
2T
x'(k) is an approximate for the first derivative of x(k) at sample k, sampling rate of 12 The maximum of G, is N2/z2 at the
and w o is the fundamental frequency. Refenkg to Figure 2, the frequency f = N / 2 . f o .
transfer functions of the two filters in the Mann-Morrison
algorithm are
G,(z)= 1 (13a)
1
G, ( 2 )= -( z - 2-I )
2w0T
To find the frequency response, substituting z in (13) by eJoT
yields
sin UT
Gl=l, S,=O; G,=- , 0, =90".
WOT
Since the angle difference between the phase shifts of the
two linear filters is 90 degrees for any frequencies, according to 1 2 3 4 5 6
Mlltlples offdrdamerlallrequency
(10) and (11) the upper and lower boundaries of the normalized
Figure 3 Frequency response of Mann-Morrison Algorithm
sin wT
magnitude variation band will be maximum(1,-) and
DOT
ininimum(1,- sin wT 1, respectively.
w0T
If we use N to represent the sample rate, G2 can be written f
z
1
r-
as

i // I
where f is the input signal frequency, and fo is the fundamental //I/
N
frequency. From (14), G, reaches its maximum value - when
2z 0 1 2 3 4 5 6
Mlltiples offundamerlallrequeny
f = N/4. f , . By taking the sampling rate as 12, the frequency
Figure 4 Frequency response of Prodar algorithm
response of the Ma-Morrison algorithm is calculated and shown
 101

4.3 Full cycle DFT

For the full cycle DFT,

2 N-l 27r( N - i)
G,(z) = -Xsin(- )z"
N i=o N
The two filters represented by (16) and (17) are oflen referred to as
cosine and sine filters, respectively. The magnitude responses of
the cosine and sine filters are shown in Figure 5. The frequency Mwiples of Wamerlaltequency
response of the full cycle DFT is computed by substituting (16) Figure 7 Magnitude responses of the half cycle Sine and Cosine filters
and (17) into (8) and (9) and is shown in Figure 6. It can be seen,
by comparing Figures 5 and 6, that the upper boundary of the
magnitude estimates by the full cycle DFT is very close to
max(Gl,G2), and the lower boundary of the magnitude estimates
by the full cycle DFT is very close to min(G, ,G2) .

0 1 2 3 4 5 6
Mtitiples of furdamerial fequercy
Figure 8 Frequency response of the half cycle DFT algorithm

I 4.5 Cosine Filter

The cosine filter relaying algorithm uses two cosine filters
O2
-I described by (16) for estimating the magnitude of sinusoids [ 5 ] .
I
0
1 2 3 4 5 6 One of the two filters is delayed by a quarter fundamental
Mtitiples of M a m e r t a l fequency frequency cycle. The two filters are
Figure 5 Magnitude responses of the full cycle Sine and Cosine filters

The magnitude responses of the two filters are the same and have
already been show1 in Figure 5. The frequency response of the
cosine filter relay algorithm is shown in Figure 9.

1 2 3 4 5 0
Mtitipleso f m a m e d a l fequency
Figure 6 Frequency response of the full cycle DFT algorithm

4.4 Half cycle DFT

In the half cycle DFT,.
Midttlplcs of hmdamerlal feqvemy
4 27r(N/2 - i) )z..i
N/2-1
GI( z )= - C COS( Figure 9 Frequency response of the cosine filter algorithm
N i=n N
4 N/Z-1 2 4 N j 2 - i) ) z i 4.6 Least Square
G2(z)= -
N I=o
C
sin(
N
In the Least Square (LS) relaying algorithm [GI, the two
filters are described as
N-l
The magnitude responses of the two filters are shown in Figure 7. G,(z)= x A ( l , i ) z - '
The frequency response of the half cycle DFT is obtained by n=n
substituting (18) and (19) into (8) and (9) and is shown in Figure N-l
8. G2(z)= x A ( 2 , i ) z - '
t=n
102

where A is a 2xN constant matrix. The magnitude responses of the state condition, the normalized magnitude estimates will fall
two filters are shown in Figure 10. The frequency response of the within the normalized var
LS relay algorithm is shown in Figure 1 1. boundaries are described by ( 8 ) and (9), respectively. The time it
takes for an relaying algori
and reach the steady state condition
associated linear filters and
16-
values of the state variables as well. During a transient period,
14- I
I\ \
especially at the first few samples immediately following the fault
inception, the magnitude estimates of the deterministic relaying
algorithms depend on the pre-fault sampled values, and the

