Abstract A hybrid method which combines four different

algorithms for low-frequency oscillation parameter estimation is

presented in this project. The first algorithm is discrete Fourier
transform (DFT) which is used to detect low-frequency oscillation.
The second one is FIR window filter which is employed in filtering
unwanted components, and Window Correction Factor (WCF) is
provided to eliminate the side effect of using FIR window filter. The
third one is the prototype algorithm which is proposed to estimate
low-frequency oscillation parameters, and the last one is the least
square error (LSE) method which is utilized to enhance the
computational accuracy by changing the window size. This hybrid
method has been implemented by Java language and is under test in
Taiwans power system. To demonstrate the performance of this
hybrid method, Matlab/Simulink simulation and field test results are
provided in this project.
Index Terms - low-frequency oscillation, discrete Fourier
transform, FIR window filter, least square error method.
I. INTRODUCTION
ince low-frequency oscillation has been detected in
Taiwans power system (Taipower system) in 1984, related
researches have become important in Taiwan. This is because
low-frequency oscillation causes transmission capacity
reduction, equipment damage, and even system instability.
To prevent the damage caused by low-frequency oscillation,
several generators in the Taipower system have been installed
with power system stabilizers (PSS). However, with the
increase of load and shortage of transmission capacity, low-
frequency oscillation may occur in the future. Therefore, it is
essential to monitor and estimate low-frequency oscillation
parameters.
In the last ten years, many algorithms have been proposed
for low-frequency oscillation parameter estimation, including
DFT [1]-[4], Kalman filter [5], Prony method [6]-[9], and
wavelet transform [7], each having its own features.
Easy implementation, efficient execution, and high noise-
resistance are the main advantages of DFT, but the waveform
of low-frequency oscillation is not a periodic sinusoidal wave.
DFT is not suitable for low-frequency oscillation parameter
estimation.
Kalman filter is a technique for estimating an unknown state
of a dynamic system which has additive noise, but model
analysis makes it inconvenient for low-frequency oscillation
parameter estimation.
Prony method is a very efficient tool and is able to estimate
damping, amplitude, frequency, and phase from sequential
data. However, it is very difficult to compute the parameters
through this method [7] and it is well known to perform poorly
when the signal is imbedded in noise [8], [10].
Wavelet transform is suitable in recognizing the features of
signal in both time and frequency domains, but wavelet
analysis also has its own limitations. A correct wavelet is the
key to correct analysis, unfortunately however, a damping
sinusoid is not a wavelet.
The hybrid method we presented in this project is
composed of four parts, and each part has its own function.
The first part is DFT which is used to detect low-frequency
oscillation. The second part is the FIR window filter which is
employed in filtering unwanted components while Window
Correction Factor is provided to eliminate the side effect of
FIR window filter. The third part is the prototype algorithm
which we proposed to estimate low-frequency oscillation
parameters, and the last part is the LSE method which is
utilized to enhance the computational accuracy by changing
the window size.
The organization of this paper is as follows. We describe
each part of this hybrid method in section II. Next, the test
results are provided in section III. Finally, we give the
conclusions of the research in section IV.
II. THE HYBRID METHOD
More than 10 synchronized phasor measurement units
(PMUs), which were produced by the local ADX Company,
have been installed into the backbone of Taipower system, and
each PMU uploads its time-tagged phasor data at a rate of 20
Hz to the central servers in Taipower building and Taipower
Company Power Research Institute. This quantity of
information is enough for Taipower Company to monitor the
situation of the Taipower system before doing some
applications. Low-frequency oscillation parameters estimation
is one of these applications.
A. Detector
The difference in phase angle between PMUs is used to
detect low-frequency oscillations. Since any small disturbance
JAVA
Implementation of Real-Time Low Frequency Oscillation Estimating
Algorithm Using JAVA
: NSC 94-2622-E-214-015-CC3
: 94 8 1 95 7 31
:
:
:
S
would cause change in the phase angle, how to detect low-
frequency oscillations is useful to prevent worthless
computation.
To design the detector, we divided the start of low-
frequency oscillations into two types. The first type is that
oscillations start with severe swings in phase-angle difference.
Usually, this type of oscillation is caused by serious
disturbances. Therefore, the first type is detected while the
difference of the maximum and minimum difference in phase
angle in 1 second is over the threshold.
The second type of oscillation is occurred spontaneously
and the amplitude of oscillation is from small to large.
Sometimes, this type of oscillations is difficult to be detected
by the method used to detect the first type, because the
frequency range of low-frequency oscillation is 0.2 Hz to 3.0
Hz [11].
To detect the second type of low-frequency oscillations, we
use DFT with three different fundamental frequencies which
are 0.5 Hz, 1 Hz, and 2 Hz, respectively, to compute the
amplitudes of phase-angle difference in 1 sec. If one of these
three amplitudes is over the threshold, the low-frequency
oscillation is detected.
B. Prototype Algorithm
Since the FIR window filters we designed have a close
relationship with the prototype algorithm we provided, we
have to introduce the prototype algorithm before the FIR
window filter.
When low-frequency oscillation occurs in a power system,
we assume that all the components in the signal, the difference
in phase angle between PMUs, can be expressed as follows:
noise X
e X t k f e X k x
dc
t k
op
p
i
i i
t k
i
i
+ +
+ + A =
A
=
A

