Lithium-Ion Battery Capacity Prediction via Reconstruction-Error-Guided Hierarchical Decomposition and Frequency-Specific Modeling

Abstract

Accurate lithium-ion battery capacity prediction is critical for reliable battery management. However, measured capacity sequences contain slowly varying evolution, local oscillations, and short-term fluctuations, which makes direct raw-sequence prediction difficult. This study proposes an ICEEMDAN-guided fine-grained VMD and frequency-specific TCN–LR framework, named IFVMD-FSTLR. ICEEMDAN first performs coarse separation, retaining the residual as a slowly varying low-frequency component and aggregating the remaining oscillatory components. VMD then refines the aggregated component into compact higher-frequency modes. These components are treated as signal-scale representations rather than direct indicators of specific electrochemical mechanisms. Guided by the reconstruction-error analysis, TCN is used to model the refined higher-frequency components, while linear regression is used for the smoother low-frequency component. The final prediction is obtained by summing the component-wise forecasts. Experiments on the NASA and CALCE datasets are independently repeated five times, with mean ± standard deviation used to characterize result variability. The RMSE, MAE, and MRE are below 0.0033 Ah, 0.0026 Ah, and 0.18% on NASA, and below 0.0044 Ah, 0.0033 Ah, and 0.62% on CALCE, respectively. Overall, IFVMD-FSTLR achieves lower average errors in most evaluated cases.