IFS is ECMWF's global numerical weather prediction system. As part of the past series of European projects MACC and now of CAMS, IFS has been developed to represent optionally chemical processes in the troposphere and in the stratosphere. Flemming et al. (2015) provide a detailed description of the modeling of chemical processes in the IFS, and Inness et al. (2015) describe the data assimilation aspects.

For the work presented here, the version of IFS used is Cycle 43R1 (see documentation at https://www.ecmwf.int/en/forecasts/documentation-and-support/changes-ecmwf-model/ifs-documentation, last access: 7 December 2018). The model is run globally at a resolution of T511 (about 40 km) on the horizontal and with 60 levels on the vertical extending up to the top of the stratosphere. The chemical package used originates from the TM5 chemistry and transport model (Huijnen et al., 2010). It has been fully integrated into the IFS code and comprises 54 tracers and 120 reactions focusing on tropospheric-ozone–CO–NMVOC–NOx chemistry. In the configuration used here, stratospheric ozone is modeled with a simple linearized scheme. Aerosols are represented using the scheme described by Morcrette et al. (2009), which includes five species: dust, sea salt, black carbon, organic carbon, and sulfates. Tracers are transported using the semi-Lagrangian scheme available in IFS with a mass fixer activated in order to minimize mass nonconservation.

During the study period, IFS has been run twice daily (5-day forecasts) assimilating a range of satellite chemical data on top of the full list of meteorological satellite and non-satellite data that ECMWF uses for its medium-range weather forecasts. Table 1 indicates the satellite data streams actively assimilated for the experiments presented here. As a result, IFS forecasts benefit from all these observations to afford a realistic representation of large scales for weather parameters as well as, to some extent, for chemical variables (species assimilated).

IFS used the MACCity emission dataset updated for the year 2017. Biogenic emissions of volatile organic compounds (VOCs) were taken from a climatology of a multiyear Model of Emissions of Gases and Aerosols from Nature (MEGAN) simulation. Daily emissions from biomass burning were derived from satellite retrieval of fire radiative power (FRP) from the MODIS instruments by the Global Fire Assimilation System (GFAS; Kaiser et al., 2012). The observed fire emissions from the day before the forecast start are used for all 5 days of the forecast. Desert dust and sea salt emissions were simulated online for each time step based on the IFS meteorological fields and the land use.

As part of CAMS, the chemical configuration of IFS benefits from routine detailed evaluations. Validation reports are produced quarterly and can be found here: http://atmosphere.copernicus.eu/quarterly_validation_reports (last access: 7 December 2018). The report for the period March–May 2017 provides insight on the overall performance of the runs that are also presented here. Further information about the IFS code can be obtained from Vincent-Henri Peuch (vincent-henri.peuch@ecmwf.int) and on the website https://www.ecmwf.int/en/about/what-we-do/environmental-services/copernicus-atmosphere-monitoring-service (last access: 7 December 2018).

CHIMERE is a regional chemistry transport model used for analysis, scenarios, and forecast (Menut et al., 2013a). When used in the forecast mode, the model provides local-scale information (to be compared with data from numerous air quality networks) or regional-scale information (e.g., the French PREV'AIR and the CAMS systems). CHIMERE is an open-source model, freely distributed at http://www.lmd.polytechnique.fr/chimere/ (last access: 7 December 2018). In this version, CHIMERE is used in off-line mode at a spatial resolution of 0.25∘ (about 25 km). It is forced by pre-calculated hourly meteorological fields for the dynamics and by several emissions fluxes for the chemistry. The emissions are pre-calculated or online estimated in the model with anthropogenic emissions (MEIC 2010), biogenic emissions with the online MEGAN (Guenther et al., 2006), mineral dust (Menut et al., 2013b), and biomass burning emissions (Turquety et al., 2014). The gas-phase chemistry is calculated using the MELCHIOR2 mechanism, and the aerosols are represented using a distribution of 10 bins, from 40 nm to 40 µm to describe both number and mass well. The chemical boundary conditions are provided by the LMDz-INCA model for gas and particles (Szopa et al., 2009), except for mineral dust, which is extracted from global GOCART simulations (Ginoux et al., 2001). Further information about the implementation of the model for air quality forecasts in China can be obtained from Ronald van der A (avander@knmi.nl) at KNMI and on the website http://www.lmd.polytechnique.fr/chimere/CW-download.php (last access: 7 December 2018).

