Abstract
This paper examines the use of manifold learning in the context of electric power system transient stability analysis. Since wide-area monitoring systems (WAMSs) introduced a big data paradigm into the power system operation, manifold learning can be seen as a means of condensing these high-dimensional data into an appropriate low-dimensional representation (i.e., embedding) which preserves as much information as possible. In this paper, we consider several embedding methods (principal component analysis (PCA) and its variants, singular value decomposition, isomap and spectral embedding, locally linear embedding (LLE) and its variants, multidimensional scaling (MDS), and others) and apply them to the dataset derived from the IEEE New England 39-bus power system transient simulations. We found that PCA with a radial basis function kernel is well suited to this type of power system data (where features are instances of three-phase phasor values). We also found that the LLE (including its variants) did not produce a good embedding with this particular kind of data. Furthermore, we found that a support vector machine, trained on top of the embedding produced by several different methods was able to detect power system disturbances from WAMS data.
Petar Sarajčev
Full Professor | Department of Power Grids and Substations
Dino Lovrić
Associate Professor | Department of Theoretical Electrical Engineering and Modelling
Associate professor at the Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture in Split, with reserch focused on the development of numerical models of grounding systems in various types of soil, particularly in scenarios involving the dissipation of alternating current and transient currents caused by lightning strikes or switching overvoltages, also involved in developing models of dynamic and transient processes in power systems using modern numerical methods.