Graph neural network based topology recognition and operational state assessment for complex power-grid automation control

Abstract

Accurate topology awareness is a prerequisite for automation-control-oriented monitoring of complex power grids that support electrified transportation loads, distributed energy resources, and frequent switching operations. This paper proposes an edge-gated graph neural network (GNN) framework for joint topology recognition and operational state assessment. Multi-source supervisory control and data acquisition (SCADA) measurements, phasor measurement unit (PMU) streams, switch states, line parameters, and load profiles are synchronized and converted into an attributed power-grid graph. A convolutional neural network (CNN)-style measurement heatmap is used only to expose input saliency, whereas the main inference is performed by topology-aware graph message passing. The edge-gated backbone fuses bus features, branch attributes, and a 12-step temporal window to estimate branch-status probabilities, voltage states, loading levels, and alarm classes. Experiments on the IEEE 33-bus, 118-bus, and 300-bus systems show a topology F1-score of 98.1%, a voltage-state mean absolute error (MAE) of 1.9%, and an inference latency below 6 ms per snapshot. Automated replay and integration tests further demonstrate an end-to-end processing time below 14 ms for the 300-bus case. These results indicate that the framework can serve as a fast validation layer between topology processors, state estimators, and automation-control applications.