U.S. residential buildings account for a significant share of national energy consumption, highlighting their potential for energy-savings. Accurately predicting building energy consumption and understanding the impact of household features are, therefore, crucial for effective energy management, conservation efforts, and the development of energy policies. However, most existing models predicting U.S. residential energy consumption tend to focus on particular regions, limiting their generalizability across the entire country. In addition, many studies have overlooked the significant variations in the drivers of energy consumption between different types of residential buildings, resulting in a lack of separate prediction models for different residential building types. Moreover, when analyzing the impact of household features on building energy consumption, most studies provide a holistic measure of feature importance without sufficient interpretability. To address these gaps, this study uses the Residential Energy Consumption Survey dataset and three tree-based machine learning algorithms to develop separate energy use intensity (EUI) prediction models for two typical U.S. residential building types. The results demonstrate that the LightGBM-based prediction model performs best for apartments, while the CatBoost-based prediction model performs best for single-family houses. Furthermore, the study applied SHapley Additive exPlanations to analyze the impact of household features on energy consumption. The results reveal that total square footage, space heating with natural gas, climate conditions, and building age are the common key features influencing EUI for both building types. Based on these findings, this study provides general and targeted energy-saving recommendations for both building types, serving as a valuable guide for new building design and retrofitting of existing buildings.

中文翻译：

---

使用机器学习和 SHAP 对不同住宅建筑类型进行能耗预测和家庭特征分析：迈向节能建筑

美国住宅建筑占全国能源消耗的很大一部分，凸显了其节能潜力。因此，准确预测建筑能耗并了解家庭特征的影响对于有效的能源管理、节能工作和能源政策的制定至关重要。然而，大多数预测美国住宅能源消耗的现有模型往往关注特定地区，限制了其在整个国家的普遍适用性。此外，许多研究忽视了不同类型住宅建筑之间能源消耗驱动因素的显着差异，导致缺乏针对不同住宅建筑类型的单独预测模型。此外，在分析家庭特征对建筑能耗的影响时，大多数研究提供了特征重要性的整体衡量标准，但缺乏足够的可解释性。为了解决这些差距，本研究使用住宅能源消耗调查数据集和三种基于树的机器学习算法，为两种典型的美国住宅建筑类型开发单独的能源使用强度 (EUI) 预测模型。结果表明，基于 LightGBM 的预测模型对于公寓表现最佳，而基于 CatBoost 的预测模型对于单户住宅表现最佳。此外，该研究还应用 SHapley 加法解释来分析家庭特征对能源消耗的影响。结果表明，总面积、天然气供暖、气候条件和建筑年龄是影响两种建筑类型 EUI 的共同关键特征。基于这些发现，本研究为这两种建筑类型提供了一般性和有针对性的节能建议，为新建建筑设计和现有建筑改造提供了有价值的指导。