To address the limitations of existing real-time energy consumption prediction models for electric vehicles—particularly in environmental perception and dynamic calibration mechanisms—this paper proposes an energy prediction framework that integrates environmental perception with reinforcement learning. First, to enhance the model’s ability to perceive complex driving conditions, a road condition perception algorithm is designed based on contrastive learning and coupled reinforcement learning, incorporating a multi-scale image feature fusion mechanism. This enables the effective extraction of environmental features highly correlated with vehicle energy efficiency, thereby improving perception accuracy. Second, a Markov-based real-time energy estimation model is constructed and embedded into a reinforcement learning framework. By data-driven updates of the Q-function and cumulative rewards, the prediction model achieves self-evolution and adaptive optimization. Meanwhile, a scene-aware prioritized experience replay mechanism is introduced to emphasize key scenarios such as sudden slope changes and aggressive acceleration/deceleration, further enhancing the model’s learning efficiency and generalization capability under complex conditions. Finally, a scenario-aware prioritized sampling strategy improves training data quality, significantly boosting the convergence and training efficiency of the reinforcement learning process. Experimental results demonstrate that the proposed method exhibits superior robustness and stability across diverse driving scenarios and vehicle types, achieving a MAE below 0.2%, RMSE below 0.3%, and R2 above 99.5%, outperforming existing models such as Transformer, Informer, Mamba, and LSTM.