Journal of Theoretical and Applied Information Technology
st
31 December 2011. Vol. 34 No.2
© 2005 - 2011 JATIT & LLS. All rights reserved.
ISSN: 1992-8645
www.jatit.org
E-ISSN: 1817-3195
LIGHTNING STRIKE MAPPING FOR PENINSULAR
MALAYSIA USING ARTIFICIAL INTELLIGENCE
TECHNIQUES
1
M.K.HASSAN, 2R. Z. ABDUL RAHMAN, 3A..CHE .SOH., 4M.Z.A..AB KADIR
1,2,3,4
Department of Electrical and Electronics, Faculty of Engineering,
University Putra Malaysia, 43400, Serdang,
Selangor Darul Ehsan, Malaysia
E-mail: [email protected], [email protected], [email protected],
4
[email protected]
ABSTRACT
This research focuses on artificial intelligence (AI) techniques on mapping the lightning strike area in
Peninsular Malaysia. Three AI techniques such as fuzzy logic, neural network and neuro-fuzzy techniques
are selected to be explored in classifying the characteristics of lightning strike which are based on; level of
strike (high, medium, low) and category of lightning (positive cloud-to-ground, negative cloud-to-ground,
flash). Nine predefined areas in Peninsular Malaysia were chosen as a case study. The analysis was carried
out according to twelve months lightning data strikes which had been made available by Global Lightning
Network (GLN). All three AI techniques have successfully demonstrated the ability to mapping and
classify lightning strikes. Each technique has shown very good percentage of accuracy in term of
determining the area and characterizing the lightning strikes. The finding of this research can be made use
in risk management analysis, lightning protection analysis, township planning projects and the like.
Keywords: Lightning Strike, Classification, Fuzzy Logic, Neural Network, Neuro-Fuzzy
1. INRODUCTION
[2, 3]. The isokeraunic level is approximately
200 thunderstorm days a year. The lightning
ground flash density is about 15-20 strike per
km2 per year.
Lightning strike comes about every day in the
world. The lightning strike towards the surface
on earth has been estimated at 100 times every
second. Thus, almost every governments suffer
major loses because of this phenomenon every
year. It also would cause horrific injury and
fatality to humans and animals. The lightning
may affect almost every organ system as the
current passes through the human body taking
the shortest pathways between the contact points.
There are 25.9% of lightning strike occurrences
for victims who have sheltered under trees or
shades, whereas 37% at open space area. Head
and neck injury are two common areas which
have an effect on the lightning strike victims
with 77.78% and 74% respectively. Only
29.63% of the cases presented with ear bleeding
[1]. United State National Lightning Safety
Institution reported that Malaysia has highest
lightning activities in the world whilst the
average-thunder day level for Malaysia’s capital
Kuala Lumpur within 180 - 260 days per annum
Lightning has an extremely high current, high
voltage and transient electric discharge. It is
transient discharge of static electricity that serves
to re-establish electrostatic equilibrium within a
storm environment [1]. Malaysia lies near the
equator and therefore it is categorized as prone to
high lightning and thunderstorm activities [2].
Observations performed by the Malaysian
Meteorological Services indicate that thunders
occur 200 days a year in Malaysia.
Thunderstorms have been suspected to have
caused between 50% and 60 % of the transient
tripping in the transmission and distribution
networks for Tenaga Nasional Berhad (TNB),
Malaysia’s electric power provider. The main
reason could be short of precise and consistent
202
Journal of Theo
oretical and
d Applied In
nformation Technology
T
y
st
31 December 2011. Vol. 34
4 No.2
© 2005 - 2011 JAT
TIT & LLS. All rig
ghts reserved.
ISSN: 1992-8645
ww
ww.jatit.org
lightniing data in Malaysia
M
to enable
e
throughh
studies on lightning and its mitigation [5].
E
E-ISSN:
1817-31995
2. METHODOLO
M
OGY
P
reggion for Peninsular Malaysiia
2.1 Predefined
Lightnning dischargee occurs wheen the electricc
currennt flows due to
t exert a puull on chargedd
particlles of positive and negative. There
T
are threee
types of lightning discharge
d
(i) clloud-to-groundd
(ii) clloud-to-cloud (intercloud) (iiii) intracloud.
Cloudd-to-ground den
noted as C2G happens whenn
the chharge at the clo
oud strike the ground
g
surface.
Aboutt 25% of liightning strikkes are C2G.
Majorrity of C2G lightning wiill appear ass
growinng ball [5, 6]. Cloud-to-clouud (intercloud))
denoteed as C2C occu
urs when the negative
n
chargee
and poositive charge of different clouds attracted.
Meanw
while, intraccloud lightninng dischargee
happenns when somee charge at thhe base of thee
cloud attract the possitive charge att the top of thee
cloud.
The predefined reegion of Peninnsular Malaysia
has been
b
divided innto nine as illusstrated in Figurre
1. Each
E
column and
a
row is characterized
c
a
as
follow
ws:
1. A1 and B1
B is South West
W
denoted as
a
(SW).
2 A1 and B22 is South denooted as (S).
2.
3 A1 and B3
3.
B is South East
E
denoted as
a
(SE).
4 A2 and B11 is West denoted as (W).
4.
5 A2 and B22 is Central dennoted as (C).
5.
6 A2 and B33 is South Eastt denoted as (E
6.
E).
7 A3 and B1
7.
