Datasets:

---

LLMDH

/

OpenScience

like 0

Follow

LLM - Digital Humanities 16

ERROR: type should be string, got "https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Correspondence to: Christer Johansson (christer.johansson@aces.su.se) \n10 Abstract. As air pollution is regarded as the single largest environmental health risk in Europe it is important that \ncommunication to the public is up-to-date, accurate and provides means to avoid exposure to high air pollution levels. Long- \nas well as short-term exposure to outdoor air pollution is associated with increased risks of mortality and morbidity. Up-to-\ndate information on present and coming days’ air quality help people avoid exposure during episodes with high levels of air \n15 \npollution. Air quality forecasts can be based on deterministic dispersion modelling, but to be accurate this requires detailed \ninformation on future emissions, meteorological conditions and process oriented dispersion modelling. In this paper we apply \ndifferent machine learning (ML) algorithms – Random forest (RF), Extreme Gradient Boosting (XGB) and Long-Short Term \nMemory (LSTM) – to improve 1-, 2- and 3-day deterministic forecasts of PM10, NOx, and O3 at different sites in Greater Abstract. As air pollution is regarded as the single largest environmental health risk in Europe it is important that \ncommunication to the public is up-to-date, accurate and provides means to avoid exposure to high air pollution levels. Long- \nas well as short-term exposure to outdoor air pollution is associated with increased risks of mortality and morbidity. Up-to- as well as short term exposure to outdoor air pollution is associated with increased risks of mortality and morbidity. Up to\ndate information on present and coming days’ air quality help people avoid exposure during episodes with high levels of air \n15 \npollution. Air quality forecasts can be based on deterministic dispersion modelling, but to be accurate this requires detailed \ninformation on future emissions, meteorological conditions and process oriented dispersion modelling. In this paper we apply \ndifferent machine learning (ML) algorithms – Random forest (RF), Extreme Gradient Boosting (XGB) and Long-Short Term \nMemory (LSTM) – to improve 1-, 2- and 3-day deterministic forecasts of PM10, NOx, and O3 at different sites in Greater date information on present and coming days’ air quality help people avoid exposure during episodes with high levels of air \n15 \npollution. Air quality forecasts can be based on deterministic dispersion modelling, but to be accurate this requires detailed \ninformation on future emissions, meteorological conditions and process oriented dispersion modelling. Improving 3-day deterministic air pollution forecasts using machine \nlearning algorithms Christer Johansson1,2, Zhiguo Zhang3, Magnuz Engardt2, Massimo Stafoggia4, Xiaoliang Ma3 \n1Department of Environmental Science, Stockholm University, Stockholm, Sweden \n5 \n2Environment and health administration, SLB-analys, Stockholm, Sweden \n3 KTH Royal Institute of Technology, Dept. of Civil and Architectural Engineering, Stockholm, Sweden \n4 Department of Epidemiology, Lazio Region Health Service, Rome, Italy Christer Johansson1,2, Zhiguo Zhang3, Magnuz Engardt2, Massimo Stafoggia4, Xiaoliang Ma3 \n1Department of Environmental Science, Stockholm University, Stockholm, Sweden \n5 \n2Environment and health administration, SLB-analys, Stockholm, Sweden \n3 KTH Royal Institute of Technology, Dept. of Civil and Architectural Engineering, Stockholm, Sweden \n4 Department of Epidemiology, Lazio Region Health Service, Rome, Italy Correspondence to: Christer Johansson (christer.johansson@aces.su.se) \n10 In this paper we apply \ndifferent machine learning (ML) algorithms – Random forest (RF), Extreme Gradient Boosting (XGB) and Long-Short Term \nMemory (LSTM) – to improve 1-, 2- and 3-day deterministic forecasts of PM10, NOx, and O3 at different sites in Greater Stockholm, Sweden. 20 \nIt is shown that the deterministic forecasts can be significantly improved using the MLs but that the degree of improvement of \nthe deterministic forecasts depends more on pollutant and site than on what machine learning (ML) algorithm is applied. Deterministic forecasts of PM10 is improved by the MLs through the input of lagged measurements and Julian day partly \nreflecting seasonal variations not properly parameterised in the deterministic forecasts. A systematic discrepancy by the Stockholm, Sweden. 20 \nIt is shown that the deterministic forecasts can be significantly improved using the MLs but that the degree of improvement of \nthe deterministic forecasts depends more on pollutant and site than on what machine learning (ML) algorithm is applied. Deterministic forecasts of PM10 is improved by the MLs through the input of lagged measurements and Julian day partly \nreflecting seasonal variations not properly parameterised in the deterministic forecasts. A systematic discrepancy by the deterministic forecasts in the diurnal cycle of NOx is removed by the MLs considering lagged measurements and calendar data \n25 \nlike hour of the day and weekday reflecting the influence of local traffic emissions. For O3 at the urban background site the \nlocal photochemistry not properly accounted for by the relatively coarse Copernicus Atmosphere Monitoring Service ensemble \nmodel (CAMS) used here for forecasting O3, but compensated using the MLs by taking lagged measurements into account. The machine learning models performed similarly well for the sites and pollutants. Performance measures like Pearson \nl ti\nt\n(RMSE)\nb\nl t\nt\n(MAPE)\nd\nb\nl t\n(MAE)\n30 deterministic forecasts in the diurnal cycle of NOx is removed by the MLs considering lagged measurements and calendar data \n25 \nlike hour of the day and weekday reflecting the influence of local traffic emissions. For O3 at the urban background site the \nlocal photochemistry not properly accounted for by the relatively coarse Copernicus Atmosphere Monitoring Service ensemble \nmodel (CAMS) used here for forecasting O3, but compensated using the MLs by taking lagged measurements into account. The machine learning models performed similarly well for the sites and pollutants. Correspondence to: Christer Johansson (christer.johansson@aces.su.se) \n10 Performance measures like Pearson deterministic forecasts in the diurnal cycle of NOx is removed by the MLs considering lagged measurements and calendar data \n25 \nlike hour of the day and weekday reflecting the influence of local traffic emissions. For O3 at the urban background site the \nlocal photochemistry not properly accounted for by the relatively coarse Copernicus Atmosphere Monitoring Service ensemble \nmodel (CAMS) used here for forecasting O3, but compensated using the MLs by taking lagged measurements into account. The machine learning models performed similarly well for the sites and pollutants. Performance measures like Pearson correlation, root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE), \n30 \ntypically differed less than 30% between ML models. At the urban background site, the deviations between modelled and \nmeasured concentrations (RMSE errors) are smaller than uncertainties in the measurements estimated according to correlation, root mean square error (RMSE), mean absolute percentage error (MAPE) and mean absolute error (MAE), \n30 \ntypically differed less than 30% between ML models. At the urban background site, the deviations between modelled and \nmeasured concentrations (RMSE errors) are smaller than uncertainties in the measurements estimated according to 1 1 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. recommendations by the Forum for Air Quality Modeling (FAIRMODE) in the context of the air quality directives. At the \nstreet canyon sites modelled errors are higher, and similar to measurement uncertainties. Further work is needed to reduce \ndeviations between model results and measurements for short periods with relatively high concentrations (peaks). Such peaks \ncan be due to a combination of non-typical emissions and unfavourable meteorological conditions and may be difficult to \nforecast. We have also shown that deterministic forecasts of NOx at street canyon sites can be improved using MLs even if \n5 \nthey are trained at other sites. For PM10 this was only possible using LSTM. recommendations by the Forum for Air Quality Modeling (FAIRMODE) in the context of the air quality directives. At the \nstreet canyon sites modelled errors are higher, and similar to measurement uncertainties. Further work is needed to reduce \ndeviations between model results and measurements for short periods with relatively high concentrations (peaks). Such peaks \ncan be due to a combination of non-typical emissions and unfavourable meteorological conditions and may be difficult to 5 forecast. Key words: Dispersion modelling, random forest, XGboost, LSTM, neural network, PM10, O3, NOx, GAM Correspondence to: Christer Johansson (christer.johansson@aces.su.se) \n10 We have also shown that deterministic forecasts of NOx at street canyon sites can be improved using MLs even if \n5 \nthey are trained at other sites. For PM10 this was only possible using LSTM. An important aspect to consider when choosing ML is that the decision tree based models (RF and XGB) can provide useful \noutput on the importance of features that is not possible using neural network models like LSTM, and also that training and \noptimisation is more complex with LSTM, which could be important to consider when selecting ML algorithm in an An important aspect to consider when choosing ML is that the decision tree based models (RF and XGB) can provide useful \noutput on the importance of features that is not possible using neural network models like LSTM, and also that training and \noptimisation is more complex with LSTM, which could be important to consider when selecting ML algorithm in an \noperational forecast system. A random forest model is now implemented operationally in the forecasts of air pollution and\n10 An important aspect to consider when choosing ML is that the decision tree based models (RF and XGB) can provide useful \noutput on the importance of features that is not possible using neural network models like LSTM, and also that training and \noptimisation is more complex with LSTM, which could be important to consider when selecting ML algorithm in an operational forecast system. A random forest model is now implemented operationally in the forecasts of air pollution and \n10 \nhealth risks in Stockholm. Development of the tuning process and identification of more efficient predictors may make forecast \nmore accurate. operational forecast system. A random forest model is now implemented operationally in the forecasts of air pollution and \n10 \nhealth risks in Stockholm. Development of the tuning process and identification of more efficient predictors may make forecast \nmore accurate. Key words: Dispersion modelling, random forest, XGboost, LSTM, neural network, PM10, O3, NOx, GAM 15 2 2 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 1 \nIntroduction Studies have used ML to predict both hourly and daily average concentrations \nof particulate matter (PM) as well as gaseous air pollutants using meteorological and traffic data (e.g. Quadeer et al., 2020; Di \net al., 2019; Thongthammachart et al., 2021; Kamińska, 2019; Chuluunsaikhan et al., 2021; Doreswamy et al., 2020; Castelli \net al., 2020; Stafoggia et al., 2020; Stafoggia et al., 2019). In addition, a combination of ML, LUR, dispersion modelling, Application of machine learning models (ML) to predict outdoor air quality is getting more and more popular (Rybarczyk and \n20 \nZalakeviciute, 2018; Iskandaryan et al., 2020). Studies have used ML to predict both hourly and daily average concentrations \nof particulate matter (PM) as well as gaseous air pollutants using meteorological and traffic data (e.g. Quadeer et al., 2020; Di \net al., 2019; Thongthammachart et al., 2021; Kamińska, 2019; Chuluunsaikhan et al., 2021; Doreswamy et al., 2020; Castelli \net al., 2020; Stafoggia et al., 2020; Stafoggia et al., 2019). In addition, a combination of ML, LUR, dispersion modelling, ground-based and satellite measurements have been used to obtain temporally and spatially distributed concentrations (Shtein \n25 \net al., 2020; Staffogia et al., 2019; Brokamp et al., 2017; Di et al., 2019). Although good prediction results have been achieved \nusing machine learning models, the challenges of forecasting air pollution concentrations in a longer-term horizon such as a \nday or even several days have not been investigated and very few studies have combined deterministic models and ML in \nforecasting air pollution levels of a few hours/days in the future. ground-based and satellite measurements have been used to obtain temporally and spatially distributed concentrations (Shtein \n25 \net al., 2020; Staffogia et al., 2019; Brokamp et al., 2017; Di et al., 2019). Although good prediction results have been achieved \nusing machine learning models, the challenges of forecasting air pollution concentrations in a longer-term horizon such as a \nday or even several days have not been investigated and very few studies have combined deterministic models and ML in \nforecasting air pollution levels of a few hours/days in the future. ground-based and satellite measurements have been used to obtain temporally and spatially distributed concentrations (Shtein \n25 \net al., 2020; Staffogia et al., 2019; Brokamp et al., 2017; Di et al., 2019). 1 \nIntroduction CHIMERE, EMEP and \nMATCH are part of the Copernicus Atmosphere Monitoring Service (CAMS, atmosphere.copernicus.eu/) to predict air pollution over Europe (Horàlek et al., 2019). The uncertainties in the output of the deterministic models include uncertainties \n15 \nin the input, such as emissions, model algorithms and parameterisations. In urban areas detailed knowledge of the emissions \nis crucial, and there may be important non-linear relationship between the concentration of contaminants and emission. Another \nmethod widely used to obtain spatio-temporal estimates of air pollutant concentrations without detailed knowledge of \nemissions is Land use regression (Hoek et al., 2008). pollution over Europe (Horàlek et al., 2019). The uncertainties in the output of the deterministic models include uncertainties \n15 \nin the input, such as emissions, model algorithms and parameterisations. In urban areas detailed knowledge of the emissions \nis crucial, and there may be important non-linear relationship between the concentration of contaminants and emission. Another \nmethod widely used to obtain spatio-temporal estimates of air pollutant concentrations without detailed knowledge of \nemissions is Land use regression (Hoek et al., 2008). pollution over Europe (Horàlek et al., 2019). The uncertainties in the output of the deterministic models include uncertainties \n15 \nin the input, such as emissions, model algorithms and parameterisations. In urban areas detailed knowledge of the emissions \nis crucial, and there may be important non-linear relationship between the concentration of contaminants and emission. Another \nmethod widely used to obtain spatio-temporal estimates of air pollutant concentrations without detailed knowledge of \nemissions is Land use regression (Hoek et al., 2008). Application of machine learning models (ML) to predict outdoor air quality is getting more and more popular (Rybarczyk and \n20 \nZalakeviciute, 2018; Iskandaryan et al., 2020). Studies have used ML to predict both hourly and daily average concentrations \nof particulate matter (PM) as well as gaseous air pollutants using meteorological and traffic data (e.g. Quadeer et al., 2020; Di \net al., 2019; Thongthammachart et al., 2021; Kamińska, 2019; Chuluunsaikhan et al., 2021; Doreswamy et al., 2020; Castelli \net al., 2020; Stafoggia et al., 2020; Stafoggia et al., 2019). In addition, a combination of ML, LUR, dispersion modelling, Application of machine learning models (ML) to predict outdoor air quality is getting more and more popular (Rybarczyk and \n20 \nZalakeviciute, 2018; Iskandaryan et al., 2020). 