Development of a Deep Neural Network Model to Predict PM2.5 Concentrations Based on Key Factors Affecting Air Pollution in Bangkok

Abstract

PM_2.5 is a significant air pollution concern in Thailand, particularly in urban areas such as Bangkok, where the 24-hour average concentration frequently exceeds national ambient air quality standards during severe seasonal episodes. Road transportation is the primary source of PM_2.5 in Bangkok, accounting for 73% of total emissions, according to the Pollution Control Department. In this study, a predictive modeling framework is developed to investigate the relationship between vehicle count and the mass concentration of PM_2.5 from vehicular exhaust emissions. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are employed for vehicle detection and counting, whereas a long short-term memory (LSTM) network is used to predict traffic-induced PM_2.5 levels in urban environments. The fixed-box model was utilized as a baseline calculation, incorporating the vehicle count, meteorological parameters, and localized PM_2.5 concentration. Data was collected from a monitoring station on Vibhavadi Rangsit Road, Din Daeng, Bangkok. To ensure robust evaluation on a limited dataset, a leave-one-out cross-validation (LOOCV) strategy was applied. The results indicate that the LSTM model can stably reproduce traffic-induced PM_2.5 patterns without overfitting, achieving an R² of 0.4391, an RMSE of 2.2774 µg m^-3, and a MAPE of 26.63%. Furthermore, a supplementary analysis comparing these predictions to measured ambient data estimated that vehicular emissions contribute approximately 59% to the total PM_2.5 pollution in the study area. This aligns with the 50–73% contribution from land transport established by prior studies. These findings demonstrate the model’s ability to reliably predict traffic-related emissions. The proposed model can be applied to identify PM_2.5 sources from transportation and, alongside other models, assist in forecasting PM_2.5 levels in Bangkok to support air quality management.