| 摘要: |
| 在全球能源短缺和环境问题日益加剧的
背景下，提高公共建筑的能源效率和降低碳排放，
已成为城市规划和建筑设计领域的重要研究课题。
研究以政务区93栋公共建筑为研究对象，基于不
同公共建筑能耗数据核算其碳排放数值，运用多
种机器学习及其可解释性方法，评价多个空间形
态指标对建筑碳排放的非线性影响和交互效应。
结果表明：第一，XGBoost模型在预测公共建筑碳
排放量方面表现出最佳性能，测试数据集的决定
系数（R2）为0.896＞0.749（RF）＞0.680（CatBoost）＞
0.670（SVR）＞0.659（LightGBM）；第二，建筑面
积、建筑平均高度、容积率、开敞空间率和道路网
密度是影响公共建筑碳排放的关键因素，建筑面
积和平均建筑高度的平均绝对SHAP 值分别为
0.05 和0.01；第三，开敞空间率、道路网密度、
天空开阔度和平均周长面积比等变量对公共建筑
碳排放具有非线性的影响和明显的阈值效应，建
筑面积、容积率和平均建筑高度与公共建筑碳排
放更近似于线性关系；第四，道路网密度与容积
率的交互作用主要表现为负向，仅当道路网密度
低于0.015 km/km2时，才会与高容积率产生正向
交互作用。 |
| 关键词: 街区空间形态 公共建筑 碳排放 XGBoost模型 非线性影响 |
| DOI：10.13791/j.cnki.hsfwest.20250222002 |
| 分类号: |
| 基金项目: |
|
| Exploring the nonlinear effects of neighborhood spatial patterns on carbon emissions ofpublic buildings based on machine learning algorithms: A case study of Hefei’sAdministrative District |
|
GU Kangkang,QI Haodong,GAO Xianguang
|
| Abstract: |
| Against the backdrop of global energy shortages and escalating environmental challenges,
enhancing the energy efficiency of public buildings and reducing their carbon emissions have
become critical research topics in urban planning and architectural design. Public buildings are
major energy consumers, and their operational-phase carbon emissions are influenced by the spatial
configuration of surrounding neighborhoods. Taking Hefei’s Government District as a case study,
this research explores the nonlinear relationship and interactive effects between neighborhood
spatial form factors and public building carbon emissions. The study analyzes 93 public buildings
within the district, calculating each building’s annual average carbon emissions based on actual
electricity consumption data from 2023. To comprehensively quantify the impact of neighborhood
spatial form, the study preliminarily selected ten morphological indicators derived from the “spatial
form-microclimate-building energy consumption/carbon emissions” transmission mechanism.
These indicators encompass both intrinsic building attributes (e.g., building area, average area-toperimeter
ratio) and surrounding environmental factors (e. g., building density, green space ratio).
After multicollinearity testing using tolerance and variance inflation factors, building density and
total exterior wall area of surrounding buildings were excluded, leaving eight predictive indicators.
Five typical regression models—Support Vector Regression（SVR）, Random Forest（RF）,
Lightweight Gradient Boosting（LightGBM）, Categorical Boosting（CatBoost）, and Extreme
Gradient Boosting（XGBoost）—were compared to determine the optimal model for
predicting public building carbon emissions. The dataset underwent normalization and was
split into an 80% training set and 20% test set. Model performance was evaluated using root
mean square error (RMSE), mean absolute error (MAE), mean squared error (MSE), and
coefficient of determination (R2). To overcome the “black box” nature of complex models
and enhance interpretability, the study employed the Shapley additive explanation (SHAP)
framework, combining feature importance ranking with partial dependency plots for analysis.
Performance evaluation demonstrated that the XGBoost model achieved the highest
predictive accuracy on the test set, with an R2 of 0.896 > 0.749 (RF) > 0.680 (CatBoost) >
0.670 (SVR) > 0.659 (LightGBM). SHAP analysis revealed the relative importance of various
spatial form indicators. Building area was identified as the most influential feature, with an
average absolute SHAP value of 0.05, directly correlating with energy consumption activities and
system scale. The next most influential features were the average building height, floor area ratio,
and open space ratio. Road network density, average area-perimeter ratio, green space ratio, and
sky openness exerted relatively minor impacts, with average absolute SHAP values below 0.01.Crucially, SHAP dependency plots from locally weighted regression smoothing revealed significant nonlinearity and threshold effects for several variables. For
instance, the open space ratio positively influenced carbon emissions within the 0.64~0.70 range. The impact of floor area ratio shifts from negative to positive
around a threshold of approximately 5.08, with the positive effect increasing as the value rises. The inflection point for average building height occurs around
12.16 m. Road network density exhibits a negative impact when exceeding 0.02 km/km2, potentially due to improved ventilation reducing cooling loads. Green
space ratios exceeding 0.15 negatively impact carbon emissions, stabilizing after surpassing 0.3. The relationship between average area-to-perimeter ratio and
emissions follows an inverted U-shape: positive effects occur between 6.90 m and 16.67 m, turning negative beyond 16.67 m. Building area exhibits a clear
positive correlation with carbon emissions, particularly becoming more pronounced beyond 66 496.75 m2. In contrast, the relationships between building area,
floor area ratio, and average building height with carbon emissions are closer to linear. Additionally, SHAP interaction analysis revealed key synergistic effects
among variables. Average building height and floor area ratio exhibit a positive interaction, indicating that both at high values exacerbate carbon emissions—
likely due to poor ventilation, enhanced heat island effects, and increased vertical transportation energy consumption. Open space ratio also interacts positively
with sky openness, though in areas with low open space ratios, increasing sky openness yields minimal or slightly adverse emission reduction effects. A
negative interaction exists between road network density and floor area ratio; a positive interaction occurs only when low road network density combines with
high floor area ratio, increasing carbon emissions. Additionally, a positive interaction emerges when the average area-to-perimeter ratio is below 15 m and floor
area ratio exceeds six, particularly in conjunction with high building floor area. In summary, this study demonstrates the effectiveness of the XGBoost-SHAP
framework in predicting public building carbon emissions and deciphering the complex nonlinear effects of block spatial morphology. Key drivers include
building area, average building height, floor area ratio, and open space ratio. The findings provide valuable insights and a quantitative basis for formulating
targeted carbon reduction strategies for public buildings at the neighborhood scale. |
| Key words: neighborhood spatial pattern public building carbon emissions XGBoost model nonlinear effects |
