You are on page 1of 5

1

RMS-Energy Filter Design for Real-Time


Oscillation Detection
Matt Donnelly, Dan Trudnowski, James Colwell John Pierre Luke Dosiek
Montana Tech University of Wyoming Union College
Butte, MT USA Laramie, WY USA Schenectady, NY USA

(BPA) control center using a system-wide real-time PMU


AbstractReal-time monitoring of power-system oscillations network. It is also being implemented as part of an automated
is a concern in the operation and control of synchronous power supervisor for the DC modulation damping controller
systems. A simple and powerful approach is to use an RMS described in [1].
energy filter. As defined in this paper, an RMS energy filter
Several approaches for oscillation detection are possible.
estimates the total RMS energy of a signal in a defined frequency
band. This approach has been implemented in the operation One of the first applied to a power system is contained in [2].
control center at the Bonneville Power Administration for the This approach uses a linear band-pass filter, cascaded with an
past year, and it is being implemented in a DC modulation absolute operator, cascaded with low-pass filter. The
damping control monitoring system. This paper describes RMS approach is simple, and proved to work very well on actual-
energy filter design requirements, approaches, and alternatives. system data. According to oral interviews with retired BPA
Simulation examples demonstrate the performance.
engineers, the method in [2] was successfully used for several
Index TermsPower system dynamics, electromechanical
oscillations, forced oscillations, Phasor Measurement Units. years to trigger a data collection system to archive oscillation
events in the electromechanical frequency range. It proved to
I. NOMENCLATURE be very robust. In [3], long-windowed FFTs are employed to
identify small-amplitude sustained oscillations. The method is
ARMA = Auto-Regressive Moving Average.
applied to PMU data from several actual-system forced
BP = Band Pass.
oscillation cases. An oscillation detection approach based
FFT = Fast Fourier Transform.
FIR = Finite Impulse Response. upon Welchs periodogram and used for estimating the
IIR = Infinite Impulse Response. frequency, starting time, ending time and amplitude of an
LP = Low Pass. oscillation in noisty PMU data is presented in [4]. A detection
OD = Oscillation Detection. method for identifying the frequency of a very low-level
PMU = Phasor Measurement Unit. oscillation is presented in [5].
RMS = Root Mean Square.
sps = Samples Per Second. III. RMS ENERGY FILTER APPROACH
Some goals guiding the requirements for an OD are:
II. INTRODUCTION The OD output must have a useful unit of measure.

T HE physics governing a synchronous power system


naturally produce an oscillatory response. Also,
controllers governing the response of all types of connected
Ideally, the OD should provide the total RMS energy of a
signal in a given frequency band of interest.
The OD must be fast enough to alarm on an oscillation
apparatus can produce power system oscillations. The nature before the oscillations can cause harm.
of these oscillations widely varies and includes The OD must be robust with respect to rejecting the
electromechanical transients and forced oscillations. Real- ambient noise of the system. It must avoid false positives.
time monitoring of such activity is critical to the reliable The OD output should be high-information, and should be
operation of the system. The goal of an oscillation detector is intuitive and quantitative.
to quantitatively monitor the oscillation energy in a given These design goals led the authors to adopt an RMS Energy
frequency band for a given signal (such as the real power Filter approach based upon the method in [1]. The appraoch
flowing on a given line). In an operations environment, if the is depicted in block diagram form in Fig. 1.
oscillatory energy exceeds a threshold for a specified amount In the first stage of the filter a PMU-derived Input signal
of time, an alarm is provided to the system operator. Other (e.g., real-power or voltage magnitude) is formed, passed thru
applications of an oscillation detector include automated a down-sampling LP filter and subsequently down-sampled to
monitoring of real-time damping controllers. In this a sampling rate fd. The purpose of the LP filter in this stage is
application an OD may be used to detect failure of the to avoid aliasing from the down sampler. Down sampling
damping controller. The approach described in this paper has reduces numeric sensitivity of the BP filter. Depending on the
been implemented in the Bonneville Power Administration bandwidth of the BP filter, down sampling may not be
required, i.e. fd may be equal to the reporting rate of the input
signal.
This work was supported by the U.S. Department of Energy under grant
DE-FC26-06NT42750. In the next stage the signal is passed thru a BP filter that

