Shift in seasonal climate patterns likely to impact residential energy consumption in the United States

Observed gridded daily maximum temperature (Tmax) and daily minimum temperature (Tmin) are obtained from the Parameter-Elevation Regressions on Independent Slopes Model (PRISM) (Daly et al 2008). Simulated Tmax and Tmin are obtained by bias correcting dynamically downscaled 11-member ensemble of GCMs (see table S1, available online at stacks.iop.org/ERL/14/074006/mmedia) from Coupled Model Inter-comparison Project Phase 5 (CMIP5) (Ashfaq et al 2016) following the approach detailed in Ashfaq et al (2010), (2013). State level natural gas and electricity consumption (EIA US (2016a), (2016b)) and residential energy consumption survey (RECS) data for the 10 US census divisions (figure S1) are obtained from US Energy Information Administration (EIA). Population statistics are obtained from US Census Bureau (Census US 2016). Further description of these datasets is detailed in the supplementary information.

The multi-step methodology is detailed in the following subsections and a schematic is provided in figure S2.

2.2.1. Heating and cooling degree days

2.2.2. Econometric model

Equations (1) and (2) represent the econometric models for residential electricity and natural gas consumption. Each econometric model is based on state level observed EIA energy (electricity or natural gas) consumption, state level degree-days from aggregated PRISM meteorological observations and state level population from Census for 1990–2005

$$\begin{eqnarray}&&\begin{array}{l}\mathrm{log}\left({E}_{sch}^{smy}\right)={c}_{e}+{\alpha }_{e0}\left({\rm{H}}{\rm{D}}{{\rm{D}}}_{elec}^{smy}\right)+{\alpha }_{e1}{\left({\rm{H}}{\rm{D}}{{\rm{D}}}_{elec}^{smy}\right)}^{2}\\ \,\,\,\,\,+\,{\beta }_{e0}\left({\rm{C}}{\rm{D}}{{\rm{D}}}^{smy}\right)+{\beta }_{e1}{\left({\rm{C}}{\rm{D}}{{\rm{D}}}^{smy}\right)}^{2}\\ \,\,\,\,\,+\,{\gamma }_{e0}\left({P}_{sy}\right)+{f}_{s}+{f}_{m}+{\varepsilon }_{o},\end{array}\end{eqnarray} \tag{ 1 }$$

$$\begin{eqnarray}&&\begin{array}{l}\mathrm{log}\left({\rm{N}}{{\rm{G}}}_{sh}^{smy}\right)={c}_{ng}+{\alpha }_{ng0}\left({\rm{H}}{\rm{D}}{{\rm{D}}}_{ng}^{smy}\right)\\ \,\,\,\,\,\,\,\,+\,{\alpha }_{ng1}{\left({\rm{H}}{\rm{D}}{{\rm{D}}}_{ng}^{smy}\right)}^{2}+{\gamma }_{ng0}\left({P}_{sy}\right)\\ \,\,\,\,\,\,\,+\,{f}_{s}+{f}_{m}+{\varepsilon }_{1}.\end{array}\end{eqnarray} \tag{ 2 }$$

RECS-based percentage usage is applied to total residential electricity and natural gas consumption to obtain the relative shares of electricity $\left({E}_{sch}^{cmy}\right),$ used for space heating and cooling, and the natural gas $\left({\rm{N}}{{\rm{G}}}_{sh}^{cmy}\right),$ used for space heating for state s, month m and year y.

Similarly, ${\rm{H}}{\rm{D}}{{\rm{D}}}_{elec}^{smy},$ ${\rm{H}}{\rm{D}}{{\rm{D}}}_{ng}^{smy}$ and ${\rm{C}}{\rm{D}}{{\rm{D}}}^{smy}$ are HDD fulfilled by electricity, HDD fulfilled by natural gas and CDD for respective state, month and year. ${c}_{e}$ and ${c}_{ng}$ are constant terms, ${\alpha }_{e0},$ ${\alpha }_{e1},$ ${\beta }_{e0},$ ${\beta }_{e1},$ ${\gamma }_{e0},$ ${\alpha }_{ng0},$ ${\alpha }_{ng1},$ ${\gamma }_{ng0}$ are coefficients of respective terms as shown in the equations and ${\varepsilon }_{o},$ ${\varepsilon }_{1}$ are error terms. Additionally, we use ${f}_{s}$ and ${f}_{m}$ as fixed effects for states and months respectively. Fixed effects are used in a regression model to control for some types of omitted variable bias. The state fixed effects account for the average difference in time-invariant state characteristics. The month of year fixed effects account for common shocks to all states in a month of the year and can capture seasonal patterns unobserved in the data. In this case, inclusion of state and month fixed effects allows control for differences in the magnitude of degree days that arise (i) among states because of differences in population size and (ii) among months because of seasonal variations. It should be noted that while the month fixed effects account for intra-annual differences among the states, the differences between years arising from factors such as macroeconomic fluctuations are not incorporated. Moreover, the electricity model uses all months while summer months (June, July and August) are excluded for the natural gas model. Overall, the econometric models are able to precisely estimate the determinants of electricity and natural gas demand. The R-square values for the models are 0.99 and 0.98 respectively and there are significant t-statistics for all the coefficients at 95% significance level. The F-statistics suggest that the results are jointly as well as individually statistically significant (table S2 and S3). We also perform additional tests to rule out any temporal or spatial correlation in our datasets. First, we perform regressions using Newey-West standard errors with 45 and 34 lags for electric and natural gas models respectively. Second, we estimate these regression using (Driscoll and Kraay 1998) standard errors. These standard errors are robust to very general forms of temporal and spatial correlation with all coefficients being statistically significant at the 1% level.

