In recent years, there has been a sustained effort across many countries worldwide to reduce greenhouse gas emissions from buildings while also improving their environmental resilience. These efforts are often supported by archetype-based building stock models which can inform policymakers on the impact of design and retrofit policies on energy consumption, greenhouse gas emissions and indoor environmental conditions. As with any modelling endeavor, uncertainties are an integral part of archetype-based models, and their reduction and quantification through Bayesian calibration can result in more robust predictions. While some examples of housing stock model calibration for energy use exist, there have been no such studies focusing on the calibration of indoor temperature, despite their importance in informing policies to mitigate the increasing risk of indoor overheating. Motivated by this research gap, this thesis aims to develop a Bayesian calibration framework for archetype-based building stock models of daily, free-floating indoor temperatures during summer periods.
To address this aim, a calibration framework was developed which consists of five steps: (1) Statistical Analysis, (2) Categorical Variable Classification, (3) Stochastic Characterisation, (4) Sensitivity Analysis and (5) Bayesian Calibration. The framework was applied to the UK Housing Stock Model (UK-HSM); a well-established bottom-up, building physics model written in Python that relies on EnergyPlus for dynamic thermal simulations and can provide estimates of summer indoor temperature by dwelling construction and type, and geographic location. In Step 1, indoor temperatures, dwelling and household characteristics of 823 homes monitored during the 2011 Energy Follow-Up Survey were used to identify variables that were statistically associated with summer indoor temperatures. Step 2 identified categorical variables to classify dwellings into groups. In Step 3, probability distributions were identified for continuous model inputs for each group, based on empirical evidence where possible. In Step 4, dwellings were further classified in relation to multimodal distributions and calibration variables were selected. In the last step, a homogeneous group of 26 semi-detached dwellings out of the 193 dwellings monitored during the 2009 4M project in Leicester were calibrated using the priors identified in Step 3.
The calibration framework developed draws from examples within the literature and addresses a common limitation in previous work; it clearly defines what a homogeneous group of dwellings is, based on practical considerations of the calibration process, and provides a method for classifying dwellings according to that definition. In addition, it is the first example of a calibration framework applied to a housing stock model of summer indoor temperature. For an unseen validation period, the framework was shown to significantly improve the model’s predictive ability whilst also capturing the existing uncertainties. As part of Step 3, a novel method of identifying probability distributions that adequately describe an empirical dataset is introduced. In addition, a key methodological advancement is made in the implementation of Bayesian calibration that significantly reduced the method’s computational burden. Furthermore, the analysis conducted in Step 1 provides insights on indoor overheating that go beyond building modelling. The framework can be easily adapted for other models, while the novel methods introduced by this work can be widely used within the field of building stock modelling. Ensuring that models are calibrated efficiently and effectively, while also capturing uncertainties, would result in better-informed decision-making for the construction industry practitioners and policymakers.