Physics‐Informed Meta‐Learning With Graph Attention for Lithium‐Ion Battery Remaining Useful Life Prediction

Accurate prediction of lithium‐ion battery (LIB) remaining useful life (RUL) is important for assessing battery degradation and supporting reliable operation of energy systems. However, prediction across different operating conditions remains challenging because of complex spatial–temporal dependencies, limited degradation observations, and difficulties in incorporating degradation‐related physical knowledge. To overcome these challenges, this study presents a physics‐informed meta‐graph attention network (PIMGAN) that integrates degradation feature learning, dynamical regularization, and few‐shot adaptation. Firstly, a multi‐scale graph attention mechanism is combined with adaptive bidirectional temporal convolution to characterize correlations among heterogeneous battery measurements and their temporal evolution. Then, a physics‐guided regulator (PGR) learns a nonlinear differential operator in the latent feature space, whose residual serves as an additional regularization term during model training. Finally, a first‐order meta‐learning strategy is integrated to learn transferable model initialization parameters across multiple degradation tasks, enabling rapid adaptation to unseen battery conditions with limited target‐domain observations while avoiding explicit second‐order meta‐gradient computations. The proposed PIMGAN is evaluated on two publicly available LIB datasets under cross‐cycle and cross‐condition prediction settings. Under cross‐condition evaluation, PIMGAN achieves mean absolute percentage errors of 0.267% and 0.336% on the MIT and XJTU datasets, respectively, demonstrating its effectiveness in battery RUL prediction with limited target‐domain data.