Towards the coupling of a chemical transport model with a micro-1 scale Lagrangian modelling system for evaluation of urban NO x 2 levels in a European hotspot 3

. A multi-scale modelling system was developed to provide hourly NO x concentration fields at a building-resolving 11 scale in the urban area of Modena, a city in the middle of the Po Valley (Italy), one of the most polluted areas in Europe. The 12 WRF-Chem model was applied over three nested domains and employed with the aim of reproducing local background 13 concentrations, taking into account meteorological and chemical transformation at the regional scale with nested resolutions 14 of 15 km, 3 km and 1 km. Conversely, the PMSS modelling system was applied to simulate 3D air pollutant dispersion, due 15 to traffic emissions, with a very high-resolution (4 m) on a 6 km x 6 km domain covering the city of Modena. 16 The methodology employed to account for anthropogenic emissions relies on two different strategies. Traffic emissions were 17 based on a bottom-up approach using emission factors suggested by the European Environmental Agency with traffic fluxes 18 estimated by the PTV VISUM model in the urban area of Modena, combined with direct traffic flow measurements 19 performed between October 28 and November 8, 2016 which was used for the hourly vehicle modulation. Other 20 anthropogenic emissions were taken from the TNO-MACC III inventory at the scales resolved by the WRF-Chem model. 21 Simulations were performed for the same period whereby the traffic measurement campaign was carried out. 22


Introduction
Atmospheric pollution is one of the main risk factors for a number of pollution-related diseases and health conditions: they may occur through the appearance of harmful and carcinogenic effects on the respiratory system as well as the onset of other cardiovascular, nervous and ocular pathologies (Loomis et al., 2013, Novaes et al., 2010, PopeIII et al., 2003).These critical issues particularly affect urban areas with a higher population density: a complex mixture of pollutants is produced by inefficient combustion of fuels in internal combustion engines, power generation and other human activities like domestic heating and cooking.
One of the most critical air pollutants in terms of health effects is nitrogen dioxide (NO 2 ), whose levels in the last years exceeded national and WHO standards in many urban areas across Europe (European Environment Agency, 2016).In Italy, and more particularly in the Po Valley, despite an overall decrease in PM in the last 10 years (Bigi and Ghermandi, 2016), the urban population is still exposed to harmful levels of NO 2 .
The present research has as its goal the estimation of the air quality in the urban area of Modena, a city in the central Po Valley, in terms of NO x atmospheric concentrations.More in detail, the aim of the project is to support environmental policies, epidemiological studies and urban planning and management.
Current approaches to produce spatial maps of urban air pollution include the use of interpolation methods and land-use regression (LUR) models.However, all these techniques need a large number of in-situ observations at strategic locations to represent the full spatial and temporal pollutant variability and cannot be used to take into account turbulent atmospheric dispersion.To meet this need, a variety of micro (Moussafir et al, 2004(Moussafir et al, , Öttl, 2015) ) and local (Tinarelli et al, 1992, Bellasio and Bianconi, 2012, Cimorelli et al, 2004) scale air dispersion models have been developed in the last few years, as they can provide a high-resolution information on air pollution level within urban city area by taking into account space-time emissions distribution and local meteorological characteristics (Ghermandi et al, 2014, Ghermandi et al, 2015).
A key issue is the quantitative estimation of the different contribution to air pollution level from emissions sources located within the city urban environment and from countryside areas, also known as rural background.An approach that has been used for several years to account for both urban and rural contribution is generally called "Lenschow" approach (Lenschow et al., 2001), which envisages the influence of a city as the difference between the concentrations in the urban environment and the concentrations at rural site.This methodology is generally applied in source apportionment studies in order to estimate the primary and secondary component of PM by means of receptors models (Pirovano et al., 2015, Bove et al., 2014), or in urban impact assessments through an approach combining measurements and modelling results.In this latter case, the model is used to evaluate the contribution (effects) of the sources located within the city, while measured background concentrations are added to the simulated concentrations to account for remote sources outside the simulation domain (Ghermandi et al. 2019, Berchet et al., 2017).
Despite the flexibility in enforcing measured-based activity and modelling results, a number of criticisms related to the applicability of this methodology were recently highlighted.Following Thunis (2018) the city urban impact can be defined as the sum of three components: the Lenschow urban increment, which is the concentration difference between the city and background locations, the "city spread" meaning the impact of the city at the rural background location, and the "background deviation" that quantifies the concentration difference between city and background location when city emissions are set equal to zero.According to this definition, the "Lenschow" approach can be correctly employed only when the "city spread" and the "background deviation" are negligible or compensate each other, i.e. the urban impact in background area is close to zero and when background levels are spatially homogenous.
An alternative approach to quantify the different contributions to air pollution in the city is based on Chemical Transport Models (CTMs) which, unlike the Lenschow incremental approach, are able to generate different emission scenarios on multi spatial scale, from regional to local, which can be exploited to estimate background concentrations keeping city emissions set to zero.
Based on the advantages given by CTMs, the methodology employed in this study was a modelling activity relied on the NO x dispersion by combining two different models: the Weather Research and Forecasting (WRF) model coupled with Chemistry (Grell et al., 2005), which is able to compute concentrations fields over regional domain by considering specific emission scenarios, and the Parallel Micro SWIFT and SPRAY (PMSS, Moussafir et al., 2013, Oldrini et al., 2017) modelling suite accounting for dispersion phenomena within the urban area.
In this project, the PMSS modelling suite was used to simulate at building-scale resolution the NO x dispersion produced by urban traffic flows in the city of Modena.Conversely, the WRF-Chem model simulations were performed to estimate the NO x background concentrations on multiple domains with a nesting technique, in order to take into account emissions both at regional and local scale by excluding traffic emissions sources over the city of Modena.A similar approach was used also by Tewari et al. (2010), Wyszogrodzki et al. (2012) and Kwak et al. (2015), which developed an integrated urban air quality modelling system by coupling a CFD model with a chemical transport model in order to account not only the dispersion in buil-up areas but also to consider larger scale influences.
In the first part of the paper the modelling system which is applied to take into account both urban and background emissions sources is described.The following section describes the set-up of the models employed in the simulation, along with the method for estimation of emissions.Finally, the performance of the models in representing the meteorological mesoscale reconstruction and the pollutant dispersion is evaluated.

