Deep Learning‐Driven Modeling for Thermal Runaway Warning During Lithium‐Ion Battery Charging in Electric Vehicles

ABSTRACT Regarding the threat posed by lithium‐ion battery charging thermal runaway to electric vehicle (EV) safety applications, this paper proposes a Q‐learning optimized multimodal deep learning framework, and based on this framework, further constructs a lithium‐ion battery charging temperature prediction model for EVs. By integrating the local feature extraction capability of Convolutional Neural Networks (CNN), the temporal memory characteristics of Long Short‐Term Memory networks (LSTM), and the temporal modeling advantages of Temporal Convolutional Networks (TCN), the framework employs a Q‐learning algorithm to optimize network weights, ultimately resulting in the formation of the EV lithium‐ion battery charging temperature prediction model (QCLT) with high‐precision prediction capabilities. Experiments selected highly correlated parameters in EV charging through Pearson correlation coefficient as inputs, and validated the model using charging data from both NCM (Nickel‐Cobalt‐Manganese) and LFP (Lithium Iron Phosphate) lithium batteries. Comparative results showed that the QCLT model demonstrated superior prediction accuracy over other benchmark models. Furthermore, dynamic warning thresholds were established using the sliding window method, with additional validation through thermal runaway data under varying ambient temperatures. Constructed based on the aforementioned multimodal deep learning framework, the QCLT model can effectively predict abnormal temperature residual variations, issuing timely warning signals before thermal runaway occurs. This provides a critical time window for implementing safety protection measures, thereby reducing accident risks.