Optimizing Neural Networks for SoH Prediction in Li-Ion Batteries: CNN, GRU, and GRU+Attention with Optuna

Coronado A.M.; Zegarra F.C. · Proceedings of the 2025 IEEE 32nd International Conference on Electronics, Electrical Engineering and Computing, INTERCON 2025 (2025) · DOI: 10.1109/intercon67304.2025.11244713

Resumen

Accurate prediction of Li-Ion battery State of Health (SoH) remains essential for effective battery management systems and operational safety. This study evaluates three neural network architectures - CNN, GRU, and GRU with attention - trained on raw current and voltage data from the CC-CV charging phase. Optuna hyperparameter optimization fine-tunes each model to extract maximum performance from the dataset. Unlike previous works that focus on individual batches, this analysis utilizes all operational batches from the XJTU Battery Dataset, encompassing diverse cycling conditions across 55 LISHEN 18650 batteries. Results demonstrate that the optimal model (GRU) achieves MAPE errors as low as 0.0234 with R2 values reaching 0.9718 on the 11 cells comprising the testing dataset.

Instituciones: Universidad Nacional Tecnológica de Lima Sur