com hekim tr
Transcription
com hekim tr
A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey Rüstü YAYAR Faculty of Economics and Administrative Science Gaziosmanpasa University, Tokat, Turkey [email protected] Mahmut HEKIM Tokat Vocational School Gaziosmanpasa University, Tokat, Turkey [email protected] Veysel YILMAZ Suşehri Timur Karabal Vocational School Cumhuriyet University, Sivas, Turkey [email protected] Fehim BAKIRCI Faculty of Economics and Administrative Science Atatürk University, Erzurum, Turkey [email protected] Abstr ct In this study, the electric energy demand of Tokat province was estimated by means of ANFIS and ARIMA techniques. Seven different forecasting experiments were implemented for the subscriber groups and the consumption of electric energy which is the dependent variable. The electric energy demand of the province for the first six months of the year 2011 was estimated by means of ANFIS and ARIMA techniques. The obtained results were compared and interpreted in order to illustrate the forecasting success of these techniques. We showed that the ANFIS is more appropriate than the ARIMA in point of the forecasting of electric consumption. Keywords: Electric energy consumption, Forecasting, ANFIS, ARIMA, and Tokat. Jel odes: C22, C45, Q47 Volume 1 Number 2 July 2011 87 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI Introduction Nowadays, energy is an important source of life. By entering into the life of mankind in 1880s, the electric energy gradually became an indispensable part of modern life and industry (Şekerci Öztura, 2007). In order to generate the electric energy, which is a secondary energy source, the support of primary energy sources is needed. Electric energy is widely used in many fields, and a large part of energy resources for the benefit of the people are converted into electric energy (Demir, 1968). The electric energy demand has been continuously increasing in parallel with the growing population, urbanization, industrialization, technologic deployment, and enhancement of welfare. The spread of the usage fields of electric energy which is one of the most important parts of all types of economic activities increases the electric energy demand. Also, since the distribution network provided a great deal of the electric energy of even the smallest residential areas, the share of electric energy in the total energy consumption increased (Kılıç, 2006). The province of Tokat is geographically located between 39-51, 40-55 North latitudes and 35-27, 37-39 East longitudes, in the inner side of middle Black Sea part of Black Sea Region. Covering the 1.3% of mainland Turkey, elevation from the sea level of the province is 623m and surface area of it is 9.958km² (Governorship of Tokat, 2006). There are 12 districts of Tokat province including with the central district¹. Historical periods of the province consist of Hattie, Hittite, Phrygian, Med, Persian, Alexander the Great, Roman, Byzantine, Arabian, Danishmend, Anatolian Seljuk, Mongolian, Ilkhanid, Ottoman Governments and Emperors (Provincial Department of Environment And Foresty of The Governorship of Tokat, 2007:1). The province of Tokat became a province with the proclamation of the republic in 1923 (Turkish Statistical Institute, 2010:10). By 2010 the population of the province is 550.703 (Turkish Statistical Institute, 2010). The province has a wide potential of both agriculture and other sectors (such as tourism). In the economic structure of the industry, agriculture, livestock sector plays an important role. Particularly in the food industry, rock and land-based industries, forest products industry and in recent years, textile weaving and garment sector is the backbone of the economy in Tokat (Provincial Department of Environment And Foresty of The Governorship of Tokat, 2007:130). 88 Journal of Economic and Social Studies A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey Electric energy was firstly provided in the city center of Tokat in 1935 by the Hydroelectric power plant which was built on Aksu for the city lightening. This power plant consists of 2 tribunes with 175 horsepower (HP), and each tribune operates by 230 m³ water passing through the water channel. The power plant provided electric energy of 135 kWh for 1580 subscribers. However, the electricity production was insufficient for the city. Because of Almus Hydroelectric power plant built in 1966, Tokat had continuously the electric energy, and still it provides the electric energy of the city. Besides, transmission lines were renewed and electric energy of every part of the city was provided by using 18 transformers (Governorship of Tokat, 2006). The installed power plant in Tokat consists of Almus, Köklüce, Ataköy Hydroelectric power plants within Tokat. Almus Hydroelectric power plant became operational in 1966 and Number of Unit – Power is 3X9 MW and its installed power is 27 MWAnnual production of the plant is 100 GWh. Köklüce entered service in 1998 and Number of Unit – Power is 2X46 MW and its installed Power is 90 MW. Annual production of the power plant is 588 GWh (Electricity Generation Corporation, 2011). The construction of Ataköy Hydroelectric power plant dam was completed in 1977. Having 5.5 MW Power, its annual production is 8 GWh (VIII. Regional Directorate of State Hydraulic Works, 2011). In this study, electric energy demand are estimated by Adaptive Neuro-Fuzzy Inference System (ANFIS) and Autoregressive Integrated Moving Average (ARIMA) techniques by using the electric energy consumption data of Tokat province in the time of period between January 2002 and December 2010. Matlab 7.04 package program is used for the ANFIS model and Minitab 14.0 package program is used for the ARIMA model. Aim This study aims at contributing to the planning of supply by estimating electric demand in the future. The demand for electric energy was forecasted in Tokat province by means of the ANFIS and the ARIMA models. The generation planning of electric energy is very important because it cannot be stored, and therefore must be consumed shortly after its generation. If these kind local studies were generalized into all country, it helps the supply planning and provides a more effective usage of resources. Therefore we estimated the electric energy demand by ANFIS and ARIMA models for Tokat province. Volume 1 Number 2 July 2011 89 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI Literature Energy consumption and demand are among the most debated topics, and many studies have been made available in the literature. It is seen that the studies especially focused on the causality between the electricity consumption and economic growth and controversial results has been achieved. There are scarcely any studies on electric energy consumption and demand through ANFIS and ARIMA models. These models are often used in engineering studies. There are many studies in various fields by using the fuzzy logic method. In one of them, Tufan and Hamarat (2003) analyzed the “Aggregation of The Financial Ratios of publicly-traded companies Through Fuzzy Logic Method”. The first detailed study examining the demand for electricity with the help of econometric models is the work of Houthakker (1951). The study includes the econometric analysis of the household electricity demand through cross-section data from the period of 1937-1938 for 42 residential center in England. Another study was conducted by Fisher and Kaysen. Fisher and Kaysen (1962) using time-series and cross sectional data, examined the demand for electricity with the help of multiple regression and analysis of covariance. Electric demand was taken up in four components, including electricity demand, household electricity