State of health prediction of lithium-ion batteries with adaptive loss-based graph neural network

Accurate state-of-health (SoH) prediction of lithium-ion batteries (LiBs) is a critical step in LiBs health prognostics to ensure its reliable operation and safety. Recently, there has been an increase in graph neural network (GNN) adaptation across many fields including LiBs health prognostics, which have shown benefits of spatio-temporal-based neural networks over the traditional counterparts. However, many of these studies have been shown to rely on discharge profiles as part of their input features (health indicators, HIs), which in real life would pose a major challenge as a result of high variations in discharge profiles from different tasks. Additionally, many studies utilised traditional loss functions, where they are prone to under-fitting due to data outliers, and some functions require a form of manual hyper-parameter tuning. This paper simplifies the methodologies of previous studies in several aspects; the utilisation of the HIs extracted solely from LiB voltage and current charge profiles explores the potentials of Barron’s adaptive robust loss function by combining it with the GraphSAGE (SAGE)-based graph neural network architecture with skip-connection-enabled (SAGE-LSTM-SAGE-adaptive/SLS-adaptive) to evaluate its performance with respect to traditional loss functions and architecture. This is the first implementation of adaptive loss function and a GNN with skip-connection in the context of LiBs health prognostics. Based on minimal data from five health indicators (HIs) extracted from the aforementioned charge profiles, this approach provides higher precision and longer-term prediction stability compared to its traditional counterparts when tested on the Zhu et al. (2022) and CALCE datasets. • A skip-enabled graph neural network (GNN) is introduced for the SoH prediction. • Adaptive loss function is implemented on the GNN. • For method unification, health indicators are extracted solely from charging data. • Adaptability and accuracy of SLS-adaptive are validated for SoH prediction.