WRF-Chem is a mesoscale non-hydrostatic meteorological model (Skamarock et al., 2008) coupled “online” with chemistry that simultaneously predicts meteorological and chemical components of the atmosphere (Grell et al., 2005; Fast et al., 2006).

The model version used at the Max Planck Institute for Meteorology (MPIM), WRF-Chem-MPIM, is based on version 3.6.1 of the WRF-Chem model coupled to the gas-phase chemistry and the aerosol microphysics schemes provided by the Model for Ozone and Related Chemical Tracers (MOZART-4; Emmons et al., 2010) and the Model for Simulating Aerosol Interactions and Chemistry (MOSAIC; Zaveri et al., 2008), respectively. Aerosol sizes are represented by four consecutive bins, and the formation of secondary organic aerosol (SOA) from anthropogenic precursors is parameterized according to Hodzic and Jimenez (2011).

Two nested model domains with horizontal resolutions of 60 km (Asian continent from India to Japan) and 20 km (eastern China), respectively, are implemented. The vertical grid is composed of 51 levels extending from the surface to 10 hPa (∼30 km). A more complete description of the selected physical and chemical options is provided in the WRF and in the WRF-Chem user's guides under http://www2.mmm.ucar.edu/wrf/users/docs/user_guide_V3.6/ARWUsersGuideV3.6.1.pdf (last access: 7 December 2018) and https://ruc.noaa.gov/wrf/wrf-chem/Users_guide.pdf (last access: 7 December 2018).

The WRF-Chem-MPIM model forecasts are initialized and forced at the lateral boundaries every day by 6-hourly meteorological analysis data from the NCEP Global Forecast System (GFS) at 0.5∘ resolution. For the chemical and aerosol species, 6-hourly datasets are provided by the global operational forecasting system implemented within the Copernicus Atmospheric Monitoring Service project (Flemming et al., 2015).  More information on the model's configuration can be obtained from Idir Bouarar (idir.bouarar@mpimet.mpg.de) at the Max Planck Institute for Meteorology and on the website http://www2.mmm.ucar.edu/wrf/users/downloads.html (last access: 7 December 2018).

FMI uses the SILAM version 5.5 (Sofiev et al., 2015a, b). SILAM includes a meteorological preprocessor for diagnosing the basic features of the boundary layer and the free troposphere from the meteorological fields provided by various meteorological models (Sofiev et al., 2010). The dry-deposition scheme for particles is described in Kouznetsov and Sofiev (2012). The surface resistance model for gases is based on a modified Wesely scheme (Wesely, 1989).

The gas-phase chemistry was simulated with CBM-IV, with reaction rates updated according to the recommendations of IUPAC (http://iupac.pole-ether.fr, last access: 7 December 2018) and JPL (http://jpldataeval.jpl.nasa.gov, last access: 7 December 2018) and the terpenes oxidation added from the 2005 Carbon Bond (CB05) chemical mechanism reaction list (Yarwood et al., 2005). The sulfur chemistry and secondary inorganic aerosol formation is computed with an updated version of the DMAT scheme (Sofiev, 2000), and secondary organic aerosol formation is computed with the volatility basis set (VBS; Donahue et al., 2006), with the volatility distribution of anthropogenic organic carbon (OC) taken from Shrivastava et al. (2011).

The MACCity land-based emissions are used together with the Ship Traffic Emission Assessment Model (STEAM). The simulations include sea salt emissions as in Sofiev et al. (2011), biogenic VOC emissions as in Poupkou et al. (2010), wild land fire emissions as in Soares et al. (2015), and desert dust.

