Evaluation of Emission Factors for Particulate Matter and NO₂ from Road Transport in Sofia, Bulgaria

Abstract

Traffic-related air pollution has a significant impact on the concentration of particulate matter (PM) and nitrogen oxides (NO_x) in urban areas, but there are many uncertainties associated with the modeling of PM concentration due to non-exhaust emissions. Bulgarian weather, road surfaces and traffic conditions differ significantly from the UK’s and other EU countries’ averages, which underpin many assumptions in established models. The hypothesis is that the emission factors differ from those used to calculate traffic emissions using the EMIT model. The objective of this work is to adjust the emissions for PM and the relationship between the fractions of NO_x and PM using the hourly mean concentrations from road transport and urban background automatic air quality stations in Sofia, Bulgaria. Various already-published and newly developed methods are applied to local observations to derive functions and relations that better represent Bulgarian road and traffic conditions. The ADMS-Urban model is validated and evaluated by comparing pollutant concentrations from simulations using original and adjusted emissions, showing an improvement in results after applying functions and relationships derived from local observations. This work is part of our efforts to improve air quality modeling in urban areas in Bulgaria.

1. Introduction

Traffic-related air pollution (TRAP) is formed by a complex mixture of gases and particulate matter (PM), and it has a significant contribution to urban smog. Vehicle emissions can be categorized into three groups: exhaust, non-exhaust and evaporative. Exhaust emissions from road transport arise from the combustion of fuels such as gasoline, diesel, liquefied petroleum gas (LPG), and compressed natural gas (CNG) in internal combustion engines. Motor vehicles emit a variety of pollutants, including nitrogen oxides (NO_x), elemental carbon (EC), PM not exceeding 2.5 μm in aerodynamic diameter (PM_2.5), ultrafine particles (UFPs) not exceeding 1 μm in aerodynamic diameter, heavy metals, polycyclic aromatic hydrocarbons (PAHs), and volatile organic compounds (VOCs). Vehicles also produce non-exhaust emissions, with most important sources of airborne PM not exceeding 10 μm in aerodynamic diameter (PM₁₀) likely to be due to the re-suspension of road dust, the abrasion of the road surface and the wear of brakes and tires. Evaporative emissions are a result of vapors escaping from the vehicle’s fuel system. Evaporative emissions are important only for VOCs. Pollutant generation from vehicle emissions depends on the type of vehicle (e.g., light- or heavy-duty vehicles) and age, operating and maintenance conditions, exhaust treatment, type and quality of fuel, wear of parts (e.g., tires and brakes), and engine lubricants used. A number of harmful air pollutants are emitted directly from vehicles, such as “primary” PM and NOx. Others, such as ozone (O₃) and “secondary” PM, are formed in the atmosphere from precursors such as NOx and VOCs.

In view of the increasing motorization in the European Union [1,2], which results in growing emissions of exhaust pollutants, coordinated action is being taken across the member states to reduce the harmful effects of vehicle exhausts. One of the key steps in reducing exhaust emissions was the introduction of the EURO standards [3]. Currently, these standards cover NOx, total hydrocarbons (THCs), carbon monoxide (CO), and PM emissions [4]. As a result of the introduction of such air quality regulations and improvements in vehicular emission control technologies, exhaust emissions from motor vehicles, along with the ambient concentrations of most traffic-related pollutants, have decreased steadily over the last several decades in the EU in general and in many of the high-income countries [5]. For instance, NO_x and PM_2.5 emissions for countries with similar population sizes, like Austria, Denmark, Finland or Portugal, have diminished [6]. Medium-income countries of similar population size such as Bulgaria and other Eastern European countries (e.g., Greece, Hungary, Czechia or Latvia) have had a modest to poor decrease in the generation of exhaust and wear-related non-exhaust emissions. The most probable reason for this is the motorization profile. This profile is characterized by three distinctive features: a higher share of older, predominantly second-hand transport fleets [7], with estimates for Bulgaria of a ratio of around 1 to 10 between new and newly imported second-hand passenger cars for the decade 2010–2020 [8]; increasing motorization rates; and a shift from smaller and older towards newer but bigger vehicles, most often Sports Utility Vehicles (SUVs). Additional national and local fleet data and tendencies, providing background to the current experiment, and data comparing vehicle types, are described in inventory-oriented studies [9] and annexes to a report [10].

The quantification of motor-vehicle emissions is critical in estimating their impact on local air quality and traffic-related exposures and requires the collection of travel-activity data over space and time and the development of emission inventories. Emission inventories are developed based on complex emission models, recently described in [11]. The review paper presents the current state-of-the-art in the field of exhaust emission models and traffic simulation. Emission models could be divided according to their precision scale into macro (regional, national), meso (local), and micro (areas of a dedicated part of a city, road intersections) models [12].

Macroscopic models are mainly based on the number of vehicles; their category, load, average speed, and distance traveled; and weather conditions, among other factors [13]. The most widely used macro-emission models are the European Computer Program to Calculate Emissions from Road Traffic (COPERT) [14] developed by the Joint Research Centre; the Mobile Source Emissions Factor (MOBILE) model, which has been superseded by the Motor Vehicle Emission Simulator (MOVES) [15], both maintained by the U.S. Environmental Protection Agency (EPA); the Handbook of Emission Factors for Road Transport (HBEFA) [16], the product of a common effort by funding agencies and development partners in six countries; and the Comprehensive Emissions Inventory Toolkit (EMIT) [17], supported by the Cambridge Environmental Research Consultants Ltd. (CERC), Cambridge, United Kingdom.

Microscale models need a large amount of data related to the measurements of vehicle parameters, such as acceleration and speed, as well as road parameters such as terrain gradients and position coordinates [18]. The most popular micro-emission models are the Comprehensive Modal Emission Model (CMEM) [19], maintained by the Centre for Environmental Research and Technology; services such as the Environmentally Sensitive Traffic Management (ESTM) [20] solution, developed by Bosch Mobility; and fleet clustering and real consumption models, supported by the Swiss Federal Laboratories for Materials Testing and Research EMPA [21].

The quality of the travel-activity data (such as vehicle miles traveled, number of trips, and types of vehicles) and the complex algorithms used to derive the emission factors suggest the presence of substantial uncertainties and limitations in the resulting emission estimates [22]. In addition to this, it should be noted that estimates of PM emissions have had very limited field valuation and verification [23].

A lot of uncertainty exists in TRAP modeling, especially for PM concentration, due to non-exhaust emissions [24]. An evaluation of existing models for PM emissions from abrasion sources, such as tires, brakes, and road surface wear and resuspension was conducted by Boulter et al. [25] to develop improved prediction methods. The study shows that there are only a few detailed methodologies for predicting emissions of PM from non-exhaust sources and that there is a need for more extensive empirical data. It is also clear from the study that significant quantitative insights can be gained by analyzing measured data from heavily instrumented traffic monitoring sites, and a number of general recommendations for methods of obtaining such data were given.

There are many gaps in the official ground air quality data in Bulgaria. The National Automated System for Environmental Monitoring [26] is operated by the Executive Environment Agency, Ministry of Environment and Water. Only 34 automatic measuring stations (including 4 in forest ecosystems), 5 differential optical absorption spectroscopy systems and 9 points with manual sampling and subsequent laboratory analysis were operated during the first 6 months of 2023 (the last available report) [27]. Only 10 of these stations measure PM_2.5 and they are mostly located in cities. Sofia has five automatic and one manual sampling measuring station (only one measures PM_2.5), and there are fewer (1–2) in other cities. Another 22 measuring sensor stations are operated by Sofia Municipality [28], and a network of low-cost citizen-owned sensor suites [29] are distributed around the city—both sources of data need additional verification and regular calibration.

Concentrations of PM₁₀ and PM_2.5 (and polycyclic aromatic hydrocarbons, PAHs) are elevated in most municipalities in Bulgaria, where air quality stations are deployed. For instance, high hourly average values are sporadically registered in big cities such as Sofia, Plovdiv, Pleven and Varna, and for the period 2015–2019, both the annual average concentration and the allowed number of annual exceedances (18 times in a calendar year) of the hourly limit values of NO₂ in the city of Plovdiv were exceeded a number of times [27,30]. In spite of official monitoring not showing a significant problem with NO₂ concentration, independent diffusion tube measurements conducted by non-profit civil society organizations show high monthly averaged values [31].

Similar problems with high PM concentrations are registered in other EU countries such as Poland [32] and southeastern European countries that exclusively make up the list of the top five countries with the highest ambient PM_2.5 levels in Europe: North Macedonia has the highest exposure in all of Europe, followed by Bosnia and Herzegovina, Serbia, and Montenegro [33]. The seasonal pattern of daily PM₁₀ concentration, with a maximum in the cold period of the year (October to March), is typical for Sofia [34]. This is due to variations in meteorological conditions and the increase in the number and strength of sources. Minimum values are observed during the summer months when intense turbulent mixing and a high atmospheric boundary layer (ABL) results in lower measured concentrations. It is important to note that on a diurnal scale, the pollutant, in particular PM₁₀, mixes well at the urban ABL.

