Battery Early Prognostics Based on Pseudo Meta-Learning

Effective management of lithium-ion batteries is pivotal for energy supply systems. However, the significance of battery early prognostics is often overlooked and conventional data-driven approaches frequently fall short in precisely predicting lifespan or reconstructing degradation trajectories during the initial stages of battery life. To address these challenges, this article proposes a pseudo meta-learning (PML) neural network integrating hybrid cyclic-based and physical-informed features for battery early lifespan estimation and capacity reconstruction. A convolutional variational autoencoder is initially employed to train a robustness encoder using sufficient battery early aging data. Furthermore, PML combines features of batteries with different lifespans for checkpoint prediction from full-life battery aging data. To mitigate overfitting concerns, multiple PMLs are interconnected through a gating network, and a joint-learning strategy is introduced to fine-tune hyperparameters. In the experimental validation, comparative experiments on other battery prognostic methods are conducted, and a detailed discussion on enhancing prediction performance, considering both computational complexity and accuracy, is presented.