The grid cell size was roughly 15 km × 10 km (0.125∘ × 0.125∘) covering the whole of China, India, Japan, and several countries of Southeast Asia (7∘ N, 67∘ E) – (54∘ N, 147∘ E). The Asian forecasts are nested into the SILAM global air quality forecasts (http://silam.fmi.fi, last access: 7 December 2018), from where they take lateral and top boundary conditions. The initial conditions for each run are taken from the previous day's forecast or, in case of failure, from global computations. Detailed information about the SILAM modeling system can be obtained from Mikhail Sofiev (Mikhail.Sofiev@fmi.fi) and from Rostislav Kouznetsov (rostislav.kouznetsov@fmi.fi) and on the website of the Finnish Meteorological Institute (http://silam.fmi.fi/).

EMEP/MSC-W (hereafter referred to as “EMEP model”) is a 3-D Eulerian chemical transport model described in detail in Simpson et al. (2012). Although the model has traditionally been aimed at European simulations, global modeling has been possible for many years (Jonson et al., 2010; Wild et al., 2012). The EMEP configuration for the present study covers the east Asian domain (15–55∘ N × 90–135∘ E) with a horizontal resolution of 0.1∘ × 0.1∘ (longitude–latitude). The model uses 20 vertical levels defined as sigma coordinates. The 10 lowest levels are within the planetary boundary layer (PBL), and the top of the model domain is at 100 hPa.

Particulate matter (PM) emissions are split into elementary carbon, organic matter (OM) (here assumed inert), and the remainder, for both fine and coarse PM. The OM emissions are further divided into fossil fuel and wood-burning compounds for each source sector. As in Bergström et al. (2012), the OM / OC ratio of emissions by mass is assumed to be 1.3 for fossil-fuel sources and 1.7 for wood-burning sources. The model also calculates windblown dust emissions from soil erosion. Secondary PM2.5 aerosol consists of inorganic sulfate, nitrate, ammonium, and SOA; the latter is generated from both anthropogenic and biogenic emissions (anthropogenic SOA and biogenic SOA, respectively), using the VBS scheme detailed in Bergström et al. (2012) and Simpson et al. (2012).

Model updates since Simpson et al. (2012), resulting in EMEP model version rv4.9 as used here, have been described in Simpson et al. (2016) and references cited therein. The main changes concern a new calculation of aerosol surface area, revised parameterizations of N2O5 hydrolysis on aerosols, additional gas-aerosol loss processes for O3, HNO3, and HO2, a new scheme for ship NOx emissions, and the use of new maps for global leaf area (used to calculate biogenic VOC emissions) – see Simpson et al. (2015) for details. The EMEP model, including a user guide, is publicly available as open-source code at https://github.com/metno/emep-ctm (last access: 7 December 2018). For more details, please contact Michael Gauss (michael.gauss@met.no).

The EMEP forecasts are driven by 3-hourly meteorological forecast data from the ECMWF IFS model at 0.1∘ resolution. As for WRF-Chem, 6-hourly datasets for the chemical and aerosol species are provided by the global operational forecasting system implemented within the Copernicus Atmospheric Monitoring Service project.

LOTOS-EUROS is a three-dimensional regional chemistry transport model for simulation of trace gases and aerosol concentrations in the boundary layer. Meteorological input is obtained from an off-line model, in this study from ECMWF. The model is of intermediate complexity allowing long-term model simulations. For a detailed model description, we refer to Manders et al. (2017) and references therein.

