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.
15
,
No.
1
,
Febr
uary
20
25
, pp.
10
38
~
1050
IS
S
N:
20
88
-
8708
, DO
I: 10
.11
591/ij
ece.v
15
i
1
.
pp
1038
-
10
50
1038
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om
Airport
infrastr
uct
ure
an
d
runway
precisi
on aids f
or
forecasti
ng flight
arri
va
l
delays
Hajar
A
ll
a,
Y
ou
sse
f
B
alo
u
k
i
Labo
ratory o
f
Mat
h
em
atics,
Co
m
p
u
ter
an
d
E
n
g
in
eering
Sciences,
Dep
art
m
en
t of
Mathe
m
ati
cs an
d
Co
m
p
u
ter
S
cien
ce,
Facu
lty
of Science
s an
d
T
echn
iq
u
es, Hassan
Fir
st
Un
iv
ersity
of Settat
,
Sett
at,
Moro
c
co
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
M
a
y 25,
2024
Re
vised
Sep 5
,
2024
Accepte
d
Oct
1,
2024
Re
cent
resea
r
ch
ha
s
co
nc
entrated
on
us
in
g
mac
hine
le
arn
in
g
appr
oach
es
to
forecast
fligh
t
delays
.
T
he
m
aj
or
it
y
of
pri
or
pre
dicti
on
al
gorithms
w
e
re
base
d
on
simple
an
d
sta
ndar
d
at
trib
utes
colle
ct
ed
from
the
data
ba
se
f
rom which
t
he
data wer
e
pull
ed.
T
his
a
r
ti
cl
e
is
the
first
at
te
m
pt
t
o
pr
opos
e
no
vel
feat
ur
es
li
nke
d
t
o
ai
rpo
rt
ca
pacit
y
an
d
infr
a
struct
ur
e
.
The
t
otal
r
unw
ays,
the
total
r
unwa
y
i
ntersec
ti
on
s,
th
e
longest
r
unwa
y
le
ngth,
th
e
sh
ort
est
run
way
le
ng
t
h,
the
r
unwa
y
pr
eci
sio
n
rate,
the
t
otal
te
r
minals,
a
nd
t
he
t
otal
gates
we
re
al
l
examine
d.
I
n
this
paper,
we
s
uggest
a
n
op
ti
mize
d
mu
lt
il
ayer
per
ce
ptr
on
t
o
pr
e
dict
fligh
t
arr
ival
retards
implementi
ng
data
for
domesti
c
fli
ghts
operate
d
in
U
nited
Stat
es
ai
r
ports.
We
employe
d
data n
ormal
iz
a
ti
on
,
sam
plin
g
te
chn
iq
ues
,
a
nd
hype
r
-
pa
ram
et
er
tu
ning
to
stren
gth
e
n
the
reli
abili
ty
of
t
he
su
ggest
e
d
m
od
el
.
The
exp
e
rime
ntal
fin
dings
de
monstrate
d
that
data
nor
mali
zat
ion
,
samplin
g
ap
proach
e
s,
a
nd
B
ayesian
opti
mi
zat
ion
pro
duce
d
t
he
mo
st
accurate
mode
l
with
92.
49%
accu
rac
y.
The
ac
h
ie
ve
m
ents
of
th
e
study
we
re
c
ompa
red
to
ot
he
r
ben
c
hma
r
k
researc
h
from
li
te
ratur
e
.
The
ti
me
c
omplexit
y
f
or
t
he
pro
posed
m
od
el
was
c
ompu
te
d
a
nd
pr
ese
nted
at th
e en
d of t
he
i
nvest
igati
on.
Ke
yw
or
d
s
:
Air
port in
fr
ast
r
uctu
re
Ba
yesian o
ptimi
zat
ion
Fli
gh
t
delay
pr
edict
ion
M
ulti
la
yer
perce
ptr
on
Runwa
y pr
eci
s
ion
ai
ds
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
BY
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Hajar
Alla
Lab
or
at
or
y of
M
at
he
mati
cs,
Com
pu
te
r
and
En
gin
eeri
ng Sc
ie
nces
,
De
par
t
ment
of M
at
he
mati
cs an
d
Com
pu
te
r
Scie
nce, Facult
y
of Sciences
and
Tech
niques,
H
assan Fi
rst
Un
i
ver
sit
y o
f Set
ta
t
Sett
at
, 2
60
00, Mor
occo
Emai
l:
h.
al
la
@
uhp.
ac
.ma
1.
INTROD
U
CTION
Fli
gh
t
dela
ys
can
arise
fro
m
mu
lt
iple
s
ources
,
co
mpris
ing
e
xtreme
weathe
r,
tra
ff
i
c
ja
ms,
pilot
exp
e
riences
an
d
qual
ific
at
ions,
mainte
nan
ce
issues
or
rep
a
irs
nee
ded
on
the
ai
rcr
a
ft,
la
te
-
ar
riving
in
bo
un
d
ai
rcr
aft
or
pas
sen
ger
s
,
an
d
c
rew
sc
he
du
li
ng.
In
s
ome
cas
es,
dela
ys
ma
y
al
so
be
cau
se
d
by
st
rikes
or
la
bor
disputes,
sec
uri
ty
concer
ns
or
ai
rpo
rt
co
ns
t
ru
ct
io
n,
delay
s
du
e
to
ov
e
rbookin
g,
or
iss
ues
with
co
nn
ect
ing
fligh
ts.
Air
port
or
r
unwa
y
cl
osures
or
co
ns
tr
uction,
in
fr
ast
r
uctu
re,
a
nd
ca
pa
ci
ty
are
al
so
c
on
si
der
e
d
f
or
traf
fic
delays
.
Air
p
ort
capa
ci
ty
ca
n
be
c
onside
re
d
as
the
highest
numb
e
r
of
ai
rc
raf
t
a
nd
pas
se
ng
e
rs
that
an
a
irp
or
t
can
ha
nd
le
at
a
ny
giv
e
n
ti
me
.
Fact
or
s
that
ca
n
a
ff
ect
ai
r
port
capaci
t
y
i
nclu
de
t
he
num
ber
and
siz
e
of
r
unway
s
,
ta
xiways,
gates
,
a
nd
te
r
minals
,
as
well
a
s
t
he
ef
fici
ency
of
a
irp
or
t
operati
ons.
T
he
i
nfrastr
uctu
re
of
a
n
ai
r
port,
if it
is not d
edi
cat
ed
to
ha
ndli
ng a lot
of airc
raf
t at
th
e same
ti
me, ca
n
le
ad
to f
li
ght
delays
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
Airp
or
t i
nfr
as
t
ru
ct
ur
e
an
d
r
unway
p
reci
si
on a
id
s for
forec
ast
ing
fl
ig
ht a
r
riv
al d
el
ays
(Haj
ar
All
a)
1039
To
e
nh
a
nce the
saf
et
y
a
nd ef
fici
ency
of
ai
rcra
ft operati
ons
on
r
unwa
ys
, pr
eci
sion
aid
s are
instal
le
d
at
ai
rpor
ts
in
ord
er
to
help
pilots
al
ign
their
ai
rcr
a
ft
with
the
runway
a
nd
de
scen
d
to
a
sa
fe
la
nd
in
g.
Thes
e
ai
ds
can
i
nclu
de
e
quipme
nt
li
ke
t
he
i
ns
tr
um
e
n
t
la
nd
in
g
s
ys
te
ms
(I
L
S)
that
help
pilots
la
nd
a
n
ai
rcr
a
ft
s
afely,
especial
ly
in
poor
visibil
it
y
conditi
ons.
F
ur
t
hermo
re,
th
e
micro
wa
ve
la
ndin
g
s
ys
te
ms
(
M
LS
)
prov
i
de
pr
eci
se
gu
i
dan
ce
to
t
he
run
way
us
i
ng
ra
dio
sig
nal
s
f
or
la
nd
i
ng
in
lo
w
-
visibil
it
y
c
onditi
ons
,
and
r
unwa
y
li
gh
ti
ng
sy
ste
ms
,
w
hich
hel
p
pilots
id
entify
t
he
locat
ion
of
the
r
unway
a
nd
it
s
bounda
ries.
Ot
he
r
exa
mp
le
s
of
r
unwa
y
pr
eci
sio
n
ai
ds
include
visu
al
glide
slo
pe
in
di
cat
or
s,
r
unwa
y
al
i
gnment
in
dicat
or
li
ghts,
and
run
way
th
reshold
li
gh
ts.
T
hese
sy
ste
ms
are
de
sign
e
d
t
o
help
pilots
ma
ke
safe
a
nd
acc
ur
at
e
la
nd
i
ngs,
eve
n
i
n
c
halle
ng
i
ng
weathe
r
co
ndit
ion
s
.
T
he
area
na
vig
at
io
n
(R
NAV)
proce
dure
,
w
hich
sta
nd
s
f
or
a
rea
na
vig
at
io
n,
is
a
lso
a
pr
eci
sio
n
to
ol
us
e
d
by
pilots
to
fly
m
or
e
dir
ect
r
ou
te
s
.
R
N
AV
is
al
so
known
as
a
glob
al
posit
ion
i
ng
sy
ste
m
(G
P
S)
na
vig
at
ion,
as
it
com
m
on
l
y
use
s
G
PS
data
to
de
fine
the
posit
ion
of
the
ai
rcr
a
ft
an
d
guide
it
to
wa
rd
it
s
final
point.
It
is
ano
t
her
t
yp
e
of
nav
i
gatio
n
mode
avail
able
in
the
flig
ht
mana
geme
nt
s
ys
te
m
(
FMS)
i
ns
ta
ll
ed
on
mode
rn
ai
r
planes
.
All
th
is
pr
eci
sio
n
e
qu
i
pm
e
nt
help
s
reduce
fli
ght
ti
mes
and
f
uel
co
nsum
ption
b
y
avo
i
ding
the
possibil
it
y
of
goin
g
ar
ound.
I
n
fact,
i
f
the
r
unwa
y
is
eq
ui
pp
e
d
with
non
-
preci
sio
n
ai
ds,
the
ai
rcr
aft
will
no
t
be
a
ble
t
o
la
nd
in
ba
d
wea
the
r
a
nd
will
pe
rform
a
misse
d
a
ppr
oach
ins
te
ad
or
di
ver
t
to
a
n
al
te
rn
at
e air
por
t.
