Model Selection and Size of Impact

To choose the optimal set point, I compute the S2a test statistic described in section 2.3.4.

The final number of net wins (”wins” minus ”losses”) of each of the 24 competing models is shown in Table 3.8.

For residential electricity demand, the most accurate out of sample forecast is gen- erated by the model in column (16), which uses a set point of 65◦F to calculate both CDH and HDH, and dry bulb temperature to compute CDH. This is the set point that most researchers have considered when analyzing electricity demand in the U.S. For the commercial sector, the most accurate model is the generated using a different set point. The most accurate model uses a set point of 55◦F with a set back of 5◦F for both CDH

and HDH (column (8)). Moreover, this optimal model for commercial demand uses CDH computed with wet bulb temperature. Finally, the results for electricity consumption by the industrial sector are inconclusive. That is, no model performs statistically better than other.

For the residential sector, our results differ from those for Maryland (Ruth and Lin, 2006). They estimate models of residential use that specify degree days variables computed with dry bulb temperature with a range of set points at 1◦F intervals, daylight hours, and energy prices. The authors select the model with the highest R2, which in the case of residential demand corresponds to the model generated with a set point of 60◦F. In the case of commercial demand, Ruth and Lin (2006) estimate a constant set point of 53◦F, which as in our results, corresponds to a threshold temperature different from the standard 65◦F. However, the authors do not allow night-time variation in the set points as I do. Kaufmann et al. (2012) estimate demand for natural gas and No. 2 fuel oil instead of electricity. However, their results are appealing since they also use data from Massachusetts. They find a lower set point than the standard 65◦F: 55◦F with a set back of 5◦F for both residential and commercial natural gas consumption and 60◦F with a set back of 5◦F for distillate oil demand.

To assess the impact of climate change on electricity demand, I compute degree days from GCM output and use it to compute changes in residential and commercial demand (section 2.3.5). Ideally I would use degree hours rather than degree days. However, the lack of hourly temperature projections, prevent us from calculating CDH and HDH. For the same reason forecasts cannot incorporate set backs. In addition, I use dry bulb temperature as opposed to wet bulb temperature because forecasts for the latter variable

are not available.

Columns (1) and (2) in Table 3.9 present cooling degree hours computed with monthly mean temperature data generated by GCMs for the period 1990–2010.12 _{I present CDD}

computed with the set points selected for the optimal residential and commercial models (65◦F for residential demand and 55◦F for commercial demand). Columns (5) and (6) are the analogous for heating degree days. Figures shown in columns (3) and (4) are the estimated cooling degree days computed with an increase in GMT of 2◦C. Columns (7) and (8) are the corresponding figures for heating degree days. I use temperature data generated by GCM for the period 1990–2010 and also temperature data generated by GCM for an increase in 2◦C in GMT. This allows us to generate consistent estimates for the change in degree days variables.13 _{I compare CDD and HDD computed for the}

period 1990–2010 with those computed for an increase in GMT of 2◦C, and I observe that cooling needs increase while demand for heating services decreases. Based on these changes, I estimate that residential and commercial electricity demand increase 2.6% and 4%, respectively. These changes occur during specific seasons (columns (10) and (12)). During summer months (from Jun to August) increase demand for cooling increases elec- tricity demand 15% to 30% for the residential sector, and 10% to 13% for the commercial sector. During winter months (from September to May) a small change in heating degree days has a moderate the impact on demand.

These results are similar to Ruth and Lin (2006). Using predictions from the HadCM2 model for the state of Maryland, they find that climate change has a small impact on resi-

12_{Section 2.4 describes in detail the temperature projections used in the present study.}

13

Notice that I consider data obtained from the Weather Service of Amesbury in Massachusetts through the estimated coefficients for the optimal residential and commercial electricity demand models.

dential demand and a larger impact on commercial demand. Similarly, Amato et al. (2005) use projections from the Canadian Climate Centre Model and forecast that per capita res- idential electricity demand in Massachusetts will increase 2.1% by 2020, presenting a large intra-annual variability. In summer they forecast that in electricity consumption will rise 6.8% and decline 2.7% in winter. As Amato et al., I also find that the increase in demand for cooling in summer months is greater than the decrease in demand for heating in winter months.

Our estimation results are based on a limited set of threshold temperatures with 5◦F intervals (J = { 50◦F, 55◦F, 60◦F, 65◦F, 70◦F, 75◦F }). There is no a priori reason to restrict the analysis only to this set. Similarly, the selection of a set back of 5◦F and its timing (during night-time, between the hours of 11PM and 4AM) are arbitrary. Further investigation will investigate whether the results of the present study change if different set points and set backs are used to estimate CDH and HDH. The estimates of this paper should therefore be considered as a first approximation of the impact of climate change on electricity demand in Massachusetts.