DeepOPF-GAF: Graph Self-Attention Framework for N-1 Security-Constrained Optimal Power Flow

ICLR 2026 Conference Submission13075 Authors

Keywords: N-1 security-constrained optimal flow, graph attention mechanism, explained variance score, machine learning

TL;DR: apply graph neural network for N-1 security power flow

Abstract: N-1 Security-Constrained Optimal Power Flow (N-1 SCOPF) extends the conventional Optimal Power Flow (OPF) problem by ensuring secure and stable operation in all single-contingency scenarios. Solving OPF directly in large-scale power systems imposes a high computational burden, whereas compact approximation models, particularly multilayer perceptrons, have been introduced to improve efficiency. However, existing small-scale DNN-based models fail to meet the demands of highly dynamic grid topologies and multi-contingency solution requirements, due to their limited adaptability to topological changes and insufficient fitting capability. To bridge this gap, this paper proposes a graph self-attention-enhanced framework to optimize N-1 SCOPF solving. Specifically, a residual-based graph self-attention architecture is proposed to enable topological variation adaptation and scalable network expansion in depth and width. Furthermore, the Explained Variance Score (EVS) is introduced as a direct quantitative metric to evaluate the fitting performance of the proposed framework. Experimental results on the IEEE 9, 118, 300, and 2000-bus systems demonstrate that increasing the scale of the graph self-attention framework effectively enhances its fitting performance on N-1 SCOPF problems.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 13075