Application of DFT Filter Bank To Power Frequency Harmonic Measurement

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Application of DFT filter bank to power frequency

harmonic measurement
S.-L. Lu
Abstract: The paper presents a discrete-Fourier-transform (DFT) lter bank for power-system
harmonic measurement. It has the ability to measure harmonic contents in time and frequency
domains, and is ideal for long-term of observation harmonics. It can also overcome shortcomings
of the windowed fast-Fourier-transform (FFT) algorithm. The subject of the DFT lter bank is
closely related to multi-rate digital-signal-processing (DSP) technique. Hence, this method using the
multi-rate technique is very efcient. Finally, an actual eld test is implemented to measure powerfrequency harmonics.

Introduction

Power semiconductor devices are widely used in electric


power distribution systems. Hence, pollution due to
harmonics has worsened owing to their nonlinear load
characteristics. The study of harmonics is a very important
topic for electric power quality. Research topics include
harmonic source analysis and effects, measurement technology and elimination strategy, as well as harmonic modelling
and control [1].
There are two main scopes in power-system harmonic
measurement. One concerns stationary signals, which are
no longer constrained to integer harmonics of power
frequency but also contains noninteger harmonics (such
as subharmonic and interharmonic) measurement. The
other concerns nonstationary signals such as power-quality
transient disturbance assessment. Many different digital
techniques have been applied to measure and estimate
power-system harmonics of stationary signals. These
techniques include FFTs, least-squares method, neural
networks, and lter banks [2, 3].
The nonstationary signal property will change with time.
For example, a capacitor switching transient voltage is a
synthesis of sinusoidal components with time-varying
amplitude, frequencies or phase. Hence, a single Fourier
transform is not suitable for signal analysis, instead, timefrequency distribution is widely used for analysis. It includes
wavelet transform, short-time Fourier transforms, WingerVille transform [4, 5].
This paper examines the harmonic measurement of a
stationary signal. The FFT algorithm is a very effective
method for analysing power-system integer harmonic levels,
however the FFT may produce aliasing, spectral leakage
and picket fence effects [6], thus affecting its direct
application to this eld. However, these effects will be
suppressed by a correct sampling rate in the integer period
[7 ]. This method [7] is ideal for short-time harmonic
observation.
r IEE, 2004
IEE Proceedings online no. 20041131
doi:10.1049/ip-gtd:20041131
Paper rst received 19th March 2003 and in revised form 6th August 2004
The author is with the Department of Electronic Engineering, Kun Shan
University of Technology, Taiwan, ROC

132

A lter bank is often used for performing spectrum


analysis and/or synthesis. When a lter bank is employed to
compute the DFT of a signal, the lter bank is called a
DFT lter bank [810]. Theoretically, a full band signal can
be separated into a number of sub-bands. According to this
concept, a noise-free stationary power system signal can be
spilt into integer harmonic and noninteger harmonic
contents.
This paper focuses on measuring the integer harmonic
contents in time and frequency domains. To obtain
harmonic components requires a comb lter with a narrow
passband at multiples of the fundamental component. The
basic idea of harmonic analysis is to create a DFT lterbank system that will compute values of the harmonic
components from a discrete-time power system signal. In
this study, a lter bank which acts like a narrowband comb
lter is designed. It also has the ability to measure
harmonics in time and frequency domains, simultaneously.
The designed DFT lter bank adopts the multi-rate DSP
technique. The technique has found important application
in the efcient implementation of DSP functions. It allows
the sampling rate to be decreased or increased without
signicant, undesirable effect errors such as quantisation
and aliasing. Furthermore, the multi-rate DSP technique
leads to very efcient implementation, by allowing ltering
to be performed at a much lower rate, which greatly reduces
lter order.
Finally, the proposed DFT lter bank will be applied in
order to measure power frequency harmonics in computer
simulation and actual eld test.

Method for analysis

The DFT lter bank consists of a set of lters arranged in


parallel, as illustrated in Fig. 1. Figure 1 also shows the
steps required for partitioning a discrete-time signal x[n] into
N equally spaced sub-bands, centred at frequencies
ok 2pk/N, for k 0,1,y., N1. (Note that ok represents
the kth harmonic frequency.) For this, we may start with a
lowpass lter (LPF). By passing x[n] through the LPF, the
low-frequency part of its spectrum is extracted. To extract
any order of harmonic of x[n] (e.g. the part centred around
the frequency o ok), we can shift the desired portion of
the spectrum to the baseband (i.e around o 0) by
IEE Proc.-Gener. Transm. Distrib., Vol. 152, No. 1, January 2005

x [n]

LPF

x0 [n]

