Application of Convolutional Neural Network to Prediction of Temperature Distribution in Data Centers
Shinya Tashiro, Yutaka Nakamura, Kazuhiro Matsuda, Morito Matsuoka · 2016
We propose a model for predicting the temperature distribution in data centers by using a convolutional neural network (CNN). Changes in the temperature distribution depend on the local structure of the data center, such as equipment locations and server types. Although the various physical relations in a data center were modeled as a network in our previous work, there were no mechanisms for automatically extracting the structure of the data center. The use of a CNN is a technique for learning local structure adaptively, which allows learning complicated features, such as the various physical relations in a data center. We evaluate the performance of the proposed model by using actual data from an experimental data center. The evaluation indicates that the proposed model can predict 20-minute future temperature distributions over 48 locations in 0.42 ms, with a root mean square error (RMSE) of 0.96 degrees. This accuracy is a dramatic improvement over simple linear prediction models, and the accuracy is sufficient to allow for control of air conditioners on the basis of these temperature predictions.