A hybrid time–frequency and cross-variable interactive Mamba network for accurate state-of-charge estimation of lithium-ion batteries

The state of charge (SOC) of lithium-ion batteries is a critical variable for energy management and safety control in battery management systems. However, under wide temperature ranges and highly dynamic operating conditions, voltage–current–temperature sequences are affected by nonlinear coupling, noise interference, and long-range dependencies. These factors pose substantial challenges for end-to-end SOC estimation, including the balance between local transients and global trends, limitations in long-sequence modeling, and insufficient cross-variable interaction modeling. To address these challenges, this paper proposes a novel end-to-end SOC estimation model, termed HCFE–TFGM–VIEM. Specifically, a Hierarchical Convolutional Feature Extraction (HCFE) module is employed to extract local dynamic features from battery operating sequences. A Time–Frequency Gated Mixer (TFGM) module is then introduced to jointly model the sequence time–frequency structure through time–frequency convolution fusion and a prior-gated mechanism, improving the model’s adaptability to periodic variations and noise perturbations. Finally, an improved Variable Interaction-Enhanced Mamba (VIEM) module is proposed to explicitly characterize the interactions among multiple variables such as voltage, current, and temperature. This module alleviates the insufficient variable interaction modeling caused by the channel-independent property of the Mamba architecture and enables stable long-range dependency modeling. Extensive experiments on the public LG 18650HG2 dataset under multiple temperatures ( − 20 ° C , − 10 ° C , 0 ° C , 10 ° C , 25 ° C , and 40 ° C ) and representative driving cycles demonstrate that the proposed model achieves favorable and consistent performance across the evaluated temperature and driving-cycle conditions within the LG 18650HG2 dataset. Under the 0 ° C condition, the average MAX, MAE, and RMSE are 2.65%, 0.40%, and 0.52%, respectively, achieving the best overall performance among the models evaluated under the common protocol and showing favorable numerical results compared with the results reported for representative advanced SOC estimation methods, while demonstrating accurate and robust SOC estimation under the evaluated conditions.