SOC prediction for electric buses based on interpretable transformer model: Impact of traffic conditions and feature importance

With increasingly serious environmental issues , the adoption of low-noise, low-emission new energy electric buses in urban public transportation systems is steadily increasing. Accurately estimating the battery State of Charge (SOC) during electric bus operations has emerged as a significant research challenge to improve battery health status. This study introduces a Transformer-based model employing a masked multi-head attention mechanism, designed to precisely capture both local and long-term dependencies through multi-source data integration, thereby achieving high-accuracy SOC predictions. The model effectively integrates vehicle dynamics, driver behavior, traffic conditions, and weather data through a Gated Residual Network (GRN) and a Variable Selection Network (VSN), improving SOC prediction accuracy. The GRN captures complex relationships between features, while the VSN identifies key factors related to SOC and reduces noise interference. The Masked Multi-Head Attention Mechanism enhances the model's ability to capture long-term dependencies, further improving SOC predictions over extended time horizons. The model is trained and evaluated using real-world datasets collected in Changsha city, China, including electric bus operation data, license plate recognition data, and weather information. Experimental results demonstrate that the model effectively captures the complex relationships between multi-source data and SOC, yielding accurate SOC predictions for the 10, 20, and 30-min prediction horizons. The corresponding Symmetric Mean Absolute Percentage Errors (SMAPE) are 0.418 %, 0.220 %, and 0.301 %, respectively, which validates the precision and reliability of proposed model. • Statistical analysis of factors influencing SOC to determine model features. • Novel Transformer-based model for precise and interpretable SOC prediction. • Effective capture of short-term dynamics and long-term dependencies in SOC prediction. • Analysis of feature importance and horizons using attention weights for interpretability.