B is North West
W
denoted as
a
(NW).
8 A3 and B22 is North denooted as (N).
8.
9 A3 and B3
9.
B is North East
E
denoted as
a
(NE).
Positivve lightning taake places wheen the positivee
chargee flows instead
d of electrons.. About 5% off
lightniing is positivee in nature andd tend to comee
about during the peaak of severe thhunderstorm orr
as thee storm is decaaying [6]. Possitive lightningg
travelss outside the cloud
c
and strikkes the groundd
where there is a po
ool of a negatiive charge. Ass
more and
a more negaative charge buuilds at the basee
of thee cloud, moree and more positive
p
chargee
builds in the upper part of the clooud and on thee
groundd beneath th
he cloud. Thhe separationn
between the cloud and the grouund creates ann
electriic field. Thee pull towarrds positivelyy
chargeed particles and negativvely chargedd
particlles arise as ch
harge continuees to build up.
The air
a prevents the electric current from
m
flowinng. However, if this insulaator of air iss
overcoome, a discharg
ge will go off. This dischargee
is know
wn as the negaative lightning.
Figure 1: Predeefined Area off Peninsular
Mallaysia with thee
d
of
o lightning strike mappingg
The development
area for
f Peninsularr Malaysia wiith regards too
lightniing characterisstic is an effo
fort to providee
governnment assessm
ment for any housing andd
industtrial ventures. Such valuablle assessmentss
are; protector design
n, prone area identification,
housinng developmen
nt, theme parkk constructionn
and inndustrial suitab
bility locationss and et cetera.
The liightning strikee is classifiedd according too
level of
o strike (high, medium, low
w) and types off
lightniing (positive cloud-to-grouund, negativee
cloud--to-ground,
flash).
T
The
chosenn
classiffication techniques are (i) fuzzy
fu
logic (ii))
neurall network and (iii)
( neuro-fuzzzy.
o
Althoough, there aree regions whicch cover part of
Indonnesia (SW) annd Singapore (S
SE) and regionns
that may
m not be meaningful
m
suchh as W and NE
N
but the
t idea of thhis research iss to justify thhe
capabbility of AI techniques in lightninng
classsification for mapping
m
purposses.
The lightning
l
data was acquired from the Global
Lighttning Networkk (GLN). The data consists of
o
(i) tiime (ii) latituude (iii) longittude (iv) strikke
curreent.
The
structure
of
lightninng
charaacterization and mapping sysstem is shown in
i
Figurre 2.
203
Journal of Theo
oretical and
d Applied In
nformation Technology
T
y
st
31 December 2011. Vol. 34
4 No.2
© 2005 - 2011 JAT
TIT & LLS. All rig
ghts reserved.
ISSN: 1992-8645
ww
ww.jatit.org
E
E-ISSN:
1817-31995
2.3 AI
A Classificatioon Techniquees
i) Fuuzzy Logic
The area of ligghtning strikee is identifieed
accorrding to it latiitudes and longgitudes numbeer.
‘IF rule’ has beeen used to categorize thhe
lightnning and levells of lightning current becausse
the syystem has singgle input singlee output (SISO
O).
It is simpler to usse ‘IF rule’ raather than fuzzzy
logicc. If the system
m is dealing with
w
a multiple
inputts such as muultiple inputs multiple outpuut
(MIM
MO) system, then fuzzy loogic method is
mostt preferable forr the characterization.
Figure 2: The Structure of Lightning
F
L
Characterization Systtem
The data
d
was divideed into (i) longitude and (ii))
latitudde to identify th
he exact locatioon of lightningg
strike and (iii) lightn
ning strike currrent to classifyy
the levvel of strike cu
urrent in amperee.
The input membeership functioon for latitudde,
longiitude and outpput are displayyed in Figure 3
and 4. The outpput membershhip function is
norm
malized for thee purpose of simplicity. Thhe
equattion used for thhe normalizatioon is:
o Lightning Category
C
2.2. Classification of
Lightnning strike can
n be categorizzed into threee
differeent types wh
hich are posiitive lightningg
(+C2G
G), negative lightning
l
(-C22G) and flash.
C2G is stands for cloud to grouund discharge.
Thus, the main inp
put is the cuurrent flow off
lightniing strike. Thee value is betw
ween -180 kA
A
and 180 kA. The basic
b
‘IF rule’ used for thiss
classiffication is tabu
ulated in Table 1.
Z (true ) = zX 0.52631578
wherre Z(true) is the actual value for
fo the output of
o
x-axiis and z is the value
v
representted in the outpuut
mem
mbership function. Based on the
t membershiip
functtions the fuzzzy rules are developed. Foor
instannt, when the laatitude is A1 annd the longitudde
is B1, the output is declared ass Region Nortth
Westt or NW. All combinations
c
f
from
both inpuut
mem
mbership functtion will be mapped intto
seleccted region based on it lightninng
charaacteristics. Alll the fuzzy rulles used for thhe
regioon classificationn are summarizzed in Table 3.
Table 1: Classsification of liightning
I
Input:
Currrent (A)
<0
=0
>0
Category of
o Lightning
Negative lighhtning (-C2G)
Flaash
Positive lighhtning (+C2G)
A1
A2
A3
1
Once the category is
i identified thhen the system
m
will cllassify the lev
vel of lightningg current (low,
mediuum or high). Th
he strike currennt levels are inn
amperres between -18
80 kA to 180 kA
k as tabulatedd
in Tabble 2.
