Machine learning based prediction of automotive fuel economy

Abstract

Accurate prediction of automotive fuel economy is important for improving vehicle design, reducing operating costs, and supporting energy-efficient transportation systems. Conventional empirical and statistical approaches often fail to capture the complex nonlinear relationships among vehicle technical characteristics. This study presents a comparative evaluation of four state-of-the-art ensemble machine learning algorithms, namely Random Forest (RF), Light Gradient Boosting Machine (LGB), XGBoost (XGB), and CatBoost (CGB), for predicting vehicle fuel economy using the publicly available UCI Auto MPG dataset, comprising 398 vehicle records characterized by seven technical attributes (392 valid samples after preprocessing). The models were optimized using Grid search optimization (GSO) and evaluated through an 80:20 training-testing split combined with five-fold cross-validation. Model performance was assessed using the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). Model interpretability was investigated using SHapley Additive exPlanations (SHAP) to quantify the influence of individual vehicle attributes. Among the evaluated algorithms, CGB achieved the best predictive performance with a testing R² of 0.9204, normalized RMSE of 0.0473, normalized MAE of 0.0347, and MAPE of 9.07% on the normalized target scale, indicating promising performance on the held-out test set. SHAP analysis revealed that model year, vehicle weight, horsepower, and engine displacement were the most influential variables affecting fuel economy predictions. The study demonstrates that optimized ensemble learning methods can accurately model complex relationships between vehicle specifications and fuel economy while providing interpretable predictions. Although the findings are based on a benchmark dataset representing vehicles manufactured between 1970 and 1982, the proposed comparative framework provides a transparent comparative framework that can be extended to larger and more contemporary automotive datasets.