Page 1 of 21 The Journal of Engineering

This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Development of Interharmonics Identification

Using enhanced-FFT Algorithm
Hsiung Cheng Lin
Department of Electronic Engineering, National Chin-Yi University of Technology, Taiwan
Email: [email protected]

Abstract: The Fast Fourier transform (FFT) is a mostly used tool to measure power system harmonics. FFT, however,
is not applicable to analyze interharmonics due to spectral leakage effect. Although International Electrotechnical
Commission (IEC) standard is recommended for interharmonic measurement, the individual interharmonic frequency
and respective amplitude can not be worked out under this framework. For this reason, this paper proposes an
enhanced-FFT model to build up the relationship between interharmonic frequency and dispersed leakage energy. The
mathematical equation is thus established to find actual value of interharmonic frequency. Also, the true interharmonic
amplitude can be retrieved from the dispersed energy collection. In other words, the sampling window length is no
longer required to match the interharmonic period and the correct measurement results can be achieved. The proposed
model is developed using a simple arithmetic equation so that it is feasible for more efficient calculation for
interharmonic analysis. Performance results verify that the proposed scheme can achieve accurate, rapid and reliable
outcomes.

Keywords: FFT, enhanced-FFT, Harmonics, Interharmonics, Non-stationary

I. Introduction
With an increasing number of power electronics facilities used in industry, the power line pollution has been
seriously deteriorated due to harmonics generated in electric power system. The situation even worsens when the
applications of periodical time-varying apparatuses grow and produce interharmonics sequentially in recent years. It is
well known that interharmonics frequencies are not an integer of the fundamental components, and it is thought as the
inter-modulation between the fundamental and harmonic components in the system. Major sources have been found in
cycloconverters, wind turbine, double conversion system, time-varying loads, variable-load electric drives and
unexpected sources [1–3]. In addition to typical problems caused by harmonics, interharmonics bring new problems
such as thermal effects, cathode ray tube (CRT) flicker, saturation of current transformers, low-frequency oscillation in
a mechanical system, voltage fluctuations, subsynchronous oscillation, etc. Even under a low amplitude of
interharmonic, the above phenomena may still exist [4–6].
Although FFT is still the most popular method in harmonics analysis, incorrect results may arise if the sampling
window length is not properly chosen. When the measured waveforms contain interharmonics, FFT will suffer from
low accuracy and less computational efficiency [7–9]. For this reason, an adaptive window width approach was
announced to estimate harmonics/interharmonics [10]. However, a large initial value is required for the generic
waveform to avoid deceptively strong correlation. A long computational time during the iteration process is therefore
needed. Another way that combined Prony-based and downsampling techniques was reported for harmonics and
interharmonics measurement [11]. Unfortunately, the selection of the downsampling coefficient and estimation order is
difficult to be formulated because of sensitivity to the measured signal. The Exact Model Order ESPRIT algorithm
based on the RD plot is another approach that improved the sliding-window ESPRIT method, but its computational
time was a big concern in practice [12]. An interpolated DFT can reduce the spectral leakage and thus find the correct

IET Review Copy Only

 The Journal of Engineering Page 2 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

parameter values of signals [13]. The measurement accuracy, however, may be influenced by the location of the
interharmonic frequency component. The sliding-window ESPRIT algorithm was proposed for the frequency
estimation of interharmonics [14]. In this model, the interharmonic number is required prior to implementation, and it
may lead to spurious components, line splitting and occasional failure. Recently, a new approach using single channel
independent component analysis for both harmonics and interharmonics was reported. The orthogonal vector of the
proposed model may be unconvinced in convergence [15]. Recently, a group-harmonic power minimizing algorithm
for harmonics and interharmonics estimation was reported [16]. Although this method can achieve an accurate
measurement, a searching procedure was required to reach the solution. Some studies have applied neural network
models to carry out both harmonic/interharmonic analysis [17-19]. However, their initial parameter settings usually
rely on trial and error. This is somehow discouraged in a real application.
In deed, the presence of interharmonics poses measurement more difficulties for some reasons: (1) very low values
of interests of interharmonics (about one order of quantity smaller than harmonics); (2) variability of frequency, and
amplitude; (3) variability of the waveform periodicity; (4) great sensitivity to the spectral leakage phenomenon. To
address aforementioned issues, a guideline for interharmonics measurement based on grouping concept was suggested
by IEC 61000-4-7 standard [20]. A 5Hz frequency resolution with rectangular window is recommended to be adopted,
but individual interharmonic information is unavailable from such a measurement.

II. Enhanced-FFT (e-FFT) model

II.1 Background of Fourier transformation
A periodical waveform can be reconstructed by series harmonic components via the Fourier transform (FT)
analysis, where harmonic frequency is defined as a multiple of fundamental. Assume a waveform i s (t ) is periodical
with satisfaction of Dirichlet condition, it can be expressed as

∞
is ( t ) = ∑i e
n =- ∞
n
j 2 πnft
(1)

T
1
where in = ∫ is ( t )e − j 2πnft dt , and T(= 1 f ) is the signal period. i0 is the dc component.
T 0

For performing FT, Discrete Fourier Transform (DFT) is then introduced as

N −1
is [n] = ∑ I s [k ]WNkn (2)
k =0

1 N −1
where I s [ k ] = ∑
N n =0
is [n]W N−kn , and W N = exp( j 2π / N ) .

The Fourier fundamental angular frequency ( ∆ω ) for i s [n] with the period T is defined as
2π
∆ω = (3)
T
When the waveform is sampled using p(p>1) periods, ∆ω can be represented as
2π ω0
∆ω = = (4)
pT p
where ω 0 = 2π .
T

