III. Combined Wavelet Transform and Neural Networks Protection ...

haremboingAI and Robotics

Oct 20, 2013 (4 years and 19 days ago)

74 views


1



Abstract
--
This paper presents an analysis and simulation methodology to discuss the possible impacts of high impedance fault

(
HIF
)
.
Its object is to
propo
se a novel protection scheme
for overhead distribution
feeder
.

The wavelet transform

(
WT
)

technique has been
successfully applied in many fields. The properties of scaling and translation of WT can be used to identify stable signal an
d transient
signal. Th
e discrete wavelet transforms

(
DWTs
)
is firstly applied to extract of distinctive features of the voltage and current signals and
transform into a series of detail
s

and approximat
ions

wavelet components. Then, the coefficients of variation of the wavelet c
omponents
are calculated. This information is introduced to
training

artificial

neural networks (A
NN
)

for identifying a HIF from the switches
operations. The simulated results clearly show that the proposed technique can accurately identify the
HIF

in dist
ribution feeder
.


Index Terms

High impedance fault, Neural networks, Wavelet transforms

I.
I
NTRODUCTION

nce power system
incurs short circuit or solid ground
fault,
in general
the fault current are large enough
to
activat
e
protection devices
. In contrast
, H
IFs that occurs
without
enough
fault
current to be detectable by conventional overcorrect
relays
.
R
ussell

[1] applies
v
e
ctor
ial

approach
by monitoring
fault sensitive parameters

to identify the HIF events from low
current condition
on the protected feeders
.

However, this approach could not cover all the HIF events. Hence
,

he introduced a
multiple algorithms to detect various types of faults and the used of
a
n expe
r
t decision maker to unravel fault signal. Swift

[
2
]
showed that the behavior of HIF’s current
is affected by the surface conditions,
such as
the degree of dryness or wetness
. T
he
nonlinear
voltage
-
current characteristic of an arcing is

well known.

David

[
3
] proposed a realistic model of HIF involving non
-
linear impedance, time varying voltage sourc
es and

TA
C
S controlled switch

to obtain the signal characteristic of arcing
.

Wavelet
transform
apply
in power system
s

has been discussed
for a long time
.

Recently, study shows that w
avelet analysis filter banks can

successfully

identify HIF
s

and discrimina
te
it
from other normal event in distribution feeder

systems [4
-
5]
.

Neural networks uses
in power systems also are widely explored.
Combinations of
w
avelet transform and neural networks in transient type pattern
recognition have

been reported [6
-
7].

T
o dat
e, although much literature is available on analysis power system transient using
w
avelet transform or/and neural
network
s
, little work has been published on analysis HIFs by combination wavelet transform with neural networks. The
mainly
objective of

this

study
is
to propose a new

protection

scheme
by
combin
ed
wavelet

transform with neural
networks, which

can
detect
and identify
normal
switching operation
and HIF. In addition, the interface of wavelet transform and neural networks was
evaluated for combinat
ion each other effectively
.

II.
F
EATURE
E
XTRACTION AND
S
YSTEM
S
IMULATION

M
ost
of
approach
es for detecting
HIF

have their own limitations and are performed directly in time domain. In this study, we
present a novel approach for HIF detection and localization b
ased on
the orthonormal

wavelet transform where detection is
carried out in the time
-
scale domain
.

The subband information

of current

can be extracted from the original signal with the
aid
of multiresolution analysis (MRA).
When a fault occurs in distribut
ion feeder,
many
subband information
within current signal
may contain
with
useful fault
signatures.
Study shows that t
he sharp variations of the signals can be viewed as the features of the faults. By
recognizing
these
sharp variations, the
HIF
can be
eff
ectively
identified. The
major
object of this study is to detect and classify the HIF from

the
level 1
detail
coefficients

d
1
[n]
of MRA.

T
he

MRA functions with
strong capability to extract important features
implicitly
embedded in distribution feeder curr
ent and voltage signals even these features are very weak. To
process

the data in a digital
sense the
discrete

wavelet transform (DWT) is used.

Selecting the adequate wavelet filter is essential to discover the characteristics of the arc
ing

current and
vol
tage

signal.




