Air quality simulations based on emission inventories with and without incorporation of CEMS data for the coal-fired power sector (cases 1 and 2, respectively) were conducted to test the improvement of emission estimates. Because of the combined influences of regional transport and chemical reactions of air pollutants in the atmosphere, nonlinear relationships were found between the changes of primary emissions and ambient concentrations of air pollutants. Compared to case 1, the simulated annual average concentrations of SO2, NO2, and PM2.5 in the YRD region were 10 %, 7 %, and 6 % lower, respectively, in case 2, while that of O3 was 7 % higher, due to combined effects of emissions of volatile organic compounds (VOCs) and NOX precursors (Gao et al., 2005; Yang et al., 2012). Previous studies have shown that O3 formation in most of the YRD region is under the “VOCs-limited” regime, i.e., the generation and removal of O3 is more sensitive to VOCs and would be inhibited with high NOX concentrations in the atmosphere (Zhang et al., 2008; Liu et al., 2010; Wang et al., 2010; Xing et al., 2011). Therefore, the simulated reduced NO2 concentrations from greater NOX emission control could elevate the O3 concentration.

The model performance was evaluated with available ground observation. The hourly concentrations were observed at 230 state-operated air quality monitoring stations within YRD, and the averages of hourly concentrations of those sites were compared with the simulations in cases 1 and 2, as summarized in Table 2. Similar model performances were found for the two emission cases, with overestimation of SO2, NO2, and PM2.5 as well as underestimation of O3. The NMEs between the simulated and observed SO2, O3, and PM2.5 concentrations were all smaller than 50 % for both cases and slightly worse simulation performances were found in July compared to the other 3 months. In particular, the correlation coefficients (R) between the simulated and observed SO2 in July were only 0.17 and 0.14 for cases 1 and 2, respectively, and the NMEs between the simulated and observed NO2 were larger than 100 %. In addition, greater overestimation of SO2 and PM2.5 by the model was found in July compared to other months, likely attributable to the bias of WRF modeling. On the one hand, the simulated WS10 in the YRD region in July (2.67 m s-1) was slightly lower than the observation (2.75 m s-1). The underestimation in wind speed could weaken the horizontal diffusion and lead to overestimation in air pollutant concentrations. Compared with the results from the European Centre for Medium-Range Weather Forecasts (ECMWF, https://apps.ecmwf.int/datasets, last access: 19 April 2021), on the other hand, the simulated boundary layer height (BLH) was lower in WRF for all months. The NMBs of the WRF and ECMWF BLH in January, April, and October were around -15 %, while that in July reached -24 %. The lower BLH would limit the vertical convection and diffusion of pollutants and thereby increase the surface concentrations of air pollutants. Similar to previous studies (An et al., 2013; Liao et al., 2015; Tang et al., 2015; Gao et al., 2016; Wang et al., 2016; Zhou et al., 2017), underestimation of O3 was commonly found. The NMBs between the simulation and observation for the two cases ranged from -34.5 % to -6.4 % and NMEs from 23.1 % to 37.1 %, respectively. The underestimation in O3 likely resulted from bias in the estimation of precursor emissions. Suggested by the positive NMBs of NO2 modeling in Table 2, the NOX emissions were expected to be overestimated in the two cases, even for case 2 with the CEMS data incorporated (which reflect the emission control benefits in recent years, as discussed in Y. Zhang et al., 2019). In addition, underestimation of VOC emissions is likely due to incomplete accounting of emission sources, particularly for uncontrolled or fugitive leakage (Zhao et al., 2017). As most of YRD was identified as a VOC-limited region for O3 formation (Wang et al., 2019; Yang et al., 2021), the overestimation of NOX and underestimation of VOCs could contribute to the underestimation in O3 concentrations with air quality modeling. The simulations of both cases captured well the temporal variations of PM2.5 concentrations, with the R between the observed and simulated concentrations around 0.9.