1 \nIntroduction According to the World Health Organisation (WHO) air pollution is one of the leading causes of mortality worldwide and is \nregarded as the single largest environmental health risk (Fuller et al., 2022). Acute effects of air pollution are due to short-term \n(e.g. daily) exposures that can lead to reduced lung function, respiratory infections and aggravated asthma. According to the \n5 According to the World Health Organisation (WHO) air pollution is one of the leading causes of mortality worldwide and is \nregarded as the single largest environmental health risk (Fuller et al., 2022). Acute effects of air pollution are due to short-term According to the World Health Organisation (WHO) air pollution is one of the leading causes of mortality worldwide and is \nregarded as the single largest environmental health risk (Fuller et al., 2022). Acute effects of air pollution are due to short-term \n(e.g. daily) exposures that can lead to reduced lung function, respiratory infections and aggravated asthma. According to the \n5 \nEuropean air quality directive, information on the air quality should be made available to the public. Public information \nregarding the expected health risks associated with current or the next few days concentrations of pollutants can be very \nimportant for sensitive persons when planning their outdoor activities. There are different approaches to obtain information on the spatio-temporal variation of air pollutant concentrations - from There are different approaches to obtain information on the spatio-temporal variation of air pollutant concentrations - from \nsimple statistical models to advanced process-oriented models. Gaussian plume models are widely used in urban areas for \n10 \nestimating impacts on atmospheric concentrations from different emission sources and for health risk assessments (Munir et \nal., 2020; Johansson et al., 2009; Orru et al., 2015; Johansson et al., 2017). Eulerian chemical transport models that describe \nemission, transport, mixing, and chemical transformation of trace gases and aerosols such as e.g. CHIMERE, EMEP and \nMATCH are part of the Copernicus Atmosphere Monitoring Service (CAMS, atmosphere.copernicus.eu/) to predict air simple statistical models to advanced process-oriented models. Gaussian plume models are widely used in urban areas for \n10 \nestimating impacts on atmospheric concentrations from different emission sources and for health risk assessments (Munir et \nal., 2020; Johansson et al., 2009; Orru et al., 2015; Johansson et al., 2017). Eulerian chemical transport models that describe \nemission, transport, mixing, and chemical transformation of trace gases and aerosols such as e.g. 1 \nIntroduction Although good prediction results have been achieved \nusing machine learning models, the challenges of forecasting air pollution concentrations in a longer-term horizon such as a \nday or even several days have not been investigated and very few studies have combined deterministic models and ML in \nforecasting air pollution levels of a few hours/days in the future. In this paper we demonstrate how ML can help improve the accuracy of 1-, 2- and 3-day deterministic forecasts of particulate \n30 \nmatter (PM10, particles with an aerodynamic diameter less than 10 µm), nitrogen oxides (NOx) and ozone (O3) for urban \nbackground and street canyon sites in Stockholm, Sweden. The deterministic forecast utilises the CAMS ensemble model to In this paper we demonstrate how ML can help improve the accuracy of 1-, 2- and 3-day deterministic forecasts of particulate \n30 \nmatter (PM10, particles with an aerodynamic diameter less than 10 µm), nitrogen oxides (NOx) and ozone (O3) for urban \nbackground and street canyon sites in Stockholm, Sweden. The deterministic forecast utilises the CAMS ensemble model to 3 account for non-local sources (long-range transport). A Gaussian model is applied over the urban area of Stockholm accounting \nfor local emissions and a street canyon model (OSPM) to account for the effect of buildings on the dispersion of local traffic \nemissions along the roads in the central area of the city. We compare three different machine learning algorithms; two based \non decision trees (random forest and XG Boost) and one neural network model (LSTM). Important questions addressed are \nalso if there are systematic differences in performance depending on different pollutants and different sites. 5 \nhttps://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. account for non-local sources (long-range transport). A Gaussian model is applied over the urban area of Stockholm accounting \nfor local emissions and a street canyon model (OSPM) to account for the effect of buildings on the dispersion of local traffic \nemissions along the roads in the central area of the city. We compare three different machine learning algorithms; two based \non decision trees (random forest and XG Boost) and one neural network model (LSTM). Important questions addressed are \nalso if there are systematic differences in performance depending on different pollutants and different sites. 5 2.1 \nAir pollution measurements Input data for ML modelling are taken from four monitoring stations in central Stockholm, including one urban background \nsite (Torkel Knutssonsgatan, hereafter called UB or urban) and 3 street canyon sites (Hornsgatan HO, Folkungagatan FO and \nSveavägen SV). They are all located in central Stockholm (Figure 1). Detailed descriptions of measurement methods and sites \n10 \nare provided in Appendix A. Input data for ML modelling are taken from four monitoring stations in central Stockholm, including one urban background \nsite (Torkel Knutssonsgatan, hereafter called UB or urban) and 3 street canyon sites (Hornsgatan HO, Folkungagatan FO and Sveavägen SV). They are all located in central Stockholm (Figure 1). Detailed descriptions of measurement methods and sites \n10 \nare provided in Appendix A. Sveavägen SV). They are all located in central Stockholm (Figure 1). Detailed descriptions of measurement methods and sites \n10 \nare provided in Appendix A. Data from the UB site covers approx. 1000 days (10 April 2019 through 31 December 2021). As the OSPM-model became \noperational at a later date, the street canyon data extends over 500 days (5 August 2020 through 31 December 2021). Two \napproaches were tested to handle missing values. The first approach simply ignores data of the timestamps with missing values, \nwhereas the alternative approach substitutes the missing values with mean values of available data in the neighbourhood. 15 Sveavägen SV). They are all located in central Stockholm (Figure 1). Detailed descriptions of measurement methods and sites \n10 \nare provided in Appendix A. Data from the UB site covers approx. 1000 days (10 April 2019 through 31 December 2021). As the OSPM-model became \noperational at a later date, the street canyon data extends over 500 days (5 August 2020 through 31 December 2021). Two Data from the UB site covers approx. 1000 days (10 April 2019 through 31 December 2021). As the OSPM-model became \noperational at a later date, the street canyon data extends over 500 days (5 August 2020 through 31 December 2021). Two \napproaches were tested to handle missing values. The first approach simply ignores data of the timestamps with missing values, \nwhereas the alternative approach substitutes the missing values with mean values of available data in the neighbourhood. 15 Figure 1. Map of central Stockholm showing locations of the urban background site and the street canyons traffic sites. Base map \ncredits: © OpenStreetMap contributors. Figure 1. 2.1 \nAir pollution measurements Map of central Stockholm showing locations of the urban background site and the street canyons traffic sites. Base map \ncredits: © OpenStreetMap contributors. kholm showing locations of the urban background site and the street canyons traffic sites. Base map \nntributors. 4 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 2.2 \nThe Stockholm air quality forecast system Three different dispersion models are used to forecast concentrations considering emissions and dispersion at European, urban \nand street scale (Figure 2). The CAMS ensemble model, part of the Copernicus program was used to obtain forecasts of long-\nrange transported air pollution from outside of the Greater Stockholm model domain. Previous assessments have found the \n5 \nensemble model to be the more accurate than any individual model part of CAMS (Meteo-France, 2017; Marècal et al., 2015). CAMS regional ensemble forecasts are published once a day and each forecast covers 96 hours (4 days). Forecasted \nconcentrations representative of background air, hour by hour, are extracted from a location outside the greater Stockholm \ndomain. All regional models constituting the CAMS ensemble includes physical and chemical schemes dealing with gas phase chemistry, heterogeneous chemistry, aerosol size distribution, aqueous phase chemistry, dry deposition, sedimentation, mineral \n10 \ndust, sea salt, wet deposition, etc. An evaluation of the CAMS regional ensemble forecast in Stockholm has been performed \nby Säll (2018). The contributions to concentrations due to local emissions in the metropolitan area were performed on a 100 m resolution \nusing a Gaussian dispersion model part of the Airviro system (https://www.airviro.com/airviro/). In this modelling domain 10 (Greater Stockholm, 35 by 35 km) individual buildings and street canyons are not resolved but treated using a roughness \n15 \nparameter (Gidhagen et al., 2005). The Gaussian model is fed with meteorological forecasts from the Swedish Meteorological \nand Hydrological Institute (SMHI). A diagnostic wind model is used to account for influences of variations in topography and \nland-use on the dispersion parameters input to the Gaussian model. For details regarding uncertainties and validation of local \nmodelling see Johansson et al. (2017). (Greater Stockholm, 35 by 35 km) individual buildings and street canyons are not resolved but treated using a roughness \n15 \nparameter (Gidhagen et al., 2005). The Gaussian model is fed with meteorological forecasts from the Swedish Meteorological \nand Hydrological Institute (SMHI). A diagnostic wind model is used to account for influences of variations in topography and \nland-use on the dispersion parameters input to the Gaussian model. For details regarding uncertainties and validation of local \nmodelling see Johansson et al. (2017). Finally, the Operational Street Pollution Model (OSPM, Berkowicz, 2000), driven by forecasted meteorology from SMHI is \n20 \napplied for the street canyon sites. It has been applied earlier at Hornsgatan in Stockholm in a number of modelling studies \n(e.g. 2.2 \nThe Stockholm air quality forecast system The database \nand its applications and comparisons between modelling and measurements are described in SLB (2022). The total emissions from road traffic are based on emission factors for different vehicle types including passenger cars (diesel, gasoline, gas), buses \n10 \n(diesel, ethanol), light duty trucks <3.5 ton (diesel and gasoline) and heavy duty trucks >3.5 ton (diesel). Exhaust emission \nfactors of NOx and particles are based on HBEFA version 3.3 (Keller et al., 2017) depending on vehicles Euro class. The \nemission factors per vehicle category were weighted according to the national Swedish Transport Administration vehicle \nregistry, but the vehicle composition taken from national vehicle registry has been shown to be similar to the local fleet using real world number plate recognition measurements at Hornsgatan in campaigns during 2009 (Burman and Johansson, 2010) \n15 \nand 2017 (Burman et al., 2019). For more details, see also Krecl et al., (2017). Non-exhaust emissions of PM due to wear of \nbrakes, tyres and roads are calculated using the NORTRIP model (Denby et al., 2013) forced by the forecasted meteorology \nfrom SMHI. Information on shares of studded winter tyres is obtained from manual counting every week during the winter at \ndifferent locations in the city centre and along highways outside of the city. Road traffic emissions are calculated for all roads real world number plate recognition measurements at Hornsgatan in campaigns during 2009 (Burman and Johansson, 2010) \n15 \nand 2017 (Burman et al., 2019). For more details, see also Krecl et al., (2017). Non-exhaust emissions of PM due to wear of \nbrakes, tyres and roads are calculated using the NORTRIP model (Denby et al., 2013) forced by the forecasted meteorology \nfrom SMHI. Information on shares of studded winter tyres is obtained from manual counting every week during the winter at \ndifferent locations in the city centre and along highways outside of the city. Road traffic emissions are calculated for all roads with more than 3000 vehicles per day. Other emission sources included in the local emissions database include shipping, \n20 \nprivate and municipal heating (including burning of waste). More information about the Stockholm air quality forecast system \nis provided in Engardt et al. (2021). 2.2 \nThe Stockholm air quality forecast system Krecl et al., 2021; Ottosen et al., 2015). NOx and PM10 are modelled on all scales, whereas O3 is only forecasted by the \nCAMS ensemble model. Finally, the Operational Street Pollution Model (OSPM, Berkowicz, 2000), driven by forecasted meteorology from SMHI is \n20 \napplied for the street canyon sites. It has been applied earlier at Hornsgatan in Stockholm in a number of modelling studies \n(e.g. Krecl et al., 2021; Ottosen et al., 2015). NOx and PM10 are modelled on all scales, whereas O3 is only forecasted by the \nCAMS ensemble model. Finally, the Operational Street Pollution Model (OSPM, Berkowicz, 2000), driven by forecasted meteorology from SMHI is \n20 \napplied for the street canyon sites. It has been applied earlier at Hornsgatan in Stockholm in a number of modelling studies \n(e.g. Krecl et al., 2021; Ottosen et al., 2015). NOx and PM10 are modelled on all scales, whereas O3 is only forecasted by the \nCAMS ensemble model. 5 Figure 2. Illustration of the deterministic modelling from European scale at a resolution of 0.1° by 0.1° (ca 11 km × 6 km), via urban \nscale (100 m resolution over an area of 35 by 35 km) down to the street canyon sites. The CAMS ensemble forecast map example is \ntaken from https://atmosphere.copernicus.eu/ (accessed 1 Feb 2023). The map with the Gaussian model local forecast example is \noutput from the Airviro system (https://www.airviro.com/airviro/, accessed 1 Feb 2023) used in Stockholm. The illustration of a \nstreet canyon site is taken from https://www.wikiwand.com/en/Operational_Street_Pollution_Model (accessed 1 Feb 2023). https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure 2. Illustration of the deterministic modelling from European scale at a resolution of 0.1° by 0.1° (ca 11 km × 6 km), via urban \nscale (100 m resolution over an area of 35 by 35 km) down to the street canyon sites. The CAMS ensemble forecast map example is \ntaken from https://atmosphere.copernicus.eu/ (accessed 1 Feb 2023). The map with the Gaussian model local forecast example is \noutput from the Airviro system (https://www.airviro.com/airviro/, accessed 1 Feb 2023) used in Stockholm. The illustration of a \nstreet canyon site is taken from https://www.wikiwand.com/en/Operational_Street_Pollution_Model (accessed 1 Feb 2023). 5 For the urban scale model domain a detailed emission database is used as input for the local dispersion modelling. 2.4 \nMachine learning models As already mentioned in the introduction two decision tree based machine learning models, RF and XGB, and one deep \nlearning model, LSTM are applied. In addition, an ensemble learning approach based on a General Additive Model (GAM), \naggregating the above three learning models, is also applied to further optimise the results. One essential challenge in this study is to forecast hourly concentrations for the coming one day, two days and three days based \n10 \non historical air pollution measurement and other available information as inputs. This indicates that the essential statistical \nprediction involves time series prediction for multiple time steps, for example, 72 time steps for three days prediction. It is \nknown that a sequence-to-sequence time series prediction, implemented using LSTM or other recurrent neural networks, \nprovides a straightforward and rolling-over computational schemes. Nevertheless, training a machine learning model with One essential challenge in this study is to forecast hourly concentrations for the coming one day, two days and three days based \n10 \non historical air pollution measurement and other available information as inputs. This indicates that the essential statistical \nprediction involves time series prediction for multiple time steps, for example, 72 time steps for three days prediction. It is \nknown that a sequence-to-sequence time series prediction, implemented using LSTM or other recurrent neural networks, \nprovides a straightforward and rolling-over computational schemes. Nevertheless, training a machine learning model with p\ng\ng\np\n,\ng\ng\nmultiple outputs requires much more computational effort, but often leads to inferior prediction accuracy compared to \n15 \nrelatively simple models with only a single output dedicated for predicting output of a certain time step. Therefore, this study \nchooses, instead of more complex machine learning structure, multiple single-output machine learning models for forecasting \ndifferent air pollutants for k=1 day, 2 day and 3 day interval: multiple outputs requires much more computational effort, but often leads to inferior prediction accuracy compared to \n15 \nrelatively simple models with only a single output dedicated for predicting output of a certain time step. 2.3 \nMeteorological forecasts 6\nAs an integral part of the Stockholm air quality forecast system, meteorological forecasts for a point in central Stockholm are \n25 \ndownloaded \nevery \nmorning \nfrom \nSMHI \n(https://www.smhi.se/data/oppna-data) \nand \nMET \nNorway \n(https://docs.api.met.no/doc/). The meteorological forecasts extend over 10 days and are a combination of output from a 6\nAs an integral part of the Stockholm air quality forecast system, meteorological forecasts for a point in central Stockholm are \n25 \ndownloaded \nevery \nmorning \nfrom \nSMHI \n(https://www.smhi.se/data/oppna-data) \nand \nMET \nNorway \n(https://docs.api.met.no/doc/). The meteorological forecasts extend over 10 days and are a combination of output from a 6 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. number of regional and global numerical weather prediction models. The combination is based on statistical adjustments as \nwell as manual edits. The meteorology is initially used to drive the models of weather-dependent PM emissions and the urban- \nand street canyon air quality modelling. The forecasted meteorological data are, finally, also used as predictors in the ML \nalgorithms as detailed below. number of regional and global numerical weather prediction models. The combination is based on statistical adjustments as \nwell as manual edits. The meteorology is initially used to drive the models of weather-dependent PM emissions and the urban- \nand street canyon air quality modelling. The forecasted meteorological data are, finally, also used as predictors in the ML \nalgorithms as detailed below. 