Developing Machine Learning Models for Predicting Microscopic Emissions From On-Road Vehicles: A Preliminary Study

초록

Accurate prediction of vehicle emissions under real-world driving conditions is essential for improving air quality management and regulatory compliance. Traditional emission estimation models often fail to capture microscopic variations in driving behavior and environmental factors, limiting their applicability to dynamic road conditions. This study explores the use of machine learning techniques to develop a predictive model for microscopic vehicle emissions using real driving emission (RDE) data. Various ML models, including Random Forest, Artificial Neural Networks (ANN-MLP), Gradient Boosting Regressor (GBR), and Support Vector Machine (SVM), were applied to estimate emissions of CO2, CO, NOx, THC under different driving conditions. A moving average technique was implemented to enhance model robustness by reducing data noise and improving the representation of emission trends. The models were trained and validated using RDE data obtained from the Korean National Institute of Environmental Research (NIER), with performance evaluation based on the coefficient of determination (R²). Results indicate that increasing the moving average window improves model accuracy, suggesting that high-frequency emission data introduce noise that complicates short-term predictions. The study highlights the effectiveness of ML-driven models in capturing complex relationships between vehicle operations and emissions, outperforming traditional estimation methods.