Chapter 3 Load Forecasting Model 3.1 Introduction Reviews of the load forecasting literature in the last chapter lead to the same conclusion that the power tariff, GDP, and temperature are the key factors affecting the load consumption. The relationship between load consumption and the key factors may be specified by the general function E t = f Pt , GDPt , TEMPt t = 1, 2 T (3.1) where E t = consumption of electricity energy in month t (gwh) GDPt = gross domestic product in month (million baht) TEMPt = temperature (celsius) Pt = average power tariff in month t baht/kwh Several monthly models of electricity energy consumption will be specified, estimated, tested and selected as forecasting models for the electricity energy consumption. Electricity energy consumption will be disaggregated and modeled for the following categories 1) total consumption in the generating system 2) consumption in the northern generating system 3) consumption in the central generating system 4) consumption in the northeast generating system 5) consumption in the southern generating system 6) consumption in the PEA jurisdiction 7) consumption in the MEA jurisdiction Data on electricity energy consumption, power tariff are monthly data provided by the three power authorities. The temperature data are collected from the Meteorological Department. The monthly power tariff is derived from the annual revenue and sales of electricity energy of the relevant power authority by computing the average monthly revenue per kWh. The monthly power tariff is thus constant for a given year Data on temperature are collected from the meteorological stations in each region. Temperature days are computed for each station as the product of average daily temperatures and the number of days in a given month. The average temperature days for each region are then computed from the number of stations in that region. Data on GDP are published only on an annual basis. Since the forecasting models are monthly models, a proxy variable for monthly GDP is required. Since the relationship between money supply and GDP is well established in macroeconomic 26 theory, and the data on money supply are available on a monthly basis, the money supply is selected as a proxy variable for GDP in the monthly forecasting models. Empirical test supports the money supply/GDP hypothesis when money supply in the narrow sense,M1, can explain 97 percent of the total variations in GDP. The general function of (3.1) may now be specified as E t = f Pt ,M1t ,TEMPt (3.2) where M1t is the money supply in month t 3.2 Model Specification The simple linear form of (3.2) is specified as (3.3) E t = A + A1M1t + A 2TEMPt + A3Pt with A1 > 0 A 2 > 0 and A 3 < 0 The above model is fitted from the data between October 1991 and September 2003 with the following results E t = 3933.97 + 0.011393 M1 + 7.05076 TEMPt + 62.4324 Pt (t = –7.98) (t = 19.42) (t = 13.20) adj.R 2 = 0.9504 (t = 0.25) (3.4) Even though the simple linear model can explain 95 percent of the variations in electricity energy consumption, the price variable has a “wrong’ sign. The wrong sign of the price variable is the model’s shortcoming which needs to be addressed. 3.2.1 Distributed lag and Autoregressive Models There may be some lags in the relationships between the exogenous variables and the endogenous variable. Models with lagged variables may be specified by a general function E t = f Pt i , M1t j , TEMPt k , E t h i, j, k, h = 0, 1, 2 (3.5) Lagged relationship between variables in the model may be explained by diverse behaviors of the power consumers. As a case in point, the power tariff structure in Thailand has a Ft mechanism for the adjustment of the tariff rate when there are changes in the generation costs. The rate is adjusted every four months so a power consumer has no knowledge of the tariff rate in month t since the rate he pays in month t is actually the ‘average tariff rate’ of the last four months. 