Forecasting China's natural gas consumption based on a combination model

Availableonlineatwww.sciencedirect.comJoumal ofScience DirectNatural GasChemistryELSEVIERJournal of Natural Gas Chemistry 19(2010)493-496Forecasting China's natural gas consumptionbased on a combination modelGang Xu*, Weiguo Wang1. Dalian Institute of Chemical Physics, Chinese Academy of Sciences, Dalian 116023, Liaoning, China.2. College of Quantitative Economics, Dongbei University of Finance Economics, Dalian 116025, Liaoning, ChinaI Manuscript received March 19, 2010: revised April 28, 2010 1AbstractEnsuring a sufficient energy supply is essential to a country. Natural gas constitutes a vital part in energy supply and therefore forecastingnatural gas consumption reliably and accurately is an essential part of a countrys energy policy. Over the years, studies have shown that acombinative model gives better projected resultsed tomodel. Inthis study, we used Polynomial Curve and Moving AverageCombination Projection(PCMACP)model to esturalumption in China from 2009 to 2015. The new proposedPCMACP model shenore reliable and accurateItscentage Error (MAPE) is less than those of anymodels within the investigated range. accordingge annual growth rate will increase for the next 7 yearsthe amount of natural gas consunn will reagh 17 b0'millio2015 in CrKey wordsatural gas consumption; forecasting; combination mydel1. Introductionsearchers to create advanced natural gas consumption modelsto help make reliable and accurate projections of China,s fufossil fuel. The demand for natural gas has a significant im- policy decisione demand so as to provideNatural gas(NG)is the most efficient and clean burninggood governmentpact on national energy production, energy consumption, naThe models developed for forecasting energy consumptional economic growth and people's daily life. Natural gas tion including natural gas consumption can be split into tw:occupies 1/4 in global energy supply, furthermore, it will ex- types: the single method and the combinative method. Goddnaceed coal to become the second biggest sources of energy examples of the single method forecasting are Autoregresin the global economy and social development in 2010 [1]. Integrated Moving Average(ARIMA)model [2], cointegraChina has been paying a lot of attention to the development of tion and error-correction model (ECM)[3], time series modnatural gas. The"11th Five Year Plan"by State Development els [4], and partial Least Square Regression(PLSR)[5]. Theand Reform Committee proposed that the proportion of natu- non-linear nature of natural gas consumption tends to reduceral gas in the primary energy consumption will increase from the reliability of the single method models. For non-linearsys2.8% in 2005 to 5.3% in 2010, which means the demand for tems like natural gas consumption, combinative method fore-natural gas will almost double between 2005 and 2010. The casting gives more accurate results and is therefore more popgrowth in the demand for natural gas in China is now higher ular. Examples of combinative method forecasting are Greythan those for all the other energy sources including coal, oil, model and BP Neural Network combination model [6], dehydropower and other renewable energiesterminate and stochastic time series combination model [71Forecasting natural gas consumption constitutes a vital Bayesian combination model 181, Grey model and Artificialpart of energy policy of a country, and is also one of the most Neutral Network combination model [9],etcimportant policy tools used by the decision makers all overIn this study, we used the Polynomial Curve and Movingthe world. Therefore, it is very valuable for the academic re- Average Combination Projection (PCMACP)model toCorresponding author. Tel: 86-411-84379201; Fax: 86-411-84686597: E-mail: xugang@dicp acThis work was supported by the Youth Fund of Chinese Academy of Sciences Knowledge Innovat中国煤化工No.S2060andthe Innovation Team Project of Education Department of Liaoning Province(No. 2007T050)CNMHGCopyright@2010, Dalian Institute of Chemical Physics, Chinese Academy of Sciences. All rights reservedoi:10.1016S1003-9953(09)60100-6494Gang Xu et al Journal of Natural Gas Chemistry Vol. 19 No 5 201060enatural gas consumption in China based on histori- curve. Therefore, based on the trend extrapolation theory, weseries. The consumption data for natural gas cov- can establish the 2nd order Polynomial Curve model of naturers the periods 1995-2008. The 1995-2008 data are ob- gas consumption as followstained from Chinese Statistics Year book from 1996 to 2009and 2007-2008 data are used to evaluate the prediction abil-Gas=a+Bt+ot+sity. The rest of this paper is organized as follows: Section where, time is an independent variable, t= 1, 2,12 and"gas2 describes the used methodology; Section 3 presents the(gas consumption) is a dependent variable. We can also useforecasted results and compared them with the actual data ordinary least squares to estimate the parameter as follows(2007-2008)and those from previous studies; and the final5ooon describes the findings from the work and draws theGas=204.6630-18.33723t+3917498t2(2)Onclusions(12.99918)(-3.293087)(9.394953)[0.0000[0.0093]0.00002. MethodsThe figures in parentheses are t statistics and the figures in2. 