Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
4060
~
4078
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp4060
-
40
78
4060
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Forecast
ing
S
hort
-
term
W
holesa
le
P
rices on the
Irish S
ingle
Electri
city
Mark
et
Francesc
o Ar
ci
1
, Jan
e
Reil
ly
2
, Pe
ngfei
Li
3
,
Ke
vin
Cu
rr
an
4
,
A
mma
r Be
latreche
5
1,2,3
Ark E
ner
g
y
Consulti
ng
L
imite
d,
Unit
20
Da
i
ngea
n
Ha
ll,
N4
Axis
Cent
re
,
B
atter
y
Ro
ad, L
ong
ford,
Ir
la
ndi
a
4
Facul
t
y
of
Com
puti
ng,
Engi
n
ee
r
ing
&
Buil
t
Env
i
ronm
ent
,
Ulst
er University
,
Nort
her
n
Ir
el
and
,
U
n
it
ed
Kingdom
5
Depa
rtment of
Com
pute
r
Scie
n
ce
and
Dig
it
a
l T
ec
hnolog
ie
s,
Northum
bria
Unive
rsit
y
,
U
ni
te
d
Kin
gdom
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ma
y
2
9
, 201
8
Re
vised
A
ug
18
, 2
01
8
Accepte
d
Aug
30
, 201
8
El
e
ct
ri
ci
t
y
m
ark
et
s
are
d
iffe
r
ent
from
othe
r
m
ark
et
s
as
elec
tr
ic
i
t
y
gene
ra
ti
on
ca
nnot
be
ea
si
l
y
stored
in
subs
t
ant
i
al
amounts
and
to
avoi
d
bl
ac
kouts,
th
e
gene
ra
ti
on
of
elec
tr
ic
i
t
y
m
ust
be
bal
an
ce
d
wi
th
c
ustom
er
demand
for
it
on
a
sec
ond
-
by
-
sec
on
d
basis.
Cu
stom
ers
te
nd
to
rely
on
elec
tr
icit
y
for
da
y
-
to
-
d
a
y
li
ving
and
ca
n
not
rep
la
c
e
it
ea
sil
y
so
when
elec
tr
ic
i
t
y
pr
ices
inc
r
ea
se
,
customer
deman
d
gene
r
al
l
y
do
es
not
red
u
ce
sign
ifi
c
ant
l
y
in
th
e
short
-
te
rm
.
As
el
ec
tr
ic
i
t
y
g
ene
ra
ti
on
and
c
ustom
er
demand
m
ust
be
m
at
che
d
per
fe
c
t
l
y
sec
ond
-
by
-
sec
on
d,
and
be
ca
use
g
ene
ra
ti
on
c
annot
be
stored
to
a
c
onsidera
bl
e
ext
en
t,
cost
bi
ds
from
gene
rat
o
rs
m
ust
be
balanc
e
d
with
d
emand
e
stim
at
es
in
adva
nc
e
of
rea
l
-
ti
m
e.
Thi
s
pap
er
outl
ine
s
a
a
for
ec
ast
ing
al
gori
th
m
buil
t
on
art
if
ic
i
al
n
eur
a
l
net
works
to
pr
e
dic
t
short
-
te
rm
wholesa
le
prices
on
the
Irish
Single
E
lectr
i
cit
y
Marke
t
so
tha
t
m
ark
et
p
a
rti
ci
p
ant
s
ca
n
m
ake
m
ore
informed
tra
d
in
g
decisions.
Re
sea
rch
stud
ie
s
have
d
emons
tra
te
d
th
at
a
n
ada
pt
ive
or
self
-
ada
pt
ive
appr
oa
ch
to
fore
ca
stin
g
would
app
ea
r
m
ore
suited
to
the
t
ask
of
pre
dicting
en
erg
y
demands
in
te
r
rit
or
y
such
as
I
rel
and
.
W
e
have
ide
nt
ified
t
he
fe
at
ur
es
th
at
s
uch
a
m
odel
d
e
m
ands
and
outli
ne
i
t
h
ere
.
Ke
yw
or
d:
Ar
ti
fici
al
n
e
ur
a
l
n
et
w
orks
Ele
ct
rici
ty
m
ar
kets
Ma
chine
l
ea
rn
i
ng
Ma
rk
et
p
re
dicti
on
s
Neural
n
et
w
orks
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Kev
i
n
C
urran
,
Faculty
of Com
pu
ti
ng
, E
ngineeri
ng &
B
uilt
Environm
ent,
Ulste
r Un
i
ver
si
ty
, N
ort
he
rn Ir
el
and
,
Unit
ed
Kingdom
.
Em
a
il
:
kj
.curra
n@ulst
er.ac
.uk
1.
INTROD
U
CTION
The
inc
reasin
g
per
ce
ntage
of
el
ect
rici
ty
gen
erated
th
rou
gh
ren
e
wa
ble
sour
ces
te
nd
s
t
o
in
validat
e
th
e
assum
ption
of
correla
ti
on
bet
ween
el
ect
rici
ty
sp
ot
pr
ic
es
a
nd
t
he
pr
ic
e
of
the
m
ix
of
c
om
m
od
it
ie
s
util
i
zed
to
su
pply
ge
ne
rat
or
s
(e
.g.
gas,
coal,
oil
-
de
pe
ndin
g
on
the
ge
ner
at
in
g
a
sset
com
po
sit
ion
on
th
e
s
pecific
gr
i
d)
.
The
var
ia
ble
natu
re
of
pro
du
ct
io
n
of
re
ne
wab
le
e
ne
rg
y
sources
al
so
increases
t
he
vo
la
ti
li
ty
of
s
yst
e
m
m
arg
inal
pri
ces
(S
MPs
)
on
m
ark
et
s
base
d
on
a
m
and
at
or
y
cent
ral
pool
m
od
el
.
Europ
ea
n
c
ountrie
s
have
unde
rtake
n
s
ubsta
ntial
inv
e
stm
ents
to
boost
the
am
ou
nt
of
energy
pr
oduc
ed
th
rou
gh
re
new
a
ble
gen
e
r
at
ion
.
Ir
el
an
d
i
n
par
ti
cular
is
ai
m
ing
at
40%
of
it
s
powe
r
nee
ds
be
ing
m
et
by
re
new
a
ble
s
ourc
e
s
by
2020.
In
this
env
i
ronm
ent,
we
can
ex
pect
the
wholesal
e,
fine
gran
ularit
y
(e.g
.
hal
f
hour
ly
)
w
hole
sal
e
pr
ic
e
of
el
ect
rici
ty
to
beco
m
e
m
or
e
vo
la
ti
le
over
ti
m
e.
The
a
bili
ty
to
operate
e
ff
ec
ti
vely
on
el
ect
rici
ty
sp
ot
m
ark
et
s
reli
es
on
the
ca
pa
bili
ty
to
d
evis
e
appr
opriat
e
biddin
g
strat
egies.
These
in
tur
n
can
be
nef
it
fro
m
the
inclusion
of
a
reli
able forecast
of
s
hort
te
r
m
syst
e
m
m
arg
inal
pr
ic
es
(S
MP
s).
In
a
m
ark
et
with
a
n
i
ncr
ea
sing
pe
rcen
ta
ge
of
re
ne
wab
le
ge
ner
at
or
s
,
rel
ia
ble
forecast
s
m
us
t
necessa
rily
ta
ke
int
o
acc
ount
add
it
io
nal
fa
ct
or
s
s
uc
h
as
m
et
eor
ologica
l
forecast
s,
f
oreca
ste
d
dem
and
an
d
c
onstrai
nts
im
po
s
ed
by
netw
ork
topolo
gy
[
1
]
,
[
2].
Tra
diti
onal
tim
e
series
fo
r
ecast
ing
al
gorithm
s
(e.
g.
ba
sed
on
Au
t
oReg
ressi
ve
I
nteg
rated
Mov
i
ng
Av
e
ra
ge
m
od
el
s)
can
pe
rfor
m
reaso
na
bly
well
in
this
con
te
xt
bu
t
rel
y
on
assum
ption
s
bei
ng
m
ade
on
beh
a
vior
ov
er
dif
fer
e
nt
te
m
po
ral
windows
to
yi
el
d
co
ns
ist
ent
resu
lt
s
[3
]
,
[
4].
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Forec
as
ti
ng
Shor
t
-
te
r
m
W
ho
le
sa
le
Prices
on
the I
ris
h
...
(
Fran
ce
sco
Arci)
4061
Ther
e
is
a
s
m
al
l
nu
m
ber
of
com
pan
ie
s
pro
vid
in
g
or
w
orkin
g
on
a
ro
bust
appr
oach
t
o
forecast
in
g
bo
t
h
re
new
a
bl
e
power
ou
t
put
and
/
or
m
arg
inal
pri
ces
f
or
el
ect
rici
ty
.
Alba
S
olu
zi
oni
are
an
ind
e
pe
nd
e
nt
consulta
ncy
pr
ov
i
ding
inf
orm
at
ion
,
trai
ning
and
b
es
poke
c
on
s
ultancy
ser
vices
in
the
E
uro
pea
n
gas
a
nd
powe
r
m
ark
et
s
[
5]
. T
heir prim
ary pu
blica
ti
on to
da
te
is co
ns
ide
r
ed
a re
fer
e
nce
on
Ital
ia
n
gas & p
ow
e
r
m
ark
et
s.
They
currently
pr
ov
i
de
a
s
hort
-
te
rm
m
arg
inal
pr
ic
e
forecast
se
rv
ic
e.
