Remaining useful life prediction of lithium-ion batteries using a CEEMDAN–Bi-LSTM hybrid model

Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries is a critical task in battery management systems. To address the nonlinearity and nonstationarity of raw capacity-fading data, as well as the fluctuation disturbances induced by localized capacity regeneration, this paper proposes a hybrid prediction method that combines complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and a bidirectional long short-term memory neural network (Bi-LSTM). The proposed method first decomposes the battery capacity sequence using CEEMDAN and then separates the resulting modal components into high-frequency and low-frequency parts according to the zero-crossing rate (ZCR) criterion. Pearson and Spearman correlation analyses are further performed to verify that the low-frequency (LF) component is representative of the overall degradation trend. Subsequently, under the offline prediction setting, baseline models including LSTM, Bi-RNN, and Bi-LSTM are compared to evaluate the effectiveness of the CEEMDAN–Bi-LSTM strategy. For the online prediction setting, a multi-input multi-output (MIMO) rolling multi-step ahead forecasting strategy is adopted, where the low-frequency component (Bi-LSTM-1) is used as the primary prediction target to enhance rolling stability and improve cycle estimation near the failure threshold. Finally, extended experiments are conducted on the NASA 18650 dataset (B05/B06) to validate the applicability of the proposed method across different battery samples. Experimental results show that the CEEMDAN–Bi-LSTM approach can track the capacity-fading trend more stably under different prediction starting points in online prediction, and achieves favorable performance in both offline and online tasks.