融合时域和频域特征的服务器出风温度预测-《计算机技术与发展》

文章信息/Info

Title:: Integrating Time-domain and Frequency-domain Features for Server Outlet Temperature Forecasting

Author(s):: ZHANG Ying; TANG Yuan-xin; WANG Yan-jie; School of Computer Science and Technology,Harbin University of Science and Technology,Harbin 150080,China

Keywords:: server outlet temperature forecasting; integrating time-domain and frequency-domain features; spatial coordinates; dynamic multi-head attention; Fast Fourier Transform

摘要:: 随着数据中心能耗问题日益突出,服务器出风温度作为机房稳定运行的关键指标,直接影响数据中心的能效和设备安全。为了提高服务器出风温度预测的准确性,该文提出了一种融合时域和频域特征的服务器出风温度预测方法,该方法分为时域和频域两个模块。在时域模块中,利用带有空间坐标的深度卷积和逐点卷积,捕捉时间序列中的空间特征。在频域模块中,通过快速傅里叶变换将时序数据转换到频域,并通过软阈值处理引入稀疏性约束,降低噪声。在时域和频域模块引入动态多头注意力机制,使模型自适应地调整注意力权重。通过主谐波能量加权分别计算时域和频域模块的特征权重,加权得到预测结果。实验结果表明,该模型在数据中心温度数据集上的 MSE 和 MAE 达到了 0. 145 和 0. 168,证明了其在提高服务器出风温度预测的准确性和有效性。

Abstract:: As the energy consumption problem of data centers is becoming more and more prominent,the server outlet temperature,as a key indicator for the stable operation of server rooms,directly affects the energy efficiency and equipment safety of data centers. In order to improve the accuracy of server outlet temperature forecasting,we propose a server outlet temperature forecasting method that integrates time-domain and frequency -domain features,which is divided into two modules:time - domain and frequency - domain. In the time domain module,spatial features in the time series are captured using deep convolution with spatial coordinates and point-by-point convo-lution. In the frequency domain module, the time series data are converted to the frequency domain by Fast Fourier Transform and sparsity constraints are introduced by soft thresholding to reduce noise. A dynamic multi-head attention is introduced in the time- and frequency-domain modules to enable the model to adaptively adjust the attention weights. The feature weights of the time and frequency domain modules are computed separately by main harmonic energy weighting,and the prediction results are obtained by weighting. The experimental results show that the proposed model achieves MSE and MAE of 0. 145 and 0. 168 on the data center temperature dataset,which proves its accuracy and effectiveness in improving the prediction of server outlet temperature.