Modelling chain description
With the aim of providing hourly NO x concentration maps over the urban area of Modena, a hybrid modelling system composed by WRF-Chem, an Eulerian model and PMSS, a Lagrangian particle dispersion model, was employed in this study.The choice of this modelling chain was based on the WRF-Chem ability to simulate the emissions, transport and chemical transformations simultaneously with meteorology at large scale, and on the PMSS capability to provide high resolution air quality maps over an entire urban domain characterized by large spatial and temporal concentrations gradient, with a reasonable computation time (Ghermandi et al., 2015).The main options for physical and chemical schemes adopted here are reported in Table 1.These include the Noah Land Surface Model (Chen and Dudhia, 2001), the Yonsei University Planetary Boundary Layer scheme (Hong S.-Y., 2010), the Grell-Freitas cumulus parameterization (Grell and Freitas, 2014) activated only for the outer domain, the Lin microphysics scheme (Lin et al., 1983), and the Rapid Radiative Transfer Model (RRTM) radiation scheme (Mlawer et al., 1997) aimed to represent both shortwave and longwave radiation.Hodzic and Jimenez (2011).
Meteorological initial and boundary conditions were provided by the 6-hourly ECMWF analysis field (ERA5 dataset) with a horizontal resolution of 0.25° x 0.25°, interpolated to 37 pressure levels from 1000 to 1 hPa.Data included 3D fields of temperature, specific humidity and wind speed components.2D surface parameters such as mean sea level pressure, sea surface temperature, soil temperature and volumetric soil water content were also considered.A grid nudging on temperature and wind field has been also performed within the boundary layer in all three model configurations using as input data the ECMWF analysis.
As to land use, the Corine Land Cover (CLC) dataset was adopted after reclassifying it into the 33 USGS classes to match the WRF land use tables.Chemical initial and boundary conditions were provided by the global Model for OZone And Related chemical Tracers output (MOZART-4; Emmons et al., 2010).Biogenic emissions were calculated online by the Model of Emissions of Gases and Aerosols from Nature version 2.1 (MEGAN2.1)by Guenther et al. (2012).In addition, sea salt and dust emissions are calculated online.

PMSS modelling system description
Parallel Micro SWIFT SPRAY or PMSS (Oldrini et al. 2011(Oldrini et al. , 2017) ) is the parallelized version of the MSS modelling suite (Tinarelli et al. 2007(Tinarelli et al. , 2013) ) constituted by the individual models SWIFT and SPRAY, both used in small scale urban model (a.k.a.Micro-SWIFT and Micro-SPRAY).
Micro-SWIFT is a 3D mass-consistent diagnostic model that uses terrain-following coordinates to provide diagnostic wind, turbulence, temperature and humidity fields using data from a dispersed meteorological network.
The first step performed by Micro-SWIFT is the interpolation of the heterogeneous meteorological input data, such as surface and vertical measurements profile or regional meteorological model output, to reconstruct the 3D wind field.Then, the first computed wind field is modified in the zones around isolated or group of buildings following the approach suggested by Röckle (1990) and Kaplan and Dinar (1996) adopting a parametrization of the recirculating flow regions around, behind, over and between obstacles.Subsequently, the mass conservation constraint is imposed through the impermeability conditions on the ground and at building surfaces.Finally, a RANS flow solver can be optionally used to simulate more accurate velocity and pressure fields in built-up environments than obtained with the pure diagnostic flow model configuration.With the RANS approach, the momentum equation is introduced in the computation and the turbulent Reynolds stress tensor is modelled by a zero-order closure based on mixing-length theory and the momentum and pressure equations are solved using the fractional time step technique (Gowardhan et al., 2011).
The estimation of the turbulence needed by Micro-SPRAY to drive pollutants dispersion is diagnosed by Micro-SWIFT through the superimposition of the background turbulence, obtained by standard boundary layer parameterizations (Hanna et al., 1982) and the turbulence inside the flow zones modified by the obstacles.
Micro-SPRAY is a 3D Lagrangian Particle Dispersion Model (LPDM) (Rodean, 1996) able to simulate the advection and the diffusion of gaseous species or fine aerosol by accounting for the presence of obstacles.The dispersion of an airborne contaminant is modelled by virtual particles that follow the turbulent motion of the air as passive tracers and their spatial distribution at a certain time represents the concentration of an emitted substance.
The trajectories of the particles emitted by a source are obtained by integrating in time their velocity.This can be considered as the sum of a transport component, defined by the local averaged wind, usually provided by Micro-SWIFT, and a stochastic component, standing for the dispersion due to the atmospheric turbulence.The stochastic component is obtained by solving a 3D form of the Langevin equation for the random velocity, following Thomson's approach (Thomson, 1987).
The PMSS modelling system has been validated (Trini Castelli et al. 2017and 2018, Oldrini et al., 2017and 2019) and applied (Carlino et al., 2016, Moussafir et al., 2013) to several experiments and real cases.In this study, the performances of the PMSS modelling system are exploited to estimate urban air quality in the city of Modena in a real case scenario.

PMSS set-up
A 3D wind and turbulence field and air pollution dispersion reconstruction was performed on a 6 km x 6 km square domain covering the city of Modena (Figure 2) with the PMSS modelling suite.Given the low altitude difference between different areas of the city, a flat domain was considered and a 3D reconstruction of buildings was made by using a pre-processor: 25,600 polygons contained in the ESRI shapefile (provided by Geoportale Regione Emilia-Romagna) were transformed into approximately 146,000 triangular prisms directly usable by Micro-SWIFT.
In order to guarantee both flow and pollutant dispersion fields at a high resolution in each part of the domain, a horizontal grid step of 4 m (square cells) was chosen for both Micro-SWIFT and Micro-SPRAY models.3D fields of wind, temperature and turbulence were obtained for 20 vertical levels from 3 m up to 200 m above the ground using the Micro-SWIFT model with the RANS flow solver option activated.The 2D Cressman interpolation method wind field was also considered in the configuration.
Regarding the Micro-SPRAY model simulations the horizontal grid was chosen to be identical to that of the Micro-SWIFT model computation and the vertical grid structure consisted of 10 levels with a linear progression up to 200 m above the ground level with 3 m height for the first layer close to the soil.This arrangement leads to a configuration of 1504 x 1504 x 10 nodes and a total number of 2.26•10 7 cells.The main SPRAY parameters also included an emissions time step of 5 seconds and a synchronisation time step of 10 seconds.Concentrations are computed every hour and sampled every 10 seconds.Since the cumulative precipitation for all the examined period was less than 10 mm, the wet deposition was not included in the set-up.The main PMSS set-up parameters are reported in Table 2.

