| 7 | 0 | 6 |
| 下载次数 | 被引频次 | 阅读次数 |
针对单一的卷积神经网络(CNN)模型和长短期记忆网络(LSTM)模型在冷负荷预测中存在的稳定性和精度不足问题，本文引入注意力机制模块，构建了卷积神经网络-长短期记忆-注意力机制(CNN-LSTM-Attention)耦合神经网络模型，以提升神经网络对数据中心冷负荷的预测性能。通过对比不同模型的预测结果可知：反向传播(BP)模型、LSTM模型、传统CNN-LSTM模型及CNN-LSTMAttention模型的平均绝对误差(MAE)依次为1.62、1.14、0.88和0.70 kW，均方根误差(RMSE)依次为2.30、1.47、1.08和0.85 kW，相关系数(R2)依次为0.86、0.94、0.96和0.98。CNN-LSTM-Attention模型的预测结果与真实数据吻合度更高，显著提升了预测精度与泛化能力。
Abstract：To address the limitations of stability and accuracy in cold load prediction by single convolutional neural network(CNN) and long short-term memory(LSTM) models, an attention mechanism module is introduced in this paper. A coupled neural network model combining CNN-LSTM-Attention is constructed to enhance predictive performance for data center cooling loads. Comparative analysis of different models shows that, for back propagation(BP) model, LSTM model, traditional CNN-LSTM model and the CNN-LSTMAttention model, the mean absolute errors(MAE) are 1.62, 1.14, 0.88 and 0.70 k W, respectively; root mean square errors(RMSE) are 2.30, 1.47, 1.08 and 0.85 kW, respectively, and correlation coefficients(R2) are 0.86, 0.94, 0.96 and 0.98, respectively. The CNN-LSTM-Attention model achieves closer alignment with real data, significantly improving prediction accuracy and generalization capability.
[1]陈梦,黄翔,金洋帆,等.数据中心用间接蒸发冷却空调机组性能测试及适用性分析[J].流体机械, 2022, 50(9):7-13.
[2]黄国瑞,刘宏伟,赵耀,等.基于TRNSYS的数据中心制冷系统能耗分析[J].制冷技术, 2024, 44(3):36-43.
[3]肖紫薇,刚文杰,袁嘉琦,等.基于长短期记忆网络的短期空调冷负荷预测[J].暖通空调, 2022, 52(4):132-137.
[4]冉彤,于军琪,冯涛.基于自适应深度信念网络的商场空调冷负荷预测[J].制冷技术, 2023, 43(2):52-56, 65.
[5] HIRATA Y, AIHARA K. Improving time series prediction of solar irradiance after sunrise:comparison among three methods for time series prediction[J]. Solar Energy, 2017,149:294-301.
[6] RíOS-MORENO G, TREJO-PEREA M, CASTA?EDAMIRANDA R, et al. Modelling temperature in intelligent buildings by means of auto regressive models[J].Automation in Construction, 2006, 16(5):713-722.
[7] PETKOVI?D, PROTI?M, SHAMSHIRBAND S, et al.Evaluation of the most influential parameters of heat load in district heating systems[J]. Energy and Buildings, 2015,104:264-274.
[8] SAJJADI S, SHAMSHIRBAND S, ALIZAMIR M, et al.Extreme learning machine for prediction of heat load in district heating systems[J]. Energy and Buildings, 2016,122:222-227.
[9] MA Z, YE C, LI H, et al. Applying support vector machines to predict building energy consumption in China[J]. Energy Procedia, 2018, 152:780-786.
[10] WANG Y Z, ZHANG N Q, CHEN X, et al. A short-term residential load forecasting model based on LSTM recurrent neural network considering weather features[J].Energies, 2021, 14(10):158-163.
[11] ZHOU C, FANG Z, XU X, et al. Using long short-term memory networks to predict energy consumption of airconditioning systems[J]. Sustainable Cities and Society,2020, 55(22):956-962.
[12] KONG W, DONG Z Y, Hill D J, et al. Short-term residential load forecasting based on resident behaviour learning[J]. IEEE Transactions on Power Systems, 2018,33(1):1087-1088.
[13]李桃迎,王婷,张羽琪.考虑多特征的高速公路交通流预测模型[J].交通运输系统工程与信息, 2021, 21(3):101-111.
[14]简献忠,顾洪志,王如志.一种基于双通道CNN和LSTM的短期光伏功率预测方法[J].电力科学与工程,2019, 35(5):7-11.
[15]于兴崭,芦天亮,杜彦辉,等.基于合成图像和Xception改进模型的安卓恶意家族分类方法[J].计算机科学,2023, 50(4):351-358.
[16]马飞,涂振宇,朱松挺,等.基于改进注意力机制的LSTM水位预测模型研究[J].江西水利科技, 2023,49(3):162-166.
[17]曾丽萍,雷奥军,张泉,等.基于Trnsys的某数据中心能耗分析[J].建筑热能通风空调, 2019, 38(11):65-68.
[18]刘坤.建筑空调负荷预测及水蓄冷空调系统控制技术研究[D].北京:北京建筑大学, 2018.
[19]张晨晨.混合优化的长短期记忆神经网络短期空调冷负荷预测及其应用研究[D].青岛:青岛大学, 2022.
[20]李昱瑾,陈焕新,刘江岩.基于PSO的SVR模型在多联机功耗预测上的应用[J].制冷学报, 2019, 40(6):53-61.
[21]宋海川,顾明伟,张弘韬,等.长短期记忆神经网络用于空调系统故障时间序列分析[J].制冷技术, 2020,40(6):24-30.
[22]胡伟,文武,魏敏.改进U-Net的高分辨率遥感图像轻量化分割[J].计算机系统应用, 2022, 31(12):135-146.
[23]王宜卿,叶明树,任义成,等.基于长短期记忆神经网络的工厂冷水机组短期功率预测[J].制冷技术, 2025,45(1):55-60.
[24]陶昊园,范波,李满峰.基于改进神经网络的冷水机组故障诊断研究[J].制冷技术, 2023, 43(2):43-51.
[25]雷柯松,吐松江·卡日,伊力哈木·亚尔买买提,等.基于WGAN-GP和CNN-LSTM-Attention的短期光伏功率预测[J].电力系统保护与控制, 2023, 51(9):108-118.
基本信息:
DOI：
中图分类号:TU83;TP308;TP183
引用信息:
[1]陈瑞,曹军.基于卷积神经网络-长短期记忆-注意力机制模型的数据中心冷负荷预测研究[J].制冷技术,2025,45(03):62-68+81.
基金信息:
上海市自然科学基金（No.20ZR1413200）; 中国石化科技开发（No.320131-3）