Numerical study of the initial condition and emission on simulating PM 2 . 5 concentrations in Comprehensive Air Quality Model with extensions version 6 . 1 ( CAMx v 6 . 1 ) : Taking Xi ' an as example

A series of model sensitivity experiments is designed to explore the effects of different initial conditions and emissions in Xi’an in December 2016, which is a major city in the key area “Fen-Wei Plains” for air pollution control in China. Three methods were applied for the initial condition tests: clean initial mechanism, restart mechanism and continuous 10 simulation. In clean initial mechanism test, the sensitivity experiments C00, C06, C12, C18, and C24 were conducted according to the intercepted time periods, and the results showed that the model performance of PM2.5 was better with the delay of the start time of the intercepted time periods. From experiments C00 to C24, the absolute mean bias (MB) decreased from 51.07 μg/m3 to 3.72 μg/m3 and the index of agreement (IOA) increased from 0.49 to 0.86, which illustrates that the model performance of C24 is much better than C00. In order to explore the restart mechanism, sensitivity experiments R1120 and 15 R1124 were set according to the time of the first day for the model simulation. Although the start times of simulations were different, after a period of spin-up time, the simulation results with different start time were nearly consistent, the results showed that the spin-up time is about 27 hours. As for the continuous simulation test, CT12 and CT24 were conducted. The start time of the intercepted time periods for CT12 and R1120 were the same, and the simulation results were nearly identical. The simulation results of CT24 performed best in all the sensitivity experiments, with the correlation coefficient (R), MB, and 20 IOA reaching 0.81, 6.29 μg/m3, and 0.90, respectively. For the emission tests, the updated local emission inventory with construction fugitive dust emissions have been added and compared to the simulation results of the original emission inventory. The simulation with the updated local emissions showed a much better performance on PM2.5 modelling. Therefore, combining the method of CT24 with the updated local emission inventory can nicely improve the model performance of PM2.5 in Xi'an, the absolute MB decreased from 35.16 μg/m3 to 6.29 μg/m3 and the IOA reached 0.90. 25


Introduction
In the recent years, severe air pollution has gradually become a big challenge in China and other developing countries (Wu et al., 2014;Li et al., 2017). China released three-year action plan for cleaner air in 2018, and efforts will be focused on areas including the Beijing-Tianjin-Hebei region, the Yangtze River Delta, and the Fen-Wei Plains. As a major city in the key area "Fen-Wei Plains", Xi'an is located in the Guanzhong Basin. The city is surrounded by the Qinling Mountains to the south, and analyze the apportionment of emission categories and contributions from different source regions for atmospheric pollutions, many researchers used the CAMx model with the particulate matter source apportionment technology (PSAT) in different areas of China, including Beijing (Zhang et al., 2018), Tangshan (Li et al., 2013), Pearl River Delta region (Wu et al., 2013) and Yangtze River Delta region (Li et al., 2011). The CAMx showed good model performances for simulation of air pollution (Panagiotopoulou et al., 2016). 45 The input files for CAMx model include initial/boundary conditions, gridded and elevated point source emissions, and meteorological files (ENVIRON, 2013). The meteorology and emissions inputs can cause great uncertainty for air quality models (Tang et al. 2010;Gilliam et al., 2015). Many researchers reduced the uncertainty of meteorology through refined physical parameterizations and other techniques such as data assimilation (Sistla et al., 1996;Seaman, 2000;Gilliam et al., 2015;Li et al., 2019). A reasonable emission inventory is very important for the simulation accuracy of the air quality model. 50 Many researchers studied East Asian emissions (Kato et al., 1992;Streets et al., 2003;Ohara et al. 2007;Zhang et al., 2009;), and tried to construct emission inventories of particulate matter (PM) in China (Wang et al. 2005;Zhang et al. 2006). However, the absence of detailed information on China introduces great uncertainty into the emission inventories (Cao et al., 2011). In recent years, more and more researchers focused on constructing and updating of regional local emission inventories to improve the model performance. Wu et al. (2014) improved the model performance by adding more regional point source 55 emissions and updating the area source emissions in village and surrounding cities of Beijing. Based on that work, Yang et al. (2019) have added local datasets into the emission inventory of Guanzhong Plain, China, which was applied in simulating PM2.5 concentrations by CMAQ model in Xi'an. Numerous works indicated that construction dust emission plays an important role of the air pollution, especially in urban areas (Ni et al., 2012;Huang et al., 2014;Wang et al., 2015). In our previous study, we built a particulate matter emission inventory from construction activities at county level in Xi'an, based on an extensive 60 survey of construction activities and combining with two sets of dust emission factors for the typical city in north China (Xiao et al., 2019).
However, few studies have investigated the effects of initial condition on the simulation or prediction of PM2.5 concentrations. Therefore, the purpose of this study was to explore the effects of different initial mechanism and emissions on the model performance of PM2.5 in the CAMx model. A series of model sensitivity experiments for the initial mechanism and emissions are designed to find a suitable method for simulating PM2.5 concentrations with the reasonable initial condition and emission inventory. Not only for Xi'an, other cities may apply the similar research method using for simulating PM2.5 concentrations in the future.
The paper is organized as follows. Section 2 gives the model descriptions for WRF-SMOKE-CAMx model system, including meteorological fields, air quality model descriptions, model domain, emission inventory and processes in Sects 2.1-70 2.4. Section 3 presents the sensitivity experiments design of different initial condition and emission. Section 4 discuss the model performance of the initial condition tests and emission tests to simulate the PM2.5 concentration model in Xi'an. The conclusions are given in Section 5.

