Real-time Energy Use Predictions at the Early Architectural Design Stages with Machine Learning

It has been argued that traditional building energy simulation methods can be a slow process, which often fails to integrate into the design process of architects at the early design stages. Furthermore, studies have shown that the actual energy consumption of buildings once built and in operation is often far greater than the energy consumption predictions made during design. The difficulty of simulating real-world systems, such as the stock market or buildings, is the lack of understanding of the complex, non-linear and random interactions that take place. This is in part due to the involvement of people, whose behaviour is difficult to predict. An alternative to simulating complex systems with mathematical models is an approach based on real-world data, where system behaviour is learned through observations. Display Energy Certificates (DECs) are a source of observed building 'behaviour' in the UK and machine learning, a subset of artificial intelligence, can predict global behaviour in complex systems. In view of this, this thesis presents research that explores a method to predict and communicate the operational energy use of buildings in real-time as early design and briefing parameters are altered interactively. As a demonstrative case, the research focuses on school design in England. Artificial neural networks, a machine learning technique, were trained to predict thermal (gas) and electrical energy use of school designs based on a range of design and briefing parameters. In order to generate data for the artificial neural networks to learn from, a building characteristics dataset was developed which contains real-world data on 502 existing school buildings across England. A product of this research is a user-friendly design tool based on the psychological principles of 'flow', aimed at non-simulation experts, such as architects. The tool is named the 'SEED Tool' (School Early Environmental Design Tool)