Documentos de Académico
Documentos de Profesional
Documentos de Cultura
ABSTRACT
Due to strong traffic momentum since the sector deregulation in Turkey passenger numbers has risen at an
approximately CAGR (Compound Annual Growth Rate) of 16% between 2003 and 2013, which means three
times the real GDP (Gross Domestic Product) growth in the same period. We believe that Turkeys
geographically advantageous position, increasing attractiveness as a tourism destination, and the governments
supportive approach to the sector are the secular reasons for this average growth. A long-term forecast of the
air transportation is crucial regarding the precautions that will be taken in the future. It is also an essential
input for an investment planning. In the field of air transportation forecasting, there are a number of empirical
models, which can be classified as judgemental, causal and time series. Time series forecasting is an important
area of forecasting in which past observations of the same variables are collected and analysed to develop a
model describing the underlying relationship. The model is then used to extrapolate the time series into the
future. In this study, forecasting models for time series are reviewed. Aircraft, passenger, and cargo statistics
between 2007-2015 years in Turkey are gathered and used to construct forecasting models for each quantity.
Based on the mathematical models constructed by using genetic algorithm from heuristics, we project a 24%
increase in Turkish aircraft traffic, a 50% increase in Turkish air passenger traffic, and a 34% increase in
Turkish air cargo traffic between 2016-2020 years.
Z
Trkiyede havaclk sektrndeki yaplan dzenlemelere bal olarak hava trafiinde 2003-2013 dneminde
gl bir byme yaanm, sz konusu dnemde ortalama byme oran yllk % 16 olarak gerekleerek
GSYH (Gayri Safi Yurt i Hsla) byme orannn katna ulamtr. lkemizin sahip olduu uygun corafi
konum, turistik mekanlarnn ekiciliinin gittike artmas ve hkmet politikalarnn havaclk sektrn
destekler mahiyette srdrlmesi bu gl bymenin temel sebepleridir. Sektrde gelecee ilikin alnacak
tedbirlerde uzun vadeli tahmin yapabilme kritik neme sahiptir. Bu tahminler ayrca yatrm planlamalar iin
de etkili birer girdi durumundadr. Hava tamacl sektrnde tahmin yapabilme konusunda yargsal,
nedensel ve zaman serileri gibi snflandrlabilen ok sayda gzleme dayal model kullanlmaktadr. Ayn
deikenin gemi dneme ilikin verilerinin ilikilendirildii ve bunlardan yararlanlarak deikenin hareketini
ifade eden bir modelin oluturulduu zaman serileri gelecee ilikin tahmin yapmada olduka nemlidir. Bu
almada ncelikle tahmin yapmada kullanlan zaman serisi modelleri gzden geirilmitir. Sonrasnda kalk
ve ini yapan uak says, tanan yolcu ve kargoya ilikin Trkiyenin 2007-2015 dnemi istatistikleri derlenmi
ve bu veriler gelecee ilikin tahminleme modelinde kullanlmtr. Sezgisel yntemlerden biri olan genetik
algoritma kullanlarak kurgulanan model kapsamnda 2016-2020 dneminde uak trafiinde %24, yolcu
trafiinde % 50 ve kargo miktarnda da %34 artn olaca tahmin edilmitir.
PEHLIVANOGLU, ATIK
23
Time Series Forecasting via Genetic Algorithm for Turkish Air Transport Market
PEHLIVANOGLU, ATIK
24
Time Series Forecasting via Genetic Algorithm for Turkish Air Transport Market
In this study, the purpose was not to compare the It is possible to use different trend functions in a
performance of the main classical forecasting functional. Exponential, polynomial, power, or even
methods. Instead of complex and sometimes locally rational functions may be selected as a trend function.
fitted mathematical models, we preferred to construct Among the possible trend functions, we preferred to
a simple functional to make forecasting in Turkish air use power and second-degree polynomial function.
