State-of-health forecasting of solid oxide fuel cells using physics-informed temporal graph convolutional network

Solid oxide fuel cells (SOFCs) are electrochemical devices that convert the chemical energy of fuel directly into electricity, offering a promising clean energy alternative to combustion-based power generation. However, SOFC degradation remains a major barrier to commercialization, highlighting the need for accurate and interpretable health monitoring tools. This study develops a real-time state-of-health (SOH) forecasting model for SOFCs under Redox cycling using a physics-informed temporal graph convolutional network (TGCN) that integrates voltage and electrochemical impedance spectroscopy (EIS) data. Physically meaningful impedance features are extracted through distribution of relaxation times (DRT) analysis, enabling the network to capture electrochemical process interactions governing degradation. The proposed model achieves a root mean square error (RMSE) of 0.085, corresponding to a 37% reduction in forecasting error relative to a baseline long short-term memory (LSTM) model that excludes impedance information. It also yields consistent improvements across mean absolute error (MAE) and mean absolute percentage error (MAPE), with the coefficient of determination ( R 2 ) reaching 0.94. Trained and validated on eight Redox datasets totaling over 2.1 million samples, the model demonstrates strong generalization and real-time feasibility, with an inference latency below 3 ms. This framework enables interpretable and accurate degradation forecasting, supporting predictive maintenance and extending cell lifespan.