27 One possible hypothesis in this case is that electricity energy consumption of a power consumer depends upon the anticipated tariff rate in month t. The model may thus be specified as E t = a + b*Pˆt + c*M1t + d*TEMPt (3.6) where E t = electricity energy consumption in month t P̂t = anticipated power tariff in month t and b < 0 c > 0 d > 0 There are no direct data on the anticipated tariff rate but it may be modeled by a given behavioral assumption. As an example, the difference between anticipated tariffs in month t and month t-1 depends on the anticipation error in month t–1 or Pˆ t Pˆ t 1 = e Pt 1 Pˆ t 1 (3.7) where e > 0 Rewriting 3.7 in the form P̂t 1 vB = ePt 1 (3.8) where v = 1 e B = Backward operator Substitution of (3.8) into (3.6) yields Et = a + b*e Pt 1 + c*M1t + d*TEMPt 1 vB (3.9) Rearranging (3.9) leads to the expression E t 1 vB = a 1 vB + b*ePt 1 + c 1 vB M1t + d 1 vB TEMPt (3.10) and the hence the autoregressive form E t = J + b*ePt 1 + cM1t + c*vM1t 1 + d*TEMPt + d*vTEMPt 1 + vE t 1 (3.11) where b*e < 0, c > 0, d > 0 28 The model specification in (3.11) is estimated for the case of Thailand. The estimation results are not satisfactory since the price variable has a wrong sign. In practice, the model specification of (3.5) is not known ex ante. The “correct” specification is hidden behind the pattern of electricity consumption. It is proposed to unveil the model specification by incorporating the Box–Jenkins (BJ) and the error– correction (EC) techniques into the modeling process. 3.2.2 Concept of the Autoregressive Model The error correction concept (EC) utilizes information from the model errors to improve the model performance. The EC technique has been applied to several studies such as the Sik and Frederick residential consumers’ demand for electricity energy report.16 The BJ and EC techniques may be incorporated into the modeling procedure in the two phases. In the first phase, the model (3.2) is estimated by the simple OLS method in the simple linear form E t = A + A1Pt + A 2 M1t + A3TEMPt + z t (3.12) where z t = error term The model errors are analyzed for their systematic pattern in the second phase and used to improve the model performance. The error of (3.12) in month t is specified by z t = E t Eˆ t = E t A A1Pt A2 M1t A3TEMPt (3.13) where Ê t = estimated electricity energy consumption Similarly, the error in month t–1 is specified z t 1 = E t 1 Eˆ t 1 = E t 1 A A1Pt 1 A2 M1t 1 A3TEMPt 1 (3.14) and the difference in errors between the two months may be specified by z t = z t z t 1 = E t Eˆ t 1 = E t 1 A1 Pt Pt 1 A 2 M1t M1t 1 A3 TEMPt TEMPt 1 Julian I. Silk and Frederick L. Joutz, “Short and long–run elasticities in US residential electricity demand: a co–integration approach”, Energy Economics 19(1997), pp.493–513 16 29 or z t = E t A1Pt A 2 M1t A3TEMPt (3.15) Campbell and Perron has proposed three approaches17 in determining the specification of z t . In the first approach, z t is specified by the series j z t = az t 1 + bi z t i + ε t (3.16) i=1 The number of lags in (3.16) is determined by the autocorrelation criteria. The selected number of lags is the number that removes autocorrelation between the error terms so that the remaining error term ε t is completely random with zero expected value. In the second approach, a constant is added to (3.16) in the form j z t = C + az t 1 + bi z t i + ε t (3.17) i=1 In the third approach, a time trend is added to capture effects of the potentially significant trend variables j z t = C + az t 1 + bi z t i + ct + ε t (3.18) i=1 3.2.3 Model Specification Silk and Frederick has proposed a test to select the model specification between (3.15)–(3.17). 1 8 For the case of Thailand, the Silk and Frederick tests will not be applied to select the model specification. Instead, it is proposed to select the model specification by applying the BJ and the EC techniques in the following sequence. From the general model specification j E t = A + A1Pt + A 2 M1t + A3Temp t + bi E t i + z t (3.19) i=1 where j z t = E t Eˆ t + gi z t i + ε t i=1 17 18 Ibid., page 500 Loc.cit., page 501 (3.20) 30 there are three major components explaining the variations in E t . The first component consists of the exogenous variables Pt , M1t , and Temp t . The autoregressive term j bi E t i is the second component, and the error correction term z t is the third i=1 component of the model. Two approached may be used to estimate the (3.19) modek. In the first approach, electricity energy consumption is detrended by its own time series. This approach will be referred to as the nondetrend method. Alternatively, electricity energy consumption may be detrended by an exogenous variable such as M1 , or by a time variable. This approach will be referred to as the M1 detrend method and the time detrend method. 