Physics-informed deep learning with multi-resolution for ensemble prediction of lithium-ion battery health status

Abstract Accurate prediction of the health status is critical for the reliability and safety of lithium-ion batteries (LIBs). However, some methods do not consider physical information in battery capacity degradation and typically overlook the capacity regeneration phenomenon (CRP) in their predictions. In this study, a multi-resolution and ensemble prediction method based on physics-informed deep learning for LIBs status is proposed. Specifically, multi-resolution decomposition is performed on battery capacity degradation trends to analyze global and local features. Global degradation features are physically modeled and integrated into the deep learning model to enhance interpretability and prediction accuracy. Local features can reflect CRP, and ensemble prediction of both global and local features can enhance feature-capturing capability and prediction accuracy. Experimental results indicate that the mean absolute error and root mean square error of this method, for capacity prediction, is almost consistently within 0.01. The absolute error for remaining useful life prediction is nearly within 1 cycle, which validates the effectiveness and stability of this method for LIBs health status prediction.