Artificial Neural Network Based Prediction of Control Strategies for Multiple Air-Cooling Units in a Raised-floor Data Center

摘要

A data center cooling system consists of a hierarchy of systems with dedicated control algorithms dictating their operational states. There exists a wide range in spatial and temporal parameter space in an ensemble of non-linear dynamic systems, each executing a control task, while the global objective is to drive the overall system to an optimum operating condition i.e. minimum total operational power at desired rack inlet temperatures. Certainly, it is beneficial in optimizing workload migration at temporal scales but, solving the instability of the cooling systems operating at design points helps in understanding the whole system and make predictions to have better control strategies. Several techniques are available to realistically capture and make predictions. Datadriven modelling/Machine learning is one such method that is less expensive in terms of cost and time compared to other methods like validated CFD simulation/experimental setup.The objective of this study is to develop a control framework based on predictions made using machine learning techniques such as Artificial Neural Network (ANN) to operate multiple Computer Room Air Conditioning Units (CRAC) or simply Air-Cooling Units (ACU) in a hot-aisle contained raised floor datacenter. This paper focuses on the methodology of gathering training datasets from numerous CFD simulations (Scenarios) to train the ANN model and make predictions with minimal error.Each rack has a percentage of influence (zones) based on the placement of ACUs and their airflow behavior. These zones are mapped using steady state CFD simulation considering maximum CPU utilization and cooling provisioning. Using this map, ITE racks are targeted and given varying workload to force the corresponding ACU that is responsible for provisioning, to operate at set points. Number of such scenarios are simulated using the same CFD model with fixed bounds and constraints. Using large samples of data collected from CFD results, the ANN is trained to predict values that correspond to the activation of the desired ACU. Such efficient control network would minimize excessive cooling. The validated prediction points are used to model a control framework for the cooling system to quickly reach the operating point. These models can be used in real-time data centers provided; the training data is based on in-house sensor values.