Fo
ll
owin
g
the
In
te
r
natio
nal
Ci
vil
Av
ia
ti
on
O
rg
a
nizat
io
n
(
ICA
O)
[1]
,
a
missed
ap
proac
h
or
a
go
-
ar
o
und
is
a
n
op
e
rati
on
pe
rformed
by
a
n
ai
rcr
a
ft
by
st
oppi
ng
an
d
i
nterrup
ti
ng
t
he
a
ppr
oac
h
if
t
he
vi
su
al
ref
e
ren
ce
nece
ssary
a
nd
the
minimu
m
nee
ded
f
or
la
ndin
g
has
not
bee
n
est
a
blished
or
reac
hed.
Fi
gure
1
represe
nts
a
n
a
ircraft
go
-
a
rou
nd
pr
ocedure
.
Poor
ai
rpo
rt
in
fr
ast
r
uctu
re,
s
uc
h
as
ina
dequa
te
run
way
cap
aci
ty,
ou
t
dated
or
li
mit
ed
te
rmi
na
l
facil
it
ie
s,
and
no
n
-
pr
eci
si
on
ai
ds,
ca
n
con
t
rib
ute
to
traff
ic
de
ns
it
y
an
d
congesti
on,
w
hi
ch
le
ads
to
fligh
t
delays
.
For
this
end
a
nd
a
s
far
as
we
kn
ow,
we
offer
e
d
novel
at
t
rib
utes
that
are
relat
e
d
to
t
he
in
fr
a
struct
ure
of
t
he
ai
r
por
t
and
the
preci
sion
of
it
s
ai
ds,
w
hich
ha
ve
ne
ver
bee
n
c
ons
idere
d
in
pr
e
vious
st
ud
ie
s
,
namel
y,
num
be
r
of
r
unwa
ys
,
r
unw
ay
i
ntersecti
on
s,
l
ongest
r
unway
le
ng
t
h,
s
hortest
runway
le
ng
t
h,
run
way
pr
e
ci
s
ion
rate,
num
ber
of
te
rmin
a
ls,
an
d
num
be
r
of
gates
.
Ot
he
r
rele
van
t
fe
a
tures,
su
c
h
as:
ai
r
por
t
name,
da
y
of
week,
ai
rline
,
ta
il
nu
m
be
r
(
re
gistrati
on),
flig
ht
num
ber,
ai
r
port
of
or
i
gin
,
ai
rpor
t
of
destinat
io
n,
ar
rival
ti
me
,
dep
a
rtu
re
ti
me
,
a
rr
ival
de
la
y
(
bin
a
ry),
a
nd
dep
a
rtu
re
dela
y
(
bin
ar
y)
ha
ve
bee
n
ta
ken
from
the
B
ur
ea
u of
Tra
nsporta
ti
on Stat
ist
ic
s
database
(B
TS)
[2]
.
Figure
1. A
mi
ssed
a
ppr
oac
h op
e
rati
on (
s
our
ce:
[3]
)
This
researc
h
is
inten
de
d
t
o
de
li
ver
a
n
a
naly
ti
cal
pr
edict
ive
fr
a
me
work
th
at
minimi
zes
t
he
im
pact
of
delays
an
d
ca
nc
el
la
ti
on
s
on
pa
ssen
ger
s
,
ai
rli
nes,
a
nd
ai
r
por
t
auth
or
it
ie
s
by
pr
e
dicti
ng
t
he
arr
i
val
dela
y
of
a
par
ti
cula
r
flig
ht
based
on
new
delay
-
c
ontrib
ut
ing
fact
or
s
tha
t
wer
e
not
stu
di
ed
in
previ
ous
researc
h.
S
o
a
s
to
boos
t
the
r
obust
ness
a
nd
ef
fici
ency
of
the
model,
data
no
rmali
zat
ion
wa
s
ad
op
te
d
to
t
r
ansfo
rm
feature
s
to
be
on
a
simi
la
r
sc
al
e.
To
bette
r
unde
rstan
d
the
patte
r
ns
,
c
orre
la
ti
on
s,
a
nd
as
so
ci
at
ion
s
bet
ween
the
featu
res
a
nd
the
ta
r
get v
aria
ble, which
can
le
ad
to
b
et
te
r
predict
io
n
perf
orma
nce,
a d
at
a b
al
ancin
g
te
ch
nique w
as
exec
uted.
To
bo
os
t
the
pe
rformance
,
r
obus
t
ness,
a
nd
gen
e
rali
zat
ion
of
the
model,
we
ap
plied
a
n
op
ti
miza
ti
on
of
the
hype
rp
a
ramete
rs usin
g
B
a
yesi
an op
ti
miza
ti
on
in
s
uch a
wa
y
that t
he mo
de
l perfo
rms bet
te
r
on
unseen
dat
a.
Numer
ous
stu
dies
an
d
resea
r
ch
hav
e
bee
n
cond
ucted
on
f
li
gh
t
dela
ys
.
T
he
a
nalysis
a
xe
s
ad
dr
es
sed
the
ca
us
es
of
f
li
gh
t
delays
,
th
ei
r
c
onseq
ue
nc
es,
a
nd
meas
ures
to
a
void
the
m
from
diff
e
re
nt
perspecti
ves
.
T
he
inv
est
igati
ons
com
pr
ise
d
the
study
of
sta
ti
sti
cs,
n
et
w
ork
of
thi
ng
s
,
pro
ba
bi
li
ty
theo
ries,
operati
onal
rese
arch,
and
machi
ne
l
earn
i
ng.
Liu
et
al.
[4]
ap
plied
an
ec
onome
tri
c
m
odel
in
ord
er
t
o
pe
rform
a
n
e
m
pirical
a
na
lysis
of
flig
ht
act
ual
ai
rborne time
(
AA
T
)
i
n
the
U
S and C
hin
a.
Borsk
y
a
nd
U
nte
rb
e
rg
e
r
[5]
s
ug
gested a
d
i
ff
e
r
ence
-
in
-
diff
e
ren
c
e
f
rame
work
ba
s
ed
on
an
eco
nometri
c
a
nalysis
to
stu
dy
the
impact
of
sudd
e
n
c
ha
nges
in
mete
orolo
gical
co
nd
it
io
ns
o
n
dep
a
rtu
re d
el
ays
us
in
g
Un
it
e
d
Stat
es d
at
a
be
tween
Ja
nuar
y
2012
an
d
Se
ptembe
r
2017.
C
he
n
et
al
.
[6]
ai
me
d
to
stu
dy
ai
rc
r
aft
dela
y
distribu
ti
on
patte
r
ns
in
on
e
area
a
nd
de
monstrat
e
the
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
1038
-
1050
1040
impact
of
dela
ys
happe
ning
a
t
sever
al
ai
r
por
ts
us
in
g
a
var
ie
ty
of
vis
ualiz
at
ion
te
c
hn
i
qu
es
.
In
orde
r
to
for
ecast
the
li
kelihoo
d
of
ai
rline
delays
durin
g
ta
ke
-
off
a
nd
la
ndi
ng
op
e
rati
ons,
t
he
kernel
den
s
i
ty
f
un
ct
io
n
ha
s
bee
n
util
iz
ed
by
the
auth
ors
in
[
7]
.
Zen
g
et
al
.
[
8]
f
ocu
se
d
on
com
plex
netw
ork
th
eo
ry
an
d
the
ca
us
al
in
f
eren
ce
method
to
stu
dy
th
e
prop
a
gati
on
of
delay
s
a
nd
their
in
flue
nce
on
ai
r
traf
f
ic
co
ntro
l
syst
ems.
T
o
exa
min
e
how
extreme
weat
he
r
c
onditi
on
s
impact
punct
ua
li
ty
in
hi
gh
-
s
pe
ed
rail
an
d
a
vi
at
ion
ser
vices,
Che
n
a
n
d
Wa
ng
[
9]
util
iz
ed
both
da
ta
visu
al
iz
at
ion
an
d
sta
ti
sti
cal
analysis.
for
antic
ipati
ng
a
nd
assessi
ng
t
he
functi
onal
co
ndit
ion
of
the
ai
r
port
arr
ival
s
ys
te
m,
Ro
dr
í
gu
ez
-
Sa
nz
[
10]
sug
ges
te
d
a
tw
o
-
sta
ge
m
od
el
:
t
he
predict
io
n
pa
rt
us
in
g
a
pro
bab
il
ist
ic
Bayesian
n
et
wor
k
a
nd the
reli
abili
ty p
a
rt w
it
h a
M
ar
kov
c
hai
n
a
ppr
oach.
Stat
ist
ic
al
methods
are
ge
ne
rall
y
base
d
on
pr
ob
a
bili
ti
e
s
and
a
ppr
ox
im
at
e
meas
ur
e
m
ents,
w
hich
migh
t
res
ult
in
misl
eadi
ng
outc
om
es.
M
achi
ne
le
a
rn
i
ng
ha
s
the
ad
va
ntag
e
of
res
ulti
ng
i
n
i
ncr
ease
d
ac
cur
ac
y
and b
ei
ng ab
le
to addre
ss e
nor
mous
qu
a
ntit
ie
s of dat
a, a
utomat
ion
,
and
w
orkin
g bett
er
w
it
h
unstr
uctu
re
d data,
a
ccordin
g
to
Alla
et
al.
[11]
.
Q
u
et
al.
[
12]
us
ed
a
dee
p
le
a
rn
i
ng
te
c
hn
i
qu
e
for
ass
essing
a
nd
pro
je
ct
ing
ai
rcr
aft
delay
s
.
T
he
pr
e
dicti
on
accu
rac
y
was
8.7
per
ce
ntage
points
hi
gh
e
r
c
ompare
d
with
the
tra
diti
on
al
machine
le
a
rni
ng
te
ch
nique.
So
as
to
pr
e
dict
delaye
d
domesti
c
flig
hts
op
e
rated
by
Amer
ic
a
n
Ai
rline
s
,
Chak
rab
a
rty
[13]
hav
e
de
ploy
ed
a
gradie
nt
boos
ti
ng
cl
assifi
er
model
with
data
sampli
ng
and
hype
rp
a
ra
mete
r
tun
in
g.
The
sugg
e
ste
d
met
hod
ha
s
acc
ompli
sh
e
d
an
accu
ra
cy
of
85.
73
%
.
Fo
r
ai
r
port
del
ay
pr
e
dicti
on,
a
long
sh
ort
te
rm
me
mory
(LST
M)
neural
net
wor
k
f
rame
wor
k
usi
ng
histo
rical
fligh
t
data
fro
m
seve
ral
ai
rpor
ts
in
t
he
U.
S
.
from
2015
t
o
2018
has
bee
n
pro
pose
d
by
resear
cher
s
in
[8]
.
A
ccordin
g
to
the
ex
pe
rime
ntal
resu
lt
s,
the s
uggested
techn
i
qu
e
out
pe
rforms e
xisti
ng meth
ods r
e
ga
r
ding
reli
abili
ty and
pr
eci
sio
n.
