Meteorological and traffic effects on air pollutants using Bayesian networks and deep learning

Traffic emissions have become the major air pollution source in urban areas. Therefore, understanding the highly non-stational and complex impact of traffic factors on air quality is very important for building air quality prediction models. Using real-world air pollutant data from Taipei, Taiwan, this study integrates diverse factors, including traffic flow, speed, rainfall patterns, and meteorological factors. We constructed a Bayesian network probability model based on rainfall events as a big data analysis framework to investigate understand traffic factor causality relationships and condition probabilities for meteorological factors and air pollutant concentrations. Generalized Additive Model (GAM) verified non-linear relationships between traffic factors and air pollutants. Consequently, we propose a long short term memory (LSTM) model to predict airborne pollutant concentrations. This study propose a new approach of air pollutants and meteorological variable analysis procedure by considering both rainfall amount and patterns. Results indicate improved air quality when controlling vehicle speed above 40 km/h and maintaining an average vehicle flow < 1,200 vehicles per hour. This study also classified rainfall events into four types depending on its characteristic. Wet deposition from varied rainfall types significantly affects air quality, with TypeⅠrainfall events (long-duration heavy rain) having the most pronounced impact. An LSTM model incorporating GAM and Bayesian network outcomes yields excellent performance, achieving correlation R2 > 0.9 and 0.8 for first and second order air pollutants, i.e., CO, NO, NO2, and NOx; and O3, PM10, and PM2.5, respectively.