a. Background

Prior to the start of natural gas sector deregulation in the mid-1980s, local distribution companies were the only source of natural gas for consumers, and the price was federally regulated because of concern about monopoly power. The inefficiency of the regulated natural gas market first was recognized in the early 1970s when increasing natural gas demand resulted in supply shortages and caused prices in interstate markets to rise to federal price limits (Energy Information Administration 1997). In more general terms, this federal constraint of natural gas prices resulted in prices that were not sufficiently market driven and often correlated poorly with natural gas demand (Mariner-Volpe 2000). During these regulated times, the value of seasonal climate forecasts for managing energy resources therefore was limited, as noted by Weiss (1982). However, had federal price limits not been in place, the impact on natural gas prices of three severe winters during the late 1970s would have been much more significant. This would have required that contemporary planning depend on knowledge of the relationship between temperature and energy consumption.

A fully deregulated natural gas market, based on principles of competition and choice, was established by the Natural Gas Decontrol Act of 1989. The deregulation gave gas buyers the opportunity to bypass local distribution companies and purchase gas directly at the wellhead at deregulated prices (Jess 1997). Without federal constraint of natural gas prices, the deregulated natural gas market became much more sensitive to temperature-related demand, with anomalously cold/warm periods often producing increases/decreases in natural gas price and enhanced overall price volatility (Fig. 1). This situation provided the basic motivation for this study, which quantifies relationships between temperature and residential natural gas consumption through the development of monthly and seasonal temperature-based indices for the United States east of the Rocky Mountains.

Additional developments resulting from natural gas deregulation that further prompted this investigation were the emergence of spot and futures natural gas markets in the late 1980s, because of their respective dependence on current and anticipated temperature-related demand. The futures trading market for natural gas was established to permit hedging against the price volatility experienced following deregulation—the market allows buyers to lock in at a price for an amount of natural gas that will be delivered in the future (Mariner-Volpe and Trapmann 2003). Natural gas trading companies, which emerged in the mid-1980s, acquire significant revenue through the futures market by advance buying and selling of large amounts of gas to serve future aggregated needs of many customers (Jess 1997).

Because spot and futures prices of natural gas are highly dependent on temperature-related demand, energy trading companies use short- and long-term temperature forecasts in their decision making. In accord with this fact, the usefulness and value of these forecasts to natural gas buyers utilizing the futures market would be enhanced through improved understanding of the relationships between temperature and natural gas consumption. The temperature-based indices of residential natural gas consumption developed in this study are such quantitative tools. Although the influence of temperature on gas consumption likely is modulated by several interrelated residential characteristics—home insulation levels, desired human comfort, and personal income—the data needed to incorporate these additional influences in temperature-based statistical models currently are not available.

b. Previous research and the current investigation

Previous research into associations between winter temperature and residential energy consumption primarily used heating degree-days (HDD) to define monthly and seasonal anomalous temperature. Most of this HDD-based research utilized an HDD reference temperature of +65°F (+18.3°C), because HDDs are intended to measure the deviation from a “comfortable” temperature. Those studies included Quayle and Diaz (1980), Warren and LeDuc (1981), Guttman (1983), Downton et al. (1988), Lehman and Warren (1994), and Heim et al. (2003).

The previous work began with Quayle and Diaz (1980) revealing that total electric usage and heating oil consumption at the household level in an area covering 6500 km² around Asheville, North Carolina, correlated strongly (correlation coefficient r ≥ +0.97) with National Weather Service (NWS) airport temperature data when transformed to HDDs. This study did not extend to residential natural gas consumption because heating oil and electricity are the predominant home heating sources in the Asheville area. Warren and LeDuc (1981) used HDDs as one of the predictors in a multiple regression model of regional natural gas consumption, but because additional predictors such as gas price also were incorporated, the correlation between gas consumption and HDDs alone was not reported. HDD quantification for regional and national levels was performed by Guttman (1983) and Downton et al. (1988), who determined the effect of population shifts on population-weighted HDDs. However, in both studies, consideration of HDD application to energy consumption analyses was limited to discussion of the potential value of such research. Later, Lehman and Warren (1994) developed a single regression model that used HDDs to predict monthly natural gas consumption (r = +0.98) for the immediate vicinity of Columbus, Ohio, but this analysis did not extend beyond the local level.