IET Review Copy Only

Page 3 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

From equ. (4), the Fourier fundamental frequency ( ∆f ) is defined as

1 1 1 f
∆f = = = = s (5)
pT pN s Ts NTs N
∆ N ∆ 1
where N s = , and Ts = . It is noted that N points is sampled using the sampling rate f s .
p fs

The waveform power (P) can be expressed by the Parseval relation as [21-22]
N −1 N −1
1
P=
N
∑ is [ n]2 = ∑ I s [ k ]2
n =0 k =0
(6)

The power at the frequency f k can be expressed as

P[ f k ] = I s [k ] 2 + I s [ N − k ] 2 = 2 I s [k ] 2 (7)

where k=0,1, 2,…,N/2-1.

The amplitude of the m th harmonic at f k is then calculated as

Am [ f k ] = P[ f k ] = 2 I s [k ] (8)
where m=1,2,…,M.

It is known that interharmonics contained in a waveform is not synchronized with the fundamental. As a result, the
th
m harmonic power at f k disperses over around the f k . Based on the concept of group harmonics, all spilled
power around the adjacent harmonics can be collected into a “group power” as [20]
+τ
Pm* [ f k ] = ∑τ( A [ f
∆k = −
m k + ∆k ]) 2 (9)

where τ denotes the group bandwidth.

As above, the true harmonic amplitude can be retrieved from collection of all dispersed power as
A [ f k ] = Pm* [ f k ]
*
m (10)

II.2 e-FFT algorithm

The proposed e-FFT model is based on the FFT for suiting interharmonics measurement. According to the FFT
analysis, the relationship between harmonic frequency and dispersed energy can be classified into two cases: small and
big frequency deviation [23]. Case 1: Depiction for small frequency deviation is shown in Fig. 1(a), where the second
larger magnitude ( Am [ f k +1 ] ) at f k +1 is located at the right side of the dominant frequency at f k with
Am [ f k ] > Am [ f k +1 ] . In view of practice, however, f k may be wrongly interpreted as the dominant harmonic. The
situation is that the true frequency is equal to f k plus the “frequency deviation” ( ∆f k ) defined in equ.(11). On the
other hand, it reveals that higher Am [ f k +1 ] introduces more amount of deviation ( ∆f k ) distant from f k . Case 2:
Depiction for a big frequency deviation is shown in Fig. 1(b), where the second larger amplitude ( Am [ f k ] ) at f k is
located at the left side of the dominant frequency at f k +1 with Am [ f k ] < Am [ f k +1 ] . The f k +1 may be wrongly
interpreted as the dominant harmonic frequency in this case. In deed, the true frequency should be equal to f k plus

IET Review Copy Only

 The Journal of Engineering Page 4 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

the “frequency deviation” ( ∆f k ). Higher Am [ f k +1 ] also introduces more amount of deviation ( ∆f k ) distant from f k .
As illustrated above, both small and big frequency deviation are defined based on the generic phenomenon of
frequency deviation condition, where “small” and “big” only denotes the deviation status rather than a real value.

(a) Small frequency deviation

(b) Big frequency deviation

Fig. 1 Relation between harmonic frequency and dispersed Energy [23]

The e-FFT model is formulated from the relation between the frequency deviation amount and dispersed energy
distribution [6]. It is induced that the real frequency can be represented by the dominant frequency ( f k ) plus
“frequency deviation” ( ∆f k ), i.e., f k + ∆f k .

The Frequency Deviation Range (F.D.R.) is defined as

IET Review Copy Only

Page 5 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

+τ

∑A
∆k =1
m [ f k + ∆k ]2
∆f k = ⋅ ∆f (11)
0 +τ

∑A
∆k = −τ
m [ f k + ∆k ] + 2
∑A
∆k =1
m [ f k + ∆k ] 2

fs
where ∆f is determined by N and f s due to ∆f = , and τ = 0,1,2,3,.....
N
From the group-harmonic concept, the dispersed energy around the major harmonic can be efficiently collected
for retrieving the original amplitude [20]. Thus, the Restored Amplitude (R.A.) can be defined as

R.A.
+τ
= ∑τA [ f
∆k = −
m k + ∆k ]2 (12)

where τ = 0 ,1,2 ,3,.....

The following example is used to demonstrate the calculation of ∆f and R.A. Assume the signal, i.e.,
i (t ) = 0.25 sin( 2π ⋅ 33 ⋅ t ) + 1.0 sin(2π ⋅ 238 ⋅ t ) , contains two major harmonics located at 33Hz and 238Hz, and their
amplitudes are set as 0.3 and 1.0, respectively. Please note that f s =1.28kHz, N=256 and τ = 5 , where ∆f =5Hz. The
waveform and its spectrum using FFT are shown in Fig. 2(a) and Fig. 2(b), respectively. At both 33Hz and 238Hz, it
can be seen that some energy are dispersed around the neighbor sides considerably.

