Digital Power System Protection

DigitalPower
SystemProtection
S.R. Bhide
Associate Professor of Electrical Engineering
Visvesvaraya National Institute of Technology
Nagpur
Delhi 110092
2014
Digital power system protection
S.R. Bhide
2014 by PHI Learning Private Limited, Delhi. All rights reserved. No part of this book may be
reproduced in any form, by mimeograph or any other means, without permission in writing from the
publisher.
ISBN-978-81-203-4979-7
The export rights of this book are vested solely with the publisher.
Published by Asoke K. Ghosh, PHI Learning Private Limited, Rimjhim House, 111, Patparganj
Industrial Estate, Delhi-110092 and Printed by Star Print-O-Bind, F-31, Okhla Industrial Area
Phase I, New Delhi-110020.
To
the memomy of My Parents
Pushpa and Ramkrishna Bhide
Contents
Prefacexi
1. Evolution of Power System Protection and the Emergence

of Digital Relaying 15
1.1 Evolution of OC Relay from an Ammeter 1
1.2 Evolution of a Directional Relay from a Wattmeter 2
1.3 The Birth of Transistor 2
1.4 Start of a Revolution 3
1.5 Architecture of the Modern Digital Relay 4
2. Digital Signal Processing Basics and Architecture

of Numerical Relay 628
2.1 Introduction to Digital Signal Processing 6
2.2 The DSP Signal Processing Chain 7
2.3 Analog to Digital Converters 7
2.3.1 Quantization Error 8
2.3.2 ADC Types 12
2.4 Sampling 14
2.4.1 Need for Sample and Hold Circuit 14
2.4.2 Shannons Sampling Theorem and Aliasing 18
2.5 Anti-aliasing Filter 21
2.5.1 Design of Anti-aliasing Filter 22
2.5.2 Numerical Example 23
2.6 Functional Block Diagram of Numerical Relay 24
Review Questions27
v
viContents
3. Algorithms Based on Undistorted Single Frequency Sine Wave 2959

3.1 Mann and Morrison Algorithm 29
3.1.1 Historical Perspective 29
3.1.2 Derivation of the Sample and Derivative (Mann & Morrison)
Algorithm30
3.1.3 Instantaneous OC Relay Based on Mann and Morrison Algorithm 32
3.1.4 Simulation of the Mann and Morrison Algorithm in Spreadsheet 33
3.1.5 Simulation of Mann and Morrison Algorithm in MATLAB 35
3.2 Three-Sample Technique 38
3.2.1 Determination of Frequency of the Signal 38
3.2.2 Determination of Amplitude of the Signal 39
3.2.3 Determination of Phase of the Signal 41
3.2.4 Summary of Three Sample Algorithm 41
3.2.5 Excel Spreadsheet for Simulation of Three-Sample Algorithm 42
3.2.6 MATLAB Program for Simulating Three-Sample Algorithm 43
3.3 First and Second Derivative Algorithm 44
3.3.1 Excel Spreadsheet for Simulation of First and
Second Derivative Algorithm 47
3.3.2 MATLAB Simulation of First and Second Derivative Algorithm 48
3.4 Two-Sample Technique 51
3.4.1 Spreadsheet for Two-Sample Technique 53
3.4.2 MATLAB Simulation of Two-Sample Technique 54
Review Questions57
4. Algorithms Based on Solution of Differential Equation 6086

4.1 Differential Equation Algorithm 60
4.2 Justification for Lumped Series R-L Model 62
4.3 Excel Spreadsheet Implementation of the Differential Equation
Algorithm65
4.4 MATLAB Implementation of Differential Equation Algorithm 66
4.5 Solution of Differential Equation Algorithm Using Numerical
Integration68
4.6 MATLAB Implementation of Solution of Differential Equation
Algorithm Using Numerical Integration 71
4.6.1 Trapezoidal Rule for Numerical Integration 71
4.6.2 Simpsons Rule for Numerical Integration 72
4.6.3 MATLAB Script of Implementation of Solution of Differential
Equation Algorithm Using Numerical Integration 73
4.7 Application of Differential Equation Algorithm to
Three-Phase Line 76
4.7.1 Ground Fault Protection of Three-Phase Line
Using Phase Quantities 77
4.7.2 Ground Fault Protection of Three-Phase Line Using
Sequence Quantities 79
4.7.3 Phase Fault Protection of Three-Phase Line Using Phase
Quantities81
Contentsvii
4.7.4 Phase Fault Protection of Three-Phase Lines Using

Sequence Quantities 83
4.8 Elimination of Specific Harmonics by Integration between
Selected Limits 84
Review Questions86
5. Algorithms Based on Least Squared Error (LSQ) 8798

5.1 LSQ Technique 87
5.2 Average is Best Value in LSQ Sense 88
5.3 LSQ and Pseudo-Inverse 89
5.4 Relation between LSQ and Pseudo-Inverse 90
5.5 LSQ Algorithm by Sachdev 92
5.6 MATLAB Implementation of LSQ Algorithm 96
Review Questions98
6. Discrete Fourier Transform 99132

6.1 Introduction to Concept of Spectrum 99
6.1.1 Generalised Fourier Series 101
6.1.2 Dirichlet Conditions 102
6.1.3 Fourier Series in Trigonometric Form 102
6.1.4 Fourier Series in Exponential Form 104
6.2 Fourier Coefficients are Best in LSQ Sense 105
6.3 Summary of Fourier Series 107
6.4 Applications of Fourier Analysis 108
6.5 Fourier Series for a Recurring Pulse 108
6.6 Recurring Pulse to Non-recurring Pulse 111
6.7 Discrete Fourier Transform Development 112
6.8 Implicit Assumption behind Windowed Fourier Transform 114
6.9 Meaning of N Samples Collected in a Window 115
6.10 Redundancy of DFT 116
6.11 DFT as a Mapping 118
6.12 Numerical Calculation of DFT Coefficients 118
6.13 Sliding DFT Algorithm 125
6.13.1 Modified DFT 128
6.13.2 MATLAB Program for Sliding Modified Recursive
DFT Algorithm 129
Review Questions131
7. FFT and Goertzel Algorithm 133160

7.1 What is the Motivation for FFT? 133
7.2 FFT by Decimation-in-Time (DIT) 136
7.2.1 Numerical Problem on FFT (DIT) 145
7.3 Fast Fourier Transform by Decimation-in-Frequency 145
7.3.1 Numerical Problem on FFT (DIF) 150
7.4 MATLAB Implementation of FFT: Decimation-in-Time 150
viiiContents
7.5 MATLAB Implementation of FFT: Decimation-in-Frequency 153

7.6 Motivation for Goertzel Algorithm 155
7.7 The Goertzel Algorithm 155
7.8 Implementation of Goertzel Algorithm in MATALB 158
Review Questions159
8. Windowing and Spectral Leakage 161171

8.1 The Fourier Machine 161
8.2 Windowing without Spectral Leakage 162
8.3 Windowing Leading to Spectral Leakage 165
8.3.1 Spectral Leakage from Time Domain Viewpoint 165
8.3.2 Spectral Leakage from Frequency Domain Viewpoint 166
8.4 Windowed and Sampled Signal 168
Review Questions171
9. Introduction to Digital Filtering 172195

9.1 Introduction to Filters 172
9.2 What is a Digital Filter? 173
9.3 Why Digital Filtering? 173
9.4 FIR and IIR Filters 175
9.5 Running Average as FIR Filter: 5-Point Running Average Filter 177
9.5.1 Impulse Response of 5-Point Running Average Filter 178
9.5.2 Frequency Response of a 2-Point Running Average FIR Filter 179
9.5.3 Frequency Response of 2-Point Running Average Filter
Using Z-transform182
9.6 Frequency Response of a 3-Point Running Average Filter 183
9.6.1 Frequency Response of 3-Point Running Average Filter
9.7 Frequency Response of 4-Point Running Average Filter
9.8 MATLAB Program for Plotting Frequency Response of N-Point
Running Average Filter 189
9.9 Introduction to Infinite Impulse Response (IIR) Filters 191
Review Questions193
10. Digital Filter Design 196215

10.1 Filter Specifications 196
10.2 Relation between Linear Phase and Symmetry of Filter
Coefficients for FIR Filter 196
10.3 Design of FIR Filter Using Frequency Sampling
Method199
10.3.1 Numerical Problem on Design of FIR Filter Using Frequency
Sampling Method 200
10.4 Bilinear Transformation 204
10.4.1 Design of IIR Filter Using Bilinear Transformation 204
Contentsix
10.4.2 Warping of Frequency 204

10.4.3 Pre-Warping of Frequency During Design 206
10.4.4 Step by Step Design of Iir Filter Using Bilinear Transformation 206
10.4.5 Review of Formulas Derived 207
10.4.6 Numerical Problem on Design of IIR Filter 207
10.5 Design of a Digital Resonator Using Pole Zero Placement 209
10.5.1 Understanding the z Plane 209
10.5.2 Placement of Pole 211
10.5.3 Placement of Zeros 212
10.5.4 Expression for H(z), the Frequency Domain Transfer Function 213
10.5.5 Numerical Problem on Design of Digital Resonator 213
Review Questions215
11. Synchrophasors 216236

11.1 Introduction 216
11.2 The Phasor 217
11.3 The Synchrophasor as per IEEE Std C37.118.1-2011 218
11.3.1 The Synchrophasor in Steady State Conditions 218
11.3.2 The Synchrophasor in Transient Conditions 220
11.3.3 Synchrophasor under Dynamic Conditions 221
11.3.4 Total Vector Error (TVE) 221
11.3.5 Frequency and Rate of Change of Frequency (ROCOF) 222
11.4 Time Tagging (Stamping) of the Synchrophasor 223
11.5 Dissemination of the Time Stamp: IRIG-B Standard 224
11.6 Synchrophasor Reporting Rates 225
11.7 Block Diagram of the Synchrophasor Enabled Digital Relay
or Phasor Measurement Unit (Pmu)225
11.7.1 Two Alternatives for Sampling 225
11.8 Off-nominal Frequency Operation of the PMU 227
11.8.1 A Demonstration of Error Caused by Off-nominal
Frequency Operation 227
11.8.2 Analytical Expression for Effect of Off-nominal
Frequency Operation 228
11.8.3 A MATLAB Script to Demonstrate Off-nominal
Frequency DFT 232
11.9 Synchrophasor Applications 234
Review Questions236
12. Removal of DC Offset 237245

12.1 Why Decaying DC Offsets are Created? 237
12.2 What is the Effect of Decaying DC Offsets up on
Relays?239
12.2.1 Mimic Impedance Method of Removal of DC Offset 239
12.2.2 Immunity of LSQ Method to Decaying Dc Offset 240
12.2.3 Immunity of Differential Equation Algorithm to Dc Offsets 240
12.3 Digital Mimic Impedance Based Filter 240
xContents
12.4 Method of Partial Sums for Filtering DC Offsets 242

12.5 Characterising the Decaying DC Offset by Integration 244
Review Questions245
Appendix 247251
References 253256
Index 257259
Preface
On the one hand the electronic technology is changing at a breath taking speed. The relentless
march of Moors law is making it not only possible to put computing power into every
imaginable existing gadget but is also giving birth to entirely new gadgets and gizmos. Todays
note-book puts yesterdays supercomputer to shame. The internet and the digital computer
have become ubiquitous. Interestingly, advances in the VLSI technology have also made the
digital computer largely invisible by embedding it deeply into the innards of the system that
it is controlling. You have digital computers embedded in every imaginable object.
Thus, the current trend is to replace the mechanical parts with electronic hardware and the
electronic hardware, with software. It is no wonder, therefore, that the field of power system
protection has also been caught up in the throes of the digital revolution. The result is that, the
protective relay has been completely transformed into an intelligent electronic device (IED).
On the other hand, equally dramatic changes have taken place in the general power
system scene. The electric utility business has been deregulated in large parts of the world.
Environmental concerns are giving rise to the proliferation of non-conventional energy sources
like wind, solar, bio-mass etc. There is more variety in the sources of power generation today
than a decade back, leading to challenges and opportunities for the protection engineer.
All these factors make the present times very exciting and fascinating for the practicing
protection engineers. However, for the new comer, to put it mildly, it becomes confusing and
chaotic. This is because an appreciation of three major disciplines is required before one can
make sense of all the digital brouhaha. Firstly, one must never lose sight of the underlying
electrical principles. The various entities that are being protected; the generators, bus-bars,
transmission lines and transformers have their own idiosyncrasies, i.e., unique set of responses
and behaviours that set them apart from every other entity. These must first be thoroughly
understood. Secondly, since the implementation of all the ideas is through the digital signal
processing technology, one needs thorough appreciation of the art and science of DSP at the
level of algorithms. Thirdly, we need a machine; the DSP microprocessor to practically
xi
xiiPreface
implement the algorithms. Thus, appreciation of this physical layer of digital technology is
equally important.
It is not possible to do justice to all the three areas mentioned above in a text like this.
We have mainly attempted to introduce the reader to the middle-ware, i.e., various digital
algorithms which are at the heart of the digital protective relays.
The text is aimed at the final year undergraduate and postgraduate students of Electrical
Engineering who have already undergone courses on Power System Protection, Signals and
Systems and Microprocessors. A course on Numerical Methods will be a definite advantage.
The practicing engineers, who routinely install, operate and maintain these systems, will
also find the text useful in deepening their insight into the ideas that make the digital relays
intelligent.
The author feels indebted to all the researchers, inventors and pioneers who have enriched
the state of the art by their contributions and made this field extremely fascinating.
The author wishes to thank his teacher and guide Prof. Y.G. Paithankar for introducing this
fascinating field to him in the early 1990s. His love for the subject, enthusiasm for learning,
his simplicity and pure humanism has been very inspiring. The author will always cherish the
memories of his association with Dr. Paithankar.
The author would like to gratefully acknowledge the affection and encouragement received
from all his colleagues at the Electrical Engineering Department, VNIT, Nagpur.
Last but not the least, the numerous students who have taken the courses on DSP
Applications to Power Systems and Digital Protection, over the years, deserve special thanks
for being catalysts in enriching authors understanding of the subject and enhancing his patience.
S.R. Bhide
1
Evolution of Power System Protection and
the Emergence of Digital Relaying
1.1 Evolution of OC Relay from an Ammeter

The discipline of power system protection is, now, more than 100 years old. In the early days,
circuit breakers were tripped, by directly making use of the current carried by them through
a pickup coil. Thus, the concept of an entity called relay, separate from the circuit breaker
was yet to be evolved. The mechanism for tripping the circuit breaker was based on force
developed, due to build-up of current to a large value, such as that encountered during fault.
However, it was soon realised that it would help, if the setting of current above which the
circuit breaker tripped is adjustable at will. Thus, what was needed was a trip decision based
on precise measurement of current magnitude. Hence it was natural to evolve the protective
over-current relay from an ammeter as shown in Figure 1.1.
Figure 1.1 Evolution of an ammeter into an over-current relay.
1
2Digital Power System Protection
The technology used for making ammeters, at that time was based on force experienced
by a current carrying coil placed in a magnetic field. We call this the electromechanical
technology which made measurements through the balancing of deflecting torque against a
restoring or restraining torque. The restraining torque could be developed by a spring, by
gravity or by another current etc.
1.2 Evolution of a Directional Relay from a Wattmeter

We can trace the evolution of directional relay from wattmeter. Figure 1.2 shows that a
wattmeter endowed with tripping contacts just beyond zero in the positive direction. If the
wattmeter current coil and pressure coil are fed with relay current and voltage signals, it can
be used as a directional relay with maximum torque angle (MTA) of zero degrees.
Figure 1.2 Evolution of directional relay.
It is easy to see that electromechanical measuring instruments evolved into their new avatar
of dedicated and very sensitive and precision mechanisms called protective relays. Somewhere
along the evolutionary path, the indicating needles were dropped. It is interesting to note that
modern numerical trelays do store the instantaneous values of voltage and current and can
display them on request of the user. Thus, the measuring instrument from which the relay
has evolved is still there. This reinforces the idea that modern relays can be looked up on as
measuring instruments endowed with numerical computational and logical decision making
capabilities.
1.3 The Birth of Transistor

The transistor, invented in 1947 by William Shockley, John Bardeen and Walter Brattain,
(for which all three were awarded the Nobel Prize for Physics in 1956) in the Bell telephone
laboratories, brought phenomenal changes in electronics, signal processing, measurement and
almost all of electrical engineering. Protective relaying field was no exception to this and
within a year of the invention of the transistor, researchers started reporting relays built around
the solid-state transistor. But it took long time for this technology to mature since the early
Evolution of Power System Protection and the Emergence of Digital Relaying3
electronic devices were not as reliable as their present day counterparts. Further, user experience
and confidence in the new technology was lacking. It was only during the early 1960s that
solid-state analog technology was accepted by the industry.
Even though digital computers were around right from the early 1940s, they were very
bulky, in fact monstrous by todays standards, consumed astounding amounts of power and
lacked speed for real time applications demanded by a protective relay. Not only that, they
were so expensive that only big corporations and government defence establishments could own
them. In those days, it was unthinkable to use a digital computer in a relay. This is borne out
by the fact that early research papers on use of digital computer for protective relaying go to
great lengths to justify their use and almost sound apologetic about using digital computers for
such a mundane task as protective relaying!
1.4 Start of a Revolution

The year 1972 saw the birth of the microprocessor which was CPU on a single chip of silicon.
This opened the door to many unthinkable applications of computer to the relaying field. In fact,
it was microprocessor which spawned the current information technology revolution which is
still playing out at an unrestricted pace. One major change that happened with the introduction
of the microprocessor technology, however, was that the element of software was added to
a product which was earlier purely an item of hardware. In fact, it is the software which
makes the modern protective relay so protean. One can now implement any protection concept
that one can think of, provided one can write suitable software to do it. The hardware part
of the digital relay has become much more standardised than the software. It is the software
that distinguishes one vendors protective relay from another. This change has also been quite
slow and for many years even though microprocessors were used in the relays, they were
doing peripheral mundane things while the time critical filtering and measurement functions
were still handled by analog devices. But as the microprocessors evolved their architectures
to cater the signal processing tasks, it became possible to handover the time critical signal
processing tasks to DSP microprocessors bringing in an era of truly numerical relays. In the
modern numerical relay, once the analog signals are sampled, and numbers which represent
the instantaneous values of voltages and currents are generated by the ADC, every subsequent
operation till issuing of trip signal is done through software numerically. Thus even though
the use of microprocessor in relays is reported from the late 1970s, it was only during 1986
that the true numerical relay has emerged.
In addition to the ability of quickly disconnect a faulty element of the power system from
the rest of the system using CT/PT in conjunction with the relays and circuit breakers, there
are many routine things which need to be attended to. For example, consider a 100 MVA
transformer in a substation. We need to do the following functions in addition to protection:
1. Display the status of the circuit breakers on either side of the transformer on a control
panel
2. Metering of energy, active and reactive power, peak demand and power quality
3. Annunciation, logging and flagging
4. Archival of data
5. Keeping count of the number of circuit breaker tripping to facilitate CB maintenance

6. Keep a check on the health of the trip batteries by monitoring their voltage
7. Monitor and display the tap changer position
8. Communication with substation computer
Such other functions which are distinct from the protection per se, are called control and
automation functions. Much of the data used for control and automation is common with that
used for protection. Further, the computational processing requirement of the control functions
is very small as compared to protection and can easily be accommodated in relay hardware and
software. Thus, it was very natural to merge the control and automation functionality with the
protection functionality giving rise to what is known as a numerical terminal dedicated to
the protection,control, automation and monitoring of a particular element of the power system.
Presently vendors are offering such fully numerical terminals for protection,control, automation
and monitoring of transformers, transmission lines bus-bars, motors, etc.
1.5 Architecture of the Modern Digital Relay

Hierarchy of protection can be loosely compared with the seven layers of communication
defined in the OSI model (Open Systems Interconnection) adopted by the International Standards
Organisation as shown in Table 1.1.
Table 1.1 Seven Layers of communication

in the OSI model
Application layer
Presentation layer
Session layer
Transport layer
Network layer
Data link layer
Physical layer
On similar lines, we can conceptualise layers of protective relaying functions where

different kinds of activities take place. At the lowest level is, of course,the interface
between the high voltage, high current power system and the low voltage and low
current measurement system made possible by the current transformers and the potential
transformers. This is followed by layers of electronic hardware to convert the voltages and
currents into numbers representing their instantaneous values. Once we have the numbers
available in the memory, the software takes over and we are in the realm of algorithms.
This conceptual scheme is shown in Table 1.2.
Evolution of Power System Protection and the Emergence of Digital Relaying5
Table 1.2 Seven layers of protection
Algorithm for control, automation and monitoring functions

Algorithm for implementation of relay characteristics or protection scheme
Algorithm for extraction of information from raw samples
Analog to digital conversion
Sampling and holding
Filtering signal conditioning
Ct and pt
In this book, we will mainly focus on the top 3 layers with special emphasis on algorithms
for extraction of information useful for relaying purposes.
2
Digital Signal Processing Basics and
Architecture of Numerical Relay
2.1 Introduction to Digital Signal Processing

We live in an analog world. Temperature, pressure, velocity, speed, mass, volume, voltage,
current; are all analog quantities. Thus, it would be natural to process them in the analog domain.
However, even though analog signal processing was extensively used in the 1950s, 1960s and
early 1970s, it has been superseded by digital signal processing using digital methods. This
has come about because digital computers have proved to be much more powerful, economical
and flexible. The reason for this digital revolution is that the number of transistors that can be
fabricated on a silicon chip is doubling every 18 months in accordance with Moores law. It
is no exaggeration to say that the computing power that was available only to big government
defence establishments some 20 years back is now in the hands of a school boy in the form
of a smart-phone!
The digital computers are controlled by software which is also continuously evolving and
exhibiting exponential growth. All these factors rule out analog computers. Thus, it is inevitable
that digital computers are finding their use in every imaginable application from dish-washer
to DTH receiver. We are now in a situation where signal processing is entirely dominated by
the digital method. Thus, on the one hand, all our activities are in the analog domain and on
the other hand, all the computation, control and monitoring is to be done using the digital
computer as depicted in Figure 2.1.
Figure 2.1 The two worlds.

6
Digital Signal Processing Basics and Architecture of Numerical Relay7
2.2 The DSP Signal Processing Chain

The computer, with which we are going to process the power system signals, only understands
digital signals. Hence, we must first convert the analog signal into digital signal using a sub-
system known as analog to digital converter (ADC). These digital signals are stored as numbers
representing the instantaneous values of various voltages and currents. The digital signal
processing sub-system essentially works on these numbers using a mathematical procedure or
algorithm and produces a digital output. In some applications like processing of audio signals
we again convert these digital signals into analog form using a digital to analog converter
(DAC), say, for feeding to a speaker or a headphone. However in the digital relaying field we
are not interested in getting back the signal in analog form. Hence, the DAC is absent from
the digital relaying hardware. It is shown in Figure 2.2 for the sake of showing the complete
classical DSP signal chain. The digital relay issues a trip/restrain signal, which is essentially
a binary or digital signal. This arrangement is shown in Figure 2.2.
Figure 2.2 The classical DSP signal chain
2.3 Analog to Digital Converters

Thus, it can be seen that the ADC forms a very crucial link between the analog and the
digital worlds. Therefore, the choice of the appropriate type of ADC will decide the overall
performance of the digital relay. ADCs can be implemented using a variety of techniques. Each
technique of A to D conversion has its pros and cons. However, it is possible to talk about the
ADC at a functional level, without going into the details of its working. Figure 2.3(a) shows
a bipolar ADC with N bits in the output. The ADC can accommodate signal between Vm
and +Vm volts and produces an N bit digital code corresponding to the input analog voltage.
Figure 2.3(a) Input and output of a N-bit bipolar ADC.

Figure 2.3(b) Functional diagram of an ADC.
2.3.1 Quantization Error

The ADC consists of a quantizer followed by an encoder (or code converter) as shown in
the block diagram of Figure 2.3(b). The quantizer receives a sample of the analog signal and
converts it into its digital version. The input analog signal can take infinite number of values
while the output digital signal can take only 2N number of values, for an N-bit ADC. This,
process of quantization, which essentially maps an infinite space into a finite space, is shown
in Figure 2.4. Quantization leads to a definite loss of information, leading to an unavoidable
error, known as quantization error.
Figure 2.4 Quantization in a unipolar 3-bit ADC.
It can be seen from Figure 2.4 that for a 3-bit ADC with an input voltage range of 0 to V
volts, all voltages in the range 0 to (V/8) are assigned the same code 000. All voltages in the
range from (V/8) to (2V/8) are assigned the same code 001, and so on. Thus, the resolution
of the ADC is (V/8). As the input voltage changes from 0 to slightly more than (V/8) volts,
the code changes from 000 to 001. Thus, the LSB corresponds to a change of (V/8) volts. The
quantum of (V/8) volts is therefore called 1 LSB, and the resolution of the ADC is said to
be 1 LSB. Alternatively, ADC resolution is expressed simply as number of bits in the output.
Let us indicate the total excursion in input voltage that the ADC can handle by Vspan where
Vspan can be written as:
Vspan = Vin, maxmim Vin, minimum
Magnitude of 1 LSB for an N-bit bipolar ADC with an input span of Vspan is given by:
Vspan (Vm - ( -Vm )) 2Vm
1 LSBbipolar = N
= N
=
2 2 2N
Note that for a bipolar ADC, Vin , minimum = Vm while for a unipolar ADC, Vin, minimum = zero.
Let us assume Vin, maximum for both ADCs as +Vm. Then, the span of a bipolar ADC will
be 2Vm while that of unipolar ADC will be Vm.
If we apply a ramp to the input of the ADC, the output of the ADC will be a quantized
digital output. If we feed this to a DAC, it will generate a staircase type waveform as shown
in Figure 2.5. We define the quantization error of the ADC as:
Quantization error = output of ADC (as seen through DAC) analog input
Note that even an ideal ADC having no other imperfections cannot escape from quantization
error. We can only make the quantization error smaller and smaller by increasing the number
of bits in the ADC, but we can never completely eliminate it.
We can plot the error signal waveform as shown in Figure 2.5. The error in the output
Figure 2.5 Transfer characteristics of a 3-bit ADC and quantization error.

of the ADC can be considered as a noise signal attributable to the quantization phenomenon.
Hence, it is customary to quantify the error due to quantization in terms of signal to noise ratio
defined as ratio of signal voltage to noise voltage usually expressed in decibels.
Figure 2.5 shows that the actual transfer curve of the ADC, always hovers on the lower side
of the ideal transfer curve. Statistically it will be more appropriate if the actual transfer curve
symmetrically encompasses the ideal transfer curve. This can be easily achieved by adding a
LSB offset to all the analog levels at which the code makes transition as shown in Figure
2.6. This results in the error curve also being symmetric about the x-axis.
Figure 2.6 Transfer characteristics of a 3-bit ADC with half LSB offset and quantization error.
We are now in a position to express the quantization error in terms of signal to noise ratio.
Let us assume that a sinusoidal wave with a peak value of Vm is applied to the ADC. The RMS
value of the signal is thus Vm / 2 . Note that the signal excursion is from Vm to +Vm, which
is a change of 2Vm requiring an ADC with a span of 2Vm. Thus, the ADC is representing a
voltage range of 2Vm using N-bits.
The RMS value of the saw-tooth wave, generated because of quantization phenomenon,
which is considered as noise can be written as:
1 T/2 2
Vnoise = ve dt
T T/2
where ve, the error signal, having the saw-tooth waveform, is given by:
Ve = (VHLSB/(T/2))t, where T is the time period of the saw-tooth error waveform and VHLSB
is the magnitude of voltage corresponding to LSB given by:
1 1 2Vm Vm
VHLSB = LSB = = N
2 2 2N 2
As noted earlier, the value of 2Vm in the numerator of above equation is due to the fact that
a sine wave with a peak of |Vm| needs a bipolar ADC which can accommodate input voltage
excursion from Vm to +Vm which is actually an excursion of 2Vm.
Hence, we can write:
2
1 T/2 -VHSLB
Vnoise =
T T/2 (T/2)
t dt
VHSLB 1 T/2 2
Vnoise =
T/2
(t ) dt
T T/2
T/2
V 1 t3 V
Vnoise = HSLB = 2 HLSB
T/2 T 3 -T/2 12
Hence,
Vm Vm
Vnoise, RMS = 2 N
= ( N -1)
2 12 2 12
We can therefore write signal to noise ratio (SNR) as:

Vsignal,RMS
SNR =
Vnoise,RMS
Vm V
We have, Vsignal,RMS = and Vnoise,RMS = ( N -1)m
2 2 12
Hence, SNR is obtained as:
Vm 2( N -1) 12
SNR =
2 Vm
SNR = 2 N 1 6
Let us express it in decibels:
Vsignal,RMS
SNR dB = 20 log10
Vnoise,RMS
SNR dB = 20 log10 (2 N -1 6 )
SNR dB = 20( N - 1) log10 2 + 20 log10 ( 6 )
SNR dB = 20 N log10 2 20 log10 2 + 20 log 10 ( 6 )

SNR dB = 6.02 N - 6.02 + 7.78
SNR dB = {6.02 N + 1.76}dB
However, when sampling is done above the Nyquist sampling rate then noise gets spread
over the frequency range from 0 to (fsampling/2) and a part of the noise gets pushed to frequencies
above fsig, max. This decreases the in band noise, improving the signal to noise ratio to:
( fsampling /2)
SNR dB = 6.02 N + 1.76 + 10 log10 dB
fsignal,max
Noting that fsig,max is the bandwidth of the signal, and representing it by BW, we can write:
( fsampling /2)
SNR dB = 6.02 N + 1.76 + 10 log10 dB
BW
2.3.2 ADC Types

There are several types of ADCs. Each type of ADC has its advantages and disadvantages.
Table 2.1 lists various ADC types with some pertinent remarks.
Table 2.1 Analog to digital converter types

Type of ADC Conversion speed Remarks
1 Flash or parallel comparator Fastest N-bit ADC needs (2 N1) number of
comparators. Hence, becomes impractical
when N is large. Suitable for real time
applications like numerical relays.
2 Successive approximation Intermediate Moderate complexity. Can be built for high
value of N. Used in numerical relays.
3 Dual slope Slowest Highly accurate but slow. Not suitable for
real time applications. Used in bench top
high precision voltmeters.
4 Sigma-delta High Trades sampling rate for number of bits.
Uses noise shaping. Due to high sampling
frequency, anti-aliasing filter design is
simplified. Down sampling can be performed
to reduce sampling rate at the output. Digital
filtering can also be done to enhance signal.
Used in audio applications.
The flash type ADC

A 2-bit flash or parallel-comparator type ADC is shown in Figure 2.7. It can be seen that the
quantizer is based on 3 numbers of comparators. In general, it can be shown that a N bit flash
converter requires (2N 1) number of comparators. Hence, an 8-bit flash ADC would require
28 1 = 256 1 = 255 comparators while a 10-bit flash comparator would require (210 1)
= 1024 1 = 1023 comparators. Such large number of comparators becomes unmanageable
when fabricated on a chip because of imperfections of the analog components. Hence, flash
converters become impractical for higher number of bits.
Figure 2.7 A flash type 2-bit A to D converter.
The reason for calling it a flash converter is that the conversion from analog to digital
value in this topology is very fast since it involves only the switching time of the comparators
which is very small. In fact all other methods of analog to digital conversion are slower than
the flash method. Since all the comparators work in parallel, it is also known as parallel
comparator type ADC.
It can be seen from Figure 2.7 that the comparators have thresholds of V/22, 2V/22, and
3V/22, i.e., V/4, 2V/4 and 3V/4. Further, the output digital code generated by the quantizer is
not a natural binary code. Thus, the ADC needs a code converter or the so-called encoder.
The encoder is a simple combinatorial circuit. The encoder can be easily designed as shown
in Figure 2.7 using K-map method.
2.4Sampling
An ADC cannot handle a continuously changing analog signal. We must first sample the analog
signal and then feed one sampled analog value to the ADC at a time. Figure 2.8(a) shows the
sampling of an analog signal at 16 samples per cycle.
Figure 2.8(a) Analogue signal and its sampled version.
In the above discussion we had assumed an ideal ADC. However, a real-life ADC has
many imperfections. One such important imperfection is that all ADCs require a finite amount
of time to convert the input analog voltage into its digital value.
2.4.1 Need for Sample and Hold Circuit

Let us see what happens when we feed the analog signal directly to the ADC. In order to
avoid error, we need to hold the analog voltage constant during the conversion time. Thus,
we not only need to sample the input but also hold it constant. If the signal changes by more
than +/- LSB during the time it is being converted, it will cause erroneous output. Hence,
we should only feed one unique analog voltage to the ADC by sampling the analog voltage at
a sharply defined instant of time. We would tend to think that if we have a fast enough ADC,
this will solve the problem of finite conversion time of ADC. To illustrate how things get
more complicated, consider Figure 2.8(b) where a unipolar ADC is assumed and three analog
voltage waveforms with different frequencies but same peak value are shown. Using the rule
that signal should not change by more than LSB during conversion we see that for a given
number of ADC bits and peak value of input signal, the frequency of the signal that can be
handled without error goes down as we employ a better ADC. A better ADC is the one with
a smaller conversion time and greater number of bits. We must therefore hold the input to the
ADC constant for a time equal to conversion time of ADC.
Figure 2.8(b) shows three voltages of different frequencies LF, MF and HF having the
same peak value. It can be seen the LF < MF < HF. In a time duration equal to conversion
time Tc, each voltage changes by a different quantum. The quantum being highest for highest
frequency signal since all signals are assumed to have same amplitude. The figure also shows
the voltage corresponding to LSB. For error free operation of ADC, the input voltage should
not change by more than this value of LSB during conversion time Tc. It can be seen that
the condition that input voltage should not change by more than LSB is satisfied only by
the lowest frequency signal LF.
Figure 2.8(b) Need for sample and hold circuit.
Now if we employ a better ADC, i.e., an ADC with higher number of bits N and smaller
conversion time Tc, the maximum frequency that can be handled goes down still further than
LF. This can be seen from Figure 2.8(c).
Thus, we cannot improve upon the maximum frequency that can be handled by using ADCs
with faster conversion time and more number of bits. This can be easily appreciated by finding
out the maximum frequency that can be handled by an ADC with excellent specifications.
Consider a 16-bit ADC with a conversion time of 10ms. Let us find the maximum frequency
that can be handled by an ADC with so good specifications.
Figure 2.8(c) Maximum frequency goes down for a better ADC.
Let the signal be represented by

v = Vm sin(w t )
v = Vm sin(2p ft )
dv
= 2p f Vm cos w t
dt
dv
= 2p f maxVm
dt signal,max
(dv/dt)ADC,max i.e., the maximum rate of change of voltage that the ADC can accommodate is
given by (dv/dt)ADC,max = (Dv/Dt)ADC,maxwhere and Dv = (Vm/2N) and Dt = Tc
dv V
= Nm
dt ADC,max 2 Tc
Equating (dv/dt)signal,max and (dv/dt)ADC,max we get:
Vm
2p f maxVm =
2 N Tc
Hence,
1
f max = ( N +1)
2 p Tc
For the ADC with given specifications of N = 16 and Tc = 10ms, we get

1
f max = (16 +1)
2 p (10)(10 -6 )
f max = 0.2428 Hz
This is too low a frequency to be of any use in power system protection work! And the
frequency will become still lower if we use a better ADC! Thus we have to think of a device
in between the analog input signal and the ADC which will be able to cope up with the
high (dv/dt) rates involved. Fortunately, a simple solution to this problem exists. If we use a
sampler followed by a hold circuit, i.e., a sample and hold circuit (S/H) then we can go
for sampling much higher frequencies. This becomes possible because now it is the aperture
time of the hold circuit, Ta, that has to meet the (dv/dt) requirement rather than the ADC.
Fortunately, the aperture times of S/H circuits are many orders of magnitude smaller than the
ADC conversion times. Thus, fmax with sample and hold can be obtained by substituting Ta in
place of Tc in the expression for fmax derived earlier.
1
f max = ( N +1)
without hold circuit
2 p Tc
1
f max = ( N +1)
with hold circuit with aperture time Ta
2 p Ta
Let us rework the maximum frequency by assuming a realistic aperture time of 250 ps.
1
f max = (16 +1)
with hold circuit
2 p (250)(10 -12 )
f max = 9.714 kHz
Thus, we get improvement in maximum frequency from 0.2428 Hz to 9.714 kHz because
of simple expedient of using a sample and hold circuit ahead of the ADC. We make suitable
modification in the DSP signal chain as shown in Figure 2.9(a). Figure 2.9(b) shows an analog
signal , the sampling impulses, the sampled signal and sampled-and-held versions of an analog
signal. Note that and fsignal = (1/Tsignal) and fsampling = (1/Tsampling)
Figure 2.9(a) Modified DSP signal chain with sample and hold circuit included.
Figure 2.9(b) Sampled analog signal.
2.4.2 Shannons Sampling Theorem and Aliasing

Note that we have to sample the analog signal before feeding it to the ADC. Thus, now, there
are two frequencies in picture; the signal frequency and the sampling frequency. In real life,
we may never encounter a signal with a very well defined single frequency. Real life signals
tend to contain a band of frequencies from the lowest frequency of zero Hz, i.e., dc to some
maximum frequency fsignal,max. Hence, we talk about the signal bandwidth denoted by BW.
Thus, the question that naturally arises is: What should be the relationship between the signal
bandwidth (or maximum frequency contained in the signal) and the sampling frequency? Here,
the famous theorem known as Shannons sampling theorem comes into play. According to the
sampling theorem the minimum sampling frequency must be at least twice the maximum signal
frequency, so that there is no loss of information. The sampling theorem can be stated as:
fsampling,minimum = 2 fsignal if signal is a undistorted single frequency sinusoid
fsampling,minimum = 2 fsignal,max if the signal happens to be broad-band

fsampling,minimum = 2 BW where BW is the bandwidth = fsignal, maximum
The minimum sampling frequency stipulated by Shannons sampling theorem is known as

the Nyquist sampling rate.
We can therefore in general write that:
fsamping 2 BW
The sampling theorem is pictorially shown in Figure 2.10. It can be seen from Figure 2.10
that as dictated by the sampling theorem, there is a band of frequencies from zero to 2fsignal,max
which are forbidden for the sampling frequency.
Figure 2.10 Shannons sampling theorem.
This naturally leads to the next question: What happens if we violate the sampling theorem
and sample at a forbidden frequency, i.e., at a frequency less than that stipulated by Shannon?
We encounter the phenomenon of aliasing in such an eventuality, wherein a higher
frequency signal appears as a lower frequency signal. When aliasing takes place, we cannot
recover correct information from the signal. Thus, aliasing is to be avoided at all costs.
Figure 2.11(a) shows the effect of violating the sampling theorem. It can be seen that
the sampling is at rate less than twice the signal frequency since there are no samples during
alternate halves of the input sine wave. The wave that will be recreated from the samples will
be a low frequency aliased wave different from the input.
Figure 2.11(a) Aliasing.
Figure 2.11(b) shows a more specific example of a 50 Hz signal sampled at frequencies of

49 Hz ,51 Hz, 100 Hz and 500 Hz. This figure was generated using a small MATLAB script.
As expected, there is aliasing at sampling frequencies of 49 Hz, 51 Hz and even at 100 Hz.
Sampling at 49 Hz and 51 Hz give an aliased signal of 1 Hz with phase angles of zero and
180 respectively. At 100 Hz, since the sampling is synchronised with the zero crossing of
the waveform, we get a DC with zero magnitude as the aliased wave. However, at sampling
frequency of 500 Hz we are able to correctly recreate a 50 Hz wave, since there are 50 complete
oscillations in 1 second.
Fig. 2.11(b) A 50 Hz wave sampled at various frequencies.
In DSP based systems, sampling plays very important role as what we perceive
through the sampling process strongly depends upon the relation between the signal
bandwidth and the sampling frequency. Figure 2.11(c) shows perceived frequency Vs
the actual input frequency for a fixed sampling frequency. It can be seen that only in the
region where Shannons sampling theorem is not violated, do we correctly perceive the
input frequency. Outside this range we actually either have no perception or are working
in the illusory realm because of aliasing.
Fig. 2.11(c) Perceived frequency Vs input frequency for a fixed sampling frequency.
2.5 Anti-aliasing Filter

It would appear that in order to avoid aliasing, all that we need to do is to choose a
sampling frequency which is greater than twice the maximum signal frequency. However
in practise, it is found that a real life power system signal contains a number of unwanted
high frequency components right into the tens or hundreds of megahertz range. Thus, we
need to remove these high frequency components or at least attenuate them below LSB
magnitude before we sample the analog signal. In the absence of such a filter, the high
frequency components will get aliased back below fsampling. Hence, the low-pass filter
used for removing all frequency components above fsampling is called anti-aliasing filter.
The anti-aliasing filter has to be of analog type. Thus, the DSP signal chain gets further
modified as shown in Figure 2.12(a).
Figure 2.12(a) Addition of anti-aliasing filter to the DSP signal chain.

Figure 2.12(b) Characteristics of ideal low-pass filter characteristics.
Figure 2.12(b) shows the characteristics of an ideal anti-aliasing filter that will entirely
filter out all frequency components above fsampling/2 . It can be seen that such an ideal anti-
aliasing filter will have to have its cut-off frequency at fsampling/2 and exhibit an infinite roll-off.
However, all practical filters have finite roll-off. Hence in case of a practical anti-aliasing filter
we will have to push-back the cut-off frequency so that the magnitude of the signal falls below
LSB in the stop band as shown in Figure 2.12(c).
Figure 2.12(c) Practical low-pass anti-aliasing filter characteristics.
2.5.1 Design of Anti-aliasing Filter

Let us design a Butterworth type low-pass anti-aliasing filter, by noting that the gain of the nth
order low-pass Butterworth filter at any frequency, factual, is given by:
1
A( factual ) =
2n
f
1 + actual
fcut-off
so that when factual = fcutoff, the gain of the filter is = 1/ 2 = 0.707

For an N-bit, bipolar ADC, the magnitude corresponding to LSB is (1/2)(2Vm/2N)
To satisfy the requirement that signal magnitude should fall below LSB beyond fsampling,
the following inequality applies:
Vm 1 2Vm

fsampling
2n 2 2N
2
1+
fcut-off
If we represent the ratio [(fsampling/2)/fcut-off] by a, then we can write the above equation as:
Vm 1 2Vm

1 + (a ) 2n 2 2N
Solving the above equation gives order of the low-pass filter as:
loge (4 N - 1)
n
2 loge (a )
2.5.2 Numerical Example

Let us explain the above concept with a numerical example.
Example 2.1 Decide the order of a low-pass Butterworth type anti-aliasing filter to be
used in a DSP system with the following specifications:
(i) Number of bits in the ADC = 12
(ii) Sampling frequency = 3000 Hz
(iii) Cut-off frequency of the anti-aliasing filter = 150 Hz
Solution
Given: N = 12
fsampling 3000
2 1500
a= = 2 = = 10
fcut-off 150 150
loge (412 - 1)
n (3.61 rounded to 4)
2 loge (10)
Hence, 4th order low-pass anti-aliasing Butterworth filter should be used for the given system.
2.6 Functional Block Diagram of Numerical Relay

Typical number of input signals to a numerical relay in a three-phase power system is 6
viz., the three line to neutral voltages Van, Vbn, Vcn and the three line currents Ia, Ib, Ic.
Ideally therefore we should use six number of analog to digital converters to acquire all
the data in parallel. However for reasons of economy we may sample all the six channels
in parallel and hold the values before they are converted to digital. We can use a 6-into-1
analog multiplexer which is digitally addressable to achieve this. Further we need extensive
signal conditioning before the signals from CTs and PTs reach the sensitive electronic
circuits. Protective devices to bypass any high voltage surges to ground are required.
Similarly, clamping circuits, to limit signal excursions to safe values, so that they do not
exceed the input range of analog to digital converter are required as shown in Figure 2.13.
Numerical relays are continuously evolving further. Since the late 1980s, numerical
relays are being endowed with synchrophasor measurement capabilities. This has become
possible because of widespread and cheap availability of GPS (Geographical Positioning
Satellite) technology. A numerical relay can now act as phasor measurement unit (PMU).
Through a high speed optical fibre network data from various PMUs can be collected
and collated by a phasor data concentrator in a hierarchical manner to implement wide
area protection and control schemes. Many of the routine monitoring and control
functionalities in a substation can be easily integrated into the numerical protective relay.
Thus, the protective relay is morphing in to an intelligent electronic device (IED). Some
vendors are calling such devices as protection, control and monitoring systems rather
than mere relays. Figure 2.14 illustrates a modern numerical relay with a variety of
connectivities, inputs and outputs.
We conclude, by observing, that regardless of the apparent complexity of the modern
numerical relay, once the signals are acquired and converted to digital domain, we are left
with numbers which represent the instantaneous values of the signals. Beyond this point,
the protective relaying software takes over. These numbers will now be processed as per
the relaying algorithm which finally may come up with either a trip output or a no-trip
(restrain) output. Note that this is a continuous and un-ending process. A relaying algorithm
therefore works in an infinite loop which is typical of all embedded system software. In
the sections that follow, we will go into the details of various relaying algorithms.
Figure 2.13 Block diagram of a numerical relay.

Figure 2.14 Modern numerical relay.