In 2017, the European Court of Justice ruled against Bulgaria for exceeding PM₁₀ (daily and annual) limit values. Bulgaria is facing financial penalties as of December 2020, when the European Commission referred the country to the European Court of Justice for not implementing the 2017 court decision [35]. Additionally, in 2021, two long standing legal cases brought forward by citizens and nongovernmental organizations came to resolution in Plovdiv and Sofia. The court rulings set a 1-year deadline on both cities to achieve air quality norms and implement concrete measures to protect the health of citizens.

In search of a resolution, two atmospheric air quality management programs for Sofia were developed: one for the period 2015–2020, for PM [36,37], and another one for the period 2021–2026, which covers all harmful pollutants [38]. The preliminary findings of the two programs show that there is a significant difference in PM emissions for both periods. From the currently existing emission inventories for Sofia, it is apparent that there is a lack of informative and accessible data on the dynamics and structure of the traffic flow on the major road arteries, and this results in biassed modeling, which in turn may mislead policy- and decision-makers.

Air quality modeling can alleviate some of the deficiencies arising from the lack of official measurement and communication campaigns, but it requires a careful consideration of underlying assumptions and pre-set parameters in computational models. In a previous study [39], it was shown that there is a clear contrast in pollution mechanisms in different parts of Sofia. Transport has a substantial impact on PM concentration in the central parts of the city, where the pollution is mostly local, in proximity to the source, while the contribution of domestic heating from wood and coal burning is highest on the outskirts and in the small towns in Sofia Municipality. These local phenomena must be accounted for. In the same study, some of the deficiencies occurring when conducting an evaluation of the complex pollution processes in urban areas were defined, and concrete steps towards improving the modeling at the urban scale in Bulgaria were formulated. An example of such a step is the development of a high-quality traffic emissions inventory for Sofia [9,40]. Using a tailor-made inventory reduces the uncertainties in existing inventories and helps improve the applicability of established models to the local setting.

The atmospheric dispersion modeling system (ADMS-Urban), which is a comprehensive software suite for modeling air quality in urban areas and close to motorways, roads, and large industrial areas [41,42], was used in this study. The same model was employed in the work of Dimitrova and Velizarova [39], but the present study uses a new high-quality traffic emissions inventory, advanced street canyon modeling, and background concentration, and includes chemical transformations and high-resolution topography description. Despite some deficits of the local Gaussian-type models for air quality studies, the potential of resolving emissions and the physical environment at a fine scale makes them a very useful tool. However, the depiction of the spatial concentration pattern of regional air quality models underrepresents pollution (especially for PM) [43] due to the restriction of their spatial resolution, as they are unable to capture the high concentration gradients nearby roads with heavy traffic and from other sources near the surface. In high-fidelity modeling, accounting for the effect of line sources (e.g., major road arteries with heavy traffic) requires the usage of small spatial steps in the order of tens of meters, enabling us to register the large gradients in proximity of the segment. Pollution concentration decreases rapidly with an increase in the distance from the road, even if the effect of the street canyon is not taken into account for the simulations. It is possible to add further receptor points within the canyon in order to more accurately detect the effect of pollutant retention within the canyon under certain conditions, which is not achievable with the regional models.

Weather, road surfaces and traffic conditions in Bulgaria can also differ significantly from the UK’s and other EU countries’ averages, which underpins many assumptions in established models.

Normal vehicle ageing and wear is likely to be an important factor in transport-related emissions. In Bulgaria, the average age of cars is 16 years old, with almost 70% of all vehicles having a standard of performance below Euro 4, and more than half (56.5%) with no Euro category at all, according to recent investigations [38]. In comparison, the average age of passenger cars is 11.5 years old in the EU, 8 years old in the UK, and 8.6 years old in Switzerland [44]. Fuel type is also important. In [45], the authors show that there is little evidence of NO_x emissions reduction from all types of diesel vehicles over the past 15–20 years, and only for petrol passenger cars (including hybrids) was there any evidence for effective NO_x control. In this setting, 43.8% of cars in Bulgaria are powered by diesel engines, compared to 50.1% for petrol powertrains [38] (a rate that is slightly higher in comparison to the average split in EU countries) 42.3% for diesel and 52.9% for petrol [44]. There has been recent growth in the demand for electric vehicles in the country, but considering the relatively low purchasing power of the population, there are certain limits to replacing the car fleet with zero-exhaust-emission vehicles.

Road surface wear and dust deposits also vary by type of road and vehicle speed. The typical speed limit in Bulgarian cities is 50 km/h, but on higher-class transit roads, the limit reaches 70 km/h or 80 km/h, which increases dust emissions from the road. In addition, the intensive traffic flow through the central and densely occupied built and intersected parts of the city increases the occurrence and severity of congestions, which further deteriorates air quality. After a significant loss of passengers in the previous two decades, public transportation is slowly recovering part of the lost ridership, with a significant constraint experienced during the pandemic of COVID-19. Coverage by bicycle lanes is much lower than in major European cities, with the existing network of cycle routes being insufficiently integrated, uncomfortable, and sometimes unsafe.

Considering the societal and technological challenges outlined so far, this work has two main contributions:

• To provide a means of adjusting the emissions of PM and the rates between NO₂ and NO_x gaseous pollutants, using hourly mean concentration measurements from road transport and urban background automatic air quality stations (AQSs) in Sofia, Bulgaria. Different already-published and new methods are explored and evaluated.
• To estimate the contribution of PM₁₀ traffic emissions. To identify the contribution of the main groups of PM₁₀ sources in Sofia, the results of a receptor-oriented positive matrix factorization (PMF) analysis for a wide variety of chemical elements [34] are presented. This study covers one year (2019–2020), but selected periods in April and September 2019 are used to validate the newly developed PM₁₀ traffic emissions. These periods were chosen considering several important factors—the absence of domestic heating and additional background sources such as dust transport from the Sahara Desert, lack of precipitation and low wind speed.

The remainder of the paper is arranged as follows: the materials and methods are described in Section 2. Section 3 presents results from experimental adjustment and modeling, and an overall discussion and conclusions follow in Section 4 and Section 5.

2. Materials and Methods

2.1. Research Question and Hypothesis

The working hypothesis is that modeling at a local scale can be improved by implementing a more reliable inventory of traffic emissions if the relationships used to develop the emission factors are derived using local data that is more representative of Bulgarian conditions. To help test the hypothesis, three main factors are considered: data availability and gaps; the application of different methods to process sparse data; and testing necessary assumptions, conducting experiments, and model validation.

Data at all three traffic AQSs in Sofia for a 12-year period (from 2009 to 2020) were used for local NO₂ concentration adjustment, and data for the same period at AQS Pavlovo (traffic station) and AQS Hipodruma (urban background station) were used to estimate the proportional factors between PM₁₀ and the coarse fraction of PM (PM₁₀-PM_2.5) ($\mathit{P}{\mathit{M}}_{\mathbf{10}}-\mathit{P}{\mathit{M}}_{\mathbf{2.5}}$). For dispersion modeling and model validation against PM₁₀ obtained through PMF analysis, selected periods in April and September 2019 were used. A conceptual workflow diagram of the study, including the input data, models used and experimental adjustment of the TRAP emissions in Sofia, Bulgaria, is shown in Figure 1.

2.2. Data

2.2.1. Air Quality Data

Hourly measurements for meteorological factors and pollution concentrations are available from the National Environmental Monitoring System of the Executive Environment Agency (EEA) at the Ministry of Environment and Water at six AQSs for the period from 2009 to 2020. Three of the sites located in Druzhba, Nadezhda and Hipodruma are classified as urban background stations, and the other two, in Pavlovo and Orlov Most, are classified as transport stations. The AQS at Orlov Most was moved to a new location in Mladost in November 2015, and as a result, the data sets from both sites are shorter.

Quartile charts for NO₂ concentration from the traffic stations are shown in Figure 2. A direct comparison with the established norms for the levels of concentration of NO₂ related to the protection of human health, set by European directives and implemented into national legislation [46], can be made on the basis of the average annual limit of 40 µg/m³ and the threshold of an hourly average rate of 200 μg/m³, which may not be exceeded more than 18 times in a calendar year. Concentrations above the hourly average rate are registered at AQS Pavlovo during the first years from 2009 to 2011, and there were eight exceedances at AQS Orlov Most in 2009 (the number of exceedances is below the threshold of 18 incidents within one year). The annual average concentration is close to or above the limit for all sites. These data are used for the calculations described in Section 3.1.1.

Quartile charts for the PM₁₀ concentration from the traffic stations are shown in Figure 3. The annual limit for PM₁₀ is 40 µg/m³, and the threshold of a 24 h average rate is 50 μg/m³, which may not be exceeded more than 35 times in a calendar year. The annual concentration limit is exceeded until 2014 and slowly decreases after. The number of days with PM₁₀ concentrations above the daily limit exceeded 35 for all years. PM_2.5 concentration measurements are available only at AQS Hipodruma (an urban background station). These data are used for the calculations described in Section 3.1.2.