1
) 2 cos( ) (
(1)
where
i
X is the amplitude,
i
is the decay constant,
i
f is
the frequency,
i
is the phase ,
dc
X is the DC component of
the signal, t A is the time interval (0.05 sec.), and
t k
op
e X
A
is
used to represent the shift of the operating point. Since we will
use FIR window filter to filter
dc
X and noise, we temporarily
ignore these components in this section.
t k
op
p
i
i i
t k
i
e X t k f e X k x
i
A
=
A
+ + A =



1
) 2 cos( ) ( (2)
Since we know that
2
cos

j j
e e

+
= (3)
Then ) (k x can be expressed as
t k
op
p
i
t k f j
i
t k f j
i
t k
e X
e e X e e X e k x
i i i i i
A
=
A A A
+
+ =

1
2 2
) (
2
1
) (
(4)
For further derivations, we use the following definitions:
i
j
i
i
e
X
A

2
= (5)
t f j
i
i
e a
A
=
2
(6)
t
i
i
e b
A
=

(7)
t
e c
A
=

(8)
) Re(
i i
a z = (9)
Then (4) can be expressed as
( )
k
op
p
i
k
i i
k
i i
k
i
c X a A a A b k x + + =

=

1
) ( (10)
Therefore, we can find the following relations between ) (k x
and ) 1 ( + k x .
After some algebraic manipulations, we find
0 ) ( ) (
1 2
0
= +

+
=
p
i
i k x i C (11)
}} , 2 , { } , 2 , {
} , 2 , { } , 2 , { } 1 , {{
1 1
1 1 1
1
2 2 2
1
1 1 1



- -
- - - =
p p p p p p
b z b b z b
b z b b z b c C

(12)
where - represents the convolution operator. From (11) and
(12), we have very useful consequences for follow-up
development. We can use these two equations in two different
ways. First, if the parameters of low-frequency oscillation are
unknown, we can utilize (11) and the LSE method, which will
be described in the last subsection to get the parameters.
Second, if we already know the parameters of low-frequency
oscillation, then we can apply (12) as a FIR window filter,
which will be described in the next subsection.
Besides, from (12), it is obvious that the complexity of
getting parameters is in direct proportion to the value of p.
Larger p value denotes more difficulty to solve and more
sensitivity to noise. Therefore, in Taipower system, the value
of p is chosen to be 1 for on-line application. As such, the
waveform can be simplified as follows:
noise X
e X t k f e X k x
dc
t k
op
t k
+ +
+ + A =
A A
) 2 cos( ) (
1 1 1
1
(13)
C. FIR Window Filter
Noise and some unwanted components are always in the
signal. As these components will affect the accuracy of the
computational results, a well-designed filter is essential to get
rid of them.
Since digital filters are easy to design and implement, they
have been widely used in signal processing. To get the precise
low-frequency oscillation parameters, we have to design
specific FIR window filters to filter not only noise but also DC
component, decaying DC component, and decaying sinusoid
component, respectively.
According to (12), the FIR window filters we designed are
as follows:
}} 1 , 1 { {
1
- =
F
W W (14)
} 1 , {
2
= c W (15)
} , , {
1
1 1 1 3

