Graph Convolutional Neural Networks for Optimal Load Shedding under Line Contingency

Abstract

Power system operations under contingency need to solve large-scale complex nonlinear optimization problems in a short amount of time, if not real time. Such nonlinear programs are computationally challenging and time-consuming and thus do not scale with the size of the power system network. We apply a graph convolutional network (GCN) model, as a supervised learning model, for predicting an optimal load-shedding ratio that prevents transmission lines from being overloaded under line contingency (i.e., line tripping). In particular, we exploit the power system network topology in the GCN model, where the topology information is convoluted over the neural network. Using IEEE test cases, we benchmark our GCN model against a classical neural network model and a linear regression model and show that the GCN model outperforms the others by an order of magnitude.