ABSTRACT: The direct and indirect carbon emissions from the building sector are about 9 GT in 2020 based on the IEA data. The buildings are still off track for carbon neutrality by 2050. Hence, it is important to explore the carbon performance of buildings. The machine learning technique can provide fast and reliable energy analysis of buildings. This research applies the four types of machine learning models (full linear model, multivariate adaptive regression spline, support vector machine, and gradient boosting machine) to create the surrogate building energy models in the campus buildings of Tianjin, China. Then the sensitivity analysis based on these learning models is used to determine the key factors affecting building carbon emissions. The multivariate adaptive regression spline model performs the best in estimating carbon emissions of buildings with the R² (coefficient of determination) above 0.99. The equipment heat gains are responsible for over 40% variations of carbon emissions in most of the campus buildings. The following three important variables are heating set-point during unoccupied periods, lighting heat gains, and infiltration rate. These four variables should be carefully checked to reduce the carbon emissions of campus buildings studied in this research.
 

carbon emissions, machine learning, building energy, campus building