Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
23
,
No.
3
,
Septem
ber
2021
, pp.
1
814
~
182
4
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v
23
.i
3
.
pp
1814
-
1824
1814
Journ
al h
om
e
page
:
http:
//
ij
eecs.i
aesc
or
e.c
om
Develop
ment of
site
-
sp
ecific non
-
i
ntrusiv
e load m
on
itor
ing
fo
r
maximu
m dema
nd
co
nt
rol
Az
ha
rudin
Mukht
aruddi
n
1
,
Fakro
ul Ridz
ua
n
H
as
him
2
, Ma
t Kamil
A
w
an
g
3
,
Hu
sin
Mama
t
4
,
Hafiz
i
Z
ak
ari
a
5
1,2,3
Facul
t
y
of En
gine
er
ing, Univers
it
i
Per
ta
han
an Nasional
Ma
lay
s
ia
,
Mal
a
y
s
ia
4
School
of
A
ero
spac
e Engi
n
ee
ri
ng,
Univer
si
ti Sa
ins Ma
lay
si
a
,
Malay
si
a
5
Perm
odal
an
Na
sional
B
erh
ad
,
Malay
s
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Oct
23
,
202
0
Re
vised
J
ul
31
,
2021
Accepte
d
Aug
19
,
2021
Dem
and
-
side
lo
ad
m
ana
gemen
t
(DS
M)
req
uires
gre
ater
ro
le
-
pl
a
y
b
y
end
-
users.
To
lower
the
inve
stm
ent
for
thi
s
loa
d
m
ana
gement
co
nce
pt
,
non
-
int
rusive loa
d
m
ana
gement
(NIL
M) was
int
roduc
ed
as
th
e
solut
io
n.
How
eve
r
,
m
ost
of
the
m
at
hemati
c
al
techni
ques
used
in
N
I
LM
are
comple
x.
Thi
s
m
a
y
hinde
r
users
fro
m
ac
ti
v
ely
ta
k
e
par
t
in
th
e
ene
r
g
y
m
an
age
m
ent
eff
ort
.
Thi
s
pape
r
expl
ore
s t
he
poss
ibi
litie
s
o
f
apply
ing
ch
ang
e
point
de
te
c
ti
on
technique
s
with
hel
p
of
di
ffe
ren
t
ia
t
ion
an
d
appl
icat
ion
of
fil
te
rs
.
The
se
fil
ters
were
s
el
ecte
d
strictl
y
base
d
on
site
-
spec
ific
condi
t
i
ons.
As
par
t
of
the
NILM
implementa
t
ion,
a
new
and
pra
cti
ca
l
t
ec
hn
ique
wa
s
deve
lope
d
for
t
his
pape
r.
It
was
found
t
hat
the
develop
ed
te
chni
qu
e,
d
espit
e
it
s
sim
pli
ci
t
y
i
t
ca
n
ide
nti
f
y
th
e
e
le
c
tri
c
al
equi
pm
ent
w
hic
h
adde
d
th
e
significan
t
lo
a
d
demand.
The
per
form
ance
of
the
te
chn
iq
ue
was
found
to
be
sati
sfac
tor
y
as
compare
d
to
resul
ts r
epor
ted b
y
oth
er
r
ese
ar
che
rs.
Ke
yw
or
ds:
Dem
and
-
side
load m
anag
em
e
nt
Ed
ge dete
ct
ion
Ele
ct
rical
en
er
gy ef
fici
ency
Non
-
i
ntr
us
ive
lo
ad
m
anag
em
e
nt
Sit
e
-
sp
eci
fic
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
B
Y
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Azh
a
r
ud
i
n bin M
ukhtar
uddin
Dep
a
rtm
ent o
f El
ect
rical
an
d
Ele
ct
ro
nic
En
gi
neer
in
g
Un
i
ver
sit
i Pe
rtahan
a
n Nasi
onal
Mal
ay
sia
Kem
Su
ngai
B
esi
, 57000
K
ua
la
Lu
m
pu
r, M
a
la
ysi
a
Em
a
il
:
azharud
in@
upnm
.ed
u.m
y
1.
INTROD
U
CTION
Ele
ct
rici
ty
is
a
vital
ener
gy
s
ource
f
or
our
m
od
ern
so
ci
et
y.
Accord
i
ng
t
o
the
Wo
rl
d
Ba
nk,
in
2018
alm
os
t
89.5
%
of
the
w
or
l
d
popula
ti
on
has
acce
ss
to
el
ect
rici
ty
wh
il
e
f
or
Ma
la
ysi
a
it
is
10
0%
[1]
.
M
al
ay
sia
recorde
d
a
n
a
nnual
2.6%
i
ncrea
se
in
el
ect
ric
it
y
sal
es
and
a
3.1%
jum
ped
in
m
axi
m
u
m
de
m
and
in
2019
[2]
.
This
is
des
pite
the
introd
uction
of
e
nergy
-
e
ff
ic
ie
nt
eq
uipm
ent,
program
m
es
and
le
gislat
ion
.
T
o
re
duce
th
e
el
ect
rici
ty
acro
ss
al
l
sect
ors,
a
de
m
an
d
-
s
i
de
m
anag
e
m
ent
(D
SM
)
u
nit
ha
s
bee
n
set
up
unde
r
the
Ma
la
ysi
an
energy
com
m
i
ssion
[
3]
.
DSM
con
ce
pt
re
f
ers
to
init
ia
ti
ve
s
an
d
te
ch
no
l
og
ie
s
t
hat
enc
oura
ge
co
nsu
m
ers
to
op
ti
m
iz
e
their
energy
us
e
[
4]
.
Am
on
g
th
e
obj
ect
ives
of
DS
M
are
to
red
uc
e
the
pe
ak
load
or
pe
ak
loa
d
cl
ipp
in
g,
overa
ll
r
ed
uction o
f e
nergy
us
age
a
nd loa
d
s
hiftin
g
[5]
.
Tw
o
m
ajo
r
ap
proac
hes
f
or
l
oa
ds
m
on
it
or
in
g
are
intr
us
ive
a
nd
non
-
intr
us
i
ve
loa
d
m
anag
e
m
ent
(I
LM
and
N
ILM,
res
pecti
vely
)
c
ou
l
d
be
c
on
si
der
e
d
[
6]
.
T
he
f
orm
er
ap
proach
requires
t
h
e
e
xt
ensive
i
ns
ta
ll
at
ion
of
sens
or
s
to
be
i
ns
ta
ll
ed
on
eve
ry
el
ect
rical
m
achine.
N
ILM
is
an
i
nteresti
ng
a
ppr
oach
in
wh
ic
h
the
ag
gr
egated
data
at
the
power
s
ource
wi
ll
be
deco
m
po
sed.
F
ro
m
the
deco
m
po
sit
io
n
data,
it
is
po
ssible
to
ide
nt
ify
th
e
identit
y
and
o
pe
rati
on
al
sta
tu
s
of
each
m
achine
[
7]
,
as
i
f
a
physi
cal
sens
or
is
bein
g
instal
l
ed
instea
d.
NILM
is
al
so
kn
own
as
the
sin
gle
poin
t
m
easur
em
ent
appr
oach.
O
nc
e
the
m
achine
is
identifie
d
a
nd
it
s
c
urren
t
e
nergy
consum
ption
is
know
n,
acti
on
can be ta
ke
n o
n
the
m
achine to
ac
hieve a
ny
or all
D
SM
obje
ct
ives.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Develo
pm
e
nt
of
sit
e
-
sp
eci
fi
c non
-
intr
us
iv
e lo
ad mo
nitorin
g f
or
maxi
mum
…
(
Azha
r
udin
Muk
hta
r
uddin
)
1815
Eve
n
th
ough
s
cor
es
of
N
ILM
te
chn
iq
ues
ha
ve
bee
n
pu
t
f
or
ward
by
resear
cher
s
,
it
was
note
d
that
the
m
ai
n
pr
oble
m
is
to
hav
e
it
as
a
reli
able
so
lu
ti
on
[
8],
[9]
.
A
n
ap
par
e
nt
pro
blem
with
cur
r
ent
NI
LM
pr
ac
ti
ce
is
rel
at
ed
to
the
i
naccuracy
in
t
he
m
achine’
s
i
den
ti
ficat
io
n
a
nd
al
s
o
to
t
he
app
li
cat
io
n
of
com
plex
te
chni
qu
es
that
even
t
ually
hinders
us
e
rs
to
ta
ke
pa
rt
in
this
te
chnolo
gy
.
This
pa
pe
r
is
go
i
ng
to
pro
m
ote
a
m
et
ho
d
that
is
read
y t
o be
us
e
d by m
os
t ener
gy p
la
ye
rs
.
T
hi
s w
il
l be
the m
ai
n
co
ntri
bu
ti
on to
the
N
ILM
stud
y.
