Comparison of Exponential Smoothing Methods in Forecasting Palm Oil Real Production

Transcription

1 Journal of Physics: Conference Series PAPER OPEN ACCESS Comparison of Exponenial Smoohing Mehods in Forecasing Palm Oil Real Producion To cie his aricle: B Siregar e al 2017 J. Phys.: Conf. Ser View he aricle online for updaes and enhancemens. Relaed conen - Analysis of Palm Oil Producion, Expor, and Governmen Consumpion o Gross Domesic Produc of Five Disrics in Wes Kalimanan by Panel Regression E Sulisianingsih, M Kifiah, D Rosadi e al. - Performance Evaluaion on Oo Engine Generaor Using Gasoline and Biogas from Palm Oil Mill Effluen Irvan, B Trisaki, T Husaini e al. - Decision Suppor Model for Selecion Technologies in Processing of Palm Oil Indusrial iquid Wase Aulia Ishak and Amir Yazid bin Ali This conen was downloaded from IP address on 27/12/2017 a 22:18

2 Inernaional Conference on Compuing and Applied Informaics 2016 Inernaional Conference on Recen Trends in Physics 2016 (ICRTP2016) Journal of Physics: Conference Series 755 (2016) doi: / /755/1/ Comparison of Exponenial Smoohing Mehods in Forecasing Palm Oil Real Producion B Siregar 1, I A Buar-Buar 1, RF Rahma 1, U Andayani 1 and F Fahmi 2 1 Dep. of Informaion and Technology, Faculy of Compuer Science and Informaion Technology, Universiy of Sumaera Uara Jl. Dr. Mansur 9 Medan Indonesia 2 Dep. of Elecrical Engg, Faculy of Engineering, Universiy of Sumaera Uara Jl. Dr. Mansur 9 Medan Indonesia fahmimn@usu.ac.id Absrac Palm oil has imporan role for he planaion subsecor. Forecasing of he real palm oil producion in cerain period is needed by planaion companies o mainain heir sraegic managemen. This sudy compared several mehods based on exponenial smoohing (ES) echnique such as single ES, double exponenial smoohing hol, riple exponenial smoohing, riple exponenial smoohing addiive and muliplicaive o predic he palm oil producion. We examined he accuracy of forecasing models of producion daa and analyzed he characerisics of he models. Programming language R was used wih seleced consans for double ES (α and β) and riple ES (α, β, and γ) evaluaed by he echnique of minimizing he roo mean squared predicion error (RMSE). Our resul showed ha riple ES addiives had lowes error rae compared o he oher models wih RMSE of 0.10 wih a combinaion of parameers α = 0.6, β = 0.02, and γ = Inroducion Palm oil is one of planaion crops ha have an imporan role for he planaion subsecor. Indusrial developmen of palm oil benefi in increasing he income of farmers and sociey, producion of he raw maerial ha creae added value for he counry, he expor of CPO which generae foreign exchange, and i provides job opporuniies for more han wo million workers in he various subsysems. Palm oil planaions in Indonesia are managed in he form of small planaions and large farms. arge esae consiss of a counry esae as a Perkebunan Nusanara (PTPN) and privae esaes [1]. For each planaion, here will be monhly evaluaion processing regarding qualiy objecives. This was done o look a he performance of he producion for he monh wheher i has reached a producion arge or no. In order o achieve he producion arge, he availabiliy of informaion abou real producion in previous years plays a very imporan role. Targe producion in he fuure can be expeced only if he previous year's producion daa is available and of good qualiy. In he planaion producion arges are calculaed by looking a he producion of previous years done by manual calculaions. Therefore we need a mehod ha can provide informaion on producion forecass as a reference in deermining he company plans more realisic producion arges. Research on Palm oil producion forecass have been done by several mehods. Mehods exised included he use mehods of causal and ime-series. Research conduced in 2013 using wo mehods o compare he producion forecass Palm oil is used in he form of muliple regression mehods of causal Conen from his work may be used under he erms of he Creaive Commons Aribuion 3.0 licence. Any furher disribuion of his work mus mainain aribuion o he auhor(s) and he ile of he work, journal ciaion and DOI. Published under licence by d 1

