Automatic Recognition

of Power Quality Disturbances

S
e
n
s
o
r
s
,

E
n
e
r
g
y
,

a
n
d
Au
t
o
m
a
t
i
o
n

L
a
b
o
r
a
t
o
r
y
SEAL
APT Center
MSEE Thesis Presentation

Min Wang
Advisor: Prof. Alexander Mamishev

August 9, 2001

Outline
Overview the project and my thesis
Two new PQ event recognition
algorithms
Signal resources
Other techniques under exploration
Conclusions
What is Power Quality (PQ)?
Any deviation from a perfect sinusoidal waveform
that can results in failure or misoperation of
customer equipment

Quality of the current and voltage provided to the
customers
Providing customers with a clean sinusoidal
waveforms at 60 Hz without sags or spikes.
Providing power to allow sensitive electronic
equipment operate reliably.
Several typical PQ disturbances
Voltage Sag
Lightning Strike
Capacitor Switching
High Impedance Fault (RMS)
Harmonics
Voltage sags
Major causes: faults, starting of
large loads, and brown-out
recovery

Major consequences: shorts,
accelerated aging, loss of data or
stability, process interrupt, etc.
Capacitor switching transients
Major causes: a power factor
correction method
Major consequences: insulation
breakdown or sparkover,
semiconductor device damage,
shorts, accelerated aging, loss of
data or stability
Harmonics
Major causes: power
electronic equipment, arcing,
transformer saturation

Major consequences:
equipment overheating, high
voltage/current, protective
device operations
Lightning transients
Major causes: lightning strikes

Major consequences: insulation
breakdown or sparkover,
semiconductor device damage,
shorts, accelerated aging, loss of data
or stability
High impedance faults
(One of the most difficult power system
protection problems)
Major causes: fallen conductors, trees (fail
to establish a permanent return path)
Major consequences: fire, threats to
personal safety
Why PQ becomes important?
Proliferation of highly sensitive computerized
equipment places more stringent demands on PQ
Semiconductor industry
Computers and computer-related businesses
Variable-speed drives or robots
Programmable logic controllers
Electronic equipment results in more PQ problems
Deregulation of power industry creates more
competitive market
Why PQ becomes important?
Impact to Silicon Valley
One cycle interruption makes a silicon device worthless
Five minutes shut down of a chip fabrication plant
causes delay from a day to a week
One second of power outage makes e-commerce sites
lose millions of dollars worth of business
US PQ losses: $20 billion/year (Frost & Sullivan)

Industry Loss per voltage sag event
Paper Manufacturing $30,000
Chemical industry $50,000
Automobile industry $75,000
Equipment manufacturing $100,000
Credit card processing $250,000
Semiconductor industry $2.5 million
ITIC curve
(1996)

State of the art
PQ monitoring software and hardware are needed in both
utilities and customers
Detect, identify, and localize different PQ disturbances
Real time decision making
The topic of a general event classification (opposed to
individual fault detection) has rarely been addressed
Existing automatic recognition methods need much
improvement in terms of their versatility, reliability, and
accuracy.
The accumulation of a comprehensive PQ database will
significantly expedite the birth of the solutions
Goals and Status of this project
Goals of this research project
Enhancement of real-time power system protection
Statistical accumulations of power quality problems
Incipient fault detections
Current work -- Applying advanced signal processing
techniques to identification of power quality events
Capture the key information from the waveforms
(Feature Extraction)
Discriminate based on the features captured
(Classifier)
Error correcting / comprehensive decision making
(Post-processing)
Signal Sources
Matlab simulations
PSCAD/EMTDC
simulations
data from industrial
partners
single instrument
measurement (power
platform)
Power Quality Events
harmonics
capacitor high
frequency
switching
capacitor high
frequency
switching
voltage sudden
sag
voltage sag decay
voltage swell
high impedance
faults
motor starting
lightning strike
transformer
inrush
Feature Extractors
ambiguity plane
based class-
dependent TFR
clustered
wavelet MSD
matrix
Applications
statistical
analysis,
troubleshooting
relaying,
protection
incipient fault
detection
Classifiers
artificial neural
networks
hidden Markov
models
Scope of this thesis
Scope of this project
Overview of Algorithm I
PQ Disturbance Classification Algorithm II
Voltage Waveform
Types of PQ
Disturbances
Instantaneous
Autocorrelation
Function
Classification By
Artificial Neural
Network
Modified Fishers
Discriminant Ratio
Kernel
Ambiguity
Plane
Feature
Extraction
Low Pass Filtering
and Resampling
Time-Frequency Representations (TFR)
A chirp signal Fourier Spectrum
Optimal-kernel TFR Spectrogram Wigner TFR
Overall Strategy
Want a TFR specially for classification purpose
Why?
Spectrogram is not for the classification goal
We dont need accurate information about the signal
We only need information useful for the classification goal
Need a class-dependent, instead of signal-dependent TFR
Want a TFR specially for this classification task
Different classification task have different optimal TFRs
No TFR available we can use directly
Need Design a TFR by ourselves
Overall Strategy
Signal
A
Signal
B
Signal
C
Signal
D
(TFR
i
)
A
(TFR
i
)
B
(TFR
i
)
C
(TFR
i
)
D

