You are on page 1of 5

2016 2nd International Conference on Science in Information Technology (ICSITech)

The Use of Triple Exponential Smoothing Method


(Winter) in Forecasting Passenger of PT Kereta Api
Indonesia with Optimization Alpha, Beta, and
Gamma Parameters
Wawan Setiawan, Enjun Juniati, Ida Farida
Computer Science Department
Indonesia University of Education
Bandung, Indonesia
wawans@upi.edu, enjun@upi.edu, ida.parida@yahoo.com

Abstract—This research aims to implement Triple needed for the PT KAI Company to improve service comfort
Exponential Smoothing Methods (Winter) with the control and safety of passengers. Through good forecasting, PT KAI
parameters of alpha, beta, and gamma through techniques Company can provide services with the capacity and quality is
initialization data history with the smallest value of Mean satisfactory.
Absolute Percentage Error (MAPE). The time series data used
from Kereta Api Indonesia Ltd. (PT KAI) Bandung Indonesia There are several methods of forecasting time series are
between 2006 and 2014 for the Argo Wilis, Turangga, Mutiara commonly used include Exponential Smoothing (Winter). One
Selatan, Pasundan, and Kahuripan trains. The results of this type of Exponential Smoothing method (Winter) is Triple
study indicate that for the data fifth train fleet has a pattern of Exponential Smoothing (Winter). The advantages of this
non-stationary fluctuating trend. Initialization of the most well method can be used in forecasting for stationer data types or
done to the data is one year to produce optimal MAPE. The
data is not stationary. Because of this, the research aims to
results of forecasting with Triple Exponential Smoothing
Methods (Winter) generally have good accuracy, namely Argo
apply the method Triple Exponential Smoothing (Winter) in
Wilis is 86.60, Turangga is 70.13, Mutiara Selatan is 85.16, the system of forecasting passenger PT KAI Company.
Pasundan 90.87, and Kahuripan 88.47 percent. In general, the The performance of Triple Exponential Smoothing method
accuracy of forecasting that is used quite well. (Winter) depends on the parameters α, β, and γ. Determining
the value of the parameter α, β, and γ generally by a trial and
Keywords—triple exponential smoothing, winter, control
error. In this study, the determination of the value of the
paramater, MAPE
parameter α, β, and γ will be controlled by initialization of
I. INTRODUCTION historical data. For the purposes of initialization required a
complete data history, the necessary data in a long period. In
Forecasting is a process to predict the future state based on this study, used passenger data PT KAI Company Bandung
the size of the quantity, quality, time, and location in order to within the period January 2006 - December 2014 for this type
meet the demands of goods and services [1,2]. Forecasting is a of train fleet Argo Wilis, Turangga, Mutiara Selatan, Pasundan
series of activities to predict a state that will occur in the future and Kahuripan.
based on previous data. Forecasting is expected to assist the
institution or organization to strengthen the planning so that II. LITERATURE REVIEW AND RESEARCH METHOD
measures can be formulated well and produced the right
decisions, both strategic and influential in the long term, or a A. Type Data Forecasting
tactical decision that short-term [2]. Generally there are two types of data that is predictable, ie
quantitative and qualitative data [4]. Qualitative data is data
In a free market, the conditions are complex and dynamic
that a description or summary of statements, such as the
so forecasting the demand for much needed as a reference in
opinion of the manager, the advice of experts in a particular
making in the planning of a good production with high
field, feedback from customers or salespeople.
accuracy [3]. A good forecasting is the information that is
needed in a planning and decision-making of an organization. While quantitative data is the data in the form of numbers.
Quantitative data can be divided into two: Time Series Data,
Kereta Api Indonesia Ltd. (PT KAI) is a company engaged
and Data Cross-Sectional. Time series data is data that is
in the transportation services that have a regular schedule (time
displayed based on time, such as monthly data, daily data,
series). As the dynamics of life and the characteristics of
weekly data or other types of time. Cross-sectional data is data
passengers becomes more complex, it becomes a complicated
that is displayed by a certain time, but the data at a particular
problem of transportation in Indonesia. Therefore a plan is
time. In general, forecasting time series data can be done using