Anothe
r
s
hortc
om
ing
of
t
he
c
urren
t
NI
LM
pract
ic
es
is
the
con
ce
ntrati
on
of
t
he
stu
dies
m
os
tly
on
the
su
bject
of
resi
den
ti
al
or
el
ect
rical
app
li
ance
s.
This
ca
n
be
pro
ved
from
t
he
li
st
of
data
set
avail
ab
le
f
or
t
he
NI
LM
stu
dy
[
10
]
.
F
ro
m
the
li
st,
on
ly
abou
t
5
data
set
s
ar
e
from
so
ur
ces
oth
er
t
han
res
identia
l
or
el
ec
tric
al
app
li
anc
es.
A
lot
of
ot
her
re
searche
rs
al
s
o
us
e
d
data
tha
t
was
m
easur
e
d
in
a
re
side
nt
ia
l
bu
il
ding
or
just
m
easur
ed
thei
r
data
us
i
ng
la
b
-
based
el
ect
rical
con
s
umpti
on
set
up
[
11
]
-
[
13
]
.
Wh
i
le
al
l
the
res
earc
h
con
t
rib
uted
t
o
the
un
der
sta
ndin
g
of
NILM
,
m
or
e
researc
h
on
t
he
im
pl
e
m
entat
ion
of
NI
LM
i
nto
bu
siness
pr
em
ise
s
is
need
ed
[14]
.
T
his
stud
y
ech
oes
the
need
for
r
eal
i
m
ple
m
entat
ion
of
N
ILM
at
a
big
ger
sc
al
e.
A
ca
m
pu
s
-
wide
dat
a colle
ct
ion
s
yst
e
m
w
as sele
ct
ed
as t
he
tria
l
grou
nd to
t
he am
biti
on
.
2.
RELATE
D
W
ORKS
As
far
as
t
he
i
m
ple
m
entat
i
on
of
NILM
on
-
sit
e
is
c
oncern
e
d,
the
re
are
se
ver
al
pr
opos
al
s
m
ade
avail
able
by
di
ff
e
ren
t
researc
her
s
.
F
or
exam
p
le
,
a
five
-
sta
ge
i
m
ple
m
entat
i
on
of
NI
LM
ha
s
bee
n
pro
po
s
ed
by
a
gro
up
of
rese
arch
e
rs
i
n
[15]
.
Othe
r
t
han
tha
t,
a
th
ree
-
ste
p
gen
e
ral
fr
am
e
work
of
N
ILM
was
disc
us
se
d
in
[6]
,
[16]
.
H
ow
e
ve
r
,
auth
or
s
i
n
[
14
]
,
[
17]
pro
pose
d
a
four
-
st
ep
proce
dure
to
carry
out
pract
ic
al
NI
LM
.
This
pro
po
sal
was
c
ho
s
en
due
to
it
s
su
it
abili
ty
with
this
stu
dy
a
nd
it
is
sel
ect
ed
as
the
one
f
ol
lowed
by
this
pap
e
r.
Am
on
g
t
he rea
so
ns i
s it
s
flo
w
is lo
gical
and i
t i
s su
it
able
for
sit
e
-
sp
eci
fic
usa
ge. The
steps
are:
Data coll
ect
io
n an
d pr
e
-
pr
oce
ssing.
Eve
nt d
et
ect
io
n
-
c
hange
of s
ta
te
in
the m
achine
’s op
e
rati
on
ov
e
r
ti
m
e.
Feat
ur
e
ex
tr
act
ion
-
uniq
ue
m
achine
’s
si
gn
at
ur
e
for exam
ple in voltag
e, c
urre
nt and
powe
r.
Loa
d
ide
ntific
a
ti
on
.
All
ste
ps
will
be
disc
us
se
d
a
s
par
t
of
the
m
et
ho
dolo
gy.
Fo
r
the
data
c
ollec
ti
on
a
nd
pre
-
processi
ng
ste
p,
a
detai
le
d
ov
er
view
of
th
e
syst
e
m
is
giv
en
as
a
back
dr
op
on
the
thr
ee
scop
es
or
lim
i
ta
ti
on
s
of
this
s
tud
y.
The
li
m
i
ta
ti
on
s
are
the
el
ect
ri
cal
par
am
et
e
r
avail
able
f
or
a
naly
sis,
the
dur
at
ion
of
t
he
stu
dy,
a
nd
t
he
sa
m
pl
ing
fr
e
qu
e
ncy.
All
lim
it
at
ion
s ar
e
du
e
to
t
he
sit
e
conditi
on.
The
first
li
m
it
a
ti
on
is
on
ly
the
el
ect
rical
cur
r
ent
sign
al
wa
s
consi
der
e
d
in
t
his
stud
y.
T
his
is
the
sit
e
-
const
raint
lim
i
ta
ti
on
.
A
l
ot
of
disc
us
sio
ns
hav
e
bee
n
m
a
de
on
w
hat
ar
e
the
best
pa
r
a
m
et
ers
to
be
us
e
d
in
NI
LM.
Am
on
g
the
pr
op
os
al
s
are
el
ect
rical
po
we
r,
cu
rr
e
nt,
and
a
dm
i
tt
ance
[1
1],
[
13]
,
[18
]
.
Nev
ert
heless
,
the
decisi
on
to
use
an
el
ect
rical
current
si
gn
al
is
su
pp
or
te
d
by
auth
ors
in
[
17]
,
[
19
]
,
i
n
the
form
er,
the
auth
ors
ho
l
d
to
the
opi
nion that t
he
e
ven
t i
s
m
or
e pr
onounce
d
i
n
th
e cu
rr
e
nt w
a
ve
form
.
Anothe
r
li
m
i
t
at
ion
is
the
pe
rio
d
of
st
udy.
This
is
a
ga
in
a
sit
e
-
im
po
se
d
rest
rict
ion
.
T
he
fina
l
lim
it
at
ion
is
the
data
sam
plin
g
f
req
ue
ncy.
It
was
deci
ded
t
hat
the
sam
pli
ng
s
houl
d
be
a
t
5,
10,
15,
30
and
60
m
inu
te
s.
Ag
ai
n,
seve
ral
res
earche
rs
cam
e
ou
t
with
sev
eral
proposals
on
the
sam
pl
ing
sp
ee
d
[20],
[21]
.
Howe
ver,
in
[
22
]
it
was
m
entione
d
that
a
desirab
le
sa
m
pl
ing
fr
e
que
ncy
dep
e
nds
on
the
ty
pe
of
load
.
Fu
rt
her
m
or
e,
no
t
al
l
e
xisti
ng
data
ac
qu
isi
ti
on
syst
em
s
are
re
adily
eq
uipped
with
hi
gh
-
f
reque
ncy
data
sam
pling
capab
il
it
y.
This
sh
or
tc
om
ing
give
s
this
researc
h
to
ex
plore
wh
et
her
al
te
r
na
ti
ve
sa
m
pling
tim
e
is
su
it
able to
b
e
used i
n NI
LM.
T
his ca
n
al
s
o b
e seen
as a
no
t
he
r
c
on
tri
bu
ti
on of this
stu
dy.
A
gr
oup
of
te
chn
i
qu
e
s
ch
ose
n
as
th
e
eve
nt
detect
ion
on
the
ag
gr
e
gate
data
set
is
dis
cusse
d.
T
he
even
t
detect
ion
a
s
pro
posed
in
the
la
nd
m
ark
pap
e
r
of
NI
LM
m
ay
in
cl
ud
e
filt
erin
g,
diff
e
ren
ti
at
in
g,
pe
a
k
detect
ion
[23]
and
al
s
o
fitt
in
g
m
et
ho
d
[
11
]
.
However,
for
this
stud
y,
on
l
y
diff
ere
ntiat
io
n
an
d
filt
erin
g
wer
e
app
li
ed
.
A
det
ai
le
d
ex
planati
on
of
the
te
c
hniq
ue
ca
n
be
f
ound
in
[
8],
[
11]
.
Wh
il
e
exte
ns
ive
re
view
c
an
be
fou
nd in
[17]
.
The dat
a sam
pling
fr
e
quency i
s also i
nclu
ded in the
d
isc
us
si
on.
The
thir
d
ste
p
i
s f
eat
ur
e e
xtrac
ti
on
b
y detec
ti
ng
t
he
ed
ge
of
the f
il
te
red
differentia
ti
on
dat
a. Sin
ce the
ou
tc
om
e
of
the
filt
er
m
ay
be
m
ade
of
se
que
ntial
data
po
i
nts
change
po
i
nt
detect
ion
t
ech
ni
qu
e
was
ap
plied,
i
n
par
ti
cula
r
ed
ge
detect
ion
.
T
his
te
chn
iq
ue
is
us
e
d
to
determ
ine
on
ly
cha
nges
in
data
po
i
nt
s
[24]
.
Fi
nally
,
for
the
f
ourt
h
ste
p,
a
discu
ssio
n
about
t
he
c
onf
us
io
n
m
at
rix
a
nd
it
s
ap
plica
ti
on
in
ide
ntifyi
ng
loa
ds
is
pre
sented
.