demand, industrial electricity demand and the short and long period determinant of them. Accordingly, short run is in question if the stock of electric appliances is fixed and long run is in question if the stock is variable. In the case of industry, short run is in question if there is an assumption that technology is invariable and long run is in question if there is an assumption that technology is variable (Tak, 2002). Al-Garni and Javeed Nizami (1995) developed a model of artificial neural networks for electric energy consumption with the data from seven years such as temperature, moisture, solar radiation and population. After the comparison of the model of artificial neural networks with the regression model, it was revealed the model of artificial neural networks was a better forecasting model. Al-Garni and Abdel-Aal (1997) estimated the Electric Energy consumption for five years on the east of Saudi Arabia for the consumption on the sixth year, developing monthly ARIMA models by using the univariate and Box-Jenkins time-series analysis. When ARIMA models are compared to the abductive network machinelearning models, it is seen that ARIMA models needs less data and coefficient and give better results as well. 90 Journal of Economic and Social Studies A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey Brown and Koomey (2002) examined the increase in electric demand in their work named “Electric Use in California: Past Trends and Present Usage Patterns. They took sector (settlement, commercial, industrial, agricultural and other) of electric consumption from the previous period as data. They brought forward that there had been a great increase in the electric demand in 1990, compared to 1980 and that it stems from the increase in buildings and the tendencies in the building sector (Enduse Forecasting and Market Assessment, 2011). It can be seen that the studies on the modeling of the energy consumption and demand, which is also crucial for the economy of Turkey, accelerated in 2000s. There are not many studies, done widely on the provincial (regional) electricity consumption and demand. State institutions and organizations of which subject of activity is electricity became obliged to state the electricity consumption of the provinces on their annual report on the basis of subscriber groups after “Legislation On The Annual Reports Drawn Up By The Public Administrations” were published by the Ministry of Finance on the Official Gazette dated 17.03.2006 and became effective on 01.01.2006 (Official Gazette, 2006). These institutions conduct studies also on the monthly and yearly electric data which must be sent to Turkish Statistical Institute (TÜİK) and Ministry of Energy and Natural Resources (ETKB) by these institutions even though these data changes in the ensuing years. In the work of Terzi (1998) which analyzes the relationship between the electric consumption and economical growth for the period of 1950-1991, the relationship between the electricity consumption and of the commerce house, industry and household and economical growth; the long period relationship between the variables were determined through the Engle-Granger cointegration method and the short run dynamics were analyzed through the debugging tool. It was determined through this econometric method that the income and price elasticity were in fact inelastic. A meaningful and two-way relationship between electricity consumption and economical growth came along in the business and industry sector. Sarı and Soytaş (2004) employed the technique of generalized forecast error variance decomposition and came to the conclusion that the electricity demand and variance in national income growth are as important as employment (Lise & Van Monfort, 2005). Çebi and Kutay (2004) used artificial neural networks while estimating the long run electric energy consumption and compared the results to the Box-Jenkins models and regression technique. The results revealed that using artificial neural networks was a good forecasting method for electric energy consumption. Volume 1 Number 2 July 2011 91 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI In addition, we must specify the MAED (Model for Analysis of Energy Demand) model study of Ministry of Energy and Natural Resources, the most important study that was conducted in our country, which reveals the medium and long run general energy demand and electric energy demand in this demand. Research Method In this study, monthly electricity consumption data of Tokat province was used. This data covers the period between January 2002 and December 2010. The subscriber groups consist of private houses, industry, business firms, government agencies and other subscribers. The total electricity used in agricultural irrigation, fresh water, work-sites, temporary activities, state-owned enterprises, municipalities, internal activities, prefectures, sanctuaries, and local government lightening systems was also included in these groups. Seven different experiments were implemented for investigating the success of the ANFIS and the ARIMA models in the forecasting of the total electric energy consumption of all subscriber groups. Time Series and Box-Jenkins Forecasting Model Time Series A time series is simply a sequence of numbers collected in regular intervals (as day, month and year) over a period of time (Dikmen, 2009: 227). Time series monitors the motion of a variable in a time sequence. For example, monthly unemployment ratio, monthly increase ratio of money supply, annual inflation ratio, monthly electric use, etc. Time series can be also used as a resource of knowledge acquisition and a method for forecasting the future. While in the evaluation process, analysis is important in degrading the trend, growth trend, seasonality, cyclical, and irregular fluctuations (Bozkurt, 2007). The selection of method used for forecasting the future values depends on the estimation of purpose, type and elements of time series, amount of data, and the length of the estimation period (Asilkan & Irmak, 2009). 92 Journal of Economic and Social Studies A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey Box-Jenkins Forecasting Model Box-Jenkins method is the most widely used model for stationary time series modeling. For the implementation of Box- Jenkins method, the time series must be stationary (with constant mean, variance and autocorrelation). If the series is not stationary, it should be made stationary by taking the difference of a few times (Gujarati, 2009). Box-Jenkins method is based on the principle that each time series is a function of past values and may only be explained by means of them. Some assumptions cannot be applied based on the econometric models, but there is not any restrictive assumption for Box-Jenkins method (Bircan & Karagöz, 2003). In this method; • In contrast to the regression models that explain yt with a k number of explanatory variables of x1, x2, x3,…, xk, • The dependent variable Yt can be explained by its own past or lagged values and stochastic error terms. The most important stage of the Box-Jenkins method is the selection of the appropriate ARMA (p, q) model by examining the autocorrelation and partial autocorrelation coefficients. Experience of the researcher is very important because of this phase is not able to determine mechanically. If the time series is not stationary, artificial autocorrelations will prevent the model to determine. Non-stationary time series is transformed stationary time series by logarithmic transform or taking differences. Autocorrelation and partial autocorrelation coefficients of the distribution can be examined with the help of graphs (autocorrelation function (ACF) and partial autocorrelation