MK
On
li
ne
pro
vid
es
onli
ne
m
ark
et
intel
li
gen
c
e
serv
ic
es
t
o
pro
vid
e
cl
ie
nts
with
tim
ely
and
hi
gh
-
res
olu
ti
on
forecast
s
of
f
undam
ental
s
an
d
pri
ces
f
or
t
he
sh
ort
,
m
id an
d l
ong t
erm
h
or
iz
on.
I
t
also
offers a c
om
ple
m
entary weathe
r
se
rv
ic
e
[
6]
.
Me
te
olo
gica
s
upply
forecast
s
of
busine
ss
var
ia
bles
relat
ed
to
weathe
r
throu
gh
the
pro
vision
of
integrate
d
f
or
e
cast
ing
s
olu
ti
ons,
uniq
ue
to
each
cl
ie
nt.
M
e
te
olo
gica
s
pe
ci
al
iz
es
in
wind
a
nd
s
olar
pow
e
r
forecast
in
g
ser
vices
al
l
aro
un
d
the
w
or
ld
[7]
.
Their
f
or
eca
sts
are
util
iz
ed
by
hu
ndre
ds
of
plant
owne
rs,
powe
r
traders
an
d
gr
i
d
operat
or
s
to
op
ti
m
iz
e
their
bu
si
ness
act
ivi
ti
es.
Fr
om
pr
el
i
m
inary
con
ve
r
sat
ion
s
as
a
pote
ntial
su
ppli
er
of
re
ne
wab
le
po
wer
ou
t
pu
t
data
,
Me
te
olo
gica
is
currently
wor
kin
g
on
a
n
SMP
fo
reca
sti
ng
se
rv
ic
e
.
Me
te
og
r
oup
is
a
glo
bal
pr
i
va
te
weather
bus
iness
with
offi
ces
around
the
wo
rl
d
[8]
.
T
he
y
aim
to
co
m
bin
e
exp
e
rience
an
d
global
c
ov
e
r
age
with
l
ocal
ex
per
ti
se
t
o
offer
our
c
us
tom
ers
highly
accurate
a
nd
be
spoke
weathe
r
se
rv
ic
es.
Me
te
ogr
ou
p
has
rece
ntly
la
un
c
hed
a
por
ta
l
aim
ed
at
presenti
ng
m
et
eor
ol
og
ic
al
in
f
orm
at
ion
us
ef
ul to
en
e
r
gy
trad
er
s.
Used
ANN
-
ba
sed
load
fore
cast
ing
m
et
ho
ds
f
or
24
-
hour
-
ah
ead
peak
load
f
or
eca
sti
ng
by
us
in
g
forecast
ed
te
m
per
at
ur
e
[
9
]
.
They
pro
po
s
ed
a
one
hour
-
a
head
l
oad
f
or
ecast
in
g
m
e
thod
us
i
ng
th
e
m
os
t
sign
ific
a
nt
wea
ther
data.
In
th
e
pro
posed
f
oreca
sti
ng
m
et
hod,
weathe
r
dat
a
is
first
a
naly
zed
to
dete
rm
i
ne
the
m
os
t
cor
relat
ed
facto
rs
to
l
oa
d
cha
nges.
T
he
m
os
t
cor
relat
e
d
weat
her
data
is
then
us
e
d
i
n
trai
ning,
validat
ing
and
te
sti
ng
th
e
ne
ur
al
net
w
ork.
Co
rr
el
at
ion
analy
sis
of
weath
er
data
was
use
d
to
determ
ine
th
e
input
par
am
et
ers
of t
he neu
ral
netw
orks
a
nd they t
est
ed
it
on act
ua
l l
oad
data
fro
m
the Eg
y
ptian
Unified
Syste
m
.
O
utli
ne
a
ne
ural
netw
ork
a
ppr
oach
for
f
or
ecast
in
g
s
hort
-
te
rm
el
ect
ri
ci
ty
pr
ic
es
us
ing
a
back
-
pro
pag
at
io
n
al
gorithm
[
10
]
.
The
res
ults
ob
ta
ined
from
their
ne
ural
net
work
s
how
t
ha
t
the
neural
ne
twork
-
b
ase
d
ap
proac
h
is
m
or
e
acc
urat
e.
P
rese
nt
a
n
ANN
base
d
sh
ort
-
te
rm
load
f
oreca
sti
ng
m
od
el
fo
r
a
s
ubsta
ti
on
in
Kano,
Ni
geri
a
[
11
]
.
The
re
corde
d
daily
load
pro
file
with
a
le
ad
t
i
m
e
of
1
-
24
ho
ur
s
f
or
the
ye
ar
2005
was
ob
ta
ine
d
from
the
util
it
y
com
pan
y.
T
he
Le
ve
nb
e
r
g
-
Ma
r
qu
ard
t
optim
iz
ati
on
te
ch
nique
was
us
e
d
as
a
bac
k
-
pro
pag
at
io
n
al
gorithm
fo
r
t
he
Mult
il
ay
er
Feed
F
orwa
rd
A
NN.
T
he
f
orec
ast
ed
ne
xt
day
24
ho
ur
ly
pea
k
loads
wer
e
obta
ine
d
base
d
on
the
st
at
ion
ary
outp
ut
of
the
ANN
w
it
h
a
per
f
or
m
a
nce
Me
an
Squ
ared
Er
r
or
(M
SE)
of
5.84
e
-
6
a
nd
c
om
par
ed
favor
a
bly
with
the
act
ual
Po
we
r
util
it
y
data.
The
resu
lt
s
showe
d
that
their
te
chni
qu
e
i
s
rob
us
t
in
f
orec
ast
ing
fu
t
ur
e
l
oad
dem
and
s
f
or
t
he
da
il
y
op
erati
onal
pla
nnin
g
of
powe
r
syst
e
m
distribu
ti
on
su
b
-
sta
ti
ons in
Nige
ria.
Shor
t
-
te
rm
loa
d
f
or
eca
st
is
therefo
re
an
e
ssentia
l
par
t
of
el
ect
ric
pow
er
syst
em
pla
nn
i
ng
a
nd
op
e
rati
on.
For
ecast
ed
values
of
syst
em
load
a
ff
ect
t
he
de
ci
sion
s
m
ade
f
or
unit
c
omm
itm
ent
and
s
ecur
it
y
assessm
ent,
wh
ic
h
hav
e
a
di
rect
i
m
pact
on
op
e
rati
onal
costs
an
d
syst
e
m
secur
it
y.
Co
nv
e
ntio
nal
re
gressi
on
m
et
ho
ds
a
re
use
d
by
m
os
t
power
c
om
pan
ie
s
fo
r
l
oad
forec
ast
ing
.
Howe
ve
r,
due
to
the
nonlinea
r
relat
io
ns
hi
p
betwee
n
l
oad
and
fact
or
s
af
f
ect
ing
it
,
c
onve
ntion
al
m
et
ho
ds
a
re
no
t
suff
ic
ie
nt
en
ough
to
pr
ov
i
de
acc
ur
at
e
loa
d f
or
ecast
or to
conside
r
t
he
seasonal
var
i
at
ion
s
of loa
d.
We
belie
ve
art
ific
ia
l
neural
ne
tworks
(
A
NN)
base
d
loa
d
f
oreca
sti
ng
m
et
ho
ds
ca
n
deal
w
it
h
24
-
hour
-
ahead
loa
d
f
oreca
sti
ng
by
us
i
ng
f
oreca
ste
d
weathe
r
input
var
ia
bles,
w
hich
can
le
ad
to
hi
gh
f
or
eca
st
ing
erro
r
s
in
case
of
ra
pid
weat
her
c
ha
ng
e
s
[
12
]
,
[
13
]
.
A
NN
s
per
m
it
m
od
el
li
ng
of
com
plex
and
nonlinea
r
relat
io
ns
hi
ps
thr
ough
trai
ni
ng
with
t
he
use
of
histo
rica
l
data
an
d
ca
n
the
refor
e
be
us
e
d
in
m
odel
s
base
d
on
weathe
r
inf
or
m
at
ion
w
it
ho
ut
the
nee
d
for
ass
um
ption
s
f
or
a
ny
f
un
ct
io
nal
relat
ion
s
hip
bet
we
en
lo
ad
an
d
weathe
r
var
ia
bles.
We
ou
tl
ine
he
re
a
novel
neural
ne
twork
-
ba
sed
appr
oach
f
or
s
hort
-
te
rm
load
forecast
in
g
t
ha
t
us
es
the
correla
te
d
weathe
r
data
f
or
trai
ni
ng,
val
idati
ng
an
d
te
s
ti
ng
of
a
neura
l
network.
Correl
at
ion
analy
s
is
of
weathe
r
data
determ
ines
the
input
par
am
et
ers
of
t
he
ne
ur
al
ne
tw
orks.
The
s
uitabil
ity
of
the
pro
pose
d
appr
oach
is
il
lustrate
d
thr
ou
gh
a
n
a
pp
li
ca
ti
on
t
o
the
ac
tual
loa
d
data
of
th
e
I
rish
Ele
ct
rici
ty
Mar
ket.
This
pap
e
r
is
orga
nised
as
f
ollows:
Sect
io
n
2
pr
ov
i
des
a
bac
kgrou
nd
to
the
Si
ng
le
Ele
ct
rici
ty
m
a
rk
et
in
Ir
el
an
d,
sect
io
n,
sect
io
n
3
in
tro
du
ces
Ar
ti
fi
ci
al
Neu
ral
Ne
tworks
&
Short
-
te
rm
Load
Fo
recasti
ng,
sec
ti
on
4
pr
ese
nts t
he
s
hort
-
te
rm
f
or
eca
sti
ng
m
od
el
a
nd secti
on
5 pro
vid
es
a c
on
cl
usi
on
.
