IEEE Transactionson Power Delivery ,Vol. 6, No.

3, July 1991 11153

A DIGITAL RECURSIVE MEASUREMENT SCHEME FOR ON-LINE

TRACKING OP POWER SYSTEM HARMONICS

Adly A. Girgis W. Bin Chang Elham B. Makram

IEEE Senior Member IEEE Student Member IEEE Senior Member

Clemson University
Department of Electrical and Computer Engineering
Clemson, South Carolina 29634-0915

Abstract: This paper presents an optimal measurement scheme APPLICATION O F FFT FOR HARMONIC ANALYSIS
for tracking the harmonics in power system voltage and current The DFT and FFT algorithms have been applied to many
waveforms. The scheme does not require an integer number of useful applications in power system phasor measurements and
samples in an integer number of cycles. It is not limited to harmonic analysis [8,9]. However, misapplication of the FFT
stationary signals, but it can track harmonic with time-varying algorithm would lead to incorrect results. There are basic
amplitudes. The paper starts with a review of the common assumptions embodied in the application of DFT and FFT.
frequency domain techniques for harmonics measurement. The These assumptions are: (i) The signal is stationary (constant
common techniques are based on the discrete Fourier transform magnitude); (ii) The sampling frequency is equal to the number
(DFT) and fast Fourier transform (FFT). The paper includes of samples multiplied by the fundamental frequency assumed by
examples of pitfalls in the common techniques. The paper the algorithm; (iii) The sampling frequency is greater than
then introduces the concepts of the new scheme. The new twice the highest frequency in the signal to be analyzed; and
scheme is based on Kalman filtering theory for the optimal (iv) Each frequency in the signal is an integer multiple of the
estimation of the parameters of time-varying harmonics. The fundamental frequency. Basically, the fundamental frequency is
scheme is tested on simulated and actual recorded data sets. the reciprocal of the window length of data (T). When these
assumptions are satisfied, the results of the DFT or FFT are
KEY WORDS: Harmonics, Measurement, DFT, FFT, Kalman accurate. To show this, consider the waveform described by
Filtering. equation (1). The frequency components are 60 Hz, 180 Hz,
and 300 Hz. Thus, the signal to be analyzed is described by
INTRODUCTION
The widespread applications of electronically controlled s (t)= A, m(w0t + el)+ %ox@oot+ e,)+ As ms(%oot+ Os) (1)
loads have increased the harmonic distortion in power system
Using 64 samples at a sampling frequency of 32 x 60 Hz, the
voltage and current waveforms [I-41.As power semiconductors
results of the FFT program correctly indicated the values AI,
are switched on and off at different points on the voltage wave-
form, damped high-frequency transients are generated. If the
A3, and As at the specified frequencies. This is because the
switching occurs at the same points on each cycle, the tran-
signal analyzed satisfied all the assumptions in the FFT algo-
sient becomes periodic. This transient is non-stationary and
rithm.
its frequency is not a multiple of 60 Hz [5,6]. Consequently,There are three major pitfalls in the application of FFT;
voltage and current waveforms of a distribution or a transmis-
namely, aliasing, leakage, and picket-fence effect [IO].
sion system are not pure sinusoids, but may consist of a com-
Aliasing can be alleviated by increasing the sampling fre-
bination of the fundamental frequency, harmonics, and high-
quency (f,). However, pseudoaliasing may still occur even if
frequency transients. Furthermore, many of power system the highest frequency component is not higher than fJ2. This
loads, especially industrial loads, are dynamic in nature. This
may be caused by the presence of a fraction of a cycle of data
produces a time-varying amplitude for the current waveforms.
or the presence of a white noise in the measurements. The
For example, loads that exhibit a continuous starting and brak-
picket-fence effect occurs if the analyzed waveform includes a
ing operations generate a time-varying current amplitude. The
frequency which is not one of the discrete frequencies (an
rate of change of the current amplitude depends on the type and
integer times the fundamental). The term "leakage" refers to
size of load. the apparent spreading of energy from one frequency into adja-
Accurate measurement of power system harmonics is essen-
cent ones. It arises due to the truncation of the sequence such
tial to evaluate the harmonic distortion in both current and
that a fraction of a cycle exists in the analyzed waveform.
voltage waveforms. The design of harmonic filters relies on Data acquisition systems sampling rates are usually set at
the measurement of harmonic distortion. Most frequency fixed values, multiples of kHz. If the sampled waveform does
domain harmonic analysis algorithms are based either on the
not contain an integer number of samples per integer number
discrete Fourier transform (DFT) or on the fast Fourier trans-
of cycles, the results of the DFT algorithm will include errors.
form (FFT) to obtain the voltage and current frequency spectra
The resulting error is known as spectral "leakage" [10,11].
from discrete time samples. The necessary equations to apply
The DFT and FFT of such a sampled waveform will incorrectly
the DFT or FFT are shown in Appendix A [7]. indicate non-zero values for all of the harmonic frequencies.
Two examples will be shown here to explain how incorrect
91 WM 109-9 PWRD A paper recommended and approved results may be obtained from the FFT algorithm.
by the IEEE Transmission and Distribution Committee First, consider a signal that includes a 60 Hz component
of the IEEE Power Engineering Society for presentation (no harmonics) with a time-varying amplitude. This signal is
at the IEEE/PES 1991 Winter Meeting, New York, New shown in equation (2) as
York, February 3, 1991. Manuscript submitted
January 31, 1990; made available for printing s(t) = (A, - $e-at)msmot (2)
December 18, 1990.
where A1 = 1.0, A2 = 0.8, a = 60, a0 = 2./: <: 60 radi-
ans/second. The signal is shown graphically in Figure 1. The
FFT algorithm was applied to the samples of the waveform
described by equation (2) using a sampling rate of 32 x 60 Hz.
The results of a 32-point (one cycle) FFT algorithm indicated