3 Inernaional Conference on Compuing and Applied Informaics 2016 and ime-series mehods such as exponenial smoohing. In he es, he value of Mean Absolue Percenage Error (MAPE) obained using regression is 21:08% while he value of MAPE obained using exponenial smoohing mehod is 12.78% per year [2]. In anoher sudy, a research was conduced on he producion predicion using neural neworks wih back propagaion algorihm. Usually neural nework was uilized in image based analysis such as meer reading [3] or image recogniion [4]. The sudy used seven daa as a variable based on he qualiy of he land as well as rainfall, aliude, and slope, age of he plan, rocks, solium, and he acidiy of he soil. Experimen wih several layers o ge he bes parameers: 3 layer, 4 layer and 5 layer. The bes resuls obained in experimen wih 3 layer ieraion o 30,000, wih he learning rae by 0.9. Training resuls obained wih R2 = and RMSE = and es resuls wih R2 = and RMSE = [5]. Anoher research in 2015 prediced Palm oil producion using RBF neural nework. The bes resuls for Palm oil producion forecass obained hrough a combinaion of parameers: he number of variables = 1, he number of inpu nodes = 5, he value of learning rae = 0.75 and he maximum epoch value = 3000 wih he resuls of error MAPE = 11.75% [6]. Anoher research in 2011 uilized exponenial smoohing o forecas sea surface emperaures, he researchers prediced he nex year based on he resuls of he model were compared o analyze he characerisics of he model and also comparing he values obained using hree mehods: Single Exponenial Smoohing, Double exponenial smoohing Hol and Hol -winers. Of he hree mehods produced he bes performance on he mehod and Hol-Winers exponenial smoohing Double Hol [7]. In his sudy we compared several exponenial smooohing mehods used o predic he acual producion of palm oil such as single ES, double exponenial smoohing hol, riple exponenial smoohing, riple exponenial smoohing addiive and muliplicaive and analyze he performance. 2. Maerial and Mehods The general archiecure of his sudy can be seen in Figure Daa Collecion This secion shows he rerieval of daa, hen he daa ransformaion process o he creaion of a daa plo. The daase used in his sudy is a documen from Palm oil producion. The daa used was secondary daa, i.e daa Palm oil planaion PTPN III Planaions Sei Merani which have been colleced from The daa was sored in he form of csv and ables ha no normalizaion of daa unil he daa can be included in a daabase or daa in comma-separaed values exension *.csv. Collecing daa in his sudy was using survey echniques, performed direcly in he office Facory Palm oil (MCC) Sei Planaions Merani. The daa consised of daa from he producion of fresh frui bunches (FFB) Palm oil, palm oil producion daa and core daa obained from FFB oil ha had been processed. The nex sage was he ransformaion of daa, where he daa was normalized so ha daa values can be processed easily. These daa were convered ino a value in he range of 0.1 up o 0.9 by using equaion 1 [7]. x = 0.8 ( x a) b a (1) The nex sep was o creae a daa plo. Plo Palm oil producion daa in in Planaions Sei Merani which had been ransformed can be seen in Figure 2. Based on Figure 2, Palm oil producion saring from he year 2010 o 2014 were likely o increase. Only in early 2014 Palm oil producion declined. Based on informaion from he managemen of he farm in 2014, here was a drasic climae change, so ha no only he planaions Sei Merani are decreased, bu also almos all he planaions ha exis in PTPN III experienced he same hing. Overall i can be concluded ha he producion of oil in he planaion Palm Sei Merani has a posiive rend (increasing). 2

