Inter-annual variability in fossil-fuel CO₂ emissions due to temperature anomalies

Scientists and policymakers pay attention to the evolution of energy related carbon dioxide (CO₂) emissions as the main driver for present and future climate change. CO₂ emissions from fossil fuels and industry, as reported by Olivier et al (2015) and Le Quéré et al (2016) grew by approximately 0.5% between 2013 and 2014 and stalled in 2015. The fairly small increase in 2014 despite continued economic growth came as a surprise for many stakeholders. Jackson et al (2015) cautiously suggested that this deceleration in CO₂ emissions, which comes earlier than anticipated, may be the early sign of structural changes occurring in energy systems, in particular in China. However there could be other reasons for inter-annual variations in fossil-fuel CO₂ emissions both at the national and global levels. For instance there could be small variations in CO₂ emissions due to the year-to-year variations in the number of public holidays and a small increase can be expected on leap years, the latter effect being corrected in Le Quéré et al (2016). Meteorology is also known to affect energy consumption in many countries. While statistics for energy efficiency and final energy consumption are corrected for meteorological (e.g. mild winter) effects by some national statistics offices (e.g. in the UK and France), such corrections are not applied for energy consumption or for CO₂ emissions reported in national inventories compiled and aggregated by international institutions (UNSD 2016, IEA 2016, UNFCCC, 2016). Considine (2000) showed that inter-annual weather variations have a sizeable (up to a few %) impact on carbon emissions in the United States. However this effect has not been quantified systematically and remains unknown for many countries. Although such an impact is likely to be small when aggregated at the global scale, it needs to be understood—and corrected for—if we want to interpret changes in CO₂ emissions trends on the basis of a small number of years. In the following, we present our methodology and results on the signature of the inter-annual temperature anomalies in the time series of CO₂ emissions at the country and global scales.

2.1. CO₂ emissions

We use estimates of national fossil fuel emissions of CO₂ from EDGAR (European Commission, Joint Research Centre (JRC) and Netherlands Environmental Assessment Agency (PBL); version 4.3.2). This dataset provides yearly values of fossil fuel emissions from 1970 to 2015. CO₂ emission estimates were made by PBL and the JRC on the basis of energy consumption data for the 1970–2013 period, published by the International Energy Agency (IEA, including the correction for China published in 2016), and on 2014–2015 trends, published by BP. The estimations are also based on production data on cement, lime, ammonia and steel developed jointly by the JRC and PBL (Olivier et al 2016). The time series of these emissions are shown for the 15 largest emitters in figure 1.

Zoom In Zoom Out Reset image size

2.2. Economy

2.3. Meteorology

2.4. Population distribution

2.5. Data processing

The HDD and CDD have been interpolated onto the 5' resolution grid of the population data. The relatively coarse resolution of the ERA-interim dataset, upon which the HDD and CDD are based, is such that coastal and orographic effects on surface temperature may be missed. However this approximation is not thought to have any significant impact on inter-annual variations. The HDD and CDD have then be convoluted with the population and integrated at the country scale using a digital map of countries. As such, indexes are expected to provide a more realistic estimate of the potential energy demand for heating or cooling on a national basis than the use of un-weighted degree-day values (Taylor 1981). The results of this convolution are a time-series of annual heating-degree-day-person (HDP) and cooling-degree-day-person (CDP) for each of the 106 countries considered here. Although some of the time series span a longer period, our analysis is based on the 1990 to 2015 period (26 years).

Note that some data (either emissions or GDP) are missing for some of the world countries. Our analysis is thus based on 103 individual countries. The total of their 2015 emissions is 34.0 GtCO₂, whereas the total of EDGAR territorial emissions is 34.9 GtCO₂. Our analysis is therefore based on 97.5% of the territorial fossil fuel emissions, excluding bunker fuels. Note also that we account for leap years in a simple way: all input time series to our analysis (CO₂ emissions, GDP, HDP and CDP) are corrected by a factor 365/366 for all leap years.

Our goal is to distinguish the respective contributions of the economy and the meteorology. We therefore fit the national CO₂ emissions time series as

$$\begin{equation} \begin{array}{*{20}c} {\text{EMI}_{\text{CO}_\text{2} } (\text{yr})} \hfill & = \hfill & {\text{Int(yr)}\;\text{GDP(yr)} + \alpha \;\text{HDP(yr)}} \hfill \\ {} \hfill & \text{ + } \hfill & {\beta \;\text{CDP(yr)} + \varepsilon (\text{yr})} \hfill \\ \end{array} \end{equation} \tag{ 1 }$$

where α and β are two constants to be adjusted on the data and ε(yr) is the fit residual. Int(yr) describes the emission intensity of the economy (gCO₂ US$⁻¹). It is expected to be a function with a typical time scale that is longer than that of the GDP variations. In general, the energy efficiency of the economy increases so that Int is a time-decreasing function.