Motivated by the above high correlations obtained by Quayle and Diaz (1980) and Lehman and Warren (1994) for a single localized region, Heim et al. (2003) developed a residential energy demand temperature index (REDTI) that quantified the linkage between HDDs (based on a +65°F/+18.3°C reference temperature) and residential energy consumption on the national level and seasonal time scale. The winter REDTI correlated strongly (r = +0.75) with total U.S. residential energy consumption for 1980–2000 on the seasonal time scale and established the need for a more detailed regional analysis of the type presented below. Here, we expand on Heim et al. (2003) by developing a range of temperature-based indices with similarly high (or higher) correlations to natural gas consumption for winter monthly and seasonal time scales for nine U.S. regions east of the Rocky Mountains. Use of natural gas for home heating predominates in most of these regions during winter. Our method is to complement Heim et al. (2003) by using totals of days below certain temperature percentiles (DBP) to develop temperature-based gas consumption indices, in addition to again employing HDDs. Whereas HDDs emphasize cold anomaly magnitudes relative to a fixed reference temperature, DBP totals capture in particular the duration of cold anomalies relative to spatially varying temperature thresholds that reflect changing temperature frequency distributions across a region. This additional use of DBPs was encouraged by scientists at the National Oceanic and Atmospheric Administration/National Environmental Satellite, Data, and Information Service (NOAA/NESDIS) National Climatic Data Center [NCDC (in Asheville)] who had contributed to the Heim et al. (2003) study (J. H. Lawrimore 2004, personal communication).

Furthermore, unlike all previous studies, we assess Quayle and Diaz’s (1980) speculation that changes with time in energy conservation practices and human comfort levels would require consideration of HDD reference temperatures other than +65°F/+18.3°C. Our regional evaluations of both DBP and HDD indices therefore use a wide range of threshold and reference temperatures, respectively. These indices are quantified for 1949–2000. However, because residential natural gas consumption data are available on a regional basis only since 1989, development of temperature–gas consumption relationships and identification of associated key temperature thresholds are limited to 1989–2000. The lack of earlier regional gas consumption data anyhow is not a limitation here because, as noted above, market regulation masked the temperature consumption link for earlier years. Therefore, application of these relationships to infer gas consumption in earlier years must utilize (assume) the temperature consumption associations that prevailed following the completion of gas deregulation in the late 1980s.

The results of this study now are providing the basis for further research at NCDC to develop a quantifiable energy index that will give regional and national estimates of natural gas usage relative to monthly and seasonal averages (J. H. Lawrimore, NCDC, 2006, personal communication). This information will give decision makers statistics on consumer costs/savings that result from anomalous climate.

a. Richman–Lamb daily maximum and minimum temperature dataset

The U.S. daily maximum and minimum temperature data used to calculate the monthly and seasonal DBP and HDD indices were from the Richman–Lamb dataset for North America used in Skinner et al. (1999). The construction of this dataset followed that of a counterpart daily precipitation dataset developed earlier (e.g., Richman and Lamb 1985, 1987; Gong and Richman 1995; Montroy et al. 1998). This temperature dataset contains 766 stations across the United States and southern Canada east of the Rocky Mountains (Fig. 2), for which complete records of daily maximum and minimum temperatures are available from January 1949 through December 2000. For the United States, 98% of the stations are NWS cooperative stations and the remainder are NWS first-order stations. In developing this dataset, missing temperature values at each primary station in Fig. 2 (which totaled ∼21%) were replaced by substitutes from nearby secondary or tertiary stations as described in Richman and Lamb (1985) and Gong and Richman (1995). This missing data rate at primary stations is higher than for the counterpart daily precipitation dataset (15%). Despite its intentional gridlike distribution of primary stations, the dataset contains actual temperature observations and not objectively analyzed or otherwise interpolated values at formal grid points. The U.S. portion of the dataset was derived from the TD-3200 file of NCDC, and the Canadian data were furnished by the Atmospheric Environment Service/Meteorological Service of Canada.

The Richman–Lamb U.S. temperature stations used here (538 of 766) to quantify the regional DBP and HDD indices were those located within the 9 U.S. Petroleum Administration for Defense Districts (PADD) investigated (Fig. 3). As noted above, even though regional monthly residential natural gas consumption data are available from the U.S. Energy Information Administration only since January 1989, the DBP and HDD indices evaluated from the Richman–Lamb dataset were calculated for the longest period possible (January 1949–December 2000). The Richman–Lamb dataset has not yet been updated after 2000; the resource-demanding nature of such updates limits their occurrence to only every 8–10 yr.

