An interactive clustering method via metric learning for analyzing energy consumption patterns

Abstract

Accurate analysis of energy consumption patterns is the key to achieving the goal of sustainable building energy efficiency, and clustering analysis in data mining techniques has shown efficacy in this regard. However, traditional clustering methods suffer from several issues including lack of users’ participation and low quality of results. To address these challenges, this paper proposes an interactive clustering method based on metric learning. It allows users to interact with clustering models by manipulating initial clusters and inject their domain knowledge for quality improvement. Meanwhile, a metric learning technique is applied to learn a distance metric based on the initial clustering results to improve the quality and flexibility of clustering. Accordingly, a latent graph that imitates the proximity of consumers can be constructed to propagate the clustering intent of users to larger unseen data. In collaboration with five experts in the field of building energy consumption, case studies demonstrate that our method can effectively recognize four types of daily consumption patterns. Compared with the classical k-means method, the mean variance of data features, the Davies-Bouldin Index (DBI) has decreased by 11.19 % and 21.54 %, respectively, and the Calinski-Harabasz (CH) has improved by 10.72 %.