聖約翰技術學院改善師資案成果報告書

類別:研究


時間:自

92


1


1
日起至

92


12


31


內容:
配電系統高阻抗接地故障之研究

O


2

Furthermore, the understanding
the behavior of
extinguishing

and
r
eigniting in the arc

is important
. After the evaluation of
se
vera
l

kinds of wavelet, the Daubechies D
-
10 wavelet is prove
d to have little computational burden as well as good
pe
rformance, so it is adopted. In this study the sampling
frequency

is 3840Hz.

In order to investigate the properties of multiresolution decomposition in HIFs
,

an example

distribution

networks
in

Fig
.
1

is
setup for simulation by
using ARENE.

ARENE is a dig
ital transient network
analyzer
.
The simulations including normal load
switching, capacitor bank switching and abnormal HIF fault in the feeder.




Fig. 1 The
example d
istribution

network


After DWT
filtering
,

the

c
4
[n] indicates the approximation, d
1
[n
]~d
4
[n] shows the details and c
0
[n] represents the original
signal, as shown in Fig. 2. T
he results
are present from
Fig
.
3

to Fig.
8 individually.

Only phase A information are shown. It
shows

that the
HIF
will
generate
a
big spike
with certain amplitude
i
n the

fault phase.

Obviously, it
can be used to

identify the
HIF
. The

d
1
[n]
value
s

are near zero before the fault

occurring
.

A
fter the fault occurs the coefficient value
s

jump to a value

and
intermittence

persists for a short time
. The
a
ct
u
al
coefficient
v
alue
s

are

depende
d up
on the fault condition
.
The coefficient values
represent the spectral energy of this signal. However the energy variation of switching transient is more than less compared
to the
energy variation of the signals produced by a HIF
.




F
ig.
2 Four
-
Level Wavelet Decomposition Tree


3


Fig.
3

Current of 500kVA Load Switching


Fig.

4

Voltage of 500kVA Load Switching


Fig.
5
Current of
100kvar
Capacitor Banks Switching


Fig.
6

Voltage of
100kvar

Capacitor Banks Switching


4


Fig.

7

Curre
nt of HIF on Phase
A


Fig.
8

Voltage of HIF on Phase
A

III.
C
OMBINED
W
AVELET
T
RANSFORM AND
N
EURAL
N
ETWORK
S

P
ROTECTION
T
ECHNIQUE


In this study, Daubechies wavelet
f
un
ctions

D
-
10 with a resolution of 4
levels

is used.

T
he features for classification are
obtain
ed from the wavelet coefficient vectors. The back propagation
artificial

neural network
s

(BP
-
ANN) are powerful for
pattern recognition.
W
ith proper training, neural networks can acquire HIF detection

ability
.

The object of the BP
-
ANN is to diagnose HIF an
d
distinguish

it
from
normal

system operation events, such as load switching
and capacitor banks switching.

Fig. 9 shows that

the BP
-
ANN topology is composed of three independent

multiplayer

of

feedforward

network
s (
two hidden layers
)

which has 20 neurons
in the first
hidden

layers, 10
neurons

in the second hidden layers
and 3 neurons in the output layer
.

It is not practical to directly
input
the wavelet signals to ANN, because
it
will
incur too many
inputs

to ANN,

and increase the
difficulty in ANN con
ve
r
g
ence
.
Fortunately, utilize
the coefficient of variation
can
overcome

this disadvantage.



Fig.
9

Structure of Back Propagation Neural Network
s


Coefficient of
variation

(CV) estimates relative variation of data. Normally, indicated CV by percentage.



(1)

:
s
tandard

derivation

: mean

T
he
CV

is obtained by a series operation
of
DWT analysis which the voltages and currents are
measured at
bus A in Fig.1.

5

The calculation results o
f CV were shown in Table
I
. Essentially,
this
approach
not only reduces ANN size, but also retains
important features of the
original
signals. In this
study
the 5
cycles’

of
data window are
applied to

calculate the

CV.

H
ence,

the
correspon
ding signal of d
1
[n] for each

voltage and current are
obtained
.

Therefore,
there are
only
6
coefficients
that are fed into

the ANN
.

There are three target sets in ANN output.