Both HDP and CDP show a mean trend in addition to high frequency variations (examples in figure 1). The high frequency variations are linked to inter-annual anomalies in the temperature. A general trend is visible, which can be related to both a warming of the climate (with increasing CDP and decreasing HDP) and population growth. The question is whether the time series make it possible to identify the respective contributions of GDP, HDP and CDP on EMI_CO2. One expects that the inter-annual variations of HDP, CDP and GDP are uncorrelated, which would make this distinction possible.

To determine α and β, we use the IDL function REGRESS that provides an estimate of the linear sensitivity to HDP and CDP together with their uncertainty. Many countries show no significant values for either HDP or CDP, in which case the linear regression is applied to the other temperature-dependent parameter only. More precisely, we do not attempt a double linear fit when the inter-annual variability of HDP or CDP is less than 10% of the other. We also disregard the fit when the correlation is less than 0.15.

The economy parameter Int is also adjusted on the data. As said above, one assumes that the economy intensity varies slowly in time. We thus retrieve Int using a Savitzky–Golay filter with a time period of 10 years based on the ratio of the CO₂ emissions and the GDP, after subtracting the heating and cooling-related emissions. The retrieval of the simple model parameters is thus an iterative process that estimates Int and (α, β) successively. Four iterations are sufficient to reach convergence for the model parameters.

Figure 1 shows the various time series used for this analysis for the 15 countries with the larger fossil fuel emissions, contributing 78% of the world total in 2014. Figure S-1 in the supplementary material does the same for the 15 following countries in terms of emission, which represents an additional 12% of the global CO₂ emissions (i.e. countries in figure 1 and figure S-1 combined encompass 90% of the total fossil fuel emissions, bunker fuels excluded). Most countries show a significant increase of GDP during the analysis period. Countries from the former USSR (Russia, Ukraine, Kazakhstan) show a sharp decrease during the 90s and a regrowth afterward. Many countries show a temporary drop in emissions during 2009, a consequence of the global financial crisis. CO₂ emissions of most countries show a trend similar to that of the GDP consistently with the assumption of our analysis. The Int (Intensity) parameter is shown as a green line. Although most subplots show a decreasing trend, indicating an improvement in the energy efficiency, there are significant exceptions among developing countries.

In most cases, the combination of GDP, Int, HDP and CDP (black dotted line) components reproduces accurately the estimate of EMI_CO2 (black plain line) for both the trend and some of the higher frequency variations. An earlier version of this analysis used the CDIAC dataset (Boden et al 2010) for the CO₂ emissions. With this input, the fit was not as good for oil exporting countries such as Saudi Arabia, Venezuela and the United Arab Emirates, and the time series were somewhat suspicious. The CO₂ emission accounting for CDIAC is sensitive to uncertainties in production and export of fossil fuels. These observations led us to use the EDGAR estimates instead, which appear more consistent and are the only alternative covering the period up to 2015 on the global scale at the time of writing of this paper.

Figure 2 shows a scatter plot of the residual EMI_CO2 (yr) − Int(yr) GPD(yr) as a function of both CDP (blue) and HDP (red). When a significant correlation was found in the data analysis, the best linear fit line is shown and its slope and uncertainty are provided. The slopes of the linear fit are positive in most cases. This is somewhat expected but nevertheless provides some support to our data analysis that does not impose such a positive relationship. Note however that the uncertainties are large and often encompass the null slope (i.e. the one sigma uncertainty is larger than the best estimate of the slope). This is better seen in figure 3 that shows these slopes and their uncertainties for the 20 major emitting countries. In figure 2, the CDP (resp. HDP) datapoints are not shown when their variation is much smaller (a fraction of 0.02 or less) than that of the other. Note that, for countries for which a significant relationship with both CDP and HDP is found, the visual impression of the subplot in figure 2 and figure S-2, may be misleading as the scatter around each best fit is somewhat altered by the dependence to the other variable. Also, the relative variations of CDP and HDP differ markedly between countries; HDP is dominant for mid-latitude countries whereas CDP variations are larger for the tropical countries. These different ranges of variations explain the simple model capabilities to estimate, or not, the CO₂ emission sensitivity to heating and cooling.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Figure 3 shows that the CO₂ emissions associated with heating are on the order of 500–1000 gCO₂ per degree-day-person for mid-latitude countries. Canada, a country notable for harsh winters, shows lower values, which might be related to a better home insulation adapted to the local climate. China also shows a sensitivity that is lower than that of other mid-latitude countries, while the opposite is true for South Korea. The CO₂ emissions associated with cooling show a larger diversity. Interestingly, several subtropical countries (India, Indonesia, Brazil, Thailand) show very small sensitivities to CDP (a few tens gCO₂ CDP⁻¹) with small uncertainties, possibly indicating a small penetration of air conditioning. The two largest emitters, China and the United States, have retrieved estimates for both heating and cooling, but the cooling uncertainty range is large and encompasses the null value for the latter. This low sensitivity to CDP is surprising and contradicts earlier results. For instance, Hao et al (2016) concludes that CO₂ emissions in China are more sensitive to CDD than to HDD. For the United States, Hadley et al (2006) indicates that cooling is less energy efficient than heating, so an increase in cooling needs (and associated fossil-fuel carbon emissions) can more than offset an equal decrease in heating needs (and associated fossil-fuel carbon emissions). Our results for the cooling sensitivity are therefore suspicious.

