Real-Time Power System Frequency and Phasors Estimation Using Recursive Wavelet Transform
Real-Time Power System Frequency and Phasors Estimation Using Recursive Wavelet Transform
Real-Time Power System Frequency and Phasors Estimation Using Recursive Wavelet Transform
I. INTRODUCTION
Manuscript received October 05, 2009; revised March 19, 2010; accepted
March 23, 2011. Date of publication May 05, 2011; date of current version June
24, 2011. Paper no. TPWRD-00749-2009.
The authors are with the Department of Electrical and Computer Engineering, Texas A&M University, College Station, TX 77843-3128 USA
(e-mail: [email protected]; [email protected]).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TPWRD.2011.2135385
extract frequency due to its low computation requirement. However, the implicit data window in the DFT approach causes errors when frequency deviates from the nominal value [5]. To
improve the performance of DFT-based approaches, some adaptive methods based on the feedback loop by tuning the sampling
interval [6], adjusting the data window length [7], changing the
nominal frequency used in DFT iteratively [5], correcting the
gains of orthogonal filters [8], and tuning the weighted factor
[9] recursively are proposed. Due to the inherent limitation in
DFT, at least one cycle of analyzed signal is required, which
hardly meets the demand of high-speed response for protection
schemes. A method using three consecutive samples of the instantaneous input signal is discussed in [10]. The noise and zero
crossing issue may bring large errors to this method. On the
basis of the stationary signal model, some nonlinear curve fitting
techniques, including extended Kalman filter [11] and recursive
least-squares algorithm [12], are adopted to estimate the fundamental frequency. The accuracy is only reached in a narrow
range around the nominal frequency due to the truncation of
Taylor series expansions of nonlinear terms. Some artificial-intelligence techniques, such as genetic algorithm [13] and neural
networks [14], have been used to achieve precise frequency estimation over a wide range with fast response. Although better
performance can be achieved by these optimization techniques,
the implementation algorithm is more complex and intense in
computation.
Many techniques have been proposed to eliminate the impact of decaying dc components in phasor estimation. A digital mimic filter-based method was proposed in [15]. This filter
features high-pass frequency response which results in bringing
high-frequency noise to the outcome. It performs well when its
time constant matches the time constant of the exponentially decaying component. Theoretically, the decaying component can
be completely removed from the original waveform once its parameters can be obtained. Based on this idea, [16] and [17] utilize additional samples to calculate the parameters of the decaying component. Reference [18] uses the simultaneous equations derived from the harmonics. The effect of dc components
by DFT is eliminated by using the outputs of even-sample-set
and odd-sample-set [19]. Reference [20] hybridizes the partial
sum-based method and least-squares-based method to estimate
the dc offsets parameters. A new Fourier algorithm and three
simplified algorithms based on Taylor expansion were proposed
to eliminate the decaying component in [21]. In [22], the author
estimates the parameters of the decaying component by using
the phase-angle difference between voltage and current. This
method requires both voltage and current inputs. As a result, it
is not applicable to the current-based protection schemes.
where
is the Fourier transform of
.
A set of wavelet functions can be derived from
by dilating and shifting the mother wavelet, as will be given
where and are the scaling (dilation) factor and time shifting
(translation) factor, respectively.
A good wavelet is such a function that meets the admissibility condition and has a small timefrequency window area
[26]. We construct a mother wavelet function as expressed as
follows:
1393
( ).
,
makes the wavelet function
Setting
admissible (i.e.,.
.
One can see that the newly constructed wavelet is a complex
function whose time- and frequency-domain expressions contain real and imaginary parts. Figs. 1 and 2 give the time- and
and
, respectively.
frequency-domain waveforms of
Some performance parameters can be calculated to specify a
wavelet function [26]. The time-domain center and window
of wavelet function
are 0.99 s and 0.40 s, reradius
spectively. As one can see in Fig. 2, it features a band-pass filter
and band radius
of
with the frequency-domain center
rad and 1.38 rad. One advantage of the wavelet transform
is that the quality factor , defined as the ratio of frequency
and bandwidth
, stays constant as the observacenter
,
2.27. The complex
tion scale varies. For
wavelet exhibits good time-frequency localization characteristics. Its time-frequency window area , defined as a product of
and frequency bandwidth
, is 2.23
time window width
s.
