Meta-learning based on Hypernetworks for State of Health prediction of lithium-ion batteries under data scarcity

Accurate prediction of the State of Health (SOH) of lithium-ion batteries is crucial for ensuring their safe and reliable operation. However, existing data-driven methods rely on large amounts of data for training, resulting in limited generalization ability under data-scarce conditions. To address this issue, Model-Agnostic Meta-Learning (MAML) has been widely applied, but its two loops optimization mechanism involving inner and outer loops leads to low training efficiency and high computational burden. Therefore, this paper proposes a Meta-Learning based on Hypernetworks (MLH) method. This method extracts Health Indicators (HI) through an attention mechanism and uses a support set encoder and hypernetworks to generate prediction network parameters. Based on this, the prediction network performs gradient descent on the query set and achieves multi-module collaborative updates through end-to-end gradient propagation. Compared to MAML, the proposed method generates prediction network parameters through a single forward propagation for a single task, avoiding multi-step gradient updates within the inner loop, thus significantly reducing computational burden. Furthermore, the model can directly predict SOH requires limited data for rapid fine-tuning under different types of battery conditions. Experimental results based on the MIT, CALCE, and XJTU datasets demonstrate that, under conditions of data scarcity, MLH achieve higher accuracy and lower computational burden compared to MAML and the traditional transfer learning method in the task of predicting SOH of unseen batteries.