We use 2013–2018 hourly data for surface air PM2.5 together with SO2, nitrogen dioxide (NO2), and CO concentrations from the CNEMC network (http://106.37.208.233:20035/, last access: 18 March 2019). The network started in January 2013, with 496 sites in 74 major cities across the country (Chinese State Council, 2013b), growing to ∼1500 sites in 454 cities by 2018. PM2.5 mass concentrations are measured using the micro-oscillating balance method and/or the β-absorption method (MEE, 2012; Zhang and Cao, 2015). SO2, NO2, and CO concentrations are measured at the same sites as PM2.5. NO2 concentrations are measured by the molybdenum converter method known to have positive interferences from NO2 oxidation products (Dunlea et al., 2007). SO2 and CO are respectively measured using ultraviolet fluorescence and infrared absorption (MEE, 2012; Zhang and Cao, 2015). We applied quality control to the hourly CNEMC data following Barrero et al. (2015) to exclude severe outliers (Lu et al., 2018). There are also occasional consecutive repeats of data that may be caused by faulty instruments or reporting (Rohde and Muller, 2015; Silver et al., 2018). Here we removed values from the hourly time series when there are >24 consecutive repeats. These in whole removed 7.4 %, 7.0 %, 6.4 %, and 6.7 % of the PM2.5, SO2, NO2, and CO data respectively.

We correlated these air quality observations with meteorological observations from 839 stations distributed across China (Fig. S1 in the Supplement). The meteorological observations are compiled in the Surface Daily Climate Dataset (V3.0) released by the China National Meteorological Information Center (CNMIC; http://data.cma.cn/, last access: 16 March 2019). These include data for wind speed (WDS), precipitation (PRECIP), relative humidity (RH), and temperature (TEM). We also used the 850 hPa meridional wind velocity (V850) from the MERRA-2 reanalysis produced at 0.5∘ × 0.625∘ horizontal resolution by the NASA Global Modeling and Assimilation Office (https://gmao.gsfc.nasa.gov/reanalysis/MERRA-2, last access: 20 March 2019). We choose these meteorological variables for their strong correlations with PM2.5 identified in previous studies (Wang et al., 2014; Cai et al., 2017; Shen et al., 2017; Leung et al., 2018; Song et al., 2019; Zou et al., 2017). V850 in particular is a strong predictor of PM2.5 wintertime pollution events in the North China Plain because northerly winds (negative V850) ventilate the region with clean dry air (Cai et al., 2017; Pendergrass et al., 2019).

All data in this work are averaged over 10 d (10 d time resolution). Trend analyses use only those sites with at least 70 % data coverage for each of the 6 years from 2013 to 2018. We did sensitivity tests with data coverage thresholds changing from 70 % to 90 % and obtained similar pollutant trends. To make the most use of available data, 70 % is chosen. For the MLR model, we further average all data on a 2∘ × 2.5∘ grid to increase statistical robustness, following Tai et al. (2012) and Shen et al. (2017).

The 2013 Clean Air Action (Chinese State Council, 2013a) identified three megacity clusters as target regions for reducing air pollution: Beijing–Tianjin–Hebei (BTH; 35–41∘ N, 113.75–118.75∘ E), the Yangtze River Delta (YRD; 29–33∘ N, 118.75–123∘ E), and the Pearl River Delta (PRD; 21–25∘ N, 111.25–116.25∘ E). The more recent plan released in July 2018 (Chinese State Council, 2018b) removed PRD from the list of target regions and added the Fenwei Plain (FWP; 33–35∘ N, 106.25–111.25∘ E; 35–37∘ N, 108.75–113.75∘ E). Previous studies (Zhang et al., 2012) also identified the Sichuan Basin (SCB; 27–33∘ N, 103.75–108.75∘ E) as one of the major haze regions in China. We present analyses for these five target regions by averaging the data from all sites with more than 70 % data coverage for each of the 6 years from 2013 to 2018. The only continuous record for 2013–2018 in the FWP region is for Xi'an (13 sites). Additional FWP sites outside Xi'an started operating in early 2015 and are consistent with the Xi'an data, as will be shown below.

