Assessing Air Quality and Related Health Impacts in Urban Semi-Enclosed Railway Stations: Source attribution, Machine learning approach, and Exposure assessment

Abstract

Air pollution in transport hubs presents a growing public health challenge due to concentrated emission sources and high commuter exposure. While airports and seaports have been extensively studied, urban railway stations, particularly those operating diesel-powered trains have received comparatively limited attention. This thesis aims to identify and attribute train-related pollutant sources within and near railway terminals and to examine how environmental factors and train operations influence pollutant levels within the terminal. Furthermore, the study estimates the associated health impacts on commuters and staff, as well as the potential economic savings associated with efficient train operations.

An initial study employed low-cost monitoring (LCM) devices for fixed-site monitoring near a railway station in Dublin, Ireland, to quantify ambient PM2.5 and NO2 concentrations and disaggregate contributions from railway operations and adjacent road traffic. Sensor data were calibrated using machine learning models, including Extreme Gradient Boosting (XGBoost) and Random Forest (RF) regression. XGBoost demonstrated superior performance (NO2: R2 = 0.80, RMSE = 9.1 μg/m3; PM2.5: R2 = 0.92, RMSE = 2.2 μg/m3), highlighting its suitability for LCM calibration. Model interpretation showed that local wind conditions, atmospheric pressure, PM2.5 concentrations, and road traffic intensity were the primary drivers of NO2 variability, while raw sensor signals dominated PM2.5 predictions. A one-month monitoring campaign revealed exceedances of the annual World Health Organization (WHO) guideline values by 1.6-1.8 times for PM2.5 and 3.2-5.2 times for NO2. Wind-direction-based source apportionment identified the railway station as the main PM2.5 source and road traffic as the dominant NO2 contributor.

Following this, an examination of air quality at the same fixed monitoring site was undertaken over a longer monitoring period (June 2024-January 2025) using a statistical modelling approach. Generalised Additive Model (GAM) results demonstrated good model performance for PM2.5 (R2 = 0.81) and moderate performance for NO2 (R2 = 0.73). For NO2, background concentrations, wind direction, and wind speed emerged as the strongest drivers of variability. For PM2.5, background PM2.5 was the most influential variable, followed by meteorological factors such as temperature, humidity, and wind conditions, highlighting the important role of regional background pollution in influencing local concentrations. To examine the sources of pollution within a semi-enclosed railway station environment, an investigation of continuous fixed monitoring was undertaken, to correlate to operational train data. Fixed monitoring showed 24-hour average PM2.5 concentrations of 3.18-32.56 μg/m3 and 5.6-42.4 μg/m3, respectively, with Diesel multiple unit (DMU) trains dominated platforms exceeding WHO PM2.5 guideline values during approximately 90% of the monitoring period. NO2 24-hour averages ranged from 135.6 to 169.8 μg/m3, with 11-15% of hourly values exceeding the EU hourly guideline value. An XGBoost model was used to explain NO2 variability achieved a moderate predictive performance (R2 = 0.61; RMSE = 15.12 μg/m3) and identified ambient temperature and average train idling time as the most influential predictors. Partial dependence analysis indicated that reducing average idling time by two minutes could lower NO2 concentrations by approximately 2.5 μg/m3 per train. Operational assessments suggested that shutting down engines when dwell times exceed 30 minutes, restarting 15 minutes prior to departure, and maintaining idle speeds below 600 RPM could reduce fuel consumption by 40-45%, save over 7,000 litres of diesel per month and approximately €12,140 per month in fuel costs, and decrease NO2 and PM emissions by approximately 54%.

A complementary characterisation study of exhaust and non-exhaust sources of PM2.5 in the semi-enclosed railway station environment was also embarked upon, using chemical speciation and receptor modelling techniques. The 24-hour PM2.5 concentrations at the station ranged from 15.1 to 19.6 μg/m3, exceeding the WHO guideline value of 15 μg/m3 but remaining below the proposed EU guideline value of 25 μg/m3. Chemical analysis showed that PM2.5 was dominated by carbonaceous matter (83-90%), followed by elemental species (8-12%) and ionic species (4-5%). Positive Matrix Factorisation (PMF) analysis identified four stable sources, with good agreement between modelled and measured PM2.5 concentrations (R2 = 0.87). Train exhaust was the dominant contributor (80-85%), followed by secondary organic and inorganic aerosols (10-15%), sea salt aerosols (3-7%), and non-exhaust sources (2-3%).

Finally, health impacts were quantified using disability-adjusted life years (DALYs) and translated into economic losses for staff and commuters. The DALY analysis estimated that staff could lose up to 0.04 years (15 days) of healthy life annually, corresponding to an average monetary loss of €9,731 per staff member from combined exposure to PM2.5 and NO2. Commuters experienced a DALY loss of 0.004 years (1 day) per year, corresponding to a total economic impact of approximately €6.25 million per year across all daily users. These findings highlight the importance of implementing practical, cost-effective measures, such as reducing train idling times, to improve air quality and minimize health risks.

This study outlines the importance and value of LCM devices and complementary monitoring and modelling techniques to improve our understanding of spatio-temporal conditions, pollutant sources and health risk in and around semi-enclosed train station, emphasizing the potential for evidence-based strategies to improve air quality and sustainability in semi-enclosed railway stations.