[Paper Review] Graph neural networks for power grid operational risk assessment under evolving grid topology

Yadong Zhang, Pranav Karve+1 more|arXiv (Cornell University)|May 12, 2024

Smart Grid and Power SystemsEngineering3 citations

TL;DR

This paper proposes a graph neural network (GNN) framework to predict power grid operational risk hours in advance, using stochastic forecasts of wind, solar, and load as inputs and MILP-based optimal power flow solutions as targets. The GNN achieves high accuracy in predicting system-level, zonal, and branch-level risks (e.g., load shedding and overloading), serving as a fast, reliable surrogate for computationally expensive MILP solvers in real-time risk assessment.

ABSTRACT

This article investigates the ability of graph neural networks (GNNs) to identify risky conditions in a power grid over the subsequent few hours, without explicit, high-resolution information regarding future generator on/off status (grid topology) or power dispatch decisions. The GNNs are trained using supervised learning, to predict the power grid's aggregated bus-level (either zonal or system-level) or individual branch-level state under different power supply and demand conditions. The variability of the stochastic grid variables (wind/solar generation and load demand), and their statistical correlations, are rigorously considered while generating the inputs for the training data. The outputs in the training data, obtained by solving numerous mixed-integer linear programming (MILP) optimal power flow problems, correspond to system-level, zonal and transmission line-level quantities of interest (QoIs). The QoIs predicted by the GNNs are used to conduct hours-ahead, sampling-based reliability and risk assessment w.r.t. zonal and system-level (load shedding) as well as branch-level (overloading) failure events. The proposed methodology is demonstrated for three synthetic grids with sizes ranging from 118 to 2848 buses. Our results demonstrate that GNNs are capable of providing fast and accurate prediction of QoIs and can be good proxies for computationally expensive MILP algorithms. The excellent accuracy of GNN-based reliability and risk assessment suggests that GNN models can substantially improve situational awareness by quickly providing rigorous reliability and risk estimates.

Motivation & Objective

• Address the challenge of real-time, hours-ahead operational risk assessment in power grids with high renewable penetration and uncertain topology.
• Develop a data-driven surrogate model that bypasses the high computational cost of solving mixed-integer linear programming (MILP) problems for optimal power flow.
• Enable comprehensive risk quantification at system, zonal, and branch levels, including standalone and multi-step failure events.
• Demonstrate the feasibility and accuracy of GNNs in predicting risk metrics such as load shedding and branch overloading under evolving grid topology.

Proposed method

• Train GNNs using supervised learning on synthetic data generated from spatio-temporally correlated stochastic forecasts of wind, solar, and load demand.
• Use mixed-integer linear programming (MILP) to solve optimal power flow problems and generate ground-truth outputs (quantities of interest) for training.
• Train multiple GNN models to predict system-level, zonal, and branch-level quantities of interest (QoIs), including load shedding and branch power flows.
• Apply Monte Carlo sampling to the GNN-predicted QoIs for probabilistic reliability and risk assessment, evaluating standalone and multi-step overloading probabilities.
• Use Cholesky decomposition and copula-based sampling to generate realistic, correlated stochastic scenarios of grid variables.
• Validate GNN predictions against MILP solutions using case studies on synthetic grids with 118 to 2848 buses, including Case1354pegase and Case2848rte.

Experimental results

Research questions

• RQ1 Can GNNs accurately predict system-wide and branch-level operational risk metrics (e.g., load shedding, overloading) hours in advance without explicit knowledge of future generator on/off status?
• RQ2 How well do GNNs generalize across different grid sizes and topologies when trained on synthetic, probabilistically forecasted data?
• RQ3 What is the predictive accuracy of GNNs compared to exact MILP solutions for key reliability indicators like overloading probability and consequence cost?
• RQ4 To what extent can GNNs support fast, risk-informed decision-making by replacing computationally intensive MILP solvers in real-time grid operations?
• RQ5 How do standalone versus multi-step failure event probabilities compare in GNN-based risk assessment, and what insights do they provide for grid resilience?

Key findings

• GNNs achieved prediction accuracy within negligibly small error margins compared to MILP-based solutions for system-level and branch-level risk metrics.
• The probability of branch overloading predicted by the GNN was 1.0 for the most vulnerable branch across all time steps in Case1354pegase, matching the MILP reference solution.
• For the second and third most vulnerable branches, GNN-predicted overloading probabilities were 0.9 and 0.8 in multi-step assessments—slightly elevated but highly accurate compared to MILP results.
• The GNN-based risk quantification for branch overloading was slightly conservative but within negligible error margins, confirming its reliability as a surrogate model.
• The method successfully identified that only the most critical branches (e.g., first three) required inclusion in risk assessment, as others had overloading probabilities below 0.5.
• The framework demonstrated robust performance on large-scale synthetic grids (up to 2848 buses), proving scalability and practicality for real-world operational use.