Developed by the China Meteorological Administration (CMA), the CUACE model is used in this study to simulate air quality in Beijing. CUACE is a unified chemical weather numerical forecasting system which is independent of weather and climate models. It consists of four functional blocks: anthropogenic and natural emissions, atmospheric gaseous chemical mechanisms, atmospheric aerosol chemical mechanisms, and a numerical assimilation system. The gaseous chemical block is based on the Regional Acid Deposition Model (RADM) covering 66 gaseous species (Stockwell et al., 1990; Wang et al., 2015). The aerosol module includes a mixing scheme, clear-sky processes, dry deposition, below-cloud scavenging, and in-cloud processes. Seven aerosol species, i.e. sulfates (SF), soil dust, BC, organic carbon (OC), sea salts, nitrates (NI), and ammonium salts (AM), are considered in aerosol chemical module. The first six aerosol components were divided into 12 bins with a diameter ranging from 0.01 to 40.96 µm. Based on the mixing assumptions, the ambient size and density of aerosols in a size bin are evaluated. The optical properties of these aerosols are readily computed when the mixing state, composition, and ambient size are determined. The details of sulphur chemistry, cloud chemistry, coagulation, nucleation, condensation, etc. were depicted by Gong et al. (2003). CUACE is online coupled to the fifth-generation Penn State/NCAR mesoscale model (MM5) and Global/Regional Assimilation and PreDiction System (GRAPSE); MM5 is selected to simulate mesoscale meteorological fields in this study. For different research target and application purposes, CUACE is designed with an open interface to allow it to be easily integrated into different time and spatial scale models. A more detailed description can refer to Gong et al. (2009). The performance of CUACE was evaluated by many researchers. Wang et al. (2010) simulated dust weather occurring in April 2006 and indicated that CUACE could predict the outbreak, development, transport, and depletion processes of sand and dust storms accurately over China and the East Asian region. Li et al. (2014) evaluated air quality prediction by CUACE over Ürümqi and acquired a quite accurate forecasting of air quality levels, especially for NO2 and PM10 levels. Given the good performance in air quality prediction, CUACE has been used for haze forecasting at the National Meteorological Center of CMA and some local environmental protection agencies.

In this study, the MM5–CUACE model is configured to have three nested domains to reduce spurious boundary effects in the inner domain: a horizontal resolution of 27 km covering North China and the surrounding areas, a resolution of 9 km covering Jing–Jin–Ji (Beijing, Tianjin and Hebei), and a resolution of 3 km covering Beijing and surrounding areas (Fig. 1). In the vertical, there are a total of 35 full eta levels extending to the model top at 10 hPa, with 16 levels below 2 km.

Two periods: July and December in 2013 are selected for model integration to evaluate different seasonal impacts (summer and winter respectively) of vehicle emission on air quality. The time steps of the MM5 and CUACE models are 15 and 150 s respectively. The driving field provides the initial, lateral, and surface boundary conditions and transmits the weather background information to MM5. However, for large domain or long-term simulations, the large-scale weather situation simulated by MM5 may diverge from that of the driving field. The methods to constrain MM5 to the driving field involve frequent re-initialization, analysis nudging, spectral nudging, and scale-selective bias correction (Bowden et al., 2013). A 36 h re-initialization run is executed to simulate meteorological conditions and air quality, and the former 12 h simulation is discarded as spin-up time, which is the same as Zhang et al. (2012). The initial and boundary meteorological conditions are from T639 reanalysis data with 30 × 30 km spatial resolution and 6 h temporal resolution supplied by CMA (Xiao et al., 2010). The initial and boundary chemical conditions of the first simulation segment are based on averages from several field studies over the eastern Pacific Ocean (McKeen et al., 2002), which were used as the default profiles in WRF-Chem, and other segment initial and boundary conditions are derived from previous simulation segment. The extra 10-day run (i.e. 21 to 30 June, 21 to 30 November) was conducted to reduce the effect of chemical initial and boundary conditions.