978-1-4673-8040-9/15/$31.00 2015 IEEE


2

passes the desired band for oscillation detection. After BP electromechanical modes and generator controls. And Band 4
filtering, the signal is squared, passed thru a LP filter, and then starts at 5.0 Hz and extends to the Nyquist rate of the input
squared-rooted. The goal of the LP filter in this stage is to data. Band 4 may contain oscillatory dynamics typically
estimate the mean of the squared signal. This LP filter must be associated with generator torsional modes or dynamics
matched to the BP filter in order to achieve the desired result associated with fast-moving power electronic controllers, for
an output representing the RMS energy of the input signal in example. RMS energy filters might be designed for any
the bandwidth of the BP filter. number of frequency bands, though the authors feel that the
four bands shown in Fig. 2 provide sufficient information
without overloading the user with data.

Fig. 1: RMS energy filter.

As previously stated, the RMS energy filter is similar to the


approach in [1]. The primary difference is the use of squaring
and square-rooting versus the absolute operation in [1]. The
squaring/rooting allows for the output to have the units of total
RMS energy of the input signal in the bandwidth of the BP
filter. Other differences include the filter design approaches.
A. Filtering Background
Each of the filters in Fig. 1 is a standard linear discrete-time
filter. In its most general sense, a linear filter can be
represented as an ARMA difference equation [6]: Fig. 2: RMS energy filter settings for operation control monitoring.

A. RMS Energy Filter Requirements


= + 1 In order to guide the design of the filters for the four bands
in Fig. 2, the following design requirements were established.
1. The filter must have 90% steady-state accuracy for
bands 1 - 3 and 70% steady-state accuracy for band 4.
where x is the filter input, y is the output, k is integer time.
2. The response times must be as follows:
The ai coefficients represent the AR term and bi coefficients o Band 1: 200 sec. or less.
represent the MA part [6]. If the AR part is included, i.e. at o Band 2: 12 sec. or less.
least one ai coefficient is non-zero when i>0, then the filter is o Band 3: 6 sec. or less.
IIR. If the AR part is not included, the filter is then FIR. o Band 4: 6 sec. or less
A FIR filter has three primary advantages over an IIR filter. 3. The filter must show minimal ringing to a step or
First, FIR filters have a finite impulse response and that impulse input. That is, if a step or impulse input is
response and can be designed to have minimal ringing. An applied the output should not oscillate excessively.
IIR filters impulse response can have significant ringing Quantitatively, we define this to be more than 3 cycles
depending on the location of its poles. Second, by selecting of significant oscillation.
symmetric bi coefficients, a FIR filter can be designed to have 4. The filter must have minimum out-of-band (aka stop
a constant delay time which makes the response time exactly band) rejection, defined as 40 dB, with minimum
the same for any input. Lastly, FIR filters typically have much transitions bands.
lower sensitivity to numerical round off in the bi coefficients Although the 90% steady-state accuracy requirement may
making implementation more computationally robust. In both seem overly permissive, recall that the purpose of the OD is to
the IIR and FIR filters, numerical sensitivity can be reduced rapidly alert operations personnel to high-energy oscillations
by down sampling the data. and therefore it was deemed to be useful to sacrifice some
The primary advantage of an IIR filter over a FIR is the degree of accuracy for speed. Also because an M-class PMU
response time. For a given frequency roll-off, an IIR filter will also exhibit significant filtering above 5 Hz the steady-
will typically have a much faster response than an FIR filter. state accuracy of Band 4 is of even less concern; thus, a 70%
requirement.
IV. OPERATIONS AND CONTROL APPLICATION The response time requirements dictate the speed of
response of the RMS energy filter. Response time is defined
Interviews with prospective users of the OD led the authors
to be the amount of time for the output to estimate the total
to choose four different RMS energy filters implemented in
RMS content of the input with 90% accuracy. The response-
parallel with settings shown in Fig. 2. Frequency band #1 is
time requirements were established by considering the realistic
used to monitor very slow oscillations that typically involve
time for a filter to estimate the RMS content of a signal,
speed-governor controllers. Band 2 is tuned to oscillations
typically at least one full cycle of an oscillation at the lowest
typically observed in the inter-area electromechanical
frequency, and the operational requirements for decision-
oscillation range. Band 3 is typically associated with local
3

