Transformer and Long Short-Term Memory-Based Framework with Nonlinear Feature Selection for Simultaneous Prediction of Lithium-Ion Battery State of Health and Remaining Useful Life

Precise prediction of a lithium-ion battery’s state of health (SOH) and remaining useful life (RUL) is essential for battery-management systems to schedule operations safely and keep performance optimized throughout its service life. However, existing methods often sacrifice either accuracy or generalizability due to reliance on feature engineering, difficulty capturing long-range degradation, weak SOH-RUL coupling, and manual hyperparameter tuning. To address these issues, we develop a neural architecture that fuses a transformer encoder with a long short-term memory (LSTM) network. First, hierarchical correlation analysis selects nonlinear degradation features from cyclic data. Second, a shared-backbone network with dual output heads simultaneously regresses SOH and RUL, explicitly capturing their coupling. Finally, Bayesian optimization identifies optimal hyperparameters within a constrained computational budget. Experiments on public NASA and Oxford datasets demonstrate that, compared to multilayer perceptron, LSTM networks, and transformers, our method achieves MAE, RMSE, and MAPE reductions of up to 47.1%, 45.9%, and 47.0%, respectively, while consistently maintaining superior performance in cross-dataset validation. These findings offer a promising path toward real-time, high-precision, and transferable intelligent prognostic models for battery lifetime prediction.