Abstract As server densities increase to support the rising demand for high density computing, traditional room cooling infrastructure is struggling to keep up. Rack-mounted cooling units are increasingly being deployed in data centers. These cooling units are located closer to the heat sources (the servers) which allow them to cool more efficiently than traditional cooling infrastructure. In this work, a Multi-Input Single-Output (MISO) Adaptive Predictive Controller (APC) for rack-mounted cooling units is investigated and implemented using a low-cost general-purpose microcontroller. The proposed APC is implemented using Weighted Recursive Least Squares (WRLS) and a sub-optimal but fast algorithm based on the General Predictive Controller (GPC) approach. These are combined with a variable forgetting factor and variable prediction horizon algorithms. In addition, methods are proposed to handle stability issues arising due to practical hardware limitations. The controller is implemented on a real single rack system and challenges for practical implementation are addressed and illustrated. The proposed APC is also compared (via simulation) to a standard APC with equivalent complexity and a split-range PI controller. The results show that the proposed controller outperforms both the standard APC and the split-range PI controllers with respect to Mean Squared Error (MSE).

中文翻译：

---

数据中心冷却装置的自适应预测控制

摘要 随着服务器密度的增加以支持对高密度计算不断增长的需求，传统的房间冷却基础设施正在努力跟上。机架式冷却装置越来越多地部署在数据中心。这些冷却装置更靠近热源（服务器），这使得它们比传统的冷却基础设施更有效地冷却。在这项工作中，使用低成本的通用微控制器研究并实现了用于机架式冷却装置的多输入单输出 (MISO) 自适应预测控制器 (APC)。建议的 APC 是使用加权递归最小二乘法 (WRLS) 和基于通用预测控制器 (GPC) 方法的次优但快速算法实现的。这些与可变遗忘因子和可变预测范围算法相结合。此外，还提出了处理由于实际硬件限制而引起的稳定性问题的方法。该控制器是在一个真正的单机架系统上实现的，并且解决和说明了实际实现的挑战。所提出的 APC 还（通过仿真）与具有等效复杂性和分程 PI 控制器的标准 APC 进行了比较。结果表明，所提出的控制器在均方误差 (MSE) 方面优于标准 APC 和分程 PI 控制器。该控制器是在一个真正的单机架系统上实现的，并且解决和说明了实际实现的挑战。所提出的 APC 还（通过仿真）与具有等效复杂性和分程 PI 控制器的标准 APC 进行了比较。结果表明，所提出的控制器在均方误差 (MSE) 方面优于标准 APC 和分程 PI 控制器。该控制器是在一个真正的单机架系统上实现的，并且解决和说明了实际实现的挑战。所提出的 APC 还（通过仿真）与具有等效复杂性和分程 PI 控制器的标准 APC 进行了比较。结果表明，所提出的控制器在均方误差 (MSE) 方面优于标准 APC 和分程 PI 控制器。