Koopman-Based Capacity Prediction for Health Assessment of Lithium-Ion Batteries

Abstract Lithium-ion battery capacity loss reflects coupled electrochemical ageing, while gradual decline and local fluctuations complicate prediction. This study developed a Koopman-based network combining linear evolution of a trend-derived latent state with nonlinear residual correction. A one-sided moving average separated capacity-derived inputs, and a bounded increment updated the last observed state of health. Retrospective evaluation on 11 cells from three datasets yielded 33 matched tasks and a mean absolute error of 0.004892 ampere-hours, 8.4% below the lower persistence/drift error selected per task as an evaluation oracle. Dataset-specific reductions were 13.8% on NASA and 4.5% on TJU, while CALCE error was 0.1% higher. Both reference rules outperformed all ten comparison networks. Predicted increments were approximately drift-like rather than tracking individual cycle changes. Complementary analysis of 452 paired capacity and impedance records from four NASA cells associated declining state of health with increasing electrolyte resistance; charge-transfer resistance showed cell-dependent behaviour. The common electrolyte-resistance trend and differing interfacial responses provide electrochemical context for the capacity target. Observation-updated predictions, centred preprocessing and overlapping validation windows limit the findings to retrospective one-step assessment. The gains are dataset-dependent, and their operational value requires further validation.