2.
SIN
GLE
EL
E
CTRICIT
Y MA
RKET
The
Si
ng
le
Ele
ct
rici
ty
Ma
rk
et
(S
EM)
is
th
e
w
ho
le
sal
e
el
ect
rici
ty
m
ark
et
for
the
isl
an
d
of
Ir
el
a
nd,
regulat
ed
j
oi
ntly
by
the
CER
and
it
s
c
ounter
par
t
in
Be
lfast
,
the
Util
i
ty
Re
gu
la
tor
.
The
Com
m
issio
n
f
or
Energy
Re
gula
ti
on
(CER)
is
the
ind
e
pe
nden
t
bo
dy
res
pons
ible
fo
r
r
eg
ulati
ng
the
na
tural
gas
an
d
el
ect
rici
ty
m
ark
et
s
in
I
reland.
By
com
bi
ning
w
hat
we
r
e
two
se
par
at
e
j
uri
sd
ic
ti
onal
el
ect
rici
ty
m
ark
et
s,
the
SEM
be
ca
m
e
on
e
of
the
first
of
it
s
kin
d
in
Euro
pe
w
hen
i
t
wen
t
li
ve
on
1s
t
Novem
ber
2007
[1
4
]
.
The
SEM
is
design
ed
t
o
pro
vid
e
for
the
le
ast
cost
sou
r
ce
of
el
ect
rici
ty
gen
e
rati
on
to
m
ee
t
custom
er
dem
and
at
any
one
tim
e
acro
s
s
the
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4060
-
4078
4062
isl
and
,
w
hile
a
lso
m
axi
m
isi
n
g
lo
ng
-
te
rm
su
sta
inabili
ty
an
d
reli
abili
ty
.
The
S
EM
is
op
erated
by
SE
MO,
t
he
Sing
le
Ele
ct
rici
ty
Ma
rk
et
O
pe
rator,
a
jo
int
-
ven
t
ur
e
bet
we
en
Eir
G
rid
a
nd
S
O
NI
,
the
t
ran
sm
issi
on
sy
stem
op
e
rato
rs
in
Irel
and
a
nd
N
ort
hern
Irel
an
d
resp
ect
ively
.
S
EMO
[
15
]
is
respo
ns
ibl
e
f
or
adm
inist
ering
th
e
m
ark
et
,
inclu
di
ng
payi
ng
ge
ne
rators
f
or
thei
r
el
ect
rici
ty
gen
erate
d
an
d
in
vo
ic
in
g
s
uppliers
f
or
the
el
ec
tric
ity
they
h
a
ve bou
ght [
1
4
]
.
SEM
co
ns
ist
s
of
a
ce
ntrali
se
d
an
d
m
and
at
or
y
al
l
-
isl
an
d
wholesal
e
po
ol
(o
r
spot)
m
ark
et
,
t
hroug
h
wh
ic
h
gen
e
rat
or
s
an
d
s
uppli
ers
tra
de
el
ect
r
ic
it
y.
Gen
erat
ors
bid
int
o
this
pool
th
ei
r
own
s
hort
-
run
c
ost
s
for
each
hal
f
hour
of
the
fo
ll
ow
i
ng
day,
w
hic
h
is
m
os
t
ly
their
fu
el
-
relat
ed
operati
ng
c
os
ts.
Ba
sed
on
this
set
of
gen
e
rato
r
c
os
ts
an
d
cust
om
e
r
dem
and
for
el
ect
rici
ty
,
the
Syst
e
m
Ma
rg
inal
Pr
ic
e
(S
M
P)
f
or
each
ha
lf
-
ho
ur
tradin
g
pe
rio
d
is
determ
ined
by
SEMO,
us
i
ng
a
sta
ck
of
t
he
chea
pest
al
l
-
isl
and
ge
ner
at
or
c
os
t
bid
s
ne
cessary
to
m
eet
all
-
isl
a
nd
dem
and
[1
6
]
.
It
is
these
m
or
e
ef
fici
ent
ge
ner
at
or
s
wh
ic
h
are
gen
e
rall
y
run
to
m
eet
dem
and
in
the
half
ho
ur
i
n
wh
at
is
known
as
the
“M
ark
et
Sc
he
dule
”.
Mo
re
e
xpen
sive
or
i
ne
ff
ic
ie
nt
gen
e
ra
tors
a
re
“ou
t
of
m
erit
”
and
he
nce
they
are
not
r
un
a
nd
are
not
pai
d
SMP,
keep
i
ng
custom
ers’
bill
s
dow
n
as
s
ho
wn
i
n
Figure
1
.
Figure
1
.
The
role o
f Sy
stem
Marg
inal
Pr
ic
e
The
SMP
f
or
e
ach h
al
f
hour
is
paid
t
o
al
l
ge
ner
at
o
rs
t
hat
ar
e
nee
ded
t
o
m
e
et
dem
and
.
S
uppliers
, wh
o
sel
l
el
ect
ricity
direct
to
the
final
co
ns
um
er,
buy
their
el
ect
rici
ty
fr
om
the
pool
at
this
com
m
on
pr
ic
e,
as
il
lustrate
d
in
Figure
2
.
O
ve
rall
the
S
E
M
facil
it
at
es
the
r
unni
ng
of
the
chea
pes
t
possible
ge
ner
at
or
s
,
determ
ined
by
the
sta
ck
of
ge
ner
at
io
n
c
os
t
bi
ds
,
to
m
eet
custom
er
de
m
and
acr
os
s
the
isl
and.
This
m
and
at
ory
centrali
sed
po
ol
m
od
el
in
SEM,
in
wh
ic
h
al
l
key
gen
erat
ors
and
s
upplie
rs
m
us
t
par
ti
ci
pate,
diff
e
rs
from
m
os
t
oth
e
r
Eu
ropea
n
m
ark
et
s in
whic
h
m
os
t t
rad
e takes p
la
ce bila
te
rall
y between
ge
ner
at
or
s and supp
li
ers
. In
these
bilat
eral
m
ark
et
s
on
ly
,
a
residu
al
am
ou
nt
of
el
ect
rici
ty
i
s
traded
in
an
exch
a
ng
e
,
pr
i
m
aril
y
fo
r
bal
ancin
g
pur
po
ses
.
I
n
c
on
t
rast
al
l
key
play
ers
m
us
t
tr
ade
in
SEM,
s
o
there
is
m
or
e
trans
par
e
ncy
associat
ed
with
SEM
pr
ic
es a
nd m
ark
et
outcom
es.
Figure
2
.
Wholesal
e an
d
retai
l
m
ark
et
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Forec
as
ti
ng
Shor
t
-
te
r
m
W
ho
le
sa
le
Prices
on
the I
ris
h
...
(
Fran
ce
sco
Arci)
4063
Gen
e
rato
rs
op
erati
ng
within
the
SEM
al
s
o
receive
se
pa
rate
capaci
ty
paym
ents
wh
ic
h
co
ntribute
towa
rd
s
their
f
ixed
c
os
ts,
if
t
hey
are
a
vaila
bl
e
to
ge
ner
at
e
.
The
ca
pacit
y
pa
ym
ent
po
t
of
m
on
ey
fo
r
ge
ne
rators
is
set
ahead
of
tim
e
by
the
SEM
Com
m
i
tt
e
e
and
is
cal
cul
at
ed
base
d
on
the
relat
ively
l
ow
fixe
d
costs
o
f
a
peak
i
ng
p
la
nt.
As
a
res
ult,
the
pay
m
ents
generall
y
cov
er o
nl
y
a
po
rtion
o
f
the
fixe
d
costs involve
d
in
bui
ldin
g
m
os
t
plants.
S
uppliers
al
s
o
pay
f
or
t
hese
capaci
ty
pay
m
ents
and
an
y
oth
er
syst
em
char
ges,
w
hich
a
re
ty
pical
ly
passed
th
rou
gh
to
c
ust
om
ers.
T
o
sel
l
el
ect
rici
ty
into
the SEM p
oo
l,
ge
ner
at
ors
m
us
t
subm
it
cost
bid
s
to
SEMO
t
he
day
be
fore
the
ph
ysi
cal
trad
e
/gene
rati
on
ta
ke
s
place,
known
as
D
-
1.
Th
e
bid
s
s
ubm
i
tted
are
pr
im
aril
y
based
on
a
ge
ner
at
or’s
r
unni
ng
or
S
hort
Ru
n
Ma
rg
inal
C
os
t
(S
RM
C),
i.e
.
t
he
cost
of
eac
h
e
xtra
M
W
it
cou
ld
pro
du
ce
e
xclu
ding
it
s
fixed
costs.
The
SR
MC
ref
le
ct
s
the
opportu
nity
cost
of
the
el
e
ct
rici
ty
pro
du
ce
d,
w
hich
is
the
e
co
nom
ic
activity
t
hat
the
ge
ner
at
or
f
orgo
e
s
to
pro
du
ce
el
ect
ri
ci
ty
.
Fo
r
e
xam
ple,
i
n
the
case
of
a
gen
e
rat
or
f
uelle
d
by
gas
,
the
opportu
nity
cost
inclu
des
th
e
pr
ic
e
of
gas
on
the
day
that
it
is
biddin
g
in
,
be
cause
if
t
he
ge
ner
at
or
was
no
t
producin
g
el
ect
rici
ty
it
cou
ld
sel
l
it
s
gas
in
the
op
e
n
m
ark
et
.
