Here, we use the framework of the MEIC (Multi-resolution Emission Inventory for China, http://www.meicmodel.org, last access: 30 September 2018) to estimate China's anthropogenic emissions from 2010 to 2017. MEIC is a bottom-up emission inventory model which covers 31 provinces in mainland China and includes ∼700 anthropogenic sources. Emissions for each source in each province are estimated as follows: Emisi,j,k=Ai,j×∑mXi,j,m×EFi,j,k,m×∑nCi,j,m,n.×1-ηk,n), where i represents the province, j represents the emission source, k represents the air pollutants or CO2, m represents the technologies for manufacturing, n represents the technologies for air pollution control, A is the activity rate, X is the fraction of a specific manufacturing technology, EF is the unabated emission factor, C is the penetration of a specific pollution control technology, and η is the removal efficiency. The details of the technology-based approach and source classifications can be found in Zhang et al. (2007, 2009), Lei et al. (2011), and M. Li et al. (2017a).

The underlying data in the MEIC model are gathered from different sources. Activity rates of energy consumptions by fuel type, by sector, and by province are derived from Chinese Energy Statistics (National Bureau of Statistics, 2018a, b; National Energy Administration, 2018). Productions of various industrial products and penetration of different technologies are collected from a wide variety of statistics (for details, please refer to Lu et al., 2010; Lei et al., 2011). We also use unpublished data from the Ministry of Ecology and Environment to supplement the technology penetration data that are absent in statistics (Qi et al., 2017; Zheng et al., 2017). These data are collected from each plant by local agencies, and then managed and verified by Ministry of Ecology and Environment. The information adopted in this study include pollution control technologies, penetrations, and efficiencies for electric generators, cement factories, iron- and steel-making furnaces, and glass kilns in each province, which are used to calibrate emission control levels (i.e., C and η in Eq. 1) in the bottom-up inventory. Detailed activity rates by province for the year 2017 are not available when the emission trends presented in this work were developed. In this case, we used national activity data for the year 2017 (National Bureau of Statistics, 2018b; National Energy Administration, 2018) and downscaled these national total data to each province using the weighting factors from 2016 data at the provincial level. Unabated emission factors in MEIC are compiled from a wide range of previous studies, for instance, SO2 from Lu et al. (2010), NOx from Zhang et al. (2007), NMVOCs from Li et al. (2014), CO from Streets et al. (2006), primary aerosols (PM10, PM2.5, BC, and OC) from Lei et al. (2011), and CO2 from Liu et al. (2015a). We then override those data by local emission factors summarized in M. Li et al. (2017b) wherever available, to represent the most recent progress on emission factor developments in China.

Emissions from power plants are estimated following the unit-based approach developed by Liu et al. (2015b). In summary, we track the emissions of each unit from electricity generation, fuel quality, and the progress in emission control using unit-specific parameters. Emissions from on-road vehicles are estimated using a county-level emission model developed by Zheng et al. (2014), which resolves the spatial–temporal variability of vehicle ownership, fleet turnover (i.e., new technology penetration), and emission factors. Detailed documentation of the method and data for power plants and on-road vehicles can be found in Liu et al. (2015b) and Zheng et al. (2014), respectively.

We use index decomposition analysis (IDA) to study the driving forces of China's anthropogenic emissions from 2010 to 2017. IDA is one of the major techniques used to analyze the impact of changes in indicators on emission trends (Hoekstra and van den Bergh, 2003). The IDA method is described as follows.

