Remaining useful life prediction of lithium-ion batteries based on sequence decomposition and improved itransformer

Abstract

Accurately predicting the remaining useful life (RUL) is essential for the safe use of lithium batteries. However, achieving accurate RUL prediction remains highly challenging due to the complexities of degradation mechanisms and the impact of operational noise, particularly the capacity regeneration phenomenon. Consequently, we propose a model based on sequence decomposition and improved iTransformer for predicting the RUL of lithium-ion batteries. Initially, battery capacity is selected as the health indicator, and Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) is applied to decompose the degradation data into multi-scale components. Subsequently, the multi order Kolmogorov-Arnold network (MKAN) module is integrated to capture the nonlinear patterns of capacity regeneration. This is followed by the multi-head differential attention iTransformer, which amplifies attention to the relevant context while canceling noise. Specifically, the differential attention mechanism calculates the attention scores as the difference between two separate softmax attention maps, promoting a more focused feature representation. The method’s effectiveness is confirmed through experiments conducted on the public datasets from NASA, CALCE and Wenzhou. The results indicate that the proposed method achieves a maximum MAE of no more than 1.5% and a maximum RMSE of no more than 2% across three publicly available datasets, demonstrating improved accuracy and stability in RUL estimation.