Electric Vehicle Energy Consumption Prediction: A Physics-Informed Machine Learning Approach

Abstract

The increasing adoption of electric vehicles (EVs) as a sustainable alternative to internal combustion engine vehicles has intensified the need for accurate and interpretable energy consumption prediction models to support vehicle design, battery management, and charging infrastructure planning. This study presents a physics-informed machine learning framework for predicting EV energy consumption using a dataset of approximately 300 electric vehicle models sourced from publicly available vehicle specifications. A reduced-order physical model derived from the work–energy theorem and Newtonian mechanics was developed to bridge classical vehicle dynamics theory with data-driven modeling, incorporating vehicle mass, aerodynamic drag coefficient, and the mass-to-battery-capacity ratio as physically meaningful input features. Five regression algorithms Linear Regression, Random Forest, XGBoost, LightGBM, and Support Vector Regression were implemented and evaluated under a consistent 5-fold cross-validation framework using R², RMSE, and MAE as performance metrics. Random Forest achieved the highest predictive accuracy (R² = 0.841, RMSE = 1.593 kWh/100 km), followed by Linear Regression (R² = 0.837) and XGBoost (R² = 0.820), while LightGBM and SVR demonstrated substantially weaker performance. Feature importance analysis confirmed that battery capacity, vehicle mass, and driving range are the most influential predictors, consistent with the physics-informed framework and recent literature. The convergence between data-driven findings and physical interpretations validates the proposed approach as a robust, transparent, and scalable tool for EV energy modeling, with direct applications in sustainable transportation planning and evidence-based energy policy.