Road concentrations strongly depend on transport emissions, which vary with the time of day, the day of the week and the season. Average NO₂ and PM₁₀ concentration profiles for weekdays, Saturday, and Sunday (based on the entire 12-year period) for different traffic stations are presented in Figure 4. A diurnal variation shows a well-established binary profile with two peaks in the morning and evening hours. The dynamic is similar for all traffic stations, with the highest concentrations at AQS Orlov Most. During the weekdays (Figure 4a), the first peak concentration is at 9–11, with the second at 19–21. There is a 2 h shift in the evening peak at AQS Mladost, probably due to the location of the station in comparison with the other two. AQS Pavlovo and AQS Orlov Most are located near main transport roads, which connect different parts of the city with the city center. AQS Mladost is close to a big road with somewhat intense traffic that mainly takes local flow, joining one of the major transit traffic axes in the city “Tsarigradsko shoes”, which is also relatively close but at a lower elevation. The morning peak is smoothed on Saturdays (Figure 4b) and almost disappears on Sundays (Figure 4c).

2.2.2. Meteorological Data

The meteorological data for the studied period were provided by the National Institute of Meteorology and Hydrology (NIMH). Both soundings and surface measurements are from the CMO—an experimental base of the NIMH in Sofia and part of the official network of the World Meteorological Organization. The surface observations collect air temperature (T) and relative humidity (f) at 2 m and wind speed (WS) and wind direction (WD) at 10 m, measured with an Vaisala AWS310 [47] meteorological station. The meteorological observations are checked, cleaned, and averaged at 10 min intervals for the entire study period.

Hourly averaged values were used as an input to the ADMS-Urban model, shown in Figure 5. The amplitudes in the fields of T and f are more significant in April compared to September. The temperature is near zero in the early morning hours, increasing to above 20 °C in the afternoon, with a night-time rise at the end of the period (23–26 September). The f decreases during the day and increases above 90% at night for some days, but it never reaches saturation levels (an indication of rain or fog). The wind rose for selected days in April and September, as shown in Figure 6. The prevailing WD is from the southeast, with a low speed of less than 3 m/s.

Aerological soundings were conducted with the Vaisala MW41 [48] sounding system and RS-41SG [49] aerosondes and include data for T, f, WS and WD. The data set consists of midday soundings for periods in April and morning and midday sounding for periods in September. Vertical profiles for days with extremal discrepancies between modeled and observed PM₁₀ concentrations for the two study periods are shown in Figure 6. The largest discrepancy between modeled and observed concentration values is characterized by zero wind velocity (Figure 7a–c). A strong inversion layer is observed in all potential temperature profiles at 06:00 UTC (Figure 7b,e), and a mixed layer up to 1000 m above sea level (ASL) is observed at 12:00 UTC (Figure 7a,c,d,f). Relative humidity is below 80% on all occasions except for the morning hours on the 24, 25 and 26 September. All available profiles are presented in Supplementary Materials S1.

2.3. Modeling

2.3.1. Transport Emission Inventory Modeling

The emissions data for this study were provided by a new traffic emission inventory for the city of Sofia, developed by the authors. Emissions are calculated with EMIT using the activity data described below. To calculate exhaust emissions, data for 2010 were used; from the 2014 urban factor data set, part of the UK National Atmospheric Emissions Inventory (NAEI) was used [50]. This particular year was chosen because it best represents the fleet composition in Sofia, in terms of its average age, for the study period. The fleet approximation is based on comparison between the average age of the fleet, which in Bulgaria is around 8 to 10 years older than that in the UK. In the EMIT input data, most of the vehicles manufactured prior to the introduction of the Euro 4 standard are assumed to lack properly working catalyzers and filters, as is the case observed in practice, although there is no conclusive data available about this well-known and widely discussed phenomenon in the country. It must be noted that this approach only offers a proxy, because vehicles with first registration circa 2010 would have experienced some additional age-related degradation by the time of this study. Thus, the expected emission factors would be higher for all categories, except for newer vehicles, which hold a relatively small share of the fleet.

One of the biggest challenges is the quality of basic, publicly available traffic and fleet composition data. The data come from a variety of sources with poor accessibility, and in some cases, there is questionable integrity and many methodological inconsistencies in the data gathering approaches [9]. This is a typical administrative practice leading to fragmentation, stemming from sectorized institutional silos. In an attempt to overcome this fragmentation, the approach proposed in [40] is applied in this work. In summary, this approach attempts to fill the gaps in the data for traffic along primary and secondary streets by first collecting data for a large number of related features such as road and street categories, population density, functional analysis, spatial syntax, previous traffic, noise and NO₂ and then applying ensemble learning to generate insight.

After a series of consecutive data imputations, the annual average daily traffic distribution in the street and road network of the city of Sofia and the metropolitan municipality for 2018 and 2022 (the years with more comprehensive data availability) was estimated. The traffic data modeled for 2018 were used to calculate the exhaust and non-exhaust traffic emissions using EMIT. The chosen year is the previous closest to and most in tune with the previously described meteorological and pollution data from 2019. This period of two years is also just before the big shifts in mobility behavior around the period marked by the COVID-19 pandemic lockdowns and recoveries.

The composition of the vehicle fleet was divided into three groups—heavy, light and motorcycles—according to the available activity data. The non-exhaust emission factor data sets were taken from the European Monitoring and Evaluation Programme (EMEP) [51] for tire, road, and brake wear, and from the Department for Environment, Food and Rural Affairs/Transport Research Laboratory (DEFRA/TRL) for resuspension [25]. Detailed information about the emission factor data sets can be found in Appendix A of the EMIT User Guide [17].

The used approach also implements new route-type data following the NAEI2014 Urban 2014 version 2 detailed taxonomy available in EMIT. It is based on multi-dimensional stratification including fuel types, vehicle types, engine volume and vehicle weight categories, types of EURO standards, and types of cleaning technologies, among others [33]. These data provide the link between the traffic count data/fleet components and vehicle sub-categories with the emission factor. Active traffic data are shown in Figure 8.

2.3.2. Atmospheric Dispersion Modeling in Urban Street Canyons

ADMS-Urban is a second-generation Gaussian dispersion model with an integrated street canyon model that considers the impacts of street canyons on dispersion, turbulence and mixing induced by traffic. It incorporates the latest understanding of the boundary layer structure, using advanced algorithms for the height-dependence of wind speed (WS), turbulence, and stability. ADMS-Urban can model point, area, line, road, volume and grid sources. The model integrates with geographical information system (GIS) tools and emissions databases and supports calculation of concentrations based on a range of averaging periods, including percentiles. The GIS interface is an important visualization tool and provides a means of manipulating the spatial element of the model input data. ADMS-Urban is developed and maintained by the Cambridge Environmental Research Consultants (CERC) Ltd. More details regarding the model are available in [52].

Concentrations of NO₂, with a surface lifetime of around 1 day, are primarily related to the dispersion and the chemical transformation of local emissions. The NO_x chemical reactions take place over a relatively short time, and in order to calculate NO₂ concentrations, NO_x chemistry needs to be taken into account. ADMS-Urban has an integral chemistry model, using what is known as the “generic reaction set” [53], which includes the reactions between ozone (O₃), VOCs and the oxidation reactions of sulphur dioxide (SO₂) that lead to the formation of PM₁₀ and PM_2.5 of ammonium sulphate. The background concentrations in this study were taken from the Copernicus Atmosphere Monitoring Service (CAMS) reanalysis [54], which are required by the model to enable the use of the chemical reaction module. Only the dry deposition option was used due to a lack of precipitation during the selected simulated time periods.

The advanced street canyon module was employed in this study. This module provides a more detailed description of the dispersion within the canyon, due to the consideration of a wider range of canyon geometries, including the effects of tall canyons and asymmetries [55]. The model was also used in “network” mode, which allows for the passing of pollutants from the end of one street canyon into another canyon connected to it. The advanced street canyon module requires the specification of additional input parameters, which were obtained via the street canyon tool [55]—a Python plug-in that is designed to export street canyon geometries from within ArcGIS.

By using the time-varying emission factors option, different diurnal profiles for each day of the week and monthly profiles for each source were applied. The temporal profiles were taken from the Dutch Organization for Applied Scientific Research (TNO) [56].

The interested reader may refer to the ADMS—Urban v5.0 user guide [52] and the technical specifications [57] for more information regarding the model settings used in this study.

The modeling domain over Sofia and the locations of the AQSs (where the measurements used in this study were taken from) is shown in Figure 9a. The smaller area in the green rectangle magnified in Figure 9b shows the modeling domain near AQS Mladost employed for emission factors adjustment. Two different sites are used for model validation: AQS Mladost for NO₂ and the central meteorological observatory (CMO) in Sofia for PM₁₀.