In general, better modeling performance in the YRD region was found in case 2 than case 1. The NMBs between the simulated and observed concentrations of SO2, NO2, O3, and PM2.5 for the whole simulation period were -3.1 %, 56.3 %, -19.5 %, and -1.4 % for case 2, which were smaller in absolute value than those for case 1 at 8.2 %, 68.9 %, -24.6 %, and 7.6 %, respectively. The bootstrap sampling (Gleser et al., 1996; He et al., 2017) was further applied to test the significance of the improvements of case 2 over case 1. (A significant difference is demonstrated if the confidence intervals of given statistical indices sampled from the two cases do not overlap.) As can be seen in Table 2, the modeling performances of the concerned species in case 2 were improved significantly in most instances compared to case 1. For example, the improvement of NMB for the SO2 simulation was significant at the 99 % confidence level for July and October and 95 % for January. The improvement of NMB and NME for NO2 was significant at confidence levels of 99 % and 95 %, respectively, for April. The improvement of NMB for O3 was significant at the 95 % confidence level for January and that of PM2.5 at 95 % for April and 99 % for July. The statistical test confirms that incorporation of online monitoring data in the emission inventory can improve the regional air quality modeling for the YRD region. Besides the emission data, it should also be noted that the changes in model schemes would affect the model performance. For example, the newer version of CMAQ incorporated the chemistry schemes of bromine and iodine and was expected to influence the O3 simulation importantly. According to our recent test in the YRD region (Lu et al., 2020), the impact of CMAQ version on the simulation of difference species was inconclusive, implying the necessity of further intercomparison and evaluation studies for the region.

Figure 2 illustrates the spatial patterns of the simulated monthly SO2, NO2, O3 and PM2.5 concentrations for case 2. For a given species, similar patterns were found for different months. In general, the simulated concentrations of SO2, NO2, and PM2.5 were larger in central and northern Anhui, southern Jiangsu, Shanghai, and coastal areas in Zhejiang, where large power and industrial plants are concentrated, as shown in Fig. S2 in the Supplement. In the highly populated cities (Shanghai, Nanjing, Hangzhou, and Hefei; see their locations in Fig. 1), the simulated concentrations of pollutants were significantly larger than their surrounding areas. For example, the simulated SO2, NO2, and PM2.5 concentrations in Nanjing were 1.4, 1.3, and 1.2 times of those in its nearby cities. The analogous numbers for Hangzhou were 2.5, 1.5, and 1.3. In contrast, the simulated O3 concentrations were smaller in urban areas and larger in suburban ones. For instance, the simulated O3 in Nanjing, Shanghai, Hefei, and Hangzhou were 0.7, 0.4, 0.6, and 0.6 times of those in their surrounding areas, respectively. The spatial distributions of the simulated NO2 and O3 concentrations in Fig. 2 also indicated that O3 concentrations were less in the regions with higher NO2 concentrations, such as the megacity of Shanghai. The simulated high concentrations of NO2 in urban areas promotes titration of O3, reducing its concentrations. In addition, O3 concentrations could remain relatively high after transport from urban to the suburban areas due to relatively small emissions of NOX in the latter.