LPF

x1 [n]

jn 1

jn 1

the interpolator ensures that the input signal and output


signal are at the same sampling rate.
As for the LPF design of Fig. 1, we must use the multirate DSP technique to increase the implementation
efciency. In fact, most practical multi-rate systems employ
the multistage approach, because it allows gradual reduction or increase in sampling rate, leading to a signicant
relaxation in the digital LPF at each stage.
3

xN 1 [n]

LPF

Fig. 1

jn N 1

jn N 1

h0 mD  nxn expjnok ;
1

Proposed measuring concept

m 0; 1; 2; . . . ;

frequency-domain output {Xk [m]}

signal

Fig. 3

simple
analogue
ALLPF

analogue to
digital
conversion

h0 [n]

analysis filter

g0 [n]

xk [m]g 0[n ml ]

synthesis filter
LPF

Fig. 2

Functional block diagram of LPF corresponding to Fig. 1

where D is the decimated factor. The term h0[n] is the


impulse response of the prototype lter of LPF. The output
xk[n] corresponding to Fig. 1 can be expressed as [810]:
X
xk n N1
Xk mg0 n  mI expjnok ;
m
2
n 0; 1; 2; . . . ;
where xk[n] represent the samples of the inverse DFT
sequence corresponding to Xk[m]; g0[n] is the impulse
response of the interpolation lter and I D.
In Fig. 2, the sampling rate of the input signal is rst
reduced as far as possible by decimation, D. The desired
ltering operation is then performed at the low sampling
rate, and nally the sampling rate of the lter data is
restored back to its original rate by interpolation I. The use
of the sampling rate conversion factor at the decimator and
IEE Proc.-Gener. Transm. Distrib., Vol. 152, No. 1, January 2005

{Xk [n]}

3.2 Decimation and interpolation LPF


design using multi-rate DSP technique
A multistage design offers signicant benets in computation and storage requirements over a single stage design. An
optimum number of stages is one which leads to the least
computational effort, such as measuring the number of
multiplications per second (MPS) or the total storage
requirements (TSR) for the coefcients:
J
X
i1

x [n]exp( jnk)

DFT
Filter
Bank

Proposed concept of harmonics measurement

MPS
xk [m]

time-domain output

multiplying x[n] with the complex sinusoidal function ejnok ,


and then use the LPF for extraction of the contents.
The LPF, which is repeatedly used for extraction of
different harmonic contents of the input spectrum, is called
the prototype lter. By using the frequency-shift property of
a spectrum, all channels can share the same LPF. When the
LPF is designed as a narrowband lter, the lter bank will
act like a narrowband comb lter. As seen in Fig. 1, each
channel of the DFT lter bank must do two jobs. First, it
isolates the frequency of harmonics; and, secondly, it
computes the complex amplitude for that frequency. The
job of isolating the frequency is best performed by a
narrowband LPF.
Figure 2 shows a functional block of an LPF corresponding to Fig. 1. Xk {[m]} are the DFT samples at frequencies
ok 2pk/N, and can be expressed as[810]:
X

3.1

Figure 3 shows the proposed harmonic measuring block. It


consists of a simple front-end analogue 3 kHz anti-aliasing
LPF, an analogue-to-digital converter, and the proposed
DFT lter bank. The output of the lter bank has 50
channels in time and frequency domains, respectively.

Block diagram of DFT filter bank

Xk m

Measurement Setup

Mi Fi and TSR

J
X

Mi

i1

where Mi and Fi are the number of lter coefcients (also


lter length) and the output sampling rate for the ith stage,
respectively.
Table 1 shows the analytical results of three different
decimators using empirical formula from Herrman et al.
[11]. It is clear that a multi-stage design yields a signicant
reduction in both MPS and TSR compared with a singlestage design. The reductions are due to the wide transitions
of lters at the early stages, leading to small values of lter
coefcients.
Figure 4 shows a designed function block of a
narrowband LPF corresponding to Table 1. It consists of
both three-stage decimation and interpolation processes.
The overall decimation factor D is expressed as the smaller
factor: D D1D2D3 (where D1 25, D2 4, D3 2). The
narrowband LPF can extract the required spectrum in the
frequency domain from the decimator. Furthermore, the
overall interpolation factor I I3I2I1 (where I1 25, I2 4,
I3 2). The interpolation process is used for reconstructing
each harmonic signal back to the time domain. Hence,
interpolation lters have the same specications as decimation lters.
In Fig. 4, the required lters LPF1, LPF2 and LPF3 are
designed as narrowband nite impulse response (FIR)
LPFs. The desired specication of these FIR LPFs are also
shown in Table 1, where it can be seen that the three-stage
133

Table 1: Summary of the efficient of decimators


Number
of stages

MPS

TSR

Decimation
factor Di

Filter
length

Passband
edge (Hz)

