Cross-domain machine transfer learning for capacity aging trajectory prediction of lithium-ion batteries

Abstract

Machine transfer learning-based methods have emerged as a promising solution to capacity aging trajectory prediction for different types of batteries. However, the methods lack transferring interpretability, coupled with obvious variations in data across diverse batteries, presenting challenges to early lifespan prediction. A voltage-capacity (V-Q) curve reconstruction-based transfer learning method across source and target domains is proposed to predict the aging trajectories for various batteries. Initially, a generalized mathematical model is established for the V-Q curves of three types of batteries after a two-step transformation. Based on the model, a partial loss function for parameter optimization within deep learning is constructed. Subsequently, a Global Attention-sequence-to-sequence model is developed to reconstruct the V-Q curves in the target domain. The reconstructed data distribution in the target domain exhibits high similarity to that in the source domain, significantly reducing the detrimental impact of domain differences on transfer learning. Ultimately, a predictive model for aging trajectories, applicable to different types of batteries, is formulated with the transfer method. The model's accuracy is validated across three types of batteries. It is demonstrated that the mean absolute percentage error of aging trajectory for Lithium Cobalt Oxide battery and Ternary Lithium-ion battery are 1.09 % and 1.71 %, respectively.