This
com
m
on
te
chn
i
qu
e
is
use
d
to
s
how
t
he
perform
ance
of
the
pro
po
se
d
te
chn
iq
ues
.
A
detai
le
d
ex
plan
at
io
n
of
the
c
onfusion
m
at
rix
is
t
o
be
disc
us
se
d
in
the
m
e
tho
dolo
gy
sect
io
n.
All
analy
sis
us
in
g
each
of
the
te
chn
iq
ues
w
as
r
eal
ise
d usi
ng
Si
m
ulink
®
.
Diff
e
re
nt
from
o
ther
st
ud
ie
s
,
this
stud
y
rea
di
ly
has
the
aggreg
at
e
an
d
in
di
vidual
data
set
.
Hen
ce
the
deco
m
po
sit
io
n
ste
p
was
om
i
tted.
W
hat
this
s
tud
y
is
goin
g
to
dem
on
strat
e is
the
app
li
cat
ion
o
f
s
om
e
relat
ively
com
m
on
te
chn
iqu
es
that
co
uld
al
so
be
co
ns
i
der
e
d
to
be
us
e
d
i
n
NI
LM
.
Since
the
te
chn
i
ques
em
plo
ye
d
i
n
thi
s
stud
y
is
not
as
com
plex,
pro
blem
-
relat
ed
N
ILM
ex
plaine
d
in
the
I
ntr
oduc
ti
on
sect
io
n
c
an
be
overc
ome
.
O
n
top
of that,
actual
ope
rators
of e
nergy m
anag
em
ent can
fi
na
ll
y pr
act
ic
e NILM
as
pa
rt of
their
works.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2021
:
18
14
-
18
24
1816
3.
RESEA
R
CH MET
HO
D
Data
us
e
d
in
t
his
stu
dy
cam
e
from
a
un
iver
sit
y
ca
m
pu
s.
T
he
cam
pu
s
ha
s
about
20
buil
dings
within
250
ac
res
of
la
nd
a
rea.
M
os
t
of
t
he
buil
di
ngs
are
c
onnecte
d
to
a
ce
ntral
bu
il
di
ng
aut
om
at
ion
syst
e
m
(BAS).
Using
the
syst
e
m
,
a
certai
n
l
evel
of
c
on
tr
ol
s
on
sel
ect
ed
e
le
ct
rical
equ
ip
m
ent
is
avail
able.
It’s
al
so
has
a
ver
y
extensi
ve
m
o
nitor
i
ng
ca
pa
bili
ty
m
ai
nly
on
HVAC,
energy
m
on
ito
ri
ng
a
nd
sa
f
et
y/
secur
it
y
aspects
.
Mon
it
ori
ng
on
the
m
ai
n
el
e
ct
rical
su
pply
for
eac
h
m
ajor
buil
ding
i
nclud
e
s
pa
ram
et
e
rs
s
uc
h
as
vo
l
ta
ges,
currents
, p
owe
r
fact
or,
al
l
ty
pe
s
of
el
ect
rical
p
owe
r
as w
el
l
t
he
sta
tus
o
f
sel
ect
ed
ci
rc
uit
br
eakers
. H
owe
ve
r,
at
the
eq
uip
m
ent
le
vel,
on
ly
el
ect
rical
curren
t
is
m
on
it
or
e
d,
exce
pt
f
or
la
r
ge
loa
ds
su
c
h
as
chill
ers
that
ha
ve
m
or
e
m
on
it
or
i
ng
p
a
ram
et
ers.
H
ence
only
cu
r
ren
t
data is co
nsi
der
e
d
in this
stud
y a
nd
it
g
i
ves
the
first sc
op
e
of
this stu
dy.
The
syst
em
h
as a central m
onit
or
in
g
syst
em
,
o
r w
orkstat
io
n (
WS)
,
locate
d
in a f
aci
li
ti
es office.
Fro
m
the
WS,
a
n
op
erator
can
m
onit
or
a
real
-
tim
e
sit
uation
of
t
he
cam
pu
s
inclu
ding
da
ta
on
el
ect
rical
par
am
et
ers.
The
data
can
a
lso
be
rec
orde
d
in
the
W
S
at
a
m
ini
m
u
m
1
-
seco
nd
sam
pli
ng
f
re
qu
e
ncy.
The
sp
a
n
of
th
e
data
colle
ct
ion
is
lim
it
ed
by
the
a
vaila
ble
disk
s
pace.
H
oweve
r
,
durin
g
the
vi
sit
to
evaluate
th
e
syst
e
m
,
the
WS’
disk
s
pace
was
low
and
a
lot
of
el
ect
rical
m
on
it
ori
ng
point
s
wer
e
unavail
able.
The
refore
,
the
data
colle
ct
ion
po
i
nts
(buildin
g
an
d
m
achines)
,
the
sam
pling
fr
e
quency,
a
nd
the
recordi
ng
s
pa
n
m
us
t
be
caref
ully
sel
ect
ed.
Du
e
to
this
,
th
e
seco
nd
sco
pe
of
this
stu
dy
is
the
lim
it
at
ion
of
the
spa
n
of
data
colle
ct
ion
to
only
three
days.
The
rec
ordin
g
was
sta
rted
at
0700
on
day
-
1
and
en
de
d
at
1555
on
day
-
3.
Howev
e
r,
the
analy
sis
will
on
ly
be
done o
n data st
arted
on
day
-
1
and en
de
d
at
t
he
end
of the
bu
siness
hour
on
day
-
2.
An
a
dm
inist
rati
ve
buil
ding
was
c
ho
se
n
f
or
t
he
data
colle
ct
ion
.
It
ha
s
one
el
ect
ri
ci
ty
m
a
in
switc
hboa
rd
(M
SB)
w
hich
i
n
tur
n
s
upplies
the
bu
il
di
ng.
F
our
pieces of
m
achiner
y
that
dr
a
w
their p
ower f
r
o
m
the
sam
e
MSB
wer
e
sel
ect
ed
.
The
o
pe
rati
on
of
the
buil
ding
is
m
a
inly
0800
t
o
17
00
with
a
sm
al
l
pr
ese
nce
of
24
-
hour on
-
dut
y sec
ur
it
y pe
rs
on
al
s
.
The
sel
ect
ion
was
base
d
on
the
data
avail
a
ble
on
the
m
a
chine.
T
hey
al
so
re
pr
es
ent
a
sign
ific
a
nt
portio
n
of
the
powe
r
delive
re
d
by
the
M
SB.
On
e
of
t
he
m
a
chines
was
a
water
-
c
ool
pac
kag
e
ai
r
-
c
ondit
ion
in
g
un
it
(
WCPU),
ano
t
her
was
th
e
coo
li
ng
to
we
r
m
oto
r
(CT)
a
nd
the
oth
e
r
two
wer
e
water
pu
m
ps
(CD
W
1
an
d
CD
W2
)
.
Ma
ny
of
the
li
te
rature
agr
eed
on
f
our
-
cl
assifi
cat
io
n
of
loa
ds
[
6],
[8
]
,
[
22
]
.
All
the
ab
ov
e
-
m
ent
ion
e
d
loads
can
be
c
at
egorised
as
t
ype
I
I
m
ulti
-
sta
te
m
achines
(
al
so
known
as
finite
sta
te
m
a
chines
(FSM
).
It
has
a
rep
eat
a
ble sw
it
chin
g patt
ern.
Du
e
to
t
his,
it
i
s easie
r
t
o
ide
nt
ify
.
The
data
gathe
red
i
n
this
stu
dy
can
be
div
i
de
d
into
t
w
o
ty
pe
s:
the
global
data
w
hich
wa
s
sam
pled
at
the
MSB
,
a
nd
the
data
ta
ken
from
fo
ur
el
ect
rical
m
achines.
Both
ty
pe
s
of
data
we
re
sam
pled
at
e
ver
y
5,
10,
15,
30
a
nd
60
m
inu
te
s,
ov
e
r
t
hr
ee
days
.
T
he
sam
pl
ing
f
re
quency
is
sc
op
e
num
ber
th
ree.
The
decisi
on
of
the
sam
pling
tim
e
was
m
ade
bas
ed
on
the
a
vaila
bili
ty
and
ca
pab
il
it
y
of
the
syst
em
as
well
as
du
e
to
a
r
evie
w
done
i
n
[22]
m
entioni
ng that
a d
esi
ra
ble sa
m
pl
ing
fr
e
quen
cy
d
epe
nds
on
the ty
pe of
loa
d.
Figu
re
1
s
how
s
an
exam
ple
5
-
m
inu
te
interv
al
of
cu
rr
e
nt
re
adin
g
at
ph
ase
-
Y
f
or
the
MSB
,
or
wh
at
is
al
so
kn
own
as
aggre
gate
el
ect
rical
current d
a
ta
.
