A remaining useful life prediction method for lithium-ion batteries based on improved transformer and stochastic process

Accurate prediction of lithium-ion battery remaining useful life (RUL) is crucial for ensuring safety, optimizing maintenance, and improving the reliability of electric vehicles. However, the battery RUL prediction suffers from long range dependence and large uncertainty. Thus, this paper proposes a model based on the combination of the extended long and short-term memory network (xLSTM), Transformer and Wiener Process (xLSTM-Transformer-WP), aiming to improve the reliability. The method first incorporates xLSTM as an encoder of the transformer architecture to capture the long-range dependence and temporal characteristics. The xLSTM-Transformer model is used as a drift function to model the dynamic behavior of battery degradation. In addition, based on the first hitting time (FHT), the probability density function (PDF) of the RUL of lithium-ion batteries is derived to quantify the uncertainty. This study calculate the drift coefficient and diffusion coefficient, and optimize the hyperparameters of the xLSTM-Transformer using the Bayesian optimization. Finally, the effectiveness of the proposed method is verified by comparison experiments with two sets of lithium-ion battery data and other existing popular methods. The experimental results show that the xLSTM-Transformer-WP model can significantly improve the accuracy of battery RUL prediction and provide a reliable uncertainty assessment.