5 2.4 \nMachine learning models Therefore, this study \nchooses, instead of more complex machine learning structure, multiple single-output machine learning models for forecasting \ndifferent air pollutants for k=1 day, 2 day and 3 day interval: 𝜌̂𝑖,𝑗(𝑑, 𝑡) = mlearn_model (𝜌̃𝑖,𝑗(𝑑−𝑘, 𝑡), 𝜌̅𝑖,𝑗\n𝑆(𝑑−𝑘, 𝑡), 𝜌̌𝑖,𝑗(𝑑, 𝑡), 𝑊(𝑑, 𝑡), 𝐶(𝑑, 𝑡) ) where 𝜌̂𝑖,𝑗(𝑑, 𝑡) is predicted concentration value of the pollutant j for day d and time t at the location i, and 𝜌̃𝑖,𝑗(𝑑, 𝑡) is the \n20 \ncorresponding real measurement; 𝜌̅𝑖,𝑗\n𝑆(𝑑, 𝑡) uses a set S to represent several statistical measures, including maximum, \nminimum, 25% quantile and 75% quantile of the measured concentration data during the past 24 hours until 𝜌̃𝑖,𝑗(𝑑, 𝑡), and the \nmeasurement dataset can be represented by a set, i.e. {𝜌̃𝑖,𝑗(𝑑, 𝑡), 𝜌̃𝑖,𝑗(𝑑, 𝑡−1), 𝜌̃𝑖,𝑗(𝑑, 𝑡−2). . . . }. 𝜌̌𝑖,𝑗(𝑑, 𝑡) is the one day \npredicted concentration value using deterministic physical model. 𝑊(𝑑, 𝑡) represents the weather condition predicted for day d and time t. 25 \nFigure 3 demonstrates the prediction horizon and lagged information horizon for the case of one day prediction. To build \nconsistent statistical machine learning models with a fixed rolling horizon, a new measurement point at current time (d, t) will \nlead to an additional prediction for one day ahead, i.e. the predicted value at (d+1,t). In the case, the measurement statistics \n𝜌̅𝑖,𝑗\n𝑆(𝑑, 𝑡) will be based on one day preceding measurement data of (d, t), leading to a lagged rolling horizon described in the \nfigure. 30 7 7 Figure 3. Illustration of the machine learning modelling framework for one-day prediction based on available datasets. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure 3. Illustration of the machine learning modelling framework for one-day prediction b Figure 3. Illustration of the machine learning modelling framework for one-day prediction based on available datasets. This study has applied both LSTM and two conventional supervised learning models, RF and \nlearning cores to carry out supervised learning using the same input and output training dataset This study has applied both LSTM and two conventional supervised learning models, RF and XGB, as the essential machine \nlearning cores to carry out supervised learning using the same input and output training dataset. In fact, an ensemble approach \nbased on all three models is also applied to predict air quality for different days. 2.4 \nMachine learning models The conventional models require nontrivial \n5 \neffort to prepare input feature data as they don’t fit as easily with time series data as RNN. To make a fair comparison with \nboth types of models, LSTM model in this case is only based on the same type of input as other two models. It is well known \nthat LSTM can learn the temporal correlation of different ranges. Nevertheless, this study applies the data to a simple LSTM \nstructure, without taking advantages of its full potential. In principle, the measurement data at (d, t) may provide hourly update based on all three models is also applied to predict air quality for different days. The conventional models require nontrivial \n5 \neffort to prepare input feature data as they don’t fit as easily with time series data as RNN. To make a fair comparison with \nboth types of models, LSTM model in this case is only based on the same type of input as other two models. It is well known \nthat LSTM can learn the temporal correlation of different ranges. Nevertheless, this study applies the data to a simple LSTM \nstructure, without taking advantages of its full potential. In principle, the measurement data at (d, t) may provide hourly update 5 of predicted values within the prediction horizon i.e. from (d,t+1) to (d+1,t). Nevertheless, it is our future work to extend the \n10 \nmodel structure and improve prediction using latest real-time information. In addition to the measured air pollution time series data itself, the forecasted meteorological conditions for the prediction day \nd (or d+1 or d+2) and calendar information such as weekday, hour etc. are also applied as input features. Moreover, the air \npollutant concentrations predicted by the deterministic models is also used as inputs to the MLs. model structure and improve prediction using latest real time information. In addition to the measured air pollution time series data itself, the forecasted meteorological conditions for the prediction day \nd (or d+1 or d+2) and calendar information such as weekday, hour etc. are also applied as input features. Moreover, the air \npollutant concentrations predicted by the deterministic models is also used as inputs to the MLs. Fel! Hittar inte referenskälla. summarizes the methodological framework of machine learning and associated computational \n15 \nexperiments for air pollution prediction. Fel! Hittar inte referenskälla. 2.4 \nMachine learning models summarizes the methodological framework of machine learning and associated computational \n15 \nexperiments for air pollution prediction. Fel! Hittar inte referenskälla. summarizes the methodological framework of machine learning and associated computational \n15 \nexperiments for air pollution prediction. 8 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Th i\ni\nl d\nh d\ni i i f\nf PM\nNO\nd O\nl\nh\nh h d\ni i i f\nb The input includes the deterministic forecasts of PM10, NOx and O3, to evaluate how much the deterministic forecasts can be\nimproved by the ML algorithms. In the computational experiments, data-driven forecasting models are trained for one urban \nbackground site and three street canyon sites separately, and different machine learning models are trained and tested separately\nfor predicting various air pollution concentrations coming 1-day (0 – 24 h), 2-day (25 – 48 h) and 3-day (48 – 72 h) periods. Table 1 presents detailed explanation of the essential input features that are applied in the computational experiments. All \nmachine learning models are implemented in python using existing machine learning libraries including “scikit-learn” and\n“tensorflow” (also implemented using “pytorch”) for conventional machine learning models and deep learnings models\nrespectively. The detailed implementation can be referred to the code provided. The input includes the deterministic forecasts of PM10, NOx and O3, to evaluate how much the deterministic forecasts can be \nimproved by the ML algorithms. In the computational experiments, data-driven forecasting models are trained for one urban \nbackground site and three street canyon sites separately, and different machine learning models are trained and tested separately \n5 \nfor predicting various air pollution concentrations coming 1-day (0 – 24 h), 2-day (25 – 48 h) and 3-day (48 – 72 h) periods. Table 1 presents detailed explanation of the essential input features that are applied in the computational experiments. All \nmachine learning models are implemented in python using existing machine learning libraries including “scikit-learn” and \n“tensorflow” (also implemented using “pytorch”) for conventional machine learning models and deep learnings models \nrespectively. The detailed implementation can be referred to the code provided. 10 background site and three street canyon sites separately, and different machine learning models are trained and tested separately \n5 \nfor predicting various air pollution concentrations coming 1-day (0 – 24 h), 2-day (25 – 48 h) and 3-day (48 – 72 h) periods. 2.4 \nMachine learning models Table 1 presents detailed explanation of the essential input features that are applied in the computational experiments. All \nmachine learning models are implemented in python using existing machine learning libraries including “scikit-learn” and \n“tensorflow” (also implemented using “pytorch”) for conventional machine learning models and deep learnings models \nrespectively. The detailed implementation can be referred to the code provided. 10 9 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Table 1. Measured and forecasted air pollutant concentrations used as input data (features) in the ML modelling of pollutant \nconcentrations at the urban background site (UB) and at the street canyon sites (SC). NOx and PM10 are modelled at both UB and \nSC. Ozone is only modelled at UB. For periodic input data, using sine and cosine values can remove discontinuities and create \nconsistent distance measures, thereby improving model accuracy. Category \nShort \nnames \nDescription \nDeterministic features \nNOx_nday_local \nPM10_nday_local \nn=1, 2, 3 \nDeterministic 1-day, 2-day and 3-day forecast of contributions from \nlocal emissions based on urban scale Gaussian modelling \nNOx_nday_regional \nPM10_nday_regional \nO3_nd_regional \nn=1, 2, 3 \nDeterministic 1-day, 2-day and 3-day forecast of contributions based \nfrom non-local emissions based on CAMS ensemble model (regional \nbackground) \nAutocorrelation features \nNOx_lagXX \nPM10_lagXX \nO3_lagXX \nXX = 24, 48, 72 \nXX hour lagged air pollutant concentrations based on autocorrelation \nand prediction time span. Statistical features \nNOx_Sta_dXX \nPM10_Sta_dXX \nO3_Sta_dXX \nSta=avg., median, min, \nmax, Q1, Q3 \nXX = 24, 48, 72 \nAverage, median, minimum, maximum, quantiles 1 and quantiles 3 of \nlagged air pollutant concentrations in rolling XX hour periods. Time features \nTime \nTime_sin \nTime_cos \nTime= year, julianday, \nmonth, weekday, day, hour \nJulian day of the year (1, 2, 3, … 365), sine and cosine of 2*pi*day/365. Day of the week (1, 2, 3, … 7), sine and cosine of 2*pi*day/7. Hour of the day (0, 1, 2, … 23), sine and cosine of 2*pi*hour/24. Year \nMonth \nDay \nMeteorological features \nwind_direction \nwind_direction_cos \nwind_direction_sin \nWind direction[0, 360) at 10 m in central Stockholm, sine and cosine \nof (2*pi/360)*wind direction \npressure; temperature; \nprecipitation; cloudiness \nPressure (10 m); Temperature (10 m) \nwind_speed \nWind speed (10 m) \nrelative_humidity \nRelative humidity \nboundary_layer_height \nBoundary layer height for central Stockholm 5 2.5 \nStatistical performance indicators Several common performance metrics have been selected for comparing the prediction results of different machine learning \nmodels including Pearson correlation (r) and normalised error measures: mean average error (MAE), mean absolute percentage \nerror (MAPE) and root mean squared error (RMSE). These measures have also been recommended for air quality model \nbenchmarking in the context of the Air Quality Directive 2008/50/EC (AQD) by Janssen and Thunis (2022). 10 Mean absolute error: Mean absolute error: 𝑀𝐴𝐸(𝑦, 𝑦̂) = 1\n𝑛∑ \n𝑛\n𝑖=1\n|𝑦𝑖−𝑦̂𝑖| 𝑀𝐴𝐸(𝑦, 𝑦̂) = 1\n𝑛∑ \n𝑛\n𝑖=1\n|𝑦𝑖−𝑦̂𝑖| 10 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. where 𝑦̂𝑖 is the predicted value of the 𝑖-th sample, and 𝑦𝑖 is the corresponding true value for total 𝑛 samples. q\n𝑅𝑀𝑆𝐸(𝑦, 𝑦̂) = √1\n𝑛∑ \n𝑛\n𝑖=1\n(𝑦𝑖−𝑦̂𝑖)2 MAE and RMSE were normalised by diving by the mean of the measured concentrations, hereafter called nMAE and \n5 \nnRMSE. Mean absolute percentage error: \n𝑛 Mean absolute percentage error: 𝑀𝐴𝑃𝐸(𝑦, 𝑦̂) = 1\n𝑛∑ \n𝑛\n𝑖=1\n|𝑦𝑖−𝑦̂𝑖|\n|𝑦𝑖| 𝑀𝐴𝑃𝐸(𝑦, 𝑦̂) = 1\n𝑛∑ \n𝑛\n𝑖=1\n|𝑦𝑖−𝑦̂𝑖|\n|𝑦𝑖|\n \n0 10 10 Pearson correlation coefficient: Pearson correlation coefficient: r(𝑦, 𝑦̂) =\n∑\n \n𝑛\n𝑖=1 (𝑦𝑖−𝑦𝑖̅)(𝑦̂𝑖−𝑦̂𝑖̅)\n√∑\n \n𝑛\n𝑖=1 (𝑦𝑖−𝑦𝑖̅)2√∑\n \n𝑛\n𝑖=1 (𝑦̂𝑖−𝑦̂𝑖̅)2 r(𝑦, 𝑦̂) =\n∑\n \n𝑛\n𝑖=1 (𝑦𝑖−𝑦𝑖̅)(𝑦̂𝑖−𝑦̂𝑖̅)\n√∑\n \n𝑛\n𝑖=1 (𝑦𝑖−𝑦𝑖̅)2√∑\n \n𝑛\n𝑖=1 (𝑦̂𝑖−𝑦̂𝑖̅)2 The model quality indicator (MQI): In order to properly assess model quality it is necessary to consider measurement uncertainty. In the FAIRMODE community, \n15 \nthe modelling quality indicator (MQI) is used to assess if a model fulfils certain objectives (Janssen and Thunis, 2022). It is \ndefined as the ratio between the model bias at a fixed time (i), quantified by the RMSE, and a quantity proportional to the \nmeasurement uncertainty as: In order to properly assess model quality it is necessary to consider measurement uncertainty. In the FAIRMODE community, \n15 \nthe modelling quality indicator (MQI) is used to assess if a model fulfils certain objectives (Janssen and Thunis, 2022). It is \ndefined as the ratio between the model bias at a fixed time (i), quantified by the RMSE, and a quantity proportional to the \nmeasurement uncertainty as: 𝑀𝑄𝐼(𝑖) = \n√1\n𝑛∑\n \n𝑛\n𝑖=1 (𝑦𝑖−𝑦̂𝑖)2\n𝛽√1\n𝑛∑\n \n𝑛\n𝑖=1 𝑈(𝑦𝑖)2\n= 𝑅𝑀𝑆𝐸\n𝛽𝑅𝑀𝑆𝑈\n \n20 20 20 U(yi) is the expanded 95th percentile measurement uncertainty and β is a coefficient of proportionality (Janssen and Thunis, \n2022). The value of β determines the stringency of the MQI and is set equal to 2, allowing thus deviation between modelled \nand measured concentrations as twice the measurement uncertainty. The uncertainty of the measurements (RMSU) was \ncalculated for the mean of the measurement concentrations as: \n25 𝑈(𝑦𝑖) = 𝑈𝑟(𝑅𝑉)√(1 −∝2)2(𝑦𝑖2) +∝2 𝑅𝑉2 𝑈(𝑦𝑖) = 𝑈𝑟(𝑅𝑉)√(1 −∝2)2(𝑦𝑖2) +∝2 𝑅𝑉2 Here 𝑈𝑟(𝑅𝑉) and ∝ are parameters that depend on pollutant and RV is a reference value, here taken to be 200, 50 and 120 µg \nm-3, corresponding 𝑈𝑟(𝑅𝑉) was 0.24, 0.28 and 0.18 and ∝ was 0.25, 0.20, 0.79 for NO2, PM10 and O3 respectively (Janssen \nand Thunis, 2022). In our case we have calculated NOx, not NO2, but we used the same settings of the parameters for NOx as \nrecommended for NO2. It should be noted that another important source of error when comparing model results with \n30 \nmeasurements is associated with the spatial representativeness of a measurement station for comparison with the model. This Here 𝑈𝑟(𝑅𝑉) and ∝ are parameters that depend on pollutant and RV is a reference value, here taken to be 200, 50 and 120 µg \nm-3, corresponding 𝑈𝑟(𝑅𝑉) was 0.24, 0.28 and 0.18 and ∝ was 0.25, 0.20, 0.79 for NO2, PM10 and O3 respectively (Janssen \nand Thunis, 2022). The model quality indicator (MQI): In our case we have calculated NOx, not NO2, but we used the same settings of the parameters for NOx as \nrecommended for NO2. It should be noted that another important source of error when comparing model results with \n30 \nmeasurements is associated with the spatial representativeness of a measurement station for comparison with the model. This Here 𝑈𝑟(𝑅𝑉) and ∝ are parameters that depend on pollutant and RV is a reference value, here taken to be 200, 50 and 120 µg \nm-3, corresponding 𝑈𝑟(𝑅𝑉) was 0.24, 0.28 and 0.18 and ∝ was 0.25, 0.20, 0.79 for NO2, PM10 and O3 respectively (Janssen \nand Thunis, 2022). In our case we have calculated NOx, not NO2, but we used the same settings of the parameters for NOx as \nrecommended for NO2. It should be noted that another important source of error when comparing model results with \n30 \nmeasurements is associated with the spatial representativeness of a measurement station for comparison with the model. This 11 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. is due to the mismatch between the model grid resolution and the location of the monitoring station. But in this paper we are \nmainly interested in comparing the results of the deterministic model with the results using the different MLs together with the \ndeterministic model output. is due to the mismatch between the model grid resolution and the location of the monitoring station. But in this paper we are \nmainly interested in comparing the results of the deterministic model with the results using the different MLs together with the \ndeterministic model output. 3 \nResults The focus of this paper is to compare the deterministic forecasts of NOx, PM10 and O3 with the forecasts based on the different \n5 \nmachine learners which also include the deterministic forecasts as input variables (features). As described above we have made \ndeterministic and ML forecasts for hourly mean concentrations for the coming 72 hours, based on 1-day, 2-day and 3-day \nmeteorological forecasts for one urban background site (NOx, PM10 and O3) and three street canyon sites (NOx and PM10). We \nalso compare results separately for the urban background site and the street canyon sites. 3.1.1 \nImportance of features - urban background The relative importance of different features depending on model (RF or XGB), pollutant (PM10, NOx, O3) and forecast period \n(1-day, 2-day and 3-day) is shown in plots in Appendix B. It should be noted that the local deterministic models (Gauss and \nOSPM) use the same meteorological data to forecast concentrations, so when the meteorological variables are important features for the MLs, it indicates that the deterministic models don’t capture all processes related to those variables. In summary \n15 \nregarding importance of features for urban background: \nNOx. Lagged 24-hour mean concentrations, calendar data, wind speed and local deterministic forecasts are among the top-10 \nmost important variables using RF and XGB, but it can be noted that the deterministic forecast is not the most important feature \nfor any model. Of the calendar features hour is most important reflecting the importance of regular, diurnal variations in traffic features for the MLs, it indicates that the deterministic models don’t capture all processes related to those variables. In summary \n15 \nregarding importance of features for urban background: \nNOx. Lagged 24-hour mean concentrations, calendar data, wind speed and local deterministic forecasts are among the top-10 regarding importance of features for urban background: \nNOx. Lagged 24-hour mean concentrations, calendar data, wind speed and local deterministic forecasts are among the top-10 \nmost important variables using RF and XGB, but it can be noted that the deterministic forecast is not the most important feature \nfor any model. Of the calendar features hour is most important reflecting the importance of regular, diurnal variations in traffic \nemissions. 