To obtain the center frequency
of the band-pass filter,
which is defined as the frequency in which the function reaches
the maximum magnitude, we have the Fourier transform for the
dilated wavelet function
1394
Let
(4)
Denoting
th order
(3)
where
(5)
where ,
, and
represent the frequency, amplitude, and
phase angle of the th order harmonic, respectively. Denoting
the absolute phase angle of the th order harmonic at sample
as
, one can see that frequency
represents the rate of change of . For simplicity, the sampling
is neglected when expressing variables for the rest
period
of this paper.
in the timefrequency doTo represent the input signal
main, apply RWT in scale using (4). As derived in the Appendix we have the following expression:
From (2) and (3), we obtain
(6)
From (6), one can see that the wavelet transform coefficient
contains information on the input signal in both cosine
and
in (7a) and (7b)
form and sine form, denoted as
(given in the Appendix), respectively, multiplied by weighting
and
in (8a) and (8b) (given in the
factors, denoted as
Appendix), respectively.
represent the initial estimate of frequency variable,
Let
and rewrite (8a) using the first-order Taylor series expansion.
That is
(9a)
1395
can be calculated by
In (10b), the wavelet coefficient
, it can be calcuusing recursive (4). For weighting factor
using (8a) and (8b) and (9a)
lated with estimated frequency
and (9b). Solving (10b), we obtain vector variable
. Then,
:
we can derive the following formula for
(11a)
After we estimate the frequency adjustment, update the freand iterate the aforementioned approxquency with
imation procedures until either the frequency change reaches
0.001 Hz, or a maximum
the cutoff valuel; for example,
number of iterations denoted as is performed. As a result, the
real frequency can be estimated at the last iteration. Then, the
and phase angle
can be estimated by the folamplitude
lowing equations:
(11b)
(11c)
where
.
and
as
For simplicity, denote
and
, respectively. Then, we rewrite the equation as follows:
(9b)
Then, (6) can be expressed as follows:
(10a)
where
and
.
in a series of scales
,
Applying RWT to
that can be
we obtain a series of coefficients
expressed in (10a). Rewrite those equations in matrix form as
shown in the equation at the bottom of the page. For simplicity,
we represent the previous matrix in vector form. At sample ,
we have the following equation:
(10b)
..
.
..
.
..
.
..
.
where
or
.
The flowchart as given in Fig. 3 illustrates the implementation procedures for the proposed frequency, magnitude, and
phase estimation algorithm. In practice, a low-pass filter with
appropriate cutoff frequency is applied for eliminating high-frequency components in voltage and current measurements. As a
result, the order of harmonic components can be limited within
the range of cutoff frequency. For example, if a third-order Butterworth low-pass filter with a cutoff frequency of 320 Hz is
used to prefilter input signals, in this case, the maximum order
5). Generally,
of harmonics will be limited to five (i.e.,
multiples of the nominal frequency (i.e.,
we select
60 Hz,
represents the order of harmonics asan initial estimate to start iterations. To achieve high accuracy, scale facare required to cover all of the frequency
tors
components of the signal being analyzed. Therefore, we select
. Extensive simulations show
that the proposed algorithm can converge to the real value within
three iterations. It should be noted that if only the fundamental
is taken into
frequency component is of interest (i.e., only
the iteration loop), the dimension of scale factors and weighting
. Obviously, if the input signal
matrix will be reduced to
only contains the fundamental frequency component, the solved
and
will be some numbers
variables
close to zero, and then the parameters of those harmonics are
meaningless.
..
.
..
..
.
..
.
1396
= 65 Hz.
= 65 Hz.
that the window length can be shortened to 0.2 cycles if the sampling rate is 70 samples per cycle (i.e., 4.2 kHz) or higher.
Let us consider a case when the fundamental frequency deviates to 65 Hz and performs the algorithm to estimate frequency,
magnitude, and phase. Relationships between frequency error,
TVE, and two variables and , are shown in Figs. 5 and 6,
respectively, in which the signal sampling rate is simulated from
50 to 150 samples per cycle while the window length changes
from 0.25 to 1 cycle. One can see that the proposed algorithm
achieves high accuracy and fast convergence. Simulations performed in Section V also show that for a broad range of frequency deviation, such as 55 Hz65 Hz, the algorithm can converge to the real value within three iterations. Besides, the sampling rate has barely any effect on the accuracy once it reaches
50 samples per cycle (i.e., 3 kHz for the 60-Hz power system)
or higher. Compared to the conventional DFT-based methods,
this algorithm can shorten the window length to a quarter cycle.
Let us now consider the computation burden of the proposed
algorithm. If we use 3 kHz sampling frequency and 0.25 cycle
data window as the case performed in convergence analysis and
performance tests, it approximately requires 6000 multiplicamultitions and 5796 summations. Only
summations are used for
plications and
(where
5), and
computing RWT coefficients
5184 multiplications and summations for matrix inverse computation when three iterations are performed.
