• Login
• Admin

Wate, Parag (2020) Emulation based uncertainty and sensitivity analysis for complex building performance simulation. PhD thesis, University of Nottingham.

PDF (Thesis - as examined) - Repository staff only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader Download (5MB) Request Thesis

Abstract

Uncertainties in building performance simulation arise from the propagation of both epistemic (e.g. due to poorly or un- observed input parameters) and aleatory (e.g. due to occupants' stochastic interactions) uncertainties. A good framework for the quantification and decomposition of uncertainties in dynamic building performance simulation should: (i) simulate the principle deterministic processes influencing heat flows in buildings and the stochastic perturbations to them, (ii) quantify and decompose the total uncertainty into its respective sources, and the interactions between them, and (iii) achieve this in a computationally efficient manner. In this thesis, a new framework, named as Emulation based Uncertainty and Sensitivity Analysis (EmUSA) framework, is introduced which does just that. Two case studies, a monozone office building for the frameworks' proof of principle and the other multizone residential building for it's extension are being investigated to perform Uncertainty and Sensitivity Analysis for the Complex case of Building Performance Simulation (BPS) i.e. the Stochastic BPS (S-BPS). The detailed development of this new framework for emulating both the mean and the variance in the response of a stochastic building performance simulator (EnergyPlus co-simulated with a multi-agent stochastic simulator called No-MASS) is presented, for heating and cooling load predictions.

The effectiveness of these emulators, applied to a monozone office building, has been demonstrated and evaluated to quantify and decompose prediction uncertainties. With a range of 25-50 kWh/m2, the epistemic uncertainty due to envelope parameters dominates over aleatory uncertainty relating to occupants' interactions, which ranges from 6-8 kWh/m2, for heating load predictions. However, the converse is observed for cooling loads, which vary by just 3 kWh/m2 for envelope parameters, compared with 8-22 kWh/m2 for their aleatory counterparts. This is due to the correspondingly larger stimuli provoking occupants' interactions. Sensitivity indices corroborate this result, with wall insulation thickness (0.97) and occupants' behaviours (0.83) having the highest impacts on heating and cooling load predictions respectively. This new emulator framework (including training and subsequent deployment), for the case of monozone office, achieves a factor of c.30 reduction in the total computational budget, whilst overwhelmingly maintaining predictions within a 95% confidence interval, and successfully decomposing prediction uncertainties.

The second case of a multizone residential building demonstrates the applicability and readiness of this new framework for the uncertainty problems (pertaining to the building scale) of increasing scope and complexity. An increased scope is analysed by investigating the impact of uncertainties in detailed thermophysical (e.g. conductivity, density, solar transmittance, etc.) properties of opaque and transparent building envelope materials, along with the mostly inexact infiltration rate. For the increased complexity, a greater degree of stochastic occupants' interaction freedom (e.g. use of lights) along with the stochastic phenomena of occupants' behaviours (e.g. presence, windows and shades interactions) was enabled. To this end, the framework has shown a ready extensibility and applicability by incorporating a parameter screening step as a pre-step to the Uncertainty Quantification (UQ) workflow for the purpose of segregating the most influential input parameters from the non-influential ones. For the extended case, the framework is now overall (training and subsequent deployment) moderately computationally efficient as compared to the proof of principle case with the costs reduced to a low fraction (0.25) of its equivalent classical counterpart, with predictivity coefficient Q2 > 0.95. As the occupants' behaviours are relatively constrained during heating season, the effects of uncertainties in wall insulation conductivity (sensitivity index = 0.8428) and thickness (0.0771) dominate the heating load predictions than the stochasticity due to occupants' interactions (0.0010). The similar insignificance of stochastic phenomena (0.0130) as compared to brickwork solar absorptance (0.5032) and wall insulation conductivity (0.3714) has been observed during cooling season, due to the acceleration of short- and long-wave radiation phenomena which then dampens the occupants' interaction to restore their thermal comfort.

Item Type:	Thesis (University of Nottingham only) (PhD)
Supervisors:	Robinson, Darren Iglesias, Marco Coors, Volker
Keywords:	Gaussian process emulator, Building performance, Stochasticity, Uncertainty quantification and decomposition
Subjects:	T Technology > TH Building construction
Faculties/Schools:	UK Campuses > Faculty of Engineering > Built Environment
Item ID:	60208
Depositing User:	Wate, Parag
Date Deposited:	31 Jul 2020 04:40
Last Modified:	31 Jul 2020 04:40
URI:	https://eprints.nottingham.ac.uk/id/eprint/60208

Actions (Archive Staff Only)

Edit View