We used the CHIMERE regional chemical transport model in this study, which is designed to produce daily forecasts of tropospheric trace gas and aerosol pollutants and make long-term simulations at a range of spatial scales (Menut et al., 2013). We use the CHIMERE model v2013b over East Asia (18 to 50∘ N and 102 to 132∘ E) with a resolution of 0.25∘ following the configuration in Ding et al. (2015). The CHIMERE simulation was driven by operational meteorological data from the European Centre for Medium-Range Weather Forecasts (ECMWF) with a horizontal resolution of 0.25∘. Atmospheric variables were simulated in eight layers from the surface to 500 hPa. Tropospheric photochemistry is represented using the reduced MELCHIOR chemical mechanism (Derognat et al., 2003), including about 120 reactions and 44 gaseous species. An aerosol module accounting both for inorganic and organic species of primary or secondary origin is included according to Bessagnet et al. (2004). Boundary conditions for the model domain were derived from monthly mean climatology based on the second-generation Model for OZone And Related chemical Tracers (MOZART) (Horowitz et al., 2003) for gases, the Laboratoire de Météorologie Dynamique Zoom – Interaction avec la Chimie et les Aérosols (LMDz-INCA; Folberth et al., 2006) for nitrate and ammonium, and the Georgia Tech/Goddard Global Ozone Chemistry Aerosol Radiation and Transport (GOCART; Ginoux et al., 2001) for other aerosols. At default, NOx emissions are speciated as 9.2 % of NO2, 0.8 % of HONO, and 90 % of NO in the CHIMERE model (Menut et al., 2013), following the Generation of European Emission Data for Episodes (GENEMIS) recommendations (Friedrich, 2000; Kurtenbach et al., 2001; Aumont et al., 2003). Open-access satellite-derived and bottom-up inventories that provide up-to-date emissions over East Asia were selected to drive the model in this study, which will be detailed in Sect. 2.2.

The satellite-derived NOx emissions were estimated by the algorithm DECSO (Daily Emission estimates Constrained by Satellite Observations) v5 using an extended Kalman filter (Mijling and van der A, 2012; Ding et al., 2015, 2017b). DECSO uses one forward model run of a CTM to calculate the response of NO2 concentrations to both local and nonlocal NOx emissions. Daily OMI NO2 observations retrieved with the DOMINO version 2 algorithm (Boersma et al., 2011) are used as a constraint to update emissions. The DECSO emission data are available at www.globemission.eu (last access: 20 March 2018).

The bottom-up anthropogenic NOx emissions were taken from the MIX inventory (Li et al., 2017a), a mosaic Asian anthropogenic emission inventory under the international collaboration framework of the Model Inter-Comparison Study for Asia (MICS-Asia) and the Task Force on Hemispheric Transport of Air Pollution (TF HTAP). The MIX inventory is developed for the years 2008 and 2010 by an integration of state-of-the-art regional emission inventories for all major anthropogenic sources in 29 countries and regions over Asia. The emissions of China integrated in the MIX inventory are derived from the Multi-resolution Emission Inventory for China (MEIC: http://www.meicmodel.org, last access: 20 December 2017) compiled by Tsinghua University. The anthropogenic emissions together with the biogenic emissions, which were computed automatically in the CHIMERE model using the global MEGAN (Model of Emissions of Gases and Aerosols from Nature) model (Guenther et al., 2006), were adopted as the bottom-up inventory. We refer to this combination as the MIX inventory for brevity hereinafter. Note that monthly emissions for all inventories above were provided at the spatial resolution of 0.25∘ × 0.25∘.

Both inventories show comparable spatial distributions at national and regional scales, but distinctions between urban and rural areas (see Sect. 3.1). The strength of the MIX inventory is that it includes detailed source-category information (e.g., power plant and transportation sector) for emissions, which is useful for driving atmospheric models and designing emission mitigation policies but is not included in DECSO. The advantage of the DECSO inventory is that emissions are timely updated (as soon as the satellite observations are available); while bottom-up inventories usually lag behind a few years and are outdated by the time they become available. In addition, the spatial information in DECSO is based on OMI NO2 observations, while MIX relies on spatial proxies like gross domestic product (GDP) to allocate emissions due to the lack of data. An in-depth comparison between inventories has been described by Ding et al. (2017).