Anthropogenic Emissions
Following the regional emissions inventory database produced by Arpae Emilia-Romagna, the local environmental agency (INEMAR 2013), the road traffic in Modena contributes up to the 60% of the total emissions in terms of NO x , while the domestic heating and industrial combustion represent only the 15% and 14% of the total amount.Based on this percentage distribution, the methodology employed to account for anthropogenic emissions rely on two different strategies: a citytailored emission estimate, to describe traffic emissions at micro-scale resolution in the urban area of Modena, and an emission inventory estimate, more suitable to account for emissions at large-scale area, used as an input for the chemical transport model in order to estimate the contribution of all the SNAP (Selected Nomenclature for Air Pollution) emission categories throughout Europe.
The anthropogenic emissions used for the parent and the nested WRF-Chem domains were taken from the TNO-MACC III inventory, available on a regular grid with a horizontal resolution of 0.125° x 0.0625°, which contains emissions for air pollutants such as NO x , SO 2 , NMVOC, NH 3 , CO and primary particulate matter (PM 2.5 and PM 10 ).The inventory is based on nationally reported emissions for specific sectors and spatially distributed with proxy data such as the population density for urban emissions or the road network for non-urban emissions.The main developments with respect the version II of the inventory (Kuenen et al. 2014) includes improved emissions and trends for the international sea shipping, improved wood consumption estimation and more detailed spatial distribution, as well as improved industrial emissions apportioning achieved through the use of CORINE land cover data instead of population density data as a default item.
TNO-MACC III emissions are provided as annual totals, therefore each SNAP category was scaled to take into account monthly, daily variation (weekend or weekday) and hour of the day (diurnal cycle), as suggested by Kuik et al. (2018) Ntziachristos and Samaras (2016) were introduced in the package and a series of functions to automatically compute for each road of the network the total NO x emissions were implemented.
The Tier 3 methodology, defined in the European guidelines EMEP/EEA (Ntziachristos and Samaras, 2016) for the estimate of exhaust emissions from road transport was adopted, and total exhaust emissions were calculated as the sum of hot emissions (when the engine is at its normal operating temperature) and emissions during transient thermal engine operation (termed 'cold-start' emissions).As stated before, the road network considered in the micro-scale simulation is contained in the urban area of Modena, therefore only urban driving situation was considered in the emissions estimation.
EF values were computed for each vehicle fleet category considering the flow speed estimated by PTV VISUM, and then EFs were mathematically weighted according to fleet composition to obtain for each road two EFs, one representing the light vehicles and one on behalf of duty vehicles.
Hot emissions computation included also correction values applied to the baseline emission factors to account for different vehicle age for passenger cars and light commercial vehicles, also called emissions factors degradation due to vehicle age (Ntziachristos and Samaras, 2016).Actual average annual mileage (AAAM) for gasoline and diesel cars in the Italian vehicle stock, respectively AAAM gasol and AAAM diesel , were estimated with the formulas suggested by Caserini et al. (2013).
Considering age as the number of years between the current year and the year of vehicle purchase, the formulas can be written as follow: = 20.817The hot EFs obtained following the EMEP/EEA guidelines for the city of Modena are in general in line with the hot EFs computed using the HBEFA methodology (v4.1) for similar European countries.The main differences in absolute terms regard the Heavy Duty Truck category, where the HDT EFs estimated in Modena are 4.28 and 5.34 g km -1 for petrol and diesel respectively, while the average hot EFs for the same vehicles category are 6.75 and 3.97 g km -1 .The variability between diesel HDT EF for the six European countries considered is also large and the maximum of these EFs is greater than the EF considered in Modena.On the other hand, diesel LDV EFs are very similar to each other and equal to 1.47 g km -1 for Modena and 1.58 g km -1 for the average EF between the six countries.By contrast, petrol LDV EFs differ of about 0.37 g km -1 (0.83 g km -1 for Modena and 0.46 g km -1 for the average of the other countries).Finally, Motorcycle and Passenger Cars EFs estimated in Modena agree very well with the average EFs estimated following the HBEFA methodology: 0.28 (petrol), 1.02 (diesel) and 0.05 (CNG) g km -1 are the PC EFs estimated in Modena, while the HBEFA respective average is 0.16, 0.94 and 0.18 g km -1 .Motorcycle EFs are almost the same, 0.15 g km -1 for Modena and 0.16 g km -1 for the HBEFA average.

Model Simulations
Following the setup described in the previous section, simulations were performed for the period between 28 October 2016 and 8 November 2016, the same period whereby the traffic measurement campaign was carried out.These days were characterized by weather condition typical for autumn in the central Po Valley (Bigi et al., 2012, Thunis et al., 2009) with a very little atmospheric circulation due to recurrent thermal inversions at low altitude, low mixing layer heights and persistent foggy and hazy events which lasted also during day time.Recurrent wind calm episodes and high-pressure conditions enhance persistence and homogenization of air masses on a regional scale: the characteristic climate conditions, along with the strong anthropic pressure in the area, lead to long-lasting high concentrations of pollutants also at remote rural sites (Bigi et al., 2017, Masiol et al., 2015, Tositti et al., 2014).
Low rainfall rate (cumulative precipitation lower than 10mm), mean wind speed lower than 2 m s -1 and daily average temperature from 7.4 to 13.6 °C characterized the meteorological condition in Modena during the investigated period.
WRF-Chem was performed using a 2 days spin-up period to ensure consistency with meteorological and chemical fields, as well as PMSS was executed using 6 hours of spin-up at the beginning of the simulation.

