Predicting Power System Frequency Health Index with PMUs and Graph Attention Networks

This paper develops a power system frequency health index prediction model using the graph attention network (GAT). A range of pre-contingency grid operating conditions at different loading levels and diverse contingency information serve as training inputs to infer the grid health in terms of system frequency. The frequency nadir is leveraged for the frequency health indicator and its corresponding labels for training samples. The dynamic attention mechanism is incorporated in the graph convolutional network (GCN) model to learn the varying correlation levels between substations/buses effectively. The extended IEEE 118-bus model is adopted to evaluate the frequency health index prediction performance of the proposed GAT and baseline methods. The prediction accuracy of the GAT is 12% higher than that of the baseline method. It is demonstrated that the GAT model is much more robust than the GCN model against erroneous phasor measurement unit (PMU) measurements and partial observability from PMUs.