Review Questions
1.
What do you mean by resolution of an ADC?
2.
What do you mean by 1 LSB, LSB w.r.t. ADCs?
3.
In a 10-bit ADC with input range of 10 V to 10 V, what is the value of 1 LSB and
LSB?
4.
In an ADC with input range of +/5 V, a resolution (1 LSB) of 10 mV is required.
What is the minimum number of bits it must have?
5.
Find the signal to noise ratio (in dB) due to quantization in an DSP system which has
ADC of 16-bits and sampling is done at the Nyquist rate. What will be the error if
sampling is done at 10 times the Nyquist rate?
6.
Find the SNR (in dB) due to quantization in a DSP system with the following
specifications:
Number of bits in the ADC = 12, signal bandwidth = 1000 Hz and sampling frequency
= 5000 Hz.
7.
How does one decide the minimum sampling frequency in a DSP system?
8.
A power system signal is given by:
v(t ) = 100 sin(2p 50t + 45) + 50 sin(2p 100t + 25) + sin(2p 150t + 15)
Suggest minimum sampling frequency, in hertz to avoid aliasing errors.
9. What, if any, are the disadvantages of sampling at very high frequencies , much above
the Nyquist sampling rate?
10.
Give at least two examples where aliasing is encountered in daily life.
(Hint: Movies, Table fan)
11.
The frequency spectra of signal is as shown in Figure Q.11. If the signal is ideally
sampled at intervals of 1 ms, then state with justification whether this will lead to
aliasing. Sketch the frequency spectra of the sampled signal.
Figure Q.11
12.
Find the maximum frequency that can be sampled without using hold circuit for a DSP
system with the following specifications:
Conversion time of ADC = 5 s and number of bits in the ADC = 16
13.
Find the order of low-pass anti-aliasing filter of Butterworth type for a DSP system
with the following specifications:-
ADC resolution = 12-bits , sampling frequency = 6000 Hz and cut-off frequency = 300 Hz
14.
Provide a frequency domain explanation of the aliasing phenomenon.
3
Algorithms Based on Undistorted
Single Frequency Sine Wave
3.1 Mann and Morrison Algorithm [26,27]

In this section, we will discuss the algorithms for digital protection of power systems. One may
feel awe and wonder at the variety of approaches that the researchers have taken. It seems that
the search for algorithmic solutions to problems in the area of power system protection is an
un-ending one. Inventions and discoveries made in other fields like communication, statistics,
artificial neural networks, fuzzy logic, genetic algorithms and expert systems have fuelled
research in this area.
3.1.1 Historical Perspective

Even though this algorithm was not the first among algorithms for digital protection, it was the
first to get wide publicity. The algorithm proposed by M. Ramamoorthy, which was based on
Fourier analysis, in the CIGRE Journal can be considered as the first algorithm to be proposed
for digital protection of power system.
It should be kept in mind that in the early 70s when these algorithms were first proposed,
the microprocessor had not burst on the scene and the digital revolution which was to be
fuelled by the availability of cheap microprocessors was just round the corner. Computers were
a very costly affair. The confidence levels of practising engineers about the reliability of the
new technology were very low to put it mildly. In those days researchers had to justify their
idea of using computer for protection purposes by citing additional and possibly more important
duties that the computer could do. However, as is well-known, the digital revolution in the late
70s totally changed the rules of the game. The microprocessor just became a commodity and
another component of a bigger computing system. As hardware prices plummeted, it was not
long before engineers realised that software has become a key part of the whole and as the
clich goes, rest is history.
29
3.1.2 Derivation of the Sample and Derivative (Mann & Morrison) Algorithm
In this algorithm, we model the signal in a very simplistic manner. It is assumed that the signal
can be represented as a pure sinusoid whose amplitude as well as frequency is constant during
the period under consideration as shown in Figure 3.1.
Figure 3.1 Mann and morrison algorithm.
Thus, voltage samples collected by the relay are assumed to belong to a signals of the
type:
v(tk ) = Vm sin(w tk + q v ) (3.1)
where we assume Vm, w and qv to be constant. Further we assume that the frequency w is
known. Thus in the above equations we have two unknowns, namely the peak value of the
signal and its phase angle. Hence, we need one more equation to solve for the unknowns. Let
us generate another equation by taking derivative of the first equation giving:
v (tk ) = wVm cos(w tk + q v ) (3.2)
from which we get:
v (tk )
= Vm cos(w tk + q v ) (3.3)
w
Combining Eqs. (3.1) and (3.3), we get:
2
v (tk ) (3.4)
Vm2 = (v(tk ))2 +
w
Hence, we can find the amplitude of voltage signal as:
v (tk )
2
Vm = (v(tk ))2 + (3.5)

w
Similarly, we can write for the current signal as:
i(tk ) = I m sin(w tk + q i ) (3.6)
Algorithms Based on Undistorted Single Frequency Sine Wave31
Differentiating Eq. (3.6) at instant tk we get:

i (tk ) = w I m cos(w tk + q i ) (3.7)
Combining Eqs. (3.6) and (3.7) we get:
2
i (t ) (3.8)
I m2 = (i(tk ))2 + k
w
Hence, we can find the amplitude of current signal as:
i (tk )
2
I m = (i(tk )) +
2

w
Dividing Eq. (3.1) by Eq. (3.3) we get:
v(tk ) (3.9)
tan(w tk + q v ) =
v (tk ) /w
v(tk )
w tk + q v = tan -1 (3.10)
v (tk )/w
v (t k )
q v = tan -1 ' - w tk (3.11)
v ( t k )
w
Similarly, we can get:
i(tk )
q i = tan -1 - w tk (3.12)
i (tk )

w
The derivatives can be computed numerically using the central difference formula as:
(vk +1 - vk -1 )
vk = (3.13)
2 Dt
and (ik +1 - ik -1 ) (3.14)
ik =
2t
Thus, computation of one derivative requires three samples. Now, we cannot base
the trip decision on a single estimate (which involves only three samples), we have to
continuously keep estimating the peak and the phase angle and hence the phasor, in order to
be sure about what way the phasor is really going. We should, thus, base our trip decision
on a sufficiently large number of estimates rather than a single estimate. This will make
the decision more robust and provide some measure of noise immunity as discussed in the
following section.
Let us summarise the results of the Mann and Morrison algorithm as shown in Table 3.1.
Table 3.1 Mann and Morrison algorithm: Summary of results
Model of voltage signal v(tk ) = Vm sin(w tk + q v )
Model of current i(tk) = Im sin(w tk + qi)
v (tk )
2
( v (t k )) 2
+ w
Vm =
(vk +1 - vk -1 )
where vk =
2 Dt
v(tk )
qv = tan -1 - w tk radians
v (tk ) / w
i (tk )
2
(i(tk )) +
2

w
Im =
(ik +1 - ik -1 )
where ik =
2 Dt
i(tk )
qi = tan -1 - w tk radians
i (tk ) / w
3.1.3 Instantaneous OC Relay Based on Mann and Morrison Algorithm

We can implement a variety of relays based on the Mann and Morrison algorithm viz., an
over-current relay, directional over current relay or a distance relay. Every relay requires first to
compute from the instantaneous values of the input quantities, the phasor for the input quantity.
Since the Mann and Morrison algorithm estimates the phase angle and the peak value, the RMS
value being given by Peak Value/ 2, we are essentially estimating the phasor. Note that this
is possible only because of the underlying assumption of pure sine waveform.
Let us consider the example of implementing an instantaneous over-current relay based on
this algorithm. Obviously we will have to wait for three samples to make one estimate of the
phasor. Afterwards, however, on the occurrence of every new incoming sample we can modify
our estimate by sifting-in the newest sample and shifting-out the oldest sample.
Thus we have a calculation window of three samples, which keeps sliding over the signal.
In practise we will not issue a trip signal as soon as the RMS value is more than the pickup
value because, the increase in RMS value could, in all probability, be due to a noise spike.
In order to avoid, maloperation, we set up a fault counter and increment it every time
the RMS value is more than the threshold and decrement it if the RMS value is less than the
threshold. In case the counter crosses a predetermined threshold, we can be sure that there
indeed is a fault on the system and issue trip command to the circuit breaker.
The threshold value to be used for the counter will be based on severity of noise at a
particular location. The general logic of such an implementation is shown in Figure 3.2.
Figure 3.2 Instantaneous OC relay based on Mann and Morrison algorithm.
3.1.4 Simulation of the Mann and Morrison Algorithm in Spreadsheet

Spreadsheet software provides a very convenient environment for simulation of numerical
relaying algorithms. We can quickly visualise and plot the relevant information. Spreadsheets
are famous for their capability of performing what if analysis.
For the sake of demonstrating the working of the algorithm, we take a single phase system
as shown in Figure 3.3. Spreadsheet shown in Figure 3.4 shows the samples of voltage and
Figure 3.3 Simulation of SIR based on Mann and Morrison algorithm.

34

Digital Power System Protection

reactance as seen from the relay location.

Figure 3.4 Spread sheet for simulation of Mann and Morrison algorithm.

samples of current, slides through these samples and computes the apparent resistance and
current which are generated for a cycle. A window of three samples of voltage and three
3.1.5 Simulation of Mann and Morrison Algorithm in MATLAB

The system shown in Figure 3.3 is simulated with the help of the program written in MATLAB
is shown below:
% Mann and Morrison Algorithm
clear;clc
%=================== Settings =================
% System frequency and Sampling freuency
f=50; Over_samp_factor =100; % OSF = (fsamp/2fsig)
% Voltage Amplitude and Phase angle

Vm = 110; Phi_v=30;% deg
% Line R-L Parameters

R_perKm =.01; XbyR =10;Line_Length = 100 ; %km
%==============================================
% =============== End Of settings=============
R = R_perKm * Line_Length ;
X = R * XbyR ;
w = 2 * pi * f;
L=X/w;
Z = sqrt(R^2 + X^2);
Im = Vm / Z ;
Phi_i = (atan2(XbyR,1))*(180/pi);
% v = Vm sin(2 * pi * f * t + Phi_v)
% i = Im sin(2 * pi * f * t + Phi_v - Phi_i)
%=================================================
R ; L ; X ; Z ; %<== For error calculation
Vm; Phi_v ; %<== For error calculation
Im; Phi_i; %<== For error calculation
%=================================================
fsamp = 2* f * Over_samp_factor;
dT = 1/ fsamp ;
T =1/f;
k=0;
for t = 0 :dT:T
k=k+1;
v(k)= Vm * sin(2 * pi * f * t + (Phi_v)*(pi/180) );
i(k)= Im * sin(2 * pi * f * t + (Phi_v - Phi_i)*(pi/180));
end
%plot(v):holdon
%plot(i)
no_of_samples=numel(v);
% for n=1:no_of_samples
% fprintf( v = %f i =%f \n ,v(n),i(n))

% end
for k = 2 : no_of_samples-1
v_dash(k)=(v(k+1)-v(k-1))/(2*dT);
i_dash(k)=(i(k+1)-i(k-1))/(2*dT);
Vm_est(k)=sqrt(v(k)^2 + (v_dash(k)/w)^2);
Im_est(k)=sqrt(i(k)^2 + (i_dash(k)/w)^2);
Phi_V_est_rad(k) = (atan2(((v(k))/((v_dash(k))/w)),1))-(w*k*dT);
Phi_V_est_deg(k)= (Phi_V_est_rad(k))*(180/pi);
Phi_I_est_rad(k) = (atan2(((i(k))/((i_dash(k))/w)),1))-(w*k*dT);
Phi_I_est_deg(k)= (Phi_I_est_rad(k))*(180/pi);
Zmag_est(k) = Vm_est /Im_est;

Zang_est_rad(k)= Phi_V_est_rad(k)- Phi_I_est_rad(k);
Zang_est_deg(k)= Phi_V_est_deg(k)- Phi_I_est_deg(k);
R_est(k) = Zmag_est(k)*cos(Zang_est_rad(k));
X_est(k) = Zmag_est(k)*sin(Zang_est_rad(k));
L_est(k) = X_est(k)/w;
end
fprintf(==============================================================
=======================================================\n)
fprintf(Samp v i v_dash Vm_est i_dash Im_est Phi_V_est
Phi_I_est R_est L_est\n)
fprintf(True Value = %5.2f %5.4f %5.2f %5.4f
%5.2f% 5.4f\n,Vm,Im,Phi_v,Phi_v-Phi_i,R,L )
fprintf(==============================================================
=======================================================\n)
for n = 1:no_of_samples-1
if(n==1)
fprintf(%2i | %f %f\n,n,v(n),i(n))
continue
end
fprintf(%2i | %f %f| %f| %f| %f| %f| %f| %f| %f| %f\n,n,v(n),i(n),v_
dash(n),Vm_est(n),i_dash(n),Im_est(n),Phi_V_est_deg(n),Phi_I_est_
deg(n),R_est(k),L_est(k) )
end
% Vm_est
% Im_est
%
% Phi_V_est_rad
% Phi_V_est_deg
% Phi_I_est_deg
Figure 3.5 Output of MATLAB simulation.

3.2 Three-Sample Technique [20,65]

Three-sample algorithm, kind of, extends Mann and Morrison algorithm. It removes the
condition in Mann and Morrison algorithm that frequency should be known. Recall that a single
frequency sinusoid given by v = Vm sin(wt + q) or v = Vm cos(wt + q) has three unknowns
associated with it viz., amplitude Vm, frequency w and phase angle q. This algorithm enables
us to estimate all the three unknowns using only three samples. It is possible, because, to solve
for three unknowns we need only three equations. Three samples that we gather from the signal
are the known quantities on the LHS of the three equations in three unknowns. The signal and
its sampling are shown in Figure 3.6.
Figure 3.6 Sampled signal for three-sample algorithm.
The model of the signal assumed for this algorithm is:

v = Vm sin(w t + q ) (3.15)
v = Vm sin(2p f t + q ) (3.16)
Let the three signal samples, starting from an arbitrary index k be:
vk = A sin(2p f (k Dt ) + q ) (3.17)
vk +1 = A sin(2p f (k Dt + Dt ) + q ) (3.18)
vk + 2 = A sin(2p f (k Dt + 2 Dt ) + q ) (3.19)
3.2.1 Determination of Frequency of the Signal

Now, consider vk + vk+2
vk + vk + 2 = A[sin(2p f (k Dt ) + q ) + sin(2p f (k Dt + 2 Dt ) + q )] (3.20)
If we consider,
X = 2p f ( k Dt + Dt ) + q and Y = 2p f Dt
then
X + Y = 2 p f (k Dt + Dt ) + q + 2 p f Dt = 2 p f (k Dt + 2 Dt ) + q (3.21)
X - Y = 2 p f (k Dt + Dt ) - 2 p f Dt = 2 p f (k Dt ) + q (3.22)
Hence vk + vk+2 can be written as:
vk + vk + 2 = A [sin( X - Y ) + sin( X + Y )] (3.23)
We can use the following trigonometric identity:
[sin( X - Y ) + sin( X + Y )] = 2 sin( X ) cos(Y ) (3.24)
and we can write:
vk + vk + 2 = A [2 sin( X ) cos(Y )] (3.25)
vk + vk + 2 = A [2 sin(2 p f (k Dt + Dt ) + q ) cos(2 p f Dt )] (3.26)
However,
A sin(2 p f (k Dt + Dt ) + q ) = vk +1 (3.27)
Hence, we can write
vk + vk + 2 = [2 vk +1 cos(2 p f Dt )] (3.28)
vk + vk + 2
cos(2 p f Dt ) = (3.29)
2 vk +1
vk + vk + 2 (3.30)
2 p f Dt = cos -1
2 vk +1
1 v + vk + 2
f= cos -1 k (3.31)
2 p Dt 2 vk +1
Thus, since Dt, the sampling time period which is 1/fsampling is known, the unknown
frequency can be estimated with the help of three consecutive samples.
3.2.2 Determination of Amplitude of the Signal

Now, in order to estimate the peak value A of the signal, consider (vk+1)2 (vk)(vk+2)
or, ( A sin(2p f (k Dt + Dt ) + q ))2 - ( A sin(2p f (k Dt ) + q ))( A sin(2p f (k Dt + 2 Dt ) + q )) (3.32)
( A2 )((sin(2p f (k Dt + Dt ) + q ))2 - (sin(2p f (k Dt ) + q ))(sin(2p f (k Dt + 2 Dt ) + q )) (3.33)
1 - cos 2q
Using the trigonometric identity (sin q )2 =
we can write 2
1 - cos(2(2p f (k Dt + Dt ) + q )) (3.34)
(sin(2p f (k Dt + Dt ) + q )2 =
2
further, if we write
(2p f (k Dt + 2 Dt ) + q ) = X (3.35)
and (2p f (k Dt ) + q ) = Y (3.36)

then we can write
(sin(2p f (k Dt ) + q ))(sin(2p f (k Dt + 2 Dt ) + q )) = sin X sin Y (3.37)
Using trigonometric identity
cos( X - Y ) - cos( X + Y )
sin X sin Y = (3.38)
2
X - Y = (2p f (k Dt + 2 Dt ) + q - (2p f (k Dt ) - q )) = 2p f (2 Dt ) (3.39)
X + Y = (2p f (k Dt + 2 Dt ) + q + (2p f (k Dt ) + q )) = 2p f (2k Dt + 2 Dt ) + 2q (3.40)
Hence, we can write
cos(2p f (2 Dt )) - cos(2p f (2k Dt + 2 Dt ) + 2q )
(sin(2p f (k Dt ) + q ))(sin(2p f (k Dt + 2 Dt ) + q )) =
2
(3.41)
Hence, we can now write
1 - cos(2(2p f (k Dt + Dt ) + q ))
(vk +1 )2 - (vk ) (vk + 2 ) = A2
2
cos(2p f (2 Dt )) - cos(2p f (2 k Dt + 2 Dt ) + 2q ) (3.42)

-
2
or, (vk +1 )2 - (vk ) (vk + 2 )
1 - cos(2(2p f (k Dt + Dt ) + q )) - cos(2p f (2 Dt )) + cos(2(2p f (k Dt + Dt ) + q )) (3.43)

= A2
2
= A2 1 - cos 2p f (2 Dt ) (3.44)
2
Using the trigonometric identity cos (2 X ) = 2(cos ( X )) - 1
2
1 - (2(cos(2p f ( Dt )))2 - 1) (3.45)

= A2
2
2 - (2(cos(2p f ( Dt )))2 )
= A2 (3.46)
2
(1 - (cos(2p f ( Dt ))2 )
= A2 2 (3.47)
2
(vk +1 )2 - (vk ) (vk + 2 ) = A2 (sin(2p f ( Dt )))2 (3.48)
(vk +1 )2 - (vk ) (vk + 2 ) = A sin(2 p f Dt ) (3.49)
(vk +1 )2 - (vk ) (vk +2 )

A= (3.50)
sin(2 p f Dt )
Since frequency f has already been evaluated, the amplitude A can be estimated.
3.2.3 Determination of Phase of the Signal

To estimate the phase angle we can use the equation for the first sample as:
vk = A sin(2p f (k Dt ) + q ) (3.51)
which can be manipulated as:
vk
= sin(2p f (k Dt ) + q ) (3.52)
A
v
2p f (k D t ) + q = sin -1 k (3.53)
A
v
q = sin -1 k - 2p f (k Dt ) (3.54)
A
In the above expression vk, A, k and Dt all are known. Hence, phase angle q can be estimated.
3.2.4 Summary of Three Sample Algorithm

Measurement of amplitude, phase and frequency from three-sample is summarised in Table 3.2.
Table 3.2 Summary of three-sample algorithm
Signal v = Vm sin(w t + q )
vk = A sin(2p f (k Dt ) + q )
Samples vk +1 = A sin(2p f (k Dt + Dt ) + q )
vk + 2 = A sin(2p f (k Dt + 2 Dt ) + q )
1 v + vk + 2
Frequency f f (k ) = cos -1 k
2 p Dt 2 vk +1
(vk +1 )2 - (vk ) (vk + 2 )

Amplitude A A(k ) =
sin(2 p f Dt )
v
Phase q q (k ) = sin -1 k - 2p f (k Dt )
A
42

Digital Power System Protection

Figure 3.7 Spreadsheet for simulation of Three-sample algorithm.

3.2.5 Excel Spreadsheet for Simulation of Three-Sample Algorithm
3.2.6 MATLAB Program for Simulating Three-Sample Algorithm

% --------------------------------------------------------------------
-------
%3 Samples Algorithm for measurement of amlitude, phase and frequency
% Signal is modelled as:-
% x = A cos( wt +theta) ; dT=1/fsamp ;
% cos(w dT)=[x(0)+x(2)] / [ (2 x(1))] ;
% w = acos(x(0)+x(2)/ (2 x(1)))/w dT ;
% A =sqrt[(x(1)^2 -x(0)x(2))]/sin(wdT) ;
% theta = acos[ x(0)/A] ;
%------------------------------------------------------------
clear,clc
%===== Settings =================
A = 100+round(rand*10) ; fsig = 50+(0.5-rand) ; theta= 30+round(rand*10);
Over_sampling_factor = 10;
%=================================
w =2*pi*fsig;
fsamp =2*fsig*Over_sampling_factor;
dT = 1/ fsamp ; T=1/fsig;
k=0;
% Let us generate samples of signal
for t=0:dT:T
k=k+1;
x(k)=A * cos((2*pi*fsig*t) + (theta)*(pi/180));
end
plot(x)
for n=2:(numel(x))-1
w_est(n)=(acos( (x(n-1)+x(n+1) )/(2*x(n))))/(dT);
f_est(n)=w_est(n)/(2*pi);
A_est(n)=(sqrt(x(n)^2-x(n-1)*x(n+1)))/(sin(w_est(n)*dT));
t = (n-1)*dT;
theta_est(n)=acos((x(n))/(A_est(n)))- ((w_est(n))*t);
% Why (n-1)?
% Because when n=1; t=0;
theta_est_deg(n)=(theta_est(n)*(180/pi));
err_A(n)=((A_est(n)-A)/A)*(100);
err_w(n) =(( w_est(n)-w )/(w))*(100);
err_theta(n)=((theta_est_deg(n)-theta)/(theta))*(100);
end
% myQ=acos(x(1)/A_est(2))
% myQQ=myQ*(180/pi)
fprintf(=====================3 Sample Technique
===========================\n)
fprintf(The true values are:- \n)
fprintf( A=%f fsig=%f theta=%f \n,A,fsig,theta )
fprintf(======================================================\n)
fprintf(======The Estimated values and Errors are =================\n)

fprintf(n Omega freq Ampltd PhAng Err_A%% Err_w%% Err_Ph_Ang%% \n)
fprintf(=====================================================\n)
for n=2:numel(x)-1
fprintf(%d %8.3f %8.3f %8.3f %8.3f %8.3f %8.3f %8.3f\n,n,w_est(n),f_
est(n),A_est(n), theta_est_deg(n),err_A(n),err_w(n),err_theta(n) )
end
fprintf(\n\n)
for p=1:3
fprintf(Problem No. %d \n,p)
fprintf(The samples of a single frequency signal are as shown below.\n)
fprintf(Make at least three estimates of the amplitude,phase and frequency\n)
fprintf( 0 1 2 3 4 5 6 \n)
for n=p+1:p+7
fprintf( %f ,x(n) )
end
fprintf(\n)
fprintf(===================================================== \n)
fprintf(\n\n)
end
===============================================
The true values are:-
A=104.000000 fsig=50.195001 theta=39.000000
===============================================
======The estimated values and errors are ======================================================
n Omega freq Ampltd PhAng Err_A% Err_w% Err_Ph_Ang%
=======================================================================================
2 315.384 50.195 104.000 39.000 0.000 0.000 0.000
3 315.384 50.195 104.000 39.000 0.000 0.000 0.000
4 315.384 50.195 104.000 39.000 0.000 0.000 0.000
5 315.384 50.195 104.000 39.000 0.000 0.000 0.000
6 315.384 50.195 104.000 39.000 0.000 0.000 0.000
7 315.384 50.195 104.000 39.000 0.000 0.000 0.000
8 315.384 50.195 104.000 39.000 0.000 0.000 0.000
3.3 First and Second Derivative Algorithm [11,48]

Samples of voltage collected by the relay are assumed to belong to a signals of the type:
v(tk ) = Vm sin(w tk + q v ) (3.55)
where we assume Vm, w and qv to be constant. Further we assume that the frequency w is
known. Thus in the above equations, we have two unknowns, namely the peak value of the
signal and its phase angle. Hence, we need one more equation to solve for the unknowns. Let
us generate another equation by taking derivative of the first equation giving:
v (tk ) = wVm cos(w tk + q v ) (3.56)

v (tk ) = -w Vm sin(w tk + q v )
2
(3.57)
Hence, we can find the amplitude of the voltage signal as:
v (t ) 2 v (t ) 2
Vm = (3.58)
+
k k
2
w w
Similarly, we can write for the current signal as:
i (t ) 2 i (t ) 2
I m = k + 2k (3.59)
w w
Further, on the lines of Mann and Morrison algorithm, it can be easily shown that:
v (tk )/w 2
q v = tan -1 - - w tk (3.60)
v (tk )/w
Similarly, we can write for the current signal:
i '' (t )/w 2
q i = tan -1 - ' k - w tk
(3.61)
i (tk )/w
The derivatives can be computed numerically as shown below:

(v - v ) (i - i ) (3.62)
vk = k +1 k -1 and ik = k +1 k -1
2 Dt 2 Dt
v(k + 1) - 2v(k ) + v(k - 1) (3.63)

v (k ) @
Dt 2
i(k + 1) - 2i(k ) + i(k - 1) (3.64)

i (k ) @
Dt2
The above results are summarised in Table 3.3.

Table 3.3 First and second derivative algorithm: Summary of results
Model of voltage signal v(tk ) = Vm sin(w tk + q v )
Model of current i(tk ) = I m sin(w tk + q i )
v (t ) 2 v (t ) 2
Vm = k
+
k
2
w w

Vm (vk +1 - vk -1 )
where, vk =
2 Dt
v(k + 1) - 2 v(k ) + v(k - 1)
v (k ) @
Dt2
v (tk )/w 2
qv q v = tan -1 - - w tk radians
v (tk )/w
i ' (t ) 2 i '' (t ) 2
I m = k + 2k
w w
Im (ik +1 - ik -1 )
where, ik =
2 Dt
i(k + 1) - 2i(k ) + i(k - 1)
i (k ) @
Dt 2
i (tk )/w 2
qi q i = tan -1 - - w tk radians
i (tk ) /w

Algorithm

Figure 3.8 Spreadsheet for implementation of first and second derivative algorithm.

3.3.1Excel Spreadsheet for Simulation of First and Second Derivative

3.3.2 MATLAB Simulation of First and Second Derivative Algorithm

% ----------------------------------------------------------------
% First and Second Derivative Algorithm
% -----------------------------------------------------------------
clear ; clc ; clf

%================================================
mag_z_set = 30 ; % Relay Reach Setting
theta_n = pi/2 ; % Relay MTA setting
%=================================================
%--------------Numerical Problem Setting --------------------------------
f = 50 ; % System frequency
OverSamplingFactor = 20 ; % fsamp = 2 * OSF * fsig
km = 100 ; % kilometers
R = 0.037 ; % ohms per km for 345kV line
XbyR_ratio = 9.9189 ; % for 345 kV line
%=================================================
f_samp = 2*OverSamplingFactor*f ;
X = R*XbyR_ratio ;
L = X/(2*pi*f) ;
Z = sqrt( R*R + X* X) ;
line_ang_deg = (180/pi)*atan2(X,R) ;
w = 2* pi* f ;
T = 1/f ;
dT = 1/f_samp ;
Vm = 100 ;
Im = Vm / Z ;
Qv_deg = 30 ;
Qv = Qv_deg*(pi/180) ;
Qi = Qv - atan2(X,R) ;
Qvdeg = Qv * (180/pi) ;
Qi_deg = Qi*(180/pi) ;
k=0;
sss=50;
%---------------------- Generation of Samples ----------------------------
for tim = 0:dT: T
k =k+1 ;
t(k) = tim ;
v(k) = Vm * sin( 2* pi* f * t(k) + Qv) ;
i(k) = Im * sin( 2* pi* f * t(k) + Qi) ;
end
% ---------------------------------------------------------------------
samples =k ;
for n = 2:samples-1
%---------- Estimation of voltage phasor -------------------------------
v_d(n) = ( v(n+1)-v(n-1)) / (2*dT) ;% First derivative
v_dd(n) = (v(n+1)-2*v(n)+ v(n-1))/(dT*dT) ;% 2nd Derivative
Vm_est(n) = sqrt( (v_d(n)/w)^2 + ( v_dd(n)/(w*w) )^2 ) ;
num = (v_dd(n))/((w*w)) ;
den = v_d(n)/w ;
y = num/den;
theta_v_est(n) = (atan2(- y,1))-((w*(n-1)*dT));
theta_v_deg_est(n)=theta_v_est(n)*(180/pi);
err_Vpc(n)=(((Vm_est(n))-(Vm))/(Vm))*100;
err_theta_v_deg(n)=(((theta_v_deg_est(n))-(Qv_deg))/(Qv_deg))*100;
%----------------------------------------------------------------------
%---------- Estimation of Current phasor -------------------------------
i_d(n) = ( i(n+1)-i(n-1)) / (2*dT) ;% First derivative
i_dd(n) = (i(n+1)-2*i(n)+ i(n-1))/(dT*dT) ;% 2nd Derivative
Im_est(n) = sqrt( (i_d(n)/w)^2 + ( i_dd(n)/(w*w) )^2 ) ;
numi = (i_dd(n))/((w*w)) ;
deni = i_d(n)/w ;
y = numi/deni;
theta_i_est(n) = (atan2(- y,1))-((w*(n-1)*dT));
theta_i_deg_est(n)=theta_i_est(n)*(180/pi);
err_ipc(n)=(((Im_est(n))-(Im))/(Im))*100;
err_theta_i_deg(n)=(((theta_i_deg_est(n))-(Qi_deg))/(Qi_deg))*100;
%----------------------------------------------------------------------
R_est(n)=(Vm_est(n)/Im_est(n))*cos(theta_v_est(n)-theta_i_est(n));
X_est(n)=(Vm_est(n)/Im_est(n))*sin(theta_v_est(n)-theta_i_est(n));
err_R_pc(n)=(((R_est(n))-(R))/(R))*100;
err_X_pc(n)=(((X_est(n))-(X))/(X))*100;
end
% ------------------------------------------------------------------
fprintf(------------------First and Second Derivative Algo.-------------
----------\n)
fprintf(--------------------------------------------------------------
------------\n)
fprintf(Vm Vm_est(n) Error in Vm Ph Ang Estimate_of_ph Error in Ph\n)
fprintf(--------------------------------------------------------------
------------\n)
for n=2:numel(Vm_est)
fprintf(%f %f %f %f %f %f\n,Vm,Vm_est(n),err_Vpc(n),Qvdeg, theta_v_
deg_est(n),err_theta_v_deg(n))
end
%----------------------------------------------------------------------
fprintf(--------------------------------------------------------------
------------\n)
fprintf(Im Im_est(n) Error in Im Ph Ang Estimate_of_ph Error in Ph\n)
fprintf(--------------------------------------------------------------
------------\n)
for n=2:numel(Im_est)
fprintf(%f %f %f %f %f %f\n,Im,Im_est(n),err_ipc(n),Qi_deg, theta_i_

deg_est(n),err_theta_i_deg(n))
end
fprintf(--------------------------------------------------------------
------------\n)
fprintf(R R_est err_R_pc X X_est err_X_pc \n)
fprintf(--------------------------------------------------------------
------------\n)
for n=2:numel(R_est)
fprintf(%f %f %f %f %f %f\n,R ,R_est(n), err_R_pc(n), X, X_est(n),
err_X_pc(n) )
end
% First and Second Derivative Algo.
Vm Vm_est(n) Error in Vm Ph Ang Estimate_of_ph Error in

Ph
100.000000 99.670624 0.329376 30.000000 30.057713 0.192377

100.000000 99.702694 0.297306 30.000000 30.058660 0.195534
100.000000 99.733704 0.266296 30.000000 30.053867 0.179558
100.000000 99.760621 0.239379 30.000000 30.043808 0.146026
100.000000 99.780814 0.219186 30.000000 30.029469 0.098229
100.000000 99.792312 0.207688 30.000000 30.012252 0.040841
Im Im_est(n) Error in Im Ph Ang Estimate_of_ph Error in

Ph
271.105781 270.273043 0.307164 54.243028 54.302015 0.108746

271.105781 270.186930 0.338927 54.243028 54.299300 0.103742
271.105781 270.108992 0.367675 54.243028 54.291073 0.088573
271.105781 270.046867 0.390591 54.243028 54.278134 0.064720
271.105781 270.006650 0.405425 54.243028 54.261750 0.034516
271.105781 269.992288 0.410723 54.243028 54.243529 0.000924
R R_est err_R_pc X X_est err_X_pc
0.037000 0.036244 2.042321 0.366999 0.366992 0.001959

0.037000 0.036279 1.948958 0.366999 0.367226 0.061793
0.037000 0.036384 1.664476 0.366999 0.367438 0.119546
0.037000 0.036550 1.216498 0.366999 0.367607 0.165617
0.037000 0.036760 0.648868 0.366999 0.367717 0.195458
0.037000 0.036994 0.017362 0.366999 0.367756 0.206118
3.4 Two-Sample Technique [25]

Let the signal be
v(k ) = Vm sin(w tk ) (3.65)
then we can write
v(k + 1) = Vm sin(w (tk + Dt )) (3.66)
v(k + 1) = Vm sin(w tk + wDt ) (3.67)
v(k + 1) = Vm sin(w tk ) cos(wDt ) + Vm cos (w tk ) sin (wDt ) (3.68)
Using Eq. (3.65), we can substitute the value of v(k)
v(k + 1) = v(k ) cos(wDt ) + Vm cos(w tk ) sin (wDt ) (3.69)

The above can be simplified to:
v(k + 1) - v(k ) cos(wDt )

= Vm cos(w tk ) (3.70)
sin(wDt )
Combining Eqs. (3.65) and (3.70), we get:
(v(k + 1))2 - 2 v(k )v(k + 1) cos wDt + (v(k ))2 cos2 (wDt )
(v(k ))2 + = Vm2 (3.71)
sin (wDt )
2
Thus, we can write Vm as:
(v(k ))2 sin 2 (wDt ) + (v(k ))2 cos2 (wDt ) + (v(k + 1))2 - 2 v(k )v(k + 1) cos(wDt )
Vm2 = (3.72)
sin 2 (wDt )
Simplifying further, we get:
(v(k ))2 + (v(k + 1))2 - 2v(k )v(k + 1) cos(wDt )

Vm2 = (3.73)
sin 2 (wDt )
Thus, Vm can finally be written as:
(v(k ))2 + (v(k + 1))2 - 2v(k )v(k + 1) cos(wDt ) (3.74)

Vm =
sin 2 (wDt )
Note that all the terms on the RHS of Eq. (3.74) are known.
Similarly,
(i(k ))2 + (i(k + 1))2 - 2i(k )i(k + 1) cos(wDt ) (3.75)

Im =
sin 2 (wDt )
Next, we will investigate, how to find the phase angle between voltage and current.
Let i(k ) = I m sin(w tk + q ) (3.76)

i(k ) = I m sin(w tk ) cos(q ) + I m cos(w tk )sin(q ) (3.77)
i(k + 1) = I m sin(w (tk + Dt )) cos(q ) + I m cos(w (tk + Dt ))sin(q ) (3.78)
i(k + 1) = I m [sin((w tk ) cos(wDt ) + cos(w tk ) sin(wDt )] cos(q )

+ I m [cos(w tk (cos(wDt ) - sin(w tk ) sin(wDt )] sin(q ) (3.79)
We have v(k ) = Vm sin(w tk ) , hence,

v( k )
sin(w tk ) = (3.80)
Vm
Similarly, we have shown above that:
v(k + 1) - v(k ) cos(wDt )

= Vm cos(w tk ) (3.81)
sin(wDt )
v(k + 1) - v(k ) cos(wDt ) (3.82)

cos(w tk ) =
Vm sin(wDt )
By manipulating the above equations, we can show that:
i(k )v(k ) + i(k + 1)v(k + 1) - (i(k )v(k + 1) + i(k + 1)v(k )cos(wDt )

q = cos -1 (3.83)
I mVm sin 2 (w Dt )
Summary of two-sample algorithm is shown in Table 3.4.
Table 3.4 Summary of two-sample algorithm
Signal model of voltage v(k ) = Vm sin(w tk )
Signal model of current i(k ) = I m sin(w tk + q )
(v(k ))2 + ( v( k + 1))2 - 2 v(k )v( k + 1) cos(wDt )

Vm
sin 2 (wDt )
(i(k ))2 + (i(k + 1))2 - 2i(k )i(k + 1)cos(wDt )

Im
sin 2 (wDt )
i(k )v(k ) + i(k + 1)v(k + 1) - (i(k )v(k + 1) + i( k + 1)) v( k )cos(wDt )

q cos -1
I mVm sin 2 (wDt )
3.4.1 Spreadsheet for Two-Sample Technique
Figure 3.9 Spreadsheet for two-sample algorithm.

3.4.2 MATLAB Simulation of Two-Sample Technique

% --------------------------------------------------------------
% Two-Sample Algorithm
% ---------------------------------------------------------------
clear ; clc ; clf
%==============================================
%=====================================================
%============= Numerical Problem Setting ===================
%=====================================================
X = R*XbyR_ratio ;
L = X/(2*pi*f) ;
Z = sqrt( R*R + X* X) ;
w = 2* pi* f ;
T = 1/f ;
dT = 1/f_samp ;
Vm = 100 ;
Im = Vm / Z ;
Qv_deg =0 ;% Theta_V = 0
Qi = 1 * Qi ; %In the algorithm ph. angle of
current is taken positive
Qvdeg = Qv * (180/pi) ;
Qi_deg = Qi*(180/pi) ;
k =0;
sss =50;
%========== Generation of Samples ================
for tim = 0:dT: T
k =k+1 ;
t(k) = tim ;
v(k) = Vm * sin( 2* pi* f * t(k) ) ;
i(k) = Im * sin( (2* pi* f * t(k) ) + Qi) ;

% both +Qi and -Qi give +Qi as angle
% it is not possible to distinguish bet. lead or lag
end
% ===============================================================
samples =k ;
for n = 1:samples-1
%---------- Estimation of voltage phasor -------------------------------
Vm_est(n)=sqrt((((v(n))^2)+((v(n+1))^2)-(2*v(n)*v(n+1)*cos(w*dT)))/
((sin(w*dT))^2));
Im_est(n)=sqrt((((i(n))^2)+((i(n+1))^2)-(2*i(n)*i(n+1)*cos(w*dT)))/
((sin(w*dT))^2));
%---------------------------------------------------------------
%---------- Estimation of Current phasor -------------------------------
num=(i(n)*v(n))+(i(n+1)*v(n+1))-(((i(n)*v(n+1))+(i(n+1)*v(n)))*cos(w*
dT));
den=(Im_est(n)*Vm_est(n)*((sin(w*dT))^2));
Ph_i_est(n)=acos(num/den);
Ph_i_est_deg(n)=(Ph_i_est(n))*(180/pi);
%---------------------------------------------------------------
%-------- Error in estimation of voltage and current phasors
err_Vm(n)=(((Vm_est(n))-(Vm))/(Vm))*100;
err_Im(n)=(((Im_est(n))-(Im))/(Im))*100;
err_Ph_i_deg(n)=(((Ph_i_est_deg(n))-(Qi_deg))/(Qi_deg))*100;
R_est(n)=(Vm_est(n)/Im_est(n))*cos(Ph_i_est(n));
X_est(n)=(Vm_est(n)/Im_est(n))*sin(Ph_i_est(n));
err_R_pc(n)=(((R_est(n))-(R))/(R))*100;
err_X_pc(n)=(((X_est(n))-(X))/(X))*100;
end
% ---------------------------------------------------------------
fprintf(------------------Two Sample Technique -----------------------
\n)
fprintf(--------------------------------------------------------------
------------\n)
fprintf(Vm Vm_est(n) Error in Vm \n)
fprintf(--------------------------------------------------------------
------------\n)
for n=1:numel(Vm_est)
fprintf(%f %f %f \n,Vm,Vm_est(n),err_Vm(n))
end
%---------------------------------------------------------------
fprintf(--------------------------------------------------------------
------------\n)
fprintf(Im Im_est(n) Error in Im Ph Ang Estimate_of_ph Error in Ph\n)
fprintf(--------------------------------------------------------------
------------\n)
for n=2:numel(Im_est)
fprintf(%f %f %f %f %f %f\n,Im,Im_est(n),err_Im(n),Qi_deg,Ph_i_est_
deg(n),err_Ph_i_deg(n))
end
fprintf(--------------------------------------------------------------
------------\n)
fprintf(R R_est err_R_pc X X_est err_X_pc \n)

fprintf(--------------------------------------------------------------
------------\n)
for n=2:numel(R_est)
fprintf(%f %f %f %f %f %f\n,R ,R_est(n), err_R_pc(n), X, X_est(n),
err_X_pc(n) )
end
Two Sample Technique.
Vm Vm_est(n) Error in Vm
100.000000 100.000000 0.000000

100.000000 100.000000 0.000000
100.000000 100.000000 0.000000
100.000000 100.000000 0.000000
100.000000 100.000000 0.000000
100.000000 100.000000 0.000000
100.000000 100.000000 0.000000
100.000000 100.000000 0.000000
100.000000 100.000000 0.000000
100.000000 100.000000 0.000000
Im Im_est(n) Error in Im Ph Ang Estimate Error in Ph

_of_ph
271.105781 271.105781 0.000000 84.243028 84.243028 0.00

271.105781 271.105781 0.000000 84.243028 84.243028 0.00
271.105781 271.105781 0.000000 84.243028 84.243028 0.00
271.105781 271.105781 0.000000 84.243028 84.243028 0.00
271.105781 271.105781 0.000000 84.243028 84.243028 0.00
271.105781 271.105781 0.000000 84.243028 84.243028 0.00
271.105781 271.105781 0.000000 84.243028 84.243028 0.00
271.105781 271.105781 0.000000 84.243028 84.243028 0.00
271.105781 271.105781 0.000000 84.243028 84.243028 0.00
271.105781 271.105781 0.000000 84.243028 84.243028 0.00
R R_est err_R_pc X X_est err_X_pc
0.037000 0.037000 0.000000 0.366999 0.366999 0.000000

0.037000 0.037000 0.000000 0.366999 0.366999 0.000000
0.037000 0.037000 0.000000 0.366999 0.366999 0.000000
0.037000 0.037000 0.000000 0.366999 0.366999 0.000000
0.037000 0.037000 0.000000 0.366999 0.366999 0.000000
0.037000 0.037000 0.000000 0.366999 0.366999 0.000000
0.037000 0.037000 0.000000 0.366999 0.366999 0.000000
0.037000 0.037000 0.000000 0.366999 0.366999 0.000000
0.037000 0.037000 0.000000 0.366999 0.366999 0.000000
0.037000 0.037000 0.000000 0.366999 0.366999 0.000000
Review Questions
1. The voltage and current samples at a relay location are shown below. The system
frequency is known to be 50 Hz and the sampling frequency is 10 kHz. Using Mann
and Morrison algorithm find the apparent resistance and reactance as seen from the
relay location.
Sample set no. 1
v i
(i) 55.000000 8.887405
(ii) 57.965138 8.682345
(iii) 60.873070 8.468716
(iv) 63.720929 8.246729
Sample set no. 2
(i) 66.505903 8.016604
(ii) 69.225243 7.778567
(iii) 71.876266 7.532854
(iv) 74.456357 7.279707
Sample set no. 3
(i) 76.962967 7.019376
(ii) 79.393625 6.752117
(iii) 81.745931 6.478195
(iv) 84.017563 6.197880
Sample set no. 4
(i) 86.206280 5.911448
(ii) 88.309922 5.619183
(iii) 90.326413 5.321371
(iv) 92.253762 5.018309
Sample set no. 5
(i) 94.090069 4.710293
(ii) 95.833519 4.397630
(iii) 97.482394 4.080626
(iv) 99.035065 3.759595
Sample set no. 6

(i) 100.490000 3.434855
(ii) 101.845764 3.106724
(iii) 103.101019 2.775527
(iv) 104.254525 2.441591
2.
What are the assumptions in the Mann and Morrison algorithm?
3.
Derive the Mann and Morrison algorithm.
4.
What are the drawbacks of using numerical derivative?
5.
Assuming that amplitude of signal is known, can you find its frequency by using an
algorithm similar to Mann and Morrison algorithm.
An ac signal v = Vm sin(2 pft + f) has three unknown Vm, f and f. Thus, we need
6.
three equations to solve for the three unknown. Using this idea can you derive an
algorithm?
7.
Three-sample algorithm assumes a sine waveform. Derive three-sample algorithm when
cosine waveform is taken as reference.
8.
What could be the applications of three-sample algorithm in general and in the field
of power system protective relaying?
9.
Discuss the importance of measurement of power system frequency.
10.
What are the assumptions underlying three-sample algorithm?
11.
What is the minimum number of samples required for making one estimate each of
amplitude, phase and frequency, assuming the signal to be undistorted sine wave?
12.
What happens to the estimation using three-sample algorithm if one of the samples
happens to be zero or very low in magnitude?
13.
Compare and contrast three-sample algorithm with Mann and Morrison algorithm.
14.
Can three-sample algorithm be applied under non-stationary conditions?
15. Numerical problem on three-sample technique:
(i) The samples of a single frequency signal are as shown below:
Make at least three estimates of the amplitude,phase and frequency
0 1 2 3 4 5 6
56.642460 26.917181 5.442939 37.270267 65.449321 87.221739 100.456286
=================================================================
The sampling frequency is 1003.90002 Hz.
(ii) The samples of a single frequency signal are as shown below:
Make at least three estimates of the amplitude,phase and frequency
0 1 2 3 4 5 6
5.442939 -37.270267 65.449321 87.221739 100.456286 103.857472 97.092364
=================================================================
The sampling frequency is 1003.90002 Hz.
16.
Derive the first derivative and second derivative algorithm.
17.
What are the assumptions underlying the first and second derivative algorithm?
18.
What could be the motivation for such algorithms involving higher derivatives?
19.
Derive the two-sample algorithm. What are the advantages of using this algorithm?
20.
What are the assumptions underlying the two-sample algorithm?
21.
Develop a spreadsheet for simulating the first and second derivative and two-sample
algorithm.
22.
Write MATLAB programs for simulating the first and second derivative algorithm and
two-sample algorithm.
4
Algorithms Based on Solution of
Differential Equation
4.1 Differential Equation Algorithm [20,47,56]

