For electric grids with uncertain demand and renewable generation, repeated probabilistic power flow studies can be too expensive for fast planning loops. I developed multi-fidelity GNN surrogates that learn from a combination of lower-cost and higher-fidelity simulations while respecting the native graph structure of power networks.

Key contributions

• Built graph-native surrogate models tailored to power-network topology rather than generic tabular approximations
• Used multi-fidelity learning to reduce dependence on expensive high-fidelity simulations
• Studied performance under high-dimensional uncertainty and modest topology variation

Impact

This work supports faster probabilistic analysis, risk assessment, and planning for increasingly dynamic power systems where repeated simulation under uncertainty is a core operational requirement.