Equation (1) can be converted to a matrix form using the following formula: Emisi,j,k=Ai,jxEη. The technology distribution factors Xi,j,m in Eq. (1) are assembled into the row vector x, and the relevant unabated emission factors EFi,j,k,m are assembled into a diagonal matrix E. The column vector η represents the average removal efficiencies weighted by the penetration rates Ci,j,m,n of all types of pollution control technologies. According to Eq. (2), over a given period of time, any changes in emissions can be decomposed into their component driving factors using Eq. (3). ΔEmisi,j,k=ΔAi,jxEη+Ai,jΔxEη+Ai,jxΔEη+Ai,jxEΔη, where Δ is the difference operator. The four multiplicative terms in Eq. (2) are converted into four additive terms in Eq. (3). Each additive term represents the contribution of one driving factor to the changes in emissions, while all other factors are kept constant. For example, Δη is the change in pollutant removal efficiencies and the last term in Eq. (3) represents the change in total emissions caused by end-of-pipe abatement measures, with the activity range Ai,j, technology distribution x, and unabated emission factor E assumed to be constant.

Technically, the decomposition of four factors in Eq. (3) has 4!=24 unique first-order decomposition results. In this study, we use the average of all possible first-order decompositions (Dietzenbacher and Los, 1998) in the analysis of emission drivers. By way of illustration, one of the 24 possible decompositions is shown in Eq. (4). ΔEmisi,j,k=Emisi,j,k(t1)-Emisi,j,k(t0)=ΔAi,jx(t1)E(t1)η(t1)+Ai,j(t0)ΔxE(t1)η(t1)+Ai,j(t0)x(t0)ΔEη(t1)+Ai,j(t0)x(t0)E(t0)Δη. The decomposition analysis generates a four-dimensional array with dimensions that represent the year (2010–2017), province (size =31), emission source (size >700), and emission drivers (i.e., A, x, E, and η in Eq. 2). This means that for each pollutant, the year-to-year change in emissions can be attributed to the drivers of A, x, E, and η by source and by province. A is the activity effect (e.g., fuel combustion), and the other three factors constitute the overall effect of air pollution control. This study is mainly concerned with source contributions rather than province contributions; hence, we sum the 4-D array of the decomposition analysis results along the province dimension and perform the following analysis at the country scale.

We adopt atmospheric observations to evaluate and validate the emission trends estimated in this study. We use NO2 column retrievals from the DOMINO V2 product (Boersma et al., 2011) and SO2 column retrievals from the Ozone Monitoring Instrument (OMI) V3 product (Krotkov et al., 2015). The 2010–2017 trends of satellite observations are calculated over eastern China, where anthropogenic sources are dominant relative to natural sources and are compared against the emission trends of NOx and SO2, respectively. Several recent papers have used satellite retrievals to infer recent trends in emissions from East Asia or China. These results are summarized in this study and compared to our emission estimates. We also collect surface-level SO2, NO2, and PM2.5 concentration data from national air quality monitoring stations (http://106.37.208.233:20035/, last access: 30 September 2018) for the period of 2013–2017. These real-time monitoring stations were established in 2013 and had the ability to report hourly concentrations of criteria pollutants from over 1400 sites in 2017.

China's anthropogenic emissions are estimated to have declined by 62 % for SO2, 17 % for NOx, 27 % for CO, 38 % for PM10, 35 % for PM2.5, 27 % for BC, and 35 % for OC since 2010 (Table 1). Most of these emission reductions have been achieved since 2013 when the Clean Air Action was enacted and implemented. SO2 and NOx are the only air pollutants that were incorporated into national economic and social development plans with emission reduction targets in China. The 12th 5-Year Plan required the total national emissions of SO2 and NOx to be cut by 8 % and 10 % from 2011 to 2015, respectively, while the actual reductions were much larger than planned due to the more stringent pollution control requirements implemented after 2013. During this period, our estimates suggest that CO emissions decreased by 23 %, whereas CO2 emissions were flat, reflecting China's improved combustion efficiency and emission control. The years since 2013 also observed a sharp drop in particulate emissions, in contrast with the flattening emissions observed before 2013. This trend is more evident for coarse particles because they are more easily removed by end-of-pipe abatement measures. Given that China's economy is growing rapidly, China's emissions are decoupling from population, economic, and energy consumption growth (Fig. 2). China's gross domestic product grew by 7.6 % per year from 2010 and achieved 67 % growth by 2017; however, China's emissions flattened out from 2010 to 2013, followed by a significant decrease after 2013 according to our calculations. In contrast, NMVOC emissions are estimated to have increased by 11 % and NH3 emissions remained flat from 2010 to 2017; these trends were mainly due to the absence of effective emission control measures.

