Intelligent control technology for liquid-cooled data centers based on a dual-drive model integrating mechanism and data models

Amidst the expanding infrastructure of the information age, cooling systems represent a substantial and growing portion of global data center energy consumption. Researchers have explored various methods to reduce energy consumption in these systems, with model-based intelligent optimization control being the most commonly employed approach. The most prevalent models in this field are thermodynamic mechanism models and data-driven models, yet each of these standalone models has its own advantages and limitations. Therefore, this study innovatively develops a dual-drive gray-box model that integrates mechanism and data models. It supplements data with a mechanism model possessing strong generalization capabilities to form a fusion sample library, and trains the fusion library using optimal data-driven algorithms. Finally, a proprietary optimization control logic is used and a lot of energy saving is realized under the premise of meeting the chip safety. Comparing mechanism-based models, data-driven models, and the dual-drive model reveals that the proposed dual-drive model combines the strong generalization capability, low data cost requirements, and high optimization reliability of mechanism-based models with the fast optimization efficiency and high model accuracy within the data collection range of data-driven models. Finally, experiments validate that the optimization control based on dual-drive model can maintain chips within safe upper limits while achieving substantial energy savings—up to 69.5% in Test Case 1. And comparing different cases, we can get the influence of resource scheduling uniformity on model-based optimal control logic. This unique optimization control framework based on the dual-drive model proposed by this paper resolves industry challenges in practical intelligent control implementation, offering novel insights for modeling data center cooling systems.