Module 9 : Numerical Relaying II : DSP Perspective

Lecture 37 : Estimation of System Frequency

Objectives
In this lecture,

We will introduce the concept of DFT leakage, and use it to estimate magnitude and phase angle errors
due to change

in system frequency.

Error in Phasor Estimation due to Change in System Frequency

So far while discussing the phasor estimation problem, we have assumed that frequency of the power
system remains fixed at it's nominal value (
). Hence, we have also fixed the sampling
frequency to

, with N-point DFT being used for estimation.

However, during disturbance and even in steady state to a certain extent, the frequency varies. Thus, we
expect the phasor estimation under constant frequency assumption to be erroneous. Under such
situations, how good is our estimate of the phasor? We now plan to answer this question. As a by product
of the analysis, we will also develop a frequency estimation technique which can be used in under
frequency and rate of change of frequency relays. To simplify presentation, the analysis is developed
.
gradually. First, we determine the DFT of complex exponential signal at frequency

DFT of Complex Exponential

Let the signal be given by
(1)
For this signal let N-sample be captured in P-cycles. P can be a positive integer or even a positive real
number. Then, sampling interval (

) is given by

, sampling speed by

previous lectures, we had taken P = 1. The discrete sample at

expression:

. In the

is given by the following

(2)
Thus,

DFT component corresponding to frequency

, is given by

(3)

It should now be fairly easy to compute the summation in (3)

Let

, which is a constant once P and m are fixed.

(4)

DFT of Complex Exponential (contd..)

Then

(5)

(6)

Note that MATLAB defines

.

DFT of Complex Exponential (contd..)

. This convention differs from our convention viz.

DFT of Complex Exponential (contd..)

Fig 37.1(a) shows the envelope of response for
as a function of
, where
is treated as a
to be a discrete number, this plot is sampled at
continuous variable. Since in DFT we have to restrict
discrete points (
). The plot indicates that fundamental is extracted correctly as
expected. At all harmonic frequencies, the DFT gain i.e.

is zero. Infact, fig 37.1(a) is identical to

the frequency response of full-cycle Fourier algorithm. To be specific, we have chosen

, N =

12.
Fig 37.1(b) shows the envelope of response for 3 different frequencies 49, 50, 51 Hz. The sampling rate
is fixed at 12 samples per cycle at nominal frequency of 50 Hz. Thus, at 49 Hz, P = 49/50 and for 51Hz
waveform P = 51/50. It is seen from the figure that magnitude response is more or less identical when
Hz.
frequency deviation is within
Finally, fig 37.1(c) shows a set of response when frequency deviation from nominal is large enough i.e.
when
= 40, 50 and 60 Hz. Now it can be seen easily that, if we sample the envelope in fig 37.1(c)
), then the P = 0.8, and gain at 50 Hz is a
(not the time domain signal) of 40 Hz signal at 50 Hz (
finite non-zero value different from unity. This effect is known as DFT leakage. It can be said that the
energy in frequency bin
has leaked into frequency bin
.
Remark 1: As

varies from

, with a constant sampling frequency

proportionately. This is because we always obtain N-samples in

cycles of

Hz waveform.

, we note that P varies

seconds which is equivalent to

Remark 2: From eqn. (6), it is clear that DFT gain is unity at

. Thus, the curve of fig 37.1(a),

slides to the left when 'P' reduces below 1 and it slides to the right when
our waveform to be a 50 Hz signal, we always sample the envelope at

. However, if we assume
.

In DSP literature, DFT leakage is considered to be undesirable. It means that we wrongly interpret a 40
or a 60Hz signal as a 50 Hz signal. However, by a little more analysis, we will show that deviation small
enough from nominal frequency can be easily estimated from corresponding phase characteristic of DFT.
This not only allows us to build underfrequency and rate of change of frequency relays but it simplifies
hardware as sampling rate need not follow the system frequency. Consequently, no-zero crossing
detectors are required. From the relaying perspective, it turns out to be a boon in disguise.
As seen in fig 37.1(b), the DFT magnitude leakage for

Hz frequency deviation around nominal

= -

frequency say 50Hz, is practically negligible. For f = 51 Hz, it equals

0.0065. Thus, we conclude that effect of magnitude leakage on estimation of phasor magnitude can be
neglected. However, the phase angle of DFT tells another story. Note that
For

and

Thus,
Thus an error of is

introduced in phase computation due to 1 Hz frequency deviation.

DFT of real COSINE signal

So far, we have concentrated on DFT response of complex exponential. In practice, we are interested in
DFT of real sine (or) cosine signal and not the complex exponential signal. However, deriving the
response for real cosine (or) sine signal from response of complex exponential is not very difficult. From
the fact that
and by following similar steps as in the case of complex

exponential, we get

where P is the number of cycles of the signal e.g. 1,

1.25, 2 etc, in which N-samples for DFT are obtained.

