Finding Underlying Structure in Electric Load Data Using Independent Component Analysis
Masaomi Yanagida, Atsushi Ishigame, Atsuhiro Koyama, Nobuo Umeda, K. Yoshida · IEEJ Transactions on Power and Energy · 2008
In this paper we analyze electric load data using Independent Component Analysis (ICA) approach. ICA is a method for finding unknown signals from observed ones. Given that observations are linear mixture of independent signal sources, then we can estimate signal sources from just only observations using ICA method. There are some papers studying time-series data with ICA. They apply ICA to economic data, climate data and topic identification. One of the advantages of using ICA is its ability to capture data structures. Among many ICA algorithms, we adopt Complexity Pursuit (CP). CP is based on maximizing nongaussianity and temporal structure of signals. We aim at finding underlying structure of electric load in Kansai Electric Power Co., Inc. As a result, 13 components are extracted from 2004 to 2006 data. Then we explore the meanings of components, and separate weeks based on each component's fluctuation. The groups are related to each season.