In this algorithm, we model the transmission line as a series lumped RL circuit as shown in
Figure 4.1 for a single phase line with zero fault resistance (bolted fault).
Figure 4.1 Series lumped R-L model of faulted transmission line.
It may be noted that we have ignored the shunt capacitance of the line and it is assumed
that fault resistance is zero. Note that v t) and i t) are the voltage and current available at the
relay location through CT and PT which are not shown in Figure 4.1. Hence, both the CT and
the PT ratio may be considered to be equal to 1. Figure 4.2 shows the sampling instants and
sampled values of voltage and current.
If we apply KVL around the loop formed because of fault, at instant tk , we can write:-
di(tk )
v(tk ) = R i(tk ) + L
dt
It can be easily seen that in the above differential equation, the knowns are v(tk) and i(tk)
, and the unknowns are R and L, i.e., the resistance and the inductance of the transmission
line from relay location upto the fault point. The rate of change of current term, even though
60
Algorithms Based on Solution of Differential Equation61
Figure 4.2 Voltage and current samples.
not measured directly, can be computed from the current measurements. In order to solve for
the two unknowns, we need two equations. The second equation can be generated by applying
KVL at another instant of time say (k + 1)th instant as shown below:
di(tk )
v(tk ) = R i(tk ) + L (4.1)
dt
di(tk +1 )
v(tk +1 ) = R i(tk +1 ) + L (4.2)
dt
Let us represent di(t)/dt by i(t) and rewrite Eqs. (4.1) and (4.2) as:
v(k ) = R i(k ) + Li (k ) (4.3)
v(k + 1) = R i(k + 1) + Li (k + 1) (4.4)
Note that when the numerical values of the derivative of the current are inserted in the
above equations, they are no longer differential equations but become simultaneous algebraic
equations with constant coefficients.
We can write the above two equations in matrix form as:
v ( k ) i( k ) i ( k ) R
= (4.5)
v ( k + 1) i(k + 1) i (k + 1) L
Which gives us:
-1
R i(k ) i ' ( k ) v( k )
= (4.6)
L i(k + 1) i ' (k + 1) v(k + 1)
R 1 i (k + 1) -i (k ) v(k )
L = i(k )i (k + 1) - i (k )i(k + 1) -i(k + 1) i(k ) v(k + 1) (4.7)

Finally we can express R and L as a function of various samples of voltage and current
and computed derivatives of current
i (k + 1)v(k ) - i (k )v(k + 1)
R = (4.8)
i(k )i (k + 1) - i (k )i(k + 1)
i(k )v(k + 1) - i(k + 1)v(k )
L= (4.9)
i(k )i (k + 1) - i (k )i(k + 1)
Note that the numerical derivatives can be calculated using central difference formula as:
v(k + 1) - v(k - 1)
v (k ) = (4.10)
2 Dt
i(k + 1) - i(k - 1) (4.11)
i ( k ) =
2 Dt
where Dt is the sampling time period given by 1/fsampling.
The block-diagram showing the input to the algorithm and the block required for
implementation of specific distance relay characteristics is shown in Figure 4.3.
Figure 4.3Block-diagram showing inputs to differential equation algorithm and block for
implementation of specific distance relay.
It can be seen from Figure 4.3 that after estimating the apparent R and L as seen from
the relay location, we will have to check whether the apparent impedance lies within the trip
region of the relay characteristics being implemented. Thus, in addition to the differential
equation algorithm, we will have to write another algorithm for implementing specific distance
relay characteristics.
4.2 Justification for Lumped Series R-L Model

Before we proceed further, let us find out how far we are justified in representing the transmission
line as a lumped series R-L circuit. Let us calculate the shunt and series parameters of a length
of 100 km, 345 kV line. We will use the data from Table 4.1 which happens to be for a 60
Hz power system.
Table 4.1 Data of a 60 Hz power system
230 kV 345 kV 500 kV 765 kV 1100 kV
R /km 0.050 0.037 0.028 0.012 0.005
XL 0.488 0.367 0.326 0.329 0.292
XL/R 9.76 9.9189 11.6428 27.4166 58.4
1
tan (xL/R) deg 84.15 84.2430 85.09 87.91 89.019
S s/km 3.371 4.518 5.20 4.978 5.544
Zc 380 285 250 257 230
Succeptance, S = 2pfC = (2)(p)(60)(C) = 4.518 106 S/km
4.518 10 6
Hence, C= = 0.01198436721 micro-farad (mF)
(2)(p )(60)
Consider 100 km length of line, with total capacitance lumped at one end.
Ctotal = (100) (0.01198436721) mF = 1.2 mF
1 1
Zshunt = XC = = = j2210.486 W
jw C j ((2)(p )60(1.2 10 6 ))
For p model, of shunt capacitance is connected at each end. Hence, XC becomes double,
i.e., j 4420.97 W.
Figure 4.4 Entire shunt capacitance modelled at sending end with fault on.
Figure 4.5 p model of the transmission line with fault on.
Table 4.2
Consider of the 100 km length of the 345 kV line modelled by:
Series lumped model Total capacitance lumped at sending p model. with capacitance
end lumped at each end
XL = (100) (0.367) = j 36.7 W Zseen = (3.7 + j36.7)||(j 2210.486) Zseen = (3.7 + j36.7)||(j 4420.97)
R = (100) (0.037) = 3.7 W (3.7 + j 36.7) ( - j 2210.486) (3.7 + j 36.7) ( - j 4420.97)
= =
Zseen = (3.7 + j 36.7) W (3.7 + j 36.7)( - j 2210.486) (3.7 + j 36.7)( - j 4420.97)
= (3.825 + j 37.313) W = (3.7622 + j 37.00403) W
|Zseen|=36.886 84.243 Zseen = 37.5087 84.1455 Zseen = 37.19484.19468

% Error in lumped R-L model w.r.t % Error in lumped R-l model w.r.t.
this model this model
(36.886 37.5087) (36.886 37.194)
100 = 1.66 % 100
(37.5087) (37.194)
= 0.8280 %
Thus, we see that the error introduced because of ignoring shunt capacitance is less than
1%, which is not at all serious. Hence, our representation of the transmission line as a series
R-L circuit is justified.
4.3 E

Equation Algorithm

Figure 4.6 Excel spreadsheet implementation of differential equation algorithm.

xcel Spreadsheet Implementation of the Differential
4.4 MATLAB Implementation of Differential Equation Algorithm

% ----------------------------------------------------------------
% Numerical Problem on Differential Equation Algorithm
% -----------------------------------------------------------------
clear ; clc ; clf

%=================================================================
%=================================================================
%============= Numerical Problem Setting ===========================
%===============================================================
X = R*XbyR_ratio ;
L = X/(2*pi*f) ;
Z = sqrt( R*R + X* X) ;
w = 2* pi* f ;
T = 1/f ;
dT = 1/f_samp ;
Vm = 400*sqrt(2) ;
Im = Vm / Z ;
Qv_deg = -15 ;
Qvdeg = Qv * (180/pi) ;
Qideg = Qi*(180/pi) ;
k =0 ;
sss = 50 ;
for tim = 0:dT: T
k = k+1 ;
t(k) = tim ;
v(k) = Vm * sin( 2* pi* f * t(k) + Qv) ;
i(k) = Im * sin( 2* pi* f * t(k) + Qi) ;
end
samples = k ;
m = 0 ;
for n = 2:samples-1
m = m+1 ;
i_dash(m) = ( i(n+1)-i(n-1)) / (2*dT) ;

end
fprintf(================================================= \n)
disp(Numerical Problem)
fprintf( A Mho Relay with following settings is to be implemented \n) ;
fprintf( Z Setting = %f Angle Z setting = %f\n , mag_z_set , (180/pi)* theta_n)
fprintf( Use differential algorithm to make two estimates apparent R and L
\n)
fprintf( Given that the system freq is %f \n,f )
fprintf( and given that the sampling freq is %f \n,f_samp )
%fprintf(\n Sampling Interval =%f Sampling Freq = %f\n , dT , 1/dT)
%---------------------------------------------------------------------
--
fprintf(The Voltage and Current Samples are :- \n);
fprintf(Voltage Samples )
fprintf(\n==================================================\n);
nn =8 ;
for kk = 1 : nn
fprintf( %f | ,v(kk)) ;
end
fprintf(\n===================================================\n);
fprintf(\nCurrent Samples)
fprintf(\n===================================================\n);
for kk = 1 :nn
fprintf(%f | ,i(kk)) ;
end
fprintf(\n===================================================\n);
% ===== The DEA Algorithm

% den = i(k)i(k+1)-i(k+1)i(k)
% R(k) = i(k+1)v(k)- i(k)v(k+1)/den
% L(k) = i(k)v(k+1) - i(k+1)v(k) /den
%(numel(v))-4
num_calc =8 ;
for k=2:num_calc
den(k) = (i(k)*i_dash(k+1) )- ( i(k+1)*i_dash(k) ) ;
RR(k) = ((i_dash(k+1)*v(k))-(i_dash(k)*v(k+1)))/(den(k)) ;
LL(k) = ((i(k)*v(k+1))-(i(k+1)*v(k)))/(den(k)) ;
err_R_pc(k) = (((RR(k))-(R))/(R))*100 ;
err_L_pc(k) = (((LL(k))-(L))/(L))*100 ;
end
R,RR,err_R_pc
L,LL,err_L_pc
%--------- Printout the true values -------------------

fprintf(\n R = %f X = %f Z = % f , R,X,Z ) ;
fprintf( Line Angle = %f \n, line_ang_deg ) ;

fprintf( X/R ratio = %5.2f \n , X/R ) ;
%---------------------------------------------------------------------
=========================================================================================
Numerical Problem
A mho relay with following settings is to be implemented:
Z setting = 30.000000 and Angle Z setting = 90.000000
Use differential algorithm to make two estimates apparent R and L.
Given that the system frequency = 50.000000 and the sampling frequency = 200000.000000
The Voltage and Current Samples are :
Voltage Samples
=========================================================================================
146.410162 | 145.551682 | 144.692844 | 143.833648 | 142.974098 | 142.114195 | 141.253941 | 140.393338 |
=========================================================================================
Current samples
=========================================================================================
1513.693429 | 1514.078498 | 1514.459831 | 1514.837427 | 1515.211286 | 1515.581406 | 1515.947787 |
1516.310427 |
=========================================================================================
R = 0.0370
RR = 00.03760.03760.03760.03760.03760.03760.0376
err_R_pc = 01.55811.55811.55811.55811.55811.55811.5581
L = 0.0012
LL = 00.0012 0.0012 0.00120.00120.00120.00120.0012
err_L_pc = 1.0e003 *
00.16450.16450.16450.16450.16450.16450.1645
R = 0.037000, X = 0.366999, Z = 0.368860, Line angle = 84.243028

X/R ratio = 9.92
4.5
Solution of Differential Equation Algorithm Using Numerical
Integration
We use the same model of the faulted transmission line, as before, as shown in Figure 4.7.
We have already seen that the differential equation relating the voltage and current at the relay
location v(t ) = R i(t ) + L di(t )/dt gets converted into an algebraic equation when we substitute
the numerical value of the derivative term. However, a derivative has the problem that it tends
to amplify noise because of its high-pass nature. This is very easy to see. Let the signal be
v = Vm,signal sin(wsignalt) + Vm,noise sin(wnoiset) (4.12)
Note that wnoise >> wsignal and Vm,noise << Vm,signal
Figure 4.7 Model of faulted transmission line.
Thus, the noise amplitude may be smaller than the signal amplitude to start with, but since
the noise frequency is many orders of magnitude higher than the signal frequency, after taking
derivative it will dominate over the signal as can be seen from the expression for derivative,
v= (wVm) cos(wsignalt) + (wnoiseVm,noise) cos(wnoiset) (4.13)
Look at the amplitude of the noise. It is amplified by wnoise. This is problematic because
real life signals are always replete with noise. Thus, in the process of differentiation the noise
will get amplified. This may happen so much so that many times the noise will mask out the
signal of interest. Thus, wherever possible we should try to avoid taking derivative. This leads
to implementation of the differential equation algorithm using numerical integration. We start
with the basic equation relating the relay voltage and current:
di
v = Ri + L (4.14)
dt
Integrating both sides between two instants of time t1 and t2 and then between t3 and t4,
we get:
t2 t2 t2
di(t ) (4.15)
v(t )dt = R i(t ) dt + L dt
dt
t1 t1 t1
t4 t4 t4
di(t ) (4.16)
v(t )dt = R i(t ) dt + L dt
t3 t3 t3 dt
Which can be written as:

t2 t2 t2
v(t )dt = R i(t ) dt + L di(t ) dt (4.17)

t1 t1 t1
t4 t4 t4
(4.18)
v(t )dt = R i(t ) dt + L di(t ) dt
t3 t3 t3
Note that the last term in each equation does not involve finding the numerical integration
since it reduces simply to:
t2
di(t ) dt = i(t )t = i(t2 ) - i(t1 )

t2
1
(4.19)
t1
t4
di(t ) dt = i(t )t = i(t4 ) - i(t3 )

t4
(4.20)
3
t3
Note, further, that the other integrals can be evaluated numerically. Thus, when we
substitute the numerical values of the integrals, we get algebraic equations. To simplify the
derivation further, let us use the following representation for the numeric values of the integrals:
t2 t2 t2

v(t )dt = E i(t )dt = A di(t )dt = B (4.21)
t1 t1 t1
t4 t4 t4

v(t )dt = F i(t ) dt = C di(t )dt = D (4.22)
t3 t3 t3
Thus, we can write the two differential equations as the following two simultaneous
algebraic equations with constant coefficients:
E = R A + L B (4.23)
F = R C + L D (4.24)
Let us represent the system of equations in matrix form as shown below:
E A B R
F = C D L (4.25)

Thus, we can find the unknowns R and L as follows:
-1
R A B E (4.26)
L = C D F

R 1 D -B E (4.27)
L = ( AD - BC ) -C A F

Solving the above, we get:
( DE - BF )
R= (4.28)
( AD - BC )
( AF - CE )
and L= (4.29)
( AD - BC )
Note that A, C, E and F are the values returned by the numerical integrals defined
above.
4.6MATLAB Implementation of Solution of Differential Equation

Algorithm Using Numerical Integration
We can implement the numerical integration using a number of numerical techniques. However
we restrict ourselves to the two very well-known techniques viz., trapezoidal rule and the
famous Simpsons rule of integration. We have to appreciate that numerical integration , like
any other numerical method, is not perfect. It involves errors. Depending upon how far the
waveform being integrated conforms to the underlying assumption on which the algorithm is
based, the errors will be more or less. Next we take up the derivation of the two methods of
numerical integration.
4.6.1 Trapezoidal Rule for Numerical Integration
Figure 4.8 Trapezoidal rule of integration.
Referring to Figure 4.8, the area under the curve from x0 to x1 is

x1
f ( x )dx @ Area of square + Area of traiangle (4.30)

x0
1
= ( x1 - x0 ) y0 + ( x1 - x0 )( y1 - y0 ) (4.31)
2
1 1
= ( x1 - x0 ) y0 - ( x1 - x0 ) y0 + ( x1 - x0 ) y1 (4.32)
2 2
1 1
= ( x1 - x0 ) y0 + ( x1 - x0 ) y1 (4.33)
2 2
1
= (( x1 - x0 )( y0 + y1 )) (4.34)
2
substituting x1 x0 = h we get:
h
= ( y0 + y1 ) (4.35)
2
Similarly, the next adjacent segment has area:

h
= ( y1 + y2 ) (4.36)
2
and so on till the last segment having area
h
= ( yn -1 + yn ) (4.37)
2
Hence we can write:
xn
f ( x ) dx = 2 [ y0 + y1 + y1 + y2 + y2 + + yn -1 + yn -1 + yn ]
h
(4.38)
x0
xn
f ( x ) dx = 2 [ y0 + 2[ y1 + y2 + + yn -1 ] + yn ]
h
(4.39)
x0
We can put it in words as Trapezoidal rule of integration, i.e.,

xn
h
f ( x )dx = 2 [first sample + last sample + 2(sum of remaining samples)] (4.40)
x0
4.6.2 Simpsons Rule for Numerical Integration

The interpolating curve in case of trapezoidal rule was a straight line. Instead, if we use
other interpolating curves, we get other numerical integration formulas. Simpsons 1/3 rule
uses parabola as the interpolating curve and results in the following formula for integration
as follows:
xn
h
f ( x )dx = 3 [ y0 + y1 + y3 + + yn -1 + 2 y2 + y4 + + yn -2 ) + yn ] (4.41)
x0
Note that xn = x0 + nh;

where n is the number of sub-intervals into which the interval of integration is divided.
For Simpsons rule n should be an even number as shown in Figure 4.9. We can put the
Simpsons rule for numerical integration into words , for ease in recall as Simpsons 1/3 Rule
of Integration, i.e.,

Figure 4.9 Simpsons rule for numerical integration.
x n
h
f ( x )dx = [first sample + 4(sum of odd indexed samples) + 2(sum of even indexed samples)]
x0 3
(4.42)
4.6.3MATLAB Script of Implementation of Solution of Differential Equation

Algorithm Using Numerical Integration
% DEA Implementation by Numerical Integration
% -------------------------------------------------------
% Faulted line is modelled as a series R-L circuit.
% v(k) = R i(k) + L di(k) /dt ...1
% v(k+1) = R i(k+1) + L di(k+1)/dt ...2
% In (1) and (2) the only unknowns are R and L.
% Integrating between t1,t2 and t3,t4 to get two equations
% ---------------------------------------------------------
% I_t1_t2(v) = R I_t1_t2(i) + L [i(t2)- i(t1)]
% I_t3_t4(v) = R I_t3_t4(i) + L [i(t4)- i(t3)]
% A =R B +LC
% D =R E +LF
% Where A , B , D , E are numerical integrals
% C = i(t2) - i(t1) and F = i(t4) - i(t3)
% [A ] = [B C ] [R ]
% [D ] = [E F ] [L ]
% [R ] = [ F/d -C/d ] [A ]
% [L ] = [ -E/d B/d] [D ]
% d = 1/(BF-EC)
% R = AF-CD /BF-CE
% L = BD-AE/ BF-CE
% ----------------------------------------------------------
clf ;
clear ,
clc
% ----------------------------------------------------------
f = 50 ;
OverSamplingFactor = 500 ;
Vm = 110 ;
R =1 ;
XbyR = 20 ;
% ----------------------------------------------------------
fsampling = 2 * f * OverSamplingFactor ;
delta_t = 1/fsampling ;
L = XbyR /(2*pi*f) ;
X = 2*pi*f*L ;
phi = atan2(X,R) ;
phi_deg = phi*180/pi ;
theta = phi ; % Switching angle
tau = L/R ;
X = 2*pi*f*L ;
Z = R+j*X ;
Zmag = abs(Z) ;
Im = Vm/Zmag ;
phi = atan2(X,R) ;
% v = Vm sin(2*pi*f*t)
% i = Imsin(2*pi*f*t + theta - phi) - Im exp(-t / tau) sin(theta -phi);
k =0 ;
%delta_t = 1e-5 ;
T = 1/50 ;
h = delta_t ; % for Simpsons Rule
% ------- Generation of samples of v and i ----------------------
% -----------------------------------------------------------------
for t=0:delta_t:T
k = k +1 ;
v(k) = Vm * sin(2 * pi * f * t ) ;
i_ss(k) = Im*sin((2*pi*f*t)-phi) ;
i_tran(k) = Im*(exp(-t/tau))*(sin(theta-phi)) ;
%i(k) = Im*sin((2*pi*f*t)-phi)- Im*(exp(-t/tau))*(sin(theta-phi));
i(k) = i_ss(k) - i_tran(k) ;
end % end of for loop

% -----------------------------------------------------------------
subplot(3,1,1)
plot(v) ;
grid on ;
xlabel(Sample No. )
ylabel(v )
axis([1 length(v) -Vm Vm ])
title(Solution of Differential Equation Algorithm )
subplot(3,1,2)
plot(i);grid on
axis([1 length(v) -2*Im 2*Im ])
xlabel(Sample No. )
ylabel(i )
subplot(3,1,3)
plot(i_tran)
axis([1 length(v) -Im Im])
xlabel(Sample no.)
ylabel(DC offset in current)
% --------------------------------------------------------------------
% ------- Creation of two arrays of v() and i() for integration -----
% --------------------------------------------------------------------
samp = 64 ;
for n = 1:samp
v1(n) = v(n) ;
v2(n) = v(n+samp) ;
i1(n) = i(n) ;
i2(n) = i(n+samp) ;
end
v1_intg = my_simpson(v1,h) ;
v2_intg = my_simpson(v2,h) ;
i1_intg = my_simpson(i1,h) ;
i2_intg = my_simpson(i2,h) ;
A = v1_intg; B = i1_intg;C = i1(samp)-i1(1) ;
D = v2_intg; E = i2_intg; F =i2(samp)-i2(1) ;
% ----------------------------------------------------------------
den = B*F - C*E ;
RR = ( A*F - C*D) / (den ) ; % Computed value of L
LL = (B*D - A*E) / (den ) ; % Computed value of R
XX = 2 * pi * f * LL ;
Error_R = ((RR-R)/(R))*100 ;
Error_L = ((LL - L)/(L))*100 ;
% ------------------------------------------------------------------
fprintf(Sampling frequency = %f \n, fsampling);
sprintf(Samples per integration interval %d,samp)
sprintf(Setting R = %f Setting L = %f Setting X=%f,R,L,X)
sprintf(Computed R = %f Computed L = %f Computed X=%f,RR,LL,XX)
sprintf(Error in R = %f %% Error in L = %f %%,Error_R,Error_L)
% input: samp and delta_theta || output : value of integral

% my_simpson(samp,delta_theta)
% samp is array of samples , delta_theta is sampling interval in radians
% ----- code for the function [integral]=my_simpson(samples,h)
% % % function [integral]= my_simpson(samples ,h)
% % % % start
% % % L = length(samples);
% % % first=samples(1);
% % % last = samples(L);
% % % even= 0 ; odd = 0;
% % % for k = 2:2:L-2
% % % even = even + samples(k);
% % % odd = odd + samples(k+1);
% % % end
% % % integral= (h/3)* ( first + last + 4*odd + 2 *even );

% % % % end of function my_simpson
% % %-------------------------------------------------------------------
Sampling frequency = 50000.000000

Ans = Samples per integration interval 64
Ans = Setting R = 1,000000 Setting L = 0.063662 Setting X = 20.000000
Ans = Computed R = 0.979276 Computed, L = 0.063329, Computed X = 19.895521
Ans = Error in R = 2.072414%, Error in L = 0.522396 %
Figure 4.10 MATLAB output.
4.7Application of Differential Equation Algorithm to Three-

Phase Line
Differential equation algorithm when applied to Three-phase line needs careful consideration.
This is because there are a total of 11 shunt faults that can take place on the three-phase line.
Further, whatever happens in one-phase of a three-phase line depends upon what is happening
in other two-phase because of the inductive and capacitive coupling between the phases.
The protection of a three-phase line is organised in the form of two distinct protections
viz., ground-fault protection and phase-fault protection. Appropriate voltage and current signals
need to be selected from the three voltage signals (va, vb, vc) and three current signals (ia, ib,
ic) available at the relay location to implement these two protections. Thus, six numbers of
measuring units (M.U.) are required for implementing protection against all 11 shunt faults
that can occur on a three-phase line as shown in Figure 4.11.
Figure 4.11 Organisation of distance protection of a three-phase line.
4.7.1 Ground Fault Protection of Three-Phase Line Using Phase Quantities

We will first take up ground fault protection. Figure 4.12 shows phase a to ground fault at a
distance x from the sending end ( relaying end). We are interested in finding out the voltage/
current pair whose ratio correctly measures the impedance of the faulted line.
We will adopt the model shown in the Figure 4.12 for the sake of developing differential
equation algorithm for three-phase line. It can be seen that we are again neglecting the shunt
capacitance. Initially, we develop equations based on the series distributed parameter model
and then lump the parameters.
Voltage drop across a small length dx of phase a of the line, denoted by dVaa can be
written as:
di di di
dVaa = ( Ra dx )ia + ( La dx ) a + ( Lab dx ) b + ( Lac dx ) c (4.43)
dt dt dt
Figure 4.12 Model of three-phase line for developing DEA for three-phase lines.
Assuming that there is a single line to ground bolted fault at length x from the sending
end, the voltage across length x can then be written as:
di di di
Vaa = ( Ra x )ia + ( La x ) a + ( Lab x ) b + ( Lac x ) c (4.44)
dt dt dt
di di di
Vaa = x ( Ra )ia + x ( La ) a + x ( Lab ) b + x ( Lac ) c (4.45)
dt dt dt
d L L
Vaa = x ( Ra )ia + x ( La ) ia + ab ib + ac ic (4.46)
dt La La
Thus, voltage of phase a at the relay location not only depends upon the current in phase
a but also upon currents in the other phases as well as the self-inductance of phase a and
mutual inductances between the phases a and b, a and c.
Note that vaa = va for line to ground fault on phase a. Hence, we can write:
diq L L
va = x ( Ra )i p + x ( La ) where ip = ia and iq = ia + ab ib + ac ic (4.47)
dt La La
Similarly, for b-g fault and c-g fault we can write:
diq L L
vb = x ( Rb )i p + x ( Lb ) where ip = ib and iq = ab ia + ib + bc ic (4.48)
dt Lb Lb

diq L L
vc = x ( Rc )i p + x ( Lc ) where ip = ic and iq = ca ia + cb ib + ic (4.49)
dt Lc Lc
In each of these equations, there are two unknowns (xRa) and (xLa) and thus we again have
di(t )
equation of the form v = R i(t ) + L albeit the current terms i(t) and di(t)/dt are different.
dt
Thus, we can follow the same methodologies that we followed for the single phase case.
Figure 4.13 shows the input and output of the phase a ground fault measuring unit. Similar
measuring units in block-diagram form for b to g and c to g faults will have to be set
up to monitor and provide protection against b-g and c-g faults.
Figure 4.13 Phase a ground fault measurement unit of a three-phase line.
Note that we have exclusively used all the phase quantities in the above measurement.
We can instead make use of sequence (symmetrical component) quantities. We will take this
up in a later section.
4.7.2 Ground Fault Protection of Three-Phase Line Using Sequence Quantities

We can write the voltage Va at the relay location for a bolted line to ground fault at distance x as:
di di di
va = ( Ra x )ia + ( La x ) a + ( Lab x ) b + ( Lac x ) c (4.50)
dt dt dt
We note that, for a well designed symmetric line:
Ra = Rb = Rc = R; Laa = Lbb = Lcc = LS = Self inductance of each phase of the three-phase
line and Lab = Lbc = Lca = LM = Mutual inductance between phases of the three-phase line
Hence we can write:
di di di
va = ( Rx )ia + ( LS x ) a + ( LM x ) b + ( LM x ) c (4.51)
dt dt dt
di d
va = ( Rx )ia + ( LS x ) a + ( LM x ) (ib + ic ) (4.52)
dt dt
Now, we have ia + ib + ic = 3 i0 (ib + ic) = 3i0 ia
di d
va = ( Rx )ia + ( LS x ) a + ( LM x ) (3i0 - ia ) (4.53)
dt dt
di di di
va = ( Rx )ia + ( LS x ) a + 3 xLM 0 - xLM a (4.54)
dt dt dt
di di
va = (Rx )ia + x ( LS - LM ) a + 3 xLM 0 (4.55)
dt dt
We have positive and zero sequence inductance of the line L1 and L0 given by:
L1 = LS LM (4.56)
L0 = LS + 2 LM (4.57)
L0 L1 = 3 LM (4.58)
di di
va = ( Rx )ia + x (L1 ) a + x ( L0 - L1 ) 0 (4.59)
dt dt
d L - L1
va = ( Rx )ia + x ( L1 ) ia + 0 i0 (4.60)
dt L1
d
va = x R ia + xL1 (ia + ki0 ) (4.61)
dt
L - L1
where, k = 0 (4.62)
L1
Hence for a three-phase line, for a-g fault we get an equation which is exactly of the form
di(t ) ; albeit the terms in the i(t) and di(t ) are different but known currents.
v(t ) = Ri(t ) + L
dt dt
Hence we can use the differential equation algorithm developed for single phase line.
For catering to b-g and c-g faults we can similarly prove that:
d
vb = xRib + x L1 (ib + ki0 ) (4.63)
dt
d
vc = x R ic + x L1 (ic + k i0 ) (4.64)
dt
Figure 4.14Input and output of ag ground fault Figure 4.15Input and output of bg ground fault
measuring unit. measuring unit.
Figure 4.16 Input and output of c g ground fault measuring unit.
4.7.3 Phase Fault Protection of Three-Phase Line Using Phase Quantities

Consider a phase a to phase b fault at a distance x from the relaying end as shown in
Figure. 4.17.
Figure 4.17 Phase a to phase b phase fault: Development of relaying quantities.

We can write:
dia di di
va = x ( Ra )ia + x ( La ) + x Lab b + x Lac c (4.65)
dt dt dt
dia di di
vb = x ( Rb )ib + x ( Lab ) + x Lb b + x Lbc c (4.66)
dt dt dt
di di
va - vb = x ( Ra )ia - x ( Rb )ib + x ( La ) a - x ( Lab ) a
dt dt
di di di di
+ x ( Lab ) b - x ( Lb ) b + x ( Lac ) c - x ( Lbc ) c (4.67)
dt dt dt dt
R di di di
va - vb = x Ra ia - b ib + x ( La - Lab ) a + x ( Lab - Lb ) b + x ( Lac - Lbc ) c (4.68)
Ra dt dt dt
For a fully transposed line it can be assumed that Lac Lbc, hence the last term in the
above expression vanishes, giving:
R di L - Lb dib
Va - Vb = x Ra ia - b ib + x ( La - Lab ) a + ab (4.69)
Ra dt La - Lab dt
R di L - Lab dib
Va - Vb = x Ra ia - b ib + x ( La - Lab ) a - b (4.70)
Ra dt La - Lab dt
R d L - Lab
Va - Vb = x Ra ia - b ib + x ( La - Lab ) ia - b ib (4.71)
Ra dt La - Lab
diq
Va - Vb = x Ra (i p ) + x ( La - Lab ) (4.72)
dt
Rb Lb Lab
where i p = ia - ib and iq = ia - (ib ) (4.73)
Ra La - Lab
L - Lab R
Note that for a symmetric line b and b will be equal to 1 and the above equation
reduces to: La - Lab Ra
d (ia - ib )
Va - Vb = x Ra (ia - ib ) + x ( La - Lab ) (4.74)
dt
Figure 4.18 Inputs and outputs of a-b phase fault measuring unit.
4.7.4 P
hase Fault Protection of Three-Phase Lines Using Sequence
Quantities
We have seen that during a-b phase fault, the voltage (Va Vb) can be expressed as:
d (ia - ib )
dt
diq
Va - Vb = x Ra (i p ) + x ( La - Lab ) (4.76)
dt
Rb L - Lab
where i p = ia - ib and iq = ia - b
L - L (ib ) (4.77)
Ra a ab
L - Lab R
Note that for a symmetric line, b and b will be equal to 1 and the above
equation reduces to: La - Lab Ra
d (ia - ib )
dt
We have: Ra = R, La = LS, the self-inductance and Lab = LM, the mutual inductance
d (ia - ib ) (4.79)
Va - Vb = x R(ia - ib ) + x ( LS - LM )
dt
The positive sequence inductance L1 can be written as: L1 = LS LM
d (ia - ib )
Va - Vb = x R(ia - ib ) + x ( L1 ) (4.80)
dt
Figure 4.19 Inputs and outputs of measuring unit for a-b phase fault.
4.8 E
limination of Specific Harmonics by Integration between
Selected Limits
Again consider the basic differential equation between voltage and current at the relay
location
diq
v = R ip + L (4.81)
dt
Referring to Figure 4.20, if we choose the first limits of integration as 0 to a and second
limit as (p/3) to ((p/3) + a) we get the following two simultaneous equations:
Figure 4.20 Elimination of third harmonic by choosing proper limits of integration.
a /w a /w a /w
v dt = R i p dt + L diq (4.82)
0 0 0
p p p
+a w +a w +a w
3 3 3
v dt = R i p dt + L diq (4.83)
p / 3w p / 3w p / 3w
Adding the above two equations, we get:

p
+a w p
+a w p
a /w 3
a /w 3
3 + p
vdt + vdt = R i p dt + i p dt + L iq w - iq [0] + iq w - iq 3w
0 p /3w 0 p /3w

(4.84)
Similarly, we can choose another interval b to generate another equation:
p
+ b w
p
+ b w p
+b
p
b /w b /w
3
3
b
vdt + vdt = R i p dt + i p dt + L iq w - iq [0] + iq w - iq 3w
3
0 p / 3w 0 p / 3w

(4.85)
Let us make the following substitutions:
p p
+a w
+a w 3
a /w 3 a /w

E= vdt + vdt; A = i p dt + i p dt (4.86)
p / 3w 0 p / 3w
0

p
3 + p
B = iq - iq [0 ] + iq - iq (4.87)
w w 3w

p
+b w p
+b w

b /w 3 b /w 3

F= vdt + vdt; C = i p dt + i p dt (4.88)
0 p / 3w 0 p / 3w

p
b 3 +b p
D = iq - iq [0 ] + iq - iq (4.89)
w w 3w
Then we can express the two equations in matrix form as:

E = R A + L B (4.90)
F = R C + L D (4.91)
E A B R
= (4.92)
F C D L
Hence, R and L can be found out as:

-1
R A B E
L = C D F (4.93)

Review Questions
1. Derive the differential equation algorithm.

2. What is the difference between solution of DEA by numerical derivative and numerical
integration?
3. Do we model the signal at the relay location or the power system in differential
equation algorithm?
4. How do we model the system in differential equation algorithm?
5. Can lumped series model of an EHV transmission line be justified?
6. What are the problems inherent in the numerical derivative approach?
7. How numerical integration approach is better than numerical derivative approach?
8. Derive the trapezoidal method of numerical integration.
9. What is the curve underlying the trapezoidal rule?
10. Derive the Simpsons 1/3 rule for numerical integration.
11. What is the underlying curve for the Simpsons 1/3 rule?
12. State names of other methods of numerical integration.
13. Explain how the distance protection of a three-phase line is organised?
14. Why positive sequence impedance is preferred for setting of the distance relay?
15. What do you mean by ground faults?
16. What do you mean by phase faults?
17. Derive the inputs to be given to a ground fault distance measuring unit.
18. Derive the inputs to be given to a phase fault distance measuring unit.
19. What is the difference between a distance measuring unit and a distance relay?
20. What is the typical value of the zero sequence compensation factor (L0L1)/(L1)?
21. State the relationship between LS, LM and L1, L2, L0.
22. What do you mean by transposed line?
23. What do you mean by fully transposed line?
24. What do you mean by symmetric line?
25. What is the motivation for integration between specific limits method of implementing
the DEA?
5
Algorithms Based on
Least Squared Error (LSQ)
5.1 LSQ Technique

First step in any relaying or control activity is measurement. It is a fact of life that measurement
is never perfect. It always involves errors. We can never completely eliminate errors. Thus, it
is impossible to find the true value. We have to live with errors.
Gurudev Ravindranath Tagores observation about error is very pertinent:
If you shut the door on error, truth also will be shut out
Thus, measurement is the art of keeping the baby while throwing out the bath water! We
can think of errors as gateways to measurement rather than hindrance to measurement.
While we cannot eliminate errors, we can certainly minimise them. The branch of
mathematics known as statistics provides us with very powerful methods of minimising errors.
LSQ technique is one such method of minimising the errors. It is to be noted that even though
we do not know the magnitude of the error, we have a good idea about its statistical nature.
And this is fortunate, because otherwise we would have been left groping in the dark with
a blind-fold. Knowledge of statistical nature of error neither removes the blind-fold nor the
darkness but provides us with the blind-mans stick.
The statistical nature of error that we will assume is:
random with zero mean
Gaussian
additive (and not multiplicative)
Before delving into numerical relaying algorithms based on LSQ technique, we will see
in some detail theory behind it and its relationship with average and pseudo-inverse.
The least squared error (LSQ) also called as least sum of squares of errors (LSE) can be
traced back to Gauss who first used it to predict the orbits of planets. It has been widely used
over the years and offers us a tool for estimating values when the signal is subject to noise.
87
5.2 Average is Best Value in LSQ Sense

Right from our early school days, we have been taught to trust the average of readings obtained
in an experiment more than the individual observation. Why is average so universally accepted
as the true value?
The following investigation will throw some light on the relationship between average
value and value estimated by minimising the sum of squares of errors.
In order to find value of an unknown x, let us say, two measurements were made. During
each measurement we got a different result.
Let, x = 2 first measurement
and x = 3 second measurement
Now, we are faced with the problem of estimating the true value. Of course we cannot
know the true value. We can only minimise the error. We will try to minimise the sum of
squares of errors instead of the individual errors. We acknowledge that both measurements
include additive errors. Thus, we can write the measurements as:
x = 2 + e1 e1 is the error in the first measurement
x = 3 + e2 e2 is the error in the second measurement
Which gives:
e1 = x 2
e2 = x 3
Hence, we can write the sum of the squares of the errors, denoted by SSE as:
SSE = e12 + e22 = ( x - 2)2 + ( x - 3)2
We are interested in finding the value of x which minimises the sum of the squares of
errors. Hence, this value will be that for which, the following two conditions are met:
d (SSE)
(i) = 0 and
dx
d 2 (SSE)
(ii) = positive
dx 2
Let us find the value of x for which the first derivative of SSE w.r.t. x is zero, i.e.,
d d2
(( x - 2)2 + ( x - 3)2 ) = 0 and 2
(( x - 2)2 + ( x - 3)2 ) > 0
dx dx
2( x - 2) + 2( x - 3) = 0
10 5
2 x - 4 + 2 x - 6 = 0 fi 4 x - 10 = 0 fi 4 x = 10 fi x = fi x = = 2.5
4 2
Algorithms Based on Least Squared Error (LSQ)89
d2 d
Further, (( x - 2)2 + ( x - 3)2 ) = (4 x - 10) = 4 which is > 0
dx dx
5
Hence, x = = 2.5 is the value at which the sum of the squares of errors is indeed minimum.
2
5 3+2
Now, it is easy to see that x = = = 2.5
2 2
Thus, the estimated value with minimum error that we arrived at by minimising the sum
of the squares of the errors is the same as the mean or the average value. Hence, there is
greater significance to the average value as a value which minimises the sum of squares of
errors, which is the closest that we can approach the truth.
5.3 LSQ and Pseudo-Inverse

Now, we will show how the problem illustrated above can be elegantly handled by using
pseudo-inverse method. We write the two equations x = 2 and x = 3 as matrix equation in the
following form:
1 2
1 [ x ]11 = 3
2 1 2 1
In this system of equations, we have a situation where we have more equations than the
number of unknowns. This is called an over-determined system. We cannot invert the coefficient
1
matrix since it is not square. In order to convert the coefficient matrix into square, we
12 1
pre-multiply both sides by its transpose, which is [1 1], thus we get:
1 2
[1 1] 1 [ x ] = [1 1] 3

-1
1
If we multiply both sides by [1 1] then we get:
1
-1 -1
1 1 1 2
[1 1] [1 1] 1 [ x ] = [1 1] 1 [1 1] 3
1
-1
1 2
[1][ x ] = [1 1] [1 1] 3
1
-1
1 1
Note that [1 1] = [2]-1 = = [0.5]
1 2
1 2
[ x ] = [1 1]
2 3
1 2
[ x ] = [1 1]
2 3
1 1 2
[ x] =
2 2 3
2 3
[ x] = +
2 2
[ x ] = 2 = 2.5
5

Representing the problem in symbolic fashion, we started with the matrix equation:
[A]21 [x]11 = [c]21
T
Pre-multiplying both sides by [A]
[A]T[A]21[x]11 = [A]T [c]21
Pre-multiplying both sides by {[A]T[A]21}1
{[A]T[A]21}1 {[A]T[A]21}[x]11 = {[A]T[A]21}1[A]T[c]21
[I]11[x]11 = {[A]T[A]21}1[A]T[c]21
[x]11 = [A+][c]21
where [A+] = {[A]T[A]21}1[A]T is called the pseudo-inverse of the rectangular matrix A whose
number of rows is greater than number of columns. Pseudo-inverse has properties very similar
to inverse. Some of the properties of the pseudo-inverse are given below:
(i) (P P+) P = P
(ii) (P+ P) P+ = P+
(iii) (P+ P) = (P+ P)T
(iv) P P+ = (P P+)T
Pseudo-inverse can also be determined for under-determined system as:
[A]T {[A]21[A]T}1
5.4 Relation between LSQ and Pseudo-Inverse

We present a more formal proof that pseudo-inverse technique amounts to minimising in the
LSE sense. We will use the symbolism of statistical signal processing to denote:
[Z] is a column vector of measurements
[ x ] x is the vector of estimates of true value of [x] of which [Z] is the measurement
[n] is the vector of noise terms (errors)

[H] is the matrix representing the process
The measurement process is defined by the following equation:
[ Z ]m 1 = [ H ]m n [ x ]n 1 + [ v]m 1
Here our aim is to find x such that error is minimised. We can write
[ v]m 1 = [ Z ]m 1 - [ H ]m n [ x ]n 1
Sum of squares of errors can be written as:

n1
n
SSE = (v12 ) + (v22 ) + + (vm2 ) = [n1 n m ] = [n T ][n ]
2

n m
[n T ][n ] = ([ Z ] - [ H ][ x ])T ([ Z ] - [ H ][ x ])
Noting that: - ([ Z ] - [ H ][ x ])T = ( Z T - x T H T )
[n T ][n ] = ( Z T - x T H T )([ Z ] - [ H ][ x ])
(SSE)11 = ( Z T [ Z ] - [ x ]T [ H T ][ Z ] - Z T [ H ][ x ] + [ x ]T [ H T ][ H ][ x ])
Note: [ x ]T [ H T ] [ Z ] = Z T [ H ][ x ]
SSE = ( Z T [ Z ] - 2[ x ]T [ H T ][ Z ] + [ x ]T [ H T ][ H ][ x ])
As already seen d(SSE)/d x = 0 and d2(SSE)/d( x)2 > 0 for minimising the sum of squares
of errors. Thus, differentiating the SSE and equating it to zero gives:
d (SSE)
= ( - 2[ H T ][ Z ] + 2[ H T ][ H ][ x ]) = 0
d ( x )
([ H T ][ H ][ x ]) = [ H T ] [ Z ]
([[ H T ][ H ]]-1[ H T ][ H ][ x ]) = [[ H T ][ H ]]-1[ H T ][ Z ]

-1
[ x ] = [ H T ][ H ] [ H T ][ Z ]
[ x ] = [ H ]+ [ Z ]
+ -1
where [ H ] = [[ H ][ H ]] is the pseudo-inverse of H.
T
Thus, minimising the sum of squares of errors directly involves pseudo-inverse.

d 2 (SSE)
Further, let us find out if = positive?
dx 2
d 2 (SSE)
then 2
= [ H T ][ H ]
dx
T
Now, square matrix of the type [H ][H] is always positive definite. Hence both, the first
order and second order conditions for minima are met.
Hence, we can say that the pseudo-inverse method minimises in the LSE or LSQ sense.
Thus, we are justified in calling the algorithm by Sachdev as an LSQ/LSE algorithm.
5.5 LSQ Algorithm by Sachdev [50]

In the classical curve fitting technique as shown in Figure 5.1, the measured values of the
function are given by Y0, Y1, Y2... corresponding to the independent variable values x0, x1,
x2,.... Assuming that the true values of the function f(x) are y0, y1, y2,... then the errors are:
Figure 5.1 Curve fitting using LSQ Technique.
e1 = y0 Y0 ; e1 = y1 Y1; e2 = y2 Y2. We search for a function y = f(x) which minimises

the sum of squares of errors, SSE, instead of minimising the individual errors.
SSE = ei2 = (e0 )2 + (e1 )2 + (e2 )2 +

i
Sachdevs algorithm is based on the concept of curve fitting. On the basis of apriori
knowledge, we know the nature of the waveform of the fault current but we do not know
exact values of the parameters. We know that the fault current will consist of a decaying DC
offset, a fundamental and a few harmonics. Thus, we try to fit the samples of fault current to
the model given by:
y(t) = f(t) = I0 et/t + Im1sin(w0t + q0) + Im3 sin(3w0t + q3) + Im5 sin(5w0t + q5) +
Assuming, for the sake of illustration, that the fault current consists of:
(i) a decaying DC offset
(ii) no even harmonics

(iii) harmonics up to third harmonic
We can write the expression for the current as:
y(t) = f(t) = I0 e-t/t + Im1 sin( w0t + q0) + Im3 sin(3w0t + q3)
Using the Taylors series expansion of ex, we get:
x2
e x =1 + x + +
2!
I0t I0t 2
writing (I0et/t ) using only three terms of the expansion as: I 0 e - t / t @ I 0 -
+ 2.
t 2t
Thus, the model of the fault current with which we will be trying to fit the samples of the
fault current becomes:
I0t I0t 2
+ 2 + Im1 sin(w0t + q1) + Im3 sin(3w0t + q3)
y(t) = f(t) = I 0 -
t 2t
Further we will expand the sin(w0t + q0) and sin(3w0t + q3) terms to get:
I0t I0t 2
i(t ) = I 0 - + 2 + Im1 sin(w0t) cos(q1) + Im1 cos(w0t) sin(q1)
t 2t
+ Im3 sin(3w0t) cos(q3) + Im3 cos(3w0t) sin(q3)
By Rearranging, we get:
I0t I0t 2
i(t ) = I 0 - + 2 + Im1 cos(q1) sin(w0t) + Im1 sin(q1) cos(w0t)
t 2t
+ Im3 cos(q3) sin(3w0t) + Im3 sin(q3) cos(3w0t)
Again by rearranging, we get:
i(t) = I0 + Im1 cos(q1) sin(w0t) + Im1 sin(q1) cos(w0t)
I0t I0t 2
+ Im3 cos(q3) sin(3w0t) + Im3 sin(q3) cos(3w0t) - + 2
t 2t
Let us denote i(t) since it is a sample value by s(t).
s(t) = I0 + Im1 cos(q1) sin(w0t) + Im1 sin(q1) cos(w0t)
I 0t I 0t 2
+ Im3 cos(q3) sin(3w0t) + Im3 sin(q3) cos(3w0t) - + 2
t 2t
In the above equation we have seven unknowns. Hence, we need to generate a total of
seven equations to be able to solve for the seven unknowns. This is shown below with unknown
quantities in each equation shown in bold.
S(t1) = 1. I0 + sin(w0t1) Im1 cos(q1) + cos(w0t1) Im1 sin(q1)

I I
+ sin(3w0t1) Im3 cos(q3) + cos(3w0t1) Im3 sin(q3) t1 0 + (t1)2 02
t 2t
S(t2) = 1. I0 + sin(w0t2) Im1 cos(q1) + cos(w0t2) Im1 sin(q1)

I I
+ sin(3w0t2) Im3cos(q3) + cos(3w0t2) Im3 sin(q3) t2 0 + (t2)2 02
t 2t
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
S(t7) = 1. I0 + sin(w0t7) Im1 cos(q1) + cos(w0t7)Im1 sin(q1)
I I
+ sin(3w0t7) Im3 cos(q3) + cos(3w0t7)Im3 sin(q3) t7 0 + (t7)2 02
t 2t
In each equation there are seven knowns, which can be pre-computed. For example, in the
first equation these are:
a11 = 1, a12 = sin(w0t1), a13 = cos(w0t1), a14 = sin(3w0t1), a15 = cos(3w0t1), a16 = t1 and a17 = t12
The seven unknowns are:
x1 = I0, x2 = Im1 cos(q1), x3 = Im1 sin(q1), x4 = Im3 cos(q3),
I I
x5 = Im3 sin(q3), x6 = 0 , x7 = 02
t 2t
The amplitudes and phase angles of the fundamental and the third harmonic can be found
as shown below:
I m1 = x22 + x32 I m 3 = x42 + x52

x x
q 1 = tan -1 3 q 3 = tan -1 3
x2 x4

We will simplify the representation by writing as:
s(1) = a11 x1 + a12 x2 + a13 x3 + a14 x4 + a15 x5 + a16 x6 + a17 x7
Using the matrix notation:

s(1) a11 a12 a13 a14 a15 a16 a17 x1
s(2) a21 a22 a23 a24 a25 a26 a27 x
2
s(3) a31 a32 a33 a34 a35 a36 a37 x3

s(4) = a41 a42 a43 a44 a45 a46 a47 x4
s(5) a51 a52 a53 a54 a55 a56 a57 x5

s(6) a61 a62 a63 a64 a65 a66 a67 x6
s(7) x
a71 a72 a73 a74 a75 a76 a57 7
For simplicity, we will write the equation as follows:

[S]71 = [A]77 [X]71
Since [A] matrix is square we can invert it and write:
[X]71 = [A1]77 [S]71
If we take m number of samples such that m > 7 then [A] matrix becomes rectangular and
we can use pseudo-inverse to solve for the unknowns.
[S]m1 = [A]m7[X]71
[X]71 = [A+]7m [S]m1
Figure 5.2 shows how LSQ algorithm can be used to implement any distance relay.
Figure 5.2 Implementation of LSQ algorithm.