We construct a stepwise multiple linear regression (MLR) model to quantify the effect of meteorology on PM2.5 variability. The model fits the deseasonalized and detrended 10 d PM2.5 mean time series on the 2∘ × 2.5∘ grid to the five deseasonalized and detrended 10 d mean meteorological variables (WDS, PRECIP, RH, TEM, and V850). The deseasonalized and detrended time series are obtained by removing the 50 d moving averages from the 10 d mean time series (Tai et al., 2010). This focuses on synoptic scales of variability and avoids aliasing from common seasonal variations and long-term trends between variables (Shen et al., 2017).

Separate fits of PM2.5 to the meteorological variables are done for each 2∘× 2.5∘ grid square and season (DJF, MAM, JJA, and SON). The fit has the form 1Yd,i(t)=∑k=15βi,kXd,i,k(t)+bi, where Yd,i(t) is the deseasonalized and detrended PM2.5 time series for grid square and season i, and Xd,i,k(t) is the corresponding time series for the deseasonalized and detrended meteorological variable kϵ[1,5]. We fit the regression coefficients βi,k and the intercept bi. The regression is done stepwise, adding and deleting terms based on their independent statistical significance to obtain the best model fit (Draper and Smith, 1998). The fits and the selected meteorological variables differ by location and season, but with regional consistency (Table S1 in the Supplement). For meteorological variables not in the final MLR model, the regression coefficients βi,k in Eq. (1) are zero.

We use the MLR model to remove the effect of meteorological variability from the 2013–2018 PM2.5 trends, including not only the 10 d synoptic-scale variability but also any interannual variability and 6-year trends. This makes the standard assumption that the same factors that drive synoptic-scale variability also drive interannual variability (Jacob and Winner, 2009; Tai et al., 2012). We thus apply Eq. (1) to the meteorological anomalies Xa,i,k, obtained by removing the 6-year means of the 50 d moving averages from the 10 d mean time series. The anomalies calculated in this manner are deseasonalized but not detrended. This yields the meteorology-driven PM2.5 anomalies Ym,i: 2Ym,i(t)=∑k=15βi,kXa,i,k(t)+bi. Consider now the PM2.5 anomaly Ya,i for grid square and season i obtained by deseasonalizing but not detrending the PM2.5 data (by removing the 6-year means of the 50 d moving averages) in the same way as for the meteorological variables. The residual anomaly Yr,i, after removing meteorological influence from the MLR model, is given by 3Yr,it=Ya,i(t)-Ym,i(t). The residual is the component of the anomaly that cannot be explained by the MLR meteorological model, and we will refer to it as the meteorology-corrected data. It includes noise due to limitations of the MLR model and other factors but also a long-term trend over the 6-year period that we can attribute to changes in anthropogenic emissions. The same approach was recently applied by Li et al. (2019) to separate anthropogenic and meteorological drivers of ozone trends in China.

Figure 1 shows annual mean observed PM2.5 concentrations from the CNEMC over China for 2013 and 2018 and the linear regression trends on the 2∘ × 2.5∘ grid based on the PM2.5 anomalies Ya,i(t), including effects of both changing emissions and meteorology. In 2013, PM2.5 across most of China was well above the Chinese national air quality standard (annual mean of 35 µg m-3). BTH and FWP (Xi'an) had the highest PM2.5 among the five target regions, with annual average concentrations of 108±34 µg m-3 (standard deviation describes variability in the annual average across sites in the region) and 108±11 µg m-3 respectively, followed by SCB (71±17 µg m-3), YRD (67±12 µg m-3), and PRD (47±7 µg m-3). PM2.5 decreased dramatically from 2013 to 2018, by 34 %–49 % for the five target regions. Mean 2018 concentrations were 55±13 µg m-3 in BTH, 62±4 µg m-3 in FWP (Xi'an), 40±6 µg m-3 in SCB, 40±7 µg m-3 in YRD, and 31±5 µg m-3 in PRD.

