A Fourier graph neural network for SOH estimation of lithium-ion batteries simultaneously considering spatio-temporal features

Lithium-ion batteries have been commonly applied in electric vehicles and renewable energy systems. To ensure the operating reliability and efficiency of lithium-ion batteries, it is imperative to estimate their state of health (SOH) online. Among various SOH estimation methods, data-driven methods have been given considerable attention. However, they often rely on time series neural networks, failing to capture inter-series (spatial) dynamics and intra-series (temporal) dependencies within health feature sequences. In this study, a Fourier graph neural network (FourierGNN) is proposed for SOH estimation of lithium-ion batteries. A hypervariable graph is constructed to represent the spatial and temporal correlations of multivariate features related to capacity degradation of lithium-ion batteries. The node dependencies in the hypervariable graph are refined into fully connected node-to-node dependencies, thus addressing the uncertainty and compatibility issues in spatiotemporal modeling, and establishing adaptive spatiotemporal dependencies. Experimental results show that the FourierGNN model performs well on multiple datasets. Compared with three existing neural networks, FourierGNN model averagely achieves about 35% and 52% reduction in MAE and RMSE errors on NASA B5 battery, respectively. In addition, when the models trained by NASA B5 battery dataset are directly applied to other batteries, FourierGNN can reduce the average MAE and RMSE errors about 50% and 46%, respectively.