5.6 MATLAB Implementation of LSQ Algorithm

% LSQ algorithm by Sachdev
% Algo is based on modelling of the signal.
% If we model the signal to contain
% decaying DC offset, funda and 3rd harmonic :-
% i(t)= I0 exp(-t/tau)+ Im1 sin(w0 t + theta1 ) + Im3 sin(3
w0 t + theta3)
% Unknowns are : Im1 , theta1 , Im3 , theta3 = 4 unknowns
% DC offset is modelled using three terms of expansion = 3 unknown
% Total num of unknowns in this case = 7 unknowns
% Therefore we need at least 7 sampling instants to form 7 equations
% [s] = [a ] * [ X ]
% [X ] = inv(a) * [s] if [a] is not square then
% [X ] = pinv(a) * [s]
clear;clc
% --------------------------------------------------------------
%----------------------------- Settings -----------------------
%---------------------------------------------------------------
samples = input( Enter the no. of samples = ) ;
f = 50 ;
Vm = 110 ;
I0 = 10 ;
theta1 = pi/3 ;
theta3 = pi/6 ;
R = 1 ;
XbyR = 20 ;
MaxHarmonic = 3 ;
%---------------------------------------------------------------
%------------------------ End of Settings -------------------
%---------------------------------------------------------------
fsig_max = f * MaxHarmonic ;
OverSamplingFactor = MaxHarmonic *20 ;
% 2 is the minimum value ;
fsamp = 2* fsig_max * OverSamplingFactor ;
delta_t = 1/fsamp ;
X = R * XbyR ;
Zmag = sqrt(R^2+ X^2) ;
w0 = 2*pi*f ;
L = X / w0 ;
tau = L/R ;
Im1 = Vm/ Zmag ;
Im3 = Im1/3 ;
%---------------------------------------------------------------
% NumOfUnknowns will depend upon model of signal assumed

NumOfUnknowns = 7 ;
%----- Sampling the signal to generate samples for processing
by LSQ ----
for col=1:NumOfUnknowns
for row=1:samples
t = row*delta_t;
s(row,1)= I0*exp(-t/tau)+Im1*sin(w0*t+theta1)+Im3*sin(3*w0*t+the
ta3);
a(row,1)= 1 ;
a(row,2)= sin( w0 * t);
a(row,3)= cos( w0 * t);
a(row,4)= sin(3*w0 * t);
a(row,5)= cos(3*w0 * t);
a(row,6)= t;
a(row,7)= t*t;
end
end
% --------Samples of signal generated --------
% ------- LSQ Algorithm Starts ------------
x = pinv(a)*s;
% ------- ------------------------- ------------
% x = inv(a) * s if a is 7 x 7
x2 = x(2,1);
x3 = x(3,1);
x4 = x(4,1);
x5 = x(5,1);
IIm1 = sqrt( (x2^2) + (x3^2) );

theta1_calc = atan2(x3,x2);
IIm3 = sqrt( (x4^2) + (x5^2) );
theta3_calc = atan2(x5,x4);
err_Im1 = ((IIm1-Im1) / (Im1))*100 ;
err_theta1 = ( (theta1_calc-theta1)/ ( theta1)) * 100;
fprintf( Peak Value of Funda=%f Estimated Peak of Funda=%f
\n,Im1 , IIm1);
fprintf( theta1 =%f Estimated theta1
=%f \n,theta1,theta1_calc);
fprintf( Error in Estimation of Funda Peak =%f
%% \n,err_Im1 );
fprintf( Error in Estimation of Phase Ang. of Funda =%f
%% \n,err_theta1);
fprintf( Im3=%f IIm3=%f theta3=%f theta3_calc=%f\n,Im3 ,
IIm3,theta3,theta3_calc)
plot(s)
%---------------------------------------------------------------
%---------------------- End of Program ----------------------
%--------------------------------------------------------------%
Enter the no. of samples = 9

Peak Value of Funda = 5.493138 Estimated Peak of Funda = 5.468352
theta1 = 1.047198, Estimated theta1 = 1.044800
Error in Estimation of Funda Peak = 0.451218 %
Error in Estimation of Phase Ang. of Funda = 0.228963 %
Im3 = 1.831046 IIm3 = 1.831287 theta3 = 0.523599 theta3_calc = 0.523735
>>
Review Questions
1. Show how average value is related with LSQ technique.

2. Show how pseudo-inverse is related with LSQ technique.
3. Give the geometric significance of LSQ technique.
4. Show that Fourier coefficients fit the given waveform using the Fourier series in the
LSQ sense.
5. Define left pseudo-inverse and right psudo-inverse.
6. When are we required to use left pseudo-inverse?
7. In order to apply the LSQ technique, we must have apriori knowlege about the nature
of the signal. Discuss.
8. It is known apriori that a certain fault current contains decaying dc offset, fundamental
and fifth harmonic. Develop a LSQ algorithm to extract the various frequency
components from the fault current signal.
9. What are the various variations of the LSQ technique?
10. What is total least squares?
11. What is recursive least squares?
12. How Kalman filter is related with least squares method?
13. What is weighted least squares?
6
Discrete Fourier Transform
6.1 Introduction to Concept of Spectrum

Mankind has always been fascinated by periodically repeating events like sun-rise and sun-set,
onset of various seasons, phases of the moon, tides in the sea, celestial events like eclipses,
sun-spots, and sighting of comets. Thus, curiosity about periodic events predates scientific
study of the subject.
Sir Issac Newton introduced the scientific term spectrum using the Latin word for an
image. Modern English language has the word spectre meaning ghost or apparition and
the corresponding adjective spectral. Every school-boy knows what VIBGYOR shown in
Figure 6.1 stands for in case of light and GAMUT in case of musical sounds.
Figure 6.1 Concept of spectrum.
For the sake of completeness, we list the frequencies of various visible colours (in terra-
hertz) and audible sounds (in Hz) as shown in Tables 6.1 and 6.2 respectively, which by the
way, shows that we are almost deaf as well as blind since we can perceive only a small fraction
of sound and light spectra.
99
Deaf we may be, but it can be seen from Table 6.2 that all of us, whether we are musically
trained or not, have the ability to correctly decode the ratio of two frequencies. The ratio of
frequency of adjacent (pure) notes on the piano is exactly 21/7, so that the eighth note (first
note in the next octave) is double in frequency as compared to the first note! In fact that is the
origin of the word octave (Latin for eight). This is a tribute to our ears, which over millions
of years, evolved as specialised instruments for doing real-time frequency analysis of very high
complexity. In fact our ears have evolved dedicated hardware to do frequency analysis before
sending the signal to the brain.
Thus, sound, which is information in time domain, encoded in time varying pressure wave
in air, is actually converted into a frequency domain signal by the mechanism of the ear.
What we finally perceive is not the instantaneous variations of pressure but a tone, i.e., a
frequency with a certain magnitude which we perceive as weak or strong. This is exactly the
job of an instrument called spectrum analyser. Thus, we have a spectrum analyser for sound
in the audible frequency range of 20 Hz to 20 kHz.
It is interesting to note that it is also possible to trick our ears into believing what we
want it to believe if we program the sounds correctly. We can create the illusion of a stereo
sound or surround sound by manipulating the time of arrival of sound signals. Similarly, we
can substantially reduce perception of noise in a recording by using some more tricks like
DolbyTM techniques.
Similarly our eyes perceive the time domain light signals as colours, i.e., frequencies with
their respective strengths (lightness or darkness of the colour) before passing it to the brain.
We cannot tell the instantaneous value of the light signal. Thus, our eyes constitute a spectrum
analyser working in the visible frequency range.
This is a fascinating topic but we are constrained to leave it here and concentrate on
electrical engineering!
Table 6.1 Spectrum of light

Infra-red Visible Spectrum Ultra-violet
Red Orange Yellow Green Blue Indigo Violet
< 400 THz 400 THz 484 THz 508 THz 526 THz 631 THz 651 THz 668 THz > 668 THz
700 nm 400 nm
Table 6.2 Spectrum of sound

Infra- Ultra-
Audible spectrum
sonic sonic
Sa Re Ga Ma Pa Dha Ni Sa
Do Re Mi Fa So La Ti Do
Less 470 Sa 2 1/7
Re 2 1/7
Ga 2 1/7
Ma 2 1/7
Pa 2 1/7
Dha 2 1/7
470 2 Greater
than 20 Hz Hz Hz Hz Hz Hz Hz Hz than 20
Hz kHz
Today our thinking is so much dominated by digital computers that it is difficult to imagine
Fourier analysis being done by any other method than by using digital computers. However,
Discrete Fourier Transform101
using computers for Fourier analysis is of comparatively recent origin. For tens of years,
scientists have done Fourier analysis using mechanical, acoustic or optical methods. Prior to
1965 such devices were in wide use for Fourier analysis. However, with the publication of
the FFT algorithm by Cooley and Tuckey in 1965, all other methods have been superseded by
digital computer oriented methods.
6.1.1 Generalised Fourier Series

The generalised Fourier series is based on the concept of orthogonal functions. Let
f0 (t ), f1 (t ), f2 (t ), f3 (t ),, fn (t ) be a set of mutually orthogonal functions. Orthogonality is
defined as follows:
Two functions f(t) and g(t) are orthogonal in the interval a to b if
b
f (t ) g(t ) = 0
a
with the additional property that
b b
f (t ) f (t ) = K and g(t ) g(t ) = K
a a
if K = 1 the functions are called orthonormal.

According to the theory of generalised Fourier series, we can express a given function as
a linear combination of the above orthogonal functions as follows:
f (t ) = C0f0 (t ) + C1f1 (t ) + C2f2 (t ) + C3f3 (t ) + + Cnfn (t ) + + Cf (t )
Generalised Fourier analysis consists in finding the values of the coefficients C0, C1, C2,...,
Cn. On the other hand synthesis works in opposite direction, i.e., given coefficients, we can
synthesise a time domain function as illustrated in Figure 6.2.
Figure 6.2 Basic idea of frequency analysis and synthesis.

Now, how to find the coefficients C0, C1, C2 etc.? This is where the property of orthogonality
becomes extremely useful. If we multiply both sides of above equation by f0 and integrate
between a and b, we get:
b b b b
f (t )f0 (t )dt = C0 f0 (t )dt + C1 f0 (t )f1 (t )dt + + Cn f0 (t )fn (t )dt +
2
a a a a
Because of orthogonality, all integrals of the type

b
fm (t )fn (t )dt = 0 and
a
b b
fn (t )fn (t )dt = fn (t )dt = k0 (for orthonormal functions k0 = 1).
2
a a
Thus, we get:
b
f (t )f0 (t )dt = k0 C0
a
b
1
C0 =
k0
f (t )f0 (t ) dt
a
In general,
1 b
Cn = f (t )fn (t ) dt
kn a
Thus, multiplying by fn(t) and integrating between a and b has the effect of filtering
out all the other terms except the n the term.
6.1.2 Dirichlet Conditions

The conditions under which it is possible to write Fourier series for a periodic function are
known as Dirichlet conditions.
They require that the periodic function in each period:
1. Have a finite number of discontinuities
2. Have a finite number of maxima and minima
T
3. Be absolutely convergent, i.e., f (t ) dt <
0
Fortunately, most signals encountered in practice obey Dirichlet conditions.
6.1.3 Fourier Series in Trigonometric Form

The Fourier series can be written in the trigonometric form as:
f (t ) = a0 + a1 cos(w 0 t ) + a2 cos(2 w 0 t ) + a3 cos (3 w 0 t ) + + an cos(n w 0 t )
+ b1 sin(w 0 t ) +b2 sin(2 w 0 t ) + b3 sin(3 w 0 t ) + + bn sin(n w 0 t )
This can also be written in compact notation as:

f (t ) = a0 + ( an cos(n w 0 t ) + bn sin(n w 0 t ))
n =1
Note that the above form is phase angle free. The phase angle is not explicitly stated but
is encoded in the ratio of bn and an. Another form which is sometimes used is a cosine wave
with phase angle, explicitly stated. To develop that form consider:
Cn cos(n w 0 t + q n )
i.e., Cn cos(n w 0 t + q n ) = Cn cos(n w 0 t ) cos(q n ) - Cn sin(n w 0 t )sin(q n )
or, Cn cos(n w 0 t + q n ) = Cn cos(q n ) cos(n w 0 t ) - Cn sin(q n )sin(nw 0 t )
Let an = Cn cos(q n ) and bn = -Cn sin(q n )
b
Let C0 = a0 , then Cn = (an2 ) + (bn2 ) and q n = - tan -1 n (Note: - ve sign)
an
Then, Cn cos(n w 0 t + q n ) = an cos(n w 0 t ) + bn sin(n w 0 t )
Hence, alternate representation of the Fourier series will be:

f (t ) = C0 + Cn cos(n w 0 t + q n )
n=0
The coefficients of the Fourier series can be found by evaluating the following integrals:
1T
0 T f (t )dt
a = is the average value of the function over a cycle
0
2T
an = f (t )cos(n w 0 t )dt
T0
2T
bn = f (t )sin(n w 0 t )dt
T0
The above simple expressions for the coefficient become possible because of the orthogonal
nature of the following set of functions, which form the basis functions of Fourier series:
{1, cos(w0t), cos(2w0t), cos(3w0t),, cos(nw0t), sin(w0t), sin(2w0t), sin(3w0t),, sin(nw0t)}
The orthogonal relationship between the basis functions is captured in Table 6.3.
Table 6.3 Orthogonal relationship between basis functions
1 dt sin(mw0t)dt sin(nw0t)dt cos(mw0t)dt cos(nw0t)dt

T
T
sin(mw 0t ) zero
2
=p zero zero zero
0
T
T
sin(nw 0t ) zero zero
2
=p zero zero
0
T
T
cos(m w 0t ) zero zero zero
2
=p zero
0
T
T
cos(nw 0t ) zero zero zero zero
2
=p
0
6.1.4 Fourier Series in Exponential Form

We have the trigonometric form of Fourier series as:
f (t ) = a0 + a1 cos (w 0 t ) + a2 cos (2 w 0 t ) + a3 cos (3 w 0 t ) + + an cos (n w 0 t )

+ b1 sin (w 0 t ) + b2 sin (2 w 0 t ) + b3 sin (3w 0 t ) + + bn sin(n w 0 t )

f (t ) = a0 + (an cos(n w 0 t ) + bn sin(n w 0 t ))
n =1
By De Movires theorem we have:
1 1
cos(nw 0 t ) = (e jnw 0 t + e - jnw 0 t ) and sin(nw 0 t ) = (e jnw 0 t - e - jnw 0 t )
2 2j
Hence, we can write f(t) as:
(e jnw 0 t + e - jnw 0 t ) (e jnw 0 t - e - jnw 0 t )

f (t ) = a0 + an + bn

n =1 2 2j
a - jb
n jnw 0 t a + jbn - jnw 0 t
f (t ) = a0 + n + n
2
e e
n =1 2
an - jbn a + jbn
Let a0 = C0 , = Cn~ and n = C-~n
2 2
1 2 1
Cn~ = an + bn2 = Cn
2 2
where Cn is the coefficient in the cosine-phase angle form of Fourier series. f(t) can now be
expressed in compact form

f (t ) = [Cn~ e jnw t ]
0
n =-
Where the coefficients Cn~ are obtained by:

1T 1T
Cn~ =
T0
f (t )cos(nw 0 t )dt - j f (t )sin(nw 0 t )dt
T0
1T
Cn~ = f (t ) [cos(nw 0 t ) - jsin(nw 0 t )] dt
T0
1T - jnw t
Cn~ = f (t )e 0 dt
T0
6.2 Fourier Coefficients are Best in LSQ Sense

Fourier analysis can be visualised in terms of curve fitting. Given an arbitrary waveform,
we are trying to find which sine and cosine curves can be used to fit the given curve. The
Fourier coefficients encode the magnitudes and phases of these sines and cosines.
We have the Fourier series:

n =1
Fourier analysis is a search for the values of and bn coefficients which fit the given
curve with the help of sines and cosines. We should try to find the values of an and bn which
are optimal in the LSQ sense. Assuming N number of terms in the series, we can express
the error as:
N
error = f (t ) - a0 + (an cos(n w 0 t ) + bn sin(n w 0 t ))
n =1
Sum of squared errors can be written as:
T 2
N
SSE = f (t ) - a0 + (an cos(n w 0 t ) + bn sin(n w 0 t )) dt
0 n =1
For minimising the SSE, we impose the conditions that:
d (SSE) d (SSE) d (SSE)

= 0; = 0; and =0
da0 dan dbn
T
d (SSE) N
= 0 = 2 f (t ) - a0 + (an cos(n w 0 t ) + bn sin(n w 0 t )) dt ( -1)
da0 0 n =1
T T T N
0 = 2 f (t )dt - 2 a0 dt - 2 (an cos(n w 0 t ) + bn sin(n w 0 t ))dt
0 0 0 n =1
Note that each of the term within the summation sign integrates to zero. Hence, we are
left with:
T T
0 = 2 f (t )dt - 2 a0 dt
0 0
T T
0 = f (t )dt - a0 dt
0 0
T T
a0 dt = f (t )dt
0 0
T
a0 T = f (t )dt
0
1T
a0 = f (t )dt
T0
.... which is the same expression derived earlier. Hence, a0 is best in
the LSQ sense.
To find the best value of an: d SSE = 0

d an
T
d (SSE) N
= 0 = 2 f (t ) - a0 + ( an cos(n w 0 t ) + bn sin(n w 0 t )) (cos(nw 0 t ))dt
dan 0 n =1
T T T
0 = f (t )cos(nw 0 t )dt - a0 cos(nw 0 t )dt - an (cos(n w 0 t ))2 dt
0 0 0

T N N T
- bn sin(n w 0 t )cos(n w 0 t ) dt - bn cos(m w 0 t )cos(n w 0 t ) dt
0 n =1 0 m =1, m n
All these integrals evaluate to zero:

T
a0 cos(nw 0 t )dt = 0
0
T
T
N N
n b sin( n w 0 t )cos( n w 0
t ) dt = bn cos(m w 0 t )cos(n w 0 t ) dt = 0
0 n =1 0 m =1, m n
T
T
and (cos(n w 0 t )) dt =
2
; thus we are left with:
0 2
T
T
0 = f (t )cos(nw 0 t )dt - an
0 2
2T
an = f (t )cos(nw 0 t )dt
T0
which is the same expression as obtained earlier.
Similar analysis gives:

2T
bn = f (t )sin(nw 0 t )dt
T0

which is the same expression as obtained earlier. Thus an and bn are indeed optimal in LSQ
sense. Thus, if we do curve fitting with the restriction that only sines and cosines are to be
used to fit the curve, we again arrive at Fourier coefficients!
6.3 Summary of Fourier Series
1. Trigonometric Form of Fourier Series (Phase Angle Free Form)

n =1
1T 2T 2T
a0 = f (t )dt
T0
an = f (t )cos(n w 0 t )dt
T0
T0
2. Trigonometric Form of Fourier Series (Phase Angle Free Form)
a0
f (t ) = + (an cos(n w 0 t ) + bn sin(n w 0 t ))
2 n =1
2T 2T
an = f (t )cos(nw 0 t )dt
T0
T0
n = 1 to
3. Trigonometric Form of Fourier Series with Cosine Wave and Phase Angle

f (t ) = C0 + Cn cos(n w 0 t + q n )
n=0
b
q n = - tan -1 n
C0 = a0 Cn = (an2 ) + (bn2 ) an
(Note: ve sign)
4. Exponential Form of Fourier Series

f (t ) = [Cn~ e
jnw 0 t
] Complex Coeff. Cn~ = Cn~ e jq n
n =-
b
q n = - tan -1 n (Note: ve sign)
an
an - jbn an + jbn
a0 = C0 = Cn~ = C-~n
2 2
1T - jnw t
Cn~ = f (t )e 0 dt
T0
6.4 Applications of Fourier Analysis
Fourier analysis is useful in a very large number of fields. In fact the double helix form
of DNA was discovered in 1962 through X-ray diffraction techniques and Fourier analysis.
NASA improves the clarity and detail of picture of celestial objects taken in space by means
of Fourier analysis. Fourier analysis has found application in plasma physics, semiconductor
physics, microwave acoustics, seismography, oceanography, radar mapping, medical imaging,
chemical analysis and biology to name a few areas.
6.5 Fourier Series for a Recurring Pulse

Figure 6.3 shows a recurring or periodic pulse with a time period of T second and amplitude
of V0. For the sake of simplicity, pulse is taken symmetric about the t = 0 point. The width of
the pulse is a second. We will use this pulse as a vehicle for transition from Fourier series
to Fourier transform.
Figure 6.3 Recurring or periodic pulse.

The pulse is defined by:

a a
f (t ) = V0 ; <t <
2 2
T a
= 0; <t <-
2 2
a T
= 0; <t <
2 2
Figure 6.4 The sinc( ) function drawn using MATLAB ezplot(sin(x)/(x) ).
First let us find the Fourier series representation for the pulse and plot its amplitude and
phase spectrum. The complex Fourier coefficients Cn~ are given by:
T
Cn~ = f (t ) e - jnw 0 t dt
0
a/2
V0
Cn~ = e - jnw 0 t dt
T -a / 2
a/2
V e - jnw 0 t V0 e - jnw 0 a / 2 - e jnw 0 a / 2
Cn~ = 0 =
T - jnw 0 - a / 2 T ( - jnw 0 ) 1
1 V0 2 e jnw 0 a / 2 - e - jnw 0 a / 2
Cn~ =
T n 2 p f0 1 2j
Figure 6.5 The sinc() function shown in more detail.
V0 e jnw 0 a / 2 - e - jnw 0 a / 2 w a
Cn~ = ; multiplying and dividing by n 0 , we get:
np 2j 2
V0 n w 0 a sin(nw 0 a/2)
Cn~ =
np 1 2 (nw 0 a/2)
V0 a sin(nw 0 a/2)
Cn~ =
T (nw 0 a/2)
V0 a sin( x ) nw 0 a
Cn~ = where x =
T ( x) 2
V0 a
Cn~ = sinc( x )
T
sin( x ) sin( x )
where sinc( x ) is defined as . It can be shown that the value of is 1 when x = 0.
( x) ( x)
The c in sinc() comes from the Latin cardinalis. The full name of the sinc() function
is sinus-cardinalis. The plot of sinc() function using MATLAB ezplot() function is shown in
Figure 6.4. In Figure 6.5 plot of Cn~ and magnitude and phase spectrum of sinc() function are shown.
6.6 Recurring Pulse to Non-recurring Pulse

Because of the constraints imposed by finite computing power we cannot process infinite
duration or (infinitely) recurring signals. The signals that we process, thus, can be said to
exist during a finite window of time. We are forced to ignore everything outside this window.
Thus, even if we are processing recurring signals they appear to be non-recurring! Since the
Fourier theory is valid for infinitely recurring signals, strictly speaking, we cannot apply it to
any of power system signals. However, Fourier technique is too valuable a tool to let go of.
What, then, should we do?
We modify the Fourier theory itself so that it can be applied to non-recurring signals. We
do this by using a simple mathematical trick. Consider what will happen as we stretch the time
period T to infinity while keeping the on time a constant as shown in Figure 6.6. Under
such a situation there will be no more instances of the pulse and we are left with a single,
non-recurring pulse of duration a.
Figure 6.6 Transition from recurring to non-recurring pulse as T .
It can be shown that as T , the discrete frequency spectrum remains the same in shape
sin(nw 0 a/2)
but assumes the envelope of the spectrum of the recurring pulse. The envelope of
sin(w a/2) (nw 0 a/2)
becomes . Note that in the first expression we have w0, a discrete frequency
(w a/2)
whereas in the second expression we have w the continuous frequency variable.
The mathematical treatment of the non-recurring pulse can thus be derived from the
recurring pulse by letting the time period T approach in the limit. The changes that take place
in the representation as we transit from recurring to non-recurring pulse are listed in Table 6.4.
Table 6.4
Recurring pulse Single pulse
Time period Time period T is finite Time period T
Frequency Fundamental frequency is (1/T) Fundamental frequency ffunda is zero
Nature of Spectrum is discrete (spaced at ffunda Spectrum is continuous (spaced at zero Hz
frequency from each other) from each other)
spectrum
Summation/ Fourier sum Fourier integral

Integration 1
f (t ) = Cn~e jw n t
0
f (t ) = jw t
F ( jw )e dt
n =- 2p -

1T - jnw t
Coefficients Cn~ = f (t )e 0 dt F ( jw ) = f (t )e - jw t dt
T0 -
Last row of Table 6.4 shows that the difference between the frequency spectrum of the
recurring and non-recurring versions of a pulse is not in the shape. Both versions have the
same shape. The difference is that the recurring pulse has discrete (discontinuous) frequency
spectrum while the non-recurring pulse has continuous frequency spectrum. Thus, we can make
the following observations:
1. The shape of the continuous amplitude and phase spectra for a non-recurring f(t) is
identical with the envelopes of the amplitude and phase line spectra for the same pulse
recurring.
2. All frequencies are present in the continuous amplitude spectrum since F(jw) is defined
for all w from to +.
3. Lightning, which is a very short duration pulse in time, makes its presence felt
irrespective of the frequency of the radio station to which we may be tuned. (We do
not hear the lightening crashes in our mobile phones because, we are not listening to
the raw audio signal, the phone processes the audio signal to filter out noise using
DSP techniques.)
We draw an important inference from the above discussion that in case of non-recurring
pulse, we have to use Fourier integral instead of the Fourier summation. Fourier integral is also
popularly known as Fourier transform. Note that the Fourier transform that has been described
is a continuous transform. As can be expected, this continuous transform will have to give way
to discrete transform. But this will be taken up in the next section.
6.7 Discrete Fourier Transform Development [5,13,19,22,33,52,61]

We have the exponential form of the Fourier series

f (t ) = [Cn~ e jnw t ]
0
n =-
where the coefficients Cn~ are obtained by:
1T 1T
Cn~ =
T0
f (t )cos( nw 0 t ) dt - j f (t )sin(nw 0 t ) dt
T0
1T
Cn~ = f (t )[cos(nw 0 t ) - jsin(nw 0 t )] dt
T0
1T - jnw t
Cn~ = f (t )e 0 dt
T0
Table 6.5
Fourier sum (Recurring signal) Fourier integral (Windowed signal)

1
0
f (t ) = F ( jw )e
jw t
dt
n =- 2p -
Fourier Coefficients Fourier Coefficients

1T - jnw t
Cn~ = f (t )e 0 dt F ( jw ) = f (t )e - jw t dt
T0 -
We develop the DFT by noting the expression for F(jw) for windowed signal:

F ( jw ) = f (t )e - jw t dt
-
Note that w refers to angular frequency of the signal in the continuous domain, i.e., wsignal.
wsignal = 2pfsignal, where fsignal is the signal frequency in the continuous domain. Note that time
t can take only discrete values t = nDt; where Dt = 1/fsampling in the discrete domain and fsignal
can take only discrete values corresponding to mffunda. Thus changing over to the discrete
domain where X(m) represents the Fourier coefficient at the mth harmonic frequency, we can
write the discrete version of the Fourier integral as:
N -1
- j 2p fsignal nDt
X ( m) = x (n)e
n=0
- j 2p m ffunda n
N -1
x (n)e
fsampling
X ( m) =
n=0
Since we are taking N samples, it implies that fsampling = N ffunda

- j 2p m ffunda n
N -1
X ( m) = x (n)e
Nffunda
n=0
Thus, we get the expression for DFT as:

N -1 - j 2p mn
X ( m) = x (n)e N
n=0
We summarise the transition from CFT to windowed DFT as shown in Table 6.6.
Table 6.6 Transition of windowed DFT from CFT

Fourier sum Fourier Integral Discrete Fourier Inverse
(recurring continuous signal) (windowed continuous signal) (windowed sampled signal)

j 2p mn
0
1 jw t 1 N -1
n =-
f (t ) =
2p
F ( jw )e dt x ( n) = X ( m)e N
- N m =0
Fourier coefficients Fourier coefficients Fourier coefficients
- j 2p mn
1T - jnw t
N -1
Cn~ = f (t )e 0 dt F ( jw ) = f (t )e - jw t dt X ( m) = x ( n) e N
T0 - n=0
In the DFT expression, the following points are worth noting:

1. N = Number of samples
2. m = index for frequency domain sample, i.e., the order of the harmonic
Both m and n run from 0 to (N1) and n = index for time domain sample
3. The multiplier 1/N is included in IDFT but not in DFT definition
4. X(m) is the frequency domain variable, the complex Fourier coefficient, representing
the frequency content at the mth harmonic frequency
5. x(n) is the time domain variable, nth sample of the signal
Note, however this does not mean that maximum order of harmonic in N1. The highest
order of harmonic is limited to N/2. Harmonics from (N/2) + 1 to N1 are in fact complex
conjugate of harmonics from (N/2)1 to 1.
6.8 Implicit Assumption behind Windowed Fourier Transform

Implicit assumption in DFT is that there are infinite repetitions of the window, both, before
as well as after the window. Fourier analysis cannot but represent the signal in any other
form. Thus, if we collect eight signal samples of 1,1,1,1,1,1,1,1 in the window, then implicit
assumption causes DFT to add infinite repetitions of the window, both before and after the
window, as shown in Figures 6.7 and 6.8.
In Figure 6.7, the implicit assumption causes DFT to infer that the signal is a DC signal.
This is represented on the frequency spectrum as a single line at a frequency of zero hertz.
In Figure 6.8, the implicit assumption causes DFT to represent the signal as a sinc()
function in the frequency domain.
Figure 6.7 Fourier interpretation of windowed signal.
Figure 6.8 Another example of Fourier interpretation of windowed signal.
6.9 Meaning of N Samples Collected in a Window

Let us say, we have a window in which (N = 12), samples of a signal sampled at 600 Hz have
been collected. This statement contains a lot of information! As per the implicit assumption in
DFT, it automatically means that these 12 samples constitute a full cycle of the signal. Hence,
time period of the signal, Tsignal can be written as:
1 12 1
Tsignal = N Tsampling = N Dt = N = = s
fsampling 600 50
Hence, frequency of the signal fsignal = 1/Tsignal = 50 Hz.
It is worth noting that Fourier analysis assumes that we know the frequency of the signal
apriori because otherwise we would not have been able to decide the sampling frequency.
The primary purpose of Fourier analysis is not frequency measurement. Its purpose is to
estimate magnitudes of various frequency components. The frequencies of these components
automatically are tied up with the sampling frequency and therefore known beforehand (apriori).
Thus we can write:
1 1 fsampling
Tfunda = N Tsampling =N fsampling = Nffunda ffunda =
ffunda fsampling N
This is illustrated in Figure 6.9.
Figure 6.9 Meaning of collecting 12 samples in the window.
6.10 Redundancy of DFT

All the N number of DFT terms are not unique! Only half the terms are unique. The second
half of the DFT terms are complex conjugates of corresponding first half of the terms. This
will become clear when we try to find the relationship between X(m) and X(N m).
We have:
N -1 - j 2p mn
X ( m) = x (n)e N
n=0
Hence, we can write X(N m) as:

N -1 - j 2p ( N - m ) n
X ( N - m) = x ( n) e N
n=0
N -1 - j 2p ( N ) n j 2p m n
X ( N - m) = x ( n) e N e N
n=0
N -1 + j 2p m n
X ( N - m) = x ( n) e - j 2 p n e N
n=0
Note that e - j 2p n = 1 for n = 0, 1, 2, 3,

Hence,
N -1 + j 2 p mn
X ( N - m) = x ( n) e N
n=0
Thus, the only difference between X(m) and X(N m) is the sign of the exponents, which
is reversed. Hence, X(N m) terms are complex conjugate of X(m) terms. Thus, the magnitudes
of X(N m) and X(m) will be identical but their phase angles will be negative of each other.
X(N m) = {X(m)}* or {X(N m)}* = X(m) We are assuming here that the signal x(n) is real.
Figure 6.10 Spectrum of DFT.
This is depicted in Figure 6.10 as well as in Table 6.7.

Table 6.7
For N = 8; assuming that the signal x(n) is real.
m=0 m=1 m=2 m=3 m=4 m=5 m=6 m=7
Nm=8 Nm=7 Nm=6 Nm=5 N m =4 Nm=3 Nm=2 Nm=1
X(0) = X(1) = X(2) = X(3) = X(4) = X(5) = X(6) = X(7) =
X(8)* X(7)* X(6)* X(5)* X(4)* X(3)* X(2)* X(1)*
Note: X(8) does not exist and X(0) is always real. Since X(4)* = X(4), X(4) is also real.
Remaining Fourier coefficients are complex.
6.11 DFT as a Mapping

Mathematicians sometimes look at DFT purely as a transformation of one vector consisting of
real samples of signal into another vector consisting of the complex values of the coefficients.
Thus, we have N real numbers at the input of this process and we get N complex numbers
at the output of this process as shown in Figure 6.11 for N = 8.
Figure 6.11 DFT as a mapping of N real numbers into N complex numbers.
6.12 Numerical Calculation of DFT Coefficients

Example 6.1 A power system signal with nominal frequency of 50 Hz was sampled at
600 Hz. The following 12 samples were obtained.
Sample no. 0 1 2 3 4 5
Value 253.000 275.2891 161.2628 77.0000 74.0526 65.7109
Sample no. 6 7 8 9 10 11
Value 33.0000 31.9212 35.9474 11.0000 29.2628 89.0788
Find the values of:

1. the DC offset,
2. magnitude and phase angle of fundamental, and
3. magnitude and phase angle of second harmonic component.
Solution We shall demonstrate the solution of the above numerical problem by (a) hand
calculation, (b) through a spreadsheet and (c) by using a MATLAB program and by executing
fft(), the MATLAB function for DFT .
(a) Computation of DFT by Hand

In the expression for DFT; putting N = 12 and running n from 0 to N 1:
n = N -1
m = N -1
m =0 X ( m) = x (n)e - j 2 p mn/N
n=0
(i) For DC component m = 0

n =11 n
- j 2p 0
X (0) = x ( n) e 12 = x (0) + x (1) + x (2) + x (3) + x (4) + x (5) + x (6) + x (7) + x (8)
n=0
+ x (9) + x (10) + x (11)

X(0) = 2 53.000 + 275.2891 + 161.2628 + 77.0000 + 74.0526 + 65.7109 + 33.0000
+ 31.9212 + 35.9474 11.0000 29.2628 + 89.0788 = 1056
Note: The DFT formula does not include the multipliers of 1/N for DC term
and 2/N for other terms. This is to speed-up the DFT calculation by avoiding
an extra division. The additional division is time consuming and hence avoided.
The result is not wrong because we can always factor in the 1/N and 2/N terms.
Sometimes these may also get cancelled out during division. For example, if
we are finding impedance then 2/N factor in voltage and 2/N factor in current
will get cancelled.
Actual value of DC offset is (1/N)* X(0) = (1/12)*(1056) = 88 (Answer)

(ii) For fundamental m = 1 and n = 0 to 11 in the general expression for DFT:
n =11 n =11 n =11 p

- j ( )n
X (1) = x (n)e - j 2 p 1n /12
= x (n)e - j p n/6
= x ( n) e 6
n=0 n=0 n=0

n =11 p p p p p
- j n - j 0 - j 1 - j 2 - j 3
6 6 6 6 6
X (1) = x (n)e = x (0)e + x (1)e + x (2)e + x (3)e
n=0
p p p p p p
- j 4 - j 5 - j 6 - j 7 - j 8 - j 9
6 6 6 6 6 6
+ x (4)e + x (5)e + x (6)e + x (7)e + x (8)e + x (9)e
p p
- j 10 - j 11
6 6
+ x (10)e + x (11)e
On the Casio fx-991ES and similar calculators, we can easily evaluate the above
equation by noting that re jq = rq . Hence, we can write the expression for X(1) as
follows:
n =11 p
- j ( )n -1.p -2.p -3.p
X (1) = x ( n) e 6 = x (0) + x (1)
6 + x (2)
6 + x (3)
6
n=0
-4.p -5.p -6.p -7.p -8.p

+ x (4) + x (5) + x (6) + x (7) + x (8)
6 6 6 6 6
-9.p -10.p -11.p
+ x (9) + x (10) + x (11)
6 6 6

Converting radians to degrees for convenience:
n =11 p
- j n
6
X (1) = x (n)e = x (0) + x (1)- 30 + x (2)- 60 + x (3)- 90 + x (4)- 120
n=0
+ x (5)- 150 + x (6)- 180 + x (7)- 210 + x (8)-240 + x (9)-270 + x (10 )

- 300 + x (11)-330

n =11 p
- j n
6
X (1) = x (n)e = 253 + 275.2891- 30 + 161.2628- 60 + 77.0000- 90
n=0
+74.0526- 120 + 65.7109- 150 + 33.0000- 180 + 31.9212- 210 + 35.9474

- 240 -11.0000 - 270 - 29.2628- 300 + 89.0788- 330
The steps to calculate the above for CASIO FX-991ES calculator are as follows:
1. Press [Shift] + [Mode/Setup] and then [3] for degrees
2. Press [Mode/Setup] and then [2] for complex mode
3. Press [253] + [275.2891] [Shift][()] when you do this appears on the LCD
then enter 30
4. The LCD shows the following string for steps up till now:
253 + 275.2891 30 press [+]
5. Go on entering other terms in a similar fashion finally pressing [=]
6. If all is well you will see the answer on the LCD as:
461.999789 396.0000809i
2 2
C1 = (461.999789 396.0000809 i) = (a1 - jb1 )
N 12
Hence,
a1 = (461.999789/6) = 76.99996483 77 (Answer)
and b1 = (396.000809/6) = 66.00001348 66 (Answer)
(iii) To find a2 and b2 of 2nd harmonic for m = 2 and n = 0 to 11 in the general expression
for DFT:
n =11 n =11 n =11 p
- j n
3
X (2) = x (n)e- j 2p 2 n /12 = x (n)e- jp n / 3 = x ( n) e
n=0 n=0 n=0
n =11 p p p p p
- j n - j 0 - j 1 - j 2 - j 3
3 3 3 3 3
X (2) = x ( n) e = x (0)e + x (1)e + x (2)e + x (3)e
n=0
p p p p p
- j 4 - j 5 - j 6 - j 7 - j 8
3 3 3 3 3
+ x (4)e + x (5)e + x (6)e + x (7)e + x (8)e
p p p
- j 9 - j 10 - j 11
3 3 3
+ x (9)e + x (10)e + x (11)e

n =11 p
- j n -1.p -2.p -3.p
x (n) e
3
X (2) = = x (0) + x (1) + x (2) + x (3)
n =0 3 3 3
-4.p -5.p -6.p -7.p -8.p

+ x (4) + x (5) + x (6) + x (7) + x (8)
3 3 3 3 3
-9.p -10.p -11.p

+ x (9) + x (10) + x (11)
3 3 3

n =11 p
- j n
3
X (2) = x ( n) e = 253 + 275.2891- 60 + 161.2628- 120 + 77.0000- 180
n=0
+ 74.0526- 240 + 65.7109- 300 + 33.0000- 360 + 31.9212- 420 + 35.9474
- 480 - 11.0000- 540 - 29.2628- 600 + 89.0788- 660
Result on the calculator will be:

X(2) = 330 264.00005011i
2 2
C2 = (330 - 264.00005011i) = (a2 - jb2 )
N 12
Hence,
330
a2 = = 55 (Answer)
6
264.00005011
b2 = = 44 (Answer)
6
Let us tabulate the results:
DC Fundamental Second harmonic

a0 = 88 a1 = 77 b1= 66 a2 = 55 b2 = 44
|c1| = (a1 )2 + (b1 )2 |c2| = (a2 )2 + (b2 )2
|c1| = (77)2 + (66)2 |c2| = (55)2 + (44)2

|c1| = 101.414999 |c2| = 70.43436661
1
q1 = tan (b1/a1) q2 = tan1(b2/a2)
q1 = tan1(66/77) q2 = tan1(44/55)
q1 = 40.60129465 q2 = 38.65980825
(b) A spreadsheet for DFT calculation

The samples of the signal given in Example 6.1 are entered in an Excel spreadsheet. The first
three harmonics are computed in the spreadsheet as shown in Figure 6.12. It can be seen that
the results of the spreadsheet match with hand calculation.
The answers obtained from the spreadsheet are shown in the tabular form:
DC a1 b1 a2 b2 a3 b3
88 77 66 55 44 33 22
(c) MATLAB program for DFT calculation

% Numerical computatio of DFT in MATLAB
clear , clc
fsig = 50 ;
N = 12 ;
fs = N * fsig ;
dt = 1/fs ;
% -----------------------------------------------------------
%---------- Samples of signal ------------------------------
%------ x(o) to x(11) are written as x(1) to x(12) since
matrix /vector index of zero
% ----- is not allowed in MATLAB ----------------------------
x(1)= 253.0000 ; x(2) = 275.2891 ; x(3) = 161.2628;
x(4)= 77.0000 ; x(5) = 74.0526 ; x(6) = 65.7109 ;

Figure 6.12 12 point DFT computation using Excel spreadsheet.

