Dynamic Energy Saving Operation for Data Center Through Sequential Optimization of Server Workload Allocation and Air Conditioning

This paper presents an energy-efficient operational strategy for data centers by optimizing server workload allocation and air conditioning settings based on power consumption models for servers and air conditioners. These models were developed using control parameters accessible to operators of both systems and included static pressure difference as an explanatory variable. This approach allowed the models to be used generically (not sitespecific), ensuring their applicability across various installation environments. The models achieved low regression prediction errors of 2.0 % for servers and 5.7 % for chiller-based air conditioners. Over a two-year period, energy savings were validated by simulating the workload patterns of a commercial data center in a container-based environment equipped with highperformance CPU servers and a chiller-based air conditioner. Results demonstrated significant energy savings during both hot and cold seasons:$\mathbf{3 7 \%}$for air conditioners,$\mathbf{2 3. 3 \%}$for servers, and$\mathbf{2 6 \%}$for the entire data center in hot season; and$\mathbf{3 3 \%}$for air conditioners,$\mathbf{2 3. 5 \%}$for servers, and$\mathbf{2 3 \%}$for the whole of the data center in cold season. These savings were achieved through a sequential optimization approach using Bayesian optimization for workload distribution and air conditioning control at$\mathbf{1 0}$-minute intervals.