function (PACF). When the autocorrelation coefficients are seen to be approaching zero exponentially, AR model must be applied; while the partial autocorrelation coefficients are realized to be approaching the same level mentioned above, then MA model must be used; if both of these approach zero exponentially; ARMA model must be applied in this situation. In ARMA model, the degree of AR is determined by the number of partial autocorrelation coefficients (p), while the degree of MA is determined by the number of autocorrelation coefficients (q) (Önder & Hasgül, 2009: 65–66). Volume 1 Number 2 July 2011 93 (PACF). When the autocorrelation coefficients are seen to betoofapproaching zerozero exponentially, Autocorrelation andand partial autocorrelation coefficients the distribution cancan be be examined (PACF). When the autocorrelation coefficients are seen be approaching exponentially, Autocorrelation partial autocorrelation coefficients of the distribution examined l autocorrelation function Autocorrelation and partial autocorrelation coefficients of the distribution can be examine AR model must be applied; while the partial autocorrelation coefficients are realized to be with the help of graphs (autocorrelation function (ACF) and partial autocorrelation function AR model must be applied; while the partial autocorrelation coefficients are realized to be with the help of graphs (autocorrelation function (ACF) and partial autocorrelation function ching zero exponentially, with the help of partial graphsautocorrelation (autocorrelationcoefficients function (ACF) and partial autocorrelation functio Autocorrelation and ofbemust the distribution can be approaching theto same level mentioned above, thenthen MA model be used; if both of examined these (PACF). When the autocorrelation coefficients are seen to approaching zero approaching the same level mentioned above, MA model must be used; ifexponentially, both of these (PACF). When the autocorrelation coefficients are seen to be approaching zero exponentially, icients are realized be (PACF). When the autocorrelation coefficients are seen to be approaching zero exponentiall withmodel the help of graphs (autocorrelation function (ACF) autocorrelation function approach zero exponentially; ARMA model must be applied inand this situation. AR must be applied; while the partial autocorrelation coefficients areare realized to to be be approach zero exponentially; ARMA model must be applied inpartial this situation. model must be applied; while the partial autocorrelation coefficients realized be used; ifAR both ofmodel these AR must be applied; while the partial autocorrelation coefficients are realized to b (PACF). When the autocorrelation coefficients are seen to be approaching zero exponentially, Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI approaching the same level mentioned above, then MA model must be used; if both of these approaching thethe same level mentioned above, then MA model must be be used; if both of of these situation. AR approaching same level mentioned above, then MA model must used; if both model must bethe applied; while partial autocorrelation coefficients areautocorrelation realized to bethe In ARMA model, the degree of ARMA AR is the determined by number ofthis partial autocorrelation approach zero exponentially; model must bethe applied in in this situation. In approach ARMA model, degree ofARMA AR is model determined by the number ofsituation. partial zero exponentially; must be applied approach zero exponentially; ARMA model must be applied in this situation. approaching the same mentioned above, then MAbymodel must beofused; both of these coefficients (p), (p), while thelevel degree of MA is determined the number autocorrelation coefficients while the degree of MA is determined by the number of if autocorrelation of partial autocorrelation approach zero exponentially; ARMA model must be applied in this situation. (q)model, (Önder &the Hasgül, 2009: 65–66). ARMA the degree of AR is is determined byby thethe number of of partial autocorrelation coefficients (q) (Önder °ree Hasgül, 2009: 65–66). In ARMA model, of AR determined number partial autocorrelation umbercoefficients ofIn autocorrelation In ARMA model, the degree of ARthese is determined by the number of partial autocorrelatio ARMA models consists of four models, are AR, MA, ARMA and ARIMA. These coefficients (p), while the degree of MA is determined by the number of autocorrelation coefficients (p), while the degree of MA is determined by the number of autocorrelation coefficients (p), while the of (Demirel MA is determined by the numberautocorrelation of autocorrelatio In ARMA model, of AR isthese determined by the number of partial models will(Önder bethe explained in thedegree following: & et al.,MA, 2010). ARMA models consists ofdegree models, are are AR, MA, ARMA and ARIMA. These coefficients (q) &four Hasgül, 2009: 65–66). ARMA models of four models, these AR, ARMA and ARIMA. These coefficients (q)consists (Önder & Hasgül, 2009: 65–66). coefficients (q) (Önder & Hasgül, 2009: 65–66). coefficients (p), while the degree of MA is determined by the number of autocorrelation models will be explained in the following: (Demirel & et al., 2010). models will be explained in the following: (Demirel & et al., 2010). MA and coefficients ARIMA. These (q) (Önder &ofHasgül, 2009: 65–66). ARMA models consists four models, these areare AR, MA, ARMA andand ARIMA. These ARMA models consists of four models, these AR, MA, ARMA ARIMA. These ARMA models consists of four models, these are AR, MA, ARMA and ARIMA. The AR (p) Model AR models (p)models Model bebe explained in in thethe following: (Demirel && et et al.,al., 2010). AR (p) will Model will explained following: (Demirel 2010). models will consists be explained in the following: et al.,ARMA 2010). and ARIMA. These ARMA models of four models, these(Demirel are AR,&MA, In AR (p) model, Y value is the linear function of stochastic error term and weighted models will be explained (Demirel al., 2010). t in In AR (p)(p) model, Yt value is the the linear function of & stochastic errorerror term and and weighted AR Model In AR (p) model, Yt value is following: the linear function ofetstochastic term weighted AR Model ARaggragates (p) Model of past values in p period of the series. AR (p) model is shown as follows: aggragates of past values in p period of the series. AR (p) model is shown as follows: aggragates of past values in p period of the series. AR (p) model is shown as follows: error termAR and (p)weighted Model In AR of ofstochastic error term andand weighted Y Φt 1 Ymodel, Φ1 2YYΦ tt ... Φ linear δplinear at (1) Y(p) Yvalue ...ispYisthe Φ δlinear function atfunction (1) In (p) stochastic error term weighted t AR tΦ 11model, t t22Y t ppthe hown as follows: 2 value In AR (p)Yt model, Yvalue isYt the function of stochastic error term and weighte t aggragates of past values in p period of the series. AR (p) model is shown as follows: aggragates of past values in p period of the series. AR (p) model is shown as follows: (1) (1) aggragates of past values is in the p period of the series.ofAR (p) modelerror is shown follows: In AR model, Yt Yvalue stochastic term asand weighted Y(p) t Φ Yt1Y1 t Φ ... ...