COMPARE AND DETERMINE THE APPROPRIATE CLUSTERING ALGORITHM FOR DIFFERENT BUILDING LOAD PROFILE TYPES

Abstract

Cluster analysis and load profiling are commonly used in the building sector as tools to analyse building energy so that building stakeholders will have a better idea on the type of energy efficiency measures to implement. However, as proven by past research works, the choice of clustering algorithms and determination of a suitable cluster number are difficult issues that have to be considered carefully. Therefore, this research paper aims to provide a guideline for building stakeholders in choosing the appropriate clustering algorithm to apply to their building data. The objectives of this study are: (1) To carry out an extensive literature review to provide a comprehensive understanding on the clustering algorithms; (2) To determine the appropriate clustering algorithm for specific building data; and (3) To develop a general guideline to choosing which algorithm to apply based on the overall analysis of all the buildings. The approach used in this paper is to run the clustering algorithm over a dataset with a range of different values using Python and its libraries. Different partitions will be generated and the partition with the highest cluster validity index (CVI) will be considered to be the best partition, and the corresponding number of clusters is noted. Following this, the cluster plot and load profile are generated. Visual analysis of the cluster plot and load profiles are carried out to determine the “correct” number of clusters. If the computed number of clusters is the same as the observed number of clusters, the algorithm is concluded to be suitable for that building type. The three main findings of this research are (1) any clustering algorithm applied to distinct and spherical clusters will achieve the same results; (2) HDBSCAN applied to distinct and non-convex will achieve the best results; (3) K-means applied to non-distinct and convex data will achieve the best results. Building stakeholders can reference this guideline and apply the appropriate clustering algorithm to their building data that will generate the best results for analysis and subsequent implementation of measures.