Anthropogenic PM2.5 (fine particulate matter with an aerodynamic diameter smaller than 2.5 µm) is known as a robust indicator of mortality and other negative health effects associated with ambient air pollution. PM2.5 components originate not only from primary emissions but also from secondary formations through various precursors (e.g., SO2, NOx and volatile organic compounds – VOCs). Regional haze with extremely high PM2.5 concentrations (exceeding the WHO standard tenfold) has become the primary air quality concern in China, especially over northern China (e.g., L. T. Wang et al., 2014; W. Wang et al., 2014; Han et al., 2015; Sun et al., 2015). To control PM2.5 pollution and improve the overall air quality, a series of strict pollution control strategies have been implemented by the government since 2010, including the “Guiding Options on Promoting the Joint Prevention and Control of Air Pollution to Improve Regional Air Quality” (The Central Government of the People's Republic of China, 2010) and the Atmospheric Pollution Prevention and Control Action Plan (The Central Government of the People's Republic of China, 2013), in which the government stated that environmentally related equipment (for flue-gas desulfurization, selective catalyst reduction, exhaust dust removal, etc.) are mandatory for both industries and vehicles. In addition to long-term pollution control strategies, different emergency measures under different pollution alerts were also implemented occasionally. For example, the production of large industrial sources (coal burning and cement) was limited to reduce emissions, construction sites were restricted to prevent fugitive dust pollution, and traffic restrictions were implemented on even- and odd-numbered license plates. These emission control strategies were stricter and implemented more often in northern China than anywhere else in winter, when haze events occur more frequently. These control strategies were expected to reduce both the concentrations of significant precursors (e.g., SO2 and NOx) and the emissions of PM2.5.

Despite these strict emission control strategies, the ambient PM2.5 concentrations in major cities still fluctuated during the wintertime from year to year. For example, the overall January PM2.5 concentrations in 74 cities generally decreased from 2015 to 2016, but the concentrations in January 2017 were still higher than those in 2016 (China National Environmental Monitoring Center “Ambient Air Quality Monthly Report 2015-01/2016-01/2017-01”, http://www.cnemc.cn/jcbg/kqzlzkbg/, last access: 7 May 2019). While annual emission reduction trends were expected from 2015 to 2017, the overall increase in the surface concentrations observed in January 2017 contradicted these expectations, thereby indicating that other factors (especially meteorological conditions) in addition to emissions may play important roles. Some studies have attempted to investigate the variability of air pollution and the effects of climate change on wintertime air pollution by using statistical data. Li et al. (2016) indicated that wintertime fog–haze days across central and eastern China have a close relationship with the East Asian winter monsoon. Zuo et al. (2015) concluded that the significant weakening and strengthening of the Siberian high and East Asian trough, respectively, are the two main factors for the occurrence of extreme warm and extreme cold events over China in winter, when warm air boosts air pollution. In addition to utilizing statistical methodology, it is necessary to distinguish the roles of emissions and meteorology to further investigate the driving factors of interannual air pollution changes.

Regional air quality models are important tools for scientifically understanding the formation of haze events, technically constructing forecasts and evaluating the effects of control strategies. For regional modeling studies, emission inventories are important for reflecting the emission inputs into the atmosphere. Generally, an emission inventory is based on a “bottom-up” methodology, thereby relying on the statistics of energy activity, emission factors, etc. However, uncertainties in energy statistics can cause variations in the emission estimates (Zhao et al., 2017; Hong et al., 2017; Zhi et al., 2017). For regional modeling applications, the total emissions based on statistics are spatially and temporally distributed according to relevant factors (He, 2012). Nevertheless, the occasional emission control strategies implemented in winter can cause large uncertainties in not only total emission estimations but also in spatial and temporal allocations, which leads to large biases in the model simulations.