b. Petroleum Administration for Defense Districts

The U.S. Department of Energy (DOE) divides the United States into five PADDs for planning purposes and organization of historical energy data and further subdivides the populous PADD 1 (entire U.S. East Coast; Fig. 3) into three subdistricts: PADD 1X (New England), PADD 1Y (mid Atlantic), and PADD 1Z (southeastern United States). The PADD structure was designed to reflect demand/supply dynamics for petroleum and natural gas and especially the time required for national supply (via pipelines) to meet regional demand. By grouping areas that are similarly distanced from their energy import source, each PADD should have relatively equal sensitivities to winter demand “spikes” based on distance from that source (http://www.eia.doe.gov/oiaf/servicerpt/fuel/gas_appb.html). We performed the analyses that follow within the PADD structure (subject to slight modification) in the belief that it will increase the utility of the results for DOE and energy companies. This approach is consistent with Changnon et al.’s (1999, p. 60) recommendation that “to be effective in interactions with ‘users’ of atmospheric information . . . one has to learn about the industry . . . and how the material is presented [to it].”

This study considers the three PADDs east of the Rocky Mountains but also subdivides the large PADD 2 (Midwest) and PADD 3 (Gulf Coast) into four and two subdistricts, respectively. As shown in Fig. 3, the resulting subdistricts of PADD 2 are PADD 2A (Ohio River Valley), PADD 2B (Great Lakes region), PADD 2C (northern Great Plains), and PADD 2D (central Great Plains), and the subdistricts of PADD 3 are PADD 3A (central Gulf Coast and Arkansas) and PADD 3B (Texas). These subdivisions reflect winter-climate spatial uniformity (e.g., Iowa in PADD 2C rather than 2D; Figs. 4, 5), a desire to minimize subdistrict size variability (e.g., Ohio in PADD 2A rather than 2B), and proximity to import source (e.g., same Ohio placement; http://www.eia.doe.gov/pub/oil_gas/natural_gas/analysis_publications/deliverability/pdf/appa.pdf). Separate DBP and HDD indices were calculated for each of these areas (henceforth termed PADDs 1X, 1Y, 1Z, 2A, 2B, 2C, 3A, and 3B).

c. Residential natural gas consumption data

State-level monthly residential natural gas consumption data were obtained from the Energy Information Administration (http://tonto.eia.doe.gov/dnav/ng/ng_cons_top.asp) for January 1989–December 2000. As noted above, consumption data for earlier years (beginning with January of 1973) were available only on a national basis and anyhow reflected regulation-related, nonweather influences on gas prices and consumption. Total monthly residential natural gas consumption since 1989 in each of the nine PADDs in Fig. 3 was calculated by aggregating state-level consumption data. To remove the seasonal cycle, these individual monthly totals were expressed as anomalies from the 1989–2000 calendar monthly means. Seasonal consumption in each PADD was obtained by summing the consumption data for the season months involved (November–February and December–February).

Figure 4 indicates the importance of natural gas for residential heating in each PADD. The Great Lakes region (PADD 2B) has the largest total consumption (∼24% of national) and per capita consumption, which reflects the high dependence of a large population on gas for home heating during the cold Great Lakes winters. High per capita consumption also characterizes the much less populated but similarly cold northern-tier PADD 2C (northern Great Plains), along with PADDs 2D (central Great Plains) and 2A (Ohio River Valley) that border PADDs 2B–2C immediately to the south. In contrast, the much lower per capita consumption in the other northern-tier PADD (1X: New England) reflects a much reduced dependence on gas for residential heating there. Indeed, per capita consumption in New England is only slightly larger than in PADD 3A (central Gulf Coast and Arkansas), where total consumption is one-third that of New England. Many homes in New England are heated by other types of energy—fuel oil (especially), propane, and electricity.

d. Temperature index and gas consumption time series

Monthly and seasonal DBP and HDD indices were calculated for three daily temperature measures (maximum, minimum, and mean) for each 3-month (December–February) and 4-month (November–February) winter season during 1949–2000 and then linked to form discontinuous 1949–2000 time series for each temperature measure/time scale/season length. November was added to the traditional winter season because early-winter temperature–gas consumption relations potentially can influence energy trading decisions of the type discussed at the outset; early cold promotes speculation about late-season gas storage/availability/price levels (D. J. O’Brien 2003 and 2004, personal communication). Employment of the full suite of daily temperature measures accommodated the widest range of meteorological conditions that potentially could impact gas consumption. For example, a dry, cloudless day with very light winds under an anticyclone center likely would have a higher maximum temperature anomaly than minimum temperature anomaly because of nocturnal radiative cooling.

