This research utilized the AERMOD dispersion model system, with PM_2.5 observations used as model inputs and for model performance testing. Figure 1 provides an overview of the modeling system and relevant input data. The study focused on an area 2.5 km by 2.5 km centered at Chulalongkorn University, Bangkok, Thailand. Figure 2 shows the location of the modeling domain, which lies mostly in the Pathumwan District. The reasonings for site selection are based on the fact that the area is highly commercialized with high traffic activity yet presents mixed land use of educational institutes, hospitals, temples, residential buildings, and recreational space. So there exists a clear motivation to develop a tool that can assess air pollution health risk and support air quality management.

An emissions inventory, classified by vehicle type, was developed at 1-h time resolution for primary PM_2.5 from road vehicle exhaust for each road segment located in the study area for the period of 2018–2020. The road segments are meant for representing parts of the road with different traffic volumes. The multiyear period provides a larger set of traffic volume observations to represent traffic inhomogeneities. The emissions rate ER _i for each vehicle type i for a given road section was calculated as. E R i   g   P M 2.5 h r = A D i   ( v e h i c l e s ∙ k m   t r a v e l e d h r ) ∗ E F i g   P M 2.5 v e h i c l e s ∙ k m   t r a v e l e d where AD _i represents the traffic activity data for vehicle type i and EF _i represents the emissions factor for vehicle type i. Traffic activity data comprises traffic volume per time and an average traveled distance. For a given road segment, we assume the average travel distance equals the length of the road segment. The total PM_2.5 emissions for a given road section is the summation of emissions rates across all vehicle types i = 1…n, where n is the total number of vehicle types. We categorized four vehicle types, namely, personal cars (PC), light-duty vehicles (LD) including vans and pick-up trucks, heavy-duty vehicles (HD) including buses and large trucks, and other vehicles (OT) including motorized three-wheelers. The details of traffic activity data, including the data sources and treatment of hourly variability, are provided in the Supplementary Material. The PM_2.5 EFs categorized by vehicle type were those from the Atmospheric Brown Cloud emissions inventory manual (Shrestha et al., 2013). For PC, LD, and HD, the statistics for the case of good control-EURO I and EURO II were used. The EF used for PC was that for gasoline vehicles because 64.2% of passenger cars in Bangkok use gasoline fuel (Department of Land Transport, 2020). For other types (OT), the EF used was that for motorcycles (4-stroke with control). The EFs were 0.007, 0.15, 0.72, and 0.05 g km^–1 for PC, LD, HD, and OT, respectively. The emissions rates for the four vehicle types are tabulated for each road section in the Supplementary Material.

The American Meteorological Society (AMS)/U.S. Environmental Protection Agency (EPA) Regulatory Model (AERMOD) is a steady-state plume dispersion model that accounts for planetary boundary layer turbulence structure and is capable of handling both surface and elevated sources over both simple and complex terrain (Cimorelli et al., 2004). We used AERMOD View™ software, version 10.0.1, to conduct our simulations. The pollutant type was PM_2.5. The simulation was limited to primary particulate matter.

The sources in the model were configured as line area sources because the lanes of emitting vehicles constitute a continuous line source confined in the area within the width of the road. The size of each line area source was determined by the width of the road, ranging from 10 m to 45.5 m for 2-lane to 6-lane roads, respectively. Only the main roads were designated as emissions sources. We ignored vehicle emissions from alleys and minor streets due to lack of traffic data. We assumed that these roads were far less busy than the main roads, carrying insignificant hourly average traffic volumes. The assigned emissions rates were calculated as described in the previous section. For each road segment, we set up four individual sources for the four emitting vehicle types. Each source was assigned a mean emissions rate and hourly variability factors. Ground-level Road segments were assigned their actual mean ground elevations, while expressway segments were assigned an elevation of 15 m above the actual mean ground elevation below the segment.

