Physics-Informed Decoupled Machine Learning for Context-Aware EV Range Optimization and Multi-Objective Driver Advisory

Abstract

Auxiliary heating, ventilation, and air conditioning (HVAC) systems can reduce electric vehicle (EV) driving range by over 20%, yet prevailing machine learning estimators often suffer from temporal data leakage, uninterpretable black-box structures, and lack real-time driver feedback. To address these challenges, this study presents a physics-informed decoupled machine learning framework integrated with a multi-objective Pareto Human–Machine Interface (HMI) advisory system. Powertrain traction power is estimated using a HistGradientBoosting regressor incorporating a mechanistic Vehicle Specific Power (VSP) feature, while cabin thermal dynamics are modeled via a regularized Random Forest regressor enriched with a Newtonian thermal decay function. Evaluated across an empirical 55-trip dataset using a 5-Fold GroupKFold cross-validation protocol, the traction and thermal models achieved out-of-sample accuracy of R2 = 0.9869 (MAE = 0.71 kW) and R2 = 0.8656 (MAE = 0.25 kW), respectively. Feature attributions were verified using SHAP analysis. An onboard Pareto optimization loop dynamically balances range extension against passenger thermal discomfort to deliver actionable driver recommendations. Multi-trip evaluation indicates that a representative 30% auxiliary load suppression yields average net energy savings of 5.21% entirely through software-driven guidance.