Predicting Near-Road Exposure from Vehicular Emission in Urban Areas Using the LSTM Technique

Emissions arising from vehicular traffic are a significant contributor to air pollution in urban environments, causing severe health effects in pedestrians and near-road residents. Thus, it is crucial to estimate pollutant concentrations along the roadside in urban areas. However, machine learning models such as random forest (RF) and support vector regression (SVR) often fail to consider the time dependency pattern in air quality models. Thus, this study forecast pollutant concentrations from vehicular emission using the long short-term memory (LSTM) method at urban traffic intersections. Furthermore, shorter monitoring intervals of 5 min were considered to account for dynamic traffic flow and atmospheric changes. The pollutants considered in this study were CO, NO2, PM2.5, and PM10, estimated from traffic volume data categorized by vehicle type and the existing environmental conditions. The results indicated that shorter intervals provided satisfactory results in forecasting vehicular pollution. Moreover, better prediction accuracy was achieved using LSTM than with RF and SVR. The R2 value varied between 0.78 and 0.94, and mean absolute error (MAE) and root mean square error (RMSE) were less than 9 and 13 μg/m3, respectively. The findings of the study indicate that the proposed LSTM model accurately predicts the concentrations arising from vehicular activities. Furthermore, assessing models over shorter periods effectively captured the variation in pollution trends. This work illustrates the significance of integrating deep learning techniques with experimental monitoring studies to investigate traffic-related air pollution in urban areas.