1

DSP and Power System Protection

Richard W. Wall
University of Idaho
Department of Electrical Engineering
Moscow, ID 83844-1023
January 30, 2000
(208) 882-8334 [email protected]

Why talk about Power Systems?

This article is intended to demonstrate how DSP is used in a particular

industry for real-time instrumentation. It is my hope that those in other areas who
have similar instrumentation needs can leverage from the digital filtering
experience and successes in the power system protection area.

What is power system protection?

The good news is that the electrical power grid did not fail because of any
Y2K glitches. This we all know, but do you know why? It was because our
power systems have protection. This year I am on sabbatical with a company
that makes power system protective relays. For those outside the power
industry, power system protection is not money you pay to the godfather to keep
your lights on. It is the art and science of the application of devices that monitor
the power line currents and voltages and generate signals to deenergize faulted
sections of the power grid. The goal is to minimize damage to equipment and
property and maintain the delivery of electrical energy to the consumer.
Fifty years ago, electromechanical protective relays were used almost
exclusively. In general, these devices use torque generated by AC currents to
magnetically close a set of mechanical contacts. The contacts are held open or
“restrained” by a mechanical spring much like the common circuit breaker we use
in our homes. In reality, these devices were vastly more complicated.
Frequently the phase relationships of currents and voltages allowed the relay to
determine the direction of the fault relative to the relay. This makes the relays
selective resulting in deenergizing only the parts of the power grid that are
absolutely necessary to isolate the faulted section.

Computer entry into the power protection

Computer based relaying was experimented with in the 50’s and 60’s but
were not commercially viable because of the size, expense, and reliability of
early computers. Along came the 70’s and the microcontroller revolution, which
of course, changed all of that. Believe it or not, the microprocessor-based relays
used those same torque relationships that their predecessor electromechanical
relays used. Instead of using magnetic flux to generate the torque, the
microprocessor relay computed the torque. Modern microprocessor-based relays
still use the magnitude and phase of the 60 Hz (50 Hz in Europe) fundamental
 2

power voltage and currents to compute torque like quantities that determine the
existence of faults.
One of the challenges then (and continues to be today) is how to reliably
and efficiently convert sampled analog signals to magnitude and phase needed
for the torque equations. Since the fundamental component is the only signal of
interest, all other signals, whether they be harmonics, arcing noise or transients
generated by exciting the natural modes of the electrical network, are considered
noise that corrupts the signal of interest. Schweitzer and Houi reviewed seven of
the more common algorithms used to convert a time sequence to a time varying
complex vector. Only two of these algorithms are based on orthogonal basis set
decomposition similar to the Fourier transform. Discussion in the article will be
limited to only one of these algorithms because of its simplicity, efficiency, and
performance, the DFT.

Switching domains

The discrete Fourier transform (DFT) is a digital filtering algorithm that

computes the magnitude and phase at discrete frequencies of a discrete time
sequence. Fast Fourier transforms are computationally efficient algorithms for
computing DFTs. FFTs are useful if we need to know the magnitude and/or
phase of a number individual or band of frequencies. Jack Crenshaw told us all
about Fourier transforms, DFTs and FFT in previous ESP in a series of articles
spanning Oct. ’94, through Mar. ’95. But DFTs are simply FIR digital filters and
Crenshaw told us “more about filters” (June ‘96), “filters, the very last word”
(Sept. ’96) and “filters – a few more words” (Nov. ’96). After Jack said all there
was to say, Don Morgan told us about the “fundamentals of FIR design” in a
sequence of four ESP articles starting in June ’97. With that plethora of
background information, we can jump right into the application.
For protective relaying, we don’t really need to extract the magnitude and
phase of every signal contained in the sampled data stream, as is the case of an
FFT. This is especially true if there is only one signal present to begin with. If we
use a DFT for the only signal we’re interested in, we have both a conversion
algorithm and a band bass filter.

