impact of climate change on electricity demand of ... - Editorial Express

6 downloads 17040 Views 304KB Size Report
May 22, 2013 - energy demand of the representative building can be exactly calculated. ... Hourly electricity loads were obtained from the website of the Energy .... separate dummies for public holidays that fall on weekdays and the eve of ...
IMPACT OF CLIMATE CHANGE ON ELECTRICITY DEMAND OF SINGAPORE Tilak K. Doshi, Principal Fellow, Energy Studies Institute, National University of Singapore, [email protected] Abhishek Rohatgi, Research Associate, Energy Studies Institute, National University of Singapore, [email protected] Nahim Bin Zahur, Energy Analyst, Energy Studies Institute, National University of Singapore, [email protected] Yuen Kah Hung, Energy Analyst, Energy Studies Institute, National University of Singapore, [email protected]

Introduction Climate change might have consequences such as higher average temperatures and greater humidity. The impact of such weather fluctuations on future electricity demand will affect the electricity generation capacity required in the future, a major concern for policymakers throughout the world given the expense and time required to make investments in the electricity sector. However, the relationship between weather fluctuations and electricity demand has not been established in the Singapore context to date. Singapore’s electricity consumption has grown from 37,709 GWh in 2009 to 41,725 GWh in 2011 [1]. The industrial sector is the most energy intensive followed by commercial and service sectors. Given Singapore’s equatorial tropical climate, a significant proportion of electricity is consumed for cooling and in particular airconditioning. This suggests electricity consumption could vary with changes in temperature. As the figure below illustrates, household energy consumption peaks during the middle of the year when the temperature is highest. To ensure supply side reliability, therefore, it is imperative to estimate the relationship between electricity demand and climate in order to provide a framework for forecasting how climate change in the future might affect electricity demand. Figure 1 Singapore’s average monthly electricity consumption and temperature, 2011

Source: Singapore Energy Statistics 2012 [1] Existing studies carried out in different parts of the world have found that weather fluctuations have an impact on the demand for electricity [2] [3] [4]. There are two main approaches that have been used to study the relationship between climate variables (in particular, ambient temperature) and electricity demand. The first relies on building energy simulation software packages (such as the DOE-2.1E building simulation software developed by Lawrence Berkeley National Laboratory) in which a detailed model of a representative building can be set up [5]. Using these simulation programs, the impact of changing any of the input variables (including temperature) on the energy demand of the representative building can be exactly calculated. A study using this approach found that for Switzerland, a 4.4 degree Celsius increase in temperature relative to the reference case would lead to a 33-44% decrease in the annual heating energy demand for Swiss residential buildings [6]. However, while this approach has proven useful for conducting detailed micro-level analyses of the energy use characteristics of individual buildings, 1

it has some shortcomings when it comes to the analysis of aggregate energy consumption. This is because even within specific sectors such as the residential sectors, actual buildings may vary considerably from the sample building considered in the simulation programs, while at the economy-wide level different sectors can have quite distinct energy usage patterns. Thus in this paper we utilize the alternative approach, which involves using econometric techniques to find out the relationship between electricity demand and climate variables, using historical data on both electricity consumption and temperature (and other climate variables). This approach is well-suited to the study of energy use patterns and relationships at the aggregate level. Crowley and Joutz [3] specify hourly load models of electricity consumption in the Pennsylvania-New Jersey-Maryland (PJM) region of the United States using seasonal time series regression analysis, with temperature-based variables among the explanatory variables; this allows the effect of temperature on electricity demand to be identified for different hours of the day. A key insight from this study is that the effect of temperature on peak electricity demand is different from the effect on average demand. In [4], the monthly electricity consumption of 20 air-conditioned office buildings in Hong Kong is regressed on a set of variables including a temperature variable, the building envelope heat gain and the internal load density; the inclusion of the latter two variables controls for differences across buildings, allowing for a better estimate of the impact of temperature. In this paper we use an hour-by-hour modeling approach to estimate the relationship between electricity demand and weather variables in Singapore. The specification of the weather component of the model is detailed and allows us to estimate the impact of short-run fluctuations in both temperature and humidity (including lagged effects from a few hours back), seasonal differences in how temperature affects electricity demand, and long-run effects of temperature on demand. By estimating separate models for each hour of the day, we are able to capture the effect of weather variables on electricity demand at different times of the day. Crowley and Joutz [3] demonstrate, in the context of the U.S., that the effect of temperature on peak electricity demand is different from the effect on average demand. From a policy perspective it is crucial to understand the effect of climate change on peak energy demand as well as average energy demand in order to adequately plan future generation capacity. We find that temperature changes have a significant positive impact on electricity demand. Short-run temperature elasticities range from 0.3-0.5 depending on the hour of the day, while there are additional adjustments to temperature change in the long-run with elasticities ranging from 0.2-0.8. Both short-run and long-run temperature elasticities are larger during the night than in the day, a result of the fact that the residential load takes up a much bigger proportion of the overall load at night. This indicates that the effect of future climate change on peak demand might be less than the effect on average electricity demand. There are seasonal differences in how temperature affects electricity demand, with the impact marginally higher during the warmer months of the year. We also find that increases in humidity have a small but noticeable upward impact on electricity demand. Our paper thus contributes to the understanding of how climate change impacts on electricity demand in countries with equatorial climate.