Thus,

(7)

We have already deduced that

Similarly it can be shown that

Thus,

(8)
With P

1,

, it can be seen that

while,

and

Hence, following approximation can be made to equation (8).

(9)
We conclude that even with real sine or cosine signal, any reasonable deviation of the signal from the
nominal value, leads to negligible magnitude leakage. However, phase angle error for 1Hz deviation in
frequency is approximately
which is not negligible.

DFT of real COSINE signal

Fig 37.2 shows the envelope of magnitude response of
per eqn. (9). For these plots, we have used N = 32 and

for different frequencies 40, 50, 60 Hz as

= 160 Hz.

Estimation of Frequency
Now our aim, is to develop a method of estimating power system frequency using the recursive DFT

approach to phasor estimation. We plan to show that in the moving window approach the phase angle
estimated by recursive DFT approach rotates at a speed proportional to the deviation from the nominal
frequency. In turn, this deviation can be assessed by measuring the rate of change of phase angle.
In the previous section, we have derived the DFT leakage for complex exponential and real cosine signals.
We will now generalize the DFT computation of complex exponential so that it can handle moving window
concept.
Generalized DFT of Complex Exponential
Following the methodology used while generalizing DFT, we can write generalized DFT for

window

(window number corresponds to the sample number of first sample in the window) as:
(10)
where
Substituting,

in (9), we get

where

Summation

is the DFT corresponds to window number 1. Substituting from eqn. (8),

(11)

we obtain
If m = P, then it is clear that DFT

is stationary i.e. independent of window number. In relaying

applications where we are interested in fundamental phasor estimation, we choose

. Hence, we

can expect obtain stationary DFT in (10) when m = P = 1 i.e. when signal corresponds to fundamental
frequency.
With the knowledge of the generalized DFT of the complex exponential, now we can derive the
generalized DFT expression for real cosine signal.

Generalized DFT of Real Cosine Signal

For a real cosine signal, with moving window concept eqn (7) can be generalized as follows:
(12)

(13)

Following the similar reasoning as outlined in the previous section of the DFT of real cosine signal, we can

neglect the second term in (13). This leads to following approximation.

(14)

Observe that P = m = 1, we obtain stationary DFT because

deviates from nominal frequency i.e

, then

= 1. However, when frequency

, the DFT estimate for nominal frequency, will

start rotating along with the window. From, (14) we can derive that

Since fundamental phasor estimated for window number w' is given by

, we deduce that

(15)

is the sampling time interval. Summarizing, if we set sampling frequency for a sinusoidal signal

where

to provide N-samples per cycle at the nominal frequency

, (say 50 or 60 Hz), and keep the sampling

frequency invariant of the actual frequency of sinusoid, then the phasor estimated by moving window
approach rotates at a speed proportional to
. This rotation will be in anti-clockwise direction if
>
0 i.e.

>

. Conversely, it will be in the clockwise direction if

< 0 i.e.

<

Generalized DFT of Real Cosine Signal (contd..)

If we monitor, this phase rotation, then from the proportionality relationship of (15), we can estimate the
frequency '. If
denotes phase-angle, then from (15), we can obtain the rate of change of
as
follows:
(16)

Discussion
There are many advantages associated with the above algorithm. The method is not based on zero
crossing time and it is immune to noise and harmonics. Instead of using single phase quartiles, one can
estimate the positive sequence component and derive frequency from it. Such an approach will use all the
three phase voltages and hence will have better noise rejection properties. At harmonics of nominal
and hence, the above approach will reject frequencies m
completely.
frequency
Measurement of Frequency
To measure

accurately, we integrate eqn. (16) until

equal to 0.5 radians. Let T' be the time to

accumulate this change:

If we assume that frequency computation will be further averaged over four measurements to smoothen
is given by
out noise, then time to compute deviation

Thus, a frequency deviation of 1 Hz is detected in 0.32sec while a frequency deviation of 0.1 Hz is

detected in 3.2sec. This is an illustration of standard speed versus accuracy conflict in relaying. This
technique can be recommended for development of under frequency and rate of change of frequency
relays.

Recap

In this lecture we have learnt the following:

We introduced DFT leakage. It was shown that when fixed sampling frequency is used, typically Nsamples in 1 - cycle

at nominal frequency, then, DFT leakage is zero. This means that there are no magnitude and phase angle
errors in estimation. However, when the system frequency deviates from the nominal (of the order
say), then errors introduced in estimating the amplitude of the signal is negligible. However, now phase
angle errors are not negligible.
It was shown that if the frequency deviates from the nominal value, with constant sampling frequency,
the phasor starts rotating at a speed proportional to it. This can be used for frequency estimation.