x(7)= 33.0000 ; x(8) = 31.9212 ; x(9) = 35.9474 ;

x(10)= -11.0000 ; x(11) = -29.2628 ; x(12)= 89.0788 ;
%---------------- DFT calculation ----------------------------
% Initialization of coeffs. to zero before start of
calculations
for m = 0:N-1
X(m+1)= 0 ;
end
%------------- Actual Caluculation of DFT -------------------
for m = 0:N-1
for n=0:N-1
X(m+1) = X(m+1) + x(n+1) * exp(-j*2*pi*m*n/N);
end
end
% -----------------------------------------------------------
% Check calculated values against MATLABs FFT function
X_FFT = fft(x);
for m = 0:N-1
DFT_Error(m+1) = abs(X(m+1) - X_FFT(m+1));
end
max_error = max(DFT_Error)
fprintf(Calculated values of DFT\n)
for m = 0:N-1
fprintf( X(%d) =%f +j %f \n,m+1,real(X(m+1)), imag(X(m+1)));
fprintf( X_FFT(%d)=%f +j %f \n\n,m+1,real(X_FFT(m+1)),imag(X_
FFT(m+1)));
end
% Check using MATLABs built in FFT() function
a0_calc = X_FFT(1)/N ;
a1_calc = real(X_FFT(2)*(2/N)) ;
b1_calc = -imag(X_FFT(2)*(2/N)) ;
fprintf( a0_calc=%f \n,a0_calc )
fprintf( b1_calc=%f \n,b1_calc )
fprintf( a3_calc=%f \n,a3_calc)
---- Output of the program ---------------------
max_error = 5.8524e-013
Calculated values of DFT

X(1) = 1056.000000 + j 0.000000
X_FFT(1) =1056.000000 + j 0.000000
X(2) = 461.999979 + j -396.000081

X_FFT(2) = 461.999979 + j -396.000081
X(3) = 330.000000 + j -264.000050

X_FFT(3) =330.000000 + j -264.000050
X(4) = 198.000000 +j -132.000000

X_FFT(4) = 198.000000 +j -132.000000
X(5) =-0.000000 +j -0.000173

X_FFT(5)=0.000000 +j -0.000173
X(6) =0.000021 +j 0.000081

X_FFT(6)=0.000021 +j 0.000081
a0_calc = 88.000000
a1_calc = 76.999996
b1_calc = 66.000013
a2_calc = 55.000000
b2_calc = 44.000008
a3_calc=33.000000
b3_calc=22.000000
6.13 Sliding DFT Algorithm [35,41]

Computing DFT is not an one-time affair. We have to continuously compute DFT as samples
keep streaming-in. Fortunately, considerable amount of computational burden can be saved
since DFT calculation during any window is related to DFT, already computed for the previous
window. This becomes possible because, in the two consecutive windows, we are operating
on two sets of samples which are almost same, except for the newest sample that we include
and the oldest sample that we discard. We will modify the expression for DFT to account for
the fact that we have to implement the DFT in a continuous fashion over contagious windows.
The modified expression for DFT during kth window is as follows:
n = N + k -1 - j 2 p (n-k )m
X k ( m) = x (n)e N
n=k
where k is the window number and m is the order of the DFT; m = 0 DC, m = 1
fundamental etc.
Let us run through a few values of the indices to verify that the basic DFT expression
remains same.
Let N = 8
window k n = k to (N + k1) (nk)
0 07 07
1 18 07
2 29 07
3 3 10 07
4 4 11 07
Figure 6.13 Sliding DFT.
DFT during the (k)th window can be written as:

n = N + k -1 - j 2 p (n-k )m
X k ( m) = x ( n) e N
n=k
DFT for the (k + 1)th window can then, be written as:

n= N +k - j 2 p ( n - k -1) m
X k +1 (m) = x (n)e N
n = k +1
If we start the above summation from n = k, instead of n = k + 1, we will get an extra term:
- j 2p ( -1) m j 2p m
x ( k )e N = x ( k )e N
Thus, if we modify the lower limit of DFT for (k + 1)th window to (n = k) then we will
j 2p m
get the above extra term, hence we can keep the sum unaltered by subtracting x (k )e N
from
the sum with lower limit (n = k) instead of (n = k + 1). Hence, we can write it as:
n = N + k - j 2p ( n - k -1) m j 2p m
X k +1 (m) = x (n)e - N
N
x ( k ) e
n = k
If we take out the last term of the above summation corresponding to n = (N + k), we will
have to reduce the upper index of the summation to (N + k 1) and the term corresponding
to n = (N + k) will be:
- j 2p ( N + k - k -1) m - j 2p ( N -1) m
x ( N + k )e N = x ( N + k )e N
- j 2p ( N -1) m - j 2 p ( N ) m j 2p m
x ( N + k )e N = x ( N + k )e N e N
j 2p m
- j 2p m
= x( N + k ) e e N
j 2p m
= x(N + k ) e N since e - j 2p m is equal to 1 for all values of m.
n = N + k -1 - j 2 p ( n - k -1) m j 2p m j 2p m
N - x ( k )e N
X k +1 (m) = x (n)e N
+ x ( N + k ) e
n = k
n = N + k -1 - j 2 p ( n - k ) m j 2p m j 2p m j 2p m

X k +1 (m) = x (n)e N e N
+ x( N + k ) e N - x ( k )e N
n = k
j 2p m n = N + k -1 - j 2p ( n - k ) m
X k +1 (m) =e N x ( n) e N
+ x ( N + k ) - x (k )
n = k
j 2p m

X k +1 (m) = e N
{X k (m) + x( N + k ) - x(k )}
j 2p m

DFT(m) k +1 = e N
{DFT(m)k + [Newest sample - Oldest sample ]}
Consider the sliding DFT computation of an undistorted sine wave shown in Figure 6.14.
In the first window, the samples from x(0) to x(7) will be used for computing DFT(1)1. Next
we slide the widow by one sample discarding x(0) and including x(8). Let us find how the
DFT(1)2 is related with DFT(1)1.
j 2p 1

DFT(1)2 = e 8
{DFT(1)1 + [ x(8) - x(0)]}
Figure 6.14 Sliding DFT computation.
However, in this case, since x(8) = x(0), magnitude of DFT remains unaffected but there
is an apparent phase shift of p/4 radians or 45 as the window slides along even if the signal
is stationary. Thus:
jp
DFT(1)2 = e4 {DFT(1)1 }
This is an unwanted artefact and provides motivation for modifying the DFT further as
described in the next section.
6.13.1 Modified DFT

The recursive DFT expression shows that with every new computation of the sliding window, a
phase shift of j2pm/N is added to the phasor. Thus, there is a strong motivation for modifying
the DFT to get rid of this apparent phase shift. The DFT can be modified as follows:
Computation of the mth order DFT during the (k + 1)th window is given by:
j 2p m
{X k (m) + x( N + k ) - x(k )}
X k +1 ( m ) = e N
Consider the computation of the fundamental, i.e., m = 1:

j 2p
X k +1 (1) = e N {X k (1) + x ( N + k ) - x (k )}
j 2p ( k +1)
It can be seen that DFT during (k + 1)th window accumulates a phase shift of e N
j 2p ( k +1)
We can get rid of this phase shift of e N in the computation of the (k + 1)th window
- j 2p ( k +1)
by multiplying with e N (1) as:
. Hence, we define a modified DFT X k +1
- j 2p ( k +1)
(1) = X (1) e
X N
k +1 k +1
- j 2p ( k +1)
Multiplying both sides of equation for DFT by e N , we get:
- j 2p ( k +1) - j 2p ( k +1) j 2p
X k +1 (1)e N =e N e N {X k (1) + x( N + k ) - x(k )}
- j 2p ( k )
(1) = e
X k +1
N
{X k (1) + x( N + k ) - x(k )}
- j 2p ( k ) - j 2p ( k )
(1) = X (1)e N + [ x ( N + k ) - x (k )]e N
X k +1 k

Using the definition for the modified DFT, it can be seen that:
- j 2p ( k )
X k (1)e N (1)
=X
k
Hence, we can write the modified recursive sliding DFT as:

- j 2p ( k )

X k +1 (1) = X k (1) + [ x ( N + k ) - x (k )] e N .

6.13.2 MATLAB Program for Sliding Modified Recursive DFT Algorithm

% Sliding Modified Recursive DFT Algorithm
% First 12 samples are generated from 200*sin(2*pi*50*t+pi/4)
% Next 12 samples are generated from 1000*sin(2*pi*50*t)
% sampling is done at 12 samples/cycle of 50 Hz i.e., at
600 Hz
% Sliding Modified Recursive DFT is applied to track the
signal magnitude
clear , clc
fsig = 50 ; T=1/fsig;
N = 12 ;
fs = N * fsig ;
dt = 1/fs ;
% -----------------------------------------------------------
%---------- Lets generate samples of signl ----------------
Im1 = 200 ; Im2 = 1000;
phi1 = 45*(pi/180) ; phi2 = 0 ;
n = 0 ;
for t=0:dt:(T-dt)
n = n+1;
x(n)=Im1*sin(2*pi*fsig*t+phi1);
end
fprintf(\nAt the end of first loop n = %f \n, n )
fprintf(First Twelve Samples\n)

disp(x(1:12))
% first 12 samples are from 50Hz signal with peak value of 200
for t = T:dt:(2*T-dt)
n = n+1 ;
x(n)= Im2*sin(2*pi*fsig*t+phi2);
end
fprintf(Next 12 Samples\n)
disp(x(13:24))
% Next 12 samples are from 50 Hz signal with peak value
of 1000
%--------------------------------------------------------
%---------------- DFT calculation ----------------------------
% Initialization of coeffs. to zero before start of
calculations
for m = 0:1 % m is harmoic number 0=dc , 1 = fundamental
X(m+1)= 0 ;
end
%----- Plotting of samples of signal--------------------------

subplot(2,1,1)
stem(x)
xmin1=0;xmax1=25;ymin1=-1000;ymax1=1001;
axis([xmin1 xmax1 ymin1 ymax1])
grid on
xlabel(Sample Number -->)
ylabel(Sample value)
title(Composite Signal ( discontinuity at sample no.13))
% -----------------------------------------------------------
%------------- Actual Caluculation of DFT ------------------
m = 1 % Only fundamental
for n=0:N-1
X(m+1)= X(m+1)+ ( x(n+1)*exp(-j*2*pi*m*n/N));
end
% ----------------------------------------------------
fprintf(Calculated value of DFT\n)
% Let us represent the sliding windowed DFT with X_hat(window
no. = k)
% Note X(1) is DC component and X(2) is fundamental

X_hat(1)=X(2);
for k=1:12
Newest = x(N+k);
Oldest = x(k) ;
X_hat(k+1)= X_hat(k)+( (Newest - Oldest)*(exp(-j*2*pi*k/N)));
end
X_hat_mag =(2/N)*abs(X_hat)
% ----- Plotting of Sliding DFT -----
subplot(2,1,2)
stem(X_hat_mag)
xmin1=0;xmax1=13;ymin1=0;ymax1=1001;
grid on
xlabel(Sliding Window Number --->)
ylabel(Sliding DFT Magnitude)
title(Sliding DFT)
% ------------- End of program ------------------------
Figure 6.15 Sliding modified recursive DFT algorithm in MATLAB.
Review Questions
1. What are the conditions to be fulfilled by a signal for being represented by Fourier
series?
2. What are the Dirichlet conditions?
3. State Fourier series in trigonometric form without phase angle.
4. State Fourier series in trigonometric form as a cosine wave with phase angle.
5. State Fourier series in trigonometric form as a sine wave with phase angle.
6. State the relationship between coefficients of Fourier series in Q.4 with Fourier series
in Q.3.
in Q.3.
8. Derive expression for Fourier series in complex exponential form.
in Q.3.
10. Prove that the an and bn coefficients of the Fourier series are optimal in LSQ sense.
11. Compare the expressions for CFT and DFT and point out similarities between them.
12. If the DFT has N frequency terms, what is the highest harmonic frequency in terms
of multiples of fundamental?
13. In a protective relay hardware, 8 samples at a sampling frequency of 1000 Hz are
collected in the DFT window. What is the frequency of the fundamental, 2nd, 3rd,
4th harmonic etc.?
14. All samples inside the DFT window have same magnitude. What frequency content
is indicated?
15. Only one sample is inside the DFT window, at the centre is non-zero, all other samples
are zero. What frequency content is indicated?
16. If there are N samples in time domain, how many samples will be there in the
frequency domain?
17. What is the effect of increasing number of samples in a given window on the precision
of DFT?
18. What is special about X(0)?
19. What is special about X(N/2); assuming N is even.
20. Explain the statement: all N DFT coefficients are not unique.
21. If x(n) are all real, is it guaranteed that X(n) will all be real?
22. Why conventional Fourier series theory fails when we try to apply it to non-recurring
waveform?
7
FFT and Goertzel Algorithm
7.1 What is the Motivation for FFT?
The expression for DFT is normally written in the form shown below:
n = N -1
X (m) mm == 0N -1 = x (n)e - j 2p mn/N
n=0
Let us represent ej2p/N = WN (Note that negative sign is included in WN) and rewrite expression
for DFT as shown below:
n = N -1
X (m) mm == 0N -1 = x (n)WNmn
n=0
For N = 8, we can write:
n=7
X (m) mm == 07 = x (n)W8mn
n=0
This is actually a matrix equation since we have a N 1 column vector on LHS and
product of a (N N) matrix with a (N 1) column vector on the RHS. Hence, we can write
the equation as:
133
W80 0 W80 1 W80 2 W80 3 W80 4 W80 5 W80 6 W80 7

X (0) x (0)
X (1) W81 0 W811 W81 2 W813 W81 4 W81 5 W81 6 W81 7 x (1)
X (2) W82 0 W82 1 W82 2 W82 3 W82 4 W82 5 W82 6 W82 7 x (2)
X (3) W 3 0 W831 W83 2 W833 W83 4 W83 5 W83 6 W83 7 x (3)
= 8
X (4) W84 0 W84 1 W84 2 W84 3 W84 4 W84 5 W84 6 W84 7 x (4)
X (5) W85 0 W851 W85 2 W853 W85 4 W85 5 W85 6 W85 7 x (5)
X (6) x (6)
W86 0 W86 1 W86 2 W86 3 W86 4 W86 5 W86 6 W86 7
X (7) x (7)
W87 0 W871 W87 2 W873 W87 4 W87 5 W87 6 W87 7
Thus, we can see that an 8-point DFT involves 8 8 = 64 multiplications. In general

an N-point DFT involves N2 multiplications. As N goes on increasing, N 2 increases in an
exponential manner. Since multiplication takes time to perform on the DSP microprocessor, the
computation time of DFT sharply increases as number of sample points increase. Therefore, it is
highly profitable to somehow cut-down the number of multiplications. The following observation
provided a solution to the problem of cutting down the number of multiplications. An N-point
DFT requires N2 multiplications and an (N/2) point DFT requires (N2/4) multiplications. If,
somehow, we could breakup an N-point DFT as a simple function of two (N/2) point DFTs
then the number of multiplications required would be (N2/4) + (N2/4) = 2(N2/4) = N2/2
multiplications instead of N2 multiplications. Thus, we would be able to reduce the number of
multiplications from N2 to N2/2, i.e., multiplications are reduced to half. Now, we can apply
the same argument and break down N/2 point DFT in terms of two N/4 point DFTs and get
further reduction in number of multiplications. Let us apply this divide and conquer policy to
an 8-point DFT, as shown below:
8-point DFT direct implementation 82 64 multiplications
8-point DFT [4-point DFT] + [4-point DFT] 16 + 16 = 32 multiplications
8-point DFT [2-point] + [2-point] DFT + [2-point] + [2-point] DFT 4 + 4 + 4
+ 4 =16 multiplications
Thus, by breaking down the DFT into smaller and smaller DFTs, we could reduce the
number of time-consuming multiplication operations from 64 to 16. N is very large, typically
1024, in case of image processing applications and saving becomes substantial. We did not
count the number of additions involved. It can be shown that the number of additions required
also goes down substantially if we apply the approach discussed above. This process of breaking
down the DFT into smaller and smaller DFTs is called decimation. If we do decimation-in-
time domain, the DFT algorithm is called decimation-in-time FFT. The alternative approach
is decimation-in-frequency. Thus, FFT is nothing but a method of implementing DFT so as
to speed it up. FFT is not different from DFT as far as the input and output are concerned as
shown in Table 7.1.
FFT and Goertzel Algorithm135
Table 7.1 DFT and FFT

Multiplications Additions
DFT N2 N(N 1)
FFT N/2 log2 N N log2 N
Table 7.2 shows the substantial saving in computational effort offered by FFT assuming
N = 213 = 8192.
Table 7.2 Substantial saving in computational effort offered by FFT

N = 8192 Multiplications Additions
DFT N 2 = 67108864 N(N 1) = 67100672
FFT N/2 log2 N = 53248 N log2 N = 106496
Saving DFT multiplications DFT additions
@ 1260 @ 630
FFT multiplications FFT additions
The FFT algorithm was invented by J.W. Cooley and J.W. Tuckey in 1965. In those days,
digital computer capabilities were rather limited in terms of processing speed and memory
storage etc. Hence, FFT algorithm had a dramatic impact upon the scientific community.
In fact, it started a revolution in signal processing. Many applications where one would not
think of applying FFT (because of the high cost) could now benefit from FFT. Eventually, all
other methods of doing Fourier analysis were superseded by digital computer based FFT. It is
interesting to note that like most other mathematical inventions even FFT can be traced back
to Carl Fredrich Gauss, the 18th century mathematical genius.
Consider an 8-point DFT. We have taken out the relevant portion of the matrix equation
for extraction of fundamental. Another way to look at it is as a dot product or inner product
of two vectors. One vector consists of the 8 weights and the other vector is composed of
the 8 samples in time domain. Weights for computation of fundamental can be visualised
as shown in Figure 7.1.
Figure 7.1 Visualisation of weights for computation of fundamental in an 8-point DFT (contd.).
Figure 7.1 Visualisation of weights for computation of fundamental in an 8-point DFT.
7.2 FFT by Decimation-in-Time (DIT) [5,13,19,22]

The idea of decomposing a bigger DFT into smaller ones is called decimation. It is shown
conceptually in Figure 7.2 for an 8-point DFT. Note that a 1-point DFT of a signal is the signal
itself, so it is a trivial operation, hence most of the books stop at the 2-point DFT.
Figure 7.2 Decimation of an 8-point DFT.

In the decimation-in-time approach, we express the N-point DFT as a combination of

N/2 point DFT of even indexed samples and another N/2 point DFT of odd indexed samples.
Note that for N/2 to be an integer N must be even. Further it helps if N is a power of 2.
Mixing of even indexed and odd indexed samples in this fashion is perfectly valid since
in a summation the order in which terms are summed does not matter. The idea is shown
conceptually in Figure 7.3
Figure 7.3 Combining two N/2 point DFTs.
To develop the decimation-in-time FFT algorithm, we start with the expression for DFT:
n = N -1 n = N -1
X (m) mm == 0N -1 = x (n)e - j 2p mn/N = x (n)WNmn
n=0 n=0
Decomposing the sum as that of even indexed and odd indexed time domain samples:
N N
n= -1 n= -1
2 2
X (m) mm == 0N -1 = x (2n)e - j 2p m (2 n )/N
+ x (2n + 1)e - j 2p m (2 n +1) /N
n=0 n=0
N N
n= -1 n= -1
2 2
X (m) mm == 0N -1 = x (2n)e - j 2p m (2 n )/N
+e - j (2p m )/N
x (2n + 1)e - j 2p m (2 n)/N
n=0 n=0
Consider the first part of the sum. By comparing it with standard DFT expression, we can
conclude that, the expression represents an N/2 point DFT of the even indexed time domain
samples.
N
n= -1 (2 n)
2 - j 2p m N
x (2 n)e N This is the
2
point DFT of even indexed samples
n=0
Consider the summation in the second part of the expression for DFT. Again it can
be seen that it is another N/2 point DFT, but this time, of the odd indexed time domain
samples.
N
n= -1
2 N
x (2 n + 1)e - j 2p m (2 n)/N This is the
2
point DFT of odd indexed samples
n=0
Recall the definition of WN; WN = ej2p/N. Hence the remaining term, i.e., ej(2pm)/N can be
written as WNm. Thus, what we have done is the decomposition of the N-point DFT as:
[mth term of N-point DFT] = [N/2 point DFT(even indexed samples)]+ WNm [N/2
point. DFT(odd indexed samples)]
Thus, N-point DFT is not simply a sum of two N/2 point DFTs, we have to use a weighting
factor of WNm before combining the two DFTs. Using more compact notation we can write:
N
- j 2p mn /
- j 2p m (2 n )/N 2
e =e = WNmn
/2
and hence the expression for DFT becomes:

N N
n= -1 n= -1
2 2
X (m) mm == 0N -1 = /2 + WN
x (2n)WNmn m
x (2 n + 1)WNmn
/2
n=0 n=0
Further saving in labour is possible by noting the relationship between:

N N
X (m); m = 0 to - 1 fi X (0) to X - 1 and
2 2
N N N
X m + ; m = 0 to - 1 fi X to X ( N - 1)
2 2 2
N N N
m= -1 n= -1 N N n = -1 N
N 2 2 m + n m + 2 m + n
X m +
2 m =0
= x (2n)WN/2 2 + WN 2 x (2 n + 1)WN/2 2
n=0 n=0
N
m+ n

Consider WN/2 2
N N - j 2p nN/2
m + n n
WN/2 2
= WN( mn ) 2
/2 WN/2 = WN( mn )
/2 e
N/2
- j 2p nN/2
e N/ 2 = e - j 2p n = 1 (always)
Hence,
N
m + n
WN/2 2
= WN( mn
/2
)
N
m +
Similarly, considerWN 2
N N
m +
WN 2
= WN( m )WN 2
N

Consider WN 2
N - j 2p N/2

WN 2 = e N
= e - jp = -1
N
m +
Hence, WN 2
= -WN( m )
i.e., weights for the second half of DFT terms are negative of those for the first half. Thus,
we can write:
N N N
m= -1 n= -1 n= -1
N 2 2 2
X m +
2 m =0
= x (2n)WN( m/2) n - WN( m ) x (2n + 1)WN( m/2) n
n=0 n=0
We can rewrite the above expression as:

N N
n= -1 n= -1
2 2
X (m) m =( NN -1) = x (2n)WN( m/2) n - WN( m ) x (2n + 1)WN( m/2) n
m =( ) n=0 n=0
2
Collecting the expressions for the DFT together:

N N
N n= -1 n= -1
m = -1 2 2

X ( m) m = 0 2 = x (2 n)WNmn
/2 + WNm x (2 n + 1)WNmn
/2
n=0 n=0
N N
n= -1 n= -1
2 2
X (m) m =(NN-1) = /2 - WN
x (2n)WNmn ( m)
x (2 n + 1)WNmn
/2
m = n=0 n=0
2

N
n= -1
2
If we let /2 = Am = DFT of all even indexed samples
x (2n)WNmn
n=0
N
n= -1
2
and x (2n + 1)WNmn
/2 = Bm = DFT of all odd indexed samples
n=0
N
m = -1

then we can write X ( m) m = 0 2 = A(m) + WNm B(m)
and X (m) mm == NN -/ 21 = A(m) - WNm B(m)
Hence for an 8-point DFT, we can write the following:

X(0) = A(0) + W 80 B(0)
X(1) = A(1) + W 81 B(1)
X(2) = A(2) + W 82 B(2)
X(3) = A(3) + W 83 B(3)
and the second half of the terms as:
8
m = 0; X m + = X(0 + 4) = X(4) = A(0) W 80 B(0)
2
8
m = 1; X m + = X(1 + 4) = X(5) = A(1) W 81 B(1)
2
8
m = 2; X m + = X(2 + 4) = X(6) = A(2) W 82 B(2)
2
8
m = 3; X m + = X(3 + 4) = X(7) = A(3) W 83 B(3)
2
Figure 7.4 Development of FFT (decimation-in-time).
We take a look at the whole process in a step by step manner in Figures 7.5 and 7.6.
We can continue to follow the same mathematical logic to break down the 4-point DFT
into a combination of two 2-point DFTs and each 2-point DFT as a combination of two 1-point
DFTs. However, instead, we illustrate graphically, in Figures 7.7 and 7.8 how a 4-point DFT
is synthesised from two 2-point DFTs. Figure 7.9 shows the synthesis of a 2-point DFT using
two 1-point DFTs. Figure 7.10 depicts the concept of a butterfly. Finally Figure 7.11 combines
all these steps into a full-fledged flow-chart of an 8-point FFT with decimation-in-time.
Figure 7.5 Development of FFT(decimation-in-time) in a step-by-step manner.
Figure 7.6 Development of FFT (decimation-in-time) continued.


Figure 7.10 Concept of butterfly.

Figure 7.11 The FFT (decimation-in-time) signal flow graph.
It is to be noted that the original sequence of the indices which was in natural progression,
i.e.,
01234567
has become scrambled. It is now, 0 4 2 6 1 5 3 7.
Table 7.3 shows the binary representation of the indices.
Table 7.3 Binary representation of indices

Decimal 0 1 2 3 4 5 6 7
Binary 000 001 010 011 100 101 110 111
Bit reversed 000 100 010 110 001 101 011 111
Decimal 0 4 2 6 1 5 3 7
Thus, it is seen that the sequence of indices for the FFT can be easily generated by
bit reversal of the naturally ordered indices. To take advantage of this property, DSP
microprocessors include facilities for bit reverse addressing. Note that bit-reversal is not the
same as bit complementation.
7.2.1 Numerical Problem on FFT (DIT)

In order to appreciate the simplicity and beauty of FFT, it is worthwhile to solve a numerical
problem by hand calculation.
Example 7.1 Find the 8-point FFT(DIT) of the following signal samples collected in a
window of 8-samples.
x(0) = 650.0000 x(1) = 391.4214 x(2) = 350.0000 x(3) = 274.2641

x(4) = 150.0000 x(5) = 108.5786 x(6) = 50.0000 x(7) = 574.2641
Solution We can solve the problem in a visual manner by writing the calculations on the
signal flow graph of the FFT(DIT) as shown in Figure 7.12.
The answers can be seen to be:
X(0) = 400.0000 X(1) = 400 j399.8792

X(2) = 800.0000 j800.0000 X(3) = 1199.9390 j1199.8790
X(4) = 0.0 X(5) = 1199 + j1199.879
X(6) = 800.0000 + j800.0000 X(7) = 400.0000 + j399.8792
Thus, it can be seen that X(1) = X(7)*, X(2) = X(6)* and X(3) = X(5)* while there are no
conjugate entries for X(0) and X(4).
7.3 Fast Fourier Transform by Decimation-in-Frequency

To develop the decimation-in-frequency FFT algorithm, we again start with the expression for
DFT:
n = N -1 n = N -1
X (m) mm == 0N -1 = x (n)e - j 2 p mn/N = x (n)WNmn
n=0 n=0
Decomposing the sum as that of first-half and second-half of the time domain samples:
N N
n= -1 n= -1 N
2 2 N - j 2p m 2 + n /N
X (m) mm == 0N -1 = x ( n) e - j 2p m ( n )/N
+ x + n e
n=0 n=0
2
e - j 2p m ( n) / N = WNmn
N
- j 2 p m + n /N
2
e = e - j 2 p mN/2 N e - j 2p mn/N
Figure 7.12 Numerical problem on FFT(DIT).

N
- j 2p m + n / N
2
e = e - j 2p m/2 e - j 2p mn/N ; in the first term the constant 2 in numerator is not
cancelled with that in the denominator so that the term remains in standard form WN = ej2p/N.
Hence,
N
- j 2p m + n / N
2
e = W2mWNmn
N N
n= -1 n= -1
2 2 N
X (m) mm == 0N -1 = x (n)WNmn + W2m x + n WNmn
2
n=0 n=0
In order to introduce further simplification by way of recursion, consider the even indexed
terms of the DFT, i.e., X(2m):
N N
n= -1 n= -1
2 2 N
X (2m) mm == 0N / 2 -1 = x (n)WN(2 m ) n + W2(2 m ) x + n WN(2 m ) n
n=0 n=0
2
N N
n= -1 n= -1
2 2 N
X (2m) mm == 0N / 2 -1 = x (n)WNmn/ 2 + W2(2 m ) x + n WNmn/ 2
n=0 n=0
2

What is W2(2 m ) ?
- j 2p 2 m
W2(2 m ) = e 2 = e - j 2p m = 1
Hence, we get:
N N
n= -1 n= -1
2 2 N
X (2m) mm == 0N / 2 -1 = x (n)WNmn/ 2 + x + n WNmn/ 2
n=0 n=0
2

We can club the summation as:
N
n= -1
2 N mn
X (2m) mm == 0N/2 -1 = x (n) + x 2 + n WN/2
n=0
The above is an N/2 point DFT of sum of signals from the two halves, i.e., for an 8-point
DFT
[x(0) + x(4)], [x(1) + x(5)], [x(2) + x(6)] and [x(3) + x(7)]
But remember that the above expression gives only even indexed terms of the DFT. We
need to find expression for remaining half of the DFT terms, i.e., the odd indexed terms of
the DFT.
Consider the expression for DFT derived earlier:

N N
n= -1 n= -1
2 2 N
X (m) mm == 0N -1 = x (n)WNmn + W2m x + n WNmn
n=0 n=0
2
Hence, consider odd indexed terms of the DFT:

N N
n= -1 n= -1
2 2 N
X (2m + 1) mm == 0N / 2 -1 = x (n)WN(2 m +1) n + W2(2 m +1) x( + n)WN(2 m +1) n
n=0 n=0 2
What is WN(
2 m +1) n
?
WN(2 m +1) n = WN2 mnWNn = WNmn/ 2WNn
What is W2(2 m +1) ?
- j 2p 2 m - j 2p 1
W2(2 m +1) = W22 mW21 = e 2 e 2 = e - j 2p m e - jp = 1 ( -1) = -1
Hence, we get:
N N
n= -1 n= -1
2 2 N
X (2m + 1) mm == 0N / 2 -1 = x (n)WNmn/ 2WNn - x + n WNmn/ 2WNn
n=0 n=0
2
Simplifying further we get:

N
n= -1
2 N mn n
X (2m + 1) mm == 0N / 2 -1 = x (n) - x + n WN / 2 WN
n=0 2
Alternate way of writing the above expression could be:
N
n= -1
2 N n mn
X (2m + 1)mm == 0N / 2 -1 = x (n)WN - x + n WN WN / 2
n
n=0 2
The above is a weighted N/2 point DFT of difference of signals from the two halves,
weighted by factor WNn , i.e., for an 8-point DFT:
[x(0) x(4)], [x(1) x(5)], [x(2) x(6)] and [x(3) x(7)]
Thus, on the input (time domain) side we have to club the first half of samples with the
second half of samples and on the output side (frequency domain) we get a sequence of even
indexed DFT terms followed by odd indexed DFT terms. If we compare this with decimation-
in-time FFT then by analogy we can call this as decimation-in-frequency FFT. Hence, the term
called decimation-in-frequency FFT.
The process of breaking down the bigger DFT into DFT of sum and difference of smaller
sequences is continued till there is only one sample left in the sequence. The idea can be
grasped intuitively by Figures 7.13 to 7.15. Finally Figure 7.16 shows the full-fledged signal
flow graph for the FFT with decimation-in-frequency.
Figure 7.13 Development of decimation-in-frequency 8-point FFT.
The process of breaking down the DFT can be continued further till we get to one point
DFT which is the signal itself. This is shown for an 8-point FFT in Figures 7.14 and 7.15.

Figure 7.14 Breaking down a 4-point DFT into two number of 2-point DFTs.
Figure 7.15 Further breaking down of 2-point DFT into 1-point DFTs.
7.3.1 Numerical Problem on FFT (DIF)

We repeat the problem given in Section 7.2.1 and solve it by decimation-in-frequency as shown
in Figure 7.17 to get identical results.
7.4 MATLAB Implementation of FFT: Decimation-in-Time

% Program for testing my FFT DIT
clear, clc
x=[1 2 3 4 5 6 7 8];
XX = myFFTDIT(x)
YY = fft(x), % For checking results against MATLABs fft()
function
% the user created function myFFTDIT() is include here for ease in
documentation
% function [X] = myFFTDIT(x)
% Implementation of decimation-in-time FFT
% N = numel(x);
% if (N == 1 )
% X = x ;
% else
% xeven = x(1:2:N-1);
% xodd = x(2:2:N);
% WNm=exp( (-j * 2 * pi / N)*(0:(N/2)-1 ) );
Figure 7.16 Fully developed 8-point FFT signal flow graph.

Figure 7.17 Numerical problem on FFT(DIF).

%
% Am = myFFTDIT(xeven);
% Bm = myFFTDIT(xodd);
% X1 = Am + Bm .* WNm;
% X2 = Am - Bm .* WNm;
% X = [X1 X2]
%
%
% end
XX =
Columns 1 through 6
36.0000 4.0000 + 9.6569i 4.0000 + 4.0000i 4.0000 + 1.6569i 4.0000 4.0000 1.6569i
Columns 7 through 8
4.0000 4.0000i 4.0000 9.6569i
YY =
Columns 1 through 6
36.0000 4.0000 + 9.6569i 4.0000 + 4.0000i 4.0000 + 1.6569i 4.0000 4.0000 1.6569i
Columns 7 through 8
4.0000 4.0000i 4.0000 9.6569i
7.5 MATLAB Implementation of FFT: Decimation-in-Frequency

% -------------- FFT, Decimation-in-Frequency Algorithm
% --------------------------------------------------------------
%---------- The idea is to form two sequences of size N/2.
%---------- First sequence is formed by summing respective
elements of
%---------- first and second half.
%---------- Second sequence is formed by difference of the
above two, but
%---------- weighted by a weight W = exp( -j 2 pi n / N )
%----------- N/2 point DFT can be found for each of these
sequences ordered as
%----------- even indexed DFTs and odd indexed DFTs.
%------ This process is recursively continued till N/2 becomes 1

%------ 1-point DFT is the signal itself. This terminates the
recursion
%----------- Assumption is that N is a power of 2.
clear, clc
N = 8 ;
x =[ 1 2 3 4 5 6 7 8] ; % for testing
x
XX = myFFT(x);% FFT output in XX() is scrambled.
%-------------------- XXX() collects unscrambled FFT
XXX(1)=XX(1);% 0
XXX(5)=XX(2);% 4
XXX(3)=XX(3);% 2 Note:
XXX(7)=XX(4);% 6 Output of FFT

DIF is
XXX(2)=XX(5);% 1 scrambled.
XXX(6)=XX(6);% 5
XXX(4)=XX(7);% 3
XXX(8)=XX(8);% 7
XXX
YY = fft(x) % Compare the results with MATLABs fft() function
% The user defined function myFFT(x) is included here for ease

in documentation
% -------------------------------------------------------------
% --------- function [ X ] = myFFT( x )
% ----------Decimation in Freq FFT Algo
% --------------------------------------------------------------
% N = numel(x);
% if ( N == 1 )
% X = x ;
% else
%
% WNn = exp((-j*2*pi/N)*(0:(N/2)-1));
% first = x(1:(N/2))
% second= x((N/2)+1:N)
%
% sum = first + second ;
% diff = first - second ;
% diffW = diff .* WNn ;
% X1 = myFFT(sum);
% X2 = myFFT(diffW);
%
% X=[X1 X2];
%
%
% end
XXX = Columns 1 through 6
36.0000 4.0000 + 9.6569i 4.0000 + 4.0000i 4.0000 + 1.6569i 4.0000 4.0000 1.6569i
Columns 7 through 8
4.0000 4.0000i 4.0000 9.6569i
YY = Columns 1 through 6
36.0000 4.0000 + 9.6569i 4.0000 + 4.0000i 4.0000 + 1.6569i 4.0000 4.0000 1.6569i
Columns 7 through 8
4.0000 4.0000i 4.0000 9.6569i
7.6 Motivation for Goertzel Algorithm

This algorithm was published by Gerald Goertzel in 1958. Its utility stems from the fact that
in many applications, we may not be interested in computing the entire frequency spectrum
of a signal. In such situations the time and efforts taken by FFT in terms of computation
resources may not be justifiable. For example, in majority of power system applications, one
is interested in finding only the fundamental phasor, since the steady state theory is based on
fundamental quantities. In some algorithms like transformer protection we may be interested
in a few more harmonics like 2nd, 3rd and 5th but almost never in the full spectrum. Goertzel
algorithm offers a more efficient method of computation in such circumstances. A popular
application of this algorithm is recognition of the DTMF tones produced by the buttons pushed
on a telephone keypad.
7.7The Goertzel Algorithm [5,13,19,22]

For the sake of development the Goertzel algorithm, we take the case of a 4-point DFT:
n =3 - j 2p m n
X (m) mm ==30 = x (n)e 4
n=0
- j 2p - j 2p mn
Let e 4 = W4 hence, e 4 = W4mn
Hence, we can write the 4-point DFT as:
n =3
X (m) mm ==30 = x (n)W4mn
n=0
Let us say, we are interested in finding only the fundamental component; X(1), then
m = 1 and we can write:
n =3
X (1) = x (n)W4n
n=0
expanding the summation:
X (1) = x (0)W40 + x (1)W41 + x (2)W42 + x (3)W43
since W40 = 1; we can write:
X (1) = x (0) + x (1)W41 + x (2)W42 + x (3)W43
X (1) = x (0) +W41 ( x (1) + x (2)W41 + x (3)W42 )
X (1) = x (0) +W41 ( x (1) + W41 ( x (2) + x (3)W41 ))
X (1) = x (0) +W41[ x (1) + W41 ( x (2) +W41 x (3))]
The algorithm can be implemented as shown in Figure 7.18.
Figure 7.18 Flow chart of Goertzel algorithm for finding fundamental component of DFT.
The algorithm can be expressed more succinctly as shown in Figure 7.19.
Figure 7.19 Flow chart of Goertzel algorithm for finding fundamental component of DFT.
We can also appreciate the Goertzel algorithm with the help of the signal flow graph
shown in Figure 7.20.
Figure 7.20 Signal flow graph of Goertzel algorithm.

We can run through the signal flow graph by following entries as in Table 7.4.
Table 7.4 Entries to signal flow graph
Input to adder Output
Step Time x(n) Output of delay block Output of adder block Input to delay block
1 0 x(3) 0 x(3) W41 x (3)
2 1 x(2) W41 x (3) x(2) + W41 x (3) W41 ( x (2) + W41 x (3))
3 2 x(1) W41 ( x (2) + W41 x (3)) x (1) + W41 ( x (2) + W41 x (3)) W41 ( x (1) + W41 ( x (2)
+W41 x (3)))

4 3 x(0) W41 ( x (1) + W41 ( x (2) x (0) + W41 ( x (1)
+W41 x (3))) +W41 ( x (2) + W41 x (3)))
(4-point fundamental DFT
complete)
7.8 Implementation of Goertzel Algorithm in MATALB

%---------------------------------------------------------------
%----------------- Goertzel Algorithm -------------------------
%---------------------------------------------------------------
clear,clc
N=8 ; WN1 = exp(-j*2*pi* 1 /N);
X1=0;
x=[0 1 2 3 4 5 6 7] ;
for k=N:-1:1
X1= X1 * WN1 + x(k)
end
X1
%------- Lets check our result with MATLAB fft()
XX= fft(x);
XX1=XX(2) % First term of FFT is index 1 and not zero and is
the DC term, hence XX(2)
X1 = 7
X1 = 10.9497 4.9497i
X1 = 9.2426 11.2426i
X1 = 2.5858 14.4853i
X1 = 5.4142 12.0711i
X1 = 10.3640 4.7071i
X1 = 9.6569 + 4.0000i
X1 = 4.0000 + 9.6569i
X1 = 4.0000 + 9.6569i Final value of X1 found by Goertzel algorithm
XX1 = 4.0000 + 9.6569i Value of X1 found by MATLABs FFT function
Review Questions
1. What is the motivation for development of the FFT algorithm?

2. For an N-point DFT, what is the number of multiplications and number of additions?
3. Explain the concept of decimation.
4. Show how you will break down a 16-point DFT.
5. What is the number of multiplications required for FFT of 16 real samples?
6. What is the number of additions required for the FFT of 16 real samples?
7. Why FFT had a dramatic impact on signal processing?
8. What is the MATLAB function for performing FFT?
9. Perform FFT of [1, 2, 3, 4] and comment upon it.
10. Sketch the signal [1 2 3 4] and its recurring version.
11. Derive the decimation-in-frequency FFT algorithm.
12. Sketch the signal flow graph for an 8-point decimation-in-frequency FFT algorithm.
13. Why is it called as decimation-in-frequency?
14. What is special about the sequence of output samples in DIF FFT?
15. Write binary codes for the indices of samples of the output in DIF FFT and show
how they can be obtained from natural binary sequence.
16. What are the numerical values of the following weights?
[W8]0, [W8]1, [W8]2 [W8]3

[W4]0, [W4]1
[W2]0
19. What is the butterfly in the decimation-in-frequency algorithm of FFT?
20. Convert the FFT-DIT algorithm into a MATLAB program.
21. What do you mean by the term decimation?
22. What do you mean by decimation-in-time?
23. What do you mean by decimation-in-frequency?
24. What is the difference between DFT and FFT?
25. A signal has 8 samples. What will be the number of samples in the output of the
FFT?
26. Normally inputs to the FFT are real numbers. What can you say about the nature
(real/complex) of FFT of such signals?
27. How will you extract the magnitude spectrum from the FFT?
28. How will you extract the phase spectrum from the FFT?
29. Is FFT algorithm implemented recursively suitable for real-time applications?
30. What is the motivation for development of Goertzel algorithm?
31. Develop Goertzel algorithm for second harmonic extraction.
32. Develop Goertzel algorithm for mth harmonic extraction.
33. Calculate the number of multiplications and additions required by the Goertzel
algorithm.
8
Windowing and Spectral Leakage
8.1 The Fourier Machine

Fourier transform can be depicted as a machine extracting the various frequency components
and collecting them into individual bins as depicted in Figure 8.1. If the sampling frequency
Figure 8.1 The Fourier machine.
161
is an integral multiple of the signal frequency and the window length is adjusted to contain
integral number of cycles of the signal then the various frequency components get collected
in the correct bin. However, if the sampling frequency is not an integral multiple of the signal
frequency or if the window length is such that it does not contain integer number of cycles
of the signal, all the bins collect something, even if the signal has only single frequency
component. Thus, the frequency components are said to leak into those bins where they are
not supposed to show up.
8.2 Windowing without Spectral Leakage

In order to perform signal processing, we have to pick samples of signal for a finite duration,
say Twindow. Mathematically, this is equivalent to multiplying the signal with a rectangular
window of magnitude equal to 1 inside the window and zero outside the window. Thus, the
signal that we process is not the original signal but product of the original signal and the
window signal. It is therefore natural that the spectral characteristics of the windowed signal
will be different from that of the original signal. Since there is no escape from windowing,
we have to live with the fact that the spectrum of the windowed signal will be different from
that of the original signal. The next best thing that we can do is to strive for a window which
causes minimum contamination in the spectral content of the windowed signal from that of
the original signal. This motivates the study of various window shapes.
The following points should be kept in mind in order to appreciate the topic of windowing
and spectral leakage.
1. Windowing cannot be avoided.
2. It is helpful to think in terms of frequency bin, which is easy to imagine if we
visualise the Fourier machine which is a mechanistic way of looking at the Fourier
transform.
3. By windowing we are introducing a new signal, i.e., the window signal into the picture
in addition to the signal being processed. It is natural to expect that this additional
signal will show up in some way. In order that, windowing has least undesirable effect,
we will have to give very careful thought to the shape of the window. Intuitively, the
gentler the slope of the window at the edges, the better it is.
4. In time domain, windowing is simple multiplication between the samples of the signal
and samples of the window function.
5. In frequency domain, windowing amounts to convolution between the frequency spectra
of signal and frequency spectra of window function.
6. Because of convolution in frequency domain, it can be expected that frequency
spectrum of the windowed signal will smeared. Thus, a solitary line in the spectrum
of a signal may get transformed into a broader spectrum. This is obviously going to
cause confusion in the frequency domain, thus, leading to errors.
7. Rectangular window causes a drastic effect in time domain, because it multiplies
everything outside the window by zero. Hence, equally drastic effect in the frequency
domain can be expected.
Windowing and Spectral Leakage163
8. Sampled data system has vision only at integral multiples of fsampling/N, i.e., at analysis
frequencies of mfsampling/N; where m is from 0 to N/2. Thus, we have vision only
at discrete points in the frequency domain, up to a maximum frequency of fsampling/2
at m = N/2. At all other points in the frequency domain, we remain blind.
9. Frequency response of DFT is zero at integral multiples of fundamental. And these
are the points at which the DSP system has vision.
10. Frequency response of DFT is non-zero at non-integral multiples of fundamental, but
at these points DSP does not have vision.
It is to be noted that we are performing the mathematical operation of multiplication
between signal and window in the time domain. Multiplication in time domain has effect of
convolution in frequency domain. Hence, spectrum of time windowed signal can be found by
convolution of spectra of signal and window function.
Spectrum of the rectangular function is the well-known sinc() function. Hence, the spectrum
of the rectangular windowed signal is obtained by convolution of sinc() function with spectrum
of the signal.
Figure 8.2 The sampled and windowed signal in time and frequency domains.
Figure 8.2 shows that in spite of sampling and windowing, the spectrum of the signal
has remained intact! Careful observation of Figures 8.2 and 8.3 will reveal that this feat was
possible only because of the following:
1. Selection of the window size to match one complete cycle of the signal.
2. Selecting sampling frequency as an integral multiple of signal frequency.
Figure 8.3 Convolution between window spectra and signal spectrum.
The convolution operation between the window spectrum and signal spectrum can be
verified with the help of Figure 8.4.
It can now be easy to appreciate that in real life these conditions can seldom, if ever, be
satisfied. This raises the natural question: What will happen if:
1. a fraction of cycle of signal appears in the window?
2. sampling frequency is not an integral multiple of the signal frequency?
The answer to both above questions is that in such a circumstance there will be spectral
leakage. We take up spectral leakage in the next section.
Figure 8.4 Discrete convolution between window spectra and signal spectra.
8.3 Windowing Leading to Spectral Leakage [5,13,19,22]

Any phenomena in signal processing, in general, can be viewed from two alternate viewpoints:
from frequency domain viewpoint and from time domain viewpoint. Both the views give
different insight into the phenomena, making its understanding clearer and deeper. Let us first
view the spectral leakage from time domain viewpoint.
8.3.1 Spectral Leakage from Time Domain Viewpoint

Consider the windowing of an undistorted 50 Hz signal. If we adjust the window length to
exactly 20 ms as shown in Figure 8.5 then there is exactly one cycle of the signal inside the
window. The implied windowed signal which is obtained by extending the signal, on both
sides, outside the window, also happens to be a smooth undistorted signal. Hence, additional
spurious frequency components are not generated. Next consider that the window remains fixed
at 20 ms but the signal frequency drops to 45 Hz. Note that the signal is undistorted. It can
be seen from Figure 8.5 that the implied windowed signal is no longer undistorted and severe
distortions have appeared at the window edges! This is bound to cause additional frequency
components in the frequency spectrum of the windowed signal. These additional spurious
frequency components are referred to as spectral leakage in the signal processing literature.
Figure 8.5 Spectral leakage from time domain viewpoint.
8.3.2 Spectral Leakage from Frequency Domain Viewpoint

Figure 8.6 shows that when the window length is exactly equal to time period (T) of the signal
then there is no spectral leakage. However, if there is a mismatch between the window length
and signal time period then there is spectral leakage. It can be seen from Figure 8.6 that a 50
Hz signal (T = 20 ms) viewed through a 20 ms rectangular window does not cause any spectral
leakage, however a 45 Hz (T = 1/45 = 22.2222 ms) undistorted signal processed through the
same rectangular window of 20 ms causes spectral leakage in all the bins. In practice it is not
possible to match the window length precisely with the fundamental time period of the signal.
Hence, some degree of spectral leakage is unavoidable. The next best thing that we can do is
to search for a window shape which will minimise the spectral leakage.
From Figure 8.5, it is clear that the basic cause of spectral leakage is the abrupt change
in the signal magnitude at the edges of the rectangular window. Smoother window edges give
rise to lesser leakage. In the literature on signal processing, many window shapes have been
reported, however the theme common to all window shapes is to smoothen out the transition
Figure 8.6 Frequency domain view of spectral leakage.
to zero at the edges of the window. If we look at the spectrum of a window function, we
will notice that in general there is an inverse relationship between width of main lobe and
height of side lobe, i.e., side lobe ripple. Ideally we would like the width of main lobe to be
zero and side lobe ripple to be zero as well, which makes it an impulse at 0 Hz. However,
such a window shape (impulse) in frequency domain corresponds to infinite window length
in time domain, obviously an impractical proposition. Thus, we have to live with ill effects
of windowing, and the best that we can do is to choose a window shape which will minimise
spectral leakage. Table 8.1 lists a few popularly used windows.
Table 8.1 Popularly used windows
N -1 2p n
1. Hanning window = 0.5 - 0.5cos
n = 0 w ( n)
N
N -1 2p n
2. Hamming window = 0.54 - 0.46 cos
n = 0 w ( n)
N
n
N /2
=
n = 0 w ( n)
N/2
3. Triangular window
N -1 n
=2-
n=
N w ( n)
+1 N/2
2
Many other windows are reported in the literature, which are named after their inventors.
Some of these are: Bartlett, Blackmann, Chebyshev, Gaussian, Kaiser, Nuttal, Parzen and
Tuckey.
Figure 8.7 shows three windows in time and frequency domain. It can be observed that
in the frequency domain, the peak of rectangular window is shown normalised to 1. This is
known as processing gain or coherent gain of the window. This is in fact the DC signal gain
of the window. For all other windows, the gain is reduced because the window smoothly goes
to zero near the boundaries. The processing gain of the triangular window and the Hamming
window are 0.5 and 0.5328 respectively. This is a known bias on spectral amplitudes of the
windowed signal.
8.4 Windowed and Sampled Signal

Figures 8.8 and 8.9 show a signal which is windowed and sampled by using a triangular
and Hamming window respectively. It can be seen that in the middle of the window, the
amplitude is remains relatively unchanged while at the edges the signal magnitude is smoothly
brought to zero.
Figure 8.7 Various windows and their spectra.