As a sensitivity analysis, we have perturbed the time series of HDP and CDP, and applied the same regression tools. One test uses the fully randomized time series of HDP/CDP, while another swapped the values of each odd and its following even years. The latter method retains the long-term trend in the time series but de-correlates the year-to-year variations. In both cases, the correlations are much less than those shown above and no consistent CO₂ emission sensitivity to the temperature is found. This simple analysis provides strong evidence that the correlations and retrieved sensitivity are not spurious.

Figure 4 is an analysis of the annual growth rate of CO₂ emissions for the same countries as in figure 1 and figure 2. It shows the various components that contribute to the growth rate (in a given year relative to the previous year), according to our simple model. The components are the GDP, the energy intensity, heating and cooling. Note that the ranges vary markedly between countries so that we use different scales that are indicated on the right side within each subplot. This figure shows that the inter-annual variations of the meteorology have a small but discernible influence on the CO₂ emission growth rate. This influence appears particularly significant for Germany, South Korea and Japan (figure 4), as well as the United Kingdom, Italy, France and Poland (supplementary figure S-3). The model however is far from perfect as the actual growth rate (diamond symbols) sometimes differs significantly from the sum of the components (horizontal white bar). Some uncertainty on the input data may also contribute to the discrepancies. For the seven countries cited above, for which the relative contribution of heating/cooling is the largest, there is a rather good agreement between the modelled CO₂ growth rate and its true value. One conclusion that can be derived from this figure is that the strong decrease of Germany's emissions in 2014 was for a large part due to a mild winter, confirming previous findings (EEA 2015). Indeed, the winter of 2014 was near record warmth in Europe and its impact on the emissions can be seen on the figure 4 subplot for Germany, but also for the United Kingdom, France, Italy and Poland (figure S-3).

Zoom In Zoom Out Reset image size

The results derived at the country level can be aggregated to analyse the impact of inter-annual meteorology on the CO₂ emissions at the global scale. For this objective, we use the heating and cooling sensitivity convolved by the HDP and CDP time series. The uncertainties are also aggregated assuming that the errors for the various countries are independent. When the estimate was unsuccessful, we assume that there is no CO₂ emission related to heating or cooling. There are also some countries for which the GDP data is incomplete so that our analysis could not be applied to them, but these countries contribute a very small fraction of the total emissions. The results of the aggregation are shown in figure 5. According to our analysis, heating related emissions contribute slightly above 3 ± 0.5 GtCO₂ per year, which is about 10% of the total fossil fuel emissions. Overall, the share of cooling and heating emissions are broadly consistent with their share in final energy consumption in 2010 according to Labriet et al (2015) (10.4%) and the share estimated by Isaac and van Vuuren (2009) with a much more detailed model (nearly 3 GtCO₂ for domestic heating and cooling only). There are inter-annual variations of a few tenths of GtCO₂. Cooling related emissions, around 0.45 ± 0.35 GtCO₂ per year, are much smaller than those of heating but show a significant positive trend. Note however that the general trend for the cooling emissions is 0.4 GtCO₂ from 1990 to 2014, which is the order of magnitude of the interannual variations of the heating emissions. Both population increase and global warming contribute to the positive trend on the cooling emissions. Conversely, the impact of global warming on the heating-related emissions and that of the population growth are opposite in sign and tend therefore to compensate. This explains that the relative trend on the cooling-related emissions is much larger than that of the heating-related emissions (figure 5). Penetration of air conditioning plays a role as well, but cannot be captured by our simple modelling approach. However, even in more detailed modelling attempts, such as the one conducted by Isaac and van Vuuren (2009) for the building sector, this is not expected to play a major role on heating and cooling induced fossil-CO₂ emissions over the period covered by our analysis (see their figure 7).