Bi
sandu
et
al.
[14]
hav
e
re
co
mmende
d
a
de
ep
rec
urren
t
ne
ur
al
net
work
(
DRN
N)
m
odel
in
o
rd
e
r
to
analyze
an
d
sol
ve
fli
gh
t
dela
y
predict
io
n
iss
ue
s.
T
he
sug
ges
te
d
met
hod’s
e
ff
ic
ie
nc
y
a
nd
c
ompu
ti
ng
ti
me
wer
e
com
par
e
d
to
e
xisti
ng
ben
c
hma
r
k
a
ppr
oac
he
s.
In
order
to
hel
p
i
n
decisi
on
-
ma
king
an
d
pre
dicti
ng
ai
r
traf
fic
delays
,
Ni
bar
e
ke
an
d
Laassi
r
i
[1
5]
perform
ed
anal
ys
is
on
a
fligh
t
da
ta
set
us
in
g
decisi
on
tr
ee
,
n
aï
ve
Ba
yes,
and
li
nea
r
re
gr
ession
.
Th
e
cal
culat
ion
an
d
c
omparis
on
of
accurac
y,
e
rror,
an
d
sco
re
me
tric
s
hav
e
ge
ne
rate
d
decisi
on
trees
as
the
best
m
od
el
a
nd
n
aï
ve
Ba
yes
as
th
e
wea
kest
on
e
.
H
uo
et
al.
[
16]
have
c
ho
s
en
fi
ve
methods
,
w
hic
h
are
naï
ve
bayes,
lo
gisti
c
re
gressi
on,
k
-
nea
r
est
neig
hbors,
rand
om
forest,
and
decisi
on
t
r
ees
to
forecast
ai
rcr
a
f
t
delays
at
H
ong
K
ong
In
te
r
national
Air
port.
F
or
est
imat
ing
ai
rcr
a
ft
de
pa
rtur
e
dela
ys
,
Kh
a
n
et
al.
[
17]
ha
ve
propose
d
a
new
model
us
i
ng
var
i
ous
ne
ur
al
netw
ork
a
lgorit
hm
s
c
ombine
d
wit
h
dif
fer
e
nt
samplin
g
te
ch
ni
qu
es
.
By
e
xa
minin
g
var
ia
bl
es
that
are
in
r
el
at
ion
with
de
la
ys
su
c
h
as
w
eat
her
data,
operati
on
s
in
ai
r
ports
gr
ound,
ca
pacit
y
f
or
dema
nd
a
nd
fl
ow
c
on
t
ro
l
qual
it
ie
s,
Esma
e
il
zadeh
an
d
Mokhta
rim
ousav
i
[
18]
dev
el
op
e
d
a
s
uppo
rt
vecto
r
m
achine
(SVM)
model
to
in
ves
ti
gate
the
nonlinear
c
onnecti
on
of
the
ai
r
del
ays
in
the
thre
e
bi
ggest
ai
rpor
ts
i
n
Ne
w
Y
ork
ci
ty.
Alla
et
al.
[19
]
ex
pe
rim
ented
with
gr
a
dient
boos
ti
ng,
li
near
regressio
n,
e
xtreme
gradie
nt
boos
ti
ng,
ra
nd
om
f
or
e
st
,
a
nd
decisi
on
trees
al
gorith
ms
in
order
to
foreca
st
the
arr
ival
ti
me
of
a
s
pecific
flig
ht.
rand
om
f
orest
was
the
m
os
t
s
uc
cess
fu
l
model,
with
t
he
bi
gg
est
accu
racy
of
98.11%
compa
red to t
he othe
r
ones.
The
c
onte
nt
of
this
pa
per
i
s
arr
a
nged
as
outl
ined:
s
ect
ion
2
e
xami
ne
s
the
st
ud
ie
s
an
d
ef
f
or
ts
cond
ucted
in t
he
area
of
flig
ht
d
el
ay
est
imat
ion. Secti
on 3
hi
gh
li
ghts t
he
m
et
hod
p
r
ov
i
ded in this r
esea
rc
h,
the
al
gorithms
us
e
d,
as
well
a
s
al
l
the
featu
res
a
nalyze
d
an
d
pr
opos
e
d
so
as
t
o
boost
the
me
thod’s
perform
ance.
Sect
ion
4
pro
vi
des
the
em
piri
cal
resu
lt
s
of
the
s
uggeste
d
t
echn
i
qu
e
as
w
el
l
as
the
ti
me
com
plexity
co
mputed
in this
w
o
r
k.
S
ect
ion
5 discus
ses the c
oncl
usi
on
,
v
ie
wpoint
s,
a
nd possi
ble
fu
t
ur
e
de
velo
pme
nts.
2.
METHO
D
2.1.
Problem
st
at
e
ment
Passen
ge
rs
ma
y
e
xp
e
rience
s
ever
e
dif
ficult
y
as
a
res
ult
of
fligh
t
delays
,
su
c
h
as
misse
d
co
nn
ect
io
ns
,
missed
a
ppoint
ments,
a
nd
un
plan
ned
over
ni
gh
t
sta
ys
.
Airl
ines
can
proac
ti
vely
al
ert
cus
tomers
a
nd
pr
ov
i
de
al
te
rn
at
e
ch
oices,
su
c
h
as
re
-
bookin
g
on
a
nothe
r
ai
rcr
a
ft,
i
f
dela
ys
are
predict
ed
in
a
dva
nce.
Fli
ght
de
la
ys
ca
n
al
so
be
c
os
tl
y
f
or
ai
rlines,
r
esulti
ng
in
i
nc
reased
us
e
of
fu
el
,
c
us
to
mer
c
harges
pai
d
for
the
an
noye
d
an
d
dissati
sfied
pas
sen
ger
s
,
stuffi
ng
overti
me
,
an
d
ot
her
operati
on
al
c
os
ts.
Airl
ines
can
ta
ke
preve
ntive
act
io
ns
t
o
minimi
ze
thes
e
costs,
s
uc
h
as
adju
sti
ng
fl
igh
t
sc
hedules
or
opti
mizi
ng
gro
und
operat
ion
s
,
w
hich
ca
n
al
so
impro
ve
c
us
to
mer
sat
isfact
io
n
a
nd
lo
yalty
.
I
n
orde
r
t
o
a
vo
i
d
sa
fety
iss
ues
e
ngen
de
re
d
by
the
stre
s
s
an
d
fati
gu
e o
f
pilot
s
an
d
the
c
rew
in
ge
ner
al
,
a
fte
r
ex
per
ie
ncin
g
delays
,
ai
rlines
may
ta
ke
prec
autions
to g
ua
r
antee
that
their
c
re
w
mem
ber
s
are
well
-
reste
d
a
nd
that
al
l
safet
y
protoc
ols
a
re
f
ollow
e
d
i
f
tra
ff
ic
delays
are
known
in
ad
van
c
e.
F
or
this
reas
on,
we
deci
ded
i
n
our
stu
dy
to
de
velo
p
a
pre
dicti
ve
model
th
at
al
lows
pas
s
eng
e
rs,
ai
rlines,
a
nd
a
irp
or
t
ma
na
gers
to
be
a
war
e
of
dela
ys
s
o
as
to
ta
ke
pr
oa
ct
ive
act
ion
s
and
meas
ur
e
s.
The
ob
je
ct
ive
is
to
pro
vid
e
tra
vele
rs
with
high
fl
exibili
ty
a
nd
pe
ace
of
mi
nd
wh
il
e
al
s
o
assi
sti
ng
ai
rline
s
i
n
m
ore
su
ccess
fu
ll
y
di
recti
ng
t
heir
operati
ons
a
nd
i
mpro
ving
c
us
tome
r
sat
isfact
ion.
F
ur
t
her
m
or
e,
it
will
help
ai
rpor
t
a
uthoriti
es
a
nd
le
ader
s
in
the
decisi
on
-
ma
ki
ng
pr
ocess.
W
e
sta
rted
by
e
xt
racti
ng
histo
rical
fligh
t
i
nfo
r
mati
on
from
the
BTS
data
base.
T
he
data
un
derwen
t
meti
culo
us
prepa
rati
on
,
se
gm
e
ntati
on
an
d
nor
mali
zat
ion
,
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
Airp
or
t i
nfr
as
t
ru
ct
ur
e
an
d
r
unway
p
reci
si
on a
id
s for
forec
ast
ing
fl
ig
ht a
r
riv
al d
el
ays
(Haj
ar
All
a)
1041
ens
ur
in
g
it
s
re
adiness
f
or
e
xa
minati
on.
F
rom
the
dataset
a
t
hand,
we
de
ri
ved
dif
fer
e
nt
a
tt
ribu
te
s
that
de
scribe
the
perf
orman
ce
of
e
very
fli
gh
t.
T
o
boos
t
the
accu
rac
y
of
our
m
od
el
a
nd
as
fa
r
as
we
know
,
we
s
ug
gested
novel
feat
ur
es
that,
acco
rd
i
ng
to
ai
r
tra
nsporta
ti
on
orga
nizat
ion
s
a
nd
ass
ociat
ion
s
,
are
extremel
y
im
porta
nt
and
le
a
d
t
o
ai
r
traff
ic
dela
ys
.
A
fter
t
hat, w
e
seg
me
nted
the
f
inal data
as
f
ol
lows
: 7
0%
f
or
trai
ni
ng
a
nd 30%
f
or
te
sti
ng
.
We
uti
li
zed
the
m
ulti
la
yer
pe
rcep
t
ron
(
MLP
)
to
tra
in
our
model.
To
e
nsure
ef
fe
ct
ive
trai
ni
ng
of
the
pro
po
se
d
s
ys
t
e
m,
hy
perpara
m
et
er
tun
i
ng
was
adopted
us
in
g
the
Ba
yesian
o
ptimi
zat
ion
a
ppr
oac
h.
We
de
ci
ded
to
perf
or
m
a
da
ta
sam
plin
g
usi
ng
the
s
ynth
et
ic
min
or
it
y
oversa
mp
li
ng
te
chn
i
qu
e
(SM
O
TE)
co
mb
i
ne
d
wit
h
Tom
e
k
li
nk
s
.
We
e
xten
sivel
y
e
xami
ned
a
nd
as
sesse
d
t
he
eval
uation
of
the
pro
pose
d
model'
s
perf
ormance
against
dif
fer
e
nt
met
rics.
We
en
de
d
t
he
st
ud
y
with
a
c
ompl
ex
c
omp
utati
on.
Fig
ure
2
off
ers
a
s
um
m
ar
y
of
t
he
arch
it
ect
ure
of
the g
e
ne
ral str
uctu
re
of
our
a
ppr
oach.