Figure 8.8 Triangular windowed and sampled signal.
Figure 8.9 Hamming widowed and sampled signal.

Review Questions
1. Explain the concept of frequency bin.

2. What factors fix the position of the bin on the frequency axis?
3. What do you mean by spectral leakage?
4. In practise is it possible to avoid spectral leakage?
5. When is it possible to have a situation, where, in spite of rectangular window, there
is no spectral leakage?
6. What do you mean by windowing?
7. Why windowing is unavoidable?
8. Why rectangular shape gives rise to a drastic window?
9. List various window shapes and their mathematical functions.
10. Discuss: In time domain ideal window is that with infinite length.
11. Discuss: In frequency domain the ideal window is an impulse (The dual of the above
statement).
12. Define processing gain of window (also known as coherent gain) and list its value for
various windows.
9
Introduction to Digital Filtering
9.1 Introduction to Filters

What is common among the following:
1. Your sun-glasses
2. Water/Tea/Coffee dispensing machine
3. TV or any kind of radio receiver including your mobile telephone
4. Spike buster
5. DC power supply
6. Noise cancelling microphone
7. Lightening arrestor
8. Mosquito net
9. An antenna
10. Tendency of students to tune-out teachers words or advice?
A little thought will convince you that all these things involve the basic act of filtering. But
what is filtering? We intuitively understand filtering. Giving preferential treatment to one entity
as compared to some other entity could be one general definition. Thus, we can see that the
concept of filtering is ubiquitous. Filtering is not limited to the domain of electrical engineering
but can take place in almost any domain like chemical, optical or even psychological! In this
chapter we are going to concentrate upon a special version of filters called digital filters.
Everywhere you see, you come across filters. Life will be very difficult without filters.
We are going to study the digital versions of electrical filters. The speciality of these digital
filters is that they receive a stream of numbers as input and give out another stream of numbers
as output according to some algorithm. Thus, in the true spirit of GIGO philosophy, it is
numbers-in and numbers-out.
Thus, the world around us is full of filters. Filters help us to make sense out of the deluge
of information that floods us (the system). With everything going digital, it was inevitable that
filters will also take on a digital avatar! So we are confronted with digital filters.
172
Introduction to Digital Filtering173
9.2 What is a Digital Filter?

The term digital filter refers to a computational process by which a sequence of numbers,
usually samples of an analog signal is transformed into a second sequence of numbers as
shown in Figure 9.1.
Figure 9.1 Digital filter as a number-in, numbers-out system.
It can be proved that the computational process that goes on inside the box corresponds to
some kind of filtering operation like high-pass, low-pass, band-pass or band-stop filtering and
differentiating, integrating, convolution or co-relation. In fact every act of processing on time
domain samples creates some or the other effect in the frequency domain, be it intentional or
unintentional. We can also implement other types of digital filters which cannot be classified
into one of the basic four types listed above, such as adaptive filters used for noise cancelling
and beam steering. For example we can devise an adaptive digital filter for picking out an
EEG signal buried deep in 50 Hz line noise. We can design a filter for predicting tomorrows
share price index or the electrical load on the power system during the next hour and so on.
Keeping in mind that an antenna is a spatial filter, we can design it in such a way as to have a
null in the direction of an interfering signal, without, of course physically moving the antenna!
(Adaptive beam steering).
Thus, any mathematical operation is an act of filtering. Even if you do not do anything to
samples of a signal and just pass them from input to output, it is an all-pass filtering operation.
9.3 Why Digital Filtering?

Before we venture into detailed discussion about digital filters, it will be pertinent to ask; what
motivates us to go for digital filtering? There are certain advantages of using digital filtering
which are listed in Table 9.1.
Table 9.1 Advantages of digital filter over analog filter

Attribute Analog filter Digital filter
1 Consistency Analog filters exhibit deterioration in Once designed the digital filters exhibit
their performance because of drift in rock-steady performance as there is no
the component values with time and ageing effect and temperature plays no
temperature. role.
2 Repeatability Every piece of the filter has slightly Once designed the billionth piece will
different characteristics from the perform exactly as the prototype. Hence,
prototype because of component performance is perfectly repeatable and
tolerances. Hence, every piece needs to each instance of the filter need not be
be tested and checked. checked for performance!
3 Precision It is limited by the precision of the analog Precision can be increased to an arbitrarily
parts. Increasing precision beyond a high value (at least theoretically) by
point becomes very costly and hence increasing the number of bits used to
impractical. represent the values. This can be done
either in hardware by using a CPU with
large word length or can be done in
software.
4 Cost They turn out to be costly as compared to Because of advancements in VLSI
digital filters. techniques, they are turning out to be more
and more economical. As per Moores
law, capabilities of semiconductors
are doubling every 18 months. This
is continuously driving the costs in
downward direction.
5 Very low It becomes problematic to process very No such limitation in case of digital
frequency low frequency signals as component filters. They can handle low frequency
values and physical size of components signals with ease.
becomes impractical.
However, there are some drawbacks of digital filters. These are listed in Table 9.2.
Table 9.2 Disadvantages of digital filter over analog filter

Attribute Analog filters Digital filters
1 Very high Analog filters can process very It is difficult to process signals in the gigahertz
frequency high frequency signals upto the range (on a real time basis) as clock speeds need
signals gigahertz range. to be in tens and hundreds of gigahertz range. At
present such high clock speeds are not available.
2 Power Low power consumption High power consumption as compared to analog
consumption technique. Further, power consumption goes on
increasing with clock speed.
3 Simplicity Analog filters are simple to Digital filters are more complex as compared to
visualise. analog filters.
9.4 FIR and IIR Filters [5,13,19,22]

Digital filters are classified as finite impulse response (FIR) and infinite impulse response
(IIR). The basic philosophies of FIR and IIR are shown in Table 9.3, and Table 9.4 shows the
comparison between FIR and IIR filters.
Table 9.3 Basic philosophies of FIR and IIR

FIR filter IIR filter
Present output = function of (present input and Present output = function of (present input and
p number of past inputs) Q number of past inputs, and P number of past
outputs)
Coefficients remain constant w.r.t. time Coefficients remain constant w.r.t. time
Table 9.4 Comparison between FIR and IIR Filters
Characteristics FIR IIR
1 Number of computations to get the same Most Least

filtering effect
2 Non-recursive implementation possible Yes No
3 Recursive implementation possible Yes Yes
4 Can simulate analog filter prototype No Yes
5 Linear phase Yes, provided impulse No
response (coefficients)
exhibit either even or odd
symmetry
6 Stability Guaranteed if implemented Has to be ensured by
non-recursively design
7 Amount of hardware memory required Most Least
8 Sensitivity to filter quantization Very low High for direct form but
can be reduced by cascade
or parallel form
9 Probability of over-flow Very low Can be high for direct
form but can be reduced
by cascade form or parallel
form.
10 Difficulty of quantization noise analysis Least Most
11 Hardware filter complexity Simple Moderate
12 Ease of design Simple Moderately complex
13 Adaptive filtering possible Yes Yes
14 Suitability for use in numerical protective More suitable since it Less suitable since it
relays quickly forgets past (pre- remembers all past inputs,
fault) inputs and therefore therefore responds to pre-
is more affected by post fault conditions.
fault condition.
The IIR filters can be naturally implemented using recursion because of the inherent
feedback. Hence, they are sometimes referred to in the literature as recursive filters. There
is no apparent recursion in the FIR filters and are naturally implemented using non-recursive
procedures. Hence, they are referred to in the literature as non-recursive filter. However, it can
be shown that both FIR as well as IIR filters can be expressed recursively or non-recursively.
Hence, it is better to classify them as FIR and IIR instead of recursive and no-recursive.
FIR filters are the simplest of all. Hence, we first take up the explanation of FIR filters.
In FIR filters as shown in Table 9.3, the present sample of output is produced by linearly
combining the present input with a finite number of past input samples. FIR filters do not use
any of the past output samples, they use past input samples only.
Present FIR filter output = f[present input, past inputs]
y(n) = f[x(n), x(n 1), x(n 2),, x(n k)]
Figure 9.2 shows a block diagram representation of FIR and IIR filters. The block diagram
also suggests the hardware that will be required in case the filter is realised in hardware.
Figure 9.2 Block diagrams of FIR and IIR filters suggesting hardware implementation.
Thus, the block-diagram serves a dual purpose of a flowchart as well as hardware diagram of
the digital filter! It can be seen that multiplier, adder, delay (MAD) and memory are the basic
elements from which any digital filter of the FIR or IIR type can be constructed.
The outputs of the FIR and IIR filters, as can be seen from the block diagram, will be
given by:
FIR y(n) = b(0) x(n) + b(1) x(n 1) + b(2) x(n 2) + + b(k) x(n k)
IIR y(n) = b(0) x(n) + b(1) x(n 1) + b(2) x(n 2) + + b(Q) x(n Q)
a(1) y(n 1) a(2)y(n 2) a(P) y(n P)
9.5 Running Average as FIR Filter: 5-Point Running Average Filter

We will first get familiar with the symbolism associated with digital filters and then take up
simple mathematical analysis to show that the whatever processing is occurring can indeed be
called as a filter.
Consider a FIR filter whose output is the running average of present input sample and
previous four input samples. Thus, we can write the present output sample y(n) as:
x (n) + x (n - 1) + x (n - 2) + x (n - 3) + x (n - 4)
y ( n) =
5
Which can be written as:
1 1 1 1 1
y ( n) = x (n) + x (n - 1) + x (n - 2) + x (n - 3) + x (n - 4)
5 5 5 5 5
Which can be depicted in the form of a signal flow graph as shown in Figure 9.3.
The questions that naturally arise are:
Why is it called a filter? Does it work like a filter?
Which kind of filter?
Why is it classified as FIR?
Figure 9.3 Block diagram of 5-point running average filter.

We will answer all the above questions as we progress. Let us write the equation for y(n)
as follows:
y(n) = h(0) x(n) + h(1) x(n 1) + h(2) x(n 2) + h(3) x(n 3) + h(4) x(n 4)
Which can be written in the compact form as:
k =4
y ( n) = h( k ) x ( n - k )
k =0
Does the above equation look familiar? It surely does, as we can easily see that it is the
operation of convolution. It is a convolution between two sequences h( ) and x( ). This is a
very important result because it gives us a procedure for implementing digital filter and provides
an insight into its working.
9.5.1 Impulse Response of 5-Point Running Average Filter

Let us investigate the working of the FIR filter by applying a unit impulse signal to it. The
impulse signal is a very useful signal for the purpose of interrogating various digital systems,
since it reveals the information about the inner working of the digital system.
Figure 9.4 shows a unit impulse applied to the filter. We can see that the response to the
Input Output
Time x(n) x(n 1) x(n 2) x(n 3) x(n 4) y(n)
0 1 0 0 0 0 1/5 = h(0)
of impulse
Ts 0 1 0 0 0 1/5 = h(1)
Duration
response
2Ts 0 0 1 0 0 1/5 = h(2)

3Ts 0 0 0 1 0 1/5 = h(3)
4Ts 0 0 0 0 1 1/5 = h(4)
5Ts 0 0 0 0 0 0
6Ts 0 0 0 0 0 0
Figure 9.4 Computation of impulse response of 5-point running average filter.

impulse lasts for a finite duration after which it becomes zero. Hence, it is named as finite
(time) impulse response or FIR filter. We can also see that the impulse causes the coefficients
of the filter to be outputted one-by-one. Since the coefficients are all that can be known about
a filter, the impulse response reveals everything about the filter. The impulse response is
sketched in Figure 9.5.
Figure 9.5 Impulse response of 5-point running average filter.
9.5.2 Frequency Response of a 2-Point Running Average FIR Filter

We have not yet proved that the running average results into low-pass filtering. So it is time
now to investigate the frequency response of the running average system. For the sake of
simplicity, we will consider a running average system which takes average of just two samples;
the present sample and the sample just previous to it. Hence, y(n) can be written as:
1 1
y(n) = x (n) + x (n - 1)
2 2
jw nT
Let x (n) = e signal s be the signal with normalised amplitude (amplitude = 1) where wsignal
is the actual frequency of the signal and Ts is the sampling interval which is the reciprocal of
the sampling frequency fsampling. It will be much more convenient to express the frequency in
per unit rather than actual frequency in radians, since all signal frequencies in DSP have to
be limited to a value less than half the sampling frequency in hertz or to p in terms of pu
radian frequency. We define per-unit frequency as:
fsignal,actual fsignal,actual,max fsampling /2 1
f pu = fi f pu,max = = =
fsampling fsampling fsampling 2
Hence, wpu,max can be written as:
1
w pu,max = 2p f pu,max = 2p =p
2
Note that the following three are all equivalent:
fsignal,max fsampling/2 0.5 pu (maximum pu hertz freq.) p (maximum pu radian freq.)
1 fsignal
j 2 p fsignal n j 2p n
j 2p fsignal nTsampling fsampling fsampling j 2p ( f pu ) n jw pu n
x ( n) = e =e =e =e =e
Dropping the subscript pu from wpu but keeping in mind that w is per unit radian frequency,
we can write the expression for x(n) as:
x (n) = e jw n
where, w is the pu radian frequency of the signal.
Since output is the average of present sample and previous samples, so we can write:
1 jw n 1 jw ( n -1)
y ( n) = e + e
2 2
1 jw n 1 jw n 1 - jw
y ( n) = e + e e
2 2 2
1
y(n) = (1 + e - jw )e jw n , but e jw n = x (n)
2
Hence we can write:

1
y(n) = (1 + e - jw ) x (n)
2
y ( n) 1
= (1 + e - jw ) = H (w ); which is complex
x (n) 2
Hence we can write:

H (w ) = H (w ) H (w )
1
H (w ) = (1 + e - jw )
2
Taking out ej(w/2) common, we get:
1
H (w ) = (e j (w /2) + e - j (w /2) )e - j (w /2)
2
But we know that
1 j (w /2) w
(e + e - j (w /2) ) = cos
2 2
w
H (w ) = cos e - j (w /2)
2
Hence, the frequency versus magnitude response is a cosine curve and phase versus
frequency is linear, with negative slope, in the fourth quadrant as shown in Figure 9.6. Thus,
phase shift is always negative as the output of the filter always lags the input because of the
delay inherent in the filter operation.
Figure 9.6 Frequency response (magnitude and phase response) of 2-point running average filter.
Consider signal samples corresponding to lowest and highest frequencies. At lowest

frequency, i.e., w = 0
x (n) = e jw n = [e j 0 0 e j 0 1 e j 0 2 e j 0 3 e j 0 4 ]
= [1 1 1 1 1]
For the 2-point running average filter, we can find the output by running through the
sequence and finding the average of every two consecutive points.
(1 + 1) (1 + 1) (1 + 1)
y(n) = * *
2 2 2
y(n) = [* 1 1 1 *]
Thus, the lowest frequency signal, or DC is passed without any attenuation.

Now, let us first generate the samples of signal at highest possible per unit frequency of
p.
x (n) = e jp n = [ e jp 0 e jp 1 e jp 2 e jp 3 e jp 4 ]
x (n) = e jp n = [1 -1 1 -1 1]
Again, calculating the 2-point running average of the sequence, we get another sequence:
(1 - 1) (1 - 1) (1 - 1)
y(n) = * *
2 2 2
Hence, we get:
y(n) = [* 0 0 0 *]
Thus, the signal at the highest (per unit) frequency is totally attenuated to zero.
9.5.3 Frequency Response of 2-Point Running Average Filter Using

Z-transform
Let us analyse the 2-point running average filter using the Z-transform method. For the 2-point
running average filter, we have:
1 1
y(n) = x (n) + x (n - 1)
2 2
Taking Z-transform of both sides:
1 1
Y (z) = X ( z ) + z -1 X ( z )
2 2
1
Y ( z ) = (1 + z -1 ) X ( z )
2
Y (z) 1
= H ( jw ) = (1 + z -1 )
X (z) 2
Substituting z = ejw
1
H ( jw ) = (1 + e - jw )
2
1
H ( jw ) = (e + jw /2 e - jw /2 + e - jw /2 e - jw /2 )
2
1
H ( jw ) = (e + jw /2 +e - jw /2 )e - jw /2
2
e + jw /2 +e - jw /2 - jw /2
H ( jw ) = e
2
w
H ( jw ) = cos e - jw /2
2
w
H (w ) = cos is the magnitude response and H ( jw ) = e - jw /2 is the phase angle response.
2
These responses are exactly same as those derived without using Z-transform.
9.6 Frequency Response of a 3-Point Running Average Filter

Let us investigate the magnitude-frequency and phase-frequency response of a 3-point running
average filter. For a 3-point running average filter we can write:
y(n) = Average of present and last two samples
1 1 1
y(n) = x (n) + x (n - 1) + x (n - 2)
3 3 3
Let x (n) = e jw n
Hence, we can write y(n) as:
1 1 1
y(n) = e jw n + e jw ( n -1) + e jw ( n - 2)
3 3 3
1 1 1
y(n) = e jw n + e jw n e - jw + e jw n e - jw 2
3 3 3
1
y(n) = [1 + e - jw + e - j 2w ] e jw n
3
but e jw n = x (n); hence we can write:
1
y(n) = [1 + e - jw + e - j 2w ]x (n)
3
y ( n) 1
= H ( jw ) = 1 + e - jw + e - j 2w
x (n) 3
[1 + cos(w ) - j sin(w ) + cos(2w ) - j sin(2w )]

1
H ( jw ) =
3
H ( jw ) = [(1 + cos(w ) + cos(2w )) - j (sin(w ) + sin(2w ))]

1
3
(1 + cos(w ) + cos(2w ))2 (sin(w ) + sin(2w ))2
H ( jw ) = +
9 9
1
H ( jw ) = (1 + cos(w ) + cos(2w ))2 + (sin(w ) + sin(2w ))2
3
1 (1 + 2 cos(w ) + (cos(w ))2 + 2(1 + cos(w ) cos(2w )) + (cos(2w ))2
H ( jw ) =
3 + (sin(w ) + sin(2w ))2 )
1 [1 + 2 cos(w ) + (cos(w ))2 + 2 cos(2w ) + 2 cos(2w ) cos(w )

H ( jw ) =
3 + (cos(2w ))2 + (sin(w ))2 + 2 sin(2w )sin(w ) + (sin(2w ))2 ]
1
H ( jw ) = 3 + 4 cos(w ) + 2 cos(2w )
3
Recall that for a 2-point running average filter the cut-off frequency was p which can be
written as 2p/2. One can expect that the cut-off frequency for a 3-point running average filter
will be 2p/3. Hence, substituting w = 2p/3, we get
1
H ( j) = 3 + 4 cos(w ) + 2 cos(2w )
3
2p 1 2p 2p
H j = 3 + 4 cos + 2 cos(2)
3 3 3 3
2p 1 -1 -1
H j = 3+4 +2
3 3 2 2
2p 1
H j = 3 - 2 -1 = 0
3 3
To find H(jw):
H ( jw ) = (1 + cos(w ) + cos(2w )) - j (sin(w ) + sin(2w ))
(sin(w ) + sin(2w ))
H ( jw ) = - tan -1
(1 + cos(2w ) + cos(2w ))
(sin(w ) + 2 sin(w ) cos(w ))

H ( jw ) = - tan -1
(1 + cos(w ) + (2 cos(w )) - 1)
2
(sin(w ) + 2 sin(w ) cos(w ))

H (jw ) = - tan -1
(cos(w ) + (2 cos(w )) )
2
(sin(w )(1 + 2 cos(w )))

H (jw ) = - tan -1
(cos(w )(1 + 2 cos(w )))
(sin(w )
H (jw ) = - tan -1
cos(w )
H (jw ) = - tan -1 [tan (w )]
H (jw ) = -w
Hence,
2p 2p
H j =-
3 3
The frequency response of the 3-point running average filter is shown in Figure 9.7. Note
that at DC, i.e., w = 0;
1 1 1
H ( jw ) = 3 + 4 cos(w ) + 2 cos(2w ) = 3 + 4 .1 + 2 .1 = 9 =1
3 3 3
Figure 9.7 Frequency response of 3-point running average filter.
9.6.1 Frequency Response of 3-Point Running Average Filter Using

Z-transform
Let us analyse the 3-point running average filter using the Z-transform method. For the 3 point
running average filter, we have:
1 1 1
y(n) = x (n) + x (n - 1) + x (n - 2)
3 3 3
Taking Z-transform of both sides:
1 1 1
Y (z) = X ( z ) + z -1 X ( z ) + z -2 X ( z )
3 3 3
1
Y ( z ) = (1 + z -1 + z -2 ) X ( z )
3
Y (z) 1
= H (w ) = (1 + z -1 + z -2 )
X (z) 3
Substituting z = ejw
1
H (w ) = (1 + e - jw + e - j 2w )
3
jw
Taking out common factor e , we get:
1
H (w ) = (e jw + 1 + e - jw )e - jw
3
1 e jw + e - jw - jw
H (w ) = 1 + 2 e
3 2
1
H (w ) = (1 + 2cos(w ))e - jw = Magnitude angle
3
1
Thus, magnitude of output = (1 + 2cos(w )) and phase angle = w.
3
However, (1 + 2cos(w)) cannot be taken as magnitude since as cos(w) becomes 1, the
1
term becomes negative. Magnitude of (1 + 2cos(w )) can be written as:
3
1
H (w ) = (1 + 2cos(w ))2
3
1
H (w ) = (1 + 4 cos(w ) + 4(cos(w ))2
3
1
H (w ) = (1 + 4 cos(w ) + (2 cos(2w ) + 2))
3
H (w ) =
1
((3 + 4 cos(w ) + (2 cos(2w )))
3
and H (w ) = -w
Note that we have used the following:
1 + 2cos(w ) = (1 + 2cos(w ))2
Let us find the frequency at which |H(w)| is zero by solving the following:
1 + 2 cos(w ) = 0
2 cos(w ) = -1
1
cos(w ) = -
2
1
w = cos -1 -
2
2p
w= = 120 ( deg.)
3
This is the same as derived earlier without using Z-transform.
9.7 F
requency Response of 4-Point Running Average Filter Using
Z-transform
Output of the 4-point running average filter will be given by average of present sample taken
with the last three samples:
1 1 1 1
y ( n) = x (n) + x (n - 1) + x (n - 2) + x (n - 3)
4 4 4 4
Taking Z-transform of both sides
1 1 1 1
Y (z) = X ( z ) + z -1 X ( z ) + z -2 X ( z ) + z -3 X ( z )
4 4 4 4
1
Y (z) = (1 + z -1 + z -2 + z -3 ) X ( z )
4
Y (z) 1
H (z) = = (1 + z -1 + z -2 + z -3 )
X (z) 4
Substituting z = ejw, we get
Y (z) 1
H ( jw ) = = (1 + e - jw + e - j 2w + e - j 3w )
X (z) 4
Y (z) 1
H ( jw ) = = (1 + cos(w ) - j sin(w ) + cos(2w ) - j sin(2w ) + cos(3w ) - j sin(3w ))
X (z) 4
Y (z) 1
H ( jw ) = = [{1 + cos(w ) + cos(2w ) + cos(3w )} - j{sin(w ) + sin(2w ) +sin(3w )}]
X (z) 4
1
H ( jw ) = (1 + cos(w ) + cos(2w ) + cos(3w ))2 + (sin(w ) + sin(2w ) +sin(3w ))2
4
[sin(w ) + sin(2w ) +sin(3w )]

H ( jw ) = tan -1
[1 + cos(w ) + cos(2w ) + cos(3w )]
As the expressions for magnitude and phase response of the 4-point running average digital
filter is somewhat involved and not easy to sketch manually, let us write a small MATLAB
script to study the magnitude and phase response as shown in Figure 9.8. The program is
shown below:
Figure 9.8 Frequency response of 4-point digital filter.
% --------------------------------------------------------------
% ----- Freq Response of 4-Point Running Average Digital Filter
% --------------------------------------------------------------
clear , clc,clf
k=0;
step=pi/100;
for w = 0 : step : pi
k = k+1;
freq(k) = w;
x = 1 + cos(w)+ cos(2*w)+ cos(3*w);
y = sin(w) + sin(2*w) + sin(3*w);
H(k) =sqrt(x*x + y*y);
Ang_H(k) = atan2(-y, x);
end
subplot(2,1,1)
plot(freq,H)
title(Magnitude response of 4-point running average digital filter)
xlabel(Frequency , radians , w,max,pu=pi)
subplot(2,1,2)
plot(freq,Ang_H)
title(Phase response of 4-point running average digital filter)
xlabel(Frequency , radians , w, max , pu = pi)
In the next section, we write a MATLAB program for plotting the frequency response of
an N-point running average filter.
9.8MATLAB Program for Plotting Frequency Response of

N-Point Running Average Filter
%--- Frequency Response of N-point Running Average ( R.A.) Filter
% --------------------------------------
%--- For Nth order filter
% x = 1+cos(w)+cos(2w)+...+cos(N-1)w
% y = sin(w)+sin(2w)+...+sin(N-1)w
% |H(jw)| = sqrt( x*x + y*y )
% Ang(H(jw) = atan( -y x)
% In this program first the formulas for x and y are built-up

%- as strings. And then the strings are evaluated using the
% eval( ) function.
% Input :- N = Number of points in the running average
% Output :- (1) Magnitude response of the N-point R.A. filter

% (2) Phase response of the N-point R.A. filter
% (3) Formuals leading to |H(jw)| and Ang(H(jw))
clear,clc,clf
%--------------------------------------------------------
%------ Enter the number of points in the R.A. filter ---
N = 10;
%--------------------------------------------------------
x= 1
y=0
w=1;
for n=1:N-1
s=num2str(n);
ss=strcat(+cos(,s,,*,w,))
sss= strcat(+sin(,s,,*,w,))
x = [ x strcat(ss)]
y = [ y strcat(sss)]
end
x
y
step = pi/100;
k=0;
for w=0:step:pi
k =k+1;
ang(k)=w;
xx=eval(x);
yy=eval(y);
magH(k)=(sqrt(xx*xx+yy*yy))/(N);
phH(k)=atan2(-yy,xx)
end
ang_deg=(180/pi).*ang;
xmin=0;xmax=180;ymin=0;ymax=1;
xmin1=0;xmax1=180;ymin1=-pi;ymax1=pi;
subplot(2,1,1)
plot(ang_deg,magH)
axis([xmin xmax ymin ymax])
title(Magnitude Response of Running Average Filter)
xlabel( Freuency in pu deg ( max=pi rad))
subplot(2,1,2)
plot(ang_deg,phH)
title(Phase Response of Running Average Filter)
xlabel( Freuency in pu deg (max =pi rad) )
x;
y;
mag_of_H =[|H(jw)| =sqrt ( x ) ^2 + ( y )
^2];
ph_of_H =[ ang(H(jw) =atan2 (,-,( y ,),/,x ,)]
formula1=num2str(mag_of_H)
formula2=num2str(ph_of_H)
fprintf( |H(jw)| = %s \n ,formula1 )
fprintf( Angle(H(jw)) = %s \n,formula2 )
text(1,10,formula1)
text(1,0.8,formula2)
text(1,11,No. of points=)
text(40,11,num2str(N))
% subplot(2,1,2)
% text(1,
x = 1 + cos(1*w) + cos(2*w) + cos(3*w) + cos(4*w) + cos(5*w) + cos(6*w) + cos(7*w) + cos(8*w) + cos(9*w)
y = 0 + sin(1*w) + sin(2*w) + sin(3*w) + sin(4*w) + sin(5*w) + sin(6*w) + sin(7*w) + sin(8*w) + sin(9*w)


Figure 9.9 Frequency response of 10-point running average filter.
9.9 Introduction to Infinite Impulse Response (IIR) Filters

The infinite impulse response (IIR) filters are given by the following expression:
Q P
y ( n) = bk x (n - k ) - ak y(n - k )
k =0 k =1
The above expression can be interpreted as:

Present output = W
eighted sum of present input, past Q input samples and past
P output samples
Thus, there is a feedback from output to input. We can construct a signal flow graph based
on the above expression as shown in Figure 9.10. The weights denoted by bk and ak are called
the coefficients of the filter. The filter coefficients entirely decide the filter behaviour; hence
design of filter means finding bk and ak coefficients which will give the desire response.
We need to first answer the question: Why is it called infinite impulse response (IIR)?
Let us first make it clear that IIR stands for Infinite-time Impulse Response. Thus, the
response to an impulse signal, at least theoretically, exists for infinite-time. The impulse signal
sequence in the discrete domain is a finite signal of one-sample duration, which is preceded
and succeeded by an infinite string of zeros as shown below for a unit-impulse:
[ from 0 0 0 01 0 0 0 0 0 to ]
Figure 9.10 IIR filter block diagram.
The response of the IIR filter to the impulse will be of the form:
[from 0 0 0 0 y1 y2 y3 y4 y5 to ]
where y1, y2, are all finite values. The basic cause for continuation of the impulse response
for infinite time is the feedback that exists in the IIR filter. This feedback causes the signal to
get recycled again and again even after the input has vanished. Note also that the basic reason
for instability of IIR filters also stems from the feedback. FIR filters do not have feedback.
Hence, impulse response of FIR filter lasts for a finite time and there is no possibility of the
FIR filter becoming unstable.
We repeat the time domain expression for output of the IIR filter below, for convenience.
Q P
y ( n) = bk x (n - k ) - ak y(n - k )
k =0 k =1
The block-diagram of Figure 9.10 can be used to directly implement an IIR filter in either
hardware or software form. Hence, this method of implementation or the topology of the IIR
filter is called direct form IIR filter.
Taking the Z-transform of the above equation, gives us:
Q P
Y (z) = bk z - k X (z) - ak z - k Y (z)
k =0 k =1
Rearranging further gives:
P Q
Y ( z ) + Y ( z ) ak z - k = X ( z ) bk z - k
k =1 k =0
P Q
Y ( z ) 1 + ak z - k = X ( z ) bk z - k
k =1 k =0
Thus, we get the expression for H(z), the transfer function of the filter in frequency domain:
Q
Y (z)
bk z - k
= H ( z ) = k = 0P
X (z) -k
1 + ak z
k =1
b0 + b1 z -1 + b2 z -2 + b3 z -3 + + bQ z -Q
H (z) =
(1 + a1 z -1 + a2 z -2 + a3 z -3 + + aP z - P )
Polynomial of order Q in z
H (z) =
Polynomial of order P in z
Factored polynomial of order Q in z
H (z) =
Factored polynomial of order P in z
( z - q1 )( z - q2 )( z - q3 ) ( z - qQ )
H (z) =
( z - p1 )( z - p2 )( z - p3 ) ( z - pP )
Note that we can express the transfer function in various ways. For example we can write:
H(z) = H1(z).H2(z)Hk(z)
The above form can be implemented using a cascade of lower order transfer functions.
Alternatively we can express H(z) as:
k l
H ( z ) = 1 + H I ( z ) + H II ( z )
i =1 i =1
where HI(z) are k first order transfer functions and HII(z) are l second order transfer functions.
This type of representation leads to parallel form of IIR filter realisation. A question that needs
to be answered at this stage is: What is the motivation for alternate forms of implementation
when we can implement the filter easily in the direct form? The answer to this question is
not intuitively obvious. The answer lies in the error analysis of various topologies and their
sensitivity to quantization of coefficient values. It can be shown that direct form is not always
the best, from the above consideration and cascade and parallel form prove to be better under
certain situations. Hence, large numbers of topologies have been developed in the literature
on digital filtering.
Review Questions
1.
FIR stands for
(a) Finite magnitude impulse response
(b) Finite duration impulse response
(c) Finite duration, finite magnitude impulse response
(d) Finite magnitude, infinite duration impulse response.
2.
FIR filter makes use of
(a) Samples of past input only
(b) Samples of past output only
(c) Samples of past input as well as samples of past output
(d) Samples of past input along with present sample of input.
3.
IIR filter makes use of
(a) Samples of past input only
(b) Samples of past output only
(c) Present sample, samples of past input as well as samples of past output
(d) Samples of past input along with present sample of input.
4.
What are the basic hardware elements required for implementation of FIR or IIR filters?
5.
Discuss why FIR has finite-time and IIR has infinite-time impulse response.
6.
Why IIR filters are referred to as recursive filters?
7.
Why FIR filters are called non-recursive?
8. Show how an FIR filter can be implemented recursively.
9. FIR filters are referred to as filters with only zeros and a Qth order pole at the origin.
Justify/Explain.
10. IIR filters have both poles as well as zeros. Explain.
11. FIR filters are free from possibility of becoming unstable. Explain.
12. Explain why IIR filters need to be carefully designed in order to avoid instability.
13.
Explain intuitively how running average can have low-pass filtering effect.
14.
By corollary, which mathematical operation will cause high-pass filtering?
15.
What will be the effect of increasing the number of points in the running average?
16.
Analyse a 4-point running average filter and find is cut-off frequency and phase shift
at cut-off frequency.
17.
Perform the analysis of 2-point and 3-point running average filters using Z-transform.
18.
State whether the running average filter is of FIR type or IIR type.
19.
Explain the concept of per unit frequency.
20.
What is the maximum per unit radian frequency allowable in digital signal processing?
21.
What is the maximum per unit hertz frequency allowable in digital signal processing?
22.
What happens if a (pu) frequency greater than p appears at the input of a digital
processing system?
23.
Write the equation in time domain for output of a 2-point running average filter.
24.
Write the equation in time domain for output of a 3-point running average filter.
25.
Write the equation in time domain for output of a n-point running average filter.
26.
Convert above equations into Z-domain equations.
27.
Convert above equations into transfer functions of respective filters.
28.
Write a MATLAB script to plot the magnitude and frequency response of an n-point
running average filter.
10
Digital Filter Design
10.1 Filter Specifications

The design of filters starts with specifications of the desired filter in terms of its:
( a) Magnitude Vs frequency response
(b) Phase Vs frequency response
In terms of magnitude Vs frequency response, the main types of the filters are:
(i) Low-pass (LP) filter
(ii) High-pass (HP) filter
(iii) Band-pass (BP) filter
(iv) Band-stop (BS) filter
FIR filters can be designed to exhibit linear phase response while linear phase response
is not possible with IIR filters. Advantage of linear phase response is that it does not distort
the signal waveform. The phase response of the digital filters is not specified as rigorously as
the magnitude response.
Figure 10.1 shows ideal magnitude response of various filters.
Detailed filter specifications of a practical low-pass filter are shown in Figure 10.2.
The specifications of interest as shown in Figure 10.2 are pass band deviation dp, stop
band deviation ds, pass-band edge frequency fp and stop band edge frequency fs. The peak pass
band ripple is given as Apass = 20 log(1 + dp) dB and the minimum stop band attenuation is
given as Astop = 20log(ds).
10.2 R
elation between Linear Phase and Symmetry of Filter
Coefficients for FIR Filter [52]
Before we delve into design procedure for FIR filters, the following points need to be kept
in mind:
196
Digital Filter Design197
Figure 10.1 Ideal magnitude response of LP, HP, BP and BS filters.
Figure 10.2 Detailed specifications of LP digital filter.

1. Infinite combinations of filter coefficients are possible; theoretically making it possible

to design an infinite variety of filters.
2. However, since we are mostly interested in designing FIR filters with linear phase,
certain symmetry constraints get imposed on the FIR filter coefficients.
3. It can be shown that constraint of linear phase demands that filter coefficients exhibit
either even or odd symmetry.
4. Depending upon coefficient symmetry (even/odd) and filter order (even/odd), there
are four types of filters labelled as type-I, type-II, type-III and type-IV as shown in
Table 10.1 and Figure 10.3 .
Table 10.1 Four types of FIR filters

Filter Impulse Response Filter Order Phase Amplitude End-point Candidate
Type Symmetry M Offset Response Zeros Filters
Type-I Even Even Zero Even None All
Type-II Even Odd 0 Even z = 1 LP, BP
Type-III Odd Even p/2 Odd z = 1 BP
Type-IV Odd Odd p/2 Odd z=1 HP, BP
Figure 10.3 Impulse response symmetry patterns for type I, II, III and IV FIR filters.
10.3Design of FIR Filter Using Frequency Sampling Method

[13,22,52]
Let there be N samples of the entire frequency range from 0 to fsampling, then the ith frequency
sample will be located at a frequency of fi = i(fsampling/N) with i ranging from 0 to (N1).
The relation between impulse response and frequency response is as follows:
Frequency response = DFT (Impulse response)
Hence, taking inverse DFT, i.e., IDFT of both sides of above equation:
IDFT(Frequency response) = IDFT ( DFT (Impulse response))
= Impulse response = Coefficients of (FIR) filter
Hence, the impulse response h(k) can be written as:
h(k) = IDFT(H(fi)); k = 0 to N1; where H(fi) denotes the desired frequency response of
the filter. Further, since we always want the filter to have linear phase, we have to place certain
constraints on the symmetry that h(k) should exhibit. Thus even before we find actual values
of h(k), we know apriori that it will have certain symmetry. Let us assume even symmetry
about the mid point k = m/2. The frequency response of this type of filter can be written as:
H(f ) = Ar(f )ejpmfT
where the magnitude response Ar(f ) is a real even function that specifies the desired magnitude
response, i.e., |Ar(f)| = A(f).
Hence, we can write the expression for IDFT as:
N -1 j 2p i k
1
h( k ) =
N
H ( fi )e N
i =0
Note: The differences between DFT and IDFT formulae are as follows:
1. 1/N factor is included in IDFT formula. No such factor is in DFT formula.
2. Sign of exponent is positive in IDFT and negative in DFT formula.
We can write h(k) as:

N -1 j 2p ik
1
h( k ) =
N
Ar ( fi )e- jp fiT e N
i =0
which can be written as:

N -1 jp mif s T j 2p ik
1
h( k ) =
N
Ar ( fi )e N e N
i =0
Where fs = fsampling = 1/T; hence fsT = 1. Thus, we get:

N -1 jp mi j 2p ik
1
h( k ) =
N
Ar ( fi )e N e N
i =0
N -1 j 2p ik jp mi
1
h( k ) =
N
Ar ( fi )e N
i =0
N -1 j 2p i ( k 0.5 m )
1
h( k ) =
N
Ar ( fi )e N
i =0
N -1
1 2p i(k - 0.5 m) 2p i(k - 0.5 m)
h( k ) = Ar ( fi ) cos + j sin
N i =0 N N
We are designing a filter with all h(k) values real. Hence, the j sin() terms cancel out, giving:
N -1
1 2 p i (k - 0.5m)
h( k ) =
N
Ar ( fi ) cos N
i =0
Further, treating the i = 0 term separately, we can write:

N -1
Ar (0) 1 2 p i (k - 0.5m)
h( k ) =
N
+
N
Ar ( fi ) cos N

i =1
Using the symmetry property of DFT, i.e., the contributions of the ith and (Ni)th terms are
the same. Thus, we can sum half of the terms and double the result. Further recalling that b(k)
= h(k) for FIR filters and noting that m the order of the filter is given by:
m = N 1; where N is the number of time/frequency domain samples
N = m + 1; where m is the order of the filter
We get the final expression for filter coefficients as:
m
floor
k = m -1 Ar (0) 2 2
2 p i ( k - 0.5m )
k = 0 [ bk ] = +
(m + 1) (m + 1)
Ar ( fi ) cos
m +1
i =1
Note that for type-I filter, the order of the filter m is even, in which case floor (m/2) = m/2.
But for type-II filter, m is odd, hence m/2 will not be an integer but the floor() function of
MATLAB will generate an integer, thus allowing the summation.
10.3.1 N
umerical Problem on Design of FIR Filter Using Frequency Sampling
Method
Let us illustrate the design of FIR filter using the frequency sampling method with the help
of a numerical problem.
Example 10.1 Design a low-pass FIR filter to mimic an ideal low-pass filter with a cut-off
frequency of fsampling/4. Assume order of the filter to be 20.
Solution Let us design a type-I filter. We have order of the filter m = 20. We have m = N
1; where N = number of frequency domain samples. Hence N = m + 1 = 20 + 1 = 21. Thus,
number of frequency domain samples = 21. Let us first sketch the desired frequency response
of the filter as shown in Figure 10.4.
Figure 10.4 Frequency response of the desired low-pass filter.
For constructing Figure 10.4, following steps may be followed:

1. Draw 21 equally spaced points on the x-axis which represents frequency.
2. Mark this entire range from 0 to 20 as fsampling.
3. Mark half of this range, at the 10th sample point as fsampling/2.
4. Mark half of the above, at 5th sample point as fsampling/4.
Now, given the above frequency response we are interested in finding out the b(k)
coefficients of the FIR filter in the block diagram of the filter shown in Figure 10.5.
Thus, we are given information in frequency domain and are required to transform it to
time domain. Hence, inverse Fourier transform is called for as shown below:
Frequency domain Time domain : Inverse Fourier transform
Time domain Frequency domain : Fourier transform

Figure 10.5 FIR filter block diagram.
We can write, for the b(k) coefficients as:
m
floor
2
A (0) 2 2 p i (k - 0.5m)
[b(k )] = r +
(m + 1) (m + 1)
Ar ( fi ) cos
m +1

i =1
From the sketch of the desired frequency response we can see that:
i = 5 i = 20
Ar ( fi ) = 1 and Ar ( fi ) =0
i = 0 i=6
Further, we have m = 20, hence 0.5m = 10. Thus, we can write the above expression as:
1 2 10 2 p i(k - 10)
b( k ) = + Ar ( fi )cos
21 21 i =1 21
Let us write a small MATLAB script to compute the filter coefficients and verify that the
coefficients indeed implement the desired filter by finding the FFT of the coeffiients:
clear , clc
%--------------------------------------------
%---------Order of filter -----------------
m=20;
%-------------------------------------------
for k = 0 : m
fpu(k+1)= k/m;
b(k+1)=(1/(m+1));
for ii =1:5
b(k+1)=b(k+1) + (2/(m+1))* cos( 2 *pi*ii*(k-10)/(m+1)) ;
end
fprintf( b(%d)= %f \n,k,b(k+1))

end
%------------------------------------------------------------------
subplot(2,1,1)
stem(b)
xlabel(sample number)
ylabel(h(n)= Impulse response =b(k))
title(Design of low pass FIR filter with fc = f/4 )
subplot(2,1,2)
stem(fpu,abs((fft(b)))) For checking the frequency response from
the computed coeffs.
xlabel(Freuency in pu )
ylabel(Magnitude of output A(f))
b(0) = 0.032480 b(1) = 0.038188 b(2) = 0.028817

b(3) = 0.047619 b(4) = 0.026427 b(5) = 0.065171
b(6) = 0.024916 b(7) = 0.106999 b(8) = 0.024078
b(9) = 0.318607 b(10) = 0.523810 b(11) = 0.318607
b(12) = 0.024078 b(13) = 0.106999 b(14) = 0.024916
b(15) = 0.065171 b(16) = 0.026427 b(17) = 0.047619
b(18) = 0.028817 b(19) = 0.038188 b(20) = 0.032480
Figure 10.6 Design of FIR filter by frequency sampling method.