The

1
st

target

is for fault detection. It
only
h
as

two

statuses

as shown in Table

I

:

0


means
no fault and

1


indicate
HIF

exists in feeder
. The
2
nd

target

is used to distinguish HIFs from normal system
operations
like load and capacitor switching.

This
stage
may have three
statuses
:

0


means
normal load
switching
,


1


represents
capacitor

banks switching

and

2


shows

HIF. The
3
rd

target

possess four status
is used to indicate the
individual
fault phase.

0


means
no fault,

1


shows

a HIF on phase

A
,

2


indicate

a HIF on phase
B
, and

3


represents
a HIF on pha
se
C.

The proposed
protection

logic for distinguishing normal system operation event from a HIF by combining the wavelet
transform with neural network
s

as

shown in Fig.

10.

Normally, t
he
logic of

protection scheme is activated
if one of

the three
phases

cu
rrent

exceeds the preset threshold
.
.
However, the

fault current of HIF
may

more than less the
threshold setting. Hence, the

wavelet

transform analysis of voltages


and
currents
are requir
ed
.
After that the
d
1
[n]

values

of each phase can

obtain
. The
n, the

obtained
CV
of the
wavelet signal feed to the ANN
is good to used
to
distinguishing

load switching,
capacitor

banks switching and HIF.

There are
600 cases
of
load switching, capacitor bank
s switching and HIF

events

were
s
imulated

with
the ARENE
and Matlab
software, and 500 of those were used for training of the neural network
s

and the rest 100 cases were used for testing the network
.

The
part of
test results of switching and HIF events are

given
in
T
able

II
.

All
the results satisfy the requirement
s

of intelligent
device with general protection function and HIF detection capability.
In this paper
,

a number of evaluations were executed

varying

with

one or two hidden layers as

well as varying the number of neurons in each hidden

layer.
The performance of ANN
with different architectures
is

summarized in
T
able

III
. It is
shown
that almost all of them give well performance
.

IV.
C
ONCLUSION


A
new topology
,
linking wav
elet transform
s

and neural network
s
,

for HIF detection

has been presented.

Such a
new intelligent
protection scheme
not only possessed
general overcurrent protection
function but also
satisfies in
HIF detection and
identification

capability
.
In addition,
t
he
study found
that the coefficients of variation have
great advantage
s on linking wavelet
transform and neural network
s

and the performance of neural network
s

is perfect.
W
e recommend that these procedures be cited
for development HIF

protection scheme.