Gen
e
rato
r
bid
s
al
so
incl
ud
e
a
ge
ner
at
or’
s
st
art
-
up
c
os
ts,
w
hich
a
re
c
os
ts
it
faces
if
it
ne
eds
to
be
t
urne
d
on
after
a
pe
rio
d
of
inact
i
vity
,
as
well
as
ge
ner
at
or
no
-
lo
ad
costs
w
hich
are
(m
os
tl
y
fu
el
)
c
os
ts
w
hich
a
re
ind
if
fer
e
nt to
outp
ut levels.
The
ge
ner
at
or
s
subm
it
these
bid
s
to
SEMO
up
unti
l
Gate
Cl
os
ure,
cu
rre
ntly
at
10:0
0am
on
D
-
1.
So
ft
war
e
is
the
n
run
by
SEM
O
to
determ
ine
a
Ma
rk
et
Sched
ule
w
hich
f
oreca
sts
the
SMP
for
each
half
hour
tradin
g
per
i
od
for
the
f
ollow
i
ng
day.
H
ow
e
ver,
no
s
oft
wa
re
can
predict
with
c
om
plete
accuracy
w
ha
t
will
happe
n
in
reali
ty
:
real
-
t
i
m
e
factor
s
s
uc
h
as
a
cha
ng
e
i
n
wind
ge
ne
rati
on
or
c
us
tom
er
dem
and
,
wh
i
ch
can
aff
ect
SMP,
m
us
t
be
acc
ount
ed
for.
F
or
this
reas
on,
SEM
O
c
om
plete
s
two
m
or
e
s
of
tw
are
runs
ref
le
ct
ing
t
he
reali
ty
of
w
hat
happe
ne
d
in
ge
ner
at
or
disp
at
ch,
on
e
on
the
day
after
t
he
tr
adin
g
da
y
(D
+
1),
an
d
a
no
t
her
fou
r
days
after
(
D+
4),
to
cal
culat
e
the
final
SMP
fo
r
eac
h
ha
lf
hour
of
t
he
tra
ding
day.
T
his
D+4
pr
ic
e
is
the
one
that
is
paid
t
o
gen
e
rato
rs
a
nd
pai
d
by
s
uppl
ie
rs.
T
he
Ma
r
ke
t
Sche
dule
id
entifi
es
the
lo
west
c
os
t
s
olu
t
ion
at
w
hic
h
ge
ner
at
i
on
ca
n
m
eet
de
m
and
for
each
half
hour
tradi
ng
per
io
d.
It
ra
nk
s
gen
e
rato
rs
with
the
lowes
t
bid
s
first
un
ti
l t
he q
uan
ti
ty
n
ee
de
d for the
d
em
and is m
et
-
see bl
ue
s
ha
ded b
a
rs i
n
Figure
3
.
T
he
m
arg
inal
ge
nerat
or
nee
de
d
to
m
eet
the
dem
a
nd
set
s
the
SM
P
for
that
trading
per
io
d.
The
oth
e
r
ge
ne
rators
who
ha
ve
su
bm
itted
SRM
C
bid
s
lowe
r
than
this
pri
ce
are
deem
ed
to
be
“i
n
m
erit”
and
will
al
so
be
sc
heduled
t
o
r
un.
All
gen
e
rato
rs
who
hav
e
s
ub
m
itted
bid
s
w
hi
ch
ar
e
higher
t
han
t
his
pri
ce
(
SMP)
are d
eem
ed
to b
e “o
ut o
f
m
erit
” and
w
il
l no
t
be
sche
dule
d
to run
-
see the
gr
ee
n
ba
r
in
Fi
gure 3
. Th
ese t
end
to
be old
or ine
ff
i
ci
ent p
la
nts.
Figure
3
.
Ma
r
ke
t sche
du
le
All
ge
ner
at
ors
who
ha
ve
s
ub
m
it
te
d
a
bid
w
hic
h
is
unde
r
the
SMP
earn
a
pro
fit,
know
n
a
s
“i
nf
ram
arg
inal
re
nt”,
on
the
di
ff
ere
nce
betw
een
t
heir
SRM
C
bid
offe
r
a
nd
th
e
SMP
.
T
hi
s
is
il
lustrate
d
in
r
e
d
sh
a
ded
ba
rs
i
n
the
gr
a
ph.
Th
e
plant
that
set
s
the
m
arg
inal
p
rice
in
a
half
hour,
i.e.
t
he
on
e
with
the
hi
gh
est
run
ning
costs
a
m
on
g
tho
se
th
at
are
schedule
d
to
r
un,
does
no
t
recei
ve
an
y
infr
a
-
m
arg
in
al
ren
t.
H
ow
e
ve
r,
this
is
ty
pical
ly
a
p
eakin
g
pla
nt
w
hich,
w
hile
it
has
high
sho
rt
-
run
c
os
ts,
has
l
ow
fixe
d
c
os
ts.
Hen
c
e
it
s
c
os
t
s
are
cov
e
re
d
th
rou
gh
the
SMP
a
nd
the
c
apacit
y
paym
ents
it
receives
.
Infra
-
m
arg
inal
re
nt
is
nee
ded
f
or
m
os
t
gen
e
rato
rs
that
are
r
un,
incl
uding
e
ff
ic
ie
nt
m
od
e
r
n
gas
pla
nts
an
d
wi
nd
f
ar
m
s,
becau
se
w
hile
su
c
h
pla
nt
s
ha
ve
relat
ively
low
run
ning
co
sts
(
SRM
C),
they
hav
e
m
uch
hi
gher
fixe
d
c
os
ts
w
hich
t
he
(
rel
at
ively
low)
ca
pacit
y
paym
ent
do
es
no
t
f
ully
cover.
W
it
hout
in
fr
a
-
m
arg
inal
r
ent,
it
would
no
t
be
eco
no
m
ic
to
bu
il
d
m
od
ern
eff
ic
ie
nt
po
we
r
plants
or
wind
fa
rm
s,
threatenin
g
secu
rity
of
el
ect
rici
ty
su
p
ply
an
d
dri
vi
ng
higher
pr
ic
es
in
the lo
ng
-
r
un.
W
i
nd
fa
rm
s
are
an
exam
ple
of
el
ect
rici
ty
gen
e
r
at
ors
t
hat
ha
ve
ver
y
low
SRM
C
-
the
wind
is
fr
ee
-
an
d
s
o
ty
pi
cal
ly
they
receive
a
highe
r
r
at
e
of
in
fr
a
-
m
a
rg
i
nal
ren
t
th
a
n
ot
her
el
ect
ric
it
y
gen
erato
rs,
wh
ic
h
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4060
-
4078
4064
in
tu
r
n
is
needed
t
o
pay
f
or
t
heir
m
uch
hi
gher
fi
xed
costs
.
I
f
a
ge
ne
rato
r
was
disp
at
c
he
d
m
or
e
t
han
it
was
sche
du
le
d
to
in
the
Ma
r
ket
Sche
du
le
,
f
or
exam
ple
to
co
m
pen
sat
e
for
a
no
t
her
(c
hea
pe
r)
ge
ner
at
or
no
t
bein
g
bro
ught
onli
ne
du
e
t
o
a
net
w
ork
fail
ure
or
“
const
r
ai
nt”,
it
is
“co
ns
trai
ne
d
on
”
.
T
his
m
ea
ns
it
receives
i
ts
bid
cost
to
c
om
pen
sat
e
for
t
he
extra
M
W
it
m
us
t
pr
od
uce,
th
ough
it
does
not
receive
inf
ra
-
m
arg
ina
l
ren
t
.
Gen
e
rato
rs
w
ho
we
re
ori
gina
ll
y
included
in
the
Ma
rk
et
Sc
hedule,
bu
t
no
t
act
ually
ru
n
f
or
reas
ons
ou
ts
ide
of
their
c
on
t
ro
l,
f
or
exam
ple
du
e
to
a
net
work
fau
lt
,
a
re
sai
d
t
o
be
“c
onstrai
ne
d
off”.
They
r
ecei
ve
the
SM
P
le
ss
their
bid,
i.e.
the
inf
ra
-
m
arg
inal
ren
t
they
would
hav
e
r
e
cei
ved
in
the
m
ark
et
had
the
y
been
run.
C
onstrai
nts
costs also
cove
r
co
s
ts ass
ociat
ed wit
h
“re
ser
ve
”. T
his is whe
re,
to
ensu
re the
conti
nu
e
d
s
e
cur
it
y o
f
the
syst
e
m
,
for
exam
ple
in
the
eve
nt
of
a
gen
e
rato
r
trip
ping,
s
om
e
gen
erat
or
s
a
re
in
structed
t
o
r
un
at
lower
le
vel
s
than
ind
ic
at
ed
in th
e Ma
rk
et
Sch
e
du
le
. T
his m
ea
ns
ther
e is
sp
a
r
e g
enera
ti
on
ca
pacit
y avail
able (r
eser
ve)
wh
i
ch
can
be
quic
kly
bro
ught
onli
ne
if
need
e
d.
T
o
m
a
intai
n
the
dem
and
-
s
upply
balance,
this
rese
r
ve
m
eans
that
so
m
e
gen
e
rato
rs
will
be
co
ns
trai
ne
d
do
wn
w
hile
oth
e
rs
m
ay
be
co
ns
trai
ned
on
/
up,
agai
n
le
ading
t
o
the
act
ual
disp
at
c
h dev
ia
t
ing
from
the Mark
et
Sche
dule
[1
4
].
3.