3.2.4 Nondetrend Method The time series of electricity energy consumption are analyzed for autocorrelations and partial autocorrelations by the BJ technique. The stationary forms of the series for the four regions of EGAT, PEA, and MEA are identified and specified as EGAT 1 B 1 B12 EGATt 1 B 1 B12 CENTRALt 1 + 0.5417B + 0.3722B 2 = 0.2300 + 1 0.5187B 1 0.6813B12 u t + 0.1563B3 1 B 1 B12 NORTHEASTt = 0.3375 + 1 0.4815 u t 1 + 0.4058B 1 + 0.5910B12 + 0.3384B24 1 B 1 B12 SOUTHt = 0.0785 + 1+0.2169B11 u t 1 + 0.5750B + 0.3561B 2 + 0.1881B3 1 B 1 B12 NORTHt = 0.2417 + 1 0.6396B12 u t MEA 1 B 1 B12 MEAt = 0.5095 + 1 0.5302B 1 0.8657B12 u t PEA 1 B 1 B12 PEAt = 0.0575 + 1 0.5108B 1 0.7346B12 u t = 0.1886 + 1 0.6028B 1 0.5663B12 u t 31 3.2.5 M1 and Time Detrend Method The exogenous variable M1 is used to detrend the electricity energy demand by regressing electricity energy demand on M1. Series of the differences between actual and estimated electricity energy consumption are then identified for the AR and MA terms by the BJ technique. The patterns of AR and MA are then used as guidelines for the specification of the electricity energy demand model. Alternatively, a time variable may be used to detrend the electricity energy demand in the same manners as the M1 variable. The M1 regression results for the three power authorities are summarized below EGAT EGATt = 1,928.7580 + 0.0117M1t 1 adj R 2 = 0.9089 (t = 15.06) (t = 38.82) CENTRAL t = 1,692.2730 + 0.0084M1t 1 (t = 17.29) adj R 2 = 0.8977 (t = 36.42) NORTHEASTt = 55.0927 + 0.0012M1t 1 adj R 2 = 0.9035 (t = 3.93) (t = 37.61) SOUTH t = 84.8092 + 0.0010M1t 1 adj R 2 = 0.9402 (t = 9.50) (t = 48.76) NORTH t = 96.5830 + 0.0010M1t 1 adj R 2 = 0.8834 (t = 7.35) (t = 33.84) MEA MEA t = 1,349.2320 0.0021M1 + 0.0050M1t 2 adj R 2 = 0.8183 (t = 28.63) (t = –3.12) (t = 7.27) PEA PEA t = 483.0679 + 0.0080M1t 2 adj R 2 = 0.9329 (t = 6.17) (t = 44.30) Series of the differences between actual and estimated electricity energy consumption are then identified for the AR and MA terms by the BJ technique. The identification results from the software PROC ARIMA in SAS/ETS for the three power authorities are summarized below. 32 EGAT 1 0.8487B 1 0.9747B12 EGATt = 0.6152 + 1 0.3616B 1 0.6619B12 u t 1 0.5356B 0.2075B 1 0.9755B CENTRAL = 0.6785 + 1 0.6847B u 1 0.6133B 1 0.7194B NORTHEAST = 1.0672 + 1 + 0.2115B u 1 0.8849B 1 0.9676B SOUTH = 0.03691 + 1 0.4305B 1 0.0683B u 1 0.9029B 1 0.9805B NORTH = 0.0215 + 1 0.4337B 1 0.7038B u 2 12 12 t 12 t 17 t t 12 12 t t 12 12 t t MEA 1 0.3264B 0.2231B 1 0.6190B MEA 2 12 t = 4.8132 + u t PEA 1 0.3912B 0.2592B 1 0.9130B PEA 2 12 t = 1.9078 + 1 0.5783B12 u t When time is used as the detrend variable, electricity energy consumption is regressed on time with the following results EGAT EGATt = 3,904.0808 + 35.4386TIME t (t = 49.86) adj R 2 = 0.9139 (t = 40.18) CENTRAL t = 3,117.9401 + 25.2871TIME t adj R 2 = 0.8930 (t = 49.50) (t = 35.64) NORTHEASTt = 261.3636 + 3.8005TIME t adj R 2 = 0.9344 (t = 36.06) (t = 46.55) SOUTH t = 255.1707 + 3.1359TIME t adj R 2 = 0.9682 (t = 62.40) (t = 68.07) NORTH t = 269.6064 + 3.2151TIME t adj R 2 = 0.9160 (t = 38.46) (t = 40.72) MEA MEA t = 1,818.5050 + 8.4902TIME t (t = 52.88) (t = 21.92) adj R 2 = 0.7592 33 PEA adj R 2 = 0.9480 PEA t = 2,011.1310 + 24.3753TIME t (t = 50.44) (t = 51.09) As in the case of M1, series of the differences between actual and estimated electricity energy consumption are identified for the AR and MA terms by the BJ technique. The results are summarized below EGAT 1 0.4837B 0.3665B 1 0.9826B EGAT = 0.1145 + 1 + 0.1347B + 0.1900B 1 0.6743B u 1 0.9467B 1 0.9882B CENTRAL = 0.0270 + 1 0.4811B + 0.1405B 1 0.7047B u 1 0.4850B 0.2210B 1 0.9502B NORTHEAST = 0.0378 + 1 0.5060B u 1 0.9083B 1 0.9809B SOUTH = 0.0020 + 1 0.4185B 1 0.1230B 0.6626B