To address the limitations of existing real-time energy consumption prediction models for electric vehicles — Particularly their insufficient environmental perception and lack of dynamic calibration mechanisms — this study proposes an energy consumption prediction model integrating environmental perception with reinforcement learning. First, to enhance the model’s capability of understanding complex driving conditions, a road-condition perception algorithm is designed based on contrastive learning and coupled reinforcement learning, together with a multi-scale image feature fusion mechanism. This design effectively extracts environment features highly correlated with vehicle energy efficiency, thereby improving perception accuracy under non-stationary operating conditions. Second, a Markov-based real-time energy efficiency estimation model is constructed and mapped into a reinforcement learning framework. A temporal consistency regularization term based on discounted future energy consumption (where the $Q $-function is used solely as an energy-response evaluator) is introduced to achieve self-calibrated optimization, significantly enhancing prediction robustness and adaptability in dynamic scenarios (without generating control outputs). Meanwhile, a scenario-aware prioritized experience replay mechanism is incorporated to strengthen the model’s ability to recognize and learn from key driving conditions such as slope mutations and rapid acceleration/deceleration events, further improving feature extraction and generalization in complex environments. Finally, a scenario-aware prioritized sampling strategy is employed to optimize the distribution of training samples, improving the convergence rate and efficiency of the reinforcement learning process. Experimental results demonstrate that the proposed method exhibits excellent robustness and stability across two vehicle types and multiple simulated driving scenarios, achieving an MAE below 0.2%, an RMSE below 0.3%, and an $R^2 $ above 99.5%. Compared with existing Transformer, Informer, Mamba, and LSTM models, the proposed approach reduces average prediction error by approximately 40% $\sim $ 70% and improves convergence speed by about 30%, yielding significantly higher prediction accuracy under complex driving conditions.