Energy consumption prediction of range-extended electric vehicles using dependency-guided physics-regularised neural networks

Abstract

The reliable prediction of energy consumption for range-extended electric vehicles under unseen driving scenarios remains challenging due to variations in driving conditions and vehicle operating characteristics. This paper proposes a two-stage framework integrating a difference-gated long short-term memory (dLSTM) model with a dependency-guided physics-regularised neural network (DG-PRNN) model. The dLSTM captures short-horizon temporal variations in vehicle operating signals, while DG-PRNN introduces dependency-guided sparse connectivity and physics regularisation to improve prediction robustness and provide a more interpretable input structure. The model is trained using NEDC data and evaluated under unseen WLTC conditions, driver-in-the-loop scenarios, and an additional validation experiment using independently calibrated battery and engine component models derived from BYD vehicle data. Compared with data-driven and physics-regularised baselines, DG-PRNN reduces large-error risks in energy consumption prediction under WLTC and maintains accurate component-level power prediction under additional validation with different component-model conditions. In WLTC evaluation, DG-PRNN reduces the electric and fuel P 95 error levels by 1.017 kWh/100 km and 1.83 L/100 km, respectively, compared with the data-driven baseline. The proposed framework provides a lightweight and interpretable approach for robust energy consumption prediction under unseen driving scenarios.