Accurate and efficient prediction of the remaining useful life (RUL) of lithium-ion batteries is critical for ensuring their safe operation, preventing catastrophic accidents, and mitigating potential economic losses. Health indicators (HIs) are capable of characterizing the health state of batteries, thereby providing a reliable basis for RUL prediction. To further improve the accuracy of RUL prediction, this study proposes a novel online prediction model for lithium-ion batteries, which integrates multi-HI optimization and residual correction. In this paper, two types of HIs with excellent sensitivity to battery degradation are extracted, namely the constant current charge time (CCCT) and the time interval of equal discharge voltage drop (TIEDVD). To enhance the discriminability of these HIs, a stacked auto-encoder (SAE) is employed for feature refinement. Additionally, a one-dimensional convolutional neural network (1DCNN) is first used to establish the mapping relationship between battery health states and optimized HIs. On this basis, a hybrid model (LSTM-1DCNN) is developed by modifying a long short-term memory (LSTM) network with the pre-trained 1DCNN, aiming to achieve high-precision RUL prediction. Case studies are conducted using two publicly available lithium-ion battery datasets to verify the effectiveness of the proposed model. The results demonstrate that the SAE effectively preserves and enhances the degradation information of CCCT and TIEDVD, and the hybrid LSTM-1DCNN model significantly improves RUL prediction accuracy. Compared with traditional models such as SSALSTM-ARIMA and AUKF, the RMSE values of the proposed model for the B05, B06, and CS2₃6 datasets are reduced by at least 28. 06%, 41. 32%, and 57. 63%, respectively.
Su et al. (Wed,) studied this question.