In this study LOTOS-EUROS version 1.10 was used to simulate air quality over China. The configuration is described by Timmermans et al. (2017), who adopted this version of the model to investigate the origin of fine particulate matter across China using a source apportionment technique. Through a one-way nesting procedure a simulation over east China was performed on a resolution of 0.25∘ longitude by 0.125∘ latitude, approximately 21 km by 15 km. This domain is nested in a larger domain covering China almost entirely with a resolution of 1∘ longitude by 0.5∘ latitude, approximately 84 km by 56 km. Chemical boundary conditions for the coarse resolution domain were taken from the CAMS global modeling framework (Flemming et al., 2015) and include trace gasses and aerosols. In the vertical, the model used a boundary layer approach with five layers: a surface layer of 25 m, a well-mixed boundary layer, two reservoir layers, and a layer for the free troposphere. The boundary layer height therefore defines the vertical structure of the model, and is here taken from the meteorological input. More details about the code can be obtained by contacting Renske Timmermans (renske.timmermans@tno.nl) at TNO or by consulting the website https://lotos-euros.tno.nl/ (last access: 7 December 2018).

WRF-Chem-SMS hosted at the Shanghai Meteorological Service is based on WRF-Chem (Grell et al., 2005) version 3.2. The Regional Acid Deposition Model version 2 (RADM2; Chang et al., 1989) is used to represent gas-phase chemistry. ISORROPIA II is implemented to treat thermodynamic equilibrium for inorganic aerosols (Fountoukis and Nenes, 2007), and the Secondary ORGanic Aerosol Model (SORGAM) (Schell et al., 2001) is used to parameterize secondary organic aerosol formation. A Madronich TUV scheme is applied for photolysis (Madronich and Flocke, 1999; Tie et al., 2003). The model domain covers the eastern region of China with horizontal resolutions of 6 km and 28 vertical layers. Biogenic emissions are calculated online using MEGAN (Guenther et al., 2012). The multi-resolution emission inventory for China (MEIC inventory, http://www.meicmodel.org/ (last access: 7 December 2018); Li et al., 2014; Liu et al., 2015) for the year 2010 is used to represent anthropogenic emissions.

The modeling system is initialized and forced at the lateral boundaries every day by 6-hourly data from the NCEP GFS at 0.5∘ resolution. For chemical species, a previous modeling result is used for initial conditions. MOZART-4 historic data are employed as the gaseous chemical lateral boundary, and a real-time forecast of dust from the WRF-Dust model is employed as a dust lateral boundary every 6 h. More detailed information can be found in Zhou et al. (2017) and by contacting Jianming Xu (metxujm@163.com) at the Shanghai Meteorological Service.

A regional air quality operational forecasting system was developed at Nanjing University, China, on the basis of the WRF–CMAQ model. The versions adopted for the WRF (Weather and Forecasting) and CMAQ (Community Multiscale Air Quality) models are V3.5 and V4.7.1, respectively. Two nested domains with horizontal resolutions of 36 and 12 km are adopted for the forecasts. The outer domain covers the entire continental region of China as well as surrounding countries in east Asia. The inner domain mainly focuses on the densely populated area of eastern China. The number of grid points adopted for the WRF model are 170×130 and 202×226, respectively with 51σ layers in the vertical (12 layers below 1.5 km a.g.l.) between the surface and the model top at 50 hPa. The CMAQ model is applied to the same domains but with three grid cells removed at each lateral boundary of the WRF domains. Overall, 15 vertical layers are selected from the 51 WRF layers, including about 8 layers in the boundary layer and 7 layers in the free troposphere.

Anthropogenic emissions are supplied off-line from the MIX inventory (Li et al., 2017). Terrestrial biogenic emissions are calculated off-line using MEGAN v2.04 (Guenther et al., 2006). Sea salt emissions are incorporated into the AERO4 aerosol module and calculated online in CMAQ. Windblown dust is derived online from the WRF-Dust model. Open biomass burning emissions are not considered here. It should be noted that the anthropogenic emissions are not fixed in this system but are automatically adjusted every week according to the system performance in the past week. The adopted scaling factors are determined from the deviation between the weekly averaged calculated and observed concentrations of SO2, NOx, CO, PM2.5, and PM10 in 334 Chinese prefectures.