We focused on 2015 as the most recent year with available DECSO emission estimates and in situ measurements, but we used the MIX inventory for 2010 because the year 2015 is not available yet. However, the use of the 2010 MIX inventory without scaling is not expected to bring significant bias, as the similarity of NOx emissions for 2010 and 2015 has been reported by both the bottom-up inventory MEIC (Liu et al., 2016a) and the satellite-derived inventory DECSO. For the period of 2010–2012, the NOx emissions of China experienced a rapid growth. A sharp decline in NOx emissions was observed in the years of 2013–2015, with a peak around 2012 (Liu et al., 2016a). As a result, the inventory for 2010 is comparable to that for 2015, even though there is a 5-year lag. Figure 1 compares DECSO NOx emissions for 2015 (a) and 2010 (b), which are consistent in both total amount (21.5 vs. 21.6 Tg) and spatial distribution (r = 0.83). Figure 1 further displays the spatial distributions of the MIX NOx emissions for 2010 (c). These emissions are significantly higher (39 %) than the DECSO inventory when averaged over the model domain.

An air quality simulation using the CHIMERE model was conducted for the full year 2015. Pollutant concentrations including NO2 were simulated based on the 2015 DECSO and the 2010 MIX NOx inventories, respectively. Note that the 2010 MIX inventory for other species was used together with both NOx inventories. Because of the inconsistency between the emission sectors used in the DECSO and the MIX inventories and that in SNAP (Selected Nomenclature for Air Pollution) 97, which is internally used in the CHIMERE model, we adopted the sector mapping table as discussed in Ding et al. (2015). The concentration in the lowest model layer (from the ground up to 20 m) was used for validation against surface NO2 observations in this study. Figure 2 illustrates the annual mean surface NO2 simulation using both inventories. Large enhancements are found over industrial regions, in particular northern China, the North China Plain, and the Yangtze River Delta. The model run based on the MIX inventory (Fig. 2b) shows overall larger concentrations than that based on the DECSO inventory (Fig. 2a).

The real-time hourly NO2 concentrations as well as other major air pollutants are continuously recorded by the Ministry of Environmental Protection (MEP) in China and are publicly accessible from the year 2013 onwards (Zhang and Cao, 2015). We obtained the hourly in situ measurements from a total of 1413 air quality monitoring sites of the MEP network for 323 major cities over the model domain. The majority of those monitoring sites have been placed in the city center and are named urban assessing stations in the official document (MEP, 2013). These are meant to evaluate the overall level and trend of air quality for areas with the highest concentrations and highest population exposure. The placement criteria of urban assessing stations laid down in the legislation (MEP, 2013) ensure that the measurements are representative for urban areas. Stations are required to be well distributed within the developed area of the city and not too close to stationary emission sources (50 m) or roads (10–100 m depending on the traffic flow). The minimum number of monitoring sites required per city depends on both the urban population and city size, i.e., at least one station for an area of ∼ 50 km2 (Table 1). In addition, for areas with the concentration exceeding grade II of the national ambient air quality standard (i.e., the annual mean NO2 concentration of 40 µg m-3; MEP, 2012), the minimum required number of monitoring sites is increased by 50 %. MEP also operates other types of measurement sites, including regional and background stations to assess the background air pollution levels and pollutant transport, and source impact and traffic stations close to emission sources. However, only megacities like Beijing and Guangzhou operate such non-urban stations. The fact that urban observations dominate should be kept in mind when comparing the observations with the model results. The horizontal resolution of the model is limited to 0.25∘, which will cause representativeness errors (biases) when comparing the measurements from city stations with the mean of a grid box of the simulations, which can also include rural areas. Note that only the measurements for the dates with 24 h valid measurements (larger than 0) are used for the following analysis in this study.

