Professional Documents
Culture Documents
1 s2.0 S2352710221002552 Main
1 s2.0 S2352710221002552 Main
1 s2.0 S2352710221002552 Main
A R T I C L E I N F O A B S T R A C T
Keywords: Data centers are large facilities housing numerous IT equipment and supporting infrastructure. Frequent vari
Artificial neural network ations in IT load, continuous removal/addition/replacement of IT equipment for business requirement, cooling
CFD modeling equipment, air supply settings, design layout, etc. make Data centers dynamic. Such complexities lead to
Datacenter
overcooling and increased energy consumption. To reduce the energy consumption of the data center, a real-time
Predictive model
Thermal management
control framework based on various thermal parameters inside the data center is imperative. Accurate prediction
of various variables affecting the thermal behavior of the data center, especially for the small-time horizon, using
computational fluid dynamics (CFD) simulations requires a large number of computational resources and
physical time, making them unfeasible for real-time control of the data centers. Data-driven modeling especially,
the Artificial Neural Networks (ANN) can be potentially helpful in such cases. This study aims to examine the
ANN-based model with Multi-Layer Perceptron (MLP) to predict thermal variables such as rack air temperature
inside data centers. The ANN-based models for the rack and facility-level system were trained and validated on
the experiments and validated CFD data. The optimum delay for each case was found using cross-correlation
between the input and output parameters of the ANN. The response of the multi-input multi-output ANN
model was validated using R-value and mean square error (MSE). R-value for all the cases was approximately
0.99. This study recommends the use of ANN models for fast and accurate prediction of thermal parameters for
real-time energy-efficient control of the data center system.
1. Introduction Power consumed by cooling units shares a large part of the total
energy utilized by the data center [5]. Due to time-varying job arrivals
A big data-warehousing facility consisting of several servers (hun on the servers and cold-hot aisle arrangements, the airflow inside the
dreds to thousands), cooling units, and IT equipment in a controlled data center is highly dynamic. However, the overall power consumption
environment is commonly known as a data center. These facilities are of the data center can be reduced by efficient cooling strategies. The
today’s one of the critical assets of internet giants and enterprises. Data knowledge of dynamic thermal parameters inside the data center is
centers are increasing rapidly in number with more than 4600 coloca imperative for the development of such control mechanisms.
tions worldwide and 39% of them situated in the USA [1]. Data centers Since the experimental investigation of the entire flow field is un
were considered for around 1% of global electricity use by Masanet et al. feasible and time-consuming, researchers performed computational
[2] with a prediction of a threefold increment within the next decade. fluid dynamics (CFD) analysis to understand data center airflow with
According to Jones [3], data centers used an estimated 200 TW-hours various parameters such as cold air supply-return temperature and rack
(TWh) in 2018 and contributed around 0.3% to overall carbon emission. inlet-outlet air temperatures [6–14]. CFD models require large compu
The environmental concerns and higher power cost has triggered the tational resources and processing-time due to complex airflow inside the
alarm by the environmental protection agencies. Estimations by Andrae data center. These restrictions make CFD models unfeasible for real-time
and Edler [4] suggest that usage of global electricity by data centers will forecasting required for the optimum cooling strategy.
rise to around 3–13% in 2030 compared to 1% in 2010. Instead, the data-driven models trained on the data obtained using
* Corresponding author.
E-mail addresses: anasrsaiyad@gmail.com, anashusen.saiyad@mtech2016.iitgn.ac.in (A. Saiyad), patelasif643@gmail.com (A. Patel), yogeshsfulpagare@gmail.
com (Y. Fulpagare), atul.bhargav@iitgn.ac.in (A. Bhargav).
https://doi.org/10.1016/j.jobe.2021.102397
Received 11 September 2020; Received in revised form 2 March 2021; Accepted 7 March 2021
Available online 17 March 2021
2352-7102/© 2021 Elsevier Ltd. All rights reserved.
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
Fig. 1. CAD model used for (a) rack (three racks) and (b) facility level (eight racks from R1 to R8) CFD simulations. The computational domain consists of a cold aisle
in the front of the rack with a cold air inlet from the bottom (No plenum). A hot aisle at the backside of the racks with hot air outlet to the ceiling. The middle rack (in
(a) & R7 in (b)) is divided into four compartments. Each rack was divided into four compartments in (b) [11], [39].