The system provides a forecast every day for the next 192 h. The NCEP GFS's products at 00:00 UTC are used for the initial and boundary conditions of the WRF model with a resolution of 0.5∘ and with a 3 h interval. For the CMAQ model, the boundary conditions are created using ideal profiles, and the chemical initial fields are initialized from the previous forecasting. In addition, hourly averaged observed concentrations of SO2, NO2, CO, O3, PM2.5, and PM10 from 1415 national control air-quality-monitoring sites are assimilated into the initial fields using an optimal interpolation method (Lorenc, 1981). More information on the code can be obtained from Fei Jiang (jiangf@nju.edu.cn) at Nanjing University. Information on WRF–CMAQ is also available on the website http://carbon.nju.edu.cn/cn/ (last access: 7 December 2018) and https://www.epa.gov/cmaq/cmaq-models-0 (last access: 7 December 2018).

The Community Multiscale Air Quality (CMAQ) model is a 3-D Eulerian chemical transport model that explicitly simulates emissions, gas-phase, aqueous, and mixed-phase chemistry, advection and dispersion, aerosol thermodynamics and physics, and wet and dry deposition. A detailed description and an evaluation of the CMAQ model are available in the papers by Byun and Schere (2006), Foley et al. (2010), and Appel et al. (2017). Several studies have applied the CMAQ model to study the air quality in China. For example, Zheng et al. (2015) used the WRF–CMAQ model to study the impact of heterogeneous chemistry during the January 2013 haze episode. Hu et al. (2016) performed a 1-year retrospective simulation using the WRF–CMAQ model to study the O3 and particulate matter formation with a detailed evaluation. Here the CMAQ version 5.0.2 is adopted and includes the CB05 (Yarwood et al., 2005) to represent the gas-phase chemistry. The fifth-generation modal CMAQ aerosol model (aero5) is adopted to formulate the aerosol chemistry and dynamics (Carlton et al., 2010).

In this version, CMAQ is used in an off-line mode. It is forced by pre-calculated hourly meteorological fields for the dynamics and by several emissions fluxes for the chemistry. Meteorology fields that drive chemical transport are produced by the SMS-WARMS. The SMS-WARMS has been extensively evaluated and provides weather predictions in eastern China. The modeling domain consists of 760 by 600 horizontal grids at 9 km resolution, with 51 layers in the vertical. As a subdomain of the SMS-WARMS run, the CMAQ domain consists of 430 by 370 horizontal grid cells at 9 km resolution. In the vertical, 26 layers are applied.

The anthropogenic emissions are based on the monthly HTAP v2 dataset (http://edgar.jrc.ec.europa.eu/htap_v2/, last access: 7 December 2018) (Janssens-Maenhout et al., 2015) for the year 2010. As suggested by operational forecasting results, the HTAP NOx, SO2 emissions are adjusted to account for rapid economic growth in the region. Biogenic emissions are estimated by MEGAN version 2.10 (Guenther et al., 2012). Currently, dust and biomass burning emissions are not included.

For the SMS-WARMS model forecasts, the NCEP GFS output at 0.5∘ is used as a background for the ADAS data assimilation scheme, which ingests many local observations (e.g., radar and buoys), and to provide lateral boundary conditions. The chemical boundary conditions are currently based on the default vertical profiles of gaseous species and aerosols in CMAQ that represent clean-air conditions. For more details, please contact Ying Xie (yxie33@outlook.com) at the Shanghai Meteorological Service. The CMAQ code available on the U.S. EPA modeling site https://github.com/USEPA/CMAQ/ (last access: 7 December 2018).

We first compare the March 2017 monthly mean concentrations of different chemical species calculated by seven models (IFS, LOTOS-EUROS, EMEP, SILAM, WRF-Chem-MPIM, WRF-Chem-SMS, and CHIMERE) with surface measurements reported at different sites in the eastern part of China (http://pm25.in/).