ΦYpΦ Ytplinear δfunction at (1) 1Φ Φ Y δ at t Y 2Y t2Y 2Y YY ptt− Y Y Φ Y Y Y Y 1t t 2 p Φ Y Φ Φ p t − p t p p Y Φ Φ Y ... Φ Y δ at (1)the Φ21follows: Φ2 ,...,is(1) tt−−22 ,..., 2 2,..., pthe aggragates ofabove values series. model is shown1 , as 1in tpast 1model, t11, t t 2,..., t pthe is values of(p) past observation, In the In the above model, istpof the values ofAR past observation, ,..., In the above model, , tt−−tp11t2,period ispthe the values past observation, ,,..., is Φ Ytp isthe Φ Φ2of Yt 2for ... values Φobservation, δδis atδthe (1) the coefficients the observation, is value the constant and Φ1 coefficients 1Yvalues t 1 p Yt of p past observation, and and atvalue isatthe error term. is constant the constant value is the error the values of ation, coefficients , Φ2 ,..., for isfor the YtpYtp Φ pΦterm. Yt Y1past Y1t past Φ Φ Y 2 Φ Φ 1 2 t t 2 at is the error term. In In thethe above model, , Y,t 1 ,..., is the values of past observation, , ,..., ispΦthethe 1Φ 2Φ Y Y above model, ,..., is the values of past observation, , ,..., t p theterm. above model, , t 2 ,..., is the values of past observation, 1 , 2 ,..., isp is th e and at is the In error δ is δ Y Φ MA(q) Model coefficients forfor thethe values of past observation, the constant value and at is the error term. Y Y MA(q) Model Φ Φ coefficients values of past observation, is the constant value and at is the error term. t p p 1 , t 2 ,..., 1 , at2 ,..., In the above model, is the valuesδofis past observation, is theterm coefficients for the tvalues of past observation, the constant value and is the error δ is the constant value and at is the error term. Model coefficients for the values past observation, In MA (q)MA(q) model, Yt value isofthe linear function of average pastpast errorerror terms in qinperiod MA(q) Model In MA (q) model, Y value is the linear function of average terms q period MA(q) Model MA(q) Model t backwards. MA (q ) model is shown as follows: backwards. MA (q ) model is shown as follows: error terms in In qModel period MA (q) model, Yt value is the linear function of average past error terms in q MA(q) In MA (q) linear of average past error terms in in q (2) period YtMA period t (q) model, at 1aYtt- 1Y1value a(q) t -... is... linear at-function Y at atMA - 22model athe In model, value is average past error terms q period tis q qq afunction tY - 1 2 2the tas - qfunction backwards. shown follows:of of In MA (q) model, is the linear average past error terms in q(2) perio t value backwards. MA (q ) model is shown as follows: backwards. MA (q ) model is shown as follows: (2) backwards. MAYt(qvalue ) model is asfunction follows:of average past error terms in q period In MA model, theshown linear Y(q) t a1 at1-aa1 t- 1 is ... (2)(2) at a at 2 at2-a2ta a ......q atq-aqt - aq t Y q (2) 2 a a a 1 ,21,,… t q Y at a a 2 , ,… t -1 2 ,……, t - qthe error backwards. MA (qt ,) model follows: t , , t -1t -1,is t tshown -t1-2 ,……, 2 as t - 2is t terms, -q In the above model, is theq error terms, ,is qthe is coefficients the coefficients ( In the above model, q error Yterms and at 1average at - 1 average 2of at - 2the ... q at - q (2) isthe series. terms t and error … , of isofthe coefficients at -qaseries. q q aist athe at t -1at -1at -2at -2 of the 1 2 t q , ,at, a,t -1 ,……, isa the error terms, ,1 ,… , , isthe coefficients In In thethe above model, 2 at -2 error terms, coefficients above model, - q the q the 2, … is the error terms, the above , , ,……, ,……, tis is the error terms, ,1 ,,… , isis is the coefficien In In the above model, model ARMA (p,q) Model is the average of the series. of error terms and ARMA (p,q) Model a is the average of the series. of of error terms and a a a t q q the coefficients of error terms and m is the average of the series. series. error terms ,……, of the is the error terms, 1 , 2 , … , is the coefficients In the above model,andt , t -1is, thet -2average ismost the average of thepreocess series. of error terms andthe ARMA model, the most stochastic preocess models, is the linear function of past ARMA (p,q) Model model, stochastic models, is the linear function of past ARMA (p,q) Model ARMA (p,q) Model observations and error terms. ARMA (p,q) model is generally shown as follows: observations and ARMA (p,q)error Modelterms. ARMA (p,q) model is generally shown as follows: linear function of past ARMA (p,q) Model ARMA model, the most preocess thethe function of(3) past Y Φ Y Φ YΦ ... Φstochastic - atδ -at θ 1 atθ-models, θ atθ- 22is θ qlinear atθ-linear Y Φ Y stochastic ... amodels, at -... ... the ARMA model, is t t 22 Y p YΦ t p Yt δ 11 qq a t - q n as follows: ARMA t 1 t 1model, 1 t 1 2the 2 most p preocess t -12models, 2 is thetmost stochastic preocess linearfunction functionof (3) ofpast pa observations and error terms. ARMA (p,q) model is generally shown as follows: ARMA(3) model, the most stochastic preocess models,is isgenerally the linearshown function past observations andand error terms. ARMA (p,q) model asofas follows: θ q at -q ARMA observations error terms. ARMA (p,q) model is generally shown follows: the most the as linear of past (3)(3) Yt model, Yt1Y ΦY2error Yt2Y2 tterms. 2Ystochastic ... ... Φ YtpYpt preocess (p,q) δ -δat θmodels, aθt1generally θ aist2-shown θfollows: aθtq-qafunction observations ARMA model Yt Φ 1Φ Y Φ ...pΦ -δat a2θt- 2a... ... 1 tand 1is -1a t qΦ 1Y 1Y Y p -1a2 θ tΦ - qa Φ p Φ p Y Y Φ Y Φ Φ Φ Y at θ ... θ t p Φ t t 2 t p observations and error terms. ARMA (p,q) model is generally shown as follows: 1 2 t 1 t 2 t 1 ,t 1 , ,..., 2 ,..., t 2 is the ppast t p past 1 t -1values, 2 t -2 q 2t ,..., -q observation , 1 ,,..., is the In third equation, is the observation values, is the ( In third equation, Φ pYt Φ1Yt 1 Φ2Yt 2 ... Φ p Yt p δ - at θ1 at -1 a θa2aat - 2a (3) a ... at -θ2 q at -aq t -q(3) at -the 1 , Φ2 ,..., t , t -1 δ is δthe q Φ error es, Φcoefficient thepast t ,, t -1t -,2 ,……, Y constant, is forisfor past observation values, Yt Y1 ,t Y1t Y2 t,..., is the constant, ,……, is the error coefficient observation values, Φ Φ Y Φ t p p Φ Φ 1 2 t p p is observation values, , 1 the is InaInthird equation, 2 ,..., Y is Φ pisthethe Y Y2 t,..., thepast past , ,..., third equation, pthe 2 Y 1 ,Φ 2 ,..., at -2 istthe past observation values,F1,values, F2values, ,...,Fp is Φ third equation, ,..., is the pastobservation observation is th third equation, q q Yt-1,,t Y1t-2,, ..., t - q InIn t-p ,……, is the error 1 2 1 2 , , … , is the coefficients of error terms. terms and at -qat -Φ , past , for … past ,observation isYthe coefficients of error terms.a , aat ,aaat t……,a terms and Yvalues, Yt 1observation -1a,t -1at -2a,……, δ Φ Φ t values, p p qathe is the ercoefficient is the constant, t 2 t 2 δ 1 2 is the constant, is error coefficient for a a a , observation ,..., values, isvalues, the is past observation values, ,..., is the Incoefficient third equation, thethe constant, , t , ,t -1t-q, ,……, error forfor past observation t t-1, t-2 t - q the t -,2 ,……, δ is constant, isis the err coefficient past ror terms and q1, qq2, … ,qq is the coefficients of error terms. a a a a 1 2 q t q t t -1 t 2 δ ARIMA (p,d,q) Model , 1 ,,past … coefficients of error terms. terms andand 2 , observation ARIMA (p,d,q) Model , is, the isq the coefficients error terms. terms isof the constant, , ,……, is the error coefficient for values, 1 ,,… 2, … is the coefficients of error terms. , terms and 1 , 2 , … , q is the coefficients of error terms. terms and To ARIMA make a non-stationary timetime series stationary, one one or two times the the difference-making (p,d,q) Model To make a (p,d,q) non-stationary series stationary, or two times difference-making ARIMA Model ARIMA (p,d,q) Model process is carried out and the result are shown with d. The model that is applied to the process is carried out and the result are shown with d. The model that is applied to series the series es the difference-making ARIMA (p,d,q) Model To make a non-stationary time series stationary, one or two times the difference-making 94make Journal of or Economic and Socialthe Studies Toto make a non-stationary time series stationary, oneone twotwo times difference-making hat is applied the series To a non-stationary time series stationary, or times the