Meteorological fields
For the WRF model configuration, we chose the rapid radiative transfer model (RRTM; Mlawer et al., 1997) and the Dudhia 85 for longwave and shortwave radiation options (Dudhia, 1989), WSM3 cloud microphysics (Hong et al., 2004), the YSU scheme (Hong et al., 2006), the Kain-Fritsch (new Eta) cloud parameterization (Kain, 2004), and 5-layer thermal diffusion scheme (Dudhia, 1996). The meteorological initial and boundary conditions were derived from the National Centers for Environmental Prediction (NCEP) global final analysis data (FNL), with spatial resolution of 1° ×1° and temporal resolution of 6-h. The simulation was performed for the period of November 20, 2016to January 20, 2017 The simulation effect of daily average temperature (T2) and relatively humidity (RH2) simulated by WRF model in domain 3 were primarily validated by the observation data at 7 monitoring stations in Xi'an, which the station map was shown in Fig.2.
Some statistical parameters of Appendix A were used to evaluate the model performance and shown in Table 1, and the time series was shown in Fig 3. The ME, R and RMSE of daily average T2 are 1.37℃, 0.80, 1.65℃, respectively, and the simulation shows cooling bias of -0.95℃. The ME and RMSE of daily average RH2 are 6.77% and 8.30%. The correlation coefficient of 95 the relatively humidity is 0.71, which is reasonable. RH2 was slightly overestimated as the MB is 6.22%.
In previous studies, Yang et al. (2019) used WRF to drive CMAQ model for winter air quality in Xi'an, and the model evaluations for winter in 2016 showed that the MB, ME, R and RMSE of T2 were -2.83℃, 2.83℃, 0.89, 3.29℃, respectively.
And the MB, ME, R and RMSE of RH2 were 9.59%, 10.63%, 0.71, 13.43%, respectively. Wu et al. (2010) used the fifthgeneration NCAR/Penn State Mesoscale Model (MM5) as meteorological driver for the Nested Air Quality Prediction 100 Modeling System (NAQPMS), and the statistical results showed that the MB and R of T2 were 2.1℃ and 0.84 and those of RH2 were -15.8% and 0.65.
Compared with previous studies, T2 and RH2 in this study have lower MB, ME, and RMSE. The R of T2 is slightly lower than previous studies, while the R of RH2 is higher. Thus, the meteorological simulation of this study is reasonable.

Air quality model descriptions
The CAMx model is a state-of-the-science air quality model, which is developed by the Ramboll Environ (http://www.camx.com). In this study, the PPM advection scheme (Colella and Woodward, 1984) is used for horizontal 120 diffusion, and the K-theory is selected for vertical diffusion. The Regional Acid Deposition Model (RADM-AQ) (Chang et al., 1987) scheme as the aqueous-phase oxidation, ISORROPIA (Nenes et al., 1999) as inorganic aerosol thermodynamic equilibrium, and CB05 (Yarwood et al., 2005) as the gas-phase chemical mechanism, and the Euler-Backward Iterative (EBI) solver with Hertel's solutions (Hertel et al., 1993) is used in the model system. The resistance model for gases (Zhang et al., 2003) and aerosols (Zhang et al., 2001) in dry deposition module, and scavenging model for gases and aerosols (Seinfeld and 125 Pandis, 1998) in wet deposition module is chosen in this study. The CAMx model forecasted the next 48 hours' PM2.5 concentrations in clean initial mechanism testing and will be described more in Sec. 2.3. On the first day, CAMx used the results of the ICBCPREP which can prepare simple, static CAMx initial condition (IC) and boundary condition (BC). On the following days, it used the different initial conditions of the sensitivity experiments.

