Work

Generative Data Augmentation and Sparse-Attention Informer for Cross-Temperature SOH Estimation of Lithium-Ion Batteries

Year: 2026

Type: article

Abstract: Accurate state-of-health (SOH) estimation is essential for the safe, reliable, and cost-effective operation of lithium-ion battery systems. However, practical SOH modeling remains challenging because battery degradation data are often limited, degradation trajectories vary substantially across temperatures, and high-capacity deep sequence models typically require large, labeled datasets and considerable computational resources. To address these limitations, this study proposes a data-efficient GAN-Informer framework for lithium-ion battery SOH estimation. In the proposed method, a generative adversarial network is employed to augment the feature-space degradation distribution and improve the effective coverage of limited training samples, while an optimized Informer network with sparse multi-head attention learns long-range dependencies between aging-sensitive features and SOH trajectories. Laboratory cycling experiments were conducted under three representative temperature conditions: 20 °C, −20 °C, and 40 °C. The results show that the proposed framework achieves accurate capacity-tracking performance under both within-temperature and cross-temperature prediction settings. Compared with the Box–Cox-LSTM baseline, GAN-Informer consistently reduces MAE and RMSE across all tested datasets; for example, the RMSE decreases from 0.2235 to 0.0847 on CX1 with 50% training data and from 0.2385 to 0.1187 in the CX1-to-CX3 transfer scenario. These findings demonstrate that adversarial data augmentation combined with sparse-attention sequence modeling provides a promising route for robust, small-sample, and cross-temperature battery health estimation.

Source: Batteries

DOI: 10.3390/batteries12100405