= b z b W (16)
where
1
W is designed to filter noise and DC component, and
window size is M ,
2
W is designed to filter decaying DC
component,
3
W is designed to filter decaying sinusoid
component, and the chosen Blackman window is
F
W .
By these FIR window filters, we can separate decaying DC
component and decaying sinusoid component, respectively
from signal. Since there are two different decaying
components in the signal and the decay constant of these two
components should not be the same, one of them will fade
faster than the other one. Therefore, to prevent the LSE
method from encountering singular matrix or noise sensitivity,
we estimate the parameters of these components separately.
The steps we take to separately estimate the parameters of
low-frequency oscillation are as follows:
Step 1. When low-frequency oscillation has been detected, the
signal ) (k x will be filtered by
1
W to get rid of noise
and DC component and to get ) (
~
1
k x . Therefore, the
main components in the ) (
~
1
k x are the decaying DC
component and the decaying sinusoid component.
Step 2. The initial value of c is 1.
Step 3. ) (
~
1
k x is filtered by
2
W to get rid of the decaying DC
component and to get ) (
~
2
k x . Therefore, the main
component in the ) (
~
2
k x is the decaying sinusoid
component.
Step 4. By utilizing (11) and the LSE method, the parameters,
1
b and
1
z can be estimated.
Step 5. ) (
~
1
k x is filtered by
3
W to get rid of the decaying
sinusoid component and to get ) (
~
3
k x . Therefore, the
main component in the ) (
~
3
k x is the decaying DC
component.
Step 6. By utilizing (11) and the LSE method, the parameter
c , can be estimated.
According to the experiments, after repeating Step 3 to Step 6
three times, the values of
1
b ,
1
z , and c will be converged.
Where ) (
~
1
k x , ) (
~
2
k x , ) (
~
3
k x can be obtained as follows:

=
+ =
1
0
1 1
) ( ) ( ) (
~
M
m
m k x m W k x (17)

=
+ =
1
0
1 2 2
) (
~
) ( ) (
~
m
m k x m W k x (18)

=
+ =
2
0
1 3 3
) (
~
) ( ) (
~
m
m k x m W k x (19)
Then, according to the definitions (6-9) given above, we
can estimate the parameters of the signal by means of the
following equations.
t
z
f
A
=

2
) ( cos
1
1
1
(20)
t
b
A
=
) ln(
1
1
(21)
t
c
A
=
) ln(
(22)
However, other problems encountered are phase shift and
amplitude modulation, which are the side effects of using FIR
window filter. To eliminate these problems, Window
Correction Factor (WCF) is presented.
From (13), the signal ) (k x is filtered by
1
W to get
) (
~
1
k x in (17), and then ) (
~
1
k x is filtered by
2
W to get ) (
~
2
k x
in (18). Therefore, ) (
~
2
k x can be expressed as

=

=
+
=
M
m
m m k k
M
m
m m k k
m W b a b a A
m W b a b a A k x
0
1 1 1 1 1
0
1 1 1 1 1 2
) (
) ( ) (
~
(23)
where
}} 1 , { } 1 , 1 { { - - = c W W
F
(24)
According to (5), (10) and (23), we obtain
1
2
1
2 2
1
1 1
1
WCF ) 1 (
) (
~
) 1 (
~

+
=

a
k x k x b a
A (25)
1 1
2 A X = (26)
) (
1 1
A angle = (27)
where
1
WCF is defined as follows:

=
=
M
m
m m
m W b a
0
1 1 1
) ( WCF (28)
By WCF, the side effects of using FIR window filter can be
totally eliminated, and the amplitude and phase of low-
frequency oscillation can be estimated precisely.
D. LSE Method
We utilize (11) and LSE method to estimate the parameters
of low-frequency oscillation. This is because LSE method is
easy to implement and the data window size of LSE method
can be adjusted according to the frequency of low-frequency
oscillation. Since the frequency range of low-frequency
oscillations is between 0.2 Hz to 3 Hz, it is certain that a larger
window size is needed for obtaining more accurate estimation
when the frequency is lower than 1 Hz. Therefore, from (18),
and using the consequences of (11) and (12), we obtain
0 ) 2 (
~
) 1 (
~
2 ) (
~
2
1
1 2 1 2 1
= + + +

k x b k x z k x b (29)
Rearrange (29) as
(
(
(
(

+
+
+
=
(

(
(
(
(

+ +
+ +
+
) 1 (
~
) 3 (
~
) 2 (
~
) 2 (
~
) 3 (
~
) 2 (
~
) 1 (
~
) 1 (
~
) (
~
2
2
2
2
1
2 2
2 2
2 2
L k x
k x
k x
P
P
L k x L k x
k x k x
k x k x

(30)
where L is the data window size, and
2
1 1
b p = (31)
1 1 2
2 b z p = (32)
Here, we can simplify (30) to the following:
B AP = (33)
By the LSE method, P can be obtained as
B A A A P
T 1 T
) (