Model Evaluation
The statistical performance analysis considered multiple statistical indicators regardless of the model's application since each one has its advantages and disadvantages and it is not possible to identify a unique exhaustive index of quality.
The   Generally, modelled hourly 2 m temperatures reproduced by WRF-Chem at 1 km resolution (d03 domain) are consistent with observations at all the stations: r is between 0.68 and 0.87, except for two stations, Persico Dosimo and Bologna urbana, where the correlation is respectively very good (0.96) and not particularly high (0.59).The minimum value can be attributed to the difficulties of the model in representing urban meteorological dynamics where land use data for a large urban area may not be extremely accurate and some local phenomena can be missed.This could be the case of the Bologna city environment.By contrast, the maximum correlation observed at the Persico Dosimo station could in this case reached by a chance where hourly measurements were available only for one fourth of the investigated period.However, the good correlation between modelled 2 m temperature and observations for a large number of stations shows that the WRF-Chem model represented the observed meteorological variability quite well.
The model on average tends to be positively biased with a MB smaller than +1°C for most of the stations where only two of them exceed +2°C of MB, respectively at the Reggio Emilia urbana and Colorno stations.On the other hand, the minimum MB is -1.6°C, achieved at the Porto Venere station on to the Ligurian Sea shore.These results are in the same range as the MB that Gsella et al. ( 2014) found using MM5, WRF and TRAMPER meteorology models for the same area.Simulated hourly wind speed at the d03 domain generally express performance in line with similar case study in literature (Kuik et al., 2016, Mar et al., 2016, Gsella et al., 2014) since the great majority of the stations exhibits a MB between -0.5 m s -1 and +0.5 m s -1 , range suggested by Malm et al. (2009) and by European Environmental Agency (EEA) guidelines (EEA, 2011).Only the stations of Loiano and La Spezia are outside these limits with respectively a MB of -0.99 m s -1 and +1.43 m s -1 : the former is located in the Tuscan-Emilian Apennines at about 700 m above the sea level and the latter is close the Ligurian sea and then characterized by a strong influence of land-sea breeze, similarly to the temperature monitoring station of Porto Venere.The large bias found at these two stations suggests that the model might have difficulties in simulating the wind field in mountainous areas and close to the sea where complex orography and local breeze characterize the territory, however for the rest of the stations the MB values are consistent with the reference benchmarks proposed in literature.
Another statistical indicator suggested by the EEA guidelines (2011) and by Malm et al. (2009) is the Root Mean Square Error, for which the recommended benchmark for wind speed is less than 2 m s -1 .As for the MB, RMSE of modelled wind speed values are below 2 m s -1 at all stations besides Loiano and La Spezia, where RMSE are 3.76 m s -1 and 2.14 m s -1 , respectively.Nonetheless, modelled wind speed at the vast majority of the stations is in line with the benchmark for a mesoscale meteorological reconstruction.
In Figure 8 the performance of WRF-Chem in reproducing wind speed for the three different resolutions is shown: RMSE is plotted as a function of the MB.As the resolution is increased, the model tends to show lower MB in absolute terms.
Similarly, the RMSE generally tends to moderately decrease as well with increasing resolution.It is therefore possible to conclude that in increasing model resolution from 15 km to 1 km there is a slight improvement of performance also for wind speed reconstruction.Although the model performances in reproducing the wind directions are not outstanding, they are in line with other case studies within the Po Valley, where MAE was between 42° and 93° (Gsella et al., 2014, de Meij et al., 2009) and RMSE was between 127° and 148° (Gsella et al., 2014), confirming the difficulties in modelling the wind fields in this area characterized during winter and fall time by stagnant conditions and low wind speed.

WRF-Chem nitrogen oxides
Since the role of WRF-Chem in this study was to estimate NO x (NO + NO 2 ) concentrations due to emissions on the regional scale that may affect the background air quality in Modena, modelled hourly NO x concentrations were compared with observations at 10 rural background sites (8 of them from the Arpae Emilia-Romagna network and 2 of them from Arpa Lombardia network, see Table 3 and Figure 6) within the d03 WRF-Chem domain.
Modelled NO x concentration at 1 km resolution is biased negatively for 8 stations with a minimum MB equal to -18.1 µg m -3 (-62% of NMB) and biased positively for 2 stations with a maximum MB of +7.8 µg m -3 (+30% of NMB).In addition, for each reference station, the fraction of predicted values within a factor of two of observations was computed, also referred as FAC2.The corresponding average value over all stations is 56% (minimum 30% at the station of Ostellato and maximum 76% at the stations of Schivenoglia), in accordance with the reference value suggested by Chang and Hanna (2004), greater or equal to 50%.
In order to test which level of model spatial resolution gives better results, observed NO x concentrations were compared with modelled concentrations at 15 km (d01) and 3 km (d02) resolutions in terms of FAC2 and NMB (Figure 10).The model at 15 km resolution presents on average the highest FAC2 with respect other configuration, 62% (maximum equal to 82% at the stations of Schivenoglia and minimum equal to 39% at the station of Ostellato), conversely the NMB presents its higher variability, from -58.9% to 122.2% (respectively -17.5 µg m -3 and +5.7 µg m -3 of MB).
Among the other configurations, WRF-Chem at 3 km resolution shows better performance with an averaged FAC2 between all the stations equal to 61% (79% of maximum at Langhirano station and 37% of minimum at the station of Ostellato) and the smallest variability between all the stations in terms of NMB, from -56.6% to 28.7% (respectively -16.8 µg m -3 and +1.3 µg m -3 of MB).The comparison showed that generally the WRF-Chem model with a horizontal resolution of 3 km is better suited to reproduce NO x concentrations at rural background sites considering an emission inventory of 14 km x 7 km resolution.
Increasing resolution to 1 km, the model tends to decrease the number of predicted concentrations within a factor of two of observations, despite the variability of the NMB being approximately the same as the configuration at 3 km.It is also worth remarking that nonetheless the modelled NO x concentrations are on average in agreement with the values suggested by Chang and Hanna (2004) for model performance evaluation in terms of NMB and FAC2, not all considered stations satisfy these statistical indicators, this might mean that a more detailed emissions estimation such as an improved emissions distribution in the area should be implemented to achieve better results.

Micro-scale wind fields
High resolution winds in the urban area of Modena were estimated by the means of Micro-SWIFT (the 3D mass-consistent diagnostic model that composes PMSS), initialised with the meteorological fields reproduced by WRF-Chem at hourly time step.Simulated winds reproduced by Micro-SWIFT were compared with observation at three different meteorological sites located within the urban center of the city.These stations are respectively placed on top to the Geophysical Observatory tower, at 42 m height above the ground and placed in the historical part of the city (referred as OSS), above the public hospital to the Est of the historical city center at about 20 m height (referred as POL), and on top to the municipality building at 40 m height, to the West of the historical city center (MOD station), the latter used also to validate the WRF-Chem meteorology.Figure 4 depicts the position of these three stations.
The time series reported in Figure 11 show the comparison between modelled and measured hourly wind speed for the three urban meteorological sites.Notwithstanding a few remarkable overestimations on November 3, 5 and 7, mostly visible at the OSS and MOD stations, modelled data reproduced observed trend quite well for all the three locations.MOD and OSS modelled time series show also a very similar behaviour (with a general less pronounced overestimation for MOD on November 3, 5 and 7) due to the location of the sensor, both above 40 m and thus characterized by the same meteorological input and not affected by the presence of buildings.The satisfactory performances showed by the time series regarding wind speed are also confirmed by the statistical metrics.
The MB is less than 0.2 m s -1 for all the three stations and the RMSE is between 0.93 and 1.24 m s -1 .FB and FAC2 are also in line with the values found during the validation of the PMSS modelling suite in urban environment (Oldrini et al., 2019).
Table 4 summarizes computed metrics.In Figure 12 hourly simulated wind directions are compared with hourly observed wind for the same three locations.As for regional wind field evaluation, wind directions were poorly reproduced by Micro-SWIFT, respectively with MAE equal to 125° for the MOD station, 124° for OSS and 114° for POL.By contrast the performance of Micro-SWIFT evaluated in terms of FB and NMSE are similar to the results obtained in urban environment by Oldrini et al. (2019) using the same model: for MOD site the FB and NMSE are respectively equal to 0.21 and 0.76, FB is 0.23 for both OSS and POL locations and NMSE is 0.80 and 0.70 at OSS and POL stations.
This poor behaviour in wind direction reconstruction can be partly attributed to the input data used by Micro-SWIFT, which reflects the bias given by the regional forecast model in wind direction estimation.This is mainly due to the difficulty of the meteorological models in reproducing wind fields during the situations with very little atmospheric circulation (stagnant condition) and low wind speed, as occur red in this case study (Gsella et al., 2014, de Meij et al., 2009).