Zoom In Zoom Out Reset image size

Figure 6 is similar to the subplots in figure 4 but for the total of the 108 countries considered in our analysis. This total differs from the world total given in EDGAR because it does not include emissions from the few countries excluded from our analysis for lack of GDP data, and the so-called 'bunkers'. This comes in addition to the slight differences with the national time series provided by CDIAC. As a result, our growth rates differ slightly from those of CDIAC. In particular our growth rates for the countries considered are 2.2, 1.1 and −0.1% for 2012/2013, 2013/2014 and 2014/2015, respectively, whereas they are 1.3, 0.8 and 0.1% for the CDIAC 'World' emissions discussed in Le Quéré et al (2015, 2016). Figure 6 shows that the increase in emissions is mostly fuelled by the economic growth quantified by the GDP, partly compensated by a decrease in the emission intensity of the economy. The inter-annual variations of the GDP are large and represent the main driver of variations in CO₂ year-on-year emission growth rate. Heating and, to a lesser extent, cooling also contribute to the inter-annual variation of the growth rate. There was a significant decrease of the emission growth rate in 2014, which was attributed to the slowdown of the Chinese economy and a gradual shift from coal to renewable energy (Green and Stern 2015, Jackson et al 2015). This slowdown was recently confirmed with the release of 2015 emissions (Le Quéré et al 2016). The heating-related emissions of 2014 are significantly smaller than those of 2013 by 150 ± 50 MtCO₂, while the cooling emissions are also smaller by ≈70 MtCO₂. The sum of these decreases represents about 0.5% of total emissions. In 2015, heating related emissions decrease by an additional 80 MtCO₂, partly compensated by an increase of the cooling emissions (figure 5). Our analysis thus indicates that meteorology was a significant factor, explaining about half of the 2013 to 2014 emission growth reduction. Note that the true emission growth (diamond symbol) for the three most recent years is well reproduced by our simple model (white horizontal bar) while it is not the case for a few earlier years. The discrepancies for these years confirm that other factors for the CO₂ growth rate have had a significant influence that is unaccounted for in our simple analysis. Nevertheless, our analysis indicates a significant contribution of meteorological effects to the decrease of the CO₂ emission growth rate from 2013 to 2014, and a small but no significant impact from 2014 to 2015.

Zoom In Zoom Out Reset image size

Our analysis confirms that time series of national CO₂ emissions from 1990 to 2015 exhibit a close relationship with gross domestic product (GDP) modulated by the slow-changing carbon intensity of the economy. The inter-annual variations in CO₂ emissions that are not explained by the change in GDP, are positively correlated to HDP and/or CDP for many countries, with slopes of typically 500–1000 gCO₂ (degree-day-person)⁻¹ but with significant country-to-country variations. The inter-annual anomalies of the meteorology contribute significantly to year-on-year variations in the CO₂ emission growth rates, in particular for mid-latitude countries.

We acknowledge several shortcomings and uncertainties in our modeling and results. First, the input time series convey some uncertainties and we have not attempted to quantify their impact on the results. Second, there may be hidden correlations between the time series that are not properly accounted for in the modeling. For instance, the meteorology anomaly may impact the GDP and emissions in ways that are not related to heating and cooling. An example is the precipitation anomalies that drive the hydro-electricity potential, and therefore the need to use coal power plants as a substitute. Another example is the fact that (bad or good) weather is known to affect consumption and GDP. Finally, we have hypothesized that the heating and cooling emissions are proportional to the HDP and CDP with arbitrary definition of temperature thresholds. In practice the typical usage of heating and cooling devices may change with time: on the one hand, the surface heated or cooled per inhabitant may increase; on the other hand the average home insulation may improve; finally the fossil carbon content of the energy may change over time together with the use of fossil fuels in replacement of traditional biomass or the development of renewables and/or electrification combined with low-carbon sources of electricity. Clearly, there are more accurate ways to quantify the carbon content of heating and cooling, but these would require the use of detailed statistics at the country scale.

Despite these admitted limitations, we have used the HDP and CDP-dependent emission sensitivity inferred from the national time series to aggregate the emission anomalies at the global scale. The heating- and cooling-related emissions of 2014 are significantly smaller than those of 2013 by 220 ± 60 MtCO₂, which represents about 0.5% of the total emissions. Thus, about half of the change in emission from 2013 to 2014 can be explained by temperature factors. Changes in heating- and cooling-related emissions are marginally smaller in 2015 as compared to 2014 and represent a rather small contribution to the limited growth in 2015.