We present the sectoral trends of China's emissions in Figs. 3 and 4. The most important sector identified by our estimates is the industrial sector, which is the dominant source of SO2, NOx, PM10, PM2.5, and CO2 emissions during 2010–2017, accounting for average values of 60 %, 38 %, 57 %, 50 %, and 53 % of total emissions, respectively. The industrial sector is the driver of changes in 2010–2017 emissions for these pollutants, except for NOx and CO2, and it also drives down CO and BC emissions. The power sector, though accounting for more than half of burning coal, is not a dominant contributor to the emissions of any pollutant. The reason for this is that upgrading plants with pollution control equipment in the 11th 5-Year Plan (2006–2010) significantly reduced SO2 and particulate emissions from power plants (Liu et al., 2015b), and the remaining part is not compared to industrial emissions. With upgraded emission standards and the spread of the ultralow emission technique, the new emission limit values have further driven down power plant emissions, which is the dominant driving force of the decrease in NOx emissions, while industrial combustion sources lack an effective control on NOx.

The residential sector is the dominant source of CO, BC, and OC, to which it contributes average values of 40 %, 49 %, and 79 % to national emissions, respectively, and is the second-most important source of PM10 (31 %) and PM2.5 (38 %); its relative contributions of these components have increased while industrial emissions have considerably decreased. The residential sector drives OC emissions down and contributes to the reductions of CO, NMVOCs, and particulate matter. The transportation sector accounts for 17 % of CO emissions, 19 % of NMVOC emissions, and 31 % of NOx emissions. The increase in fuel consumption drives up transport CO2 emissions, with an increase of 43 % from 2010 to 2017 that is faster than that of any other emission source sector. Solvent use is a major contributor to the increase in NMVOC emissions. Solvent emissions are estimated to have increased by 52 % since 2010, making them the largest contributor (36 %) to NMVOC emissions in 2017, while the share of this sector was only 27 % in 2010. The agricultural sector is the dominant source of NH3 emissions, as it contributes to 93 % of total emissions. NH3 emissions have remained constant because agriculture and rural activities showed small interannual variations.

The effect of air pollution control can partially or totally offset the additional emissions caused by growing activity rates, and the combination of pollution control and activity growth entirely determines China's emission pathways (Fig. 5). Based on the drivers of emissions, we can classify air pollutants into two categories, namely, activity-driven increasing pollutants and pollution control-driven decreasing pollutants (Fig. 5). NMVOCs and NH3 belong to the former category. Their emissions have continued to increase at a constant rate from 2010 to 2017, primarily driven by activity growth. Assuming that activity rates are frozen at their 2010 levels (Fig. 3), the NMVOC emissions could decrease by 21 % from 2010 to 2017 due to the emission controls on residential and transport sectors. These emission reductions were far outweighed by the growing use of solvent for paints, coatings, and chemical industry, which consequently drove up their total emissions. The solvent used for paints increased by 110 % from 2010, which is attributed to the increasing demand to coat buildings, cars, and machinery due to the rapid increase in the area of newly built houses (+52 %) and the production of vehicles (+54 %). The solvent used in chemical industry also rose at a fast rate due to the increase in industrial production (e.g., ethylene production grew by 28 % since 2010), which makes them the second largest contributor to NMVOC growth after paints. For NH3, the lack of control measures has caused its emissions to correlate well with activity; thus, its emissions do not decline, unlike the regulated pollutants that have experienced progressive emission control.

