Performance comparison between KNN and NSGA-II algorithms as calibration approaches for building simulation models

In this paper, a study of calibration methods for a
thermal performance model of a building is
presented. Two calibration approaches are evaluated
and compared in terms of accuracy and computation
speed. These approaches are the 푘 Nearest Neighbour
(KNN) algorithm and NSGA-II algorithm.
The comparison of these two approaches was based
on the simulation model of the Birmingham Zero
Carbon House, which has been under continuous
monitoring over the past five years. Data from
architectural drawings and site measurements were
used to build the geometry of the house. All building
systems, fabric, lighting and equipment were
specified to closely correspond to the actual house.
The preliminary results suggest that the predictive
performance of simulation models can be calibrated
quickly and accurately using the monitored
performance data of the real building. Automating
such process increases its efficiency and consistency
of the results while reducing the time and effort
required for calibration. The results show that both
NSGA-II and KNN provide similar degree of
accuracy in terms of the results closeness to
measured data, but whilst the former outperforms the
latter in terms of computational speed, the latter
outperforms the former in terms of results wide
coverage of solutions around the reference point,
which is essential for calibration.