Researchers have proposed a Fourier graph neural network for estimating the state of health of lithium-ion batteries while simultaneously considering spatial and temporal feature relationships. The model, called FourierGNN, is designed to improve online battery health estimation by capturing both inter-series dynamics and intra-series dependencies in battery degradation data.
Lithium-ion batteries are widely used in electric vehicles and renewable energy systems, where reliable operation depends on knowing how battery health changes over time. State of health, or SOH, estimation helps battery management systems assess degradation, plan maintenance, and support safe and efficient operation. If SOH is estimated inaccurately, a system may overestimate usable capacity, underestimate aging risks, or make suboptimal control decisions.
Data-driven SOH estimation methods have attracted considerable attention because they can learn degradation patterns from measured battery data. However, many existing methods rely mainly on time-series neural networks. According to the article, these approaches can fail to capture inter-series, or spatial, dynamics and intra-series, or temporal, dependencies within health-feature sequences. That limitation can reduce performance when degradation behavior depends on relationships among multiple features as well as changes over time.
The new study addresses this gap by introducing FourierGNN for lithium-ion battery SOH estimation. A central part of the method is the construction of a hypervariable graph that represents spatial and temporal correlations among multivariate features related to capacity degradation. In practical terms, this graph allows the model to treat battery health features as connected signals rather than isolated time-series inputs.
The study further refines node dependencies in the hypervariable graph into fully connected node-to-node dependencies. This is intended to address uncertainty and compatibility issues in spatiotemporal modeling while establishing adaptive dependencies across the feature space. By doing so, the model can learn how different health indicators interact and evolve, which may be especially useful when degradation patterns are complex or differ across datasets.
The Fourier component of the model adds another dimension to this representation. By working with signal patterns in a way that can capture relationships beyond simple time-domain trends, the Fourier graph neural network is positioned to identify degradation information that may be difficult for conventional neural networks to extract. The result is a framework that combines graph-based structural learning with frequency-informed modeling for battery SOH estimation.
The researchers validated the proposed model using three publicly available battery datasets. This multi-dataset validation is important because SOH models often perform well on the specific cells used for training but struggle when applied to other batteries or datasets. A method intended for online battery management needs both accuracy and generalization if it is to be useful across electric vehicle or energy storage applications.
Experimental results reported in the paper suggest that FourierGNN performs strongly on multiple datasets. Compared with three existing neural networks, the FourierGNN model achieved average reductions of about 35% in mean absolute error and 52% in root mean square error on the NASA B5 battery. In addition, when models trained on the NASA B5 battery dataset were directly applied to other batteries, FourierGNN reduced average mean absolute error and root mean square error by about 50% and 46%, respectively.
These results indicate that explicitly modeling spatiotemporal relationships among battery health features can improve both accuracy and transferability. Further validation will still be needed across wider battery chemistries, operating conditions, aging profiles, and real-world battery management systems. Even so, the study offers a strong indication that Fourier graph neural networks could help make lithium-ion battery SOH estimation more robust, supporting safer and more efficient operation of electric vehicles and renewable energy storage systems.
Reference
Author:
Wanglin Liu a b , Jindong Tian a b , Xiaoyu Li b , Yong Tian b , Guang Li c
Title of original paper:
A Fourier graph neural network for SOH estimation of lithium-ion batteries simultaneously considering spatio-temporal features
Article link:
https://www.sciencedirect.com/science/article/pii/S2773153725000519
Journal:
Green Energy and Intelligent Transportation
DOI:
10.1016/j.geits.2025.100301
Affiliations:
a Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen 518132, China
b College of Physics and Optoelectronic Engineering, Shenzhen University, Shenzhen 518060, China
c School of Engineering, The University of Manchester, Manchester, M13 9PL, UK
Green Energy and Intelligent Transportation
Experimental study
Not applicable
A Fourier graph neural network for SOH estimation of lithium-ion batteries simultaneously considering spatio-temporal features
31-Dec-2025