difference-makin process is carried outout andand thethe result areare shown with d. d. The model that is applied to to thethe series process is carried result shown with The model that is applied series process is carried out and theseries resultstationary, are shown one with or d. two The model that difference-making is applied to the seri To make a non-stationary time times the process is carried out and the result are shown with d. The model that is applied to the series A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey ARIMA (p,d,q) Model To make a non-stationary time series stationary, one or two times the differencemaking process is carried out and the result are shown with d. The model that is applied to the series stationary by differencing is called as non-stationary linear stochastic model or integrated model shortly (Bircan & Karagöz, 2003). Figure 1. Box-Jenkins Procedure Box-Jenkins process operates as follows: (Dobre & Alexandru, 2008: 157). Plot Series Is it Stationary? Static? Identify Possible Model Difference “Integrate” Series Seriler Diagnostic OK? Make Forecast Adaptive Neuro-Fuzzy Inference System (ANFIS) Fuzzy Inference System (FIS) consists of three conceptual components: fuzzy rule base, data base and inference. In this system, the fuzzy rules and membership functions of input and output variables are determined by the user. The most important step is to set the membership degrees of input and output variables. FIS techniques aim at providing of significant inferences by using the linguistic rules (Ross, 2004). Fuzzy systems do not have the skill to learn things, so they heavily depend on expert opinion. Adaptive Neuro-Fuzzy Inference System (ANFIS), a hybrid model, was first developed by Jang in 1993 in order to overcome this problem. This system has combined the learning skill by artificial neural networks with inference skill of expert opinion based FIS models (Jang, 1993). It adjusts the membership functions of input and output variables and generates the rules related to input and output, automatically. ANFIS can produce all the rules by using the dataset and enables the researchers to interpret these rules. Therefore, it is the widely used model in the studies of classification and estimation. Volume 1 Number 2 July 2011 95 Adaptive Neuro-Fuzzy Inference System (ANFIS), a hybrid model, was first developed by Jang in 1993 in order to overcome this problem. This system has combined the learning skill by artificial neural networks with inference skill of expert opinion based FIS models (Jang, 1993). It adjusts the membership functions of input and output variables and generates the rules related to input and output, automatically. can produce all the rules by using the Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & ANFIS Fehim BAKIRCI dataset and enables the researchers to interpret these rules. Therefore, it is the widely used model in the studies of classification and estimation. In an ANFIS modelmodel consisting of two output,the thesetset rules In an ANFIS consisting of twoinputs inputsand and one one output, of of rules is as is as follows (Jang, 1993): follows (Jang, 1993): Rule 1 : If x is A1 and y is B1 , then f1 p1 x q1 y r1 Rule 2 : If x is A2 and y is B 2 , then f 2 p 2 x q 2 y r2 where xxand the the inputs, Ai andABand are the fuzzythe sets, fi are sets, the outputs the within the Bi are fuzzy fi are within the outputs where andy are y are inputs, i i , q and r are the design fuzzy region specified by the fuzzy rule, and parameters p fuzzy region specified by the fuzzy rule, and parameters pi,i qii and rii are the design parameters thatduring are determined duringprocess. the training Thearchitecture ANFIS architecture that areparameters determined the training Theprocess. ANFIS to implement these to implement these two rules is shown in Figure 2, in which a circle indicates a whereas fixed two rules is shown in Figure 2, in which a circle indicates a fixed node, a square Figure 2. 2. Structure Structure of of an an ANFIS. ANFIS. Figure Figure 2. Structure of an ANFIS. node, whereas a square indicates an adaptive node (Jang, 1993). indicates an adaptive node (Jang, 1993). Figure 2. Structure of an ANFIS. In the first layer, each node produces membership grades to which they belong to each In the first layer, each node produces membership grades to which they belong to each of the In layer, each produces membership grades totooutputs which Inthe thefirst first layer, eachnode nodesets produces membership grades whichthey they belong toeach eachof ofthe the of the appropriate fuzzy using membership functions. The of thisbelong layer areto appropriate fuzzy sets using membership functions. The outputs of this layer are the fuzzy appropriate fuzzy sets using membership functions. The outputs of this layer are the fuzzy appropriate fuzzy sets using membership Thebyoutputs of this layer are the fuzzy the fuzzy membership grade of the inputs,functions. which are given membership grade of the inputs, which are given by membership grade of the inputs, which are given by membership grade of the inputs, which are given by 1 O (4) (4) OOii1i1 A AAii i(((xxx))) (4) (4) 1 11 B ( ) O y (5) (5) OOii i BBii i 222((yy)) (5) (5) (small, large, etc.) areare the the linguistic Where, and the y arecrisp the crisp inputs ith node, Ai andBBi (small, xx and yx are inputs to iito th node, A large, etc.) Where, crisp toto th AAii iand and large, etc.) Where, xand andyyare arethe the crispinputs inputs ithnode, node,membership andBBii i(small, (small, large, etc.)are are, the thelinguistic linguistic Where,linguistic and mB labels characterized by appropriate functions mA i i A and B , respectively. labels characterized by appropriate membership functions i and i , ,respectively. A B labels by appropriate membership functions A and B respectively. labelscharacterized characterized by appropriate membership functions ii ii respectively. In the the second layer,layer, every node incomingsignals signals sends the product product out. In the second every nodemultiplies multiplies the incoming and and sendssends the prodIn In the second second layer, layer, every every node node multiplies multiplies the the incoming incoming signals signals and and sends the the product out. out. Each node output represents the firing strength of a rule. uct out. Each node output represents the firing strength of a rule. Each node output represents the firing strength of a rule. Each node output represents the firing strength of a rule. 2 O (6) OOii2i2 w wwii i AAAii (((xxx))) BBBii (((yyy))) (6) (6) (6) i i In the third layer, the main objective is to calculate the ratio of each iith rule’s firing strength In Inthe thethird thirdlayer, layer,the themain mainobjective objectiveisistotocalculate calculatethe theratio ratioof ofeach each th ithrule’s rule’sfiring firingstrength strength wi is is taken taken as as the the normalized normalized firing firing to the the sum sum of of all all rules’ rules’ firing firing strength. strength. Consequently, Consequently, w to wi is taken as the normalized firing to the 96 sum of all rules’ firing strength. Consequently, Journal of iEconomic and Social Studies strength strength strength w Oi333 (7) w wwii i O (7) Oi i wwii i w1 w (7) ww11ww222 A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey In the third layer, the main objective is to calculate the ratio of each ith rule’s firing strength to the sum of all rules’ firing strength. Consequently, wi is taken as the normalized firing strength Oi3 = wi = wi w1 + w2 (7) In the fourth layer, the nodes are adaptive nodes. The output of each node in this layer is simply the product of the normalized firing strength and a first order polynomial (for