2

Data The dataset used consists of hourly electricity loads, dry-bulb temperature and humidity in Singapore for a 10year period (2003-2012). Hourly electricity loads were obtained from the website of the Energy Market Company of Singapore [7] while the weather data was provided by the National Environment Agency, Singapore [8]. The load series is divided into 24 separate daily time series corresponding to 24 hours of the day. In a similar fashion we obtain separate daily time series for the temperature and humidity at each hour of the day. The effect of economic activity is captured by including quarterly real GDP (at 2005 S$ prices), obtained from [9]. Figure 2 shows the temperature time series for hours 2, 8, 14 and 20 (corresponding to 2 AM, 8 AM, 2 PM and 8 PM) respectively. As a result of the equatorial location of Singapore, the temperature remains relatively steady over time, not straying beyond the 23°C to 34°C range. Nevertheless some patterns can be observed. The temperature is highest during the afternoon, is somewhat lower in the early evening and lowest late at night and early in the morning. The temperature is also most variable (on a day-to-day basis) during the afternoon. Figure 2 Temperature time series for hours 2, 8, 14 and 20 Temperature Hour 2

Temperature Hour 8 36

Temperature (degrees Celsius)

Temperature (degrees Celsius)

36

32

28

24

03

04

05

06

07 08 Year

09

10

11

32

28

24

12

03

04

05

Temperature Hour 14

07 08 Year

09

10

11

12

10

11

12

Temperature Hour 20 36 Temperature (degrees Celsius)

36 Temperature (degrees Celsius)

06

32

28

24

03

04

05

06

07

08

09

10

11

12

32

28

24

03

Year

04

05

06

07

08

09

Year

Another feature in the temperature time series shown above is the presence of seasonal fluctuations. Although Singapore’s temperature and relative humidity remains rather uniform throughout the year, there are nonetheless two main seasons that corresponds to certain months of the year. Usually the Northeast Monsoon months of December to February are the cooler and rainier part of the year. This is followed by a relatively hotter and drier Inter-monsoon period in March, and then the Southwest Monsoon season from April to October which is the hottest part of the year. A second inter-monsoon period occurs in November before the Northeast Monsoon starts and the cycle repeats itself. This pattern can be observed clearly in Figure 3 below: there are distinct (albeit small) differences in temperature across the different seasons for almost each hour of the day.

3

Temperature (degrees Celsius)

Figure 3 Average seasonal temperature in different hours of the day 32 30 28

Northeast Monsoon  (Dec‐Feb)

26

Intermonsoon period  (Mar, Nov)

24

Southwest Monsoon  (Apr‐Oct)

22 20 1 3 5 7 9 11 13 15 17 19 21 23 Hour of the Day

Figure 4 shows the electricity load time series for hour 2, 8, 14 and 20. The consistent growth in the load value indicates the presence of trend dynamics in the time series and reflects Singapore’s GDP growth during the 2003 – 2012 period. Looking at the yearly profiles, electricity load is low at the beginning of the year, increases and then again decreases towards the end of the year. Hence, the load profile also has seasonal characteristics consistent with those observed for temperature. The seasonality is observed at daily, weekly and monthly level. The load is high during the day and low during night time. Weekdays have higher load than weekends due to more economic activity. Similarly, higher temperature leads to higher demand and hence the electricity load is higher during months with higher average temperature (corresponding to the April-October Southwest Monsoon period). Figure 4 Electricity load time series for hours 2, 8, 14 and 20