0.5
0
Table 2: Strrike level definition
Classs
Low
w
Meddium
Highh
1
2.5
4
5.5
Input Varriable “Latitudde”
Rangee
[-60 kA,+60
k
kA]
[-120 kA, -60 kA] and [+60kA,
+120 kA]
kA, -120 kA]] and [+120
[-180k
kA, +180 kA]
204
7
Journal of Theoretical and Applied Information Technology
st
31 December 2011. Vol. 34 No.2
© 2005 - 2011 JATIT & LLS. All rights reserved.
ISSN: 1992-8645
www.jatit.org
E-ISSN: 1817-3195
ii) Neural Network
B1
B2
B3
The implementation of neural network for
classification problems is dependable on their
structure and functions. By considering a set of
classes, the objective in classification is an
assignment of a random sample to one of this
class with minimum probability error. Each
sample is described by a set of parameter which
then forms a vector, usually referred as the
feature vector. The development of such
classification system can be achieved as a result
of neural network training so that it produces the
output which corresponds to one of these classes.
However, the training sample must have similar
form as its input that is belong to the same class.
1
0.5
0
99
100.5
102
103.5
105
Input Variable “Longitude”
Figure 3: Input membership function for
fuzzy logic technique
The ability of neural network to correctly
classify the test sample is subjected to it
generalization ability. Back propagation method
has been used to train the data from Global
Lightning Network (GLN) database. The
network consists of an input layer, one hidden
layer and an output layer. The input layer
consists of 2 input neurons which represent the
longitude and latitude as it input, hidden layer
with 16 hidden neurons while the output layer
represent the 9 output regions that need to be
classified their distributions. Each neuron in the
hidden layer and output layer has a bias which is
connected via weight matrix to the previous
layer. The number of hidden layer depends on
the performance index of the system. Trial and
error approached has been implemented to get an
accurate number of hidden layers for the system.
It begins with a modest number of hidden
neurons and gradually increasing the number if
the network fails to reduce the error. A much
used approximation for the number of hidden
neurons for a three layered network is;
SW S SE W C E N WN NE
1
0.5
0
0
1
Output Variable “Region”
Figure 4: Output membership function for
fuzzy logic technique
Table 3: Fuzzy rules for the region
classification
No.
of
Rule
1
Input
Latitude
Longitude
A1
B1
2
3
A1
A1
B2
B3
4
5
6
7
A2
A2
A2
A3
B1
B2
B3
B1
8
9
A3
A3
B2
B3
Output
Region
North West
(NW)
North (N)
North East
(NE)
West (W)
Central (C)
East (E)
South West
(SW)
South(S)
South
East(SE)
N =1/2(j + k) +P,
(2)
where N is the number of neuron, j and k are the
number of input neurons and P is the number of
patterns in the training set. As the pattern is set
to 1 the number of hidden layer begins with 6
and amplifies to 16.
Optimum weight was calculated using WidrowHoff Rule. The algorithm of the Widrow-Hoff
Rule shows that the network weights are moved
along the negative of the gradient of the
performance function. The derivative of the
205
Journal of Theo
oretical and
d Applied In
nformation Technology
T
y
st
31 December 2011. Vol. 34
4 No.2
© 2005 - 2011 JAT
TIT & LLS. All rig
ghts reserved.
ISSN: 1992-8645
ww
ww.jatit.org
sigmoid function evaluated
e
as the
t
net outputt
value for that neuro
on which cann be expressedd
mathematically as:
= ( tk – yk).f’ (yk-in)
requiires may be too large, ressulting in slow
perfoormance. Usuaally the default value of moost
comm
mercial neural network packkages are in thhe
rangee 0.1-0.3 provviding a satisfactory balancce
betw
ween the two reeducing the learning rate maay
help improve conveergence to a loocal minimum of
o
the error
e
surface. For this proj
oject, the initial
learnning rate is sett at 0.1. The difficulty arisees
whenn training interrnal layer weigghts as no targget
is avvailable. The soolution lies in propagating thhe
errorr back, layer by layer, frrom the outpuut
succeessively to bacckwards layer.. The jth hiddeen
neuroon receives ann error signal as
a the weighteed
sum of the followinng layer error signals,
s
and theen
mmation with thhe derivative of
o
multiiplies this sum
inputt to the neuron by using equaation (5):
(3))
put for the kth output neuron.
where, tk target outp
These deltas values of are used to fine-tune theirr
input connection weights accoording to thee
follow
wing formula:
Wik new = wik old + α.δk.zj
E
E-ISSN:
1817-31995
(4))
In thiis technique, the initial weight
w
for alll
connecction is initiallized to 1.0. There
T
are 1766
connecctions weight, which is the input
i
to hiddenn
connecction (32 weiights) and hiddden to outputt
connecction (144 weiights). The connnection of alll
layers is illustrated in
n Figure 5 beloow.
δj = f’ (zk-in) x δk.wik
(55)
Thenn, this signal will
w be used to adjust the inpuut
weigght wij to the hidden neuroon zj which is
compputed using thee following equuation
Wij new = wij old + α . δi
(66)
Finallly, neuron usses its error siignal to train it
assocciated weight,, then passes it back to all
a
neuroons to which it
i is connected in the previouus
layerr. This two stepp process is reepeated over thhe
trainiing set elemennts until the nettwork convergees
and produces the desired responnse. It must be
b
notedd that all weighht wet set to 1.00 initially.
iii) Neuro-fuzzy
N
Neurro-fuzzy algorrithm is deriived from tw
wo
reliabble artificial inntelligence techhnique known as
a
fuzzyy logic and neural netw
work algorithm
m.