u 1 0.4645B 0.2574B 1 0.9764B NORTH = 0.0183 + 1 0.6429B u 2 12 7 10 t 12 t 12 7 t 12 t 2 12 12 t t 12 6 12 t 2 12 12 t t MEA 1 0.5144B 0.4016B 1 0.9942B MEA 2 12 t = 0.0154 + 1 0.8228B12 u t PEA 1 0.4925B 0.2075B 1 0.6213B PEA 2 12 t = 0.2414 + u t The AR and MA terms identified by the BJ technique may be used as guidelines for electricity energy demand model specifications. The guidelines for model specification are summarized in Table 3.1 t 34 Table 3.1 Identifications of AR and MA Terms Nondetrend Method Power Authority AR, month M1 Detrend Method EGAT 1, 12, 13 MA , month 1, 12, 13 CENTRAL 1, 12, 13 1, 12, 13 1 11 1, 12, 13 1, 12, 13 1, 12, 13 1, 6, 7, 12, 13 12 1, 12, 13 1, 12, 13 1, 2, 12, 13, 14 12 MEA 1, 2, 3, 4, 12, 13, 14, 15, 16 1, 2, 12, 13, 14, 24, 25, 26, 36, 37, 38 1, 2, 3, 4, 12, 13, 14, 15, 16 1, 12, 13 1, 2, 12, 13, 14 1, 12, 13 1, 12, 13 – 1, 12, 13 1, 12, 13 1, 2, 12, 13, 14 1, 2, 12, 13, 14 12 PEA 1, 2, 12, 13, 14 1, 2, 12, 13, 14 NORTHEAST SOUTH NORTH AR, month 12 7 1, 2, 12, 13, 14 1, 12, 13 MA , month 1, 12, 13 Time Detrend Method 12 AR, month 1, 2, 12, 13, 14 1, 12, 13 MA , month 7, 10, 12, 19, 22 1, 7, 12, 13, 19 12 – 3.3 Model for Electricity Energy Consumption Data on electricity energy consumption are collected from EGAT, MEA for the period January 1991 through September 2003. The PEA data are available only for the period October 1991 through September 2003. Electricity energy consumption at the EGAT level will be modeled for all of the regions, the central region, the northeastern region, the northern region, and the southern region. Due to constraints on the data availability, the MEA and PEA models will not be disaggregated into regions. 3.3.1 Nondetrend Model From the correlation analysis of EGAT electricity energy consumption and the M1 at various lags it is found that the pairwise correlation of the M1 at lag 1 and electricity energy is higher the others. The M1 at lag 1 is then selected to be one of the exogenous variables in the forecasting model of EGAT electricity energy consumption. EGAT The forecasting model of EGAT electricity energy consumption may be written in the form EGATt = A + A1EGATt 1 + A 2 EGATt 12 + A3EGATt 13 + A9 M1t 1 + A10TEMPt + A11PRICE t + z t z t = θ1z t 1 + θ12 z t 12 + θ13z t 13 + ε t (3.21) 35 where TEMPt is the mean of the cumulative average daily temperature in Thailand in month t and PRICE t is the average consumer electricity price per KWh in month t. The forecasting model of EGAT electricity energy consumption in the central region may be written as CENTRAL t = A + A1CENTRAL t 1 + A 2CENTRAL t 12 + A3CENTRAL t 13 + A9 M1t 1 + A10TEMP_Ct + A11PRICE_PEA t + z t z t = θ1z t 1 + θ12 z t 12 + θ13z t 13 + ε t (3.22) where TEMP_Ct is the mean of the cumulative average daily temperature in month t in the central region of PEA distribution system and PRICE_PEA t is the average PEA consumer electricity price per KWh in month t In the northeast region, the forecasting model of EGAT electricity energy consumption can be written as NORTHEASTt = A + A1NORTHEASTt 1 + A 2 NORTHEASTt 2 + A 3 NORTHEASTt 3 + A 4 NORTHEASTt 4 + A5 NORTHEASTt 12 + A6 NORTHEASTt 13 + A7 NORTHEASTt 14 + A8 NORTHEASTt 15 + A9 NORTHEASTt 16 + A10 M1t 1 + A11TEMP_NE t + A12 PRICE_PEA t + z t z t = θ1z t 1 + ε t (3.23) where TEMP_NE t is the mean of the cumulative average daily temperature in month t in the northeast region of PEA distribution system. The forecasting model of EGAT electricity energy consumption in the south region may be written as SOUTH t = A + A1SOUTH t 1 + A 2SOUTH t 2 + A 3SOUTH t 12 + A 4SOUTH t 13 + A5SOUTH t 14 + A 6SOUTH t 24 + A 7SOUTH t 25 + A8SOUTH t 26 + A9SOUTH t 36 + A10SOUTH t 37 + A11SOUTH t 38 + A12 M1t 1 + A13TEMP_St + A14 PRICE_PEA t + z t z t = θ11z t 11 + ε t (3.24) where TEMP_St is the mean of the cumulative average daily temperature in month t in the south region of PEA distribution system. Finally the forecasting model of EGAT electricity energy consumption in month t in the north region may be written in as 36 NORTH t = A + A1NORTH t 1 + A 2 NORTH t 2 + A 3 NORTH t 3 + A 4 NORTH t 4 + A5 NORTH t 12 + A 6 NORTH t 13 + A 7 NORTH t 14 + A8 NORTH t 15 + A9 NORTH