making to mitigate an issue detected by an OD.


The requirement related to ringing is intended to minimize 0
false positives. Many IIR filters have under-damped poles

Gain (dB)
resulting in an oscillatory response to a step or impulse input. -20

Such an input is common in the power-system data; for -40


example, a capacitor switching or a PMU outlier data point
-60
due to a communication error. 10
-2
10
-1
10
0

Freq. (Hz)
B. RMS Energy Filter Design
0.2
Using the above design requirements, many combinations

Pulse response
Delay = 4.7 sec.
of standard filters were tested. All filter designs presented in 0.1

this section are linear-phase MA filters, i.e. constant time-


0
delay FIR filters. The Matlab Signal Processing Toolbox [7]
was employed in the design process. Filters were designed for -0.1
PMU data operating at 30 sps, 60 sps, or 120 sps. The results 0 1 2 3 4 5 6 7 8 9 10
Time (sec.)
to follow are for the 30 sps data.
1) Band 1 Design
Using Fig. 1 as a reference, the down-sampling filter is Fig. 4: Gain and Pulse response for the 5-sps to 1-sps down-sample filter.
executed in two steps. First the data is passed through the MA
filter in Fig. 3 and down sampled to 5 sps. It is then passed
through the filter in Fig. 4 and down sampled to 1 sps. The 0

purpose of these two filters is to prevent any aliasing from the

Gain (dB)
-20 BP
down sampling. Note that both these filters have nearly no LP 2

ringing in the pulse response. Also note the excellent rejection -40

(over 40 dB) in the stop band. -60


0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
The 1-sps signal is then passed through the Fig. 5 BP filter, Freq. (Hz)
squared, passed through the Fig. 5 LP filter, and finally rooted.
0.4 0.02
The pulse response of the BP filter has one significant cycle of

LP 2 Pulse Response
BP Delay = 83 sec.
BP Pulse Response

ringing which is minimal for a BP filter. The LP filter is used 0.2 LP Delay = 50 sec. 0.015

to average the squared signal after the BP filter. It is critical 0 0.01


that the corner of this filter is below the lowest corner of the -0.2 0.005
BP filter as shown in Fig. 5. Both filters have excellent
-0.4 0
rejection above 40 dB in their respective stop bands. 0 20 40 60 80 100 120 140 160 180
The steady-state accuracy requirement is directly related to Time (sec.)

the amount of pass-band ripple in each filter. A pass-band Fig. 5: Gain and pulse response for the BP and LP filters for Band 1.
ripple of 1 dB or less will satisfy the requirement. Each of the
filters is well below this level. The total delay of all the filters
together is approximately 140 sec. which is less than the 2) Band 2 Design
required 200 sec. For Band 2, the input signal is first low-pass filtered
. through the MA filter in Fig. 3 and down sampled to 5 sps.
Again, the goal of the filter is to prevent aliasing from the
down sampling. The 5-sps signal is passed through the Fig. 6
0
BP filter, squared, passed through the Fig. 6 LP filter, and
finally rooted. Note that these filters meet the ringing and
Gain (dB)

-20
stop-band requirements. The total delay of all the filters
-40
together is approximately 11 sec. which is less than the
-60
-2 -1 0 1
required 12 sec
10 10 10 10
Freq. (Hz) 3) Bands 3 and 4 Design
0.2
No down sampling is utilized for Bands 3 and 4. The input
signals are directly input into the BP filter. For Band 3, the
Pulse response