Figure
2. Se
quentia
l
w
or
kf
l
ow
of
t
he
pro
posed mo
del
2.2.
Data
c
ollec
tio
n
Historical
flig
ht
rec
ords
f
or
non
-
sto
p
dome
sti
c
fligh
ts
wit
hin
the
U
nited
Stat
es
for
the
year
2019
wer
e
obta
ine
d
from
t
he
BT
S
[
2]
.
Data
on
run
way
dime
ns
io
ns
,
ju
nctio
ns
,
a
nd
oth
er
per
ti
ne
nt
detai
ls
from
about
10
6
ai
r
ports
a
cr
os
s
the
U.S.
wer
e
retrieve
d
f
rom
the
Fe
deral
Av
ia
ti
on
A
dmi
nistrati
on
(
F
AA
)
[
20]
ai
rpor
t
databas
e.
Mo
reover
,
i
nformat
ion
on
nav
i
gationa
l
ai
ds
,
e
quipme
nt,
an
d
facil
it
ie
s
at
these
ai
r
ports
wa
s
acce
ssed
from
the
ai
r na
vig
at
ion
data
base
we
bs
it
e
[
21]
.
2.3.
Data
p
rep
roce
ssing
Var
i
ou
s
data
minin
g
an
d
m
achine
le
ar
ning
meth
odol
og
i
es
can
be
util
iz
ed
to
un
c
ove
r
intrig
uing
insig
hts
an
d
pa
tt
ern
s
f
rom
e
xtensi
ve
data
ba
ses
[
22]
.
T
he
data
prep
r
oce
ssing
operati
ons
pr
e
par
e
t
he
input
dataset
f
or
the
fo
ll
owin
g
data
minin
g
act
io
ns
.
T
he
y
al
so
con
t
rib
ute
to
e
nh
a
ncin
g
a
nd
boos
ti
ng
the
a
ccur
ac
y
and
performa
nc
e
of
mac
hin
e
le
arn
in
g
syst
ems,
pa
rtic
ularl
y
i
n
cl
assifi
cat
ion
,
acc
ordin
g
to
[
23]
.
In
this
pa
pe
r,
we
a
dopted
tw
o pr
e
processi
ng tech
niques:
da
ta
clea
ning a
nd
data no
rmali
zat
ion
.
In
d
at
a
c
le
ani
ng
,
t
he
pr
ocess
of
data
cl
eani
ng
c
onsist
s
of
t
he
el
imi
nation
of
duplica
te
el
ements,
the
handlin
g
of
mi
ssing
i
nf
ormat
ion,
the
co
rr
ect
i
on
of
i
nconsist
ent
values
,
an
d
the
pro
per
f
or
matt
ing
of
data
,
s
o
it
can
be
rea
dy
a
nd
prepa
red
t
o
be
anal
yze
d
a
nd
us
e
d,
acc
or
ding
to
[11]
.
It
is
an
imp
ort
ant
phase
in
t
he
data
analysis p
r
oces
s.
It
gua
ran
te
es
that
the
co
ncl
us
io
ns
o
btaine
d
f
rom
the d
at
a
are
co
rr
ect
a
n
d
de
pe
nd
a
ble.
In d
at
a
n
ormal
iz
at
ion
,
n
ormal
iz
at
ion
adj
us
ts
the
ra
ng
e
of
at
tribu
t
e
val
ues
t
o
fit
withi
n
a
ne
w
scal
e.
T
his
ki
nd
of
appr
oach
is
cr
ucial
f
or
cl
ass
ific
at
ion
meth
od
s
beca
us
e
it
en
ha
nces
the
le
ar
ning
proc
ess
a
nd
e
nsur
es
th
a
t
at
tribu
te
s
with
higher
val
ues d
o no
t
dominate
those
with lo
w
er
values, as
hi
gh
li
ghte
d i
n
[
23]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
1038
-
1050
1042
2.4.
Fe
at
ure
s
sel
ection
Database
f
eat
ur
e
s
us
e
d
t
o
pr
e
dict
arr
i
val
delays
f
or
domesti
c
flig
hts
op
e
rated
i
n
Un
it
es
Stat
e
s
A
ir
ports,
we
use
d
data
for
th
e
yea
r
2019.
We
e
xtracted
t
he
sta
ti
sti
cal
inf
ormat
ion
f
r
om
the
BTS
[2]
.
T
he
dataset
co
mpri
sing
al
l
the
relevan
t
i
nfo
rmat
ion
a
bout
the
f
li
gh
t
is
s
umma
rized
i
n
Ta
ble
1.
Pro
po
se
d
f
e
at
ur
es
are
us
e
d
by
i
nter
national
organ
iz
at
io
ns
a
nd
ass
ociat
ion
s
are
w
orkin
g
to
en
han
ce
ai
r
travel
sa
fety
and
eff
ic
ie
nc
y, as
well
as to
imp
r
ov
e
the
passe
nger
experie
nce
and fi
gh
t a
gain
st fli
gh
t
dela
ys
.
Table
1.
A
naly
sis of da
ta
base
featur
e
s
Featu
re
Categ
o
ry
Descripti
o
n
Day
of
m
o
n
th
Nu
m
eric
al
The d
ay
of the
m
o
n
th
in wh
ich
the fli
g
h
t was execu
ted
Day
of wee
k
Nu
m
eric
al
The d
ay
of the w
ee
k
du
ring
which
the
tr
ip
was
ex
ecut
ed
Carrie
r
co
d
e
Nu
m
eric
al
The airlines
d
esig
n
atio
n
Tail
n
u
m
b
er
Nu
m
eric
al
The
airlines
regis
tr
atio
n
/m
atriculatio
n
Flig
h
t nu
m
b
er
Nu
m
eric
al
The n
u
m
b
er
o
f
the
fligh
t
Origin
Categ
o
rical
The airpo
rt
o
f
orig
in
Destin
atio
n
Categ
o
rical
The airpo
rt
o
f
des
t
in
atio
n
CR
S_
DEP
Nu
m
eric
al
The p
rog
rammed d
ep
arture
ti
m
e
Actu
al_
DEP
Nu
m
eric
al
The true dep
artur
e
tim
e
DEP
d
elay
Bin
ary
1
if
the fligh
t is del
ay
ed
on
dep
arture
0
if
no
t
CR
S_
ARR
Nu
m
eric
al
The p
rog
rammed a
rr
iv
al ti
m
e
ARR
delay
Bin
ary
1
if
the fligh
t is del
ay
ed
on
ar
rival, 0 i
f
n
o
t
(T
h
e dep
en
d
en
t variab
le in o
u
r
r
esear
ch
)
Distan
ce
N
u
m
eric
al
The d
istan
ce in
m
i
les b
etween th
e ai
r
p
o
rt
o
f
o
rigin
and
the airpo
rt
o
f
d
estin
atio
n
The
IC
AO
[1]
ha
s
est
a
blish
ed
sta
ndar
ds
a
nd
recomme
nded
pr
act
ic
es
(S
AR
Ps)
f
or
ai
rlines
a
nd
ai
rpor
ts
t
o
gua
ran
te
e
sa
fe
a
nd
ef
fici
ent
op
e
rati
on
s
.
Air
tr
aff
ic
ma
na
gem
ent,
ai
r
port
operati
on
s
,
a
nd
a
irli
ne
safety
a
re
am
ong
the
to
pics
c
ov
e
re
d
by
t
hes
e
SA
RPs
.
T
he
ICA
O
[
1]
al
so
colla
borates
with
mem
be
r
sta
te
s
t
o
pu
t
these
guide
li
nes
int
o
act
io
n,
imp
rove
the
safet
y
a
nd
e
ff
i
ci
ency
of
ai
r
tr
ans
port,
a
nd
m
anag
e
fligh
t
de
la
ys.
The
E
urop
ea
n
Un
i
on
(E
U)
ha
s
propose
d
a
"
passe
ng
e
r
ri
gh
t
s"
po
li
cy
t
hat
e
sta
blishes
guid
el
ines
for
ai
rlines
to
fo
ll
ow
in
t
he
case
of
ai
rc
raf
t
dela
ys
,
c
ancell
at
ion
s,
or
refuse
d
bo
ard
i
ng.
Passe
ng
e
rs
a
re
enti
tl
ed
to
com
pensat
ion,
assist
ance,
an
d
re
fun
ds
in
par
ti
cula
r
case
s.
Air
ports
Co
un
ci
l
I
nter
national
(
ACI
)
[
24]
has
create
d
t
he
"ai
rpor
t
ser
vice
qual
it
y"
(
ASQ
)
pro
gr
am
to
ass
ess
co
nsume
r
con
te
ntme
nt
w
it
h
ai
rpo
rt
ser
vi
ces.
The
i
niti
at
ive
s
olici
ts
passe
nger
feedbac
k
on
ma
ny
areas
of
the
ai
rpor
t
ex
per
ie
nce,
su
c
h
as
chec
k
-
in,
se
cur
it
y
,
bo
a
r
ding,
on
-
ti
me
ar
rival,
an
d
flig
ht
dela
ys
.
Air
ports
ma
y
e
mp
lo
y
t
his
in
f
ormat
ion
to
im
p
r
ove
the
pas
s
eng
e
r
exp
e
rience
by
identif
ying
a
re
as
f
or
im
pro
ve
ment.
T
hroug
h
it
s
aw
ar
ds
f
or
the
yea
r
20
22
in
the
U
S,
the
F
AA
[20]
has
e
sta
blishe
d
t
he
ai
r
por
t
imp
roveme
nt p
r
ogram
(
A
IP)
, w
hic
h
fun
ds
a
num
be
r
of
init
ia
ti
ves
an
d
pro
je
ct
s
su
c
h
as
t
he
bu
il
di
ng
of
ne
w
a
nd
upgr
a
ded
ai
r
port
in
fr
ast
r
uctu
re,
r
epairs
t
o
run
way
s
a
nd
ta
xi
way
s
,
mainte
na
nce
of airfiel
d co
mpo
nen
ts
su
c
h
a
s li
gh
ti
ng
or sig
ns, an
d
t
he
purc
ha
se of air
port e
qu
i
pm
e
nt.
In
order
t
o
me
et
the
exc
essiv
e
grow
t
h
in
Br
azi
li
an
pa
ssen
ge
r
t
raffic
a
nd
t
he
extra
dema
nd
ge
ner
at
ed
by
the
World
Cup
2014,
a
lot
of
project
s
wer
e
e
sta
blish
ed
no
ta
bly
t
he
co
ns
tr
uction
of
a
new
ai
rpo
r
t
in
the
nearb
y
to
wn of
Sã
o G
onçal
o d
o Amara
nte,
w
hich was
desi
gnat
ed
t
o
se
rv
e
the cit
y of Nata
l
[25]
. Acco
rd
i
ng to
Kenne
dy
[
26]
,
intersect
ions
betwee
n
r
unw
ays
are
e
xpect
ed
to
e
xp
e
rien
ce
add
it
io
nal
de
la
ys
in
the
N
at
ion
al
Airs
pace
S
ys
te
m
as
a
res
ult
of
i
ncr
ease
d
w
ai
t
per
io
ds
.
