Data-driven Approaches to Improving EUI Prediction Accuracy for Benchmarking

 Approximately 40% of global energy consumption and 42% of global carbon emissions are  attributed to the building sector. Data-driven benchmarking, known for its precision and efficiency,  is crucial in supervising energy use and incentivizing efficiency over time. A vital component of  this approach, energy use intensity (EUI) prediction, currently relies heavily on multi-linear  regression. However, this method has been widely questioned for its reliability in predicting EUI.  Consequently, numerous studies have introduced machine learning algorithms to improve the  accuracy of EUI predictions with data-driven benchmarking. Despite these efforts, there remains  a lack of comparative performance analysis across various algorithms. This thesis  comprehensively compares 9 machine learning algorithms—including multi-linear regression,  LASSO regression, Ridge regression, random forest, gradient boosting, XGBoost, support vector  machines, and LightGBM — on a national energy dataset comprising 66,050 buildings across 18  building types and seven climate zones. It examines the accuracy of different ML algorithms in  EUI prediction and explores the relationships between algorithm accuracy and factors such as  building type and climate zone. The thesis then demonstrates the application steps in the use of the  most accurate algorithms for benchmarking with greater accuracy than linear regression. This  research serves as a reference to enhance the reliability of future data-driven benchmarking by  selectively employing the most appropriate algorithms based on building type and climate.