Paper Conference

Proceedings of Building Simulation 2017: 15th Conference of IBPSA


Interactive Building Design Space Exploration Using Regionalized Sensitivity Analysis

Torben Østergård1,2, Rasmus Lund Jensen1, Steffen Enersen Maagaard2
1Department of Civil Engineering, Aalborg University, Aalborg, Denmark
2MOE, Consulting Engineers, Aarhus, Denmark First author’s e-mail address Nomenclature Di KS2 maximum distance between two cumulative distributions for ith parameter Di-AB Di between cumulative distribution sets SA and SB Dij Di for jth repetition in the TOM method EE method of elementary effect (Morris’ SA) FF factoring fixing (SA setting based on total effects) FM factor mapping (SA setting) FP factor prioritization (SA setting based on main effects) FR factor ranking (SA setting based on total effects) J number of repetitions in TOM method KS2 Kolmogorov-Smirnov two-sample statistics N total number of Monte Carlo simulations Normative model Danish simulation software Be10 based on ISO 13790 (here combined with regression model for daylight) Overtem-perature thermal comfort penalty output in normative model [kWh/m² floor area] PCP parallel coordinate plot (for real-time analysis) PEAR Pearson’s product-moment correlation coefficient Q number of simulations in random selected subset RSA regionalized sensitivity analysis SA sensitivity analysis SRC standardized regression coefficients (linear regression) SA set of all simulations SB set of behavioural simulations meeting all criteria SN set of non-behavioural simulations Si first order effect (Sobol’s variance-based SA) ST total effect (Sobol’s variance-based SA) SATOR comparable SA measure based on TOR [0
100%] SATOM comparable SA measure based on TOM [0
100%] TOR proposed RSA method used for real-time analysis – both inputs and outputs TOM proposed RSA method to rank inputs according to sensitivity towards multiple outputs

Abstract: Monte Carlo simulations combined with regionalized sensitivity analysis provide the means to explore a vast, multivariate design space in building design. Typically, sensitivity analysis shows how the variability of model output relates to the uncertainties in models inputs. This reveals which simulation inputs are most important and which have negligible influence on the model output. Popular sensitivity methods include the Morris method, variance-based methods (e.g. Sobol’s), and regression methods (e.g. SRC). However, such methods only address one output at a time, which makes it difficult to prioritize and fixate inputs when considering multiple outputs. In this work, the primary outcome is a novel sensitivity method denoted TOM, which relies on Kolmogorov-Smirnov two-sample (KS2) statistics to rank inputs due to their influence on multiple outputs. A secondary method, denoted TOR, provides a real-time sensitivity measure when exploring data with the interactive parallel coordinate plot (PCP). The latter is an effective tool to explore stochastic simulations and to find high-performing building designs. The proposed methods help decision makers to focus their attention to the most important design parameters. As case study, we consider building performance simulations of a 15.000 m² educational centre with respect to energy demand, thermal comfort, and daylight.
Pages: 726 - 735