Figure 3a displays the heterogeneous spatial distribution of monitoring sites at the scale of the model grid cell. The over 1000 monitoring sites are allocated to a total of 594 grid cells based on their geolocations. The sites belonging to the grid cells with one, two, and three sites account for 17, 21, and 22 % of the total, respectively (Fig. 3b). We calculated the averaged distance between monitoring sites by averaging individual pairwise distances for every two stations in the same grid cell. Because most monitoring sites are urban stations and are clustered in the city areas, which are often much smaller than the area of a grid cell (∼ 600 km2), the averaged distance is rather small with an average of 3.6 km for all grid cells as shown in Fig 3c. For megacities with significantly larger built-up areas and thus more monitoring sites, the distribution of sites is more homogeneous over the grid cell and results in a lager distance between stations. The average distance increases from 5 km for grid cells with only one pair of stations to 11 km for those with over eight stations.

In our analysis, we excluded in situ measurements from cities with unexpected discrepancies between urban and suburban stations. Because only large cities potentially place the monitoring sites outside urban areas related to the rapid expansion of built-up areas, we classified stations as urban and suburban by visually inspecting satellite imagery from Google Earth for large cities with over four stations. We calculated the annual mean NO2 of each station. When the NO2 concentration of urban stations is less than that of suburban stations, the measurements behave differently than expected. The cities (four in total) detected to have unexpected measurements are labeled as “unselected” and discarded from the validation dataset. Note that suburban stations presenting higher NO2 levels than urban stations but close to large emission sources, e.g., industrial park and airport, are understandable and thus are not excluded from the database. Figure 4 presents the daily average surface NO2 abundance for the city of Xi'an. Only the dates with 24 h valid measurements (lager than 0) are used for the time series illustration here. The expected enhancement in winter highlighted by both urban (red line) and suburban (blue line) stations has not been detected for the urban station with lower annual mean NO2 abundance than suburban stations (black line), which provides further support for excluding the Xi'an measurements from the model evaluation.

NO2 concentrations are measured using commercial chemiluminescence analyzers (Zhang and Cao, 2015), which are subject to a systematic overestimation of ambient NO2 concentrations (Steinbacher et al., 2007). NO2 is catalytically transformed into NO by a molybdenum converter and subsequently measured with chemiluminescence. However, other reactive oxidized nitrogen compounds (NOz) such as peroxyacetyl nitrate (PAN) and nitric acid are also partly converted to NO, resulting in an overestimation of the measured NO2.

We applied a correction factor proposed by Lamsal et al. (2008) to account for the interferences of other oxidized nitrogen compounds, based on the modeled ratio of NO2 to NOz. The correction factor (CF) was calculated from the local chemical concentrations as follows: CF=NO2NO2+∑AN+0.95×PAN+0.35×HNO3, where ∑AN is the sum of all alkyl nitrate concentrations.

Figure 5 shows the seasonal means of the correction factors determined with concentrations of the interfering species predicted by the CHIMERE model driven by the DECSO inventory. Consistent with the findings in Europe (Huijnen et al., 2010) and the US (Lamsal et al., 2008), the correction factor (difference with the ideal value of 1.0) is largest over polluted urban regions, where NOx is a larger fraction of total oxidized nitrogen compounds. The correction factor tends to be closer to unity in winter, when the NOx photochemistry is slower and thus NOx has a larger relative contribution to total oxidized nitrogen compounds. The correction factor derived from simulations with the MIX inventory (not shown) shows a similar pattern to Fig. 5, but with a larger number of values close to 1 related to the larger emissions. Hourly correction factors for individual hours of each day during the year for all individual stations have been applied to the in situ measurements. It is difficult to quantify the accuracy of the correction factors and errors, as the collocated measurements of other oxidized nitrogen compounds are not publicly available. We used the standard deviation of the daily means of correction factors within a season as a measure of its uncertainty. The average standard deviations for all sites are 10 %, which is comparable to the uncertainty level pointed out by the study of McLinden et al. (2014).