10.4 Bilinear Transformation

IIR filters are based upon analog filters. We start with an analog filter and transform its s
domain transfer function into the z domain transfer function of the corresponding digital filter.
Bilinear transformation is used for converting s domain transfer function into its z domain
equivalent. Hence, we will first derive the bilinear transform.
We have z = esT which can be written as:
2
sT

sT 2
1+ + +
e sT / 2
2 2!
z = esT / 2 esT / 2 = =
e - sT / 2 sT
2

sT 2
1- + -
2 2!
If we neglect higher order terms, we can write:
1 + (sT/2) 2 + sT
z@ =
1 - (sT/2) 2 - sT
Hence, we can write the following transformation as:
2 ( z - 1)
s=
T ( z + 1)
10.4.1 Design of IIR Filter Using Bilinear Transformation

If we have an analog filter prototype whose s domain transfer function is known, we can
transform it to z domain using bilinear transform. We can write time domain difference
equation for the filter from the z domain transfer function very easily. Thus, we can implement
an IIR filter. However, there is one more factor to be taken into account. As we migrate from
s domain to z domain, the frequency gets transformed in a very non-linear fashion. The basic
reason for this is that in the analog domain, frequency range is from zero to infinity, while,
in the digital domain, the frequency range is from zero Hz to (fsampling/2)Hz or on a per unit
basis from zero to p per unit. Thus, when we do the bilinear transform, the entire frequency
range from zero to infinity (Hz) is compressed into a range from zero Hz to (fsampling/2)Hz or
zero to p(pu Hz). This is known as warping of the frequency.
10.4.2 Warping of Frequency

To distinguish between frequency in the analog and digital domains, we use the following
symbols:
W : Analog frequency
w : Digital frequency
The relation between analog and digital frequencies can be derived as follows:
Noting that: s = jW and z = ejw
We have the bilinear transform given by:
2 ( z - 1) 2 (1 - z -1 )
s= =
T ( z + 1) T (1 + z -1 )
2 (1 - e - jw )
s = jW =
T (1 + e - jw )
which can be simplified as follows:
2 e - jw /2 (e jw /2 - e - jw /2 )
jW =
T e - jw /2 (e jw /2 + e - jw /2 )
2 (e jw /2 - e - jw /2 )
jW =
T (e jw /2 + e - jw /2 )
2 (e jw /2 - e - jw /2 ) 1
W= jw /2
T 2j (e + e - jw /2 ) / 2
w
sin
2 2
W=
T w
cos
2
2 w
W= tan
T 2
WT
w = 2 tan -1
2
2 w
W max = tan max
T 2
As digital frequency w reaches its maximum value, i.e., wmax = p, it gets translated as:
2 p
W max = tan = infinity
T 2
Thus, the entire range of analog frequencies from 0 to gets compressed into digital
frequencies from 0 to p; giving rise to the warping effect as shown in Figure 10.7.
Figure 10.7 Warping of frequency during bilinear transformation.
10.4.3 Pre-Warping of Frequency During Design

We can defeat the warping by pre-compensating the frequency as follows:
The bilinear transform has the effect if we design at an analog frequency of W, we get
effect at the digital frequency of 2tan1(WT/2). Hence, if we want effect at frequency of W,
we should design at pre-warped frequency of (2/T)tan(W/2). This will give effect at W since
2 tan1[(2/T)tan(W/2).(T/2)] = W.
To summarize:
2 W
(i) Pre-warped frequency at which filter is designed = tan
T 2
(ii) Effective (digital domain) frequency at which effect will be obtained = W
10.4.4 Step by Step Design of Iir Filter Using Bilinear Transformation

We can summarise the steps in the design of IIR filter using bilinear transform as follows:
1. Start with analog prototype of the filter to be designed. Widely used forms of analog
filters are Butterworth, Chebyshev, Elliptical etc. Analog filter data books contain
details about these filters.
2. Write down the s domain transfer function of the prototype analog filter.
3. Pre-warp the analog cut-off frequency WC to (2/T)tan(WC/2).
4. Re-write the s domain transfer function of the prototype filter in terms of the pre-
warped cut-off frequency.
5. Convert from s domain transfer function to z domain transfer function using bilinear
transformation, i.e., substitute
2 ( z - 1) 2 (1 - z -1 )
s= =
T ( z + 1) T (1 + z -1 )
6. Express the H(z) in terms of z1 so that implementation in terms of delay elements

becomes obvious.
7. Draw block diagram of the IIR filter in terms of multipliers, adders and delay elements.
The filter can now be implemented either in software or hardware.
10.4.5 Review of Formulas Derived

We have derived the formulas as shown in Table 10.2.
Table 10.2
2 ( z - 1) 2 (1 - z -1 )
Bilinear transform s= =
T ( z + 1) T (1 + z -1 )
WT
W
Analog frequency to digital frequency
w w = 2 tan -1
2
2 w
w
Digital frequency to analog frequency
W W= tan
T 2
Pre-warping 2 W
Pre-warping W tan
T 2
10.4.6 Numerical Problem on Design of IIR Filter

EXAMPLE 10.2 Design an IIR digital filter based on first order low-pass Butterworth filter
prototype with 3 dB cut-off frequency wc = 0.25 p; by using bilinear transformation.
Solution First order low-pass Butterworth filter with cut-off frequency of Wc has the s
domain transfer function H a (s) = 1(1 + s/Wc ).
As explained earlier if we design the filter to have a cut-off frequency of Wc, because of
bilinear transform the actual cut-off frequency will be (2/T)tan1(W/2).
Hence, we have pre-warp Wc to (2/T)tan(Wc/2). Hence,
2 0.25p 2 p 2 0.828
Wc ( pre-warped ) = tan = tan = 0.414 =
T 2 T 8 T T
Thus, the transfer function of the prototype filter which was H a (s) = 1(1 + s/Wc ) becomes:
1 1
H a (s) = =
s sT
1+ 1+
0.828 0.828
T
2 z 1
Applying bilinear transformation, i.e., substituting, s =
T z +1
1
H (z) = ; which can be simplified as shown in the following steps:
2 ( z - 1) T
1+
T ( z + 1) 0.828
1
H (z) =
2( z - 1)
1+
0.828( z + 1)
1
=
0.828( z + 1) + 2( z - 1)
0.828( z + 1)
0.828( z + 1)
=
0.828z + 0.828 + 2 z - 2
0.828( z + 1)
=
2.828z - 1.172
0.828( z + 1)
=
2.828( z - 1.172 / 2.828)
0.2927( z + 1)
=
( z - 0.414)
0.2927(1 + z -1 ) z
=
(1 - 0.414 z -1 ) z
0.2927(1 + z -1 )
=
(1 - 0.414 z -1 )
Now, let us think about implementing the filter. We have:
Y ( z ) 0.2927(1 + z -1 )
H (z) = =
X ( z ) (1 - 0.414 z -1 )
Y ( z )(1 - 0.414 z -1 ) = X ( z )0.2927(1 + z -1 )
Y ( z ) - 0.414 z -1Y ( z ) = 0.2927 X ( z ) + 0.2927z -1 X ( z )
Using the following inverse Z-transformations; where inverse Z-transform is denoted by

Z 1, i.e.,
Z 1[z1 Y(z)] = y(n 1); Z 1[Y(z)] = y(n); Z 1 [z1X(z)] = x(n 1); Z 1[X(z)] = x(n).
Thus, we can transform the above z domain equation into a time domain difference
equation as shown below:
y(n) - 0.414 y(n - 1) = 0.2927 x (n) + 0.2927 x (n - 1)
y(n) = 0.414 y(n - 1) + 0.2927 x (n) + 0.2927 x (n - 1)
The above time domain equation can be directly implemented with the help of multiplier,
adder and delay blocks, as shown in Figure 10.8.
Figure 10.8 Numerical on design of IIR filter using bilinear transformation.
10.5 Design of a Digital Resonator Using Pole Zero Placement

We will now take up design of a digital resonator using placement of z plane poles and
zeros of the transfer function. The response of the resulting digital system is analogous to that
of a tuned circuit. The motivation in taking up this exercise is to demonstrate that there are a
variety of approaches available for designing digital filters.
10.5.1 Understanding the z Plane

In order to understand the z plane, let us start with the definition of z:
z = r ejwTs
We have w = 2pf ; where f is the frequency of the signal and Ts = (1/fs) where Ts is the
sampling time period which is the reciprocal of fs the sampling frequency. Hence, we can write:
1
j2 p f
z = re fs
j 2p f pu
z = re
jw pu
z = re
Normally, the subscript pu is dropped from wpu. But we should keep in mind that w is per
unit radian frequency whose maximum value is p. Thus,
z = r ejw
Remembering that w is an angle, say q which varies from 0 to p.
z = r ejq
z = r e jq (angle from 0 to p)
The frequency response is found by evaluating the transfer function H(z) on the unit circle,
i.e., |z| = 1 or r = 1. Hence, z = ejq. Thus, we see that W which is a linear distance (in
the vertical direction) on the s plane, becomes an angle q on the z plane as shown in
Figure 10.9. Thus, as frequency tends to infinity in the s domain, the angle tends to p in
the z domain as shown in Table 10.3.
Table 10.3 Comparison between s and z planes

W w
Analog frequency in s domain Per unit frequency in the z domain z = r ejw
W0 w0 z = ej0 = 1
W infinity wp z = ejp = 1
Figure 10.9 Mapping from s plane to z plane.

Mapping of z plane from s plane from Figure 10.9 is shown in Table 10.4.
Table 10.4 Mapping of z plane from s plane
s plane z plane
Left half Inside unit circle
+/ j w axis Circumference of unit circle
Zero frequency z=1
Infinite frequency z = +/ p or 0.5 fs
10.5.2 Placement of Pole

The aim is to design a digital resonator with a resonant frequency of F0 as shown in
Figure 10.10.
Figure 10.10 Frequency response of the digital resonator.
Observation of the desired frequency response shown in Figure 10.10 leads to the following
conclusions:
1. The transfer function must contain a pole.
2. The pole should be at frequency F0 q0.
3. The pole should be located inside unit circle on the z plane, i.e., r < 1 for stability.
Now, the mapping is as follows:
W = 0 to infinity analoge frequency (Hz)
w = 0 to p digital frequency (pu)
f = 0 to fs /2 signal frequency (Hz)
F0
q0 = p
f s /2
2p F0
q0 =
fs
The z domain transfer function must have a pole at z = rejq 0, where r < 1
Now, since we want real coefficients, poles must always appear in conjugate pairs. Hence,
we can now directly write the expression for the z domain transfer function as:
Numerator
H (z) =
( z - re jq0 )( z - re - jq0 )
10.5.3 Placement of Zeros

Desired frequency response is zero at zero frequency as well as at very high frequencies.
Thus, we have to locate zeros at zero frequency and infinite frequency. At some intermediate
frequency, the response peaks because of the pole.
Thus, we also have to ensure that H(z) has a null at w = 0, i.e., z = ejw = 1.
Hence, z =1 is a zero of the transfer function.
We also have to ensure that H(z) has a null at w = p , i.e., at the highest frequency.
z = ejp = 1
Hence, z = 1 is also a zero of the transfer function. Thus, the transfer function that we
are trying to synthesise can be written as:
b0 ( z - 1)( z + 1)
H (z) =
( z - re jq0 )( z - re - jq0 )
b0 ( z 2 - 1)
H (z) =
z 2 - z r (e jq0 + e - jq0 ) + r 2
b0 ( z 2 - 1)
H (z) =
z 2 - 2 z r (cos(q 0 )) + r 2
If we design the gain to be unity at w = w0, then it will be automatically less than 1 at all
other frequencies. The gain has already been designed to be zero at w = 0 and w = p.
Imposing this constraint on H(z), we can write:
b0 e2 jq0 - 1
H ( z ) z = e jq = 1 =
e2 jq0 - 2e jq0 r cos (q 0 ) + r 2
which gives b0 as:
e2 jq0 - 2re jq0 cos(q 0 ) + r 2
b0 =
e2 jq0 - 1
The value of r should be nearly equal to 1 and on the lower side. The empirical formula
for r is given by:
DF p
r =1-
fs
where DF is change in frequency from F0 at which the gain is 3 dB down.
10.5.4 Expression for H(z), the Frequency Domain Transfer Function
b0 ( z 2 - 1)
H (z) =
z 2 - 2 z r (cos(q 0 )) + r 2
b0 e2 jq0 - 1
H ( z ) z = e jq = 1 =
e2 jq0 - 2e jq0 r cos(q 0 ) + r 2
e2 jq0 - 2re jq0 cos(q 0 ) + r 2

b0 =
e2 jq0 - 1
DF p
r =1-
fs
10.5.5 Numerical Problem on Design of Digital Resonator
Example 10.3 Design a digital resonator for a frequency of 200 Hz. Assume sampling
frequency to be 1200 Hz and DF = 6 Hz.
Solution We have:
2p F0 200 2p p
q0 = = 2p = =
fs 1200 6 3
We have
DF p 6p
r =1- =1- = 0.9843
fs 1200
e2 jq0 - 2re jq0 cos(q 0 ) + r 2

b0 =
e2 jq0 - 1
e2 jq0 - 2(0.9843)e jp / 3 cos (p / 3) + (0.9843)2

b0 =
e2 j p / 3 - 1
b0 = 0.0156
Hence, H(z) can be written as:

b0 ( z 2 - 1)
H (z) =
z 2 - 2 z r (cos(q 0 )) + r 2
0.0156( z 2 - 1)
=
z 2 - 2 z (0.9843)(cos(p / 3)) + (0.9843)2
0.0156( z 2 - 1)
H (z) =
z 2 - 2 z (0.9843)(0.5) + 0.9688
0.0156( z 2 - 1)
=
z 2 - (0.9843) z +0.9688
0.0156(1 - z -2 )
=
1 - (0.9843) z -1 + 0.9688z -2
Y (z) 0.0156(1 - z -2 )
= =
X ( z ) 1 - (0.9843) z -1 + 0.9688z -2
Y ( z ) - (0.9843) z -1Y ( z ) + 0.9688z -2Y ( z ) = 0.0156 X ( z ) - 0.0156 z -2 X ( z )
Taking inverse Z-transform of the above equation, we get:

Y ( z ) - y ( n)
z -1Y ( z ) - y(n - 1)
z -2Y ( z ) - y(n - 2)
X ( z ) - x ( n)
z -2 X ( z ) - x (n - 2)
y(n) - (0.9843) y( n - 1) + 0.9688 y(n - 2) = 0.0156 x (n) - 0.0156 x (n - 2)

y(n) = 0.0156 x (n) - 0.0156 x (n - 2) + (0.9843) y(n - 1) 0.9688 y(n - 2)
The above difference equation leads to the direct form implementation as shown in
Figure 10.11.
Figure 10.11 Direct form implementation of digital resonator (numerical problem).

Review Questions
1.
Sketch the ideal low-pass, high-pass, band-pass and band-stop filter magnitude
responses.
2.
Why phase response of digital filters is always negative?
3.
What are the advantages of having linear phase response?
4.
What constraints get imposed on the FIR filter impulse response (filter coefficients)
when it is required to have linear phase response?
5.
Describe the frequency sampling method of FIR filter design.
6.
Compare FIR and IIR filters.
7.
Which type of filter (FIR/IIR) is more suitable for numerical protective relaying
application?
8.
Design a low-pass FIR filter with a cut-off frequency of fsampling/6. Assume order of
the filter to be equal to 32.
9.
In IIR filter, state whether the magnitude of the impulse response is infinite or its
duration is infinite.
10.
What is the basic reason for infinite impulse response?
11.
Write expression for direct time domain implementation of the IIR filter.
12.
Write the frequency domain transfer function, H(z) of the IIR filter.
13.
What are the various alternate ways in which H(z) can be expressed?
14.
What is the motivation for implementing the filter in other than direct form?
15.
Derive bilinear transformation.
16.
What is the range of digital frequencies in Hz and in (pu) radians?
17.
What do you mean by warping of frequency?
18.
How frequency warping can be pre-compensated by pre-warping?
19.
State transfer functions of low-pass Butterworth, Chebyshev and elliptical filters in the
s domain.
20.
Numerical problems:
(i) Design an IIR digital filter based on first order low-pass Butterworth filter prototype
with 3 dB cut-off frequency wc = 0.1p ; by using bilinear transformation.
(ii) Design a digital resonator for a frequency of 100 Hz. Assume sampling frequency
to be 1500 Hz and DF = 4 Hz.
11
Synchrophasors
11.1Introduction [1,4,7,9,12,18,32,35,36,37]
Synchrophasors are marching inexorably into the field of power systems and seem to
have caught the imagination of the protection engineers. They have become the mainstay
of the so called smart-grid and penetrated considerably into the existing power system
transmission network. They have made feasible the concept of wide area protection, control
and monitoring. It is said that If SCADA system is akin to X-Ray then synchrophasor is
the MRI! One single factor which has made the synchrophasor revolution possible is the
affordability GPS technology.
Synchrophasors combine the concept of phasor with that of synchronisation in time. The
phasors were conceptualised by Charles Proteus Steinmetz in 1893. In 1988 synchrophasor
measurement units were invented by Arun G. Phadke and James S. Thorp at Virginia Tech
University in the U.S.A. Thus, Steinmetzs technique of phasor calculation has evolved
into the calculation of real time phasors that are synchronised to an absolute time reference
distributed by the global positioning system (GPS). Early prototypes of the PMU were
built at Virginia Tech. Macrodyne Corporation (U.S.A.) built the first commercial phasor
measurement unit (PMU) in 1992 almost a hundred years after Steinmetzs conceptualisation
of phasor. The concept of synchrophasor was standardised by IEEE in 1995 as IEEE Std
13341995 which was reaffirmed in 2001. A working group was established in 2001 to
update the standard. The new synchrophasor standard was issued in 2005 as IEEE Std
C37.1182005. This was modified in 2011 and broken up into two standards C37.118.1
which specified the measurement part of the PMU and C37.118.2 which specified the
communication aspects. The new standard now includes specifications of PMU output under
transient conditions, about which its previous version was silent. In the following sections
we will first visit the concept of phasor and then develop the idea of synchrophasor as
defined by the IEEE.
216
Synchrophasors217
11.2The Phasor
An electrical signal represented by the equation:
vx = 2 VRMS cos (2p f t + f ) where w = 2 p f
or vx = 2 VRMS cos (w t + f )
can be written in exponential form as:
vx = Re( 2 VRMS e j (w t + f ) )
vx = Re{ 2(VRMS e jf )e j w t }

vx
= Re 2
VRMS e jf e j w t

Instantaneous value Phasor
Thus, while defining the phasor, we drop the term ejwt in the above expression and consider

only the RMS value and phase angle to express the phasorV as:

V = VRMS e jf
with the tacit understanding that frequency w is constant. In phasor calculations we replace

the signal vx by the phasorV . Thus, we have made the reversible phasor transformation denoted
by P( ).

P (vx ) V
Yields
and the inverse operation is:

P -1 (V ) vx
Yields
P ( 2 VRMS cos (w t + f )) VRMS e jf

Yields
and the inverse operation is:
P -1 (VRMS e jf ) 2 VRMS cos(w t + f )

Yields
The reason, for popularity of phasors, is that in the steady state, it is very convenient to
deal with the signal represented in the phasor form rather than in instantaneous form. We
should note, therefore, that phasor is inherently a steady state concept. Conventional phasor
representation is valid for single frequency undistorted sinusoids of constant amplitude, i.e.,
for stationary signals. In a power system, immediately after occurrence of faults or switching
events or during a power swing, signals are not stationary. During these times, it is difficult to
apply classical concept of phasor. However, the recent modification to the IEEE synchrophasor
standard for synchrophasor C37.118.1, provides definition of the synchrophasor under non-
stationary conditions, i.e., when both its amplitude and phase angle are being modulated by a
low frequency signal.
Figure 11.1 shows, the snap-shot of a length Vm rotating at w rad/s. As the length Vm rotates,
its instant to instant projection along the horizontal axis is the instantaneous value of the phasor.
If the length Vm happened to be aligned with the horizonal axis at t = 0 then we associate

a phase angle of 0 with it. Thus, expressing the phasor as V1 = (Vm / 2 )0 . Suppose at time
t = 0 the rotating length Vm happened to be at an angle f with the horizontal axis, then we

associate a phase angle of f with it, thus, expressing the phasor as V2 = (Vm / 2)f . Hence,
the instantaneous value, the peak value, the RMS value and the phase angle are all linked
with the same entity, i.e., the sinusoidal waveform, the phasor captures all this information in
a compact form.
Figure 11.1 Phasor representation.
The phasor diagram is a snap-shot of the rotating phasor of magnitude Vm / 2 rotating

with angular speed w = 2pf. Note that the instantaneous value of the signal is 2 times the
projection of the rotating phasor on the horizontal axis.
11.3The Synchrophasor as per IEEE Std C37.118.1-2011

The phase angle f of the phasor is determined by comparing it with a fictitious cosine waveform
which was at its peak at some absolute instant of time which is taken as reference (t = 0).
The synchrophasor technology has become possible because of the ability, through GPS, to
precisely refer to the t = 0 instant, all over the power system and at all times. The concept of
a phasor is a strictly steady state concept. However, in practise we encounter situations where
the amplitude as well as phase angle are continuously varying according to a low-frequency
signal. The IEEE Standard for synchrophasor now covers both the steady-state and transient
conditions. We first take up synchrophasor definition under steady state conditions followed
by that under transient conditions.
11.3.1The Synchrophasor in Steady State Conditions
Figure 11.2 shows the reference cosine wave, given by vx = Vm cos(w t ), i.e., a cosine wave

with a phase angle of zero, with which any given 50 Hz wave will be compared for its
synchrophasor representation. As defined in the IEEE Std C37.118-2011, the time count starts
from t = 0 instant, which is defined as 00 hr : 00 min : 00 sec on January 1st, 1970. Precise
time-keeping has been maintained from this instant with the help of highly precise 1 pulse per
Synchrophasors219
second signal. At the start of every second (which can be traced back to this t = 0 instant),
a signal known as the 1 PPS signal is broadcasted from the geographical positioning system
satellites, and is available throughout the power system over widely separated geographical
locations, indeed throughout the globe. A small fist sized GPS antenna connected to a receiver
is used to pick up this signal and decode it. Figure 11.2 also shows the phasor diagram for the
reference wave, which will be expressed mathematically as:
V V
Vref = m 0 = m e j 0
2 2
Figure 11.2 Reference cosine wave for synchrophasor representation.
Figure 11.3 shows a signal with amplitude of Vm and phase angle of f leading (i.e.,
positive phase angle). It can be seen that the phase angle of the measured waveform is given
by measuring the occurrence of its peak with respect the rising edge of the 1 PPS signal.
Figure 11.3 Measured signal with peak value Vm and phase angle f.
11.3.2The Synchrophasor in Transient Conditions

In the general case when the amplitude is a function of time expressed as Vm(t) and the sinusoid
frequency is also a function of time f(t), we can define the function g(t) = f(t) f0 where f0 is
the nominal frequency and g(t) is the difference between the actual and nominal frequencies.
Note that frequency, by definition, is rate of change of phase, i.e., w = df /dt
The sinusoid during transient condition can then be expressed as:
v(t ) = Vm (t ) cos(w (t ) t + fo )
Note that {w(t)t} represents the phase which continuously changes with time while f o is the
fixed phase offset.
v(t ) = Vm (t ) cos(f (t ) + fo )
We have
d
w= (f (t ))
dt
d (f (t ))
2p f =
dt
d (f (t )) = 2p f dt
d (f (t )) = 2p f dt
Phase change with time = 2p f dt
Thus, when the input frequency is varying, we can write the input signal as:
(
v(t ) = Vm (t ) cos 2 p f dt + fo )
v(t ) = Vm (t ) cos (2 p ( g(t ) + f )dt + f )
0 o
v(t ) = Vm (t ) cos (2 p f dt + 2 p g(t )dt + f )

0 o
The synchrophasor representation for the above signal is given by:

V (t ) j 2 p g(t ) dt + fo
V= m e
2
Note that unlike the classical phasor, the synchrophasor magnitude as well as phase angle,
both are now functions of time.
For the special case when Vm(t) = Vm is constant and g = Df is a constant offset from the
nominal frequency g(t )dt = D f dt = D f t , so the synchrophasor is simply:
V
V = m e j (2 p D f t + f0 )
2
Synchrophasors221

Note the V will now rotate at uniform rate Df, the difference between the actual and off-
nominal frequency.
11.3.3 Synchrophasor under Dynamic Conditions

Under dynamic conditions, the three phase signal may be represented by
va = Vm[1 + Kx cos(wmt)] cos[w0t + Kacos(wmt p)]
vb = Vm[1 + Kx cos(wmt)] cos[w0t + Kacos(wmt p) 2p/3]
vc = Vm[1 + Kx cos(wmt)] cos[w0t + Kacos(wmt p) + 2p/3]
where Vm is the amplitude of the input signal, w0 is the nominal power frequency, wm is the
modulating frequency in radians per second and fm = wm/2p is the modulating frequency in
Hz, Kx is the amplitude modulation factor and Ka is the phase angle modulation factor. The
phase a positive sequence signal corresponding to the above 3-phase input is given by
Va1 = Vm[1 + Kx cos(wmt)]cos[w0t + Kacos(wmt p)]
Thus, the PMU should produce a positive sequence phasor measurement of:
V [1 + K cos (w t )]
Va1 = m x m
K a cos(w m t p )
2
Note that we can write, t = nT; where n is integer and T is phasor reporting rate, giving
positive sequence a phase phasor as:
V [1 + K x cos (w m nT ) ]
Va1 (nT ) = m K a cos(w m nT p )
2
It may be noted that during steady state the synchrophasor expression did not explicitly
involve power system frequency w0 or time t. It was implicitly implied. However during
dynamic conditions, wm, i.e., the rate at which the synchrophasor frequency is being modulated
and nT, the discrete time at which synchrophasors are reported is explicitly a part of the
synchrophasor expression.
11.3.4 Total Vector Error (TVE)

The theoretical values of the synchrophasor and that estimated by PMU may have differences
in both magnitude as well as phase angle. The IEEE Standard combines both errors by defining
total vector error (TVE). The TVE is defined as:
2 2

Vr ( n ) - Vr ( n ) + Vi (n) - Vi (n)
TVE (n) =
(Vr (n))2 + (Vi (n))2

whereVr (n) and Vi (n) are the sequence of estimates given by the PMU under test and Vr (n)
and Vi(n) are the sequence of theoretical values of the input signal.
As can be seen from Figure 11.4, the maximum error is 1% when the phase angle error is zero
and maximum error in angle is tan -1 (.01OP/OP) = tan -1 (.01) = 0.572938698 @ 0.573.
Figure 11.4 Total vector error.
11.3.5 Frequency and Rate of Change of Frequency (ROCOF)

As per the synchrophasor standard, the PMU must be capable of calculating and reporting
frequency as well as rate of change of frequency (ROCOF). The following definitions are used
for this purpose:
Given a sinusoidal signal:
x (t ) = X m cos[y (t )]
The frequency is defined as:
1 dy (t )
f (t ) =
2p dt
and the rate of change of frequency (ROCOF) is defined as:
df (t )
ROCOF (t ) =
dt
Synchrophasors are always computed in relation to the system nominal frequency f 0. If the
argument of the cosine is represented as:
j (t )
y (t ) = w 0 t +j (t ) = 2 p f0 t +
2 p
then the frequency can be expressed as:
d j (t )
f (t ) = f0 + = f0 + D f (t )
dt 2 p
where Df(t) is the deviation of frequency from nominal.
Synchrophasors223
Hence, ROCOF can be written as:
d 2 j (t ) d
ROCOF(t ) = = [ D f (t )]
dt 2 2 p dt
Frequency in phasor measurements may be reported as the actual frequency f(t) or the
deviation of frequency from nominal, Df(t). In steady state conditions, Df(t) can be represented
as a scalar number.
11.4 Time Tagging (Stamping) of the Synchrophasor

What sets apart the synchrophasor from phasor is called time tag or time stamp. According
to the IEEE Standard, the synchrophasor measurement is to be tagged with UTC (universal
time coordinated) time corresponding to the time of measurement. This shall consist of three
numbers, i.e., a second-of-century (SOC) count, a fraction-of-second count, and a time status
value. The SOC count shall be a 4-byte binary count in seconds from midnight (00 hr:00 min)
of January 1st, 1970 to the present second. Leap seconds shall be added or deleted from this
count as necessary to keep it synchronised with UTC. Insertion of a leap second results in two
successive seconds having the same SOC count which are differentiated by the leap second
bit in the FRACSEC word as defined further. This SOC time stamp is the same as used by
the UNIX operating system and is similar to those used by other computer operating systems
including DOS, MAC OS and network time protocol (NTP).
The second shall be divided into an integer number of subdivisions by the TIME_BASE
integer which is 224. The fraction of second = (fraction of second count (FRACSEC)/224). The
fraction-of-second count shall be zero when it coincides with the 1s rollover. The structure of
the time stamp is shown in Figure 11.5.
Figure 11.5 Structure of the synchrophasor time stamp.
The SOC counter started filling up at 00:00 on January 1st, 1970 and is getting incremented
every second. One may wonder how long it will take for the counter to become full and hence
over-run in the next second. It can be calculated as:
Seconds sec min hr day
= = (60)(60)(24)(365) = 31536000
Year min hr day year
Capacity of the SOC counter = 232 = 4294967296

Number of years after which SOC counter will become full:
Capacity of SOC conuter 4294967296
= = 136.1925... years.
Seconds/year 31536000
Thus, it will be 136 years from January 1st, 1970, i.e., 2106 A.D., by the time SOC counter
over-runs. In all probability long before this happens, the technology will have changed and
hopefully will not cause something like the infamous Y2K problem during roll-over from 1999
to 2000.
The exact number of seconds, including the fractional part, since January 1st, 1970 can be
computed from the time stamp as:
FRACSEC
Exact number of seconds = SOC +
224
The PMU has to communicate the measurements with other devices such as phasor data
concentrators. Some protocol has to be used for such communications. The protocol allows
for necessary identifying information, such as the PMU IDCODE and status to be embedded
in data frame for proper interpretation of the measured data. Four message types are defined,
i.e., data, configuration, header, and command. The first three message types are transmitted
from the PMU and the last, i.e., command is received by the PMU. Configuration is a machine
readable message describing the data that the PMU sends and providing calibration factors.
Header is human-readable descriptive information sent from the PMU but provided by the
user. Commands are machine-readable codes sent to the PMU for control and configuration.
Information can be stored in any convenient form in the PMU but when transmitted it shall be
formatted as frames. Figure 11.6 shows the configuration of a data frame sent out by a PMU.
Figure 11.6 Data frame transmission order.
11.5 Dissemination of the Time Stamp: IRIG-B Standard

IRIG-B (Inter Range Instrumentation Group, B is for utility industry) is fully described in
the IRIG Standard 200. Time is provided once per second in seconds through day of year
in a binary coded decimal (BCD) format and an optional second of the year count. IRIG-B
signal may be provided in a level shift, a 1 kHz amplitude modulated signal or in a bi-phase
Manchester modulated format. The IRIG-B amplitude modulated format is the most commonly
used. The Manchester format is more compatible with fibre optic and digital systems and
provides complete synchronisation without additional signals.
The time code format is:
<sync> SS: MM: HH: DDD <control> <binary seconds>
where <sync> is the on-time sync marker
SS is the second of the minute [00 to 59] (60 during leap second)
MM is the minute of the hour (00 to 59)
HH is the hour of the day in 24 hour format (00 to 23)
Synchrophasors225
DDD is the day of the year (001 to 366)

<control> is a block of 27 binary control characters
<binary seconds> is a 17-bit second of the day in binary
11.6 Synchrophasor Reporting Rates

As per the IEEE Std C37.118, the phasor measurement unit (PMU) shall support data reporting
(by recording or output) at sub-multiples of the nominal power-line (system) frequency.
Standard reporting rates for the 50 Hz and 60 Hz systems are listed in Table 11.1.
Table 11.1 Synchrophasor data reporting rates
System frequency 50 Hz 60 Hz
Reporting rates in frames per second 10 25 10 12 15 20 30
11.7Block Diagram of the Synchrophasor Enabled Digital Relay

or Phasor Measurement Unit (Pmu)
Figure 11.7 shows the architecture of a synchrophasor enabled digital relay. It can be seen
that it has been evolved from the block diagram of the digital relay given in Chapter 2, by
adding the GPS clock and a phase locked oscillator locked to the 1 PPS signal given by the
GPS clock. The GPS receiver also provides the DSP microprocessor with the time stamp using
IRIG-B protocol.
A phasor measurement unit will have very similar organisation. To difference between a
synchrophasor enabled digital relay and a PMU is that the PMU just computes the phasors,
time stamps them, assembles the data frames and either stores them (records them) or transmits
them to a remote entity which may be a phasor data concentrator or a digital computer, whereas
a digital protective relay is tasked with taking a trip/restrain decision after computing the
synchrophasors.
11.7.1 Two Alternatives for Sampling

We need to give careful thought to the effect of sampling frequency on the phasor computation,
because in a real life power system the signal frequency is almost never exactly equal to
the nominal frequency of 50 Hz (or 60 Hz). If we lock the sampling frequency to the
nominal frequency of 50 Hz, as we do in synchrophasors, we will end up with inaccurate
measurement due to the leakage effect. This inaccurate measurement will have to be corrected
or compensated by accounting for the actual signal frequency.
If we wish to get rid of this error (rather than compensating it) then we will have to lock
the sampling to the actual power system frequency by actually measuring it. Thus in both
cases extra work, than that indicated by the DFT algorithm is needed. Both the approaches
are shown schematically in Figures 11.8 and 11.9. As already mentioned in synchrophasor
Figure 11.7 Block diagram of a synchrophasor enabled relay or phasor measurement unit.
Synchrophasors227
approach, we keep the sampling frequency locked to a fictitious cosine wave whose positive
peak synchronised with 00 hr: 00 min : 00 sec on January 1st, 1970 as shown in the figure.
Figure 11.8 Sampling locked with a fictitious 50 Hz Signal via GPS.
Figure 11.9 Sampling locked with system frequency.
11.8 Off-nominal Frequency Operation of the PMU

It is well-known that the power system frequency is never exactly equal to the nominal
frequency of 50 Hz or 60 Hz. The power system frequency is kept very close to the nominal
value and in the worst case can vary within a few hertz around the nominal frequency for a short
duration. However, the sampling is kept locked to the fictitious 50 Hz cosine wave starting at
t = 0 instant. This can be expected to cause some error in the DFT computation, which needs
to be investigated and accounted for. We first demonstrate the effect of off-nominal frequency
operation by carrying out a 4-point DFT computation by hand and then follow it up with a
small MATLAB script and then develop analytical expression for the same.
11.8.1 A
Demonstration of Error Caused by Off-nominal Frequency
Operation
Let the nominal frequency be 50 Hz and the sampling be done at four times the nominal
1 1
frequency, i.e., at 200 Hz. Hence, the sampling time period = Dt = = = (5)10 -3 s.
fsampling 200
Now let us consider two phasors, both of value 1 0 but one having frequency of
50 Hz and another of 51 Hz and sample both of them at 200 Hz for one cycle thus drawing
four samples of each. From these four samples let us evaluate the DFT and compare the
computed and actual values, thus finding errors, if any. The samples of the two signals
are shown in Table 11.2.
Table 11.2 Samples of 50 Hz and 51 Hz signal

Sample number n t = n Dt v50 = (1) cos(2 p 50 n Dt) v51 = (1) cos(2 p 51 n Dt)
0 0 1 1
1 (1)(5.103) 0 0.03141
2 (2)(5.103) 1 0.9980
3 (3)(5.103) 0 0.0941
The 4-point DFT of the 50 Hz signal is given as:

3 - j 2 p 1n 3 - jp n 3
v50 ( n) e 4 = v50 ( n) e 2 = v50 (n) (-90(n))
n =0 n =0 n =0
V50 = v50 (0) ( -90(0)) + v50 (1) ( -90(1)) + v50 (2) ( -90(2)) + v50 (3) ( -90(3))
V50 = 1 ( -90(0)) + 0 ( -90(1)) - 1 ( -90(2)) + 0 ( -90(3))
V50 = 1 ((0)) - 1 ( -180) = 2 0
2 2
Hence, actual (peak) value of V50 = V50 = 20 = 10 .
N 4
Thus, there is no error, either in magnitude or in phase angle when the 50 Hz phasor is sampled
at 4 50 Hz frequency. Let us do similar calculations on the samples of the 51 Hz signal as
shown below:
3 - j 2 p 1n 3 - jp n 3
v51 (n)e 4 = v51 (n)e 2 = v51 (n) ( -90(n))
n=0 n=0 n=0
V51 = v51 (0) ( -90(0)) + v51 (1) ( -90(1)) + v51 (2) ( -90(2)) + v51 (3) ( -90(3))
V51 = 1 ( -90(0)) - 0.03141 ( -90(1)) - 0.9980 ( -90(2)) + 0.0941 ( -90(3))
V51 = 2.001965 ( 3.5946)
2 2
Hence, actual (peak) value of V51= V51 = 2.001965(3.5946) = 1.00098273.5946 .
N 4
Thus, there is error, both in magnitude and in phase angle when the 51 Hz phasor is sampled
at 4 50 Hz frequency. It can be seen that the error in magnitude is only 0.098% which
is less than 0.1% while the error in phase angle is 3.5946. Thus, phase angle error in
synchrophasors when off-nominal frequency signals are encountered is more serious than
the error in magnitude.
11.8.2 Analytical Expression for Effect of Off-nominal Frequency Operation

We can write the expression for phasor as:
N -1 - j 2pn
2
X=
N
x[ n Dt ]e N
n=0
Synchrophasors229

where X is the one-cycle phasor estimate,
1 1
Dt = =
fsampling N f nom
where fnom is the nominal frequency (50 Hz or 60 Hz). It is assumed that N is even. Let the
sample belong to an off-nominal frequency f . Hence, from the definition of a phasor we can
express the sample as:

x[ nDt ] = 2 Real [ X e j 2p f nDt ]

where X is the true phasor value of the samples.
j 2p f n
x[ nD t ] = 2 Real X e N fnom
The one cycle phasor estimate can then be written using the DFT equation for an off-
nominal frequency signal as:
2 N -1
j 2p f n - j 2p n
X=
N
2 Real X e Nfnom e N
n=0

Now for any phasor P , the Real ( P) can be written as

P + P*
Real (P) =
2
Hence, we can write the above equation during kth window as:
2 k + N -1 N fnom * N fnom N

j 2p f n - j 2p f n - j 2p n

X=
2N n=k
X e + X e e
k + N -1 - j 2p n 1 k + N -1 - j 2 p f n - j 2p n
j 2p f n
1
X= X e N f nom e N + X * e N fnom e N
N n=k N n=k
1 k + N -1 1 k + N -1 - j 2 p f n - j 2 p n
j 2 p f n j 2p n
-
X = X e N f nom N + X * e N fnom N

N n=k N n=k
j 2p n f k + N -1 - j 2p n f +1
1 k + N -1 N -1
f nom + 1 X * e N fnom
X = X e
N n=k N n=k
n n
j 2p f k + N -1 - j 2p f
1 k + N -1 -1
N f nom f +1
X = X e + 1 X *
e
N nom
N n=k N n=k
Let n k = q, hence (n = q + k) and when n = k; n k = q = 0 and when n = k + N 1;

n k = q = N 1, hence we can write the above equation as:
q+k q+k
j 2p f - j 2p f
1 N -1 N fnom -1 1 * N -1 N fnom +1
X = X e + X e
N q =0 N q =0
k q k q
j 2p f N -1
j 2p f - j 2 p f +1 N -1
- j 2p f
1 -1
N f nom
-1
N f nom + 1 X * e N fnom
+1
N f nom
X = X e
e e
N q =0 N q =0
Using closed form summation formula:

Nq
sin
2 j ( N -1) 2
N -1 q

jq n
(e ) =
q
e
n=0
sin
2
We can write the estimate of phasor for the off-nominal frequency signal as:
N 2p f
k
sin - 1
N fnom -1
j 2p f f 1 2p f
2 N nom j ( N -1) 2 N f -1

First term = X e e nom
1 2p f
N sin f - 1
2 N nom
N -2 p f
k
- j 2 p f +1 sin 2 N f + 1 -1 2 p f
j ( N -1) 2 N f +1
Second term = X * e
N f nom nom nom
e
-1 2 p f
N sin f + 1
2 N nom
Simplifying them as:

p f
j 2p f
k
sin - 1 p f
N fnom
-1
1 f nom j ( N -1) N fnom -1
First term = X e e
p f
N sin - 1
N f nom
Noting that
f f - f nom w - w nom f w + w nom
- 1 = = and + 1 =
f nom f nom w nom f nom w nom
Synchrophasors231
w - w nom
j 2 p k w -w nom sin p p w -w nom
N w w nom e
j ( N -1)
w nom
First term = X e nom N
p w - w nom
N sin
N w nom
Similarly,
w + w nom
- j 2 p k w +w nom sin p - j ( N -1) p w +w nom
w w nom e w
nom nom
Second term = X e
N N
p w + w nom
N sin
N w nom

Hence, expression for the off-nominal frequency DFT in terms of the true DFT X can be
written as:
j 2p k w -w nom - j 2p k w +w nom
N w nom N w nom
X = AXe + BX * e
w - w nom
sin p w -w nom
w nom
p
j ( N -1)
where A= e
N w
nom

p w - w nom
N sin
N w nom
w + w nom
sin p w +w nom
w nom - j ( N -1) N
p
and B = e
w
nom

p w + w nom
N sin
N w nom
The variation of magnitudes of A and B as fnom changes from 45 Hz to 55 Hz, i.e., a
change of 5 Hz as shown in Figures 11.10 and 11.11.
Figure 11.10 Variation of A for off-nominal frequency.