transport market. There is a significant difference between second
degree polynomial and a power function and it is the
3. MATHEMATICAL MODELLING nature of trend. Polynomial function shows an
increasing character. Therefore, its forecasting nature
Time series including aircraft, passenger, and cargo shows the existence of development in the future of
can be decomposed of three sub functions such as industry. The power function shows more stable
trend, seasonal, and irregularity component. A trend character although there are slightly increase from one
component shows a long term movement in a time time to next time. The descriptions of each function
series. It is the underlying direction showing an and their characteristic trend are shown in Figure 1.
upward or downward tendency of a time series. A
seasonal component describes any regular A seasonal component contains periodic fluctuations.
fluctuations. It is the component of variation in a time Periodicity can be described as the following:
series which is dependent on the time of year. The last
component named irregular or noise component is that (2)
left over when the other components of the series have
been accounted for. It includes random fluctuations in
a time series. A functional for air transportation time where is the time and is the periodicity. Air
series can be defined as the sum of a trend ( ), a transportation time series does not show a fully
seasonal ( ), and an irregularity ( ) functions like periodic behavior due to long term movement.
the following [9]: However, a proper value of and amplitude of a time
series can be found by using the extreme points. For
(1) any time series, the maxima points are defined by
using the following condition;
4
x 10
10 if
8
then (3)
number of A/C
2
The time difference between the maxima is the period
of a time series. A mathematical model of a seasonal
0 component can be constructed by using a simple
0 50 100 150 200 sinusoidal function like the following:
time
(a) Polynomial function:
4
x 10
(4)
8
The second functional contains a trend function such In our study, the design vector depending on the
as a second degree polynomial, a periodic function selected functional for genetic algorithm optimization
like sinusoidal, and an irregular function such as process is as follows; and
logistic distribution as follows: . These are also called as
chromosome in GA process.
(11)
The objective function is usually a least-squares
where function given by
(12)
(15)
(13)
(14) where is ith value of a time series including data
and is ith value of the computed response obtained
In this second functional: is the coefficient of from the forecasting model. In applications, the
leading term, is another coefficient of the second following features selected by our experience shown
term, and is the constant term. Similar to the in Table 1 are used in GA optimization processes:
previous description, is amplitude, is a period,
is a phase shift, and is a scale factor of Table 1. The GA optimization features
irregularity. The coefficients in each combination
can be determined via a proper optimization process Feature Data
such as genetic algorithm (GA). Population size 50
Fitness scaling Rank
As a stochastic method, GA is an emergent Selection Stochastic uniform
optimization algorithm mimicking of the natural Reproduction Elite:2; Crossover:0.8
evolution, where a biological population evolves over Mutation Gaussian; scale:1.0, shrink 1.0
generations to adapt to an environment by selection, Crossover Heuristic:; ratio:1.2
recombination based on crossover operations, and Max. generation 500
mutation. When GA is applied to optimization
problems, fitness, individuals and genes usually In GA process, the fitness scaling converts the raw
correspond to an objective function value, a design fitness scores that are returned by the fitness function
to values in a range that is suitable for the selection
PEHLIVANOGLU, ATIK
26
Time Series Forecasting via Genetic Algorithm for Turkish Air Transport Market
function. The selection function uses the scaled fitness where 1 is a user defined scale factor, is other user
values to select the parents of the next generation. The defined shrink factor, u is the random number
selection function assigns a higher probability of generated from standard normal distribution. U and L
selection to individuals with higher scaled values. The are the upper and lower bound indexes for the genes.
selected fitness scaling option, rank, scales the raw Mutation rate is not applicable to this type of operator
scores based on the rank of each individual instead of because all genes in chromosome are mutated. The
its score. The rank of an individual is its position in current generation is labeled with g, the maximum
the sorted fitness scores. generation is G, a child is Mc, and the individual
selected for mutation is indexed with P.
The selection function chooses parents for the next
generation based on their scaled values from the Obviously, there are other different options for GA
fitness scaling function. An individual can be selected operators including fitness scaling, selection, mutation
more than once as a parent, in which case it and crossover operations. Different combinations of
contributes its genes to more than one child. The them may provide different performances. The
selection option, stochastic uniform, lays out a line in purpose of this study is to provide a reasonable
which each parent corresponds to a section of the line mathematical modelling for forecasting. Therefore, the
of length proportional to its scaled value. The effects of operator selections were not discussed in
algorithm moves along the line in steps of equal size. this study.