This section will

Muliplesoffuxlamerialkequery relay algorithms whose e selected in such a way
Figure 10 Magnitude responses of the linear filters in the LS algorithm two linear filters involved are linear time-invariant The section
starts by deriving the transfer functions of the two linear filters in
the KF relay algorithm and then studies the frequency responses of
the nonlinear KF relaying algorithms. Both two-state and three-
state KF relaying algorithms will be studied The details about the
KF relay algorithms will not be included in the section, and the
readers may refer to reference [7, 81. As in the previous section,
the sampling rate is assumed to be 12
5.1 Two-state Kalman filter relaying algorithm
A. Transferfunctions of the two linear$lters Gl(z) and G 2 ( z )
When we use two states to represent the input signal, the
two linear filters in the KF relaying algorithm may be described as
Mullplesoffilndamentalkequemy
Figure 11 Frequency response of the LS relay algorithm

5 . FREQUENCY RESPONSE OF KALMAN FILTER RELAY ALGORITHMS where CD and H are the state transition matrix and measurement
matrix, and K is the Kalman gain vector
In addition to the deterministic algorithms, KF algorithms
Two models have been suggested to represent the input
were also proposed for phasor estimation in Microprocessor relay
signal [7, 81. Each model corresponds to a different combination
[7]. The common feature between KF and the detenninistic of the state transition matrix and measurement matrix H. One
algorithms is that they all fit in the block diagram shown in Figure
model has the following state transition inatrix and measurement
1 The main difference between the two kinds of algorithms is that
matrix
the parameters of the two linear filters in the KF relaying
algorithm is designed based on the statistic characteristics of the 27c 27c
signal while, in deterministic algorithms, the parameters of the cos- -sin-
linear filters are constant and have nothing to do with the mput
signal characteristics. Another difference between KF and the sin- cos-
deterministic algorithms is that the two linear filters in the
deterministic algorithms are linear time-invariant FIX filters while In the other model,
the two linear filters in the KF relaying algorithm are W filters
and, depending on the model selected to represent the signal, could
be time-varying.
The magnitude estimates of KF algorithms to sinusoidal
@=[A P] (26a), H , = [cos(wkT) -sin(wkT)] (26b)

inputs were mvestigated in [9], where the magnitude estimates for For the model described by (25), the steady-state Kalman
a sinusoidal input were found through simulations by actually gain K is a constant vector, thus equation (24) represents linear
applying such a sinusoid to KF algorithms. The characteristics of time-invariant filters But for the model described by (26), the
KF algorithms at different samples following the fault inception steady-state Kalman gain K is a time-varying vector whose
was studied. elements are independent of the state variables x, and x 2 ;
It should be emphasized that the frequency response is a therefore, equation (24) represents linear time-vaying filters. I
concept describing the characteristics of systems (or algorithms) Substituting (25) into (24) and rearranging the resulting
under sinusoids at steady state conditions. For nonlinear relaying equations, we get
algorithms, the frequency response is represented by the
normalized magnitude estimates of sinusoidal inputs corresponding
to different frequencies. Followmg the inception of a fault, the
outputs of an relaying algorithm will first experience a transient
period and then enter the final steady state condition. In the steady where
 103

2n the inputs contain less noise, they will have a lower noise rejecting
a,, = -(1- K,)sin- capacity but a shorter response time.
N '
2n 2n 2n 2n
a,, = s i n - - K , c o s - - , =cos-+K,sin-,
N N N N
K, and K , are the two elements of vector K
Taking z transform of (27) and manipulating the resulting
equations, we have the following relations

where G,(z) and G,(z) are the transfer functions of the two
linear filters in the two-state KF relaying algorithm, and 02/

