Graph Neural Network Surrogates for Scalable Risk Quantification in Power Grid Operation

Abstract

The electric power grid is a critical infrastructure whose reliable and efficient operation underpins modern society and economic prosperity. However, increasing uncertainty in both electricity supply and demand has introduced significant operational challenges. On the supply side, the growing penetration of renewable energy sources such as wind and solar introduces variability tied to meteorological conditions, while on the demand side, distributed energy resources and shifting consumption patterns complicate forecasting and control. Traditional optimization-based approaches, including stochastic, chance-constrained, and robust formulations, provide frameworks for uncertainty-aware decision-making, but they fall short in delivering explicit, failure mode-specific risk quantification. Sampling-based risk assessment offers a more comprehensive alternative by explicitly modeling the distributions of uncertain variables and evaluating thousands of scenarios through deterministic power system models. Despite its accuracy, this approach is hampered by the computational burden of solving large-scale optimization problems repeatedly. This dissertation addresses these challenges by developing surrogate models for unit commitment and optimal power flow, by leveraging machine learning techniques that exploit the graph structure of power networks. Objectives of this dissertation include (1) developing surrogate models for solving unit commitment problem efficiently, (2) predicting zonal and system-level QoIs without explicit UC solving, (3) building surrogate models for OPF under intact and varying topologies, and (4) quantifying operational risk in power systems using the developed surrogate models. The proposed methodology delivers high-resolution, computationally efficient risk evaluation, enabling operators to balance reliability and cost under uncertainty. The resulting methodology enhances situational awareness, strengthens resilience, and supports the reliable integration of renewable energy.