Parameter prediction optimization of data center’s heat dissipation system using machine learning algorithms

Abstract

In this study, an optimization control method based on knowledge and operational data was investigated for a centralized water-cooling heat dissipation system in a data center. First, a numerical model describing the heat transfer and thermodynamic processes of the heat dissipation system was established, and five thermal characteristic parameters were extracted to identify the system. The thermal characteristic parameters, used as inputs to the model, were identified in real time using the mechanism formula from the operational data. Subsequently, the operational data of the heat dissipation system in a real data center were used to validate the numerical model. The model accuracy was within 15 %. The water flow rate of the actual system was optimized, showing that the average PUE of four typical cases under the optimized conditions were 1.12, 1.25, 1.13 and 1.3, respectively, with an improving rate between 4.3 % and 12.4 %. Next, considering the reasonable value ranges of the five thermal characteristic parameters and four direct input parameters, 62,208 groups of working conditions were designed, and the optimized water flow rate and the corresponding COP and PUE for each working condition were obtained using the numerical model. Finally, a prediction model that combined the heat transfer mechanisms was established based on an optimized dataset. Machine learning algorithms were applied to establish neural networks between the nine inputs and three optimized operating parameters. Based on the prediction model and thermal characteristic parameter identification module, software was developed to suggest optimized water flow rates and predict the PUE and COP in real time from the operational data.