A Hybrid Transformer Network Optimized by Kalman Filter for State of Health Prediction on Lithium-Ion Battery

Abstract

State of Health (SOH) is a critical metric for evaluating the efficient and reliable operation of lithium-ion batteries (LIBs), although it cannot be directly measured. Accurate SOH prediction throughout the entire lifecycle of LIBs remains a significant challenge, primarily due to severe signal fluctuations and complex degradation mechanisms. In this paper, a novel hybrid Transformer-based architecture for SOH prediction is introduced, termed KF–SAMformer–GRU, which integrates a Kalman filter (KF) optimizer, a sharpness-aware minimization Transformer (SAMformer) model, and a multi-layer gated recurrent unit (GRU). To overcome the challenges of multivariate long-term forecasting, we innovatively integrate SAMformer to extract robust feature indicators, actively mitigating data distribution shifts via sharpness-aware minimization. Furthermore, reversible instance normalization (RevIN) is first introduced to tackle non-stationarity in multi-source datasets, effectively eliminating uncertainty and significantly boosting generalization capability. Complementing this, the KF mechanism is uniquely employed to fuse multi-dimensional features, reducing computational overhead while accelerating training. Finally, the multi-layer GRU precisely refines the mapping between SOH metrics and predicted values. Experimental validation on the NASA and CALCE datasets demonstrates the superiority of our approach, achieving a MAPE of 0.01, an RMSE of 0.91%, and an R2 of 0.99. Notably, the method accurately captures phenomena such as battery capacity regeneration, exhibiting superior performance at peaks and valleys while maintaining high computational efficiency.