A modified TimeGAN-based data augmentation approach for the state of health prediction of Lithium-Ion Batteries

Lithium-ion batteries are critical components of zero-emission electro-hydrogen generators (GEH2), where accurate performance prediction is essential for ensuring optimal operation and enabling effective predictive maintenance. Data-driven models have become increasingly prominent for predicting the State of Health (SOH) of lithium-ion batteries due to their high accuracy and reduced development time. However, in hybrid systems like GEH2, where the battery frequently remains inactive while the fuel cell supplies most of the power, the available battery data is limited. This data scarcity presents a significant challenge for achieving accurate SOH prediction. To address this challenge, we propose a novel data augmentation approach that integrates Time-series Generative Adversarial Network with a Transformer and a Gated Recurrent Unit to enhance data availability and improve prediction accuracy. This new approach enhances the model’s ability to capture long-term temporal dependencies within multivariate battery parameters while effectively addressing irregular time intervals, a common challenge in real-world batteries datasets. We evaluated the proposed approach using real-world industrial datasets from four distinct GEH2 batteries and two additional batteries from the publicly available NASA dataset. The performance of SOH prediction was assessed using a Long Short-Term Memory (LSTM) model trained on augmented data generated by various data augmentation techniques. The results consistently demonstrate that our approach outperforms all competing methods, highlighting its superior ability to enhance data for lithium-ion batteries. These findings highlight the effectiveness of our approach in enhancing predictive accuracy and robustness, making it highly suitable for real-world battery applications. • A modified TimeGAN-based data augmentation method for Lithium-ion batteries. • Proposal of a hybrid structure combining a Transformer and a Gated Recurrent Unit. • Incorporation with Long Short-Term Memory for batteries’s SOH prediction. • Application to real-world batteries datasets in zero-emission electro-hydrogen generators. • A comparison study with conventional AI-based data augmentation models.