Robust state-of-charge estimation of lithium-ion batteries via a parallel informer-long short term memory framework

Accurate and robust state-of-charge (SOC) estimation is essential for the safe and efficient operation of lithium-ion battery management systems under dynamic operating conditions. This study proposes a parallel global-local temporal modeling framework that integrates an Informer branch for long-range dependency extraction with an LSTM branch for short-term transient characterization. The two branches independently learn complementary temporal representations from terminal voltage, current, and temperature measurements, which are subsequently integrated through a nonlinear concatenation-based regression head to generate the final SOC estimate. The proposed method was first evaluated on the CALCE datasets using FUDS data for training and DST data for testing, thereby establishing an intra-dataset cross-condition evaluation protocol. Under this setting, the model achieved an RMSE of 1.80%, an MAE of 1.33%, and an R 2 of 0.9961, outperforming representative recurrent, convolutional, Transformer-based, and hybrid baseline models. Ablation experiments further demonstrated the complementary contributions of the Informer and LSTM branches and confirmed the effectiveness of the parallel connection topology and nonlinear feature-fusion strategy. Additional evaluations across temperatures from 0 to 50 °C and under different initial SOC levels indicated stable SOC trajectory tracking under the investigated distribution shifts. Independent validation was also conducted on the McMaster LG 18650HG2 datasets using three random seeds, yielding mean MAE, RMSE, and R 2 values of 0.826%, 1.247%, and 0.9979, respectively. These results demonstrate that collaborative modeling of long-range dependencies and local transient dynamics provides an effective data-driven approach for accurate SOC estimation across the evaluated operating conditions and battery datasets.