Power System Failure Identification Using Gaussian Processes
Thomas Catanach, California Institute of Technology
Identifying and characterizing the effects of a power line failure helps operators avoid blackouts by letting them adjust power flow to avoid future faults. To do this, voltage phasors are measured at a limited set of buses outfitted with phasor measurement units (PMUs). This work seeks to develop a method to use this data to infer possible line failures through the tools of machine learning and Bayesian inference. Simulated time series are analyzed to extract key frequency features and then these features are used to build a Gaussian process classifier. Several methods are used to find the posterior distribution of the classifier, including different stochastic methods such as Hamiltonian Monte Carlo and deterministic methods such as Expectation Propagation. Gaussian process classifiers become computationally intractable for large datasets so the efficacy of parallelization and sparse Bayesian learning are explored to reduce the computation time. This classifier is applied to a 37-bus test system used in previous studies and their results are compared. The insights gained are then used to determine the next stages of this work.
Abstract Author(s): Thomas A. Catanach