Cross-Chemistry Meta-Learning with Frequency-Domain Augmentation for Robust Capacity Retention Prediction in Vanadium Redox Flow Batteries

  • Shim, Young-Dae; 
  • Yim, Junghwan; 
  • Liu, Nian; 
  • Kwon, Cheong Hoon; 
  • Choi, Hae-Jin; 
  • 외 1명
Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Vanadium redox flow batteries (VRFBs) are promising for long-duration energy storage, but reliable capacity-retention forecasting is difficult because available VRFB degradation data are scarce and operating protocols vary widely. To address this domain-mismatch problem, this article proposes a cross-chemistry meta-learning framework that meta-trains on lithium-ion battery degradation data and adapts to VRFB capacity trajectories using only early-cycle observations. A cycle-based frequency-domain data augmentation approach based on CEEMDAN decomposition is introduced to increase trajectory diversity while preserving the global degradation trend; the four NASA lithium-ion cell datasets are augmented to 4,000 synthetic trajectories for meta-training. The target domain consists of VRFB datasets collected from prior studies under heterogeneous current densities (100, 120, 160, and 200 mA/cm²), and the same augmentation procedure is applied. A weighted ensemble of five MLP predictors, trained with Bayesian hyperparameter optimization, is integrated into the meta-learning pipeline to further enhance robustness. Across all target VRFB datasets, the proposed method (Meta-Ensemble-Aug) achieves strong generalization, outperforming conventional models trained on single-dataset baselines.

키워드

Battery Capacity; Meta-Learning; Predictive maintenance; Vanadium-Redox Flow Battery
제목
Cross-Chemistry Meta-Learning with Frequency-Domain Augmentation for Robust Capacity Retention Prediction in Vanadium Redox Flow Batteries
저자
Shim, Young-Dae; Yim, Junghwan; Liu, Nian; Kwon, Cheong Hoon; Choi, Hae-Jin; Choi, Seung-Kyum
DOI
10.1109/ACCESS.2026.3737729
발행일
2026
유형
Article
저널명
IEEE Access
권
14
페이지
160377 ~ 160394

파일 다운로드

Thumbnail