Multimodal temperature prediction for lithium-ion battery thermal runaway using multi-scale gated fusion and bidirectional cross-attention mechanisms

Lithium-ion batteries are widely used in electric vehicles and energy storage devices. However, their high energy density poses a risk of thermal runaway, potentially leading to fires or explosions. Accurate temperature prediction can effectively mitigate the risk of thermal runaway, ensuring the safe operation of the battery. This study proposes a data-driven model for early prediction of lithium-ion battery temperatures during the thermal runaway phase. Specifically, the model utilizes time-series two-dimensional (2D) thermal images and one-dimensional (1D) temperature data as multimodal inputs. A Convolutional Neural Network (CNN) is used to effectively encode the sample data, while a bidirectional cross-attention mechanism module aligns feature domains across different modalities. Subsequently, a gate recurrent unit (GRU)-based multi-scale gated fusion module integrates these features, enabling the model to fully leverage the battery historical data. Finally, accurate battery temperature prediction is achieved through effective feature metrics extracted by the fully connected layer. Under a limited dataset of thermal runaway samples, multiple ablation experiments were designed to validate the effectiveness of the proposed modules. Compared to conventional algorithms such as CNN-GRU, the proposed model exhibits enhancements of 91.07 % and 75.45 % in terms of MSE and MAE, respectively, demonstrating its superior predictive performance.