基于奇异谱分析的CNN-BiLSTM短期空调负荷预测模型  

CNN-BiLSTM Short-term Air Conditioning Load Prediction Model Based on Singular Spectrum Analysis

作　　者：杨心宇 任中俊 周国峰 易检长 何影 YANG Xinyu;REN Zhongjun;ZHOU Guofeng;YI Jianchang;HE Ying(School of Environmental and Municipal Engineering,North China University of Water Resources and Electric Power,Zhengzhou 450046,China;Shenzhen Secom Technology Co.Ltd.,Shenzhen 518071,Guangdong,China;Guangdong Building Equipment Intelligent Control and Operation and Maintenance Engineering Technology Research Center,Shenzhen 518071,Guangdong,China;School of Mechanical Engineering,Tongji University,Shanghai 200082,China)

机构地区：[1]华北水利水电大学环境与市政工程学院,郑州450046 [2]深圳市紫衡技术有限公司,广东深圳518071 [3]广东省建筑设备智慧控制与运维工程技术研究中心,广东深圳518071 [4]同济大学机械与能源工程学院,上海200082

出　　处：《建筑节能（中英文）》2024年第3期64-73,共10页Building Energy Efficiency

基　　金：厦门市建设科技计划资助项目“厦门市公共建筑绿色低碳运营管理与技术研究”(XJK2022-1-6)。

摘　　要：空调负荷的精准预测对建筑空调系统优化控制具有重要意义。为提高空调负荷预测精度,提出了一种基于奇异谱分析(SSA,Singular Spectrum Analysis)的卷积神经网络(CNN,Convolutional Neural Network)和双向长短时记忆网络(BiLSTM,Bidirectional Long Short Term Memory)短期空调负荷预测模型。使用皮尔森相关系数选取与空调负荷高相关性特征。针对空调负荷的波动性和随机性,采用SSA将空调负荷分解为多个分量,同时将各个分量带入CNN-BiLSTM模型进行预测,该模型利用了CNN的特征提取和BiLSTM的双向学习能力,并将各个分量预测结果进行重构。通过不同建筑类型的空调数据对该模型进行验证分析,发现所提出模型在预测办公建筑空调负荷中RMSE、MAPE和MAE为19.47RT、14.72RT和2.33%,在预测商业建筑空调负荷中RMSE、MAPE和MAE为82.5RT、34.21RT和0.87%。结果表明,所提出的模型具有普适性且精度较高,可进行推广应用。Accurate prediction of air conditioning load is of great significance for optimizing control of building air conditioning systems.In order to improve the accuracy of air conditioning load forecasting,a short-term air conditioning load forecasting model based on Singular Spectrum Analysis(SSA,Singular Spectrum Analysis) Convolutional Neural Network(CNN,Convolutional Neural Network) and Bidirectional Long and Short Memory Network(BiLSTM,Bidirectional Long Short Term Memory) is proposed.Pearson correlation coefficient is applied to select features with high correlation with air conditioning load.In response to the volatility and randomness of air conditioning load,SSA is used to decompose the air conditioning load into multiple components,and each component is brought into the CNN-BiLSTM model for prediction.This model utilizes the feature extraction of CNN and the bidirectional learning ability of BiLSTM,and reconstructs the prediction results of each component.The model is validated and analyzed using air conditioning data from different building types,and it is found that the RMSE,MAPE,and MAE of the proposed model are 19.47RT,14.72RT,and 2.33% in predicting office building air conditioning loads,and 82.5RT,34.21RT,and 0.87% in predicting commercial building air conditioning loads.The results indicate that the proposed model has universality and high accuracy,and can be widely applied.

关 键 词：空调负荷预测 双向长短时记忆网络 奇异谱分析 卷积神经网络 

分 类 号：TU831[建筑科学—供热、供燃气、通风及空调工程]