(a) Waveform (b) Spectrum

Fig. 2 Spectrum of v(t ) using FFT

At 33Hz, the F.D.R. beyond 30Hz can be calculated using equ. (11) as
0.17 2 + 0.037 2 + 0.0152 + 0.00782 + 0.0057 2
∆f k = ⋅∆ f
0.037 2 + 0.0412 + 0.0482 + 0.066 2 + 0.14 2 + 0.17 2 + 0.037 2 + 0.0152 + 0.00782 + 0.0057 2
0.175
= ⋅ 5 ≈ 2.5 (Hz) (13)
0.175 + 0.175
R.A.
= 0.037 2 + 0.0412 + 0.0482 + 0.066 2 + 0.14 2 + 0.17 2 + 0.037 2 + 0.0152 + 0.00782 + 0.0057 2
≈ 0.24 (14)

IET Review Copy Only

 The Journal of Engineering Page 6 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

This interharmonic frequency (33Hz) is found equal to 30Hz ( f k ) plus 2.5Hz ( ∆f k ), close to the actual frequency
value (33Hz). Its R.A. is 0.24 that is close to the actual amplitude value (0.25).

At 238Hz, the F.D.R. beyond 235Hz can be calculated using equ. (11) as

0.752 + 0.212 + 0.12 2 + 0.086 2 + 0.066 2

∆f k = ⋅∆ f
0.069 2 + 0.087 2 + 0.12 2 + 0.19 2 + 0.512 + 0.752 + 0.212 + 0.12 2 + 0.086 2 + 0.0662
0.8
= ⋅ 5 ≈ 2.9 (Hz) (15)
0.57 + 0.8

R.A.
= 0.069 2 + 0.087 2 + 0.12 2 + 0.19 2 + 0.512 + 0.752 + 0.212 + 0.12 2 + 0.086 2 + 0.066 2
≈ 0.98 ≅ 1.0 (16)

This interharmonic frequency (238Hz) is found equal to 235Hz ( f k ) plus 2.9Hz ( ∆f k ), very close to the actual
frequency value (238Hz). Its R.A. is 0.98 that is very close to the actual amplitude value (1.0).
The τ is set as 5 for the above case. In reality, the selection of group bandwidth ( τ =1~5) should consider
harmonics locations to avoid overlapping between each other. Based on this principle, the e-FFT model is formulated
by the following rule:

f1 − f 2 < 4∆f ⇒ τ =1
4∆f ≤ f1 − f 2 < 6∆f ⇒ τ =2
6∆f ≤ f1 − f 2 < 8∆f ⇒ τ =3
8∆f ≤ f1 − f 2 < 10∆f ⇒ τ =4
f1 − f 2 ≥ 10∆f ⇒ τ =5
where f1 and f 2 are assumed as two arbitrary near major harmonics in the waveform.

The flowchart of the proposed e-FFT model is shown in Fig. 3, and its performance procedure is demonstrated as
follows [24].

IET Review Copy Only

Page 7 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Start

Determine f s , N
Sample signal

FFT

Number of major
harmonics=M

Define f1 , f 2

YES
f1 − f 2 < 4∆f τ =1
?

NO

YES
4∆f ≤ f1 − f 2 < 6∆f
?
τ =2
?

NO

YES
6∆f ≤ f1 − f 2 < 8∆f τ =3
?

NO

YES
8∆f ≤ f1 − f 2 < 10∆f τ =4
?

NO

?τ =5

Find ∆f k , R.A.

Exclude f1
M=M-1

NO
M=0 ?

YES
Stop

Fig. 3 Flowchart of the proposed e-FFT model

IET Review Copy Only

 The Journal of Engineering Page 8 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

1. Select f s , N and sample the signal.

2. Perform FFT.
3. Determine the number (M) of major harmonics.
4. Define the biggest amplitude and second big amplitude at f1 and f 2 , respectively.
5. Check if f1 − f 2 < 4∆f . If yes, select τ =1 and go to Step 10. Otherwise, go to next step.
6. Check if 4∆f ≤ f1 − f 2 < 6∆f . If yes, select τ =2 and go to Step 10. Otherwise, go to next step.
7. Check if 6∆f ≤ f1 − f 2 < 8∆f . If yes, select τ =3 and go to Step 10. Otherwise, go to next step.
8. Check if 8∆f ≤ f1 − f 2 < 10∆f . If yes, select τ =4 and go to Step 10. Otherwise, go to next step.
9. Select τ =5.
10. Calculate ∆f k , R.A.
11. Exclude the f1 component (the biggest amplitude) that has been identified, and M=M-1.
12. Check if M=0. If yes, the procedure stops. Otherwise, go back to Step 4.

III. Model verification

III.1 Model performance
To verify the proposed model, the waveform s (t ) is tested.
s ( t ) = a1 sin ( 2π ⋅ f1 ⋅ t ) + ai1 sin ( 2π ⋅ f i1 ⋅ t ) + ai 2 sin ( 2π ⋅ f i 2 ⋅ t ) + ai 3 sin ( 2π ⋅ f i 3 ⋅ t )
+ ai 4 sin( 2π ⋅ f i 4 ⋅ t ) + ai5 sin( 2π ⋅ f i 5 ⋅ t ) (17)
where a1 = 1.0 is the amplitude of the fundamental, and its respective frequency is f 1 = 50Hz . The amplitudes of
interharmonics are ai1 = 0.3, ai 2 = 0.4, ai 3 = 0.2, ai 4 = 0.2, ai 5 = 0.3 , and their respective frequencies are
f i1 = 68Hz, f i 2 = 96Hz, f i 3 = 134Hz, f i 4 = 183Hz, f i 5 = 253Hz .

The waveform of s (t ) shown in Fig. 4(a), and its spectrum using FFT is shown in Fig. 4(b). It is obvious that a
considerable dispersed power always comes with interharmonics and thus causes incorrect results.

