Deep Learning for Lithium-Ion Battery State of Health Estimation via Explainable Feature Screening

Abstract

Accurate state of health estimation is essential for the safe and reliable operation of lithium-ion battery management systems, yet practical prediction remains difficult because battery degradation involves strong nonlinearity, local capacity regeneration, and long-range temporal dependence. This study proposes an Informer encoder and Newton-Raphson-optimized extreme learning machine framework (InE-NRBO-ELM) that integrates random forest and SHapley Additive exPlanations feature screening (RF-SHAP) with an Informer encoder and a Newton-Raphson optimizer for extreme learning machine regression. Nineteen candidate health features are first evaluated to retain the most informative inputs and suppress redundant interference. The Informer encoder then learns long-range degradation representations, while the Newton-Raphson optimizer jointly optimizes the extreme learning machine regularization coefficient and hidden node number to improve regression stability. Experiments on the NASA and CALCE datasets show that the proposed method accurately tracks both gradual degradation and local regeneration behavior. Under a 60% training ratio, the model achieved mean absolute percentage error of 0.4801%, mean absolute error of 0.00342, and root mean square error of 0.00460 across eight batteries. Within the controlled NASA and CALCE test conditions, benchmark, robustness and ablation studies demonstrate stable performance and the complementary roles of input screening, temporal representation, and adaptive regression optimization.