A transfer learning and graph neural network-based method for SOH prediction of lithium-ion batteries in energy storage systems

Abstract Accurate state of health (SOH) estimation of lithium-ion batteries is vital for battery management systems (BMSs) in energy storage systems. However, due to intrinsic inconsistencies among individual cells, SOH prediction across varying conditions remains challenging. This paper proposes a novel SOH estimation framework that integrates transfer learning and graph neural networks (GNNs). Health indicators are extracted from voltage, current, and temperature curves, and those with strong correlation to SOH are selected via Pearson analysis. A K-nearest neighbor graph is constructed to model the interdependencies among selected indicators. The Graph sample and aggregate-based GNN is then employed to capture both local and global spatial relationships in the feature space, enabling the model to learn richer, structure-aware representations compared to conventional methods. To enhance generalization, the model is pretrained on a source domain and fine-tuned on a target domain via partial parameter freezing. The proposed approach is validated on two heterogeneous datasets, consistently outperforming several baselines. Experimental results demonstrate the method’s superior generalization, spatial feature learning capability, and practical potential for real-world BMS deployment.