Deep Learning-Powered Lifetime Prediction for Lithium-Ion Batteries Based on Small Amounts of Charging Cycles

Abstract

The accurate lifetime prediction of lithium-ion batteries (LIBs) is essential to the normal and effective operation of electric devices. However, such estimation faces huge challenges due to the nonlinear capacity degradation process and uncertain LIBs' operating conditions. This article proposes a novel end-to-end deep learning (DL) model, namely, a dual-stream vision transformer with the efficient self-attention mechanism (DS-ViT-ESA), to predict the current cycle life (CCL) and remaining useful life (RUL) of the target battery. The local and global spatiotemporal features are effectively captured via the vision transformer (ViT) with the efficient self-attention (ESA) mechanism based on small amounts of charging cycles. Meanwhile, by serving the differences between each cycle as the supplementary model input, the inner cycle and cycle-to-cycle aging information could be extracted and fused by a dual-stream structure to enhance prediction accuracy. Experiments exhibit that the proposed model only needs 15 charging cycles (about 1%~3% along the trajectory to failure) while ensuring the lifetime prediction accuracy (RUL error: 5.40%, CCL error: 4.64%, and early lifetime prediction error: 2.16%). Meanwhile, the model also shows the effective zero-shot generalization capacity for the charging strategies not appearing in the training dataset.