(a) Waveform of s (t )

IET Review Copy Only

Page 9 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

(b) Spectrum of s (t ) using FFT

Fig. 4 Analysis of s (t ) with ∆f =5Hz

III.1.1 Analysis with ∆f =5Hz

The parameters of e-FFT model are set as f s =1.28kHz, N=256, i.e., ∆f =5Hz. The interharmonics analysis
using e-FFT model is carried out by the following five cases from equ. (17). According to the rule of group bandwidth
( τ ) selection, τ is chosen as follows.

(i) For f i1 = 68Hz ,

f i1 − f 1 = 68 − 50 < 4∆f (= 20) ⇒ τ =1

(ii) For f i 2 = 96 Hz ,
4 ∆f ( = 20) ≤ f i 2 − f i1 = 96 − 68 < 6∆f ( = 30) ⇒ τ =2

(iii) For f i 3 = 134 Hz ,

6∆f ( = 30) ≤ f i 3 − f i 2 = 134 − 96 < 8∆f ( = 40) ⇒ τ =3

(iv) For f i 4 = 183Hz ,

8∆f ( = 40) ≤ f i 4 − f i 3 = 183 − 134 < 10∆f ( = 50) ⇒ τ =4

(v) For f i 5 = 253Hz ,

f i 5 − f i4 ( = 253 − 183) ≥ 10∆f ( = 50) ⇒ τ =5

Consequently,

Case 1: ai1 = 0.3, f i1 = 68Hz , τ = 1

0.212
∆f k = ⋅∆ f
0.0752 + 0.17 2 + 0.212
0.21
= ⋅ 5 ≈ 2.7 (Hz) (18)
0.186 + 0.21

IET Review Copy Only

 The Journal of Engineering Page 10 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

R.A.
= 0.075 2 + 0.17 2 + 0.212
≈ 0.28 (19)

The measured frequency is equal to f k = 65Hz plus ∆f k = 2.7Hz, i.e., 67.7Hz, very close to the real one (68Hz).
On the other hand, the measured amplitude is about 0.28 that is also close to the real one (0.3).

Case 2: ai 2 = 0.4 , f i 2 = 96Hz , τ = 2

0.093 2 + 0.0412
∆f k = ⋅∆ f
0.0612 + 0.37 2 + 0.093 2 + 0.0412
0.10
= ⋅ 5 ≈ 1.0 (Hz) (20)
0.38 + 0.10

R.A.
= 0.0612 + 0.37 2 + 0.093 2 + 0.0412
≈ 0.39 (21)

The measured frequency is equal to f k = 95Hz plus ∆f k = 1.0Hz, i.e., 96Hz, same as the real one (96Hz). The
measured amplitude is about 0.39 that is very close to the real one (0.4).

Case 3: ai 3 = 0.2 , f i 3 = 134Hz , τ = 3

0.18 2 + 0.022 2 + 0.0068 2

∆f k = ⋅∆ f
0.023 2 + 0.029 2 + 0.055 2 + 0.18 2 + 0.022 2 + 0.0068 2
0.18
= ⋅ 5 ≈ 3.66 (Hz) (22)
0.066 + 0.18

R.A.
= 0.023 2 + 0.029 2 + 0.055 2 + 0.18 2 + 0.022 2 + 0.0068 2
≈ 0.19 (23)

It is found that the measured frequency is equal to f k = 130 Hz plus ∆f k = 3.66Hz, i.e., 133.66Hz, almost same
as the real one (134Hz). The measured amplitude is about 0.19 that is very close to the real one (0.2).

Case 4: ai 4 = 0.2 , f i 4 = 183Hz , τ = 4

0.15 2 + 0.039 2 + 0.02 2 + 0.012 2

∆f k = ⋅∆ f
0.015 2 + 0.023 2 + 0.039 2 + 0.12 + 0.15 2 + 0.039 2 + 0.02 2 + 0.012 2
0.155
= ⋅ 5 ≈ 2.87 (Hz) (24)
0.115 + 0.155

R.A.
= 0.015 2 + 0.0232 + 0.039 2 + 0.12 + 0.15 2 + 0.039 2 + 0.02 2 + 0.012 2

10

IET Review Copy Only

Page 11 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

≈ 0.19 (25)

The above results indicate that the measured frequency is equal to f k = 180 Hz plus ∆f k = 2.87Hz, i.e.,
182.87Hz, almost same as the real one (183Hz). The measured amplitude is about 0.19 that is very close to the real one
(0.2).

Case 5: ai 5 = 0.3 , f i 5 = 253Hz , τ = 5

0.232 + 0.07 2 + 0.042 2 + 0.0312 + 0.024 2

∆f k = ⋅∆ f
0.012 2 + 0.018 2 + 0.028 2 + 0.0512 + 0.152 + 0.232 + 0.07 2 + 0.042 2 + 0.0312 + 0.024 2
0.249
= ⋅ 5 ≈ 3.06 (Hz) (26)
0.158 + 0.249

R.A.
= 0.012 2 + 0.0182 + 0.0282 + 0.0512 + 0.152 + 0.232 + 0.07 2 + 0.042 2 + 0.0312 + 0.024 2
≈ 0.3 (27)

As above, the measured frequency is equal to f k = 180 Hz plus ∆f k = 2.87Hz, i.e., 182.87Hz, almost same as
the real one (183Hz). The measured amplitude is about 0.19 that is very close to the real one (0.2).
The measured spectrum using e-FFT is shown in Fig. 5, indicating no dispersed power around
harmonics/interharmonics.

