Low-frequency-data-driven energy consumption prediction for battery electric vehicles: Integrating continuous trip segments and multi-task learning

Battery electric vehicles (BEVs) have yet fully crossed the barriers of range anxiety. Existing energy consumption prediction for BEVs typically relies on single-task learning and requires high-frequency (≥1Hz) vehicle data combined with environmental/road information. This raises infrastructure costs and privacy concerns. In this study, a novel framework using only low-frequency vehicle data (0.1Hz) is proposed for BEV energy consumption prediction. First, a Continuous Trip Segment Division (CTSD) algorithm is developed to implicitly capture environmental context, and then a novel Transformer-Conv1D Multi-Task Learning (TCMTL) model is built to enhance feature representation and generalization capability. The TCMTL model could simultaneously predict energy consumption (main task) and travel distance (auxiliary task). Real-world experiments demonstrate that the proposed approach reduces prediction errors by 37.1 % MAE and 31.2 % RMSE compared to the machine learning methods in existing research. Optimal results were obtained using 15-min trip segments and balanced task weights. Model interpretability analysis reveals that multi-task learning enhances performance by expanding feature representation space and systematically reallocating parameter importance toward driving inputs and vehicle response metrics. This study provides a cost-effective, privacy-preserving solution for BEV energy management.