Machine Learning Approaches for Predicting CO2 Emission Rates from Light Vehicle Engine

This research assesses the predictive accuracy of three supervised machine learning algorithms - Multiple Linear Regression (MLR), Support Vector Machine (SVM), and Neural Networks (NN) - for modeling CO2 emission rates of a light-duty vehicle engine typical of the fleet in Lima-Metropolitana. The models were trained and tested on a dataset of 23,370 measurements obtained from engine dynamometer testing under varied speed and load conditions. Key predictor variables included air mass flow, intake air pressure, intake air temperature, engine speed, and torque. The results indicate high predictive performance for all three techniques, with Root Mean Square Error (RMSE) values approaching zero and R-squared (R2) coefficients near unity. For 75% of the data, the maximum prediction errors were 1.11% for MLR, 1.67% for SVM, and 1.00% for NN. Notably, while Neural Networks demonstrated the lowest error in this range, they also exhibited a higher maximum absolute error of 7.85%. A key advantage of the NN approach was its reduced requirement for data preprocessing and feature engineering compared to MLR and SVM.