4

Figure 5 shows the autocorrelation function (ACF) and partial autocorrelation functions (PACF) for the electricity load in hour 2. The ACF and PACF values are used to determine the dynamics of the time series. The stochastic component of a time series can be described by either autoregressive (AR) terms or moving average (MA) terms or both. The autocorrelation function decreases gradually in Figure 5 while partial correlation factors have abrupt changes (spikes) in the values. This suggests that an AR model can be used to describe the stochastic behavior of the electricity load time series (see, for instance, [10]). Although the ACF and PACF are shown only for hour 2, a very similar pattern is observed for the remaining 23 hours as well. The exact AR terms are selected using the minimum Akaike Information Criterion (AIC) and the Schwarz criterion (SIC) during the modelling phase (discussed later). Figure 5 Autocorrelation function (ACF) & partial autocorrelation function (PACF) of electricity load, hour 2

Figure 5 shows that electricity demand is highly autocorrelated across days. To decide whether to model electricity demand in levels or in differences, we conducted unit root tests on each of the load time series as well as the other major variables. Table 1 shows the results of the Augmented Dickey-Fuller (ADF) unit root test on the electricity load time series for hour 2. The t-statistic for the ADF test shows that the probability of having a unit root in the load time series is less than 1%. Hence, the unit root hypothesis is rejected and the electricity load for hour 2 can be modeled in levels. A similar conclusion is reached for the electricity demand in each of the other hours, while unit roots are also not present in any of the other variables that we consider. Table 1 Unit root test for electricity load in hour 2 Null Hypothesis: EL_02 has a unit root Exogenous: Constant, Linear Trend Lag Length: 28 (Automatic - based on SIC, maxlag=29) Augmented Dickey-Fuller test statistic Test critical values:

1% level 5% level 10% level

*MacKinnon (1996) one-sided p-values.

5

t-Statistic -4.75938 -3.96057 -3.41105 -3.12734

Prob.* 0.0005

Methodology A time-series econometric model is used to estimate the relationship between electricity demand and weather variables. The regressors in the model include temperature and humidity as the weather variables of interest, GDP and a trend term to explain long-run changes in electricity demand, seasonal variables and autoregressive terms. A number of econometric approaches can be used to model hourly loads (see [11] for several examples). The most common strategy is to express each of the hourly series (in our case, electricity load, temperature and humidity) as a single consecutive series that runs from hour 1 to hour 24 on each day followed by hours 1 to 24 on the next day, and so on. While this strategy requires estimating only a single equation for electricity demand, it suffers from a major disadvantage. By treating electricity demand as a single continuous series, it imposes the same behavioural model on electricity use in different hours of the day, whereas intuitively one would expect electricity use at different times of the day to be governed by quite different factors and dynamics. Adequately capturing these effects while modelling electricity load as a single continuous series poses considerable challenges. Instead, we use an hour by hour modelling strategy where the load for each hour of the day is treated as a separate variable and modelled with its own equation. Ramanathan et al. (1997) [12] demonstrated that this approach is superior in explaining and forecasting short-run loads compared to the alternative strategy of modelling the load as a continuous series. Although there is some loss of information with this approach, since we are only explaining day-to-day changes in the electricity load (for each hour) rather than hour-to-hour changes, this is not a significant issue, since for each hour we still have 10 years of daily data to work with. The major benefit of this approach is that it gives us the flexibility to test whether electricity demand indeed behaves differently across different hours and thus allows for more precise estimates of the impact of weather variables on electricity use. The 24 different equations for each hour together form a system of seemingly unrelated regressions or a SUR system. Such a system can be consistently estimated using Ordinary Least Squares (OLS); in fact, as is well known, estimating each equation separately using OLS is equivalent to estimating the entire system using OLS [13]. Although there are potential efficiency gains from systems estimation (for instance by using Feasible Generalized Least Squares (FGLS)), these are not likely to be large, especially since the regressors that we use are quite similar across equations. The general model for each hour h can be described as: ∑

Log



∑ (1)

is the Time subscripts are denoted using h and t, where h gives the hour and t the day of the observation. total electricity demand, while represents j different weather variables (including temperature and humidity). captures the effect of a change in the quarterly GDP on electricity consumption, while the coefficient on the trend term accounts for any remaining deterministic trend in the electricity consumption series. Dummy variables are used to capture the demand effects of the day of the week and the occurrence of public holidays; we include separate dummies for public holidays that fall on weekdays and the eve of public holidays. The lagged polynomial operator, , captures lagged effects in hourly electricity demand. Seven autoregressive lags are added to the model to remove serial correlation between the residuals. This is because current weather and electricity load could be similar to that of the previous day(s), and electricity load could be similar to that of the previous week. is a white noise random disturbance term. Note that the coefficients are indexed with h even when the independent variable is the same across hours (e.g. with GDP or with public holidays); thus if the effects of such variables are indeed different across the hours, the difference can be captured in the model. Meteorological parameters are of particular interest in this study. consists of current and lagged humidity and temperature variables. The weather component of equation (1) can be written down in an equation (for each hour h) as follows: ∑ (2) is the temperature for the current hour h, while is the temperature i hours before. is the relative humidity for the current hour h, while is the relative humidity i hours Similarly, before. We include temperature and humidity variables for the past three hours because (as will be illustrated later) they typically add explanatory power to the models: one reason could be that it takes time for buildings to absorb 6

