Sequence-aware energy consumption prediction for electric vehicles using pre-trip realistically accessible data

Energy consumption (EC) prediction plays a crucial role in reducing range anxiety and operation scheduling as well as optimization of electric vehicles. Current methods predominantly treat a trip as a singular entity or assume unrealistic inputs (e.g., driving trajectories or speed profiles of a trip) that are not accessible prior to departure for EC prediction. This study proposes a sequence-dependence aware deep learning methodology for EC prediction using pre-trip realistically accessible data. Sequence modeling architectures are employed to capture the nuanced variations and dependencies among adjacent segments rather than relying on coarse-grained average features. This study highly emphasizes pre-trip accessible data in reality for trip EC prediction, improving upon unrealistic assumptions that presuppose access to future speed profiles per second throughout a trip. Large-scale field datasets are utilized for model development, covering 2.2 million kilometers of driving from eight cities and four different vehicle models. The results demonstrate that the proposed sequence-dependence aware deep learning methodologies outperform existing methods in both prediction accuracy and interpretability, highlighting the efficacy of incorporating sequence dependencies in EC prediction. This study also quantifies the influence of various factors on EC at the segment level, providing a more granular analysis and understanding of energy efficiency. The results provide accurate and realistic EC predictions and understanding for electric vehicles that are applicable in real practice.