t 16 + A10 M1t 1 + A11TEMP_N t + A12 PRICE_PEA t + z t z t = θ12 z t 12 + ε t (3.25) where TEMP_N t is the mean of the cumulative average daily temperature in month t in the north region of PEA distribution system. Similarly, from the regression analysis, the M1 at lag 0 and lag 2 are selected to be the exogenous variables i the forecasting model of MEA electricity energy consumption and the M1 at lag 2 is the exogenous variable in the forecasting model of PEA electricity energy consumption. MEA The forecasting model of MEA electricity energy consumption may be written in the form MEA t = A + A1MEA t 1 + A 2 MEA t 12 + A3MEA t 13 + A 4 M1t + A5 M1t 2 + A6 TEMP_C t + A 7 PRICE_MEA t + z t z t = θ1z t 1 + θ12 z t 12 + θ13z t 13 + ε t (3.26) PEA The forecasting model of PEA electricity energy consumption may be written as PEA t = A + A1PEA t 1 + A 2 PEA t 12 + A3PEA t 13 + A 4 M1t 2 + A 6 TEMP + A 7 PRICE_PEA t + z t z t = θ1z t 1 + θ12 z t 12 + θ13z t 13 + ε t (3.27) The parameters in the forecasting models are estimated by PROC Model in SAS software. The equation in forecasting model of EGAT electricity energy consumption can be written as 37 EGATt = 430.0440 + 0.7129EGATt 1 + 0.7754EGATt 12 0.6407EGATt 13 (t = –2.24) (t = 13.32) (t = 15.35) (t = –9.32) + 0.0021M1t 1 + 1.1989TEMPt 160.0250PRICEt + zˆ t (t = 4.96) (t = 4.72) (t= –1.57) ẑ t = 0.3769z t 1 0.3792z t 12 0.2196z t 13 (t = –4.31) (t= –3.83) where adj R 2 = 0.9850 . (t = –2.20) The signs of an exogenous variables are correct as expected. The price variable has the correct sign and is not significant at 0.5 significant level but is significant at 0.13 significant level. The parameters of part errors Z t 1 , Z t 12 and Z13 are significant and can be used the error–correction model. The equations of forecasting models are summarized in Table 3.2 Table 3.2 Estimation Results of Nondetrend Electricity energy Consumption Model by Power Authorities EGAT : adj R 2 = 0.9850 EGATt = 430.0440 + 0.7129EGATt 1 + 0.7754EGATt 12 0.6407EGATt 13 (t = –2.24) (t = 13.32) (t = 15.35) (t = –9.32) + 0.0021M1t 1 + 1.1989TEMPt 160.0250PRICEt + zˆ t (t = 4.96) (t = 4.72) (t= –1.57) ẑ t = 0.3769z t 1 0.3792z t 12 0.2196z t 13 (t = –4.31) (t= –3.83) (t = –2.20) EGAT Central : adj R 2 = 0.9803 CENTRAL t = 545.6500 + 0.6868CENTRAL t 1 + 0.7170CENTRAL t 12 (t = –2.53) (t = 12.60) (t = 13.15) 0.5873CENTRAL t 13 + 0.0019M1t 1 + 1.3475TEMP_Ct (t = –8.59) (t = 5.29) 210.2510PRICE_PEA t + zˆ t (t = –2.14) ẑ t = 0.3106z t 1 + 0.2414z t 12 (t = –3.58) (t = –2.54) (t = 4.87) 38 Table 3.2 (Continued) EGAT Northeast : adj R 2 = 0.9846 NORTHEASTt = 0.0412 + 0.2612NORTHEASTt 1 + 0.1941NORTHEASTt 3 (t = 1.17) (t = 3.55) (t = 2.72) + 0.8621NORTHEASTt 12 0.2344NORTHEASTt 13 (t = 21.45) (t = –2.99) 0.2035NORTHEASTt 15 + 0.0003M1t 1 + 0.1140TEMP_NE t (t = –2.75) (t = 5.90) (t = 5.69) 74.6667PRICE_PEA t + zˆ t (t = –4.68) ẑ t = 0.3047z t 12 (t = –3.08) EGAT South : adj R 2 = 0.9895 SOUTH t = 205.6990 + 0.3448SOUTH t 1 + 0.2403SOUTH t 2 (t = –4.95) (t = 4.96) (t = 4.96) + 0.4449SOUTH t 12 0.4438SOUTH t 13 + 0.2490SOUTH t 24 (t = 6.31) (t = –6.59) (t = 4.55) + 0.0001M1t 1 + 0.2831TEMP_St + 4.6599PRICE_PEAt + zˆ t (t = 4.42) (t = 5.96) (t = 0.35) 39 Table 3.2 (Continued) EGAT North : adj R 2 = 0.9821 NORTH t = 132.0760 + 0.2219NORTH t 1 + 0.1450NORTH t 2 (t = –4.25) (t = 3.05) (t = 3.25) + 0.6065NORTH t 12 0.2322NORTH t 13 0.0003M1t 1 (t = 11.26) (t = –3.24) (t = 6.84) + 0.2252TEMP_N t 13.0464PRICE_PEA t + zˆ t (t = 7.53) (t = –0.73) MEA : adj R 2 = 0.9616 MEA t = 276.8910 181.8220PRICE_MEA t + 0.6314MEA t 1 + 0.5108MEA t 12 (t = –2.40) (t = –4.09) (t = 12.77) (t = 9.00) 0.4283MEA t 13 + 1.0642TEMP_Ct + zˆ t (t = 7.02) (t = 7.01) ẑ t = 0.3183z t 1 (t = –3.49) PEA : adj R 2 = 0.9842 PEA t = 759.1100 + 0.5505PEA t 1 + 0.1654PEAt 12 + 0.0021M1t 2 (t = –3.68) (t = 9.64) (t = 3.02) + 22.1426PRICE_PEA t + 1.1862TEMPt + zˆ t (t = 0.23) ẑ t = 0.3314z t 1 + 0.2442z t 12 (t = –3.71) (t = 2.51) (t = 5.77) (t = 7.58) 40 3.3.2 Detrend Model: M1 The forecasting models of electricity consumptions can be written in the form EGAT EGATt = A + A1EGATt 1 + A 2 EGATt 12 + A3EGATt 13 + A 4 M1t 1 + A5TEMPt + A 6 PRICE t + z t z t = θ1z