1 153$Ol.O001991 IEEE
0885-8977/91/0700-
1154

that the waveform includes harmonics of the 60 Hz and a total The aforementioned pitfalls raised the need for an optimal
harmonic distortion of 33.18 percent. When a 64-point (two estimation technique capable of tracking harmonics with time-
cycles) FFT algorithm was applied, the results indicated fre- varying magnitudes. The concepts of signal modeling and the
quencies multiples of 30 Hz with a dominant frequency of 90 estimation algorithm are based on Kalman filtering. State
Hz. The total harmonic distortion was incorrectly indicated to variable modeling of current and voltage waveforms and the
be 31.85 percent. The FFT results up to 420 Hz are shown in principles of the optimal estimation algorithm are described in
Table 1. the following sections.

OPTIMAL ESTIMATION OF HARMONICS

1.o The Kalman filtering approach provides a means for opti-
mally estimating phasors and the ability to track time-varying
0.5 parameters [13-161. The derivations of Kalman filtering equa-
tions are well described in several excellent references [13].
0.0 However, the derivations described here are intended to famil-
iarize the power engineers with state variable representation of
-0.5 phasor quantities, and the necessary system and measurement
-1.o equations of the K h a n filter recursive algorithm for the opti-
mal tracking of harmonics with time-varying magnitudes.

0 20
40 60 80 100 State Variable Representation of a Signal With a Constant or a
TIME(MSEC) Time-Varvinp M a w
Figure 1. A 60 Hz waveform with changing magnitude Consider a signal with a frequency o and a magnitude of
Al(t), where Al(t) represents a combination of a constant value
A second example shown here is the analysis of a pure 60 plus a time-varying component. Considering a reference rotat-
Hz waveform sampled at a sampling frequency of 2 kHz. The ing at O , the noise-free signal may be expressed as:
signal is then described by
s (t) = A(t) ms(at + 0) = A(t) ad m m t - A(t) sine sinat (4)
s (t) = A ms(oot + e) (3)
Let xl be A(t)cosO and x2 be A(t)sine; therefore, each xl and
The results of 32- and 64-point FFT analysis of the wave- x2 includes two components. One component is constant but
form described by equation (3) are shown up to 420 Hz in Table unknown. The other component may be time-varying. The
1. These results indicate (incorrectly) a total harmonic distor- variables xl and x2 represent the in-phase and quadrature-phase
tion of 8.89 and 15.74 percent, respectively. components and referred to as state variables [13]. This leads
TABLE 1 to the following state equations
FFT ANALYSIS OF WAVEFORMS DESCRIBED
BY EQUATIONS (2) AND (3)
k+l k k
Freq.(Hz)
where wl and w2 allow the state variables for random walk
(time variation). The measurement equation would include the
0 signal and noise and it can be represented as:
30
60 0.47687
0.07901 0.10843 - -
0.97394
90 0.1 29 12 0.06169
120 0.10839 0.02899
150 0.05337 0.01838
where vk represents the high frequency noise.
180 0.06180 0.01334
210 0.03526 0.01047
240 0.04447 0.00863 State Variable Rewesentation of a Signal that Includes n
270 0.02685 0.00737 Harmonics (Model 1)
300 0.03525 0.0241 1 0.00480 0.00645 A noise-free current or voltage signal s(t) that includes n
330
360
390
0.02954 I 0.02195
0.02020
0.01876
I 0.00373 I 0.00576
0:00522
0 00479
I harmonics may be represented by

