Sie sind auf Seite 1von 6

1110

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO. 4, OCTOBER 2000

Computation of the Short-Term Flicker Severity Index


Thomas Keppler, Neville Watson, Member, IEEE, and Jos Arrillaga, Fellow, IEEE

AbstractShort-term flicker severity ( ) is an important


electric power quality index defined by the International Electrotechnical Commission (IEC).
results from the statistical
evaluation of flicker over short periods of time producing an
objective measure for flicker originating from various types of
sources. This paper reviews previous methods for computing
and proposes a new iterative algorithm. The algorithm is
computationally efficient and shown to comply with the IEC
standard.
Index TermsElectric power quality, flickermeter, flicker
severity.

II. STATISTICAL EVALUATION OF FLICKER


A. Cumulative Probability Functions
Random flicker occurrences require monitoring over a period
of time representative of the considered disturbance. During the
observation time the IFL will vary widely and therefore its statistical evaluation is appropriate.
that
For a time domain signal, the cumulative probability
a signal level is exceeded during a period of time is defined
as
total time where signal level

I. INTRODUCTION

HE HUMAN eye is very sensitive to light flicker but irritation depends on lighting conditions and the individual.
Flicker monitoring should therefore be based on continuous
voltage measurements indicating experienced flicker consistently and objectively.
The International Electrotechnical Commission (IEC) has
published design specifications for a flickermeter [1], [2]
which evolved from a proposal by the Union Internationale
dElectrothermie [3].
The IEC flickermeter can be divided into two main parts. The
first part realizes an electrical model of the lamp-eye-brain chain
relating voltage amplitude fluctuations to flicker sensation given
as the instantaneous flicker level (IFL) [1]. The flickermeter thus
quantifies flicker on the basis of human irritation. The second
part consists of an on-line statistical evaluation of the IFL [2].
The evaluation produces a representative value of short-term
in 10 minute intervals which reduces the
flicker severity
amount of data considerably and simplifies comparison to limit
values.
This paper resulted from experience gained in designing a
DSP based flickermeter. Space does not permit the presentation
of the full details and therefore we focus on the implementation
of the statistical evaluation method which unlike the IFL measurement (for example [4][6]) has not received much attention.
calculation by classification of the IFL is well defined
The
in the IEC document [2]. To reduce classifier sizes but maintain
accuracy interpolation techniques have been proposed.
A review of the existing methods, however, revealed inconsistent performance. A new evaluation algorithm, which works
reliably and more efficiently is described in this paper.

Manuscript received December 8, 1996; revised April 25, 1997. This work
was financially supported by Trans Power NZ.
The authors are with the Department of Electrical and Electronic Engineering, University of Canterbury, Christchurch, New Zealand (e-mail:
T.Keppler@elec.canterbury.ac.nz).
Publisher Item Identifier S 0885-8977(00)10298-5.

(1)

reflects the distribution of


A graphical representation of
magnitudes and is referred to as cumulative probability function
(CPF).
The CPF of the instantaneous flicker level hence gives the
percentages of observation time for which flicker levels have
been exceeded.
B. Short-Term Flicker Severity Index
The IEC flickermeter monitors regular and irregular types of
flicker from different types of sources. This includes household
or similar small appliances and industrial loads. The shapes
of the CPF curves of the resulting IFLs are varied and do not
follow a standard type of distribution. To simplify comparisons
the short-term flicker severity has been defined for their
characterization.
A number of gauge points on the CPF are used to derive the
in (2). Short-term refers to obshort-term flicker severity
servation periods of 10 minutes.
(2)
where
th weighting coefficient
CPF curve level being exceeded for % of the observation period.
Initially five gauge points on the CPF curve and corresponding weighting factors have been chosen, so that
becomes

(3)
in the case of loads that produce a
Implausible results for
constant disturbance and follow an on/off duty cycle, led to an
,
extension of the definition in (3). The gauge point levels
, and
are determined from subsidiary gauge points by

08858977/00$10.00 2000 IEEE

KEPPLER et al.: COMPUTATION OF THE SHORT-TERM FLICKER SEVERITY INDEX

Fig. 1.

