Data mining for energy analysis of a large data set of flats

To improve the energy efficiency of a large building stock, authority planners and designers need to identify which buildings consume most energy and why. For this purpose, this paper provides a data mining-based methodology for setting decision-making rules to identify patterns of energy consumption for a large data set of flats and evaluate the potential effects achievable by retrofitting actions. The calculated normalised primary energy demand (E _PDn) and the geometrical, thermo-physical and heating system attributes of 92 906 flats are analysed. Firstly, an accurate statistical description of the building stock and its main technological features is provided. Secondly, a supervised classification algorithm to rank flats as ‘low’, ‘medium’ or ‘high’ E _PDn is developed based on the flats’ attributes. To classify E _PDn, reference threshold values are set between the attributes. These values will benefit authority planners and designers when setting performance objectives. Finally, the high-E _PDn flats are analysed in depth through an unsupervised classification algorithm. Thus, intrinsic properties and hidden dependencies are discovered. Moreover, a manageable number of real reference flats representative of the entire high-consumption class are identified. These real reference flats can be used to study the causes of high-E _PDn and propose different energy retrofit actions.