Ap-Gnn: Unsupervised Adaptive Distribution Grid-Level Representation Learning

Recent episodes of extreme natural events have challenged the ability of power grids to supply demand. Given the increase in the frequency and severity of these events, new methods to evaluate the level of security of power systems are needed. By introducing the emerging deep learning concepts such as attention mechanism for graph neural networks (GNNs) into the power system analysis, we develop a novel approach that systematically classifies this level of security along multiple dimensions. In particular, in contrast to the traditional risk-neutral reliability assessment procedures which focus only the impact of routine failures, our attention-based distribution grid-level representation learning model (AP-GNN) also allows us to simultaneously address the consequences of high impact low probability (HILP) events and to perform unsupervised classification of the distribution grid expansion plans in a computationally efficient manner. Furthermore, we discuss a new tractable resilience metric called Uniqueness Scores which systematically accounts for the key topological characteristics of the heterogeneous distribution grid networks. Our extensive numerical experiments on 54-bus system indicate that the proposed AP-GNN framework is highly competitive both in terms of classification performance and computational efficiency, thereby opening further paths for integration of the state-of-the-art deep learning and artificial intelligence tools to resilience quantification of power systems.