Table 3 summarizes the absolute and relative changes of the simulated monthly concentrations of the concerned air pollutants in cases 3–5 compared to the base case (case 2). The average contributions of the power sector to the total ambient concentrations of SO2, NO2, and PM2.5 for the four simulated months are estimated at 10.0 %, 4.7 %, and 2.3 %, respectively, based on comparison of cases 2 and 5. The contributions to the concentrations were close to those of emissions at 10.7 %, 6.6 %, and 1.6 % for the three species (as indicated in Table 1), respectively. The larger power sector contribution to the ambient PM2.5 concentrations than to primary PM emissions reflects high emissions of precursors of secondary sulfate and nitrate aerosols. In general, limited contributions from the power sector were found for all concerned species except SO2, which is attributed to the gradually improved controls in the sector. The further implementation of the ultra-low emission policy in the sector, therefore, is expected to result in limited additional benefits for air quality. As shown in Table 3, the absolute changes of the simulated SO2, NO2, O3, and PM2.5 concentrations in case 3 compared to case 2 were all smaller than 1 µg m-3 for the 4 months. Larger changes were found for primary pollutants (SO2 and NO2) than for those of secondary ones (O3 and PM2.5): the simulated monthly concentrations of SO2 and NO2 were 2.7 %–6.1 % and 2.0 %–2.9 % lower, while PM2.5 was only 0.1 %–1.3 % lower and O3 0.8 %–2.2 % higher, respectively. Much larger benefits were found when the ultra-low emission policy was broadened from the power sector to the industrial sector (case 4), which is attributed to the dominant role of industry in air pollutant emissions in the YRD region (Table 1). The simulated monthly concentrations of SO2, NO2 and PM2.5 were 1.5–2.0, 2.5–3.7, and 4.6–6.5 µg m-3 lower compared to the base case, respectively, or reduction rates of 32.9 %–64.1 %, 16.4 %–22.8 %, and 6.2 %–21.6 %. In contrast, the simulated O3 concentration was 0.8–4.8 µg m-3 higher, with growth rates ranging 2.6 %–14.0 %. As mentioned earlier, the YRD was identified as a VOC-limited region, and reducing NOX emissions without any VOC controls would enhance O3 concentrations. Currently, CEMSs do not report VOC concentrations in the flue gas, and the ultra-low emission policy does not include a VOC limit, either. In order to alleviate regional air pollution including O3, coordinated controls of NOX and VOC emissions are urgently required. These would include measures to reduce large sources of VOCs, notably in industries other than the power industry such as the chemicals and refining industry and in solvent use (Zhao et al., 2017).

The relative changes in the simulated pollutant concentrations varied by month, due to the combined influences of meteorology and secondary chemistry, and larger relative changes were found for SO2 and PM2.5 in summer. As shown in Table 3, for example, the average simulated PM2.5 concentrations in July were 0.4 and 6.5 µg m-3 lower, respectively, under cases 3 and 4 compared to case 2, with the larger reduction than other 3 months. This could result partly from the faster response of ambient concentrations to the changed emissions of air pollutants with shorter lifetimes in summer. The formation of secondary pollutants like PM2.5 would be enhanced in summer, with more oxidative atmospheric conditions under high temperature and strong sunlight. Moreover, the relatively low concentrations in summer also contributed to the largest percentage changes in SO2 and PM2.5 simulation for the season.

Figures 3 and 4 illustrate the spatial distributions of the relative changes of simulated pollutant concentrations in cases 3 and 4 compared to case 2, respectively. As shown in Fig. 3, the overall changes across the region due to ultra-low emission controls in the power sector only were less than 10 % for primary pollutants SO2 and NO2 and 5 % for secondary pollutants PM2.5 and O3. Larger changes in simulated SO2 concentrations were found in central and northern Anhui as well as central and southern Jiangsu, with relatively concentrated distribution of coal-fired power plants. The changes of simulated SO2 and NO2 in Shanghai were tiny, due to few remaining power plants subject to the ultra-low emission policy and thus few emission reductions. Compared to case 2, the SO2 and NOX emissions in case 3 were estimated to be 2.2 % and 0.8 % lower, respectively, for Shanghai, i.e., much smaller than for other provinces (6.1 % and 2.5 % for Anhui, 9.5 % and 4.4 % for Jiangsu, and 5.5 % and 2.7 % for Zhejiang). The results suggest that the potential of emission reduction and air quality improvement is limited from implementation of more stringent control measures in the power sector alone, particularly in highly developed cities where air pollution controls have already reached a relatively high level.

