One-Step-Ahead Lithium-Ion Battery Capacity Tracking Using Multidimensional Health Features and MDRSN-BiGRU-AM

Abstract

Accurate cycle-level battery-capacity tracking can support condition-based battery management, but it must be distinguished from state-of-health estimation and prospective long-horizon remaining-useful-life forecasting. This study formulates a bounded one-step-ahead task in which capacity at target cycle t is estimated using only health information extracted from completed cycles t−L to t−1. Four indirect features are constructed from voltage, current, and time-domain charge/discharge signals. Pearson correlation analysis across eight cells and feature-level ablation identify constant-current discharge time as the dominant capacity-related proxy under fixed-current protocols, while the remaining charging- and voltage-related features provide smaller complementary gains. An MDRSN-BiGRU-AM model is evaluated using chronological within-cell tracking, controlled M1–M4 architecture ablation, repeated seeded runs, and sensitivity analyses. The complete M4 model obtains average RMSE values of 0.0075 Ah on CALCE and 0.0116 Ah on NASA, corresponding to reductions of 17.6% and 15.9%, respectively, relative to M1 (BiGRU) within the architecture ablation. These results support bounded one-step capacity tracking under the evaluated laboratory protocols. Generalization to previously unseen cells, performance relative to fully specified conventional baselines, local capacity-increase regimes, and variable field conditions was not established and remains unresolved. EOL thresholds are used only as descriptive visual references, and no environmental or economic benefit is quantified.