Fusion of Model-Based and Data-Driven Approaches for SOC Estimation of Low-Temperature Lithium-Ion Batteries

This paper presents a fusion strategy for State of Charge (SOC) estimation of lithium-ion batteries under lowtemperature conditions (-10° C) by combining model-based and data-driven approaches. The model-based method is used to capture the physical polarization characteristics, while the nonlinear dynamics is obtained by long short-term memory (LSTM) through operational data. A hybrid model is further proposed by integrating dual extended Kalman filter-estimated SOC as auxiliary input to LSTM, addressing the limitations of standalone methods in low-temperature scenarios. Experimental results under Urban Dynamometer Driving Schedule (UDDS), United States 06 (US06), and Dynamic Stress Test (DST) show that the fusion model showcased advantages that the root mean square error (RMSE) can be reduced to 0.53%-0.91% compared with the pure model-based (3.85%-7.00%) and pure data-driven (1.12%-1.91%) approaches. This work highlights the effectiveness of synergizing physical constraints with data-driven adaptability for robust SOC estimation in cold environments.