and dissipate heat, and thus energy consumption could respond to changes in outside temperature with some lag. We also include interaction terms of the current temperature with the monsoon seasons in Singapore. The dummy variables and represent the Inter-monsoon (March and November) and Southwest monsoon (April to October) seasons respectively; both of these are zero in the case of the Northeast monsoon season is the additional impact of temperature on electricity demand during the Inter(December to February). Thus . Finally, in monsoon period relative to the North-east Monsoon season, and a similar interpretation holds for  order to capture possible long run effects of temperature on electricity demand, the 90-day moving average of past in equation (2). The purpose is to temperature is also included as one of the temperature variables, _ smoothen out day-to-day fluctuations in the hourly temperature while the coefficient measures the impact of a long-term temperature change on the electricity load in addition to the short-run effects captured by the coefficients to . Using the estimated coefficients from the hour-by-hour models, we can compute elasticities of electricity demand with respect to the variables of interest. Since GDP is included in equation (1) in logarithmic form, the GDP elasticity is simply the coefficient .1 The temperature and humidity elasticities require some further calculation: since they are not expressed in logarithms, the coefficients have to be multiplied with the average value of the variable to obtain the elasticities. We can calculate the humidity elasticity for each hour h by the following equation:    ∑

_

_

(3)

Similarly, the short-run temperature elasticity for hour h in the North-east monsoon season (when the dummies and  are zero) can be described by the following equation: _

_

_

   ∑

_

(4)

The short-run temperature elasticities for the other seasons can be easily calculated by adding   and   to ∑ in equation (4) and replacing the average temperature of the Northeast monsoon season with the average temperature from the appropriate season. Finally, we can calculate the “long-run” elasticity of electricity demand with respect to the moving average of temperature, as in equation (5) below. _

_

_

_

   

(5)

Note that this is not the same as the true long-run temperature elasticity of demand, because the latter would also encompass all the short-run effects captured in equation (2), as well as additional effects resulting from the presence of the autoregressive terms included in equation (1). Due to the complexity of the model, the true long-run temperature elasticity is difficult to compute, and so we calculate the elasticity with respect to the moving average temperature instead. This elasticity value measures how electricity demand responds to a long-run change in the temperature after controlling for short-run fluctuations in temperature, and thus captures only long-run adjustments to changes in temperature. Thus, both short-run and long-run effects can be estimated using the proposed model. Short-run effects describe the relationship between electricity load and quickly changing variables like temperature and humidity. Long-run effect are captured by slowly changing variables like GDP and the 90 day moving average of the past temperature values.

Results Using the modelling strategy described above, we estimated separate regression equations for all 24 hours of the day. Table 2 shows model statistics for one such equation, that for Hour 2. The high R2 value is expected given the high autocorrelation in the load series. In the case of temperature and humidity, due to multicollinearity between the current variable and its lags, the coefficients for some of the variables for certain hours turn out to be insignificant. However, joint F-tests of the current temperature and its lags were conducted, and they confirmed a positive and significant relationship between temperature and electricity load. A similar process was conducted for the humidity variables as well. The seasonal interaction terms are not statistically significant for hour 2, but are statistically significant for a number of other hours. 1 Note that is the percentage change in electricity demand due to a percentage change in GDP, which is equivalent to the GDP elasticity of electricity demand.

