Environment-Aware Reinforcement Learning-Based Energy Consumption Prediction Model for Electric Vehicles

Abstract

As electric vehicles (EVs) gain growing popularity worldwide, demands for real-time and precise energy-consuming prediction have increased correspondingly. Targeting at limitations of existing models in environmental perception and dynamic calibration, this research put forward a novel model for energy-consuming prediction. This model integrated environmental perception with reinforcement learning. Specifically, at first, a road-condition perceiving approach deeply coupled with reinforcement learning was designed, and a linear multi-scale attention encoder was constructed. The aim was to extract multi-granularity environmental features related to energy efficiency and thus enhance the model’s representational capabilities under complicated dynamic driving situations. Second, a real-time energy efficiency estimation model was developed under a Markov decision process, which was also mapped to the reinforcement learning framework. Based on temporal-difference learning, the data-driven Q function was iteratively updated, and constant calibration of energy estimation was realized. Finally, a prioritization mechanism for causal-structure-based Kullback–Leibler (KL) divergence scenarios was proposed to enhance the sampling efficiency in cases of critical incidents such as slope variations and abrupt speed-accelerating/decelerating, while strengthening the robustness and generalization of the model under complicated conditions. Results confirmed the superior stability and robustness of the proposed approach across multiple operating conditions and vehicle types. Specifically, the mean absolute error (MAE) was below 12%; the root mean-squared error (RMSE) exhibited a value under 1.8%; and the R² value exceeded 99.5%. All these demonstrated its significantly improved efficiency over Transformer, Informer, Mamba, and long short-term memory (LSTM) models. EVs’ actual energy consumption in the real world was also compared with that in speed profile (EVECS) dataset, presenting an MAE below 1.15% and a RMSE under 1.65%, which further verified its excellent generalization.