Electro-Thermo-Mechanical Coupling Modeling and Internal Stress Prediction of Lithium Batteries Using an Attention-Enhanced Long Short-Term Memory Network

To address the problem that the internal stress of lithium-ion batteries is difficult to measure directly under high-rate operating conditions, which restricts safety evaluation, this paper proposes an electro-thermo-mechanical coupling modeling method (MA-LSTM-ETM) based on an attention mechanism-enhanced long short-term memory network (LSTM) to achieve high-precision intelligent prediction of internal stress. Firstly, an electro-thermo-mechanical coupling mathematical model is constructed based on multiphysics theory, and labeled multiphysics time-series data under a wide rate range of 0.5 C–5 C and a full temperature range of −10 °C–40 °C are generated via COMSOL simulation. Furthermore, a multi-head attention mechanism-enhanced long short-term memory network (LSTM) is designed to adaptively extract the nonlinear mapping relationship between input features such as voltage, current, and surface temperature, and internal temperature and diffusion-induced stress. Experimental results demonstrate that the proposed method exhibits excellent performance under various rate and temperature conditions. In terms of global average indicators, compared with the standard long short-term memory network (LSTM), BP neural network, and CNN-LSTM, the root mean square error (RMSE) of internal stress prediction is reduced by 38.9%, 58.0%, and 27.5%, respectively. Within the full temperature range, the mean absolute error (MAE) of internal temperature prediction ranges from 0.25 °C to 0.65 °C, and the error is below 0.30 °C under normal and high-temperature conditions. This study provides a new technical approach for online internal state perception and intelligent safety management of lithium-ion batteries.