Because of missing months in the 1948/49 (November, December) and 2000/01 (January, February) winters, those months and associated seasons were not included in the DBP and HDD time series. In a similar way, the 1989–2000 time series of monthly and seasonal natural gas consumption did not include November and December of 2000 or the 2000/01 season, respectively, when correlated with the same DBP and HDD time series segments.

e. Calculation of “days below percentile” indices

The above DBP index time series contain monthly and seasonal totals of days on which a daily temperature measure (maximum, minimum, or mean) fell below site-specific and month-specific or season-specific temperature percentile thresholds. Separate monthly and seasonal DBP indices were calculated for 3-month (December–February) and 4-month (November–February) winters for the 2nd through 80th temperature percentiles at two percentile intervals for each daily temperature measure.

For the monthly time scale, separate percentiles for a given temperature measure were based on the ranking (in ascending order) of all daily values for that calendar month during 1949–2000. For example, the 40th percentile of daily maximum temperature for January for each station was calculated by arranging the 1612 January daily maximum temperature values in ascending order and determining which temperature threshold exceeded the lowest 40% of the ranking (Wilks 1995, p. 23). Because the percentiles were calculated separately for each calendar month, the monthly DBP time series are measures of monthly anomalous temperature from which seasonal cycle effects are removed. Seasonal DBP index time series were developed similarly by summing the monthly totals of days below the above percentiles over each 3- and 4-month winter season during 1949–2000.

Spatial patterns of the 20th, 40th, 60th, and 80th percentiles of January daily mean temperature are displayed in Fig. 5 as examples. Here, higher percentiles represent warmer temperatures at every location because they are calculated from daily values ranked in ascending order. Also, the same percentile represents much colder temperatures in northern areas than farther south because of the strong poleward winter temperature gradient. This regional difference served as additional motivation for developing indices based on percentiles, because they can provide information on regional differences in human sensitivity to cold temperatures. To be specific, if residents of southern locations with warmer winter climates are more sensitive to occasional cold, temperature thresholds for home heating in these warmer locations likely are higher than they are in typically colder locations. Use of percentiles accommodates such a difference in human perception, because the same percentile value represents warmer temperatures in southern PADDs than in northern PADDs (Fig. 5).

The four steps used to determine “optimal” monthly and seasonal DBP indices were

1. 2nd–80th percentile values for 1949–2000 were calculated at 2-percentile intervals for each daily temperature measure (maximum, minimum, and mean) for each winter month and season, for each of the 538 Richman–Lamb stations (Fig. 2) in the nine PADDs (Fig. 3);

2. 1949–2000 time series of PADD-averaged monthly and seasonal totals of days on which the daily temperature fell below each of the above percentiles were produced for each percentile and daily temperature measure for 3- and 4-month winters;

3. each time series was “scaled” from 0 to 100 using the two-step process Heim et al. (2003) developed as part of the REDTI—this scaling first involved subtracting the minimum value in each time series from every member value, which set the minimum in the resulting time series at zero, and then each value in that resulting time series was multiplied by the number that converted its maximum to 100; and

4. the monthly and seasonally scaled DBP indices with the highest correlation to regional natural gas consumption for 1989–2000 were selected to represent each PADD (see below).

The scaling procedure in step 3 was designed to simplify and facilitate year-to-year comparisons of DBP index values, because zero is the resulting minimum value of each final index time series and 100 is the assigned maximum value. Because constant (but different) numbers are 1) subtracted from and 2) multiplied with each member of the original index time series in this scaling process, the correlations with natural gas consumption are unaffected.

f. Calculation of “heating degree day” indices

The second method used to quantify the relationships between monthly/seasonal temperature and residential natural gas consumption was the more traditional calculation of HDD totals. These also were computed for individual months and seasons for the above 3- and 4-month winters. For a given day, HDDs are defined as the number of degrees by which a daily temperature measure (maximum, minimum, or mean) falls below a given reference temperature, and they can be summed over days, months, seasons, or any other extended time period. HDD values were calculated using a wide range of reference temperatures (from 0°C/+32°F to +40°C/+104°F at 2°C/3.6°F intervals) for each above temperature measure. This approach accommodated the substantial winter climate variation over the large study region, explored the potential utility of reference temperatures other than the traditional +65°F/+18.3°C, and maximized correlations with gas consumption.

Spatial distributions of average January HDD totals for 0° and +18°C daily mean reference temperatures are presented in Fig. 6 as examples. The mean January monthly HDD totals are much larger for the northern stations, at which colder winter climates are experienced, and also are considerably higher for the +18°C reference temperature than for 0°C. The substantial variability from north to south in winter monthly HDD totals for a given reference temperature was one motivation for considering a wide range of reference temperatures in the development of the HDD indices.