7

Table 2 Regression results and statistics for Hour 2 Model Dependent Variable: @LOG(EL_02) Method: Least Squares Date: 07/08/13 Time: 14:53 Sample (adjusted): 4/07/2003 12/31/2012 Included observations: 3557 after adjustments Convergence achieved after 16 iterations Variable C @TREND LOG(GDP) PUBLIC_HOLIDAY PUBLIC_HOLIDAY _ON_WEEKDAYS EVE_OF_PUBLIC_HOLIDAY SATURDAY SUNDAY MONDAY WEDNESDAY THURSDAY FRIDAY HUMIDITY_HR_02 HUMIDITY_HR_01 HUMIDITY_HR_24(-1) HUMIDITY_HR_23(-1) TEMP_HR_02 TEMP_HR_01 TEMP_HR_24(-1) TEMP_HR_23(-1) TEMP_MOVING_AVG INTER_MONSOON*TEMP_HR_02 SW_MONSOON*TEMP_HR_02 AR(1) AR(2) AR(3) AR(4) AR(5) AR(6) AR(7) R-squared Adjusted R-squared S.E. of regression Sum squared resid Log likelihood F-statistic Prob(F-statistic) Inverted AR Roots

Coefficient 4.0663 0.0000 0.2568 0.0004 0.0049 0.0069 0.0109 -0.0141 -0.0334 0.0028 0.0037 0.0043 0.0003 0.0007 -0.0005 0.0006 0.0039 0.0088 -0.0086 0.0135 0.0275 0.0000 0.0002 0.7223 0.1357 -0.0055 -0.0042 -0.0007 0.0330 0.0401

0.9765 0.9763 0.0154 0.8361 9813.4 5053.4 0.0000

Std. Error 0.6985 0.0000 0.0639 0.0023 0.0025 0.0014 0.0010 0.0009 0.0007 0.0007 0.0009 0.0009 0.0002 0.0004 0.0004 0.0002 0.0015 0.0028 0.0028 0.0014 0.0048 0.0001 0.0002 0.0170 0.0208 0.0209 0.0209 0.0209 0.0208 0.0169

Mean dependent var S.D. dependent var Akaike info criterion Schwarz criterion Hannan-Quinn criter. Durbin-Watson stat

.95 -.12+.56i

.50+.44i -.49+.22i 8

.50-.44i -.49-.22i

t-Statistic 5.8219 4.0415 4.0182 0.1542 1.9560 5.0655 11.4532 -16.0749 -46.4474 3.8733 4.1960 4.5434 1.1378 1.8421 -1.2852 3.0175 2.5276 3.1701 -3.1064 9.3085 5.6876 0.3607 0.9172 42.5672 6.5117 -0.2640 -0.2019 -0.0357 1.5877 2.3705

Prob. 0.0000 0.0001 0.0001 0.8775 0.0505 0.0000 0.0000 0.0000 0.0000 0.0001 0.0000 0.0000 0.2553 0.0655 0.1988 0.0026 0.0115 0.0015 0.0019 0.0000 0.0000 0.7183 0.3591 0.0000 0.0000 0.7918 0.8400 0.9715 0.1124 0.0178 8.2654 0.1000 -5.5009 -5.4488 -5.4823 1.9954

-.12-.56i

Using the estimated models, Figure 6 shows the estimated and actual results of the proposed model for hour 2. The model is able to capture the time series dynamics and the residuals appear to be white noise. Figure 7 shows the ACF and PACF of the residuals for the hour 2 model for 20 lags, and illustrates that the residuals do not display significant autocorrelation. Figure 6 Actual values, fitted values and residuals for the hour 2 model

Figure 7 Autocorrelation function (ACF) & partial autocorrelation function (PACF) of residuals for hour 2 model

9

Table 3 Coefficients of the models for 2 AM, 8 AM, 2 PM and 8 PM Deterministic components Constant Trend Line Log of GDP Saturday Sunday Monday Wednesday Thursday Friday Public holidays Public holidays that fall on weekday Eve of public holidays

2 A.M. 4.0663 0.0000 0.2568 0.0109 -0.0141 -0.0334 0.0028 0.0037 0.0043 0.0004 0.0049 0.0069

8 A.M. 3.1604 0.0000 0.4166 -0.0891 -0.1758 -0.0097 0.0011 -0.0012 -0.0015 -0.0256 -0.1242 0.0009

2 P.M. 3.4905 0.0000 0.4123 -0.1428 -0.2350 -0.0019 -0.0015 -0.0043 -0.0048 -0.0473 -0.1698 -0.0121

8 P.M. 3.3946 0.0000 0.4010 -0.1074 -0.1295 -0.0047 -0.0006 -0.0029 -0.0132 -0.0231 -0.0905 -0.0152

Meteorological components Temperature, current hour Temperature, 1 hour ago Temperature, 2 hours ago Temperature, 3 hours ago 3-month moving average of temperature at current hour Relative humidity, current hour Relative humidity, 1 hour ago Relative humidity, 2 hours ago Relative humidity, 3 hours ago Current Temperature × Inter-monsoon2 Current Temperature × Southwest Monsoon2

