Interpretable Multi-Feature Fusion for Lithium-Ion Battery State-of-Health Prediction Using ICEEMDAN-Autoformer-BiTCN

Abstract

Accurate state-of-health (SOH) prediction is essential for the safe and cost-effective operation of lithium-ion battery systems. However, current data-driven methods still struggle to jointly extract multiscale degradation information, capture long-range dependencies, fuse heterogeneous health indicators, and provide transparent model decisions. This study proposes an interpretable multi-feature SOH prediction framework that integrates improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), Autoformer, and a bidirectional temporal convolutional network (BiTCN). ICEEMDAN adaptively decomposes capacity-related sequences into intrinsic mode functions and residual components, thereby suppressing noise and representing degradation dynamics at multiple temporal scales. Autoformer captures long-term trends and periodic patterns through series decomposition and auto-correlation, whereas BiTCN models local dynamic fluctuations through bidirectional temporal convolutions. The decomposed components and health indicators are then fused for battery health-state prediction. Shapley additive explanations (SHAP) are used to quantify feature contributions, and interval prediction is introduced to evaluate predictive uncertainty. Experiments on CALCE lithium-ion battery datasets show that the proposed ICEEMDAN-Autoformer-BiTCN model outperforms the tested baseline combinations, achieving an MAE of 6.6646 and an R2 of 0.9896 under ICEEMDAN decomposition. SHAP analysis further indicates that residual components, SOH-related information, cycle number, and internal resistance make consistent contributions across battery samples. These results suggest that the proposed framework improves prediction accuracy while providing a more interpretable and uncertainty-aware approach to battery health assessment.