Combined Sensitivity Ranking of Input Parameters and Model Forms of Building Energy Simulation

Abstract

Sensitivity analysis has gradually become an intrinsic part of uncertainty analysis for the identification of key factors affecting the prediction of building performance. Traditional methods include variance-based methods, screening-based methods, and meta-model based methods. However, state-of-the art uncertainty analysis can now explicitly include hidden discrepancies in the models that we use, which typically are not characterized as exposed parameter uncertainty. These structural discrepancies have been characterized as “model form uncertainty”, similar to “model discrepancy” in the statistical realm. Another major source of uncertainty is the scenario of use that the building is subjected. In this paper, we will regard the latter as a special form of model form uncertainty. The reason for this is that scenarios (in weather and occupancy for example) are represented as time series inside the simulation and their role does not fundamentally differ from the uncertainty resulting from physical model simplifications. Then the imminent question is how to rank the sensitivity of both types of uncertainty, i.e. in input parameters and in model form. The need to do so is justified by the argument that spending effort in model improvement that turns low-fidelity modules into higher fidelity ones thus reducing model form uncertainty needs to be justified against the effect of parameter uncertainty. In other words, if the effect of parameter uncertainty is dominant over model form uncertainty, it makes more sense to concentrate on reducing parameter uncertainty. The latter may in some cases be achieved by performing additional measurements for the most sensitive parameters. Both the development of a higher fidelity model as well as conducting better parameter uncertainty quantification are costly. Any investment should therefore be driven by inspecting their relative importance which will drive the prioritization of either approach and single out the parameters or model improvements that have the highest impact on resulting uncertainty in the outcomes of the model. This paper proposes a new sensitivity analysis method that applies group lasso with discrete categorical variables and sliced Latin Hypercube sampling. By applying it on a case study building, we make several important observations, for instance, the sensitivity of infiltration