T
he
need
t
o
que
ue
wh
il
e
wait
in
g
f
or
a
n
i
nte
rs
ect
ing
runway
to
cl
ea
r
mi
gh
t
ca
us
e
consi
der
a
ble
de
la
ys
.
Delays
c
an
al
s
o
be
ca
use
d
by
wait
in
g
t
o
tra
verse
on
e
of
the
runways
wh
il
e
ta
xiin
g.
T
he
I
nter
national
Air
Tra
nsport
As
so
ci
at
ion
(IAT
A)
[
27]
has
la
un
c
he
d
the
"
A
irp
or
t
Coll
aborati
ve
Decisi
on
-
M
a
kin
g
(
ACD
M
)
"
i
niti
at
ive
to
opti
mize
the
inf
ormat
ion
e
xc
hange
bet
wee
n
ai
r
ports
,
ai
rlines
a
nd
ai
r
traf
fic
c
on
t
ro
l,
with
the
obje
c
ti
ve
of
minimi
zi
ng
f
li
ght
delays
a
nd
e
nhanci
ng
the
use
of
a
irp
or
t
resou
rces
s
uc
h
as
ai
rcr
a
ft
gates
a
nd
pa
r
king.
Acc
ordi
ng
to
t
he
inte
rn
at
io
nal
fe
de
rati
on
of
Air
Traffic
Con
tr
ollers
’
A
sso
ci
at
ions
(
IFATCA
)
[
28]
,
a
mar
ked
incre
ase
in
the
num
ber
of
ap
proac
h
cat
eg
or
ie
s
stud
ie
d
has
occ
urre
d
over
the
past
fe
w
yea
rs.
T
his
i
s
m
os
tl
y
ca
us
e
d
by
the
a
pp
li
c
at
ion
of
cutti
ng
-
e
dge
te
c
hnol
og
ie
s
includi
ng
t
he
GP
S
a
nd
t
he
R
NAV.
S
uc
h
te
chn
i
qu
e
s
al
lo
w
pilots
t
o
la
nd
with
bette
r
pr
e
ci
sion
an
d
c
onfide
nce
,
lowe
rin
g
t
he
po
s
sibil
it
y
of
missed
ap
proa
ches
or
go
-
a
r
ound
proce
dur
es
by
giv
in
g
pr
eci
se
a
nd
re
li
able
gu
i
dan
ce
and
posit
ion i
nfo
rma
ti
on
, w
hich
c
an
r
e
du
ce
the
occ
urren
ce
of
flig
ht d
el
a
ys
.
The
ICA
O
[
1]
has
produce
d
ai
rpo
rt
desig
n
sta
ndar
ds
that
rec
ommen
d
minimu
m
r
unway
le
ngths
dep
e
ndin
g
on t
he
siz
e an
d
ty
pe
o
f
ai
rc
raf
t e
xpect
ed
to
opera
te
in
the airp
ort
. Th
ese sta
nd
a
rd
s ca
n
c
on
t
rib
ute to
gu
a
ra
nteei
ng
that
ai
r
ports
a
r
e
eq
uippe
d
to
acce
pt
a
wide
range
of
ai
rc
ra
ft,
mi
nimizi
ng
the
c
ha
nce
of
dela
ys
du
e
to
ca
pacit
y
li
mit
s.
Simi
la
r
ly,
the
I
AT
A
[
27]
has
pro
duc
ed
the
A
DR
M
,
w
hich
offe
rs
r
ecomme
ndat
io
ns
on
ai
rpor
t
plan
ning
an
d
de
velo
pme
nt,
i
nclu
ding
r
unwa
y
le
ng
th
an
d
oth
e
r
in
fr
ast
r
uctu
re
re
qu
i
reme
nts.
Airpor
ts
can
e
nsure
that
th
ei
r
facil
it
ie
s
are
e
nh
a
nce
d
f
or
safet
y
a
nd
e
ff
ic
ie
nc
y
by
fol
lowing
t
hese
s
ta
nd
a
rd
s
,
wh
ic
h
ca
n
help
re
du
ce
fli
gh
t
dela
ys
a
nd
pro
vid
e
a
bet
te
r
pa
ssen
ge
r
exp
e
rience
.
T
hi
s
ex
plains
w
hy
we
opte
d
to
sta
y
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
Airp
or
t i
nfr
as
t
ru
ct
ur
e
an
d
r
unway
p
reci
si
on a
id
s for
forec
ast
ing
fl
ig
ht a
r
riv
al d
el
ays
(Haj
ar
All
a)
1043
fo
c
us
e
d
on
a
chievin
g
the
ob
je
ct
ives
a
nd
po
li
ci
es
est
ablishe
d
by
th
e
internati
onal
orga
nizat
ion
s
and
associat
ions
previ
ou
sl
y
e
xpla
ined
by
crea
ti
ng
ne
w
feat
ur
es
that
f
ulfi
ll
the
nee
ds
i
n
te
rms
of
ai
rport
infr
a
struct
ur
e
and
buil
dings,
th
e
mane
uveri
ng
a
rea
of
the
aerodr
om
e
(ru
nw
a
ys
,
ta
xiwa
ys
,
i
ntersecti
ons,
a
n
d
gates,)
,
a
nd
th
e
ai
ds
us
e
d
to
op
e
rate
the
fligh
ts
safely
an
d
eff
ic
ie
ntly
.
Ta
ble
2
outl
ines
t
he
feat
ur
e
s
tha
t
were
pro
po
se
d
i
n
thi
s stu
dy.
Table
2.
Desc
ription o
f pro
pos
ed feat
ur
e
s
Featu
re
Categ
o
ry
Descripti
o
n
Ru
n
way
L
en
g
th
Nu
m
eric
al
The d
istan
ce of the
m
o
st u
sed
r
u
n
way
Nu
m
b
er
o
f
Ru
n
way
s
Nu
m
eric
al
The n
u
m
b
er
o
f
r
u
n
way
s in
a
sp
ecific
airpo
rt
Ru
n
way
s Intersect
io
n
s
Nu
m
eric
al
The n
u
m
b
er
o
f
r
u
n
way
intersectio
n
s
p
o
ts in
a
sp
ecific
a
irpo
rt
Ru
n
way
s Precisio
n
Rate
Bin
ary
1
if
the run
way
is
p
recisio
n
-
aid
s eq
u
ip
p
ed
(
IL
S,
RNAV
,
and
GNSS)
0
if
no
n
-
p
recisio
n
-
aid
s eq
u
ip
p
ed
(
VOR, DM
E,
NDB,
an
d
L
OCAT
OR)
Nu
m
b
er
o
f
Gates
Nu
m
eric
al
The n
u
m
b
er
o
f
air
p
lan
e
p
arkin
g
in
a
sp
ecific
ai
rpo
rt
Nu
m
b
er
o
f
T
e
rm
in
als
Categ
o
rical
The n
u
m
b
er
o
f
pas
sen
g
er
ter
m
in
als in
a
sp
ecific
air
p
o
rt
2.4.1.
Runw
ay
leng
th
The
le
ngth
of
a
r
unwa
y
can
aff
ect
flig
ht
de
la
ys
i
n
dif
fer
e
nt
ways.
Alth
ough
lo
ng
e
r
r
unway
s
al
lo
w
for
la
r
ge
r
a
nd
heav
ie
r
ai
rc
raf
t
to
ta
ke
off
an
d
la
nd,
wh
ic
h
can
i
ncr
ease
th
e
capaci
ty
a
nd
dema
nd
at
a
n
a
irp
or
t
,
it
al
so
means
a
lo
ng
e
r
ti
me
to
vacate
a
nd
cl
ea
r
t
he
r
unwa
y,
wh
ic
h
causes
dela
ys
for
f
ollow
i
ng
fligh
ts.
Howe
ver,
if
a
runway
is
to
o
sh
ort
f
or
a
par
t
ic
ular
ai
rcr
a
ft,
the
la
nd
i
ng
or
ta
ke
-
off
distan
ce
will
al
so
be
sh
ort
,
and
it
can
le
a
d
to
a
run
way
e
xc
ur
si
on.
As
a
preve
ntive
act
io
n,
that
ai
rc
raf
t
may
need
t
o
be
div
erte
d
to
a
nothe
r
ai
rpor
t,
ca
us
in
g
a
delay
.
A
ddit
ion
al
ly
,
i
nclement
weat
her
ca
n
al
s
o
cau
s
e
flig
ht
dela
ys,
par
ti
cula
rly
if
t
he
run
w
ay
is
no
t
l
ong
e
no
ugh
for
a
n
ai
rc
raf
t
to
safely
ta
ke
off
or
la
nd
i
n
poor
visibil
it
y
c
ondi
ti
on
s,
e
sp
eci
al
ly
i
f
the ru
nw
a
y
is
wet or sli
ppe
ry, which
r
es
ults
in a lon
ger b
ra
king
distance.
2.4.2.
Nu
m
ber
of
r
un
w
ays
The
numb
e
r
of
r
unwa
ys
at
an
ai
r
port
can
hav
e
a
n
im
pa
ct
on
fligh
t
delays
.
Ha
ving
nume
r
ou
s
runways
al
l
ows
an
ai
r
port
to
serv
e
m
or
e
ai
r
traff
ic
a
nd
re
duce
delays
ca
use
d
by
c
onge
sti
on
,
ena
blin
g
a
ircraft
to
ta
ke
-
off
a
nd
ar
rive
simult
a
neousl
y,
t
her
e
by
inc
reasin
g
the
ai
r
port
’s
ov
erall
capaci
ty
.
Th
ough
a
n
i
nc
reased
numb
e
r
of
r
unway
s
ca
n
al
so
resu
lt
i
n
a
c
ompli
cat
ed
a
rch
it
ect
ur
e
f
or
the
a
irp
or
t.
I
n
ge
neral
,
the
m
or
e
runw
a
ys
an
ai
r
port
has
and
t
he
m
ore
com
plex
it
s
la
yout,
the
gr
eat
er
the
pote
ntial
fo
r
del
ays
du
e
t
o
r
unwa
y
intersect
io
ns
.
2.4.3.
Runw
ay
s
i
nt
ersec
tions
Fli
gh
t
dela
ys
c
an
be
increa
se
d
with
r
unwa
y
intersect
io
ns
.