2.4. Experimental Design of Emissions Adjustment

2.4.1. Study Approach for NO₂ Concentration Adjustment

By default, if NO₂ emissions are not included explicitly, ADMS-Urban assumes that a certain percentage of the NO_x emissions are in fact due to NO₂. This percentage is modeled as an empirical relationship between hourly NO₂ and NO_x based on a third-degree polynomial produced by Derwent and Middleton [58]. To develop the polynomial, the authors of that paper analyzed hourly concentrations of NO_x, sorting the data into 10 bins of increasing values in parts per billion (ppb). The corresponding NO and NO₂ concentration values were indexed according to the sorted NO_x data. The NO₂ concentration values corresponding to each bin were then averaged, and a curve was interpolated between NO₂ mean concentrations and upper NO_x bin limits. The relationship was derived as

[NO₂] = 2.166 − [NO_x] (1.236 − 3.348A₁₀ + 1.933A₁₀² − 0.326A₁₀³),

(1)

where the square brackets indicate the hourly average concentration in ppb, and A₁₀ = log([NO_x]). This function has limits in the range 9.0 ppb < [NO_x] < 1141.5 ppb. Below 9.0 ppb of NO_x, the efficiency of [NO₂]/[NO_x] ratio was limited to 0.723. Above 1141.5 ppb of NO_x, the [NO₂]/[NO_x] ratio was limited to 0.25. The derived polynomial function is based on one year (May 1991 to June 1992) of data collected at a fixed monitoring location in London.

A follow-on study [59], based on a larger amount of data (1991–1997), investigated the validity of new empirical relationships at several sites. The point of significant difference from previous studies was to consider the ratio between NO₂ and NO_x concentrations as a dimensionless yield quantity: Y = [NO₂]/[NO_x], where [NO_x] = [NO] + [NO₂] in ppb, and 0 ≤ Y ≤ 1 is dimensionless by construction. The data were sorted in 10 ppb bins by NO_x concentration, and a fourth-order polynomial was fit to the log-transformed upper NO_x bin limits (independent variable) and average NO₂ values corresponding to the NO_x bins. The upper bin limits were used in defining the yield to stabilize the result by avoiding division by small values. The least-squares polynomial fit to the processed data, recommended for urban areas outside London, was reported as

Y₂ = −3.083 + 7.472 A₁₀ − 5.116A₁₀² + 1.382A₁₀³ − 0.129A₁₀⁴,

(2)

The polynomials used in both cases are highly contextual, as they have been fitted on data reflecting the pollution mechanisms in the UK. When modeling pollution dispersion in Sofia, the relationship between NO_x and NO₂ needs to be adjusted based on local dynamics. To do this, a similar method is used to obtain an empirical relationship between NO₂ and NO_x using the hourly mean concentration measured at the traffic stations in Sofia. The NO_x data are split into 10 ppb bins and log-transformed. The corresponding NO₂ concentrations are averaged for each bin, and the bin yield is calculated as the ratio of the average NO₂ to the upper NO_x bin limit. It is important to note that, due to their fixed end points, the bins contain a variable number of values, so the average NO₂ concentration will be a variable-quality estimate of the true mean. Even though, in principle, the standard error of the mean can be computed for each bin and an interval regression can be used to learn the relationship between NO₂ and NO_x (see, e.g., [60]), a regular point-value regression is used here, with some extensions considered as part of future efforts. Hourly concentrations for NO₂ do not order in the same way as NO_x values due to a number of variables that influence the formation of NO₂, such as turbulent mixing, temperature, light intensity, and the presence of ozone, among others.

2.4.2. Study Approach for PM Emissions Adjustment

The primary emitters of particulate pollution—industry and traffic—generate PM predominantly in the fine particle size range (PM_2.5), while emissions from activities such as construction work and those from natural sources, such as pollen, contribute predominantly to coarse particle concentrations (particles greater than 2.5 µm in aerodynamic diameter). As discussed in Section 3.1.2. a major contribution arises from surface soil and dust on paved areas becoming resuspended due to the wind flow or due to turbulence induced by passing road traffic. This makes the estimation of resuspension contributions particularly important for completing a traffic emissions inventory. However, estimating the non-exhaust sources, including the contribution of resuspension, is very challenging due to the lack of direct measurements and various impracticalities related to the necessary road measurements. To address this issue, a methodology to calculate the total traffic emissions of PM₁₀ was developed by [25].

The total emissions attributable to the non-exhaust sources can be expressed as the sum of the abrasion emissions (brake, tire, and road wear) and a resuspended component:

$${\mathrm{E}}_{\mathrm{total}}={\mathrm{E}}_{\mathrm{tire}}+{\mathrm{E}}_{\mathrm{brake}}+{\mathrm{E}}_{\mathrm{road}}+{\mathrm{E}}_{\mathrm{resusp}},$$

(3)

Emissions for tire, brake and road surface wear can be estimated from EMEP emission factors and traffic fleet data. Coupled with an estimate of total PM, Equation (3) can be rearranged to obtain an estimate for resuspension. The approach taken to estimate the contribution of resuspension relies upon several fundamental assumptions. Firstly, it was assumed that PM₁₀ and NO_x behave similarly in the atmosphere and therefore the roadside incremental values of the two, which are calculated as the difference in concentration measurements between a traffic station and an urban background station [61,62], are closely related. This enables the estimation of the total PM₁₀ emissions from traffic ($E_{P{M}_{10}}$) based on roadside increments of PM₁₀ (∆PM₁₀) and NO_x (∆NO_x) and calculated NO_x emissions ($E_{N{O}_x}$), as in the following equation:

$${\mathrm{E}}_{{\mathrm{PM}}_{10}}={\mathrm{E}}_{{\mathrm{NO}}_{\mathrm{x}}}\left(\frac{\Delta {\mathrm{PM}}_{10}}{\Delta {\mathrm{NO}}_{\mathrm{x}}}\right)$$

(4)

A second assumption in the initial calculations is that the roadside increment of PM_2.5 was solely attributable to vehicular exhaust sources and that non-exhaust emissions were largely confined to the coarse particles fraction (${\mathrm{PM}}_{\mathrm{coarse}}={\mathrm{PM}}_{10}-{\mathrm{PM}}_{2.5}$). Apportioning total PM₁₀ emissions into PM_2.5 and PM_coarse fractions according to the ratios between their observed concentrations allows a value for total PM_coarse (i.e., non-exhaust PM) to be calculated as

$${\mathrm{E}}_{{\mathrm{PM}}_{\mathrm{coarse}}}={\mathsf{\beta }\mathrm{E}}_{{\mathrm{PM}}_{10}}$$

(5)

where β can be calculated as the slope of the line of best fit between concentration measurements for PM_coarse and PM₁₀.

Finally, the following relationships for the PM₁₀ and PM_2.5 components of the road traffic emissions are used in ADMS-Urban:

$${\mathrm{E}}_{{\mathrm{PM}}_{\mathrm{coarse}}}={\mathrm{E}}_{\mathrm{tire},\mathrm{coarse}}+{\mathrm{E}}_{\mathrm{brake},\mathrm{coarse}}+{\mathrm{E}}_{\mathrm{road},\mathrm{coarse}}+{\mathrm{E}}_{\mathrm{resusp}}$$

(6)

$${\mathrm{E}}_{{\mathrm{PM}}_{2.5}}={\mathrm{E}}_{\mathrm{exhaust}}+{\mathrm{E}}_{\mathrm{tyre},2.5}+{\mathrm{E}}_{\mathrm{brake},2.5}+{\mathrm{E}}_{\mathrm{road},2.5}$$

(7)

where ${\mathrm{E}}_{\mathrm{x},\mathrm{coarse}}$ are the coarse parts of the respective emissions, ${\mathrm{E}}_{\mathrm{x},2.5}$ are their fine fraction parts and ${\mathrm{E}}_{{\mathrm{PM}}_{2.5}}$ is the total PM_2.5 emissions.

2.4.3. Source Apportionment Technique

The majority of non-exhaust PM derives from the resuspension of material already deposited on the road surface (between wheel tracks, on the kerbside or in the pores of the asphalt) due to tire shear, vehicle-generated turbulence and the action of the wind [63]. There are several sources that contribute to road dust, such as particles from brakes, tires, and road wear; particles deposited from exhaust emissions; and particles from nearby environments, including fugitive loading from construction sites, pavements, unpaved roads, dry and wet deposition from the atmosphere, salt residue during freezing periods, traction sand, and the deposition of pollen and plant matter [64]. Because of the heterogeneity of dust sources, it is difficult to distinguish between “direct” wear emissions (tire, brake or road wear) and “resuspended” wear emissions, and it is therefore hard to separate their relative contribution to atmospheric PM levels [65].

Source apportionment (SA) is the technique that relates a source emission (an activity sector or an area) to the ambient air concentration of a pollutant. Positive matrix factorization (PMF, also known as non-negative matrix factorization) is the most frequently used factor analysis technique. It solves a weighted factorization problem with non-negativity constraints using known experimental uncertainties as input data, thereby allowing for the individual treatment of matrices. For SA, measured concentrations and their uncertainties are used to solve the mass balance equation X = G × F + E, where X is the chemical composition matrix, G is the source contribution matrix, F is the factor profile matrix, and E is the matrix of residuals. The most recent versions of the EPA-PMF [66] software solve the factorization problem using the conjugate gradient method and contain routines to estimate the optimal number of factors, to test the rotational ambiguity and introduce constraints. The uncertainty and stability of the solution are estimated by bootstrapping and displacement methods [67].