Figure 7a shows the calculated and observed surface concentrations of carbon monoxide (CO). We first note the substantial differences that exist between the individual model forecasts, probably reflecting differences in the adopted emissions or in the atmospheric production resulting from the oxidation of volatile organic compounds in the planetary boundary layer. Observations indicate that CO concentrations are generally higher than 900 ppbv, except near the southeastern coast and in the southwestern part of the country, where the values are as low as 500 to 700 ppbv. The models show considerably lower values, ranging from about 300 to 500 ppbv. The regions with the highest mean concentrations are located in the North China Plain (NCP), where values higher than 1200 ppbv are recorded. Relatively high values (close to 1000 ppbv) are also found in some urban areas (e.g., Hong Kong) near the south coast of the country.

The models provide a rather different picture: most of them substantially underestimate the CO concentrations, in particular WRF-Chem-SMS, WRF-Chem-MPIM, EMEP, and LOTOS EUROS. Higher concentrations are derived by SILAM and IFS. These models, however, produce peak concentrations in the region of the Sichuan Basin in contrast with the observations. Only IFS reproduces the high concentrations observed in northern China, probably because in this particular model the initial conditions are constrained by assimilated observations. Clearly, the performance of the models regarding the calculation of CO concentrations is not satisfactory. The discrepancies may be attributed to an underestimation of CO emissions, to errors in the lateral boundary conditions, or indirectly to an underestimation of the emissions for primary hydrocarbons.

In the case of NO2 (Fig. 7b), the observations show that the surface concentrations are highest in the northeastern portion of China with a few urban hot spots. These patterns are reproduced well by the EMEP, SILAM, and IFS models. The other models also produce high concentrations in urban areas but with values that are lower than those provided by the monitoring stations.

The mean surface ozone concentrations derived from measurements are lowest (about 20 ppbv) in the central part of China and highest (30–40 ppbv) near the east coast (Shanghai region), the south coast, and the western part of China. Since nitrogen oxides tend to titrate ozone, the models that predict high NO2 concentrations derive the lowest ozone values (EMEP, SILAM, IFS). The high NO2 concentrations predicted by EMEP are probably related to the large emissions used as shown in Fig. 1. CHIMERE, WRF-Chem-SMS, and to a lesser extent WRF-Chem-MPIM overestimate the mean ozone concentration during March. All models, however, produce a minimum in the ozone concentrations in northeastern China, a pattern that is not visible in the observational data (Fig. 7c).

Finally, in the case of PM2.5 (Fig. 7d), the measurements suggest the presence of high concentrations (higher than 80 µg m-3) in the region between Beijing and Shanghai. High abundances of PM2.5 are derived in this region by IFS, SILAM and to a lesser extent by LOTOS-EUROS, EMEP, CHIMERE, and WRF-Chem-SMS. Interestingly, most models produce another marked hot spot in the region of the Sichuan Basin, while the observations suggest a less pronounced maximum with a more limited geographical extent.

We now focus on the time period during which the most intensive comparison between models has been performed. We first examine the time evolution of surface ozone, NO2, and PM2.5 produced by the different models for the time period ranging from 1 to 15 March 2017 and for the three large metropolitan areas: Beijing, Shanghai, and Guangzhou. In Fig. 8, we compare the median concentrations of the three species with the median values derived from the different measurements provided by the network of instruments deployed in the three cities. The median model values are represented by the red curves, while the shaded areas highlight the dispersion of the calculated concentrations around the median values.

Beijing. Here the predictions of the PM2.5 concentrations follow the observations very closely. Two events with relatively high aerosol loads are visible, the first one between 2 and 5 March and the second one on 11 March. In the case of NO2, the models reproduce the daily variability reported by the monitoring stations fairly well, but on average, they slightly overestimate the concentrations values. The high concentrations appearing between 2 and 5 March and between 10 and 11 March are well captured by the median of the models. Finally, the models reproduce the diurnal variability in the ozone concentrations, but they underestimate these concentrations by typically 20 µg m-3.

