Data center PUE optimization based on machine learning

We aim at reducing the power usage effectiveness (PUE) metric of data centers with machine learning methods. At the present stage, industry such as Google considers only a small number of features and the impact of a single variable on the PUE metric, lacking the analysis of the coupling characteristics between features. In addition, machine learning methods have high requirements on the quality and quantity of data, and it is easy to be interfered by signal noises when implementing machine learning methods in practice. There are few specific optimization cases in both academia and industry at this stage. In this paper, we improve the current method of optimizing the PUE metric using neural networks by increasing the feature dimension, so as to increase the prediction accuracy, which is higher than that of the PUE predicting model built by Google. We use statistical methods to approximate the coupling characteristics between features with historical samples, and integrate them into sensitivity analysis to obtain more accurate results. We propose a cooling system parameter setting method based on sensitivity analysis. Based on the data and infrastructure in a Tencent Data Center located in Tianjin, we implement an experiment of parameter optimization of the cooling system, and the effectiveness of the proposed method is demonstrated by the experimental results.