Journal article
Clustering district heat exchange stations using smart meter consumption data
Transport, Department of Technology, Management and Economics, Technical University of Denmark1
Transport Economics, Transport, Department of Technology, Management and Economics, Technical University of Denmark2
Department of Technology, Management and Economics, Technical University of Denmark3
Sustainability, Department of Technology, Management and Economics, Technical University of Denmark4
Energy Systems Analysis, Sustainability, Department of Technology, Management and Economics, Technical University of Denmark5
Dynamical Systems, Department of Applied Mathematics and Computer Science, Technical University of Denmark6
Department of Applied Mathematics and Computer Science, Technical University of Denmark7
AffaldVarme Aarhus8
CITIES - Centre for IT-Intelligent Energy Systems, Centers, Technical University of Denmark9
Contrary to electricity smart meter data analysis, little research regarding district heat smart meter data has been published. Previous papers on smart meter data analytics have not investigated autocorrelation in smart meter data. This paper examines district heat smart meter data from the largest district heat supplier in Denmark and autocorrelation is identified in the data.
The K-Means algorithm is not able to take autocorrelation into account when clustering. We propose different data transformation methods to enable K-Means to account for this autocorrelation information in the data by using wavelet transformation and autocorrelation features. Our results show that the K-Means yield acceptable clustering results for district heat data when clustering normalized data, inclusion of autocorrelation improves the clustering.
The clusters on normalized data are similar to the wavelet transformed clusters, where the autocorrelation has been accounted for. The clustering achieved with the autocorrelation transformation yields finer clusters through accounting for autocorrelation. We are not able to statistically show a difference between the transformations.
All transformations result in shadowing clusters, but the autocorrelation transformation generates fewer shadow clusters and reduce the number of dimensions from 744 to 24, resulting in a dramatic reduction in K-Means runtime.
Language: | English |
---|---|
Year: | 2019 |
Pages: | 144-158 |
ISSN: | 18726178 and 03787788 |
Types: | Journal article |
DOI: | 10.1016/j.enbuild.2018.10.009 |
ORCIDs: | Tureczek, Alexander Martin , Nielsen, Per Sieverts and Madsen, Henrik |