In order to measure the performance of the individual models involved in the present intercomparison, we have calculated statistical measures of the model results for the chosen period of 1–15 March 2017. These measures include the mean bias (BIAS), the mean normalized bias (MNMBIAS), the root mean square error (RMSE), the fractional gross error (FGE), and the correlation coefficient for ozone, NO2, and PM2.5 (Table 4). They apply to the data for the 37 cities considered in the MarcoPolo–Panda forecast system. The same statistical measures are also provided for the ensemble median.

When examining the mean bias of the ensemble median, the values are equal to -14.7, -3.0, and +3.7 µg m-3 for ozone, NO2, and PM2.5, respectively, to be compared to mean concentration values of the order of 50 µg m-3 for these three different species. Table 4 shows that in the case of ozone, individual models are characterized by biases ranging from -25.8 (SILAM) to +13.2 µg m-3 (WRF-Chem-SMS), with the smallest absolute value equal to 5.9 µg m-3 (CHIMERE) The corresponding numbers range from -20.7 µg m-3 (LOTOS-EUROS) to +11.2 µg m-3 (EMEP) with the smallest absolute bias of -2.0 µg m-3 (IFS) for NO2. For PM2.5, they range from -4.7 µg m-3 (LOTOS-EUROS) to +39.6 µg m-3 (IFS) with the smallest absolute value equal to -2.0 µg m-3 (CHIMERE). In general, during the period chosen for the intercomparison, the models underestimate the ozone and NO2 concentrations and overestimate the concentration of PM2.5. The table also shows that the RMSE for the median values for ozone, NO2, and PM2.5 are 32.8, 21.8, and 30.2 µg m-3, respectively. With some exceptions (CHIMERE and IFS for ozone, LOTOS-EUROS, for PM2.5), these values are lower than the RMSE derived by individual models. The highest values for RMSE are 44.7 µg m-3 (WRF-Chem-MPIM) in the case of ozone, 34.4 (LOTOS EUROS) in the case of NO2, and 71.3 (IFS) in the case of PM2.5. The smallest RMSE are equal to 27.0 µg m-3 (CHIMERE) in the case of ozone, 23.1 µg m-3 (IFS) in the case of NO2, and 27.3 µg m-3 in the case of PM2.5 (LOTOS-EUROS). The correlation coefficient for the ensemble median is of the order of 0.6 for the three species, which in most cases is higher than the values derived from individual model forecasts. There are a few exceptions, however. The correlation coefficients are higher in the forecast of ozone by CHIMERE (0.70) and IFS (0.72), in the case of NO2 by IFS (0.65), and in the case of PM2.5 by SILAM (0.66) and LOTOS-EUROS (0.64). Table 5 summarizes the models that have achieved the best performance from the point of view of the mean bias, the RMSE, and the correlation coefficient.

The time evolution of predicted concentration values at Beijing by five different models involved in the intercomparison is provided in Fig. 9 for the period of 1–15 March 2017. An examination of the figure shows that, during most days, the daytime height of the PBL reaches 2500–3000 m with an exception on 2 to 5 March, when the height does not exceed 1000 m. Interestingly, during this period, the observed concentration of particulates, of NO2, and of SO2, strongly influenced by surface emissions, are significantly higher than during the following days. During the same days, the nighttime concentration of ozone is relatively low. On 10 March, one also observes high surface concentrations of emitted species and a low concentration of nighttime ozone, even though the calculated PBL height is not particularly low. One should mention here that, in two models (i.e., EMEP and LOTOS-EUROS), the information on the PBL is deduced from the IFS forecast, while in other models (such as WRF-Chem-MPIM and WRF-Chem-SMS), the PBL height is derived independently. In the case of WRF-Chem-MPI, however, the calculation of the PBL height makes use of meteorological data provided by the IFS model.