Two real simulations based on default emission of CUACE and the improved emission with a vehicle emission inventory with high temporal–spatial resolution (hereafter referred to as HTSVE) are carried out to evaluate the accuracy of pollutant concentrations simulated by CUACE and analyse the influence of HTSVE on Beijing air quality (hereafter referred to as SIM1 and SIM2). The methods of investigating the contribution rate to ambient pollution level (or source apportionment), based on an air quality numerical model, include source sensitivity simulations using the brute force method (also referred as zero-out method) or the decoupled direct method, air pollution tagged method, and the adjoint method (An et al., 2015; Burr and Zhang, 2011; Zhang et al., 2015). With comprehensible physical and chemical processes, the adjoint method has a significant advantage in source apportionment compared to sensitivity simulations or the tagged method. However, complicated mathematics and a large amount of data processing and programming result in a limited number of available regional-scale air quality adjoint models at present. Recently, An et al. (2015) developed an adjoint of the aerosol module in CUACE. The development of gaseous adjoint module of CUACE is needed for wider applications of source apportionment or source assimilation. The tagged method tracks the contribution of pollutant from specific sources and undergoes explicit atmospheric processes, but it is not able to simulate indirect effects and oxidant-limiting effects. With the ability to simulate indirect effects and relatively simple model runs, source sensitivity analysis is widely used in source attribution. However, significant source variations may result in misunderstandings due to non-linearity and changes in atmospheric background concentrations. In a previous study, the impact of local Beijing emission on air pollution is almost linear in source sensitivity analysis (An et al., 2007). Sensitivity analysis is suitable for investigating the contribution of vehicle emission in Beijing due to limited change of emission in this study. The vehicle emission contribution (VEC) to ambient pollutant concentration is computed based on the sensitivity analysis method of switching vehicle emission on (SIM2) and off (here after refer to SIM3) in Beijing. This method keeps atmospheric background pollution level basically steady, which has a significant effect on the chemical conversion because of relatively limited changes in emission. Meanwhile the effect of vehicle emission on secondary pollution, e.g. secondary aerosols which become important components of PM in Beijing (Huang et al., 2014), was considered. The formula for VEC is as follows: VEC=CSIM2-CSIM3CSIM2×100%, where C represents pollutant concentration. In fact, the regional transports of pollutants obviously has an effect on VEC; we calculate relative vehicle emission contribution (RVEC), which does not consider pollutant regional transports, in Eq. (2): RVEC=CSIM2-CSIM3CSIM2-CSIM4×100%, where SIM4 represents the simulation of switching off all emission sources in Beijing. All simulation test schemes are listed in Table 1.

CUACE has an independent pollution emission module, which contains natural and anthropogenic emissions including many gas and particle matter emissions (Gong et al., 2009). Anthropogenic emissions of SO2, NOx, CO, volatile organic compounds, PM2.5, PM10, BC, OC, etc. used in the emission module were developed by CMA based on the INTEX-B inventory, the emissions database for global atmospheric research (EDGAR), and an environmental statistics database. Gridded INTEX-B inventory covers 22 countries and regions in East Asia with a resolution of 0.5∘ × 0.5∘ and is separated into industry emission, power station emission, residential emission, and vehicle emission (Zhang et al., 2009). EDGAR is a joint project of the European Commission Joint Research Centre and the Netherlands Environmental Assessment Agency. The environmental statistics database is supplied by the Environmental Protection Agency. Some old data were corrected or updated according to the variation rate of anthropogenic emissions from environmental statistics database. Finally, the emission inventory was pretreated by SMOKE for detailed temporal and spatial distribution. Hourly emissions were obtained for CUACE model input. The emission inventory is a key factor in air quality numerical simulation. Annual emissions of CO, NOx, SO2, and PM2.5 in CUACE in Beijing are 3149.5, 173.8, 158.2, and 79.0 kt respectively. Comparing different research (Table 2) found that there are many uncertainties of inventories, especially for CO and NOx emissions, but it is difficult to identify which one is more accurate. With rapid economic development and the adjustment of energy structure, anthropogenic emissions have a significant variation in recent years in North China. However, the database of emission inventory in previous studies (Table 2) is from before 2010, which is the main reason for the differences between CUACE emission and others. For example, the Beijing municipal government has commenced strict traffic restriction since 2008. The number of vehicles in Beijing increased about 8 % in 2013. The change of vehicle emission may be responsible for NO2 emission variation. Except for dates of basic data, the methods of establishing inventory, emission factors, and basic data source would result in significant differences of emission inventory.