Fig. 5 Spectrum of s (t ) using e-FFT with ∆f =5Hz

The comparison between FFT and e-FFT is concluded in Fig. 6. As can be seen, the results from e-FFT model are
almost identical to the real values for either amplitude or frequency identification, but traditional FFT is unable to
achieve a correct analysis except at fundamental component (50Hz). The maximum error for amplitude and frequency
estimation using FFT is up to 30% and 4.41%, respectively. By contrast, the maximum error for amplitude and
frequency estimation using e-FFT is only 6.67% and 0.44%, respectively.

11

IET Review Copy Only

 The Journal of Engineering Page 12 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

1.2

1 Real value FFT e-FFT

0.8

Amplitude
0.6

0.4

0.2

0
50 68 96 134 183 253
Frequency (Hz)
(a) Amplitude measurement

300 Real value FFT e-FFT

250
Measured result (Hz)

200

150

100

50

0
50 68 96 134 183 253
Harmonics/Interharmonics
(b) Frequency measurement
Fig. 6 Comparisons of FFT and e-FFT with ∆f =5Hz

III.1.2 Analysis with ∆f =2.5Hz

The parameters of e-FFT model are set as f s =1.28kHz, N=512, i.e., ∆f =2.5Hz. The interharmonics analysis
using e-FFT model is carried out by the following five cases from equ. (17). According to the rule of group bandwidth
( τ ) selection, τ is chosen as follows. The spectrum of s(t ) using FFT is shown in Fig. 7. It is obvious that a
considerable dispersed power always comes with interharmonics and thus causes incorrect results.

12

IET Review Copy Only

Page 13 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

Fig. 7 Spectrum of s (t ) using FFT with ∆f =2.5Hz

(i) For f i1 = 68Hz ,

6∆f (= 15) ≤ f i1 − f1 = 68 − 50 < 8∆f (= 20) ⇒ τ =3

(ii) For f i 2 = 96 Hz ,
f i 2 − f i1 = 96 − 68 ≥ 10∆f (= 25) ⇒ τ =5

(iii) For f i 3 = 134 Hz ,

f i 3 − f i 2 = 134 − 96 ≥ 10∆f (= 25) ⇒ τ =5

(iv) For f i 4 = 183Hz ,

f i 4 − f i 3 = 183 − 134 ≥ 10∆f (= 25) ⇒ τ =5

(v) For f i 5 = 253Hz ,

f i 5 − f i4 (= 253 − 183) ≥ 10∆f (= 25) ⇒ τ =5

Case 1: ai1 = 0.3, f i1 = 68Hz , τ = 3

0.06 2 + 0.022 2 + 0.015 2

∆f k = ⋅∆ f
0.036 2 + 0.057 2 + 0.29 2 + 0.06 2 + 0.022 2 + 0.015 2
0.07
= ⋅ 2.5 ≈ 0.47 (Hz) (28)
0.3 + 0.07

R.A.
= 0.036 2 + 0.057 2 + 0.29 2 + 0.06 2 + 0.022 2 + 0.015 2
≈ 0.3 (29)

The measured frequency is equal to f k = 67.5Hz plus ∆f k = 0.47Hz, i.e., 67.97Hz, almost same as the real one
(68Hz). On the other hand, the measured amplitude is about 0.3 that is also same as the real one (0.3).

Case 2: ai 2 = 0.4 , f i 2 = 96Hz , τ = 5

13

IET Review Copy Only

 The Journal of Engineering Page 14 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

0.2 2 + 0.074 2 + 0.044 2 + 0.0312 + 0.023 2

∆f k = ⋅∆ f
0.027 2 + 0.035 2 + 0.0512 + 0.087 2 + 0.32 + 0.2 2 + 0.074 2 + 0.044 2 + 0.0312 + 0.0232
0.22
= ⋅ 2.5 ≈ 1.0 (Hz) (30)
0.32 + 0.22

R.A.
= 0.027 2 + 0.035 2 + 0.0512 + 0.087 2 + 0.3 2 + 0.2 2 + 0.074 2 + 0.044 2 + 0.0312 + 0.023
≈ 0.39 (31)

The measured frequency is equal to f k = 95Hz plus ∆f k = 1.0Hz, i.e., 96Hz, same as the real one (96Hz). The
measured amplitude is about 0.39 that is very close to the real one (0.4).

Case 3: ai 3 = 0.2 , f i 3 = 134Hz , τ = 5

0.16 2 + 0.048 2 + 0.03 2 + 0.022 2 + 0.018 2

∆f k = ⋅∆ f
0.00008 2 + 0.00014 2 + 0.000312 + 0.0012 + 0.00912 + 0.16 2 + 0.048 2 + 0.03 2 + 0.022 2 + 0.018 2
0.17
= ⋅ 2.5 ≈ 1.57 (Hz) (32)
0.1 + 0.17

R.A.
= 0.00008 2 + 0.00014 2 + 0.000312 + 0.0012 + 0.00912 + 0.16 2 + 0.048 2 + 0.032 + 0.022 2 + 0.018 2
≈ 0.196 (33)

The measured frequency is equal to f k = 132.5Hz plus ∆f k = 1.57Hz, i.e., 134.07Hz, almost same as the real
one (134Hz). The measured amplitude is about 0.196 that is very close to the real one (0.2).

