建筑能耗预测的机器学习回归模型研究

Study on Machine Learning Regression Model for Prediction of Building Energy Consumption

目的使用回归模型对建筑样本进行机器学习训练,筛选出适用于预测建筑能耗的模型。方法以香港地区1923个商用建筑作为研究对象,通过网络信息收集的方式获取建筑的物理参数、使用参数以及环境参数,采用PCA算法对高维度数据进行降维处理,并对相关数据进行了归一化处理;选取13种常见的机器学习回归模型作为建筑能耗预测模型,采用平均绝对误差MAE、绝对中位差MAD和决定系数R^(2)作为模型性能评价指标,采用StratifiedKFold分层采样法对总样本进行划分,并对划分后样本进行机器学习训练。结果Bagging、XGBoost、Random Forest、Extra Trees集成学习回归模型对建筑能耗预测的准确性远优于其他9种模型,其中XGBoost有最小的MAE(6.47)和MAD(2.95),Random Forest有最大的R^(2)(0.97)。结论5种集成学习回归模型中除了分类算法外,Bagging、XGBoost、Random Forest、Extra Trees 4种模型对建筑能耗预测较优。XGBoost对数据较为完整的建筑能耗预测准确度最高,Extra Trees对于数据缺失严重的建筑预测准确度优于XGBoost。

The purpose of this paper is to choose the appropriate machine learning regressions as building energy consumption prediction model.1923 commercial buildings in Hong Kong taken as samples,buildings′parameters on their physical features,utilization features and environmental features are collected by Internet.And high dimension data in parameters are adjusted by PCA algorithm.13 kinds of regressions were chosen as prediction models of building energy consumption,with index MAE,MAD and R^(2) to evaluate their performance.The whole samples were divided by StratifiedKFold algorithm,and then trained by machine learning regressions.Prediction results of the building energy consumption by Bagging,XGBoost,Random Forest and Extra Trees model,which are based on ensemble learning,were more accurate than that of other models,XGBoost got minimum value on MAE(6.47)and MAD(2.95)and Random Forest got maximum value on R^(2)(0.90801).In addition,the main parameters that affect on building energy consumption are building area,stand alone building,wind parameter in 500 meter height,rental status,point solar radiation,height and number of floors.In addition to the classification algorithm,among the 5 ensemble learning regression models involved in this article,Bagging,XGBoost,Random Forest and Extra Trees perform good at building energy consumption predict.Among them,XGBoost is the best regression to predict the energy consumption of building with complete data,while Extra Trees perform better at samples with lack of data.