The five steps used to determine the optimal monthly and seasonal HDD indices were

1. raw HDD totals were computed for 1949–2000 for each reference temperature (from 0°C/+32°F to +40°C/+104°F, at 2°C/3.6°F intervals) for each daily temperature measure (maximum, minimum, and mean) for each winter month and season for each of the 538 Richman–Lamb stations (Fig. 2) in the nine PADDs (Fig. 3),

2. the raw monthly HDD totals were converted to monthly anomalous HDD totals to remove seasonal cycle effects and unadjusted seasonal HDD totals were retained to form seasonal HDD time series,

3. 1949–2000 time series of PADD-averaged monthly and seasonal HDD totals were produced for each daily temperature measure and threshold for 3- and 4-month winters,

4. the resulting HDD time series were scaled from 0 to 100 using the same two-step process as described for the DBP index time series, and

5. the monthly and seasonally scaled HDD indices with the highest correlation to regional natural gas consumption for 1989–2000 were selected to represent each PADD (see below).

g. Regression and correlation analyses

To understand the variability in regional natural gas consumption, the full range of above time series of monthly and seasonal DBP and HDD indices for the nine PADDs were regressed against residential natural gas consumption time series for each PADD from January 1989 to December 2000. The goal was to determine which indices were best correlated with regional gas consumption on winter monthly and seasonal time scales. Because the study focused on the winter season as a whole, the monthly-scale correlation analyses pooled the data for the above 3 or 4 calendar months and did not treat those months separately. A linear model was employed because scatterplots suggested the relationships either were strongly linear (e.g., r ≥ +0.75) or exhibited no higher-order polynomial association. Although these regression/correlation analyses were performed for a relatively short (11 yr) period, the results presented below exhibit considerable robustness through spatial and temporal consistency and standard statistical significance testing.

Because the monthly DBP and HDD indices developed were measures of monthly anomalous temperature, they were correlated with anomalous monthly gas consumption. The seasonal DBP and HDD indices conversely are unadjusted measures of seasonal temperature and therefore were correlated with total seasonal gas consumption. The optimal DBP base percentiles and HDD reference temperatures producing the highest correlations with natural gas consumption in each PADD then were compared to determine regional patterns of index consistency and correlation strength. The magnitude and importance of PADD-to-PADD and regional differences in optimal DBP base percentiles/HDD reference temperatures were inferred from plots of correlation versus base percentile/reference temperature.

a. Regional patterns of maximum DBP and HDD index correlations with residential natural gas consumption

Consistent regional patterns are evident in Fig. 7 for the maximum DBP and HDD index correlations with residential natural gas consumption on the monthly time scale. In particular, the optimal DBP and HDD indices for most northern PADDs (e.g., 1Y, 2A, 2B, 2C) have substantially higher monthly maximum correlations with gas consumption than those for most southern PADDs (e.g., 3A, 3B) and the central Great Plains (PADD 2D). The maximum correlations for the northern PADDs are identical, or nearly so, for the DBP and HDD indices. Especially high maximum monthly correlations (r from +0.89 to +0.91) consistently characterize PADDs 2A (Ohio River Valley) and 2B (Great Lakes region). The lower monthly maximum correlations for the more southern PADDs (r from +0.55 to +0.75) particularly characterize PADD 3A (central Gulf Coast and Arkansas; r from +0.55 to +0.60) and vary more between DBP and HDD indices than farther north.

These lower southern PADD monthly-scale correlations stem from weak correlations in January and, to a lesser extent, February. Figure 8 documents this situation for the most pronounced case of PADD 3A. Here, the 36th (optimum monthly) percentile DBP correlation with gas consumption is strong for November and December (r = +0.87 and +0.80, respectively) but is much weaker during January and February (r = +0.33 and +0.54). This pattern characterizes all of the above southern and adjacent PADDs with relatively low monthly DBP and HDD index correlations with gas consumption (not shown). The especially low southern January correlations suggest that residences there may be heated nearly continuously during midwinter (despite relatively warm daytime temperatures)—presumably because of high human sensitivity to cold temperatures in this area of otherwise comfortable climates. Such behavior would weaken the dependence of January natural gas consumption on temperature. Deficient insulation may similarly inflate gas consumption relative to temperature, especially in Januaries when strong northerly winds penetrate deep into the South.