The results from a SA study in Sofia covering one year (January 2019–January 2020) are presented in [34]. The data were analyzed using the EPA-PMF 5.0 model. More than 200 samples were studied, and PM chemical composition and the application of the PMF method led to the identification of eight factors, six of which have the most significant contribution to the PM₁₀ mass concentration in Sofia. Sources of ambient PM₁₀ have been grouped into eight categories: Resuspension (RES), Secondary (SEC), Biomass burning (BB), Traffic (TR), Industry (IND), Nitrate-rich (N), Fuel-oil burning (FUEL), and Mixed SO₄²⁻ (SO₄). The analysis showed that the resuspension factor is the main contributor to the total PM₁₀ mass (25%), followed by BB (23%), SO₄ (19%), SEC (16%), TR (9%), IND (4%), N (4%), and FUEL (0.4%) in Sofia.

The “resuspension” category includes contributions from soil and road dust resuspended into the surface air layer from the wind and the mechanically induced resuspension of air particulate from the vehicles. This factor has seasonal variation, with a minimum at the beginning of 2019 (January–February) and a maximum in the middle of December 2019, when stagnant weather conditions with prolonged inversion and the presence of fog led to high PM₁₀ concentrations. The relatively high impact of resuspension (33%) is connected to the dryer periods. In 2019, significantly lower monthly precipitation amounts were measured. After June and in December 2019 and January 2020, the total monthly precipitation was less than 60% of the climatic norm. The traffic factor accounts for 9% of the measured PM₁₀ and has a weak seasonal pattern, with maximums in the cold periods in Sofia (13 μgm⁻³). More details related to the method, data sampling, analyses of chemical composition in PM₁₀ and selection of different categories based on specific elements can be found in [34].

3. Results

3.1. Results from Experimental Adjustment

3.1.1. Local NO₂ Concentration Adjustment

Data at the three traffic AQSs in Sofia for a 12-year period (from 2009 to 2020) were used (see Figure 2) to fit the polynomial regressors. To identify the polynomial that most adequately describes the data, a model selection study was set up with polynomials from the first to the fourth order with all possible combinations of terms. The models are denoted as [i₁, i₂, …, i_n], where i_j are the powers of monomials having non-zero terms. Thus, for example, [0, 2, 4] denotes $\mathrm{y}={\mathsf{\beta }}_0+{\mathsf{\beta }}_2{\mathrm{x}}^2+{\mathsf{\beta }}_4{\mathrm{x}}^4$. Each polynomial was fit on data from individual stations, as well as data from all stations combined. For the purposes of this study, adequacy is quantified by the coefficient of determination, R² and the Akaike information criterion (AIC).

Goodness of fit is measured through the adjusted R², which penalizes the performance score of the model based on its complexity. The AIC is also computed to provide a measure of the complexity of each polynomial with respect to the available training data. In this model selection study, only polynomials with a non-zero intercept were tested due to the nature of the data established in an exploratory data analysis. Note that some polynomials, such as [1, 2, 3] fitted to the data from AQS Orlov Most, also perform well (adjusted R² = 0.985, AIC = −356.2) without an intercept term, although such cases form a minority.

Table 1. Summary of model selection study. Best performing models are bolded for each data set.

	All Stations		Pavlovo		Mladost		Orlov Most
Active Terms	R2 adj.	AIC	R2 adj.	AIC	R2 adj.	AIC	R2 adj.	AIC
[0, 1]	0.92	−249.4	0.86	−130.6	0.84	−102.8	0.94	−268.2
[0, 1, 2]	0.92	−247.9	0.89	−136.8	0.96	−142.7	0.95	−275.0
[0, 2]	0.91	−242.6	0.89	−138.0	0.90	−119.1	0.90	−238.9
[0, 1, 2, 3]	0.99	−369.1	0.98	−201.0	0.98	−166.0	0.99	−372.4
[0, 2, 3]	0.93	−257.2	0.88	−136.0	0.94	−135.0	0.96	−296.1
[0, 1, 3]	0.92	−247.4	0.88	−134.5	0.95	−137.2	0.95	−279.0
[0, 3]	0.88	−220.1	0.88	−134.9	0.93	−131.1	0.85	−209.9
[0, 1, 2, 3, 4]	0.99	−371.2	0.98	−200.3	0.98	−164.0	0.99	−370.4
[0, 2, 3, 4]	0.98	−348.4	0.98	−201.0	0.98	−164.5	0.99	−367.4
[0, 1, 3, 4]	0.99	−355.9	0.98	−202.2	0.98	−165.4	0.99	−369.5
[0, 1, 2, 4]	0.99	−363.7	0.98	−202.1	0.98	−165.9	0.99	−371.8
[0, 3, 4]	0.95	−276.7	0.89	−140.2	0.94	−132.0	0.98	−329.3
[0, 2, 4]	0.94	−262.2	0.88	−136.2	0.94	−132.0	0.97	−306.6
[0, 1, 4]	0.92	−247.6	0.87	−132.6	0.94	−132.4	0.95	−283.2
[0, 4]	0.83	−198.6	0.85	−126.5	0.94	−134.0	0.78	−188.2

shows the results from the model selection investigation.

To verify the value added by increasing the order of the polynomial for each data set, a complexity-oriented F-test was also performed. The F-statistic represents the result of a pairwise hypothesis testing procedure, where the null hypothesis is that the more complex model of the pair does not provide a significantly better fit than the simpler one. The p-value for the test is also reported to help gauge the statistical significance of the test result. The hypothesis testing was limited to full q-degree polynomials, that is, polynomials of degree q, which have non-zero coefficients for all terms of degree q-r with r = {0, 1, 2, …, q}. The results of the F-tests are summarized in Figure 10. Each tile in the figure shows the performance of the polynomials on the data set indicated in the respective title. The main diagonal in each tile contains the adjusted R² for reference to the overall performance of the model. The upper triangle shows the values of the F-statistic, which are colored in shades of blue, with darker colors signifying more desirable results. The lower triangle contains the p-values for the corresponding F-statistic as the cumulative probability of this statistic under an appropriate F-distribution. The definition of all statistics and metrics are given in Supplementary Materials S2.

The critical significance level was set to 5%, with test results indicated as significant (p-value < 5%) highlighted in green and the rest left in white. It can be quickly seen that despite the increases in R² for more complex polynomials, the improvements may not be statistically significant to the chosen level and may often be far from it. This is always the case for comparisons between third- and fourth-order polynomials. In one case that did not feature in the comparisons, the [0, 1, 3, 4] model performed better than the [0, 1, 2, 3] one, but these could not be compared in an F-test because both have the same number of coefficients. The minor improvement in R² (in the order of 10⁻⁴) was neglected. In view of these results, a full third-order polynomial was selected for the rest of the work in this paper. To provide a visual assessment of the quality of fit, Figure 11 shows each data set with the corresponding third-degree polynomial superimposed as a black line.

As an additional comparison, the functional maps developed by Derwent and Middleton [58] in Equation (1) and by Dixon et al. [59] in Equation (2) for use in London produce results that severely misrepresent the behavior of NO₂ concentration in Sofia. That is, if these expressions are evaluated on the available NO_x data, their results are, perhaps unsurprisingly, around three orders of magnitude worse than those from the customized polynomials.

Table 2. Comparative performance between two established models and the polynomial developed in this study. RMSE in µg/m³.

Air Quality Monitoring Site	According to [59] (Yield)	This Study (Yield)	According to [58] (NO2)	This Study (NO2)
All stations	10.9	0.013	8457.2	6.0
Pavlovo	5.4	0.018	3476.4	4.5
Orlov Most	10.9	0.013	8457.5	6.0

summarizes the performance, in terms of RMSE, of the two models and compares it to the performance of the models developed specifically for this study. As expected, RMSE scales strongly with sample size, but despite this, the custom polynomials achieve much better results in both the NO₂ and yield representations. This serves as a reminder that modeling assumptions must be challenged and relaxed whenever possible to obtain meaningful results. Using ADMS-Urban for other cities will likely require a similar approach.

3.1.2. Local PM Concentration Adjustment

As already mentioned in the introduction, only one AQS in Sofia measures PM_2.5, and it is an urban background station. The two traffic AQSs measure only PM₁₀ concentration. The air quality stations at Pavlovo (traffic station) and Hipodruma (background station) were used in this study for two reasons. Firstly, the sites are located in relative proximity to each other (~2.6 km; see Figure 9a), which makes their observation more relevant to the local pollution dynamics. Secondly, all pollutants (except PM_2.5 at AQS Pavlovo) and meteorological variables were available at both stations, which becomes useful in estimating PM_2.5 concentrations at the traffic station, AQS Pavlovo. Roadside increments of PM₁₀ (∆PM₁₀) and NO_x (∆NO_x) were calculated using data for 2019. Entries with missing or erroneous values for ∆PM₁₀ and ∆NO_x, as well as those with an ∆PM₁₀/∆NO_x value larger than 0.1, were removed, corresponding to the data-cleaning steps recommended in [25]. A representative incremental ratio used in Equation (4) was computed as the slope of the zero-intercept line of best fit to the data. Figure 12 depicts the cleaned data and line of best fit. The final ingredient for estimating the total adjusted PM₁₀ emissions in Equation (4) are the total NO_X emission (E_NOx) for each road source, which were calculated using EMIT.