The other air pollutants followed distinct emission pathways before and after 2013 (Fig. 5). According to our estimates the emissions of these pollutants slightly increased (e.g., SO2 and NOx) or remained flat (e.g., CO and particulate matter) during 2010–2013 because emission mitigation just counterbalanced the additional emissions caused by growing activities. China's fuel combustion increased by 15.2 % from 2010 to 2013 (Fig. 6), and its industrial production increased by 14–35 % in different industries. During this period, China's clean air actions mainly focused on upgrading emission standards for the power and industrial sectors. These measures effectively offset the growth in activities but were not stringent enough to reverse the growing trends in emissions; therefore, air pollutant emissions remained stable from 2010 to 2013.

After 2013, emissions of all air pollutants except NMVOCs and NH3 are found to have reduced as a result of pollution controls. China's fuel combustion and industrial production have flattened out since 2013 (Fig. 6), while high-efficiency mitigation measures have been increasingly implemented in all emission source sectors, as required by the Clean Air Action. Scenario analysis suggests that the effect of pollution control rapidly removes air pollutants and consequently drives down China's emissions (Fig. 3). Assuming that pollution control is frozen at 2010 levels, SO2 emissions in 2017 could increase by 167 % compared to the actual data, NOx and TSP emissions could increase by 38 % and 111 %, respectively, and other pollutants could see increases of 23–66 %. The different reduction rates of air pollutant emissions are determined by the source sector distributions and emission mitigation efforts of each sector. For example, this decrease is most notable for SO2 (emissions are estimated to have decreased by 59 % from 2013 to 2017) because the dominant source sectors (i.e., power and industry) both significantly reduced their emissions. The decrease in emissions is smallest for NOx (21 % of emissions cut from 2013 to 2017 based on the analysis) because the power sector was the major contributor to emission reduction but only accounted for one-third of total emissions. To understand the underlying drivers of emission reduction, we decompose the avoided emissions due to pollution control (i.e., the sum of contributions from x, E, and η) into sectors (Fig. 7) to identify the main drivers underlying key source categories. We select the year of 2017, which exhibited the largest reduction in emissions according to our calculations, to perform this analysis.

1.

The power sector. The generation of electricity from hydrocarbon fuels in China has increased by 33 % since 2010, which has led to increases of 1.2 Tg SO2 and 1.7 Tg NOx in 2017 compared with their levels in 2010 (Fig. 7). Mitigation efforts have yielded reductions of 7.1 Tg SO2 and 6.1 Tg NOx and thus totally offset the emissions caused by growing activities. The reduction of emissions was achieved through the ultralow emission standard. To fulfill the stringent standards, flue gas desulfurization (FGD) and selective catalytic reduction (SCR) systems have been increasingly installed at utilities in coal-fired power plants, with penetration rates reaching >95 % in 2017. Of the current power plants, 71 % have operated close to the design performance of ultralow emission levels (China, 2018).

Comparison of trends in PM2.5 precursor emissions with satellite and ground-based PM2.5 concentrations are presented in Fig. 8a. All data are normalized to the year 2013 because it is the only year that all data are available. Satellite-derived PM2.5 concentrations presented a relatively flat trend during 2010 and 2013 (e.g., Fontes et al., 2017; Liang et al., 2018; Lin et al., 2018), corresponding to small variations in emissions of different precursors estimated for the same period. Satellite-based PM2.5 concentrations over China decreased by 18 % from 2013 to 2015 (Lin et al., 2018), in good agreement with the trend in surface PM2.5 concentrations over 74 cities. During 2013–2017, surface PM2.5 concentrations over 74 cites decreased by 35 %, while emissions of PM2.5 precursors over China presented various changing rates. We estimate a faster decrease in SO2 emissions than the observed surface PM2.5 concentrations, while the estimated decreasing rates of NOx and NH3 emissions were slower than the observed PM2.5 concentrations. This phenomenon was qualitatively confirmed by an observed large decrease in sulfate and increased relative contribution of nitrate in PM2.5 compositions from 2013 to 2017 (Shao et al., 2018). However, the quantitative relationship between PM2.5 concentrations and precursor emissions will require further studies with chemical transport modeling.

