LAUSR

Abstract Grouping techniques are frequently used in urban studies as a means for partitioning building stock heterogeneity. The choice of grouping techniques and selection indices can have a significant influence on the segmentation of building stock and associated identification of representative buildings. To date, this issue has not been systematically examined nor justified. The current paper compares different grouping and partition selection techniques using residential building data from Germany. Three grouping techniques are investigated: supervised, unsupervised and semi-supervised. The unsupervised approach is addressed through three clustering algorithms in this work: agnes, diana and partition around medoids. The semi-supervised approach consists of typology-based seed buildings coupled with a distance-based grouping, while the supervised approach relies on classification rules. The resulting partitions are assessed through multiple criteria: internal indices (CH, Dunn2, Silhouette), external index (F-measure) and impact on heating demand modelling (Heating Demand Error). Results show that the algorithms and the selection indices impact the choice of representative buildings to be modelled. Moreover, considering the F-measure, similarities between the three techniques results were observed on some of the groups. Parameters to account for when selecting a grouping technique are discussed and include the number of groups, group uniformity, and compactness/separation.