Details

Autor(en) / Beteiligte

• Ji, Jiezhou
• Yu, Donghui
• Yang, Dewang
• Shen, Xiao
• Liu, Haoran
• Zhao, Linqiang
• Wang, Tao
• Zeng, Yu
• Pan, Qiuhui

Titel

Predictive Control based on Transformer as Surrogate Model for Cooling System Optimization in Data Center

Ist Teil von

• 2023 International Conference on Mobile Internet, Cloud Computing and Information Security (MICCIS), 2023, p.36-42

Ort / Verlag

IEEE

Erscheinungsjahr

2023

Quelle

IEEE Electronic Library Online

Beschreibungen/Notizen

• This paper develops a kind of Model Predictive Control (MPC) based on the Transformer proxy model. Traditional MPC usually makes predictions via linear models which are unable to understand the interaction between various input variables. In this paper, we utilize the Transformer to approximate the nonlinear model, and capture the law of time series data. Specifically, we exploit the idea of dynamic programming to optimize the model to predict the cost of the trajectory, and recommend the control operation at each time point to determine the optimal strategy, and finally achieve the purpose of optimization. Experiments conducted on the data room energy-saving control problem demonstrated that, our Transformer based MPC model can predict the temperature change more accurately and is more robust to the field environment change. Moreover, compared with the previous Proportion Integration (PI) control, the MPC developed by us can make the temperature in the machine room is stable around 22°C, and the power consumption is reduced by 37.40%.