As for modeled processes, the model computed not only transport and dispersion but also deposition loss. The deposition calculation followed the theory described in Cimorelli et al. (2004). The particle size distribution used for emissions had a 0.99 fine particle fraction and a mean mass diameter of 0.284 microns, as calculated from the aerosol number size distribution measurement data from an urban site by Schneider et al. (2015). That site’s size distribution was selected due to its similarity to the traffic-dominated environment of the Bangkok CBD.

The discrete receptors in the model were set at 4 monitoring points, namely, Chulalongkorn Hospital, the Chulapat 14 Building, the Chulalongkorn University (CU) dormitory canteen, and Chulalongkorn Demonstration Primary School Gate 5, which we will refer to as Chula Demonstration School. In addition, grid receptors were set at a uniform spacing of approximately 130 m along each road segment over the entire modeling domain.

The main input data for the AERMOD simulation included meteorological data and terrain data. To account for PM_2.5 originating from outside the modeling domain, we input advected-in PM_2.5 concentrations. Descriptions for all data elements are provided below.

Meteorological conditions were calculated from the Weather Research and Forecasting Model (WRF) model. The meteorological data files, both for surface data and upper air data, were generated by the U.S. EPA’s Mesoscale Model Interface Program (MMIF) at the coordinate 13.73944 °N 100.5237 °E for the period of 2020. The data files were processed by the U.S. EPA’s AERMET meteorological pre-processor software, AERMET View™, version 10.2.1. The surface wind speeds and directions (blowing from) are summarized as wind rose charts in Supplementary Figure S2 in the Supplementary Material. The winds have seasonal patterns as shown for the summer (March–June), the rainy season (July–October), and the winter (November–February).

For the terrain data, we utilized the Shuttle Radar Topography Mission SRTM1 products at a resolution of 1 arc-second (30 m) with global coverage. The data were obtained via a terrain processor and downloaded through the AERMOD View program. The modeled area, in the city center, had low-lying and relatively flat terrain, with most of the area at elevations of 10–20 m and the whole area ranging from 69 m above mean sea level (MSL) to a few meters below MSL.

Since the actual PM_2.5 levels in the modeling domain were influenced by both local emissions and those emitted elsewhere beyond the domain boundary, we augmented the predicted PM_2.5 levels with advected-in PM_2.5. Advected-in PM_2.5 concentrations, which are referred to in the AERMOD View program as background concentrations, were introduced to the model by referring to the hourly PM_2.5 measurements at upwind locations according to the specific wind direction at each hour. The methodology for advected-in PM_2.5 is described in the Supplementary Material.

The statistical indices chosen for model testing included fractional bias (FB), correlation coefficient (R), fraction within a factor of two (FAC2), and normalized mean square error (NMSE). These indices followed the recommendations of the U.S. EPA’s “Guideline on Air Quality Models” (EPA, 2017). Past model studies have discussed the relevance and meaning of these statistical indices (Chang and Hanna, 2004; Gibson et al., 2013).

For model performance evaluation, the required PM_2.5 measurement data in the study area for 2020 were downloaded from www.CUsense.net, which provides collections of PM_2.5 data measured by various agencies. Measurement at the Chulalongkorn Hospital monitoring station (station ID: 50t), conducted by the Pollution Control Department, followed the official federal equivalent method (Pollution Control Department, Ministry of Natural Resource, 2010). The other three monitoring points in the Chulalongkorn University area used light-scattering sensors to measure PM_2.5 (Chunitiphisan et al., 2017; 2018). Table 1 summarizes the locations of the 4 monitoring stations.

To communicate the highest acute health risk of PM_2.5 (Brook et al., 2010), the model results are presented as the maximum PM_2.5 concentration found for each receptor point. Figure 3 shows the highest 24-h average PM_2.5 concentration for each point in the modeling domain for the year 2020. The spatial distribution of PM_2.5 in the modeling domain was very similar regardless of the time-averaging period (24-h and 1-h) or the percentile threshold. Maximum values and annual averages for the year 2020 are summarized in Table 2. The values in the domain are presented for various points among the grid receptors, including the point with the highest annual average value, a point with the mean annual average value, and the point with the lowest annual average value. The receptors of interest include the hotspot found at Phong Rama Intersection (13.746559° N, 100.520794° E) and the four monitoring points described in Table 1.

