Segmentation of Load Profiles by Time Series Feature Extraction
by Carlos Quesada
In the context of Work Package 2 of the WHY project, University of Deusto is building a causal model of the energy demand at household level. The objective of this model is to understand, in a quantitative manner, the decision-making process that leads a user to consume electricity and his/her reactions to changes in energy policies. In order to determine the types of households that exist in terms of energy consumption, a segmentation of load profiles is currently being carried out.
Different data sets of household energy load profiles, obtained from public sources, WHY partners, and stakeholders, are being analysed using a methodology based on time series feature extraction. First, a set of features that best describe the load profiles (such as statistical moments, quantiles, seasonal aggregates, load factors, entropy, etc.) is selected. Then, these features are computed for the entire data set, creating a feature space that is smaller than the original time series space. Finally, a clustering in the feature space is performed.
Some initial results have been obtained using the Low Carbon London data set, publicly available on the London Datastore website. This data set contains energy consumption readings for a sample of 5,567 London households, measured between November 2011 and February 2014. A set of features containing aggregated means of different periods (mainly hours, days of the week and months) has been computed, and an unsupervised clustering based on PCA and k-means has been performed. The images to the right show the average hourly load profiles of two clusters.