Decomposed kinetics of electric vehicle energy consumption rate: A multi-method approach

This work introduces an integrated Residual Physics-Informed Machine Learning (RPIML) framework for estimating the electric vehicle (EV) energy consumption rate. The energy consumption rate is critical for range prediction and energy management, yet remains challenging due to the complex interplay between deterministic physical forces and nonlinear dynamic losses. Traditional data-driven approaches often suffer from time-series data leakage and struggle to balance predictive fidelity with the computational constraints of vehicular edge computing. This work bridges the gap between static, physics-based models and conventional black-box machine learning models. Utilising high-fidelity chassis dynamometer data from nine diverse vehicles provided by the Argonne National Laboratory, the study evaluates four distinct architectures: a static physics baseline, 1-stage black-box models, 2-stage cascaded (grey-box) models, and the RPIML framework. To mitigate data leakage and ensure reliable generalisation, the methodology employs a 9-fold Leave-One-Vehicle-Out (LOVO) cross-validation protocol. Statistical validation via Wilcoxon and Friedman/Nemenyi tests reveals that conventional 1-stage black-box models fail to outperform the physics baseline when deprived of deterministic kinetic variables. In contrast, both the 2-stage cascaded and RPIML architectures significantly outperform the physical baseline. Furthermore, a multi-objective Pareto-frontier analysis utilising a Model Complexity Index (MCI) identifies the optimal model architectures that balance maximum theoretical accuracy with relative computational complexity. It is important to note that this study relies on a 2012–2015 laboratory chassis dynamometer dataset; therefore, these findings provide a foundational framework that requires further real-world calibration. Ultimately, the findings demonstrate the theoretical viability of the RPIML framework as an efficient and physically interpretable proof of concept. This establishes a methodological foundation for future research into real-time energy estimation, illustrating the theoretical potential to reduce reliance on direct physical dynamometer sensors.