Abstract
The operation of buildings is a major source of greenhouse gas emissions, accounting for 27% of energy-related carbon dioxide emissions globally. There is large potential for reduction in these emissions through energy efficiency improvements and optimised operation of buildings.
Models of building energy consumption are used to form site-specific descriptions of energy consumption for a single building or site, which may include residential, commercial or industrial buildings. The energy consumption is modelled to a particular timescale, for example half-hourly or daily energy consumption, and may incorporate external influences such as weather data. These models can be used to provide insights into a building’s operation, make predictions of energy usage, and measure the impact of energy efficiency and demand reduction initiatives.
Statistical data-driven methods for building energy modelling avoid the time and costs associated with gathering detailed information of a particular building, for example material properties and dimensions, that are required for more complex physics-based simulation methods. A statistical data-driven building energy model typically aims to model the relationship between historical load and external variables such as time-of-day and outdoor temperature. However the existing data-driven methods often fail to capture these effects in a way that represents the physics of the system, leaving room for improvement.
Overall, this thesis aims to make improvements to data-driven building energy models by incorporating key aspects of the physics of building systems, yielding improved accuracy and interpretability. One area of improvement is capturing the physics of solar gain effects. Chapter 4 presents the novel Time-of-Week Solar and Temperature (TOWST) model which builds upon the popular and well-established Time-of-Week and Temperature (TOWT) model by incorporating solar terms that effectively infer the orientation of the building and the actual solar gain received by each glazing façade. The TOWST model accounts for the opposite effect of solar gain on the energy demand depending on whether the building is heating or cooling. Results show a reduction in mean squared error of 30%–72% compared to TOWT, across a range of locations and building types.
In part motivated by the heating and cooling distinction in the TOWST model, Chapter 5 presents the development of a novel hidden Markov model of building energy consumption. Hidden Markov models are useful for modelling the energy consumption of buildings as they capture the underlying operating states, and the differences in characteristics of energy consumption among these states. Example states may include heating, cooling, occupied and unoccupied. The novel method presented is a ‘nested’ hidden Markov model that captures the occupancy effects and thermal effects in a separable manner.
Chapter 6 presents a novel method for producing counterfactual predictions from hidden Markov models of building energy consumption. The counterfactual aims to predict what a building’s energy consumption would have been had an energy efficiency improvement not taken place, used to measure the actual savings achieved by such an improvement. Hidden Markov models present nuanced challenges for modelling the counterfactual. The novel method presented in Chapter 6 utilises statistical sampling techniques and information drawn from the data of the building post-improvement.
In addition to the primary aims of developing models to be used for assessing energy efficiency savings over a long period, methods for assessing short term savings are also explored. When facing constraints due to short term peaks in demand, electricity system operators employ demand response events, where customers are asked to reduce demand and are financially rewarded for their contribution. Measuring this actual contribution is not straightforward as it requires determining what a site’s energy consumption would have been in the absence of a demand response event. Chapter 2 presents a comprehensive literature review of models for short-term demand response events, providing recommendations for which methods are likely to be best suited for application in the New Zealand context.