在绿色建筑设计中, 传统节能方案的选择依赖设计师经验, 难以考量技术、成本等多维因素, 决策效率受限。该文提出基于经验知识的建筑节能方案智能决策模型。通过构建含有147个绿色星级认证建筑的案例库, 利用案例推理方法, 根据初步设计已知属性计算余弦相似度进行相似案例检索, 为目标案例提供技术方案参考; 通过比较分析4种机器学习算法, 以确定的技术方案为输入参数, 建立基于机器学习算法的增量成本预测模型; 以夏热冬冷地区建筑为例对所构建的模型进行有效性验证。结果表明：基于案例推理的建筑节能方案智能决策模型可找到最相似案例并重用其技术方案, 提高了决策效率; 基于极端梯度提升算法的增量成本预测模型的预测准确率最高, 为72.41%。该研究所构建的智能决策模型可提高建筑节能方案的科学性与决策效率, 为绿色建筑设计提供重要支持。

Objective: To achieve energy-saving and emission reduction in buildings, green building design is increasingly gaining attention. However, traditional design methods often rely heavily on the designer's experience, which complicates the consideration of multidimensional factors such as technical strategies and costs, thus limiting decision-making efficiency. Mining tacit knowledge to support green building design decisions and improve decision-making efficiency presents a significant challenge. Methods: This study proposes a two-stage intelligent decision-making model for energy-saving building strategies based on tacit knowledge. The first stage employs a case-based reasoning (CBR) model to determine energy-saving technical strategies. A case library containing 147 green-certified buildings provides reference strategies using attributes from the preliminary design phase, such as building type, structure, number of floors, height, orientation, shape coefficient, floor area, and green certification level. Cosine similarity helps retrieve relevant cases and identify technical strategies like window-to-wall ratios, heat transfer coefficients of the building envelope, heat pump loads, and renewable energy use. The second stage involves an incremental cost prediction model that uses machine learning algorithms. A 2∶8 split of the case library into test and training sets enables comparison across four machine learning algorithms: artificial neural network, extreme gradient boosting (XGBoost), support vector machine, and random forest. Each model's prediction accuracy, precision, and F₁ score (the harmonic mean of precision and recall) are evaluated. The model takes the technical strategies identified in the first stage and the known information from the preliminary design phase as input feature parameters. The import_plot module analyzes feature importance to eliminate redundant features. The two-stage model is validated on buildings from regions with hot summers and cold winters. Results: Findings indicate the following: (1) The CBR model effectively identifies and reuses the most similar energy-saving technical strategies, thereby improving decision-making efficiency. Most target cases achieve a similarity greater than 0.8 in the case library. (2) Among the machine learning models, the XGBoost-based incremental cost prediction model exhibits the highest accuracy, achieving 72.41%. (3) By applying the synthetic minority oversampling technique to balance samples and remove outliers, the prediction accuracies for four types of costs reach approximately 70%. However, the prediction accuracy for the fifth type of incremental cost is lower owing to varying owner preferences and requirements. Conclusions: The proposed two-stage intelligent decision-making model successfully integrates the CBR model with machine learning algorithms. The proposed model optimizes the use of limited known information available during the preliminary design stage to predict both technical strategies and incremental costs. This model enhances the scientific rigor and efficiency of energy-saving decision-making, providing significant support for green building design.