Class 1 Class 2 Class 3
Classifier
TFR
1

TFR
2
TFRi
Signal
TFR
n
TFR
i
Ambiguity Plane
Any TFR can be generate from ambiguity plane
TFR= Fourier Transform {AF .* Kernel}
Design our TFR by smoothing Ambiguity Function
(AF) with some form of kernel
Ambiguity Function:
V=[ v
1
v
2
v
3
v
n-2
v
n-1
v
n
]
2
1
0
1
[ , ] { [ , ]} [ , ]
N
j n
N
n
n
A n n e
N
t
q
q
q t t t

=
= 9 = 9

F
q and t are discrete Doppler and lag respectively
*
[ , ] [ ] [(( )) ]
N
n v n v n t t 9 = +
Ambiguity Plane


(d)
(f)
(e)
Figure 8 AP corresponding to Harmonics
20 40 60 80 100 120 140 160
20
40
60
80
100
120
140
160
Figure 9 AP corresponding to capacitor fast switching
20 40 60 80 100 120 140 160
20
40
60
80
100
120
140
160
Figure 10 AP corresponding to capacitor slow switching
20 40 60 80 100 120 140 160
20
40
60
80
100
120
140
160
Figure 11 AP corresponding to voltage sudden sag
20 40 60 80 100 120 140 160
20
40
60
80
100
120
140
160
Figure 12 AP corresponding to voltage gradual decay
20 40 60 80 100 120 140 160
20
40
60
80
100
120
140
160
Figure 13 AP corresponding to voltage swell
20 40 60 80 100 120 140 160
20
40
60
80
100
120
140
160
(a)
(f) (e)
(d) (c)
(b)
To extract features, class-
dependent kernels need to be
designed for smoothing the
ambiguity plane.
Intuitive Feature Extraction From AP
11
2, 1 2
32
1
12 1, 1 1
21 22
31 3, 1 3
1,1 1, , 1
2
2 1,
1 , 1
n
n n
n n
n n n n
n n n
n
n
n
n n
n
ap ap ap
ap ap
ap ap ap
a
ap
ap ap
ap
ap
ap
p ap ap
ap ap
AP
ap

(
(
(
(
=
(
(
(
(
(

11
2, 1 2
32
1, 1
2
n n
n n
n
ap
ap ap
ap
ap
ap
AP

(
(
(
(
(
(

=
(
(
(
2, 1
2
11
32
2
1, 1
n
n
n
n n
ap
ap
ap
ap
ap
ap

(
(
(
(
(
(
(
(
(

How to Choose
Feature Points
From the AP ???
Design class-dependent TFR from AP
Features for a pattern recognition task should
Maximize the separability of signals from different
classes
Maximize the similarity of signals from the same
class
Select those points whose
Between-class variances are largest
Within-class variances are smallest
Design the Kernel
( ) ( )
2
1 1
( ) 2
1
| [ , ] [ , ] |
[ , ]
( [ , ])
C C
i j
ij
i j
C
c
i
w A A
FDK
q t q t
q t
o q t
= =
=