Table 1
ANN model parameters for the rack-level thermal load as an input.
Parameter Details
2
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
Fig. 4. Comparison of CFD and ANN response of rack air temperatures for thermal load input in the rack-level model (Here ‘nn’ represents ANN model results and
‘Sim’ represents CFD simulation results; ‘T_IN_R2_1’ depicts a first thermocouple placed at the inlet of the second server simulator from the bottom and similarly
for others.).
3
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
Fig. 5. Regression plot for thermal load variation in the rack-level system.
Table 2
ANN model parameters for rack-level tile air velocity as input based upon
simulation results.
Parameter Details
step-index.
ANN model for each case was trained using the feed-forward back
propagation method. Experimental as well as CFD data was divided for
each case into two parts. The models were trained using the first part.
Then, the models were validated based on predictions generated using
the remaining part. The low-pass Butterworth filter [37] was used for
Fig. 6. Cold air inlet velocity variation at the perforated tile for rack- the experimental data to reduce the effect of noise.
level system. The validation of the predicted results was performed based on two
[ ] parameters:
Xk+1 = Xk − J T J + μI − J T e (2)
1. R-value – Also known as the coefficient of correlation, it indicates the
where e is a vector of network errors, J is a Jacobian matrix that contains degree of relationship between two variables. R-value ranges be
the first derivatives of network errors with respect to the weights and tween − 1 and 1. The R-value of 1 indicates that both the variables
biases, μ is a scalar used to control the performance function, and κ is the are in perfect unison, whereas there is no relationship for 0 value.
4
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
Fig. 7. The response of rack temperature for air velocity variation in the rack-level system (here ‘nn’ represents ANN model results and ‘Sim’ represents CFD
simulation results).
2. Mean square error (MSE) – It measures the average squared differ heat generation rate of the servers on the Rack air inlet and outlet
ence between the estimated value by the model and the actual value temperatures was studied [8,11,13,35].
(equation (3): In actual servers, a part is occupied by the electronic components,
and the remaining part is void to allow the cooling airflow. Fans are
1 ∑N
( )2
placed at the exit of the server to develop a pressure gradient across the
MSE = ypredicted − yactual (3)
N i=1 server and thus, to facilitate cooling air to pass through the servers at the
desired flow rate. Modeling based upon actual internal geometry re
Cross-correlation between input and output variables was used to
quires a large computational cost and is not required for the analysis of
decide the optimum delay for the ANN model in each case. Some pa
the server as a whole. Therefore, in CFD modeling, the servers were
rameters have been assigned recommended values by the Neural
modeled as porous regions to mimic the actual flow physics. The
Network toolbox of MATLAB® for all the cases.
pressure-drop of the cold air due to internal components of the server
was incorporated by the viscous resistance and inertial resistance for the
3. Thermal modeling details
porous media. The exit of the servers was modeled as a fan boundary
condition to mimic the actual fan effects of the server simulator. The
The laboratory has a raised floor plenum chamber and ceiling return
server fan curve provided by the manufacturer [38] was used at the fan
type configuration. The cold air was supplied by the CRAC units under
boundary condition to relate the pressure drop and flow rate in
the plenum. At various desired locations in the cold aisle, porous tiles
teractions at the boundary. Each server simulator contributed equally to
were placed to allow cold air to enter into the racks. The hot air leaving
the total heat-load of a rack. CFD simulations were performed using the
from the rear side of the servers was collected and returned from the
standard k-ε model for incompressible flow along with the Boussinesq
ceiling to the inlet of the CRAC units.
hypothesis. A temporal discretization of 2nd order with a time step of
42U (1U = 4.45 cm) server simulator rack house 4 server simulators,
10− 3 s and a 2nd order upwind scheme was adopted for flow quantities.
each of height 10U. The server simulators were consisting of heaters and
four fans mounted at the back end. Each server simulator had a control
panel for heating load and a fan speed control dial. T-type thermocouple 4. Methodology and ANN models for various cases
sensors were placed at the front and rear side of the racks in the vicinity
of the servers to measure the Rack air inlet and outlet temperatures. The The ANN models were tested for four datasets available from vali
airflow rate at the perforated tile was measured using a tool having an dated CFD and experimental data on the rack and facility-level data
array of anemometers. In various sets of experiments, the effect of center. The optimum neural network for each case is decided based upon
variation in the CRAC supply air temperature, supply airflow rate, and various parameters such as auto-correlation, cross-correlation, overall
regression (R), and Mean Square Error (MSE). Delay for the network is
5
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
Fig. 8. Input and response variation trend for Rack-level experiment. (Here ‘nn’ represents ANN model results and ‘exp’ represents experimental results; ‘WM_ … ’
represent thermocouple location).