1. Polynomial Curve modelthe square brackets are p values. Durbin-Watson statistic1.541523 and the adjusted R reaches 0.985368, which showAs one of the trend extrapolation methods, Polynomial that there is a very good fit between the predicted results of4@0e model is used widely and efficiently. As new meththe model and the original datatrend extrapolation methods were developed and detailedThe predicted results obtained by the 2 OPC are givenvaluations of forecast accuracy and utility were conducted. in Table 1. As seen from the table, thebsolute percentage error(MAPE) from 1995 to 2006 is 4.41%0, whichmethods particularly useful for small-area projections. The forecasting errors, But the 2nd opC only simulates the trenddefining characteristic of trend extrapolation method is that and ignores the residual series between the trend value and thefuture values of any variable are determined solely by its historical values [10]established in the Moving Average(MA)model to predict thePolynomial cibest modeled by a polynomialresiduals. Finally, the 2nd oPC and Ma models are combined300. They may be second order equations in the form ofto establish the Polynomial Curve and MA Combination ProY=aX2+bX+c, resulting in a parabolic shape, where r is jection model to further improve the prediction accuracy ofthe dependent variable, X is the independent variable, a is the modelthe slope for the nonlinear trend and b is the slope for thelinear trend. The coefficients of a second order polynomialTable 1. Model values and forecast errorsCurve model (2nd OPC) can be estimated using ordinary leastModel value*Error(%)squares(OLs)regression techniques7.16Polynomial curve model is an important forecast tech183.65849ue to predict natural gas consumption. Figure I illustrates2193.99401increasing over the period215.33710210.91426-2.05from 1995 to 2006. The natural gas consumption is expressedby the ascending curve without additive seasonality, furthe274.5447more, the shape of the curve is similar to 2nd order polynomial292.18069308.685015.65339,454293569452539728586413.04049486.93617476.97073200661.05120548.73596本Unit100 million cubic meters (MCM).in absolute10percentage error (MAPE)=4.41%, root mean square error(RMSE)=13, 193, mean absolute error (MAE)=11. 873, Theilinequality coefficient =0.021: MAPE =100RMSE=VMAE=2I9t-ytl/n,Theil inequalitycoefticient01中国煤化工199519961997199819992000200120022003200420052006CNMHG1995-996↑997η9981999200〔z0r2002032004200于2006YearJournal of Natural Gas Chemistry VoL. 19 No 5 20102. 2. Moving Average modelWe carried out the test of autocorrelations for the residualsin MA(6)by the Q statistic and correlogram. From Figure 2In time series analysis, Moving Average(MA)model is both autocorrelation value and partial correlation value are ina common approach for modeling univariate time series modthe critical interval. and it also shows that residuals are a ranIs. The notation MA(@) refers to the moving average model dom disturbance series without autocorrelation. In addition,of order qaccording to the Breush-Godfrey Lagrange Multiplier (Lm)test of the residual series, F statistic is 0. 1279 and related pXt=A+Et+01Et-1+.+eqEt-q(3) value is 0.8930 Tx R2 statistic is 0. 4904 and related p valuewhere, u is the mean of the series, 61,,8g are the parameters is 0.7825; both of the p values are more than 0.05. As a resultvalue of g is called the ord % are white noise error terms. The Ho can not be rejected and there is no serial correlation in theof the model ander of the ma model [11residual sericause began with the test of variable's stationarity be-To sum up, both Q statistic and LM test show that theerecause Ma models are used in time series analysis to deno serial correlation in the residuals of Ma(6). The adscribe stationary time series. Augmented Dickey-Fuller test justed R- reaches 0.989978 and this model can be used forwas applied to residual series of MA(2). The t statistic the prediction of natural gas consumption in the future2.680631, and p value is 0.0125, which is less than1.977738 at the 5% significance level, It can be found thatthe null hypothesis of unit root can be rejected for residual se- Date: 1 1/15/09 Time: 16: 11ries at the 5% significance level; it also implies that residualSample:1995-2006series seem to be stationary. So, we conclude that residualseries are integration of order zero ((O))Q-statistic probabilities adjusted for 1 ARMA term(s)accordance with the correlogram of residuals inAutocorrelationPartialAC PAC O-Stat. ProbMA (2), in particular, with the support of the trial calculationMA(2)model was the appropriate one, furthermore, MA(2)10.1030.1030.16132-0.137-0.149047780489was estimated below3-0.037-0.285223720.327A(2)=u+0.848173-24-0.401-0410561500.1325-0.149-0.2816.14550.189(13.19135)60.176-0.092700850.220[0.000070.296-0.016995290.1278-0.055-0423100820.184The figure in parentheses is t statistic and the figure in square90.053-0.16110.2380.249brackets is p value. Since p value is less than 0. 