A hybrid data-driven method for lithium-ion battery capacity and remaining useful life early prediction

Accurately predicting the capacity and remaining useful life (RUL) of lithium-ion batteries during the early cycles is crucial for battery management systems (BMS). Therefore, this paper proposes a hybrid data-driven model to capture the capacity degradation characteristics and improve early prediction performance. Initially, the original battery capacity data are decomposed using the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) method to enhance the analysis of local fluctuations and overall degradation trends. Additionally, an adaptive weighting mechanism, designed using the improved crested porcupine optimizer (ICPO), assesses the contribution of each subseries to prediction results and then inputs the weighted subseries into the Pyraformer to improve accuracy. Subsequently, the Pyraformer model can capture multi-resolution temporal dependencies and achieve effective small-sample learning by integrating the pyramidal attention module (PAM) and coarse-scale construction module (CSCM), making the proposed model more adept at both single-step and multi-step early prediction. Experimentally, when trained on only the initial 20% of the original data, the model achieved maximum R2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R<^>{2}$$\end{document} values of 0.996 (single-step) and 0.994 (multi-step), with corresponding minimum AE of 0 and 1, respectively. These results demonstrate that the ICEEMDAN-ICPO-Pyraformer model can effectively alleviate the impact of capacity regeneration and achieve superior accuracy and robustness in capacity and RUL early prediction across different output step sizes.