We compare the modeled surface NO2 with the corrected in situ measurements throughout China. In general, the spatial distribution of annual mean NO2 concentrations from the CHIMERE model simulations is well in line with that from in situ measurements, with a correlation coefficient of over 0.7. However, the modeled NO2 is biased compared to ground measurements. The differences of annual mean NO2 concentrations between simulations and measurements are given in Fig. 6. The CHIMERE simulations with the DECSO inventory show considerably lower NO2 concentrations than the in situ measurements, with a negative difference for nearly 90 % of all grid cells. Conversely, the CHIMERE simulations with the MIX inventory are generally higher than the in situ measurements for grid cells corresponding to large cities: A positive bias is found for 70 % of the grid cells with over four monitoring sites.

Grid cells are classified into five categories, i.e., mountainous, northern, < four stations, densely located, and main sample, and the corresponding scatter plots of corrected measurements against simulations are shown in Fig. 7. We define a grid cell as “mountainous” where the average elevation is higher than 1000 m and the standard deviation of elevations is over 15 % of the mean, based on the topographic data from the 30 arcsec global land topography “GTOPO30” archived by the US Geological Survey (available at https://lta.cr.usgs.gov/GTOPO30, last access: 10 July 2017, rescaled to 0.05∘). The grid cells higher than 45∘ N are classified as “northern”. The grid cells with less than four measurement stations are classified as “< four stations”. The grid cells with only densely located stations (see definition later in this section) are classified as “densely located”. Note that the priority of the category of mountainous, northern, < four stations, and densely located is from high to low when we perform the classification in this study. For instance, for grid cells that meet the criteria of both mountainous and northern, we classify them as mountainous. The remaining grid cells are classified as “main sample”. The results for the daytime period (08:00–19:00 LT) are displayed separately. The correlation coefficient, regression slope, and root-mean-square error for the individual categories compared to measurements are given in Table 2.

Significant regional differences are found. The small slope over mountainous regions could be related to model limitations to resolve cities in the valleys. Furthermore, we may expect difficulties for the model in describing NO2 concentrations over complex terrain. For mountainous regions, the lower slope may also be related to the large uncertainties in the meteorological parameters associated with the difficulties in resolving the characterization of small-scale orography in the ECMWF model (Beljaars et al., 2004). The errors on meteorological parameters, such as mixing height and temperature (Hongisto, 2005) and wind fields (Minguzzi et al., 2005), can introduce biases for air quality simulations. In addition, the accuracy of the DECSO algorithm highly relies on the appropriate wind fields because DECSO performs trajectory analysis to account for NOx transport away from the source when calculating the sensitivity of concentrations to emissions (Mijling and van der A, 2012). In this way, uncertainties in meteorological parameters are amplified in the DECSO inventory, resulting in a worse agreement for mountainous areas (r = 0.51) compared to the MIX inventory (r = 0.77).

The CHIMERE model accurately reproduces the spatial variability in NO2 for northern grid cells with a high correlation coefficient of 0.92 and 0.81 for the simulations with the DECSO and the MIX inventories, respectively, but with a large negative bias. The bias could be related to model uncertainties in NOx sinks for high latitudes (Ding et al., 2017b), indicated by the sensitivity studies of modeled NO2 columns to errors in chemical parameters associated with NOx sinks (Lin et al., 2012; Stavrakou et al., 2013). Additionally, the bias is particularly significant for the simulations with the DECSO inventory, showing a slope of merely 0.20. This could be further explained by the general underestimation of NOx emissions caused by the bias in NO2 tropospheric columns of DOMINO v2 for this area (Ding et al., 2017b), partly due to a bias in the calculation of air mass factor for retrievals at large solar zenith angles by the radiance transfer model (Lorente et al., 2017) and possibly biases in the estimated stratospheric background.