t 1 + θ12 z t 12 + θ13z t 13 + ε t (3.28) EGAT : Central CENTRAL t = A + A1CENTRAL t 1 + A 2CENTRAL t 2 + A3CENTRAL t 12 + A 4 CENTRAL t 13 + A5CENTRAL t 14 + A 6 M1t 1 + A 7 TEMP_C t + A8 PRICE_PEA t + z t z t = θ12 z t 12 + ε t (3.29) EGAT : Northeast NORTHEASTt = A + A1NORTHEASTt 1 + A 2 NORTHEASTt 12 + A3 NORTHEASTt 13 + A 4 M1t 1 + A5TEMP_NE t + A6 PRICE_PEA t + z t z t = θ 7 z t 7 + ε t (3.30) EGAT : South SOUTH t = A + A1SOUTH t 1 + A 2SOUTH t 12 + A3SOUTH t 13 + A 4 M1t 1 + A5TEMP_St + A 6 PRICE_PEA t + z t z t = θ1z t 1 + θ12 z t 12 + θ13z t 13 + ε t (3.31) 41 EGAT : North NORTH t = A + A1NORTH t 1 + A 2 NORTH t 12 + A 3 NORTH t 13 + A 4 M1t 1 + A5TEMP_N t + A6 PRICE_PEA t + z t z t = θ1z t 1 + θ12 z t 12 + θ13z t 13 + ε t (3.32) MEA MEA t = A + A1MEA t 1 + A 2 MEA t 2 + A3MEA t 12 + A 4 MEA t 13 + A5MEA t 14 + A6 M1t + A 7 M1t 2 + A8TEMP_C t + A9 PRICE_MEA t + ε t (3.33) PEA PEA t = A + A1PEA t 1 + A 2 PEA t 2 + A3PEA t 12 + A 4 PEA t 13 + A 5PEA t 14 + A6 M1t 2 + A7 TEMPt + A8 PRICE_PEA t + z t z t = θ12 z t 12 + ε t (3.34) Estimations of the M1 detrend model for the three power authorities are summarized in Table 3.3. Table 3.3 Estimation Results of M1 Detrend Electricity energy Consumption Model by Power Authorities EGAT : adj R 2 = 0.9850 EGATt = 430.0440 + 0.7129EGATt 1 + 0.7754EGATt 12 0.6407EGATt 13 (t = –2.24) (t = 13.32) (t = 15.35) (t = –9.32) + 0.0021M1t 1 + 1.1989TEMPt 160.0250PRICEt + zˆ t (t = 4.96) (t = 4.72) ẑ t = 0.3769z t 1 0.3792z t 12 0.2196z t 13 (t = –4.31) (t = –3.83) (t = –2.20) (t = –1.52) 42 Table 3.3 (Continued) EGAT Central : adj R 2 = 0.9825 CENTRAL t = 1,063.6700 + 0.3444CENTRAL t 1 + 0.3209CENTRAL t 2 (t = –4.22) (t = 4.84) (t = 4.53) + 0.6274CENTRAL t 12 0.3364CENTRAL t 13 (t = 10.27) (t = –4.56) 0.1742CENTRAL t 14 + 0.0023M1t 1 + 2.0572TEMP_Ct (t = –2.19) (t = 5.58) (t = 6.40) 237.3940PRICE_PEA t + zˆ t (t = –2.08) ẑ t = 0.2757z t 12 (t = –2.52) EGAT Northeast : adj R 2 = 0.9840 NORTHEASTt = 96.2250 + 0.3307NORTHEASTt 1 + 0.7160NORTHEASTt 12 (t = –2.78) (t = 4.95) (t = 14.34) 0.2617NORTHEASTt 13 + 0.0003M1t 1 + 0.1917TEMP_NE t (t = –3.62) 31.9091PRICE_PEA t (t = –1.53) (t = 6.85) (t = 6.31) 43 Table 3.3 (Continued) EGAT South : adj R 2 = 0.9906 SOUTH t = 32.4147 + 0.7879SOUTH t 1 + 0.9225SOUTH t 12 (t = –1.47) (t = 14.02) (t = 18.72) 0.7687SOUTH t 13 + 0.0001M1t 1 + 0.0503TEMP_St (t = –10.86) (t = 2.58) (t = 2.00) + 0.5819PRICE_PEA t + zˆ t (t = 0.08) ẑ t = 0.3599z t 1 0.4353z t 12 0.2773z t 13 (t = –4.01) (t = –4.55) (t = –2.85) EGAT North : adj R 2 = 0.9811 NORTH t = 90.5757 + 0.4793NORTH t 1 + 0.6462NORTH t 12 (t = –3.49) (t = 6.40) (t = 11.66) 0.3411NORTH t 13 + 0.0002M1t 1 + 0.1621TEMP_N t (t = –4.16) (t = 6.11) (t = 6.22) 9.9956PRICE_PEA t + zˆ t (t = –0.66) ẑ t = 0.2435z t 1 (t = –2.31) PEA : adj R 2 = 0.9842 PEA t = 831.9630 + 0.3112PEA t 1 + 0.1878PEAt 2 + 0.3991PEA t 12 (t = –3.55) (t = 4.05) (t = 3.16) (t = 5.26) 0.1675PEA t 13 + 0.0020M1t 2 + 1.2988TEMPt (t = –2.08) (t = 6.11) + 14.6278PRICE_PEA t (t = 0.12) (t = 5.79) 44 3.3.3 Detrend Model: Time The forecasting models of electricity consumption may be written as EGAT EGATt = A + A1EGATt 1 + A 2 EGATt 2 + A3EGATt 12 + A 4 EGATt 13 + A5EGATt 14 + A5TIME t + A 6 TEMPt + A 7 PRICE t + z t z t = θ7 z t 7 + θ10 z t 10 + θ12 z t 12 + θ19 z t 19 + θ 22 z t 22 + ε t (3.35) EGAT : Central CENTRAL t = A + A1CENTRAL t 1 + A 2CENTRAL t 12 + A3CENTRAL t 13 + A 4TIME t + A5TEMP_C t + A 6 PRICE_PEA t + z t z t = θ1z t 1 + θ7 z t 7 + θ12 z t 12 + θ13z t 13 + θ19 z t 19 + ε t (3.36) EGAT : Northeast NORTHEASTt = A + A1NORTHEASTt 1 + A 2 NORTHEASTt 2 + A3 NORTHEASTt 12 + A 4 NORTHEASTt 13 + A5 NORTHEASTt 14 + A6 TIME t + A 7 TEMP_NE t + A8PRICE_PEA t + z t z t = θ12 z t 12 + ε t (3.37) EGAT : South SOUTH t = A + A1SOUTH t 1 + A 2SOUTH t 12 + A 3SOUTH t 13 + A 4TIME t + A5TEMP_St + A 6 PRICE_PEA t + z t z t = θ1z t 1 + θ6 z t 6 + θ7 z t 7 + θ12 z t 12 + θ13z t 13 + ε t (3.38) 45 EGAT : North NORTH t = A + A1NORTH t 1 + A 2 NORTH t 2 + A3 NORTH t 12 + A 4 NORTH t 13 + A5 NORTH t 14 + A6TIME t + A 7 TEMP_N t + A8PRICE_PEA t + z t z t = θ12 z t 12 + ε t (3.39) MEA MEA t = A + A1MEA t 1 + A 2 MEA t 2 + A3MEA t 12 + A 4 MEA t 13 + A5MEA t 14 + A6 TIME t + A7 TEMP_Ct + A8 PRICE_MEA t + z t z t = θ12 z t 12 + ε t (3.40) PEA PEA t = A + A1PEA t 1 + A 2 PEA t 2 + A3PEA t 12 + A 4 PEA t 13 + A 5PEA t 14 + A6 TIME t + A7 TEMPt + A8 PRICE_PEA t + ε t (3.41) Estimations of the time detrend model for the three power authorities are summarized in Table 3.4. Table 3.4 Estimation Results of Time Detrend Electricity energy Consumption Models by Power Authorities EGAT : adj R 2 = 0.9853 EGATt = 280.2028 + 0.3329EGATt 1 + 0.3516EGATt 2 + 0.7127EGATt 12 (t = 0.69) (t = 4.42) (t = 4.86) (t = 11.71) 0.3868EGATt 13 0.3082EGATt 14 + 14.9263TIME (t = –5.01) (t = –4.03) + 2.2613TEMPt 649.9750PRICE + zˆ t (t = 5.60) (t = –2.57) ẑ t = 0.1845z t 7 + 0.2030z t 10 0.2419z t 12 (t = 2.09) (t = 2.32) (t = –2.36) (t = 4.49) 46 Table 3.4 (Continued) EGAT Central : adj R 2 = 0.9792 CENTRAL t = 51.1803 + 0.8732CENTRAL t 1 + 0.8580CENTRAL t 12 (t = 0.24) (t = 24.69) (t = 19.07) 0.8602CENTRAL t 13 + 4.9390TIME + 0.8612TEMP_Ct (t = –17.70) (t = 3.09) (t = 3.55) 265.9370PRICE_PEA t + zˆ t (t = –1.88) ẑ t = 0.4279z t 1 0.4683z t 12 0.2667z t 13 (t = –5.25) (t = –5.26) (t = –2.80) EGAT Northeast : adj R 2 = 0.9885 NORTHEASTt = 75.0712 + 0.3392NORTHEASTt 1 + 0.1347NORTHEASTt 2 (t = 2.00) (t = 5.21) (t = 2.08) + 2.3671TIME + 0.3807TEMP_NE t 65.2448PRICE_PEA t + zˆ t (t = 8.78) (t = 9.62) (t = –3.36) ẑ t = 0.8443z t 12 (t = 15.42) EGAT South : adj R 2 = 0.9917 SOUTH t = 29.3433 + 0.7997SOUTH t 1 + 0.8341SOUTH t 12 (t = –1.38) (t = 17.99) (t = 15.09) 0.8375SOUTH t 13 + 0.8572TIME + 0.1477TEMP_St (t = –14.74) (t = 4.62) (t = 4.14) 31.8448PRICE_PEA t + zˆ t (t = –3.15) ẑ t = 0.3992z t 1 0.1687z t 6 0.4256z t 12 0.2649zt 13 (t = –4.67) (t = –2.22) (t = –4.43) (t = –2.80) 47 Table 3.4 (Continued) EGAT North : adj R 2 = 0.9872 NORTH t = 76.8115 + 0.2605NORTH t 1 + 0.1396NORTH t 2 (t = –2.21) (t = 4.09) (t = 2.25) + 2.3190TIME + 0.3970TEMP_N t 60.0873PRICE_PEA t + zˆ t (t = 9.54) (t = 10.62) (t = –3.48) z t = 0.8128z t 12 (t = 13.60) MEA : adj R 2 = 0.9602 MEA t = 27.8191 + 0.4026MEA t 1 + 0.3921MEA t 2 + 0.6602MEA t 12 (t = –0.16) (t = 5.59) (t = 5.60) (t = 11.05) 0.3897MEA t 13 + 0.3135MEA t 14 + 4.5058TIME (t = –5.28) (t = –4.27) (t = 4.61) + 1.1600TEMP_Ct 328.1980PRICE_MEA t + zˆ t (t = 6.30) (t = –3.46) z t = 0.2577z t 12 (t = –2.42) PEA : adj R 2 = 0.9821 PEA t = 254.4270 + 0.5691PEA t 1 + 0.4142PEA t 12 0.3280PEA t 13 (t = –0.78) (t = 8.83) (t = 5.11) (t = –4.04) + 10.9533TIME + 1.8337TEMPt 394.4580PRICE_PEA t (t = 4.76) (t = 6.17) (t = –2.17) 3.4 Summary of Estimation Results Three different electricity energy consumption model specifications : nondtrend, M1 detrend, and time detrend, have been estimated and tested for the case of Thailand. All of the three specifications have high explanatory power. The adjusted R2 for the nondetrend models range from 0.9616 to 0.9895. The error correction terms 48 are found to enhance the explanatory power of the detrend model significantly, with the exception of model for EGAT south, EGAT north, and PEA. Non–price exogenous variables in the nondetrend models have expected signs that are statistically significant at the 0.05 level. The price variable has expected sign and is significant at the 0.05 level for the EGAT central, the EGAT northeast, and the MEA models. The remaining nondetrend models have expected price signs that are not significant with the exception of the EGAT south and the PEA models, where the price signs are positive but not statistically significant. The adjusted R2 for the M1 detrend models are between 0.9678 to 0.9906. The error correction terms enhance the explanatory power of all models significantly, except for the EGAT northeast, MEA, and PEA models. The statistically significant non- price variables have expected signs. The price variable has the expected negative sign but significant only for the EGAT central and MEA models. The exception is the EGAT south and the PEA models where the price variable has positive sign but not statistically significant. The adjusted R2 for the time detrend models range from 0.9602 to 0.9917. The error correction terms are found to enhance the explanatory power of the models significantly, except for the PEA