Early-stage remaining useful life prediction for lithium-ion batteries based on geometric output construction

The early estimation of the batteries’ Remaining Useful Life (RUL) is important for manufacturers and customers. Technically challenges arise from not only the complexity and nonlinearity of the electrochemical systems but also the lack of clear physical meanings corresponding to the application-specific end-of-life definition. In this paper, we propose a battery RUL prediction algorithm based on data-driven and geometric construction, with a particular emphasis on improving the output of the network. Validated with 180 commercial lithium-ion batteries, our method achieves a typical error of only 4.32%. Even if the training set is tuned far away from the testing set, an accuracy of 5.17% can still be achieved. In addition, the proposed output optimization method can also be integrated into existing data-driven strategies and datasets without further tuning. For instance, by simply adjusting the output mode, an additional 1.8% accuracy improvement can be obtained on the Massachusetts Institute of Technology (MIT) dataset. In contrast to most works focusing on input features and network structures, this paper emphasizes the importance of constructing better outputs for improving the accuracy and generalization of data-driven models, thereby advancing the practical application of data-driven methods in the field of batteries.