A Remaining Useful Lifetime Prediction Method based on Mixture Maximum Correntropy and Improved Unscented Particle Filter for Lithium-ion Battery

Prediction of remaining useful life (RUL) provides a useful tool to guarantee the safety and health management of lithium-ion batteries (LIBs). In order to improve the prediction accuracy for long-term prediction, a new RUL prediction method (for short, MMCUPF- BiLSTM) is proposed based on the mixture maximum correntropy based unscented particle filter (MMC-UPF) and bidirectional long short-term memory (BiLSTM) for LIBs with various prediction terms. Firstly, the degradation model is trained iteratively by the MCC-UPF and the prediction errors for battery capacity are calculated. Then, the errors are reconstructed by using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to remove the small fluctuations. Finally, using the error-correction mechanism, the multi-step-ahead RUL prediction results based on the trained model are corrected by predicted errors using the BiLSTM method. In the experiment, three LIBs are used to test the prediction accuracy for long-, medium-, and short-term predictions. The experimental results show that proposed MMCUPF-BiLSTM has much higher prediction accuracy and stable performance for different prediction terms and batteries.