From energy signature to cluster analysis: comparison between different clustering algorithms

The energy audit on the existing buildings has become a priority in the last years, as consequence of the adoption of the European Directives about building energy efficiency. In particular, in Italy, public buildings are often the most inefficient ones among the stock and, thus, those with the highest potential for improvements. Many methods can be applied to perform an energy diagnosis; one of them is \u201cEnergy Signature\u201d simplified method, ES, described in the Annex B of the technical standard EN 15603:2008. The ES can actually be seen as a very simplified model of the building, based on a linear regression between energy consumption and degree days in a set of reference periods. If applied year after year, the ES allows a fast detection of system faults, changes of use pattern, and to assess the efficacy of different energy management strategies or retrofitting interventions, discounting the effect of weather variations. When the stock of buildings is large, individual energy audits can be too much onerous and time consuming and building simulation impracticable. For this reason, ES can be combined with clustering techniques in order to identify groups of buildings with similar behaviour among which a reference cases can be identified and deeply investigated either experimentally or through detailed building simulation (BS). In this respect, ES and clustering can be seen as the key element to allow the extension of BS also to the analysis of building stocks. In this work, ES and different clustering techniques have been used to analyse a set of 41 schools in the province of Treviso, north of Italy, pointing out the buildings features most affecting their energy signatures through multiple linear regressions. A comparison between two non-hierarchical clustering algorithms, K-means and K-medoids, has been conducted. Particular attention has been paid to the approaches for the evaluation of closeness of schools in the same group and the identification of the reference school for each set. As final outcome of this research, the impact of the clustering algorithms is discussed, in order to assess to which extent the selection of the schools with the most representative energy signatures can be affected by the choice of the data mining techniques