978-1-5090-1721-8/16/$31.00 ©2016 IEEE 198


2016 2nd International Conference on Science in Information Technology (ICSITech)

methods Moving Average, Exponential Smoothing, the original data showed seasonal fluctuations stable. While the
Autoregressive Moving Average (ARMA) and so on. As for multiplicative models are used if the plot of the original data
the cross-section data forecasting can use regression method, showed seasonal fluctuations variations. The equation (2) to (5)
both simple regression or multiple regression. used for additive model.
B. Data Stationary a. To estimate the component level.
Based on the time series, there are two kinds of status, ie
the data is stationary and unstationary. Stationary data is data = ( − ) + (1 − )( + ) (2)
that the average value does not change significantly with time.
Dara statsioner a stable data. The data unstationary is data that b. To estimate the trend component.
the average changes significantly with time. The data changes
can occur due to the effects of the trend, seasonal (season = ( − ) + (1 − ) (3)
effect) or the influence of the cycle. Stationarity status data is
very important to know to determine more about the methods c. To estimate seasonality component.
of forecasting what it will be used.
Some forecasting methods that can be used for data = ( − ) + (1 − ) (4)
stationary are Naïve Method, Simple Moving Averages and
Autoregressive Moving Average (ARMA). For data stationary d. For the forecast period to p of the specific data.
which containing the pattern of the trend can use the
Exponential Smoothing Holt Method, simple regression and Ŷ = + + (5)
ARMA (Box-Jenkins) [10]. For data stationary containing
seasonal patterns can use Decomposition Method, Triple As for the multiplicative models, by the following equation
Exponential Smoothing (Winter) and ARMA. As for the (6) until (9).
influence of cyclical data, can use Decomposition Methods,
a. To estimate the component level.
Econometric Models, Regression, and ARMA [4].
A pattern type of data is a major consideration in selecting = + (1 − )( + ) (6)
the most appropriate forecasting methods to be used. These
patterns may represent the characteristics that recurred
simultaneously with the passage of time or may represent a b. To estimate the trend component.
turning point that are not periodic [3]. Therefore it takes a
knowledge to recognize patterns of data in a series of data in = ( − ) + (1 − ) (7)
order to avoid mistakes in the selection of forecasting methods.
c. To estimate seasonality component.
C. Forecasting Errors
There are several tools that are often used to calculate the = + (1 − ) (8)

forecasting error. In this study we used MAPE (Mean Absolute
Percentage Error) [5]. MAPE stated percentage forecasting
errors in the results of the actual demand during a given period d. For the forecast period to p of the specific data.
that will provide information on the percentage of the level a
mistake. Mathematically, MAPE is expressed as in (1). Ŷ =( + ) (9)

where
= ∑ (1)
L = level estimate (influenced by the amount of α)
T = the trend estimate (influenced by the amount of β)
The criteria used for the MAPE value presented as Table I.
S = seasonality estimate (influenced by the amount of γ
TABLE I. CRITERIA OF MAPE p = seasonal period
l = length
No Value of MAPE Description Y = the actual data
1 < 10% Very good
2 10% - 20% Good
Ỹ = Value forecast for the coming period
3 20% - 50% Enough
4 > 50% Bad One of the problems in using winter is determine the value
for α, β and γ parameters which will minimize MAPE. The
The smaller value generated by the tools, indicate the approach for determining this value is usually done by trial and
better forecasting method. error. In the exponential smoothing, the initial value is needed
due to the value of L t - 1 is not available then the value forecast
D. Triple Exponential Smoothing (Winter) Method for any yet [9]. Therefore it takes an initialization to get the
The Triple Exponential Smoothing Methods (Winter) has original value. To initialize this Winter forecasting methods,
two models of computation, that is additive and multiplicative the data can be used for at least a complete seasonal (ie. L
manner. Additive model calculations performed if the plot of period) to determine the initial estimate of the seasonal index,

199
2016 2nd International Conference on Science in Information Technology (ICSITech)

St-l, and need to assess the trend factor from one period to the
next using the following equation (10) to (12).