The above regional contrast is reduced greatly or is absent for the seasonal time scale (Fig. 7), primarily because seasonal DBP and HDD index maximum correlations with central Great Plains and southern PADD gas consumption (r from +0.81 to +0.97) are much higher than their monthly counterparts. For most northern PADDs, the increases in maximum correlation from the monthly to seasonal time scale are constrained to be relatively small by the aforementioned high maximum monthly correlations there. The much higher maximum seasonal correlations for the central Great Plains and southern PADDs stem from reduced impacts of the above low January correlations because of the seasonal aggregation of individual winter months. This applies especially to PADD 3A (central Gulf Coast and Arkansas), for which the variance explained increases strikingly from 30%–36% (monthly) to 66%–94% (seasonal) and surprisingly approaches that obtained for the northern PADDs with the strongest seasonal relationships (again, 2A and 2B, 86%–94%). Furthermore, these PADD 3A increases are largest when November is included in the season, which further offsets the above effect of January. Maximum correlation increases from the monthly scale to the seasonal scale across the South are smallest for PADD 1Z (southeastern United States), for which monthly correlations are higher than for PADDs 3A, 3B, and 2D.

Especially on the monthly time scale, the maximum correlations in Fig. 7 for PADD 1X (New England) are relatively low when compared with other northern PADDs that experience similarly lengthy and cold winters. For PADD 1X, the monthly (seasonal) maximum correlations are between only +0.64 and +0.69 (+0.62 and +0.80). These low maximum correlations likely result from the greatly reduced dependence on natural gas for home heating in New England, as indicated by PADD 1X’s low per capita consumption values in Fig. 4. To be specific, homes in the northern, more rural, and colder areas of the PADD (Maine, New Hampshire, and Vermont) primarily are heated with fuel oil, whereas homes in the southern, more urban, and warmer areas of the PADD (southern New York, Connecticut, Rhode Island, Massachusetts) more often are heated with natural gas. For example, in 2003, fuel oil was used to heat 79.2% of Maine homes but only 41.4% of Rhode Island homes farther south (see www.census.gov/acs). Thus, total PADD gas consumption is less sensitive to PADD-averaged temperature indices for PADD 1X than for other northern PADDs where gas is by far the dominant residential heating fuel. An attempt to address this situation through state population weighting of gas consumption in PADD 1X produced negligible change in the maximum correlations in Fig. 7. This outcome may reflect the higher-density housing, urban heat island effects, and coastal temperature amelioration in the more populated southern areas.

Tables 1 and 2 complement Fig. 7 by indicating the daily temperature measure and DBP percentile/HDD reference temperature that yielded each of Fig. 7’s maximum DBP/HDD index correlations with natural gas consumption. The regional patterns of DBP percentile/HDD reference temperature involved are addressed in the next two sections. Table 3 further documents the dependence of the maximum gas consumption–temperature correlations in Fig. 7 on the daily temperature measure.

Two striking regional patterns are evident in Table 3. First, all optimal DBP and HDD indices for the far northern PADDs 1X (New England), 2B (Great Lakes region), and 2C (northern Great Plains) are based on daily maximum temperature, albeit by small correlation differences in all cases (≤0.04). This situation could result from daily minimum temperatures in these northern regions often being so cold that residences may be heated at sufficient but conservative rates throughout many winter nights, especially during December–February. If so, gas consumption may be less sensitive to absolute overnight temperature minima than to low daytime maxima that affect peoples’ living conditions more strongly. On the other hand, all optimal DBP/HDD indices for the far southern PADDs 1Z (southeastern United States), 3A (central Gulf Coast and Arkansas), and 3B (Texas) are based on daily minimum or mean temperature, with the correlations for indices based on daily maximum temperature being less by 0.08–0.11 (PADDs 1Z, 3A) and 0.03–0.05 (3B). This result could reflect that winter daily maximum temperatures in these regions often are too warm to require much daytime heating and thus influence gas consumption.

b. Regional patterns of DBP index base percentile

The small range of gas–temperature correlations for the three daily temperature measures (Table 3) indicates that the gas–temperature associations obtained are robust and are not strongly influenced by the meteorological conditions reflected in daily maximum and minimum temperatures. This finding permitted Fig. 9 to use daily mean temperature to illustrate the regional patterns of the percentiles for which DBP indices correlated most strongly with residential natural gas consumption. The DBP-index information in Table 3 shows that the correlations for daily mean temperature are within 0.04 of the largest correlation for each PADD. Thus, the major features of the percentile results in Fig. 9 reflect substantially their counterparts for DBP indices based on daily maximum and minimum temperature (not shown).