0.0039 0.0088 -0.0086 0.0135 0.0275

0.0021 0.0035 -0.0016 0.0067 0.0092

0.0002 0.0054 0.0004 0.0041 0.0050

-0.0002 0.0046 -0.0022 0.0075 0.0105

0.0003 0.0007 -0.0005 0.0006 0.0001 0.0002

0.0004 0.0001 0.000 0.0004 0.0004 0.0008

0.0004 0.0003 0.0002 0.0002 0.0005 0.0009

-0.0001 0.0009 -0.0002 0.0006 0.0003 0.0006

Autoregressive parameters Lag 1 Lag 2 Lag 3 Lag 4 Lag 5 Lag 6 Lag 7

0.7223 0.1357 -0.0055 -0.0042 -0.0008 0.0330 0.0401

0.4301 0.2196 0.0546 0.0404 0.0457 -0.0262 0.0311

0.3565 0.2126 0.0696 0.0213 0.0046 0.0427 -0.0451

0.5337 0.1810 0.0688 0.0092 0.0055 0.0449 -0.0415

Table 3 above gives the estimated regression results for four representative hours (2 A.M., 8 A.M., 2 P.M. and 8 P.M.) using 2003 to 2012 data. Electricity consumption varies over the days of a week. The model specifies dummies for each day of the week except Tuesday which is a normal working weekday. Compared to Tuesday, electricity consumption during the weekend is markedly lower because most companies in Singapore adopt a fiveday work week, operating from Monday to Friday. For example, the electricity consumption on Sunday at 2 P.M. is 23.5% less than that on Tuesday at 2 P.M. In addition, public holidays have an effect on electricity consumption. Generally less electricity is consumed during public holidays than otherwise, as shown by the negative coefficients of the Public Holidays dummy at 8 A.M, 2 P.M. and 8 P.M. The effect is more pronounced when a public holiday falls on weekday. For example at 2 P.M., 17% less electricity is consumed on a public holiday that falls on a weekday, as compared to one that does not. The amount of economic activity, measured by the GDP, will influence electricity consumption. The coefficient of the log of GDP is positive across all four hours, an expected result since more economic activity will result in more electricity being consumed. The coefficient at 2 A.M. is smaller than the coefficients of the other three hours, indicating that economic activities occurring at 2 A.M. are less energy intensive than those that take place during waking hours.

2 The Inter-Monsoon season corresponds to the months of March and November, while the South-West Monsoon season corresponds to the months from April to October.

10

Given the widespread use of air-conditioning in Singapore to cool building interiors, coefficients of the temperature variables are expected to be positive. However, due to multicollinearity between the current temperature and its lags, the coefficients for some of the variables for certain hours turn out to be insignificant. In each case, though, the sum of the coefficients on the temperature variables is positive, indicating that an increase in the current and lagged temperatures has a positive net impact on electricity demand. Instead of using month dummies to account for weather variation over the year, the study amalgamated months corresponding to the different seasons in Singapore to form the monsoon dummies, which then forms an interaction term with the current temperature. The results show that for the same increase in temperature, the electricity load during the Inter-monsoon and Southwest Monsoon seasons increase more than the Northeast Monsoon season. This is probably due to the greater demand for cooling as the weather is hotter during these seasons. Finally, a 1°C rise in the 90-day moving average temperature would raise electricity consumption by between 0.5% and 2.75%, confirming the existence of long-run responses to sustained changes in temperature. In summary, both the short-term and long-term temperature variables are found to have a positive and significant impact on electricity load. This is a major finding, especially since the temperature in tropical Singapore fluctuates minimally compared to temperate regions where similar studies have been conducted.3 Similarly, relative humidity is expected to have a positive relationship with electricity consumption. When relative humidity rises, air conditioners work harder to dehumidify the building interior. Although some of the relative humidity variables are not significant, joint F-tests of the variables confirm a positive and significant relationship between relative humidity and electricity load. Elasticities Elasticities of electricity demand in Singapore with respect to temperature and other key variables are of particular interest in this study. Figure 8 shows the short-run temperature elasticities for different hours of the day for the Southwest Monsoon period from April to October, as well as the average electricity load for each hour. The temperature elasticities are in the 0.3-0.5 range, suggesting that a 1% increase in temperature would lead (in the short-run) to an increase in electricity load by between 0.3-0.5%. An interesting observation is that temperature elasticities tend to be higher at night-time than in the day-time, even though the temperature itself is lower. The most probable explanation for this is that the residential load forms a significantly higher proportion of the overall load during the night than in the day, and one expects residential energy consumption to be more responsive to fluctuations in temperature than industrial energy use. A contributing factor is the common practice in Singapore of leaving the air-conditioning on during the night. These effects appear to be more than sufficient to counter the fact that consumers have a greater ability to adjust their behaviour in the daytime than in the night when they are asleep.4