4 Inernaional Conference on Compuing and Applied Informaics 2016 Figure 1. General Archiecure 3

5 Inernaional Conference on Compuing and Applied Informaics Exponenial Smoohing Mehod Double exponenial smoohing (DES) Hol. In DES Hol mehod, he smoohed rend componen separaely using differen parameers, namely α and β. In his echnique he value of he rend can be smoohed by using differen weighs. However, hese wo parameers need o be opimized so he search for he bes combinaion of parameers is more complicaed han using only one parameer. In addiion, he componens of season in his echnique are no aken ino accoun. Here is he process underaken Double exponenial smoohing hol Hol. Sep 1: Deermine he iniial iniializaion value. Value is saionary, and he rend is beginning o use equaion 2 o equaion 3 [8]. 1 S X aau S (X1 X2... X) (2) T ( X X ) ( X X ) ( X... X ) Sep 2: Iniialize he value of he parameers a and b where each range is beween 0-1. In his sudy, we used he mehod of rial and error o produce he value of α and β are opimal based on he value Roo Mean Square Error (RMSE) of he mos minimum. Values α and β are opimal will be deermined direcly by he applicaion program ha has been designed. Sep 3: Calculaion of saionary of daa can be done using Equaion 4 [8]. S = αx + (1 α)(s 1 + b 1 ) (4) \ Sep 4: Calculaion of rend daa can be done using he equaion 5 [8]. T = β(s + S 1 ) + (1 β)b 1 (5) Sep 5: Once he daa is saionary and rendy value has been esablished, hen he forecass can be found using equaion 6 [8]. F +m = S + b m (6) Sep 6: Then calculae he forecas errors using RMSE and MAPE. If MAPE produced no he smalles hen repea sep wo and so on unil we go he smalles error value. Sep 7: Finding he forecass for he nex few periods, for example, for 12 monhs. By using he S parameer (level), and T (rend) of he las daa specify he lengh of he forecas period p. By re-using Equaion 6, se he value of m = 1, o find he forecass one period or one monh following he nex. Suppose for 1 year ahead or equal o 12 periods ahead hen se he value of m wih m = 1, m = 2 and m = 3 and so on up o m = Triple exponenial smoohing Addiive and Muliplicaive. From he analysis ha has been done before, i was found ha he amoun of oil producion Palm Planaions Sei Merani from he firs monhs of 2010 unil he final monhs of 2014 was influenced by rends and seasonal facors. I is seen from Figure 2, in he sevenh monh he number of producion Palm oil spiked higher hen monh-a monh afer he demand will decrease and hen going back upward as i approached he seasonal period in he nex year. This siuaion repeaed every year, which means influenced by seasonal facors. Because he daa is affeced by seasonal facors, he oher mehods ha can be implemened are Triple exponenial smoohing mehod. The following processes are carried ou in Triple exponenial smoohing (TES) Addiive and Muliplicaive. (3) 4

6 Inernaional Conference on Compuing and Applied Informaics 2016 Sep 1: Deermine he iniial iniializaion value. ie saionary value, rend and seasonally iniial value using equaions 2, 3 and 7 [8]. X k I, k 1,2,..., (7) S Sep 2: Iniialize he value of he parameers α, β and γ where each range is beween 0-1. Sep 3: Calculaion of TES daa is saionary wih addiive and muliplicaive mehod can be performed using equaions 8 and 9 [8]. Adiive : S X I ) (1 )( S T ) (8) ( 1 1 X Muliplicaive: S ( 1)( S 1 T 1 ) (9) I Sep 4: Calculaion of rend daa can be performed using equaions 10 and 11 [8]. Adiive : T ( S S 1 ) (1 ) T (10) 1 Muliplicaive: T ( S S 1 ) (1 ) T (11) 1 Sep 5: For he calculaion of he iniial value I, in one cycle of he firs season (12 firs period), can be performed using Equaion 12 and 13 [8]. Adiive : I ( X S ) (1 ) I (12) X Muliplicaive: I ( 1 ) I (13) S Sep 6: Wih all hree parameer values ha have been obained, hen he forecass can be found by Equaion 14 [8]. F ( S T m) I (14) m m Sep 7: Then calculae he forecas errors using RMSE and MAPE. If MAPE produced no he smalles hen repea sep wo and so on unil you ge he smalles error value. Sep 8: Finding he forecass for he nex few periods, for example, for 12 monhs. By using he S parameer (level), T (rend), and I (seasonal) of he las daa specify he lengh of he forecas period p by re-using Equaion Oupu Afer obaining he approximae value of he desired period, he nex sep is o normalize he daa, which reurns he daa ino he form of he iniial uni using Equaion 15. x = (x 0,1)(d maks d min ) 0,8+ d min (15) Then he resuls of normalize will be processed ino he form of plos and ables o faciliae he user in observing he forecas. Implemenaion of his mehod is done using he sofware R Programming. 5

