Abstract Recent years have witnessed increased demands on battery systems due to the rapid global transition to renewable energy storage, necessitating reliable performance analysis under diverse conditions, ranging from low-power edge computing and Internet of Things (IoT) applications to high-throughput materials screening in digital twin simulations. However, obtaining sufficient, high-quality battery performance and status-monitoring data remains challenging, particularly with limited or noisy datasets. To address these, this article proposes the PLT-GRU, a physics-informed lightweight Transformer-GRU algorithm for few-shot battery SoH Estimation. That is, a few-shot cross-domain adaptation is developed within the Transformer-GRU architecture to reduce the need for a complete set of different target battery degradation data and improve the predictive robustness. In addition, lightweight architecture is proposed to minimize parameter counts to reduce inference time and memory overhead, reducing reliance on considerable prior knowledge or more intricate architectures. Then, the physics-informed capacity–resistance soft coupling is proposed to gather additional data for model training by capturing crucial physical priors, including internal resistance, Constant Current Charge Test (CCCT), Constant Voltage Charge Test (CVCT), and battery capacity, improving stability and interpretability with few-shot data. In order to demonstrate the advantage of the proposed PLT-GRU, the experiment conducted partial-cycle experiments on three kinds of lithium-ion batteries with limited data to detect their complex degradation patterns. The experimental results showcase that the proposed PLT-GRU reduces complexity by 94.93% while maintaining high prediction accuracy, effectively adapting to new battery types using few-shot target battery datasets. The proposed method attains average root-mean-square error (RMSE) and mean absolute error (MAE) of 5.87% and 4.30% for the CS2 battery series, 4.26% and 3.01% for the CX2 series, and 1.43% and 1.06% for the NASA battery series.
Yin et al. (Mon,) studied this question.