In addition to the uncertainties in emission inventories, deficiencies in the model chemistry can also cause model uncertainties. Increasing numbers of observations have revealed that anthropogenic emission-relevant aerosol species, such as sulfate, nitrate and ammonium (denoted as SNA), are the predominant inorganic species in the wintertime PM2.5 in China (Y. S. Wang et al., 2014; Yang et al., 2015). Various reaction paths during haze events have also been proposed (e.g., Zheng et al., 2015; Cheng et al., 2016; Wang et al., 2016; Li et al., 2017; Moch et al., 2018; Wang et al., 2018; Shao et al., 2019). For example, Moch et al. (2018) used a 1-D model and revealed the importance of aqueous-phase chemistry of HCHO and S(IV) in cloud droplets by forming a S(IV)-HCHO adduct, hydroxymethane sulfonate. Shao et al. (2019) implemented four heterogeneous sulfate formation mechanisms (via H2O2, O3, NO2 and transition metal ions on aerosols) into GEOS-Chem model, which partially reduced the modeled low bias in sulfate concentrations. However, a scientific consensus regarding the importance of the reaction paths has not yet been reached, partially due to the uncertainties in aerosol liquid water content, pH, ionic strength, etc. The original WRF/Chem model with either the Goddard Chemistry Aerosol Radiation and Transport (GOCART; Chin et al., 2000, 2002) or the Model for Simulating Aerosol Interactions and Chemistry (MOSAIC) 4-bin (MOSAIC-4BIN) aerosol scheme failed to reproduce the highest PM2.5 concentrations; it is assumed that this failure is due to missing heterogeneous and aqueous reactions. In Chen et al. (2016; hereafter Chen16), we added three heterogeneous reactions (SO2-to-H2SO4 and NO2/NO3-to-HNO3 reactions) to the WRF/Chem model based on the MOSAIC-4BIN aerosol scheme. Although the reaction paths may still not be comprehensively understood, the new MOSAIC-4BIN aerosol scheme significantly improved the simulation of sulfate, nitrate and ammonium on polluted days in terms of the concentrations of those species and their partitioning.

Data assimilation (DA), that is, the combination of observations with numerical model output, has been proven to be skillful at improving aerosol forecasts (e.g., Collins et al., 2001; Pagowski et al., 2010; Liu et al., 2011, 2016; Zhang et al., 2016). Liu et al. (2011; hereafter Liu11) implemented DA on aerosol optical depth (AOD) estimates within the National Centers for Environmental Prediction (NCEP) Gridpoint Statistical Interpolation (GSI) three-dimensional variational (3DVAR) DA system coupled with the GOCART aerosol scheme within the Weather Research and Forecasting with Chemistry (WRF/Chem) model (Grell et al., 2005). Schwartz et al. (2012; hereafter S12) and Jiang et al. (2013; hereafter Jiang13) extended the above system to assimilate surface PM2.5 and PM10. The evaluation results demonstrated improved aerosol forecasts from the DA system in studies over East Asia and the United States.

Following Liu11, S12 and Chen16, we updated the GSI–WRF/Chem system by changing from the GOCART aerosol scheme to the MOSAIC-4BIN aerosol scheme to better characterize the complex PM2.5 pollution in China. We applied the updated system to assimilate PM2.5 concentrations of January 2015, 2016 and 2017 for two purposes: (1) to reproduce the PM2.5 output by the DA system and (2) to investigate the different impacts of meteorological conditions and emissions on the PM2.5 pollution in different years. In this paper, Sect. 2 provides descriptions of the model, observations and methodology and addresses the updated GSI–WRF/Chem-coupled DA system with the MOSAIC-4BIN aerosol scheme. In Sect. 3, the assimilation results for the PM2.5 concentrations from January 2015, 2016 and 2017 are presented and compared with surface observations (PM2.5 total mass) to evaluate the DA system. In contrast to previous applications emphasizing the forecast skill improvement achieved by the DA system, we fully utilized reanalysis data to investigate the driving factors of pollution and to differentiate the roles played by meteorological conditions and emissions in different years by analyzing the reanalysis data and model simulations. The results are given in Sect. 4, and the conclusions are given in Sect. 5.