Trends in NOx and SO2 emissions are generally consistent with satellite- and ground-based observations during 2010–2017 (Fig. 8b and Table 2). Specifically, rapid decreases in SO2 and NOx emissions after 2013 are confirmed by satellite-based observations. We estimate that NOx and SO2 emissions in eastern China decreased by 21 % and 59 % during 2013–2017 respectively, lower than the 30 % and 73 % decreases in OMI observed NO2 and SO2 columns for the same region and time period. Surface SO2 concentrations decreased by 57 % over eastern China for the period of 2013–2017, in good agreement with the estimated emission trend. In contrast, surface NO2 concentrations only decreased by 9 % for the same time, significantly lower than the estimated trend in NOx emissions.

Different trends between emissions and concentrations could be attributed to many factors. First, temporal and spatial patterns of emissions and concentrations are impacted by variations in meteorology, atmospheric transport, and chemical reactions. Interannual variabilities can result in remarkable variations in column and surface concentrations (Uno et al., 2007), which may partly explain the disagreement between changes in emissions and observations for a signal year (e.g., the year 2011 in Fig. 8b). Surface observations are more sensitive to surface emissions than high-stack emissions and satellite-based column observations are more visible to high-stack emissions due to different transport patterns, in which both will contribute to the differences when comparing trends. In addition, chemical partitioning of NOx into NO2 may also contribute to discrepancies between trends in emissions and observations (Lamsal et al., 2011, Valin et al., 2011). Taking the above factors into account, inverse modeling (IM) approaches were developed to derive top-down emissions constrained by observations. Table 2 presented recently published estimates on top-down emission trends over China using IM approaches. As shown in Table 2, the discrepancies between bottom-up and top-down emission estimates are not always narrowed, indicating uncertainties from other aspects might exist.

Second, uncertainties in observations can also contribute to the discrepancies. In situ observations are usually thought to be more accurate. However, surface NO2 concentrations obtained from national monitoring networks relied on chemiluminescence measurements, which can significantly overestimate NO2 concentrations (Lamsal et al., 2010) and then contribute to discrepancies between emissions and surface observations. Satellite retrievals are subject to larger uncertainties, for instance, tropospheric SO2 columns are quite uncertain due to difficulties in isolating anthropogenic SO2 signals from ozone and volcanic SO2 (Krotkov et al., 2006). Most uncertainties in satellite retrievals are systematic and canceled out when trends are compared. However, influences of aerosols on satellite NO2 retrievals (Lin et al., 2015) may impact the reliability of the NO2 column trend due to the large decrease in aerosol concentrations over the discussed period. SO2 columns are less sensitive to small and near-surface emissions (C. Li et al., 2017), which may lead to an underestimation of the SO2 budget in China for most recent years and a disagreement between emission and SO2 column trends when high-stack emissions (e.g., power plants) were significantly reduced.

Last but not least, emissions estimates are uncertain due to incomplete knowledge of underlying data (Zhao et al., 2011; M. Li et al., 2017b). Similar magnitudes of uncertainties are expected considering previous work (e.g., Zhang et al., 2009; Lei et al., 2011; Lu et al., 2011; M. Li et al., 2017b) since similar methodologies and data sources are used. In general, uncertainties are smaller for species whose emissions are dominant by large sources (e.g., SO2 and NOx) but larger for species whose emissions are mainly contributed by scattered emitting sources (e.g., BC and OC). Many of the uncertainties in bottom-up emissions are also systematic and may have less impact on emission trends (Lu et al., 2011), but non-compliance with regulations due to lack of inspection will lead to differences between estimated and real-world efficiencies of emission control facilities (e.g., Wang et al., 2015) and impact the validity of estimated emission trend. Specifically, the effectiveness of the measures targeting the small and scattered emitting sources (e.g., phase out small high-emitting factories and eliminate small coal-fired industrial boilers) are difficult to validate, which may lead to higher uncertainty ranges in emission estimates for most recent years.