0.1
Delay = 0.82 sec.
BP and LP 2 filters are shown in Fig. 7. These filters meet the
steady-state accuracy, ringing, stop-band requirements. The
0
total delay is approximately 4.5 sec. which is below the 6 sec.
requirement. The Band 4 filters are shown in Fig. 8. Again,
-0.1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 the requirements are met. The pass-band ripple is larger
Time (sec.)
because the steady-state accuracy requirement is reduced to
Fig. 3: Gain and Pulse response for the 30-sps to 5-sps down-sample filter. 70%. The total delay of these filters is 1.45 sec. well below
the required 6 sec.
4

= 100 + 12cos 2 0.3 + 7cos 2 0.7 + 5cos 2 4


10 40 3
BP
0 = 0
LP 2
otherwise
Gain (dB)

-20 The RMS content of this signal for the four bands defined
-40
above are:
Band 1 RMS content = 0
-60
12 7
Band 2 RMS content = 2+ 2 = 9.8
0 0.5 1 1.5
Freq. (Hz)

0.4 0.04
Band 3 RMS content = 5
2=3.5

LP 2 Pulse Response
BP Delay = 5 sec.
BP Pulse Response

0.2 LP Delay = 5 sec. 0.03


Band 4 RMS content = 0
0 0.02
The results are shown in Fig. 9. The raw signal is shown in
-0.2 0.01 the 1st plot while the Band 1 thru 4 RMS filter outputs are
shown in plots 2 thru 5. Note the estimated RMS content
-0.4 0
0 1 2 3 4 5 6 7 8 9 10 matches the expectation. Also note the response times match
Time (sec.)
those of the filter designs.
Fig. 6: Gain and pulse response for the BP and LP filters for Band 2.

150

BP

x(t)
100
0
LP 2
Gain (dB)

50
-20 0 10 20 30 40 50 60

-40 10 Band 1
RMS

-60 5
0 1 2 3 4 5 6 7
Freq. (Hz) 0
0 10 20 30 40 50 60
0.5 0.04
LP 2 Pulse Response

BP Delay = 3.27 sec.


BP Pulse Response

10 Band 2
LP Delay = 1.25 sec.
RMS

5
0 0.02
0
0 10 20 30 40 50 60

-0.5 0 10 Band 3
0 1 2 3 4 5 6 7
RMS

Time (sec.) 5

Fig. 7: Gain and pulse response for the BP filter and LP filter 2 for Band 3. 0
0 10 20 30 40 50 60

10 Band 4
RMS

0 5
Gain (dB)

-20 0
BP 0 10 20 30 40 50 60
LP 2 Time (sec.)
-40

-60
0 5 10 15 Fig. 9: Example 1 results.
Freq. (Hz)

1 0.08 V. DAMPING CONTROLLER SUPERVISION APPLICATION


LP 2 Pulse Response

BP Delay = 0.93 sec.


BP Pulse Response

0.5 LP Delay = 0.52 sec. 0.06 For the real-time damping controller [1], supervisory
0 0.04
detection was limited to Bands 3 and 4. But, an effort to
improve the response time of the Band 3 and 4 filters was
-0.5 0.02 desired. The goal was to disable the controller if it is observed
-1 0 to be participating in oscillations above its design frequency
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (sec.)
range. IIR filters were chosen for their speed, but they had to
be carefully chosen to meet the ringing requirements.
Fig. 8: Gain and pulse response for the BP filter and LP filter 2 for Band 4.
Chebyshev Type 2 filters were selected for the BP component.
C. Example They offer a flat pass band and fast roll-off that adequately fit
the parameters required. It is noted, the Band 3 BP and Band
As an example, the following signal is applied at the input 4 LP have more overlap than the FIR filters allowed. This was
of Fig. 2 deemed acceptable, due to the nature of the oscillation
detection. If the detector responded with an alarm in either
5

band or both, the result is the same; shut the controller down. response times match the expected and are much faster than
Also, the low-pass FIR filters were left unchanged as IIR the FIR design.
filters could not be designed that out-performed them.
150
1) Bands 3 and 4 Design
Once again, no down sampling is utilized for Bands 3 and 4.

x(t)
100
The input signals are directly input into the BP filters. For
50
Band 3, the BP and LP 2 filters are shown in Fig. 10. These 0 10 20 30 40 50 60
filters meet the steady-state accuracy, ringing, stop-band 10 Band 3
requirements. The total delay is approximately 1.26 sec.

