A few-shot and uncertainty-aware method for health status prediction of lithium-ion batteries

Lithium-ion batteries are critical energy storage devices, and their safe, reliable, and efficient operation relies on accurate battery health prognostics. However, accurate prognostics with limited labeled data remain challenging, and conventional transfer learning methods often struggle to provide reliable uncertainty quantification without computationally intensive fine-tuning. To bridge this gap, this article proposes a probabilistic meta-learning method for few-shot battery prognostics. Specifically, a lightweight convolutional neural network is employed to extract compact intra-cycle degradation features from multi-channel charging profiles, and these features are integrated into a shared-encoder attentive neural process with an asymmetric multi-task mechanism. By modeling battery degradation as a stochastic process and using the degradation information shared by state of health (SOH) and remaining useful life (RUL), the proposed method enables joint probabilistic prediction of SOH and RUL, rapid adaptation to unseen batteries through single-forward-pass inference, and efficient deployment without online fine-tuning. Experiments on two public datasets demonstrate that the proposed method achieves accurate few-shot health status prediction with quantified uncertainty. Requiring only a single forward pass and no online fine-tuning, the method offers a balance between predictive performance and computational efficiency for battery health monitoring.