Micro-scale NO x concentrations
NO x background concentrations estimated with WRF-Chem in the urban area of Modena were added to the NO x simulated concentrations reproduced with PMSS modelling system performed considering exclusively road traffic emissions.Modelled concentrations were compared with observations at two urban stations: the first one at a traffic site, located in the proximity of a busy street close to the urban ring road, named "via Giardini", and the second one at background site, within a public park to the West of the historical city centre, named "parco Ferrari" (Figure 4).
In Figure 13, the hourly NO x concentrations predicted by PMSS in combination with WRF-Chem at 3 km and 1 km resolutions (labelled as "d02+ PMSS" and as "d03+ PMSS") are compared through scatter plots for both urban traffic and background stations.In this figure, the solid line represents perfect agreement with observations and within the dashed lines modelling results and observations agree with a factor of two.
Most of the modelled data are within a factor of two of observations, especially for the urban traffic site in both WRF-Chem configurations, whereas for the urban background station an under estimation is more noticeable.It is also worth noting that the results of PMSS combined with WRF-Chem at 3 km and at 1 km resolution depict similar behaviour and relative scatter plots are very comparable to each other, with a slightly less pronounced underestimation for WRF-Chem at 3 km resolution.
Modelled hourly NO x concentrations are biased negatively in both urban stations: at the "via Giardini" traffic site the MB of simulated NO x by PMSS and WRF-Chem at 1 km resolution is -15 µg m -3 , which corresponds to -15% of NMB, for the same model configuration the MB at the "parco Ferrari" background site is -30 µg m -3 (-41% NMB).Modelled hourly concentrations correlate reasonably well with observations in both sites, with r equal to 0.48 at the traffic station and 0.43 at the background station.
The performance of the models generally increases when hourly NO x concentrations reproduced by PMSS are combined with the results of WRF-Chem at 3 km resolution, the MB at the "via Giardini" traffic station is -4 µg m -3 (-4% NMB) and -18 µg m -3 (-25% NMB) at the "parco Ferrari" measurement station.
Despite an improvement in term of MB with WRF-Chem at 3 km, its combination with PMSS doesn't particularly affect the r between modelled and observed concentrations (0.47 at "via Giardini" traffic station and 0.44 at "parco Ferrari" background site).
A quantitative estimation of the agreement between simulated and observed concentrations was also assessed following the statistical metrics proposed by Hanna and Chang (2012) for urban dispersion model evaluation.FB, NMSE, FAC2 and NAD were computed for both the urban stations located in Modena and for both the combination of WRF-Chem (3 km and 1 km resolution) with PMSS.Table 5 summarizes all the computed statistics.
Following Hanna and Chang (2012) the reference acceptance criteria for the aforementioned metrics in urban dispersion model evaluation are as follows:  |FB| < 0.67, i.e. the relative mean bias is less than a factor of ~2  NMSE < 6, i.e. the random scatter is less than 2.4 times the mean  FAC2 > 0.30, i.e. the fraction of predicted concentrations within a factor of two of observed concentrations exceeds 0.30  NAD < 0.50, i.e. the fractional area for errors is less than 0.5 The statistical analysis shows that PMSS combined with WRF-Chem at both d02 and d03 domains fulfill the acceptance criteria defined by Hanna and Chang (2012).Regarding the FB, the results are always less than the threshold of 0.67, in particular at urban traffic site the outcomes of this metric are particularly good with values equal to 0.04 (PMSS + d02) and 0.16 (PMSS + d03).At urban background station the results are larger than the previous one (0.29 for PMSS + d02 and 0.52 for PMSS + d03) indicating that the models tend to underestimate more the mean concentrations but nevertheless in well agreement with the reference benchmark.As far as the NMSE is concerned, the models show their best performances with scores largely lower than the acceptance benchmark ( 6), with a maximum value of 1.15 at urban background station for PMSS + d03 (minimum value at traffic station for PMSS + d03 equal to 0.48), meaning that predicted values very rarely differ strongly from observations.
Regarding the FAC2 and the NAD there is a significant agreement between model results and relative acceptance criteria at both urban stations and for both model configurations.Minimum and maximum FAC2 are equal to 0.59 and 0.72 achieved respectively at the urban background station for PMSS + d03 and at urban traffic station for PMSS + d02 (the lower limit proposed by Hanna and Chang is 0.30).For the same locations and model configuration the maximum and the minimum NAD are 0.35 and 0.24 respectively (upper limit proposed by Hanna and Chang is 0.50).
The statistical analysis, despite supporting a moderately better behaviour when the resolution of WRF-Chem is 3 km, also shows that the metrics of modelled NO x concentrations are comparable between the two WRF-Chem resolutions (1 km and 3 km), without a clear difference of one of the two.
Time series analysis (Figure 14) shows that modelled NO x concentrations agree quite well with observations.In particular for the first five days of simulations (considering the combination with WRF-Chem at 3 km resolution as a reference), between October 28 and November 1st observations are reproduced well and daily peaks are modelled with quite good accuracy, especially for urban traffic station where the MB between October 28 and November 1 is +7µg m -3 and NMB is about 1% (at the background station the MB is -14 µg m -3 and NMB is -1%).
By contrast, between November 1 and 2 observed NO x concentrations tend to be overestimated since the WRF-Chem contribution during the central hour of days exceeds observed NO x concentrations.This situation is particularly evident at the "parco Ferrari" station.Furthermore, between November 2 and 3, PMSS failed to capture the diurnal cycle in observed concentrations: on November 2 in the central hours of the day observations are largely overestimated with an increasing trend up to the afternoon peak.A possible explanation to this episode can be addressed to an underestimation of the PBL height during the central hours of the day, where its modelled value doesn't exceed 190 m (average daily maximum during the whole simulation is 600 m) and during only 6 hours the modelled PBL height is greater than 100 m.
On the other hand, on November 3, when observed concentrations reach values over 200 µg m -3 , the models underestimate the observed values.Between November 4 and November 6 observed NO x concentrations were lower than in the previous days (hourly maximum always lower than 200 µg m -3 ) mostly due to the rainfall which occurred on November 5 (8 mm) and 6 (1 mm), where the combination of the two models was able to reproduce the hourly pattern well, with a slight overestimation at the traffic station.
On November 7 and 8 extremely high peaks occurred with values that exceeded 400 µg m -3 at urban traffic station and 250 µg m -3 at urban background station.Despite an underestimation of absolute observed NO x concentrations experienced on these two days, the shape of the diurnal cycle was captured quite well by the models.Finally, in order to study which part of day primarily affected the general underestimation of the models and to investigate the WRF-Chem contribution to the total NO x concentrations during the day, the variation of observed and predicted NO x concentrations by hour of the day was assessed.The comparison of mean observed NO x daily cycle shows very similar behaviour between the "via Giardini" and "parco Ferrari" stations (Figure 16): two main peaks occur on average during the day, one between 08:00 and 09:00 a.m.(equal to 150 µg m -3 at traffic site and 105 µg m -3 at background site) and the second, on average greater of about 20-30 µg m -3 than the former, around 07:00 p.m. (about 170 µg m -3 at traffic site and 140 µg m -3 at background site).At "parco Ferrari" station another peak occurs in the early morning (around 01:00 a.m.) less pronounced than the other two, with a mean concentration of about 70 µg m -3 .
For the purpose of further investigate the modelled concentrations, in Figure 15 is shown the diurnal cycle of traffic emissions used for WRF-Chem and the average diurnal cycle used for PMSS in working and non-working days.There is a large correspondence between the traffic temporal cycle of WRF-Chem and PMSS in working days.For the former, the of about two hours.A more detailed description of the PBL evolution during the day, through for example ceilometer observation, could help to improve the concentration estimation.
At urban traffic site the mechanism seems to be the same but large traffic emissions affected less the underestimation during the morning peak.
The NO x underestimation could also be associated to the aged traffic fluxes estimated by the PTV VISUM model, which simulation reference year is 2010 and then affected by economic recession whereby all western countries have fallen since 2007.The economic crisis caused a slower growth of gross national product and an increase in unemployment which, among the other consequences, led to a reduction in traffic fluxes in the European business centres.For these reasons traffic fluxes estimated by PTV VISUM in the city of Modena may be affected by an underestimation with respect the real traffic occurred during the simulated period.
It is also worth noting that the PTV VISUM simulation doesn't include traffic fluxes of urban public transport, therefore the vast majority of urban buses emissions are not taken into account in the simulation.Due to the presence of schools nearby the "via Giardini" monitoring station, the morning rush hours (especially between 08:00 and 09:00 a.m.) are characterized by an intense flow of public buses and the consequent omission of public transport emissions can influence and contribute to the morning NO x underestimation.
In addition, traffic fluxes in secondary streets seem to be modelled worse than busier roads and at the urban background station the NO x underestimation can also be attributed to a rough estimation of NO x sources around the area where traffic emissions are modulated according the measurement data collected at "via Giardini" street.Moreover, other NO x emissions sources are simulated by the WRF-Chem model through the use of the TNO-MACC inventory which level of detail is about 7 km and thus not as exhaustive as the traffic emissions, and these other NO x sources are expected to have larger influence at urban background than at urban traffic sites.Besides this, looking at the NO x concentrations daily cycle, the contribution of WRF-Chem to the total simulated concentrations seems to be modest compared to direct effect of traffic emissions.In particular, when the resolution of WRF-Chem is 1 km, its share doesn't present a strong hourly trend but it is rather flat, with an average contribution during each hour of simulation in the order of 20 µg m -3 , at both traffic and background sites.On the other hand, when the resolution of WRF-Chem is 3 km, its rate to the total concentrations is greater than the one a 1 km resolution (average hourly contribution about 30 µg m -3 ), with an hourly trend more marked during the two daily peaks.This difference in contribution estimate could be explained considering the position of the WRF-Chem computational cells.In the case of 1 km resolution the computational cells over "parco Ferrari" and "via Giardini" stations, even if they present very similar concentrations, are different and located both inside the PMSS domain.Conversely, for d02, both the stations are within the same computational cell and the cell itself is no longer entirely contained in the PMSS domain but it extends to the West, beyond the PMSS borders.This latter configuration leads to account within the PMSS domain part of the TNO traffic emissions occurring outside the Modena urban area, originally excluded from the PMSS computation, causing a homogenization and an increasing in NO x concentrations over the entire WRF-Chem cell.The final outcome is that the contribution of WRF-Chem using d02 to the urban concentrations is larger and also the daily trend during the two daily peaks appears more pronounced.
More accurate traffic modulation across the city and a finer spatial resolution of non-traffic related exhaust emissions, as well as a more accurate description of the planetary boundary layer height during the day, would better fit the 1km grid of WRF-Chem used in d03 and can certainly contribute to achieve better results also at urban background station.Green lines show the mean daily cycle of planetary boundary layer height modelled by WRF-Chem and used in the micro-scale dispersion.
Solid lines represent the daily mean cycle, meanwhile shaded area show the variability between 25 th and 75 th percentiles.