Figure 11.11 Variation of B for off-nominal frequency.
Further it can be shown that the locus of all phasor estimates, for a given phasor amplitude
and variable phase angle which is a circle at the nominal frequency happens to be an ellipse
at off-nominal frequency.
11.8.3 A MATLAB Script to Demonstrate Off-nominal Frequency DFT

In this MATLAB script we demonstrate the effect of off-nominal frequency upon the DFT
magnitude calculation when the DFT computation is based on the samples drawn at multiples
of nominal frequency. In this script we sample a 50 Hz, a 45 Hz and a 55 Hz signal at
4 50 Hz. We keep the amplitude of the signal constant but change its phase from 0 to 360.
As expected the 50 Hz signal traces out a circle but the 45 Hz and 55 Hz signals trace out
ellipses. Thus, off-nominal frequency operation of DFT whose sampling is locked to nominal
frequency gives rise to error in phasor calculation.
% Demonstration of off-nominal frequency operation of DFT
% Comparison between DFT of 50 Hz sig and signals of 45 Hz and 55 Hz
% all sampled at 4*50 Hz
clf , clear , clc
f0 = 50; f1 = 45; f2 = 55; N = 4;
fsamp = N*f0;
dT = 1/fsamp;
phase = 0 ;
ph_step = pi/36;
p = 0;
for phase = 0 : ph_step : 2*pi
p = p+1;
ph_angle(p) = phase;
n = 0;
for k = 0 :N-1
n = n+1;
v50(n) = 100*cos((2*pi*f0*k*dT)+(phase));
voff_nom1(n) = 100*cos((2*pi*f1*k*dT)+(phase));
voff_nom2(n) = 100*cos((2*pi*f2*k*dT)+(phase));
end
Synchrophasors233
v50;
voff_nom1;
V50spectrum = fft(v50);
V50_complex(p) = (2/N)*V50spectrum(2);
V50mag(p) = abs(V50_complex(p));
voff_nom1spectrum = fft(voff_nom1);
voff_nom1_complex(p) = (2/N)*voff_nom1spectrum(2);
voff_nom1mag(p) = abs(voff_nom1_complex(p));
err_mag(p) = voff_nom1mag(p)-100;
voff_nom1angle(p) = (angle(voff_nom1_complex(p)))*(180/pi);
voff_nom2spectrum = fft(voff_nom2);
voff_nom2_complex(p) = (2/N)*voff_nom2spectrum(2);
voff_nom2mag(p) = abs(voff_nom2_complex(p));
err_mag(p) = voff_nom2mag(p)-100;
voff_nom2angle(p) = (angle(voff_nom2_complex(p)))*(180/pi);
end
figure(1)
polar(ph_angle,V50mag)
title(Variation magnitude w.r.t. phase for 50Hz signal sampled at
4*50 Hz)
figure(2)
polar(ph_angle,voff_nom1mag)
4*50 Hz)
figure(3)
polar(ph_angle,voff_nom2mag)
4*50 Hz)
Figure 11.12 Variation of magnitude w.r.t. phase for 50 Hz signal sampled at 4 50 Hz.
11.9 Synchrophasor Applications

Power system is spread out over a very large geographical area. Normally majority of protective
relays make use of locally available information. Carrier aided schemes make use of information
at the remote end to take a more robust decision. Now, with proliferation of synchrophasor
technology, information from diverse locations can be made available anywhere in the system
to enhance the relaying decision. This opens up a large number of novel applications which
can provide wide area protection and control. Some of these applications are as follows:
1. State estimation: The present state estimation approach consists in measuring active and
reactive power flows and voltage magnitudes at substations, and then communicating them to
a central site for processing. The data is acquired in a time window of seconds to minutes.
Synchrophasors235
Thus, the calculated state is at best an approximation to an averaged system state. Hence, the
estimates that are produced by the state estimation program are called static state estimates.
Synchronised phasor measurements of positive sequence bus voltage and currents most
naturally lead to state estimates. In the utopian view of the power system, every location is
equipped with synchrophasor measurements and information from all is available at a central
location. Thus, we have is a real-time state measurement rather than state estimation. However,
it is not necessary to have synchrophasor at every location to perform state estimation. It
can be shown that we can take measurements from a few key locations and perform state
estimation. This is possible because we can divide the network into observable locations where
synchrophasor are located and unobservable locations where synchrophasor are not located
but estimates can be indirectly calculated. From estimated values at observable locations, we
can make close estimates at unobservable locations.
2. Power system control: Prior to synchrophasor measurements, all control in power systems
used local measurements and a mathematical model of the larger power system was used to
arrive at control decisions. It was recognised that such controllers are rarely optimum and
could produce completely unacceptable response to system phenomenon when the models are
inaccurate. Synchrophasor measurements can improve such control actions. Typical applications
to problems where the control objectives were global in nature, i.e., for example an HVDC
controller may be called upon to damp electromechanical oscillations between two widely
separated areas of the power system. The high speed synchronised phasor measurements
provided by PMU offers a very attractive alternative to the problem of power system controllers.
Remote measurements could be brought to the controlled device at high speed, and the remote
inputs used as feedback signals in the controller. Phasor data are being used increasingly by
individual utilities to manage real-time grid operations.
3.Wide area power system protection: An additional protection area where phasor
measurements play a role is that of adaptive security and dependability. The existing protection
system has redundant primary protection coupled with multiple backup schemes. The resulting
system is highly dependable in that virtually every fault is ultimately cleared. The trade-off
is false trips which are tolerated. As the system has evolved, however, the fact that false
trips during large disturbances exacerbate the disturbance and allow it to cascade has been
recognised. The solution is to adaptively alter the security dependability balance during times
of stress. Remote phasor measurements can be used to determine when this happens. The
mechanism is to alter the relaying logic which normally lets any relay which sees fault to trip
the breaker to logic that demands a vote such as two out of three. This controlled dependability-
security system is an example of adaptive relaying. Other adaptive relaying applications
include distance relay zone adjustments, transformer protections, reclosing operations, out-of-
step relays, etc. Another popular innovation in modern protection systems is the increasing use
of remedial actions schemes (RAS) now renamed as system integrity protection (SIPS). These
schemes use wide area measurements to identify system conditions under which extraordinary
actions must be taken to avoid major system outages. Since the synchrophasor measurement
provides very detailed and precise time synchronised measurements of the state of the electrical
grid, the analysis of past electricity blackouts has indicated that monitoring synchrophasors
has the potential to detect system instability much earlier than current supervisory control and
data acquisition (SCADA) systems. This will prevent large-scale blackouts by giving system
operators more time to respond to disruptive events enabling them to take corrective action
before system stability is affected. Synchrophasor projects involving the installation of hundreds
of phasor measurement units (PMUs) are currently taking place throughout the globe.
4.Forensic analysis: The dictionary meaning of the word forensic is relating to the
application of science to decide questions arising from crime or litigation. In case of major
system events like blackouts, the data provided by PMUs becomes the basis on which the
disturbance can be analysed. Since, PMUs collect and store high volumes of high-speed time
synchronised data about the conditions across an interconnection, those data can be compiled
quickly and analysed to determine the sequence of events which caused the disturbance. Phasor
data were essential in the recent investigations of blackouts in Europe and the USA.
5. Smart grid: At the transmission and generation level, synchrophasor systems are the single
most effective technology element to realise and implement smart grid, because synchrophasor
systems collect, distribute and analyse critical data and convert it in real time into information
and insight that improve grid automation and operation.
Review Questions
1. Explain the concept of a phasor.

2. What is the difference between a phasor and a synchrophasor?
3. Discuss the importance of a global reference time signal for the synchrophasor.
4. What are the various ways in which standard time signal can be disseminated?
5. Describe the GPS system.
6. Describe the time stamp generated by the PMU.
7. What is the effect of off-nominal frequency operation of the power system upon the
computation of the synchrophasor?
8. Calculate the error in calculation of phasor when the actual frequency is 51 Hz. Assume
that the sampling clock is locked to 50 Hz and sampling at 24 samples per cycle of
50 Hz.
9. Repeat question 8 when the actual system frequency is 49 Hz.
10.
What are the applications of the synchrophasor technique in a large power system?
11.
What is the relationship of the synchrophasors with the state estimation of the power
system?
12.
How does the data generated by the PMUs help in forensic analysis of the power
system event?
13.
What are the standard synchrophasor reporting rates in a 50 Hz system?
14.
For a synchrophasor reporting rate of 10 times per second, what is the maximum
deviation in frequency that can be reported without error?
15.
How frequency and rate of change of frequency are defined in the synchrophasor
standard?
16.
Explain the concept of total vector error (TVE).
12
Removal of DC Offset
12.1Why Decaying DC Offsets are Created? [61]

A typical power system is predominantly inductive in nature. If we ignore the shunt capacitances,
the system can be represented as a series R-L circuit. It is well known that whenever there is
a switching or sudden change in an R-L circuit, it is accompanied by a transient which dies
down at a rate decided by the time constant of the R-L circuit. For the sake of relaying, we
model the system as shown in Figure 12.1. The voltage is expressed as v(t) = Vm sin(w t + a);
where a is the switching angle which decide the fault instant with respect to the zero crossing
of the voltage waveform.
Figure 12.1 Model of power system.
The current in the above circuit is given by writing the KVL around the loop as:
di(t )
v(t ) = R i(t ) + L
dt
237
Substituting the value of v(t), i.e.,

di(t )
Vm sin(w t + a ) = Ri(t ) + L
dt
Solving the above differential equation for i(t) gives:
i(t ) = I m sin(w t + a - f ) - I m sin(a - f )e - t / t

Vm wL L
Where I m = ; Z = R 2 + X 2 ; f = tan -1 and t =
Z R R
Thus, it can be seen from the expression for i(t) that when a f = np; n = 0, 1, 2,,
sin(a - f ) = sin( np ) = 0, hence there will be no DC offset. The DC offset will be maximum
when a - f = n p /2. It may be noted that I m sin(w t + a - f ) is the steady state fault current
and I m e - t / t sin(a - f ) is the transient fault current also called as the decaying DC offset. It
is well known that decaying DC term will become insignificant after a time equal to 5 time
constants after the instant of disturbance. Hence, DC offset is important only when the current
values immediately after the fault are to be used as in case of high speed distance relays. For
many slow applications, it may not be necessary to take cognizance of the DC offset. The
way we have modelled the power system as shown in Figure 12.1, it is clear that there cannot
be any DC offset in the voltage. However, in practise DC offset can exist in voltage as well.
Figure 12.2(a) shows steady state waveforms of voltage and current. The current lags the
voltage by 90 in the steady state. In Figure 12.2(b), the instant of switching is chosen at the
positive going zero crossing of voltage waveform giving rise to maximum dc offset in the
current.
Figure 12.2 Decaying DC offset in current.

Removal of DC Offset239
12.2What is the Effect of Decaying DC Offsets up on Relays?

[33,34]
The expression for current immediately following the fault can be written as:
i(t ) = iAC (t ) + iDC (t )
where
iAC = I m sin(w t + a - f )
iDC = - I m e - t / t sin(a - f )
Hence, the RMS value of the current can be written as:
I RMS = I AC
2
+ I DC
2
Thus, it can be seen that the RMS value of current with DC offset is more than the RMS
value of the pure ac current. This can cause fast acting over current as well as distance relays
to over-reach. Further, presence of decaying DC offset is troublesome for the DFT algorithm.
The classical DFT algorithm which is based on Fourier analysis gives correct results when a
constant DC is present in the signal but gives erroneous results when decaying DC offset is
present. This is because while a constant DC (in time domain) shows up at only zero frequency
(in frequency domain), a decaying dc offset in the time domain has a wider frequency spectrum
and thus corrupts all the DFT bins. This can be seen from the Fourier transform of eat, denoted
by FT( ), as shown below:
1
FT(e -a t ) =
a + jw
Numerous researchers have worked on this problem and there is a large body of literature
on this topic. We discuss three approaches is the next sections. Interested reader is urged to
follow the literature on this subject.
12.2.1 Mimic Impedance Method of Removal of DC Offset

Traditionally dc offset has been successfully dealt with by connecting mimic impedance in the
CT secondary circuit which has the same L/R ratio as the primary circuit. Figure 12.3 shows the
response of mimic impedance to the exponentially decaying part of the CT secondary current.
The net voltage developed across the mimic impedance can be written as:
di(t )
v(t ) = KR i(t ) + KL
dt
t/t
substituting the decaying dc offset given by I0e ; we get:
d I 0 e-t / t
vmimic = KR I 0 e - t / t + K L
dt
1
vmimic = KR I 0 e - t / t - KL I 0 e - t / t
t
Figure 12.3 Response of mimic impedance to decaying DC offset.
Noting that t = L
R R -t / t
vmimic = KR I 0 e - t / t - KL I 0 e
L
vmimic = KR I 0 e - t / t - KR I 0 e - t / t = 0
Thus, the mimic impedance completely rejects the decaying dc offset. However, one must
remember that this complete rejection was made possible because the mimic circuit parameters
had been exactly matched with the time constant of the decaying DC current. This may not
always be possible and may lead to incomplete rejection of the DC offset.
12.2.2 Immunity of LSQ Method to Decaying Dc Offset

Readers will recall that in the derivation of the LSQ algorithm, we had assumed the signal
consisting of a decaying DC offset in addition to the fundamental and harmonics. The curve-
fitting done by the LSQ algorithm was under this explicit assumption. Hence, LSQ algorithm
can successfully filter out the DC offset. Therefore, presence of decaying DC offset in the
relaying signals is not detrimental to the operation of the LSQ algorithm.
12.2.3 Immunity of Differential Equation Algorithm to Dc Offsets

In the differential equation algorithm, we use the instantaneous values of voltage and current
samples. The equation v(t ) = R i(t ) + L (di/dt ) is valid without any regard to the presence of
a decaying DC offset. Hence, methods based on solution of differential equation (DEA)
inherently accommodate the DC offset.
12.3 Digital Mimic Impedance Based Filter [3]

In the mimic impedance method the CT secondary current signal is passed through the mimic
impedance giving the voltage across the mimic impedance as:
di(t )
v(t ) = KR i(t ) + KL
dt
the equation can be discretised as:

di(tk )
v(tk ) = KR i(tk ) + K L
dt
i(t ) - i(tk -1 )
v(tk ) = KR i(tk ) + K L k
Dt
Taking Z-transform of the above equation gives:
KL
V (z) = K R I (z) + [ I ( z ) - Z -1 I ( z )]
Dt
K L I (z)
V (z) = K R I (z) + [1 - Z -1 ]
Dt
{
V (z) = K R I (z) +
L I (z)
R Dt } L
[1 - Z -1 ] ; Noting that = t
R
{
V (z) = K R I (z) +
t I ( z)
Dt }
[1 - Z -1 ] ; representing
t
Dt
= t1
V ( z ) = K R I ( z ){1 + t 1[1 - Z -1 ]}
V (z)
= K R {1 + t 1[1 - Z -1 ]}
I (z)
V (z)
= H ( z ) = K R {(1 + t 1 ) - t 1 Z -1 ]}
I (z)
which is written as:
V (z)
= H ( z ) = K m {(1 + t 1 ) - t 1 Z -1 ]}
I (z)
where Km = K R can be thought of as a gain term.
The gain Km is set in such a way that at rated frequency, the filter gain will be one.
Substituting Z = e j w T = e j 2 p f T = e j 2 p 50 T ; we can write:
Gain at 50 Hz = K m {(1 + t 1 ) - t 1 e - j 2 p 50 T ]}
where T is the time period corresponding to the sampling frequency which can be expressed
as T = (1/fsampling). Hence,
- j 2 p 50 /fsampling
K m {(1 + t 1 ) - t 1 e }=1
2 p 50 2 p 50
K m (1 + t 1 ) - t 1 cos + j t 1 sin = 1
fsampling fsampling

2 2
2 p 50 2 p 50
Km (1 + t 1 ) - t 1 cos + t 1 sin =1
fsampling fsampling
1
Km =
2 2
2 p 50 2 p 50
(1 + t 1 ) - t 1 cos + t 1 sin
fsampling fsampling
It can be easily appreciated that the mimic filter is a high-pass filter will therefore amplify
any noise contained in the CT secondary current.
12.4 Method of Partial Sums for Filtering DC Offsets [12,40]

A signal containing a decaying DC offset, fundamental and upto (N/2 1) order harmonics
can be represented in its discrete form as:
N
-1
2 2p k
Sk = A0 r + k
Am sin
N
m + jm

m =1
where k = 0 to (N 1), Am = Vm /Z, A0 = Amsin(a f) and r = e(Dt/t).

If we define a partial sum PS1 as:
PS1 = S1 + S3 + + SN1
then we can write PS1 as:
N / 2 (( N / 2) -1)
N /2
2p (2k - 1)
PS1 = A0 r 2 k -1 + Am sin m + jm
k =1 k =1 m =1 N
It can be shown that:
N / 2 (( N / 2) -1)
2p (2k - 1)
sin
N
m + jm = 0

k =1 m =1
Thus, giving
N /2
PS1 = A0 r 2 k -1
k =1
which can be written in the form highlighting the underlying geometric series as:
N /2
PS1 = A0 r -1 r 2 k
k =1
N /2
PS1 = A0 r -1 (r 2 ) k
k =1
Using closed form formula for summation of geometric series, we get:

N
+1
-1 (r 2 ) 2 - r2
PS1 = A0 r
r2 - 1
r N r2 - r2
PS1 = A0 r -1
r2 - 1
rNr - r
PS1 = A0
r2 - 1
r (r N - 1)
PS1 = A0
r2 - 1
Similarly, we can define another partial sum PS2 as:

PS2 = S0 + S2 + S4 + + SN2
which can be expressed as:
N N N
-1 -1 -1
2 2 2 2p (2k )
PS2 = A0 r 2 k + Am sin N
m + jm

k =0 k = 0 m =1
Using similar reasoning, the second term in the above equation turns out to be zero, giving:
N
-1
2
PS2 = A0 r 2 k
k =0
Putting in the form of sum of geometric series as:

N
-1
2
PS2 = A0 (r 2 )k
k =0
(r 2 ) N/2 - (r 2 )0
PS2 = A0
r2 - 1
rN - 1
PS2 = A0
r2 - 1
From equations for PS1 and PS2, we can write:
PS1
r=
PS2
r2 - 1
and A0 = PS2
(r N - 1)
Having the values of r and A0, the samples can be modified as:
Sknew = Sk A0rk; k = 0 to N 1
This, new set of samples, is free from decaying DC offset and can be used for further
processing.
12.5 Characterising the Decaying DC Offset by Integration [6]

Let us express the fault current as shown below:
p
i(t ) = I 0 e - t / t + ( I k sin(k w1 t + q k ))
k =1
where I0 is the peak value of the dc offset, t is the time constant of the decaying dc offset,
k is the harmonic order, Ik is the peak value of the kth harmonic and qk is the phase angle of
the kth harmonic component and p is the maximum harmonic order. If the above expression
is integrated for one time period T of the fundamental, then second term evaluates to zero,
and we are left with integral of the decaying dc offset term as shown below:
t t
p
i(t )dt = I
0 e -t / t
+ ( I k sin(k w1 t + q k )) dt
t -T t -T k =1
t
-t / t
= I0 e dt = - I 0t [e - t / t ]tt -T
t -T
= - I 0t (e - t / t - e - (t -T ) / t ) = - I 0t (e - t / t - e - t / t eT / t )
= - I 0t e - t / t (1 - eT / t )
Let us call the above result of the integral as Z(t). Hence,
Z (t ) = - I 0t e - t / t (1 - eT / t )
Now, the expression for Z(t + Dt), which is the integral of the dc component Z(t) extended
by a small interval of time Dt can be written as:
Z (t + Dt ) = - I 0t e - (t + Dt ) / t (1 - eT / t )
Z (t + Dt ) = - I 0t e - t

/t
(1 - eT / t) e -Dt / t
Z (t )
Thus, we can write Z(t + Dt) in terms of Z(t) as:
Z (t + Dt ) = Z (t )e Dt/t
Simplifying further,
Dt Z (t + Dt )
- = ln
t Z (t )
Thus, the time constant of the decaying dc offset t is found out as:
Dt
t =-
Z (t + Dt )
ln
Z (t )
and I0 can be found out as:

Z (t )
I0 = -t / t
-t e (1 - eT / t )
Thus, once I0 and t are found out, we can completely characterise the decaying dc offset
and compensate the raw signals to extract decaying offset free part of the signal as explained
earlier.
Review Questions
1.
Clearly state the difference between (constant) dc offset and decaying DC offset.
2.
Sketch the frequency spectra of a decaying dc offset, thus explain why it corrupts all
the DFT bins.
3.
DFT cannot filter out the effect of decaying dc offset while it does so for constant
dc offset. Explain.
4.
Describe how decaying dc offsets are created in ac power system.
5.
What is the problem caused by existence of decaying dc offsets?
6.
Explain whether decaying DC offset is present in output of CT or output of PT or
both.
7.
Out of current and voltage signals which one is more likely to exhibit decaying dc
offset?
8.
Explain the effect of decaying DC offset on OC relays.
9.
Explain the effect of decaying dc offset on distance relays.
10.
Explain the mimic impedance method of filtering decaying dc offset.
11.
Explain the digital implementation of the mimic impedance for filtering out decaying
dc offset.
12.
Explain why the mimic impedance acts as a high-pass filter. What are the implications
of this?
13.
Explain the method of partial sums to filter out decaying dc offset.
14.
Explain how dc offset can be characterised by integration over the fundamental time
period, and this can be used to filter out the decaying dc offset.
Appendix
A.1The GPS system

GPS (geographical positioning system) is a system of satellites owned and operated by the U.S.
Department of Defense for the purpose of navigation throughout the world. Being a defense
system it is very robust and is designed in such a way that it will provide navigation signal
throughout the world without fail in the worst possible scenario. This includes all weather
conditions, system degradation and interference. The launching of all the 24 satellites of the
GPS system was completed in 1994. It became fully operational in 1995. As of December
2012 there are 32 satellites in eight orbital planes. At any location at least four satellites are in
view at any time. Hence, even sites with restricted sky-view keep receiving the GPS signals.
Figure A.1 Orbits of the GPS satellites.
Range and time is broadcast on three L-band frequencies. Time can be derived from
the coarse/acquisition (C/A) code transmitted at 1575 MHz. All satellites broadcast at the
same frequencies. Signals are encoded using code division multiple access (CDMA) allowing
messages from individual satellites to be distinguished from each other based on unique
encodings for each satellite (that the receiver must be aware of). Two distinct types of CDMA
247
248Appendix
encodings are used, i.e., the coarse/acquisition (C/A) code, which is accessible by the general
public, and the precise (P(Y)) code, which is encrypted so that only the U.S. military can
access it. The signal can be received by a simple omnidirectional antenna. The spread-spectrum
technique that is used, makes the signal resistant to interference. At the heart of the GPS is
a ground based cesium clock ensemble that itself is referenced to the universal coordinated
time (UTC), which is the world time standard. Each satellite provides a correction to the UTC
time that the receiver automatically applies to the outputs. With continuous adjustments, the
time accuracy is limited only by short term signal reception whose basic accuracy is 0.2 s.
Baseline accuracy can be improved by advanced decoding and processing techniques. For
general applications a 0.5 s is a safer figure to depend on. The short term frequency accuracy
of the received signal is of the order of 10E-11.
One might wonder how ordinary public throughout the world can access the U.S. Department
of Defense satellite but access issues were largely resolved in the 1990s with agreements to
add new civilian frequencies and culminating in termination of the selective availability signal
degradation of the C/A code in the year 2000. The U.S. Department of Defence and Department
of Transportation have committed to make GPS available to civilian users at all-time except
in national emergency.
A.2 Comtrade
COMTRADE (Common format for Transient Data Exchange for power systems) is a file
formatfor storing oscillography and status data related to transient power system disturbances.
COMTRADE files are typically generated by Intelligent Electronic Devices (IEDs), such
as a protective relay, in electrical substations. Analysis tools can download the COMTRADE
file and calculate useful information related to the power system. COMTRADE files from
multiple substations can be used collectively to perform forensic analysis of large scale power
disturbance events (e.g., blackouts) to determine the root cause of the disturbance, help to
improve system protection and guide future mitigation strategies.
The COMTRADE file format has been standardised by the Power System Relaying
Committee (PSRC) of the IEEE Power & Energy Society as C37.111. The most widely used
version of the COMTRADE standard is C37.111-1999. This version specified a file format that
consisted of multiple file types designated by the assigned file extensions of *.CFG, *.INF,
*.HDR, and *.DAT. The *.DAT file contains the digitised sample data in an ASCII text format.
The *.CFG file contains configuration data as in the *.DAT file including information such as
signal names, start time of the samples, number of samples, min/max values, and more. Only
the *.CFG and *.DAT files were mandatory. Although the values of the digitised samples in the
*.DAT file are viewable without the *.CFG file, the value of the data is greatly diminished as
it would be very difficult to fully reconstruct the meaning of the data without the *.CFG file.
As applications for using COMTRADE grew the limitations of the file format began to
cause problems. For instance, the North American Synchro Phasor Initiative (NASPI) sponsored
by the North American Electric Reliability Corporation (NERC) wanted to use the COMTRADE
file format to exchange synchrophasor data. The C37.1111999 version of the COMTRADE
format only supported ASCII encoded data, supported limited analog and digital status data
Appendix249
types, and did not convey any power system configuration information. This would have
made large scale exchange of synchrophasordata more difficult. Beginning in 2011, the IEEE
PSRC began work on a new version of the COMTRADE standard that has been published as
C37.111-2013. This version of the standard supports a variety of new features including XML
encoding configuration and data, a single file type (*.CFF) that combines all the C37.111-1999
types into a single file, additional data types for 32-bit binary and floating point values and a
partial binary encoding to reduce file size.
The IEEE Power Quality Subcommittee of the IEEE Power & Energy Society has also
developed a file format called the Power Quality Data Interchange Format (PQDIF) that is
similar to COMTRADE in structure but is used primarily to convey power quality data instead
of transient disturbance data.
There are numerous COMTRADE viewers available from many commercial companies as
well as free viewers like Comcalc Pro, TOP, GTPPLOT and others.
A.3 ANSI/IEEE Standard Device Numbers

In industrial drawings, protective relays are generally referred to by standard device numbers.
Letters are sometimes added to specify the application (IEEE Std C37.2-2008).
Device Numbers (the more commonly used ones are in bold)

1 Master Element
2 Time Delay Starting or Closing Relay
3 Checking or Interlocking Relay
4 Master Contactor
5 Stopping Device
6 Starting Circuit Breaker
7 Rate of Change Relay
8 Control Power Disconnecting Device
9 Reversing Device
10 Unit Sequence Switch
11 Multifunction Device
12 Overspeed Device
13 Synchronous-speed Device
14 Underspeed Device
15 Speed or Frequency-Matching Device
16 Data Communications Device
20 Electrically Operated valve (solenoid valve)
21 Distance Relay
23 Temperature Control Device
24 Volts Per Hertz Relay
250Appendix
25 Synchronising or Synchronism-Check Device

26 Apparatus Thermal Device
27 Undervoltage Relay
30 Annunciator Relay
32 Directional Power Relay
36 Polarity or Polarising Voltage Devices
37 Undercurrent or Underpower Relay
38 Bearing Protective Device
39 Mechanical Conduction Monitor
40 Field (over/under excitation) Relay
41 Field Circuit Breaker
42 Running Circuit Breaker
43 Manual Transfer or Selector Device
46 Reverse-phase or Phase-Balance Current Relay
47 Phase-Sequence or Phase-Balance Voltage Relay
48 Incomplete-Sequence Relay
49 Machine or Transformer Thermal Relay
50 Instantaneous Overcurrent Relay
51 AC Time Overcurrent Relay
52 AC Circuit Breaker
53 Field Excitation Relay
55 Power Factor Relay
56 Field Application Relay
59 Overvoltage Relay
60 Voltage or Current Balance Relay
62 Time-Delay Stopping or Opening Relay
63 Pressure Switch
64 Ground Detector Relay
65 Governor
66 Notching or jogging device
67 AC Directional Overcurrent Relay
68 Blocking or out of step Relay
69 Permissive Control Device
74 Alarm Relay
75 Position Changing Mechanism
76 DC Overcurrent Relay
78 Phase-Angle Measuring Relay
79 AC-Reclosing Relay
Appendix251
81 Frequency Relay
83 Automatic Selective Control or Transfer Relay
84 Operating Mechanism
85 Pilot Communications, Carrier or Pilot-Wire Relay
86 Lockout Relay
87 Differential Protective Relay
89 Line Switch
90 Regulating Device
91 Voltage Directional Relay
92 Voltage and Power Directional Relay
94 Tripping or Trip-Free Relay
B Bus
F Field
G Ground or generator
N Neutral
T Transformer
References
[1] Adamiak, Mark, Bogdan Kasztenny, and William Premerlani, Synchrophasors:

definition, measurement, and application, Proceedings of the 59th Annual Georgia
Tech Protective Relaying, Atlanta, GA, 2005.
[2] Adelson, Ronald M., Rapid power-line frequency monitoring, Digital Signal
Processing 12.1, 2002: 111.
[3] Benmouyal, Gabriel, Removal of DC-offset in current waveforms using digital mimic
filtering, Power Delivery, IEEE Transactions on 10.2, 1995: 621630.
[4] Benmouyal, Gabriel, E.O. Schweitzer, and Armando Guzman, Synchronised phasor
measurement in protective relays for protection, control, and analysis of electric power
systems, 2004: 814820.
[5] Bracewell, R.N., The Fourier Transform and Its Applications, McGraw-Hill, 1986.
[6] Cho, Yoon-Sung, et al., An innovative decaying DC component estimation algorithm
for digital relaying,Power Delivery, IEEE Transactions on 24.1, 2009: 7378.
[7] De La Ree, Jaime, et al., Synchronised phasor measurement applications in power
systems, Smart Grid, IEEE Transactions on 1.1, 2010: 2027.
[8] Elmore, Walter A., Protective Relaying: Theory and Applications, ABB, 1994.
[9] Ferrer, Altvue H.J. and E.O. Schweitzer, III (Eds.), Modern Solutions for Protection,
Control, and Monitoring of Electric Power Systems, Schweitzer Engineering
Laboratories, 2010.
[10] Gilbert, J.G. and R.J. Shovlin, High speed transmission line fault impedance
calculation using a dedicated minicomputer, Power Apparatus and Systems, IEEE
Transactions on 94.3, 1975: 872883.
[11] Gilchrist, G.B., G.D. Rockefeller, and E.A. Udren, High-speed distance relaying
using a digital computer I-system description, Power Apparatus and Systems, IEEE
Transactions on 3, 1972: 12351243.
253
254References
[12] Guo, Yong, Mladen Kezunovic, and Deshu Chen, Simplified algorithms for removal
of the effect of exponentially decaying DC-offset on the Fourier algorithm, Power
Delivery, IEEE Transactions on 18.3, 2003: 711717.
[13] Hays, H.H., Schaums Outline of Theory and Problems of Digital Signal Processing,
Tata McGraw-Hill, New Delhi, 2004.
[14] Holbert, Keith E., G.I. Heydt, and Hui Ni, Use of satellite technologies for power
system measurements, command, and control, Proceedings of the IEEE 93.5, 2005:
947955.
[15] Hope, G.S. and V.S. Umamaheswaran, Sampling for computer protection of
transmission lines, Power Apparatus and Systems, IEEE Transactions on 5, 1974:
15221534.
[16] Horowitz, S H., A.G. Phadke, and J.S. Thorp, Adaptive transmission system
relaying, Power Delivery, IEEE Transactions on 3.4, 1988: 14361445.
[17] Horton, John W, The use of Walsh functions for high speed digital relaying, IEEE
Publication 75CH1034-8, 1975: 19.
[18] IEEE Standard for Synchrophasor Measurements for Power Systems, IEEE Std
C37.118.1TN 2011 and IEEE Std C37.118.2TN 2011.
[19] Ifeachor, Emmanuel C. and B.W. Jervis, Digital Signal ProcessingA Practical
Approach, Pearson, 2007.
[20] Jeyasurya, B. and W.J. Smolinski, Identification of a best algorihthm for digital
distance protection of transmission lines, Power Apparatus and Systems, IEEE
Transactions on 10, 1983: 33583369.
[21] Johns, A.T. and M.A. Martin, Fundamental digital approach to the distance protection
of EHV transmission lines, Proceedings of the Institution of Electrical Engineers,
Vol. 125, No. 5, IET Digital Library, 1978.
[22] Lyons, Richard G., Understanding Digital Signal Processing, 3rd ed., Prentice Hall,
2010.
[23] Madhav Rao, T.S., Power System Protection: Static Relays with Microprocessor
Application, 2nd ed., Tata McGraw-Hill, New Delhi, 2001.
[24] Mahmood, Mahir K., Janan E. Allos, and Majid AH Abdul-Karim, Microprocessor
implementation of a fast and simultaneous amplitude and frequency detector for
sinusoidal signals, Instrumentation and Measurement, IEEE Transactions on 34.3,
1985: 413417.
[25] Makino, J. and Y. Miki, Study of operating principles and digital filters for protective
relays with digital computer, IEEE PES Winter Power Meeting, 1975.
[26] Mann, B.J. and I.F. Morrison, Digital calculation of impedance for transmission line
protection, Power Apparatus and Systems, IEEE Transactions on 1, 1971: 270279.
[27] Mann, B.J. and I.F. Morrison, Relaying a three-phase transmission line with a digital
computer, Power Apparatus and Systems, IEEE Transactions on 2, 1971: 742750.
[28] Martin, K.E., et al., IEEE Standard for synchrophasors for power systems,Power
Delivery, IEEE Transactions on 13.1, 1998: 7377.
References255
[29] Mason, C. Russell, The Art and Science of Protective Relaying, Wiley, 1956.
[30] McInnes, A.D. and I.F. Morrison, Real time calculation of resistance and reactance
for transmission line protection by digital computer, IEEE Trans, 1971: 1623.
[31] McLaren, Peter Grant and M.A. Redfern, Fourier-series techniques applied to distance
protection, Proceedings of the Institution of Electrical Engineers, Vol. 122, No. 11,
IET Digital Library, 1975.
[32] North American Synchrophasor Initiative: Synchrophasor System Benefits Factsheet:
http://www.naspi.org.
[33] Paithankar, Y.G. and S.R. Bhide, Fundamentals of Power System Protection, PHI
Learning, New Delhi, 2010.
[34] Paithankar, Y.G., Transmission Network Protection, Theory and Practise, CRC Press,
1997.
[35] Phadke, A.G. and J.S. Thorp, Synchronised Phasor Measurements and Their
Applications, Springer, 2008.
[36] Phadke, A.G. and Bogdan Kasztenny, Synchronised phasor and frequency measurement
under transient conditions, Power Delivery, IEEE Transactions on24.1, 2009: 8995.
[37] Phadke, A.G. and J.S. Thorp, History and applications of phasor measurements,Power
Systems Conference and Exposition, 2006, PSCE06, 2006, IEEE PES, IEEE, 2006.
[38] Phadke, A.G., et al., Synchronised sampling and phasor measurements for relaying
and control,Power Delivery, IEEE Transactions on 9.1, 1994: 442452.
[39] Phadke, A.G., J.S. Thorp, and M.G. Adamiak, A new measurement technique
for tracking voltage phasors, local system frequency, and rate of change of
frequency,Power Apparatus and Systems, IEEE Transactions on5, 1983: 10251038.
[40] Phadke, A.G., T. Hlibka, and M. Ibrahim, A digital computer system for EHV
substations: analysis and field tests,Power Apparatus and Systems, IEEE Transactions
on 95.1, 1976: 291301.
[41] Phadke, Arun G. and James S. Thorp, Computer Relaying for Power Systems, 2nd
ed., John Wiley & Sons, 2009.
[42] Pratap, Rudra, Getting Started with MATLAB 7, Oxford University Press (Indian
Edition), 2006.
[43] Rahman, M.A. and B. Jeyasurya, A state-of-the-art review of transformer protection
algorithms, Power Delivery, IEEE Transactions on 3.2, 1988: 534544.
[44] Ram, Badri, Power System Protection and Switchgear, Tata McGraw-Hill, 2011.
[45] Ramamoorty, M., Application of digital computers to power system protection,
Journal of the Institution of Engineers, India, 52, 1972: 235238.
[46] Ranjbar, A.M. and B.J. Cory, Algorithms for distance protection, IEEE Conf. on
Developments in Power System Protection, Institution of Electrical Engineers, London,
1975.
[47] Ranjbar, A.M. and B.J. Cory, An improved method for the digital protection of
high voltage transmission lines,Power Apparatus and Systems, IEEE Transactions
on 94.2, 1975: 544550.
256References
[48] Rockefeller, G.D. and E.A. Udren, High-speed distance relaying using a digital
computer II-test results,Power Apparatus and Systems, IEEE Transactions on 3,
1972: 12441258.
[49] Rockefeller, George D., Fault protection with a digital computer,Power Apparatus
and Systems, IEEE Transactions on 4, 1969: 438464.
[50] Sachdev, M.S. and M.A. Baribeau, A new algorithm for digital impedance
relays, Power Apparatus and Systems, IEEE Transactions on 6, 1979: 22322240.
[51] Scarborough, James Blaine, Numerical Mathematical Analysis, Oxford and IBH
Publishing, 1955.
[52] Schilling, Robert Joseph and Sandra L. Harris, Introduction to Digital Signal
Processing, 2nd ed., Cengage Learning, 2011.
[53] Segui, Trinidad, et al., Fundamental basis for distance relaying with parametrical
estimation, Power Delivery, IEEE Transactions on 16.1, 2001: 99104.
[54] Sidhu, T.S., et al., Discrete-Fourier-transform-based technique for removal of decaying
DC offset from phasor estimates, IEEE Proceedings-Generation, Transmission and
Distribution 150.6, 2003: 745752.
[55] Singh, Lankeshwar Prakash, Digital Protection: Protective Relaying from
Electromechanical to Microprocessor, New Age International, 1997.
[56] Smolinski, W.J., An algorithm for digital impedance calculation using a single PI
section transmission line model, Power Apparatus and Systems, IEEE Transactions
on 5, 1979: 15461551.
[57] Soman, S.A., Power System Protection Course, NPTEL Website.
[58] Strang, Gilbert, Linear Algebra and Its Applications, 4th ed., Cengage Learning, 2006.
[59] Thorp, J.S. and A.G. Phadke, A microprocessor based three-phase transformer
differential relay,Power Apparatus and Systems, IEEE Transactions on 2, 1982:
426432.
[60] Thorp, J.S., et al., Limits to impedance relaying, Power Apparatus and Systems,
IEEE Transactions on 1, 1979: 246260.
[61] Van Valkenburg, M.E., Network Analysis, Prentice Hall Inc., 1974.
[62] Warrington, AR van C. (Ed.), Protective Relays Their Theory and Practise: Volume
I, 1962.
[63] Warrington, AR van C. (Ed.), Protective Relays Their Theory and Practise: Volume
II, 1969.
[64] Wilson, Robert E., Uses of precise time and frequency in power systems, Proceedings
of the IEEE 79.7, 1991: 10091018.
[65] Yu, Sun-Li and Jyh-Cherng Gu, Removal of decaying DC in current and voltage
signals using a modified Fourier filter algorithm,Power Delivery, IEEE Transactions
on 16.3, 2001: 372379.
[66] Ziegler, Gerhard, Numerical Distance Protection, Siemens, Publicis, Germany, 2006.
Index
LSB offset, 10 Coherent gain, 168

1 LSB, 9 Comtrade, 248
1 PPS signal, 219 Convolution, 164
Cooley and Tuckey, 135
Curve fitting, 92
Adaptive filters, 173

ADC types, 12 Decaying DC offset, 237, 238
Algorithms, 4 Decimation, 134
Aliasing, 19 Decimation-in-frequency, 134
Analog to digital converter (ADC), 7 Decimation-in-time FFT, 134
ANSI/IEEE standard device numbers, 249 Design of digital resonator, 213
Anti-aliasing filter, 21 Design of a digital resonator using pole zero
Architecture of digital relay, 4 placement, 209
Average value, 88 Design of IIR filter using bilinear transformation,
204
DFT bins, 239
Differential equation algorithm, 60, 240
Band-pass filter, 197
Digital filter, 173
Band-stop filter, 197
Digital filter design, 196
Bilinear transformation, 204, 206
Digital Filtering, 172
Bipolar ADC, 7
Digital mimic impedance based filter, 240
Bit reversal, 144
Digital signal processing, 6
Butterworth filter, 22
Digital to analog converter (DAC), 7
Direct form IIR filter, 192
Dirichlet conditions, 102
Cascade form IIR filter, 193 Discrete Fourier transform (DFT), 99, 112
Characterising the decaying DC offset by Distance relay, 95
Integration, 244 Distance relay characteristics, 62
Coefficient symmetry, 198 Dual slope ADC, 12
257
258Index
Effect of decaying DC offset, 239 IIR filters, 175

Effect of off-nominal frequency operation on Impulse response of digital filter, 178
phasor measurement, 228 Infinite impulse response (IIR), 191
Embedded system software, 24 Instantaneous over-current relay, 32
Encoder, 8 Instantaneous value, 217
Evolution of a directional relay, 2 Intelligent electronic device, 24
Evolution of OC relay, 1 IRIG-B, 224
Fault counter, 32 Least squared error (LSQ), 87

FFT, 133 Least sum of squares of errors, 87
Filter coefficients, 191 Linear phase, 198
Filter frequency response, 197 Low pass filter, 197
Filtering DC offsets, 242 LSQ, 240
Filter specifications, 196 LSQ sense, 88
FIR filter block diagram, 176 LSQ technique, 87
First and second derivative algorithm, 44 Lumped R-L circuit, 60
Flash or parallel comparator, 12 Lumped series R-L model, 62
Flash type ADC, 13
Forbidden frequency, 19
Forensic analysis, 236 Mann and Morrison algorithm, 29
Fourier machine, 162 Method of partial sums, 242
Fourier series in exponential form, 104 Mimic impedance, 239
Fourier series in trigonometric form, 102 Modified DFT, 128
FRACSEC, 223
Frequency bin, 162
Frequency measurement, 38 Non-recurring pulse, 111
Frequency response, 183, 185, 187 Numerical integration, 68
Frequency response of digital filter, 179 Numerical relay, 24
Frequency response using Z-transform, 182
Frequency sampling method, 199, 200
Off-nominal frequency operation of the PMU, 227
Orthogonal functions, 101
Generalised Fourier analysis, 101 Orthonormal, 101
Generalised Fourier series, 101 Over-determined system, 89
Goertzel algorithm, 133, 155 Over-reach, 239
GPS, 216, 218, 247
GPS antenna, 219
Ground fault protection, 79 Parallel comparator type ADC, 13
Ground fault protection of three-phase line, 77 Parallel form of IIR filter realisation, 193
Perceived frequency, 20
Phase fault protection of three-phase lines using
Hamming window, 168 sequence quantities, 83
Harmonic elimination, 84 Phasor, 217
High-pass filter, 197 Phasor data concentrators, 224
Phasor measurement unit (PMU), 24
PMU, 216, 225
Ideal anti-aliasing filter, 22 Power system control, 235
IIR filter block diagram, 176 Pre-warping of frequency, 206
Index259
Processing gain, 168 State estimation, 234

Protection, control and monitoring systems, 24 Stationary signals, 217
Pseudo-inverse, 89, 90 Statistical nature of error, 87
Statistics, 87
Successive approximation, 12
Quantization, 8 Synchrophasor, 216, 218
Quantization error, 8 Synchrophasor applications, 234
Quantizer, 8 Synchrophasor enabled digital relay, 225
Synchrophasor in steady-state, 218
Synchrophasor in transient conditions, 220
Rate of change of frequency (ROCOF), 222 Synchrophasor measurement, 24
Rectangular window, 162 Synchrophasor reporting rates, 225
Recurring pulse, 108 Synchrophasor under dynamic conditions, 221
Redundancy of DFT, 116 Synthesis of signal, 101
Removal of DC offset, 237
ROCOF, 223
Running average, 177
Three-phase line protection, 76
Time stamp, 224
Time stamping of synchrophasor, 223
Sachdevs algorithm, 92
Time tagging of synchrophasor, 223
Sample and derivative algorithm, 30
Total vector error (TVE), 221
Sample and hold circuit, 14
Trapezoidal rule, 71
SCADA, 216
Triangular window, 168
Second-of-century (SOC) count, 223
Two-sample technique, 51
Sequence quantities, 79
Shannons sampling theorem, 18
Shunt capacitance, 60
Sigma-delta type ADC, 12 Warping of frequency, 204
Signal to noise ratio, 10 Wide area power system protection, 235
Simpsons rule, 72 Windowed Fourier transform, 114
Sinc() function, 111 Windowing, 161
Smart grid, 236
SOC count, 223
Spectral leakage, 161, 165, 166 Z-plane, 209
SSE, 88 Z-transform, 185

Digital Power System Protection

Uploaded by

Copyright:

Available Formats

Digital Power System Protection

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Digital Power System Protection

Uploaded by

Copyright:

Available Formats

DigitalPower

1. Evolution of Power System Protection and the Emergence

2. Digital Signal Processing Basics and Architecture

3. Algorithms Based on Undistorted Single Frequency Sine Wave 2959

4. Algorithms Based on Solution of Differential Equation 6086

4.7.4 Phase Fault Protection of Three-Phase Lines Using

5. Algorithms Based on Least Squared Error (LSQ) 8798

6. Discrete Fourier Transform 99132

7. FFT and Goertzel Algorithm 133160

7.5 MATLAB Implementation of FFT: Decimation-in-Frequency 153

8. Windowing and Spectral Leakage 161171

9. Introduction to Digital Filtering 172195

10. Digital Filter Design 196215

10.4.2 Warping of Frequency 204

11. Synchrophasors 216236

12. Removal of DC Offset 237245

12.4 Method of Partial Sums for Filtering DC Offsets 242

1.1 Evolution of OC Relay from an Ammeter

Figure 1.1 Evolution of an ammeter into an over-current relay.

1.2 Evolution of a Directional Relay from a Wattmeter

Figure 1.2 Evolution of directional relay.

1.3 The Birth of Transistor

1.4 Start of a Revolution

5. Keeping count of the number of circuit breaker tripping to facilitate CB maintenance

1.5 Architecture of the Modern Digital Relay

Table 1.1 Seven Layers of communication

On similar lines, we can conceptualise layers of protective relaying functions where

Table 1.2 Seven layers of protection

Algorithm for control, automation and monitoring functions

2.1 Introduction to Digital Signal Processing

Figure 2.1 The two worlds.

2.2 The DSP Signal Processing Chain

Figure 2.2 The classical DSP signal chain

2.3 Analog to Digital Converters

Figure 2.3(a) Input and output of a N-bit bipolar ADC.

Figure 2.3(b) Functional diagram of an ADC.

2.3.1 Quantization Error

Figure 2.4 Quantization in a unipolar 3-bit ADC.

Figure 2.5 Transfer characteristics of a 3-bit ADC and quantization error.

We can therefore write signal to noise ratio (SNR) as:

SNR dB = 20( N - 1) log10 2 + 20 log10 ( 6 )

SNR dB = 20 N log10 2 20 log10 2 + 20 log 10 ( 6 )

2.3.2 ADC Types

Table 2.1 Analog to digital converter types

The flash type ADC

Figure 2.7 A flash type 2-bit A to D converter.

Figure 2.8(a) Analogue signal and its sampled version.

2.4.1 Need for Sample and Hold Circuit

Figure 2.8(b) Need for sample and hold circuit.

Figure 2.8(c) Maximum frequency goes down for a better ADC.

Let the signal be represented by

For the ADC with given specifications of N = 16 and Tc = 10ms, we get

f max = 9.714 kHz

Figure 2.9(b) Sampled analog signal.

2.4.2 Shannons Sampling Theorem and Aliasing

fsampling,minimum = 2 fsignal,max if the signal happens to be broad-band

The minimum sampling frequency stipulated by Shannons sampling theorem is known as

Figure 2.10 Shannons sampling theorem.

Figure 2.11(a) Aliasing.

Figure 2.11(b) shows a more specific example of a 50 Hz signal sampled at frequencies of

3.3.1Excel Spreadsheet for Simulation of First and Second Derivative