PhyGNN: Physics guided graph neural network for complex industrial power system modeling

In multi-dimension time series (MTS) tasks within industrial scenarios, several challenges arise due to the difficulty of establishing physical models, the scarcity of high-quality data, and the high demands for model accuracy, robustness, and interpretability. Traditional physical models and pure neural networks exhibit certain limitations in dealing with these challenges. Physics informed neural networks (PINN) have emerged to alleviate these issues. However, in complex industrial power systems (CIPS), classical PINNs present new challenges. The physical laws governing CIPS are vast and extremely intricate. If these laws are converted into loss terms, the loss function becomes complex, redundant, and hard to optimize, even generates conflicting gradient directions and pathological optimization curvature. To address this challenge, we propose a physics guided graph neural network (PhyGNN). One advantage of graph structures is their natural representation of complex systems like CIPS. PhyGNN utilizes this capability as a bridge to integrate physical information directly into the model architecture rather than embedding it into the loss function. Specifically, the spacecraft power system (SPS) is selected as a case study, which is a typical CIPS. First, its physical model is constructed, which includes eight subsystems and deploys diverse fidelity strategies. Then, the physical knowledge of this model is embedded into the proposed PhyGNN. Finally, various comparative experiments and visual analyses are performed on our dataset XJTU-SPS. Overall, the core contribution of this work lies in a physics guided GNN method. Meanwhile, it also contributes a comprehensive physical simulation model for power systems, and a dataset of spacecraft power systems.