a first order Sugeno model). Thus, the outputs of this layer are given by Oi4 = wi f i = wi ( pi x + qi y + ri ) (8) where wi is the ith node’s output from the previous layer. Parameters pi, qi and ri are the coefficients of this linear combination and are also the parameter set in the consequent part of the Sugeno fuzzy model. In the fifth layer, there is only one single fixed node. This single node computes the overall output by summing all the incoming signals as follows f = ∑w i i fi ∑w (p x + q = ∑w i i i i i y + ri ) (9) i Accordingly, the defuzzification process transforms each rule’s fuzzy results into a crisp output in this layer. In this study, the ANFIS was trained by hybrid learning algorithm which is highly efficient in training the ANFIS. This learning algorithm adjusts all parameters {ai, bi, ci} and {pi, qi, ri} to construct the ANFIS output match the training data. When the premise parameters ai, bi and ci of the membership functions are fixed, the output of the ANFIS becomes as follows: f = w1 f1 + w2 f 2 = w1 ( p1 x + q1 y + r1 ) + w2 ( p 2 x + q 2 y + r2 ) (10) Where, p1 , q1 , r1 , p 2 , q 2 and r2 are the adjustable resulting parameters? The least squares method is widely used to easily identify the optimal values of these parameters (Jang, 1993). Volume 1 Number 2 July 2011 97 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI ARIMA and ANFIS Applications We benefited from autocorrelation and partial autocorrelation functions of the related series in order to obtain ARIMA models. The appropriate models for applications were investigated based on the monthly electric data between years 2002–2010. After choosing the models, the monthly electric values pertaining to the period between the first half of 2010 and second half of 2011 were estimated by the models. Then, the forecasting values relating to the second half of 2010 were compared with the real values of the same period. Seven different models based on the subscriber groups were tested for the investigations. They are total electric use (Model 1), electric use in private houses (Model 2), electric use in industrial organizations (Model 3), electric use in business firms (Model 4), electric use in government agencies (Model 5), electric use in other subscriptions (Model 6), and electric use in business firms-government agencies-other subscriptions (Model 7). In the implemented experiments, we observed that there is no ARIMA model appropriate for Model 1 and 7. Experiment 1. Analysis of total energy use (Model 1) In the construction stage of an appropriate ARIMA model related to the consumption of total energy use, we determined that the change of energy consumption versus months is non-stationary. The difference of the monthly energy series was taken once in accordance with the autocorrelation and partial autocorrelation function graphs. When the difference was once taken, the series had got stationray character in model validation stage. However, the ARIMA techniques generated through tests are to pass the suitability test in order to be used for future forecastings. Thus, it is indicated that the autocorrelation coefficients of the forecastings generated through the ARIMA techniques show a trend of systematicity in the suitability test. Despite all those efforts put forward, no model was detected as appropriate for this purpose. When the ANFIS model was trained by the training dataset for 1000 iterations, it found three rules for this experiment. The results obtained by ANFIS model are shown in Figure 3, where the test result of the model is red line and the forecasting of the next six month period is green line as follows: 98 Journal of Economic and Social Studies A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey Figure 3. The ANFIS output for Model 1. As seen in Figure 3, there is a small decrease possibility, but the stationary case will be lasting in the next six months. The comparative results of ARIMA and ANFIS techniques are given in Table 1: Table 1. Forecasting of total electric use (Model 1) Month Jul.10 Real ARIMA Forecasting ANFIS MSE Forecasting 42,0229 50,9282 ug.10 51,4488 50,9275 ep.10 59,9116 50,9268 ct.10 43,4847 50,9261 ov.10 46,3052 50,9254 Dec.10 48,5912 Jan.11 o appropriate model has been found 50,9240 50,9233 Mar.11 50,9226 pr.11 50,9219 May.11 50,9212 Jun.11 50,9205 Number 2 July 2011 40,4117 50,9247 eb.11 Volume 1 MSE 99 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI As seen in Table 1, the ARIMA could not achieve any forecasting result. However, when the ANFIS was tested by six-month testing dataset, its mean square error (MSE) ratio was 40.4117. The forecasting obtained by the ANFIS for the next six months was close to the obtained test results of the ANFIS, so it can be observed that there was approximately a stationary case. Experiment 2. Analysis of the electric use in private houses (Model 2) In the construction stage of an appropriate ARIMA model related to the consumption of electric energy use in private houses, we determined that the change of energy consumption versus months is non-stationary. The difference of the energy series is once taken in accordance with the autocorrelation and partial autocorrelation function graphs. When the difference taken, the series showed stationary haracter in the model validation stage. As a result of the tests, ARIMA(1,1,2) was chosen as the model. When the ANFIS model was trained by the training dataset for 4000 iterations, it found eight rules for this experiment. The results obtained by ANFIS model are shown in Figure 4, where the test result of the model is red line and the forecasting of the next six month period is green line as follows: Figure 4. The ANFIS output for Model 2. 100 Journal of Economic and Social Studies A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey As seen in Figure 4, there is a constant increase in the next six months in Model 2. The comparative results of ARIMA (1,1,2) and ANFIS techniques and the real values for the period between 2002-2010 are given in Table 2: Table 2. Forecasting of the electric use in private houses (Model 2) ARIMA(1,1,2) ANFIS Month Real Jul.10 18,9146 21,1999 22,0048 ug.10 22,9082 21,3377 22,4716 ep.10 32,8206 21,4178 ct.10 19,9593 21,5264 ov.10 20,9913 21,6209 23,8837 Dec.10 23,7073 21,7224 24,3547 Jan.11 21,8205 24,8256 eb.11 21,9202 25,2962 Mar.11 22,0191 25,7668 pr.11 22,1184 26,2372 May.11 22,2175 26,7077 Jun.11 22,3167 27,1781 Forecasting MSE 24,0842 Forecasting 22,9416 23,4125 MSE 21,3410 As seen in Table 2, the MSE ratio was 24.0842 for ARIMA (1,1,2) and 21.3410 for ANFIS in the result of forecasting implemented by the six-month testing data for Model 2. Thus, it can be said that the ANFIS provided a more successful forecasting than the ARIMA. Experiment 3. Analysis of the electric use in industry (Model 3) We experienced that the monthly energy change is originally unsuitable for the validation of a model in the set stage of an appropriate ARIMA model to be applied in electric use in industry. After the analysis on the autocorrelation and partial autocorrelation function graphs of the series employed in industrial electric use, the difference was taken twice. Then, the series was made suitable for the analysis. After a few testing, the model was selected as ARIMA (1,2,1). Volume 1 Number 2 July 2011 101 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI When the ANFIS model was trained by the training dataset for 1500 iterations, it found two rules for this experiment. The results obtained by ANFIS model are shown in Figure 5, where the test result of the model is red line and the forecasting of the next six month period is green line as follows: Figure 5. The ANFIS output for Model As seen in Figure 5, the