1 730 450

34 609

200

34 609

96 850

1 454

25

69

10

0.0008

0.0001

0.005

1 385

0.0025

0.0001

0.005

62 200

588

Stopband
edge (Hz)

Normalized
transition width

Passband
ripple

Stopband
ripple

0.0001

0.0001

0.005

25

69

500

0.0498

0.0001

0.005

173

10

0.02

0.0001

0.005

346

0.01

0.0001

0.005

analysis filter
10 kHz
x [n]exp( jnk)
LPF1

I3
2

400 Hz
D1
25

LPF3

100 Hz

100 Hz
D2

LPF2

I2
4

50Hz
D3

LPF3

I1

LPF2

25

400 Hz

xk [m]

LPF1

xk [m]g 0[n ml ]

10 kHz

synthesis filter

Fig. 4

Proposed multi-rate 3-stage LPF corresponding to Fig. 2

decimation process achieves ltering with a very short


bandwidth and a low computational load in both MPS and
TSR.
In Fig. 1, every channel lter has a maximum bandwidth
of 2 Hz and a transition band from 2 to 3 Hz. The
maximum passband ripple and stop band ripple must be
0.0001 and 0.005, respectively. To achieve this specication
with a single lter and no decimation, an equiripple lter of
length 34609 is required, a very long lter with high
computational load in both MPS and TSR.
This study uses a total decimation factor of 200 to
decimate the signal in three stages (D1 25, D2 4,and
D3 2). The rst stage must employ an LPF with a
normalised cutoff of 1/25. The required lter length 69 is
obtained from the Parks-McClellan algorithm [10]. The
second- and third-stage lter lengths required are 173 and
346, respectively. The total number of lter length is 588.
The computation rate has also been reduced dramatically,
because the rst-stage lter runs at 1/25 the sampling rate of
the input signal, and the second- and third-stage lters run
at 1/100 and 1/200 of this rate, respectively.

20 ms time resolution. They also represent X1[m], X5[m],


X7[m], X11[m] and X13[m] of Fig. 3.

4.2

Computer simulation and field test

The following two cases measure power frequency harmonics using the proposed DFT lter bank.

4.1

Case I: Computer simulation

Figure 5 shows a computer simulation 3-phase 220 V 60 Hz


6-pulse AC/DC converter original current waveform with
10kHz sampling rate. Figure 6 shows some split harmonic
components with 0.1 ms time resolution from the current
waveform. They are 1st , 5th, 7th, 11th, and 13th
harmonics, also x1[n], x5[n], x7[n], x11[n] and x3[n] of
Fig. 3. Owing to space limitation, only these are shown.
Figure 7 shows the corresponding split harmonic spectral
magnitude samples of the DFT lter bank at different
harmonic contents with 4 Hz frequency resolution and
134

40
30

current, A

20
10
0
10
20
30
40
4.90 4.91 4.92 4.93 4.94 4.95

4.96 4.97 4.98

4.99

5.00

time, s

Fig. 5

A 6-pulse AC-DC converter current waveform

Case II: Actual field test

In this case, data acquisition is implemented by a


commercial digital instrument, and sent to a PC by IEEE488 GPIB. In order to avoid picket-fence effect of a DFT,
the fundamental power frequency measurement is required
before adjusting the sampling rate of the instrument. It is
also closely related to the DFT lter bank design.
In this case, the power source frequency is 60 Hz. Figure
8 shows an actual recorded nonlinear load current with
10 kHz sampling rate. Figure 9 shows some split harmonic
components with 0.1 ms time resolution in time domain,
also x1[n], x3[n], x5[n], x7[n] and x11[n]. Figure 10 shows the
corresponding split harmonics spectral magnitude samples
of the DFT lter banks at different harmonics with 4 Hz
frequency resolution and 20 ms time resolution. They are
X1[m], X3[m], X5[m], X7[m] and X11[m] of Fig. 3. It is clear
IEE Proc.-Gener. Transm. Distrib., Vol. 152, No. 1, January 2005

5.00

5 th

0
4
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99

5.00

5
7 th

0
5
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99

5.00

11 th

0
5
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99

5.00

2
0
3 rd
2
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99 5.00
2
0

5 th

2
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99 5.00
0.5
7 th

0.5
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99 5.00
0.2
11 th
0
0.2
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99 5.00