Input variables 1. Server heat generation rate – 1 4.1.1. Thermal load variation
Output variables 1. Server rack air inlet temperatures – 8 This ANN model consists of one input as thermal load variation in the
2. Server rack air outlet temperatures – 8
Delay 30 steps
second server simulator (Fig. 3) and output consisting of 16 rack air
Number of hidden layers 1 temperatures as mentioned in Table 1 (single input multiple outputs)
Number of neurons 10 with a time-step of 1 min. Cross-correlation is obtained between thermal
load variation (input) and rack air temperatures (output). From the
cross-correlation result, a delay of 30 steps is selected for the ANN model
determined based upon the autocorrelation data of the outputs and
and 10 neurons in the hidden layer. The available CFD data was divided
cross-correlation data of the inputs and the outputs. The optimum
into two parts as shown in Fig. 3. The initial part was assigned for
number of neurons is decided based on network performance.
training the neural network, and the remaining part for validation data.
The response of rack air temperatures is shown in Fig. 4 for selected
locations. As shown in Fig. 5, the overall Regression (R) value is 0.9958.
4.1. Rack level CFD data
Also, the Mean Square Error (MSE) value for this case is 0.58. Values of R
and MSE indicate that neural network response follows the trend and the
Two different cases considered are:
predicted values are closer to the actual values.
1. The rack air temperature variation for thermal load input inside the
4.1.2. Cold air inlet velocity variation
servers
In this case, the ANN model consists of one input (cold air inlet ve-
2. The rack air temperature variation for cold air inlet velocity input at
locity variation Fig. 6) and 16 rack air temperatures as output Table 1
the tile
(single input multiple outputs) with a time-step of 1 min. The optimum
network for this case was found to have a single layer of 10 neurons with
Two different neural networks were created for these two cases. The
20 steps delay based on the correlation values (Table 2).
middle rack as shown in Fig. 1 was divided into four compartments, each
The initial part of the data was assigned for training in the neural
depicting a server simulator. Two thermocouples were placed at the
6
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
Fig. 10. (a) input air velocity at the tile, (b) input heat generation rate, and (c) input cold air temperature variations at facility level CFD simulation data.
network (up to 355 min) and the remaining part for validation. The
Table 4 response of rack temperatures is shown in Fig. 7. In this case, the pre
ANN model parameters for facility-level model based upon simulation results. dicted output has a strong correlation with the target value of overall
Parameter Details Regression (R) as 0.9984, and MSE equal to 0.76.
Input variables 1. Tile air inlet temperature – 1
2. Tile air flow rate - 1
3. Facility heat generation rate – 1
4.2. Rack-level experimental data
Output variables 1. Rack air inlet temperatures – 32
2. Rack air outlet temperatures – 32 For the ANN model based on experimental results, the training data
Delay 5 steps was generated [17]. Each set of data consist of an input and output
Number of hidden layers 1
matrix. The input matrix has a thermal load on the server simulator
Number of neurons 10
(Fig. 8 (a)), and the output matrix has the rack air temperatures as
shown in Table 3. ANN model was generated for the prediction of
temperatures using all data points with a time-step of 1 min. These
7
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
Fig. 11. Response of rack temperature for the facility-level system (CFD). The thermocouple designated as ‘T_IN_R4_1’ depicts a first thermocouple placed at the inlet
of the first server simulator from the bottom, and similarly for others. (Here ‘nn’ represents ANN model results and ‘Sim’ represents CFD simulation results).
Table 5
ANN model parameters for facility-level model based upon experimental results.
Parameter Details
Table 6
R-value and MSE for all the cases.