In case 4, where both power plants and selected industrial sources meet the ultra-low emission requirement, the average reduction rates of simulated SO2 and NO2 concentrations compared to case 2 were above 40 % and 25 %, respectively, for the whole region, and the changes of secondary pollutants O3 and PM2.5 were also significantly larger than those of case 3 in most of the region. The relative changes of SO2 were found to be more significant than other species, as the SO2 concentrations are greatly affected by primary emissions. Due to the large number and wide distribution of industrial plants throughout the YRD, moreover, there was little regional disparity in the changed ambient SO2 levels. Compared to other areas, the relatively less reduction in the simulated NO2 in central YRD resulted in significant enhancement of O3 concentrations (note that much more reduction in NO2 resulted in similar enhancement of O3 in southern Anhui for October). The comparison implies that the O3 formation in central YRD was more sensitive to NOX emission abatement than other VOC-limited regions in the YRD. The result suggests a particularly great challenge of O3 pollution control in central YRD, and more efforts on VOC emission abatement would be required for those developed areas.

Figure 5 illustrates the spatial distributions of PM2.5 concentrations for the base case (case 2) and the differences of cases 3 and 4 compared to the base case. The reduction of PM2.5 concentrations from the implementation of the ultra-low emission policy in the power sector was less than 1 µg m-3 over the YRD region (Fig. 5b). Larger reductions (above 0.4 µg m-3) were found in northern Anhui and northern and southern Jiangsu provinces, as those regions are the energy base of eastern China, with abundant coal mines and power plants with large installed capacities. With the policy expanded to certain industrial sectors, the simulated average PM2.5 concentrations were 5.8 µg m-3 lower for the whole region (Fig. 5c). In particular, the difference was greater than 10 µg m-3 along the Yangtze River, as there are many industrial parks located along the river containing a large number of big cement, iron and steel, and chemical industry plants. Stringent emission controls at those plants would result in significant benefits in air quality for local residents.

We further calculated the fractions of the population with different annual average PM2.5 exposure levels in cases 2–4, as shown in Fig. 6. Compared to case 2, slight differences in the population distribution by exposure level were found in case 3, while the differences were much more significant in case 4. The population fractions exposed to the average annual concentrations of PM2.5 smaller than 35, 35–45, and 45–55 µg m-3 were estimated to grow from 14 % in case 2 to 21 % in case 4, from 11 % to 16 %, and from 16 % to 30 %, respectively (note that 35 µg m-3 is the annual PM2.5 concentration limit in the current National Ambient Air Quality Standard for China). Accordingly, the fraction exposed to PM2.5 concentrations larger than 55 µg m-3 declined from 59 % to 33 %. The implementation of ultra-low emission policy on both power plants and industry sources thus proved an effective way in limiting the population exposed to high PM2.5 levels.