The
c
urre
nt
value
is repres
ented
by
the v
e
rtic
al
axis
w
hile
the
nu
m
ber
s
of
dat
a are on
the
ho
rizo
ntal axis. T
he
l
at
te
r
axis
m
ay
also b
e re
pr
ese
nted wit
h t
he
tim
e o
f
the
day. A
gen
e
ral
patte
r
n
can
be
seen
re
peated
for
eac
h
day.
O
n
the
f
irst
day,
the
c
urre
nt
rea
ding
is
arou
nd
50
A
be
fore
abru
ptly
cl
i
m
b
ed
to
a
bout
20
0
A
within
30
m
inu
te
s
from
0800
ho
ur
s
.
It
hove
rs
ar
oun
d
250
A
be
fore
pl
unged
150
A
at
13
00
and
sta
ys
the
re
f
or
an
ho
ur
.
It
then
s
hoots
ba
ck
to
a
highe
r
value
of
a
bout
250
A
at
ab
out
1400
up
t
o
17
00.
T
he
sam
e
patte
rn
rep
eat
s
for
the
seco
nd
day,
al
beit
lowe
r
cu
rrent
m
agn
it
ud
e
and
a
certai
n
dro
p
at
arou
nd 15
00 d
ue
to
un
known
r
eas
ons.
Figure
1. A
n
e
xam
ple o
f
c
urr
en
t rea
ding
on
ph
a
se
-
Y of t
he M
SB
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Develo
pm
e
nt
of
sit
e
-
sp
eci
fi
c non
-
intr
us
iv
e lo
ad mo
nitorin
g f
or
maxi
mum
…
(
Azha
r
udin
Muk
hta
r
uddin
)
1817
To
proces
s
the
featur
e
extrac
ti
on
a
non
-
el
ec
tric
al
data
was
e
m
plo
ye
d
in
t
his
stud
y,
s
pec
ific
al
ly
the
tim
e.
The
tim
e
con
strai
nt
is
be
tween
0800
to
1730
hour
s
.
An
y
data
beyo
nd
the
se
hours
wer
e
not
co
nsi
der
e
d
for
the
featu
re
extracti
on.
In
this
w
ork,
t
he
ou
t
put
from
the
eve
nt
det
ect
ion
was
di
r
ect
ly
us
ed
to
extract
featur
e
s
pr
es
um
ably
or
igi
nated
f
r
om
sp
eci
fic
el
ect
rical
equi
pm
ent.
The
da
ta
will
be
go
i
ng
th
rou
gh
a
se
r
ie
s
of
processes
sta
rting
with
de
riva
ti
on
.
T
his
first
process
s
eeks
t
he
m
agn
it
ud
e
of
i
ncr
ease
of
current
from
tim
e
n,
t
n
,
to
the
oth
e
r,
t
n+s
,
wh
ere
s
is
the
sam
pling
f
reque
ncy.
At
this
sta
ge,
d
et
ec
ti
on
of
the
c
ha
ng
e
s
in
data p
oi
nts
is
m
ade.
Nex
t,
th
e
outc
om
e
of
the
der
i
vatio
n
i
s
the
n
s
ubj
ect
e
d
t
o
a
filt
er.
Th
e
filt
er
is
a
high
-
pass
filt
er,
w
it
h
th
e
m
agn
it
ud
e
of
the
filt
er
is
se
le
ct
ed
based
on
the
co
nd
it
io
n
of
lo
ads
.
It
will
al
low
a
c
ertai
n
value
of
the
der
i
vation
t
o
pa
ss.
T
he
outc
om
e
of
the
filt
er is then
c
heck
e
d
f
or
a
posit
ive
ed
ge
. T
his step
is desi
gn
e
d b
ecaus
e
the outp
ut of th
e filt
er m
ay
s
pan
m
or
e tha
n on
e
sam
ple p
oin
t. S
ubseq
ue
ntly
, th
is edg
e is
us
e
d
to
determ
ine t
he
load
i
den
ti
ficat
ion
.
To
gauge
the
perform
ance
of
the
sel
ect
ed
m
et
ho
ds
of
di
sag
gr
e
gation,
[
24
]
-
[
26
]
disc
usse
d
se
ver
a
l
relat
ed
in
dicat
or
s
.
This p
a
per
will
us
e
preci
sion
, r
ecal
l,
an
d
F
β
score
as
the
perform
ance
m
at
rices.
Term
s
us
e
d
in
al
l
of
the
m
at
rices
are
par
t
of
w
hat
is
cal
le
d
a
c
onf
us
io
n
m
a
trix
[
25
]
.
P
recisi
on
f
or
thi
s
pa
per
is
de
fined
as
the
rati
o
of
al
l
detect
ed
e
dg
e
s
(tr
ue
cha
nge
po
i
nt,
TP
)
ov
erall
change
points,
i
nclu
ding
no
n
-
e
dges,
or
false
change
point,
FP.
Mat
hem
ati
cal
ly
[24]
,
(1)
Re
cal
l,
or
al
so
know
n
as
se
nsi
ti
vity
,
is
the
portio
n
of
the
total
TP
t
hat
is
co
rr
ect
ly
ide
nt
ifie
d.
T
hat
m
eans
this
m
atr
ix
m
us
t
ta
ke
i
nto
acco
unt
al
l
edg
es
.
This
te
rm
is
kn
own
a
s
false
non
-
e
dg
es
(F
N
).
O
ne
way
to
unde
rstan
d
thi
s
is
to
i
m
agine
if
al
l
edg
es
w
e
re
detect
ed
,
F
N
val
ue
m
us
t
be
zer
o.
Hen
c
e
the
recall
giv
es
th
e
rati
o
in
w
hich
how
m
any
TP
is
detect
ed
out
of
the
total
act
ual
cha
ng
e
po
i
nt
or
e
dg
e
.
It
is
cal
culat
ed
us
i
ng
th
e
fo
ll
owin
g
e
qua
ti
on
[24]
,
(2)
No
ti
ce
t
hat
pre
ci
sion
a
nd
reca
ll
hav
e
dif
fer
e
nt
at
trib
utes.
F
β
sco
re
is
giv
e
n
by
the
f
ollo
wing
e
qu
at
i
on
[24]
,
with
β
is
1
for
a
ha
rm
onic
m
ean
betwe
en
recall
an
d
preci
sion
[
27]
.
Hen
ce
,
both
re
cal
l
and
preci
sion
are
giv
e
n
the
sam
e
w
ei
ghta
ge
.
(3)
4.
RESU
LT
S
AND DI
SCUS
S
ION
In
t
his
sect
io
n,
t
he
data
is
pr
e
sente
d
a
nd
a
naly
zed
ac
cordin
g
to
the
fou
r
-
ste
p
im
plem
entat
ion
discusse
d i
n
t
he
previ
ou
s
sect
ion
.
4.1.
E
vent de
tection
on
aggr
e
g
ated
elec
tri
cal curren
t da
ta
Figure
2
s
how
s
the
Day
-
1
da
ta
for
al
l
ag
gr
e
gated
c
urre
nt
data.
No
ti
ce
al
l
the
posit
ive
a
nd
ne
gative
current
c
ha
nge
s
.
All
t
hese
c
hanges
w
ould
be
s
ubj
ect
e
d
t
o
se
ver
al
proc
esses
to
qual
if
y
them
as
even
ts
as
discusse
d
in
the
m
e
tho
dolo
gy
sect
ion
.
In
a
dd
it
io
n
to
that,
the
processes
wer
e
done
f
or
diff
e
re
nt
sam
pl
ing
fr
e
qu
e
ncies.
T
he per
form
ances of th
e
pr
oce
sses are
assess
ed usin
g
t
he
se
le
ct
ed
pe
rfor
m
ance m
at
rices.
The
s
ubseq
ue
nt
Figure
3
(a)
is
the
der
i
vative
of
the
ag
gregate
d
c
urrent
m
easur
em
ent
us
in
g
a
5
-
m
inu
te
sam
pli
ng
fr
e
que
ncy.
Be
cause
the
ti
tl
e
of
this
paper
is
to
ca
p
th
e
m
axi
m
u
m
de
m
and
on
ly
posit
ive
der
i
vation
was
go
i
ng
t
o
be
i
nspect
ed
.
N
otice
that
the
high
est
abso
l
ute
de
rivati
on
is
ab
out
22.
26
A
over
a
5
-
m
in p
eri
od. I
t
is ab
out 1
0% o
f
the m
axi
m
u
m
MSB’s c
urren
t
of
229.8
7 A.
Ba
sed
on
t
he
s
it
e
conditi
on
,
f
il
te
rs
wer
e
sel
ect
e
d
so
t
hat
de
rivati
ves
of
only
la
rg
e
r
tha
n
5%,
10%.
15%
a
nd
20%
wer
e
al
lowe
d
t
o
pass.
T
he
res
ult
is
s
hown
in
Fig
ur
e
3
(
b)
is
due
t
o
the
10
%
filt
er.
The
fi
lt
ered
data
wer
e
the
n
subj
e
ct
ed
to
an
edg
e
detect
ion
schem
e.