Figure 2 shows the 2013–2018 relative trends of annual mean PM2.5 for the five target regions along with the corresponding trends of SO2, NO2, and CO concentrations measured at the same sites. Also shown in the bottom panels are the MEIC trends in emissions of primary PM2.5, SO2, NOx, NH3, and CO for 2013–2017. The PM2.5 observations show steady decreases for BTH, YRD, and SCB. PRD flattens out in 2015–2017 before decreasing again in 2018. FWP (Xi'an) decreases sharply by 47 % from 2013 to 2015 but rebounds in 2015–2017 before decreasing again in 2018. Trends at other FWP sites that became operational in early 2015 are similar to Xi'an. We argue in Sect. 3.3 that the 2015–2017 flattening at PRD and the anomalous 2013–2015 sharp decrease and 2015–2017 rebound at FWP are driven by meteorology.

We see from Fig. 2 that only SO2 has a decrease steeper than PM2.5, indicating that SO2 emission controls have been a major driver of the PM2.5 trend (Lang et al., 2017; Shao et al., 2018). The overall SO2 decrease for the five regions is 57 %–76 % from 2013 to 2018. The SO2 decrease is quantitatively consistent with the decrease in SO2 emissions estimated by MEIC (Zheng et al., 2018). This drastic cut of China SO2 emissions is due to installation of scrubbers at coal-fired power plants (Siwen et al., 2015; Karplus et al., 2018; Silver et al., 2018), elimination of small coal boilers, improvement of coal quality (Zheng et al., 2018), and switch from residential coal to cleaner fuels (Zhao et al., 2018). We also see a significant decrease in CO of 18 %–43 % for the five regions from 2013 to 2018, again consistent with the MEIC and suggesting a reduction in organic PM2.5 emissions. Primary PM2.5 emissions in the MEIC decreased at a rate comparable to or steeper than CO. Trends in China PM2.5, SO2, and NO2 presented here are consistent with previous studies (Silver et al., 2018; Ma et al., 2019) that cover a shorter time period than 2013–2018.

Figure 3 shows the time series of monthly mean PM2.5 for the five target regions, illustrating the seasonal and interannual variability. All regions show winter maxima that can be mostly attributed to meteorology, including shallower mixing depth, lower precipitation, and increased stagnation in winter (Wang et al., 2018). Residential heating emissions in winter also contribute to the seasonality in China, north of about 33∘ N (covering BTH and FWP in this study; J. Liu et al., 2016; Xiao et al., 2015). There is a large interannual variability, particularly in winter, that must be largely driven by meteorology. Studies for BTH have shown that high PM2.5 in winter months is associated with weak southerly winds, low mixing depths, and high relative humidity (Zhang et al., 2014; Chang et al., 2016; Li et al., 2018; Shao et al., 2018). The relatively clean 2017–2018 winter was due in part to a higher frequency of northerly flow and associated ventilation (CMA, 2018; Yi et al., 2019). In addition, particularly aggressive actions by the government to restrict coal use that winter may have played a role in reducing PM2.5 levels (Zhang et al., 2019).

Figure 4 shows the correlations of 10 d PM2.5 concentrations with the individual meteorological variables used in the MLR model. Correlation coefficients r as low as 0.3 are statistically significant, more so when consistent across a region. Wind speed is negatively correlated with PM2.5, as would be expected from ventilation, except in areas of the north where wind promotes dust formation (Lyu et al., 2017; Wang et al., 2004). Precipitation is also generally negatively correlated with PM2.5, as one would expect from scavenging (Chen et al., 2018). The positive correlation between precipitation and PM2.5 over northern China in spring is likely a result of high RH associated with precipitation in adjacent days.

