Measuring and Modelling the Concentration of Vehicle-Related PM2.5 and PM10 Emissions Based on Neural Networks

Abstract

The urban environment near the road infrastructure is particularly affected by traffic emissions. This problem is exacerbated at road junctions. The roadside concentration of particulate (PM2.5 and PM10) emissions depends on traffic parameters, meteorological conditions, the characteristics and condition of the road surface, and urban development, which affects air flow and turbulence. Continuous changes in the structure and conditions of the traffic flow directly affect the concentration of roadside emissions, which significantly complicates monitoring and forecasting the state of ambient air. Our study presents a hybrid model to estimate the amount, concentration, and spatio-temporal forecasting of particulate emissions, accounting for dynamic changes in road traffic structure and the influence of meteorological factors. The input module of the model is based on data received from street cameras and weather stations using a trained convolutional neural network. Based on the history of emission concentration data as input data, we used a self-learning Recurrent Neural Network (RNN) for forecasting. Through micromodeling, we found that the order in which vehicles enter and exit an intersection affects the concentration of vehicle-related emissions. Preliminary experimental results showed that the proposed model provides higher accuracy in forecasting emission concentration (83–97%) than existing approaches.

Keywords

• vehicle-related emissions
• recurrent neural networks (RNN)
• convolutional neural network (CNN)
• traffic
• deep learning