Thesis

Sensitivity analysis and Bayesian calibration of building energy models

Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2016
Thesis identifier
  • T14302
Person Identifier (Local)
  • 201282515
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • The current state of the art of Building Energy Simulation (BES) lacks of a rigorous framework for the analysis, calibration and diagnosis of BES models. This research takes this deficiency as an opportunity for proposing a strongly mathematically based methodology serving such purposes, providing: a better consideration of the modelling uncertainties, means to reduce BES model complexity without oversimplification, and methods to test and select different modelling hypotheses depending on field observations. Global Sensitivity Analysis (GSA),Gaussian Process Regression (GPR) in a quasi-Bayesian set up and Markov ChainMonte Carlo (MCMC) methods are the foundations upon which the proposed framework is built. It couples deterministic BES models and stochastic blackbox models, thus having the physical and probabilistic representation of real phenomena complementing each other. It comprises four phases: Uncertainty Analysis, Sensitivity Analysis, Calibration, Model Selection.The framework was tested on a series of trials having increasing difficulty. Relatively simple preliminary experiments were used to develop the methodology and investigate strengths and weaknesses. They showed its capabilities in treating measurement uncertainties and model deficiencies, but also that these aspects inuence the estimation of model parameters. More detailed experiments were used to fully test the efficacy of the method in analysing complex BESmodels. Novel techniques, based on Bootstrap and Smoothing with Roughness Penalty, for the determination of the uncertainties of multidimensional model inputs, were introduced. The framework was proven effective in adequately simplifying BES models, in precisely identifying parameters, causes of discrepancies and improvements, and in providing clear information about which model was the most suitable in describing the observed processes.This research delivers a powerful tool for the analysis, diagnosis and calibration of BES models, which substantially improves the current practice and that can be already applied to solve many practical problems, such as the investigation of energy conservation measures, model predictive control and fault detection.
Resource Type
DOI
Date Created
  • 2016
Former identifier
  • 9912521586702996

关系

项目