@phdthesis{01H3YHXBD56BWM4TZC1NTJM172,
  abstract     = {{With the most recent, unprecedented energy crisis ongoing, in which the strain on energy supply and reserves has skyrocketed the energy pricing and the demand for renewable energy systems, popular attention has been drawn once more to energy use in buildings. Since the first energy crisis in the early 70s and later the awareness of climate change, the building energy sector has been revolving around energy conservation and energy efficiency such that they are no longer just buzzwords. Nevertheless, the road towards net zero energy in buildings is far from completed and further strains are needed.

To do so, simulation models have become increasingly prominent tools to aid decision-making processes and building energy policy making as they allow for the quick evaluation of competing policy options concerning the best energy conservation recommendations in the building sector. 
However, a large number of large-scale statistical studies in different countries on the gap between the real and regulatory calculated building energy use, from the last decade, revealed that the regulatory calculation methods (i.e. simplified (white-box) Building-Stock Energy Models) largely overestimated the real energy use of existing, old residential buildings (thus dwellings where energy conservation measures are most needed), inflated true energy savings and undermined national energy policy making.

The regulatory calculation methods thus prove to be inaccurate predictors of the real energy use in residential buildings. The prediction errors (i.e. the gap between the real and the theoretical energy use in buildings) vary largely from one house and household to the other and above all, the predictions are not accurate on average either. Also the predicted energy savings are rarely achieved.

This PhD-dissertation contributes to research on the gap between the real and theoretical energy use in buildings, focusing on methods to increase the reliability of bottom-up Building-Stock Energy Models using available data-driven techniques. Using data from more than 250,000 Flemish single-family houses, this research builds further on existing studies on the gap between simplified (white-box) regulatory calculation methods and real energy use in buildings by contributing results for Flanders. It further tries to identify suitable data-driven models (white-box/black-box/grey-box) that allow for reliable, accurate and fast predictions of the energy use and energy savings in buildings that can be used at building stock level, aid decision-making processes and allow for building energy policy making. Additionally, this research studies a number of model evaluation techniques for Building-Stock Energy Models during model development and model validation that must assure reliable and robust model results and inferences and thus model quality.

The first part of this PhD-dissertation situates the research in the broader context of energy use in buildings, the existing regulatory performance assessment methods and the gap between the real and regulatory calculated building energy use (Chapter 1). Further, relevant background literature about the use of large-scale building datasets in stock models, the types of building stock models for predicting building energy use and other performance indicators is presented and the reader is introduced to the types and the treatment of uncertainty in BSEMs (Chapter 2).

In Chapter 3, a number of model evaluation techniques for Building-Stock Energy Models are studied that must assure reliable and robust model results and inferences and thus model quality. The chapter proposes a methodology (that is scalable) to apply Uncertainty (UA) and Sensitivity Analysis (SA) to BSEMs with an emphasis on important methodological aspects: input parameter sampling procedure, minimum required building stock size and number of samples needed for convergence and proves the importance of executing a UA-SA in well-thought-out fashion. Also (i) the performance of common UA-SA methods was studied in order to (ii) recommend appropriate (thus reliable, robust and computationally efficient) methods for the the aimed UA-SA target: parameter screening, ranking or indices (based on the evaluations in (i)).

Using data from more than 250,000 Flemish single-family houses, Chapter 4 builds further on existing studies on the gap between simplified (white-box) regulatory calculation methods and real energy use in buildings by contributing results for Flanders. The chapter also describes the stock datasets that are then used in the following three modelling chapters, Chapter 5, 6 and 7, as input and validation data. Results from statistical analysis showed that the overestimation of the real energy use in buildings for space heating and domestic hot water (and thus also the total energy use) was exceedingly large for existing single-family houses compared to studies from other EU countries. The Flemish EPC labels proved to be very poor indicators of the real energy use in residential buildings.

Chapter 5 examines to what extent available aggregated variables explain the real annual energy use in buildings using classical statistical linear regression models and addresses the problem of multicollinearity and the importance of bootstrapped confidence intervals for model quality control. Further, based on the regression coefficients, inferences are drawn about possible causes for the gap between real and regulatory energy use. The results showed that statistical linear models explained only a fraction of all variability and indicated that a significant extent of multicollinearity had to be corrected. For most models, half of the variability has been left unexplained and has to be attributed to variables that were not available, the fact that the data were insufficiently accurate or that the model (structures) were not good enough.

Similarly to Chapter 5, Chapter 6 examines to what extent available aggregated variables explain the real annual energy use in buildings, yet using common black-box machine learning regression techniques such as gradient boosting regression trees and support vector regression. Similar to the results for the linear regression models, the results for the machine learning models showed that only a fraction of all variability could be explained. Half of the variability has been left unexplained and has to be attributed to variables that were not available, inaccurate data or the fact that the model (structures) were not good enough.

Last, Chapter 7 presents a white-box model identification procedure that allows to develop simple dynamic white-box models for the modelling of the individual buildings within bottom-up BSEMs that predict common BSEM outputs (i.e. yearly heat demand, peak load and heat load curve). Dynamic simple white-box RC models are established through pattern searches between the parameter estimates from simple stochastic grey-box models and aggregated white-box measurement data. The identified models showed great performance for predicting the heat demand in residential buildings at stock level. Not only were they able to accurately predict the average and the variability in the yearly net energy use for space heating within the boundaries of the modelled building stock, they also showed accurate predictions of the peak load and a median heat load curve ánd performed relatively well outside the boundaries of the modelled building stock and for different weather data.

Combined, the conclusions from the different chapters, summed up in the eight and last chapter, demonstrate that the the developed (grey-box trained) simple white-box models are assumed to work well within dynamic building stock energy modelling frameworks that allow to study future policy plans. As such, the author believes that the future of building stock energy studies and pathways towards net zero lie in bottom-up BSEMs with grey-box trained simple white-box models under the hood.}},
  author       = {{Van Hove, Matthias}},
  isbn         = {{9789463557290}},
  language     = {{eng}},
  pages        = {{XXVI, 475}},
  publisher    = {{Ghent University. Faculty of Engineering and Architecture}},
  school       = {{Ghent University}},
  title        = {{Increasing reliability of bottom-up building-stock energy models using available data-driven techniques}},
  year         = {{2023}},
}