= (34)
By (31) and (32), we can obtain
1 1
P b = (35)
1
2
1
2b
P
z = (36)
Now, these three subsections can be summarized as illustrated
in the flow chart in Fig. 1.
Fig. 1. Computation flow chart
III. TEST RESULTS
In this section, we present some simulation and field test
results to demonstrate the performance of this hybrid method.
The simulation tests are performed by Matlab/Simulink, and
the field measurement data are provided by ADX Company.
A. Matlab/Simulink Simulation Tests
To demonstrate the online performance of this hybrid
method, we simulate a two-machine transmission system by
Matlab/Simulink, as shown in Fig. 2. Two simulated
ADX3000s, which are phasor measurement units (PMU)
produced by ADX Company are installed in Bus 1 (B1) and
Bus 2 (B2), respectively, and the length of the transmission
line is 700 km. The hybrid method is performed by Embedded
Matlab Function, and the input data are the phase difference
between B1 and B2. A single-phase ground fault occurs at 0.1
sec and is cleared at 0.2 sec, and the total simulation time is 10
sec. The parameters used in Fig. 2 are provided in Table 1.
Two different cases are provided in this paper. One is the
unstable case, and the other is the stable case.
Fig. 2. Simulation Structure for real time low-frequency oscillation estimation
Table 1
Parameters used in the simulation system
Case 1: These two machines turn off the PSS, and the fault
occurs at the middle of the transmission line. Fig. 3(a) shows
the phase difference between B1 and B2, and Figs. 3(b), (c),
and (d) show the frequency, decay constant, and amplitude of
oscillation, respectively. From Fig. 3(a) and 3(d), the phase
difference between B1 and B2 increases rapidly after the fault
occurs. Fig. 3(b) and Fig. 3(c) show an oscillation frequency
of about 0.9Hz and positive decay constant, respectively. From
these figures, it is clear that the system is unstable.
Case 2: Machine M1 turns on the PSS, but M2 still turns off
the PSS, and the fault occurs at 550 km from the distance of
B1. Fig. 4(a) shows the phase difference between B1 and B2,
and Figs. 4(b), (c), and (d) show the frequency, decay constant,
and amplitude of oscillation, respectively. In this case, the
system is stable, because the decay constant is negative.
Moreover, the computation stopped when the amplitude of
oscillation is less than the threshold.
From these two cases, we can find that the hybrid method is
suitable for online application and the decay constant is a good
index to show the tendency of the system. Therefore, damping
ratio, as shown in (37), is a very important index in field
application.
Fig. 3. Simulation and computation results (unstable)
Fig. 4. Simulation and computation results (stable)
( )
% 100
2
2
1
2
1
1