Therefore, the performance
p
o neuro-fuzzzy
of
classsifier is certainnly reliable. Siimilar to fuzzyy,
theree are two typees of neuro-fuuzzy models, (i)
(
Sugeeno and (ii) Mamdani. Suugeno model is
derivved to generatte fuzzy ruless from a giveen
inputt-output data set and it has thhe form of “Iff x
is A and y is B, then z is f(x,yy)”, which is a
typiccal fuzzy rule inn a model. Whhere A and B arre
fuzzyy sets in the anntecedent. The order of Sugenno
modeel is representeed by the outpuut of the system
m,
z. The
T
output iss normally presented
p
in a
polynnomials form of
o input variables x and y. Thhe
coeffficient of x andd y of Sugeno output model is
know
wn as weightedd average of coonsequents. Thhe
weigght will providee a smooth efffect, mature annd
g of lightning strike system
Figurre 5: Mapping
with neural
n
network
k
The layers
l
are weell connected whilst everyy
neuronn in each layerr is connected to every otherr
neuronn in adjacent forward
f
layer. Learning rate,
α effeectively contro
ols the size of
o the step iss
neededd in multidimeensional weighht space whenn
each weight
w
is modiified. If the sellected learningg
rate iss too large then
n the local minnimum may bee
over stopped
s
constan
ntly, resulting in oscillationss
and slow convergencce to lower errror state. If thee
learninng rate is too low,
l
the number of iterationss
206
Journal of Theo
oretical and
d Applied In
nformation Technology
T
y
st
31 December 2011. Vol. 34
4 No.2
© 2005 - 2011 JAT
TIT & LLS. All rig
ghts reserved.
ISSN: 1992-8645
ww
ww.jatit.org
E
E-ISSN:
1817-31995
efficieent gain schedu
uler [4]. Thus, the
t weight willl
minim
mize the deefuzzification process inn
Mamddani model.
layerrs. That represents its membeership functionns
and fuzzy rules. As shown inn Figure 6, thhe
neuroo-fuzzy Area Classifier
C
has several layers.
The Sugeno modeel increases computationall
efficieency and suitaable for linearr and adaptivee
techniques. It also guarantees
g
a coonstancy of thee
outputt surface, whiich is a compplementary forr
mathematical exam
mination. Due
D
to thiss
advanttage, the Sug
geno model offers
o
the bestt
preference for comp
putation modellling. There aree
quite a number off neuro-fuzzy classificationn
applications worldw
wide; one of thhem is numerall
handw
written recogniition which was
w carried outt
by All-Alawi [9]. The
T system haas successfullyy
achievved 98.5% off accuracy. Another
A
neuro-fuzzy classification
n method whhich combinedd
fuzzy--gaussian neuraal network is applied
a
in facee
recognnition system [10]. The system
s
scoredd
100% accuracy on 100 images with
w 10 classes.
For ligghtning classiffication, there are researchess
using the artificial in
ntelligence metthod to classifyy
the areea and the lev
vel of strikes for
f multifractall
characcterization, an
nd fuzzy claassification off
lightniing strike mapss [11]. The claassifier used K-meanss classifier which
w
describbed a novell
approaach for measurrement of phyysical lightningg
strike mapping. Thee data was accuumulated from
m
Canaddian Lightning Detection Netw
work (CLDN).
The liightning strikee mapping wass modeled andd
characcterized to 15 classes thhrough Davis-Boulddin criterion for statisticcal purposedd
[12,133,14].
The first layer or Layer 1 is thhe crisp inputts,
wherre each inputs; x (latitude) annd y (longitudee).
The latitude input is representedd by fuzzy seets
A1, A2 and A3, while the laatitude input is
repreesented by fuzzzy sets B1, B22, and B3. Layeer
2 is the
t layers of Innput Membersship Function or
o
the fuzzification
f
laayer. In this layer, the nodde
repreesents the fuzzzy set used in the antecedennts
of fuuzzy rules. Thee fuzzificationn determines thhe
inputt is fit to whichh fuzzy sets. The
T membershiip
functtion used is thhe triangular tyype as shown in
i
Figurre 7.
(a) Latitude
L
(b) Loongitude
Fiigure 7: Inputt membership function for
n
neuro-fuzzy
Eachh triangular membership
m
fuunctions can be
b
definned by two paraameter (a,b) ass ;
wherre a and b are the
t parameterss that control thhe
centeer and the widtth of the trianglle.
ure 6: Structurre of neuro-fuzzy classifier
Figu
The next
n
layer is thhe Layer 3, is the fuzzy rulees.
Eachh node in this laayer representss the rules of thhe
systeem. The rule for
f each node is listed under
Tablee 4.
The neuro-fuzzy
n
claassifier (NFC) in the system
m
is struuctured as mullti-layer neurall network. Thee
NFC has input outp
put layers andd three hiddenn
207
Journal of Theo
oretical and
d Applied In
nformation Technology
T
y
st
31 December 2011. Vol. 34
4 No.2
© 2005 - 2011 JAT
TIT & LLS. All rig
ghts reserved.