Summary and conclusions
In this study the authors evaluated a hybrid modelling system consisting of the chemical transport model WRF-Chem and the Parallel Micro SWIFT SPRAY (PMSS) modelling system for the urban area of Modena with the aim of providing NO x concentration maps at building-resolving scale and at hourly temporal resolution, suitable to resolve the variability of emissions and atmospheric state.The WRF-Chem model was applied over three nested domains with an increasing resolution from 15 km to 1 km passing by 3 km, in order to simulate the emissions, transport and chemical transformations at the regional scale by accounting also at the same time for meteorological phenomena at the synoptic scale.Driven by these meteorological fields, the PMSS modelling suite was run over a domain of 6 km x 6 km with 4 m grid step size, to reconstruct micro-scale wind streams inside the urban area of Modena and then to simulate the dispersion of NO x , coming from traffic emissions, by accounting for the presence of buildings.Simulated and observed NO x hourly concentrations in the urban area of Modena exhibit a large agreement, in particular for urban traffic site ("via Giardini" measurement station), where detailed traffic emissions estimation (real traffic modulation combined with a bottom-up approach) proved to be very successful in reproducing the observed NO x pattern, confirming that reasonable time modulation for traffic emissions are among the main parameters to trim for urban atmospheric dispersion.
Despite the morning rush-hour peak (between 08:00 a.m. and 09:00 a.m.) tending to be generally underestimated, the magnitude of the afternoon peak around 07:00 p.m. was well captured by PMSS with an anticipation of about two hours.At the urban background station, notwithstanding a general underestimation of the observed concentrations (more pronounced than at the urban traffic site), the analysis of hourly daily modelled concentrations shows that PMSS combined with WRF-Chem provided a daily pattern in line with observations.These features highlight the strength of this modelling chain in representing urban air quality, in particular at traffic sites, whose concentration levels make them the most critical area of the city; characteristics that chemical transport models alone cannot express, due to the coarser resolution to which they operate and to their inability to reproduce street canyons and urban structures.

