Enhancing cross-temperature state-of-charge estimation accuracy for lithium-ion batteries using multi-physics features and physical guidance

Accurate state-of-charge (SOC) estimation is critical for the safe operation of lithium-ion batteries. However, traditional deep learning methods experience reduced accuracy in cross-temperature scenarios as they require training and testing data to share the identical distribution. To address these issues, the present study introduces a cross-temperature SOC estimation method leveraging multi-physics features and physical guidance without target domain labeling. The proposed approach begins by decomposing voltage signals into meaningful physical representations using an equivalent circuit model. These representations are then integrated with expansion force to form the multi-physics features, enabling effective cross-temperature knowledge transfer. These sets of features serve as inputs for a neural network architecture composed of a backbone network, maximum mean discrepancy, domain adversarial training, and a novel physical constraint criterion. The maximum mean discrepancy and domain adversarial training techniques align the feature distributions between two domains, while the physical constraint criterion functions as a proxy labeling strategy for essential physical guidance in the unlabeled target domain. The validation results across 80 cross-temperature conditions demonstrate significant improvements in the proposed method when compared to existing methods, with an average mean absolute error of 3.86% and an average root mean square error of 5.07%.