At
this
sta
ge,
on
ly
the
inc
rem
ental
edg
es
wer
e
detect
ed.
Fig
ure
3
(c)
is
the
outc
om
e
fr
om
the
e
dg
e
detect
ion
of
Fi
gure
3
(b).
N
ote
that
t
he
horizo
ntal
a
xis
is
the
point
of
ti
m
e.
Since
t
he
sam
pling
tim
e
is
5
-
m
inu
te
,
t
he
total
sp
a
n
wi
ll
be
tra
ns
la
te
d
to
0
to
575
m
inu
te
s.
F
o
r
F
i
g
u
r
e
3
(
a
)
,
t
h
e
v
e
r
t
i
c
a
l
a
x
i
s
i
s
t
h
e
d
e
r
i
va
t
i
o
n
w
i
t
h
t
h
e
u
n
i
t
o
f
A
/
5
-
m
i
n
u
t
e
.
O
n
t
h
e
o
t
h
e
r
h
a
n
d
,
t
h
e
v
e
r
t
i
c
al
a
x
i
s
f
o
r
F
i
g
u
r
e
3
(
b
)
a
n
d
(
c
)
i
s
t
h
e
o
u
t
c
om
e
o
f
t
h
e
e
d
g
e
d
e
t
e
c
t
i
o
n
p
r
o
c
e
s
s
.
I
t
i
s
a
b
i
n
a
r
y
o
u
t
c
om
e
o
f
e
i
t
h
e
r
1
o
r
0
.
The o
utput o
f Fi
gure
3
(b)
ca
n be c
onsecuti
vely
m
ad
e o
f m
or
e than
one
po
i
nt.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2021
:
18
14
-
18
24
1818
Figure
2.
A
ggr
egated c
urre
nt
m
easur
em
ent
Figure
3.
(a
) d
erivati
on
of
M
SB’s c
urre
nt wavef
or
m
at
5
-
m
inu
te
sam
pli
ng freq
ue
ncy;
(
b) outp
ut
of
10
% f
il
te
r
a
pp
li
ed
to
t
he deri
vation res
ul
t;
(c)
the
edge
detect
ion o
utput o
f
fig
ur
e
(
b)
If
the
10
-
m
inu
te
sam
pling
fr
e
qu
e
ncy
is
sel
ect
ed
and
goin
g
thr
ough
al
l
the
pr
oce
s
ses
ex
plained
pr
e
viously
,
the
ou
tc
om
e
is
as
sh
own
in
Fi
gure
4.
From
th
e
figure,
it
can
be
seen
that
the
inf
or
m
at
ion
about
the
a
ggreg
at
e
d
data
der
i
vatio
n
has
now
be
c
om
e
le
ss.
As
the
resu
lt
,
by
a
pp
ly
in
g
t
he
sa
m
e
10
%
filt
er,
on
ly
tw
o
broa
d
passe
s
are
re
gistered
.
Con
se
quently
,
on
ly
tw
o
ed
ge
s
are
detect
ed
as
exh
i
bited
in
Figure
4
(
c).
This
certai
nly d
oes no
t
ref
le
ct
t
he nu
m
ber
of sig
ni
ficant ed
ges
th
at
can be
visu
a
ll
y sc
ann
e
d
f
r
om
Figu
re
2.
W
it
h
su
c
h
ou
t
com
es,
it
is
safe
to
co
nclu
de
that
analy
sis
in
vo
l
ving
data
s
a
m
pled
at
every
15
-
m
inu
te
and
30
-
m
inu
te
will
return
a
w
or
se
res
ult.
T
hi
s
is
la
rgel
y
bas
ed
on
the
fact
t
hat
the
de
rivati
on
f
or
data
sa
m
pled
at
bo
t
h
fr
e
qu
e
nc
ie
s
wil
l
yi
el
d
a
sm
oo
ther
outc
om
e.
Fo
r
exa
m
ple,
even
if
the
filt
er
is
re
duced
to
5%
f
or
a
10
-
m
inu
te
sam
pli
ng
f
re
qu
e
ncy,
the
ou
tc
om
e
of
the
w
ho
le
pro
cess
will
giv
e
a
res
ult
as
sho
wn
in
Fig
ure
5.
Data
sam
pled
at
60
-
m
inu
te
is
no
t
relevan
t
as
it
exceeded
the
m
ini
m
u
m
30
-
m
inu
te
sa
m
pling
fr
e
qu
e
ncy
as
pr
ac
ti
sed
by the
util
it
y com
pan
y.
Figure
6
is
the
at
tem
pt
to
visu
al
iz
e
the
edge
detect
ion
res
ults
with
res
pe
ct
to
the
or
igi
nal
MSB
’
s
current.
Fig
ure
6
(a)
-
(
d)
is
t
he
ed
ge
detect
io
n
res
ult
on
filt
er
5%,
10%
15%
,
an
d
20%
res
pe
ct
ively
.
Visu
a
ll
y
i
t
can
be
co
nclu
ded
t
hat
the
s
m
al
le
r
the
filt
e
r,
the
m
or
e
sig
nals
we
re
pass
ed
f
or
t
he
e
dge
detect
ion
pro
cess
i
s
sh
ow
n
in
Fig
ure
6.
Howe
ver
,
no
accu
rate
pe
rfor
m
ance
of
the
filt
erin
g
a
nd
e
dg
e
detect
ion
proce
ss
c
ould
be
con
cl
ud
e
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Develo
pm
e
nt
of
sit
e
-
sp
eci
fi
c non
-
intr
us
iv
e lo
ad mo
nitorin
g f
or
maxi
mum
…
(
Azha
r
udin
Muk
hta
r
uddin
)
1819
Figure
4.
(a
) D
erivati
on
of
M
SB’s c
urre
nt wavef
or
m
at
1
0
-
m
inu
te
sam
pli
ng freq
ue
ncy;
(
b) outp
ut
of
10
%
filt
er a
ppli
ed
t
o
the
d
e
rivati
on
resu
lt
;
(c
)
t
he
edge
detect
ion o
utput o
f
fig
ure
(
b)
Figure
5.
(a
) D
erivati
on
of
M
SB’s c
urre
nt wavef
or
m
at
1
0
-
m
inu
te
sam
pli
ng freq
ue
ncy;
(
b) o
utput
of
5%
f
il
te
r
a
pp
li
ed
to
t
he deri
vation res
ul
t;
(c)
t
he
edge
detect
ion o
utput o
f fig
ur
e
(b)
Data
in
Table
1
on
the
oth
e
r
hand
pro
vi
de
the
achievem
en
t
of
perform
ance.
No
te
that
f
or
filt
ers
at
10%
an
d
15%,
the
recall
, p
re
c
isi
o
n
a
nd
F
1
sc
or
e all
achieve
the
m
ark
of 1
. Th
is
m
eans
al
l
po
s
sible
si
gn
ifi
cant
changes
f
or
t
he
10%
a
nd
15
%
filt
ers
we
re
correct
ly
detect
ed
by
th
e
ed
ge
detect
ion
pro
cess.
T
her
e
for
e,
bot
h
filt
ers
are
the
best
in
t
his
ev
ent
detect
io
n
s
ta
ge.
Nev
e
rthe
le
ss,
f
or
t
he
5%
filt
er,
the
r
ecal
l
scor
e
is
0.562
5
po
i
nts,
wh
il
e
t
he
pr
eci
sio
n
is
good
at
0.818
2
po
i
nts.
F
or
the
20
%
filt
er,
the
preci
sio
n
i
s
re
gistered
on
ly
at
0.478
3
points
,
even
t
hough
t
he
recall
m
at
rix
has
a
perfect
1
point.
He
nce
,
al
though
t
he
20%
filt
er
is
c
ap
able
of co
rr
ect
ly
ide
ntifie
d
T
P
fro
m
a
ll
ev
ents, it
is far
fro
m
captu
ri
ng all
av
ai
la
ble ch
a
nge
po
i
nts.
Table
1
. Per
for
m
ance index
of e
dg
e
d
et
ect
io
n on MSB
’s
curre
nt w
a
ve
for
m
Filter
5%
10%
15%
20%
Recall
0
.56
2
5
1
.00
0
0
1
.00
0
0
1
.00
0
0
Precisio
n
0
.81
8
2
1
.00
0
0
1
.00
0
0
0
.47
8
3
F1
sco
re
0
.66
6
7
1
.00
0
0
1
.00
0
0
0
.64
7
1
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2021
:
18
14
-
18
24
1820
Figure
6.
Vis
ua
l i
nterpreta
ti
on
of ed
ge dete
c
ti
on
ou
tc
om
e as im
po
sed on t
he o
ri
gin
al
a
ggreg
at
e
d
c
urren
t
d
at
a
4.2.
Fe
ature
ext
r
act
i
on
and lo
ad
ide
nt
ifi
cat
i
on
The
in
div
id
ual
sel
ect
ed
el
ect
rical
load
data
is
as
sh
own
in
Figure
7.