Theory to application

So let’s try out the theory and see how it works. Lets assume that we are
dealing with a 60 Hz signal that is sampled synchronously. This means that the
sample interval is the inverse of an integer multiple of 60. We need to compute
the DFT for the fundamental using equation (1) where N equals to the number of
samples per fundamental cycle, k equal to one for the fundamental, and n is the
coefficient subscript. Because digital computers (like most of the world) don’t
really understand the concept of imaginary numbers, two digital filters are
required, one to get the real part and one for the imaginary part. Mathematically,
the coefficients of these filters are by determined using (2).
 3

 2 π nk  (1)
− j 
Ck n = e  N 

 2    2 π nk   2π nk  (2)
Ck n =   cos  + j sin   = Ak n + jBk n
 N   N   N 

After computing the outputs of two filters using equation (3), we have the
desired complex vector shown in (4). Remember from a distant past math class
on complex variables that addition of complex numbers is easiest using
rectangular notation while multiplication is easiest using polar notation shown in
(5). For real-time applications, the conversion back and forth between the two
notations usually requires more time than can possibly be gained. Hence, the
complex variables are usually exclusively dealt with using rectangular form until
such time as a magnitude or phase is explicitly needed. This is particularly true
for processors that must use software routines for computing transcendental
functions (trig, log, and exponential functions). To further increase speed,
magnitude threshold levels are frequently left squared and angles kept as ratios.

N −1 N −1 (3)
Yn (re)= ∑ An X n , Yn (im) = ∑ Bn X n
0 0

Yn = Yn (re) + jYn (im) (4)

 Y (im)  (5)
Yn RMS = Yn (re) 2 + Yn (im) 2 , ∠Yn = arctan n 
 Yn (re) 

Transition phase

The conversion process works fine when every thing is at steady state –
amplitude, frequency, and phase is held constant or is changing very slowly
compared to the frequency of interest. But faulted power systems happen in a
flash (pun intended) and these faults can be modeled as step changes as
illustrated in Figure 1. During the transition period, the DFT output changes at
each sample point until the algorithm processes a complete cycle’s worth of
steady state data. In this period, the algorithm-generated transient makes the
DFT results an inaccurate representation of both the magnitude and the phase
(Figure 2) for a signal that has already achieved steady state.
Figure 1 also illustrates a point made earlier. That being a whole cycles
worth of the steady data must be sampled and processed by the DFT before
steady state is achieved. This is true regardless of the sample rate. I will cover
more on this later. Another observation from Figure 1 is the magnitude scaling,
which for this case, it is not RMS but peak. To obtain the RMS value, you simply
make the multiplier in (2) √2/N instead of 2/N.
Looking at the phase output in Figure 2 is sure cause to wonder of what
value is it. Very little, in of itself, because phase has no value without a time or
phase reference. The difference between the two DFT phase results accurately
represents phase relationship of the two signals if they are at the same frequency
 4

and sampled simultaneously. The phase steps ahead, as is seen in Figure 2,

each time a new sample is processed. The value of N or the size of the DFT
determines the size of the phase step as shown in (6). The absolute value of the
phase at any time is arbitrary unless the samples are somehow synchronized
with a signal’s zero crossing.

 2π  (6)
∆φ =  
 N 

Figure 1. Sampled sine wave with four-point DFT magnitude response to a step
change.
 5

Figure 2. DFT Phase response to a step input sine wave.

Frequency Response of DFTs

FTs represent the spectrum of the sampled data with a set of discrete
frequencies evenly spaced between zero and half the sampling rate, FS, minus
one interval. The interval between the discrete frequencies is equal to the FS/2N
where N the size of the DFT or the number of coefficients determined by (1).
Figure 3 shows the frequency response of DFTs over the range of zero to 480 Hz
for N equal to four and eight. This figure also shows the effects of aliasing
around the fold-over frequency F1, F2, and F3 for the four-point filter sampled at
240 Hz and just F2 for the eight-point filter sampled at 480 Hz. The advantage of
the eight-point filter is that the filter response is zero at harmonics except the
fundamental and the (N-1)th harmonic. Odd harmonics of 60 Hz are of particular
concern because they are generated by power transformers saturation and
nonlinear loads such as switching power supplies.
 6