At each step, the algorithm allocates a parent from the
section it lands on. 4. MODEL CONSTRUCTIONS
The elitism concept is applied next to ensure that the Three different quantities such as the number of
best objective function value within a population is aircraft take-off and landed, the number of passengers,
not reduced from one generation to the next. In GA and the amount of cargo in international flights are
process, the best two chromosomes of the population taken into consideration for the future expectations in
are put into next population during reproduction Turkey. The data obtained from the official website of
phase. the General Directorate of State Airports Authority of
Turkey [10] database is used in our experiments and it
Crossover proceeds in two steps; within the first step, is the total values of Turkish airports. The time series
two individuals as parents are randomly selected from include the monthly total numbers between January
the current population via stochastic uniform selection 2007 and December 2015 which mean 9 years
function. In the second step, a new structure is formed including 108 observations. The following figures
from parents by exchanging their genes; so the show the current data for each quantity. In figures,
crossover recombines features from two parents to A/C represents the abbreviation of aircraft word; t
produce offspring. In Heuristic crossover, a child (c) is represents the time schedule based on months
generated as the following following the January 2007.
4
x 10
8
(16)
the number of A/C
g
2 1 1
G
j 1,2,..., n
3 2 ( xUj x Lj ) (17)
j 1,2,..., n
x Mc
j x Pj 3 .u
PEHLIVANOGLU, ATIK
27
Time Series Forecasting via Genetic Algorithm for Turkish Air Transport Market
6 5
x 10 x 10
12 2.5
8 2
Fitness value
6
1.5
4
2
1
0
0 50 100
t
0.5
Figure 3. The number of airline passengers in Turkey 1 10 100 500
during 108 months from Jan. 2007 to Dec. 2015. Generation (log)
x 10
5
(a) Optimization process
3
4
the amount of cargo (ton)
2.5 x 10
8
2
1.5
0.5 4
0
0 50 100
t 2
Figure 4. The amount of cargo in Turkey during 108 real data
(18)
(19)
PEHLIVANOGLU, ATIK
28
Time Series Forecasting via Genetic Algorithm for Turkish Air Transport Market
5
x 10 the best fitness reaches the value of 9,081,693.
2.5 Additionally, the fitted curve and real data are shown
Best fitness in Figure 7(b).
2
7
Fitness value
x 10
5
1.5
Best fitness
1 4
Fitness value
0.5 3
0 2
1 10 100 500
Generation (log)
1
(a) Optimization process
4 0
x 10 1 10 100 500
8
Generation
6
x 10
4 12
real data
10
the number of passengers
fitted curve
2
8
real data
fitted curve 6
0
0 20 40 60 80 100 4
t
2
(b) Curve fit of aircraft quantity
0
7 5
x 10 x 10
2.2 5
Best fitness Best fitness
2 4.5
1.8 4
Fitness value
Fitness value
1.6 3.5
1.4 3
1.2
2.5
1
2
0.8
1 10 100 500 1.5
1 10 100 500
Generation
Generation
The first functional for the future of cargo quantity in The genetic optimization process results including the
Turkey is constructed by the following expression: best fitness values versus generations are depicted in
Figure 10(a). At the end of the optimization process,
the best fitness reaches the value of 126,126.
Additionally, the fitted curve and real data are shown
(22) in Figure10(b).
5 4
x 10 x 10
6 9
6
4
5
3 4
3
2
2
1 1
1 10 100 500 0 50 100 150
Generation t
(a) Real data and general trend
(a) Optimization process
4
5 x 10
x 10 9
3
real data 8
2.5 fitted curve
7
2
6
1.5
5
1
4
0.5
3
110 120 130 140 150 160
0 future time
0 20 40 60 80 100
t (b) Zoomed future expectations
(b) Curve fit of cargo quantity
Figure 11. The future expectations for the number of
Figure 10. The optimization process and the resulted aircraft in Turkey.
curve fitting.