Select those N points that have N largest

FDK values as Feature Point Locations
Figure 12 AP corresponding to voltage gradual decay
20 40 60 80 100 120 140 160
20
40
60
80
100
120
140
160
(
(
(
(
(
(
(
(
(
(
(
(
(
(
(
(
(

12
11
10
9
8
7
6
5
4
3
2
1
f
f
f
f
f
f
f
f
f
f
f
f
Feature Vector
Overview of Algorithm II
V /I Waveforms
Types of PQ
Disturbances
2nd Level
Wavelet
Decomposition
and Reconstruction

WTCs > T
9 Level
Multiresolution
signal Decomposition
(MSD)
Classification by
Neural Networks
Feature Extraction
From MSD Matrix
NOT PQ
Disturbances
no
(For first 3 level WTCs)
Moving Average
Filter
yes
Wavelet Analysis
A mathematical tool for signal analysis
A wavelet is a short-term duration wave, which grows and
decays essentially in a limited time period
RP of Morlet Mother Wavelet
Daubechies 4 Mother Wavelet
Tell us how weighted average of certain other functions
vary from one averaging period to the next
Why Use Wavelets?
Fourier Analysis
Periodic time functions
Wide bandwidth for short term transients
Not consider frequencies that evolve in time
Suffer from certain annoying anomalies
Gibbss phenomenon
Aliasing (With FFT)
FFTs computation complexity -- O(N*log
2
N)
Wavelet Analysis
Choose desirable frequency and time characteristics
Use short windows at high frequencies and low windows at low
frequencies
Basic functions employ time compression or dilation
Freedom in the choice of mother wavelet
WTs computation complexity -- O(N)
Wavelet Transform
A family of scaling functions and wavelet functions
are generated by dilating and shifting the mother
wavelet and scaling function



Calculate wavelets and scaling coefficients based on
following inner products


*
, , ,
( ), ( ) ( ) (2 )
j
j k j k j k
w x t t x t t k = < > =
}
/ 2
,
( ) 2 (2 ), ,
j j
j k
t t k j k eZ
/ 2
,
( ) 2 (2 )
j j
j k
t t k | |
*
, , ,
( ), ( ) ( ) (2 )
j
j k j k j k
v x t t x t t k | | = < > =
}
Wavelet MSD
WT decomposes a signal into different scales with multiple
levels of resolution by dilating a single mother wavelet
Decomposes a signal into its detailed and smoothed
versions


h(n)
g(n)
2
2
c
1
(n)
c
0
(n)
h(n)
g(n)
2
2
c
2
(n)
d
1
(n)
d
2
(n)
Wavelet MSD
Wavelet MSD
Conventional Methods
Use WTCs in a certain row --- lose information on the other scales
Use energy values ||d
i
||
2
of all rows --- their weights are not equal
c
0
d
1
d
2


d
M

c
M

Wavelet MSD
Basic ideas
Divide WTCs in the MSD matrix into disjoint clusters
Each cluster contributes one feature
More important frequency/scale ranges have a larger number of clusters
WTCs producing more features are more important in classification
One possible result of grouping clusters (24 features)
Scale m WTC's in MSD Matrix
1 {d
11
, d
12
, d
13
, , d
1,259
}
2 {d
21
, d
22
, d
23
, , d
2,133
}
3 {d
31
, d
32
, d
33
, , d
3,70
}
4 {d
41
, d
42
, , d
46
}{d
47
, d
48
, , d
4,38
}
5 {d
51
, d
52
, , d
5,22
}
6 {d
61
, d
62
, , d
6,9
}{d
6,10
, d
6,12
, , d
6,14
}
7 {d
71
, , d
73
}{d
74
, d
75
}{d
76
}{d
77
, , d
7,10
}
8 {d
81
, d
82
, d
83
}{d
84
, d
85
}{d
86
}{d
87
, d
88
}
9 {d
91
, d
92
, d
93
}{d
94
}{d
95
}{d
96
, d
97
}
9 {d
91
, d
92
, d
93,
d
94
}{d
95
}{d
96
}{d
97
}
Wavelet MSD
Cluster Determination
Clusters are determined by a set of training signals
Peaks of wavelet coefficients indicate the occurrences of PQ
events
Coefficients at same position of different MSD matrices
form independent random variables
Clusters are constructed around the high peaks of wavelet
coefficients
Feature Extraction
Divide the MSD Matrix according to the cluster pattern
Determine all the d features, u
1
, , u
d,