Weighting matrix
with various scales and frequencies can
be calculated and stored in advance and can be accessed very
fast by using a table lookup method. Some mathematical techniques, such as Chelosky and LU factorization methods, can be
adopted to simplify the matrix computation [28], [29]. The computation burden can then be noticeably reduced to
multiplications and
summations. Besides, increasing
the window length has a very small effect on the total computation burden because it only increases the computation burden
of RWT coefficients while the matrix dimension stays the same.
Based on the analysis, one can see the total computation burden
is fairly low. It can satisfy the time response requirement of
time-critical applications.
IV. ELIMINATING DECAYING DC COMPONENT
Similar derivation procedures can be used to develop the algorithm for eliminating the effect of decaying dc offset. Let us
consider the following signal model that contains the exponentially decaying component
1397
and
For simplicity, denote
, respectively, and rewrite the above formula
as
and
(15a)
where
Applying
RWT to
in a series of scales
, we obtain a series of coefficients
that can be expressed as the matrix,
shown at the bottom of the page. For simplicity, we represent
the above matrix in vector form. At sample , we have the
following equation:
(15b)
where
is the signal model defined in (5),
, and represents the amplitude and time constant of dc offset, respectively.
in the timeApplying RWT in scale to represent signal
frequency domain as derived in the Appendix, we have
can be calculated
In (15b), the wavelet coefficient
, it can be
by using recursive (4). For weighting factor
and time constant
calculated with approximate frequency
by using (8a)(8b), (9a)(9b), and (14a)(14b), respectively.
Solving matrix (15b), we obtain the vector variable
.
Then, we can derive the formula to estimate
(16a)
(12)
From (12), one can see that the wavelet coefficient
conand the weighted decaying
tains the coefficient for signal
dc component. Since the time constant is unknown to , iterations are required to approximate it.
Let represent the initial estimate and rewrite (14a) (in the
Appendix) by using the first-order Taylor series expansion, and
we have
(14b)
where
..
.
..
.
..
.
..
..
.
..
.
..
.
..
.
1398
or
TABLE I
TEST RESULTS FOR NOISE TESTS
V. PERFORMANCE EVALUATION
In this section, the performance of the proposed estimation algorithm is fully evaluated with various test conditions covering
static state, dynamic state, and transient state, and the results
are compared with conventional DFT methods, improved DFTbased methods in [5][7], and the latest published techniques in
[9], [10], [20], and [21]. In the static test, a signal model containing harmonics and noise is used and the performance is verified in a wide range of frequency deviations. The dynamic test
uses the scenarios that may occur in the real power system. The
scenarios including the frequency ramp, short-circuit fault, and
power swing are simulated using appropriate signal models. In
the transient test, three-phase current outputs from the Alternative Transients Program/Electromagnetic Transients Program
(ATP/EMTP) [30] are used to verify the performance of eliminating the dc offset. All tests are performed with the sampling
50 samples per cycle, (i.e.,
3 kHz, and data
rate
0.25 cycle (12 samples).
window size
A. Static Test
A signal model containing harmonics and 0.1% (signal-to60 dB) white noise is assumed, where
noise ratio
represents the zero-mean Gaussian noise. Let
1.0 p.u.,
. The fundamental frequency varies over a wide range
from 55 to 65 Hz in 0.2 Hz steps. Frequency error and total
vector error (TVE) of the fundamental frequency component are
estimated. Comparing to the DFT-based methods in [5][7], the
algorithm can output the frequency and phasor parameters in
about 4 ms. The method using three consecutive samples of the
instantaneous signal in [9] and [10], denoted as MV, achieves
the uncertainty of 10 million Hz. But they require a higher sampling frequency (6.4 kHz and higher) and the additional time
delay (approximately two cycles) introduced by the band-pass
filtering. The results are shown in Fig. 7. The output accuracy
can be improved by extending the data window. Simulation results show that the maximum frequency error and TVE can be
is extended
B. Noise Test
The inherent noise rejection capability of the algorithm is investigated by the noise test. The signal model for the static test is
used. Let the fundamental frequency take the nominal value (60
Hz). For each level of the Gaussian noise, three data windows
(quarter cycle, half cycle, and one cycle) were applied. The test
was conducted by using the method MV except applying the
variable data windows because the MV has a fixed size of data
window. Each case was performed 10 times and the maximum
value of the frequency estimate error for both RWT and MV,
and TVE for RWT are shown in Table I. As one can expect,
the better noise rejection can be obtained by slowing down the
output response (i.e., prolonging the window span). The accuracy of RWT with one cycle window is in the same level with
that of MV. The MV requires extra delay caused by filtering.
