SC‐PINN for Modeling and Degradation Prediction of Electric–Hydrogen Hybrid Power Systems

The electric–hydrogen hybrid power supply system of data‐center inspection robots involves strong coupling among lithium‐ion batteries, proton exchange membrane fuel cells, DC buses, and energy management controllers. Accurate modeling is challenging because the system exhibits multi‐physics interactions, multi‐timescale dynamics, limited operating data, and degradation‐induced parameter drift. Pure mechanistic models require difficult parameter calibration, while purely data‐driven models often lack physical consistency and interpretability. To address these issues, this study proposed a Separate‐Combined Physics‐Informed Neural Network (SC‐PINN). The proposed architecture built separate subnetworks for the battery, fuel cell, DC bus, and degradation process, and combined them through a unified loss function. Physical constraints, including SOC dynamics, electrochemical polarization, thermal balance, power balance, hydrogen consumption, and SEI‐growth‐based degradation, were embedded into the training objective. A MATLAB/Simulink electric–hydrogen hybrid power supply model was developed, and domain‐randomized operating data were generated for training, validation, and testing. Experimental results showed that SC‐PINN achieved R 2 values of 0.9259–0.9998 across seven electrical, thermal, hydrogen‐consumption, and degradation variables and maintained competitive accuracy relative to LSTM, GRU, Transformer, MLP, XGBoost, SVR, and RF baselines. At the 1800s prediction horizon, the normalized RMSE values for SOC, battery temperature, fuel cell temperature, and fuel cell power were 0.47%, 1.45%, 4.18%, and 0.34%, respectively. Reducing the SOC measurement‐update frequency increased the SOC normalized RMSE to 0.84%, while the mean DC‐bus power‐balance residual remained comparable between the measurement‐updated and state‐propagation settings at 0.862 and 0.875 W, respectively. The cumulative hydrogen‐consumption and degradation variables exhibited greater horizon‐dependent error growth, indicating that their long‐duration performance requires independent validation. These results demonstrate that SC‐PINN provides physically consistent and interpretable joint prediction of heterogeneous subsystem states within the evaluated 1800‐s operating window.