s (t)= 2 Ai(t) as@t + ei) (7)

420 0.02568 I 0.01756 I 0.00308 I 0.00444 i=l
THD(%) 33.18 I 31.85 I 8.89 I 15.74 where
Ai(t) is the amplitude of the phasor quantity representing
These two examples clearly indicate that blindly applying the i* harmonic at time t,
DFT or FFT may lead to incorrect results in harmonic analysis.
Furthermore, for the FFT algorithm to be applied, the number B i is the phase angle of the i* harmonic relative to a refer-
of samples needs to be an integer power of two. Also, there is ence rotating at i o ,
always uncertainty in determining the actual frequency of the n is the harmonic order.
fundamental components in power system waveforms. Power As indicated in the previous subsection, each frequency compo-
system frequency may not be exactly 60 Hz during steady-state nent requires two state variables. Thus, the total number of
operation. The standard deviation of frequency drift for the state variables is 2x1. These state variables are defined as fol-
Eastern Interconnection of the United States is about 0.05 Hz lows
[W. xl(t) = A,(t) CosB,, x2(t) = A,(t) sine,,
 1155

~ 3 ( t=
) A2(t) ~ 0 ~ 0 2 , x4(t) A2(t) s i n e , If the signal includes n frequencies; the fundamental plus n-1
... ... (8) harmonics, the state variable representation may be expressed
... ... -
- -

=;[ : ;]
x2n-l(t) = &(t) cos0,. x2n(t) = &(t) sine,. a1
xl xl
These state variables represent the in-phase and quadrature-
phase components of the harmonics with respect to a rotating
reference, respectively. This may be referred to as model 1. x2 x2 a2
Thus, the state variable equations may be expressed as:
- - r
... ... ... + ... wk
r - -
xl 1 0 . 0 0 xl a1
X2ll-1 ah-1
x2 0 1 . 0 0 x2 a2
- x2n +1 &-k - a%
... = . . . . . ..a + ... where the submatrices Mi are shown below as
X2n-1 0 0 . I 0 x2n-1 ah1

1
as(ioAt) -sin(ioAt)
O 0 . O l i X2nk Mi =
- x2n -k+l - ah NioAt) cOs(ioAt)
The measurement equation can be then expressed as: The measurement equation can be then expressed as:

xl r xl 1
x2 x2

... 4 = [l 0 ... 1 03 ... + Vk

fi-1

x2n
The latter model may be referred to as model 2. This model has
It should be indicated here that Hk in this case is a t h e - constant state transition and measurement matrices. However,
varying vector. A constant Hk vector can be obtained if a sta- it assumes a stationary reference. Thus, the in-phase and
tionary reference is used in the state variable representation. quadrature-phase components represent the instantaneous values
This is explained in the following subsection. of cosinusoidal and sinusoidal waveforms, respectively.