To estimate the adjusted PM_coarse emissions, using Equation (5), β must first be calculated at the selected traffic station (AQS Pavlovo). In the case of Sofia, this necessitates an estimation of the concentration of PM_2.5 at the traffic station, which can be carried out by mapping PM_2.5 concentrations from the only AQS that measures them (AQS Hipodruma; background) onto other measured quantities at that station and then applying the map to the traffic station. This is where the second reason for the selection of AQS Pavlovo and AQS Hipodruma—the two measure the same quantities, except PM_2.5—plays an important role.

Several linear regression models were developed to serve as a map. Models of the first and second order with all interactions were fit to data for 12 measurands at AQS Hipodruma for the period from 2015 to 2020, due to the lack of data for the majority of measured factors at AQS Pavlovo for the 2009–2014 period, which precludes the modeling efforts for this period. For the first-order model, only temperature was found to be a statistically insignificant factor, whereas a total 25 terms from the second-order model, including the first-order contributions of temperature and relative humidity and second-order contributions of carbon monoxide (CO) and ozone (O₃), indicated no statistical significance. However, because all these quantities participate in significant interactions, and the amount of data is sufficient to avoid overfitting (32,859 points factoring missing data), no simplifications to the model were carried out. The two models were validated on one-year worth of data from 2023. The performance of the two models is summarized in

Table 3. Estimated coefficients for the linear regression model applied for AQS Hipodruma.

Model	Number of Terms	Adjusted R2	RMSE Training	RMSE Test
First-order	13	0.862	10.5 µg/m3	8.29 µg/m3
Second-order	91	0.923	7.84 µg/m3	6.20 µg/m3

. Due to its better performance statistics, the second-order model was chosen to predict the concentration of PM_2.5 at AQS Pavlovo. The results of the correlation tests between predicted and measured values for the test data from 2020 are shown in Figure 13. The two points around (150, 10) in Figure 13b correspond to values for CO, NO₂, SO₂, PM₁₀, and wind speed and do not seem to correspond to such PM_2.5 concentrations in the training set.

Several points need to be mentioned about the model development. First, the approach assumes that the relationship between PM_2.5 and other measured quantities will remain the same at traffic stations. There is no empirical evidence to support or refute such an assumption. Second, in contrast to the NO₂ work described in Section 3.2, no thorough model selection investigation was conducted. Third, similar to the NO₂ study, no error estimation has been included in the modeling process. All of these points are addressed in the outlook for future work.

Once the model was tested, it was used in the calculation of $\beta$ in Equation (5). The predicted hourly values for PM_2.5 were split into years and subtracted from measured PM₁₀ concentrations to determine the concentration of PM_coarse at the traffic station. A zero-intercept line was fit to the PM₁₀ versus PM_coarse data, whose slope is the $\beta$ coefficient. The values for each year are shown in

Table 4. Proportional factors $\beta$ between $P{M}_{10}$ and $P{M}_{coarse}$ for the period 2015–2020 at a traffic and a background station.

Year	${\mathit{\beta }}_{\mathit{t}\mathit{r}\mathit{a}\mathit{f}\mathit{f}\mathit{i}\mathit{c}}$	${\mathit{\beta }}_{\mathit{b}\mathit{a}\mathit{c}\mathit{k}\mathit{g}\mathit{r}\mathit{o}\mathit{u}\mathit{n}\mathit{d}}$
2015	0.34	0.38
2016	0.35	0.46
2017	0.29	0.33
2018	0.32	0.33
2019	0.36	0.45
2020	0.38	0.47

.

3.2. Results from Modeling

3.2.1. Transport Emission Inventory Modeling

EMIT [17] is a database tool for storing, manipulating and assessing emissions data from a variety of sources. It can hold data from explicit sources such as major roads, rail and industrial plants. In addition, EMIT can hold data from sources that may be too small to be considered explicitly, and instead, they are treated as average emissions on a regular grid. Source data held in this way are from minor road and commercial and domestic sources. EMIT calculates emissions from source activity data (such as traffic flow, speed and source length for roads, fuel consumption for industrial sources) using up-to-date emission factors. Calculated with EMIT emissions of PM₁₀, PM_2.5, NO_x, and NO₂ using described in Section 2.3.1, data are presented in Figure 14.

3.2.2. ADMS-Urban Model Validation

The model validation is based on comparison of 24 h averaged modeled concentrations of NO₂ and PM₁₀ with observations. Measurements for the NO₂ concentration at AQS Mladost (42.656′ N, 23.383′ E at 582 m ASL) are used. A comparison of modeled PM₁₀ concentration with observations is performed at the CMO site (42.655′ N, 23.384′ E at 586 m ASL)—the same location where the PM samples were collected and processed using the PMF method (described in Section 2.4.3)—to assess the contribution of traffic emissions. The CMO site is located approximately 200 m southwest of AQS Mladost and 170 m away from the road, while the AQS location is approximately 70 m away from the road, which leads to a small difference between PM₁₀ concentrations measured at both sites.

The output from the EMIT model was used to calculate the concentration fields without applying the methods described in Section 2.4.1 and Section 2.4.2, which were subsequently used to adjust the emissions of NO₂, PM₁₀ and PM_coarse. To investigate the contribution of categories related to traffic emissions concentration and to validate the model, two periods without residential heating and Saharan dust intrusion (18–23 April and 20–26 September 2019) were selected. The meteorological conditions for these periods are described in Section 2.2.2.

Model Validation for NO₂

The results from two model runs are presented and compared. The NO₂ emissions from EMIT are calculated based on the third-degree polynomial developed by Derwent and Middleton [58], referred to as the “original” in this study. As described in Section 2.4.1, a new polynomial was obtained (referred to as “adjusted”), derived from measured concentrations of NO₂ and NO_x at AQS Mladost, to achieve a more representative relationship between the two fractions. Commonly used statistic measures—mean bias (MB), mean error (ME), normalized (with respect to the measured value) mean bias (NMB), normalized (with respect to the measured value) mean error (NME) and RMSE—are used to estimate the model performance (for definitions of the measures, see Supplementary Materials S2). A comparison of both methods with results from the model output, applying original and adjusted polynomials, is shown in

Table 5. ADMS-Urban model validation for NO₂ concentration comparing results obtained with original and adjusted polynomial function. All units are µg/m³.

Date	AQS Mladost (Measured)	ADMS- Urban (Original)	ADMS- Urban (Adjusted)	Bias (Original)	Bias (Adjusted)	Error (Original)	Error (Adjusted)
18 April2019	14.89	20.81	17.06	5.91	2.17	5.91	2.17
19 April 2019	13.81	32.24	10.49	18.43	−3.32	18.43	3.32
20 April 2019	12.42	40.08	5.26	27.66	−7.15	27.66	7.15
21 April 2019	8.01	26.66	4.40	18.65	−3.61	18.65	3.61
22 April 2019	7.96	17.04	3.79	9.09	−4.17	9.09	4.17
23 April 2019	8.34	15.09	14.04	6.75	5.70	6.75	5.70
20 September 2019	34.94	9.14	11.17	−25.79	−23.77	25.79	23.77
21 September 2019	30.15	8.46	3.76	−21.69	−26.39	21.69	26.39
22 September 2019	29.82	9.12	10.60	−20.69	−19.22	20.69	19.22
23 September 2019	45.33	19.70	24.13	−25.63	−21.20	25.63	21.20
24 September 2019	43.82	15.72	31.66	−28.10	−12.16	28.10	12.16
25 September 2019	40.89	8.51	18.82	−32.38	−22.07	32.38	22.07
26 September 2019	39.09	16.92	11.42	−22.16	−27.66	22.16	27.66
Average	25.34	18.42	12.81	−6.92	−12.53	20.23	13.74

.

It is difficult to estimate the model performance on a case-by-case basis as the results strongly depend on meteorological conditions. In general, smaller biases and errors are achieved in April in comparison with September. For most days, biases and errors using adjusted polynomial are smaller, except for two days: 21 and 26 September. When using the custom polynomial, RMSE is reduced from 21.78 µg/m³ to 16.67 µg/m³, ME from 20.23 µg/m³ to 13.74 µg/m³ and NME from 80% to 54%.

Although transport is the main contributor to NO_x concentration, other sources such as power plants, industrial processes, and the combustion of fossil fuels, wood, and other organic matter can affect the measured concentration at AQS Mladost, depending on the direction of the mean wind flow and the stability of the atmosphere. It must also be noted that the statistical measures were calculated based on a small sample size, which introduces considerable uncertainty.

