Mitigating range anxiety in electric vehicles using transformer-based frameworks

Range anxiety is a central concern for Battery Electric Vehicle (BEV) adoption, and it is exacerbated by inaccurate estimations of State of Charge (SoC) and Remaining Driving Range (RDR), creating a critical engineering bottleneck. To bridge this gap, two models, specifically Bidirectional Gated Recurrent Unit Network Enhanced with Multi-Head Attention Mechanism (BiGRU-MHA) and Twin-Decoder-Based Multi-Scale Feature Extraction Transformer (TDMSTrans) are proposed to predict RDR and (SOC) for optimising energy management strategies and alleviating range anxiety in electric vehicle: (i) In BiGRU-MHA, a multi-head attention mechanism is proposed following the BiGRU layer to enable the selective refinement of hidden states through learned attention weights, effectively capturing both local sequential dependencies and global contextual relationships, and (ii) TDMSTrans introduces two key innovations: a twin-decoder architecture that enables multi-task learning through task-specific feature refinement for SOC and RDR, and a multi-scale feature extraction module utilising convolutional layers with varying kernel sizes to capture hierarchical temporal patterns, thereby enriching temporal representations beyond the capacity of standard transformers. Validated on real-world BEV datasets, both models demonstrate superior predictive accuracy compared to standard transformers, effectively optimising energy management strategies and mitigating the stochastic uncertainty inherent in current EV range estimation.