Remaining useful life prediction of lithium-ion batteries using a denoising autoencoder and an interval attention LSTM network

Abstract Accurate remaining useful life (RUL) prediction of lithium-ion batteries (LIBs) is essential for ensuring the safety and reliability of energy storage systems. However, the capacity regeneration phenomenon introduces nonlinear fluctuations and irregular temporal dependencies, which significantly degrade prediction accuracy. To address this challenge, a denoising autoencoder and internal attention long short-term memory (DAE-IALSTM) approach is proposed. Firstly, a DAE with stack supervised and reconstructed input is designed to suppress capacity regeneration interference by jointly reconstructing input and target information at each layer, thereby preserving degradation-related features while reducing information loss. Then, an IALSTM is developed to adaptively weight historical samples according to their sampling intervals, enabling effective modeling of both local fluctuations and global degradation trends under irregular time-series conditions. Finally, Experiments conducted on NASA and CALCE public datasets demonstrate that the proposed DAE-IALSTM achieves superior prediction performance compared with state-of-the-art methods. On the NASA dataset, the proposed method achieves zero-cycle prediction error on multiple batteries and reduces MAPE and RMSE by up to 22.14% and 21.94%, respectively, compared with the best baseline. Similar improvements are observed on the CALCE dataset, confirming its robustness and generalization capability. The proposed method provides an effective and practical solution for reliable LIB RUL prediction in the presence of capacity regeneration.