Physics-Informed Machine Learning for Energy Consumption Prediction in Heavy-Duty Electric Vehicles

Sammanfattning

Accurate forecasting of energy consumption in Heavy‐Duty Electric Vehicles (HDEVs) is crucial for reducing range anxiety and optimizing charging strategies under dynamic driving conditions. Traditional approaches such as empirical formulas, physics‐based simulators, and purely data‐driven models each have limitations in terms of accuracy, real‐time adaptivity, or robustness. This thesis presents a Physics‐Informed Machine Learning framework that incorporates vehicle dynamics and domain‐specific constraints directly into neural network training. In addition to predicting segment‐level energy use, the model simultaneously estimates critical parameters, the rolling resistance coefficient (C_rr) and aerodynamic drag coefficient (C_d), thereby ensuring interpretability while capturing complex, nonlinear behaviors. We validate our approach on a real‐world dataset comprising over 550 000 road‐segment samples from more than 5 000 routes and 20+ truck variants. After engineering a comprehensive feature set, including physics‐derived terms such as mgdcosθ and aerodynamic drag factors, we establish multiple baseline models for comparison. These include classical machine learning regressors (Linear Regression, Random Forest, XGBoost), a standard feedforward neural network, and purely physics-based methods. We then compare their performance against two Physics-Informed Neural Network (PINN) architectures. The leading PINN achieves an RMSE of 5.05 Wh and a PMAE of 17.9%, outperforming both data‐driven and static physics‐based models by 35–55%. Its coefficient predictions also align closely with real‐world observed variations in C_rr and C_d. These results demonstrate that embedding first‐principles physics within machine learning produces an energy‐prediction tool that is accurate, robust, and interpretable. This work advances predictive analytics for intelligent fleet management, digital twins, and real‐time energy optimization in heavy‐duty electric mobility.