State Variable Representation of a S ienal With T h e-V arv. The Kalman Filtering Algorithm
Magnitude Usinp a Stationarv Reference (Model 2 The equations described in the previous section are suitable
Consider the noise-free signal to be: for the Kalman filtering algorithm. The Kalman filter is a
recursive optimal estimator, well suited for on-line applica-
s (5) = A<k)m(at+ e> (11) tions. It requires a state variable model for the parameters to
NOW. consider X l k to be A(tk)COS(otk + e ) and X2k to be be estimated and a measurement equation that relates the
discrete measurement to the state variables (parameters). These
A(tk)Sin(Otk + e). At tk+l, which is tk +At, the signal may be equations are described in Appendix B[13].
expressed [13] as :
TESTING THE KALMAN FILTER ALGORITHM
S ($+I>= A($+1)m(o$ + 0) = USING SIMULATED DATA
The two K h a n filter models described in the preceeding
= ~ 1 as(aAt)
, - x2,Sin(0At) section were tested using a waveform with known harmonic
Also contents. The waveform consists of the fundamental, the third,
the fifth, the ninth, the eleventh, the thirteenth, and the
q+l
= A c $ + l > h ( q+ oAt+ 0) nineteenth harmonics. The waveform is described as

= xlk h(aAt) + x2,as(oAt) +

s (t) = 1.0 cos(ot+ 107 0.1 cOsOot+B7+ 0.08 cOs(5wt+307

Thus, the state variable representation takes the following + 0.08cos(7ot+4oq + 0.06 mS(l1oa50q
form:
+ 0.05 ax(l3ot+607+ 0.03 ax(lWt+W)
The sampling frequency was selected to be 64 x 60 Hz.
sm(oA t) as(oA t)
k+l k k
Selection of Kalman Filter Parameters
The measurement equation then becomes (i) Initial process vector ( )\6)
r 7
As the Kalman filter model started with no past measure-
ment, the initial process vector was selected to be zero. Thus,
the first half cycle (8 milliseconds) is considered to be the
initialization period.
1156

(ii) Initial covariance matrix (Pi) Testine results of Model 2

The initial covariance matrix was selected to be a diagonal A 14-state variable model was also implemented using the
matrix with the diagonal values equal to 10 pu2. stationary reference described by equations (14) to (16). Figure
(iii) Noise variance (R) 5 shows the first two components of the Kalman gain vector.
The noise variance was selected to be constant at a value Figure 6 shows the first and second diagonal elements of Pk.
of 0.05 pu2. This was based on the background noise variance The estimation of the magnitude of the fundamental and third
at field measurement. harmonic were exactly the same as those shown in Figure 2.
(iv) State variable covariance matrix (Q)
The matrix Q was also selected to be 0.05 pu2. 0.6 I i

Testine Results of Model 1

A 14-state variable model was implemented using the ro-
tating reference described by equations (7) to (10). Figure 2
shows the initialization period and the recursive estimation of
the magnitude of the fundamental and third harmonic. Figure 3
shows the Kalman gain for the fundamental component. Figure
4 shows the first and second diagonal elements of Pk. 3 0.0

.o
1*2
1 :
-
-
60-HZ
-0.2:
0
I

1020
. I

30
TIME(MSEC)
. I

40
. f

50
. I

Figure 5. Kalman gain for x l and x2 using the 14-state

- 3rdharmonic model 2

$ 0.4 10
0.2
- I - P22
0 . 0 : - I . 1 . f . f .

0 10 20 30 40 50 2
1
TIMWMSEC)
Figure 2. Estimated magnitude of 60 Hz and third harmonic
component using the 14-state model 1

0.4 I I .1
0 10 20 30 40 50
TIMWMSEC)
Figure 6. The fiist and second diagonal elements of Pk
matrix using the 14-state model 2

Discussion of Results and ComDarison of Models

The Kalman gain vector Kk and the covariance matrix Pk
reach steady-state in about half a cycle. When model 1 is used,
-0.6: . , . , . I , , , 4 the steady-state Kalman gain becomes periodic with a period of
1/60 seconds. Its variations include harmonics of 60 Hz. The
0 10 20 30 40 50
covariance matrix in the steady-state consists of a constant
TIME(MSEC)
plus a periodic component. These time variations are due to
Figure 3. Kalman gain for x l and x2 using the 14-state the time-varying vector in the measurement equation. Thus,
model 1 after initialization of the model, the Kalman gain vector of the
third cycle can be repeated for successive cycles.
10 When model 2 is selected, the components of the Kalman
- P11 gain vector and the covariance matrix become constants. In
- P22 both models, the Kalman gain vector is independent of the
measurements and can be computed off-line. As the state man-
2. sition matrix is a full matrix, it requires more computation than
1 model 1 to update the state vector.
SOFTWARE IMPLEMENTATION
A Fortran program was developed to perform the DFT and
FFT analysis on digitally recorded voltage and current wave-
forms. The selection of the DFT or FFT is based on (1) the
.1 number of samples and cycles available for analysis and (2) the
0 10 20 30 40 50 nature of the waveform (stationary or nonstationary). The FFT
TIME(MSEC) program is applied if (1) the number of samples is an integer
Figure 4. The fiist and second diagonal elements of Pk power of two over an integer number of cycles and (2) the
waveform is stationary. If the number of samples is not an
matrix using the 14-state model 1
integer power of two, but includes an integer number of cycles
 ~