ISSN: 1992-8645
ww
ww.jatit.org
The learning methhod of neuro--fuzzy basicallly
applyy standard bacck-propagationn algorithm. Thhe
back-propagation algorithm willl compare thhe
inputt-output and calculates thee difference as
a
know
wn as error. The
T error thenn is propagateed
backw
wards throughhout the network in reverseed
direcction which meeans from the output layer to
t
the input layer. During
D
the propagation,
p
thhe
neuroon nodes will be affected by
b changing thhe
neuroon activation functions. In determining
d
thhe
necesssary modificaations, the acctivation neuroon
functtion will be differentiated
d
a only then it
and
will be
b modified. They
T
are three ways of neuroofuzzyy learning proocess: (i) membbership functioon
shapees may adjust during trainingg (ii) the bias or
o
weigght between layyer 3 to layer 4 may be formeed
or terrminated (iii) the
t weight bettween layer 3 to
t
layerr 4 behave adapptively by channging its valuees.
How
wever, in thhis particulaar study, thhe
mem
mbership function shapes iss fixed and no
n
addittional or term
minated weightt. The learninng
proceess is only alloowed to changes the weight in
i
betw
ween layer 3 annd layer 4. The performance of
o
neuroo-fuzzy can be measureed with manny
methhods such as Mean
M
Percentaage Error, Rooot
Meann Square Errorr and Mean Sqquared Error. In
I
this
report,
thhe
neuro-fuzzzy
classifieer
perfoormance is deetermined by Mean Squareed
Errorr as equation below:
Table 4:
4 List of Rulees
Node
1
2
3
4
5
6
7
8
9
Antecedent
If x is A1,
is B1
If x is A1,
is B2
If x is A1,
is B3
If x is A2,
is B1
If x is A2,
is B2
If x is A2,
is B3
If x is A3,
is B1
If x is A3,
is B2
If x is A3,
is B3
and y
and y
and y
and y
and y
and y
and y
and y
and y
E
E-ISSN:
1817-31995
Consequennt
Then the area is North
W)
West (NW
Then the area is North
(N)
Then the area is North
East (NE))
Then the area is West
(W)
Then the area is Central
(C)
Then the area is East (E)
Then the area is South
W)
West (SW
Then the area is South(S)
Then the area is South
East(SE)
It is allso important to mention thatt between layerr
3 and layer 4, there are weights located in eachh
node as to establish
h learning meechanism. Thee
layer 4 also has op
ption to have several
s
hiddenn
layers in order to obttain better resuult for learning..
After the rules evalu
uation, the outpput of Layer 4
is fedd into Layerr 5 which is known ass
defuzzzification
laayer.
The
method
off
defuzzzification layerr is the Weight Average (WA))
methood with equatio
on as below:
(8)
wherre t is the targeet values, o is the
t output of thhe
systeem and N is the total record involved. In
I
otherr word, The MSE
M
measuress the average of
o
the square of thhe differencess between thhe
estim
mator (o) and the
t quantities to be estimateed
(t).
where k is the consttant as the singgleton spike inn
the ouutput memberrship function as shown inn
Figuree 7, while thee n is the num
mber of outputt
membbership functio
on in the sysstem. For thiss
system
m the n numberr is 9 since the area classifiedd
has 9 areas
can
n be classifieed, they are;
NorthW
West (NW), North
N
(N), Northeast (NE),
West (W), Centrall (C), East (E
E), Southwestt
d Southeast (SE
E).
(SW), South (S) and
ESULT AND DISCUSSION
N
3. RE
Eachh technique hass been analysedd and the resullts
is described on the following secttion:
F
Logic Analysis
A
3.1 Fuzzy
It is clearly shownn in Figure 8 that
t
majority of
o
the lightning leveel is in LOW
W level whicch
rangiing from 0 too 60 kA. Meanwhile, HIG
GH
levell of current hass the least perccentage. In Maay
20100, the percentagge of all three level of currennt
are quite
q
similar. The
T highest perrcentage in Maay
for HIGH
H
level currrent which was 27%.
Fiigure 7: Outpu
ut membership function
208
Journal of Theoretical and Applied Information Technology
st
31 December 2011. Vol. 34 No.2
© 2005 - 2011 JATIT & LLS. All rights reserved.
ISSN: 1992-8645
www.jatit.org
LEvel of lightning current (%)
100
90
80
70
60
50
40
30
20
10
0
High
Medium
Sep-10
Jul-10
Aug-10
Jun-10
Apr-10
Mei 2010
Mar-10
Jan-10
Feb-10
Dec-09
Oct-09
Nov-09
Sep-09
Low
Sep‐10
Jul‐10
‐C2G
Mar‐10
Flash
Jan‐10
+C2G
Nov‐09
Sep‐09
0
with 15919 strikes, followed by Region South
(S) with 15764 strikes. The state which included
in these regions are Wilayah Persekutuan Kuala
Lumpur, Melaka, Selangor, Johor, Negeri
Sembilan, Perak and Pahang. A high level of
lightning current which occurred in locations
such as Kuala Lumpur and Selangor have the
possibilities to cause a flash flood. This is due to
the locations which is situated in the center of
city.
Table 5: Samples of classified data
Figure 8: Percentage of lightning’s level
current for 12 month
Mei 2010
E-ISSN: 1817-3195
50000 100000
Figure 9: Graph number of lightning
incidence corresponds to the types of lightning
From the results, 90% of the lightning incidence
occurred are negative lightning. Meanwhile,
there are no flash occurred. This is because flash
occurred between the clouds and do not strike
the ground as shown in Figure 9.