05, it indicatesthat the statistic results of ma (4)are significantFigure 2. Correlogram of residuals. AC: Autocorrelation; PAC: Partial auto-correlation; Q-stat Q statistic; Prob. Probability3. Polynomial Curve and MA Combination Projection2-0.137-0.1490.47780.4893. ResultsBased on the 2n oPC model and ma(2)model. we built3自,b03om咱c28n死m.327olynomial Curve and MA Combination Projection model on shows consistently accurate results and excellent performarthe historical dates of natural gas consumption from 1995 to across all the period from 1995 to 2006. Furthermore by co2006 as followsGas =a+Bt+ot+MA(2)+EMACP model and the 2nd oPc model the statistics of the pc-e employed ordinary least squares to estimate the parameter, and the equation can be expressed asMACP model are 10., 9.087%0, 3.630%0, and 0.016 in theast=248.3293-31.86327t+4849942+MA(2)at1g99281295m+189(7.568760)(-3.361810)(7.968126)PCMACP model is more accurate and reliable than the 2nd[0.0001][0.0099][0.0000OPC modelMA(2)=+0.92606Fe可20)×09908569.220(16.94169)[0.0000pares the MAPe statistic in 2007 and 2008 with the resultsThe figures in parentheses are t statistics; the figures in square25cF060m992):.127brackets are p values. The adjusted R2 and Durbin-Watson ral network model 2l and grey Forecast model [3] arestatistic are 0.989978 and 1.739574, respectively. All the-342中国煤化工a, respectivelystatistics are significant. The model also demonstrates thatthe efficiency of the estimated parameter has been improved8“口 CNMHGAU82.184by the Ma(2)model and the 2 OPC modelmodels in the years 2007 and 200890.053-0.16110.2380.249Gang Xu et al Journal of Natural Gas Chemistry Vol. 19 No 5 2010Table 2. Comparison of forecasting measurement errors among four modelsYear Actual value of NG consumption 2ndPCMACP modelBP Neural Network model(100MCM)Fitted value Error(%) Fitted value Error(%) Fitted value Error(%) Fitted value Error(%)698.90310.10641.123447.36835.99807.857715.77111.40774429702029458.64743.234. ConclusionsFitted valueA novel PCMACP model was developed to accurately28Errorpredict China's natural gas consumption based on the datafrom the Chinese statistics Year book between 1995 and2006. The new PCMacP model showed that MAPe was3.63 and the Theil inequality coefficient was 0.016034 Forthe years 2007 and 2008 the proposed PCMACP model provided more accurate and reliable results in comparison withother existing models in the public literature that the authorcould find. Furthermore, the proposed model shows that natural gas consumption will increase from 97800 million cubic99519961997199819992000200120022003200420052006meters in 2009 to 171600 million cubic meters in 2015. Theaverage annual rate of increase in natural gas consumptionFigure 3. Comof the actual and fitted valuenatural gas coslows down a little from 12.36 between 1995 and 2008 tosumption9.97%0 between 2009 and 2015. However, continued growthat these very high rates makes it important to ensure the proBased on the above comparison, we choose the PCMACPduction and availability of this amount of natural gas to meetmodel to predict natural gas consumption. The forecasted valife's needs. Finally, the novel PCMACP model proposed canand the trend from 2009 to 2015Figure 4. Thereliably and accurately be used for forecasting natural gas conresults show that there will be a sustained and rapid growth in sumption and is therefore an excellent tool in helping thenatural gas consumption from 87800 million cubic meters in icy makers in the government to make the best decisions2009 to 171600 million cubic meters in 2015. The natural gasconsumption in 2015 is a little higher than that obtained by ReferencesP Neural Network model [12] that estition will be 147100 million cubic metersThe numbers in Figure 4 demonstrate that compared with [Il Chen X H International Petroleum Economics(Guoji Shiyouthat of 48700 million cubic meters in 2005, natural gas conJingji),2003,1l(12):34sumption in the period between 2009 and 2015 will increase [2] Ediger V S, Akar S Energy Policy, 2007, 35(3):1701by80.29%,107.34%,127.28%,15556%,18583%,21808%31 Sun Yf. Special Zone Economy( Tequ Jingji),200,92(2)and 252.33%, respectively. It also shows a strong increasingtrend in the next 7 years, however, the annual rate of increase [4] Kumar U, Jain V K.Energy, 2010, 35(4):1709in natural gas consumption slows down a little from 12.36% [5 Zhang M, Mu H L, Li G, Ning Y D. Energy, 2009, 34(9): 1396between 1995 and 2008 to 9.97 between 2009 and 201516] Fu J F, Cai G T, Zhang L Resource Developement Market(Ziyuan Kaifa yu Shichang ) 2006, 23(3): 216oZ2000[7] Lu E P. Application of Statistics and Management(Shuli Tongjguanli),2006,25(5):505[8] Chai J, Guo J E, Lu H China Population, Resources and Em1500rOmmen,2008,18(4):50[9] Lu Q, Gu PL, Qiu S M. Systems Engineering-Theory Practice(Xitong Gongcheng Lilun yu Shijian), 2008, 23(3): 24[10 Smith K s, Tayman J, Swanson D A State and Local PopulationProjections-Methodology and Analysis. New York: Springer,[11http://en.wikipediaorg/wiki/moving-average-moDel2009201020112012201320142015[12] Luo D K, Xu P. Oil-Gasfield Surface Engineering(Yougitian[13] Zhou ZB, Li L中国煤化工oration Devel到19哨9灣79§920120EW2032006ear