next six months is constantly stationary for Model 3. The comparative results of the real values in the period between 2002-2010 and ARIMA (1,2,1) and ANFIS techniques are given in Table 3. Table 3: Forecasting of the electric use in the industry (Model 3) ARIMA(1,2,1) Months Jul.10 Real Forecasting ANFIS MSE Forecasting 6,5457 11,1835 6,4739 ug.10 6,6035 8,3942 6,1838 ep.10 6,7541 9,1898 ct.10 7,0919 8,2417 ov.10 6,7017 8,1281 6,1088 Dec.10 5,9649 7,6015 6,0901 Jan.11 7,2656 6,0714 eb.11 6,8283 6,0527 Mar.11 6,4308 6,0340 pr.11 6,0049 6,0153 May.11 5,5836 5,9967 Jun.11 5,1508 5,9780 102 6,1138 6,1471 6,1275 MSE 0,3079 Journal of Economic and Social Studies A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey As seen in Table 3, the MSE ratio is 6.1138 for ARIMA (1,1,2) and 0.0379 for ANFIS in the result of forecasting made through the six-month testing data for Model 3. Thus, it can be said that ANFIS provided a more successful forecasting than the ARIMA. Experiment 4. Analysis of the electric use in business firms (Model 4) While a model related to the electric use in business firms was constructed, monthly energy change was analyzed. But, it was determined that the series was originally unsuitable to construct an appropriate model. After the analysis on the autocorrelation and partial autocorrelation function graphs, the series was observed to be appropriate for an analysis stage when the difference was once taken. After a few testing, ARIMA(1,1,1) was determined to be the most appropriate model for this experiment. When the ANFIS model was trained by the training dataset for 3600 iterations, it found four rules for this experiment. The results obtained by ANFIS model are shown in Figure 6, where the test result of the model is red line and the forecasting of the next six month period is green line as follows: Figure 6. The ANFIS output for Model 4. As seen in Figure 6, there was very quickly rise in the next six months for Model 4. The comparative results of the real values in the period between 2002-2010 and ARIMA(1,1,1) and ANFIS techniques are given in Table 4. Volume 1 Number 2 July 2011 103 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI Table 4: Forecasting of electric use in business firms (Model 4) Months Jul.10 Real ARIMA(1,1,1) Forecasting ANFIS MSE Forecasting 5,1655 6,6010 6,9789 ug.10 9,5100 7,1028 7,1073 ep.10 9,4926 7,0126 ct.10 5,8424 7,1121 ov.10 6,5950 7,1508 7,5016 Dec.10 7,4564 7,2089 7,6370 Jan.11 7,2608 7,7745 eb.11 7,3147 7,9142 Mar.11 7,3680 8,0561 pr.11 7,4215 8,2001 May.11 7,4749 8,3462 Jun.11 7,5283 8,4943 2,6646 7,2370 7,3683 MSE 2,8888 As seen in Table 4, the MSE ratio is 2.6646 for ARIMA (1,1,1) and 2.888 for ANFIS as a result of the forecasting through the six month testing data for Model 4. Thus, it can be said that ARIMA is little more successful in the implemented experiment when compared to ANFIS. Experiment 5. Analysis of the monthly electric use in government agencies (Model 5) While constructing a model for the electric use in government agencies, the graphs of monthly energy change were used in hand as a basis. It was determined that the series was originally unsuitable for model validation. After the analysis on the autocorrelation and partial autocorrelation function graphs, the series was, then, found to be suitable for an analysis stage when the difference was once taken. After a few testing, the model was selected as ARIMA (2,1,1). When the ANFIS model was trained by the training dataset for 5000 iterations, it found four rules for this experiment. The results obtained by ANFIS model are shown in Figure 7, where the test result of the model is red line and the forecasting of the next six month period is green line as follows: 104 Journal of Economic and Social Studies A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey Figure 7. The ANFIS output for Model 5. As seen in Figure 7, there is a constant decrease in the next six months for Model 5. The comparative results of ARIMA (2,1,1) and ANFIS techniques and the real values in the period between 2002-2010 are given in Table 5. Table 5: Forecasting of the electric use in government agencies (Model 5) ARIMA(2,1,1) Month Jul.10 Real Forecasting MSE ANFIS Forecasting 4,2125 3,3762 3,5093 ug.10 3,9930 3,6692 3,5053 ep.10 2,6223 3,4032 ct.10 3,0580 3,4401 ov.10 3,6314 3,3647 3,4604 Dec.10 4,2085 3,3635 3,4337 Jan.11 3,3431 3,4011 eb.11 3,3429 3,3627 Mar.11 3,3415 3,3184 pr.11 3,3467 3,2686 May.11 3,3531 3,2131 Jun.11 3,3620 3,1522 Volume 1 Number 2 July 2011 0,3909 MSE 0,3841 3,4961 3,4812 105 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI As seen in Table 5, the MSE ratio is 0.3909 for ARIMA(2,1,1) and 0.3841 for ANFIS in the result of the forecasting made through the six month testing data for Model 5. Thus, it can be said that ANFIS is little more successful in the forecasting when compared to ARIMA. Experiment 6. Analysis of electric consumption of the others (Model 6) Model testing named the others about the electric consumption of the subscriber groups for six months was done. It was determined that the series was not appropriate to construct a model in its original. Thus, the series became appropriate after the difference was once taken through the autocorrelation and partial autocorrelation function graphs. After several tests, the model were determined as ARIMA (1,1,1). When the ANFIS model was trained by the training dataset for 5000 iterations, it found three rules for this experiment. The results obtained by ANFIS model are shown in Figure 8, where the test result of the model is red line and the forecasting of the next six month period is green line as follows: Figure 8. The ANFIS output for Model 6. As seen in Figure 8, there will be a little increase in the next six months for Model 6. In Table 6, there are comparative results of the real values from the period of 2002-2010 and the techniques of ARIMA (1,1,1) and ANFIS in Model 6. 106 Journal of Economic and Social Studies A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey Table 6: Electric Consumption Forecasting of the Others (Model 6) Months ARIMA(1,1,1) Real Forecasting MSE ANFIS Forecasting em.10 7,1845 8,1967 7,4670 Ağu.10 8,4338 8,2828 7,5274 yl.10 8,2217 8,3353 ki.10 7,5329 8,3873 as.10 8,3855 8,4392 7,7085 ra.10 7,2539 8,4912 7,7689 ca.11 8,5431 7,8293 Şub.11 8,5951 7,8897 Mar.11 8,6470 7,9501 is.11 8,6990 8,0104 May.11 8,7509 8,0708 Haz.11 8,8029 8,1312 0,5540 7,5877 7,6481 MSE 0,3401 As seen in Table 6, MSE ratio is 0.3401 for ANFIS whereas it is 0.5540 for ARIMA (1,1,1) as a result of the forecasting which was done by using test data for six months for Model 6. Thus, ANFIS provided a more successful forecasting than ARIMA for Model 6. Experiment 7. The analysis of electric consumption of the Business firm, State office and others (Model 7) It was determined in the forecasting of modeling about the electric consumption of the business firm, state office and others that consumed energy is non-stationary. The series became stationary by taking the difference for once according to the autocorrelation and partial autocorrelation function graphs of the related series. The ARIMA techniques which are obtained through several experiments must pass the compliance test to be used as an forecasting model regarding the future. In the compliance test which was done for this reason, it was determined that the autocor- Volume 1 Number 2 July 2011 107 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI relation coefficients of the forecasting errors of the forecastings obtained through the ARIMA techniques shows a systematic tendency. The appropriate model could not be determined. When the ANFIS model was trained by the training dataset for 5000 iterations, it found three rules for this experiment. The results obtained by ANFIS model are shown in Figure 8, where the test result of the model is red line and the forecasting of the next six month period is green line as follows: Figure 9. The ANFIS output for Model 7. As seen in Figure 9, there will be a linear increase in the next six months for Model 7. In Table 7, there are the results of ANFIS model for Model 7. 