No
ti
ce
that
the
WCPU
loa
d
is
on
ly
operati
on
al
within
the
office
hour
with
a
tem
po
rar
y
sh
ut
dow
n
of
a
bout
an
hour
at
noon.
A
ppr
opr
ia
te
ly
,
the
CT
a
nd
C
D
W1
a
nd
C
D
W2
are
al
so
e
xp
ect
e
d
t
o
be
op
e
rati
onal
in
the
sam
e
p
erio
d.
H
ow
e
ve
r,
t
he
act
ual
op
e
rati
ng
tim
e
for
both
is
30
-
m
inu
te
befor
e
and
a
fter
the
norm
al
wo
r
king
hour.
The
wa
y
W
CP
U
ope
r
at
ing
is
to
r
un
at
a
base
cu
rr
e
nt
dem
and
betwee
n
20.
1
A
t
o
21
A.
It
f
luctuat
es
in
the
ra
ng
e
of
33.
4
A
to
35.
9
A.
T
hese
fluctuati
ons
a
r
e
due
to
the
cu
t
-
in
a
nd
cut
-
off
operati
on
of
a
n
a
dd
it
io
nal
c
om
pr
essor
t
o
m
ai
ntain
the
r
oom
’s
set
tem
per
at
ur
e
.
The
CT
a
nd
CD
W
s
op
e
rati
on
are
cl
os
el
y
rel
at
ed
to
each
oth
er
.
F
ro
m
Fig
ure
9,
it
ca
n
be
s
een
t
hat
the
sta
rting
a
nd
s
hu
tt
in
g
do
wn
ti
m
e
of
CT
an
d
CD
Ws
are
i
den
ti
cal
.
H
ow
ever,
due
to
the
al
te
rn
at
in
g
po
li
cy
betwee
n
CD
W
1
an
d
C
D
W2
operati
on
was
in
p
la
ce, CD
W1
would o
per
at
in
g
in the fi
rst
ha
lf o
f
t
he
office
hour
wh
il
e
CD
W2
f
or
the
ot
her
ha
lf.
It
is
i
m
po
rtant
to
rem
ark
that
the
fluctuati
on
s
in
CT
an
d
CD
W
s
dat
a
set
s
are
fou
nd
t
o
be
lo
ose
ly
cor
relat
e
d.
Stat
ist
ic
al
l
y,
t
he
a
naly
sis
sho
wed
that
the
R
2
value
betwee
n
bo
t
h
data
is s
m
al
l.
On
to
p
of
that,
the
sta
nd
ar
d
de
viati
on
f
or
C
T
is
0.
2
A,
f
or
CD
W
1
is
0.5
6
A,
an
d
f
or
C
D
W2
is
0.31
A.
It
is
al
so
can
be
not
ic
ed
that
t
he
op
erati
onal
tim
e
is
relat
ed
to
WCPU.
This
is
exp
ect
e
d
as
t
he
functi
on
of
CT
an
d
CD
W
s
are
to
c
oo
l
do
wn
the
c
om
pr
esso
rs
in
the
W
CPU
.
As
with
the
CT
-
CD
W
s
fluctuati
on
s
relat
ion
s
hi
p,
the
correla
ti
on
s
b
e
tween a
re
ver
y
low.
Figure
7.
Cu
rr
e
nt cur
ves for
WCPU, CT
, C
D
W1
a
nd C
D
W2
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Develo
pm
e
nt
of
sit
e
-
sp
eci
fi
c non
-
intr
us
iv
e lo
ad mo
nitorin
g f
or
maxi
mum
…
(
Azha
r
udin
Muk
hta
r
uddin
)
1821
Starti
ng
of
CT
and
CD
W1
w
as
reg
ist
ere
d
a
s
sign
ific
a
nt
de
rivati
ve
val
ue
and
s
ubseq
ue
ntly
detect
ed
as
incr
em
ental
ed
ge
but
bey
ond
t
he
w
orking
ho
ur.
E
ve
n
f
or
CD
W2,
th
e
off
posit
ion
at
t
=
31
5
-
m
inu
te
ha
s
on
ly
co
ntri
bu
t
ed
to
the
ne
ga
ti
ve
der
i
vation
value
in
t
he
total
MSB
curr
ent.
All
CT
an
d
CD
Ws
hav
e
bee
n
cl
assifi
ed
as
t
ype
I
I
el
ect
rical
m
achines.
H
oweve
r
,
si
nce
al
l
CT,
CD
W1
a
nd
C
D
W2
ca
nnot
be
shut
do
wn
as
long as
an
y
W
CPU is
operati
on
al
,
m
ay
cause
them
to
al
so
be
cl
assifi
e
d
as
t
ype I
V.
N
otice
that
the
CD
W2
sta
rte
d
at
t
=
315
-
m
inu
te
.
Howe
ve
r,
this
c
hange
was
not
picke
d
up
by
the
edg
e
detect
io
n
because
the
de
rivati
on
val
ue
on
t
he
total
M
SB
cu
rr
e
nt
is
ne
gative.
The
s
udde
n
inc
rem
e
nt
will
pro
bab
ly
ne
ve
r
be
re
gistered
as
a
po
sit
ive
edg
e
on
the
a
ggre
gated
cu
rrent
wav
e
f
or
m
.
This
is
because
the
po
li
cy
says
m
os
t W
CP
Us
will
b
e sw
it
ched
off
durin
g
the br
eak hour
sta
rting
at t = 31
5
-
m
inu
te
. H
ence
CD
W2
would ne
ve
r b
e i
m
po
rta
nt to be
posit
ively
iden
ti
fie
d.
In
te
re
sti
ng
ly
,
t
he
base
c
urre
nt
dem
and
is
abo
ut
the
sam
e
as
the
highest
de
rivati
on
m
agn
it
ud
e
of
the
aggre
gated
c
urren
t
data.
At
this
sta
ge
,
it
is
log
ic
al
to
a
ssum
e
that
the
fl
uc
tuati
on
s
at
th
e
ag
gr
e
gated
c
urren
t
curve
ca
n
be
tr
aced
it
s
ori
gin
to
W
CP
U’s
de
m
and
.
In
ot
her
wor
ds
,
t
he
fea
tures
of
WCP
U’
s
c
urr
e
nt
cu
r
ve
ca
n
be reveale
d fro
m
the ev
e
nt
detect
ion
done a
t t
he
MSB
’
s c
urr
ent w
a
ve
form
.
F
i
g
u
r
e
8
s
h
o
w
s
,
(
a
)
t
h
e
d
e
r
i
v
a
t
i
o
n
o
f
W
C
P
U
’
s
c
u
r
r
e
n
t
c
u
r
v
e
,
(
b
)
t
h
e
o
u
t
p
u
t
o
f
1
0
%
f
i
l
t
e
r
,
a
n
d
(
c
)
t
h
e
e
d
g
e
d
e
t
e
c
t
i
o
n
o
f
(
b
)
.
A
l
l
p
r
o
c
e
s
s
e
s
w
e
r
e
d
o
n
e
o
n
t
h
e
5
-
m
i
nu
t
e
s
a
m
p
l
i
n
g
d
a
t
a
.
A
g
a
i
n
,
t
h
e
v
e
r
t
i
c
a
l
a
x
i
s
f
o
r
F
i
g
u
r
e
8
(
a
)
i
s
t
h
e
d
e
r
i
v
a
t
i
o
n
v
a
l
u
e
,
A
/
5
-
m
i
n
u
t
e
.
A
l
l
h
o
r
i
z
o
n
t
a
l
a
xe
s
a
r
e
i
n
t
i
m
e
(
m
i
n
u
t
e
)
.
N
o
t
i
c
e
t
h
a
t
t
h
e
d
e
r
i
va
t
i
o
n
o
u
t
c
om
e
s
a
r
e
a
l
m
o
s
t
p
u
l
s
e
-
s
h
a
p
e
d
.
H
e
n
c
e
,
a
l
l
t
h
e
o
u
t
p
u
t
s
o
f
t
h
e
f
i
l
t
e
r
a
r
e
p
u
l
s
e
s
.
T
h
e
e
d
g
e
d
e
t
e
c
t
i
o
n
i
s
n
a
t
u
r
a
l
l
y
p
u
l
s
e
-
t
y
p
e
o
u
t
p
u
t
.
T
h
e
r
e
f
o
r
e
,
t
h
e
o
u
t
c
om
e
s
o
f
e
d
g
e
d
e
t
e
c
t
i
o
n
a
r
e
i
d
e
n
t
i
c
a
l
t
o
t
h
o
s
e
f
o
u
n
d
i
n
F
i
g
u
r
e
8
(b).
The
e
dg
e
detec
ti
on
pr
ocess
on
the
W
CP
U’
s
c
urren
t
data
yi
el
ded
t
he
outc
om
es
as
sh
own
in
Fig
ur
e
9.
