Predictive models of embodied carbon emissions in building design phases: Machine learning approaches based on residential buildings in China
Xiaocun Zhang,Hailiang Chen,Jiayue Sun,Xueqi Zhang
DOI: https://doi.org/10.1016/j.buildenv.2024.111595
IF: 7.093
2024-05-02
Building and Environment
Abstract:As one of the major architectural forms, residential building construction consumes considerable resources, and it has been regarded as a crucial factor for the implementation of carbon peaking and neutrality targets. Estimation of carbon emissions is essential for making informed decisions regarding building design and carbon reduction measures. Some studies have established predictive models for building operational carbon emissions. However, there are relatively few studies concerning the prediction of embodied carbon emissions. This study established a dataset containing 850 residential building samples. Based on the trade-offs and combinations of different explanatory variables, including building features and material use, this study developed two-phase models. These models, incorporating six scenarios, predicted embodied carbon intensities using twelve machine learning algorithms. For the planning design phase, an extremely randomized tree model using the building height, structural form, seismic fortification intensity, delivery form, geographical region, and material cost as inputs was favored considering R 2 as 0.821 on the testing dataset. For the preliminary design phase, an extreme gradient boosting model using both building features and material intensities provided superior performance with R 2 and MAPE values of 0.917 and 0.038, respectively. The optimal models and relevant feature importance analysis can be helpful for efficiently predicting embodied carbon emissions, and therefore, facilitate low-carbon designs.
engineering, civil,construction & building technology, environmental