The role of physics-informed neural networks in power systems analysis

Anna Varbella – ETH Zurich

We investigate Physics-Informed Neural Networks (PINNs) as unsupervised tools for solving the optimal power flow in power systems. Starting from the fully supervised PowerGraph benchmark, we systematically explore three approaches: fully supervised, fully unsupervised PINNs, and hybrid methods, examining the advantages and limitations of each. Our analysis reveals that supervised approaches encounter challenges with mixed feasible/infeasible data, while pure unsupervised methods may lack the operational precision required for practical applications. The hybrid PINN framework shows promise as a balanced solution, capable of learning from mixed training scenarios while maintaining physics consistency through power flow constraints. Results indicate that PINNs can enhance robustness and interpretability compared to purely data-driven methods, suggesting a potential pathway toward foundation models that balance computational efficiency with physical grounding. Finally, we outline how this approach can be extended to address more complex problems, including multi-period optimal power flow and unit commitment, highlighting the scalability and versatility of the PINN framework for broader power systems optimization challenges.