Case 4: ai 4 = 0.2 , f i 4 = 183Hz , τ = 5

0.05 2 + 0.024 2 + 0.017 2 + 0.0132 + 0.0112

∆f k = ⋅∆ f
0.0069 2 + 0.0089 2 + 0.014 2 + 0.028 2 + 0.18 2 + 0.05 2 + 0.024 2 + 0.017 2 + 0.0132 + 0.0112
0.06
= ⋅ 2.5 ≈ 0.6 (Hz) (34)
0.187 + 0.06

R.A.
= 0.0069 2 + 0.0089 2 + 0.014 2 + 0.028 2 + 0.18 2 + 0.05 2 + 0.024 2 + 0.017 2 + 0.013 2 + 0.0112
≈ 0.196 (35)

The above results indicate that the measured frequency is equal to f k = 182.5Hz plus ∆f k = 0.6Hz, i.e.,
183.1Hz, almost same as the real one (183Hz). The measured amplitude is about 0.196 that is very close to the real one
(0.2).

Case 5: ai 5 = 0.3 , f i 5 = 253Hz , τ = 5

0.075 2 + 0.036 2 + 0.025 2 + 0.019 2 + 0.016 2

∆f k = ⋅∆ f
0.0086 2 + 0.013 2 + 0.0212 + 0.042 2 + 0.28 2 + 0.075 2 + 0.036 2 + 0.025 2 + 0.019 2 + 0.016 2

14

IET Review Copy Only

Page 15 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

0.09
= ⋅ 2.5 ≈ 0.6 (Hz) (36)
0.28 + 0.09

R.A.
= 0.0086 2 + 0.0132 + 0.0212 + 0.042 2 + 0.28 2 + 0.075 2 + 0.036 2 + 0.025 2 + 0.019 2 + 0.016 2
≈ 0.29 (37)

As above, the measured frequency is equal to f k = 252.5Hz plus ∆f k = 0.6Hz, i.e., 253.1Hz, almost same as the
real one (253Hz). The measured amplitude is about 0.29 that is very close to the real one (0.3).
The measured spectrum using e-FFT is shown in Fig. 8, revealing no dispersed power around
harmonics/interharmonics.

Fig. 8 Spectrum of s (t ) using e-FFT with ∆f =2.5Hz

The comparison between FFT and e-FFT is concluded in Fig. 9. It can be seen that the results from e-FFT model
with ∆f =2.5Hz provides even better performance otcomes than ∆f =5Hz. Although the measurement in frequency
estimation is improved from traditional FFT, it can not achieve a correct analysis for most of amplitude estimation. The
maximum error for amplitude and frequency estimation using FFT is 20% and 1.12%, respectively. On the other hand,
the maximum error for amplitude and frequency estimation using e-FFT is only 3.33% and 0.05%, respectively.

1.2

1 Real value FFT e-FFT

0.8
Amplitude

0.6

0.4

0.2

0
50 68 96 134 183 253
Frequency (Hz)
(a) Amplitude measurement

15

IET Review Copy Only

 The Journal of Engineering Page 16 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

300 Real value FFT e-FFT

250

Measured result (Hz)

200

150

100

50

0
50 68 96 134 183 253
Harmonics/Interharmonics
(b) Frequency measurement
Fig. 9 Comparisons of FFT and e-FFT with ∆f =2.5Hz

III.1.3 Analysis with ∆f =10Hz

The parameters of e-FFT model are set as f s =1.28kHz, N=128, i.e., ∆f =10Hz. The interharmonics analysis
using e-FFT model is carried out by the following five cases from equ. (17). According to the rule of group bandwidth
( τ ) selection, τ is chosen as follows. The spectrum of s (t ) using FFT with ∆f =10Hz is shown in Fig. 10.

Fig. 10 Spectrum of s(t ) using FFT with ∆f =10Hz

(i) For f i1 = 68Hz ,

f i1 − f 1 = 68 − 50 < 4∆f (= 40) ⇒ τ =1

(ii) For f i 2 = 96 Hz ,
f i 2 − f i1 = 96 − 68 < 4∆f (= 40) ⇒ τ =1

(iii) For f i 3 = 134 Hz ,

f i 3 − f i 2 = 134 − 96 < 4∆f (= 40) ⇒ τ =1

16

IET Review Copy Only

Page 17 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

(iv) For f i 4 = 183Hz ,

4∆f (= 40) ≤ f i 4 − f i 3 = 183 − 134 < 6∆f (= 60) ⇒ τ =2

(v) For f i 5 = 253Hz ,

6∆f (= 60) ≤ f i 5 − f i4 (= 253 − 183) < 8∆f (= 80) ⇒ τ =3

Case 1: ai1 = 0.3, f i1 = 68Hz , τ = 1

0.24 2
∆f k = ⋅∆ f
0.076 2 + 0.12 2 + 0.24 2
0.24
= ⋅ 10 ≈ 6.32 (Hz) (38)
0.14 + 0.24

R.A.
= 0.076 2 + 0.12 2 + 0.24 2
≈ 0.28 (39)

The measured frequency is equal to f k = 60Hz plus ∆f k = 6.3Hz, i.e., 66.3Hz, close to the real one (68Hz). On
the other hand, the measured amplitude is about 0.28 that is also close to the real one (0.3).