In most cases, the models capture the day-to-day variability in the species concentrations relatively well. The agreement with observations is generally good in the case of PM2.5 and PM10, except in the case of the IFS model, which considerably overestimates the concentrations, mainly because of a regional overestimation of the OM emissions and a lack of a diurnal variation in the emission. The anthropogenic OM emissions in IFS are parameterized based on anthropogenic CO emissions following Spracklen et al. (2011). The relatively high CO emission in this region may require a reduced conversion factor between OM and CO emissions. The main contribution to PM overestimation of IFS came from the nighttime values (see next section). Since nighttime overestimation also occurs for NO2, a lack of vertical mixing during the night in IFS could cause the nighttime overestimation of the surface values. As already noted, the models tend to underestimate the ozone concentrations, perhaps due to a slight overestimation of the nitrogen oxide concentrations. Another possible explanation is an underestimation of the VOC sources. Routine measurements of VOCs, however, are not available. The need for such measurements, however, needs to be stressed.

The model comparison reported here also shows differences between models in the case of NO, which should probably be attributed to differences in the emissions and emission injection heights of this species and in the formulation of vertical mixing in the boundary layer. Here again, measurements of NO in addition to those of NO2 and ozone would be useful. Finally, one notes in Fig. 9 the relatively good agreement between models (with the exception of the IFS and the WRF-Chem-SMS model) regarding the time evolution of odd oxygen (Ox=O3+NO2). The models, however, slightly underestimate the absolute values of the Ox concentration.

In order to evaluate the behavior of the different models regarding their ability to reproduce the diurnal variation in the surface concentrations of ozone, NO2, and PM2.5, we have calculated the mean diurnal variations over the period of 1–15 March 2017 averaged for the 34 cities included in our analysis (3 of the 37 cities, located in the western part of the country and adopted in the MarcoPolo–Panda prediction system have not been considered in this analysis). The resulting results are shown in Fig. 10 for ozone and NO2 (expressed in µg m-3). We have added the corresponding diurnal evolution of Ox (expressed in ppbv) defined as the sum of the ozone and NO2 mixing ratios. This last chemical variable has the advantage that it is not affected by the fast interchange (null cycle) between ozone and NO2 by the reactions NO+O3, NO2 + hv, and O + O2 + M. Since this cycle tends to transfer “odd oxygen” from ozone to NO2 after sunset and from NO2 to ozone after sunrise, the Ox variable is less variable than its two components NO2 and O3 over a diurnal cycle. Figure 10 shows that, when averaging over the 34 largest Chinese cities, the diurnal variation in the ensemble median is in good agreement with the observation in the case of NO2. In the case of ozone, the median values are somewhat underestimated in late morning and in the afternoon. A similar situation is found in the case of Ox. The RMSE for ozone and NO2, also shown on the figure, is generally lower in the case of the ensemble median than for the individual models. In the case of PM2.5, however, the RMSE of the two models CHIMERE and IFS are smaller than the RMSE of the ensemble median (not shown here). The mean bias of the ensemble median for NO2 and ozone is generally smaller than that of the individual models. In the case of Ox, some models exhibit a positive bias (WRF-Chem SMS), while others (e.g., SILAM) are characterized by a negative bias.

Figure 11a, b, and c show similar estimates of the diurnal variation in the three large cities of China: Beijing, Shanghai, and Guangzhou. These graphs show that the ozone forecast from the ensemble median is lower than observed values during the entire day both in Beijing and in Shanghai. In Guangzhou, however, ozone is slightly overestimated by the prediction. In the case of NO2, the surface concentrations are overestimated in Beijing and to a lesser extent in Shanghai, with the largest overprediction occurring during nighttime, when the planetary boundary layer is very thin and vertical mixing almost shut off. At the same time, ozone is negatively biased due to its efficient titration by NOx. In the three cities, the RMSE of NO2, ozone, and Ox appear to be largest at sunset. Thus, a general issue with the MarcoPolo–Panda prediction system is the overestimation of surface NO2 and the underestimation of ozone concentrations during the nighttime.

In the case of PM2.5, one of the models involved (IFS) strongly overestimates the concentrations during nighttime but is in fair agreement with observations during daytime. This issue may again reflect a problem with the formulation of species dispersion in the planetary boundary layer. It may also be due to the lack of specified diurnal variation in the emission of primary pollutants as well as to the increased nighttime stability.