K, - ( K , cos-2n + K , sin -)z

2 n -1 -0 1 2 3 4 5 6
MJiples O f M a m e r t a l tequsncy

GLZ)= N N Figure 12 Magnitude responses of the linear filters in the

2a 2n - (29)
1 - ( ( 2 - K,)cos-- + K , in-)^ +(1 - K l )z-' two-state KF (diag(Q,)/R, = 0.5 )
N N
2n 2 n -1
K,-(K~COS--K,S~~---)Z
G,(z) = N N
2n 2n - (30)
1 -((2 - Kl )COS- + K2sin -)z I +(1- K, )z-'
N N

B. Frequency response of the two-state Ka lman $filtering relaying

algorithm
If we focus on the linear part of the KF relaying algorithm
described by (27) and do not care how its parameters are designed,
we find that the linear part of the KF relaying algorithm has the
form of typical IlR filters. Therefore, the method for analyzing the
. --\
, -,--
frequency responses of deterministic relaying algorithms is equally 1 2 3 4
MJiples ofkniamerdalfeqwncy
5 6

applicable to the analysis of KF relaying algorithms. Figure 13 Frequency response of the two-state KF relay
In the simulations, we found that the steady-state Kalman algorithm ( diag(Q,)/R,= 0.5 )
gain vector K is only dependent on the ratio of the state variable
covariance matrix Q, to the measurement noise variance Rk and
is not affected by the initial error covariance matrix P.
Assuming Qk= diagona1(0.005,0.005) and Rk = 0.01, the
steady-state Kalman filter gain K is computed as
K=[0.6003 -0.19851n. (31)
where diagonal(.) represents a diagonal matrix, and "Tr"
represents transpose.
Substituting (31) into (29) and (30), the magnitude responses
of G,(z) and G2(z) are computed and the results are shown in
Figure 12. By using (8) and (9), the frequency response or
variation band of the magnitude estimate of the two-state KF MJiples offiudamerdalhquemy
relaying algorithm is computed and shown in Figure 13. It can be Figure 14 Magnitude responses of the linear filters in the
seen that, at the fundamental frequency, the KF relay algorithm two-state KF ( ding(Qk)/Rk= 0.01 )
gives the accurate magnitude estimate.
Decreasing the ratio of diag(Qk),the diagonal elements of
Qk ,to Rk to 0.01, the steady-state Kalman gain vector is found as
K = [0.1316 - 0 . 0 0 8 0 ~ (32)
The magnitude responses of G,(z) and G2(z) are shown in
Figure 14, and the frequency response of the two-state KF relaying
algorithm is shown in Figure 15.
Figures 12 through 15 show that the KF relay algorithm with
a smaller diug(Q,)/R, ratio has a sharper frequency response,
which indicates a higher attenuation on harmonics and a longer
transient period. Thus, the KF relay algorithms are able to adjust 0"
1 2 3 4 5
their own parameters according to the noise level such that, if the Mdiiplas of Lrndamerdalfequency

inputs are badly corrupted by noise, they will have a higher noise Figure 15 Frequency response of the two-state KF relay
rejecting capacity but a longer response time; on the other hand, if algorithm ( diag(Q, )/Rk= 0.01 )
104

5 2 Three-state Kalman filter relaying algorithm

The linear filters of the three-state KF relaying algorithm Lower bovldarj
may be described as

As in the two-state Kalman filter relaying algorithm, variables xi

and x2 represent the sinusoid signal. The third state variable xj \.. \
----
0 -,
models the dc-offset in the input signal. If we choose the state I 2 3 4 5 6
Mtiiiples offundamental tequency
transition matrix and measurement matrix H [7] as
Figure 17 Frequency response of the three-state KF relay
algorithm ( diug(Qk)/Rk= 0.5 )