20 emissions. 20 \nPM10. The regional deterministic forecast is the most important feature for PM10 forecasts, both for RF and XGB and for all \nforecast days. Also lagged measurements, both average, minimum and maximum concentrations is important. Of the calendar \nfeatures the seasonal variation is reflected in the importance of the Julian day. y\np\ng\np\nregional deterministic forecasts is the dominant feature both for RF and XGB, and for all forecast days. Also lagged measured \n25 \nmaximum concentrations is of some importance. The relative humidity is important, likely reflecting that O3 concentrations \nare typically higher during dry, clear sky conditions, which may not be completely captured by the deterministic forecasts. 12 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 3.1.2 \nComparison between deterministic forecasts and MLs - urban background Figure 4 shows an example of the temporal variations in September 2021 in the forecasts with deterministic modelling and \nGAM in comparison to the observations. Similar plots are also given for individual models in Figure C1. The plots were made \nusing the Openair R package (Carslaw and Ropkins, 2012). For all pollutants the MLs tend to improve the variability in the \nobserved concentrations compared to the deterministic forecasts, but there are significant deviations. For O3 the minimum \n5 \nconcentrations observed is often not forecasted so well and for PM10 the highest concentrations is not captured by the models. concentrations observed is often not forecasted so well and for PM10 the highest concentrations is not captured by the \n \nFigure 4. Temporal variations in hourly mean NOx, PM10 and O3 concentrations at the urban background site during Sep\n2021 based on observations, deterministic forecasts and GAM. Mean of 1-, 2- and 3-day forecasts. Figure 4. Temporal variations in hourly mean NOx, PM10 and O3 concentrations at the urban background site during Septe\n2021 based on observations, deterministic forecasts and GAM. Mean of 1-, 2- and 3-day forecasts. 0 Figure 4. Temporal variations in hourly mean NOx, PM10 and O3 concentrations at the urban background site during September \n2021 based on observations, deterministic forecasts and GAM. Mean of 1-, 2- and 3-day forecasts. 10 10 13 Figure 5 shows example of deviations from observations of forecasted NOx, PM10 and O3 for all models illustrating that during \nsome hours all models systematically show large absolute deviations from the observed mean concentrations. Sometimes the \nhours with large deviation for NOx coincide with deviations for PM10 indicating some specific meteorological situation or \ncommon source that cause this deviation. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure 5 shows example of deviations from observations of forecasted NOx, PM10 and O3 for all models illustrating that during \nsome hours all models systematically show large absolute deviations from the observed mean concentrations. Sometimes the \nhours with large deviation for NOx coincide with deviations for PM10 indicating some specific meteorological situation or \ncommon source that cause this deviation. 5 Figure 5. 3.1.2 \nComparison between deterministic forecasts and MLs - urban background Absolute deviations of forecasted NOx, PM10 and O3 concentrations from observed (Obs) concentrations based on mean of \n1-, 2- and 3-day forecasts for September 2021. All data are hourly mean concentrations. Figure 5. Absolute deviations of forecasted NOx, PM10 and O3 concentrations from observed (Obs) concentrations based on mean of \n1-, 2- and 3-day forecasts for September 2021. All data are hourly mean concentrations. Figure C2 shows systematic deviations between the observed mean diurnal variations and the deterministic forecast. This is \nsignificantly improved using the MLs, especially for NOx and O3. For O3 the deterministic forecast systematically \n10 Figure C2 shows systematic deviations between the observed mean diurnal variations and the deterministic forecast. This is \nsignificantly improved using the MLs, especially for NOx and O3. For O3 the deterministic forecast systematically \n0 10 14 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. overestimates the concentrations which is mainly due to the fact that the chemical destruction of O3 in the city centre is not \nproperly accounted for by the regional CAMS model. For NOx the concentrations calculated by the deterministic model are \nsystematically shifted one hour compared to the observed concentration and this is likely associated with errors in \nparameterisation of traffic emissions, which is the most important source of NOx in Stockholm. For PM10 concentrations \nmodelled by the deterministic model are too low during the night compared to observations, but this is corrected using RF and \n5 \nXGB, but not using LSTM. overestimates the concentrations which is mainly due to the fact that the chemical destruction of O3 in the city centre is not \nproperly accounted for by the regional CAMS model. For NOx the concentrations calculated by the deterministic model are \nsystematically shifted one hour compared to the observed concentration and this is likely associated with errors in \nparameterisation of traffic emissions, which is the most important source of NOx in Stockholm. For PM10 concentrations \nmodelled by the deterministic model are too low during the night compared to observations, but this is corrected using RF and \n5 \nXGB, but not using LSTM. As can be seen in Table 2 and Figure 6 most of the statistical performance measures are improved compared to the deterministic \nforecasts of NOx and PM10 using different MLs. For NOx Pearson correlation (r) increases from 0.35-0.39 with deterministic \nforecasts to between 0.49 and 0.70 when MLs are used. 3.1.2 \nComparison between deterministic forecasts and MLs - urban background Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Table 2. Comparison of 1-, 2-, 3-day deterministic and ML forecasts for NOx, PM10 and O3 for the urban background site. r = \nPearson correlation, MAPE = mean absolute percentage error, nRMSE = normalised rootmean square error and nMAE = \nnormalised mean absolute error. All data are based on hourly mean values. NOx \n \nr \nMAPE \nnRMSE \nnMAE \n \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \nDet \n0.39 \n0.38 \n0.35 \n69% \n65% \n67% \n130% \n124% \n116% \n63% \n61% \n61% \nXGB \n0.49 \n0.53 \n0.54 \n42% \n44% \n48% \n118% \n114% \n114% \n44% \n45% \n47% \nRF \n0.54 \n0.57 \n0.60 \n37% \n38% \n37% \n115% \n112% \n111% \n41% \n41% \n41% \nLSTM \n0.70 \n0.69 \n0.66 \n50% \n59% \n54% \n99% \n99% \n101% \n43% \n47% \n46% \nGAM \n0.50 \n0.55 \n0.58 \n37% \n37% \n37% \n117% \n114% \n112% \n42% \n42% \n42% \nPM10 \n \nr \nMAPE \nnRMSE \nnMAE \n \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \nDet \n0.53 \n0.50 \n0.50 \n54% \n56% \n59% \n81% \n85% \n87% \n47% \n48% \n50% \nXGB \n0.71 \n0.65 \n0.56 \n61% \n64% \n69% \n58% \n64% \n69% \n41% \n44% \n47% \nRF \n0.74 \n0.65 \n0.60 \n55% \n74% \n78% \n56% \n63% \n66% \n39% \n45% \n46% \nLSTM \n0.71 \n0.57 \n0.50 \n47% \n54% \n60% \n62% \n73% \n79% \n42% \n49% \n53% \nGAM \n0.73 \n0.64 \n0.59 \n55% \n76% \n80% \n56% \n64% \n67% \n39% \n46% \n47% \nO3 \n \nr \nMAPE \nnRMSE \nnMAE \n \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \nDet \n0.74 \n0.71 \n0.69 \n45% \n49% \n50% \n31% \n32% \n32% \n24% \n25% \n25% \nXGB \n0.75 \n0.71 \n0.67 \n47% \n51% \n53% \n25% \n26% \n27% \n19% \n20% \n21% \nRF \n0.76 \n0.69 \n0.71 \n47% \n54% \n52% \n24% \n26% \n26% \n19% \n21% \n20% \nLSTM \n0.76 \n0.74 \n0.74 \n46% \n47% \n51% \n24% \n25% \n25% \n19% \n20% \n20% \nGAM \n0.75 \n0.66 \n0.69 \n47% \n55% \n52% \n24% \n27% \n27% \n19% \n22% \n21% Table 2. Comparison of 1-, 2-, 3-day deterministic and ML forecasts for NOx, PM10 and O3 for the urban background site. r = \nPearson correlation, MAPE = mean absolute percentage error, nRMSE = normalised rootmean square error and nMAE = \nnormalised mean absolute error. All data are based on hourly mean values. Table 2. Comparison of 1-, 2-, 3-day deterministic and ML forecasts for NOx, PM10 and O3 for the urban background site. 3.1.2 \nComparison between deterministic forecasts and MLs - urban background MAPE, nRMSE and nMAE decreases for all models and all forecast \n10 \ndays. For PM10 Pearson r increases from 0.50-0.53 with deterministic forecasts to between 0.50 and 0.74 when MLs are used. nRMSE and nMAE decreases for forecast days, but for MAPE results are not so consistent – MAPE increases slightly with \nXGB, RF and GAM, while it decrease for 1-day and 2-day forecasts using LSTM. For O3 there are small improvements looking \nat Pearson r and MAPE, nRMSE and nMAE decreases. The Pearson correlation for O3 is already relatively high and errors \nl ti\nl\nll\nith th d t\ni i ti CAMS\nd lli\n15 As can be seen in Table 2 and Figure 6 most of the statistical performance measures are improved compared to the deterministic \nforecasts of NOx and PM10 using different MLs. For NOx Pearson correlation (r) increases from 0.35-0.39 with deterministic \nforecasts to between 0.49 and 0.70 when MLs are used. MAPE, nRMSE and nMAE decreases for all models and all forecast \n10 \ndays. For PM10 Pearson r increases from 0.50-0.53 with deterministic forecasts to between 0.50 and 0.74 when MLs are used. nRMSE and nMAE decreases for forecast days, but for MAPE results are not so consistent – MAPE increases slightly with \nXGB, RF and GAM, while it decrease for 1-day and 2-day forecasts using LSTM. For O3 there are small improvements looking \nat Pearson r and MAPE, nRMSE and nMAE decreases. The Pearson correlation for O3 is already relatively high and errors relatively small with the deterministic CAMS modelling. 15 \nFigure 6 presents mean of 1-day, 2-day and 3-day statistical performances as ratios of ML to deterministic forecasts. This \nshows that NOx is consistently improved using all MLs for all statistical performance indexes, whereas for PM10 and O3 there \nare improvements in nRMSE and nMAE, but MAPE. Overall, the difference in performance between different models is small, \nless than 30%, but larger when comparing different pollutants. y\ng\nFigure 6 presents mean of 1-day, 2-day and 3-day statistical performances as ratios of ML to deterministic forecasts. This \nshows that NOx is consistently improved using all MLs for all statistical performance indexes, whereas for PM10 and O3 there \nare improvements in nRMSE and nMAE, but MAPE. Overall, the difference in performance between different models is small, \nless than 30%, but larger when comparing different pollutants. 15 https://doi.org/10.5194/acp-2023-38\nPreprint. 3.1.2 \nComparison between deterministic forecasts and MLs - urban background r = \nPearson correlation, MAPE = mean absolute percentage error, nRMSE = normalised rootmean square error and nMAE = \nnormalised mean absolute error. All data are based on hourly mean values. 16 os of statistical performances for MLs versus the deterministic hourly forecasts for the urban site. Mean of 1-day, 2-\nforecasts. e 6. Ratios of statistical performances for MLs versus the deterministic hourly forecasts for the urban site. Mean of 1-day, 2 Figure 6. Ratios of statistical performances for MLs versus the deterministic hourly forecasts for the urban site. Mean of 1-day, 2-\nday and 3-day forecasts. For the general public it is important to receive information on future pollution episodes with high concentrations. The plots \nin Figure D1 shows that statistical performances for all models is worse when concentrations higher than when the mean value \n5 \nis analysed. Pearson r is somewhat higher for PM10 and O3, but not when RF and XGB is used for NOx. MAPE is reduced for \nPM10 and NOx but not for O3. The nRMSE is both higher and lower with MLs compared to the deterministic model, while, \nfinally, nMAE is lower for NOx and PM10 using RF and XGB, but not for PM10 using LSTM. g\np\np\np\np\ng\np\nin Figure D1 shows that statistical performances for all models is worse when concentrations higher than when the mean value \n5 \nis analysed. Pearson r is somewhat higher for PM10 and O3, but not when RF and XGB is used for NOx. MAPE is reduced for \nPM10 and NOx but not for O3. The nRMSE is both higher and lower with MLs compared to the deterministic model, while, \nfinally, nMAE is lower for NOx and PM10 using RF and XGB, but not for PM10 using LSTM. 17 17 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. As can be seen in Figure 7 all MQI are below 100% indicating that deviations between model results and measurements are \nsmaller than the estimated uncertainties in the measurements. 3.1.2 \nComparison between deterministic forecasts and MLs - urban background It can also be seen that LSTM is somewhat more efficient in \nreducing MQI, from 68% to 60% for NOx and O3 from 40% to 29%, while RF and XGB provides no improvement for NOx, \nbut both PM10 and O3 shows slightly lower MQI with RF and XGB compared to the deterministic forecast. 5 Figure 7. MQI based on hourly mean concentrations for the whole test period for NOx, PM10 and O3 of the urban site. Mean of 1-, \n2- and 3-day forecasts. Figure 7. MQI based on hourly mean concentrations for the whole test period for NOx, PM10 and O3 of the urban site. Mean of 1-, \n2- and 3-day forecasts. Figure 7. MQI based on hourly mean concentrations for the whole test period for NOx, PM10 and O3 of the urban site. Mean of 1-, \n2- and 3-day forecasts. 3.2.1 \nImportance of features - street canyon sites For the street canyon sites the relative importance of different features is different for PM10 and NOx and also somewhat \ndifferent depending on ML model and street (see figures in Appendix E). There are, however, some typical features that tend \nto be more important. For PM10 Julian day, lagged measurements and deterministic forecasts are mostly among the top 5 most \nimportant features. For NOx deterministic forecasts, hour of the day and weekday are the most important, while lagged \n15 to be more important. For PM10 Julian day, lagged measurements and deterministic forecasts are mostly among the top 5 most \nimportant features. For NOx deterministic forecasts, hour of the day and weekday are the most important, while lagged \n15 \nmeasurements are less useful for the ML models. The importance of calendar data for NOx likely reflects importance of diurnal \nand weekday variations in traffic emissions not correctly captured by the deterministic forecast. Julian day likely reflects \nseasonal variations in non-exhaust emissions of PM10. Even though there are variations it is difficult see any systematic \ndifference in the features between ML for the different street sites. 20 20 3.2.2 \nComparison between deterministic forecasts and MLs - street canyon sites Comparisons between the hourly temporal variations in observations and forecasts of NOx with the GAM model in September \n2022 are shown in Figure 8 and for all models in Appendix F. One can see that the deterministic forecast tend to overestimate \nconcentrations of NOx during daytime especially for Sveavägen and this is corrected when ML modelling is being applied. 18 Corresponding plots for PM10 are shown in Figure 9. In this case the GAM overestimates concentrations on Folkungagatan \nand Hornsgatan during the end of September, but performs well otherwise, whereas the deterministic forecast overestimates \nPM10 on Sveavägen and Hornsgatan during the first half of the month. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Corresponding plots for PM10 are shown in Figure 9. In this case the GAM overestimates concentrations on Folkungagatan \nand Hornsgatan during the end of September, but performs well otherwise, whereas the deterministic forecast overestimates \nPM10 on Sveavägen and Hornsgatan during the first half of the month. Figure 8. Temporal variations in hourly mean NOx concentrations at the street canyon sites during September 2022 based\nobservations (red) and 1-day forecasts based on deterministic modelling (blue) and GAM (green). 5 Figure 8. Temporal variations in hourly mean NOx concentrations at the street canyon sites during September 2022 based on \nobservations (red) and 1-day forecasts based on deterministic modelling (blue) and GAM (green). 19 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure 9. Temporal variations in hourly mean PM10 concentrations at the street canyon sites during September 2022 based \nobservations, deterministic modelling and GAM forecasts. Mean of 1-, 2- and 3-day forecasts. Figure 9. Temporal variations in hourly mean PM10 concentrations at the street canyon sites during September 2022 based on \nobservations, deterministic modelling and GAM forecasts. Mean of 1-, 2- and 3-day forecasts. The improvement of the temporal variations of NOx and PM10 is well illustrated by comparing the mean diurnal variations in \n5 \nobservations with deterministic modelling and using the MLs, GAM shown in Figure 10 and all models shown in figures in \nAppendix G. The improvement of the temporal variations of NOx and PM10 is well illustrated by comparing the mean diurnal variations in \n5 \nobservations with deterministic modelling and using the MLs, GAM shown in Figure 10 and all models shown in figures in \nAppendix G. For all street sites, both NOx and PM10 concentrations shows systematic deviations from observations using \ndeterministic modelling, but this is corrected using the MLs, especially for NOx. The tendency that the LSTM model is not as \ngood to capture variations in PM10 at the urban site is also seen here for the street canyon sites. 