Comparison of integrated clustering methods for accurate and stable prediction of building energy consumption data

• Published in: Applied energy Vol. 160; no. C; pp. 153 - 163
• Main Author: Hsu, David
• Format: Journal Article
• Language: English
• Published: United Kingdom Elsevier Ltd 15.12.2015; Elsevier
• Subjects: algorithms; Buildings; Cluster stability; Cluster-wise regression; data collection; energy; ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATION; Energy consumption; Latent class regression; MATHEMATICS AND COMPUTING; New York; prediction; Prediction accuracy; regression analysis; New York; Prediction accuracy; Cluster-wise regression; Energy consumption; Cluster stability; Latent class regression; Buildings
• ISSN: 0306-2619; 1872-9118
• DOI: 10.1016/j.apenergy.2015.08.126

•Summary of recent work identifying sub-groups of energy consumption in buildings.•Clusterwise (or latent class) regression gives superior prediction accuracy.•K-means gives more stable clusters when the correct number of clusters is chosen.•A tradeoff between prediction accuracy and cluster stability seems to exist. Clustering methods are often used to model energy consumption for two reasons. First, clustering is often used to process data and to improve the predictive accuracy of subsequent energy models. Second, stable clusters that are reproducible with respect to non-essential changes can be used to group, target, and interpret observed subjects. However, it is well known that clustering methods are highly sensitive to the choice of algorithms and variables. This can lead to misleading assessments of predictive accuracy and mis-interpretation of clusters in policymaking. This paper therefore introduces two methods to the modeling of energy consumption in buildings: clusterwise regression, also known as latent class regression, which integrates clustering and regression simultaneously; and cluster validation methods to measure stability. Using a large dataset of multifamily buildings in New York City, clusterwise regression is compared to common two-stage algorithms that use K-means and model-based clustering with linear regression. Predictive accuracy is evaluated using 20-fold cross validation, and the stability of the perturbed clusters is measured using the Jaccard coefficient. These results show that there seems to be an inherent tradeoff between prediction accuracy and cluster stability. This paper concludes by discussing which clustering methods may be appropriate for different analytical purposes.