Machine learning-based safety early warning model for thermal runaway in traction battery packs of electric vehicles

Abstract With the rapid adoption of electric vehicles(EVs), safety hazards caused by thermal runaway of traction battery packs have become increasingly prominent, posing a serious threat to both occupants and vehicle safety. Existing thermal runaway safety warning technologies often suffer from delayed responses and frequent false alarms due to insufficient fusion of multi-source information, making it difficult to meet the safety protection requirements of real-world vehicle scenarios. To address these issues, this study proposes a safety warning model for thermal runaway in EV traction battery packs. The model introduces a hybrid algorithm that integrates Bidirectional Long Short-Term Memory (BiLSTM), Transformer, and Gated Fusion Unit (GFU) to effectively process multi-source data. A comprehensive charging dataset was constructed using numerical simulations and real-world vehicle data. Ablation studies and performance tests demonstrate that our model significantly outperforms existing benchmark models in thermal runaway prediction accuracy. Furthermore, based on the predictive output of the model, a method for determining safety warning thresholds is established. Experimental results confirm that during the charging process, when the temperature residual exceeds the set threshold, the system can trigger a warning, providing a lead time of 9.15 to 17.35 seconds. This offers valuable time for occupant evacuation and vehicle emergency response.