Correlation between RH and PM2.5 is positive over northern China, especially in winter, and negative over southern China, especially in summer. The positive correlation between PM2.5 and RH over northern China in winter has been reported by previous studies and attributed in part to the role of aqueous-phase aerosol chemistry in driving secondary PM2.5 formation (Zheng et al., 2015; He et al., 2018; Song et al., 2019; Pendergrass et al., 2019; Tie et al., 2017). The negative correlation of PM2.5 with RH over southern China likely reflects the association of high RH with precipitation and onshore wind, which facilitate PM2.5 wet removal and ventilation (Zhu et al., 2012; Leung et al., 2018).

Temperature has a positive correlation with PM2.5 year-round over most of China (Wang et al., 2014; Leung et al., 2018), even though there is no strong direct dependence of PM2.5 on temperature (Jacob and Winner, 2009). The correlation likely reflects the covariation of temperature with other meteorological variables, including wind speed, precipitation, and RH (Tai et al., 2012; Zhu et al., 2012). A possible explanation for the negative correlation with temperature in summer over the North China Plain could be the volatilization of ammonium nitrate at high temperatures (Kleeman, 2008). V850 shows strong positive correlations with winter PM2.5 over most of China, and strong negative correlations with summer PM2.5 over southern China, especially for the Pearl River Delta.

Figure 5a describes the ability of the MLR model to account for PM2.5 variability in relation to wind speed, precipitation, RH, temperature, and V850 as potential predictor variables. Results are presented as the coefficients of determination R2 (fraction of variance explained) between observed and model PM2.5 in the detrended deseasonalized time series. The R2 values have been adjusted to account for different numbers of significant explanatory terms (predictor variables). R2 values for the five target regions are 0.59 (BTH), 0.46 (YRD), 0.65 (PRD), 0.65 (SCD), and 0.41 (FWP). Figure 5b shows the meteorology-corrected PM2.5 trends after removal of meteorological variability predicted by the MLR model, i.e., the trends in the residuals Yr,i(t) in Eq. (3). The meteorology-corrected decreasing trend averaged across China is -4.6 µg m-3 a-1, 12 % weaker than in the original data, meaning that 12 % of the PM2.5 decrease in the original data is attributable to meteorology. We elaborate on this below for the five target regions.

Figure 6 shows the 10 d mean PM2.5 anomalies in the deseasonalized (but not detrended) data for the five target regions (Ya(t) in Sect. 2.3). Also shown is the meteorological component Ym(t) derived from the MLR meteorological model and the residual Yr(t) (meteorology-corrected; Eq. 3) whose long-term trend can be interpreted as being due to changes in anthropogenic emissions. The PM2.5 anomalies show large features on 10 d basis that can be mostly captured by the MLR model. The residual meteorology-corrected time series is much smoother, as depicted by the narrower 95 % confidence intervals in the anthropogenic residual trends than in the original observed trends. The meteorology-corrected trends differ by 3 % (YRD) to 27 % (SCB) from the observed trends. The YRD trend reflects a significant contribution from the December 2013 outlier, which reflects unfavorable meteorological conditions (Fig. S2) that are not adequately captured by the MLR model. If we exclude this outlier month from the time series, the observed YRD trend becomes -5.7±0.9 µg m-3 a-1 and the meteorology-corrected trend becomes -5.9±0.7 µg m-3 a-1.

Most remarkably, it appears that the 2015–2017 flattening in the PRD and 2015–2017 increase in the FWP (see Fig. 2) can be mostly attributed to meteorological variability as resolved by the MLR model rather than to emissions. The trend in the residual is more consistent with a steady 2013–2018 anthropogenic decrease in both regions. The MLR model shows that meteorology accelerated the PM2.5 decline in PRD and FWP from 2013 to 2015 and contributed partly to the 2015–2017 PM2.5 rebound over FWP. In particular, the high PM2.5 anomalies in PRD in 2013 and early 2014 are driven by anomalously low V850, and the low PM2.5 in winter 2015–2016 is associated with anomalously high southerly flow and precipitation (Fig. S4). The low PM2.5 in FWP in the winter 2014–2015 is associated with anomalously high wind speed, low RH, and low temperature, while the high anomalies in the winter 2016–2017 are associated with anomalously low wind speed, high RH, and high temperature (Fig. S5).