Explainable end-to-end prediction of remaining driving range for electric vehicles based on balanced ensemble transformer

Range anxiety is a significant challenge impeding the widespread adoption of electric vehicles. Predicting remaining driving range is crucial for alleviating this concern. However, existing prediction methods often lack both interpretability and generalizability. Hence, this study proposes a novel deep learning model, the Balanced Ensemble Transformer (BET). The BET integrates an ensemble of specialized Transformers to handle different segments of driving data, complemented by a dual-aspect explainable module for feature importance and causality analysis. Stratified sampling was used to balance the distribution of driving ranges, ensuring both high and low ranges were well-represented in training. Next, specialized Transformers were constructed, each tailored to learn and predict specific segments of the remaining driving range spectrum. Finally, a guiding feature routes which Transformer handles each prediction, optimizing accuracy across diverse scenarios. The results demonstrate that the BET outperformed the state-of-the-art models with a reduction of 20.5 % in mean absolute error and 1.4 % in mean absolute percentage error. The explainable method underscores the pivotal role of state of energy during model learning, whereas the contribution of state of health diminishes with battery degradation. Furthermore, this study demonstrates a clear causality between voltage-related features and remaining driving range. The BET offers novel insights into the significance and reliability of features and is readily transferable across different vehicle models and driving conditions.