Case 2: ai 2 = 0.4 , f i 2 = 96Hz , τ = 1

0.29 2
∆f k = ⋅∆ f
0.079 2 + 0.212 + 0.29 2
0.29
= ⋅ 10 ≈ 5.6 (Hz) (40)
0.23 + 0.29

R.A.
= 0.079 2 + 0.212 + 0.29 2
≈ 0.37 (41)

The measured frequency is equal to f k = 90Hz plus ∆f k = 5.6Hz, i.e., 95.6Hz, very close to the real one (96Hz).
The measured amplitude is about 0.37 that is close to the real one (0.4).

Case 3: ai 3 = 0.2 , f i 3 = 134Hz , τ = 1

0.12
∆f k = ⋅∆ f
0.048 2 + 0.15 2 + 0.12
0.1
= ⋅ 10 ≈ 3.8 (Hz) (42)
0.16 + 0.1

R.A.
= 0.048 2 + 0.15 2 + 0.12
≈ 0.19 (43)

17

IET Review Copy Only

 The Journal of Engineering Page 18 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

The measured frequency is equal to f k = 130 Hz plus ∆f k = 3.8Hz, i.e., 133.8Hz, almost same as the real one
(134Hz). The measured amplitude is about 0.19 that is very close to the real one (0.2).

Case 4: ai 4 = 0.2 , f i 4 = 183Hz , τ = 2

0.074 2 + 0.029 2
∆f k = ⋅∆ f
0.036 2 + 0.17 2 + 0.074 2 + 0.029 2
0.08
= ⋅ 10 ≈ 3.2 (Hz) (44)
0.17 + 0.08

R.A.
= 0.036 2 + 0.17 2 + 0.074 2 + 0.029 2
≈ 0.19 (45)

The measured frequency is equal to f k = 180 Hz plus ∆f k = 3.2Hz, i.e., 183.2Hz, very close to the real one
(183Hz). The measured amplitude is about 0.19 that is very close to the real one (0.2).

Case 5: ai 5 = 0.3 , f i 5 = 253Hz , τ = 3

0.12 2 + 0.056 2 + 0.038 2

∆f k = ⋅∆ f
0.019 2 + 0.046 2 + 0.25 2 + 0.12 2 + 0.056 2 + 0.038 2
0.14
= ⋅ 10 ≈ 3.5 (Hz) (46)
0.255 + 0.14

R.A.
= 0.019 2 + 0.046 2 + 0.25 2 + 0.12 2 + 0.056 2 + 0.038 2
≈ 0.29 (47)

As above, the measured frequency is equal to f k = 250Hz plus ∆f k = 3.5Hz, i.e., 253.5Hz, very close to the
real one (253Hz). The measured amplitude is about 0.29 that is very close to the real one (0.3).
The measured spectrum using e-FFT is shown in Fig. 11, indicating no dispersed power around
harmonics/interharmonics.

Fig. 11 Spectrum of s (t ) using e-FFT with ∆f =10Hz

18

IET Review Copy Only

Page 19 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

The comparison between FFT and e-FFT is concluded in Fig. 12. Clearly, the results from e-FFT model are close to
the real values for either amplitude or frequency identification, but traditional FFT is unable to achieve a correct
analysis except at fundamental component (50Hz). The maximum error for amplitude and frequency estimation using
FFT is up to 27.5% and 11.76%, respectively. By contrast, the maximum error for amplitude and frequency estimation
using e-FFT is only 7.5% and 2.5%, respectively.

1.2

1 Real value FFT e-FFT

0.8
Amplitude

0.6

0.4

0.2

0
50 68 96 134 183 253
Frequency (Hz)
(a) Amplitude measurement

300 Real value FFT e-FFT

250
Measured result (Hz)

200

150

100

50

0
50 68 96 134 183 253
Harmonics/Interharmonics
(b) Frequency measurement
Fig. 12 Comparisons of FFT and e-FFT with ∆f =10Hz

III.2 Determination of group bandwidth ( τ ) , sampled point (N) and sampling rate ( f s )
The group bandwidth ( τ ) may influence the measurement accuracy, and it should be chosen appropriately. The
larger group bandwidth ( τ ) can collect all dispersed power and then regain the actual amplitude more accurately.
However, a large τ may lead to a dispersed power overlapping within near major harmonics. For this reason, τ
should be chosen sufficiently large but to avoid the overlapping from neighbor harmonics. The selection rule of τ
(1-5) is thus formulated depending on the distribution range of dispersed energy.
It notes that the N is chosen as N= 2 m , e.g., 64, 128, 256, 512, 1024,….., using f s =1.28kHz, where
∆f = 20, 10, 5, 2.5, 1.25, ..... , respectively. Correspondingly, the sampling time will take 50ms, 100ms, 200ms,

19

IET Review Copy Only

 The Journal of Engineering Page 20 of 21
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

400ms, 800ms….., respectively. Clearly, an increasing N can obtain lower ∆f , but it will sacrifice for longer
sampling time. Hence, a compromise should be reached by both ∆f and sampling time.

IV. Conclusions
Although FFT is now widely applied to harmonics analysis, it can not be directly delivered to interharmonics
measurement. From the proposed e-FFT model, the dispersed energy can be collected efficiently and the interharmonic
original amplitude is thus retrieved. In addition, the interharmonic frequency can be found using a simple arithmetic
computation. The selection of group bandwidth ( τ ) has been formulated to effectively avoid overlapping between two
close interharmonics. In this model, ∆f =5Hz is chosen based on a compromise between the measurement accuracy
and sampling time. Accordingly, even a rapid change of signal variation can be responded sufficiently fast. For future
work, it is suggested to study the case that is involved in different harmonics sources. This situation may generate
unexpected sideband interharmonic frequencies that can not be resolved by current techniques.

