Multi-kernel Analysis Paradigm Implementing the Learning from Loads Approach for Smart Power Systems

Abstract

The future of electric power grid infrastructure is strongly associated with the heavy use of information and learning technologies. In this chapter, a new machine learning paradigm is presented focusing on the analysis of recorded electricity load data. The presented paradigm utilizes a set of multiple kernel functions to analyze a load signal into a set of components. Each component models a set of different data properties, while the coefficients of the analysis are obtained using an optimization algorithm and more specifically a simple genetic algorithm. The overall goal of the analysis is to identify the data properties underlying the observed loads. The identified properties will be used for building more efficient forecasting tools by retaining those kernels that are higher correlated with the observed signals. Thus, the multi-kernel analysis implements a “learning from loads” approach, which is a pure data driven method avoiding the explicit modeling of the factors that affect the load demand in smart power systems. The paradigm is applied on real world nodal load data taken from the Chicago metropolitan Area. Results indicate that the proposed paradigm can be used in applications where the analysis of load signals is needed.