A methodology for calibration of building energy models at district scale using clustering and surrogate techniques

Prediction of building energy use, when performed at urban scale, is influenced by the choice of modelling approach, as well as the quality of available data. In the case of data scarcity, one of the main limitations of current urban scale building energy simulation models, is the use of deterministic approaches to specify modelling inputs for entire classes of buildings. This modelling approach is characterised by three major shortcomings: first, data uncertainties are not comprehensively considered; second, a rigorous method of identification of groups of buildings to be populated by similar modelling parameters is not specified; and third, strategies to calibrate the developed energy models are missing. Considering these challenges, the current paper presents a non-deterministic calibration method for groups of buildings. The methodology utilises four techniques: (i) the use of clustering to identify building groups and associated representative buildings within the urban context; (ii) the application of an automatic building energy modelling approach to simulate buildings within these groups; (iii) the application of data-driven models, used as emulators of the dynamic simulation engine, in a computationally efficient manner; and (iv) the exploitation of a non-deterministic Bayesian calibration framework to identify sets of representative parameters for the building clusters. Datasets, based on 2646 buildings from the city of Geneva, Switzerland, are used as a case study. Validation with measured data, shows that energy consumption and energy intensity predictions when considered at an aggregated scale, are within +/– 5%. On an individual building basis, the validation error is within +/– 20% for approximately 70% of the buildings.