Early Uncertainty Quantification Prediction of Lithium-Ion Battery Remaining Useful Life With Transformer Ensemble Model

Early prediction of the remaining useful life (RUL) of lithium-ion batteries remains challenging due to the weak degradation information available in early-stage data. First, a feature extractor that combines convolutional neural networks (CNN) and denoising auto-encoder based Transformers (DAE-Transformers) is proposed, which can automatically extract both local and global degradation information from raw data. Second, a two-stage training ensemble method is proposed to enhance the generalization of early prediction. This method improves the stochastic weighted average (SWA) by incorporating the cosine annealing (CA) strategy, which enables adaptive adjustment of the learning rate. Last, to avoid the problem of overconfidence induced by traditional point prediction methods, we quantify the uncertainty in the RUL prediction with the aid of quantile regression methods. As mentioned above, we proceed to construct a framework that improves the performance of early-stage RUL prediction and named it CDT-CASWA. The experimental results show that the MAPE is 9.23% and 10.52% when using the first 80 cycles for prediction on the primary test set with similar distribution and on the secondary test set with dissimilar distribution to the train set, respectively. Compared to other existing methods, CDT-CASWA has advantages in generalization and accuracy.