The simulation revealed new information regarding the peak PM_2.5 point at Phong Rama Intersection, where PM_2.5 had never before been measured. The busy intersection and the expressway above this location contribute to the consistently high concentration, making it the peak PM_2.5 point for all time-averaging periods. As for the remaining observation points, the Chulalongkorn Hospital observation point was located by a roadside at a busy intersection and thus exhibited higher concentrations than the other three points located in the Chulalongkorn University campus block, farther from the main public roads.

Table 2 also provides the 1-h average values in order to convey the severity of the PM_2.5 situation at particular hours, while the 24-h average PM_2.5 concentration is the index to be compared with the ambient standard. The value at the 99th percentile was chosen for the summary in accordance with the new World Health Organization air quality guideline for 24-h average PM_2.5 concentration. The maximum 1-h average PM_2.5 concentration (100th percentile value) in the domain and at each point of interest occurs in the nighttime between 00:00 and 01:00. The PM_2.5 levels at the 100th percentile are significantly higher than the 99th percentile values, indicating the significant role of nighttime meteorology in creating unfavorable stagnant conditions. In the context of PM_2.5 regulations, the highest PM_2.5 concentration from the model agrees with the actual situation that Bangkok faces, with the 24-h average PM_2.5 concentration exceeding the Thailand national ambient standards of 50 μg m⁻³ (as of 2020) and 37.5 μg m⁻³ (scheduled to begin in May 2023). As for the annual average PM_2.5 concentration, comparing even the minimum point in the domain to the ambient standards of 25 μg m⁻³ (as of 2020) and 15 μg m⁻³ (enacted in July 2022), shows that the permitted standards are exceeded at all points in the domain.

Table 3 provides not only the maximum values but also the domain–average PM_2.5 statistics for year 2020, including the mean, standard deviation (SD), minimum, 25th percentile, 50th percentile, 75th percentile, and maximum. The summary is calculated for the whole year as well as for each of the three seasons in order to highlight the seasonality of PM_2.5 levels in Bangkok. Seasons are categorized as Summer (March–June), Rainy (July–October), and Winter (November–February). Comparing our results with those of PM_2.5 models published in the literature, Peng-in et al. (2022) presented a PM_2.5 model based on satellite-retrieved AOD for the Bangkok Metropolitan Region (BMR) with comparable statistics for the whole year. Their modeled domain–average PM_2.5 values were lower than ours, probably because our domain lies within the CBD, which experiences intense traffic activity. In terms of seasonal statistics, our domain–average values seem to be higher and more spread than those of Peng-in et al. (2022) in the peak PM_2.5 season (winter) but lower and less spread than those of Peng-in et al. (2022) in the low-PM_2.5 season (rainy season). We further explore these patterns via comparisons with observations in our domain in Section 3.3.

With the ability to calculate PM_2.5 resulting from individual source groups as well as from all sources at once, we conducted the source contribution analysis by investigating the various road types and vehicle types. Source group is a feature in AERMOD View that lets user defines grouping of multiple sources. In the first source contribution analysis, the contributions from ground-level roads, expressways, and advection were segregated. We analyzed the source contribution information for the 2020 annual average PM_2.5 for each of the 5 points of interest as shown in Figure 4A ). Higher traffic volume on the expressway yielded higher emissions rates than on at-grade roads, with the expressway contributing a notably large contribution of 32.11 percent. The Chulalongkorn Hospital point showed a higher contribution from at-grade roads (17.49 percent) because of its proximity to busy at-grade roads. At the three points located in the Chulalongkorn University campus area, the main contribution came overwhelmingly from advected-in PM_2.5 (∼99 percent), while the nearest main roads, situated 35–180 m away, contributed only 0.18–0.29 μg m⁻³ in total.