AR
TIF
ICIAL
N
EU
R
AL
NETWOR
KS &
SHO
RT
-
TE
R
M
LO
A
D
FO
REC
AS
TI
NG
In
m
achine
le
arn
i
ng
an
d
c
ogni
ti
ve
sci
ence,
a
rtific
ia
l
neural
netw
orks
(
ANNs)
are
a
fam
i
l
y
of
m
od
el
s
insp
ire
d
by
bi
ologica
l
ne
ur
al
netw
orks
(t
he
central
ne
rvo
us
syst
em
s
of
anim
a
ls,
in
part
ic
ular
the
brai
n)
a
nd
are
us
e
d
to
est
im
at
e
or
ap
pro
xim
at
e
fu
nctions
that
can
dep
e
nd
on
a
la
r
ge
num
ber
of
i
nput
s
an
d
are
ge
ne
rall
y
unknow
n
[1
7
]
.
Ar
ti
fici
al
neural
netw
orks
are
generall
y
pr
esente
d
as
syst
e
m
s
of
interconnecte
d
"
ne
uro
ns
"
wh
ic
h
exc
ha
nge
m
essages
bet
ween
eac
h
othe
r
[
18
]
.
T
he
co
nn
ect
io
ns
ha
ve
nu
m
eric
weig
hts
that
can
be
tun
e
d
base
d
on
ex
pe
rience,
m
aking
ne
ur
al
nets
a
da
ptive
to
in
pu
t
s
an
d
ca
pa
ble
of
le
a
rn
i
ng.
F
or
e
xam
ple,
a
ne
ur
al
netw
ork for
ha
ndwr
it
in
g reco
gn
it
io
n
is
def
i
ne
d by a set
of
i
nput
neur
on
s
wh
ic
h
m
ay
b
e act
ivate
d
by t
he
p
ixels
of
a
n
in
put
im
age.
A
fter
be
ing
weig
hted
and
t
ran
s
f
orm
ed
by
a
f
unct
ion
(
determ
i
ned
by
the
ne
twork'
s
desig
ner),
t
he
act
ivati
on
s
of
t
hese
ne
uro
ns
a
re
then
passe
d
on
t
o
o
t
her
ne
uro
ns
[
1
9
]
.
T
hi
s
process
is
re
peated
un
ti
l
fi
nally
,
an
ou
t
pu
t
ne
uro
n
is
act
ivate
d.
This
determ
ines
w
hich
cha
ra
ct
er
was
rea
d.
Like
oth
er
m
a
chine
le
arn
in
g
m
et
ho
ds
–
syst
em
s
t
hat
le
arn
from
data
-
neural
net
works
ha
ve
be
en
use
d
t
o
so
l
ve
a
wi
de
va
ri
et
y
of
ta
sk
s
that
are
hard
to
s
olv
e
us
in
g
ordi
nar
y
ru
le
-
base
d
pr
ogram
m
ing
,
in
cl
ud
in
g
c
om
pu
te
r
visio
n
an
d
sp
eec
h
recog
niti
on
[
20
].
Fo
r
sho
rt
-
te
rm
load
f
or
ecast
i
ng,
the
Ba
c
k
-
Pr
opa
gatio
n
N
et
work
(BP
)
ne
twork
is
t
he
m
os
t
widely
us
e
d
one.
D
ue
to
it
s
abili
ty
to
ap
pro
x
im
at
e
any
cont
inu
ous
nonlin
ear
f
unct
ion,
the
BP
netw
ork
has
extra
ordina
ry
m
app
in
g
(fo
re
cast
ing
)
a
bili
ties.
The
BP
net
work
is
a
kind
of
m
ultilay
er
feed
f
orwa
rd
ne
twork
,
and
the tra
nsfe
r
functi
on w
it
hi
n
the n
et
w
ork i
s u
su
al
ly
a n
onli
near
functi
on su
c
h
as the
S
igm
oid
f
un
ct
io
n.
T
he
ty
pical
BP
network
st
ru
ct
ur
e
fo
r
s
hort
-
te
r
m
load
foreca
sti
ng
is
a
thre
e
-
la
ye
r
netw
ork,
wit
h
the
no
nlinear
Sigm
oid
f
unct
ion
as the
tra
nsf
er f
unct
io
n
[
21
]
.
Fu
ll
y con
ne
ct
ed
BP n
et
w
orks need
m
or
e trai
ning tim
e a
nd
a
re
no
t
a
da
ptive
e
noug
h
t
o
te
m
per
at
ur
e
cha
nge
s
there
fore
s
om
e
hav
e
m
ov
e
d
to
us
in
g
non
-
f
ully
connect
ed
BP
m
od
el
s
[22
]
.
Althou
gh
a
fu
l
ly
connecte
d
ANN
ca
n
ca
pt
ur
e
the
l
oad
c
har
act
erist
ic
s,
a
non
-
f
ully
co
nn
ect
e
d
ANN
is
m
or
e
adap
ti
ve
to
res
pond
to
tem
perat
ur
e
cha
nges.
Re
su
lt
s
al
so
show
that
the
f
oreca
sti
ng
accu
r
acy
i
s
sign
ific
a
ntly
im
pr
ov
e
d
f
or
a
bru
pt
tem
per
at
ur
e
c
ha
ng
i
ng
da
ys.
The
re
is
al
so
m
erit
in
com
bin
ing
seve
r
al
su
b
-
ANNs
t
og
et
he
r
to
gi
ve
bette
r
f
oreca
sti
ng
res
ults
su
c
h
as
us
i
ng
rec
urre
nt
hi
gh
or
de
r
neural
net
works
(RH
ONN)
[
23
]
.
Du
e
to
it
s
dy
nam
ic
natur
e,
the
RHO
NN
forecast
in
g
m
o
del
can
ada
pt
qu
ic
kly
to
cha
ng
i
ng
conditi
ons
su
c
h
as
im
po
rtant
load
var
ia
ti
on
s
or
c
ha
ng
es
of
the
daily
load
patte
r
n
[
22
]
.
A
bac
k
-
pro
pa
gation
netw
ork
is
a
ty
pe
of
a
rr
ay
whic
h
can
reali
ze
nonlinea
r
m
app
in
g
f
r
om
the
i
nputs
to
t
he
outpu
ts.
The
refore
,
t
he
sel
ect
ion
of in
pu
t
var
ia
bles of a lo
ad f
or
eca
sti
ng
net
w
ork
is ver
y i
m
po
rta
nt.
I
n
ge
ne
ral, ther
e a
re tw
o
se
le
ct
ion
m
et
ho
ds.
O
ne
is
base
d
on
e
xperie
nce
an
d
t
he
oth
er
is
bas
ed
on
sta
ti
sti
cal
analy
sis
su
c
h
as
the
ARIM
A
an
d
correla
ti
on an
a
ly
sis.
Fo
r
in
sta
nce,
we
can
de
note
the
load
at
ho
ur
k
as
l(k
)
s
o
a
ty
pical
selecti
on
of
inpu
ts
based
on
op
e
rati
on
ex
pe
rience
will
be
l(k
-
1),
l(k
-
24
),
t(k
-
1)
,
wh
e
re
t(k)
is
the
temperat
ur
e
co
rr
e
s
pondin
g
to
the
load
l(k).
Un
li
ke
th
o
se
m
et
ho
ds
wh
ic
h
a
re
bas
ed
on
e
xp
e
rie
nce,
we
can
a
pp
ly
a
uto
-
co
rrel
at
ion
a
naly
sis
on
t
he
histor
ic
al
loa
d
data
to d
et
erm
i
ne
the
in
pu
t va
riables. Auto
-
c
orrelat
ion
a
nal
ysi
s
sh
ould
sho
w
that
c
orrelat
ion
of
peaks
occurs
at
the
m
ulti
ples
of
24
-
hour
la
gs.
T
his
in
d
ic
at
es
that
the
l
oa
ds
at
the
sam
e
hour
s
ha
ve
ve
ry
s
tro
ng
correla
ti
on
wi
th
eac
h
oth
e
r.
The
refo
re,
t
hey
ca
n
be
c
ho
s
en
as
i
nput
var
ia
bles.
I
n
a
dd
it
io
n
t
o
us
i
ng
conve
ntion
al
inf
or
m
at
ion
s
uc
h
as
hist
or
ic
al
loads
an
d
te
m
per
at
ur
e
as
in
put
va
riables
,
w
ind
-
sp
ee
d,
s
ky
-
co
ve
r
can
al
s
o
be
use
d.
Po
te
ntial
input
var
ia
bles
cou
l
d
be
histo
r
ic
al
loads,
hist
or
ic
al
a
nd
f
uture
te
m
per
at
ur
e
s,
hour
of
day
ind
e
x,
da
y
of
wee
k
in
de
x,
wind
-
s
pee
d,
sk
y
-
c
over,
rainf
al
l
an
d
wet
or
dr
y
days.
Th
ere
are
no
hard
-
fast
ru
le
s
to
be
f
ol
lowed
to
d
et
erm
ine
inp
ut
var
ia
bles.
Thi
s
la
rg
el
y
dep
e
nd
s
on
e
ng
i
ne
erin
g
judgm
ent
and
exp
e
rience
.
[
24
]
found
that
f
or
a
norm
al
cl
i
m
at
e
area,
histor
ic
al
load
s,
hi
storical
&
fu
t
ure
tem
per
at
ur
e
s,
hour
of
day
an
d
day
of
wee
k
ind
e
x
are
su
f
fici
ent
to
giv
e
acce
pta
ble
foreca
sti
ng
resu
lt
s.
H
ow
e
ver,
for
an
ext
r
e
m
e
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Forec
as
ti
ng
Shor
t
-
te
r
m
W
ho
le
sa
le
Prices
on
the I
ris
h
...
