Physics-Informed Neural Networks for Privacy-Preserving Model Sharing in Power Systems

Data privacy concerns in the power systems sector significantly complicate the sharing and integration of sensitive operational data among various independent entities. This challenge is particularly pronounced when developing system-wide mathematical models, as the reluctance to share sub-system models parametrized by sensitive data hinders effective system analysis and decision-making. To address this issue, this work introduces the application of Physics-Informed Neural Networks (PINNs) to develop surrogate models that accurately replicate power system dynamics without exposing sensitive data, enabling privacy-preserving model sharing. By embedding physical laws into the training process, PINNs utilize both available data and inherent system physics, making them particularly suitable for modeling complex dynamics. We propose a framework for model development, including dataset generation and integration of the physics knowledge during PINN training. As a proof-of-concept, we apply this framework to a simplified Single Machine Infinite Bus (SMIB) system. Case studies demonstrate that the trained PINN model closely follows ground-truth dynamics and consistently achieves higher accuracy compared to generic neural networks, highlighting the potential for accurate, privacy-preserving model sharing and system-wide dynamic simulations in power systems.