Parallel State Fusion LSTM-based Early-cycle Stage Lithium-ion Battery RUL Prediction Under Lebesgue Sampling Framework

Remaining useful life (RUL) prediction for lithium-ion batteries in early-cycle stage is of great significance for improving battery performance and reducing losses caused by accidental battery failure. Due to complex degradation mechanism and insufficient recorded data of testing samples, current RUL prediction methods have trouble achieving satisfactory prediction accuracy in the early-cycle stage. In addition, accumulated errors in iterative long-term capacity degradation trajectory prediction make the problem more severe. In view of the above challenge, this paper proposes a lithium-ion battery early-cycle stage RUL prediction method based on Lebesgue sampling parallel state fusion LSTM (LS-PSF-LSTM). First, similar samples are selected by similarity measurement to train a prediction network that is more effective for testing samples. Then, a multistep time series prediction algorithm called parallel state fusion LSTM (PSF-LSTM) is proposed, which makes full use of similar samples to control accumulated errors. Finally, prediction network training and testing are transferred into Lebesgue sampling framework, and LS-PSF-LSTM is proposed to further improve prediction accuracy and efficiency. Experiments using two representative lithium-ion battery degradation datasets are presented to verify the effectiveness of the proposed method. With the starting point of 100 cycles at early stage, the average percentage error of RUL prediction using MIT dataset and Tongji dataset are 3.59% and 6.71%, respectively.