(
Fran
ce
sco
Arci)
4065
weathe
r
-
c
ondit
ion
e
d
area
th
e
oth
er
i
nput
var
ia
bles
cl
a
sses
were
rec
omm
end
ed
,
be
cause
of
the
highly
nonlinea
r
relat
i
on
s
hi
p betwee
n
the
loa
ds
a
nd
the
weathe
r
c
onditi
ons.
4.
A
S
HO
RT
-
TE
RM EL
ECTR
ICITY M
A
R
KET
FO
REC
AS
TI
NG
MO
DEL
Ar
ti
fici
al
Neur
al
Netwo
r
ks
(
ANNs
)
can
on
ly
per
form
wh
at
they
wer
e
trai
ned
to
do.
T
her
e
fore,
to
achieve
s
hort
te
rm
load
f
or
e
cast
ing
,
t
he
se
le
ct
ion
of
the
trai
ning
data
is
a
cru
ci
al
on
e
.
T
he
crit
e
ria
for
sel
ect
ing
the
tr
ai
nin
g set
is th
at
the ch
a
racter
ist
ic
s o
f
al
l t
he t
rainin
g pairs i
n
the
traini
ng s
et
m
us
t be like
tho
s
e
of
the
day
to
be
f
or
ecast
e
d.
Choosin
g
as
m
any
trai
nin
g
pairs
as
possi
bl
e
is
no
t
the
correct
ap
proac
h
f
or
a
nu
m
ber
of
reas
on
s
.
O
n
reas
on
is
load
per
io
di
ci
ty
.
Fo
r
instance,
each
day
of
t
he
wee
k
ha
s
diff
e
ren
t
patt
ern
s
.
Ther
e
f
or
e,
us
ing
S
undays'
load
data
to
t
r
ai
n
the
net
wor
k
wh
ic
h
is
t
o
be
us
e
d
to
f
oreca
st
Monday
s
'
loads
would
le
ad
to
wrong
res
ults.
Also
,
as
l
oads
po
s
sess
dif
fer
e
nt
tre
nd
s
in
dif
fer
e
nt
pe
rio
ds,
rece
nt
data
is
m
or
e
us
ef
ul
th
a
n
ol
d
data.
Th
ere
for
e,
a
ver
y
la
r
ge
trai
ning
set
w
hich
incl
ud
e
s
old
data
is
le
ss
us
ef
ul
to
tra
ck
the
m
os
t recent tre
nd
s
.
To
obta
in
good
f
oreca
sti
ng
res
ults, d
ay
ty
pe
inf
or
m
at
ion
m
u
st be conside
re
d.
We can
ac
hieve
this by
const
ru
ct
in
g
di
ff
e
ren
t
A
NN
s
for
e
ac
h
day
t
ype
a
nd
fee
ding
eac
h
A
NN
the
c
orres
ponding
da
y
ty
pe
tr
ai
ni
ng
set
s
[
25]
,
[
2
6
]
.
Anot
her
way
is
to
us
e
only
on
e
A
N
N
but
co
ntain
the
day
ty
pe
inform
ation
in
the
input
var
ia
bles
[2
7
]
.
The
tw
o
m
eth
ods
ha
ve
t
heir
a
dv
a
ntage
s
a
nd
disa
dvanta
ges.
The
f
orm
er
us
es
a
num
ber
of
relat
ively
s
m
all
siz
e
networks
,
w
hile
the
la
tter
has
only
one
netw
ork
of
a
relat
ively
la
rge
siz
e.
The
da
y
ty
pe
cl
assifi
cat
ion
is
syst
e
m
dep
e
nd
e
nt
e.g.
the
load
on
Mo
nday
m
a
y
be
li
k
e
that
on
Tues
days
bu
t
not
al
ways
.
Ther
e
f
or
e,
one
op
ti
on
is
to
cl
assify
histor
ic
a
l
loads
into
cl
asses
su
c
h
as
Monday
,
T
uesd
a
y
-
Th
ur
s
day,
F
r
iday
,
Satur
day,
an
d
Sund
ay
/P
ub
li
c
ho
li
day.
T
he
Ba
ck
-
P
r
op
a
gat
ion
al
gorithm
is
widely
us
ed
in
sh
ort
-
te
rm
load
forecast
in
g
a
nd
has
s
om
e
good
feat
ur
es
s
uc
h
as
,
it
s
a
bili
ty
to
easi
ly
acc
om
m
od
at
e
weat
her
va
riables,
and
it
s
i
m
plici
t
exp
re
ssion
s
relat
in
g
inputs
an
d
outp
uts,
bu
t
it
is
al
so
a
ti
m
e
-
consum
ing
tra
ining
proce
ss
and
it
s
conve
rg
e
nce
to
local
m
ini
m
a
[2
8]
,
[
2
9
]
.
T
he
determ
inatio
n
of
the
op
ti
m
al
nu
m
ber
of
hidde
n
ne
uro
ns
is
a
cru
ci
a
l
issue
.
If
it
is
too
s
m
al
l,
the
netwo
r
k
can
not
po
ssess
suffici
en
t
inform
ation
,
and
the
re
fore
yi
el
ds
inaccu
rate f
or
e
cast
ing
res
ults. On the
o
t
her h
and, if
it
is to
o l
arg
e,
the t
raini
ng pr
ocess wil
l be
ve
ry lo
ng
[
30
].
Othe
r
key
fact
or
s
are
to
dete
r
m
ine
how
bi
g
the
pr
e
dicti
on
window
sho
uld
be.
F
or
insta
nce,
it
co
uld
po
s
sibly
be
c
ol
d
in
on
e
m
onth
so
is
t
his
va
li
d
12
m
on
th
s
la
te
r.
T
he
f
oreca
st
horiz
on
is
day
+
1
-
a
nd
f
or
rem
ai
nd
er
of
da
y.
This
is
fo
r
the
nex
t
avail
able
m
ark
et
.
The
m
od
el
m
ay
al
so
pro
vid
e
pr
e
dic
ti
on
s
for
48
/7
2
hours.
T
his
wil
l
le
ad
of
c
ours
e
to
dim
ensioned
res
ults,
bu
t
we
ass
ociat
e
a
corres
pondin
g
error
value
.
N
ot
al
l
el
ect
rici
ty
m
ark
et
s
f
ollo
w
the
sam
e
slots
so
in
pract
ic
e
we
aim
to
weathe
r
forecast
,
m
odel
netw
ork
t
opology
and
m
or
e.
S
om
e
of
the
m
a
in
facto
rs
f
or
forecast
in
g
are
dem
and
f
oreca
st,
est
i
m
at
e
d
po
wer
producti
on
capab
il
it
y
and
avail
able
inter
connecti
on
ca
pa
ci
ty
.
Ou
tl
ie
rs
include
weat
he
r
eve
nts,
s
olar
ecl
ipses
so
we
m
us
t
al
so
be
care
fu
l
not
to
fa
ct
or
into
our
m
od
el
.
The
i
niti
al
sta
ge
in
vol
ves
de
te
rm
ining
t
he
input
var
ia
bles
from
the d
em
and
,
power p
rod
uction an
d pr
ic
e
pr
e
di
ct
ion
d
at
a
w
e
dow
nlo
a
d from
SEMO
[
15
]
ca
n
s
ee
in
Ta
ble
1
.
Table
1
.
K
ey
da
ta
f
ie
lds
Variables n
a
m
e
The u
n
it of
m
easu
r
e
m
en
t
Exa
m
p
le
Tr
ad
e date
Day
o
f
m
o
n
th
1
Feb 2
0
1
6
Deliv
ery d
ate
Half
Hou
r
1
Feb 2
0
1
6
06
:
0
0
Ju
risd
ictio
n
ROI/NI
Fo
reca
st
MW
Megawatts
2
5
5
1
.9
8
So
larpo
wer
Megawatts
0
So
larpo
wer
Utili
za
tio
n
%
0
W
in
d
p
o
wer
Megawatts
2022
W
in
d
p
o
wer
Utiliz
atio
n
%
81
SMP
Euro
1
8
.9
Sh
ad
o
w Pr
ice
Euro
1
8
.80
9
9
9
9
We
plo
t
a
subs
et
of
data.
Fi
gure
4
show
s
S
olar
po
wer
pro
du
ct
io
n
in
Nor
ther
n
Ir
el
a
nd.
The
val
ue
of
horizo
ntal
axis
is
tim
e
do
m
ain
f
r
om
1
st
Feb
2016
t
o
9
th
F
eb
2016.
T
he
Re
d
li
ne
i
nd
ic
at
es
the
s
olar
powe
r
pro
du
ct
io
n
(M
W)
in
N
ort
hern
Ir
el
a
nd
a
nd
the
blu
e
li
ne
ind
ic
at
es
the
s
olar
powe
r
uti
li
zat
ion
rate
(%)
in
Northe
rn Irel
and.
In
Fig
ur
e
5
,
th
e
Re
d
li
ne
in
di
cat
es
the
wi
nd
powe
r
produ
ct
ion
(M
W)
in
N
or
t
hern
Ir
el
and
an
d
t
he
blu
e
li
ne
i
ndic
at
es
the
wi
nd
powe
r
util
iz
at
ion
rate
(%)
in
N
or
t
hern
I
relan
d.
I
n
Fi
gure
6
,
t
he
Re
d
li
ne
in
di
cat
es
the
wind
po
w
er
pro
duct
ion
(M
W)
in
the
Re
public
of
Irel
and
a
nd
t
he
blu
e
li
ne
in
dic
at
es
the
wind
powe
r
util
iz
at
ion
rate
(%)
i
n
th
e
Re
public
of
Ir
el
a
nd.
T
he
re
is
no
so
la
r
powe
r
pr
oductio
n
in
t
he
Re
public
of
I
r
el
and
.
