Robustness of Machine Learning for Power Systems
Abstract
The applications of machine learning are broad and useful for a variety of industry and militaryobjectives, but are the current methods robust? Robustness requires more than accuracy in ideal conditions;it means the system is resistant to perturbations in the data, both from natural and adversarial causes. Thisresearch aims to analyze the robustness of neural networks used for power-grid fault classifications. Wefocus on data generated from simulations of the classical 9-bus model; however, these methods and resultscan be extended to more complex microgrids, such as those found on naval ships, submarines, and bases.First, we measure the effects of random and adversarial noise on the testing data and compare three networktypes. Then we test different structures by varying the number of nodes and layers. Finally, we test whetheradding noise to the training data can improve robustness. Before machine learning methods are adopted onsubmarines, we must first understand their weaknesses and potential for error. This research provides thefoundation for how to test robustness, where neural networks are at risk from random or adversarial noise,and how to modify networks to improve their robustness.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 01, 2022
- Accession Number
- AD1185010
Entities
People
- Elana P. Kozak
Organizations
- Naval Postgraduate School