The mortality and YLL caused by atmospheric PM2.5 exposure with the base case emissions (case 2) in the YRD region are shown in Table 4. The values in brackets represent the 95 % confidence interval (CI) attributed to the uncertainty of IER curves (i.e., uncertainties from other sources were excluded in the 95 % CI estimation such as air quality model mechanisms, emission inventories, and population data). With the base case emissions, the NMB of the simulated and observed annual PM2.5 concentrations (based on the four representative months) was calculated at -1.4 % for the YRD region. Therefore, the influence of the biases between the simulations and observations on the estimated health risks was negligible and thus not considered in this study. The total attributable deaths due to all diseases caused by PM2.5 exposure in the YRD region were estimated at 194 000 (114 000–282 000), with STK, IHD, and COPD causing the most deaths, accounting for 29 %, 32 %, and 22 % of the total, respectively. With larger populations in Anhui and Jiangsu (32 % and 37 % of the regional total, respectively), more deaths caused by PM2.5 exposure were found in these two provinces, at 34 % and 41 % of the total deaths, respectively. Among all the diseases, STK was found to cause the largest number of mortalities (19 600) in Anhui with PM2.5 exposure, IHD in Jiangsu (31 300), and COPD in Shanghai (4400) and Zhejiang (10 800). The total YLL caused by PM2.5 exposure in the YRD region was 5.11 million years (3.16–7.18 million years). More YLL caused by PM2.5 exposure was found in Anhui and Jiangsu, accounting for 34 % and 37 % of the total in the YRD region, respectively. YLL values caused by COPD were the largest in all the provinces, with 0.66 million, 0.19 million, 0.56 million, and 0.47 million years estimate for Anhui, Shanghai, Jiangsu, and Zhejiang, respectively. The spatial distribution of attributable deaths and YLL caused by PM2.5 exposure was basically consistent with that of population in the YRD region, with correlation coefficients of 0.94 and 0.96, respectively. As shown in Fig. 7, higher health risks attributed to PM2.5 pollution in the base case (case 2) were commonly found in the areas with larger population densities, including the areas along the Yangtze River, central Shanghai and some urban areas in Anhui. We further compared the population deaths attributable to PM2.5 exposure calculated in this study with the reported total deaths in provincial statistical yearbooks (AHBS, 2016; JSBS, 2016; SHBS, 2016; ZJBS, 2016) and found that the deaths caused by PM2.5 exposure accounted for 18 %, 14 %, 15 %, and 11 % of the total deaths in Anhui, Jiangsu, Shanghai, and Zhejiang, respectively, for 2015. The numbers were larger than the estimate (6.9 %) by Maji et al. (2018), which focused on 161 cities in China. As one of the most developed and industrialized regions in China, the YRD suffered higher PM2.5 pollution level than the national average, leading to the larger fraction of premature death due to PM2.5 exposure. Moreover, the baseline disease-specific mortality rates applied in this study (from GHDx) were commonly higher than those in Maji et al. (2018) except for LRI, resulting in the larger estimate of death rates exposed to PM2.5.

Many studies have focused on the human health risks attributable to air pollution in China, with considerable disparities between them due to different estimation methods and health endpoints selected. Figure 8 compares the estimates of premature deaths caused by PM2.5 exposure in the YRD region in this and previous studies. Relatively close results are found between studies for the same regions and periods. For example, Hu et al. (2017) and Liu et al. (2016) estimated that the premature deaths of adults (> 30 years old) due to PM2.5 exposure were 223 000 and 245 000, respectively, in 2013 in the YRD region. However, the health endpoints in these two studies were not completely consistent. COPD, LC, IHD, and CEV (cerebrovascular disease) were selected in Hu et al. (2017), while COPD, LC, IHD, and STK were chosen by Liu et al. (2016). The deaths caused by PM2.5 exposure in Shanghai were estimated at 19 000, 15 000, and 16 000 in Maji et al. (2018), Song et al. (2017), and this study, respectively. The IER model and the same health endpoints were adopted in all three studies, while the PM2.5 concentrations were derived from ground observations in the former two studies instead of air quality simulation in this study. The premature deaths attributable to PM2.5 exposure in the YRD region in 2015 were estimated at 122 000 in Maji et al. (2018) and 194 000 in this study, respectively. Besides the different baseline mortality rates adopted in the two studies as mentioned earlier, the smaller estimate by Maji et al. (2018) could also result partly from inclusion of only typical cities instead of all cities in the YRD region. There are clear disparities in estimates of premature deaths for different years. For example, the death estimates caused by PM2.5 exposure in 2015 were generally smaller than those in 2013. As the population and age distributions remained relatively stable over the 2 years (AHBS, 2016; JSBS, 2016; SHBS, 2016; ZJBS, 2016), the reduced estimated premature deaths result to some extent from emission abatement and air quality improvement. According to relevant studies of Shanghai in particular (Lelieveld et al., 2013, 2015; Liu et al., 2016; Xie et al., 2016; Hu et al., 2017; Song et al., 2017; Maji et al., 2018), the premature deaths attributable to PM2.5 exposure increased from 2005 to 2013 and then declined afterwards, reflecting the health benefit of air pollution control measures in Shanghai in recent years.

