Abstract:

Predicting the remaining useful life (RUL) of lithium-ion batteries is crucial for maintaining their safe and reliable performance. This paper addresses several challenges such as nonlinear capacity changes due to recovery and external disturbances, sparse degradation data, and the difficulty in acquiring complete lifecycle data. The study introduces a novel approach that leverages variational mode decomposition and permutation entropy to denoise and reconstruct degradation data for similar batteries, normalizing it for effective model training. Additionally, a rolling prediction strategy is employed, using sliding windows to partition and concatenate training data. It then trains a Transformer network that is proficient at capturing global dependency relationships. For predictions, the initial 10% of the target battery data is iteratively used for rolling predictions. This approach's effectiveness is initially validated using battery capacity datasets from the Center for Advanced Life Cycle Engineering (CALCE) at the University of Maryland, using a leave-one-out evaluation. Experimental results demonstrate that the prediction method yields strong performance metrics, with an average relative error of 2.21% for RUL across four batteries. Additional validation is conducted with battery data from the National Aeronautics and Space Administration (NASA), specifically battery B0005, to assess model generalization. Battery B0005 achieves an RUL relative error of 2.34%, further confirming the method's effectiveness.

Key words: lithium-ion battery, VMD, PE, transformer, rapid life prediction