The WRF/Chem settings are very similar to those of Chen16, although Chen16 focused on the SNA aerosols in the North China Plain during October 2014; in addition, several heterogeneous reactions were newly added to the original chemistry modules to improve the SNA simulation performance. The DA system used herein was based on the NCEP GSI system extended by Liu11 and S12. We assimilated surface PM2.5 observations, and the only difference is that the MOSAIC-4BIN aerosol scheme (32 PM species) was chosen for the WRF/Chem model instead of the GOCART aerosol scheme. Thus, the 3-D mass mixing ratios of those MOSAIC species at each grid point composed the analysis (or control) variables in the GSI 3DVAR minimization process.

Here, only a brief summary of the WRF/Chem configuration is provided below, prior to a description of the updated GSI DA system and the settings used in this work. The most important differences are noted, e.g., the forward operator for observations in the GSI system.

This section presents the results from the NO_DA and assimilation experiments outlined above. In slight contrast to S12 and Jiang13, our purpose was to reproduce the spatio-temporal variations in the surface PM2.5 within the reanalysis dataset rather than to provide the inorganic carbon (IC) of aerosol fields for improving forecasts.

Figure 3 shows the observed and modeled monthly averages of the surface PM2.5 for January 2015, 2016 and 2017. Nine regions are illustrated as rectangles in the figure: the North China Plain (NCP), northeastern China (NEC), the Energy Golden Triangle (EGT), Xinjiang (XJ), the Fenwei Plain (FWP), the Sichuan Basin (SB), central China (CC), the Yangtze River Delta (YRD) and the Pearl River Delta (PRD). Both the observations and the model show that high values are mostly observed in NCP, FWP, SB and CC. In the NO_DA case, the model results are overpredicted in SB, NCP and CC for all 3 months, while the overestimations are more severe in SB. The NO_DA case generally overestimates (underestimates) the surface PM2.5 in NCP, SB and CC (XJ and FWP) in the 3 years, potentially indicating that the 2010 emissions are not appropriate for the 2015–2017 simulations with overestimations (underestimations). As discussed in Sect. 2.1, the large area of overestimation is consistent with the national reductions in SO2, NOx and PM2.5 anthropogenic emissions (Zheng et al., 2018); however, the underestimations in XJ and FWP also indicate the introduction of new emission sources to these two regions.

Compared to the NO_DA case, the CONC_DA assimilation experiment effectively reproduces the spatial distribution of surface PM2.5 for the 3 months in terms of the relatively higher values observed in NCP, SB and CC and in some “hotspots” (NEC, FWP and XJ), which are closer to the observations.

Basic statistical measures, including the bias (BIAS), standard deviation (SD), root-mean-square error (RMSE) and correlation coefficient (CORR), were applied to evaluate the experiments. Figure 4 shows the time series of the BIAS, SD and RMSE for all the data used in the entire domain. The statistics were calculated for each 1 h DA cycle. After quality control, the number of PM2.5 observations used in the DA process differed; the number of observations was normally approximately 500–520 but reached a minimum of 320–450 occasionally due to the data availability. From the time series, we can see that the BIAS, SD and RMSE are greatly improved in the CONC_DA case. The maximum BIAS values are approximately 50 µg m-3 for January 2015 and approximately 80 µg m-3 for 2016 and 2017 in NO_DA, while they are reduced to approximately ±5 µg -3 in CONC_DA. The SD and RMSE are also reduced by at least 50 % most of the time.

Figure 5 shows the spatial distributions of the error statistics (BIAS, RMSE and CORR) at each observational site (with more than two-thirds valid data in the month) in January 2015, 2016 and 2017. We start with 2015 and then address the differences with comparisons in 2016 and 2017. In 2015 in the NO_DA case, the surface PM2.5 concentrations are generally overestimated by 20–60 µg m-3 in eastern China (NCP, SB, CC, PRD and YRD) but are underestimated in NEC, FWP, EGT and especially XJ. The high and low BIAS values in eastern and western China are greatly corrected in CONC_DA. Consistent with the BIAS changes in CONC_DA, the RMSE and CORR distributions in eastern China and NEC are also greatly improved; the RMSE is reduced by at least 50 %, and the CORR increases to almost above 0.8–0.9. The inhomogeneous distributions of the BIAS in NO_DA in 2016 and 2017 are very similar to those in 2015 (overestimated in eastern China but underestimated in NEC, EGT and XJ). However, the high biases in CC and PRD and the low biases in XJ are even larger in 2016 and 2017. Similar to the comparisons between NO_ DA and CONC_DA for the year 2015, improvements are generally achieved for almost all the regions in both 2016 and 2017. The statistics for the nine regions are listed in Table 3.