Classification of a time domain signal.

1111

Fig. 3.

Class representation by harmonic average.

determine the minimum required , the quantization errors


shall be derived.
points lie at the
For linear spacing (used in Fig. 2), the
instantaneous flicker levels given by
(5)
where

Fig. 2. Cumulative Probability Function (CPF) of the signal in Fig. 1 for the
observation time T .

averaging which yields the smoothed values in (4). The calcuis unchanged except that the respective smoothed
lation of
values are used in (3).

th signal level
minimum instantaneous flicker level
maximum instantaneous flicker level
number of points of the CPF.
do not coincide with
Usually the gauge point probabilities
calculated by (1), but lie between known
the probabilities
therefore
values (see Fig. 3). The desired gauge point levels
have to be chosen representatively of the corresponding class.
of the
Choosing the harmonic average
produces symmetric
respective neighboring points for the
relative errors and yields the lowest maximum error for a given
number of classes.
The maximum relative quantization error for class is
(6)

(4)

Substituting (5) into this last relationship yields

For completeness long-term flicker severity shall be mentioned. This index is simply calculated from 12 successive
values and is not of interest here.
(7)
III. EXISTING EVALUATION METHODS
A. Classification Method
To reduce computation, only a limited number of points of
classes, which
the CPF can be calculated. This introduces
can be linearly or logarithmically spaced. Fig. 1 shows a time
domain signal being classified into 7 equally spaced classes.
indicates, the duration the signal exceeds
Each
the lower limit of the corresponding class . The CPF, resulting
for the observation time , is shown in Fig. 2.
Quantization errors occur when the gauge points on the
discretized CPF are determined. Increasing will increase the
accuracy of the method and the computation time. In order to

cumulative probaFor logarithmically spaced classes the


bilities have to be calculated for the instantaneous flicker levels
(8)
Substitution into (6) yields the maximum error
(9)
The logarithmic classifier makes the error independent of the
class and therefore all gauge point levels are determined with
the same accuracy, regardless of their location. Evaluation of (7)

1112

Fig. 4.

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO. 4, OCTOBER 2000

Linear interpolation within classes.

and (9) shows that for given ,


, and
, the linear classifier produces much larger errors than the logarithmic classification. For this reason linear classification shall not be regarded
further.
Solving (9) for yields the number of CPF points required
:
to meet a certain error

Fig. 5.
Bisectional search for
approximations for P . P
P
achieved accuracy.

0 j (k

1 1 1 P represent successive
1 1 1 3) is a measure of the

= 1

Assuming a gauge point independent maximum error allows


simplification of (13) to
(14)

(10)

B. Classification Enhanced by Interpolation


Instead of using the harmonic average to represent a class,
by interpolating beit has been proposed to determine the
tween known values [2]. A gain in accuracy would allow the use
of less classes leading to faster computation.
The worst case error of this method can be derived from
represents the true value and
is obtained by linear
Fig. 4.
interpolation. The maximum deviation from the true value
equals the width of the class. The relative maximum error for
logarithmically spaced points using (8) becomes
(11)
Finally the number of required CPF curve points as a function
of the desired accuracy can be derived from (11)

IEC specifications demand errors of less than 5% for the


computation [2]. Consequently the error in estimating each
should be less than 10% due to
of the gauge point levels
(14).
Consider for example a realistic IFL signal range of 6 orders of magnitude. For the derived error limit, equations (10)
and (12) evaluate to 70 and 146 points respectively. Hence the
logarithmic classification with interpolation requires more than
twice as much calculation. This contradicts the initial assumption that the interpolation would improve the accuracy for a
given .
Reference [2] states that logarithmic classification complemented by linear interpolation typically keeps errors below
0.5%. This would require the excessive calculation of 2772
CPF curve points for the same IFL signal range.
The contradictory statements about the performance of the
interpolation result from different assumptions. For fairly linear
CPFs, the errors were in fact reduced dramatically by the interpolation. The errors derived in (11) however take into account
arbitrarily shaped curves, because nonlinear CPFs are common.