1157
and the waveform is stationary, the DFT algorithm is applied. 2
The algorithm then computes the frequency components in VA VB VC 1
polar form.
When the number of samples does not satisfy the basic
assumption in the DFT/FFT algorithm, a Kalman filtering based
algorithm must be applied.
TESTING ON ACTUAL RECORDED DATA
Two cases of actual recorded data are reported here. The
first case represents a large industrial load served by two
parallel transformers totaling 7500 KVA [SI. The load consists
of four production lines of induction heating with two single-
phase furnaces per line. The induction furnaces operate at 8500 0 10 20 30 40 50
Hz and are used to heat 40-ft steel rods which are cut into TIMFi(MSEC)
railroad spikes. Diodes are used in the rectifier for converting
Figure 8. Actual recorded voltage waveform of phase A, B,
the 60 Hz power into dc and SCRs are used in the inverter for
converting the dc into single-phase 8500 Hz power. The wave- and C
forms were originally sampled at 20 kHz. A program was
written to use a reduced sampling rate in the analysis. A care- 1.2 ,
ful examination of the current and voltage waveforms indicated 1.0 -
that the waveforms consist of (1) harmonics of 60 Hz and (2) a 60-Hz
decaying periodic high-frequency transients. The high-fre- 0.8

:::
quency transients were measured independently for another
purpose [6]. The rest of the waveform was then analyzed for
harmonic analysis. Using a sampling frequency that is a
multiple of 2 kHz, the DFT was then applied for a period of 3
cycles. The DFT results were as follows. 3 0.2
Freq.(Hz) Mag. Angle (rad.) 0.0
60 1.0495 -0.20 0 10 20 30 40 50
300 0.1999 1.99 TIME(MSEC)
420 0.0489 -2.18 Figure 9. Estimated magnitude of the fundamental, fifth,
660 0.0299 0.48 and seventh harmonics for phase A current
780 0.0373 2.98
1020 0.0078 -0.78
1140 0.0175 1.88
0.2 , I i

- llthharmonic
s^

E
The Kalman filter, however, can be applied for any number l_l_ 13thharmonic
of samples over a half cycle. If the harmonic has time-varying c5
magnitude, the Kalman filter algorithm would track the time
variation after the initialization period (half a cycle). Figures 0.1
7 and 8 show the three-phase current and voltage waveforms
recorded at the industrial load. Figures 9 to 11 show the recur-
sive estimation of the magnitude of the fundamental, fifth, and
seventh harmonics; the eleventh and thirteenth harmonics; and
3
the seventeenth and nineteenth harmonics, respectively, for 0 . 0 : . , . I . I a I . I
phase A current. The same harmonic analysis was also applied 0 10 20 30 40 50
to the actual recorded voltage waveforms. Figure 12 shows the TIME(MSEC)
recursive estimation of the magnitude of the fundamental and Figure 10. Estimated magnitude of the eleventh and
fifth harmonic for phase A voltage. No other voltage harmon-
thirteenth harmonics for phase A current
ics are shown here due to their negligible small value.