This study focuses on vehicle sources and their influence. HTSVE based on NRT traffic data was used to replace the vehicle emission in CUACE emission module to analyse its effects on air quality simulation. The detailed description of vehicle emission with high temporal–spatial resolution and comparison with vehicle emission in CUACE emission module are presented in Part 1. The contribution of major species from vehicle emission is presented in Table 3. The vehicle emission of NO, NO2, and HC from HTSVE is higher while that of CO and PM2.5 is lower than from CUACE.

The accuracy of air quality simulation based on numerical model greatly relates to mesoscale meteorological simulation. Although good performance of MM5 had been obtained in previous studies, our results should be evaluated due to variable performance under different regional, seasonal, and physical parameterization conditions. Based on statistical analysis, the 2 m temperature root mean square error (RMSE) and correlation coefficient (R) are 3.4 K and 0.81 in July and 3.8 K and 0.87 in December. MM5 can capture temporal and spatial variation of near-surface temperature effectively. The 2 m specific humidity RMSE and R are 2.4 g kg-1 and 0.56 in July and 0.9 g kg-1 and 0.82 in December, which indicates that basic temporal and spatial variation of near-surface specific humidity are simulated by MM5. The 10 m wind speed RMSE and R are 1.4 m s-1 and 0.37 in July and 1.7 m s-1 and 0.57 in December. The RMSE was 1–4 K for 2 m temperature, 1–2 g kg-1 for 2 m specific humidity, and 1–4 m s-1 for 10 m wind speed in most studies (Han et al., 2008; He et al., 2013, 2014; Jiménez-Guerrero et al., 2008; Kioutsioukis et al., 2016; Papalexiou and Moussiopoulos, 2006; Miao et al., 2008). In this study, MM5 presents the essential features of the local circulation over Beijing as seen from the above analysis and its performance observed here is comparable to other studies generally. The details of meteorological evaluation are provided in the Supplement. The statistic parameters are depicted in He et al. (2014).

NO2 and PM2.5 are of major concern as they are susceptible to vehicle emission. Intervals of simulated and observed daily mean near-surface NO2 and PM2.5 concentrations averaged over nine sites during two periods are shown in Fig. 2. CUACE underestimates the NO2 concentration significantly, especially during serious pollution periods. Due to the increasing emission of HTSVE (Table 2), the NO2 concentration from SIM2 increases 31.8 and 11.1 % in July and December respectively, resulting in significant improvement to the previous underestimates. The RMSEs of NO2 daily mean concentration decrease 17.6 and 10.9 % in two periods when HTSVE is used. Temporal correlation coefficients of NO2 daily mean concentrations for SIM1 and SIM2 are 0.80 and 0.79 respectively in December, which indicates CUACE can reproduce NO2 time trends accurately. However, low correlation (0.21 and 0.12 for SIM1 and SIM2 respectively) in July reflects the complexity of air quality numerical simulation. Simulated PM2.5 daily mean concentration is basically consistent with observed value. A minor difference of PM2.5 concentration is observed between SIM1 and SIM2 due to fewer vehicle emission changes (Table 3). Based on temporal correlation analysis, SIM2 improves PM2.5 time trends slightly, with correlation coefficients of 0.75 and 0.77 in SIM1 and 0.76 and 0.78 in SIM2. Compared with SIM1, the RMSE of PM2.5 daily mean concentration slightly decreased in SIM2. It is obvious that simulated PM2.5 concentration is more accurate than simulated NO2 concentration in July; similar phenomena was found in previous studies (Roustan et al., 2011; Wu et al., 2011). CUACE's ability is evaluated through the comparison of model grid and site station values; however, this method has several uncertainties because local information is involved. It should be noted that the lifetime of ambient NO2 is shorter than that of ambient PM2.5 due to the different chemical processes, and local characteristics are more significant for NO2. The grid average concentration of NO2 simulated by CUACE weakens the sub-grid local characteristics and results in poor performance of NO2 simulation compared with PM2.5. The uncertainty of the emission inventory increases with the spatial resolution of the numerical model. Although vehicle emission was replaced with HTSVE, the uncertainty of emission inventories of other sectors in Beijing and all emissions in surrounding areas is still an important reason for the bias of pollutant concentrations. Seasonal differences in CUACE performance are found in this study, with accurate simulation in winter, and this may relate to meteorological conditions, especially wind field bias as mentioned above. The uncertainty of the photochemical reaction, which is more significant in summer, might result in a large bias compared to the performance of NO2 in winter. Overall, the performance of CUACE is comparable with other studies in Beijing (Gao et al., 2011; Wu et al., 2011). Because SIM2 had a better performance, it is used as a baseline scenario in the flowing analysis.

