A spatiotemporal graph attention network-based SOH estimation method for lithium batteries in cross-temperature scenarios

Accurately estimating the state of health (SOH) of lithium batteries in complex scenarios remains highly challenging for battery management systems. Traditional data-driven methods, while effective within constant or limited temperature ranges, do not address the complex issues arising in dynamic temperature environments. Therefore, a dataset covering different temperatures was established to study the correlations of battery SOH changes across a wide temperature range. A graph neural network-based SOH estimation method was proposed, which achieves accurate SOH trend predictions for batteries at other temperatures using 25 °C battery data. First, 27 relevant battery degradation features were comprehensively summarized and proposed. The contribution of each feature was quantified via an extreme randomized tree model. The recursive feature elimination method based on SHAP thresholds was subsequently used to iteratively remove unimportant features. By calculating the covariance matrix between the features within a dynamic feature window range, a dynamic minimum spanning tree (MST) graph structure was constructed via the Kruskal algorithm. Finally, an ST-GAT based on a GNN was introduced, and an improved asynchronous successive halving algorithm was used to alleviate the hyperparameter search issues caused by numerous GNN hyperparameters. The experimental results revealed that the proposed method achieved high accuracy and robustness across multiple standard error metrics.