Electro-thermal feature fusion with temporal difference regularized Mamba network for lithium-ion battery temperature prediction

Accurate prediction of lithium-ion battery temperature is essential for ensuring safety and optimizing battery thermal management systems (BTMS). However, conventional data-driven methods relying solely on electrical signals often struggle to characterize electro-thermal coupling behaviors and maintain stable long-term temperature evolution under complex operating conditions. To address these challenges, an electro-thermal feature fusion framework with temporal difference regularized Mamba network (ETD-Mamba) is proposed. Electro-thermal features, including electrical dynamic characteristics, heat generation proxies, energy accumulation descriptors, and ambient thermal information, are constructed to provide physically interpretable representations of battery temperature evolution. A temporal difference regularization term is introduced to constrain inconsistent temperature variations and enhance temporal stability. The Mamba-based selective state space architecture enables efficient long-sequence modeling with favorable computational efficiency. Experimental results under −20 °C and 10 °C conditions demonstrate that ETD-Mamba outperforms GRU, LSTM, Transformer, TCN, and Informer, achieving RMSE values of 0.52 °C and 0.25 °C, respectively. Cross-cycle and cross-temperature validation further demonstrate the transferability of the proposed framework within the investigated dataset under varying ambient conditions. During inference, the model only requires electrical measurements and ambient temperature information without continuous battery temperature sensing. The proposed method provides an accurate and reliable solution for lithium-ion battery temperature estimation.