The number of lightning incidence occurred in
each month are analyzed corresponds to their
region. Table 5 show the sampled of classified
data for lightning strike for each month. From
the analysis results, October 2009 has the highest
number of lightning incident with 94628 strikes
followed by April 2010 with 87459 strikes. The
least number of lightning occurred in February
2009 with 7056 strikes. In October 2009, the
lightning strike the most in Region Central (C)
209
Latitude
Longitude
3.90398
66
99.87663
98
3.49327
48
100.8141
976
4.44858
7
99.44333
57
3.31488
1
100.9970
393
4.43594
84
99.44902
7
6.74426
04
99.46520
51
6.67659
29
99.37091
31
3.43408
79
100.7442
15
3.21939
51
100.8315
325
1.86789
69
102.1172
289
3.99088
99.98812
79
2.13234
26
101.7601
432
2.12730
15
101.7917
015
3.20616
96
100.8866
052
3.23967
67
100.8903
621
Curre
nt
(A)
2450
0
3280
0
9410
0
3300
0
7880
0
4110
0
1380
0
2020
0
2370
0
4870
0
3370
0
1870
0
1620
0
1800
00
4180
0
Level
of
Curre
nt
Types
of
Lightni
ng
Regi
on
Low
-C2G
W
Low
-C2G
C
Mediu
m
-C2G
W
Low
-C2G
C
Mediu
m
-C2G
W
Low
-C2G
NW
Low
-C2G
NW
Low
-C2G
C
Low
-C2G
C
Low
-C2G
S
Low
-C2G
W
Low
+C2G
S
Low
-C2G
S
High
-C2G
C
Low
-C2G
C
Journal of Theoretical and Applied Information Technology
st
31 December 2011. Vol. 34 No.2
© 2005 - 2011 JATIT & LLS. All rights reserved.
ISSN: 1992-8645
www.jatit.org
The obtained results are then mapped into the
Malaysian map using Google Earth as shown in
Figure 10. This figure shows the locations where
the lightning strike and Figure 11 shows the
mapping of lightning characteristics for one
month data. The light blue, purple and red icons
indicate low, medium and high level of current
respectively. By clicking the icon, a dialog box
as shown in figure below appeared. It tells user
the lightning current value, level of current and
also the region its corresponds to. Based on the
mapping, most of the lightning occurred at the
west coast of Peninsular Malaysia. Most of the
states in the west coast of Peninsular Malaysia
are catogarized as a developed states. Thus, a
high population density and plenty of industrial
location attract the lightning strike.
E-ISSN: 1817-3195
The developed classifier program are able to
classify the lightning parameters according to the
desired characteristics. Based on the statistical
analysis, 90 % percent of lightning incidence are
negative lightning. Meanwhile, the majority of
the levels of lightning current are Low level.
Areas that are situated in the west coast of
Peninsular Malaysia have a higher number of
lightning incidence compared to the east coast.
There are some limitations found in mapping the
lightning characteristics using Google Earth.
3.2 Neural Network Analysis
The data is classified into nine regions as
follows;
• Southern Regions:
¾ S1-latitude =10 to 2.50 & longitude =
990 to 100.50
¾ S2-latitude =10 to 2.50 & longitude =
100.50 to 103.50
¾ S3-latitude =10 to 2.50 & longitude =
103.50 to 1050
• Center Regions:
¾ C1-latitude =2.50 to 5.50 & longitude =
990 to 100.50
¾ C2-latitude =2.50 to 5.50 & longitude =
100.50 to 103.50
¾ C3-latitude =2.50 to 5.50 & longitude =
103.50 to 1050
• Northern Regions:
¾ N1-latitude =5.50 to 70 & longitude =
990 to 100.50
¾ N2-latitude =5.50 to 70 & longitude =
100.50 to 103.50
¾ N3-latitude =5.50 to 70 & longitude =
103.50 to 1050
Figure 10: Mapping of the lightning strike
locations
The system stops at 250 epochs with MSE of
0.0462 as shown in Figure 12.
Mean Square Error
0.3000
0.2500
0.2000
0.1500
0.1000
0.0500
0.0462
0.0000
0
100
200
300
Epoch
Figure 12: The Mean Square Error change
with the change in epoch value
Figure 11: Mapping of the lightning
characteristics
210
Journal of Theo
oretical and
d Applied In
nformation Technology
T
y
st
31 December 2011. Vol. 34
4 No.2
© 2005 - 2011 JAT
TIT & LLS. All rig
ghts reserved.
ISSN: 1992-8645
ww
ww.jatit.org
From initial analysis of the dataa, the averagee
lightniing strikes forr peninsular Malaysia
M
are 199
strikess per hour. Th
his number will
w give aboutt
456 sttrikes per day. The frequencyy of the strikess
shownn that the Pen
ninsular of Malaysia
M
reallyy
need a system that give
g
an informaation about thee
classiffied lightning strikes
s
distributtion in order too
create and design a suitable prottection system
m
that can
c
handle th
he lightning strikes withoutt
givingg interruption to
t the system and
a equipmentt
in the utility area. Th
he lightning strrikes at all ninee
regionns are low February becauuse it can bee
considdered as pacifi
fic monsoon which
w
does nott
hit thee peninsula areaa.