Figure 4B ) provides a contribution analysis for various vehicle types as emissions sources. This analysis provides further insight to which vehicle types dominate the local PM_2.5 emissions. HD showed the highest contribution at the Phong Rama Intersection because of the significant volume of heavy-duty trucks on the expressway. For the remaining points, LD showed a slightly higher contribution than HD because of the high LD traffic volume on nearby roads (averaging 18% of total traffic volume) compared to that of HD (3% on average). It is worth pointing out that the EF for HD was much higher than other vehicle types: 5 times that of LD, 14 times that of OT, and 103 times that of PC.

Here we test the model by comparing the simulation results with PM_2.5 observations as described in Section 2.5. The analysis is presented for the year 2020. Along with the statistical indices in Table 4, we present the comparison in a time-series plot, a scatter plot, and a quantile–quantile plot (Q–Q plot). We provide these visualizations for two representative receptor points, namely, Chulalongkorn Hospital and the Chulapat 14 Building as shown in Figure 5. The statistics for the 24-h average observations and predicted values are presented as a whisker box plot in Figure 6, with the data presented for the whole year of 2020 as well as separated by season. Here the median is marked by the middle line in the box, while the mean is marked by a cross sign. The bottom and the top of the box are defined by the P25* and P75* values, respectively, which are calculated from the 25th percentile (P25) and the 75th percentile (P75). Let us define P75–P25 as the interquartile range (IQR). P25* is then the larger of two values: P25 − 1.5(IQR) and the minimum value over all data. Similarly, P75* is the smaller value of two values: P75 + 1.5(IQR) and the maximum value over all data. The two ends of the whisker drawn as straight lines are the overall minimum and maximum. Solid dots represent outliers.

The results for the Chulalongkorn Hospital indicate better model accuracy than those for the Chulapat 14 Building and the other two nearby sites. Temporal fluctuation in PM_2.5 concentration is generally captured well in the time-series plots, with some tendency of overprediction especially in the rainy season. The whisker box plot also highlights the particularly poor modeling of the PM_2.5 concentration for the rainy season, with boxes for observed and modeled data hardly overlapping. Nevertheless, the overall correlation is considered acceptable given the R values ranging from 0.71 to 0.94 for different receptor points. Model overprediction is confirmed by the FB values ranging from −0.22 to −0.10. The Q–Q plots show that the predicted PM_2.5 concentrations have a distribution of values resembling that of the observed PM_2.5 concentrations; thus, the data points lie along a relatively straight trendline. However, the slope of the line shows a value slightly greater than unity, reflecting the overprediction of modeled PM_2.5 concentration. The acceptability of model accuracy is indicated by the scatter plots, which show data points clustering along the 1:1 line. The points lie well within the area bounded by the 2:1 line and the 1:2 line, as indicated by FAC2 values of 0.99 for Chulalongkorn Hospital and 0.78 for the Chulapat 14 Building. As presented in Table 4, the Chulalongkorn Hospital site showed the lowest NMSE value (0.08) among the points of interest, indicating the highest model accuracy. We interpret this enhanced accuracy as being due to the proximity of this site to a main road for which the emissions were well documented. The three sites in the Chulalongkorn University campus area showed poorer accuracy, especially when observed PM_2.5 concentrations fell below 20 μg m⁻³. Because these sites were located farther from emissions sources, they received hardly any contribution from the dispersed plumes from the nearest main road sources. Therefore, at low observed PM_2.5 values, PM_2.5 levels were instead dominated by the advected-in contribution. We suspect that this model–observation disparity may be due in part to the model’s assumptions about advected-in PM_2.5.

Another commonly reported PM_2.5 pollution statistic is the number of days in each year that exceed the 24-h average ambient standard. Table 5 provides a comparison between the modeled and observed exceedance counts for the year 2020 for each of the four monitoring points as well as at the hotspot location. The modeled counts are higher than the observed counts at all sites, in line with the overprediction discussed above. The whisker box plot in Figure 6 also highlights this issue, showing several high outliers for modeled winter values in particular.