= ( + + + ⋯ + ) (10)

( ) ( ) ( )
= + + ⋯ + (11)

= (12)
Fig. 2. Graph of Argo Wilis Passengers
As for initialization index Additive Winter season is Sk =
Yk-Li where k = 1, 2, 3, ... l.
The method of Winter effectively used for estimate
ionospheric delays at UKM station with the accuracy level of
78-96% [5]. Javedani H. et al. showed that the winter
multiplicative models can do best estimate in predicting the
amount of demand for electricity [8]. Further research
Permanasari, et al. showed that winter with α = 0.07, β = 0.01
and γ = 0.01 is able to provide a fairly good prediction and
error rate is quite small in predicting [8]. Fig. 3. Graph of Turangga Passengers

E. Research Method
The research method shown in Fig. 1.

Fig. 4. Graph of Mutiara Selatan Passengers

Fig. 1. Research Method

Based on the Fig. 1 in a simple, forecasting method


implemented consists of five (5) stages as follows:
Fig. 5. Graph of Pasundan Passengers
1. Collect data on passengers PT KAI in a period of several
years.
2. Conduct a study of passenger data model of PT KAI.
3. Calculate forecasting error (MAPE).
4. Determine the value of the parameter α, β and γ using
winter.
5. Forecasting passenger of PT KAI.

III. RESULS AND DISCUSSION Fig. 6. Graph of Kahuripan Passenge


A. History Data Model
Fig. 2 to 6 show seasonal trends in the data for all fleets.
The data used in this study is data on the number passenger
Elements of the trend can be seen from the movement of the
of Argo Wilis, Turangga, Mutiara Selatan, Pasundan and
Kahuripan trains. The data used are the monthly period for 9 lines up and down and repeated every year. Because the data
years in the period 2006 to 2014 so that there are 108 periods. is containing seasonal trends, it can be defined as non-

200
2016 2nd International Conference on Science in Information Technology (ICSITech)