3

The daily range of temperature for Singapore is from a minimum of 23°C to 26°C to a maximum 31°C to 34°C. This effect results in higher temperature elasticities in the US in the daytime than in the night, which is the opposite of what we find for Singapore: see Crowley and Joutz [3]. 4

11

0.6

6000

0.5

5000

0.4

4000

0.3

3000

0.2

2000

0.1

1000

Load (MWh)

Elasticity

Figure 8 Short-run temperature elasticities and average load (Southwest Monsoon)

0

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Hour of the Day Temperature Elasticity

Load

There are also differences in short-run temperature elasticities across the seasons, as Figure 9 shows. The temperature elasticity is highest in the South-West Monsoon season from April to October (when the temperature is highest), somewhat later during the Inter-monsoon months of March and November, and lowest in the North-East Monsoon season from December to February (when the temperature is lowest). These differences in elasticities, although statistically significant for most hours of the day, are relatively small, which is unsurprising given that the differences in average temperatures across the seasons are quite small in the first place (see Figure 3). Figure 9 Short-run temperature elasticities in different seasons 0.6 0.5

Elasticitiy

0.4

South‐west Monsoon  Temperature Elasticity

0.3 0.2

Northeast Monsoon  Temperature Elasticity

0.1

Inter‐monsoon  Temperature Elasticity

0 1

3

5

7

9

11 13 15 17 19 21 23

Hour of the Day

The pattern of elasticities rising at night-time is seen even more dramatically for “long-run” temperature elasticities, as shown in Figure 10 below.5 Elasticities from 10 pm to 6 am range from 0.5 to 0.8, while elasticities from 7 am to 9 pm are less than 0.4 and below 0.2 for large parts of the day. The fact that all of these elasticities are positive suggests that long-run responsiveness of electricity demand to temperature changes does exist and is distinct from short-run responses. Such responses might involve either habit changes that take time to develop (e.g. what setting to use on the air-conditioner every night) or purchases of appliances (e.g. purchasing an extra fan), though the latter is less likely given that our elasticity is based on a 90-day moving average and not a longer time period.

5 As discussed earlier, this elasticity only captures long-run adjustments to temperature and should not be confused with the standard definition of long-run elasticity, which encompasses both short-run and long-run effects.

12

The larger values of the night-time elasticities are (as with the short-term elasticities) likely a reflection of the temperature-sensitive residential load accounting for a greater proportion of the overall load at night. Figure 10 Long-run temperature elasticities and average load 0.9

6000

0.8

Elasticity

0.6

4000

0.5

3000

0.4 0.3

2000

0.2

Load (MWh)

5000

0.7

1000

0.1 0

0 1 2 3 4 5 6 7 8 9 101112131415161718192021222324 Hour of the Day Long‐Run Temperature Elasticity

Load

Figure 11 below presents humidity elasticities for different hours of the day. Humidity elasticities tend to be smaller in size than temperature elasticities and are typically around 0.09-0.1, falling to less than 0.08 between 10 AM and 3 PM. To a lesser extent they too exhibit the now-familiar pattern of lower daytime elasticities, which can again be explained by the residential load becoming “less important” in the daytime.

0.12

6000

0.1

5000

0.08

4000

0.06

3000

0.04

2000

0.02

1000

0

Load (MWh)

Elasticity

Figure 11 Humidity elasticities and average load

0 1 2 3 4 5 6 7 8 9 101112131415161718192021222324 Hour of the Day Humidity Elasticity

Load

Finally, although not the focus of our study, GDP elasticities are of considerable interest and can be calculated from our models as well. The GDP elasticities shown in Figure 12 range between 0.2 in the night-time to 0.4 in the daytime; the fact that they are less than one is consistent with the fact that energy intensity in Singapore has declined between 2003 and 2012 [14]. In contrast to the weather elasticities, the GDP elasticity is higher during the day than during the night, and the explanation is precisely the opposite: the industrial and commercial load, which we expect to be particularly responsive to economic activity, form a greater proportion of the daytime load than the night-time load, whereas the residential load (which is influenced less by economic activity) forms a greater proportion of the night-time load. 13