In order to pave the way for future research, we now explore the model limitations encountered in utilizing the AERMOD model to produce satisfying model predictions with acceptable accuracy. Given the lack of traffic activity information with continuous time coverage, we made several assumptions based on a 1-week sample of traffic activity observed in the modeling domain as well as a monthly traffic dataset from a past study. Actual traffic may not exhibit the same hourly pattern across different roads, but modeling these differences remains an aspect for future development. Also missing from the traffic data is the additional count of motorcycles, which can be rather substantial in the domain. These known limitations of the model would be expected to result in emissions inventory underestimation. Another category of emissions that the model did not account for was non-exhaust emissions, including road surface wear, road vehicle tire wear, and brake wear (EEA, 2019). To our knowledge, there has not been a local study of the EFs of these non-exhaust emissions. We decided to ignore them because road conditions, environmental conditions, and vehicle conditions in Bangkok may differ drastically from the studied conditions in European countries, lead to high uncertainty when applying the EMEP/EEA EFs. However, these non-exhaust emissions, may potentially play a significant role. A modeling study for the city of Banská Bystrica, Slovakia, by Salva et al. (2021) showed that including non-exhaust PM_2.5 emissions resulted in a total emissions rate up to 2.5 times that of exhaust emissions alone. Another limitation is the lack of chemical production calculation, causing the secondary formation of PM_2.5 to go unaccounted for. However, this omission likely did not significantly affect the predicted levels. A previous source apportionment study in the BMR suggested that secondary aerosol formation may contribute ∼0.1–1 percent of ambient concentrations (Chuersuwan et al., 2008). However, recent study suggested that secondary organic carbon may contribute up to approximately 60 percent of PM_2.5 (Rattanapotanan et al., 2023). Another source sector missing from the calculation is cooking which may be of importance at certain hours because there are numerous restaurants in the study area. There has never been a study to account for such emission rates to our knowledge.

Regarding the source contribution analysis, we found the dominant influence in the model was advected-in PM_2.5. Our assumption of a two-box model (described in the Supplementary Material) and referencing of upwind roadside monitoring data might have introduced an excessive PM_2.5 level that led to an overall overestimation by the model despite the several factors discussed above that contribute toward underestimation. The burden of adding advected-in PM_2.5 to achieve good agreement with observed PM_2.5 levels is a trade-off for using a dispersion model such as AERMOD, which is relatively easier to implement than a regional chemical transport model. Future work may explore the ways in which including non-exhaust emissions might lower the influence of advected-in PM_2.5 on the model.

To investigate the contributions of local sources, we compare our findings with those of a previous study in a similar setting. In our comparison with Salva et al. (2021)’s AERMOD simulations in a city in Slovakia, we focus on the relative contributions of exhaust emissions from various vehicle types. Salva et al. (2021) assigned the EFs based on fuel type (petrol vs. diesel) and technology (EURO2 to EURO6), while we chose a simpler treatment with a single EF value for each vehicle type due to lack of actual vehicle fleet data. Comparing the emissions factors that we used with those in Salva et al. (2021) for the same vehicle types for the technology (EURO-X) with the highest population, we found that our EFs were higher for all vehicle types. Our EF for HD was 103 times that for PC, while our AADT_PC:AADT_HD, traffic volume ratio lay in the range of 9–117 with a mean value of 35. Accordingly, the model shows HD emissions surpassing PC emissions for almost all represented road segments in the domain. Salva et al. (2021) found that among the exhaust emissions sources for the study domain, PC provided the largest contribution, while we found that HD and LD provided the largest contributions for the receptor points dominated by expressway and ground-level urban roads, respectively. This difference between study domains is in line with the technological reality of trucks in Thailand: mostly EURO2 or at best EURO4.