Spatial distribution of pollutant concentration relates to pollutant emission distribution and meteorological condition. The spatial distribution of pollutant concentration from CUACE is basically consistent with site observations (Fig. 3). The mean wind in urban Beijing is the southwesterly wind in July, and it drives local pollutant transports from the southwest to the northeast. The high NO2 concentration is located in northeastern Beijing, while two regions with high PM2.5 concentration appear in the west and city centre (Fig. 3a and b). The spatial distribution of NO2 is different from that of PM2.5 because of emission sources distribution differences with one high-emission area on the inner Fifth Ring Road for NO2 and two high-emission areas on the west Sixth Ring Road and inner Third Ring Road for PM2.5 (Fig. 4). High concentrations present in high emissions or the downwind. The mean concentrations of NO2 and PM2.5 are 29.8 and 91.3 µg m-3 in July. Urban Beijing is dominated by northwesterly wind in December, and pollutant concentration distribution is obviously different from that in July. NO2 concentration is high in southeastern Beijing, and gradually decreases outward (Fig. 3c). High PM2.5 concentration is mostly located in western and southeastern Beijing (Fig. 3d). A significant difference in NO2 distribution between July and December and a slight difference of PM2.5 are found due to the combined effect of wind fields and emission distributions. The mean concentrations of NO2 and PM2.5 are 42.8 and 136.4 µg m-3 in December respectively.

VEC to ambient pollutant concentration is analysed through comparison simulation with and without vehicle emission (SIM2 and SIM3 respectively). Probability density function (PDF) is a good way to describe the total representation. The PDF of instantaneous VEC in two periods is shown in Fig. 5. The maximum frequencies of VECs to NO2 in July and December are 55–60 and 50–55 % respectively. The frequencies of VECs to NO2 from 15 to 60 % in December are larger than in July (Fig. 5a), which indicates that a large contribution presents in summer while a small contribution presents in winter. Based on one-way analysis of variance, the difference in VECs to NO2 in summer and winter is significant. This may relate to seasonal differences of meteorological condition and pollutant emission. In summer, high temperature and strong solar radiation lead to strong atmosphere oxidation ability, and therefore it is easy to convert from NO to NO2, which results in large contribution to NO2 concentration. Meanwhile, the high rate of NO2 emission from vehicle (Table 3) is another reason for the large contribution to ambient NO2 concentration in summer. The VEC to PM2.5 is considerably lower than that to NO2. The maximum frequencies of VECs to PM2.5 in July and December are 0–5 and 5–10 % respectively. Different from NO2, the mean VEC to PM2.5 in summer is smaller than that in winter, with a significant difference from one-way analysis of variance. Relative humidity in summer is higher than that in winter, and high relative humidity is conductive to gas–particle conversion processes of other emission sources (Yao et al., 2014), which may be one of the reasons for small VECs to PM2.5 in summer. The strong turbulence mixing in summer makes rapidly vertical exchange and transport of pollutant in boundary layer and finally results in small VECs to PM2.5 in summer. Wind field variation is another reason for seasonal change of VECs to PM2.5, which will be investigated in the following part.

As the local transports of pollutants, the VEC in Beijing depends on wind field and spatial distribution of vehicle emission. Wind dependency maps of VECs to NO2 and PM2.5 are shown in Fig. 6. High VECs to NO2 in July appeared in southerly wind with 3–4 m s-1 and in northerly wind with 6–7 m s-1 in December. Due to the difference in lifetime between NO2 and PM2.5, the wind dependency map of PM2.5 is quite different from that of NO2. High VECs to PM2.5 in July and December appeared in northerly wind due to high vehicle emission of particle matter in the northeastern part of the city (Jing et al., 2016). The dominant wind is the southwesterly wind in July and northwest in December (Fig. 3), which bring a small VEC to PM2.5 in summer. Significant regional transport, which is analysed in the next section, is one of the reasons for relatively small VECs to PM2.5 in summer.