Figure 8 depicts the ratio of the simulated annual mean surface NO2 concentrations to the corrected in situ measurements sorted by the number of stations located in the same grid cell from small to large. It is interesting to note that the ratio is small for grid cells with less than four stations, but increases along with the increase in the number of stations from four to nine, ranging from 0.6 to 1.0 and 0.9 to 1.8 for the simulations with the DECSO and the MIX inventories, respectively. The trend in the ratio suggests that the representativeness of in situ measurements for the average NO2 levels of a grid cell improves with increasing numbers of stations or city size. For grid cells with less than four stations, the model with its limited spatial resolution cannot be expected to accurately resolve the spatial gradient of pollutants towards the city center in relatively smaller urban areas. Similarly, the in situ measurements for stations located close together are expected to be less representative of the grid cell mean compared to the homogeneously distributed stations. This is in agreement with the tendency that grid cells with lower average measurement station distances (< 10 km) tend to show lower ratios (< 0.5) in Fig. 8, in particular for grid cells with a larger number of stations. We select the grid cells with over four stations but lower average distances than the 10 % percentiles in Fig. 3c and name them densely located. We analyze the performance of the model in those grid cells with only densely clustered stations (labeled with “L” in Fig. 6). Not surprisingly, the simulations for those grid cells show larger discrepancies compared to the measurements, and also the correlation coefficient of simulations with the DECSO inventory drops down to a rather low value of 0.55 (Table 2).

We exclude grid cells in the special categories discussed above (i.e., mountainous, northern, < four stations, and densely located stations) to draw conclusions on the ability of the model to reproduce the measurements. Statistical values (correlation, slope, root-mean-square error) for the remaining grid cells (main sample) are given along with the plots in Fig. 7. A slope of 0.74 and 1.3 is found for the simulation with the DECSO and the MIX inventories, respectively. As mentioned before, the majority of stations are located in urban, populated, and polluted areas and the model resolution of 0.25∘ will not be enough to represent the existing NO2 gradients, thus we may expect a negative representativity offset in the modeled surface concentration, even for the main sample obtained after data screening (Irie et al., 2012; Lin et al., 2014). We select grid cells with over four stations, which potentially place one or two stations in background areas, to give a rough estimate of the offset. The background stations are defined as stations located far away from urban areas on the basis of a visual inspection of satellite imagery from Google Earth. Not surprisingly, the measurement from the background station which is expected to better represent the grid cell mean is smaller than the average value of measurements from all stations located in the same grid cell. The ratios of annual mean measurements from the background stations to the mean of corresponding measurements from all stations range from 0.64 to 0.86, with an average of 0.74. That is, the average negative representativity offset may reach 25 %. The ratio is closer to 1 in winter (0.83 on average) due to the reduced spatial gradients in NO2 caused by a longer NOx lifetime, which will be discussed in detail in Sect. 3.2. Thus, the slopes of 0.74 and 1.3 for DECSO and MIX actually indicate a slightly negative and more significantly positive bias, respectively.

A positive bias may indicate an overestimation of NOx emissions, or errors in the spatial downscaling of the bottom-up emission totals, although biases from the description of the chemistry, transport, and removal processes in the model cannot be ruled out. The overestimation of the MIX results over large cities are consistent with previous findings that regional inventories like MIX have large positive biases in urban areas (Zheng et al., 2017). The reason for the positive biases will be discussed in detail later in this section. Uncertainties in the DECSO results may be attributed to biases in the OMI tropospheric NO2 column densities, or representation errors introduced by the projection of the CTM onto the measured NO2 satellite footprint (Ding et al., 2017b). OMI NO2 observations have been reported to be systematically smaller than those from ground-based measurements (e.g., MAX-DOAS) over polluted regions (Shaiganfar et al., 2011; Ma et al., 2013; Ialongo et al., 2016) due to their different spatial representativeness (Irie et al., 2012; Lin et al., 2014) and uncertainties in NO2 vertical column retrieval, including the shielding effect of aerosols (Shaiganfar et al., 2011) and the varying observation geometry (Vasilkov et al., 2017). In addition, the fact that the adopted model resolution is not sufficient to accurately model nonlinear effects in the NO2 loss rate may contribute to the negative bias (Valin et al., 2011).

