Topology-Based Machine-Learning for Modeling Power-System Responses to Contingencies

• Activity Number: 16 - How Statistics and Data Science Help to Quantify Resilience of Power Systems
• Type: Invited
• Date/Time: Monday, August 3, 2020 : 10:00 AM to 11:50 AM
• Sponsor: Section on Statistics in Defense and National Security
• Author(s): Brian W Bush*
• Companies: NREL
• Keywords: machine learning; topological data analysis; power systems; simulation

Abstract

Using measurements of the topological features of power-system operation, we develop a machine-learning model to estimate the impact of contingencies (e.g., component outages) on a power system. The model, which predicts impact metrics such loss of load served, is trained on the simulated response to contingencies, given features measured from a graph-theoretic representation of the power system, topological metrics for that graph, static properties of the power-system components (buses, lines, transformers, generators, and loads), and dynamic properties (i.e., power-flow solutions) of the pre-contingency system. Training datasets comprise ensembles of thousands of power-flow solutions for several moderately sized, generic system models; open-source power-system analysis tools are employed. We utilize high-dimensional visualizations to assess the performance of the model and we summarize strengths and weaknesses.