Model Validation for PM₁₀

The same period was used for validating the model predictions for PM₁₀ concentrations. Observed concentrations of PM₁₀ and those calculated with the PMF model (described in Section 2.4.3) are shown in

Table 6. Comparison of obtained with PMF model (PMF) and measured (meas) concentrations, and contribution of different categories * to the total PM₁₀ (PMF) concentration. All units are µg/m³.

Date	PM10 (PMF)	PM10 (meas)	RES	TR	BB	IND	FUEL	N	SO4	SEC
18 April 2019	20.38	15.23	5.14	3.28	6.49	0.05	0.25	0.00	4.93	0.37
19 April 2019	21.41	17.58	4.91	3.87	4.67	0.11	0.24	0.00	2.80	4.84
20 April 2019	18.44	18.97	3.54	2.31	6.96	0.00	0.24	0.00	0.71	4.99
21 April 2019	19.59	20.66	6.26	2.32	4.22	0.68	0.19	0.00	0.00	7.08
22 April 2019	27.13	24.16	6.83	3.57	5.16	1.44	0.31	0.00	3.07	6.79
23 April 2019	28.63	27.55	3.34	3.31	9.40	2.18	0.22	0.17	6.87	3.13
20 September 2019	19.17	21.82	6.94	2.08	1.41	0.08	0.22	0.30	2.79	5.34
21 September 2019	19.20	24.06	5.93	2.68	4.09	0.15	0.16	0.07	6.08	0.04
22 September 2019	25.73	28.00	4.17	2.32	5.34	2.58	0.21	0.02	9.95	1.15
23 September 2019	29.91	32.42	7.84	4.84	4.04	1.09	0.03	0.28	9.57	2.20
24 September 2019	27.73	33.88	2.80	3.84	7.74	2.27	0.02	0.00	6.18	5.12
25 September 2019	29.73	30.79	2.90	3.28	2.80	2.00	0.05	0.00	12.08	6.74
26 September 2019	19.67	22.54	2.82	1.26	0.00	0.43	0.14	0.12	10.15	6.13

* Resuspension (RES), Secondary (SEC), Biomass burning (BB), Traffic (TR), Industry (IND), Nitrate-rich (N), Fuel oil burning (FUEL), and Mixed SO₄²⁻ (SO₄).

, together with the contribution of different categories to the total PM₁₀ concentration (based on data obtained with the PMF model).

All categories were selected based on the chemical composition of PM10 samples, and only RES and TR relate to transport. As the main purpose of this study is to investigate the contribution of transport to air pollution and to provide adjusted emission factors, representative of the conditions in Bulgaria, sources not related to traffic (SEC, BB, IND, N, FUEL and SO₄) were not considered.

Numerical results from ADMS-Urban were compared against the sum of the non-exhaust and exhaust traffic sources (RES and TR) for PM₁₀. Model results were extracted at the exact location of CMO, where measurements were collected. Two different runs were performed, one with PM emissions calculated with the original EMIT emission factors and another one using ${\mathrm{E}}_{{\mathrm{PM}}_{10}}$ and ${\mathrm{E}}_{{\mathrm{PM}}_{\mathrm{coarse}}}$ calculated based on the method described in Section 2.4.2. The 24 h average concentrations, as well as an average for the entire period calculated with both models, are shown in

Table 7. ADMS-Urban model validation for PM₁₀ related to transport comparing results obtained with original adjusted emission factors and PMF model at CMO site. All units are µg/m³.

Date	PMF Model at CMO	ADMS- Urban (Original)	ADMS- Urban (Adjusted)	Bias (Original)	Bias (Adjusted)	Error (Original)	Error (Adjusted)
18 April2019	8.42	8.86	9.04	0.43	0.62	0.43	0.62
19 April 2019	8.78	8.13	8.24	−0.65	−0.55	0.65	0.55
20 April 2019	5.85	8.01	8.06	2.16	2.22	2.16	2.22
21 April 2019	8.57	5.83	5.87	−2.75	−2.70	2.75	2.70
22 April 2019	10.40	5.85	5.89	−4.55	−4.51	4.55	4.51
23 April 2019	6.65	6.56	6.63	−0.09	−0.02	0.09	0.02
20 September 2019	9.02	4.97	5.20	−4.05	−3.82	4.05	3.82
21 September 2019	8.61	5.37	5.28	−3.24	−3.33	3.24	3.33
22 September 2019	6.49	7.92	7.87	1.43	1.38	1.43	1.38
23 September 2019	12.68	6.44	6.19	−6.24	−6.49	6.24	6.49
24 September 2019	6.64	7.51	7.35	0.87	0.71	0.87	0.71
25 September 2019	6.18	6.59	6.42	0.42	0.24	0.42	0.24
26 September 2019	4.07	6.74	6.72	2.67	2.65	2.67	2.65
Average	7.87	6.83	6.83	−1.05	−1.05	2.27	2.25

. For the models, the same statistical measures and nomenclature as those employed for the validation of NO₂ are used. There is no significant difference between the results obtained with both runs. The models slightly underpredict PM₁₀ concentration, with MB −1.05 µg/m³ and a positive bias for some days. The ME is 2.27 µg/m³ for the calculation with the original emissions map and 2.25 µg/m³ with the adjusted one. The largest error is for 23 September, following 22 April and 20 and 21 September. The other measures also indicate a similar performance, with an RMSE of 2.92 µg/m³, NMB of −13.30% and NME of 28.55% for the simulation with adjusted emissions, and an RMSE of 2.90 µg/m³, NMB of −13.28% and NME of 28.86% for the simulation with original emissions.

The coefficient calculated for 2019 (see Section 3.1.2) was used to estimate ${\mathrm{E}}_{{\mathrm{PM}}_{\mathrm{coarse}}}$ emissions for numerical calculations. The coefficient varies significantly from year to year as many factors (meteorological conditions, traffic flow, and anthropogenic activity, among others) affect air pollution. The emission inventory used in the study was developed for 2018, while the meteorological conditions correspond to 2019 as measurements for the model validation were available for that did not change significantly from 2018 to 2019.

3.2.3. Spatial Distribution of NO₂ Concentrations

EMIT provides emission estimations for each segment of primary and secondary roads in the study domain. A comparison of the spatial distribution of NO₂ using original and adjusted emissions and the difference in concentration patterns for concentration fields averaged over the entire period is shown in Figure 15. The emissions calculated with the original map provide lower concentrations over the primary roads and higher concentrations over the secondary roads. The adjusted emissions lead to an increase in the NO₂ fraction in close proximity to the roads and a reduced concentration within one hundred meters along the roads with intensive traffic, which is in agreement with experimental observations. The difference in simulated concentrations using original and adjusted emissions is up to 20 µg/m³ over the road and 10 µg/m³ far away from it.

Maps with a 24 h average difference between model results obtained using original and adjusted emissions are shown in Supplementary Materials S3. The same pattern as described for the mean field is typical for all days, except for the 23 and the 24 September, when there is no difference over the main road in the area: Tsarigradsko Shose. As can be seen from Table 5, the maximum concentrations for the period were measured at AQS Mladost for these days, at 45.33 µg/m³ and 43.82 µg/m³, respectively. Simulations using adjusted emissions also provide the highest concentrations for the same dates, but there is no correspondence in the 24 h average concentration values obtained using the original emissions. These two days are characterized by very high relative humidity (near saturation) during the afternoon hours and a decrease in incoming solar radiation and temperature. Vertical profiles (see Supplementary Materials S1.) show a strong inversion layer and calm conditions (wind speed less than 1 m/s).

3.2.4. Spatial Distribution of Concentrations of Different PM Fractions

Similar average concentration field maps for PM₁₀, using original and adjusted emissions and the difference in concentration patterns, are shown in Figure 16. Adjusted emissions increase the concentration over the road segments with the most intensive traffic. The difference in simulated concentrations using original and adjusted emissions is up to 5 µg/m³ over the road and is insignificant away from it. Maps of a 24 h average concentration difference between model results obtained using original and adjusted emissions are shown in Supplementary Materials S4. Differences over the road segments with the most intensive traffic are higher for the same days, 23 and 24 September, with a concentration approximately 8 µg/m³ higher for the simulation using adjusted emissions.

The same pattern can be found in the spatial distribution of PM_2.5 using original and adjusted emissions. The concentration patterns for the two models, averaged over the study period, and their difference are shown in Figure 17. Adjusted emissions increase the concentration over the road segments with the most intensive traffic. The difference in simulated concentrations using original and adjusted emissions is up to 5 µg/m³ over the road and insignificant away from it. The main reason is that small particles are associated with exhaust traffic emissions and coarse particles with non-exhaust emissions. Maps with a 24 h averaged difference between model results obtained using original and adjusted emissions are shown in Supplementary Materials S5. Differences over the road segments with the most intensive traffic are higher for the same days, 23 and 24 September, with a concentration approximately 7 µg/m³ higher for the simulation using adjusted emissions.

