摘要
为精准预测建筑施工全生命周期碳排放,研究基于生命周期评价(LCA)框架,明确了系统边界与区域范围,将碳排放划分为材料生产、运输、施工建造、运营维护4个阶段,系统剖析了各阶段碳源。通过构建三级机制确定人工、能源等碳排放因子,对比4种测算方法后,选用碳排放因子法建立计算公式。研究选取Lasso回归、支持向量机(SVM)、XGBoost和人工神经网络(ANN)等4种模型,将161组样本划分为训练集、测试集和验证集,并以决定系数R2、均方差(MSE)、平均绝对误差(MAE)及训练成本作为评估指标。结果表明,Lasso回归因线性特性而预测效果最差;SVM表现良好,具有一定泛化能力;XGBoost与ANN拟合优异,测试集R2分别达0.992和0.989,但ANN的训练成本显著更高。
To obtain accurate predictions of carbon emissions throughout the entire construction lifecycle,this study utilized a life cycle assessment(LCA)framework that delineated system boundaries and regional parameters.Carbon emissions were systematically categorized into four phases:material production,transportation,construction,and operation/maintenance,with a detailed analysis of carbon sources at each stage.A three-tier mechanism was established to determine emission factors for labor and energy consumption.After comparing four calculation methods,the carbon emission factor method was employed to formulate the computational formula.Four machine learning models—Lasso regression,support vector machines(SVM),XGBoost,and artificial neural networks(ANN)—were adopted.The 161 sample datasets were divided into training,test,and validation sets,with evaluation metrics including root mean square error(R2),mean squared error(MSE),mean absolute error(MAE),and training costs.The results indicate that Lasso regression exhibits the least prediction accuracy owing to its linear nature;SVM demonstrates commendable performance with moderate generalization capacity.Conversely,XGBoost and ANN display superior fitting outcomes,attaining test set R²values of 0.992 and 0.989,respectively.However,ANN requires significantly higher training costs.
作者
孙超
余少乐
张玉建
SUN Chao;YU Shaole;ZHANG Yujian(China Communications Second Highway Bureau,Dongmeng Engineering Co.,Ltd.,Xi’an,710100,Shaanxi,China;China State Construction Engineering Corporation Limited,Shanghai 200135,China)
出处
《上海建材》
2025年第6期29-32,共4页
Shanghai Building Materials