Model domain 130
Three-nest domains were designed for the WRF model ( Fig. 4), with a horizontal resolution of 27 km × 27 km (D1), 9 km × 9 km (D2) 3 km × 3 km (D3), respectively. The biggest domain (D1) covered the most parts of China, the second domain (D2) includes Shaanxi Province, Shanxi Province, Henan Province and the inner domain (D3) focused on the 11 cities in Fenwei Plain, including Xi'an. The CAMx has only one domain and the settings are the same as those in the D3 domain, while focusing on Xi'an as one sensitivity test area for initial conditions and emissions. To reduce the boundary effects, the CAMx 135 model cut down the outermost grid of WRF model and used the variable of the center grid in WRFCAMx module, thus, the CAMx model was three grid cells smaller than the WRF model in D3 domain. The vertical resolution of WRF was 37 layers from the ground to 5 hPa at the top, and 14 layers were extracted by the WRFCAMx module, which can convert the WRF output files into the data format for CAMx model.

Emission Inventory and Processes 145
The SMOKE version 2.4 (Houyoux and Vukovich, 1999) model was used to improve the Fen-wei emissions, especially Xi'an local emissions, and provide the gridded emissions for CAMx model in this study. Based on the emission inventories of previous study (Yang et al., 2019), this study added the emissions quantity of PM2.5 from construction fugitive dust in Xi'an to update the local emission inventories. The emission inventories in this study including: 1. The regional emissions in East Asia and the local emissions in Guanzhong plain were obtained from Wu et al. (2014) and 150 Yang et al. (2019). Major industrial emissions had a little adjusted according the annual report in this study. The emission inventory in city-level is presented in Table.2.
2. Construction fugitive dust emissions in Xi'an, based on the survey data of construction projects in Fig. 5, were collected in the previous study (Xiao et al.,2019), indicated as "local area source". This is new dataset at the county level and update in the 2017. The basic data included the location and area of each construction project. Also we replenished the missing 155 construction data and correct the error information with Google Earth and other geographic information tools to get more accurate location information. According to statistics, there were 1595 construction projects in Xi'an in 2017, with 86.1km 2 of the total construction area. The construction area in the main urban (Xincheng, Beilin, Lianhu, Yanqiao, Weiyang and Yanta) was about 62.2km 2 , accounting for 7.5% of the total area in the main urban. The distribution of the construction fugitive dust emissions in Xi'an is shown in Fig. 6.
We took the statistics-allocation approach to generate gridded area source emissions, which was to allocate the total emissions to each horizontal model grid according to the related spatial factors. In this study, the Land Scan 2015 Global Population Database (Dobson et al., 2000) was used as a spatial factor of population to allocate the emissions. For the construction fugitive dust emissions, we used the area of each construction project as the weight in the surrogate calculation, allocated the input construction projects data to the target polygons (map of administrative division in Xi'an at the county level) 165 based on weighted spatial overlap of the input data and target polygons. And the spatial results provide to the SMOKE model as spatially allocated factor. The horizontal and vertical allocation of point source emissions were assigned from their longitude-latitude coordinates and the Briggs algorithm (Briggs, 1972;1984), respectively. While the temporal variation and chemical species allocation were based on profile files in SMOKE model. Table 2, the NOx emissions ranged from 352.0 kt yr -1 to 758.5 kt yr -1 between 2008 in Zhang et al. (2009) to 170 2017 in this study. For PM10 emissions in Shaanxi Province, the emissions also increased from 474.0 kt yr -1 to 830.0 kt yr -1 .

As shown in
PM10 emission in this study are higher than others, because including the construction fugitive dust. Other emission species such as NOx, SO2, NH3, VOCs and CO are little higher in this study than previous studies.