D
Figure 3. Frequency response of a 4 and 8 point DFT with sampling rate of 240
Hz and 480 Hz respectively.
A colleague once commented that when your only tool is a hammer, every
thing tends to look like a nail. The same is true here. The DFT we are using is
looking only for 60 Hz and any energy that is passed by the filter characteristics,
regardless of the actual frequency, is aliased to appear that it is energy at 60 Hz.
But in this case, aliasing is our friend as well as our nemesis. The zeros at
harmonics on the high side of the Nyquist frequencyii (FS/2) work to our benefit.

Points to ponder

At this point we can draw some conclusions. When it comes to

DFTs, more is not necessarily better, sometimes it’s jut more. Higher order DFTs
provide greater harmonic rejection but do not decrease the algorithm transient
time. If you don’t believe me, look at Figure 4, which also bears out the claim that
a complete cycle of the fundamental must be sampled to achieve steady state
results. However, if we expect harmonics, then clearly higher order filters will
help reduce the both magnitude and phase errors.
 7

Figure 4. Magnitude responses for a 4, 8, and 16 point DFT to step change of

sine wave amplitude.
What if the signal being sampled is not exactly at 60 Hz? Well, the
magnitude will change according to the filter response attenuation for that
frequency as shown in Figure 3. DFT magnitude will not change much if the
frequency is not too far off. We will notice that the frequency increments no
longer adhere to (6). In fact, the difference in phase increment is exactly
proportional to the frequency difference. Say that the actual frequency is 59 Hz.
This means that the frequency difference is 1 Hz or 2π radians / second. If the
sampling rate is 240 Hz, then the phase shift will be off by (2π/240) radians /
sample.
We can solve this another way too and then it becomes a frequency
meter! Say we know that the system is operating in steady state and we
calculate the measured phase step (from the DFT output) and the expected
phase step according to (6). Then the actual frequency of the sampled signal is
the fundamental frequency ± the difference frequency. Mathematically, if it works
as shown in (7) through (9). When calculating the phase difference in (7), be
sure to consider the case when successive iterations are on opposite sides of the
2π / zero radian boundary. Do not expect the accuracy of such approach to
compare favorably with conventional zero crossing detectors.
 2π 
δφ =   − (φ n − φ n−1 ) (7)
 N 
 Fs 
δ f = δ φ  , where Fs is the sample rate. (8)
 2π 
f ACTUAL = δ f + f 0 (9)
 8

The fly in the ointment

With the transient response time fixed by the fundamental frequency and
the errors resulting from harmonics eliminated by the zeros of the DFT filter
response, what’s left? Usually higher harmonics and high frequency oscillations
due to exciting natural resonance modes in the power system network are
removed by analog filtering prior to sampling. Even though the DSP filter has a
zero at DC, power system faults frequently generate other low frequency
components commonly called DC offset. It is not really DC but a slowly decaying
exponential superimposed on the AC signal as shown in Figure 5. Also shown in
this figure are the DFT magnitudes for the signal without the superimposed
exponential and the offset AC signal. If the DFT for the non-offset AC signal is
considered optimal then the other DFT is what we are stuck with.

Figure 5. DFT filtering of a fully offset sine wave.

One readily observes that the response of the DFT to the offset sine wave
approaches the steady state value sooner and could concluded that the offset is
to our benefit. Unfortunately, protective relays could interpret the higher
magnitude as a fault that results in an incorrect operation. Today’s power
systems are operated so closely to designed capability that frequently a slow
correct operation is preferable to fast but possibly incorrect operations. As the
engineer, you have a choice to make.
 9