0.15
1.2
- 17thharmonic
0.8 h _yy_ 19thharmonic
2
h

0.4

I -:::
-0.8
1 ' 1 . 1 '

0 10 20 30 40 50 0 10 20 30 40 50
TIME(MSEC)
m(Msm Figure 11. Estimated magnitude of the seventeenth and
Figure 7. Actual recorded current waveform of phases A, B,
nineteenth harmonics for phase A current
and C
 1158

CONCLUSIONS
1.5 , - I
This paper presented the basic assumptions in the DFT and
h
FFT algorithms and the principles of Kalman filtering in track-
ing the time variation of power system harmonics. The
1.0 - - pitfalls in the FFT are illustrated by two examples. The first
- 60-Hz
5thHmonic
example is the FFT analysis of a waveform with exponential
varying magnitude but a single frequency. The FFT results
incorrectly indicated the presence of harmonics and high total
harmonic distortion. In the second example, a typical sam-
pling frequency of 2 W z was used. Due to leakage, the FFT
results showed incorrectly non-zero values for the harmonics.
There is no doubt that the Kalman filtering algorithm is more
0 10
20 30 40 50 accurate and is not sensitive to a certain sampling frequency.
TIME(MSEC) As the Kalman filter gain vector is time-varying, the estimator
can track harmonics with time-varying magnitudes.
Figure 12. Estimated magnitude of the 60 Hz and fifth Two models are described in this paper to show the flexi-
harmonic for phase A voltage bility in the Kalman filtering scheme. There are many applica-
tions, where the results of FFT algorithms are as accurate as a
The second case represents a continuous dynamic load (an Kalman filter model. However, there are other applications
amusement park ride). The load consists of two six-pulse where a Kalman filter becomes superior to other algorithms.
drives for two 200 horsepower dc motors. The current wave- Implementing linear Kalman filter models is relatively a simple
form of one phase is shown in Figure 13. The harmonic task. However, state equations, measurement equations, and
analysis using the Kalman filter algorithm is shown in Figure covariance matrices need to be correctly defined.
14. It should be noted that the current waveform was continu-
ously varying in magnitude due to the dynamic nature of the ACKNOWLEDGEMENT
load. Thus, the magnitude of the 60 Hz and harmonics were The financial support of Duke Power Company for the
continuously varying. The total harmonic distortion experi- development of digital measurement techniques for power
enced similar variation. system harmonics and the financial support of NSF for the
development of Kalman filtering techniques are acknowledged.
The authors appreciate the help and support of Duke Power
1.2 -/ I engineers, John Dalton and Ray Catoe, in supplying actual
recorded data. Also, the contribution of Michael Clapp to this
project is greatly appreciated.
REFERENCES
IEEE Working Group on Power System Harmonics,
"Power System Harmonics: An Overview", I E E E
Transactions on Power Auparatus and Systems. Vol.
PAS-102, NO. 8, August 1983, ~p.2455-2459.
J. F. Fuller, E. F. Fuchs, and D. J. Roesler, "Influence of
Harmonics on Power Distribution System Protection",
-1.24, , I , , , , , ,I JEEE Transactions on Power Deliverv, Vol. PWRD-3, No.
2, April 1988, pp.549-557.
I. Arrillaga, D. A. Bradley, and P.S. Bodger, P o w e r
0.1 0.2 0.3 0.4 0.5
System Harmonics, John Wiley & Sons, 1985, New
-SEC)
York.
Figure 13. Current waveform of a continuous varying load
S . B. Davan and A. Straughen, Power Semiconductor
Circuits, John Wiley & Sons, 1985, New York.
A. A. Girgis, M. C. Clapp, E. B. Makram, I. Qiu, J. G.
1.0- Dalton, and R. C. Catoe, Jr., "Measurement and
Characterization of Harmonic and High Frequency
0.8 -
Distortion for a Large Industrial Load", I E E E
Transactions on Power Delivery, Vol. PWRD-5. No. 1,
January 1990, pp.427-434.
A. A. Girgis and J. Qiu, "Measurement of the Parameters
of a Slowly Time-Varying High-Frequency Transient",
0.4 IEEE Transactions on Instrumentation and Measurements,
Vol. IM-38, No. 6, December 1989, pp.1057-1063.
0.2 L. C. Ludeman, Fundamentals of Digital Siena1
Processing, Harper and Row, 1986, New York.
J. S . Thorp, A. G. Phadke, and K. I. Karimi, "Real-Time
0.0
Voltage Phasor Measurements for Static-State
0 :1 0:2 Estimation", IEEE Transactions on Power Auuaratus and
W ( S W S v s t e m s , Vol. PAS-104, No. 11, November 1985,
Figure 14. Magnitude of dominant frequencies and harmonic pp.3099-3 106.
distortion of waveform shown in Figure 13 using
the Kalman filtering approach
 1159