E
E-ISSN:
1817-31995
Figurre 14 shows thhe lightning sttrike of cloud to
t
grounnd (-CG) is thhe highest typpe of strike thhat
occurrred in Peninnsular area. This
T
shows thhat
mostt of the strike is made of negative
n
chargees
that travel
t
from thee cloud to the ground when it
is initiated by the positive chargge on the eartth
surfaaces.
N
A
Analysis
3.3 Neuro-fuzzy
The neuro-fuzzy results is divvided into tw
wo
sectioons, (i) Traininng result, (ii) Data
D Analysis
T
Resultt
(i) Training
Figuree 13, describess the lightningg characteristicc
accordding to the leveel of current foor all months. Itt
showss that the MEDIUM level almostt
manippulates or be th
he most commoon strike for alll
monthh. The LOW an
nd HIGH level basically justt
happenn for several times in a month.
m
So thiss
show that the com
mmon strikes happen
h
in ourr
countrry in peninsu
ular area is a MEDIUM
M
lightniing strike.
Figurre 15 and Tablle 6 represent the neuro-fuzzzy
classsifier system. The graph as
a illustrated in
i
Figurre 15 is the training
t
error of neuro-fuzzzy
classsifier after 250 epoch trainingg. At epoch 2330
the weight
w
changedd in small valuues, which causse
the constant
c
traininng error for thee last 20 epoch.
Thuss, the learning was terminateed at epoch 2330
with result of Meean Squared Error
E
(MSE) of
o
0.04662. The learninng rate and moomentum rate is
seleccted as 0.022 and 0.3 accordingly
a
a
as
recom
mmended by Kamiyama
K
[15]].
Number of Strike
800000
600000
Low
Med
High
400000
200000
0.3000
0.2000
0
MSSE
9 10 11 12
1 1 2 3 4 5 6 7 8
0.1000
Month
Figuree 13: The strik
ke level for Oct 2009-Sept 20100
0.0000
Strike Type
e for Year 2009/2010
0
0
0
50
100
150
200
250
Epoch Index
Fiigure 15: Traiining Error off neuro-fuzzy
classifier
Tablle 6: Neuro-fu
uzzy classifierr training resu
ult
Data Set
Data
Entryy
MSE
Area
Analyysis
Num
mber of Strike
Figu
ure 14: The lig
ghtning strikee’s type from
Oct 20
009-Sept 20100
Trainning
Settinng
211
Traininng
4449 daata
(Randoom pick from all month databasee)
0.0462
NorthW
West:
E
East:
5.57%
2.45%
South
Wesst:
North: 2.00%
0.22%
North East:
E
3.34% South: 6.46%
South East: 7.13%
West: 2.45%
2
%
Centrall: 53.01%
Learninng rate: 0.02
Momenntum rate: 0.3
Accuraacy: 73.5 %
Journal of Theoretical and Applied Information Technology
st
31 December 2011. Vol. 34 No.2
© 2005 - 2011 JATIT & LLS. All rights reserved.
ISSN: 1992-8645
www.jatit.org
E-ISSN: 1817-3195
(ii) Data Analysis
From the analysis, lightning were mainly strikes
at the Central region with an average of 5172
strikes during September 2009 to October 2010.
The Southern region was identified the second
largest lightning strikes area with an average of
4096.5 strikes per year. Meanwhile, the lowest
area was identified at Northern East region with
average of 435.7 strikes per year. However, the
lightning strikes were very minimal during
November to March 2010.
ACKNOWLEDGEMENT
I would like to express sincere gratitude to WSI
Corp and Centre of Excellence for Lightning
Protection (CELP) UPM Collaborate to provide
Global Lightning Data especially for Peninsular
Malaysia. An appreciation also goes to CELP
researchers in Faculty of Engineering, Universiti
Putra Malaysia and research grant from UPM
under RUGS (91864) for their support
throughout the project.
The level of lightning strike has shown that
Peninsular Malaysia has received a relatively
low strike with average of 17380.8 strikes per
month. May 2010 was considered a good month
because it has a well balance of lightning strike
level distribution. The Low strike recorded at
38.86%, Medium at 34.51% and High at 26.63%.
The medium and high Strike was very minimal
during November 2009 until February 2010.
However, the high Strikes were extended until
April 2010.
REFERENCES
[1]. K. Srinivasan and J.Gu. Lightning as
Atmospheric
Electricity,
IEEE
CCECE/CCGEI, Otawa, May 2006.
[2]. L.M. Ong, and H. Ahmad. Lightning Air
Terminals Performance Under Conditions
Without Ionization And With Ionization,”
Institute Of High Voltage And High
Current, 2003. UTM Skudai, Malaysia.
[3]. National Lightning Safety, USA, 2010
[4]. Bjarne Hansen, Dr. M. E. El-Hawary, State
Of The Art of Neural Networks In
Meteorology. Technical University Of Nova
Scotia, March 10, 1997.
[5]. N. Abdullah, M. P. Yahaya and Dr. N. S.
Hudi. Implementation and Use of Lightning
Detection Network in Malaysia, 2nd IEEE
International Conference on Power and
Energy (PECon 08), Johor Baharu,
Malaysia.2008.
[6]. Negative
and
positive
lightning,
http://www.weatherimagery.com/blog/positi
ve-negative-lightning. 2010
[7]. A.R Hileman, "Insulation Coordination for
Power Systems.1999.Marcel Dekker, Inc.