Sensitivity experiments design
A set of model sensitivity experiments under different initial conditions and emissions are designed in this study. Three methods were applied for the initial condition tests: using the clean initial condition files as clean initial mechanism, using the 185 restart files as restart mechanism and the continuous simulation. For the emission tests, we compared the simulation results of the original emission inventory and the updated local emission inventory with construction fugitive dust emissions. The configurations of the simulation sensitivity experiments are shown in Table 3, and the time period for each initial condition experiments are shown in Fig. 7.

ICON test for using the clean initial condition files 190
For using the clean initial condition files, icbcprep module used a clean-troposphere vertical profile to generate the initial concentration fields for each day of simulation. The output files of CAMx were initialized at 13:00 UTC. The CAMx model forecasted the next 48 hours' PM2.5 concentrations in each cycle simulation. By extracting data from simulated results based on different time periods (0~24h, 6~30h, 12~36h, 18~42h and 24~48h, respectively) shown in Fig. 7 (a), we conducted the sensitivity experiments C00, C06, C12, C18, and C24, and explore the influence of different time periods on the simulation 195 effect of PM2.5. For the sensitivity experiment C00, the data of the period for the first 24 h of output file was cut and merged to analyze. And for C06, the first 6 h of data was spin-up time, we cut and merged the data of the period from 19:00 UTC to 18:00 UTC in the second day. C12, C18 and C24 were used the same method to extract and merge data, and the spin-up time of them was the first 12 h, 18 h and 24 h of data, respectively.

ICON test for using the restart files 200
The meteorological data of the period 12~36 h was cut to estimate the PM2.5 concentrations by restart mechanism of CAMx model. For using the restart files, icbcprep module also used clean initial concentration fields at the beginning of the first-day simulation. And gridded three-dimensional instantaneous concentrations of all species on all grids were written at the end of the simulation to allow for a model restart. Then ICON used the 24-h forecast results from the day before as the initial conditions for the following days shown in Fig. 7 (b). The first day of the simulation starts at 12:00 UTC, and the following 205 days starts at 00:00 UTC. In order to explore how long is the spin-up time can eliminate the error caused by the initial value, the sensitivity experiments R1120 and R1124 were set according to the time of the first day for the model simulation, which began on November 20 th , 2016 and November 24 th , 2016, respectively.

ICON test for continuous simulation
For the continuous simulation, sensitivity experiments CT12 and CT24 were set according to the start time of the intercepted 210 time periods, which were started at 00:00UTC and 12:00UTC, respectively, shown in Fig. 7 (c). For CT12, the meteorological data of the period 12~36 h was cut and merged to one file. The period 24~48 h was cut and merged for CT24. Also we built the continuous emission files by SMOKE model. During the simulation, there was no interruption and finally generated a longterm sequence simulation result for each start time.

Emission test for different emission inventories 215
Based on the initial condition tests, we selected the best method to do the emission sensitivity experiments. For the emission tests, we compared the simulation results of the original emission inventory (sensitivity experiments Enc) and the updated local emission inventory with the construction fugitive dust emissions (sensitivity experiments Ec). Table 3. The simulation experiment configurations. C00-C24, R1120, R1124, CT12 and CT24 were used to investigate the 220 impact of simulation methods, start time and extracted time period. The impact of different emission inventory was investigated by Ec and Enc. Method C, R, CT presented for the methods of the clean initial condition mechanism, restart mechanism and continuous simulation. Emission inventory nc and c presented for the original emission inventory and the updated local emission inventory with the construction fugitive dust emissions, respectively.

Results and discussion
In this study, we collected the observations in December 2016, and evaluate the model performance and improvement. The model ability from both the meteorological field and the daily PM2.5 simulations in Xi'an is evaluated in this study.

Model performance of the initial condition tests 235
There are 13 NSAQ Stations in Xi'an which marked as squares in The CT Station is the city background station, which is in northern urban Xi'an.