Tricks of the trade

For the power industry, it is certainly in their best interest to reduce the
response of the DFT due to the offset. One trick frequently employed is to use
only the coefficients to compute the real part of the DFT shown in (3) that are
generated by the cosine function. This is sometimes called a Cosine filter. Figure
7 shows the frequency response of the Cosine filter compared to the DFT filter.
Note that the Cosine filter favors higher frequencies and attenuates the
frequencies close to zero. This is good when trying to filter out a slowly decaying
exponential. There is also a computational advantage to eliminating the multiply
and accumulate instructions associated with imaginary term.
Note also from Figure 7 that the Cosine filter matches the response on the
DFT at 60Hz so there is no amplitude compensation required. However, off-
frequency signals will be more affected by the Cosine filter frequency response
than for DFT filters. One solution is to adjust the sampling rate to be an integer
number of the fundamental. This can accomplished by measuring the period
with a zero-crossing detector. Adjustments to the sampling period should be
slow so to track only the power system frequency changes and not frequencies
generated by transients.iii

Figure 6. Frequency response of an eight-point Cosine DFT filter

The trick is to use the result generated by the Cosine filter for both the real
and imaginary parts of the complex vector. This is accomplished by making the
 10

most recent Cosine filter output the real term and the output that has been
delayed a quarter of the period of the fundamental the imaginary term as shown
in (10) and (11). Both the real and the imaginary terms now have identical
frequency responses.
N −1
 2   2π n  (10)
Yc n = ∑ An X n , An =   cos 
0 N  N 
Yn = Ycn + jYcn− N / 4 (11)
Since we know that the DFT of the pure sine wave is the desired output
we can make it our evaluation reference. By computing the absolute difference
between reference output and the outputs of the DFT filter and the Cosine filter,
we can see the improvement. This is done in Figure 7 labeled D1 and D2
respectively. The difference for the Cosine filter response has reduced
overshoot and achieves an overall smaller difference. The cost of the improved
offset rejection is that the filter transient is extended by the time equal to one
quarter of the period of the fundamental. This is not obvious from Figure 5
because it is difficult to differentiate the signal transient from the algorithm
transient.

Figure 7. Differences of a DFT and Cosine filter for an offset sine wave
compared to a pure DFT for a sine wave without the offset.

Wrap up
Remember the leader of the Jodi Foster movie, “Contact”. The camera is
supposedly starts at planet earth and the sound track plays what seems to be the
simultaneous audio from all radio and TV broadcasts. The camera take a path
through inter stellar space that leads further from our reality and all the while the
 11

audio becomes more focused on fewer and fewer broadcasts. Finally, we’re left
with a single radio transmission of a young girl on a ham radio. I feel that this
article has taken a similar path.
The idea here is that we’re interested signals at one frequency only and
we needed an algorithm that quickly and accurately computes the magnitude and
phase of that signal. We can take advantage of aliasing to cancel harmonics if
we don’t expect that the signal will contain energy that is also passed by the
aliasing. DSP tricks can improve performance but always come at a price. It is
the responsibility of the designer to understand the application sufficiently to
know where compromises are tolerable to achieve the desired performance.

Final words of caution

The voltages on lines that deliver power make them lethal. Relays cannot
operate fast enough to prevent serious injury or death to someone coming in
contact with an energized power line. In an emergency situation, never assume
that relays have operated and the lines are deenergized. One of the most
difficult conditions to detect is a distribution line that broken and fallen to the
ground. The fault current is so small that most relays cannot sense the fault.
Always assume power lines are energized and treat them accordingly.
i
E.O. Schweitzer and D. Hou, “Filtering for Protective Relays”, 47th Annual Georgia Tech
Protective Relay Conference, Atlanta GA. April 28-30, 1993. This article is available for
download in PDF format at http://www.selinc.com/techpprs.htm.
ii
Digital Filtering: an Introduction, Edward P. Cunningham, Houghton Mifflin Co., 1992, ISBN 0-
395-53989-7
iii
R.W. Wall and H.L. Hess, “Design of Microcontroller Implementation of a Three Phase SCR
Power Converter”, Journal of Circuits, Systems, and Computers, Vol. 6. No. 6. Mar. 1997, pp.
619-633.