lo O e ‘J

the steady-state Kalman gain K is a constant vector; thus equation

(33) is linear time-invariant. In (34a), 7 is a time constant and
can be determined based on the transmission line parameters [7].
For a specified time constant z , the steady-state Kalman
gain vector only depends on the ratio of the state variable
covariance matrix Qk to the measurement noise variance Rk and
is not affected by the initial error covariance matrix P. Mtiiiples ofiundamentaltequency
Selecting the state covariance matrix Figure I8 Magnitude responses of the linear filters in the
Qk= diugunuZ(0.005,0.005,0.00S), three-state KF ( dmg(Q, )/Rk= 0.01 )
the measurement noise variance Rk = 0.01 , and the time constant
z = 0 01 s, the steady-state Kalman gain vector is found as
K = [0.4442 - 0.3144 0.2596r (35)
The magnitude responses of the two linear filters in the three-state
KF relay algorithm are shown in Figure 16 and the Gequency
response of the three-state KF relay algorithm is shown in Figure
17
Reducing the ratio of diug(Q, ) , the diagonal elements of
Qk,to Rk as 0.01, we found the steady-state Kalman gain vector
K = [0.1288 -0.0150 0.0309p’ (36)
1 2 3 4 5 6
The magnitude responses of the two linear filters are shown in Mtiiiples offundamentalfrequemy
Figure 18 and the Gequency response of the KF relaying algorithm Figure 19 Frequency response of the three-state KF relay
is shown in Figure 19. algorithm ( diag(Qr)/R, = 0.01 )
Comparing Figures 17 and 19, again we found that a smaller I
value of diug(Q,)/Rk corresponds to a sharper frequency response
implying a higher noise rejecting capacity and a longer transient 6. CONCLUSIONS I
period In phasor-based microprocessor relays, the m
phase estimate algorithms are nonlinear, therefore, the frequency
responses of these algorithms can not be analyzed by directly
applying the method defined in the linear time-invariant systems.
To study the frequency response, the paper proposed a genenc
model to represent the phasor-based relaying algorithms The
central part of the generic model is two linear filters. Based on the
generic model, the paper studied the magnitude responses of
deterministic relay algorithms and KF relay algorithms. The
deterministic algorithms studied include Mann-Morrison, Prodar,
full cycle DFT, half cycle DFT, Cosine filter, and Least Square
the KF relay algorithms, the transfer functions of the associated
linear filters were derived with both two-state and three-state
Mllliples oflindamerlalfrequeny models being considered.
Figure 16 Magnitude responses of the linear filters in the The paper shows that the magnitude responses of the
three-state KF ( diug(Q,)/R,= 0.5 ) nonlinear relaying algorithms can be represented by the normalized
 105