7 CIENT SIDE SERVER SIDE Inernaional Conference on Compuing and Applied Informaics 2016 R Programming R Shiny Package Server.R Inpu Folder Inpu Oupu Ui.R Oupu Foder Web Browser User Inerface Figure 3. Process in R Programming End User Based on Figure 3, here are wo pars in he process of forecasing sysem using he R sofware, he server side and clien side. The enire core code of he implemenaion process of exponenial smoohing is sored in a file named server.r. The process of daa inpu using csv file exension sored in he inpu folder, which is hen called direcly by server.r. While user inerface.r handle he process of how he oupu is displayed o he user. For he oupu of his sysem, he R language provides an easy way o inegrae server.r and ui.r use shiny library. The resuls of exponenial smoohing configuraion can be accessed direcly by he web browser by he user. 3. Resuls Several ess were conduced o deermine he performance of he sysem in forecasing realizaion Palm oil producion using exponenial smoohing mehod. Based on he draf research seps, hen each daa will be processed using hree mehods already menioned. Tesing parameers for all hree mehods wih differen combinaions of parameers were used. Besides being able o manually selec a combinaion of parameers, he sysem can also look for opimizing he parameers ha produce he smalles error. By execuing he equaions exponenial smoohing models wih he help of R sofware, a comparison was made beween he resuls obained form he model wih he resuls of observaions for each mehod. The daa on producion of fresh frui bunches (FFB), a consan value of α, β and γ are deermined by he sysem are respecively α = 0.7 and β = 0.01 for Hol DES mehod, α = 0.6, β = 0, 02 and γ = 0.02 for TES addiive mehod and α = 0.5, β = 0.02 and γ = 0.02 for TES muliplicaive mehod. The parameer opimizaion resuls using hese hree mehods can be seen in Figure 4. Figure 4 shows ha he model DES hol, TES addiives and mulipikaif was very precise, especially for some of he iniial daa. Forecas daa generaed by addiive and muliplicaive mehod of ime January 2010 unil December 2014 was very precise compared observaion daa. Daa FFB producion is a seasonal cycle of daa will be repeaed he following year, he muliplicaive and addiive models ry o predic he repeiion of he producion cycle. The analysis was conduced based on he accuracy of he model DES, TES TES addiive and muliplicaive FFB producion daa using saisical insrumens such as he Roo Mean Square Error (RMSE), Mean Absolue Error (MAE), Mean Absolue Percenage Error (MAPE) [8]. RMSE, MAE and MAPE refers o he magniude of he error rae (errors) of an esimae, he smaller he value of hese insrumens, he beer he forecass were made. The resuls of he evaluaion of he forecasing models each daase wih all hree mehods menioned can be seen in Table 1. 6

8 Inernaional Conference on Compuing and Applied Informaics 2016 Figure 4. DES Hol, TES Adiive, and TES Muliplicaive for producion Daase Producion of FFB Producion of Palm Oil Producion of Palm Kernel Table 1 Error esimaion for differen forecasing models Exponenial smoohing Hol (α=0.7;β=0.01) Adiive (α=0.6;β=0.02; γ=0.02) Muliplicaive (α=0.58;β=0.02 ; γ=0.02) Esimaed Error MAE RMSE MAPE Hol (α=0.74;β=0.01) Adiive (α=0.7;β=0.01 ; γ=0.03 ) Muliplicaive (α=0.64;β=0.02 ; γ=0.03 ) Hol (α=0.75;β=0.01) Adiive (α=0.68;β=0.01 ; γ=0.03 ) Muliplicaive (α=0.72;β=0.01 ; γ=0.21 ) An insrumen of error such as MAE, RMSE, and MAPE (%) of he addiive model is smaller han Hol and muliplicaive models for daa in seasonal cycles on he producion daa Planaions Sei Merani. Model riple exponenial smoohing (Hol-Winers addiive and muliplicaive) has good precision, especially he addiive mehod for daa ha are seasonal cycle such as producion daa FFB, palm oil, and palm kernel as shown in Table 4. From Table 4 he exponenial smoohing models, which is he mehod TES addiive is he bes mehod which produces he smalles error in forecasing for he daa ha is periodic wih each value RMSE = 0:09 o producion daa FFB, RMSE = 0.10 for he daa producion of palm oil, and RMSE = 0.10 for he daa producion of palm kernel. 7