Figure 7 shows time series of VECs to NO2 and PM2.5 daily mean concentrations in main urban areas (within the Sixth Ring Road) in two periods. The VEC not only changes with seasons, which is consistent with Cheng et al. (2007), but also changes with time. Time series of regional mean VECs of 49.8–60.0 % to ambient NO2 concentration in July, with a mean contribution rate of 55.4 %. In December, regional mean contribution on NO2 concentration decreases to 28.5–57.9 % at different days, with a mean contribution rate of 48.5 %. The VEC to ambient PM2.5 concentration is less than 10.3 and 13.6 % at different times, with mean contribution rate of 5.4 and 10.5 % in July and December respectively. The change of VECs to PM2.5 between July and December is most caused by meteorological condition in two periods. With different lift time of PM2.5 and NO2, PM2.5 concentration is more affected by regional transports, while NO2 concentration is more affected by local emissions. Therefore the contribution with time variation for PM2.5 is different from that for NO2. Except for wind field, pollution level is an important factor in VECs. It is obvious that low VECs present in serious pollution, while high VECs present in a low pollution concentration level, especially for NO2 (Fig. 8). The absolute contribution of vehicle emission increases in severe pollution mostly because of adverse dispersion condition. However, pollutant regional transport is enhanced in severe pollution, which results in a negative correlation between VEC and pollution concentration level. The VEC has a significant spatial variation; a previous study found that PM2.5 had a larger contribution from vehicle emission in urban than in suburban (13.0–16.3 % vs. 5.1 %) (S. W. Wu et al., 2014). Figure 9 shows the spatial distribution of the mean contribution rate of vehicle emission in two periods. Vehicle emission contributes 26.0–76.4 and 22.9–66.4 % of NO2 at different regions in July and December. A significant effect of vehicle emission on the ambient NO2 concentration level is found in southeastern and northeastern Beijing. VECs to PM2.5 are 1.2–15.4 and 2.4–24.4 % in July and December. The large contribution appears in the northeast in both summer and winter, which is vastly different from the distribution of NO2 contribution.

As can be seen from Table 4, receptor source apportionment and numerical sensitivity analysis are two main methods to compute the VEC to ambient pollutant concentration; additionally, VEC has significant uncertainties from previous studies. In summary, vehicle emission contributes 4–17 and 22 % to PM2.5 concentrations based on receptor source apportionment and numerical simulation methods and 56–74 % to NOx concentrations based on the numerical simulation method. The differences of the vehicle emission contribution to PM2.5 with the different methods are relatively large. The uncertainties of VEC are related to sampling or simulation time, location, analysis method, and weather conditions. The results from receptor source apportionment (chemical mass balance, PMF, etc.) only represent the characteristics of receptor point and can be applied to primary pollutants (Cheng et al., 2015); however, it is different from numerical sensitivity analysis which normally describes the regional characteristics and applies for primary and secondary pollutants. The uncertainty of emission source in a numerical model may be the main reason for significant differences of VECs in previous numerical studies. Though the simulation in this study is relatively short, our results are still comparable with previous studies. Small differences between our study and previous studies can be attributed to different analyzing periods and methods.

In this study, the rates of NO2 and PM2.5 from vehicle emission account for 55.1 and 22.3 % in July and 53.9 and 20.6 % in December (Table 3) of total emission. Because of the effect of pollutant regional transports, the contribution rate of vehicle emission on ambient pollutant concentration is lower than the rate of vehicle emission in total emissions. The difference between these two rates became significantly larger with more contribution of outside emission, which implies the importance of weather condition. In order to avoid the effect of weather on analysis results, the relative contribution of vehicle emission on pollutant concentrations is analysed in the following section.