Regional bottom-up inventories tend to have large positive biases in urban areas. Those inventories usually downscale local emissions from regional totals (provincial totals are used in the MEIC/MIX inventory for China) and distribute them to grid cells using spatial proxies (e.g., population density and GDP). However, the spatial proxies may not match the locations of the individual emitting sources, especially for industrial plants located far away from urban centers that tend to have a larger population density and GDP (Liu et al., 2017). Such a decoupling will result in an overestimation of emissions over urban areas, which has been proven by the comparison of proxy-based regional inventory with high-resolution urban inventories developed from the extensive use of information of individual emitting sources (Zheng et al., 2017).

In order to better compare the spatial distributions of the two inventories and identify the sensitivity of model performance to spatial distributions of emissions, we further evaluate the impact of the spatial distribution of emissions on simulating NO2 by applying the same spatial proxy for NOx emissions in both inventories. We scale the total amount of emissions of the 2015 DECSO inventory over the domain adopted in this study to that of the 2010 MIX inventory but kept the DECSO spatial distribution (hereinafter referred to as the corrected DECSO inventory; see Fig. 2). We then compare the modeled NO2 using the corrected DECSO inventory with in situ measurements in Fig. 6c. It is interesting to see that many high values in the MIX simulation are not reproduced by the simulation with the corrected DECSO inventory. We further assess the simulation results with the corrected DECSO inventory in Fig. 7c. The simulation with the MIX inventory tends to cluster the pollutants more over urban areas than that with the corrected DECSO inventory, indicating that the modeled NO2 is sensitive to the spatial distribution of emissions. The large bias in the MIX inventory is reduced significantly, with a slope decreasing from 1.3 to 1.0, which suggests an improvement of the distribution of emissions between urban and suburban or rural areas.

Note that due to the lack of the 2015 inventory, the use of the 2010 MIX emissions for other species including SO2, CO, and non-methane volatile organic compounds (NMVOCs) in both the MIX and the DECSO simulations may introduce uncertainties in simulating NO2. The anthropogenic emissions of SO2, CO, and NMVOCs for China have been reported to decrease by 2, 5, and increase by 21 % from 2010 to 2015, respectively (Li et al., 2017b). In gas-phase chemistry, the principal sink of NOx is oxidation to HNO3. The influence of the growth in NMVOCs on the oxidizing power of the atmosphere is partially compensated for by the reduction in CO, as CO and hydrocarbons play similar roles in depleting oxidants following the HOx–NOx–CO–hydrocarbon chemical mechanisms (Jacob, 2000). Additionally, SO2 influences NO2 concentrations by forming aerosols, concentrations of which have an impact on photolysis rates and thus photochemical reaction rates associated with NOx (Mailler et al., 2017). However, the emission changes are rather small compared to the uncertainties in bottom-up estimates, which are even smaller than the discrepancies among estimates from different bottom-up inventories. Thus we believe the uncertainties arising from the use of the 2010 inventory are not significant. A sensitive analysis will be further expected to quantify the influence of emissions of other species on simulated NO2.

Figure 9 compares the monthly mean NO2 concentrations simulated by the CHIMERE model using the two inventories with the in situ measurements. The spatial correlation between the modeled NO2 concentrations and the in situ measurements shows a weak dependence on season, which is slightly worse in summer (July). The correlation coefficients range from 0.64 (July) to 0.73 (January) and from 0.80 (July) to 0.83 (January) for the simulations with the DECSO and the MIX inventories, respectively. A possible explanation for the somewhat higher correlation in January is the smaller model error in winter than in summer, as indicated by previous findings in both China (Lin et al., 2012) and Europe (Huijnen et al., 2010). This may be related to the difficulties in resolving the more active NOx photochemistry in summer by the model. For instance, the model with a horizontal resolution of 0.25∘ is not able to fully resolve the spatial gradients of NO2 close to strong emission sources, but such an impact is smaller in winter than in other seasons, as the NO2 gradients in a grid cell are smeared out due to the longer NOx lifetime in winter.

