A Hybrid Method for Future Capacity and RUL Prediction of Lithium-Ion Batteries Considering Capacity Regeneration

Accurate prediction of remaining useful life (RUL) is critical to the reliability and safety of lithium-ion batteries. However, challenges frequently arise when using the measured data for RUL prediction, such as degradation data being significantly influenced by noise and difficulties in estimating uncertainty induced by capacity regeneration. To address this issue, a hybrid prediction method to predict battery future capacity and RUL is proposed by combining the adaptive variational modal decomposition (AVMD), permutation entropy (PE), long short-term memory (LSTM) network and Bayesian neural network (BNN). Specifically, the AVMD algorithm is employed to decompose the battery capacity data into the aging trend sequence at low frequencies and the noise and capacity regeneration sequences at high frequencies. AVMD adaptively optimizes the number of decomposition stages and balancing parameters through kernel estimation for mutual information and the relative energy density gradient as the objective function. PE is utilized to adaptively filter the high-frequency and low-frequency sequences while eliminating the noise sequence. The prediction models based on LSTM and BNN are then respectively developed to forecast the aging trend sequence and capacity regeneration sequence. The proposed hybrid method demonstrates broad applicability and minimal prediction error as verified by the application on lithium-ion battery dataset.

Keywords: Lithium-ion batteries, Remaining useful life, Adaptive variational modal decomposition, Permutation entropy, Long short-term memory network, Bayesian neural network.