摘要
建筑自动化技术的广泛应用产生了大量的建筑运行数据。这类数据存在复杂的非线性关系、噪音多、冗余度高,因此建模分析难度较大。采用近100组不同类型建筑的实测数据为研究对象,对其短期能耗进行预测分析,进而形成具有普适性的预测方法。针对整体预测过程,设计了特征工程和预测模型建立两方面内容。在特征工程方面,研究了基于主成分分析和卷积自编码器的线性和非线性特征工程方法。在预测模型建立方面,比较了传统的线性回归、极度梯度提升决策树和神经网络算法。通过分析近100组不同类型建筑的实测数据,量化了相关方法在短期建筑能耗预测中有效性和可靠性。实验结果表明,基于一维卷积自编码器的特征工程方法可以有效提升模型的泛化性能,同时也可加快模型的收敛速度。
With the wide adoption of building automation technologies,massive amounts of building operation data are being collected.It is very challenging to perform accurate analysis for such data,due to the existence of complicated nonlinear relationships,noisy observations and high redundancy.This study aims to tackle the above-mentioned challenges and develop generic solutions by analyzing approximately100 buildings with different functionalities and operating characteristics.More specifically,the prediction methodology proposed consists of two main steps,i.e.,feature engineering and predictive modeling.The principal component analysis-and convolutional autoencoder-based feature engineering methods are proposed for performance comparison.Three predictive modeling algorithms,i.e.,generalized linear regression,extreme gradient boosting trees,and artificial neural networks,are utilized for predictive modeling.The generalization performance has been tested and validated through the actual operational data from approximately 100 buildings.It is shown that the convolutional autoencoder-based method can produce high-quality features for short-term building energy predictions while enhancing the convergence speed of different supervised learning algorithms.
作者
王振亚
范成(指导)
李达生
曾妍洁
刘明辉
WANG Zhen-ya;FAN Cheng;Lee Da-sheng;Tseng Yan-chieh;LIU Ming-hui(College of Civil Engineering,Shenzhen University,Shenzhen 518000,Guangdong,China;Taipei University of Technology,Taipei 100,Taiwan,China;Minghsin University of Science and Technology,Hsinchu 300,Taiwan,China)
出处
《建筑节能》
CAS
2020年第7期100-107,145,共9页
BUILDING ENERGY EFFICIENCY
基金
广东省哲学社会科学规划项目(GD18YGL07)
深圳市哲学社会科学规划项目(SZ2019D014)
深圳大学-台北科技大学学术合作项目(2019003)
关键词
建筑能耗预测
特征工程
卷积神经网络
非线性建模
极度梯度提升决策树
building energy predictions
feature engineering
convolutional neural networks
nonlinear modeling
extreme gradient boosting trees