C. Dynamic Test
1) Frequency Ramp: The following synthesized sinusoidal
signal with a frequency ramp is used to perform the frequency
ramp tests
1399
To investigate the effect of prefiltering on the algorithm dynamic performance, a third-order Butterworth low-pass filter
with a cutoff frequency of 320 Hz is used to process the input
signals. Fig. 12 shows the result of the amplitude step test. Compared to Fig. 9, which shows the transient behavior without
signal prefiltering, one can see that the low-pass filter enlarges
the overshoot and undershoot, and slows the response from 4
to 10 ms though it is still faster than the DFT-based methods
[5][7] and instantaneous sample-based methods [9], [10].
3) Modulation: A sinusoidal modulation signal model is
used to simulate the transient progress of voltage and current
signals during the power swing. Its amplitude and phase angle
are applied with simultaneous modulation as shown in the
following expression:
1400
TABLE II
TEST RESULTS FOR MODULATION TESTS
VI. CONCLUSIONS
This paper proposes a new wavelet function and its recursive
wavelet transform. The method allowing real-time estimating
of power system frequency, magnitude and phase while eliminating the impact of decaying dc component based on RWT
is proposed. The algorithm features rapid response and accurate results over a wide range of frequency deviations. It uses
only a quarter cycle of input signals for outputting frequency,
and magnitude and phase results for a signal contaminated with
harmonics. The sampling rate and observation window size can
be chosen to meet selected applications requirements. The analysis of the algorithm convergence characteristics indicates that
the higher the sampling rate, the shorter the computation data
window and the faster the rate the method outputs phasor, and
vice-versa. The decaying dc component can be completely removed by estimating its parameters using RWT. The performance of the proposed algorithm is evaluated under a variety
of conditions including static state, dynamic state, and transient
state. Comparing other techniques results demonstrates the advantages. Computation burden analysis indicates that the computation requirement is moderate. Thus, this approach can satisfy the time-critical demand of the real-time applications in
power systems.
TABLE III
TEST RESULTS FOR DECAYING DC OFFSET
APPENDIX
The RWT coefficient of a given signal
is expressed as
D. Transient Test
A 230 kV power network is modeled in EMTP to generate
waveforms for testing the performance when eliminating
decaying dc offset. A three-phase fault is applied and the
three-phase currents are used as input signals. Fig. 13 shows
the phase-A current waveform. One can see that the signal is
contaminated with decaying dc component and high frequency
noise during the beginning of postfault. The third-order Butterworth low-pass filter with a cutoff frequency of 320 Hz is
used to attenuate the high-frequency components. Parameters
estimation for the steady state (twenty cycles after the fault
occurs) is used as a reference to measure the TVEs.
As shown in Table III, the results are compared with the conventional full-cycle DFT (FCDFT), half-cycle DFT (HCDFT)
methods, least error square method (LES), simplified algorithm
(SIM3) in [21], and hybrid method (HM) in [20]. In Table III,
is the time (in cycles) when the TVEs are measured. For the high
accuracy, the algorithm was adjusted to a three-quarter cycle
window span. The results show that the accuracy is comparable
to those of LES, SIM3, and HM methods while the proposed algorithm requires a shorter data window, which results in faster
response.
Denoting
, we have
where
(7a)
(7b)
(8a)
(8b)
Similarly, for signal
coefficient
Denoting
, we have
where
(13)
(14a)
The true value of frequency, amplitude, and phase angle at the
output sample for the modulated signal model can be computed as
(17a)
(17b)
(17c)
REFERENCES
[1] A. G. Phadke and J. S. Thorp, Computer Relaying for Power Systems.
New York: Wiley, 1988.
[2] M. V. V. S. Yalla, A digital multifunction protective relays, IEEE
Trans. Power Del., vol. 7, no. 1, pp. 193201, Jan. 1992.
[3] A. G. Phadke, T. Hlibka, and M. Ibrahim, A digital computer system
for EHV substation: Analysis and field tests, IEEE Trans. Power App.
Syst., vol. PAS-95, no. 1, pp. 291301, Jan./Feb. 1976.
1401
1402
[30] Power Syst. Relay. Committee, EMTP reference models for transmission line relay testing report. 2001. [Online]. Available: http://www.
pes-psrc.org
Jinfeng Ren (S07) received the B.S. degree from Xian Jiaotong University,
Xian, China, in 2004, and is currently pursuing the Ph.D. degree at Texas A&M
University, College Station, TX.
His research interests are new algorithms and test methodology for synchrophasor measurements and their applications in power system protection
and control as well as new digital signal-processing techniques for power
system measurement and instrumentation, and automated simulation methods
for multifunctional intelligent electronic devices testing.