摘要
为解决数据挖掘方法在建筑能耗预测领域应用过程中面临的数据量不足问题,本文提出了基于深度迁移学习的住宅建筑能耗预测方法,即迁移学习一维卷积神经网络(TL-1D-CNN),该方法结合了深度学习模型的特征提取能力与迁移学习对预测目标数据需求小的特点。基于三栋住宅楼的能耗数据,以其中两栋住宅楼的能耗数据作为源域数据,第三栋住宅楼的部分能耗数据作为目标域数据,研究了该方法在目标楼栋数据不足情况下的预测性能。结果表明,对比长短期记忆神经网络(LSTM)模型在数据完备情况下的预测精度,TL-1D-CNN在供冷季的预测能耗平均绝对百分比误差(MAPE)高0.14%,均方误差(MSE)高0.44 W/m^(2),供暖季MAPE高0.58%,MSE高0.06 W/m^(2),模型性能基本接近。
The insufficiency of data restricts the application of data mining in the domain of residential building energy consumption prediction.To solve the problem,a method based on deep transfer learning for residential building energy consumption prediction is proposed,i.e.TL-1D-CNN(Transfer Learning-One Dimension-Convolutional Neural Network);the feature extraction capabilities of deep learning and low demand of target data of transfer learning are combined in the method.Relying on energy consumption data of three residential buildings,two of them are chosen as source domain and the third building energy consumption data is made as target domain,this paper studies the prediction performance of the method under insufficient data volume.The results show that,comparing the prediction accuracy of the long short term memory(LSTM)model with complete data,TL-1D-CNN has a 0.14%higher mean absolute percentage error(MAPE)in the cooling season,0.44 W/m^(2) in mean square error(MSE),0.58%higher MAPE in the heating season,and 0.06 W/m^(2) in MSE,model performance is basically close.
作者
陈恕
魏子清
岳宝
丁云霄
郑春元
翟晓强
CHEN Shu;WEI Ziqing;YUE Bao;Ding Yunxiao;ZHENG Chunyuan;ZHAI Xiaoqiang(Institute of Refrigeration and Cryogenics,Shanghai Jiao Tong University,Shanghai 200240,China;Guangdong Midea HVAC Equipment Co.,Ltd.,Foshan 528311,Guangdong,China)
出处
《制冷技术》
2022年第5期46-51,共6页
Chinese Journal of Refrigeration Technology
关键词
住宅建筑能耗预测
数据挖掘
深度迁移学习
微调技术
Residential building energy consumption prediction
Data Mining
Deep transfer learning
Fine-tuning
作者简介
翟晓强(1972-)男,教授,博士。研究方向:建筑能耗预测领域的数据挖掘。联系地址:上海市闵行区东川路800号上海交通大学,邮编200240。联系电话:021-34206296。E-mail:xqzhai@sjtu.edu.cn。
