Early prediction of lithium-ion battery capacity and remaining useful life based on cycle-consistency learning and an improved transformer

Accurate prediction of the capacity degradation trajectory and remaining useful life (RUL) of lithium-ion batteries during early operational stages is critical for ensuring safety and reliability. However, the significant dispersion of capacity degradation trajectories, the asynchrony in degradation stages, and the cumulative errors from iterative prediction hinder accurate early-stage predictions. To address these limitations, this study proposes a novel RUL prediction framework. First, to mitigate the impact of significant dispersion in the capacity degradation trajectories of training samples, we propose training the prediction network by selecting similar samples. Cycle-consistency learning is introduced to select effective similar samples in the alignment subspace, which can minimize the impact of unsynchronized degradation stages on the similarity metric. Finally, to minimize the cumulative errors caused by iterative predictions, we propose a novel Transformer encoder enhanced by a denoising autoencoder and gated convolutional unit. The proposed method is validated on three datasets, including two widely used public datasets. Using the first 100 cycles as inputs, the method achieves root mean squared error (RMSE) of 0.0046, 0.0069, and 0.0072 Ah for capacity prediction and absolute percentage error (APE) of 1.36 %, 1.87 %, and 2.98 % for RUL prediction. Comparison with traditional prediction methods and existing similar sample selection methods proves that the method can achieve the best evaluation results on all indicators.