I
n
ai
r
ports,
run
way
i
ntersecti
ons
are
spots
wh
e
re
tw
o
or
more
r
unwa
ys
cro
ss
or
j
oin
.
The
y
ca
n
ca
use
traf
fic
dela
ys
if
ai
rc
raf
t
a
re
no
t
a
ble
to
ta
ke
off
or
la
nd
on
the
int
ersecti
ng
run
w
ays
at
the
sa
m
e
ti
me,
due
to
safety.
Let
us
t
ake
t
he
exa
mpl
e
in
Fig
ur
e
3
of
tw
o
intersect
in
g
r
unwa
ys
,
R1
0
a
nd
R0
9.
T
o
t
ake
off
on
R
10,
ai
rc
raf
t
A
mu
st
c
ross
R
09,
bu
t
ai
rc
raft
B
is
auth
or
iz
e
d
t
o
l
and
on
R0
9.
F
or
that,
ai
rc
raf
t
A
is
force
d
t
o
wait
an
d
hold
po
si
ti
on
to
gi
ve
wa
y
to
ai
rcr
a
ft
B
for
safety
reas
ons.
Figure
3. Tra
ffi
c o
pe
rati
ons
on tw
o
i
ntersect
ing
r
unwa
ys
(Source:
[
29]
)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
1038
-
1050
1044
2.
4
.
4.
Runw
ay
s
p
reci
sion
r
at
e
Runwa
y
preci
sion
ai
ds
a
re
sy
ste
ms
a
nd
e
qu
i
pm
e
nt
t
hat
assist
pilots
in
na
vig
at
in
g
safely
a
nd
correct
ly
duri
ng
ta
ke
-
off
an
d
la
nd
in
g.
In
acc
ur
aci
es
in
the
i
nformat
ion
pro
vid
e
d
by
t
he
non
-
pr
eci
sio
n
r
unwa
y
ai
d
can
le
ad
t
o
er
r
or
s
in
a
ppr
oac
h
a
nd
la
nd
i
ng.
T
he
di
sc
rep
a
ncies
ca
us
e
go
-
ar
ound
s
an
d
miss
ap
proac
h
op
e
rati
ons th
at
contrib
ute to
fl
igh
t dela
ys
.
2.
4.
5
.
Nu
m
ber
of
g
ates
Fli
gh
t
de
la
ys
c
an
al
so
be
a
ff
e
ct
ed
by
the
nu
mb
e
r
of
gates
at
an
ai
rpo
rt.
I
t
is
crit
ic
al
and
man
dato
ry
for
a
n
ai
rpo
rt
t
o
ha
ve
a
s
uffici
ent
num
be
r
of
pa
rk
i
ng
s
pace
s
to
ma
nage
th
e
num
be
r
of
fli
gh
ts
that
are
pl
ann
e
d
to
ta
ke
-
off
an
d
la
nd.
I
f
t
her
e
are
not
e
noug
h
gates,
fligh
ts
may
ha
ve
t
o
wait
f
or
one
t
o
become
a
vaila
ble,
causin
g delays
.
2.
4.
6
.
Nu
m
ber
of
t
er
mi
na
ls
Air
ports
with
more
te
rmina
ls
te
nd
to
welco
me
more
flig
ht
s
and
passe
nge
rs.
T
his
can
le
ad
to
m
ore
congesti
on,
lo
nger
secu
rity
li
ne
s,
an
d
more
pote
ntial
f
or
del
ays.
Ne
ver
t
heless,
ai
r
ports
with
fe
w
te
r
minal
s
can
so
meti
mes
ex
pe
rience
flig
ht del
ays,
es
pecial
ly if
one
or m
ore are
u
ns
er
vice
able.
2.5.
M
ultil
aye
r
p
erce
p
tron
The
M
LP
is
a
popula
r
a
nd
basic
ne
ural
netw
ork
generall
y
use
d
f
or
cl
assifi
cat
io
n
pr
ob
le
m
s.
Ba
sic
al
ly,
it
is
a
fee
d
-
forw
a
r
d
ne
ur
al
netw
or
k
c
ompose
d
of
man
y
pe
rcep
t
r
on
[
11]
.
Wit
h
on
e
or
more
hi
dd
e
n
la
yer
s,
t
he
MLP
is
ge
ner
al
l
y
em
ployed
f
or
patte
r
n
rec
ogniti
on,
cl
ass
ific
at
ion
,
pr
e
di
ct
ion
,
an
d
functi
on
appr
ox
imat
io
n
[
30]
.
T
he
in
put
la
ye
r
nodes
receive
t
he
i
nput
data,
wh
il
e
the
ou
t
pu
t
la
ye
r
nodes
ge
ne
ra
te
the
netw
ork’s
pred
ic
ti
on
s.
T
he
ne
uro
ns
i
n
each
l
ayer
us
e
act
iva
ti
on
f
un
ct
io
ns
t
o
c
al
culat
e
a
w
ei
gh
te
d
s
um
of
their
inputs a
nd produ
ce
a
non
-
li
ne
ar
ou
t
pu
t.
The
M
LP
f
ollo
ws
the
pr
ocess
of
m
ulti
plica
tio
n,
s
um
mati
on,
an
d
act
ivati
on
us
ed
in
ne
ural
netw
ork
s
[11]
, as
expre
s
sed b
y
(
1):
=
(
∑
∗
+
)
=
0
(1)
w
he
re
ref
e
rs
t
o
the
−
ℎ
in
pu
t
w
he
re
ranges
fro
m
0
t
o
inputs
.
ind
ic
at
es
t
he
weig
ht
matri
ce
s
f
or
both
the
hidde
n
a
nd
outp
ut
la
ye
rs,
with
s
panni
ng
from
0
to
n
inputs.
de
no
te
s
the
bias
te
rm
.
re
presents
th
e
act
ivati
on
func
ti
on
.
sig
nifies
the outp
ut
valu
e.
M
ulti
la
yer
p
er
ceptr
on
is
wi
de
ly
use
d
f
or
a va
riet
y
o
f
a
ppli
cat
ion
s.
Alla
et
a
l.
[11]
use
d
a M
LP
ne
ural
netw
ork
with
s
el
ect
ive
trai
ning
f
or
the
predi
ct
ion
of
dela
ys
on
ar
rival.
To
est
imat
e
the
co
eff
ic
ie
nt
of
sof
t
so
il
consolidat
io
n,
Ph
am
et
a
l.
[31]
c
ombine
d M
LP a
nd
bi
og
e
ogra
phy
-
ba
sed
opti
miza
ti
on
(B
BO).
The
pro
pose
d
method,
M
LP
-
BB
O,
ha
d
t
he
bi
gg
e
st
pr
e
di
ct
ive
performa
nce
with
the
l
ow
est
r
oot
mean
squa
re
e
rror
(R
M
SE
)
of
0.397
co
mp
a
red
with
oth
e
r
models.
Mu
barek
a
nd
Ad
al
i
[
32]
a
dopte
d
a
M
LP
for
f
ra
ud
detec
ti
on
.
The
pro
pose
d
m
odel
re
vealed
t
hat
it
was
t
he
m
os
t
accurate,
with
the
great
est
d
e
gr
ee
of
accurac
y
of
99.
47%
us
i
ng
ni
ne
sel
ect
ed
fe
at
ur
es.
For
dr
ought
f
or
eca
sti
ng,
Z
ulifq
a
r
et
al.
[
33]
a
pp
li
e
d
an
d
te
ste
d
the
ML
P
in
seve
ral
cl
imat
ologica
l
sta
ti
on
s
sit
uate
d
in
the
no
rthern
area
an
d
Pa
ki
sta
n.
T
he
m
odel
was
able
to
pre
dict
dro
ught
c
onditi
on
s
with
di
ff
e
ren
t
ti
me
scal
es
an
d
highe
r
a
ccur
ac
y.
T
o
de
te
rmin
e
t
he
wa
rmin
g
and
cal
min
g
re
qu
i
reme
nts
of
energ
y
-
e
ff
ic
ie
nt
buil
dings,
X
u
et
al.
[34
]
ha
ve
util
iz
ed
a
nd
opti
mize
d
an
M
LP
method
us
i
n
g
diff
e
re
nt
opti
m
iz
at
ion
al
gorith
ms.
Ra
dh
a
kri
sh
na
n
et
al.
[35
]
ha
ve
de
velo
pe
d
a
n
M
LP
m
odel
f
or
pr
e
dicti
ng
mec
han
ic
al
ve
ntil
at
or
set
ti
ng
s
by
changin
g
the
hi
dd
e
n
la
yer
s
a
nd
c
ompari
ng
the
resu
lt
s.
T
he
be
st
model wa
s the
on
e
w
it
h t
hree
hidden
laye
rs.
In
the
prese
nt
s
tudy,
the M
L
P w
as
im
pleme
nt
ed
for
a
ssessin
g
a
nd f
oresee
in
g
t
he
occurr
en
ce
of f
li
gh
t
delays
us
in
g
ne
w
featu
res.
W
e
op
te
d
f
or
the
M
LP
for
ma
ny
reas
ons:
i)
Be
cause
M
LPs
hav
e
few
pa
ra
mete
rs,
they
ca
n
be
e
mp
lo
ye
d
by
in
div
id
uals
with
ou
t
pr
e
vious
e
xp
e
rience
,
an
d
their
impleme
ntati
on
te
c
hn
i
ques
are
easy
to
un
der
st
and
[36]
;
ii
)
MLPs
ha
ve
t
he
c
apab
il
it
y
t
o
be
util
iz
ed
acr
os
s
div
e
rse
fiel
ds
f
or
s
olv
in
g
a
va
riet
y
of
prob
le
ms
[
36]
;
ii
i)
M
LPs
serv
e
as
to
ols
for
discri
minat
ion
,
recog
niti
on
of
patte
r
ns
,
empirical
m
odel
ing
,
and
ma
ny
othe
r
a
pp
li
cat
io
ns
[
36]
;
i
v)
W
he
n
a
ppli
ed
t
o
simi
la
r
issue
s,
MLP
s
oft
en
ou
t
perform
sta
nd
a
r
d
sta
ti
sti
cal
app
r
oach
e
s
[36]
;
v)
W
hile
tra
di
ti
on
al
li
near
models
str
uggl
e
to
m
odel
data
with
nonl
inear
pro
per
ti
es,
MLPs
ca
n
ef
fec
ti
vely
capt
ur
e
both
li
nea
r
a
nd
no
nlinear
i
nteracti
ons
[36]
;
a
nd
vi)
MLPs
are
eff
ect
ive
in
e
xt
racti
ng str
uctu
r
al
o
r
p
at
te
r
n
c
ha
racteri
sti
cs fr
om
both
sta
ti
c and dy
namic
da
ta
[36]
.