The estimated numbers of aircraft amount versus
From the figures and best fitness values we can months forecasted by Eq. (19) are shown in Figure11.
observe that again the second functional based on Additionally; the minimum, the maximum, an
polynomial fits better than the first functional based average, and the total number of aircraft in each year
on power terms. The amount of cargo carried by are tabulated in Table 2. On Figure 11(a), the real data
Turkish airline industry is in a rising trend. belong to the past and the forecasted data are depicted
together. Figure 11(b) shows the zoom of future time
5. THE FUTURE OF TURKISH AIR interval. According to the forecasting model, it is
TRANSPORT MARKET: 2016-2020 expected to increase % 5.5 in average in the number of
aircraft take-off and landed in each year. At the end of
Optimization results show that a forecasting model five-year term, the number of aircraft is going to be
based on a second degree polynomial, a periodic around 824,310. This increase means %24
function, and an irregular function fits better and can development in total.
be taken into consideration for future expectations. It
is possible to use Eq. (19) to make the forecast for the Table 2. The future estimations for the number of
number of aircraft in Turkey. Similarly, Eq. (21) can aircraft in Turkey
be utilized to make forecast for the number of
passengers. Finally, Eq. (23) gives a chance to make Year Min. Max. Average Total
estimations for the amount of cargo in Turkish air 2016 39,002 71,039 55,419 665,020
transportation market. For that purpose, we need to 2017 41,827 74,397 58,804 705,650
establish a time horizon. A five-year term including 2018 44,632 77,716 62,323 747,870
2016-2020 years and 60 months in total is selected as 2019 47,961 81,235 65,373 784,480
the long-term time horizon. 2020 51,468 84,178 68,693 824,310
PEHLIVANOGLU, ATIK
31
Time Series Forecasting via Genetic Algorithm for Turkish Air Transport Market
The estimated numbers of passengers carried by millions. This increase means %50 development in
airlines versus months forecasted by Eq. (21) are total. This result also shows that the increase in
shown in Figure12. In addition to this figure, the passenger volume is twofold bigger than the increase
minimum, the maximum, an average, and the total in aircraft amount.
number of passengers in each year are tabulated in
Table 3. The estimated amount of cargo carried by airlines
6
x 10 versus months forecasted by Eq. (23) is depicted in
16
Figure13. In addition to this figure, the minimum
14
real data amount of cargo, the maximum amount of cargo, an
future average amount of cargo, and the total amount of
the number of passengers
7
x 10 2
1.6
1.5
1.4
the number of passengers
1.2
0.5
1
0
0 50 100 150
0.8 t
(a) Real data and general trend
0.6
5
x 10
3.2
0.4
110 120 130 140 150 160 3
future time
the amount of cargo (ton)
2.8
(b) Zoomed future expectations
2.6
Figure 12. The future expectations for the number of
2.4
passengers in Turkey.
2.2
Table 3. The future estimations for the number of
2
passengers in Turkey
1.8
Year Min. Max. Average Total 1.6
2016 5,332,164 11,104,841 8,396,401 100,756,813
1.4
2017 6,225,148 1,200,908 9,329,795 111,957,547 110 120 130 140 150 160
2018 7,259,834 13,025,124 10,340,635 124,087,629 future time
2019 8,315,333 14,138,984 11,426,657 137,119,891 (b) Zoomed future expectations
2020 9,425,498 15,315,232 12,604,857 151,258,295
Figure 13. The future expectations for the amount of
According to the forecasting model, it is expected to cargo in Turkey.
increase an approximately % 10.6 in the number of
passengers in each year. At the end of five-year term,
the number of passengers is going to be above150
PEHLIVANOGLU, ATIK
32
Time Series Forecasting via Genetic Algorithm for Turkish Air Transport Market
Table 4. Future data for the amount of cargo in analysis for air transport demand forecasting. Journal
Turkey (Tons) of Air Transport Management, 39 (2014) 1-11.
Year Min. Max. Average Total
[6] Xie, G., Wang, S., and Lai, K.K. (2014). Short-
2016 147,428 237,417 195,877 2,350,524 term forecasting of air passenger by using hybrid
2017 168,037 252,653 212,692 2,552,313 seasonal decomposition and least squares support
2018 177,027 276,130 229,188 2,750,261 vector regression approaches. Journal of Air Transport
2019 198,780 290,006 245,035 2,940,420 Management, 37( 2014) 20-26.
