Remaining useful life prediction of lithium-ion batteries based on data denoising and improved transformer

Abstract

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries (LIBs) is essential in improving the safety and availability of energy storage systems. However, the capacity regeneration phenomenon of LIBs occurs during actual usage, seriously affecting the accuracy of LIBs' RUL prediction. This study proposes a RUL prediction method of LIBs based on mode decomposition and an improved transformer. Firstly, to mitigate the impact of capacity degradation, we use the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method to decompose the battery capacity degradation into multi-scale component sequences. However, some noise remains in the high-frequency data output by CEEMDAN decomposition. To minimize noise impact on the accuracy of the prediction results, a single high-frequency data is then decomposed into multiple rich-featured subsequences using the variational mode decomposition. Finally, an improved transformer model extracts global and local features from these subsequences to improve the RUL of LIBs prediction accuracy. The proposed method is validated on two widely used public datasets, NASA and CALCE. Experimental results show that the proposed method has lower errors in some evaluation metrics. Compared to the four state-of-the-art methods, the proposed method improves the R-squared metric by 23.37 % and 39.81 %, respectively.