A method and analysis of predicting building material U-value ranges through geometrical pattern clustering. (December 2021)

• Record Type: Journal Article
• Authors: Klemp, S.; Abida, A.; Richter, P.
• Is Part Of: Journal of building engineering. Volume 44(2021)
• Publication Date: 2021-12
• Subjects: Predictive clustering -- Evaluation of building clustering -- Unsupervised learning -- Value range prediction -- Building data preprocessing; Building -- Periodicals; 690.05
• DOI: 10.1016/j.jobe.2021.103243
• Languages: English
• ISSNs: 2352-7102

Abstract

Abstract: For the optimization of the energy consumption in buildings, a calibrated model is of paramount importance. To calibrate the model, initial value ranges for unknown parameters must be defined which is often done through manual tuning and engineering methods. These values are often inaccurate or not available, thus set arbitrarily. Therefore, in this paper, we examine the possibility of defining thermodynamic value ranges by clustering geometrical building patterns. Two issues are analyzed by the method, building pattern clustering via machine learning and the predictive ability of geometrical clusters. The method involves testing multiple clustering algorithms on features extracted from calibrated commercial buildings. The algorithms are either executed on an untransformed or Box-Cox transformed feature space and then evaluated by their geometrical patterns and U-value ranges. For the assesment of U-value ranges, two new evaluation indices are introduced. The shared nearest neighbor algorithm turns out to be the most promising one for clustering geometrical data, reducing initial U-value ranges by 50% on average. In some applications, it might be undesirable to use the shared nearest neighbor algorithm, as data points are assigned as noise. For these cases a Box-Cox transformation of the data is necessary. Without a transformation, other algorithms were not able to determine any geometrical patterns. The method shows the possibility of determining unique U-value ranges by using geometrical data only. The application of such machine learning approaches enables saving time in determining initial value ranges and further the possibility of accelerating calibrations, as smaller value ranges are used.