Performance analysis and prediction of single-phase immersion cooling data center

With the rapid expansion of data centers, the need to address the challenge of high energy consumption has become increasingly urgent. The single-phase immersion cooling (SPIC) system has attracted widespread attention. To fully understand the impact of energy-consuming devices, this study systematically examined their effects on the average GPU temperature and power usage effectiveness (PUE). Then, a machine learning model is established to minimize PUE under different ambient temperatures. Results show that, increasing the power of energy-consuming devices results in an increase in PUE, with an optimal coolant pump power (1.5 kW) that yields the lowest average GPU temperature. Among the devices assessed, the coolant pump exerts the most significant influence on the average GPU temperature, followed by the dry cooler, coolant pump power, and ambient temperature. The developed machine learning model demonstrates a high level of precision, with a mean absolute percentage error (MAPE) of 4.86% and a regression coefficient (R2) of 0.9324. The SPIC system is shown to adequately meet the heat dissipation requirements in China. Under the optimal operational strategy, the PUE fluctuates between 1.04 and 1.13. And for the life cycle cost, the future cost would be decreased by 10-48 % after optimization in different regions.