Time series prediction for thermal management in server applied with deep learning

The portion of cooling power reaches up to one third of total power budget in the operation of data centers. Recently researchers pay more attention on thermal management for improving energy efficiency in the data centers as well as the servers. To the best of existing literatures, the complex interactions of cooling and heating resources is the main problem encountered in the design of server fan control. Moreover, to achieve better control performance, the thermal information of time series should be taken into account as well. Therefore, this work proposes time series prediction of server temperatures by utilizing two multivariate time series prediction models, named MTSMFF and MsA. An automatic data acquisition system has been designed for collecting massive scale dataset. To ensure the robustness of trained results, the dataset is divided into training data and validation data. These two models are trained with the training data, and MsA provides the best performance in accuracy with the validation data. The experimental results show that the time series prediction of MsA performs splendidly for the trend of the server temperatures.