3.2.2 \nComparison between deterministic forecasts and MLs - street canyon sites For all street sites, both NOx and PM10 concentrations shows systematic deviations from observations using \ndeterministic modelling, but this is corrected using the MLs, especially for NOx. The tendency that the LSTM model is not as \ngood to capture variations in PM10 at the urban site is also seen here for the street canyon sites. The improvement of the temporal variations of NOx and PM10 is well illustrated by comparing the mean diurnal variations in \n5 \nobservations with deterministic modelling and using the MLs, GAM shown in Figure 10 and all models shown in figures in \nAppendix G. For all street sites, both NOx and PM10 concentrations shows systematic deviations from observations using \ndeterministic modelling, but this is corrected using the MLs, especially for NOx. The tendency that the LSTM model is not as \ngood to capture variations in PM10 at the urban site is also seen here for the street canyon sites. 10 20 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure 10. Mean diurnal variations in hourly mean observations, deterministic and GAM forecasts of NOx and PM10 for the street \ncanyon sites. Mean of 1-, 2- and 3-day forecasts. Figure 10. Mean diurnal variations in hourly mean observations, deterministic and GAM forecasts of NOx and PM10 for the street \ncanyon sites. Mean of 1-, 2- and 3-day forecasts. 5 For all streets statistical performance of NOx forecasts are improved using the MLs as shown for all forecasts in Table 3. The \nimprovement in terms of Pearson correlation (r), MAPE, nRMSE and nMAE is very similar for the MLs but differ between \nstreets, with forecasts for Hornsgatan showing higher r and lower relative errors compared to the other streets. 0 For all streets statistical performance of NOx forecasts are improved using the MLs as shown for all forecasts in Table 3. The \nimprovement in terms of Pearson correlation (r), MAPE, nRMSE and nMAE is very similar for the MLs but differ between \nstreets, with forecasts for Hornsgatan showing higher r and lower relative errors compared to the other streets. 0 For all streets statistical performance of NOx forecasts are improved using the MLs as shown for all forecasts in Table 3. Figure 11 clearly illustrates the improvements of all statistical performance indexes for NOx at all street canyon sites and for \n5 \nMLs. The errors (MAPE, nRMSE, nMAE) are reduced by between 30% and 60% and the Pearson correlation coefficients \nincrease by between 30% and 50%. 3.2.2 \nComparison between deterministic forecasts and MLs - street canyon sites The \nimprovement in terms of Pearson correlation (r), MAPE, nRMSE and nMAE is very similar for the MLs but differ between \nstreets, with forecasts for Hornsgatan showing higher r and lower relative errors compared to the other streets. 10 21 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Table 3. Comparison of 1-, 2-, 3-day deterministic and ML forecasts for NOx for the street canyon sites. r = Pearson correlation, \nMAPE = mean absolute percentage error, nRMSE = normalised rootmean square error and nMAE = normalised mean absolute \nerror. All data are based on hourly mean values. Folkungagatan \n \nr \nMAPE \nnRMSE \nnMAE \n \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \nDet \n0.48 \n0.49 \n0.47 \n107% \n118% \n120% \n108% \n109% \n106% \n72% \n73% \n73% \nXGB \n0.65 \n0.64 \n0.63 \n67% \n73% \n76% \n74% \n75% \n75% \n47% \n50% \n50% \nRF \n0.66 \n0.65 \n0.65 \n64% \n73% \n81% \n71% \n74% \n77% \n45% \n49% \n53% \nLSTM \n0.64 \n0.61 \n0.62 \n65% \n60% \n79% \n72% \n74% \n74% \n46% \n46% \n50% \nGAM \n0.66 \n0.65 \n0.65 \n65% \n75% \n81% \n73% \n75% \n77% \n46% \n51% \n53% \nSveavägen \n \nr \nMAPE \nnRMSE \nnMAE \n \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \nDet \n0.46 \n0.53 \n0.44 \n159% \n161% \n163% \n137% \n136% \n134% \n99% \n98% \n97% \nXGB \n0.69 \n0.68 \n0.66 \n59% \n57% \n59% \n68% \n69% \n71% \n41% \n41% \n41% \nRF \n0.73 \n0.73 \n0.73 \n51% \n51% \n50% \n65% \n65% \n65% \n37% \n38% \n37% \nLSTM \n0.71 \n0.69 \n0.66 \n58% \n60% \n64% \n68% \n69% \n71% \n41% \n41% \n43% \nGAM \n0.72 \n0.71 \n0.71 \n52% \n51% \n49% \n65% \n67% \n66% \n38% \n39% \n37% \nHornsgatan \n \nr \nMAPE \nnRMSE \nnMAE \n \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \nDet \n0.53 \n0.56 \n0.55 \n80% \n69% \n73% \n82% \n79% \n80% \n55% \n52% \n54% \nXGB \n0.80 \n0.81 \n0.81 \n45% \n45% \n44% \n52% \n51% \n50% \n32% \n32% \n32% \nRF \n0.79 \n0.79 \n0.81 \n42% \n43% \n43% \n52% \n53% \n50% \n31% \n32% \n31% \nLSTM \n0.77 \n0.76 \n0.76 \n48% \n51% \n51% \n57% \n57% \n56% \n36% \n36% \n36% \nGAM \n0.80 \n0.80 \n0.82 \n42% \n43% \n43% \n51% \n51% \n50% \n31% \n32% \n31% \n \nFigure 11 clearly illustrates the improvements of all statistical performance indexes for NOx at all street canyon sites and for \nMLs. The errors (MAPE, nRMSE, nMAE) are reduced by between 30% and 60% and the Pearson correlation coefficients \nincrease by between 30% and 50%. Table 3. 3.2.2 \nComparison between deterministic forecasts and MLs - street canyon sites Comparison of 1-, 2-, 3-day deterministic and ML forecasts for NOx for the street canyon sites. r = Pearson correlation, \nMAPE = mean absolute percentage error, nRMSE = normalised rootmean square error and nMAE = normalised mean absolute \nerror. All data are based on hourly mean values. Figure 11 clearly illustrates the improvements of all statistical performance indexes for NOx at all street canyon sites and for \n5 \nMLs. The errors (MAPE, nRMSE, nMAE) are reduced by between 30% and 60% and the Pearson correlation coefficients \nincrease by between 30% and 50%. 22 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. igure 11. Ratios of statistical performances for MLs versus the deterministic hourly forecasts for NOx at the stree\nMean of 1-day, 2-day and 3-day forecasts. igure 11. Ratios of statistical performances for MLs versus the deterministic hourly forecasts for NOx at the street canyon sites. Mean of 1-day, 2-day and 3-day forecasts. Figure 11. Ratios of statistical performances for MLs versus the deterministic hourly forecasts for NOx at the street canyon sites. Mean of 1-day, 2-day and 3-day forecasts. l performances for MLs versus the deterministic hourly forecasts for NOx at the street canyon sites. ay forecasts. Comparison between the statistical performance measures for MLs and deterministic forecasts for PM10 shows somewhat \n5 \nvariable results depending on statistical measure, street and ML. Person r values increase slightly in most cases and the \nnormalised RMSE and MAE are lower for most MLs and streets, but not always, while MAPE often increase using the MLs \n(Table 4 and Figure 12). Errors measured as nRMSE decrease by between 10% and 30%, whereas errors measured as MAPE \nmostly increase slightly and nMAE is about unchanged. Pearson r increase at Folkungagatan for all MLs (10% - 30%) but \nshow somewhat varying results for Sveavägen and Hornsgatan. 10 23 Table 4. Comparison of 1-, 2-, 3-day deterministic and ML forecasts for PM10 for the street canyon sites. r = Pearson correlation, \nMAPE = mean absolute percentage error, nRMSE = normalised rootmean square error and nMAE = normalised mean absolute \nerror. All data are based on hourly mean values. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Table 4. 3.2.2 \nComparison between deterministic forecasts and MLs - street canyon sites Comparison of 1-, 2-, 3-day deterministic and ML forecasts for PM10 for the street canyon sites. r = Pearson correlation, \nMAPE = mean absolute percentage error, nRMSE = normalised rootmean square error and nMAE = normalised mean absolute \nerror. All data are based on hourly mean values. Folkungagatan \n \nr \nMAPE \nnRMSE \nnMAE \n \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \nDet \n0.32 \n0.30 \n0.34 \n121% \n112% \n119% \n115% \n116% \n115% \n56% \n57% \n56% \nXGB \n0.41 \n0.30 \n0.34 \n122% \n134% \n121% \n85% \n102% \n83% \n52% \n63% \n54% \nRF \n0.36 \n0.39 \n0.41 \n134% \n121% \n129% \n89% \n82% \n75% \n52% \n52% \n49% \nLSTM \n0.47 \n0.43 \n0.34 \n102% \n115% \n141% \n82% \n77% \n83% \n58% \n53% \n58% \nGAM \n0.37 \n0.34 \n0.39 \n132% \n123% \n127% \n88% \n95% \n77% \n52% \n57% \n50% \nSveavägen \n \nr \nMAPE \nnRMSE \nnMAE \n \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \nDet \n0.42 \n0.40 \n0.40 \n98% \n100% \n95% \n92% \n92% \n92% \n55% \n56% \n54% \nXGB \n0.42 \n0.31 \n0.45 \n122% \n124% \n109% \n76% \n92% \n73% \n51% \n58% \n49% \nRF \n0.49 \n0.27 \n0.40 \n113% \n125% \n114% \n67% \n99% \n74% \n45% \n57% \n50% \nLSTM \n0.51 \n0.49 \n0.46 \n90% \n106% \n109% \n67% \n67% \n68% \n47% \n48% \n49% \nGAM \n0.45 \n0.28 \n0.41 \n115% \n121% \n111% \n71% \n93% \n75% \n46% \n56% \n49% \nHornsgatan \n \nr \nMAPE \nnRMSE \nnMAE \n \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \n1-day \n2-day \n3-day \nDet \n0.40 \n0.36 \n0.30 \n81% \n80% \n87% \n113% \n116% \n118% \n59% \n60% \n62% \nXGB \n0.46 \n0.30 \n0.37 \n84% \n103% \n91% \n89% \n110% \n89% \n56% \n67% \n59% \nRF \n0.42 \n0.21 \n0.33 \n85% \n115% \n94% \n91% \n130% \n90% \n57% \n73% \n59% \nLSTM \n0.49 \n0.40 \n0.34 \n77% \n84% \n93% \n82% \n85% \n89% \n56% \n59% \n64% \nGAM \n0.45 \n0.25 \n0.34 \n84% \n107% \n92% \n88% \n114% \n89% \n56% \n68% \n58% Table 4. Comparison of 1-, 2-, 3-day deterministic and ML forecasts for PM10 for the street canyon sites. r = Pearson correlation, \nMAPE = mean absolute percentage error, nRMSE = normalised rootmean square error and nMAE = normalised mean absolute \nerror. All data are based on hourly mean values. 5 24 12. Ratios of statistical performances for MLs versus the deterministic hourly forecasts for PM10 at the street can\nof 1-day, 2-day and 3-day forecasts. e 12. Ratios of statistical performances for MLs versus the deterministic hourly forecasts for PM10 at the street canyon sites. 2. 3.2.2 \nComparison between deterministic forecasts and MLs - street canyon sites Ratios of statistical performances for MLs versus the deterministic hourly forecasts for PM10 at the street canyon sit Figure 12. Ratios of statistical performances for MLs versus the deterministic hourly forecasts for PM10 at the street canyon sites. Mean of 1-day, 2-day and 3-day forecasts. l performances for MLs versus the deterministic hourly forecasts for PM10 at the street canyon sites. ay forecasts. As pointed out before it is important to assess statistical performance measures for periods with high concentrations. Similar \n5 \nto what is seen for the urban site the statistical performances for all models are much worse for the hourly mean concentrations \nthat are higher than the mean values and the pattern is also similar for the different streets. As pointed out before it is important to assess statistical performance measures for periods with high concentrations. Similar \n5 \nto what is seen for the urban site the statistical performances for all models are much worse for the hourly mean concentrations \nthat are higher than the mean values and the pattern is also similar for the different streets. 25 3 2 3\nMQI street canyon sites\nhttps://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 3.2.3 \nMQI street canyon sites Figure 13 illustrates that deviations between model results and measurements compared to the uncertainties of the \nmeasurements for all pollutants and street canyon sites. For NOx relative uncertainties decreases using the MLs compared to \nthe deterministic forecast, while for PM10 results varies, but there is no systematic improvement using MLs compared to the \ndeterministic model. 5 5 dete\nst c\node . 5\n \nFigure 13. MQI for NOx and PM10 forecasts at street canyon sites. Mean values for 1-, 2- and 3-day forecasts. Figure 13. MQI for NOx and PM10 forecasts at street canyon sites. Mean values for 1-, 2- and 3-day forecasts. Figure 13. MQI for NOx and PM10 forecasts at street canyon sites. Mean values for 1-, 2- and 3-day forecasts. 10 26 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 3.3 \nGeneralisation of street canyon modelling Until now, the model performance is evaluated using training and testing data from three single sites respectively. In Stockholm \nas well as in other cities most of the streets do not have any monitoring station. This is of course due to resource constraints \nbut also associated with the fact that the EU Air Quality Directives regulates the number of monitoring sites required in a city \ndepending on the level of air pollution and number of inhabitants. The monitoring stations should provide information for both \n5 \nareas where the highest concentrations of air pollutants occur and other areas that are representative of the exposure of the \ngeneral population. Less resources is required if this information can be achieved by accurate enough modelling. Until now, the model performance is evaluated using training and testing data from three single sites respectively. In Stockholm \nas well as in other cities most of the streets do not have any monitoring station. This is of course due to resource constraints \nbut also associated with the fact that the EU Air Quality Directives regulates the number of monitoring sites required in a city \ndepending on the level of air pollution and number of inhabitants. The monitoring stations should provide information for both \n5 \nareas where the highest concentrations of air pollutants occur and other areas that are representative of the exposure of the \ngeneral population. Less resources is required if this information can be achieved by accurate enough modelling. We therefore analyze the generalization capacities of the models, with the expectation that we can achieve certain prediction \nperformance of one site without having any measurement data. Computational experiments were carried out through cross-\nvalidation, which combines training and testing data coming from different measurement sites. For the street canyon sites, four \n10 \ncombinations of training datasets were applied to evaluate the generalization abilities of different ML models. We therefore analyze the generalization capacities of the models, with the expectation that we can achieve certain prediction \nperformance of one site without having any measurement data. Computational experiments were carried out through cross-\nvalidation, which combines training and testing data coming from different measurement sites. For the street canyon sites, four \n10 \ncombinations of training datasets were applied to evaluate the generalization abilities of different ML models. 3.3 \nGeneralisation of street canyon modelling We therefore analyze the generalization capacities of the models, with the expectation that we can achieve certain prediction \nperformance of one site without having any measurement data. Computational experiments were carried out through cross-\nvalidation, which combines training and testing data coming from different measurement sites. For the street canyon sites, four \n10 \ncombinations of training datasets were applied to evaluate the generalization abilities of different ML models. validation, which combines training and testing data coming from different measurement sites. For the street canyon sites, four \n10 \ncombinations of training datasets were applied to evaluate the generalization abilities of different ML models. Figure 14 shows mean of 1-day, 2-day, and 3-day forecasted NOx concentrations for the three street canyon sites based on \ntraining the models on the other streets. It shows that the forecast is improved compared to the deterministic forecast for g\ny,\ny,\ny\nx\ny\ntraining the models on the other streets. It shows that the forecast is improved compared to the deterministic forecast for \nHornsgatan and Sveavägen, but not so much for Folkungagatan. For Hornsgatan the correlation is 0.55 using the deterministic \n15 \nforecast whereas the MLs gives correlations between 0.61 and 0.67 and all errors decrease slightly using the MLs. For \nSveavägen the correlation is 0.48 using the deterministic forecast whereas the MLs gives correlations between 0.62 and 0.63 \nand here all errors decrease substantially using the MLs. But for Folkungagatan the MLs show different results. Correlations \nare similar or even decreases, whereas errors mostly decreases depending on ML applied. Hornsgatan and Sveavägen, but not so much for Folkungagatan. For Hornsgatan the correlation is 0.55 using the deterministic \n15 \nforecast whereas the MLs gives correlations between 0.61 and 0.67 and all errors decrease slightly using the MLs. For \nSveavägen the correlation is 0.48 using the deterministic forecast whereas the MLs gives correlations between 0.62 and 0.63 \nand here all errors decrease substantially using the MLs. But for Folkungagatan the MLs show different results. Correlations \nare similar or even decreases, whereas errors mostly decreases depending on ML applied. 20 27 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure 14. Statistical performances of NOx forecasts for the streets when the MLs are trained using only data from the other streets. Mean of 1-day, 2-day, and 3-day forecasts. Figure 14. 3.3 \nGeneralisation of street canyon modelling Statistical performances of NOx forecasts for the streets when the MLs are trained using only data from the other streets. Mean of 1-day, 2-day, and 3-day forecasts. ances of NOx forecasts for the streets when the MLs are trained using only data from the other streets. day forecasts. Figure 15 shows mean of 1-day, 2-day, and 3-day forecasted PM10 concentrations for the three street canyon sites based on \n5 \ntraining the models on the other streets. It can be seen that it is not possible to find any systematic improvement of the \ndeterministic forecast for the streets using RF and XGB compared to the deterministic forecasts. But with LSTM correlations \nincrease slightly and errors decrease at all streets compared to the deterministic forecasts. Figure 15 shows mean of 1-day, 2-day, and 3-day forecasted PM10 concentrations for the three street canyon sites based on \n5 \ntraining the models on the other streets. It can be seen that it is not possible to find any systematic improvement of the \ndeterministic forecast for the streets using RF and XGB compared to the deterministic forecasts. But with LSTM correlations \nincrease slightly and errors decrease at all streets compared to the deterministic forecasts. 