2
|| ||
i
i i
U
u r v = =

Wavelet MSD
0 10 20 30 40 50 60 70 80 90 100
0
0.01
0.02
0.03
Clustering a rowvector in the MSD matrix
Cluster 1
Cluster 2
Cluster 3
Cluster 4
Classifier Neural Networks
Input: signal signatures from feature extractors
Output: class type identified
I
n
p
u
t

P
a
t
t
e
r
n
s

O
u
t
p
u
t

D
e
c
i
s
i
o
n

Weights
Input layer
Hidden layer
Output layer
Neuronal Model
Activation Function
Bipolar Sigmoid y=T(s)
Y
X
N
X
j + 1
X
1
X
j
T
1
N
n
n
X
=
s=
Neuronal Model
W
1
W
j
W
j + 1
W
N
W
n

Basic Idea of Error Backpropagation
( ) ( ) 2
1
|| ||
( , ) / 0,
( , ) / 0
Q
q q
q
nm
mj
E t z
E w u w
E w u u
=
=
c c =
c c =

y
w
y f w
df( )/dw
= ( )
slope= w(0)
w(0)
w(1)= w(0)- (df(w(0))/dw) q
Learning rate
Linear Gradient Decent
Minimum Sum-Squared
Error Methodology
Testing of Algorithm I
RESULTS OF TESTING CLASSIFICATION METHOD
(6 CLASSES CASE)
Class Tested

Correctly
Identified

Mistaken
to C 1

Mistaken
to C 2

Mistaken
to C 3

Mistaken
to C 4

Mistaken
to C 5

Mistaken
to C 6

1. Harmonics

100%

-

0%

0%

0%

0%

0%

2. Capacitor
fast switching
transients

100%

0%

-

0%

0%

0%

0%

3. Capacitor
slow switching
transients

94% 0%

6%

-

0%

0%

0%

4. Voltage
sudden sag

92%

0%

0%

1%

-

7%

0%

5. Voltage
gradual decay
sag

93%

0%

0%

7%

0%

-

0%

6. Voltage
swell

100%

0%

0%

0%

0%

0%

-

Demo of PQ Event Recognition System
PSCAD/EMTDC
Visual power system simulator
Developed by Manitoba HVDC research
center
Simulate electromagnetic transients for DC
and AC
PSCAD is the user interface
EMTDC is the simulation engine.
Similar to EMTP and ATP but faster
Complete Circuit- AC system
AC system Continued




Dranetz BMI
Power Platform 4300
Specifications
Sampling Frequency: 7kHz
Update rates: once per second (Harmonic-based
parameters updated every 5 seconds)
Voltage: 10 600 Vrms
Frequency: Fundamental range 16 - 450 Hz
Current: depend on the current probes
(TR-2520: 300A - 3000A RMS)

Partnerships
Signal databases are being built with possible
help from
R.W. Beck
Bonnevile Power Administration
SRP (Salt River Project)
University of Washington Physical Plant
American Public Power Association
Classifier HMM (Hidden Markov Model)
Initially introduced in late 60s and early 70s
Extended from Markov process
Utilized extensively in a wide range of applications
Pattern recognition, especially speech recognition
Biological signal processing, e.g., gene prediction in DNA
Artificial intelligence, image understanding
Possible advantages as a classifier
Very good classification performance
Very competitive learning speed
Requires small number of training examples
Classifier HMM
Discrete Markov process (first order)
P[q
t
=S
j
| q
t-1
=S
i
,

q
t-2
=S
k
,]=P[q
t
=S
j
| q
t-1
=S
i
]=a
ij

Example:
State 1: rain; State 2: cloudy; State 3: sunny. (This is B.)
State transition matrix,


