Unsupervised load shape clustering for urban building performance assessment

This paper presents a method to automatically cluster typical days of energy consumption in one or several buildings. The method is based on an optimized version of the Symbolic Aggregate approXimation (SAX) method. SAX is a data mining technique for clustering time series with recent applications in building fault detection and building performance assessment. The number of clusters and accuracy of SAX highly depends on two highly sensitive input variables, i.e., the word size and the alphabet size. We propose the use of the genetic algorithm NSGA-II to optimize the number of words and alphabet size of SAX subjected to three fitness objectives, i.e., maximize data accuracy and compression and minimize complexity. In addition, we propose the use of MAVT as selection method of the optimal solution. The methodology is applied to measured energy consumption data of three representative buildings on a university campus in Singapore. Potential future uses of the approach include advanced studies in fault detection and calibration of urban building performance models.