(12)
IV. NEW ITERATIVE EVALUATION METHOD

C. Discussion of Errors
As a basis for the discussion the tolerable errors in determining each of the gauge point levels shall be derived. The error
in flicker severity equals the sum of the partial derivatives
, each weighted by the small deviation
. Using (2)
of
we obtain
(13)

The inconsistent improvements of accuracy by the interpolation technique motivated the development of a new method
based on a search algorithm.
requires the 15 gauge points levels
The calculation of
. The reviewed classification methods however involve the computation of a generally much larger number of
points. After the gauge points have been identified, the remaining curve points are discarded which makes the method inefficient.
A bisectional search algorithm was found to be more efficient
in locating the gauge points. Fig. 5 represents the principle of

KEPPLER et al.: COMPUTATION OF THE SHORT-TERM FLICKER SEVERITY INDEX

Fig. 6.

1113

Bisectional search algorithm.

the method.
is the desired gauge point flicker level that possesses a cumulative probability of %. The initial search interval
and . As a first iteration step the cumulative
is limited by
of the interval is
probability at the center
determined. Because it is less than and due to the negative gradient of the CPF, the subsequent search can be restricted to the
and . The next curve point is calculated
section limited by
and the result determines
at the level
which subsection has to be considered next and so forth.
The iteration terminates once the error meets the desired accuracy. The relative error is determined by the width of the
current section. Because the cumulative probability decreases
monotonically with increasing signal level, convergence always
occurs.
The number of iterations needed depends on the shape of the
CPF curve, the size of the initial section, and the desired accuracy. In order to minimize the total number of iterations, the
widths of the initial sections are minimized, using points from
a list of previously determined CPF curve points. The pseudo
code of the algorithm is listed in Fig. 6.
V. RESULTS
A. Comparison of Errors
Figs. 7 and 8 represent two different types of IFL waveforms,
their CPFs, and the errors in determining the gauge point levels.
The IFL in Figs. 7(a) and 8(a) were produced by a flickermeter
subjected to rectangular flicker and a sequence of sinusoidal
flicker respectively. The CPF gauge points were found by the
described techniques. The true CPF used to estimate the errors, was established by the iterative algorithm configured to
produce an error below 10 %.
The logarithmic classification and the iterative algorithm
were set up to give an error of less than 10% and thus

Fig. 7. Evaluation of rectangular flicker. (a) Time domain IFL and resulting
CPF. (b) Error for each gauge point level.

the results of the error estimation in Figs. 7(b) and 8(b) are
consistent.
Linear interpolation reduces the errors of the logarithmic
classification (for a constant number of classes) in case of
Fig. 7(b). However in Fig. 8(b) errors are increased and even
exceed the 10% limit value. This example shows that the linear
interpolation does not improve the accuracy consistently for
various CPF curve shapes. Even more complex interpolation
techniques (cubic or spline interpolation) are not suitable to
approximate the CPF due to its stochastic nature.
B. Compliance Testing
The IEC standard [2] provides reference signals to test a flickermeter and its statistical evaluation for compliance.

1114

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 15, NO. 4, OCTOBER 2000

Fig. 9.

vs. frequency of rectangular flicker.

accuracy regardless the range of the instantaneous flicker level,


this method can work with any range.
The range testing for the classification methods was eliminated by updating the number of classes used on expiry of each
observation period. The use of the minimum number of classes
given by (10) ensures that the specified accuracy is maintained.
C. Computational Considerations

Fig. 8. Evaluation of a sequence of sinusoidal flicker. (a) Time domain IFL


and resulting CPF. (b) Error for each gauge point level.

