Prediction of Cascading Failures and Simultaneous Learning of Functional Connectivity in Power System

Abstract

The prediction of power system cascading failures is a challenging task, especially with increasing uncertainty and complexity in power system dynamics due to integration of renewable energy sources (RES). Given the spatio-temporal and combinatorial nature of the problem, physics based approaches for characterizing cascading failures are often limited by their scope and/or speed, thereby prompting the use of a spatiotemporal learning technique. This paper proposes prediction of cascading failures using a spatio-temporal Graph Convolution Network (GCN) based machine learning (ML) framework. Additionally, the model also learns an importance matrix to reveal power system interconnections (graph nodes/edges) which are crucial to the prediction. The elements of learnt importance matrix are further projected as power system functional connectivities. Using these connectivities, insights on vulnerable power system interconnections may be derived for enhanced situational awareness. The proposed method has been tested on a modified IEEE 10 machine 39 bus test system, with RES and action of protection devices.