A Graph Attention Network Framework With Hybrid Node Representations for Lithium‐Ion Battery State of Health Prediction

Accurate prediction of the state of health (SOH) of lithium‐ion batteries is essential for ensuring the safety and efficiency of energy storage systems. This study proposes a multilevel SOH prediction framework named GAT‐HNR. The framework integrates physically interpretable health feature (HF) extraction, an improved stacking ensemble, and graph attention‐based fusion with hybrid node representations. Five categories of HFs are extracted from battery cycling data, and a sliding‐window smoothing technique is applied to reduce short‐term fluctuations and measurement noise. In the first layer, multiple regression models are trained using these features, and the top five performing models are selected as base learners to provide robust and complementary predictions. During the fusion stage, each GAT node concatenates a base learner's prediction with the five HFs to form a hybrid node representation, allowing the network to jointly capture both intermodel complementarity and complex feature interactions. Results on the NASA battery‐ aging dataset show that, compared with a node configuration without hybrid node representations, GAT‐HNR lowers the root‐mean‐square error (RMSE) and mean absolute error (MAE) by 58.6% and 60.0%, respectively.