State-of-Charge Estimation Considering Cell-to-Cell Differences Using Two-Stage Transfer Estimation

Accurate state of charge (SOC) estimation for lithium-ion batteries is a vital challenge hindered by inherent cell-to-cell variations and the lack of labeled data in target cells. To address the limitations, this paper proposes an unsupervised two-stage transfer estimation (TSTE) framework, integrating a fuzzy long short-term network (FLSN) that utilizes residual connections, parallel LSTM branches, and a mixed-Gaussian fuzzy block to model the nonlinear and noisy dynamics of the battery. TSTE operates in two distinct stages. First, it pre-trains a source-domain SOC estimator and an error corrector designed to map the discrepancy between source and target domain measurements. Second, a cyclic adversarial training scheme adapts the pre-trained SOC estimator for the target domain by delivering domain-invariant representations. This adaptation process is performed without any target-domain SOC labels. Experiments on self-collected and public datasets demonstrate the superiority of TSTE. Under diverse operating conditions, initial SOCs, current noise levels, temperatures, and the distinct battery type, the proposed method outperforms comparative approaches with an overall mean absolute error of 0.82%. Ablation studies verify the contributions of each architectural component of TSTE. Overall, TSTE provides a robust and high-precision solution for unsupervised, cell-to-cell SOC estimation, effectively addressing the challenge of label scarcity.