variation band of magnitude estimates in the frequency domain. 4. J.G. Proakis and D.G. Manolakis, Digital Signal Processing,
The upper and lower boundanes of the band define the region of Prentice-Hall, 1996.
the normalized magnitude estimates of the relays for sinusoidal 5. E.O. Schweitzer, D. Hou, “Filtering for Protective Relays,”
inputs with different frequencies. When the phase shift between 19Ih Annual Western Protective Relay Conference, Spokane,
the two linear filters in a relaying algorithm is 90’ apart, the Washington, Oct. 20-22, 1992.
upper boundary of its magnitude response is equal to the maximum 6. M.S. Sachdev and M.A. Baribeau, “A New Algorithm for
of the magnitude responses of its associated linear filters, and the Digital Impedance Relays,” IEEE Transactions on Power
lower boundary of its magnitude response is equal to the minimum Apparatus and Systems, Vol. PAS-98, No. 6, Nov.5ec. 1979,
of the magnitude responses of the linear filters. The value of the pp.2232-2240.
upper boundary at a harmonics indicates an algorithm’s capability 7. A.A. Girgis and R.G. Brown, “Application of Kalman
in eliminating the harmonics, and a value of zero means that the Filtering in Computer Relaying of Power Systems,” IEEE
harmonics is completely eliminated. Transactions on Power Apparatus and Systems, Vol. PAS-
For all the relaying algorithms studied in the paper except 100, NO.7, July 1981, pp.3387-3397.
the Mann-Momson and Prodar, the upper and lower boundaries of 8. A.A Girgis, W.B. Chang, and E.B. Makram, “A digital
their magnitude response are equal to 1 at the fundamental Recursive Measurement Scheme for On-line Tracking of
frequency; thus these algorithms provide accurate magnitude Power System Harmonics,“ IEEE Transactions on Power
estimates to fundamental frequency sinusoidal inputs. Delivery, Vol. 6, No. 3, July 1991, pp.1153-1160.
The normalized magnitude variation bands in the KF 9. G. Benmouyal, “Frequency Domain Characterization of
relaying algorithms are dependent on the signal statistic Kalman Filters as Applied to Power System Protection,” IEEE
characteristics. A smaller ratio of the state variable covariance to Transactions on Power Delivery, Vol. 7, No. 3, July 1992, pp.
the measurement noise variance corresponds to a sharper frequency 1129-1138.
response, which indicates a higher harmonics rejecting capacity but
a longer transient period in the time domain.
Liancheng Wang was born in Shandong, P.R. China, on
7. ACKNOWLEDGMENT December 16, 1963. He obtained his B.S. and M.S. degrees in
The author wishes to acknowledge Elmo Price and Lifeng Electrical Engineering from Shandong University of Technology,
Yang at ABB Power Automation & Protection Division for their Jinan, China, in 1983 and 1986, respectively. He obtained his
helpful discussions and support. Ph.D. degree in Electrical Engineering from Clemson University in
1995. From 1986 to 1991, he was an assistant professor in the
8. REFERENCES Electrical Engineering Department, Shandong University of
Technology. In 1996, he joined ABB Power Automation &
1. IEEE Tutorial, Computer Relaying, 79EHO148-7-PWR, 1979. Protection Division as a senior engineer in Coral Springs, FL.
2. IEEE Tutorial, Microprocessor Relays and Protection Before joining ABB, he was a senior application engineer with
Systems, 88EH0269-1-PWR, 1988. LCG Consulting in Los Altos, CA. His area of interests includes
3 . A.G. Phadke and J.S. Thorp, Computer Relaying for Power microprocessor relays, signal processing, and power system
Systems, John Wiley & Son Inc., 1988. stability. He is a Senior Member of the IEEE.
I06

Discussion

R. Grover Brown and Adly A. Girgis: The

author presented interesting discussion on
the frequency response of some digital The amplitude of the p-th harmonic is computed as:
relaying algorithms. We offer the following
comments.

Especially in the power-system relaying

application, the Kalman filter is operating in where a = 2yN, and NI stands for the number of samples per
il ”transient” mode. The filter gains are cycle.
time-varying, and this is an essential feature The delay m is selected freely from the interval m = 1 up to
of the filter in this application. In this case. m = NJ4. The higher values provide better overall filtering
but introduce an extra delay into the measurement process.
an analysis based on steady-state gains (if The unique (i.e. not depending on the phase of a signal)
such analysis exists) is absolutely frequency response of the estimator (D3) is a
’

