Global–local attention network and value-informed federated strategy for predicting power battery state of health

Predicting the health of electric vehicle (EV) batteries is critical for extending their lifespan and ensuring driving safety. Existing deep learning approaches leverage nonlinear relationships within battery monitoring signals to estimate battery health. However, these methods struggle to adaptively capture both local and global information, limiting their ability to represent complex aging patterns accurately. Some approaches employ federated learning to enhance model generalization for different batteries while maintaining data privacy. Yet, the standard average aggregation method used in these approaches constrains prediction accuracy and learning efficiency. To address these challenges, we propose a Triplet Attention-informed Federated Learning (TAFL) framework, which integrates a Global-Local Attention Network (GLAN) with a Value-informed Federated Strategy (VIFS) for precise state-of-health (SOH) predictions. GLAN incorporates both local and global attention mechanisms to strengthen feature representation, while VIFS introduces a dynamic model selection based on upload time and an attention-based aggregation strategy to improve both prediction accuracy and learning efficiency. Experimental results demonstrate that TAFL achieves an average mean absolute error (MAE) of 0.59% and a root mean square error (RMSE) of 0.62%. Tests on batteries under various operating conditions further highlight TAFL’s superior generalization, noise resistance, and efficiency.