Systematic Assessment of Machine Learning for Building Performance Analysis: Applications, Synthesis, and Way Forward

Abstract

Building science is a complex research domain that draws upon multiple disciplines, including physical and engineering sciences, architecture, economics, and the social sciences. As buildings become “smarter”, significant amounts of data are being collected about building systems and their users. Traditionally, statistical methods (e.g., correlation matrices, analysis of variance, and linear regression) have been used to analyze the collected data, quantify performance metrics, and guide building design and operation practices. However, limitations in the stated methods motivated the increasing adoption of machine learning (ML) algorithms, which have shown significant potential in dealing with complex and non-linear relationships among variables while requiring less intervention to develop the models compared to traditional statistical methods. While ML algorithms are gaining increased attention in the building science research community, it remains unclear how ML techniques perform in comparison to traditional statistical methods, and whether the observed shift is justified. Limited benchmarking efforts can be found in the literature comparing both approaches, showing mixed results and a lack of consistency in their research methods. This dissertation aims to provide a holistic assessment of ML algorithms in building performance studies to understand the current state of research and guide future efforts and applications of ML techniques. This work will (i) evaluate the premise of ML techniques in diverse building performance applications, (ii) compare and contrast the capabilities of ML techniques to traditional statistical methods, (iii) combine and synthesize the findings with those of previous studies, and (iv) provide recommendations/guidelines on when to use ML or traditional statistical methods for different building performance applications and contexts. Four case studies are conducted of which three case studies are on occupant comfort domain and one case study is on building energy. Additionally, a framework based on a Python programming language engine is developed and used for systematic analysis and benchmarking of the methods in conducted case studies. To further evaluate the generalizability of both, the proposed framework and the methods, two additional case studies are conducted in areas beyond the building science domain. The first one on in the road safety domain focusing on the prediction of accident severity, and the second case study is in the supply chain domain focusing on backorder predictions.

Date of Award	Apr 2023
Original language	American English
Supervisor	Raja Jayaraman (Supervisor)