meaningless in assessing the estimator’s feature of this algorithm in the context of the
performance to do its intended job, namely The upper and lower boundaries for algo
that of estimating the post-fault steady-state structure as in Fig.2a become one line for
voinge and current, but doing 50 during the the structure as in
Fig.Dl (compare
transient period before the steady-state Fig.5 and Fig.
condition is reached. D2). The unique
frequency re-
An error in Sec 5.1 is also worthy of sponse facilitates
mention. The two models represented by for example di-
Eqs. (25 a,b) and Eqs (26 a,b) are equivalent rect compensa-
tion for the fre-
in that there is a linear transformation that
takes one state representation into the other. Fig.D2. The frequency response of the ampli- quency deviations
relative frequency, p u
tude eshmator D3 for m = l and N~= 20. if the actual fre-
The two model’s yeild equivalent results quency is known.
when the results are properly interpreted. [Dl] Kasztenny B and Rosolowski E , Two new measuring algorithms for
generator and transformer relaying, IEEE PES WM-98 paper PE-205-
Both models have time-varying gains. This PWRD-0-11-1997, Tampa, FL, February 1-5, 1998.
is obvious by noting that the error
covariance matrix changes with each step of
the estimation process. L. Wang: The author would like to thank all discussers for
their interesting comments and providing him the
opportunity to explain some important issues not covered in
B.Kasztenny (Texas A&M University, College Station, the paper. Responses to the comments are as follows.
TX). This discusser would like to congratulate the author on Regarding the comments by Professor Brown and
his interesting paper and offers the following comment. Professor Girgis, the claim that “Especially in power system
The diagram presented in Fig.2a is not the only way of relaying application, the Kalman filter is operating in a
combining the two orthogonal components into the estimate transient mode” is partially true and applicable only for
of amplitude. Another method is shown in Fig.D1 [Dl]. protections that are designed to operate during that period,
The algorithm e.g., zone 1 protection. However, there also exist other
works well for relaying applications where the KF (Kalman Filter) works at
the fundamental its steady state. It is in these applications that the two
.3& orthogonal filters in the KF described by (25) or (34)
frequency com-
ponent (p = 1) become time-invariant IIR filters and the frequency domain
1 and higher har- method is equally applicable to the analysis of the KF.
monics (p = To give an idea as to the duration of the KF transient
2,3, ...) and splits period, a power system was simulated based on EMTP and
Fig.Dl. Block diagram of the amplitude esti- the KF gains were calculated. The system simulated
mator providing uruque frequency response. the measured
signal x in the contains two parallel lines with sources at both ends. The
orthogonal components xpc and xps. The pair of orthogonal lines are 200 miles long and rated at 500 KV. The data
filters F,, and F,, is selected freely and should serve particu- generated by EMTP was played back to a relay and captured
lar application. For example the full-cycle Fourier filters may by the relay with a sampling
- _ per
- rate of 20 samples _ cycle.
_
be used as follows. T h e mmnntatinnc uwre hacprl nn the rantiired Anta inctearl nf
 107

I
directly using the EMTP data. The three-state KF was 180

selected for this study and the resulting KF gains were 160 /-
shown in Figure 20. It is found that, following the inception
of a fault, the three KF gains experience a transient period
and then settle down at the stead state. As shown in Figure G 100
20, the transient period of the KF for the studied system is
0)
about 3/4 of a cycle. Therefore, if the relay makes its trip
decision before 3/4 cycle, the frequency response of KF 40
cannot provide any information as to its anti-noise capacity.
However, if the trip decision is made after 3/4 cycle, the KF
has already changed to linear-time-invariant IIR filters and
2:ot-,MULTIPLES OF FUNDAMENTAL FREQUENCY
10

the powerfid tool-frequency domain method-is well Figure 22: Phase shift between the two IIR filters of
suited for this situation. the KF (sampling rate N=20)

d
2 1
4
fI 0
g -1

-2

-3

0 0.5 1 1.5 2
TIME IN CYCLES
Figure 23: Frequency response of the three-state KF
Figure 20: Kalman Gains of the three-state KF (sampling rate N=20)

The steady-state KF gain vector shown in Figure 20 is To find out how the KF responds to faults if steady
state KF gains are utilized, fault current magnitudes were
K = [0.3503 -0.4537 0.4443p computed using both time-varying KF gains and steady state
Based on the gain vector, the magnitude responses of the
KF gains. The results were also compared with those
obtained by using Full-cycle DFT. Different fault inception
two orthogonal IIR filters in the three-state KF filter are
angles were considered to cover a wider range of fault
shown in Figure 21, and the phase shift between the two conditions. With phase A-to-ground (A-G) faults at the
orthogonal filters are given in Figure 22. It is found that, at middle (100 miles) of one of the parallel transmission lines
the fundamental frequency, the gains of the two IIR filters in on the system described above, the fault current magnitudes
the three-state KF are equal to 1 and the phase shift between were illustrated in Figures 24-28. The current magnitude
them is 90 degrees. Thus, the three-state KF can provide estimates were expressed in per unit with the steady state
accurate magnitude estimate for fundamental frequency fault current as the base.
sinusoid. The frequency response of the KF was obtained Figures 24-28 show that the rising speed of the current
based on (8) and (9) and shown in Figure 23. magnitudes is extremely fast for the time-varying KF gains.
For example, corresponding to 0 fault inception angle, it
takes 1 sample for the current magnitude to jump to 0.85 pu
from the pre-fault load current 0.38 pu, and it only takes 3
samples for the current to reach its steady-state value.
When steady-state KF gains are used instead, the
converging speed of the current magnitude becomes slower
but still faster than using the full-cycle DFT algorithm.
This shows that IIR filters do not have to take infinite time
to settle down as their name suggested. The transient
period in the outputs of IIR filters depends on the locations
of their poles in the transfer functions as well as their initial
I conditions. Ideally, if the initial conditions applied happen
2 4 6 8 10
MULTIPLES OF FUNDAMENTAL FREQUENCY
to be the steady state solutions at the instant when the IIR
Figure 21: Magnitude responses of the three-state KF filters are applied, the IIR filters would not experience any
(sampling rate N=20) transient period at all.
108