Tables 5 and 6, respectively, summarize the avoided premature deaths and YLL by disease and region that would result from implementation of the ultra-low emission control policy and thereby reduced PM2.5 pollution in the YRD region. If only the coal-fired power sector met the ultra-low emission limits (case 3), nearly 305 premature deaths would be avoided compared to the base case emissions in 2015, with a tiny reduction rate of only 0.16 %. If the policy is strictly implemented for selected industrial sectors as well (case 4), 10 651 premature deaths could be avoided with a reduction rate at 5.50 %. The largest numbers of avoided premature deaths were found in Anhui and Jiangsu, accounting collectively for 88.2 % and 68.7 % of the total avoided deaths in cases 3 and 4, respectively. The greatest impacts from reduced PM2.5 concentrations were found for STK, of which the avoided deaths were calculated at 85 and 2848 in cases 3 and 4, respectively. The health effects of emission control policies in the YRD region have been investigated in previous studies. Using the IER model, Dai et al. (2019) chose the premature deaths from IHD, CEV, COPD, and LC as health endpoints and found that the Clean Air Action Plan would avoid 3439 deaths caused by PM2.5 exposure in Shanghai, which is more than those in both case 3 and case 4 in this study (5 and 1185, respectively). Applying environmental health risk and valuation methods, Li and Li (2018) found that 15 709 premature deaths attributable to air pollution could be avoided in 2015 if the PM2.5 concentrations in Jiangsu province were assumed to meet the National Ambient Air Quality Standard (GB3095-2012, 35 µg m-3 as the annual average). The estimate is much more than those calculated in case 3 and case 4 (177 and 4114 deaths, respectively). The larger health benefits estimated in those two studies result from their assumption of emission control measures covering a much wider range of sectors including energy, industry, transportation, construction, and agriculture, while only the ultra-low emission policy was assumed for the power and industry sectors in this study. The comparisons illustrate that the health benefits from emission control in the power sector alone is limited, and that controls in other sectors are essential. In addition, the different methods and inconsistent data sources partly led to the discrepancies. For the particle exposure estimation, as an example, Dai et al. (2019) adopted the BENMAP-CE model (Environmental Benefits Mapping and Analysis Program – Community Edition; Yang et al., 2013) to simulate the ambient PM2.5 concentrations, while Li and Li (2018) used the average of monitored PM2.5 concentrations. As shown in Table 6, the avoided YLL values for case 3 and case 4 were estimated at 8744 and 316 562 years, respectively, compared to the base case, confirming again the greatly improved health benefits from implementation of ultra-low emission policy for the industry sector in addition to the power sector. The largest avoided YLL values were found in Anhui and Jiangsu in the YRD region, accounting collectively for 86 % and 65 % of the total avoided YLL in cases 3 and 4, respectively. Compared to case 3, the fractions of Shanghai and Zhejiang to total YRD for both avoided deaths (Table 5) and YLL (Table 6) values were clearly higher in case 4, implying a greater health benefit of emission controls at industry sources in these relatively industrialized urban regions. The reduced PM2.5 concentrations led to the largest avoided YLL of COPD in both cases (3118 and 119 300 years in cases 3 and 4, respectively).

Figure 9 illustrates the spatial distributions of the avoided deaths and YLL from the ultra-low emission policy in the YRD region. When the policy was implemented only for coal-fired power plants, the health benefits were small and the regional differences relatively insignificant, with the avoided deaths and YLL smaller than 10 persons and 100 years, respectively, for all of the grid cells (Fig. 9a and b). When the policy was implemented both in power and industry sectors, more avoided deaths (> 40 person per grid cell) and YLL (> 400 years per grid cell) were found in northern Anhui, southern Jiangsu, central Shanghai, and northern Zhejiang (Fig. 9c and d). The spatial correlation coefficient between the avoided YLL in case 4 and population was 0.93, indicating that the implementation of the emission control policy would lead to greater health benefits for areas with intensive economic activity and dense populations.