current, A

time, s

5.00

Some spilt harmonic contents

35
1 st

current, A

current, A

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

5.0

8
6
4.0

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

5.0

7 th
0
4.0

3
2
1
0
4.0

5 th
4.1

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

5.0

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

5.0

Fig. 7

4.1

4.2

4.3

4.4

4.5 4.6
time, s

4.7

4.8

4.9

5.0

0.2

10
8
6
4
2
0
2
4
6
8
4.96 4.97 4.98

4.99

time, s

Fig. 8

A 60 Hz actual nonlinear load current waveform

IEE Proc.-Gener. Transm. Distrib., Vol. 152, No. 1, January 2005

5.00

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

3 rd
4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

5.0

7 th
4.1

4.2

4.3

4.4

4.5

4.6

4.7

4.8

4.9

0.1
0
4.0

5.0

5 th

0.2
0
4.0

5.0

5.0

11 th
4.1

4.2

4.3

4.4

4.5 4.6
time, s

4.7

4.8

4.9

5.0

DFT spectral samples versus time

that the magnitude of high order harmonic content does not


have any xed value in the actual case.
5

DFT spectral samples versus time

10
4.90 4.91 4.92 4.93 4.94 4.95

0
4.0

Fig. 10

13 th
0
4.0

1 st
4.1

0.4

11 th
0
4.0

Some spilt harmonic contents

15
10
5
0
4.0

current, A

current, A

10

current, A

30
4.0

current, A

5
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99

Fig. 9

current, A

13 th

current, A

Fig. 6

current, A

current, A

time, s

current, A

current, A

50
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99

current, A

1 st

current, A

10
0
1 st
10
4.90 4.91 4.92 4.93 4.94 4.95 4.96 4.97 4.98 4.99 5.00

current, A

current, A
current, A
current, A
current, A
current, A

50

Conclusions

In Table 1, FIR LPFs are characterised by sharp transitions


between the passbands and stopbands, and by passbands
which are very small compared with the sampling rate. As a
result, narrowband FIR LPFs require a very large number
of coefcients. This poses a problem in the design and
implementation of such lters, because they are highly
susceptible to nite wordlength effects. Further, large
coefcient storage requirement and computational effort
are required. The multi-rate DSP application can overcome
these problems successfully.
Using the DFT lter-bank structure, we only need to
measure the fundamental power frequency for different
harmonic contents, because the frequencies of the higherorder harmonic components are exactly integer multiples of
the fundamental power frequency. The power frequency
and sampling rate were 60 Hz and 10 KHz, respectively. If
the power frequency is not exactly 60 Hz, the sampling rate
of power system signal must be adjusted.
135

This paper adopts a ltering viewpoint for harmonics


analysis, because it allows us to build upon our knowledge
of the frequency response. Band-partitioning is employed to
improve windowed FFT algorithm performance, such as
spectral leakage, picket fence effect and aliasing effects.
Furthermore, the DFT lter bank can measure harmonic
contents in both time and frequency domains up to the 50th
harmonic, simultaneously. Hence, it is ideal for power
system harmonic measurement and long-term harmonic
observation.
6

References

1 Arrillaga, J., Bradley, D.A., and Bodger, P.S.: Power system


harmonics, (John Wiley & Sons, New York, 1985)
2 Heydt, G.T.: Electric power quality, (Stars in a Circle Publications,
West LaFayette, IN, USA, 1991)
3 Sun, H., Allen, G.H., and Cain, G.D.: A new lter-bank conguration for harmonic measurement, IEEE Trans. Instrum. Meas., 1996,
45, (3), pp. 739743

136

4 Gu, Y.H., and Bollen, H.J.: Time-frequency and time-scale domain


analysis of voltage disturbances, IEEE Trans. Power Deliv., 2000, 15,
(4), pp. 12791283
5 Wright, P.S.: Short-time fourier transforms and WignerVille
distributions applied to the calibration of power frequency harmonic
analyzers, IEEE Trans. Instrum. Meas., 1999, 48, (2), pp. 475478
6 Girgis, A.A., and Ham, F.M.: A quantitative study of the pitfalls in
the FFT, IEEE Trans. Aerosp. Electron. Syst., 1980, AES-16, (4),
pp. 434438
7 Lu, S.L., Lin, C.E., and Huang, C.L.: Power frequency harmonic
measurement using integer periodic extension method, Electr. Power
Syst. Res., 1998, 44, pp. 107115
8 Vaidyanathan, P.P.: Multirate systems and lter banks, (PrenticeHall, Englewood Cliffs, NJ, 1993)
9 Crochiere, R.E., and Rabiner, L.R.: Multirate digital signal
processing, ( Prentice-Hall, Englewood Cliffs, NJ, 1983)
10 Oppenheim, A.V., and Schafer, R.W.: Discrete-time signal processing, ( Prentice-Hall, NJ, 1999)
11 Herrman, O., Rabiner, L.R., and Chan, D.S.K.: Practical design rules
for optimum nite impulse digital lter, Bell Syst. Tech. J., 1973, 52,
pp. 769799

IEE Proc.-Gener. Transm. Distrib., Vol. 152, No. 1, January 2005

You might also like