2020 213,834 314,286 263,242 3,158,909
[7] Zhang, G.P. (2003). Time series forecasting
using a hybrid ARIMA and neural network model.
6. CONCLUSION
Neurocomputing, 50 (2003) 159175.
The purpose of this paper was to construct [8] Schlter, S. and Deuschle, C. (2010).Using
mathematical models to forecast the Turkish air wavelets for time series forecasting does it pay off ?
transport market in long-term period. For this purpose, (Discussion Papers No. 4/2010). Department of
we first presented the short review of the main Statistics and Econometrics, University of Erlangen-
classical forecasting methods and also gave a brief Nuremberg.
introduction into a mathematical modelling. For our
[9] Majani, B.E. (1987). Decomposition methods
empirical study, we chose a combination of three basic
for medium-term planning and budgeting. Makridakis
functions such as a trend function, a seasonal function,
S, Wheelwright S (edt.): The handbook of forecasting:
and an irregularity function. A genetic algorithm
A managers guide; Whiley: New York; 219237.
process was used to determine the coefficients of each
function. We tested two different forecasting [10] The General Directorate of State Airports
functional, compared the results, and saw that the Authority of Turkey, Statistics, (2016)
second combination including a second degree http://www.dhmi.gov.tr/istatistik.aspx (Accessing
polynomial as a trend function, a periodic function January 10th, 2016)
like sinusoidal, and an irregular function such as
logistic distribution fits better. The results also
confirmed that the Turkish air transport market is an VITAE
emerging market. According to the forecasting based
on the constructed mathematical models, we project a Assoc. Prof. Dr. Y. Volkan PEHLIVANOGLU
24% increase in Turkish aircraft traffic, a 50%
increase in Turkish air passenger traffic, and a 34% He was born in 1973 in Erzurum, Turkey. He
increase in Turkish air cargo traffic between 2016- graduated as an aeronautical engineer, and received
2020 years. MSc degree from the Technical University of Istanbul.
His PhD degree is from Old Dominion University of
Norfolk with the thesis titled Hybrid intelligent
7. REFERENCES
optimization methods for engineering problems. He
[1] Airbus (2014).Global Market Forecast Flying is currently the chair of Technical Programs
on Demand 2014-2033. http://www. airbus.com/ Department in NCO (Non-Commissioned Officer)
company/ market/forecast/ (Accessing August 1st, College since 2014.
2015)
His dominant scientific interest focuses on:
[2] Boeing (2014).Current Market Outlook 2014-
optimization methods, aerodynamics, and path
2033.http://www.boeing.com/assets/pdf/commercial/c
planning.
mo/pdf/Boeing_Current_Market_Outlook_2014.pdf
(Accessing August 1st, 2015)
Assist. Prof. Dr. llhan ATIK
[3] Ince, B. (2014). Turkish Aviation Sector.
Garanti Securities Research. http://www. He was born in 1971 in Ankara, Turkey. He graduated
garantiyatirim.com.tr/arastirma/Turkish_Aviation_Sec as an industrial engineer, and received MSc and PhD
tor_Feb2014.pdf (Accessing September 15th, 2015) degrees from Dumlupnar University of Ktahya with
[4] Yeung, J.H.Y., Cheung, W., Fung, M.K., Zhao, the thesis titled Strategic budgeting applications on
X., Zhang, M. (2010). The air cargo and express oligopolistic markets. He is currently the chair of
industry in Hong Kong: economic contribution and Military Sciences Department in NCO College since
competitiveness. Int. J. Shipp. Transp. Logist.2, 321- 2011.
345.
His dominant scientific interest focuses on: logistics,
[5] Xiao, Y., Liu, J. J., Hu, Y., Wang, Y., Lai, optimization methods, vocational and technical
K.K., and Wang, S., (2014). A neuro-fuzzy education planning.
combination model based on singular spectrum
PEHLIVANOGLU, ATIK
33