Our PM_2.5 model for the Bangkok CBD can provide new insights into PM_2.5 models including previously developed models that used different techniques and tools (Ketjumpol and Sakworawich, 2020; Chalermpong et al., 2021; Kumharn et al., 2022; Peng-In et al., 2022; Thongthammachart et al., 2023). The variability in emissions rates from time-varying traffic data may provide an in-depth tool to investigate the particular incidence of extreme health risks posed by PM_2.5. Such analysis can support the decision to tackle precise sources driving PM_2.5 peaks.

Given the points discussed above, potential research that builds upon these lessons may utilize comprehensive continuous traffic surveillance with the aid of artificial intelligence. Such research may further develop this concept of emissions inventory calculation into a regional chemistry transport model to provide a complete representation of PM_2.5 sources in a wider region not limited to a small geographic domain of interest. Another opportunity to enhance the knowledge of PM_2.5 for Bangkok air quality management would be the inclusion of PM_2.5 sources other than road traffic. We note that these potential research directions emphasize the need for an updated, accurate emissions inventory at high spatiotemporal resolution for effective local air quality management.

We demonstrate the utility of the model as a decision-making support tool by studying PM_2.5 mitigation scenarios from the National Agenda Action Plan to Eradicate Particulate Matter Pollution Problem (Pollution Control Department, 2019). Two PM_2.5 mitigation scenarios from the action plan were chosen, namely, the fuel standard (FS) scenario and the shift mode (SM) scenario. The simulation period is the year 2020. The effects of each mitigation measure are reflected in the emissions rate calculation as described below. Advected-in PM_2.5 concentrations were kept the same as in the base case simulation as described in Section 2 because the result of each policy cannot be effectively estimated in a regional context. With this limitation, we do not anticipate realistic modeled PM_2.5 levels in the studied scenarios. However, even unrealistic model results can provide valuable comparative information in determining which policy results in a better improvement with regard to PM_2.5 levels.

The FS scenario refers to the measure to upgrade emissions standards to EURO6 by the year 2022; however, we note that in reality this upgrade has not been achieved as of January 2023 and also the new EURO6 standard is enforced on new vehicles only. This change was modeled by altering the emissions factors for the four types of vehicles as shown in Table 6. The EURO6 EFs are taken from Kim Oanh (2020). The resulting change in emissions rates on each road segment then depends on the traffic volume classification by vehicle type. In this scenario, we assumed 100 percent replacement of the existing fleets running in the modeling domain by new vehicles meeting the EURO6 emissions standards. The changes in PM_2.5 level (not shown) occur almost entirely as decreases with respect to the base case. The reduction is pronounced along the main road and especially along the expressway alignment because of the high total emissions in the area. The one slight increase in PM_2.5 occurs on a road segment where the traffic data shows a particularly high portion of motorized three-wheelers. The rather counterintuitive EF increase for three-wheelers under the EURO6 standard, categorized as OT, results in this small increase.

The SM scenario refers to a measure that promotes public transportation and lower-emissions modes such as bicycling. The assumption in this scenario is that PC traffic volume is lowered by 70 percent while other vehicle traffic volume is reduced by 50 percent compared to the traffic volume in the base case. The shift away from PC, including taxi usage, as well as away from three-wheeled taxi usage, would translate to higher passenger counts on mass transit electrified rail systems and, possibly, electric buses, thus result in no additional local PM_2.5 exhaust emissions. The reduced traffic volume data fed into the emissions inventory calculation results in lower overall emissions. The PM_2.5 reduction is found to be most significant along the expressway alignment, just as in the FS scenario, again because of the high PC volume on the expressway.

To compare the results of the two PM_2.5 mitigation scenarios, we consider the count of days in each year that exceed the 24-h average ambient standard 50 μg m⁻³ at the points of interest under each scenario and in the base case. As shown in Table 7, the FS scenario proved more effective at lowering PM_2.5 levels below the 50 μg m⁻³ threshold. The results at the three receptor points in the Chulalongkorn University campus area were not affected in either scenario because their dominant contribution was from advected-in PM_2.5; the results are therefore not shown.