9 Inernaional Conference on Compuing and Applied Informaics 2016 Afer he comparison of he forecass of he resuls of field observaions from he period January 2010 o December 2014 described Figure 4 and Table 1, hen we made furher forecass. In he forecass for he nex wo years ie from January 2015 o December 2016 for all hree mehods of exponenial smoohing was provided. Advanced forecasing resuls can be seen in Figure 5 (a) o (c). Figure 5. Graph forecasing for (a) FFB (b) Palm Oil (c) Palm Kernel 4. Discussions Based on he evaluaion of sysem forecasing resul i was obained ha he bes resuls o forecas FFB producion, palm oil, and palm kernel produced by he mehod of riple exponenial smoohing addiive hrough a combinaion of parameers in a row is α= 0.6, β = 0:02, and γ = 0:02 o FFB producion wih he resuls of error RMSE = 0:09, α = 0.7, β = 0:01, and γ = 0:03 o palm oil producion wih he resuls of error RMSE = 0:10, and α= 0.68, β = 0:01, and γ = 0:03 o palm kernel producion wih he resul of error RMSE = From he es resuls i is shown ha each of he daa oupu has he characerisics of ime series daa differenly ha each of he daa oupu forecass have differen parameers. The use of double exponenial smoohing mehod hol o predic wha kind of daa Palm oil producion is no appropriae because he observaion daa has a seasonal componen. Forecass for he daa conaining seasonal componens should be solved by using he mehod of riple exponenial smoohing. For furher developmen, he sudy suggesed he same case can use oher mehods ha produce resuls forecas by he smaller error value. The forecasing is a reference o be considered in seing producion arges ha required high accuracy. 5. Conclusion Compared o oher exponenial smoohing mehods, Triple exponenial smoohing addiive can provide beer forecasing of palm oil real producion wih suggesed parameers as he resul from his sudy. References [1] Baubara EA 2015 Implemenasi algorima learning vecor quanizaion pada prediksi produksi Palm oil di PT. Perkebunan Nusanara I Pulau Tiga under publicaion [2] Kacaribu RA 2013 Aplikasi Peramalan Produksi Palm oil dengan Meode Regresi Ganda dan Exponenial Smoohing. under publicaion [3] Trianoro T, Baubara FR and Fahmi F 2014 Image based waer gauge reading developed wih ANN Kohonen Proceedings of 2014 Inernaional Conference on Elecrical Engineering and Compuer Science, ICEECS [4] Siswano A, Tarigan P and Fahmi F 2013 Design of conacless hand biomeric sysem wih relaive geomeric parameers Proc. of rd In. Conf. on Insrumenaion, Communicaions, Informaion Technol., and Biomedical Engineering: Science and Technol. for Improvemen of Healh, Safey, and Environ., ICICI-BME [5] Hermanoro and Purnawan RA 2009 Prediksi produksi Palm oil berdasarkan kualias lahan menggunakan model Arificial Neural Nework (ANN) Agroeknose 4(2):

10 Inernaional Conference on Compuing and Applied Informaics 2016 [6] Jannai R 2015 Prediksi produksi panen Palm oil menggunakan jaringan saraf radial basis funcion (RBF) under publicaion [7] Salamena GG 2011 Pengujian model peramalan dere waku sea surface emperaure (SST) eluk ambon luar dengan meode exponenial smoohing Oseanologi dan imnologi di Indonesia 37(1): [8] Maranaa 2013 Peramalan jumlah produksi kakao di sumaera uara dan konsumsi kakao di Indonesia dengan pemulusan eksponensial ganda meode linier sau parameer dari brown. under publicaion [9] Siang JJ 2009 Jaringan saraf iruan & pemrogramannya menggunakan MATAB. Penerbi ANDI: Yogyakara. [10] Makridakis S, Wheelwrigh SC and McGee VE 1999 Meode dan Aplikasi Peramalan Erlangga: Jakara. 9