Application of hybrid machine learning algorithms for life cycle carbon prediction and optimization of buildings: A case study in China

Abstract

Buildings are a significant source of carbon emissions (CEs). In this work, the life cycle carbon emissions of buildings (LCCEBs) are dynamically calculated, spatiotemporal dynamic evolution laws are analyzed at the macro level, and the LCCEBs and driving factors are predicted and analyzed by integrating geographically and temporally weighted regression (GTWR) with machine learning algorithms. The results of a case study in China show the following. (1) The level of CEs in China has great spatiotemporal and geographical variation. The fitting accuracy of the GTWR prediction model can reach more than 0.75. (2) The accuracy of natural gradient boosting (NGBoost) is higher than the regression fitting accuracy of the GTWR model, especially with larger datasets. (3) The main driving factors obtained from the analysis of LCCEB driving factors using the NGBoost algorithm and SHapley additive explanation (SHAP) are CE per capita at the construction phase (ECP), construction area per capita (EAP), and carbon intensity of operation (OCI). The influence degrees and variation patterns of each factor are clarified, thereby proposing targeted measures for controlling carbon emissions in buildings. The theoretical knowledge of mining spatiotemporal patterns and driving factors of building CEs is enriched, and guidance for formulating policies and measures is provided.