A Hybrid Ensemble Approach to Predict the CO2 Emission in Transportation

Abstract

While road transport accounts for roughly a quarter of global carbon emissions, the real pressure lies with light-duty vehicles (LDVs). Canada is aiming for a net-zero future by 2050, but hitting that goal requires more than just guesswork; we need pinpoint accuracy in how we model emissions for both engineering and law-making. Our study tackles this by building a hybrid ensemble framework that combines XGBoost and CatBoost to predict CO2 output (g/km).We used the Optuna framework for Bayesian hyperparameter tuning, which helped us find the best settings without wasting processing power. We tested the model against a massive 11-year dataset from the Canadian Open Data Portal, covering over 11,300 unique vehicle models. By looking at specific details like engine displacement, how many cylinders a car has, and its actual fuel consumption across city and highway driving, the system reached a level of detail that standard models usually miss. The numbers show it worked: we hit an RMSE of 1.65 g/km and of 0.9992. It’s a practical tool for policymakers to decide where EV incentives should go and how to tighten fuel standards based on real-world data.

Keywords:

CO2 Emission Prediction, Hybrid Ensemble Model, XGboost and CatBoost, Bayesian Hyperparameter Optimization, Light-Duty Vehicles (LDVs)