The December 2022 Commercial Buildings Energy Consumption Survey conducted by the Energy Information Administration (EIA) found that space heating constitutes 32% of total building end-use energy, with cooling accounting for 9%. It has a significant impact on climate change. In this research, intelligent models were developed to predict the annual heating and cooling loads (HL and CL) of residential buildings. Eight inputs, including relative compactness, roof area, overall height, surface area, glazing area, wall area, glazing area distribution, and orientation, were used for the modeling development. The artificial intelligence (AI) models were Support Vector Regression (SVR), K-Nearest Neighbors (KNN), Random Forest (RF), Multi-layer Perceptron (MLP), Gradient Boosting (GBoost), and Extreme Gradient Boosting (XGBoost). Three scenarios of input combinations were tested: Scenario-1 (S1) with eight inputs and Scenario-2 (S2) with five inputs. Similarly, Scenario-3 (S3) with five inputs. Results indicated that, the RF was the superior algorithm in HL for S1, achieving Kling-Gupta Efficiency (KGE = 0.998) and Root Mean Square Error (RMSE = 0.501 kW h/m). XGBoost performed outstandingly in CL with KGE = 0.994 and RMSE = 0.922 kW h/m. In S2, KNN showed excellent HL with KGE = 0.945 and RMSE = 3.094 kW h/m, and RF outperformed in CL with KGE = 0.941 and RMSE = 2.727 kW h/m. In S3, XGBoost exhibited the highest efficiency for HL with KGE = 0.997 and RMSE = 0.492 kW h/m, while RF performed best for CL with KGE = 0.976 and RMSE = 1.686 kW h/m. In conclusion, S2 proved to be a logical choice, matching the efficiency of S1 with reduced error. Overall, HL predictions generally displayed superior performance compared to CL predictions.

中文翻译：

---

优化建筑能源性能预测：人工智能模型的比较研究

美国能源信息管理局 (EIA) 进行的 2022 年 12 月商业建筑能源消耗调查发现，供暖占建筑最终使用能源总量的 32%，制冷占 9%。它对气候变化有重大影响。在这项研究中，开发了智能模型来预测住宅建筑的年度供暖和制冷负荷（HL 和 CL）。模型开发使用了八个输入，包括相对紧凑度、屋顶面积、总高度、表面积、玻璃面积、墙壁面积、玻璃面积分布和方向。人工智能 (AI) 模型包括支持向量回归 (SVR)、K 最近邻 (KNN)、随机森林 (RF)、多层感知器 (MLP)、梯度提升 (GBoost) 和极限梯度提升 (XGBoost) 。测试了三种输入组合场景：具有 8 个输入的场景 1 (S1) 和具有 5 个输入的场景 2 (S2)。同样，场景 3 (S3) 有五个输入。结果表明，RF 是 HL 中 S1 的最佳算法，实现了 Kling-Gupta 效率（KGE = 0.998）和均方根误差（RMSE = 0.501 kW·h/m）。 XGBoost 在 CL 中表现出色，KGE = 0.994，RMSE = 0.922 kW·h/m。在 S2 中，KNN 在 HL 中表现出色，KGE = 0.945，RMSE = 3.094 kW h/m，RF 在 CL 中表现出色，KGE = 0.941，RMSE = 2.727 kW h/m。在 S3 中，XGBoost 在 HL 方面表现出最高效率，KGE = 0.997，RMSE = 0.492 kW h/m，而 RF 在 CL 方面表现最佳，KGE = 0.976，RMSE = 1.686 kW h/m。总之，S2 被证明是一个合乎逻辑的选择，它与 S1 的效率相匹配，同时减少了错误。总体而言，与 CL 预测相比，HL 预测通常表现出更优越的性能。