28 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 5. Statistical performances of PM10 forecasts for the streets when the MLs are trained using only data from the other streets. 1-day, 2-day, and 3-day forecasts. Figure 15. Statistical performances of PM10 forecasts for the streets when the MLs are trained using only data from the other streets. Mean of 1-day, 2-day, and 3-day forecasts. 4 \nDiscussion \n5 For PM10 the seasonal variation described \nby Julian day is the most important feature at the street canyon sites, both for RF and XGB. This indicates that the deterministic forecasts is not capable at describing impacts of meteorology and road dust emissions on PM10, even though parameterisations \n10 \nof these processes are included in the deterministic modelling system. The total mass generated by road wear is a key factor \nfor PM10 emissions and these emissions are strongly controlled by surface moisture conditions and this is taken into account \nby the NORTRIP model. But as pointed out by Denby et al (2013b) there are periods where surface wetness is not well \nmodelled and it is not known if this is the result of input data, e.g. precipitation, or of the model formulation itself. forecasts is not capable at describing impacts of meteorology and road dust emissions on PM10, even though parameterisations \n10 \nof these processes are included in the deterministic modelling system. The total mass generated by road wear is a key factor \nfor PM10 emissions and these emissions are strongly controlled by surface moisture conditions and this is taken into account \nby the NORTRIP model. But as pointed out by Denby et al (2013b) there are periods where surface wetness is not well \nmodelled and it is not known if this is the result of input data, e.g. precipitation, or of the model formulation itself. forecasts is not capable at describing impacts of meteorology and road dust emissions on PM10, even though parameterisations \n10 \nof these processes are included in the deterministic modelling system. The total mass generated by road wear is a key factor \nfor PM10 emissions and these emissions are strongly controlled by surface moisture conditions and this is taken into account \nby the NORTRIP model. But as pointed out by Denby et al (2013b) there are periods where surface wetness is not well \nmodelled and it is not known if this is the result of input data, e.g. precipitation, or of the model formulation itself. It is clear that the deterministic forecast of O3 underestimates concentrations at the urban site due to the fact that the local \n15 \nemissions of NOx influencing the photochemistry is not properly considered by the CAMS model, but this is corrected using \nthe MLs. 4 \nDiscussion \n5 Despite this the deterministic forecast is the most important feature for both RF and XGB but also lagged measured \nmean and maximum O3 concentrations improve the deterministic forecasts. Despite the fact that the configurations and traffic situations are quite similar for the street canyon sites the improvements of It is clear that the deterministic forecast of O3 underestimates concentrations at the urban site due to the fact that the local \n15 \nemissions of NOx influencing the photochemistry is not properly considered by the CAMS model, but this is corrected using \nthe MLs. Despite this the deterministic forecast is the most important feature for both RF and XGB but also lagged measured \nmean and maximum O3 concentrations improve the deterministic forecasts. Despite the fact that the configurations and traffic situations are quite similar for the street canyon sites the improvements of It is clear that the deterministic forecast of O3 underestimates concentrations at the urban site due to the fact that the local \n15 \nemissions of NOx influencing the photochemistry is not properly considered by the CAMS model, but this is corrected using \nthe MLs. Despite this the deterministic forecast is the most important feature for both RF and XGB but also lagged measured \nmean and maximum O3 concentrations improve the deterministic forecasts. Despite the fact that the configurations and traffic situations are quite similar for the street canyon sites, the improvements of the deterministic forecasts using ML differs. For NOx forecasts on Hornsgatan are more accurate (lower errors and higher r) \n20 \nthan for the other two sites, while for PM10 there is no obvious difference between the sites. The overall model quality according to the recommendations by the Forum for Air Quality Modeling (FAIRMODE) in the \ncontext of the air quality directives, is improved using the MLs resulting in uncertainties that are significantly smaller than the \nmeasurement uncertainties for all pollutants. But the forecasts of the highest concentrations including episodes with high \nconcentrations is not systematically improved for all pollutants and all statistical performance measures using the MLs\n25 than for the other two sites, while for PM10 there is no obvious difference between the sites. The overall model quality according to the recommendations by the Forum for Air Quality Modeling (FAIRMODE) in the \ncontext of the air quality directives, is improved using the MLs resulting in uncertainties that are significantly smaller than the \nmeasurement uncertainties for all pollutants. 4 \nDiscussion \n5 The performance of the MLs are quite similar for the different sites and forecast days. But there are large differences in \nimprovements for different pollutants. In general, our results indicate that MLs are more effective in improving NOx than PM10 \nand O3. For PM10 the MLs show slight improvement in r but not much improvements in relative errors. This difference in 29 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. improvement is likely associated with the different processes controlling the concentrations, such as different sources: NOx \nconcentrations being mainly due to vehicle exhaust emissions which shows regular variations from one day to the next \ndepending on day of the week and time of day, while PM10 is mainly due to road dust emissions controlled by a combination \nof variations in vehicle volumes and meteorological conditions that affect suspension of coarse particles from street surfaces \n(e g Denby et al., 2013a; Johansson et al., 2007; Krecl et al., 2021). Road dust is accumulated on the road surfaces during wet \n5 \nroad surface conditions and suspended by vehicle induced turbulence during dry conditions (Denby et al 2013a) 5 (e g Denby et al., 2013a; Johansson et al., 2007; Krecl et al., 2021). Road dust is accumulated on the road surfaces during wet \n5 \nroad surface conditions and suspended by vehicle induced turbulence during dry conditions (Denby et al., 2013a). The improvement of the forecasts of NOx with ML is partly driven by the calendar, hour, day of the week and to some degree \nalso Julian day, but different features appear as important for RF compared to XGB. For PM10 the seasonal variation described \nby Julian day is the most important feature at the street canyon sites, both for RF and XGB. This indicates that the deterministic (e g Denby et al., 2013a; Johansson et al., 2007; Krecl et al., 2021). Road dust is accumulated on the road surfaces during wet \n5 \nroad surface conditions and suspended by vehicle induced turbulence during dry conditions (Denby et al., 2013a). The improvement of the forecasts of NOx with ML is partly driven by the calendar, hour, day of the week and to some degree \nalso Julian day, but different features appear as important for RF compared to XGB. 4.1 \nComparison of different MLs This makes it hard to draw general conclusions regarding which model to \nuse. However, we find that other factors may be more important to consider than type of model – such as sources of pollutants and influence of photochemistry, characteristic of the site resulting in different features being of varying importance depending \n15 \non pollutant type of location. In this context RF and XGB can provide useful output on the importance of features that is not \npossible using LSTM. Another more practical aspect to consider when comparing the MLs is the complexity and computer resources required for \ntraining the models. In AQ literature, deep learning models such as standard LSTM and other Recurrent Neural Networks and influence of photochemistry, characteristic of the site resulting in different features being of varying importance depending \n15 \non pollutant type of location. In this context RF and XGB can provide useful output on the importance of features that is not \npossible using LSTM. Another more practical aspect to consider when comparing the MLs is the complexity and computer resources required for \ntraining the models. In AQ literature, deep learning models such as standard LSTM and other Recurrent Neural Networks (RNNs) have been explored for their prediction capacities. However, most of the studies have adopted complex neural network \n20 \nstructures, such as models of multiple outputs that mainly give convenience for data processing and automated feature \nhandling. Nevertheless, training even a simple LSTM model is computationally much more expensive than the two \nconventional machine learning models, i.e. the decision tree based models (RF and XGB) in our case. In fact, we have to resort \nto the high performance machine (The Swedish Berzelius High-performance Computer) to reduce the computational time. For the current practice in our real air quality prediction system, we prefer the two tree-based models over LSTM. But this doesn’t \n25 \ndeny the possibility that well-designed deep learning models may replace the conventional machine learning models being \nadopted in the AQ system in near future, especially when the amount of data increases. 4 \nDiscussion \n5 But the forecasts of the highest concentrations including episodes with high \nconcentrations, is not systematically improved for all pollutants and all statistical performance measures using the MLs. 25 concentrations, is not systematically improved for all pollutants and all statistical performance measures using the MLs. 25 \nWe have shown that the statistical performances of the deterministic forecasts for concentrations of NOx at the street canyon \nsites can be improved using the MLs. But for PM10 only LSTM showed systematic improvements at all sites. So again this \naccentuates the importance of testing the models not only for one pollutant. Further work is needed to improve deterministic \nforecasts of PM10 based on the training of MLs at a few monitoring stations. As discussed above the situation in Stockholm is We have shown that the statistical performances of the deterministic forecasts for concentrations of NOx at the street canyon \nsites can be improved using the MLs. But for PM10 only LSTM showed systematic improvements at all sites. So again this \naccentuates the importance of testing the models not only for one pollutant. Further work is needed to improve deterministic \nforecasts of PM10 based on the training of MLs at a few monitoring stations. As discussed above the situation in Stockholm is \ndifferent from cities in central and southern Europe since the road dust contribution is very large. It might be that results for \n30 \nPM10 is different in other cities, but we have not found any publication on this matter. different from cities in central and southern Europe since the road dust contribution is very large. It might be that results for \n30 \nPM10 is different in other cities, but we have not found any publication on this matter. 30 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 4.1 \nComparison of different MLs Several studies have compared performance of different machine learners for predicting air quality (Zaini et al., 2021). Assessing forecasts of PM10 and PM2.5 concentrations, Czernecki et al. (2021) found that XGB performed the best, followed \nby random forests and an artificial neural network model, while stepwise regression performed the worst in four Polish \nagglomerations. Likewise, Joharestani et al. (2019) found XGB to performed best of three MLs (XGB, RF and a deep learning \n5 \nalgorithm), in predicting PM2.5 in Tehran (Iran). On the contrary, LSTM was shown to outperform XGBoost for forecasting \nhourly PM2.5 concentrations (Qadeer et al., 2020), similar to what was shown by Chuluunsaikhan et al (2021). Cai et al. (2009) \nobtained more accurate predictions of CO concentrations using artificial neural network modelling compared to using multiple \nlinear regression and the deterministic California line source dispersion model. On the other hand Shaban et al. (2015) 10 concluded that a tree based algorithm (M5P) outperformed artificial neural network modelling when comparing forecasts of \n10 \ndifferent pollutants in Qatar. There may be many reasons for the different results presented in the literature, including different \ntypes of input data, different atmospheric conditions and source contributions governing the concentrations. Also different \nstatistical measures of performance has been used. This makes it hard to draw general conclusions regarding which model to \nuse. However, we find that other factors may be more important to consider than type of model – such as sources of pollutants concluded that a tree based algorithm (M5P) outperformed artificial neural network modelling when comparing forecasts of \n10 \ndifferent pollutants in Qatar. There may be many reasons for the different results presented in the literature, including different \ntypes of input data, different atmospheric conditions and source contributions governing the concentrations. Also different \nstatistical measures of performance has been used. This makes it hard to draw general conclusions regarding which model to \nuse. However, we find that other factors may be more important to consider than type of model – such as sources of pollutants concluded that a tree based algorithm (M5P) outperformed artificial neural network modelling when comparing forecasts of \n10 \ndifferent pollutants in Qatar. There may be many reasons for the different results presented in the literature, including different \ntypes of input data, different atmospheric conditions and source contributions governing the concentrations. Also different \nstatistical measures of performance has been used. We have applied different machine learning algorithms to improve 1-, 2- and 3-day deterministic forecasts of NOx, PM10 and \n30 \nO3 concentrations for a number of locations in Stockholm, Sweden. It is shown that degree of improvement of deterministic \nforecasts depend more on pollutant and monitoring site than on what ML algorithm is applied. Deterministic forecasts of NOx 5 \nConclusions We have applied different machine learning algorithms to improve 1-, 2- and 3-day deterministic forecasts of NOx, PM10 and \n30 \nO3 concentrations for a number of locations in Stockholm, Sweden. It is shown that degree of improvement of deterministic \nforecasts depend more on pollutant and monitoring site than on what ML algorithm is applied. Deterministic forecasts of NOx We have applied different machine learning algorithms to improve 1-, 2- and 3-day deterministic forecasts of NOx, PM10 and \n30 \nO3 concentrations for a number of locations in Stockholm, Sweden. It is shown that degree of improvement of deterministic \nforecasts depend more on pollutant and monitoring site than on what ML algorithm is applied. Deterministic forecasts of NOx 31 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. are improved at all sites, using all models. Pearson correlations increase by up to 80% and errors are reduced by up to 60%. For PM10 more variable results are seen likely reflection the more complicated processes controlling the road wear emissions \nwhich constitute a large fraction of PM10. For O3 at the urban background site deviation between deterministically modelled \nabsolute level is correct using the MLs, nRMSE and nMAE is reduced by on average around 20%, but there is almost no \nimprovement in the correlation and MAPE. 5 5 We have shown that it is possible to improve deterministic forecasts of NOx at street canyon sites, based on training MLs at \nother sites. But when tested for PM10 only LSTM showed some improvements of the statistical performances compared to the \ndeterministic forecast of PM10. We have shown that it is possible to improve deterministic forecasts of NOx at street canyon sites, based on training MLs at \nother sites. But when tested for PM10 only LSTM showed some improvements of the statistical performances compared to the \ndeterministic forecast of PM10. A strength of our study is that we compare forecasts based on several pollutants and base A strength of our study is that we compare forecasts based on several pollutants and base our forecasts on a combination of \ndeterministic models (that are based on the underlying physicochemical mechanisms responsible for the emissions and \n10 \ndispersion of the pollutants) and 3 different machine learning algorithms with additional variables such as measurement data, \ncalendar data and meteorological data. 5 \nConclusions And this method is evaluated at different sites and for different pollutants during