On Predicting the PUE with Gated Recurrent Unit in Data Centers

The huge energy consumption of data centers has brought great pressure to operating enterprises, power plants, and the environment. How to effectively reduce energy consumption has aroused wide attention. Prediction of data center PUE (Power Usage Effectiveness) or energy consumption is a promising way to reduce data center energy consumption. It will provide many ideas for data center to reduce energy consumption if the PUE prediction can be accurately predicted. However, in the previous researches on predicting PUE or energy consumption, there are some shortcomings such as the factors to PUE are not fully considered and lack of consideration of the time series information of energy related data. In this paper, we investigate how to predict the PUE value of data center by fully exploiting the knowledge energy consumption related historical data over time. To this end, we first collect more than 50,000 data samples with 144 energy related variables. Then, these data samples are preprocessed for normalization and feature selection. After that, considering the temporal property of the energy consumption related data, a GRU based neural network model is designed as the algorithm to train the data for generating the model for PUE prediction. Finally, extensive experiments are conducted based on the real data trace to evaluate the performance of the GRU model. The results demonstrate that our proposed model is efficient in accurately predicting the PUE value, and outperforms the baseline schemes with respect to MAE, MSE, and R-Squared.