Emulation-based uncertainty and sensitivity analysis of a multi-zone dwelling

Computer models of building thermal phenomena are increasingly coupled with the other domain specific models, for instance of occupants' stochastic behaviours, improving the plausibility of energy performance simulation results, but at the cost of increased complexity, in terms of the number of required inputs and computational expense. Model complexity may also result in unanticipated predicted behaviours, as the inputs are inherently uncertain. To improve the reliability of energy predictions under uncertainty, and thus to contribute to efforts to address the so-called energy performance gap, a pragmatic treatment of uncertainty in building energy simulation should: 1) simulate the involved thermal phenomena at an appropriate level of complexity, 2) rank the importance of and quantify the effects of parameters in the uncertain parameter space, accounting for their interactions, and 3) do this in a computationally tractable manner.

To this end, this paper introduces a new two-stage Emulation based Uncertainty and Sensitivity Analysis (EmUSA) framework. The first stage involves a systematic dimensionality reduction of the uncertain parameter space, by segregating the most influential parameters from their less-influential counterparts, using parameter screening. In the second stage a Gaussian Process emulator of a stochastic building performance simulator (EnergyPlus coupled with the Multi-Agent Stochastic Simulator No-MASS) is constructed and applied to the reduced parameter space, to efficiently perform a global uncertainty quantification study. After describing this new framework, we demonstrate its application to study both aleatory and epistemic uncertainty to inputs describing a multi-zone residential building energy model; closing with a discussion of the results.