several \nmonths with different meteorological conditions. deterministic models (that are based on the underlying physicochemical mechanisms responsible for the emissions and \n10 \ndispersion of the pollutants) and 3 different machine learning algorithms with additional variables such as measurement data, \ncalendar data and meteorological data. And this method is evaluated at different sites and for different pollutants during several \nmonths with different meteorological conditions. There is still room for improvements of this work like e g fine tuning of the models, including and excluding features, \nexpanding to other sites and making maps of spatial concentrations over a larger area. 15 There is still room for improvements of this work like e g fine tuning of the models, including and excluding features, \nexpanding to other sites and making maps of spatial concentrations over a larger area. 15 32 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 6 \nAppendix A. Description of measurement methods and sites. All measurement methods are approved for monitoring according to the EU air quality directive for NOx, O3 and PM10. PM10 \nwas measured either using an optical particle counter (Hornsgatan: OPC, Grimm EDM 180-MC) or Tapered Element \nOscillating Microbalance (Sveavägen, Folkungagatan and Urban: TEOM model, 1400AB, Rupprecht & Patashnik, Co). NOx \nwas measured using chemiluminescence (AC32M, Environnement S.A.) and O3 was measured by UV absorption (O342M, \n5 \nEnvironnement S.A.). Table A1. Description of monitoring sites. Site name \nDescription \nTraffic volume \nPhoto \nHornsgatan \nStreet canyon site. Measurements of \nNOx and PM10 on north side of street, \n3 m above ground. Street width 24 m \nand building height 24 m. 23 000 veh/day (4% \nheavy duty vehicles). Vehicle composition \nmeasured during 4 week \ncampaigns using \nautomatic number plate \nrecognition. Sveavägen \nStreet canyon site. Measurements of \nNOx, PM10 on west side of street, 3 m \nabove ground. Street width 33 m and \nbuilding height 24 m. 21 000 veh/day (7% \nheavy duty vehicles). Table A1. Description of monitoring sites. Photo 33 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Folkungagatan \nStreet canyon site. Measurements \nNOx, PM10 on west side of street, 3 m \nabove ground. Street width 24 m and \nbuilding height 24 m. 12 000 veh/day (18% \nheavy duty vehicles). Torkel \nKnutssongatan \nUrban background. Measurements of \nNOx, PM10, ozone and meteorology \non top of a 20 m high building. Ca 13 000 vehicles on \nHornsgatan road 250 m N \nof site. 34 34 7 \nAppendix B Importance of features – urban background \n \n \n \nFigure B1.Most important features for NOx forecasts using XGB and RF at the urban site \n \n \nFigure B2. Most important features for PM10 forecasts using XGB and RF at the urban site. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 7 \nAppendix B Importance of features – urban background \n \n \n \nFigure B1.Most important features for NOx forecasts using XGB and RF at the urban site 7 \nAppendix B Importance of features – urban background 7 \nAppendix B Importance of features – urban background Figure B1.Most important features for NOx forecasts using XGB and RF at the urban site \n \n \nFigure B2. Most important features for PM10 forecasts using XGB and RF at the urban site. 6 \nAppendix A. Description of measurement methods and sites. 5 Figure B1.Most important features for NOx forecasts using XGB and RF at the urban site Figure B1.Most important features for NOx forecasts using XGB and RF at the urban site \n \n \nFigure B2. Most important features for PM10 forecasts using XGB and RF at the urban site. 5 Figure B2. Most important features for PM10 forecasts using XGB and RF at the urban site. 5 35 Figure B3. Most important features for O3 forecasts using XGB and RF at the urban site. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure B3. Most important features for O3 forecasts using XGB and RF at the urban site. 36 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 8 \nAppendix C. Temporal variations in hourly mean O3, NOx and PM10 concentrations at the urban background \nFigure C1. Temporal variations of deterministic and ML forecasted NOx, PM10 and O3 concentrations together wit\ncorresponding measured concentrations at the urban background site for September 2021. Mean of 1-, 2- and 3-day forecast 8 \nAppendix C. Temporal variations in hourly mean O3, NOx and PM10 concentrations at the urban background 8 \nAppendix C. Temporal variations in hourly mean O3, NOx and PM10 concentrations at the urban background poral variations in hourly mean O3, NOx and PM10 concentrations at the urban background Figure C1. Temporal variations of deterministic and ML forecasted NOx, PM10 and O3 concentrations together with \ncorresponding measured concentrations at the urban background site for September 2021. Mean of 1-, 2- and 3-day forecasts. Figure C1. Temporal variations of deterministic and ML forecasted NOx, PM10 and O3 concentrations together with \ncorresponding measured concentrations at the urban background site for September 2021. Mean of 1-, 2- and 3-day forecasts. 5 37 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure C2. Mean diurnal variations in measured and forecasted concentrations of NOx, PM10 and O3 at the urban site. Mean of \n1-, 2- and 3-day forecasts for August – December 2021. Figure C2. Mean diurnal variations in measured and forecasted concentrations of NOx, PM10 and O3 at the urban site. 6 \nAppendix A. Description of measurement methods and sites. Mean of \n1-, 2- and 3-day forecasts for August – December 2021. 5 5 38 Appendix D. Statistical performance measures for forecasts higher than the hourly mean concentrations at the urban \nit\nhttps://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Appendix D. Statistical performance measures for forecasts higher than the hourly mean concentrations at the urban Appendix D. Statistical performance measures for forecasts higher than the hourly mean concentrations at the urban \nsite. pendix D. Statistical performance measures for forecasts higher than the hourly mean concentrations at the urb\n \nure D1. Statistical performance measures for concentrations of NOx, PM10 and O3 higher than the hourly mean value at t\nan site. Mean of 1-, 2- and 3-day forecasts. Figure D1. Statistical performance measures for concentrations of NOx, PM10 and O3 higher than the hourly mean value at the \nurban site. Mean of 1-, 2- and 3-day forecasts. 5 5 39 9 \nAppendix E. Importance of features – Street canyon sites \n \n \nFigure E1. Most important features (%) for PM10 forecasts using RF and XGB at Folkungagatan. Figure E2. Most important features (%) for PM10 forecasts using RF and XGB at Hornsgatan. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 9 \nAppendix E. Importance of features – Street canyon sites \n \n \nFigure E1. Most important features (%) for PM10 forecasts using RF and XGB at Folkungagatan. 9 \nAppendix E. Importance of features – Street canyon sites 9 \nAppendix E. Importance of features – Street canyon sites \n \n \nFigure E1. Most important features (%) for PM10 forecasts using RF and XGB at Folkungagatan. Figure E1. Most important features (%) for PM10 forecasts using RF and XGB at Folkungagatan. Figure E2. Most important features (%) for PM10 forecasts using RF and XGB at Hornsgatan. 5 Figure E1. Most important features (%) for PM10 forecasts using RF and XGB at Folkungagatan. st important features (%) for PM10 forecasts using RF and XGB at Folkungagatan. Figure E2. Most important features (%) for PM10 forecasts using RF and XGB at Hornsgatan. 5 40 Figure E3. Most important features (%) for PM10 forecasts using RF and XGB at Sveavägen. Figure E4. 6 \nAppendix A. Description of measurement methods and sites. Most important features (%) for NOx forecasts using RF and XGB at Folkungagatan. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure E3. Most important features (%) for PM10 forecasts using RF and XGB at Sveavägen. Figure E3. Most important features (%) for PM10 forecasts using RF and XGB at Sveavägen. Figure E3. Most important features (%) for PM10 forecasts using RF and XGB at Sveavägen. Figure E4. Most important features (%) for NOx forecasts using RF and XGB at Folkungagatan. g\np\n(\n)\n10\ng\ng\n \n \nFigure E4. Most important features (%) for NOx forecasts using RF and XGB at Folkungagatan. Figure E4. Most important features (%) for NOx forecasts using RF and XGB at Folkungagatan. Figure E4. Most important features (%) for NOx forecasts using RF and XGB at Folkungagatan. 5 41 Figure E5. Most important features (%) for NOx forecasts using RF and XGB at Sveavägen. Figure E6. Most important features (%) for NOx forecasts using RF and XGB at Hornsgatan. 5 \nhttps://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure E5. Most important features (%) for NOx forecasts using RF and XGB at Sveavägen. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure E5. Most important features (%) for NOx forecasts using RF and XGB at Sveavägen. Figure E6. Most important features (%) for NOx forecasts using RF and XGB at Hornsgatan. 5 Figure E6. Most important features (%) for NOx forecasts using RF and XGB at Hornsgatan. 5 42 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 10 \nAppendix F. Temporal variations in hourly mean O3, NOx and PM10 concentrations at the street canyon sites \nFigure F1. Temporal variations of hourly deterministic and ML forecasted NOx concentrations together with correspondin\nmeasured concentrations at street canyon sites for September 2021. Mean of 1-, 2- and 3-day forecasts. 10 \nAppendix F. Temporal variations in hourly mean O3, NOx and PM10 concentrations at the street canyon sites 10 \nAppendix F. 6 \nAppendix A. Description of measurement methods and sites. Temporal variations in hourly mean O3, NOx and PM10 concentrations at the street canyon sites emporal variations in hourly mean O3, NOx and PM10 concentrations at the street canyon site Figure F1. Temporal variations of hourly deterministic and ML forecasted NOx concentrations together with corresponding \nmeasured concentrations at street canyon sites for September 2021. Mean of 1-, 2- and 3-day forecasts. 43 Figure F2. Temporal variations of hourly deterministic and ML forecasted PM10 concentrations together with correspondin\nmeasured concentrations at the street canyon sites for September 2021. Mean of 1-, 2- and 3-day forecasts. Figure F2. Temporal variations of hourly deterministic and ML forecasted PM10 concentrations together with corresponding \nmeasured concentrations at the street canyon sites for September 2021. Mean of 1-, 2- and 3-day forecasts. 44 11 \nAppendix G. Mean diurnal variations in hourly mean observations, 1-day, 2-day and 3-day deterministic and ML \nf\nt\nf NO\nd PM\nf\nth\nt\nt\nit\nhttps://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 11 \nAppendix G. Mean diurnal variations in hourly mean observations, 1-day, 2-day and 3-day deterministic and ML \nforecasts of NOx and PM10 for the street canyon sites. Figure G1. Mean diurnal variations in measured and forecasted concentrations of NOx and PM10 at the street canyon sites. Mean of 1-, 2- and 3-day forecasts for August – December 2021.Shaded areas are 95% confidence intervals. Figure G1. Mean diurnal variations in measured and forecasted concentrations of NOx and PM10 at the street canyon sites. Mean of 1-, 2- and 3-day forecasts for August – December 2021.Shaded areas are 95% confidence intervals. 5 45 12 \nAppendix H. Statistical performance measures for forecasted hourly mean concentrations higher than the mean \nvalues at the street canyon sites. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. 12 \nAppendix H. Statistical performance measures for forecasted hourly mean concentrations higher than the mean \nvalues at the street canyon sites. 12 \nAppendix H. Statistical performance measures for forecasted hourly mean concentrations higher than the mean \nvalues at the street canyon sites. stical performance measures for forecasted hourly mean concentrations higher than the mea\n canyon sites. 6 \nAppendix A. Description of measurement methods and sites. 12 \nAppendix H. Statistical performance measures for forecasted hourly mean concentrations higher than the mean \nvalues at the street canyon sites. Figure H1. Statistical performance measures for forecasted NOx and PM10 hourly mean concentrations higher than the mean \nvalues at Hornsgatan. Mean of 1-, 2- and 3-day forecasts. 5 Figure H1. Statistical performance measures for forecasted NOx and PM10 hourly mean concentrations higher than the mean \nvalues at Hornsgatan. Mean of 1-, 2- and 3-day forecasts. 5 46 Figure H2. Statistical performance measures for forecasted NOx and PM10 hourly mean concentrations higher than the mean \nvalues at Folkungagatan. Mean of 1-, 2- and 3-day forecasts. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Figure H2. Statistical performance measures for forecasted NOx and PM10 hourly mean concentrations higher than the mean \nvalues at Folkungagatan. Mean of 1-, 2- and 3-day forecasts. Figure H2. Statistical performance measures for forecasted NOx and PM10 hourly mean concentrations higher than the mean \nvalues at Folkungagatan. Mean of 1-, 2- and 3-day forecasts. 5 5 47 igure H3. Statistical performance measures for forecasted NOx and PM10 hourly mean concentrations higher than the mean \nalues at Sveavägen. Mean of 1-, 2- and 3-day forecasts. ttps://doi.org/10.5194/acp-2023-38\nreprint. Discussion started: 8 February 2023\n⃝Author(s) 2023. CC BY 4.0 License. Figure H3. Statistical performance measures for forecasted NOx and PM10 hourly mean concentrations higher than the mean \nvalues at Sveavägen. Mean of 1-, 2- and 3-day forecasts. Figure H3. Statistical performance measures for forecasted NOx and PM10 hourly mean concentrations higher than the mean \nvalues at Sveavägen. Mean of 1-, 2- and 3-day forecasts. 5 Code/Data availability: Python codes and data are available here: https://zenodo.org/record/7576042#.Y9k3AXbMK71 . References Berkowicz, R.: OSPM - A parameterised street pollution model, Environmental Monitoring and Assessment, 65, 323-331, \n2020. Berkowicz, R.: OSPM - A parameterised street pollution model, Environmental Monitoring and Assessment, 65, 323-331, \n2020. Brokamp, C., Jandarov, R., Rao, M.B., LeMasters, G., Ryan, P.: Exposure assessment models f Brokamp, C., Jandarov, R., Rao, M.B., LeMasters, G., Ryan, P.: Exposure assessment models for elemental components of \nparticulate matter in an urban environment: A comparison of regression and random forest approaches. Atmos. Environ, 151, \n5 \n1–11, 2017. particulate matter in an urban environment: A comparison of regression and random forest approaches. Atmos. Environ, 151, \n5 \n1–11, 2017. Burman, L., Johansson, C.: Emissions and Concentrations of Nitrogen Oxides and Nitrogen Dioxide on Hornsgatan Street, \nEvaluation \nof \nTraffic \nMeasurements \nduring \nAutumn \n2009 \n(In \nSwedish \nOnly). SLB \nReport \n7. https://www.slb.nu/slb/rapporter/pdf8/slb2010_007.pdf, 2010. Burman, L., Johansson, C.: Emissions and Concentrations of Nitrogen Oxides and Nitrogen Dioxide on Hornsgatan Street, \nEvaluation \nof \nTraffic \nMeasurements \nduring \nAutumn \n2009 \n(In \nSwedish \nOnly). SLB \nReport \n7. https://www.slb.nu/slb/rapporter/pdf8/slb2010_007.pdf, 2010. Burman, \nL., \nElmgren, \nM., \nNorman, \nM.: \nFordonsmätningar \npå \nHornsgatan \når \n2017. 10 \nhttps://scholar.google.com/scholar_lookup?title=Fordonsm%C3%A4tningar%20P%C3%A5%20Hornsgatan%20%C3%85r\n%202017&author=L.%20Burman&publication_year=2019, 2019. Cai, M., Yin, Y., Xie, M.: Prediction of hourly air pollutant concentrations near urban arterials using artificial neural network \napproach. Transport Research Part D. 14, 32-41. doi:10.1016/j.trd.2008.10.004, 2009. Burman, \nL., \nElmgren, \nM., \nNorman, \nM.: \nFordonsmätningar \npå \nHornsgatan \når \n2017. 10 \nhttps://scholar.google.com/scholar_lookup?title=Fordonsm%C3%A4tningar%20P%C3%A5%20Hornsgatan%20%C3%85r\n%202017&author=L.%20Burman&publication_year=2019, 2019. %202017&author=L.%20Burman&publication_year=2019, 2019. Cai, M., Yin, Y., Xie, M.: Prediction of hourly air pollutant concentrations near urban arterials using artificial neural network \napproach. Transport Research Part D. 14, 32-41. doi:10.1016/j.trd.2008.10.004, 2009. Cai, M., Yin, Y., Xie, M.: Prediction of hourly air pollutant concentrations near urban arterials using artificial neural network \napproach. Transport Research Part D. 14, 32-41. doi:10.1016/j.trd.2008.10.004, 2009. Cai, M., Yin, Y., Xie, M.: Prediction of hourly air pollutant concentrations near urban arterials using artificial neural network \napproach. Transport Research Part D. 14, 32-41. doi:10.1016/j.trd.2008.10.004, 2009. Carslaw, D.C. and K. Ropkins.: Openair — an R package for air quality data analysis, Environmental Modelling & Software, \n15 \n27-28, 52–61, 2012. Carslaw, D.C. and K. Ropkins.: Openair — an R package for air quality data analysis, Environmental Modelling & Software, \n15 \n27-28, 52–61, 2012. Castelli, M., Clemente, F.M., Popovič, A., Silva, S. and Vanneschi, L.: A Machine Learning Approach to Predict Air Quality \nin California. Hindawi, Complexity, Article ID 8049504, 23 pages, https://doi.org/10.1155/2020/8049504, 2020. Python codes and data are available here: https://zenodo.org/record/7576042#.Y9k3AXbMK71 . Author contribution: ME has been responsible for the deterministic modelling and providing with monitoring data and \nmeteorological forecasts. ZZ and XM has been responsible for the ML modelling and statistical calculations. CJ, XM and ME \ninitiated and planned the project. All authors have contributed to analysing data and writing of the manuscript. 