To
c
reate
a
machine
le
ar
ni
ng
s
ys
te
m
,
weig
ht
pa
ram
et
ers
are
set
up
an
d
a
dju
st
ed
us
in
g
a
n
op
ti
miza
t
io
n
a
ppr
oach.
Accord
i
ng
to
t
he
st
udy
[
37]
,
this
act
ion
kee
ps
oc
currin
g
ti
ll
th
e
obje
ct
ive
functi
on
at
ta
ins
a
mini
mu
m
or
the
a
ccur
ac
y
reach
es
a
ma
xim
um
.
I
n
our
in
ve
sti
gation,
we
empl
oy
e
d
Ba
yesian
op
ti
miza
ti
on to
tun
e
the
hype
r
par
a
mete
rs.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
Airp
or
t i
nfr
as
t
ru
ct
ur
e
an
d
r
unway
p
reci
si
on a
id
s for
forec
ast
ing
fl
ig
ht a
r
riv
al d
el
ays
(Haj
ar
All
a)
1045
2.6.
Baye
sia
n
op
timi
z
at
i
on
The
hy
perpara
mete
r
opti
miza
ti
on
pa
rad
i
gm
i
nvolv
e
s
t
he
use
of
f
our
pe
rtin
ent
el
eme
nts:
a
n
e
sti
mator
(which
mi
gh
t
be
a
regressio
n
or
cl
assif
ic
at
ion
m
odel
with
s
om
e
sort
of
ob
je
ct
ive
f
unct
ion),
a
def
i
ne
d
searc
h
sp
ace,
a
te
c
hniqu
e
f
or
e
xplor
ing
or
opti
mizi
ng
hy
perpara
mete
r
c
ombina
ti
on
s,
a
nd
an
e
valuati
on
metr
ic
f
or
com
par
i
ng
the
ef
fecti
ven
es
s
of
var
i
ou
s
hype
r
par
amet
e
r
co
nf
i
gurati
on
s.
Ba
ye
sia
n
op
ti
miza
ti
on
m
od
el
s
com
pu
te
the
ne
xt
hype
rpara
mete
r
by
co
nsi
der
in
g
the
previo
us
outc
om
es
of
te
ste
d
hype
r
par
amet
e
r
values
avo
i
ding
num
erous
unnece
s
sary
eval
uatio
ns
.
As
a
c
ons
equ
e
nce,
the
Ba
yesian
ap
proach
can
fi
nd
the
best
hype
rp
a
ramete
r
c
ombinati
on
in
fe
wer
it
erati
on
s
tha
n
oth
e
r
opti
miza
ti
on
t
echn
i
qu
e
s
s
uc
h
as
ra
ndom
s
earch
and
gri
d
sea
rc
h.
Ba
yesian
opti
miza
ti
on
e
mp
lo
ys
t
wo
m
ai
n
co
mpo
nen
t
s
to
sel
ect
the
ne
xt
hype
r
paramet
er
config
ur
at
io
n:
a surr
ogat
e m
odel
and a
n
a
cq
uisit
ion
functi
on
[37]
.
2.6.1.
Surr
ogate m
od
el
:
G
aussi
an
p
roce
sse
s
The
s
urr
og
at
e
model
is
a
pp
li
ed
to
direct
the
searc
h
for
t
he
ta
rg
et
m
ode
l’s
global
opti
m
um.
T
he
most
commo
n
s
urrogate
m
od
el
for
obje
ct
ive
functi
on
m
od
el
i
ng
is
the
gaus
sia
n
pr
ocess
(
GP
)
,
wh
ic
h
f
ol
lows
a
normal
distri
buti
on
acc
ordin
g
t
o
(
2).
It
is
an
a
dvan
ced
pro
ba
bili
sti
c
mo
del
that
is
wi
dely
use
d
in
machine
le
arn
in
g for
re
gr
es
si
on a
nd classi
ficat
ion
pro
blems
[37
]
.
)
,
µ
|
N
(
y
D)
x,
|
(
^
^
2
=
y
p
(2)
wh
e
re
co
rr
es
ponds
to
the
hype
r
-
par
a
mete
r
co
nfi
gurati
on
sp
ace
,
a
nd
=
(
)
de
note
s
the
evaluati
o
n
ou
tc
om
e
f
or
e
a
ch hype
r
-
par
a
mete
r
value
,
²
is t
he
c
ovarian
c
e an
d
µ
t
he
m
e
an.
Af
te
r
ma
king
pr
e
dicti
on
s
,
t
he
subse
qu
e
nt
evaluati
on
points
ar
e
sel
ect
e
d
base
d
on
th
e
co
nf
i
den
ce
intervals
ge
ne
r
at
ed
by
the
BO
-
GP
m
od
el
.
Every
ne
w
dat
a
point
is
inc
orp
or
at
e
d
i
nto
t
he
dataset
,
a
nd
t
he
BO
-
GP
m
od
el
is
update
d
ac
cordin
gly.
T
hi
s
pr
ocess
is
reit
erated
s
eve
ral
ti
mes
unti
l
th
e
set
st
opping
crit
eria
are
met.
For
a
dataset
with
siz
e
,
the
BO
-
G
P
m
odel
has
a
ti
me
co
mp
le
xity
of
(
3
)
a
nd
a
s
pa
ce
co
m
plexity
of
(
2
)
.
A
sig
nific
ant
dr
a
wb
ac
k
of
the
BO
-
GP
model
is
it
s
c
ubic
ti
me
c
ompl
exity
c
oncer
ni
ng
the
num
ber
of
i
ns
ta
nces
,
wh
ic
h
im
pacts
it
s
scal
abili
ty
an
d
pa
rall
el
processi
ng
ca
pa
bili
ti
es.
Addi
ti
on
al
ly,
t
he
B
O
-
GP
model i
s mai
nly
desig
ne
d
f
or
op
ti
mizi
ng c
on
ti
nu
ous
v
a
riabl
es
[
37]
.
2.6.2.
Acq
uisi
t
ion
fu
nc
tion
To
sel
ect
the
nex
t
ca
ndidate
from
t
he
sea
r
ch
s
pace,
t
he
acqu
isi
ti
on
f
un
ct
ion
ca
n
be
de
scribe
d
to
mean
t
he
a
ntici
pated gai
n:
(
)
=
[
(
)
|
X,y
]
(3)
w
he
re
(
)
:
=
(
)
−
(
)
.
:
→
(4)
is
the
gain
f
or
unknow
n
s
ol
ution
s
.
I
n
ev
ery
lo
op,
a
n
add
it
io
nal
ca
ndidate
s
olu
ti
on
′
is
sel
ect
ed
via
the
ma
ximiza
ti
on of t
he
ac
qu
i
sit
ion
fu
nctio
n
[38]
:
′
=
∈
(
)
(5)
2.6.3.
B
ayesi
an al
go
ri
th
m
The
ke
y
sta
ges
of
the
Ba
yesi
an
opti
miza
ti
on
te
ch
nique
ar
e
dem
on
st
rated
in
Al
gorithm
1
[
38]
.
T
he
init
ia
l
ro
un
d
pro
duces
the
ba
sic
dataset
s
and
.
A
st
och
a
s
ti
c
su
r
rogate
model
of
t
he
obje
ct
ive
f
unct
ion
is
conseq
ue
ntly
dev
el
op
e
d.
F
ollow
in
g
t
hat,
a
sam
ple
is
c
hose
n
by
ma
ximizi
ng
t
he
ac
qu
isi
ti
on
f
unct
ion
.
T
he
sample
is
asse
ssed
via
the
obje
ct
ive
f
un
ct
i
on.
The
surr
og
at
e
mo
del
is
s
ub
s
eq
ue
ntly
re
v
ise
d
with
the
novel
data. T
his tec
hniq
ue wil
l co
ntinu
e
unti
l t
he maxim
um
num
ber o
f
it
erati
on
s is
met.
Algorith
m
1
.
B
ayesian
opti
miza
ti
on
al
gorith
m
Re
qu
ire:
An a
cqu
isi
ti
on fu
nc
ti
on
A
1
.
C
onstr
uct th
e prima
r
y data
set
,
us
in
g
t
he ob
je
ct
ive
fun
ct
ion
.
2
.
B
uild t
he Ga
us
sia
n p
r
ocess mo
del u
ti
li
zi
ng the
dataset
,
3
.
W
hile t
he
st
opping c
rite
ria
hav
e
not
been
met,
do
4
.
M
a
ximize
the ac
qu
isi
ti
on
functi
on:
′
=
(
)
∈
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
1038
-
1050
1046
5
.
E
valuati
on:
’
←
(
’
)
6
.
Augme
nt
the
data set
by addin
g
′
,
′
to
,
7
.
Re
bu
il
d t
he Gau
ssian
pr
ocess mo
del
for
f
u
si
ng the
expan
de
d
dataset
,
8
.
En
d
wh
il
e
3.
RESU
LT
S
AND DI
SCUS
S
ION
3.1.
Ex
peri
ment
al
s
et
up
The
simulat
io
ns
we
re
co
nduc
te
d
in
P
ytho
n
3.9.15
us
in
g
t
he
sci
kit
-
le
ar
n
li
br
a
ry
.
The
pro
gr
a
m
was
cod
e
d
on
a
n
H
P
c
ompu
te
r
with W
in
dow
s 10.
I
n
our
r
esearc
h,
dela
ye
d
flig
hts
a
re
re
pr
ese
nted
as
1
a
nd on
-
ti
m
e
fligh
ts a
s
0.
3.2.
Res
ults
and
a
n
alysi
s
Choosin
g
th
e
mo
st
ap
pro
pr
ia
te
hype
r
par
am
et
ers
possesses
an
e
norm
ous
eff
ect
on
t
he
pe
rformance
model.
A
sel
e
ct
ion
of
dif
fere
nt
hy
perpara
mete
r
values
ha
s
bee
n
c
onsid
ered
a
s
a
sea
r
ch
s
pace
to
bu
il
d
an
d
op
ti
mize
the
pr
opos
e
d
MLP
model:
i)
Hi
dden
la
yer
siz
e:
(
100,),
(
50)
,
(50,50),
(
100,1
00)
,
(10
0,50),
(
50,10
0)
;
ii
)
Acti
vatio
n
functi
on:
lo
gisti
c,
Tan
h,
a
nd
Re
LU
,
ii
i)
S
olv
e
r:
SGD
,
Ad
a
m
;
iv)
Al
ph
a:
0.0
1,
1e
-
6,
1e
-
2
;
v)
Lear
ning
rat
e:
co
nst
ant, i
nvscal
ing
,
ad
a
ptive
; an
d vi)
M
a
x
it
erati
on: 1
00
, 500, 1
000, 20
00
.