Figure 1
Figure 1 illustrates the interplay between WRF-Chem and PMSS.The WRF-Chem model was used to estimate NO x concentrations on multiple domains at different grid resolutions spanning from the European domain to the Po Valley area with a nesting technique, necessary to take into account emissions at regional scale that can affect urban air quality in Modena.Then, wind streams within the city were determined by a cascade of scales from global to buildings level: synoptic and local scale meteorological conditions in the region surrounding the city of Modena were simulated by WRF-Chem taking into account the local topography and land-use data.Driven by these mesoscale flow patterns, high resolution winds were computed in the city to account for buildings and street canyons by performing the diagnostic mass-consistent Parallel-Micro-SWIFT model.Secondly, Lagrangian dispersion simulations driven by high resolution winds were carried out with Parallel-Micro-SPRAY by estimating urban traffic emission flows.As a final step of the procedure, NO x traffic urban concentrations, simulated with Parallel-Micro-SPRAY, were added to NO x background concentrations estimated with the WRF-Chem model.Further explanations regarding the methodology used to avoid double counting of traffic emissions within Modena urban domain are reported in section 2.5.

Figure 1 :
Figure 1: Outline of the multi-model approach implemented to generate hourly NO x concentrations fields.

Figure 2 :
Figure 2: Overview of the WRF-Chem model domains on the left (Geographic coordinate system-WGS84) and PMSS investigation domain with the considered Modena street network represented as red lines on the right (UTM32-WGS84).
. A vertical emissions distribution was also taken into account by distributing the emissions of industrial sources, airports, extraction and distribution of fossil fuel into seven vertical layers, up to 750 m.In order to avoid the double counting of the traffic emissions placed inside the urban area of Modena and to better represent the spatial distribution of traffic sources in the nearby territory, a downscaling procedure was conducted for the SNAP sectors 71-75.The original TNO dataset (resolution ca.14 km x 7 km) covering the inner most WRF-Chem domain was subdivided to a finer grid with a horizontal resolution of 1 km x 1 km.Traffic fluxes of light (passenger cars and L-category vehicles) and duty (light commercial vehicle and heavy-duty trucks) vehicles on the main roads of the province of Modena at morning rush hour (07:30 -08:30, local time) for the year 2010 were provided by the Municipality of Modena and proceed from a simulation study by means of the PTV VISUM model (PTV Group, Karlsruhe, Germany http://vision-traffic.ptvgroup.com/en-us/products/ptv-visum/).These data were used as a proxy variable to assign TNO-MAC III traffic emissions over the province of Modena to the portion of the land interested by PTV VISUM road network: the more traffic fluxes were estimated for a specific road segment, the more emissions were assigned to the corresponding grid (1 km x 1 km) (Figure3).Once the downscaled grid dataset was created, a spatial surrogate function was implemented to identify the TNO-MACC III traffic emissions within the PMSS domain.This function returns zero if the territory cell is completely inside the PMSS domain, one if the territory cell is completely outside and a value between zero and one (proportional to the area outside the PMSS domain) if the territory cell crosses the domain boundaries.Finally, to exclude TNO-MACC III emissions from the PMSS domain, the spatial mask created with the surrogate function was multiplied by the downscaled traffic emission inventory.The same traffic simulation obtained with the PTV VISUM model includes also vehicle fluxes in the urban area of Modena, providing the number and the average speed for light and duty vehicles at rush hour for each segment of the urban road network, which encompasses about 1100 sections with a total length of 210 km (Figure4).

Figure 3 :
Figure 3: Total NO x traffic emissions (SNAP sector 71-75) in the province of Modena.On the left the original TNO-MACC III inventory emissions (resolution ca.14 km x 7 km).At the centre the TNO-MACC III inventory emissions downscaled to a resolution of 1 km x 1 km and distributed according to the street traffic flow estimated by PTV VISUM.On the right the TNO-MACC III inventory emissions downscaled to a resolution of 1 km x 1 km without the emissions by road traffic in the urban area of Modena.

Figure 4 :
Figure 4: PMSS computational domain (left side) and view of the radar traffic counters position (right side).Radars are represented by yellow dots, air quality urban background station by orange triangle and urban traffic station by blue triangle.Urban meteorological stations are also depicted by green dots along with their label names.
the traffic model data, a direct vehicle flow measurement campaign was carried out continuously over two weeks between October 28 and November 8, 2016, with 4 Doppler radar counters (one for each road lane) in a four-lane road in the proximity of the intersection with the urban ring road(Ghermandi et al. 2019) (Figure4).The radar traffic counters recorded the time, the length and the velocity for each passing vehicle.The captured vehicles were subdivided into two different groups according to the vehicle classes modelled by PTV VISUM: light vehicles, with measured length less or equal to 6 m and duty vehicles, with measured length greater than 6 m.Finally, to appropriately describe NO x emissions under typical vehicle flux conditions, recorded flow data were used to reproduce hourly modulation rates for the entire road network.In order to test the reliability of the hot EFs computed with the methodology previously described and to check how far these EFs are with respect the one computed considering recently emission data, based on new PEMS tests(Hausberger et al, 2019, Sjödin et al., 2018), the weighted EFs for Passenger Cars (PC), Light Commercial Vehicle (LCV), Heavy Duty Trucks (HDT) and Motorcycle in Modena were compared with the weighted average hot EFs calculated following the handbook of emission factors (HBEFA, 2019) for different European countries (Germany, Austria, Switzerland, France, Norway and Sweden).Since the HBEFA desktop application is not available for free, an extensive analysis relied on the real vehicle fleet composition in Modena was not possible.Despite this limitation, an indicative comparison between the actual hot EFs considered in the emission computation for the city of Modena and the average hot EFs for the respective vehicles category between the 6 European countries mentioned before is shown in Figure5.The green rectangles indicate the EFs used in study, while the red rectangles represent the average EFs computed between Germany, Austria, Switzerland, France, Norway and Sweden, when available.The black horizontal segments below and above the red rectangles upper limit indicate respectively the minimum and the maximum EF for each respective vehicle category.

