Comparative analysis of deep learning approaches for electrothermal and state of health estimation in lithium-ion batteries under variable thermal conditions

Due to prolonged operation and repeated charge–discharge cycles, lithium-ion (Li-ion) batteries undergo gradual degradation that impacts system reliability, safety, and lifespan. To enable precise health tracking in intelligent battery management systems (BMS) and overcome the limitations of conventional single-state estimation models, this work introduces a novel, unified data-driven framework for coupled electrothermal state prediction across multi-temperature cycling regimes (4 °C, 24 °C, and 43 °C). This study systematically compares three deep learning architectures: Long Short-Term Memory, Gated Recurrent Unit, and Transformer networks, using task-specific feature engineering that incorporates measured voltage, temperature, current, cycle number, and time. The key innovation lies in identifying regime-specific architectural superiorities aligned with underlying electrochemical dynamics: quantitative evaluation using Root Mean Square Error reveals that under low-temperature conditions (4 °C), the streamlined GRU architecture outperforms others in State of Charge (SoC) (RMSE = 4.46) and temperature (RMSE = 0.59 °C) estimation due to its robustness against noise in sluggish ion dynamics. Conversely, the Transformer model demonstrates superior accuracy in moderate and high-temperature environments (24 °C and 43 °C), achieving SoC RMSE as low as 0.82 and voltage RMSE of 0.021 V at 43 °C by capturing long-range self-attention temporal dependencies. Furthermore, an integrated analysis correlates these dynamic electrothermal predictions with impedance growth ( R e , R ct ) and capacity fade, establishing a direct link among operational state modelling, degradation assessment, and long-term State of Health (SoH) evolution. These findings provide practical, actionable guidelines for selecting temperature-adaptive deep learning models for health monitoring and predictive maintenance in next-generation BMS for electric vehicles and energy storage applications.