G. T. Heydt, A New Method for the Calculation of

" The Fast Fourier Transform
Subtransmission and Distribution System Transients The fast Fourier transform (FFT) was developed to reduce
Based on the FFT", we the computational burden of the DFT. When the period N for a
peliverv, Vol. 4, No. 3, July 1989, pp.1869-1875. waveform is an integer power of two, the determination of the
A. A. Girgis and F. Ham, "A Qualitative Study of Pitfalls Fourier coefficients can be stream lined. The number of opera-
in FFT", IEEE Transactions on Aerosvace and Electronic tion is reduced from N'%? to N x logz(N).
Svstems. Vol. AES 16, No. 4, pp.434-439, July 1980.
APPENDIX B
F. M. Ham and A. A. Girgis, "Measurement of Power
Frequency Fluctuations Using FFT", IEEE Transactions BASIC EQUATIONS AND PRINCIPLES OF THE
on Industrial Electronics, Vol. IE-32, No. 8, August KALMANmLTERINGALGORITHM
1985, pp.199-204. The state equation is expressed as:
R. K. Adams, J. M. McIntyre, and F. W. Symonds,
"Characteristics of Eastern Interconnection Line %+l = @k % -k wk (B-1)
Frequency", IEEE Transactions on Power ADvaratus and where
S v s t e m s . Vol. PAS-101, No. 12, December 1982, xk is an n by 1 process state vector at step k.
pp.4542-4547. qk is an n by n state transition matrix,
R. G. Brown, Introduction to Random Sianal Analvsis
wk represents the discrete variation of the state variables
and Kalman Filtering, John Wiley & Son, 1985, New
due to an input white noise sequence.
York.
The second term in the state equation allows for the time varia-
A. A. Girgis and T. L. D. Hwang, "Optimal Estimation of
Voltage Phasors and Frequency Deviation Using Linear tion of the state variables. It can be described by a covariance
and Non-linear Kalman Filtering: Theory and matrix Qk, where
Limitation". IEEE Transactions on Power Auuaratus and

i
Qk, i = k
S v s t e m s , Vol. PAS-103, No. 10, October 1984,
pp.2943-2951. E [wk w]
: =
0, i+k
A. A. Girgis and R. G. Brown, "Application of Kalman
Filtering in Computer Relaying of Power Systems", The observation (measurement) of the process is assumed to
; Vol. occur at discrete points in time in the form:
PAS-100, NO. 7, July 1981, pp.3387-3397.
M.S. Sachdev, H. C. Wood, and N. G. Johnson, "Kalman
Filtering Applied to Power System Measurements for where
Relaying", IEEE Transactions on Power AuDaratus and zk is an m by 1 vector measurement at step k,
S v s t e m s , Vol. PAS-104, No. 12, December 1985, H, is an m by n matrix giving the ideal (noiseless)
pp.3565-3573. connection between the measurement and the state
P. K. Dash and A. M. Sharaf, "A Kalman Filtering vector,
Approach for Estimation of Power System Harmonics", vk is an m by 1 measurement noise vector assumed to be a
Proceeding of International conference on Power System white sequence with known covariance structure and
Harmonics, 1989, pp.34-40. uncorrelated with the wk sequence.
APPENDIX A The noise is usually described by its variance, Rk. where
BASIC EQUATIONS OF DFT AND FFT ALGORITHMS
q, i=k
Discrete Fourier Transform E [vkv]: = (B-4)
The frequency content of a periodic stationary discrete time 0, i # k
signal with period N can be determined using the discrete To start the Kalman filter recursive estimation. an initial
Fourier transform (DFT). The DFT of the sequence x(n) is
process vector &) and the associated initial covariance matrix
expressed as
(Pi,) are needed. The initial covariance matrix describes, in a
statistical sense, the range of variations of the state vector x
(A-1) from the initial process vector, %. In general, the error
n =O covariance matrix, (P;), associated with an apriori estimate, 2,.
where C2 = 2n/N and k is the frequency index [7]. Considering is defined to be:
frequency index that varies from 1 to (N/2) - 1, the time
domain sequence x(n) can be written as P-k = E [ % ST]
= E[(%-$(%-<f] (B-5)
Having an apriori estimate, ft;, and the associated error covari-
ance matrix, P;, we now wish to optimally improve the
k=l
estimate using the measurement zk. This is achieved by a
linear blending of the noisy measurement and the prior esti-
where XR and XI are the in-phase and quadrature-phase compo-
mate according to the equation (B-6).
nents of X(k), respectively. The phasor quantity at the analog
frequency o k is determined as
X (ok) = XR + j XI NI . (A-3) where
fr, is the updated estimate,
The calculation of all the phasor components requires N2/2
Kk is the blending factor.
operations.
The idea is to find the particular blending factor that yields an
optimal updated estimate. This is achieved by forming first the
1160