108 Journal of Economic and Social Studies A comparison of ANFIS and ARIMA techniques in the forecasting of electric energy consumption of Tokat province in Turkey Table 7: Electric Consumption Forecasting of the Commerce Houses, State Offices and Others (Model 7) Months ARIMA Real Forecasting ANFIS MSE Forecasting em.10 16,562 17,5558 Ağu.10 21,937 17,6570 yl.10 20,336 17,7583 ki.10 16,433 17,8599 as.10 18,612 17,9616 ra.10 18,918 ca.11 o appropriate model has been found MSE 4,8571 18,0636 18,1660 Şub.11 18,2690 Mar.11 18,3730 is.11 18,4789 May.11 18,5878 Haz.11 18,7017 As seen in Table 7, there is not any forecasting through ARIMA for Model 7. MSE ratio became 4.8571 as a result of the forecasting through ANFIS which was done by using test data for six months. The estimated results of the next six months are similar to the results obtained through ANFIS but there has been an increase at the least. Conclusions We implemented different seven experiments for estimating and analyzing electric energy consumption in order to plan the production, transmission and distribution of electric energy and to determine the consequences of the events occurring in the electric market. These experiments focus on the electric energy consumption forecasting implemented by using ANFIS and ARIMA techniques including the analyses of total energy consumption, household electric consumption, industrial electric consumption, commerce house electric consumption, monthly electric con- Volume 1 Number 2 July 2011 109 Rüstü YAYAR & Mahmut HEKIM & Veysel YILMAZ & Fehim BAKIRCI sumption in state offices, electric of the others and the electric consumption of the commerce houses, state offices and the others. This is especially guiding key for the investors planning the investments in the electric sector. There has been a preparatory work for the necessary precautions regarding the electric in Tokat, revealing the electric consumption structure of Tokat province. The electric demand structure of Tokat regarding the previous consumption of Tokat province and the estimated electric energy for the future has been revealed. Although the study which was conducted through ANFIS and ARIMA techniques in the field of electric energy consumption was conducted on regional basis, it can be a pilot study for the extensive national and international studies on the energy consumption. Electric energy demands for the tested periods and the first six months of the year 2011 were estimated by using the ARIMA and the ANFIS techniques. Every forecasting experiment related to the electric consumption performed by ANFIS and ARIMA showed that ANFIS is more successful estimator. The ARIMA was more successful in only an forecasting experiment. In addition to this, an appropriate ARIMA model could not have been found for two forecasting experiments. As a result, the ANFIS is more appropriate than the ARIMA in the forecasting studies regarding the electric consumption. References Al-Garni, A.Z. & Abdel-Aal R.E. (1997). Forecasting Monthly Electric Energy Consumption in Eastern Saudi Arabia Using Univariate Time Series Analysis, Energy, Volume 22, Issue 11, 1059–1069. Al-Garni, A.Z. & Javeed Nizami, S. S. A. K. (1995). Forecasting Electric Energy Consumption Using Neural Networks, Energy Policy, Volume 23, Issue 12, 1097–1104. Asilkan, Ö. & Irmak, S. (2009). Forecasting the Future Prices of the Second Hand Automobiles Through Neural Networks, Journal of Faculty of Economics And Administrative Sciences, Volume: 14, 2, 378. Bircan, H. & Karagöz, Y. (2003). A Practice on The Monthly Exchange Rate Forecasting Through the Box-Jenkins Models, Journal of Social Sciences of Kocaeli University, Issue:6/2, 49, 51. 110 Journal of Economic and Social Studies Bozkurt, H. (2007). Analysis of Time Sequences, Ekin Publishing House, Bursa, 7–8. Brown R.E. & Koomey J.G. (2002). Electricity Use in California: Past Trends and Present Usage Patterns, Enduse Forecasting and Market Assessment, http://enduse.lbl.gov/info/LBNL-47992.pdf. Çebi, H. & Kutay, F. (2004). Forecasting of The Turkey’s Electrical Energy Consumption Until 2010 Through Artificial Neural Networks, J. Fac. Eng. Arch. Gazi Univ., Volume 19, No 3, 227-233, http://www.mmfdergi.gazi.edu.tr/2004_3/227234.pdf. Demir, A. (1968). A Study on World’s Economy, SBF Publications of Ankara University No.259 Ankara, 68. Demirel, Ö. & Kakilli, A. & Tektaş, M. (2010). Forecasting of Electrical Energy Load Through Anfis And Arima Models, Journal of Faculty of Engineering And Architecture Volume J. Fac. Eng. Arch. Gazi Uni. 25, No:3, 602. Dikmen, N. (2009). Basic Concepts and Applications of Econometrics, Nobel Publication Distribution, Ankara. Dobre, I. & Alexandru, A. A. (2008). Modelling Unemployment Rate Using BoxJenkins Procedure, Journal of Applied Quantitative Methods, 157. Electricity Generation Corporation (2011). www.euas.gov.tr. Governorship of Tokat (2006). Province Annual of 2006, Tokat, 39, 146. Gujarati, D.N.(2009). Translators Şenesen, G. G. & Şenesen, Ü., Basic Econometrics, Sixth Edition, Literature Publication, İstanbul, 378. Jang, J. S. R., (1993). ANFIS: Adaptive-Network Based Fuzzy Inference Systems, IEEE Transaction on Systems, Man, and Cybernetics, 23 (03): 665-685. Kılıç, N. (2006). A General Overview On Turkey’s Electrical Energy Consumption And Production, Izmir Chamber of Commerce, R&D Journal, July 2006, İzmir, 12. Lise, W. & Van Montfort, K. (2005). Energy Consumption And Gdp in Turkey: is There a Cointegration Relationship? June 29 – July 2, 2005, Istanbul, Turkey, 5. Official Gazette, (2006). Regulations on Activity Report by Public Administrations, 17.03.2006, Issue: 26111. Önder, E. & Hasgül, Ö. (2009). Use of Box-Jenkins Model on the Forecasting of Foreign Visitors, Time Sequence Analysis Through the Winters Method And Artificial Neural Networks Winters, Administration, Year: 20, Issue: 62, 62-83. Provincial Department of Environment and Foresty of The Governorship of Tokat, (2007). Environment Status Report of 2007 of Tokat, Tokat, 1, 130. Ross, T. J. (2004). Fuzzy Logic with Engineering Applications, Second edition, John Wiley. Şekerci Öztura, H.(2007). Today and Tomorrow of Electrical Energy In Our Country, VI. Energy Symposium of Union of Chambers of Turkish Engineers And Architects - Global Energy Policies And The Reality of Turkey, Ankara, 268. Tak S. (2002). Methods of Electrical Energy Demand Estimation and An Econometric Practice for Turkey (1970–2005), Unpublished Postgraduate Thesis, Atatürk University, Erzurum, 47. Terzi, H. (1998). Relationship Between Electricity Consumption and Economic Growth in Turkey : A Sectoral Comparision, Journal of Business and Finance, Volume: 13, Issue:144, March 1998, 62-71. Tufan, E. & Hamarat, B. (2003). Clustering of Financial Ratios of the Quoted Companies through Fuzzy Logic Method, Journal of Naval Science and Engineering, Vol. 1, No:2, 123-140. Turkish Statistical Institute (TUIK) (2011). http://report.tuik.gov.tr/reports/ rwservlet?ad nksdb2=&ENVID=adnksdb2Env&report=ikametedilen_ il.RDF&p_kod=2&p_kil1=60&p_yil=2010&p_dil=1&desformat=html. Turkish Statistical Institute (TUIK) (2009). Regional Indicators, 2009 TR 83 Samsun, Tokat, Amasya, April 2010 Ankara, 10. VIII. Regional Directorate of State Hydraulic Works - Samsun, (2011). http:// www2.dsi.gov .tr/bolge/dsi7/tokat.htm#atakoy.