model. The statistically significant non-price variables have the expected signs. The price variable has the expected negative sign which is significant at the 0.005 level for all models. Table 3.5 compares the model errors between the nondetrend, the M1 detrend, and the time detrend models. The errors of the three types of models are compatible although the standard deviations of the errors of the M1 detrend model and the time detrend models are slightly lower than the nondetrend model. Nevertheless, the errors of all the three types of model are less than three percent. The modeling exercises suggest that the model for electricity energy consumption may have the following two specifications 3 E t = CON + Ai E t i + T*TEMPt + M jM1t j + P*PRICE t 1 + u t (3.42) Et = CON + Ai Et i + T*TEMPt + B*t + P*PRICE t 1 + u t (3.43) iI j=0 iI where i is the set of lagged variables determined from the BJ technique, and u t is the residual E t Eˆ t which is the error correction term in the model. The foregoing modeling exercises are all in linear forms. In the next chapter, the double log specification will be considered as a potential forecasting model. The double log specification has an advantage over the linear specification in that the model can provide the variable elasticities directly. The modeling exercise suggests that there may be collinearity among the autoregressive variable E t i . In order to alleviate the collinearity problem, it is proposed to determine the number of autoregressive variables by the following procedure. From the BJ autoregressive model which can be specified generally as E t = a 0 + a t i E t i + ε t iI (3.44) 49 where I is the set of lagged variables identified by the BJ technique, and t is the remaining random error with zero expected value, only the autoregressive variables that are statistically significant will be retained in the model. The removal of nonsignificant variables would increase the explanatory of (3.44) by increasing its the adjusted R2. The selected variables will be specified as the autoregressive term in the electricity energy consumption model. Since collinearity may also exist in the electricity energy demand model, its severity may be further reduced by taking into account the seasonal effect by considering the signs of the autoregressive variables such as E t 12 and E t 24 , E t 1 , E t 13 and E t 25 . If these variables have the same signs, the number of the autoregressive variables may be reduced further to potentially increase the adjusted R2 . Granges et. al modeled the consumption of electricity energy in the US and found the month variables to be significant. It is proposed that the month variables are specified as dummy variables in the electricity energy consumption model for Thailand. The proposed general form of the model may thus be specified as nE t = f nE t i ; i Is , nTEMPt , nPt 1 , nM t j ; j = 0, 1, 2, 3, FEB, MAR, APR, MAY, JUN, JUL, AUG, SEP, OCT, NOV, DEC (3.45) nE t = f nE t i ; i Is , nTEMPt , nPt 1 , t, FEB, MAR, APR, MAY, JUN, JUL, AUG, SEP, OCT, NOV, DEC (3.46) Table 3.5 Model EGAT EGAT central EGAT northeast EGAT south EGAT north MEA PEA Comparison of Errors between Nondetrend, M1 detrend, and Time Detrend Models Nondetrend Model M1 Detrend Model adj.R2 error Relative error adj.R2 error Relative error Mean SD Mean SD Mean SD Mean SD adj.R2 Time detrend Model error Relative error Mean SD Mean SD 0.9850 0.9803 0.9846 –0.4707 –0.1826 –0.1451 172.7385 142.8049 18.9675 0.0195 0.0221 0.0245 0.0162 0.0164 0.0187 0.9850 0.9825 0.9840 –0.4707 –0.6389 0.0000 173.3587 132.1482 19.7737 0.0200 0.0210 0.0267 0.0162 0.0153 0.0204 0.9853 0.9792 0.9885 0.8770 0.6957 –0.4871 168.0132 146.2536 18.0601 0.0195 0.0226 0.0279 0.0151 0.0189 0.0242 0.9895 0.9821 0.9616 0.9842 0.0000 0.0000 0.1025 0.0826 11.6963 17.6736 71.1919 114.8053 0.0164 0.0254 0.0233 0.0202 0.0133 0.0195 0.0156 0.0210 0.9960 0.9811 0.9678 0.9842 –0.0240 0.0036 0.0000 0.0000 12.0356 18.1669 65.2493 113.8868 0.0172 0.0262 0.0214 0.0207 0.0140 0.0200 0.0151 0.0267 0.9917 0.9872 0.9602 0.9821 0.0675 –0.2855 –0.2013 0.0000 11.2607 16.2315 70.4354 121.7313 0.0162 0.0252 0.0225 0.0226 0.0132 0.0225 0.0169 0.0288 50