Sensitivity experiments for using clean initial condition files
Taylor nomogram (Taylor, 2001;Gates et al., 1999) is used to evaluate the accuracy of simulated PM2.5 daily concentrations 240 for NSAQ stations which was for the sensitivity experiments of using the clean initial condition files, shown in Fig 8. There are three statistical parameters to evaluate model accuracy, which are the correlation coefficient (R), normalized standard deviation (NSD), normalized root, and mean square error (NRMSE) in Taylor nomogram (Taylor, 2001;Gates et al., 1999;Chang et al., 2004). The sensitivity experiments C00, C06, C12, C18, and C24 were shown by symbols of different colors. We randomly selected 3 stations in urban Xi'an, 2 stations in county towns and a background station to show the simulation results. 245 And the "AVG" meant the average of 13 NSAQ Stations.
As shown in Fig 8, R is 0.36-0.76 for the sensitivity experiments C00, C06, C12, C18, and C24. C24's R is largest and best over all the NSAQ Stations, and C00's is lowest. The NRMSE, which measures the distance from the marker to the REF in Taylor nomogram, is smallest and best for C24, and longest for C00. For NSD, most NSAQ Stations have similar regularity, that is, the NSD values from C00 to C24 are getting closer to "1". The other statistical parameters are presented in Table 4. 250 From experiments C00 to C24, the absolute mean bias (MB) and the mean error (ME) decrease from 51.07 µg/m 3 to 3.72 µg/m 3 and from 74.09 µg/m 3 to 45.82 µg/m 3 , respectively. The absolute normal mean bias (NMB) and the normal mean error (NME) decrease from 29.73% to 2.17% and from 43.12%to 26.67%, respectively. And the index of agreement (IOA) increase from 0.50 to 0.8. In general, the model performance of C24 is better than other sensitivity experiments in clean initial mechanism tests. 255 The NSAQ Stations, which in urban city and county towns, have different model performances presented in   Chang et al. (2004) for air quality model.

Sensitivity experiments for using restart files
In order to explore the restart mechanism, sensitivity experiments R1120 and R1124 were set according to the time of the 265 first day for the model simulation. Starting from 12:00 UTC on November 24 th , the PM2.5 concentration simulation results of the two sensitive experiments, R1120 and R1124, were shown in Fig 9. At first, the results of the two sensitivity experiments were very different, and then the two lines were gradually fitted until 16:00 UTC on November 25 th . After 16:00 UTC on November 25 th , the two lines fitted almost completely. Therefore, the spin-up time of 27 hours can eliminate the error brought by the initial field for the PM2.5 concentrations in CAMx model. 270 As shown in Table 4, the model performances of the sensitivity experiments R1120 and R1124 are similar in December, 2016. For the sensitivity experiments R1120 and R1124, the R value between observations and simulations is 0.70. The mean bias (MB) and the mean error (ME) are 4.01 µg/m 3 and 49.68 µg/m 3 , respectively. The normal mean bias (NMB) and the normal mean error (NME) are 2.33% and 28.92%, respectively. The value of root mean square error (RMSE) is 67.28 and the IOA reaches 0.82. 275 Figure 9. The time series of hourly simulated PM2.5 concentrations for using the restart files during a period of the spin-up time. The red and blue lines represent the model sensitivity experiments R1120 and R1124, respectively. The begin day of R1120 for model simulation was November 20 th , 2016, and R1124 was November 24 th , 2016.

Sensitivity experiments for continuous simulation 280
As for the continuous simulation, sensitivity experiments CT12 and CT24 were conducted. Although the sensitivity experiments CT12 and R1120 use different methods to generate the initial concentration fields, the start time of the intercepted time periods for the two experiments are the same. PM2.5 concentrations of CT12 and R1120 are presented in Fig 10. As shown in Fig 10, the points lie very close to the perfect line "y=x", which indicates that the simulation results of CT12 and R1120 were nearly identical. 285 The model starting time of sensitivity experiments CT12 and CT24 are November 26 th 00:00UTC and November 26 th 12:00UTC, respectively. The concentration accumulation of CT24 is 12 hours more than that of CT12. As shown in Fig 11 there is an air pollution peak in December 2016, which CT24 matches better than CT12. The statistical parameters of CT12 and CT24 are presented in Table 4. The mean bias (MB) and mean error (ME) of CT24 results are 6.29 µg/m 3 and 42.67 µg/m 3 , respectively, a little better than the CT12 results. The root mean square error (RMSE) of CT24 results is 68.21, also slightly 290 better than the CT12 results. From CT12 to CT24, the R and IOA increase from 0.69 to 0.81 and from 0.81 to 0.90, respectively.
Thus, the sensitivity experiments CT24 has better model performance than CT12.

