Fang-Yi Chang¹, Shu-Wei Lin¹, Chia-Wei Tsai² and Po-Chun Kuo², ¹Institute for Information Industry, Taiwan and ²Southern Taiwan University of Science and Technology, Taiwan

Clustering is an useful tool in the data analysis to discover the natural structure in the data. The technique separates given smart meter data set into several representative clusters for the convenience of energy management. Each cluster may has its own attributes, such as energy usage time and magnitude. These attributes can help the electrical operators to manage their electrical grids with goals of energy and cost reduction. In this paper, we use principle component analysis and K-means as dimensional reduction and the reference clustering algorithm, respectively, and several choices must be considered: the number of cluster, the number of the leading principle components, and whether use normalized principle analysis schema or not. To answer these issues simultaneously, we use the stability scores as measured by dot similarity and confusion matrix as our evaluation decision. The advantage is that it is useful for comparing the performance under different decisions, and thus provides us to make these choices simultaneously.

Smart meter; Unsupervised; Nonparameter; Clustering; PCA; Stability; Smart Grid; Value-Add Electricity Services; Energy Saving; Energy management