The concentration maps for PM_coarse differ from the other fractions (Figure 18). This fraction is more widely dispersed when using the original emissions, and the values are higher far from the road. In the simulations with adjusted emissions, the coarse fraction is mainly dispersed near the roads due to the faster deposition of heavier particles. Maps with a 24 h average difference between model results obtained using original and adjusted emissions are shown in Supplementary Materials S6.

4. Discussion

Due to the scarcity of reliable experimental measurements supporting the evaluation of emission factors that represent traffic, roads and driving conditions in Bulgaria, advanced statistical models were developed, taking into account the influence of various factors: geographic and urban morphology, artificial surface and canyons, road infrastructure, demographics and public life, etc. The available data were processed on several different levels in order to achieve an optimal representation of traffic conditions in Bulgaria. To the authors’ best knowledge, the approach proposed here is being applied for the first time in that particular context.

Sofia was selected for this pilot study for the following reasons:

• The city of Sofia is the biggest urban area in the country, and despite the efforts made during the last few decades, the citizens in the capital are still exposed to high levels of PM₁₀ as well as other pollutants, especially NO₂. The latter is mostly related to traffic, but the evidence has somehow stayed hidden due to gaps in the monitoring of areas with more intensive traffic [26].
• Sofia as a study object is a challenging and complex urban system because of its geographical setting and due to the fast expansion of the city and the recent tendencies in compact and car-oriented urban development. Unfavorable in terms of air quality are the rising density and share of impervious surfaces, urban street canyon formation, and the interruption of “green wedges”, which have an important role in the ventilation of the city. In the previous two decades, investment priority was given to the construction of new transport infrastructure for untapping some “bottlenecks,” but this has induced demand, encouraging higher motorization and more intensive travel by private cars while increasing competition for the narrow space with the alternative mass or active mobility options [68,69,70].
• Despite the relative scarcity of reliable observational data, Sofia is well ahead in terms of experimental infrastructure in comparison to other cities in Bulgaria. Matching the different sources of data and information is a challenging endeavor in the pursuit of better knowledge, which can contribute to more appropriate decision-making.

Our previous work shows many uncertainties related to local-scale modeling, and the transport emission inventory has the highest ambiguity in air quality research [39]. The working hypothesis was that modeling at a local scale can be improved by implementing a more reliable inventory of traffic emissions if the relationships used to develop the emission factors are derived using local data more representative of Bulgarian conditions.

Our results support this hypothesis, showing an improvement in simulated concentrations for the main pollutants after model validation. Estimating NO₂ concentration with a customized polynomial, based on local measurements, shows a significant improvement with increasing near-road concentration, most notably during stable conditions (temperature inversion). As already mentioned, there are many uncertainties related to TRAP modeling, especially for PM concentration due to non-exhaust emissions. This was the reason for developing PM₁₀ and PM_coarse emissions following an established methodology but using local observations to better represent real Bulgarian road conditions and traffic structures. Emissions for PM_2.5 are obtained by subtracting PM₁₀ and PM_coarse emissions. Although model validation did not show a significant difference at the point of comparison, a significant difference was recorded above and near the major roads between the simulations using the original and adjusted emissions. For the average PM₁₀ concentration field, there was an approximately 20% increase in value over the road with the most intensive traffic, and on days with a strong inversion (e.g., 23–25 September 2019), the difference reached 8 µg/m³. Adjusted emissions show a small contribution of PM_coarse concentrations away from the road, mainly due to the higher deposition velocity of this fraction, but there was a significant increase in PM_2.5 concentration of approximately 30% for the average field and 7 µg/m³ on days with stable conditions.

These results raise a very important question about the missing observational data for PM_2.5 in Sofia and other urban areas. This pollutant is measured only in one urban background site in Sofia. The situation is the same with other Bulgarian cities. The concentration of PM_2.5 is closely related to health risk and mortality mainly from cardiovascular diseases (including ischemic heart diseases and strokes) [71]. Bulgaria faces severe air pollution, with particulate matter levels being amongst the highest in the EU [72]. However, the lack of measurements for the concentration of PM_2.5, and therefore the inability to validate numerical model predictions for this pollutant, make health risk assessment difficult and uncertain.

Furthermore, the EU Commission proposed to revise the Ambient Air Quality Directives as part of the European Green Deal. The revision aligns the air quality standards more closely with the recommendations of the World Health Organization (WHO). According to the WHO’s Air Quality Guidelines [73], the global burden of disease associated with air pollution exposure takes a massive toll on human health worldwide: exposure to air pollution is estimated to cause millions of deaths and lost years of life annually. The recommended values are for significant reduction in annual and 24 h air quality levels for PM_2.5, PM₁₀ and NO_2, pollutants closely related to TRAP emissions. A higher share of population in larger cities, which is a stable trend in Bulgaria as well, is closely related to the increase in the number of cars and human activities. Poor road quality and conditions and the large number of ageing vehicles in use contribute to the systematic exceedance of EU standards and the subsequent financial sanctions for the country.

Another very important issue is the location of the so-called traffic air quality stations. On paper, there are only four of these stations in Bulgaria—two in Sofia, one in the city of Plovdiv and one in the city of Stara Zagora. Despite it being designated as one, AQS Mladost does not conform to the EU requirements for positioning of traffic stations (e.g., traffic-oriented sampling probes have to be no more than 10 m from the kerbside) [74], and consequently, it is not representative of such a device. The lack of reliable measurements near the roads leads to severe biases in the data and the impossibility of properly validating any transport model. More precise measurements with high temporal resolution at true traffic sites are needed to assess the real traffic impacts; volume and structure; TRAP; and associated health and comfort effects for the population.

5. Conclusions

The aim of the present study was to develop a high-quality inventory of traffic emissions using an established methodology but applying local observations, in order to adjust emissions for PM as well as the rates between NO₂ and NO_x and between PM_2.5 and PM₁₀, to better represent traffic conditions in Bulgaria.

The main outcomes from the conducted study are as follows:

• Descriptive statistics based on observations at three transport stations in Sofia over an extended period (2009–2020) show a trend in the average annual concentrations of NO₂ and PM_10. A slightly negative trend (decrease in averaged annual concentration) was registered at AQS Pavlovo. The highest values were measured at AQS Orlov Most, but these decreased significantly after the station was moved to a new location in Mladost. The lowest recorded concentrations in 2019 and 2020 are most likely due to the imposed lockdown during the COVID-19 pandemic. Hourly concentrations of NO₂ (averaged for the entire period) for weekdays, Saturday, and Sunday are also presented as evidence of the diurnal profiles of traffic-related emissions.
• A new inventory of traffic emissions has been developed for the city of Sofia based on publicly available traffic and fleet data, as well as a dedicated data collection campaign to fill the gaps in secondary street traffic data. Various processing steps were applied to align diverse geometry, attribute data, and acquisition methods, utilizing the advantages of ensemble learning. This activity data were used with the EMIT model to calculate the traffic emissions. The emissions were then exploited to simulate air pollution in a specific area by the ADMS-Urban model.
• The polynomial relationship between NO_x and NO₂ was adjusted based on local dynamics at three AQSs in Sofia for a 12-year period (from 2009 to 2020), to provide a more adequate estimation of NO₂ from the ADMS-Urban model.
• Models of the first and second order with all interactions were fit to data for 12 measurands at AQS Hipodruma (where PM_2.5 observations are available) for the period from 2015 to 2020. The two models were validated on one-year worth of data from 2023. Due to its better performance statistics, the second-order model was chosen to predict the concentration of PM_2.5 at the chosen traffic station, AQS Pavlovo. Data from both transport and urban background stations were used to calculate the roadside increments and the adjusted emissions of PM₁₀ and PM_coarse.
• The ADMS-Urban model was validated and evaluated by comparing pollutant concentrations from simulations using original and adjusted emissions, showing an improvement in results after applying functions and relationships derived from local observations.

This work is part of our effort to improve modeling at the local scale and part of a framework methodology to support the management and planning of healthy urban environments and lifestyles. The methodology was developed for the city of Sofia as a pilot study but will be applicable to any other urban area in Bulgaria.

There are a number of points in the present study that will be addressed in the future. The presented work can also benefit from several extensions in the statistical and data processing parts. The first and perhaps most challenging deficiency to overcome is the lack of any sort of uncertainty in the data used in setting up the physics-based models and training the various data maps. This includes measurement and replication uncertainty, model form error, and statistical model uncertainty. Close collaboration with AQS manufacturers, operators, and service personnel as well as careful data collection and formal validation procedures can help quantify and potentially reduce all this uncertainty. Another point of improvement is the validation or elimination of some key assumptions, chiefly the one about mapping PM_2.5 from background to traffic stations. In the authors’ opinion, this can only be satisfactorily achieved by equipping the AQSs with relevant sensors and eliminating statistical modeling altogether. Sofia Municipality operates a set of equivalent sensors under the AirThings initiative [28], which could be used as an interim measure to at least validate some of these assumptions.