Given day 1(t=1) is sunny (state 3) (This is .)
What is the probability that next 7 days will be sun-sun-
rain-rain-sun-cloudy-sun? (This is O.)
If current state is known, past states is useless for
predicting future states
0.4 0.3 0.3
{ } 0.2 0.6 0.2
0.1 0.1 0.8
ij
A a
(
(
= =
(
(

Classifier HMM
Coin toss model (a HMM model)
Can only see the results, but dont know whats going on
Observation sequence: O = H H T T T H T T H H
Underlying state transition model, matrix A={a
ij
}
Probability model between observations and states, B={b
j
(k)}
Elements of an HMM
N, number of states
A & B
, initial state distribution
Complete HMM model = (A, B, )

Classifier HMM
A HMM poses three questions
Evaluation P(O| )
Given the model, probability of the observation sequence?
Forward-backward procedure
Decoding hidden state transition sequence
Given O, , best explanation of observations?
Viterbi Algorithm
Learning = arcmax P(O| )
Given O, how to adjust = (A, B, ) to approach the real
model
Expectation Maximization (EM) Algorithm
Classifier HMM
Tree structure constructed (HMT)
Black circles hidden states
White circles observations
State transition matrix A?
Probability model (B) for Wavelets
coefficients Gaussian Mixture Model

+
S
1
S
2

Classifier HMM
Pdf for the a wavelet coefficient,


Wavelet-based HMM constructed,
How to calculate (train) the HMM? -- EM algorithm
Select an initial model
0
, m=0
E step: p(S|w,
m
)
M step: set
m+1
=arg max

E
S
[ln f(w, S|)|w,
m
]
Set m=m+1. if converged, then stop; else, go to E step

, , , ,
, | , ,
1
( ) ( ) ( | ), 2 .
j k j k j k j k
L
W j k s W S j k j k
i
f w p l f w S L here
=
= =

( , , ) A B t
Classifier HMM
Maximum likelihood Classification
Compare between 2 classes



Many-class classification (finding the shortest distance
of signal likelihood)

( | )
1,
( | )
m
n
f W
if class m wins
f W

A >
1
argmin | ln ( , | ) ln ( , | ) |
J
m m d T d m
j
C f S W f S W
=
=

Postprocessing Voting Scheme

Most straightforward method for combining the
outputs of multiple classifiers
Decision is made based on votes
Multiple NNs, if
Class A receives 70% votes
Class B receives 30% votes
Then, A win
Threshold K
t
None of the classes receive more than K
t
, reject all

Postprocessing Dempster-Shafer Theory of Evidence
Mathematical Theory of Evidence (MTE)
Theory of probable reasoning and combining evidence
Provides a means of decision making
Give the degree of believe of decision made
An application example
Team 1 Team 2 Team 3
A 0 2/5 1/5
B 2/5 0 1/5
C 3/5 3/5 3/5
Belief intervals
A [6.1% 11.7%]
B [6.1% 11.7%]
C [82.2% 87.8%]
Postprocessing Rule-Based Approach
Used in artificial intelligence and expert systems
Mimic power engineers thought process
Final control process
Attempt to correct wrong decisions by some
standard rules, such as
Fast capacitor switching, wide duration of disturbance
likelihood in the highest resolution wavelet domain
These rules are from experienced PQ engineers, who
have the best embedded classification algorithms
Conclusions
The developed algorithm is new class of PQ event
classification techniques in power engineering
Software package PSCAD/EMTDC is now used
for generation of training examples
Dranetz-BMI Power Platform as a signal recorder
is introduced in research and education
Exploration on hidden Markov models and several
post-processing techniques are presented
Cause-effect type identification algorithms is
being developed
Thrust towards hardware implementation is being
considered
Acknowledgements
Professors
Alexander Mamishev
Chen-Ching Liu
Mark Damborg
APT Center
Students
Piotr Ochenkowski
Jeff Chen
Rejo Jose
My Hua
Industrial Partners