Figure 5 :
Figure 5: Comparison between weighted NO x hot EFs used in this study, computed following the EMEP/EEA guidelines (green rectangles) and the corresponding average NO x hot EFs computed between Germany, Austria, Switzerland, France, Norway and Sweden following the handbook of emission factors (HBEFA) version 4.1 (red rectangles).With horizontal black segment are also indicated the minimum and maximum HBEFA EF for the same six European countries.Please note the difference in the scale on the yaxis in each panel.
main statistical metrics employed in this study are Pearson correlation coefficient (r), Mean Bias (MB), Normalized Mean Bias (NMB), Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and the fraction of predicted values within a factor of two of observations, also referred as Factor of two (FAC2).These statistical indicators are defined as follows, with n the number of model-observation pairs, M the modelled values (with  ̅ difference of the wind direction)  :   = { min( − ,  −  + 360) ,   >  min( − ,  −  + 360) ,   <In order to evaluate the performance of the models in urban environment, the Fractional Mean Bias (FB), NormalizedAbsolute Difference (NAD) and Normalized Mean Square Error (NMSE) were also considered.They can be defined as follow: (T2), 10 m wind speed (ws10) and 10 m wind direction (wd10) meteorological fields predicted by WRF-Chem were compared against corresponding surface observations of these variables provided by 33 stations within the d03 domain (see Table3and Figure6): 18 stations (with 16 of them for T2, ws10 and wd10) belong to the RIRER (Rete idrometeo-pluviometrica integrata) Arpae-Simc network, 11 stations (with 7 for T2, ws10 and wd 10) belong to the Archivio dati idro-nivo-meteorologici ARPA Lombardia network, 4 stations belong to the Osservatorio Meteo Idrologico della Regione Liguria (OMIRL) ARPAL network.Other two stations belonging to the Osservatorio Geofisico di Modena weather network were considered for the micro-scale wind field evaluation.All the meteorological stations mentioned before are automated and realized according to WMO (World Meteorological Organization) directive.On the other hand, NO x concentrations at air quality stations are measured using chemiluminescence.With this method all the NO 2 contained in an air sample is converted to NO with a molybdenum converter, then the sample gas goes straight to the reaction chamber where NO (both the original NO contained in the air sample and the NO 2 converted to NO) will react with O 3 to form NO 2 and O 2 while emitting light.Finally, the measure of the emitted light will be proportional to the NO x concentrations in the air sample.

Figure 6 :
Figure 6: Map of the stations inside the WRF-Chem d03 domain.Site of the meteorological stations are reported by the green dots and site of air quality station are reported by the blue triangles.

Figure 7
Figure 7 shows the statistical performance of WRF-Chem in reproducing 2 m temperature.r is plotted as a function of MB for the three different model resolutions:15 km (d01), 3 km (d02) and 1km (d03).The variability of MB tends to increase by increasing the model resolution, conversely, the average r including all the stations is 0.78 for the d01 WRF-Chem domain, 0.80 for the d02 domain and 0.81 for the d03 domain, showing that the increase of the model resolution from 15 km to 1 km generally leads to slightly improve the performance of the model in reproducing 2 m temperature.

Figure 10 :
Figure 10: Factor of two (FAC2) reported in function of the Normalized Mean Bias (MB) between modelled hourly NO x concentrations and observation at 10 rural background sites for the three WRF-Chem resolutions: 15 km (d01), 3 km (d02) and 1 km (d03).

Figure 11 :
Figure 11: Hourly observed wind speed at MOD (on top), at OSS (in the middle) and at POL (on bottom) meteorological site along with hourly simulated wind speed by Micro-SWIFT, from October 28 to November 8, 2016.

Figure 12 :
Figure 12: Hourly observed wind direction at MOD (on top), at OSS (in the middle) and at POL (on bottom) meteorological sites along with hourly simulated wind direction by Micro-SWIFT, from October 28 to November 8, 2016.

Figure 14 :
Figure 14: Hourly observed concentrations of NO x at urban traffic (on top) and urban background (on bottom) measurements stations along with hourly simulated concentrations by WRF-Chem at 3 km resolution and PMSS combined with WRF-Chem at the same resolution, from October 28 to November 8, 2016.Please note the difference in the scale on the y-axis in each panel.

Figure 16 :
Figure 16: Mean daily cycle of observed NO x concentrations (black), modelled by the combination of WRF-Chem and PMSS model (red) and the only contribution of WRF-Chem (light blue), by station type (traffic or background) and by WRF-Chem resolution (3 km or 1 km).
The 2 m temperature and 10 m wind speed were captured well by the WRF-Chem model with statistical metrics in line with benchmark values suggested by the guidelines of the European Environmental Agency for meteorological mesoscale reconstruction, and with similar case studies related to the same area.Only few exceptions were observed at particular locations such as in mountainous area or close to the Ligurian sea, where complex orography and the local sea breeze strongly influenced the model bias in reproducing the meteorology.Despite, 10 m wind direction was poorly reproduced by WRF-Chem, its performance in term MAE and RMSE was in line with other cases studies focusing on the PO Valley, suffering both from the models difficulties in reproducing wind field during the situations with very little atmospheric circulation and low wind speed.Moreover, increasing WRF-Chem resolution from 15 km to 1 km resolution generally tended to slightly improve the model performance in reproducing 2 m temperature and 10 m wind speed.NO x concentrations reproduced in the Po Valley area by WRF-Chem were on average simulated reasonably well, but generally underestimated in almost all the rural background monitoring stations.The comparison with observations showed also that with an emissions inventory of 7 km horizontal resolution, the 1 km model resolution does not generally improve the results and the model configuration at 3 km resolution expressed the best performance in modelling NO x concentrations.
(Zaveri et al. 2008) chemical mechanism developed byEmmons et al. (2010), and the MOSAIC aerosol model(Zaveri et al. 2008)were used to simulate airborne pollutants over the nested domains.The first one includes 85 chemical species, 196 reactions and is consistent with the chemistry used in the global model that provides the chemical input and boundary conditions for the nested simulations.MOSAIC uses a sectional bin approach for the representation of the aerosol size distribution.The MOSAIC model predicts several aerosol species, such as sulfate, nitrate, ammonium, elemental carbon, and primary aerosols (POAs).Processes involving secondary organic aerosols (SOAs) formation were represented by the scheme based on

Table 2 :
Micro-SWIFT and Micro-SPRAY set-up.

Table 3 :
Observation sites.Locations are provided in Geographic coordinates (WGS84).Available parameters are: "T2" for temperature at 2 m height above the ground, "ws10" for wind speed and "wd10" for wind direction at 10 m above the ground and "NO x " for NO x concentrations.

Table 4 :
Statistics of hourly wind speed computed for the period between October 28 and November 8 at three urban meteorological stations.

Table 5 :
Statistics of hourly NO x concentrations computed for the period between October 28 and November 8, considering two different model configurations.