Battery State of Health Early Prediction via a Generative Adversarial Network and a Hybrid Deep Learning Model

Early prediction of the State of Health (SOH) of lithium-ion batteries is crucial for battery manufacturing and management. However, battery aging is complex and nonlinear, and scarce early data challenges long-term prediction. To address this, a data-driven SOH prediction method based on transfer learning is proposed. As SOH degradation relates to various factors, an improved TimeGAN is first employed for data augmentation and style transfer to expand the training dataset. Additionally, a hybrid architecture, ATTCN-Mamba, integrating attention mechanisms, temporal convolutional networks, and state-space models, is constructed to improve prediction accuracy. Experimental results show that the proposed model significantly outperforms traditional models in key metrics. Furthermore, with the TimeGAN-based transfer learning strategy, the model significantly enhances prediction accuracy and robustness even with substantially reduced target-domain data, with RMSE, MSE, and MAPE lower than those of the baselines without transfer learning. The method also demonstrates broad prospects for high-precision, full-life-cycle battery health management under edge-cloud architectures and cloud-based battery management systems.