Visu
al
ly
it
ca
n
be
ge
ner
al
ly
de
du
ced
that
ou
tc
om
es
of
e
dg
e
detect
ion
du
e
to
5%
,
10
%
an
d
15%
fi
lt
ers,
Figure
9
(a
)
-
(c
)
res
pecti
vely
,
are
qu
it
e
sim
ilar.
T
he
w
or
se
ou
tc
om
e
is
cl
e
arly
can
be
see
n
in
Fig
ure
9
(
d).
It
i
s
the r
es
ult o
f
e
dge
detect
ion f
or a
20%
filt
er. On
ly
tw
o
e
dge
s of the
curr
e
nt
curve
can
b
e
det
ect
ed.
Figure
8.
(a
) D
erivati
on
of
W
CPU’
s
curr
ent
wav
e
f
or
m
at
the 5
-
m
inu
te
sam
pl
ing
fr
e
quen
cy
;
(b) ou
t
pu
t
of
10%
filt
er a
ppli
ed
to
the
de
riv
at
ion
resu
lt
;
(c
)
the e
dg
e
detec
ti
on
ou
t
pu
t
of fi
gure (b
)
As
ex
pected
,
r
ecal
l
and
preci
sion
m
at
rices
fo
r
5%,
10%
and
15%
filt
ers
al
l
retur
n
1.0
000
points
as
sh
ow
n
in
Ta
bl
e
2.
It
m
eans
the
proces
s
is
able
to
capt
ure
al
l
edg
es
a
nd
perfect
ly
di
sti
ng
uis
hed
be
tween
po
sit
ive
ed
ges
an
d
non
-
e
dg
e
s.
T
he
f
ollow
i
ng
ta
ble
is
the
ta
bula
ti
on
of
the
re
su
lt
s.
F
or
the
20
%
filt
e
r,
it
is
capab
le
to
cap
ture
al
l
edg
es
it
m
anag
ed
to
detect
,
wh
ic
h
are
on
ly
two.
This
is
ref
le
ct
ed
by
the
ver
y
poor
perform
ance in
preci
sion.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
3
,
Se
ptem
ber
2021
:
18
14
-
18
24
1822
Table
2.
Per
for
m
ance index
of e
dg
e
d
et
ect
io
n on
W
C
PU’s
current
wa
vefo
rm
Filter
5%
10%
15%
20%
Recall
1
.00
0
0
1
.00
0
0
1
.00
0
0
1
.00
0
0
Precisio
n
1
.00
0
0
1
.00
0
0
1
.00
0
0
0
.10
0
0
F1
sco
re
1
.00
0
0
1
.00
0
0
1
.00
0
0
0
.18
1
8
Figure
9. Cu
rr
e
nt cur
ves for
WCPU, CT
, C
D
W1
a
nd C
D
W2
W
it
h
a
co
rr
ec
t
sel
ect
ion
of
filt
er
sel
ect
ion
,
c
hanges
in
W
CPU
’s
op
e
rati
on
ca
n
be
confide
ntly
detect
ed.
If
t
hose
cha
ng
es
are
sign
ific
a
ntly
ref
le
ct
ed
in
t
he
a
ggre
gated
c
urr
ent
data,
it
ca
n
be
per
cei
ve
d
a
s
the
featur
e
ex
t
racted
from
the equipm
ent. In
the
case
of
t
his st
ud
y,
the e
quip
m
ent it
sel
f
can
be direct
ly
ide
ntifie
d.
Table
3
is
t
he
perform
ance
in
dex
e
s
of
the
c
ha
ng
e
poi
nt
in
MSB
’s
c
urren
t
cu
rv
e
with
a
c
hange
poin
t
in
W
CP
U
’s
c
urre
nt
cu
rve
as
it
s
pr
e
dictor
.
T
he
best
pe
rfo
r
m
ance
is
w
hen
the
10%
filt
er
was
us
e
d
to
a
naly
ze
MSB
’s data
aga
inst the
outc
om
e o
f
e
dge
detect
ion
on
W
C
P
U’
s
d
at
a
us
i
ng
5%, 1
0%
a
nd
15%
filt
ers.
Table
3.
Per
for
m
ance index
bet
ween
e
dges i
n
a
ggre
gated
c
urren
t
data a
nd
edges i
n W
C
P
U
c
urren
t
data
Filter
5%
10%
15%
20%
Recall
0
.66
6
7
0
.73
3
3
0
.33
3
3
0
.26
6
7
Precisio
n
0
.40
0
0
0
.55
0
0
0
.38
4
6
0
.36
3
6
F1
sco
re
0
.50
0
0
0
.62
8
6
0
.35
7
1
0
.30
7
7
It
can
be
see
n
that
the
resu
lt
in
Table
3
an
d
in
[12]
are
co
m
par
able,
look
ing
at
the
ave
r
age
val
ues.
The
res
ult
fr
om
[1
2],
wh
ic
h
was
based
on
the
do
m
est
ic
data,
is
as
sh
own
in
Ta
ble
4.
Fu
rthe
rm
or
e,
it
was
repor
te
d
in
the
sa
m
e
pap
er
th
at
ano
the
r
study
by
Sale
rn
o
(
2018)
yi
el
ded
an
ave
rag
e
F
1
scor
e
of
0.6
7.
Table
5
is t
he
re
su
lt
as
repor
te
d
in
[2
8]
. Co
m
par
in
g t
he result
from
t
his stu
dy a
nd [28
]
als
o
s
hows
a go
od agreem
ent.
Table
4.
Per
for
m
ance index
[12
]
Ap
p
aratus
Kettle
W
ash
in
g
m
achi
n
e
Dish
wash
er
Fridg
e
Av
erage
Recall
0
.47
0
.40
0
.90
0
.96
0
.68
Precisio
n
0
.58
0
.43
0
.73
0
.97
0
.68
F1
sco
re
0
.52
0
.41
0
.81
0
.97
0
.68
Table
5.
NILM
p
e
rfor
m
ance [28
]
Cu
rr
en
t valu
e
5
0
Hz
5
0
and
150 Hz
5
0
,
1
5
0
,
2
5
0
Hz
Recall
0
.53
6
0
.63
5
0
.62
4
Precisio
n
0
.52
8
0
.65
1
0
.63
9
F1
sco
re
0
.56
6
0
.65
9
0
.64
3
0
200
400
(
a
)
0
10
20
30
40
0
0
.
5
1
0
200
400
(
b
)
0
10
20
30
40
0
0
.
5
1
0
200
400
(
c)
0
10
20
30
40
0
0
.
5
1
0
200
400
(
d
)
0
10
20
30
40
0
0
.
5
1
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Develo
pm
e
nt
of
sit
e
-
sp
eci
fi
c non
-
intr
us
iv
e lo
ad mo
nitorin
g f
or
maxi
mum
…
(
Azha
r
udin
Muk
hta
r
uddin
)
1823
5.
CONCL
US
I
O
N
Cl
early
that
us
ing
the
c
omm
on
te
chn
ique
s
propose
d
in
this
pap
e
r,
sig
nificant
ene
r
gy
-
co
ns
um
ing
equ
i
pm
ent
can
be
rev
eal
e
d
j
ust
by
a
naly
sing
t
he
a
ggre
gated
c
urre
nt
dem
and
.
T
he
perform
ance
of
the
te
chn
iq
ue
is
co
m
par
able
to
oth
er
researc
h.
T
his
te
chn
i
que
is
sit
e
-
sp
eci
fic
and
it
is
justi
fiable
as
dif
fer
e
nt
sit
e
s
m
ay
hav
e
different
el
ect
rical
load
set
ti
ng
s,
diff
e
ren
t
ene
r
gy
m
anag
em
e
nt
po
li
ci
es
an
d
eve
n
ha
ve
s
pecific
lim
it
at
ion
s,
not
un
li
ke
the
sit
e
un
de
rtake
n
unde
r
this
stu
dy.
Furthe
r
stud
y
on
a
m
or
e
com
plex
sys
te
m
is
need
e
d
in
orde
r
to
ve
rify
the
m
et
ho
d
us
ed
i
n
this
pa
per.
Of
c
ourse,
th
e
disag
gr
e
gatio
n
proce
ss
nee
ds
to
be
include
d
t
oo.
A
lo
nger
per
i
od
of
st
ud
y
is
al
so
pr
e
fer
a
ble
t
o
dev
el
op
a
m
or
e
r
obus
t
m
eth
od.
Ne
ver
t
heless,
the
m
et
ho
d
is
base
d
on
a
sim
ple
m
at
he
m
a
ti
ca
l
pr
i
nciple.
It
is
then
ap
plied
t
o
a
ver
y
s
pecifi
c
sit
e
co
ndit
ion
.
Th
e
la
cking
of
c
om
plexit
y
in
th
is
te
chn
iq
ue
it
sel
f
can
at
tract
the
op
erat
or
s
of
the
ene
rg
y
eff
ic
ie
ncy
industry.