expression for the error covariance matrix associated with the Adlv A Gireis (SM '81) received the B.S. (with distinction
updated estimate as: first class honors) and the M.S. degrees in Electrical
Engineering from Assuit University, Egypt, in 1967 and 1973,
Pk = E [ % $ ] = E r ( + Q ( 5 - f l T l 03-71 respectively. He received the Ph.D. degree in Electrical
Engineering from Iowa State University in 1981. From 1967
Now, we wish to find the particular Kk that minimizes the
to 1976, he was an Instructor in the Electrical Engineering
diagonal elements of the matrix Pk, because these elements Department, Assuit University. From 1976 to 1981, he was
represent the estimation error variances of the state vector teaching and performing research in computer relaying of
components. This particular blending factor is called the power systems at Iowa State University. From 1981 to 1985,
Kalman gain and is found to be: he was Assistant Professor of Electrical Engineering at North
Carolina State University. He is currently Professor of
\ = Pi$(qPk$+RJ (B-8) Electrical and Computer Engineering at Clemson University,
The covariance matrix associated with the optimal estimate Clemson, SC. Dr. Girgis has published numerous technical pa-
may now be computed as shown below: pers and holds four U.S. patents. He is the recipient of the
1989 McQueen Quattlebaum-Faculty outstanding achievement
(B-9) award and the 1990 Edison Electric Institute Power Engineering
Education award. His present research interests are real-time
Now, there is a means of assimilating the measurement at tk, computer applications in power system control, instrumenta-
by the use of the optimal Kk, It,, and Pi. At the next step, we tion, and protection, signal-processing, and Kalman filtering
need %Gland Pkilto make an optimal use of zk+l. First, the applications. Dr. Girgis is a member of Phi Kappa Phi, Sigma
updated estimate ?k is projected ahead via the state transition Xi, and is a registered Professional Engineer.
matrix (qk) to obtain the apriori estimate &. Thus,
W. Bin Chanp, received the Electrical Engineering diploma
%+l = @k fl (B-10) from the National Taipei Institute of Technology, Taipei,
Taiwan, Republic of China, in 1979. He received the M.S.
The error covariance matrix associated with. & is then degree in Electrical Engineering from Clemson University,
obtained by forming the expression for the apriori error Clemson, SC. in 1986. From 1979 to 1984, he was an
engineer at CTCI Corp. in Teipei, Taiwan. He is currently
(B-11) working toward his Ph.D. degree. His research interest is real-
time computer applications in power system control and
protection.
'[$+I 1=qkPk@E+q (B-12)
Elham B. Makram (SM '82) was born in Assuit, Egypt.
I t should be noted that the Kalman gain, in the usual linear She received the B.S. degree in Electrical Engineering from
recursive Kalman filter, is independent of measurements. Thus, Assiut University, Egypt in 1969. She received the M.S. and
only equations (B-6) and (B-10) need to be computed on-line. Ph.D. degrees from Iowa State University in 1978 and 1981,
The Kalman gain vector, which is the key parameters. can be respectively. From 1970 to 1976, she was an engineer in
computed off-line. power system planning in Assuit, Egypt. From 1978 to 1981,
she was a research assistant at Iowa State University. From
1981 to 1983, she was a Senior Project Engineer at Siemens-
Allis, Inc., in Raleigh, NC. From 1983 to 1985, she was an
Assistant Professor at North Carolina A & T State University.
She is presently an Associate Professor of Electrical and
Computer Engineering at Clemson University, Clemson, SC.
Dr. Makram is a senior member of IEEE, a member of ASEE and
Sigma Xi.