Given reliable PM2.5 reanalysis fields produced by assimilating surface PM2.5 (CONC_DA), the changes in the 3 years can be analyzed for not only scattered observational sites but also for different regions. To distinguish the roles of meteorological conditions and emissions in driving these changes, an analysis based on the NO_DA and CONC_DA simulations is performed. As assumed in Sect. 2.4, meteorology-driven changes can be analyzed in the NO_DA simulations with different meteorological conditions but the same emission inventory for different years; however, the changes in the reanalysis data in different years are actually the combination of all the driving forces, including meteorological conditions and emissions. By analyzing both sets of simulations, we can attempt to distinguish the roles of meteorology and emissions in determining these changes.

To analyze the complex PM2.5 pollution in China, the GSI–WRF/Chem aerosol data assimilation system was updated from the GOCART aerosol scheme to the MOSAIC-4BIN scheme, which is more appropriate for characterizing anthropogenic emission-relevant aerosol species. Three years (2015–2017) of wintertime (January) surface PM2.5 observations from more than 1600 sites were assimilated hourly using the updated 3DVAR system in the CONC_DA assimilation experiment. A parallel control experiment that did not employ DA (NO_DA) was also performed.

Both the control and the assimilation experiments were evaluated against the surface PM2.5 observations. In the NO_DA experiment, in which the 2010_MEIC emission inventory was used, the modeled PM2.5 concentrations were severely overestimated in the Sichuan Basin (SB), central China (CC), the Yangtze River Delta (YRD) and the Pearl River Delta (PRD) by 98–134, 46–101, 32–59 and 19–60 µg m-3, respectively, which indicated that the emission estimates for 2010 are not appropriate for 2015–2017, as strict emission control strategies were implemented in recent years. Meanwhile, underestimations of 11–12, 53–96, and 22–40 µg m-3 were observed in NEC, XJ and EGT, respectively. The assimilation experiment significantly reduced the high biases of surface PM2.5 in SB, CC, YRD and PRD and the low biases in NEC and XJ with biases within ±5 µg m-3.

Both the observation and the assimilation experiments showed decreasing ambient concentrations from 2015 to 2016 but increasing concentrations from 2016 to 2017 for most of the regions. To distinguish the important factors driving these changes, the reanalysis data from the assimilation experiment and the modeling results from the control experiment were analyzed. The results showed a reduction in PM2.5 of approximately 15–20 µg m-3 for the month of January from 2015 to 2016 in northern China (NCP and NEC), but meteorology played the dominant role (contributing approximately 12–21 µg m-3 of the PM2.5 reduction). The changes from 2016 to 2017 in NCP and NEC were different; meteorological conditions caused an increase in PM2.5 of approximately 12–23 µg m-3, while emission control measures caused a decrease of 1–8 µg m-3, and the combined effects still showed a PM2.5 increase for that region. The analysis confirmed that meteorology played different roles in 2016 and 2017: the higher pressure system, lower temperatures and higher PBLH in 2016 (compared with 2015) were favorable for pollution dispersion, whereas the situation was almost the opposite in 2017 (compared with 2016) and led to an increased PM2.5 from 2016 to 2017, although emission control strategies were implemented in both years. After considering the impacts of the meteorology, the analysis showed that emissions were indeed reduced from 2015 to 2016 and 2017, especially in NCP for the year 2017 (although the surface concentrations increased that year). The analysis also showed that emissions increased in XJ and FWP.

There are still large uncertainties in this approach, such as the deficiencies of forward models (including inaccurate emission inputs, uncertainties in the meteorological IC/BC and the chemistry mechanism) and the assimilation process, and the imperfection of the aerosol–meteorology feedbacks in the model simulation generated large biases in the analysis. The most straightforward approach is thus to directly estimate the emissions by data assimilation, which will be the topic of a separate study.