Figure
7
s
hows
the
dem
and
pr
edict
ion
(Mega
watt
s)
of
t
he
R
epublic
of
Ir
el
a
nd
an
d
Northe
r
n
I
relan
d.
The
Re
d
li
ne
ind
ic
at
es
the
dem
and
(M
W)
in
Re
publi
c
of
Ir
el
an
d
a
nd
t
he
bl
ue
li
ne
in
dicat
es
th
e
dem
and
(M
W)
in
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4060
-
4078
4066
Northe
rn
Irel
a
nd.
Fig
ur
e 8
s
hows
the SMP
f
or
No
rth
an
d
S
ou
t
h.
The R
ed
li
ne
ind
ic
at
es the SMP (
E
uro) in the
Re
public o
f Ire
la
nd
a
nd the
blu
e li
ne
i
nd
ic
at
e
s the sha
dow p
rice (E
uro
)
in
the Re
public o
f Irelan
d.
Figure
4. S
olar
pow
e
r pro
du
ct
ion
i
n Northe
r
n Ir
el
an
d
Figure
5.
W
i
nd Producti
on
po
wer i
n N
or
the
r
n Ir
el
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Forec
as
ti
ng
Shor
t
-
te
r
m
W
ho
le
sa
le
Prices
on
the I
ris
h
...
(
Fran
ce
sco
Arci)
4067
Figure
6.
W
i
nd Producti
on
po
wer i
n
Re
p.
of
Ir
el
an
d
Figure
7. Dem
and pre
dicti
on
in Rep
ubli
c of
Ir
el
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4060
-
4078
4068
Figure
8
.
SMP
in No
rthern
Ir
e
la
nd
a
nd
Re
public o
f
I
relan
d
Ti
m
e
series
predict
ion
is
one
of
the
m
os
t
im
po
rtant
predi
ct
ion
that
c
ollec
t
past
obse
r
vations
of
a
var
ia
ble
an
d
a
naly
ze
it
to
ob
t
ai
n
the
un
der
ly
ing
r
el
at
ion
s
hip
s
bet
wee
n
his
torical
obser
va
ti
on
s,
but
tim
e
series
has
pro
per
ti
es
su
c
h
as
nonlin
earit
y,
chao
ti
c,
non
-
sta
ti
onar
y
and
cy
cl
ic
wh
ic
h
cause
pro
blem
s.
An
adap
ti
ve
neural
net
wor
k
ba
sed
f
uzzy
inference
syst
e
m
(A
N
FI
S
)
i
s
w
her
e
t
he
le
arn
i
ng
proces
s
es
are
perform
ed
by
interl
eavin
g
th
e
op
ti
m
iz
ation
of
t
he
antece
den
t
a
nd
c
oncl
us
io
n
par
ts
par
a
m
et
ers.
The
ANFIS
m
od
el
we
are
us
in
g
is
a
Tak
agi
-
ty
pe
N
eu
r
o
-
fu
zzy
Netw
ork
wh
ic
h
com
bin
es
neural
ne
tworks
a
nd
f
uzzy
syst
e
m
s.
Fu
zzy
reasonin
g
a
nd
netw
ork
cal
c
ulati
on
will
b
e a
vaila
ble sim
ult
aneously
.
Be
fore w
e em
plo
y t
he
ANFI
S
m
et
ho
d
to
for
ecast
the d
ai
ly
e
le
ct
rici
ty
SM
P d
at
a, the raw
d
at
a n
ee
ded
to
be
pr
e
proce
ssed
to
get
the
pro
per
in
put
and
we
need
t
o
determ
ine
the
data
input
va
ri
ables.
O
ne
in
put
data
sam
ple
inp
ut
con
sist
s
of
P
r
oduction
Forecas
ti
ng
(
D
-
2),
L
oa
d
F
or
ecast
in
g
(D
-
2)
an
d
P
re
vious
P
rices
Window
(D
-
9… D
-
2). Th
e d
at
a o
f
pr
oductio
n
forec
ast
ing
a
nd
loa
d
f
or
ecast
in
g
ca
n
be
ob
ta
in
ed
f
ro
m
the
Ex
-
Ante
la
g
-
2
file
.
The
data
of
pre
vious
pri
ces
window
can
be
obta
ine
d
from
the
Ex
-
A
nte
file
s
of
la
g
-
2,
la
g
-
3,
…,
la
g
-
9.
Pr
od
uctio
n
f
oreca
sti
ng
incl
udes
9*2*48
var
i
ables,
L
oad
F
oreca
sti
ng
i
nclu
des
4*
2*48
va
riables
an
d
P
re
vious
Pr
ic
es
Win
do
w
incl
udes
7*2*48
var
ia
bles
.
O
utput
(D)
i
nclu
des
48
vari
ables
to
c
om
par
e
with
c
on
t
r
ol
data
.
This
ou
t
pu
t c
a
n
see
in
Ta
ble
2.
.
Table
2
.
Four
Day Roll
in
g
L
oad F
or
ecast
S
a
m
ple
Prod
u
ctio
n
Fo
reca
stin
g
Load
Forecastin
g
Previo
u
s Prices
W
i
n
d
o
w
Ou
tp
u
t
Co
n
trol Data (
Ou
t
p
u
t)
Data Sa
m
p
le 1
D
-
2
(9
-
d
ay
Fo
reca
stin
g
)
D
-
2
(4
-
d
ay
Fo
reca
stin
g
)
EA(
D
-
9
),
E
A
(D
-
8
)
,
…,
EA(
D
-
2)
D
EA(
D
)=
(H1,
H2
,
…)
Data Sa
m
p
le 2
D
-
3
D
-
3
EA(
D
-
1
0
),
EA(
D
-
9
),
…,
EA(
D
-
3)
D
-
1
EA(
D
-
1)
We
ex
per
im
ented
with
ot
her
al
go
rithm
to
determ
ine
the
par
am
et
ers
of
the
ANFI
S
m
od
el
(
Gr
i
d
Partit
ion
in
g,
s
ub
t
racti
ve
cl
us
te
ring
an
d
FC
M
cl
us
te
rin
g),
trai
ning
m
eth
od
(SOM
al
gorithm
,
Leve
nb
e
r
g
-
Ma
rquardt al
gorithm
,
Ba
ye
si
an
Re
gula
rizat
ion
a
nd Scaled
Conj
ug
at
e Gra
dient)
, A
R m
od
el
, s
ta
te
sp
ace
m
od
el
and A
R
IMA
X m
od
el
, N
eu
ral
Netw
ork
a
nd F
uzzy I
nf
e
re
nce
Syste
m
.
Nex
t,
we
exam
ine
our
feat
ur
e
sel
ect
ion
m
e
t
hodolo
gy.
Feat
ur
e
sel
ect
io
n
is
the
process
of
sel
ect
ing
a
su
bse
t
of
r
el
e
van
t
fe
at
ur
e
s
for
us
e
i
n
m
od
el
c
on
st
ru
ct
ion
.
Feat
ure
Sele
ct
ion
is
pl
aced
into
tw
o
m
ai
n
cat
egories, w
ra
pp
e
r
m
et
ho
ds
a
nd
filt
er
m
et
hod.
W
r
ap
pe
r
m
e
thods
e
valuate
m
ul
ti
ple
feature
s
us
i
ng p
r
oce
dures
that
ad
d
an
d/or
rem
ov
e
predic
tors
to
fin
d
t
he
op
ti
m
al
co
m
bi
nation
t
hat
m
a
xim
iz
es
m
od
el
perform
ance.
W
e
us
e
Re
c
ur
si
ve
Feat
ur
e
Elim
inati
on
with
Ba
c
kw
a
r
ds
Sele
ct
ion
in
our
featu
re
sel
ect
ion
m
od
el
a
nd
us
e
R
andom
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Forec
as
ti
ng
Shor
t
-
te
r
m
W
ho
le
sa
le
Prices
on
the I
ris
h
...
(
Fran
ce
sco
Arci)
4069
Fo
r
ecast
Me
th
od
as
t
he
f
or
ec
ast
ing
al
gorith
m
.
An
obvious
co
ncern
is
tha
t
too
fe
w
va
riables
a
re
sel
ect
ed
or
that
the
sel
ect
ed
set
of
in
put
var
ia
bles
is
not
su
ff
ic
ie
ntly
inf
or
m
at
ive.
Ha
lf
-
ho
ur
ly
SMP
it
sel
f
can
be
di
vid
ed
betwee
n
the
s
ha
dow p
rice an
d u
plift p
rice.
The
SM
P
f
ollow
s
cu
stom
er
dem
and
,
as
a
m
or
e
exp
e
ns
iv
e
sta
ck
of
ge
ne
rators
is
needed
to
m
eet
dem
and
w
he
n
it
is
high,
w
her
eas
at
lo
w
dem
and
ti
m
e
s
dem
and
ca
n
be
m
et
with
chea
per
ge
ne
rators.
Appro
xim
at
e
ly
80
%
of
the
isl
and’s
el
ect
rici
ty
gen
erati
on
c
om
es
fr
om
i
m
ported
f
os
sil
f
ue
ls,
with
m
os
t
this
in
the
f
or
m
of
ga
s
-
fire
d
ge
nerat
ion
plants,
t
hough
t
he
am
ount
o
f
ren
e
w
able
ge
ner
at
i
on
(es
pecial
ly
wind)
is
increasin
g.
Th
e
sta
rt
date
of
trai
ning
date
was
20
-
11
-
2016
an
d
the
la
st
date
of
trai
ni
ng
date
was
20
-
1
-
2017.