The fixed effect regression models established in equations (1) and (2) are subsequently applied to each of the RegCM4 ensemble members by replacing the PRISM degree days with the simulated degree days in the 25 years in the historical period (1981–2005) and the 40 years period in the future (2011–2050). For the future models, the population is kept at the 2005 level to isolate the variations in energy demand that arise solely because of climatic changes.

2.2.3. Energy data disaggregation

Across the US, residential space heating requirements are generally higher than the space cooling requirements given the higher number of HDD than CDD (figures 1(a)–(d)). Cooler temperatures associated with continental air and higher elevations drive maximum space heating requirements in the north central US (up to >3400 °C) whereas space cooling requirements peak mostly over southern US such as Florida, Texas and parts of the southwest (up to >2000 °C). The spatial variability in the degree days, along with their magnitudes are simulated exceptionally well in the RegCM4 simulations compared to the observations (figures 1(a)–(d)).

Zoom In Zoom Out Reset image size

Observations exhibit a decreasing (increasing) trend in HDD (CDD) across the US, which is also captured in the simulations (figures 1(e)–(h), S3–4). However, the simulated decrease in HDD (up to 350 °C) is substantially milder than observed (up to 550 °C) over the Rockies, and the strongest observed increase in CDD over the southwest is approximately 1.5 times higher than the simulated trends. The trends in HDD are significant in both the observations and the simulations in most of the western half and parts of the eastern half of the US. However, the increasing trend in CDD is only significant over parts of the southwest and Pacific Northwest in the observations in contrast to the simulations that also exhibits significant trends over the southeast and parts of the northeast. HDD and CDD trends in ten metropolitan regions exhibit similar characteristics in both the observations and the simulations. For instance, both datasets exhibit a decreasing trend in HDD over all ten regions, which is also statistically significant over three regions. Similarly, both datasets exhibit increasing but insignificant trends in CDD over 9 out of 10 regions (figures 1(i), S3–4). Collectively, these comparisons demonstrate that the downscaled data exhibits good skill in the simulation of the mean HDD and CDD, and first-order skill in the simulation of their historic trends. This is important because historic trends in HDD and CDD provide a precursor for future changes in energy demands in warmer climates.

Driven by the skillfulness of downscaled data in capturing the characteristics of HDD and CDD, simulations-based electric and natural gas demands also compare well with the EIA observations, particularly at low and medium demand levels (<4000 GWh for electricity and <60 000 MMcf for natural gas) (figures S5–S7). The simulated electric demand also exhibits skill at the metropolitan level where statistically significant trends are simulated across all ten metropolitan regions in both the RegCM4 simulations and the observations (figure 1(i), S5) . The trend in simulated natural gas demand compares well with the observations for 7 out of 10 regions (figure 1(i), S6). Both electricity and natural gas demand exhibit an upward trend, due to the rising population over the historical period.

To understand the impacts of climate variations on RED, we fix all other factors including population and other economic drivers that may influence energy system response. Such an approach is a standard in the future climate studies. For instance, all RCPs driven GCMs simulations only simulate climate system response to changes in representative concentration pathways. With population kept at 2005 levels, electricity demand is projected to increase across the US with the exception of some parts of western US, which exhibit a decrease of up to 7% in parts of Arizona, Nevada, and California by 2050 (figure 2(a)). The magnitude of increase across the rest of the US is up to 10%. For metropolitan regions considered in this study, an increase in electricity demand is projected for eight out of ten, ranging from approximately 0.5% for Boston-Cambridge Newton, MA to more than 7% for Miami-Fort Lauderdale-West Palm Beach, FL (figure 2(c)). In the case of natural gas, the demand is projected to decrease over most of the US (figure 2(b)) with the exception of parts of some states such as Texas, Arizona, and Florida that exhibit a strong increase. However, these strong percent increases over parts of these states primarily driven by their relatively small natural gas demand during the historical period. Given that natural gas is mainly used for space heating, the decrease in natural gas demand is driven by the changes during the winter months (i.e. September–May). RegCM4 ensemble members show a robust decrease in natural gas demand (up to 4%) across all ten metropolitan centers, which is mainly driven by a decrease in HDD (figures 2(e)–(f)). The decrease is larger during the transition months i.e. April and May in spring, and September and October in fall compared to winter months.