References
1. Testa, A., Akram, M.F., Burch, R., Carpinelli, G., et al.: 'Interharmonics: Theory and Modeling', IEEE Trans. Del.,
2007, 22, (4), pp. 2335–2348.
2. Masoud Karimi-Ghartemani, M. Reza Iravani: 'Measurement of harmonics/inter-harmonics of time-varying
frequency', IEEE Trans. Power Deliv., 2005, 20, (1), pp. 23–31.
3. Lin, H.C.: 'Identification of interharmonics using disperse energy distribution algorithm for flicker troubleshooting',
IET Science Measurement & Technology, 2016, 10, (7), pp.786–794.
4. Tayjasanant, T., Wang, W., Li, C., et al.: 'Interharmonic-flicker curves', IEEE Trans. Power Deliv., 2005, 20, (2),
pp. 1017–1024.
5. Li, C., Xu, W., Tayjasanant, T.: 'Interharmonics: basic concepts and techniques for their detection and
measurement', Electr. Power Syst. Res., 2003, 66, (1), pp. 39–48.
6. Lin, H.C.: 'Separation of adjacent interharmonics using maximum energy retrieving algorithm', IET Science
Measurement & Technology, 2016, 10, (2), pp.92–99.
7. Lin, H.C.: 'Fast tracking of time-varying power system frequency and harmonics using iterative-loop approaching
algorithm', IEEE Trans. Ind. Electron., 2007, 54, (2), pp. 974–983.
8. Aghazadeh, R., Lesani, H., Sanaye-Pasand, M., et al.: 'New technique for frequency and amplitude estimation of
power system signals', IEE Gener. Transm. Distrib., 2005, 152, (3), pp. 435–440.
9. Lobos, T., Kozina, T., Koglin, H.-J.: 'Power system harmonics estimation using linear least squares method and
SVD', IEE Proc., IEE Gener. Transm. Distrib., 2001, 148, (6), pp. 567–572.
10. Zhu, T.X.: 'Exact harmonics/interharmonics calculation using adaptive window width', IEEE Trans. Power Deliv.,
2007, 22, (4), pp. 2279–2288.
11. Chang, G.W., Chen, C.-I.: 'An accurate time-domain procedure for harmonics and interharmonics detection', IEEE
Trans. Power Deliv., 2010, 25, (3), pp. 1787–1795.
12. Jain, S.K., Singh, S.N.: 'Exact model order ESPRIT technique for harmonics and interharmonics estimation', IEEE
Trans. Instrum. Meas., 2012, 61, (7), pp. 1915–1923.
13. Wu, R.-C., Tai, C.C.: 'Analysis of the exponential signal by the interpolated DFT algorithm', IEEE Trans. Instrum.
Meas., 2010, 59, (12), pp. 3306–3317.
14. Gu, I.Y.H., Bollen, M.H.J.: 'Estimating interharmonics by using sliding-window ESPRIT', IEEE Trans. Power
Deliv., 2008, 23, (1), pp. 13–23.

20

IET Review Copy Only

Page 21 of 21 The Journal of Engineering
This article has been accepted for publication in a future issue of this journal, but has not been fully edited.
Content may change prior to final publication in an issue of the journal. To cite the paper please use the doi provided on the Digital Library page.

15. He, C., Shu, Q.: 'Separation and analyzing of harmonics and inter-harmonics based on single channel independent
component analysis', Int. Trans. Electr. Energy Syst., 2015, 25, pp. 169–179.
16. Lin, H.C.,Chen, C. H., Liu, L. Y.: 'Harmonics and Interharmonics Measurement using Group-harmonic Power
Minimizing Algorithm', Proc. of the World Congress on Engineering, London, U.K., July 2011, pp.1300–1305.
17. Lin, H.C.: 'Intelligent neural network based adaptive power line conditioner for real-time harmonics filtering', IEE
Proc., IEE Gener. Transm. Distrib., 2004, 151, (5), pp. 561–567.
18. Chen, C.I.: 'Virtual multifunction power quality analyzer based on adaptive linear neural network', IEEE Trans. Ind.
Electron., 2012, 59, (8), pp. 3321–3329.
19. Valtierra-Rodriguez, M., de Jesus Romero-Troncoso, R., Osornio-Rios, R.A., et al.: 'Detection and classification of
single and combined power quality disturbances using neural networks', IEEE Trans. Ind. Electron., 2014, 61, (5),
pp. 2473–2482.
20. IEC 61000-4-7: Electromagnetic compatibility (EMC) Part 4: testing and measurement techniques Section 7:
general guide on harmonics and interharmonics measurements and instrumentation for power supply systems and
equipment connected thereto, 2002.
21. Oppenheim AV and Schafer RW: 'Discrete-time signal processing' (Upper Saddle River, NJ: Prentice Hall, 1989)
22. Press WH, Flannery BP, Teukolsky SA, et al.: 'Numerical recipes-The art of scientific computing' (Cambridge:
Cambridge University Press,1986)
23. Lin, H.C.: 'Development of leakage energy allocation approach for time-varying interharmonics tracking', IET
Gener. Transm. Distrib., 2015, 9, (9), pp.798–804.
24. Lin, H.C., Ye, Y. C., Huang, B. J., Su, J. L.: 'Bearing vibration detection and analysis using enhanced fast Fourier
transform algorithm', Advances in Mechanical Engineering, 2016, 8, (10), pp.1–14.

21

IET Review Copy Only