Metamodelling in robust low-energy dwelling design

Abstract

Deterministic simulations are commonly used in building design to calculate for instance the energy use. Many influential parameters are however inherently uncertain. As a result, deterministic optimisa- tion unreliably predicts the impact of design measures. In order to improve the optimisation process, uncer- tainties need to be taken into account through a robust design method. Such a simulation based optimisation is often extremely time-consuming and hence unrealistic. To overcome this barrier, metamodelling is of high interest as a metamodel aims to imitate the original numerical model with a simplified fast model. In this pa- per, metamodels are constructed for the energy demand and indoor temperature of a semi-detached dwelling. Polynomial regression and multivariate adaptive regression splines (MARS) are compared as well as the number of training samples. For the current case, MARS models show to be more accurate. Furthermore, the models appeared more reliable for some outputs than for others. To ensure the reliability of the metamodels, a cross-validation strategy is proposed to construct metamodels with as less training sets as possible