Temperature Adaptive Transfer Network for Cross-Domain State-of-Charge Estimation of Li-Ion Batteries

State-of-charge (SOC) estimation plays an important role in the battery management system, which serves to ensure the safety of batteries. Existing data-driven methods for SOC estimation of Li-ion batteries rely on massive labeled data and the assumption that training and testing data share the same distribution. However, in the real world, there is only unlabeled target data and there exists distribution discrepancy caused by external or internal factors such as varying ambient temperatures and battery aging, which makes existing methods invalid. To address the challenges, a temperature adaptive transfer network (TATN) is proposed, which can mitigate domain shift adaptively by mapping data to high-dimensional feature spaces. The TATN consists of pretraining stage and transfer stage. At the pretraining stage, 2-D convolutional neural network and bidirectional long short-term memory are used for temporal feature extraction. At the transfer stage, adversarial adaptation and maximum mean discrepancy are utilized to minimize domain divergence. Furthermore, a novel label-selection method is proposed to select reliable pseudolabels. Extensive transfer experiments are performed. Notably, compared with other methods, the TATN reduces average MAE and root mean square error by $ 66\%$ and $ 78\%$ under semisupervised scenario, $ 71\%$ and $ 68\%$ under unsupervised scenario, and $ 52\%$ and $ 42\%$ at online testing. The results indicate that the TATN can achieve state-of-the-art performance in practical applications.