摘要
预测模型的输入特征变量对建筑耗热量预测性能具有较大的影响,为了进一步改进输入特征变量的选取,本文提出了稀疏自编码(SAE)方法对历史耗热量数据进行特征提取,并通过对比常规的线性化特征提取方法(主成分分析,PCA),分析了SAE特征提取方法分别对MLR、ANN和SVM预测模型精度的提升。应用某居住建筑实测数据对该方法进行了实验验证,结果表明:在测试数据集中,使用SAE方法提取到的特征值作为模型输入变量,MLR、ANN和SVM3个模型的预测性能均得到提升,相比于利用PCA特征提取的方法,CV值分别降低了3.8%、4.1%和4.2%。此外,SAE方法对模型性能的提升还表现在模型泛化性能地增强,在测试样本中的表现优于在训练样本中的表现。
The input feature variables of prediction model have a great influence on the prediction performance of building heating-energy consumption.In this context,a Sparse Auto-Encoder(SAE)method is proposed to extract features from historical data about heating-energy consumption and improve the selection of input feature variables.The effect of SAE feature extraction on improving the prediction accuracy of MLR,ANN and SVM prediction models is analyzed by comparing with the conventional linearized feature extraction method(Principal Component Analysis,PCA).The proposed method is validated by the measured data of a residential building.The results show that:the prediction performance of MLR,ANN and SVM models was improved by using the feature values extracted by SAE method as input variables in the test data set.Compared with the PCA-based feature extraction method,the CV value in the three models was reduced by 3.8%,4.1%and 4.2%,respectively.In addition,the SAE method also improved the generalization performance of model,and had better performance in the test sample than in the training sample.
作者
袁大昌
史艳霞
高俊楠
YUAN Dachang;SHI Yanxia;GAO Junnan(Tianjin University Research Institute of Urban Planning Design,Tianjin 300110;Tianjin Sino-German University of Applied Sciences,Tianjin 300350;Tianjin University,Tianjin 300350)
出处
《建筑科学》
CSCD
北大核心
2020年第8期1-6,49,共7页
Building Science
基金
国家重点研发课题“基于县域控碳体系的数据驱动型规划设计技术集成与示范应用”(2018YFC0704706)。
关键词
特征提取
稀疏自编码
耗热量预测
支持向量机
机器学习
feature extraction
sparse auto-encoder
heating-energy consumption prediction
support vector machine
machine learning
作者简介
袁大昌(1965-),男,博士,教授,联系方式:tjzdjsyy@163.com。