Acti
ons
can
va
ry
in
durati
on
and
occ
ur
c
oncurrently
,
poss
ibly
overla
ppin
g
in
ti
me
[
39]
.
To
validat
e
the
c
ho
ic
e
of
the
hype
r
par
am
et
er
op
ti
miza
ti
o
n
met
hod,
we
co
mp
a
re
d
t
he
Ba
yesian
opti
miza
ti
on
with
that
o
f
gr
i
d
searc
h
[37]
an
d
ra
ndom
s
earch
[37
]
te
ch
niques
re
ga
rd
i
ng
t
he
accu
rac
y
an
d
the
el
a
pse
d
ti
me.
Acc
or
ding
to
the
resu
lt
s
in
Table
3
,
w
e
de
du
ce
t
hat
the
Ba
yesia
n
al
gorithm
is
t
he
best
opti
miza
ti
on
met
hod
to
be
pro
po
se
d
in
t
his
stu
dy.
I
n
Table
4,
the
c
ombinati
on
of
the
best
hy
pe
rp
a
rameter
s
and
t
he
best
ac
cur
ac
y
gen
e
rated
by
the
Ba
yesian
op
ti
miza
ti
on
i
s
prese
nted
.
Our
pro
posed
m
od
el
was
e
valuated
in
te
rms
of
normali
zat
ion,
data
sam
pling,
an
d
pa
ramete
r
tu
ning.
Ta
bl
e
5
s
hows
th
e
evaluati
on
re
s
ults
befo
re
an
d
after
normali
zi
ng
th
e
data
reg
a
r
din
g
the
recall
,
the
F1
sco
re
,
the
accu
rac
y,
an
d
the
pre
ci
sion
.
It
s
ho
ws
t
he
importa
nce
of
data
no
rmal
iz
at
ion
to
e
nhan
ce
the
ef
fici
en
cy
a
nd
t
rainin
g
co
ns
ist
e
ncy
of
t
he
m
odel
bein
g
pro
po
se
d.
We
performe
d
th
e
data
bala
ncing
us
i
ng
t
he
Smo
te
-
To
me
k
te
chn
i
qu
e
,
w
hich
c
ombines
unde
r
samplin
g
a
nd
ov
e
rsam
plin
g
f
or
bette
r
sa
mpl
ing
.
Ta
ble
6
pr
ese
nts
the
e
valuati
on
fi
n
di
ng
s
with
a
nd
withou
t
data
sa
mp
li
ng.
Ba
sed
on
the
res
ults,
we
de
du
ce
that
bala
ncin
g
data
with
t
he
Sm
ote
-
T
om
e
k
te
c
hn
i
que
wa
s
very
su
cce
ssf
ul
,
w
hich
i
ncr
e
ased
the
accu
r
acy
t
o
90.
13%
.
All
the
oth
er
metri
cs
wer
e
impro
ved
c
ompare
d
with the
pre
vious
fin
dings.
T
able
7
hi
gh
li
gh
ts
the
e
ff
ect
of
opti
mizi
ng
the
s
uggeste
d
M
LP
m
odel
by
m
onit
or
i
ng
the
re
su
lt
s
befor
e
a
nd
afte
r
the
Ba
ye
sia
n
op
ti
miza
ti
on.
We
noti
ce
that
the
Ba
yesian
op
ti
miza
ti
on
ha
s
res
ulted
in
a
hi
gh
e
r
accurac
y
of
92.
49%.
All
the
oth
e
r
met
rics
wer
e
imp
rov
e
d
c
ompare
d
w
it
h
the
pre
viou
s
fin
dings
.
Fig
ur
e
4
monit
or
s
the
r
ecei
ver
oper
at
ing
c
urve
(RO
C)
f
or
t
he
final
pro
pose
d
mod
el
with
no
rmal
iz
at
ion
,
sam
pling
a
nd
op
ti
miza
ti
on. T
he
a
rea
unde
r
the c
urve is
0.923
0.
Table
3.
C
omp
ariso
n of
hy
per
-
pa
rameter
alg
or
it
hms
Bay
esian
Grid
s
ear
ch
Ran
d
o
m
s
earc
h
Accuracy
(
%)
9
2
.49
8
8
.08
8
7
.61
Elaps
ed
t
im
e
(
seco
n
d
s)
1
0
6
3
.7
9
3
9
9
4
.0
3
4
3
8
2
.3
7
Table
4.
T
he
best
hyp
e
r
-
par
a
mete
rs
a
nd acc
ur
ac
y of t
he
pr
opos
e
d mo
del
Mod
el
Hy
p
er
-
p
ara
m
ete
r
(
HP)
The b
est HP
v
alu
e
The b
est accura
cy
Multilay
er
p
er
cept
ron
Activ
atio
n
ReLU
9
2
.49
%
Alp
h
a
0
.01
Hid
d
en
layer sizes
(50
)
Lear
n
in
g
r
ate
co
n
stan
t
Max
Ite
r
1000
So
lv
er
Ad
am
Table
5.
E
val
ua
ti
on
met
rics i
n
te
r
ms
of d
at
a
normal
iz
at
ion
W
ith
n
o
rm
alizatio
n
W
ith
o
u
t no
rm
alizatio
n
Accuracy
(
%)
7
9
.40
7
5
.98
Precisio
n
7
8
.96
7
4
.00
Recall
7
7
.73
7
0
.66
F1
sco
re
7
8
.13
7
3
.22
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
Airp
or
t i
nfr
as
t
ru
ct
ur
e
an
d
r
unway
p
reci
si
on a
id
s for
forec
ast
ing
fl
ig
ht a
r
riv
al d
el
ays
(Haj
ar
All
a)
1047
Table
6.
E
val
ua
ti
on
met
rics i
n
te
r
ms
of d
at
a
b
al
anci
ng
W
ith
sam
p
lin
g
W
ith
o
u
t sam
p
lin
g
Accuracy
(
%)
9
0
.13
7
9
.40
Precisio
n
9
1
.10
7
8
.96
Recall
7
9
.93
7
7
.73
F1
sco
re
9
0
.48
7
8
.13
Table
7.
valuat
ion
metri
cs i
n
t
erms o
f hyper
pa
rameter
tu
ning
W
ith
op
tim
izatio
n
W
ith
o
u
t op
tim
izatio
n
Accuracy
(
%)
9
2
.49
9
0
.13
Precisio
n
9
2
.15
9
1
.10
Recall
8
0
.27
7
9
.93
F1
sco
re
9
2
.13
9
0
.48
Figure
4
.
ROC
curve
for t
he p
rop
os
ed
m
od
el
.
S
ource:
own ca
lc
ulati
on
3.3.
Be
nchm
ar
k
f
indi
ng
s
To
dem
onstra
te
the
eff
ic
ac
y
of
our
s
ugges
te
d
ap
proac
h,
we
co
mp
a
re
d
our
fi
nd
i
ngs
in
Table
8
t
o
tho
se
from
pre
vious
stu
dies.
The
c
omparis
on
wa
s
ma
de
ba
sed
on
t
he
rec
al
l,
the
F1
sco
re,
the
acc
urac
y,
a
nd
the preci
sio
n.
We
remark t
ha
t our su
gg
e
ste
d st
rateg
y
is
t
he most acc
ur
at
e
wh
e
n
c
ompa
re
d
to
the
othe
rs.
Table
8
.
Re
la
te
d works a
nd pr
opos
e
d met
hod com
pa
rison
Mod
el
Featu
res us
ed
Ob
jectiv
e
Metr
ic
1:
accuracy
(%)
Metr
ic
2:
F1
sco
re
(%)
Metr
ic
3:
p
recisio
n
(%)
Metr
ic
4:
reca
ll
(%)
Hen
riqu
es
et al
.
[40
]
Flig
h
t info
rm
atio
n
,
weather
data;
a
i
rc
raf
t
d
ata;
d
elay
p
r
o
p
ag
atio
n
info
rm
atio
n
To p
redict fligh
t
arr
iv
al delay
s
8
5
.63
7
9
.00
-
-
Stefano
v
ic
et al.
(ar
rival
)
[41
]
Fly
in
g
perio
d
,
trip n
u
m
b
er,
ai
rline,
d
estin
atio
n
,
o
rigin
,
clim
a
te,
ceilin
g
d
ata,
v
elo
city
of the win
d
,
win
d
directio
n
,
v
isib
ility
,
p
lan
n
ed
tim
e,
classes
To p
redict th
e
fligh
t ar
rival t
im
e
d
ev
iatio
n
f
o
r
Lithu
an
ian
air
p
o
rts
4
7
.43
5
0
.77
4
7
.43
5
6
.73
Stefano
v
ic
et al
.
(dep
arture)
[41
]
Fly
in
g
perio
d
,
trip n
u
m
b
er,
ai
rline,
d
estin
atio
n
,
o
rigin
,
clim
a
te,
ceilin
g
d
ata,
v
elo
ci
ty
of the win
d
,
win
d
d
irection
,
v
isib
ility
,
p
lan
n
ed
tim
e,
cla
ss
es
To p
redict th
e
fligh
t dep
arture
tim
e dev
iatio
n
for
Lithu
an
ian
air
p
o
rts
8
5
.65
8
7
.86
8
5
.65
9
0
.90
Pam
p
lo
n
a
et al.
[42
]
Flig
h
t
d
ata
,
d
elay
j
u
stificatio
n
cod
e
To p
redict ai
r
traf
fi
c
d
elay
s
9
1
.30
7
7
.00
8
7
.00
6
9
.00
Vo
n
itsan
o
s
et al
.
[43
]
Flig
h
t data, sam
e
-
o
rigin
-
fligh
ts co
u
n
t,
av
erage air
lin
e
del
ay
,
av
erage orig
in
d
elay
,
cancelled
(
c
lass
ification
),
a
rr
iv
al
d
elay
(
regressio
n
)
To forecast
air
fligh
t delay
s
-
5
4
.35
5
4
.37
5
3
.40
Ou
r
m
o
d
el
Flig
h
t data, dis
tan
ce,
run
way
leng
th
,
nu
m
b
er
o
f
run
way
s, r
u
n
way
s
in
tersection
s, r
u
n
w
ay
s p
recisio
n
r
ate,
n
u
m
b
er
o
f
g
ates,
nu
m
b
er
o
f
t
er
m
in
als.
To p
redict fligh
t
arr
iv
al delay
s
9
2
.49
9
2
.13
9
2
.15
8
0
.27
Evaluation Warning : The document was created with Spire.PDF for Python.