=
f
ratio damping


(37)
B. Field data tests
Since low-frequency oscillation has been detected in Tai-
power system in 1984, Taipower Company had strengthened
its transmission system. Ever since, spontaneous low-
frequency oscillation has not appeared again. However, we
still can measure some decaying oscillations when some
disturbances occur in Taipower system.
The measured data of PMUs in Taipower system can be
categorized into two types: the real-time data which is updated
at a rate of 20 Hz and historical data which are contained in
the database. Hence, to meet the on-line and off-line
Fig. 5. the diagram of Taipower system
applications and requirements, we have developed two
different versions of this hybrid method. Furthermore, these
two versions are implemented by Java language and are under
test in Taipower system.
The main purpose of providing off-line version is for
convenient in report-writing. The off-line version provides
easy-reading information for the administrative department. It
gives data such as the time when oscillation actually started,
the principal frequency and damping ratio, and the duration of
the oscillation. Since the off-line version uses historical data
for estimation, it can estimate the parameters when oscillation
occurs.
Meanwhile, the on-line version is also provided to help
central operators of Taipower Company understand the real
situation of low-frequency oscillation. The on-line version
shows real-time estimation parameters of low-frequency
oscillation. Since the on-line version uses real-time data for
estimation, it is restricted by the data rate and estimates the
parameters in a fixed window.
Fig.5 contains the map of Taiwan. Taiwan is a long, narrow
and mountainous island. Since most of its mountains are in the
east, the major development area is in the west. Accordingly,
the major electric transmission lines are almost located in the
western Taiwan as shown in the bold solid lines on the map.
Since the major loads are located in northern Taiwan and most
generating plants are in the middle and southern Taiwan,
significant power had been transferred from south to north.
Therefore, inter-area mode oscillations are the major concerns
of Taipower system.
On January 24, 2004, some faults happened in Taipower
system, which caused a little swing in Taipower system. The
field data and computation results are shown in Fig. 6, 7 and 8.
We use dotted lines to show off-line computation results and
solid lines to represent on-line computation results.
Fig. 6(a) shows the phase difference between Lung Tan and
2nd Nuclear which are in the northern Taiwan and Fig. 7(a)
shows the phase difference between Lung Chi and 2nd Nuclear
which are in the southern Taiwan. It is clear that the oscillation
in northern and southern Taipower system is smaller than the
Fig. 6. Field data and computation results (Lung Tan 2nd Nuclear)
Fig. 7. Field data and computation results (Lung Chi 3rd Nuclear)
Fig. 8. Field data and computation results (3rd Nuclear 2nd Nuclear)
oscillation between 3rd Nuclear and 2nd Nuclear as shown in
Fig. 8(a). This means that Taipower Company still has to keep
an eye on inter-area mode oscillations.
Figs. 6(b), 7(b), and 8(b) show the frequency of oscillation,
which is lower than 0.8 Hz. Figs. 6(c), 7(c), and 8(c) show the
damping ratio of oscillation, and Figs. 6(d), 7(d), and 8(d)
show the amplitude of oscillation.
IV. CONCLUSIONS
In this project, we present a hybrid method which combines
four different algorithms including DFT, FIR window filter,
LSE, and a prototype algorithm for low-frequency oscillation
parameter estimation, each part having its own function. We
also use two different kinds of test to verify this hybrid method,
and the computational results prove to be encouraging. Since
this hybrid method is used in WAMS which is built by ADX
Company in Taiwan, it follows the specifications provided by
Taipower Company and ADX Company for both on-line and
off-line applications.
V. ACKNOWLEDGEMENT
This study was supported by a joint grant from the National
Science Council, ROC, and ADX Company. (NSC 94-2622-E-
214-015-CC3)
VI. REFERENCES
[1] K. P. Poon and K. C. Lin, Analysis of Transient Stability Swings in
Large Interconnected Power Systems by Fourier Transformation, IEEE
Transactions on Power Systems, Vol. 3, No. 4, Nov. 1988, pp. 1573-
1581.
[2] K. C. Lin and K. P. Poon, Analysis of Power Systems Dynamic
Oscillations with Beat Phenomenon by Fourier Transformation, IEEE
Transactions on Power Systems, Vol. 5, No. 1, Feb. 1990, pp. 148-153.
[3] C. L. Chang, A. S. Liu, and C. T. Huang, Oscillatory Stability Analysis
Using Real-Time Measured Data, IEEE Transactions on Power
Systems, Vol. 8, No. 3, Aug. 1993, pp. 823-829.
[4] T. Hashiguchi, M. Yoshimoto, Y. Mitani, O. Saeki, and K. Tsuji,
Analysis of Power System Dynamics Based on Multiple Synchronized
Phasor Measurements, IEEE Power Engineering Society General
Meeting, 2003, Vol. 2, July 2003.
[5] P. Korba, M. Larsson, and C. Rehtanz, Detection of Oscillations in
Power Systems Using Kalman Filtering Techniques, Proceedings of
2003 IEEE Conference on Control Applications, Vol. 1, No. 1, June
2003, pp. 183-188.
[6] D. J. Trundnowski, Order Reduction of Large-Scale Linear Oscillatory
System Models, IEEE Transactions on Power Systems, Vol. 9, No. 1,
Feb. 1994, pp. 451-458.
[7] D. J. Xu, R. M. He and T. Xu, A New Approach to Power System
Electromechanical Oscillation Research, Power Systems and
Communications Infrastructures for the Future, Beijing, Sept. 2002.
[8] M. Amono, M. Watanabe, and M. Banjo, Self-testing and self-tuning
of power system stabilizers using Prony analysis, IEEE Power
Engineering Society 1999 Winter Meeting, Vol. 1, Feb. 1999, pp. 655
660.
[9] Jinyu Xiao, Xiaorong Xie, Yingduo Han, and Jingtao Wu, Dynamic
Tracking of Low-Frequency Oscillations with Improved Prony Method
in Wide-Area Measurement System, IEEE Power Engineering Society
General Meeting, 2004, June 2004, pp.1104 1109.
[10] M. Meunier, and F. Brouaye, Fourier transform, wavelets, Prony
analysis: tools for harmonics and quality of power, 8th International
Conference on Harmonics and Quality of Power 1998, Vol. 1, Oct.
1998, pp. 71 - 76.
[11] M. Klein, G. J. Rogers and P. Kundur, A Fundamental Study of Inter-
Area Oscillations in Power Systems, IEEE Transactions on Power
Systems, Vol. 6, No. 3, August 1991, pp. 914-921.