A Design-Driven Machine Learning Approach for Invariant Mining in a Smart Grid

Abstract

The study reported here was undertaken to investigate how invariants (rules) that govern the operation of a critical infrastructure can be mined automatically. Unsupervised machine learning was used to extract invariants from data generated by a testbed that includes electric power generation, transmission and distribution. The mined invariants were validated against multiple operational scenarios from the testbed to assess their suitability for future deployment in monitoring and anomaly detection. The complete list of invariants generated is available as a dataset.