Figure 12 GDP elasticities and average load 0.45

6000

0.40

Elasticity

0.30

4000

0.25

3000

0.20

2000

0.15 0.10

Load (MWh)

5000

0.35

1000

0.05

0

0.00 1 2 3 4 5 6 7 8 9 101112131415161718192021222324 Hour of the Day GDP Elasticity

Load

Conclusion An econometric model is used to estimate the impact of climate change on electricity demand in Singapore. Twenty-four separate models are constructed to capture the time-series dynamics of electricity load at every hour of the day. The model captures the effect of GDP, temperature, humidity and lagged demand variables on the hourly electricity demand of Singapore. Both short run and long run elasticities of electricity demand with respect to key variables are calculated and it is found that the temperature elasticity of demand is higher during night-time as compared to day-time. This can be explained by the higher proportion of temperature dependent load during nighttime. Similar behaviour is observed for humidity elasticities. GDP elasticities are higher during day time which is consistent with the fact the economic activity is higher during the day time as compared to night time. Our analysis suggests that it will be important for policymakers in Singapore to take the possibility of climate change into account when planning future investments in electricity generation capacity. Both the short-run and long-run temperature elasticities found in this study are significant in magnitude, implying that a sustained future increase in Singapore’s temperature will have a sizeable upward impact on electricity demand. However, any such increase in demand (in percentage terms) will be greatest during the night when the overall demand is lower, which should to a certain extent lessen the future investments in generation capacity required to cope with climate change. For a more precise understanding of these and other policy implications, this study can be used in conjunction with the different IPCC future projected climate change scenarios [15] to simulate the effect of climate change on Singapore electricity demand. The analysis will help in framing the guidelines for policymakers in Singapore to maintain a sustainable energy economy under different climate change scenarios. The analysis can also form the basis for a number of other research topics. For example, demand response and energy efficiency measures in the electricity sector are becoming important policy tools. New models can be formulated to include the effect of temperature on the energy efficiency technologies, and it is important for such models to include robust temperature elasticities.

References [1] Singapore Energy Statistics 2012, Energy Market Authority, Singapore [2] Töglhofer, C., Habsburg-Lothringen, C., Prettenthaler, F., Rogler, N. and Themessl, M. (2012). EL.ADAPT: Impacts of Climate Change on Electricity Demand. Symposium Energieinnovation, Austria, February 15-17. [3] Crowley, C. and Joutz, F.L. (2003). Hourly Electricity Load: Temperature Elasticity and Climate Change. Proceedings of the 23rd North America Conference of the USAEE/IAEE, Mexico City, October 19-21. [4] Lam, J.C., et al. (2004). Electricity use characteristics of purpose built office buildings in subtropical climates. Energy Conversion and Management 45: 829-844. [5] Guan, L. (2011). Energy use, indoor temperature and possible adaptation strategies for air-conditioned office buildings in face of global warming. Building and Environment: 1-12 14

[6] Frank, Th. (2005). Climate change impacts on building heating and cooling energy demand in Switzerland. Energy and Buildings, 37: 1175-1185. [7] Energy Market Company, Singapore (2013). Market Data. https://www.emcsg.com/marketdata/priceinformation. [8] National Environment Agency, Singapore (2013). [9] Web CEIC Data Manager (2013). Gross Value Added at Basic Prices (Constant 2005 S$ Prices) from the Ministry of Trade and Industry, last updated 22 May 2013. [10] Gujarati, Damodar N. (2004). Basic Econometrics. 4th ed. McGraw-Hill Book Co. [11] Bunn, D.W. and E.D. Farmer (1985). Comparative Models for Electrical Load Forecasting, Wiley, New York. [12] Ramanathan, R., R. Engle, C.W.J. Granger, F. Varid-Araghi and C. Bruce (1997). Short-run Forecasts of Electricity Loads and Peaks. International Journal of Forecasting 13: 161 – 174. [13] Greene, William H. (2005). Econometric Analysis. 5th ed. Prentice-Hall International. [14] Ministry of the Environment and Water Resources and Ministry of National Development. A Lively & Liveable Singapore: Strategies for Sustainable Growth. http://app.mewr.gov.sg/data/ImgCont/1292/sustainbleblueprint_forweb.pdf [15] IPCC, Climate Change 2001: Synthesis Report, Third Assessment Report, Intergovernmental Panel on Climate Change, Cambridge, 2002.

15