The seasonal difference is pronounced when comparing the magnitude of the NO2 concentrations in Fig. 9. In general, a smaller ratio between modeled NO2 concentrations and in situ measurements is detected in winter. The ratio reaches the lowest values in January, which is consistent with the general underestimation of simulations in winter as indicated by Lin et al. (2012). For simulations with the DECSO inventory, the ratio deviating more significantly from unity in winter might be due to systematic biases in the OMI NO2 observations during winter as well. Biases in OMI NO2 column densities over polluted regions are introduced by the high aerosol loading, most of which are scattering aerosols in China, as aerosols are not explicitly considered in the cloud retrieval or the air mass factor calculation in the operational NO2 product (Castellanos et al., 2015; Chimot et al., 2016; Wang et al., 2017). The aerosols' effect may be more significant in winter due to the higher aerosol concentrations and larger solar zenith angle (Ma et al., 2013). Additionally, the DECSO algorithm is more vulnerable to biased observations as a result of the smaller number of useful observations in wintertime because of the filtering of snow-covered regions. Conversely, the simulations with the MIX inventory show ratios ranging from 1.09 to 1.22 in the second half of the year. This may signal an overestimation of total emissions, as pointed out in Sect. 3.1. In addition, the assumptions used in the MIX inventory for the distribution of monthly emissions over the year may also contribute to the bias. For example, higher power and industrial emissions are assumed in the second half of the year due to larger industrial productions and thus power generations to meet the annual total production target (Li et al., 2017a).

Figure 10 presents the diurnal variability in hourly-averaged surface NO2 concentrations. The simulations with both inventories and the in situ measurements exhibit a broadly similar daily variation (r = 0.81). The distinct peak in NO2 concentrations in the morning hours (around 08:00 LT) and in the afternoon (around 20:00 LT) detected by the measurements has been well captured by the model, which can be attributed to increasing (traffic) emissions in the rush hours indicated by the Selected Nomenclature for sources of Air Pollution Prototype (SNAP) diurnal profiles of emissions (Menut et al., 2012) adopted in the CHIMERE model (grey line). Both simulations and measurements show a drop in NO2 concentrations during daytime with the same timing and amplitude, related to the varying chemical loss rate of NO2 driven by NOx photochemistry. However, the disagreement between simulated and measured values is larger at night, which may point to problems regarding the treatment of boundary layer mixing. NO2 concentrations simulated by the model cannot reproduce the observed temporal pattern at night but present constantly high values, probably caused by unrealistically low boundary layer heights and too little vertical turbulence in the model (Bessagnet et al., 2016). This has been further confirmed by the earlier evaluation of the diurnal cycle of trace gases as modeled by CHIMERE in Lampe et al. (2009).

We separately evaluated the model performance for the daytime period (08:00–19:00), when the pattern of diurnal variations simulated by the CHIMERE model is closer to what is observed by the in situ measurements. Not surprisingly, a larger negative slope of 0.64 is obtained for the simulation with the DECSO inventory compared to the surface observations, while the slope for the simulation with the MIX inventory has been reduced significantly to a value of 1.1 (Table 2) due to the tendency of overestimating NO2 concentrations during night in the model. Note that the slope close to unity for the simulation with the MIX inventory during daytime does not necessarily imply a perfect emission inventory, but still indicates a potential overestimation because we expect a slope smaller than 1 (in the range of 0.64–0.86; see Sect. 3.1) when comparing model simulations with in situ measurements which are mainly situated in populated high-concentration areas.