stationer [3]. For non-stationer data forecasting, and C. Triple Exponential Smoothing Methods (Winter)
fluctuations are varied, it is suitable to use Triple Exponential Forecasting
Smoothing Methods (Winter) multiplicative models. Table IV is the result of forecasting the number of train
B. Forecasting Model passengers for 2015 using the method of Winter with a value of
alpha, beta and gamma resulted initialization 1 year.
Forecasting models were conducted to determine the value
of the parameters α, β, and γ is right in order to produce the TABLE IV. THE RESULTS FORECAST ARGO WILIS
smallest MAPE value. Forecasting model in this research is by
changing the value of the initialization data used to determine Forecast Accuracy
No Month Actual Data
Data (%)
the value of early stage forecasting process using Triple 1 January 5,549 4,659 83.96%
Exponential Smoothing (Winter). In the testing this prediction 2 February 4,233 3,862 91.24%
model used four variations of the initialization data forecasting 3 March 3,792 4,100 91.88%
number of train passengers in 2015. Table II is the result of 4 April 4,758 3,678 77.30%
testing the forecasting model. 5 May 4,982 4,875 97.85%
6 June 4,507 5,602 75.70%
TABLE II. THE RESULTS MAPE OF FORECASTING MODEL 7 July 6,024 5,182 86.02%
8 Agust 5,470 4,921 89.96%
MAPE 9 September 4,501 4,752 94.42%
No Type of Fleet Initials Initials of Initials of 10 October 3,852 4,518 82.71%
Initials of
of Two Three Four 11 November 3,833 4,341 86.75%
One Year
Year Year Year 12 December 5,664 4,670 82.45%
1 Argo Wilis 17.85% 20.24% 20.19% 21.81% Average 86,69%
2 Turangga 20.56% 23.68% 24.97% 26.59%
The worst forecasting for Argo Wilis is in June, while the
3 Mutiara Selatan 20.07% 24.58% 25.29% 30.00% best is in May (Table IV).
4 Pasundan 24.01% 28.12% 31.05% 32.96%
5 Kahuripan 23.27% 26.93% 29.65% 30.84%
TABLE V. THE RESULTS FORECAST TURANGGA
Table II shows that the value of MAPE for the process of
Forecast Accuracy
forecasting the number of train passengers for all types of fleet No Month Actual Data
Data (%)
should use forecasting models with initial data one year. Table 1 January 3,879 4,030 96.11%
III shows an optimal value of α, β, and γ are generated for each 2 February 4,199 3,740 89.07%
fleet. 3 March 4,726 3,399 71.92%
4 April 5,700 3,556 62.39%
TABLE III. VALUE OF ALPHA, BETA AND GAMMA 5 May 5,640 3,595 63.74%
6 June 4,679 3,956 84.55%
No Train Alpha Beta Gamma 7 July 6,033 3,036 50.32%
1 Argo Wilis 0.40 0.18 0.57 8 Agust 6,071 3,726 61.37%
2 Turangga 0.54 0.10 0.92 9 September 4,989 3,377 67.69%
3 Mutiara Selatan 0.54 0.10 0.79 10 October 5,412 3,608 66.67%
4 Pasundan 0.22 0.29 0.18 11 November 5,090 3,386 66.52%
5 Kahuripan 0.24 0.24 0.46 12 December 5,812 3,560 61.25%
Average 70,13%
The value of α, β, and γ indicates a value that varies for
each train. The worst forecasting for Turangga is in July, while the
best is January (Table V). Forecasting for Turangga is worst by
1
Alpha
this method.
0.8
Beta
0.6 TABLE VI. THE RESULTS FORECAST MUTIARA SELATAN
Gamma
0.4 Forecast Accuracy
No Month Actual Data
Data (%)
0.2 1 January 5,364 6,090 86.47%
0 2 February 5,443 6,197 86.15%
Argo Wilis Turangga Mutiara Pasundan Kahuripan 3 March 5,343 6,121 85.44%
Selatan 4 April 5,797 6,684 84.70%
5 May 6,668 6,333 94.98%
Fig. 7. Graph Value of α, β, and γ 6 June 5,407 7,325 64.53%
7 July 7,886 6,333 80.31%
From Fig. 7 is observed as if there are two groups of train. 8 Agust 7,658 7,426 96.97%
Argo Wilis, Turangga, and Mutiara Selatan trains are in a 9 September 6,986 6,467 92.57%
10 October 6,192 6,700 91.80%
group, and trains Pasundan, and Kahuripan are in other group, 11 November 5,200 6,341 78.06%
especially by the value of γ. It shows the type of train 12 December 6,821 7,817 85.40%
passenger’s trend concerned. Average 85,16%
The worst Mutiara Selatan forecasting occurred in June, while
the best is in September (Table VI).

201
2016 2nd International Conference on Science in Information Technology (ICSITech)

