Research on deep learning model of lithium-ion battery life prediction based on DeepAR, LstNet and N-Beats

Lithium-ion battery lifespan prediction is a core function of battery health management systems, directly impacting user satisfaction, safety assurance, and providing guidance for battery replacement. However, two major unresolved issues persist in current research. First, the quantity and quality of comparative experiments are often insufficient, leading to a lack of comprehensive and accurate data to support the reliability of research conclusions. Insufficient comparative experiments can result in less accurate performance evaluations of prediction models, thereby limiting further research and application. Second, although historical usage data of lithium-ion batteries hold significant reference value, this data has not received sufficient attention. Further research is urgently required to uncover key issues related to battery performance, lifespan, and safety. This paper aims to predict the lifespan of lithium-ion batteries using three time-series models-DeepAR, LstNet, and N-Beats-by extracting features closely related to battery capacity from historical data, thus providing reliable comparative support for future research. The experimental results demonstrate that the $\mathbf{N}$ Beats model performed optimally on the HUST dataset, with an MAE of $\mathbf{0. 0 4 7 5}$, RMSE of $\mathbf{0. 0 6 9 0}$, and MAPE of $\mathbf{9. 4 4 1 6\%}$. On the RWTH dataset, the DeepAR model achieved an MAE of 0.0641 and a MAPE of $\mathbf{1 8. 4 3 7 9\%}$, while the N -Beats model excelled in several metrics, with a MAPE of only 12.6011%. Through comparative experiments and analysis of historical data, this study lays a solid foundation for the further development and application of lithium-ion battery health management systems.