The optimal percentiles in Fig. 9, both monthly and seasonal, show remarkable regionwide independence of season length. With the possible exception of PADD 3B (Texas), these percentiles are unaffected by the addition of November warmth to the winter and thus are especially robust. However, in particular for the seasonal time scale, the optimal percentiles are somewhat higher in most northern PADDs than in southern PADDs (Fig. 9). This result suggests that lower relative temperature thresholds (in relation to the prevailing regional climate) must be reached in warmer regions for residents to heat their homes.

For example, as suggested by Figs. 5b,c, the optimal seasonal percentile obtained for PADD 2C (northern Great Plains; 50th) corresponds to daily mean temperatures ranging from −17°C/+1°F to −7°C/+19°F across that region during January, whereas the same percentile represents a January range from +5°C/+41°F to +22°C/+71°F over PADD 3A (central Gulf Coast and Arkansas). It is clear that, for PADD 3A, 50th-percentile temperatures are too warm for most residents to heat their homes continuously, especially over southern areas of the PADD, and so a DBP index based on a lower (36th or 34th) daily mean temperature percentile has the highest seasonal correlation with residential natural gas consumption there (Figs. 8, 9). The 36th/34th percentile daily mean temperature range across PADD 3A for January is from 0°C/32°F to 10°C/50°F, as suggested by Fig. 5b. Note, however, that in an absolute sense the southern PADD 3A’s optimum percentiles correspond to temperatures that still are warmer than their northern PADD 2C counterparts. This situation may reflect southerners’ greater sensitivity to winter cold, perhaps as a result of inferior home insulation. Further such percentile–temperature information can be extracted from Fig. 5 [e.g., for Fig. 9’s monthly results for PADD 1X (New England; 60th percentile) and PADD 1Z (southeastern United States; 40th percentile)].

The context of the optimal percentiles in Fig. 9 is provided by Fig. 10, which, for the extreme northern and extreme southern PADDs, documents the dependence of DBP index correlations with residential natural gas consumption on DBP percentile. Figure 10 also uses daily mean temperature as the illustrative example, but for each PADD it also features (where different) the temperature parameter that yielded the maximum correlation in Fig. 7 and the DBP information from Table 3. Consistent with the independence of season length in Fig. 9, the correlation curves in Fig. 10 for 3-month winters generally are very similar to those obtained for 4-month winters (not shown).

All curves in Fig. 10 indicate strong (or relatively strong) correlation increases from the 2nd to about the 10th or 12th percentiles that in most cases are followed by smaller rates of correlation growth with increasing percentile until correlation maxima are reached. However, consistent with patterns in Fig. 9, the correlation maxima mostly occur at somewhat higher percentiles for the extreme northern PADDs than for the extreme southern PADDs for reasons already discussed. Furthermore, associated with this difference are progressive and pronounced correlation decreases for the southern PADDs with increasing percentile above the 40th–45th percentiles, especially on the seasonal time scale. These striking northern-versus-southern PADD correlation differences for percentiles above the 50th percentile lead to correlations for the 75th percentile, for example, being much lower for the southern PADDs than for the northern PADDs, with the explained variance involved differing by up to a factor of 2. This result provides further evidence that the associated temperatures in the northern (southern) PADDs remain cold enough (have become too warm) to require home heating.

The shape of many of the correlation curves in Fig. 10 is particularly distinctive. Except for extreme southern PADDs on the seasonal time scale, the maximum correlation is bordered broadly by correlations for surrounding percentiles that are almost as high as the maximum. It is clear that this close similarity means that many of the maximum correlations in Fig. 7a and the DBP part of Table 3 do not differ significantly from the correlations for the surrounding percentiles. However, as noted above, Fig. 9 showed that the percentiles of the maximum correlations also are remarkably independent of season length. The plateaulike features of high correlations in Fig. 10 thus document further the robustness of the temperature–natural gas consumption relations they represent. Furthermore, where the maximum correlations are very strong (e.g., r approaches or exceeds +0.80), there is high potential predictability of residential natural gas consumption from temperature. Such high potential exists for all PADDs on the seasonal scale and for most northern PADDs on the monthly scale (Figs. 7a, 10). Whether that potential is realized depends on the quality of the temperature predictions on daily-to-seasonal time scales. That issue is considered further below.

c. Regional patterns of HDD index reference temperature

Following the procedure employed for the above DBP analyses, Fig. 11 uses daily mean temperature to illustrate the regional patterns of the reference temperatures for which HDD indices correlated most strongly with residential natural gas consumption. The HDD portion of Table 3 shows that the correlations for daily mean temperature are within 0.03 of the largest correlation for each PADD. Thus, the major features of the reference temperature results in Fig. 11 reflect substantially their counterparts for HDD indices based on daily maximum and minimum temperature (not shown).