TABLE VII. THE RESULTS FORECAST PASUNDAN Forecasting is best form Pasundan and worst for
No Month Actual Data Forecast Accuracy Turangga. Extreme differences of forecasting results due to
Data (%) variable unwieldiness especially management systems of each
1 January 13,409 14,313 93.26% train are varied. But overall method used forecasting well
2 February 10,905 10,298 94.43% classified category [8].
3 March 15,319 11,278 73.62%
4 April 12,267 11,994 97.77%
5 May 13,911 13,384 96.21%
6 June 12,604 12,565 99.69%
IV. CONCLUDING REMARKS
7 July 14,220 13,870 97.54% Triple Exponential Smoothing Methods (Winter)
8 Agust 14,371 10,501 73.07% multiplicative by control the parameters α, β, and γ through
9 September 13,312 11,829 88.86% initialization data history technique give good results in system
10 October 13,380 14,834 89.13%
forecasting passenger of PT KAI Bandung Indonesia.
11 November 11,755 11,906 98.72%
12 December 13,197 11,623 88.07% The diversity in forecasting results are because of the
Average 90,87% management system of each train fleet variables that cannot be
The worst forecasting for Pasundan is in August, while the controlled.
best is in November (Table VII). The forecasting for Pasundan
is the best by this method.
ACKNOWLEDGMENT
TABLE VIII. THE RESULTS FORECASTS KAHURIPAN
This research was supported by several parties, we thank
No Month Actual Forecast Accuracy the Institute for Research and Community Services, and the
Data Data (%)
1 January 13,700 16,745 77.77%
Computational Intelligence Laboratory, Indonesia University
2 February 12,148 13,337 90.21% of Education.
3 March 15,837 14,392 90.88%
4 April 14,555 13,862 95.24%
5 May 15,764 15,624 99.11% REFERENCES
6 June 14,135 13,230 93.60% [1] M.K. Evans, Practical Business Forecasting, Blackwell Publishers Ltd,
7 July 16,705 13,586 81.33% USA, 2003.
8 Agust 15,999 12,034 75.22%
[2] R.S. Pressman, Software Engineering Software Engineering: A
9 September 15,371 15,140 98.50% Practitioner’s Approach Fifth Edition, McGraw-Hill, New Jersey, 2001.
10 October 15,406 17,484 86.51%
11 November 14,147 15,363 91.40% [3] V.L. Sauter, Decision Support Systems for Business Intelligence, 2nd
Edition, John Wiley & Sons, Inc., New Jersey, 2011.
12 December 16,032 13,123 81.86%
Average 88,47% [4] S. Makridakis, C.S. Wheelwright, and R.J. Hyndman, Forecasting:
Methods and Applications, 3rd Edition, John Wiley & Sons, Inc., New
The worst forecasting for Kahuripan is in August, while the Jersey, 1998.
best is in May (Table VIII). The forecasting of Kahuripan is the [5] N.A. Elmunim, M. Abdullah, A.M. Hasbi, and S.A. Bahari, “Short-term
second best forecasting by this method. forecasting ionospheric delay over UKM, Malaysia, using the Holt-
Winter method”, In Space Science and Communication (IconSpace),
Fig. 8 shows the results of recapitulation average 2013 IEEE International Conference on, IEEE, pp. 106-109, July 2013.
forecasting accuracy passengers of PT KAI Company for 2015 [6] K.P.G. Alekseev and J.M. Seixas, “Forecasting the Air Transport
Demand for Passengers with Neural Modelling”, in Proceedings of the
100% VII Brazilian Symposium on Neural Networks (SBRN’02). Federal
87% 85% 88% University of Rio de Janeiro, Brazil, 2002.
80% 91% [7] D.C. Montgomery, Introduction to Time Series Analysis and
Forecasting, A John Wiley &. Sons, Inc., New Jersey, 2008.
60% 70% [8] H. Javedani, M.H. Lee, An Evaluation of Some Classical Methods for
Forecasting Electricity Usage on Specific Problem. Proceedings of the
40% Regional Conference on Statistical Sciences 2010 (RCSS’10), pp. 47-56,
June 2010.
20%
[9] V. Kolar, Optimal values of alpha, beta and gamma are calculated using
0% a bounded optimization formulation, Bengaluru, Karnataka, India, 2015.
Argo Wilis Turangga Mutiara Pasundan Kahuripan [10] D. Tenne and T. Singh, Optimal Design of α-β-γ Filters, Center for
Selatan Multisource Information Fusion State University of New York at
Buffalo, 1999.
Fig. 8. Graph Average of Accuracy

202

You might also like