Graph Neural Networks for Fast Contingency Analysis of Power Systems

TL;DR

This work tackles the combinatorial challenge of fast contingency analysis in power systems by introducing physics-informed, graph-aware neural networks. It presents two architectures—Guided Dropout Neural Networks (GDNN) and Edge-Varying Graph Neural Networks (EVGNN)—and their physics-informed variants to generalize from $N{-}1$ contingencies to unseen $N{-}2$ and $N{-}3$ topologies without retraining. The proposed methods demonstrate substantial speedups over Newton-Raphson AC power flow (up to $100$–$400\times$) while maintaining competitive accuracy, with the physics-informed EVGNN achieving the best overall performance in voltage and line-flow predictions. The results suggest these topology-aware neural networks are promising for rapid, large-scale contingency screening, with practical implications for real-time grid operation and planning; future work will address data balance, interpretability, and worst-case performance guarantees.

Abstract

The successful integration of machine learning models into decision support tools for grid operation hinges on effectively capturing the topological changes in daily operations. Frequent grid reconfigurations and N-k security analyses have to be conducted to ensure a reliable and secure power grid, leading to a vast combinatorial space of possible topologies and operating states. This combinatorial complexity, which increases with grid size, poses a significant computational challenge for traditional solvers. In this paper, we combine Physics-Informed Neural Networks with graph-aware neural network architectures, i.e., a Guided-Dropout (GD) and an Edge-Varying Graph Neural Network (GNN) architecture to learn the set points for a grid that considers all probable single-line reconfigurations (all critical N-1 scenarios) and subsequently apply the trained models to N-k scenarios. We demonstrate how incorporating the underlying physical equations for the network equations within the training procedure of the GD and the GNN architectures performs with N-1, N-2, and N-3 case studies. Using the AC Power Flow as a guiding application, we test our methods on the 6-bus, 24-bus, 57-bus, and 118-bus systems. We find that GNN not only achieves the task of contingency screening with satisfactory accuracy but does this up to 400 times faster than the Newton-Raphson power flow solver. Moreover, our results provide a comparison of the GD and GNN models in terms of accuracy and computational speed and provide recommendations on their adoption for contingency analysis of power systems.

Figures (10)

• Figure 1: Illustration of the Guided Dropout technique on a NN architecture. One or more conditional neurons can be used to represent the different topologies of the grid.
• Figure 2: Illustration of the 1-hop, 2-hop, and 3-hop neighborhood shift of node A in the 9-bus network.
• Figure 3: Solution framework for the topology-aware PI-GDNN and PI-EVGNN architectures.
• Figure 4: Density distribution of the mean absolute error in voltage predictions for the GD-based and GNN-based models with and without physics included in the training under N-1 test scenarios. Note that different axes have been used for each plot to provide a clearer representation of the results.
• Figure 5: Scatter plots illustrating the relationship between actual and predicted values of voltage magnitude for GDNN and PI-EVGNN models across N-1, N-2, and N-3 contingency scenarios.