+++*

i'i +
1 1 * +*

z
w O0
4 0 9 0

0 + *
CO
+ *
+ +
1 E 0 6 +t3*+#*
+ +
0
+
0
+
K F : TIME-VARYING GAINS
K F : STEADY-STATE GAINS
5
I
I:
0 5 +*+
4
I
* FULL-CYCLE DFT 0 4 ++
031
0 5 1 1 5 2 0 3 L
I
I
0 0 5 1 1 5 2
T I M E IN C Y C L E S
T I M E IN C Y C L E S
Figure 24: A-G fault current with fault inception Figure 28: A - 6 fault current with fault inception
angle at 0"
angle at 120"

Figures 24-28 also show the effect of dc-offset on the

current magnitude estimates. In the DFT algorithm, dc-
offset causes big overshoot in the current magnitude, and it
takes a long time (more tlm two cycles) for the magnitude
estimates to settle down. Compared to DFT, the KF has an
0 K F TIME-VARYING GAINS
0 4 + K F STEADY-STATE GAINS excellent dc-offset immunity and no overshoot is
4 FULL-CYCLE DFT
experienced.
03
0 5 1 1 5 2 From the forgoing discussions, it is obvio
T I M E IN C Y C L E S
Figure 25: A-G fault current with fault inception protection is the possible application where the KF works at
angle at 30" its transient mode when the KF gains have not reached their
steady state values. As is known, protective relays also have
backup zones except the fast speed zone 1: In addition,
modern microprocessor relays have more functions than just
relaying. Zone 2, zone 3 , metering, controls (re-closing and
syn-check, et. al), and fault locations are the relay
applications where KF works at its steady state and behaves
like deterministic IIR filters. Therefore, there is no re
to doubt the significance of applying the frequency domain
method to the analysis of the Kalman Filter.
The second comment by Prof. Brown
0 0 K F TIME-VARYING G A I N S is also related to time-varying gains of the
+ + K F STEADY-STATEGAINS
+ * FULL-CYCLE D F T frequency response described in the pape
1
03'
0 0 5 1 1 5 2
applications where the KF works at
T I M E IN C Y C L E S steady state KF gains are constant.
Figure 26: A-G fault current with fault inception
angle at 60"
magnitude estimate algorithm described in Figure D1 are
equal. But the curve given in Figure D2 seems not correct.
According to Figure D1, the normalized magnitude
estimates are found as

JG, G~sin(8, - 0,

where G, , G2, , and Q2 are obtained based on (16)

0 K F TIME-VARYING GAINS
(17). Using the same sampling rate (20 sampleskycle) as in
0 4 + K F STEADY-STATE GAINS Figure D2, the frequency response of the algorithm D3 is
4 FULL-CYCLE D F T
03 shown in Figure 29. From Figure 29, the normalized
05 1 1 5 2
T I M E IN C Y C L E S magnitude estimate is equal to 1 at the fundamental
Figure 27: A-G fault current with fault inception frequency and less than 0.2 for signal frequencies larger
angle at 90" than the second harmonics.
 109

"
MULTIPLES OF FUNDAMENTAL FREQUENCY
Figure 29: Frequency response of magnitude estimator
in Fig D1 (sampling rate N=20)