[8]. Jang J.S, S. S. Neuro-fuzzy and Soft
Computing: A Computational Approach to
Learning and Machine Intelligence. New
Jersey: Prentice Hall, Upper Saddler River.
1997.
[9]. Al-Alawi, R. A Hybrid n-tupleneuro-fuzzy
Classifier for handwritten Numerals
Recognition. 0-7803-8359-1/04/2004 IEEE ,
2379-2383.
4. CONCLUSION
Fuzzy logic has been successfully applied in
order to characterize the location into eight
regions to identify the location where the
lightning strikes. Moreover, Basic ‘IF rule’ has
been successfully implemented to characterize
the other two characteristics which are type of
lightning and level of lightning current. The
fuzzy logic and ‘IF rule’ are successfully
implemented and mapped into Malaysian map
using Google Earth. The paper successfully
designs the proposed back propagation neural
networks that combine with properties of IF
THEN Rules that performs as the classifier. They
are trained and tested to classify lightning
characteristics. It also included the design of a
software tool suitable for the training and testing
NN for dataset. The results have shown
considerable degree of confidence of 97% of
accuracy in classification by referring to the
graph of MSE. It was evident that as the number
of epoch increased, the MSE will reduces with
certain parameter been initialize first. The
proposed neuro-fuzzy system has achieved
73.5% accuracy. The proposed neuro-fuzzy is
able to classify the data. Thus further study to
increase the accuracy could be done to enlighten
prospective of lightning classification with
artificial intelligence method.
212
Journal of Theoretical and Applied Information Technology
st
31 December 2011. Vol. 34 No.2
© 2005 - 2011 JATIT & LLS. All rights reserved.
ISSN: 1992-8645
www.jatit.org
[10].
Victor, I.-F.Face Recognition Using a
Fuzzy-Gaussian
Neural
Network.
Proceedings of the First IEEE International
Conference on Cognitive Informatics
(ICCI’02). 2002
[11].
A.Faghfouri, W.Kinser. Multifractal
Characterization and Fuzzy Classification of
Lightning Strike Maps. IEEE, 658-663.,
2002.
[12].
Hongxing Li, C. P.-P. Fuzzy Neural
Intelligent
Systems:
Mathematical
Foundations and the Applications in
Engineering. Florida: CRC Press LLC.
2001.
[13].
Vishwanath Hedge, Udaya Kumar.
Classification of Strokes and Identification
of Stroke location from the Measured Tower
Base Current of LPS to Indian Satellite
Launch Pad-I. The 8th International Power
Engineering Conference (IPEC 2007) ,
1161-1166.
[14].
Kamiyama, N. Tuning of Learning Rate
and Momentum Rate on Backpropagation.
Singapore ICCS/ISITA (IEEE),528-532.
1992.
213
E-ISSN: 1817-3195
Journal of Theoretical and Applied Information Technology
st
31 December 2011. Vol. 34 No.2
© 2005 - 2011 JATIT & LLS. All rights reserved.
ISSN: 1992-8645
www.jatit.org
E-ISSN: 1817-3195
AUTHOR PROFILES:
M. Z. A. Ab Kadir graduated
with B.Eng degree in Electrical
and Electronic from University
Putra Malaysia in 2000 and
obtained his PhD from the
University of Manchester,
United Kingdom in 2006 in
High Voltage Engineering.
Currently, he is an Associate Professor in the
Department of Electrical and Electronics
Engineering, Faculty of Engineering, Universiti
Putra Malaysia. To date he has authored and coauthored over 80 technical papers comprising of
national
and
international
conferences
proceedings and citation indexed journals. His
research interests include high voltage
engineering, insulation coordination, lightning
protection, EMC/EMI, keraunamedicine and
power system transients.
M.K Hassan was born in
Melaka,
Malaysia.
He
received the B.Eng (Hons)
degree in Electrical and
Electronic from University of
Portsmouth, UK in 1998,
M.Eng
Electrical
Engineering from Universiti
Teknologi Malaysia in 2001 and Ph.D degree
from Universiti Putra Malaysia in 2011.
Currently he is a senior lecturer at the
Department of Electrical and Electronic
Engineering, Universiti Putra Malaysia. His area
of interest includes; control system, automotive
control, electric vehicle, AI applications and
renewal energy.
A. Che Soh, received her
B.Eng.
degree
in
Electronic/Computer
and
M.Sc. degree in Electrical
and Electronics Engineering
from the Universiti Putra
Malaysia (UPM), Malaysia
in 1998 and 2002, respectively. In year 2011, she
received a PhD Degree at Universiti Teknologi
Malaysia (UTM), Malaysia. Currently, she is a
Senior Lecturer at the Universiti Putra Malaysia
(UPM), Malaysia. Her research interests include
artificial intelligence, process control, control
systems, robotics, and automation.
R.Z.
Abdul
Rahman
received her B.E. degree in
Electrical and Electronics
Engineering
from
the
Liverpool
John
Moores
University, United Kingdom
in 1998, and her M.E. degree
in Electrical Engineering from Universiti
Teknologi Malaysia, Malaysia in 2001.
Presently, she is a lecturer at Universiti Putra
Malaysia, Malaysia. She is currently pursuing
her PhD degree in fault detection and diagnosis
at Universiti Teknologi, Malaysia. Her research
interests include artificial intelligence, process
control, control systems, fault detection, and
diagnosis.
214