The
pr
e
proces
sing
inclu
de
d
norm
al
iz
a
ti
on
,
separ
at
io
n
of
i
nput
a
nd
outp
ut,
rem
ov
al
of
the
col
um
n
wi
th
ne
a
r
zero
va
riance
an
d
rem
ov
al
of
th
e
col
um
n
with
hi
gh
co
rr
el
at
io
n.
The
i
nputs
w
ere
["
De
li
ve
ry_D
at
e"
,
"Deli
ver
y
_Ho
ur","Deli
ver
y_
I
nter
val",
"SMP
_D_E
uro",
"
S
MP_D_Mi
nus
_6_E
uro",
"
SMP_
D_
Mi
nus
_1
3_Eu
ro
"
,
"Lo
ad
Dem
and
",
"
Power
_Pr
oductio
n_Ir
el
a
nd
"
,
"
Ou
t
pu
t
_SM
P_
E
uro"].
The
resam
pling
m
et
ho
d
is
cv
(cross
validat
io
n)
,
t
he
nu
m
ber
of
di
vid
ed
blo
c
ks
i
s
9.
T
he
W
M
m
et
ho
d
tu
ning
Gr
id
of
num
.l
abel
is
5,7,9,
11.
T
he
no
ta
ti
on
us
e
d
t
hro
ughout
the
pap
e
r
is
pro
vi
ded
i
n
Tabl
e
3
.
The
trai
ning
data
is
sh
ow
n
in
Table
4
.
Th
e
W
M
m
et
ho
ds are
show
n
i
n
Ta
ble
5
an
d
Ta
ble
6
s
hows
the
n
e
ur
al
netw
orks
m
et
ho
ds.
Table
3
.
N
om
e
nclat
ur
e
u
se
d
No
tatio
n
Meanin
g
D
Rep
o
rt
d
ate,
su
ch
a
s 1
1
/2
5
/1
6
D+2
The d
eliv
ery d
ate
o
f
pred
icted
SM
P,
su
ch
as
1
1
/2
7
/1
6
(
7
:0
0
a
m
–
6
:3
0
a
m
+1
)
S
MP
D+2
h
h
The o
u
tp
u
t (
7
:0
0
a
m
–
6
:3
0
a
m
+1
)
Deman
d
D+2
h
h
The De
m
an
d
corre
sp
o
n
d
i
n
g
to th
e outpu
t (
7
:0
0
a
m
–
6
:3
0
a
m
+1
)
Pow
er_
Irela
n
d
D+2
h
h
The p
o
wer
su
m
m
a
tio
n
of
Solar po
wer
an
d
wind
po
wer p
rod
u
ctio
n
in th
e
wh
o
le I
reland
(
7
:0
0
a
m
–
6
:3
0
a
m
+1
)
Pow
er_
UK
D+2
h
h
The p
o
wer
su
m
m
a
tio
n
of
Solar po
wer
an
d
wind
po
wer p
rod
u
ctio
n
in th
e
wh
o
le UK
m
ain
lan
d
(
7
:0
0
a
m
–
6
:3
0
a
m
+1
)
S
MP
D+1
h
h
The SMP
to
m
o
r
row (7:0
0
a
m
–
6
:3
0
am
+1
)
S
MP
D
-
5hh
The week
-
ah
e
ad
S
MP
o
f
the p
redicte
d
date
S
MP
D
-
12hh
The 2
-
week ah
e
ad
SMP of
the p
redict
ed
date
S
MP
D+1
h
h
-
1
The SMP
of
prev
io
u
s h
alf
h
o
u
r
S
MP
D+1
h
h
-
2
The SMP
of
prev
io
u
s h
o
u
r
Table
4.
T
raini
ng d
at
a
set
SMP
D
Euro
SMP
D
-
1
Euro
SMP
D
-
2
Euro
SMP
D
-
3
Euro
SMP
D
-
4
Euro
SMP
D
-
5
Euro
SMP
D
-
6
Euro
SMP
D
-
13
Euro
SMP
HH
-
1
Euro
SMP
HH
-
2
Euro
Load
De
m
an
d
Po
wer
Prod
Ir
elan
d
Po
wer
Prod
UK
Ou
tp
u
t
SMP
Euro
3
4
.11
5
6
.12
3
5
.58
3
5
.45
3
5
.45
3
3
.85
3
6
.02
3
7
.27
3
8
.56
4
0
.22
3
3
3
2
.9
9
2632
6432
2
6
.82
3
4
.96
5
3
.31
3
4
.96
3
6
.22
3
5
.67
3
3
.85
3
9
.60
3
7
.22
3
4
.11
3
8
.56
3
6
1
7
.2
2
2632
6432
3
3
.29
3
7
.35
5
2
.05
3
5
.93
3
6
.22
3
7
.65
3
3
.93
4
8
.55
4
0
.12
3
4
.96
3
4
.11
4
0
4
4
.0
4
2622
6301
3
3
.37
4
6
.83
4
8
.49
3
6
.95
4
5
.34
4
9
.53
4
1
.99
5
8
.69
4
8
.23
3
7
.35
3
4
.96
4
5
9
8
.2
6
2622
6301
4
4
.12
5
3
.00
4
5
.41
3
9
.11
5
8
.81
5
4
.65
5
0
.07
4
9
.99
5
2
.00
4
6
.83
3
7
.35
4
7
9
4
.3
2
2588
6213
3
6
.26
5
3
.00
4
2
.83
4
5
.16
5
9
.50
5
4
.65
5
0
.07
4
8
.91
5
3
.24
5
3
.00
4
6
.83
4
8
4
8
.4
4
2588
6213
3
6
.26
Table
5.
WM
Me
thods
W
an
g
and
M
en
d
el
Fu
zzy
Inf
erence S
y
ste
m
W
an
g
and
M
en
d
el
Fu
zzy
Ru
les
Nu
m
labels
RMSE
RSq
u
ared
Nu
m
L
ab
els
RMSE
RSq
u
ared
5
0
.08
0
8
5
9
7
6
3
9
1
0
.61
6
4
1
4
8
1
0
4
5
0
.08
2
4
3
6
0
2
9
5
1
0
.49
4
4
0
1
7
3
2
3
7
0
.08
3
4
8
1
1
1
3
4
1
0
.59
8
5
5
3
2
1
7
1
7
0
.08
0
3
4
6
8
1
8
5
8
0
.53
2
9
7
4
3
3
2
9
9
0
.08
2
8
2
7
0
7
4
3
3
0
.60
4
5
3
6
7
7
7
5
9
0
.06
5
2
0
3
5
2
4
7
7
0
.58
0
2
9
8
4
6
0
9
11
0
.08
3
5
1
7
3
2
0
6
0
0
.60
3
1
9
3
8
9
0
4
11
0
.06
1
5
8
8
5
9
2
1
3
0
.61
1
7
2
6
5
5
5
4
13
0
.08
2
9
7
8
3
0
4
4
4
0
.60
8
7
7
3
8
0
9
1
13
0
.06
2
8
8
6
7
1
3
6
8
0
.59
9
5
7
5
4
1
5
4
15
0
.08
1
4
1
6
3
7
7
1
3
0
.61
3
3
5
1
4
1
2
9
15
0
.06
0
9
6
2
5
8
8
1
8
0
.60
6
4
9
9
6
3
8
1
Table
6.
Ne
ur
a
l Netw
ork
Me
t
hods
Neu
ral
N
etwo
rk
Neu
ral
N
etwo
rk w
ith
Feature
Extract
io
n
Size
Decay
RMSE
RSq
u
ared
Size
Decay
RMSE
RSq
u
ared
7
0
.1
0
.05
2
7
0
5
2
4
8
9
4
0
.69
2
9
7
7
8
9
3
3
7
0
.1
0
.05
0
8
9
1
5
3
3
5
0
0
.72
0
8
9
3
3
8
4
3
7
0
.2
0
.05
3
4
7
3
1
1
1
4
0
0
.68
8
8
2
9
5
0
7
6
7
0
.2
0
.05
1
6
7
9
9
5
5
6
1
0
.71
4
0
0
1
9
1
2
8
7
0
.3
0
.05
4
5
3
9
6
3
4
1
4
0
.68
3
8
9
9
7
2
5
1
7
0
.3
0
.05
2
5
0
2
3
9
8
4
3
0
.70
5
4
3
8
4
8
2
0
7
0
.4
0
.05
5
7
3
6
1
4
6
6
8
0
.68
1
8
4
3
2
2
7
0
7
0
.4
0
.05
3
1
0
7
8
8
1
5
2
0
.69
9
1
3
6
7
9
9
7
7
0
.5
0
.05
6
9
7
1
7
9
1
8
4
0
.67
9
2
3
8
1
8
1
0
7
0
.5
0
.05
3
5
2
1
0
1
6
2
5
0
.69
3
9
2
9
1
5
9
2
9
0
.1
0
.05
2
7
1
4
5
2
6
3
4
0
.69
2
7
2
3
2
3
8
0
9
0
.1
0
.05
1
1
9
4
6
5
9
1
8
0
.71
7
5
1
9
9
8
0
1
9
0
.2
0
.05
3
4
4
6
1
1
8
6
9
0
.68
8
9
0
5
0
0
4
0
9
0
.2
0
.05
1
6
4
1
5
6
2
9
3
0
.71
4
6
2
5
6
6
8
3
9
0
.3
0
.05
4
3
7
6
6
7
3
8
8
0
.68
5
6
2
6
0
3
3
8
9
0
.3
0
.05
2
4
0
6
5
5
2
9
9
0
.70
5
4
1
8
8
3
3
6
Evaluation Warning : The document was created with Spire.PDF for Python.