Fig. 9 shows the test results obtained from simulations. The


values lie within the specified range of 1.0 0.05 and hence
all three methods comply with the standard.
The test signals consist of rectangular voltage flicker of
various frequencies and amplitudes producing IFL waveshapes
similar to Fig. 7(a). The resulting CPFs are relatively linear.
Since the classification with interpolation generally increases
accuracy in such cases it is not surprising that compliance even
for this method occurs.
The question should be asked whether the IEC test signals and
their CPFs reflect common situations representatively. Presently
representative measurements are not available to the authors and
so this is under investigation.
In addition the working range of the statistical evaluation
should be considered. Because the iterative search maintains its

for a 10 minute observation peThe computation of the


riod involves classification of 30 000 IFL samples (sampled at
the lowest possible rate of 50 Hz).
The iterative search algorithm requires all samples in memory
delayed after expiry of the observation
and produces the
period. If classification with a constant classifier size is used
memory for the class counts is needed but not for the IFL samples.
Computation times vary depending on the distribution. The
iterative search runs 5697% faster compared to a classifier with
128 classes for the CPFs considered in this paper (64 classes is
the by IEC suggested minimum).
The real-time computation of the IFL uses multirate processing optimizing it for speed and thus 60% of the DSPs
calculation. Therefore
processing power remain for the
the savings achieved by the iterative search are not crucial in
our case, however, they gain significance if higher IFL sample
rates, higher accuracies, or less powerful processors are used.
VI. CONCLUSION
Cumulative probability functions, the short-term flicker
severity index, and classification methods were reviewed.
An error assessment of the existing linear and logarithmic
classification and interpolation methods showed that interpolation does not increase accuracy consistently for various types
of flicker and hence should not be used to reduce computation
time.
A new iterative search algorithm was presented. It is more
efficient than logarithmic classification and works for any shape
of the cumulative probability function.

KEPPLER et al.: COMPUTATION OF THE SHORT-TERM FLICKER SEVERITY INDEX

REFERENCES
[1] IEC Publication 868, FlickermeterFunctional and design specifications,, Tech. Rep., IEC, 1986.
[2] IEC Publication 868, FlickermeterEvaluation of flicker severity,,
Tech. Rep., IEC, pt. 0, 1991.
[3] UIE, WG Disturbances, Flicker measurement and evaluation, Union
Internationale dElectrothermie, Tech. Rep., 2nd revised ed., 1991.
[4] W. Mombauer, Flicker simulation and minimization, in IEE Conference Publication of the 10th International Conference on Electricity
Distribution (CIRED), 1989, pp. 102106.
[5] K. Srinivasan, Digital measurement of voltage flicker, IEEE Trans.
Power Delivery, vol. 6, no. 4, pp. 15931598, Oct. 1991.
[6] L. Toivonen and J. Mrsky, Digital multirate algorithms for measurement of voltage, current, power and flicker, in IEEE PES Trans. and
Dist. Conference Proceedings, 1994, pp. 330340.

1115

Thomas Keppler received his Diplom Ingenieur Elektrotechnik from Universitt Stuttgart, Germany, in 1994. Since 1995 he has been working on a Ph.D.
at the University of Canterbury, New Zealand.

Neville Watson received his B.E. (Hons) and Ph.D. degrees in electrical engineering from the University of Canterbury, New Zealand, where he is now a
Lecturer. His interests include steady-state and dynamic analysis of ac/dc power
systems.

Jos Arrillaga received a B.E. degree in Spain and the M.Sc., Ph.D., and D.Sc.
in Manchester, where he led the power systems group of UMIST between
19701974. He has been a Professor at the University of Canterbury since
1975. He is a Fellow of the IEE, of the IEEE, and of the Royal Society of New
Zealand.

Das könnte Ihnen auch gefallen