Model performance of emission tests 300
Taylor nomogram for modelled and observed daily averaged PM2.5 concentrations for all initial condition sensitivity experiments, shown in Fig 13. The red symbols indicate the sensitivity experiment for using the clean initial condition files, the blue symbols represent the sensitivity experiment for using the restart files, and the brown symbols show the sensitivity experiment for continuous simulation. One experiment per symbol. The circles and triangles represent "bias". As shown in Fig   13, R is 0.36 ~ 0.81 in all initial condition sensitivity experiments. R of CT24 is the largest and best in all initial condition 305 sensitivity experiments. The marker of CT24 has the shortest distance to the "REF" than other initial condition sensitivity experiments, which means that the NRMSE is the smallest. The NSD of CT24 is 0.92, which determines the modeled and observed patterns have more consistent amplitude of variation. According to these statistical parameters, the sensitivity experiments CT24 have the best model performance than other initial condition sensitivity experiments.
Based on the initial condition tests, we selected the best method, CT24, to do the emission sensitivity experiments, shown The NRMSE for Enc is smaller than Ec shown in the taylor nomogram. However, the NSD of Ec, 0.92, is better than Enc, 0.74. And the bias of Enc is much larger than Ec. The other statistical parameters are presented in Table 4

Conclusions
The WRF-SMOKE-CAMx model system has been used to simulate the fine particular (PM2.5) concentrations in Xi'an in December, 2016. In this study, the emissions of construction fugitive dust in Xi'an were added in the SMOKE to update the local emission inventory. A series of model sensitivity experiments for the initial conditions and emissions are designed to 335 improve the model performance in the megacity, Xi'an.
Three methods were applied for the initial condition tests: using the clean initial condition files as clean initial mechanism, using the restart files as restart mechanism and continuous simulation. All initial condition sensitivity experiments are driven by the updated emission inventories. The emission tests are based on the initial condition sensitivity experiment which has the best model performance. 340 Comparing the model performance of PM2.5 concentrations in different model sensitivity experiments in Xi'an, we found that the model combining the method of continuous simulation with the updated local emission inventory can nicely improve the model performance. According to statistical parameters, for initial condition tests, the model performance of CT24, C24 and R1120/R1124 are the best. R ranges from 0.36 to 0.81 in all initial condition sensitivity experiments. R of CT24 is the largest and best in all initial condition sensitivity experiments. R of C24 and R1120/R1124 can reach 0.76, 0.70, respectively. 345 The MB of CT24, C24 and R1120/R1124 are lower, which are 6.29 µg/m 3 , -3.72 µg/m 3 and 4.01 µg/m 3 , respectively. The IOA of CT24, C24 and R1120/R1124 all reach above 0.8, of which CT24 is 0.9. Compared with other methods, the method of using the clean initial condition files has longer simulation time and larger data volume. Therefore, the method of continuous simulation for hindcast, which is to retrieve PM2.5 concentrations is suggested. For air quality forecast, we can give priority to the method of restart mechanism. Also, for simulating PM2.5 concentrations by CAMx model, the simulate needs the spin-up 350 time at least 27 hours. This can improve the simulation effect and reduce the simulation time.
This study updated the emissions inventory, which added construction fugitive dust emissions to the original emissions inventory. Compared to the simulation results based on the original emission inventory, a new simulation results, which were driven by the updated local emissions, showed a much better performance on PM2.5 modelling. The absolute MB decreased from 35.16 µg/m 3 to 6.29 µg/m 3 and the IOA of simulation results with the updated local emissions reached 0.90. Therefore, 355 the right addition of emissions will also help to improve the effects of simulation and forecasting.
In finally, we recommend the method of continuous simulation for hindcast, which has the best model performance of PM2.5 concentrations, and can also reduce the output of IO files to improve computing efficiency. For forecast, the method of restart mechanism is suggested, which can reach similar model performance as the continuous simulation. If the restart mechanism cannot be used due to the limitation of computing resources and storage space when forecasting PM2.5 concentrations, try to 360 extend the spin-up time as much as possible, at least 27 hours according to this work.

Author contribution
Han Xiao did the simulation and prepared the materials. Qizhong Wu designed the WRF-SMOKE-CAMx modelling system 370 for Xi'an, including emission processes. Xiaochun Yang collected the local emission inventory in Shaanxi province and help emission processes. Lanning Wang and Huaqiong Cheng help to prepare the model dataset and figure.

Acknowledgements
The National Key R&D Program of China (2017YFC0209805), the National Natural Science Foundation of China In the equations, 7 and 7 represent the simulated value and observation value of a station, respectively. represents the number of stations. ? and @ represent the average value of simulated value and observation value, respectively. 395