RMS
5
which is significantly better than the 4.5 sec. provided by the
FIR version. The Band 4 filters are shown in Fig. 11. Again, 0
0 10 20 30 40 50 60
the requirements are met. The total delay of these filters is
10 Band 4
0.57 sec. well below the 1.45 of the FIR design.

RMS
5

0 BP 0
0 10 20 30 40 50 60
LP3 Time (sec.)
-20
Gain (dB)

Fig. 12: Example 2 results.


-40

-60
VI. CONCLUSIONS
-80
1 2 3 4 5 6 7 An approach for oscillation detection based upon an RMS
Freq. (Hz)
energy filter is detailed. Specific applications are operation
0.5
BP Delay = 0.1 sec.
0.016
control center alarming and monitoring of a real-time damping
LP 3 Pulse Response
BP Pulse Response

LP Delay = 1.25 sec. control system. These approaches are being implemented at
the Bonneville Power Administration. This paper describes
0 0.008
RMS energy filter design requirements, approaches, and
alternatives. Simulation examples demonstrate the
-0.5 0 performance.
0 1 2 3 4 5 6 7
Time (sec.)
VII. REFERENCES
Fig. 10: Gain and pulse response for the BP and LP filters for Band 3 for
damping controller supervision. [1] D. Trudnowski, D. Kosterev, J. Undrill, PDCI Damping Control
Analysis for the Western North American Power System, Proceedings
of the IEEE Power & Energy Society General Meeting, July 2013.
[2] J. Hauer and F. Vakili, An oscillation detector used in the BPA power
0 HP system disturbance monitor, IEEE Trans. on Power Systems, vol. 5, no.
LP4 1, pp. 74-79, Feb. 1990.
-20
[3] L. Vanfretti, L. Dosiek, J. W. Pierre, D. Trudnowski, J. H. Chow, R.
Gain (dB)

-40 Garca-Valle, and U. Aliyu, Application of ambient analysis techniques


for the estimation of electromechanical oscillations from measured
-60
pmu data in four different power systems, European Transactions on
-80 Electrical Power, vol. 21, no. 4, pp. 16401656, 2011.
2 4 6 8 10 12 14 [4] J. Follum and J. W. Pierre, Initial results in the detection and estimation
Freq. (Hz)
of forced oscillations in power systems, Proceedings of the North
1 0.04 American Power Symposium (NAPS), 2013.
HP Delay = 0.07 sec.
LP 4 Pulse Response

[5] N. Zhou and J. Dagle, Initial Results in Using a Self-Coherence


HP Pulse Response

0.5 LP Delay = 0.5 sec.


Method for Detecting Sustained Oscillations, IEEE Trans. on Power
Systems, to be published.
0 0.02
[6] C. Philips, J. Parr, and E. Riskin, Signals, Systems, and Transforms, 5th
-0.5 Ed., Pearson Publishing, 2014.
[7] Matlabs Signal Processing Toolbox version 6.18.
-1 0 www.mathworks.com.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time (sec.)

Fig. 11: Gain and pulse response for the HP and LP filters for Band 4 for
damping controller supervision.

B. Example
Using the same test signal in (3), the results are shown in.
The raw signal is shown in the 1st plot while the Band 3 and 4
RMS filter outputs are shown in 2nd and 3rd plots. Once again,
note the estimated RMS content of Band 3 and response times
match the expected. Due to the wider filter overlap of Band 3
and 4, Band 4 shows slight RMS content too. The filter

You might also like