10 Author contribution: ME has been responsible for the deterministic modelling and providing with monitoring data and \nmeteorological forecasts. ZZ and XM has been responsible for the ML modelling and statistical calculations. CJ, XM and ME \ninitiated and planned the project. All authors have contributed to analysing data and writing of the manuscript. 10 Competing interests: The authors declare that they have no conflict of interest. Competing interests: The authors declare that they have no conflict of interest. Acknowledgements: Financial support: The project was funded by ICT – The next generation and Digital future at KTH \nRoyal Institute of Technology (contract VF 2021-0082). 15 Acknowledgements: Financial support: The project was funded by ICT – The next generation and Digital future at KTH \nRoyal Institute of Technology (contract VF 2021-0082). 15 48 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. References Chuluunsaikhan, T., Heak, M., Nasridinov, A., Choi, S.: Comparative Analysis of Predictive Models for Fine Particulate Castelli, M., Clemente, F.M., Popovič, A., Silva, S. and Vanneschi, L.: A Machine Learning Approach to Predict Air Quality \nin California. Hindawi, Complexity, Article ID 8049504, 23 pages, https://doi.org/10.1155/2020/8049504, 2020. Chuluunsaikhan, T., Heak, M., Nasridinov, A., Choi, S.: Comparative Analysis of Predictive Models for Fine Particulate \nMatter in Daejeon, South Korea. Atmosphere, 12, 1295. https://doi.org/10.3390/atmos12101295, 2021. 20 Chuluunsaikhan, T., Heak, M., Nasridinov, A., Choi, S.: Comparative Analysis of Predictive Models for Fine Particulate Matter in Daejeon, South Korea. Atmosphere, 12, 1295. https://doi.org/10.3390/atmos12101295, 2021. 20 \nCzernecki, B., Marosz, M., Jędruszkiewicz, J.: Assessment of Machine Learning Algorithms in Short-term Forecasting of \nPM10 and PM2.5 Concentrations in Selected Polish Agglomerations. Aerosol and Air Quality Research. 21, 200586, \nhttps://doi.org/10.4209/aaqr.200586, 2021. Denby, B. R., Sundvor, I., Johansson, C., Pirjola, L., Ketzel, M., Norman, M., Kupiainen, K., Gustafsson, M., Blomqvist, G., Czernecki, B., Marosz, M., Jędruszkiewicz, J.: Assessment of Machine Learning Algorithms in Short-term Forecasting of \nPM10 and PM2.5 Concentrations in Selected Polish Agglomerations. Aerosol and Air Quality Research. 21, 200586, \nhttps://doi.org/10.4209/aaqr.200586, 2021. https://doi.org/10.4209/aaqr.200586, 2021. Denby, B. R., Sundvor, I., Johansson, C., Pirjola, L., Ketzel, M., Norman, M., Kupiainen, K., Gustafsson, M., Blomqvist, G., Denby, B. R., Sundvor, I., Johansson, C., Pirjola, L., Ketzel, M., Norman, M., Kupiainen, K., Omstedt, G.: A coupled road dust and surface moisture model to predict non-exhaust road traffic induced particle emissions \n25 \n(NORTRIP). Part 1: road dust loading and suspension modelling Atmos. Environ., 77, 283-300, 2013a. Denby, B. R., Sundvor, I., Johansson, C., Pirjola, L., Ketzel, M., Norman, M., Kupiainen, K., Gustafsson, M., Blomqvist, G., \nOmstedt, G.: A coupled road dust and surface moisture model to predict non-exhaust road traffic induced particle emissions \n(NORTRIP). Part 2: surface moisture and salt impact modelling Atmos. Environ., 81, 485-503, 2013b. (NORTRIP). Part 1: road dust loading and suspension modelling Atmos. Environ., 77, 283 300, 2013a. Denby, B. R., Sundvor, I., Johansson, C., Pirjola, L., Ketzel, M., Norman, M., Kupiainen, K., Gustafsson, M., Blomqvist, G., \nOmstedt, G.: A coupled road dust and surface moisture model to predict non-exhaust road traffic induced particle emissions \n(NORTRIP). Part 2: surface moisture and salt impact modelling Atmos. Environ., 81, 485-503, 2013b. Di, Q., Amini, H., Shi, L., Kloog, I., Silvern, R., Kelly, J., Sabath, M.B., Choirat, C., Koutrakis, P., Lyapustin, A., Wang, Y., \n30 \nMickley, L.J., Schwartz. References J.: An ensemble-based model of PM2.5 concentration across the contiguous United States with high \nspatiotemporal resolution. Environment International 130, 104909, 2019. Doreswamy, Harishkumar K S., Yogesh, K.M., Gad, I.: Forecasting Air Pollution Particulate Matter (PM2.5) Using Machine spatiotemporal resolution. Environment International 130, 104909, 2019. Doreswamy, Harishkumar K S., Yogesh, K.M., Gad, I.: Forecasting Air Pollution Particulate Matter (PM2.5) Using Machine \nLearning Regression Models. Procedia Computer Science 171, 2057–2066, 2020. 49 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Engardt, M., Bergström. S. and Johansson, C.: Luften du andas - nu och de kommande dagarna. Utveckling av ett automatiskt \nprognossystem \nför \nluftföroreningar \noch \npollen. SLB \n36:2021, \n33 \npp. (In \nSwedish). https://www.slbanalys.se/slb/rapporter/pdf8/slb2021_036.pdf, 2021. Fuller\nR\nPhilip J Landrigan Kalpana Balakrishnan Glynda Bathan Stephan Bose-O’Reilly\nMichael Brauer\nJack Caravanos, Tom Chiles, Aaron Cohen, Lilian Corra, Maureen Cropper, Greg Ferraro, Jill Hanna, David Hanrahan, Howard \n5 \nHu, David Hunter, Gloria Janata, Rachael Kupka, Bruce Lanphear, Maureen Lichtveld, Keith Martin, Adetoun Mustapha, \nErnesto Sanchez-Triana, Karti Sandilya, Laura Schaefli, Joseph Shaw, Jessica Seddon, William Suk, Martha María Téllez-\nRojo, \nChonghuai \nYan.: \nPollution \nand \nhealth: \na \nprogress \nupdate. Lancet \nPlanet \nHealth, \n6, \ne535–47, \nhttps://doi.org/10.1016/S2542-5196(22)00090-0, 2022. Gidhagen, L., Johansson, C., Langner, J., Foltescu, V. L.: Urban scale modeling of particle number concentration in Stockholm. 10 \nAtmospheric Environment 39, 1711–1725, 2005. Hoek, G., Beelen, R.,, de Hoogh, K., Vienneau, D., Gulliver, J., Fischer, P., Briggs, D.: A review of land-use regression models \nto assess spatial variation of outdoor air pollution. Atmos Environ, 42, 7561-7568, doi:10.1016/j.atmosenv.2008.05.057, 2008. Horálek, J., Hamer, P., Schreiberová, M., Colette, A., Schneider, P., Malherbe, L.: Potential use of CAMS modelling results Gidhagen, L., Johansson, C., Langner, J., Foltescu, V. L.: Urban scale modeling of particle number concentration in Stockholm. 10 \nAtmospheric Environment 39, 1711–1725, 2005. Hoek, G., Beelen, R.,, de Hoogh, K., Vienneau, D., Gulliver, J., Fischer, P., Briggs, D.: A review of land-use regression models \nto assess spatial variation of outdoor air pollution. Atmos Environ, 42, 7561-7568, doi:10.1016/j.atmosenv.2008.05.057, 2008. Horálek, J., Hamer, P., Schreiberová, M., Colette, A., Schneider, P., Malherbe, L.: Potential use of CAMS modelling results \nin air quality mapping under ETC/ATNI. Eionet Report – ETC/ATNI 2019/17, ISBN 978-82-93752-21-9, 2019. 15 Horálek, J., Hamer, P., Schreiberová, M., Colette, A., Schneider, P., Malherbe, L.: Potential u in air quality mapping under ETC/ATNI. Eionet Report – ETC/ATNI 2019/17, ISBN 978-82-93752-21-9, 2019. 15 \nIskandaryan, D., Ramos, F. References and Trilles, S.: Air Quality Prediction in Smart Cities Using Machine Learning Technologies based \non Sensor Data: A Review. Appl. Sci. 2020, 10, 2401, doi: 10.3390/app10072401, 2020. Janssen, S. and Thunis, P.: FAIRMODE Guidance Document on Modelling Quality Objectives and Benchmarking (version \n3.3), EUR 31068 EN, Publications Office of the European Union, Luxembourg, ISBN 978-92-76-52425-0, doi:10.2760/41988, Iskandaryan, D., Ramos, F. and Trilles, S.: Air Quality Prediction in Smart Cities Using Machine Learning Technologies based \non Sensor Data: A Review. Appl. Sci. 2020, 10, 2401, doi: 10.3390/app10072401, 2020. on Sensor Data: A Review. Appl. Sci. 2020, 10, 2401, doi: 10.3390/app10072401, 2020. Janssen, S. and Thunis, P.: FAIRMODE Guidance Document on Modelling Quality Objectives and Benchmarking (version \n3.3), EUR 31068 EN, Publications Office of the European Union, Luxembourg, ISBN 978-92-76-52425-0, doi:10.2760/41988, \nJRC129254, 2022. 20 JRC129254, 2022. 20 \nJohansson, C., Norman, M., Gidhagen, L.: Spatial & temporal variations of PM10 and particle number concentrations in urban \nair. Environ. Monit. Assess. 127, 477–487, 2007. Johansson, C., Burman, L., Forsberg, B.: The effects of congestions tax on air quality and health. Atmos. Environ. 43, 4843-\n4854, 2009. Johansson, C., Norman, M., Gidhagen, L.: Spatial & temporal variations of PM10 and particle number concentrations in urban \nair. Environ. Monit. Assess. 127, 477–487, 2007. Johansson, C., Eneroth, K., Lövenheim, B., Silvergren, S., Burman, L., Bergström, S., Norman, M., Engström Nylén, A., \n25 \nHurkmans, J., Elmgren, M., Brydolf, M., Täppefur, M.: Luftkvalitetsberäkningar för kontroll av miljökvalitetsnormer (with \nsummary in English). SLB 11:2017 ver 2. https://www.slbanalys.se/slb/rapporter/pdf8/slb2017_011.pdf, 2017. Johansson, C. Lövenheim, B.; Schantz, P.; Wahlgren, L.; Almström, P.; Markstedt, A.; Strömgren, M.; Forsberg, B.; Nilsson \nSommar, J.: Impacts on air pollution and health by changing commuting from car to bicycle. Sci. Total Environ. 584-585, 55- Johansson, C. Lövenheim, B.; Schantz, P.; Wahlgren, L.; Almström, P.; Markstedt, A.; Strömgren, M.; Forsberg, B.; Nilsson \nSommar, J.: Impacts on air pollution and health by changing commuting from car to bicycle. Sci. Total Environ. 584-585, 55-\n63, 2017. 30 63, 2017. 30 \nJoharestani, M.Z., Cao, C., Ni, X., Bashir, B. Talebiesfandarani, S.: PM2.5 Prediction Based on Random Forest, XGBoost, \nand Deep Learning Using Multisource Remote Sensing Data. Atmosphere, 10, 373; doi:10.3390/atmos10070373, 2019. Kamińska, J. A.: A random forest partition model for predicting NO2 concentrations from traffic flow and meteorological \nconditions. Science of the Total Environment 651, 475–483, 2019. Kamińska, J. References A.: A random forest partition model for predicting NO2 concentrations from traffic flow and meteorological \nconditions. Science of the Total Environment 651, 475–483, 2019. 50 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Keller, M., Hausberger, S., Matzer, C., Wüthrich, P., Notter, B.: HBEFA 3.3. Update of NOx Emission Factors of Diesel \nPassenger Cars- Background Documentation. https://www.hbefa.net/e/documents/HBEFA33_Documentation_20170425.pdf, \n2017. Krecl, P., Harrison, R.M., Johansson, C., Targino, A.C., Beddows, D.C., Ellermann, T., Lara, C. and Ketzel, M.: Long-term \ntrends in nitrogen oxides concentrations and on-road vehicle emission factors in Copenhagen, London and Stockholm. 5 \nEnvironmental Pollution, 290, 118105, 2021. Krecl, P., Johansson, C., Targino, A.C.,, Ström, J., Burman, L.: Trends in black carbon and size-resolved particle number \nconcentrations and vehicle emission factors under real-world conditions, Atmospheric Environment, 165, 155-168, 2017. Marècal, V., Peuch, V.-H., Andersson, C., Andersson, S., Arteta, J., Beekmann, M., Benedictow, A., Bergström, R., Bessagnet, Krecl, P., Johansson, C., Targino, A.C.,, Ström, J., Burman, L.: Trends in black carbon and size-resolved particle number \nconcentrations and vehicle emission factors under real-world conditions, Atmospheric Environment, 165, 155-168, 2017. B., Cansado, A., Chèroux, F., Colette, A., Coman, A., Curier, R. L., Denier van der Gon, H. A. C., Drouin, A., Elbern, H., \n10 \nEmili, E., Engelen, R. J., Eskes, H. J., Foret, G., Friese, E.,Gauss, M., Giannaros, C., Guth, J., Joly, M., Jaumouillè, E., Josse, \nB., Kadygrov, N., Kaiser, J. W., Krajsek, K., Kuenen, J., Kumar, U., Liora, N., Lopez, E., Malherbe, L., Martinez, I., Melas, \nD., Meleux, F., Menut, L., Moinat, P., Morales, T., Parmentier, J., Piacentini, A., Plu, M., Poupkou, A., Queguiner, S., \nRobertson, L., Rouïl, L., Schaap, M., Segers, A., Sofiev, M., Tarasson, L., Thomas, M., Timmermans, R., Valdebenito, ¡., van Robertson, L., Rouïl, L., Schaap, M., Segers, A., Sofiev, M., Tarasson, L., Thomas, M., Timmerm Velthoven, P., van Versendaal, R.,Vira, J. and Ung, A.: A regional air quality forecasting system over Europe: the MACC-II \n15 \ndaily ensemble production. Geoscientific Model Development, Volume 8, issue 9, 2777–2813, 2015. Meteo-France for Copernicus.: Regional Production, Description of the operational models and of the ENSEMBLE system. Retrieved \n2018-11-20. Available \nat: \nhttps://atmosphere.copernicus.eu/sites/default/files/2018-\n02/CAMS50_factsheet_201610_v2.pdf, 2017. Munir, S., Mayfield, M., Coca, D., Mihaylova, L.S. and Osammor, O.: Analysis of Air Pollution in Urban Areas with Airviro \n20 \nDispersion \nModel—A \nCase \nStudy \nin \nthe \nCity \nof \nSheffield, \nUnited \nKingdom. References Atmosphere \n11, \n285; \ndoi:10.3390/atmos11030285, 2020. Olstrup, H., Johansson, C., Forsberg, B., Åström, C.: Association between Mortality and Short- Term Exposure to Particles, \nOzone and Nitrogen Dioxide in Stockholm, Sweden. Int J Environ Res Public Health, 16, 6, 1028-1042, 2019. Munir, S., Mayfield, M., Coca, D., Mihaylova, L.S. and Osammor, O.: Analysis of Air Pollution in Urban Areas with Airviro \n20 \nDispersion \nModel—A \nCase \nStudy \nin \nthe \nCity \nof \nSheffield, \nUnited \nKingdom. Atmosphere \n11, \n285; \ndoi:10.3390/atmos11030285, 2020. Olstrup, H., Johansson, C., Forsberg, B., Åström, C.: Association between Mortality and Short- Term Exposure to Particles, \nOzone and Nitrogen Dioxide in Stockholm, Sweden. Int J Environ Res Public Health, 16, 6, 1028-1042, 2019. Orru, H. Lövenheim, B. Johansson, C. Forsberg, B.: Estimated health impacts of changes in air pollution exposure associated \n25 \nwith the planned by-pass Förbifart Stockholm. J Expo Sci Environ Epidemiol, 1-8, 2015. Ottosen, T.-B. and Kakosimos, K. E. and Johansson, C. and Hertel, O. and Brandt, J. and Skov, H. and Berkowicz, R. and \nEllermann, T. and Jensen, S. S. and Ketzel, M.: Analysis of the impact of inhomogeneous emissions in the Operational Street \nPollution Model (OSPM). Geoscientific Model Development, 8, 3231—3245, 2015. Orru, H. Lövenheim, B. Johansson, C. Forsberg, B.: Estimated health impacts of changes in air pollution exposure associated \n25 \nwith the planned by-pass Förbifart Stockholm. J Expo Sci Environ Epidemiol, 1-8, 2015. Qadeer, K., Rehman, W.U., Sheri, A.M., Park, I., Kim, H.K. and Jeon, M.: A Long Short-Term Memory (LSTM) Network for \n30 \nHourly Estimation of PM2.5 Concentration in Two Cities of South Korea. Appl. Sci. 2020, 10, 3984, \ndoi:10.3390/app10113984, 2020. Rybarczyk, Y. and Zalakeviciute, R.: Machine Learning Approaches for Outdoor Air Quality Modelling: A Systematic \nReview. Appl. Sci. 2018, 8, 2570, doi:10.3390/app8122570, 2018. Rybarczyk, Y. and Zalakeviciute, R.: Machine Learning Approaches for Outdoor Air Quality Modelling: A Systematic \nReview. Appl. Sci. 2018, 8, 2570, doi:10.3390/app8122570, 2018. 51 https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. https://doi.org/10.5194/acp-2023-38\nPreprint. Discussion started: 8 February 2023\nc⃝Author(s) 2023. CC BY 4.0 License. Säll, B.: Evaluation and validation of Copernicus Atmosphere Monitoring Service regional ensemble forecast of air pollutants \nand birch pollen in the Stockholm region. Master thesis report 30 HP (MO9001). Department of Meteorology, Stockholm \nuniversity, 2018. Säll, B.: Evaluation and validation of Copernicus Atmosphere Monitoring Service regional ensemble forecast of air pollutants \nand birch pollen in the Stockholm region. References Master thesis report 30 HP (MO9001). Department of Meteorology, Stockholm \nuniversity, 2018. Shaban, K., B.,Kadri, A., and Rezk, E.: Urban Air Pollution Monitoring System With Forecasting Models. IEEE sensors \nJournal, 16, 2598-2606, 2016. Journal, 16, 2598-2606, 2016. 5 \nShtein, A., Kloog, I., Schwartz, J., Silibello, C., Michelozzi, P., Gariazzo, C., Viegi, G., Forastiere, F., Karnieli, A., Just, A.C. and Stafoggia, M.: Estimating Daily PM2.5 and PM10 over Italy Using an Ensemble Model. Environmental Science & \nTechnology 54, 120-128 DOI: 10.1021/acs.est.9b04279, 2020. SLB, Methods for calculating air pollution concentrations in relation to the limit values. Report in Swedish with summary in Shtein, A., Kloog, I., Schwartz, J., Silibello, C., Michelozzi, P., Gariazzo, C., Viegi, G., Forastiere, F., Karnieli, A., Just, A.C. and Stafoggia, M.: Estimating Daily PM2.5 and PM10 over Italy Using an Ensemble Model. Environmental Science & \nTechnology 54, 120-128 DOI: 10.1021/acs.est.9b04279, 2020. Technology 54, 120-128 DOI: 10.1021/acs.est.9b04279, 2020. SLB, Methods for calculating air pollution concentrations in relation to the limit values. Report in Swedish with summary in English. Environment and Health Administration of Stockholm, SLB analys, Box 8136, 104 20 Stockholm, Sweden, Report \n10 \nnr. 50:2021. https://www.slbanalys.se/slb/rapporter/pdf8/slb2021_050.pdf, accessed 30 November, 2022. Stafoggia, M., Johansson, C., Glantz, P., Renzi, M., Shtein, A., de Hoogh, K., Kloog, I., Davoli, M., Michelozzi, P., Bellander, \nT.: A Random Forest Approach to Estimate Daily Particulate Matter, Nitrogen Dioxide, and Ozone at Fine Spatial Resolution \nin Sweden. Atmosphere, 11, 239, 1-19, 2020. Stafoggia, M., Bellander, T., Bucci, S., Davoli, M., de Hoogh, K., de Donato, F., Gariazzo, C., Lyapustin, A., Michelozzi, P., \n15 \nRenzi, M., Scortichini, M., Shtein, A., Viegi, G., Kloog, I., Schwartz, J.: Estimation of daily PM10 and PM2.5 concentrations \nin Italy, 2013-2015, using a spatiotemporal land-use random-forest model. Environ. Int., 124, 170−179, 2019. Thongthammachart, T., Araki, S., Shimadera, H., Eto, S., Matsuo, T. and Kondo, A.: An integrated model combining random \nforests and WRF/CMAQ model for high accuracy spatiotemporal PM2.5 predictions in the Kansai region of Japan. Atmospheric Environment 262, 118620, 2021. 20 \nZaini, N., Ean, L.W., Ahmed, A.N., Malek, M.A.: A systematic literature review of deep learning neural network for time \nseries air quality forecasting. Environmental Science and Pollution Research, https://doi.org/10.1007/s11356-021-17442-1, \n2021. 52"