In pronounced contrast to their DBP counterparts in Fig. 9, the optimal HDD reference temperatures in Fig. 11 (both monthly and seasonal) are markedly higher for 4-month than for 3-month winters for most PADDs north of the extreme south. This season-length dependence stems from the inclusion of November warmth into the 4-month winter season, forcing the optimal HDD reference temperatures higher for most northern PADDs, and thus reduces the robustness of these HDD results relative to the DBP approach. The lowest optimal HDD reference temperatures (<10°C) in Fig. 11 are for PADDs that experience the coldest winter temperatures—1X (New England), 2A (Ohio River Valley), 2B (Great Lakes region), 2C (northern Great Plains), and 2D (central Great Plains). The remaining more southern/coastal PADDs, with milder winter climates, typically have much higher optimal HDD reference temperatures because their daily winter temperatures fall below the optimal reference temperatures of northern PADDs much less frequently. Therefore, southern PADD HDDs based on such lower reference temperatures do not capture as much temperature variability as do HDDs based on their own higher optimal reference temperatures, and thus they have lower correlations with gas consumption.

Figure 11 also contains some departures from the above patterns. Most prominent on the monthly scale are the coastal PADDs from 3A (central Gulf Coast and Arkansas) through 1Z (southeast United States) and 1Y (mid Atlantic) to 1X (New England)—having very similar optimal reference temperatures for 3- and 4-month winters, likely because of the decreased occurrence of continental polar air masses reaching the far southern and eastern PADDs during November.

The context of the optimal HDD reference temperatures in Fig. 11 is provided by Fig. 12, which, for the extreme northern and extreme southern PADDs, documents the dependence of HDD index correlations with natural gas consumption on HDD reference temperature. Figure 12 uses daily mean temperature as an illustrative example, but for each PADD it also features (where different) the temperature parameter that yielded the maximum correlation in Fig. 7b and the HDD part of Table 3. Despite the above dependence of some of the results in Fig. 10 on season length, the correlation curves in Fig. 12 for 3-month winters are very similar to those obtained for 4-month winters (not shown). The reasons for this situation are explained below.

The correlation curves in Fig. 12 are substantially similar for the extreme northern and extreme southern PADDs, with one pronounced exception for the southern PADDs discussed below. For the three northernmost PADDs on both monthly and seasonal time scales, the correlations are highly similar across the full range of HDD reference temperatures considered (from 0° to +40°C). This is true for both the illustrative example of daily mean temperature and the temperature measure that otherwise yielded the maximum correlation (daily maximum temperature), the curves for which essentially are identical. Thus, the northern PADD HDD reference temperatures that gave the maximum correlations in Figs. 7b and 11 contain minimal additional information on gas consumption above that implicit in most other reference temperatures considered. The only possible exception involves the small correlation decrease that occurs as the reference temperatures decline from +10° to 0°C. This important finding negates Quayle and Diaz’s (1980) speculation that changes with time in energy conservation practices and human comfort levels would require consideration of HDD reference temperatures other than the traditional +65°F/+18.3°C.

For the three southernmost PADDs, the monthly and seasonal correlations in Fig. 12 also are very similar for all HDD reference temperatures above a range from +12° to +15°C. Again, this characterizes the curves for the daily mean temperature example and the temperature measure that otherwise yielded the maximum correlation (daily minimum temperature). Most maximum correlations of both temperature measures with gas consumption occur between +16° and +22°C (Figs. 11, 12), a range that straddles the traditional HDD reference temperature of +65°F/+18.3°C. Thus, for the extreme southern PADDs, all HDD reference temperatures above +15°C contain highly similar information about natural gas consumption, which further underscores the continued utility of the +65°F/+18.3°C threshold for this region. Below +10°C, however, the correlation curves in Fig. 12 for the southernmost PADDs drop off substantially, much more so than for the extreme northern PADDs. This decrease, which reflects southern winter warmth, is more pronounced for daily mean temperature than daily minimum temperature where the latter gave the maximum correlation with gas consumption (Fig. 12). It is clear that HDD reference temperatures below +10°C contain reduced information about gas consumption in the southernmost PADDs.