Fig.
10

The Proposed Intelligent Overcurrent Relay logic

T
ABLE
I


C
OEFFICIENT OF
V
ARIATION


Condition

C
oefficient of variation

v
a

v
b

v
c

i
a

i
b

i
c

Load switching Case 1

1.6785

1.8589

2.1041

4.6700

5.5035

4.8317

Load switching Case 2

1.6619

1.8611

1.9641

4.6934

5.5160

4.7381

Load switching Case 3

1.7807

1.8515

1.9939

4.8244

5.5588

4.7182

Capacitor banks switching Case 1

3.3577

3.7170

4.6145

5.8787

6.1200

5.9986

Capacitor banks switching Case 2

3.0599

3.4975

4.5748

5.9847

5.9484

6.1
327

Capacitor banks switching Case 3

2.7576

3.3299

4.5145

6.0003

5.9694

6.1074

HIF for phase a Case 1

1.4275

4.7165

4.1428

5.2061

1.5183

1.5012

HIF for phase a Case 2

1.0691

4.4544

3.8577

5.3854

1.3022

1.2717

HIF for phase a Case 3

0.8879

4.2110

3.4194

5.5677

1.1655

1.1263


6

HIF for phase b Case 1

1.7244

0.7034

2.8139

1.3592

4.5012

1.3794

HIF for phase b Case 2

1.0997

0.6606

1.8929

1.2392

4.2250

1.2196

HIF for phase b Case 3

0.6477

1.0938

1.1617

0.8423

4.3044

0.8209

HIF for phase c Case 1

4.1416

4.243
6

1.2844

1.9598

1.7603

5.0935

HIF for phase c Case 2

3.7853

3.9913

1.1051

1.6596

1.3813

5.1716

HIF for phase c Case 3

3.2714

3.7957

0.9788

1.3438

1.3503

5.2610


T
ABLE
II


N
EURAL
N
ETWORK
T
EST
R
ESULT



Condition

D
esired ANN output

A
ctual
ANN output

1
st

Target

2
nd

Target

3
rd

Target

1
st

Target

2
nd

Target

3
rd

Target

Load switching Case 1

0

0

0

-
0.0028

-
0.0271

0.0090

Load switching Case

2

0

0

0

-
0.0033

0.0168

-
0.0420

Load switching Case

3

0

0

0

0.0032

-
0.0324

0.0758

Capacitor banks switc
hing Case

1

0

1

0

0.0860

1.3057

0.1488

Capacitor banks switching Case

2

0

1

0

-
0.0188

1.1522

0.3332

Capacitor banks switching Case

3

0

1

0

-
0.0126

0.9867

0.3685

HIF for phase a Case 1

1

2

1

1.0043

2.0128

1.0150

HIF for phase a Case

2

1

2

1

1.0029

2.001
4

0.9483

HIF for phase a Case 3

1

2

1

1.0024

1.9697

0.9838

HIF for phase b Case 1

1

2

2

0.9958

1.9895

1.9790

HIF for phase b Case 2

1

2

2

1.0051

2.0175

1.9470

HIF for phase b Case 3

1

2

2

1.0029

2.0279

1.9951

HIF for phase c Case 1

1

2

3

1.0000

2.0185

3.0492

HIF for phase c Case 2

1

2

3

1.0028

2.0086

3.0999

HIF for phase c Case 3

1

2

3

1.0002

1.9942

3.0119


T
ABLE
III

N
EURAL
N
ETWORK
P
ERFORMANCE



ANN size

Test pattern

C
orrect pattern

I
ncorrect pattern

C
lassification rate(%)

6/30/20/
3

100

100

0

100

6/20/20/3

100

98

2

98

6/20/10/3

100

100

0

100

6/30/3

100

99

1

99

6/20/3

100

99

1

99

6/10/3

100

100

0

100

V.
R
EFERENCES

[1]

Carl L. Benner and B. Don Russell, "Practical High Impedance Fault Detection for Distribution Feeders,"

in
Proc. 39th

Annu.
1996

Rural Electric Power
Conf
.
,
pp.
B2
-
1~B2
-
6.

[2]

A. F. Sultan, G.. W. Swift and D. J. Fedirchuk, "Detecting Arcing Downed
-
Wires Using Fault Current Flicker and Half
-
Cycle Asymmetry, "
IEEE Trans
.

Power Delivery
,
v
ol
.

9,
n
o.1,
pp.
461
-
470, Jan
.

1994.

[3]

David Chan Tat Wai and Xia Yibin
,

"A Novel Technique for High Impedance Fault Identification,"
IEEE Trans
.

Power Delivery
,

v
ol
.

13,
n
o.3,
pp.
738
-
744, July 1998
.

[4]

ASurya Santoso and Peter Hofmann, "Power Quality Assessment via Wavelet Transform Analysis,"
I
EEE Trans
.

Power Deliver
,
v
ol
.

11,
n
o. 2,
pp.
924
-
930, Apr
.
1996.

[5]

Omar A. S. Youssef, "A Wavelet
-
Based Technique for Discrimination Between Faults and Magnetizing Inrush Currents in Transformers,"
IEEE Trans
.

Power Delivery
,
v
ol
.

18,
n
o.1,
pp.
170
-
176, Jan
.

2003
.

[6]

Peilin L. Mao and Raj K. Aggarwal, "A Novel Approach to the Classification of the Transient Phenomena in Power Transformers U
sing Combined
Wavelet Transform and Neural Network,"
IEEE Trans
.

power Delivery
,
v
ol
.
11,
n
o.

4
pp.
654
-
660, Oct
.

2001
.

[7]

Fra
ncisco Martin and Jose A. Aguado, "Wavelet
-
Based ANN Approach for Transmission Line Protection,"
IEEE Trans
.

Power Delivery
,
v
ol
.
18,
n
o.4,
pp.
1572
-
1574, Oct
.

2003.