Holistic Optimization of Data Center Cooling Systems and Airflow Management

Han, Xu

Graduate Thesis Or Dissertation

Holistic Optimization of Data Center Cooling Systems and Airflow Management Public Deposited

Analytics

Télécharger le fichier PDF

Citations

Citeable URL: https://scholar.colorado.edu/concern/graduate_thesis_or_dissertations/1n79h538v

Abstract

Data centers are often overcooled to address the reliability concerns, which may lead to a lower cooling efficiency. Model-based optimization in a holistic view can be utilized to improve the cooling efficiency while meeting the stringent thermal requirements of data centers, which usually involve non-uniform thermal environment. The co-simulation of building energy simulation (BES) and computational fluid dynamics (CFD) is promising to fulfill the needs. However, the existing BES-CFD co-simulation methods are not good enough for practical use in the real world because of either huge computational costs with BES-CFD coupled models or accuracy issues with BES-ROM-CFD coupled models that integrate reduced order models (ROMs). This dissertation therefore intends to improve the simulation methods from the following aspects to allow a computationally-practical and sufficiently accurate holistic optimization of cooling systems and airflow management towards energy efficient data centers with reliable operation.

First, as CFD is the bottleneck of the BES-CFD co-simulation speed, this dissertation develops an improved Fast Fluid Dynamics (FFD) model for fast and accurate simulation of data center airflow and thermal environment. FFD, which is a simplified CFD method, has been applied for fast airflow simulation. However, few research applied FFD for optimal design and operation of data center thermal management. This dissertation improves the FFD model for data centers and conducts a comprehensive evaluation and demonstration. The FFD model is improved by 1) proposing new algorithms to solve equations for a better general quantity conservation than conventional FFD; and 2) adding new features and boundary conditions for data centers. The validation with a real data center shows that the new FFD model achieves a similar level of accuracy as CFD when compared to the experimental measurements, and is 61 times faster than CFD for the studied case.

Then, to improve the existing BES-CFD co-simulation methods, this dissertation proposes a new online BES-ROM-CFD co-simulation method. Among existing models, BES-CFD is accurate but time consuming, and BES-ROM-CFD is fast but may have accuracy issues. The proposed online BES-ROM-CFD method combines the advantages of both existing models, in which the ROM allows online learning and automatic error control. This methodology is realized by implementing a Modelica-ISAT-FFD model, which will be officially released in Modelica Buildings library to allow a broader range of applications. The new model is then comprehensively evaluated with a mixed convection case. The results show that the new model can generally control the prediction error within user-defined tolerances compared to an existing Modelica-FFD model that was validated by previous research. An annual simulation shows that the new model saves up to 95.7% of computing time against the existing Modelica-FFD model. A space heating case is also studied to demonstrate the capability of the new model to handle scenarios with feedback loop controls.

To further improve the online BES-ROM-CFD method, this dissertation proposes an adaptive online BES-ROM-CFD method, in which adaptive coupling frequencies are used to reduce the number of ROM-CFD calls during the co-simulation. It is further powered by distributed computing, which allows that BES-ROM runs on a CPU and CFD runs on a GPU in parallel during the co-simulation. Based on the new methods, an adaptive Modelica-ISAT-FFD model is developed, which is verified against an existing Modelica-FFD model with a real middle-size data center. A holistic optimization platform is proposed based on the new model, which is then evaluated and demonstrated with optimization studies. It is found that a holistic optimization with annual co-simulations for a middle size data center in the real world can be finished within a day, which is estimated to take as much as many years if using the existing Modelica-FFD model. The results also show that the holistic optimization saves the annual energy consumption by up to 48.1% while meeting the data center thermal requirements.

In addition to the off-operation model-based optimization, this dissertation further proposes a novel machine learning assisted expert system (MLES) method towards a real-time optimal control for the studied data center cooling system. Model predictive control (MPC) has been widely studied for optimal control. Though it can achieve a good performance theoretically, formulating a precise and real-time model is often not easy especially for complex systems, such as data center cooling systems involving stratified airflow. To address this problem, this dissertation proposes a novel robust and easy-to-implement model-free MLES optimal control method. First, the optimization problem is simplified to allow formulating an expert system based on expert knowledge. Then, a machine learning model trained by more than a thousand of CFD simulations is used to assist the expert system for hot spots control towards reliable operation. A MPC is also implemented as benchmark. The case studies show that the MLES could achieve a similar performance as a well-designed ideal MPC, and is much faster than MPC. An annual simulation shows that with the MLES, the energy consumption is saved by up to 64.6% and meanwhile hot spots are generally well controlled.

To conclude, this dissertation concentrates on improving the data center cooling efficiency while maintaining reliable operation in a holistic view. The outcomes include 1) an open source model for fast simulation of data center thermal environment, 2) an open source model for fast and accurate co-simulation of cooling systems and airflow management, 3) a computationally-practical and sufficiently accurate optimization platform for holistic optimization of cooling systems and airflow management, 4) a novel robust and easy-to-implement optimal control method that simultaneously considers cooling efficiency and reliable operation. The proposed models and methodologies may not be limited in data center applications, which can be expanded for studying other cooling systems involving non-uniform thermal environment in future research.

Creator