0
Research Papers

Neural Network-Based Prediction and Control of Air Flow in a Data Center

[+] Author and Article Information
Flavio de Lorenzi, Christof Vömel

 Zurich University of Applied Sciences, ZHAW, CH-8401 Winterthur, Switzerland

J. Thermal Sci. Eng. Appl 4(2), 021005 (Apr 16, 2012) (8 pages) doi:10.1115/1.4005605 History: Received August 30, 2011; Accepted November 07, 2011; Published April 16, 2012; Online April 16, 2012

As modern data centers continue to grow in power, size, and numbers, there is an urgent need to reduce energy consumption by optimized cooling strategies. In this paper, we present a neural network-based prediction of air flow in a data center that is cooled through perforated floor tiles. With a significantly smaller execution time than computational fluid dynamics, it predicts in real-time server inlet temperatures and can detect whether prevalent air flow cools the servers sufficiently to guarantee safe operation. Combined with a cooling system model, we obtain a temperature and air flow control algorithm that is fast and accurate enough to find an optimal operating point of the data center cooling system in real-time. We also demonstrate the performance of our algorithm on a reference data center and show that energy consumption can be reduced by up to 30%.

FIGURES IN THIS ARTICLE
<>
Copyright © 2012 by American Society of Mechanical Engineers
Your Session has timed out. Please sign back in to continue.

References

Figures

Grahic Jump Location
Figure 1

Floor plan of the reference data center. Picture from Ref. [10].

Grahic Jump Location
Figure 2

Cooling system model for power consumption of 40 kW, from Refs. [10-11]. Dashed lines indicate the flow path during free cooling operation and red lines mark the chiller (refrigerant: R134a).

Grahic Jump Location
Figure 3

Left: Temporal evolution of cooling system power consumption and ambient temperature for 2009 [12] . Note that without free cooling, power consumption roughly follows the ambient temperature evolution. Right: Energy consumption per year as a function of DC supply air temperature for different CRAC air flows. The solid black line corresponds to energy savings of 6%/deg increased supply air temperature.

Grahic Jump Location
Figure 4

Comparison of the temperature map measured by IBM using their Mobile Measurement Technology [16] (left panels) with CFD/HT simulation (right panels). The upper panels show a plane through rack 1 and the lower ones show a plane cutting rack 2.

Grahic Jump Location
Figure 5

Schematic view of a feed-forward neural network

Grahic Jump Location
Figure 6

Illustration: convergence with respect to the error measure F as a function of iteration step

Grahic Jump Location
Figure 7

Contour plot of the sample averaged total prediction error F depending on the number of neurons in two hidden layers

Grahic Jump Location
Figure 8

Training data sets: scatter plots of CFD measurements versus neural network predictions. The bottom panel shows the CRAC return temperature and the remaining panels compare server inlet temperatures.

Grahic Jump Location
Figure 9

Test data sets: scatter plots of CFD measurements versus neural network predictions. The bottom panel shows the CRAC return temperature and the remaining panels compare server inlet temperatures.

Grahic Jump Location
Figure 10

Racks with 12 servers each (illustration); the four additional modules at the tops hold miscellaneous equipment. Dots represent sensors. Picture from Ref. [10].

Grahic Jump Location
Figure 11

From top to bottom: Yearly evolution of ambient air temperature, CSM power consumption, CRAC supply air temperature and mass flow rate

Grahic Jump Location
Figure 12

Illustration: volumetric air flow (bottom) is adjusted according to changing power consumption of racks (top)

Tables

Errata

Discussions

Some tools below are only available to our subscribers or users with an online account.

Related Content

Customize your page view by dragging and repositioning the boxes below.

Related Journal Articles
Related eBook Content
Topic Collections

Sorry! You do not have access to this content. For assistance or to subscribe, please contact us:

  • TELEPHONE: 1-800-843-2763 (Toll-free in the USA)
  • EMAIL: asmedigitalcollection@asme.org
Sign In