Fu
rt
her
m
or
e,
a
s
the
act
ual
pe
rsons
i
n
handli
ng
the
m
at
te
r,
the
ope
rators
a
re
al
l
well
-
ver
s
ed
with
thei
r
s
yst
e
m
and
know
the
l
i
m
i
ta
ti
on
of
t
he
syst
e
m
in
or
de
r
to
fi
ne
-
t
un
e
it
s
reli
abili
ty
a
nd
rob
us
tne
ss.
Pr
act
ic
al
ly
sp
eakin
g,
for
this
sit
e,
it
is
possible
to
e
xp
la
in
the
i
ncrea
se
in
the
gl
obal
el
ect
rical
da
ta
is
du
e
to
W
CPU
at
a
10%
filt
er.
Fr
om
this
po
ss
ibil
it
y,
the
op
erator
ca
n
m
ake
an
obj
ect
ive
de
ci
sion
to
al
te
r
the
W
C
PU’s
op
e
rati
on
in
order
to
achieve
desi
ra
ble m
axi
m
u
m
dem
and
contr
ol
.
ACKN
OWLE
DGE
MENTS
This
stu
dy
was
m
ade
po
ssi
ble
with
the
F
un
dam
ental
Re
search
Grant
Schem
e
(F
RG
S/1/2
020/
TK0
/
UPNM/
03/3
)
m
ade
av
ai
la
ble
by
the
Mi
nistry
of
Higher
Ed
ucat
ion
,
Ma
la
ysi
a.
Th
e
auth
or
s
would
li
ke
to
exp
r
ess
their
appr
eci
at
ion
to
D
evelo
pm
ent
Dep
artm
ent,
Eng
inee
rin
g
Ca
m
pu
s,
Un
i
ver
sit
i
Sain
s
Ma
la
ysi
a,
Penang,
Ma
la
ysi
a,
for
pe
rm
issio
n
for
acce
ssi
ng
t
heir
syst
em
and
to
us
e
t
he
data
for
this
stu
dy
.
REFERE
NCE
S
[1]
W
orld
Bank,
“
Acc
ess
to
el
ectri
ci
t
y
(%
of
the
popula
tion).”
htt
ps://
d
at
a
.
worl
dbank.
org/
indi
c
a
tor/
EG.ELC.
AC
CS
.
ZS
?
end=
201
8&sta
rt=1990&
vie
w=c
h
art
(
acce
ss
ed
Ma
y
0
1,
2020).
[2]
Malay
s
ia
En
erg
y
Com
m
ission,
Peninsula
r
Malay
sia
El
e
ct
ri
ci
t
y
Suppl
y
Industr
y
Outlook,
2019.
Acc
essed:
Ma
y
2020.
[Online
]
.
Avail
ab
le
:
htt
ps://
ww
w.st.
g
ov.
m
y
/
en/
con
te
n
ts/fi
le
s/downlo
a
d/106/
P
eni
nsular
_Malay
s
ia
_
El
e
ctrici
t
y
_
Suppl
y
_
I
n
dustr
y
_Outlook
_
2019_compress
ed.
pdf
[3]
Malay
s
ia
Ene
rg
y
Com
m
ission,
Ene
rg
y
Ma
lays
ia
:
Streng
the
ni
ng
the
Futur
e
of
Ene
rg
y
in
Malay
s
ia.
Acc
e
ss
d:
Ma
y
01,
2020
.
[Online
]
.
Avail
able:
ht
tps:/
/w
ww
.
st.gov
.
m
y
/
e
n/c
ont
ent
s/fi
le
s/
download/
112/Energ
y
_M
al
a
y
si
a,_
Volum
e_14,
_20185.
pdf
[4]
Ene
rg
y
Marke
t
Auth
ority
of
Singapore
,
“
Dem
and
Side
Mana
gement,”
Ove
rvie
w
of
E
le
c
tric
i
ty
Mark
et
.
htt
ps://
ww
w.e
m
a.
gov.
sg/Dem
an
d_Side_Mana
ge
m
ent
.
aspx
(
ac
c
e
ss
ed
Ma
y
01,
20
20).
[5]
E.
Sarke
r
et
al
.
,
“
Progress
on
the
Dem
and
Side
Mana
gement
in
Sm
art
Grid
and
Optimizati
on
Approac
hes
,”
Inte
rnational
Jo
urnal
of Ene
rgy
Re
search
,
vol
.
4
5,
no
.
1
,
pp
.
36
-
64,
Jan
.
2021
,
d
oi
:
10
.
1002/e
r
.
5
631.
[6]
O.
Ham
id,
M.
Barba
rosou,
P.
Papage
orga
s,
K.
Preka
s,
and
C
.
T.
Sal
ame,
“
Autom
at
ic
R
ec
og
nit
ion
o
f
El
e
ct
r
i
ca
l
Loa
ds
Anal
y
zi
n
g
t
he
Chara
c
te
r
isti
cs
Para
m
et
er
s
o
f
t
he
Consu
m
ed
El
ec
tr
ic
Pow
er
Through
a
Non
-
Intrusive
Methodol
og
y
,
”
i
n
Ene
rgy
Proc
ed
ia
,
2017
,
pp
.
742
-
752
,
doi
:
10
.
10
16/j
.
eg
y
pro
.
201
7.
07.
137
.
[7]
Abubaka
r,
S.
N
.
Khali
d
,
M.
W
.
Mus
ta
fa
,
H.
S
har
ee
f
,
and
M.
Mus
ta
pha,
“
Applicati
on
of
load
m
onit
oring
in
appl
i
anc
es
ene
r
g
y
m
ana
gement
-
A
rev
ie
w,”
R
ene
wabl
e
and
Sustainabl
e
Ene
r
gy
Re
v
ie
ws
,
vol.
67,
pp.
235
-
245,
2017,
doi
:
10
.
10
16/j
.
rser
.
2016.
0
9
.
064.
[8]
M.
Ham
di,
H.
Mess
aoud,
and
N.
Bouguil
a
,
“
A
New
Approac
h
o
f
El
e
ct
ri
ca
l
Ap
pli
an
ce
Ide
n
ti
fi
c
at
ion
i
n
R
eside
n
ti
al
Buil
ding
s,
”
E
lectric
Powe
r
S
yste
ms
Re
searc
h
,
vol
.
178
,
no.
Sept
ember
2019,
p
.
10
6037,
2019,
d
oi
:
10.
1016/j.e
psr.2
019.
106037.
[9]
C.
La
m
,
Z.
Z
h
e
ng,
and
X.
Luo,
“
El
ec
tricity
Lo
ad
Dec
om
pos
it
i
on
Protot
y
pe
f
o
r
Hous
ehol
d
Ap
pli
an
ce
s:
S
y
ste
m
Design
a
nd
Deve
lopment
,
”
in
Ene
rg
y
Proce
d
ia
,
2019,
vol
.
158,
pp.
3158
-
3163.
doi:
10.
1016/j.e
g
y
pr
o.
2019.
01
.
1018.
[
1
0
]
N
I
L
M
d
a
t
a
s
e
t
s
.
A
c
c
e
s
s
e
d
J
a
n
.
3
0
,
2
0
2
1
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s
:
/
/
g
i
t
h
u
b
.
c
o
m
/
n
i
l
m
t
k
/
n
i
l
m
t
k
/
s
e
a
r
c
h
?q
=
N
I
L
M
+
d
a
t
a
s
e
t
s
[11]
Y.
Xiao,
Y.
Hu
,
H.
He,
D.
Zh
ou,
Y.
Zha
o
,
an
d
W
.
Hu,
“
Non
-
Intrusive
Lo
ad
Ide
nti
f
ic
a
ti
on
M
et
hod
Based
on
Im
prove
d
KM
Algorit
hm
,
”
IE
EE
Ac
c
ess
,
vol
.
7
,
p
p.
151368
-
1513
77,
2019
,
doi
:
10
.
1109/AC
CESS
.
2019.
2948079.
[12]
C.
Puent
e,
R
.
Pala
c
ios,
Y.
Gonzá
l
ez
-
Are
cha
v
a
la
,
and
E.
F.
S
ánc
he
z
-
Úbeda
,
“
Non
-
Intrusive
Loa
d
Monito
rin
g
(NILM)
for
Energ
y
Disaggr
egati
on
Us
ing
Soft
Com
puti
ng
Te
chn
ique
s,”
Ene
rgi
es
,
vol.
13,
no
.
12
,
p.
3117
,
2020
,
doi:
10
.
3390/e
n1
3123117.
[13]
S.
W
el
ika
la,
S.
Mem
ber
,
C.
Dinesh,
and
S.
Mem
ber
,
“
Inc
orpora
ti
ng
Appli
anc
e
Us
age
Patt
ern
s
for
Non
-
Intrusive
Loa
d
Monitori
n
g
and
Loa
d
Fore
ca
sting
,
”
IE
EE
T
rans
act
ions
on
Smar
t
Gr
id
,
vol.
10,
no.
1,
pp
.
44
8
-
461,
2019,
doi
:
10.
1109/T
SG
.
20
17.
2743760.
Evaluation Warning : The document was created with Spire.PDF for Python.