Assessment of operational carbon emissions for residential buildings comparing different machine learning approaches: A study of 34 cities in China

Abstract

As a major contributor to the energy consumption and carbon emissions of the whole society, carbon reduction from residential building operation holds significant importance for achieving the carbon peak and neutrality goals. Although research on carbon emissions of the building sector on a global and regional scale has been conducted, it is challenging to identify overall patterns representing the operational carbon emissions of residential buildings in cities. This study utilized operational energy use to analyze and characterize the relevant carbon emissions of residential buildings in 34 representative cities across China over a decade from 2012 to 2021. The average emissions were assessed as 1.0 tCO2e per capita which were relatively higher in cold regions. Based on a machine learning algorithm, regression analysis was conducted to predict the operational carbon emissions considering the features including population, economy, geography, urbanization renewable energy, and building electrification. Five models based on polynomial regression, support vector regression, random forest, gradient boosting decision tree, and extreme gradient boosting were trained. It was found that the extreme gradient boosting model exhibited the best performance for predicting operational carbon emissions per capita with R2 value of 0.948 for the testing dataset. Further analysis identified building climate zone, population, urbanization rate, and building electrification rate as the most influential features on the emissions. The predictive approach for estimating operational carbon emissions in cities can provide valuable insight into carbon reduction of the building sector.