MP-Grid: Detecting power grid outages with topological machine learning

A resilient power network is the cornerstone of a secure and economically stable society. With the frequency of power network outages rising due to extreme weather events and cyber-physical attacks, it has become imperative to detect these occurrences in the power grid promptly. To address this, we propose a novel approach for outage detection and resilience improvement of power distribution networks, leveraging the latest tools of topological data analysis which is an emerging direction in graph representation learning. Specifically, we introduce multiparameter persistent homology to smart grids, which enables capturing the finer topological patterns within the network through the utilization of multiple user-defined functions (such as bus voltage and branch currents). By using bus voltages and branch flows as filtration functions, the multiparameter persistent homology summaries capture how outages fragment the grid into disconnected regions, linking topological signatures directly to physical state changes. Our model demonstrates superior performance when compared to existing methods, with an average improvement of 2.66%, 3.73%, and 6.34% over ten other baseline models for the IEEE 37-bus, IEEE 123-bus, and 342-node LVN networks, respectively. The efficacy of our proposed topological machine learning model is also validated across large-sized realistic networks such as the IEEE 8500 bus and NREL’s synthetic San Francisco Bay Area networks. The computational efficiency and scalability exhibited by the proposed model underscore its practical utility and effectiveness in real-time detection capability. • Detecting power outages using topological features and graph learning. • Introducing multiparameter persistent homology for power grids. • Achieving scalability for large, realistic power grids.