The chemical components of PM2.5 represent the characteristics of emission source and complexity of chemical processes of pollutants in the atmosphere. Based on the sensitivity test, the VECs of BC, OC, and NI are large, while they are relatively small for SF and AM (Table 5). The VECs of BC and OC in December are approximately twice of that in July. Seasonal changes for the rates of BC and OC from vehicle emission in total emission are not apparent, which indicates that the seasonal change of VECs is unrelated to vehicle emission. Beijing is controlled by southerly wind, which results in significant regional transport. Additionally, it causes small (large) VECs of BC and OC in summer (winter). Atmospheric chemical processes and dispersion conditions are also the reason for seasonal change of different components VECs. Using MM5–CMAQ model simulation, Cheng et al. (2013) investigated the VEC to the PM2.5 and found the VECs of BC were 32.3 and 30.7 % in summer and winter respectively. Our results are comparable with Cheng et al. (2013) in winter, while they show some difference in summer.

Air pollution in Beijing is attributed not only to local emissions but also to regional transports. Using the CMAQ model, An et al. (2007) investigated the contribution to pollutant concentrations in Beijing by using emission switch on/off method; the contribution of non-local emission accounted for 15–53 % of PM2.5. Wu et al. (2011) studied the contribution to air pollution during CARE-Beijing 2006, and local emission in Beijing accounted for 65 % of SO2, 75 % of PM10, and 90 % of NO2 concentrations. Pollutant regional transport depends on atmospheric circulation and regional emission characteristics. By comparing pollutant concentrations between SIM2 and SIM4, local emissions in Beijing contribute 93.6 and 62.6 % to NO2 and PM2.5 concentrations in July and 83.8 and 76.1 % to NO2 and PM2.5 concentrations in December, which have a profound effect on RVEC.

Figure 10 depicts the spatial distribution of RVECs to NO2 and PM2.5 in July and December, and a similar distribution is found in two periods. The RVEC to NO2 is large in main southeastern and northeastern urban areas, while small in main western urban areas. Time series of regional mean RVECs to NO2 in main urban areas range from 52.3 to 63.4 % and 49.4 to 61.2 %, with means of 59.2 and 57.8 %, in July and December respectively. Different from NO2, the RVEC to PM2.5 is large northeast of main urban areas in the two periods. Time series of regional mean RVECs to PM2.5 range from 5.7 to 11.3 % and 9.9 to 16.1 %, with means of 8.7 and 13.9 %, in July and December respectively. The differences of RVECs to NO2 and PM2.5 in July and December are significant based on one-way analysis of variance. The spatial distribution of RVEC is tremendously affected by vehicle emission, as they are mostly consistent with the rate of vehicle emission in total emission (Fig. 4). As pointed out by Jing et al. (2016), the uncertainty of HTSVE is very small through multiple comparison with statistical data and real-time observation. However, the uncertainty of other sector emissions has a negative influence on the precision of RVECs, which needs more improvement for accurate environmental management. Local circulation also determines the spatial distribution of RVECs. High PM2.5 emission from vehicles is found between north Fourth Ring Road and north Fifth Ring Road (See Part 1, Fig. 9). Controlled by southwesterly wind, PM2.5 from vehicles is easily transferred out of the main urban areas, which results in low RVEC in July. However, the majority of PM2.5 from vehicles stays in the main eastern area of the city controlled by northwesterly wind, which results in high RVEC in December. Based on the zero-out method, Cheng et al. (2013) found the contribution rates to pollutant concentrations were higher than those to the emissions because near-surface emission from vehicles facilitated greater contribution to local pollutant concentrations on the ground level. Regardless of regional transports, the contribution of vehicle emission to ambient PM2.5 concentration is substantial lower than the rate of vehicle emission to total emission in this study. Our finding is seemingly in conflict with Cheng et al. (2013) but may be more reasonable for the following reasons. Different from elevated emission, PM2.5 from vehicle emission in the near-surface layer easily descends to the ground or is absorbed by vegetation, which leads to a low contribution rate to PM2.5 concentration. Secondary aerosol generated by photochemical reaction is different for different sector emissions. The VEC to SF is low in Beijing (Table 5), which indirectly causes low VEC to PM2.5. Furthermore, pollutant regional transport and the background concentration may result in lower VEC to PM2.5 than the rate of emission.