GAT-OPF: Robust and Scalable Topology Analysis in AC Optimal Power Flow With Graph Attention Networks

Abstract

As power systems rapidly grow in scale and complexity, existing data-driven methods are limited when applied to large-scale networks due to issues with low prediction accuracy and constraint violations. This paper proposes an innovative hybrid framework, GAT-OPF, which, for the first time, combines graph attention networks (GAT) with deep neural networks (DNN) to form the GAT-DNN model, designed to dynamically adapt to topology changes in the AC optimal power flow (AC-OPF) problem. A hybrid loss function is also developed, combining prediction error with a constraint violation penalty term and incorporating a dynamic Lagrange multiplier adjustment mechanism to ensure constraint compliance throughout training. The model was tested under topology changes on the IEEE 30-bus system and validated for scalability on larger systems, including IEEE 300-bus, 1354-bus, and 9241-bus systems. The results show that the proposed model significantly enhances the computational efficiency of large-scale power systems while effectively balancing high prediction accuracy and low constraint violations without post-processing, highlighting its potential for real-time optimization in large-scale power systems.