Interpretable machine learning-based carbon emission prediction in the materialization stage of buildings
Existing carbon emission calculation methods suffer from drawbacks such as excessive data requirements,complicated calculations,and applicability limited to after-action control,thereby not promoting carbon emission reduction work in the early design stage by designers.To mitigate these problems,machine learning is introduced in this paper for carbon emission calculations in buildings to help designers obtain carbon emission information at an early stage of the building design and provide reduction references.This paper has the following procedures:(1)collection and establishment of a database of carbon emissions in the materialization stage of buildings,and checking of the database quality through Mahalanobis distance;(2)establishment of four machine learning models,namely Extremely Randomized Trees(ET),eXtreme Gradient boosting,(XGboost),Multilayer Perceptron(MLP)and Support Vector Regression(SVR),based on five building characteristics,and evaluation of the model performance using three evaluation indices-R2,Mean Absolute Error(EMAE),and Root Mean Square Error(ERMSE).(3)using SHapley Additive ExPlanations(SHAP)and Partial Dependence Plot(PDP)to prove the appropriateness of applying the ET models,thereby exploring the complex relationships between each feature and carbon emission.The outcome provides new information for carbon emission reduction in buildings.The findings indicate that:(1)Each machine learning model can efficiently predict carbon emissions in buildings'physical phases.Among these,the ET model established in this paper exhibits the best performance with an R2 of 0.88,an ERMSE of 0.317 kt CO2,and an EMAE of 0.217 kt CO2.(2)Interpretable analyses of the ET model by SHAP and PDP show that the effects of the features of the ET model on the prediction results are similar to those of existing studies,which verifies the reliability and reasonableness of the prediction results of the machine learning model.(3)The impact of each feature on the carbon emission of the materialization stage of buildings is ranked from large to small,including the total building area,building height,footing area,seismic grade,and basement depth.(4)The nature of each feature's relationship with carbon emission is nonlinear,meaning that there is a positive correlation relationship between the increase in building height and the decrease of the footing area,both reducing the carbon emission in the chemical stage of buildings.The results of this paper can guide for carbon reduction in existing buildings.
environmental engineeringmaterialization stage of buildingcarbon emissionmachine learninginterpretability analysis
万方数据
维普