Zoom In Zoom Out Reset image size

We investigate the factors driving future changes in RED by defining changes in the summer-like (hereafter summer) and the winter-like (hereafter winter) conditions. Summer (winter) conditions are defined as the longest consecutive period when CDD (HDD) is greater than HDD (CDD). Figure 3 shows a comparison of accumulated degree days and degree days per day (CDD for summer and HDD for winter) in the PRISM observations and the RegCM4 simulations for summer and winter respectively. Accumulated CDD (HDD) range from 0 to >1400 (0 to >4000) °C whereas CDD (HDD) per day range from 0 to 10 (0 to 18) °C during summer (winter). The total and per day degree days are based on the timing and duration of summer and winter. Summer starts as early as April in the south and as late as June/July in the north, with a gradual decrease in the duration (>200 to <40 d) and the cooling requirements (>1400 to <200 °C). Winter starts as early as September in the north and as late as December/January in the south, with a gradual decrease in duration (>280 to <40 d) and heating requirements (>4000 to <400 °C) from north to south. This seasonal duration shift intensifies with the increase (decrease) in latitude and/or elevation for summer (winter) (figure S8)

Zoom In Zoom Out Reset image size

While electricity is used for both space heating and cooling, future changes in the characteristics of summer and winter suggest that the projected increase in the electricity demand is primarily driven by an increase in the cooling demand. A late onset of winter, particularly over the higher elevations, and an early arrival of summer shrink the length of winter conditions (figure S9) by a few days in the parts of southeast to as much as a month in the parts of the western US, reducing the HDD by 20 °C to as much as 400 °C respectively (figure 4(b)). On the other hand, there is an increase in the length of summer conditions (figure S9c) by a few degree days in the parts of Pacific Northwest and Rockies to as much as a month in most of the southwest, increasing the CDD by 20 °C to as much as 300 °C respectively (figure 4(a)). The per degree day change in heating (cooling) demand that ranges from −0.7 to 0 (0 to 1) °C is mainly driven by changes in minimum (maximum) daily temperature during winter (summer) (figures 4(c)–(f)).

Zoom In Zoom Out Reset image size

While this study only focuses on the climate driven long-term environmental variations that can potentially influence RED, there are other socioeconomic factors that may reverse, mute or amplify the projected influence of climate-driven changes. We elaborate on such an impact of socioeconomic variations by considering projected changes in population distribution (figure 5(a)) in our economic model. We use 5 yearly population projections data from Intergovernmental Panel on Climate Change Special Report on Emissions Scenarios A2 scenario, which is analogous for RCP8.5 (EPA US (2010)), and linearly interpolate it at yearly time-scale to match the future period (2011–2050). Use of both climate and population changes in our economic model enables comparisons of future changes in energy demand caused by changes in climate (figures 2(a), (b)) with those caused by changes in both population and climate (figures 5(b), (c)). With the exception of Texas, California, and Florida where future population increase is projected to be statewide, urban areas are mainly expected to experience population growth. The projected increase in RED due to population increases (figures 5(b), (c)) in these three states and urban areas across the US will outpace the projected increase in RED due to climate (figures 2(a), (b)). On the other hand, most of the rural areas are projected to witness a decline in the population (figure 5(a)); however, climate-driven increase in RED in those rural areas will overwhelm the decrease caused by population changes (figures 2, 5). Compared to climate-driven changes, our results indicate that an increase (decrease) in urban (rural) population can potentially result in as much as 10 fold increase (as low as 5 fold decrease) in residential electricity demand. Similarly, compared to climate-driven changes, increase (decrease) in urban (rural) population can potentially result in as much as 5 fold increase (as low as 5 fold decrease) in natural gas demand.

Zoom In Zoom Out Reset image size

In addition to changes in the population, economic factors can potentially exert a significant influence on the pace of technological advancements and residential energy prices. Moreover, a growing economy boosts advancement in technology, resulting in more efficient heating and cooling systems, efficient building designs etc. Likewise, factors such as supply and demand variations and inflation in the energy sector can also drive changes in residential energy prices. However, due to the absence of reliable estimates for the future energy supply and demand, economic conditions and technological innovations, no other socioeconomic factors have been considered in this study.