Comparative Evaluation of AI Techniques for Vehicle Fuel Consumption and CO2 Emission Prediction

Abstract

The transportation sector remains a major source of energy consumption and greenhouse gas emissions, making reliable vehicle-level estimates of fuel consumption and CO_2 emissions important for fleet management, regulatory assessment, and sustainable transport planning. However, predictive performance can be overstated when models are evaluated through random data splits or when target-related information leaks into the input features. This study therefore proposes a leakage-aware machine learning framework designed to assess generalization across time and data sources. Canadian vehicle records from 2015--2020 are used for model development, whereas records from 2021--2024 are reserved for temporal evaluation. External validity is examined using US EPA vehicle data. Linear regression, random forest, extra trees, support vector regression, LightGBM, XGBoost, CatBoost, and FT-Transformer are compared under a common protocol. All preprocessing steps are fitted only on training folds, and target-derived variables are excluded. PCA sensitivity, out-of-fold stacking, statistical testing, residual analysis, SHAP interpretation, and computational cost are also examined. The framework provides a reproducible basis for developing dependable vehicle-screening tools under temporal and cross-dataset distribution shifts.