Online Machine Learning With Its Applications To Data Center Energy Optimization

In recent years,with technological innovation,the introduction of digital transformation policies,and the outbreak of the epidemic,there have been more online demands,and the scale of demand has become larger,and the performance requirements of cloud servers have become more stronger.At the same time,under the background of national policies and the cloud era,as a new type of infrastructure that carries data,data centers are constantly being optimized and developed,and green energy conservation has gradually become the primary consideration in the process of data center optimization.Under the above background,in order to improve the shortcomings of high energy consumption,low expansion and high degree of customization of traditional data centers,modular data centers emerge as the times require.Compared with traditional data centers,modular data centers have the advantages of green energy saving,intelligent monitoring and management,and easy expansion.Among of them,the performance index of green energy saving is the power usage effectiveness,and the cooling system of the data center is the most primary influential factor in the power usage effectiveness.With the background of the air-condition cooling control system of the modular data center,this paper aims to improve the power usage effectiveness of the data center,and researches its technical realization.In this paper,an air-condition control strategy based on an online machine learning algorithm is proposed.First,we divide the whole system into four modules,including data acquisition module,model training module,air-condition control module and safety control module.Then,by constructing the dual network structure,the multiplied action space is converted into an additive action space,which greatly reduces the size of the action space and accelerates the convergence speed of the model.In addition,XGBoost,SVM,Linear Regression,Decision Tree are used in the cooling control system based on supervised learning for comparative experiments.algorithms such as Policy Gradient,Actor Critic,and Proximal Policy Optimization are used in the cooling control system based on reinforcement learning for comparative experiments.By comparing the experimental results,we select the supervised learning with the most stable effect as the final implementation plan.In order to test the experimental effect,a data center simulator was also designed based on the previous data,and the simulator also is used four different models such as XGBoost,Decision Tree,Linear Regression,and SVM for comparative experiments.By comparing the running results of the algorithm in two different simulator environments and real test environments,the online machine learning algorithm based on supervised learning can reduce the power usage effectiveness of the modular data center to 1.255 under the premise of ensuring equipment safety.Compared with the national standard,the power usage effectiveness is reduced by 10.3%.