Fig. 12. Input variation for the facility-level experiment. Case R-value MSE
8
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
Fig. 13. Rack air temperature response from the facility-level experiment (Here ‘nn’ represents ANN model results and ‘WM_’ represents thermocouple location).
temperatures WM10_J1, WM10_J2, WM10_J3 (sensor position in Fig. 9) model was trained to predict the rack air temperatures based on the
is shown in Fig. 8 (b)–(d). In this case, R-value was 0.97704, and the experimental data at all data points with a time-step of 1 min. These
maximum MSE of 1.12. predictions were compared with their experimental counterparts for the
same settings. The optimum neural network, in this case, has a single
4.3. Facility-level CFD data layer with 20 neurons and a delay of 5 time-steps (see Table 6).
In this case, available data was divided into four parts as shown in
CFD modeling was performed considering three variations for the Fig. 12, to have two data sets for each model. The results for selected
facility-level model: temperatures are shown in Fig. 13. The predicted results have a strong
correlation with the target values and are evident from the overall R-
1. The rack air temperature variation for thermal load input at the value of 0.9968 and MSE of 0.68.
facility
2. The rack air temperature variation for air velocity input at the tile 5. Conclusions
3. The rack air temperature variation for cold air temperature input at
the tile This study explores the usage of the Artificial Neural Networks model
using time series response techniques for the rack and facility-level
To have a general case, all three inputs were combined (Fig. 10). thermal predictions using multi-input multi-output (MIMO) data cen
Thus, the ANN model consists of 3 inputs and 64 rack air temperatures as ter system variables. Initially, the rack level CFD dataset was used for
output (Table 4) (multi input-multi output). The model was trained on testing of ANN model. This model was validated with the experimental
all data points with a time-step of 1 min. As shown in Fig. 1 (b), the dataset and then extended for the MIMO dataset from the facility level
facility-level system consisted of 8 racks with each rack divided into four CFD & experimental dataset. We observed that R-value for all the cases
compartments. Each compartment has a thermocouple at the inlet as was approximately 0.99, which shows a strong correlation between the
well as at the outlet resulting in a total of 64 temperatures. The optimum predicted value and the actual value with minimum MSE values. The
network for this case was found to have a single layer with ten neurons computational time for all the cases ranged from 10 to 30 s, which is
and five steps delay. insignificant as compared to the CFD simulation.
The responses of rack air temperatures are shown in Fig. 11 for The uniqueness of this study presents that the data center building
selected compartments and found to be in good agreement. Here, the dynamic environment can be predicted as ANN models using the real-
thermocouple designated as ‘T_IN_R4_1’ depicts a first thermocouple time and past datasets avoiding the use of computationally expensive
placed at the inlet of the first server simulator from the top, and similarly CFD models. One of the major challenges of developing a good predic
for others. The predicted output has a strong correlation with the target tion model that encompasses the dynamic building environment of the
with overall regression (R) value of 0.9987 and MSE value equal to 0.42. system can be resolved using the ANN model. Real-time prediction of
thermal parameters is imperative to reduce cooling energy and for
4.4. Facility-level experimental data optimal building energy management. Therefore, a real-time control
mechanism can be developed for the cooling system of the data center
In this case, the input matrix has cold supply air temperature, and the based on the ANN model response to optimize building energy demand
supply airflow rate as input parameters (Fig. 12), and the output matrix and management.
has 48 rack air temperatures (Table 5) (multi input-multi output). ANN
9
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
All persons who meet authorship criteria are listed as authors, and all [1] “Colocation Data Centers, Data Center Map, 2020. https://www.datacentermap.
com/datacenters.html (accessed Oct. 12, 2020).
authors certify that they have participated sufficiently in the work to [2] E. Masanet, A. Shehabi, N. Lei, S. Smith, J. Koomey, Recalibrating global data
take public responsibility for the content, including participation in the center energy-use estimates, Science (80-. ) 367 (6481) (Feb. 2020) 984–986,
concept, design, analysis, writing, or revision of the manuscript. https://doi.org/10.1126/science.aba3758.
[3] N. Jones, “How to stop data centres from gobbling up the world’s electricity,
Furthermore, each author certifies that this material or similar material Nature 561 (7722) (Sep. 2018) 163–166, https://doi.org/10.1038/d41586-018-
has not been and will not be submitted to or published in any other 06610-y.
publication before its appearance in the Journal of Building Engineering. [4] S.G.A. Andrae, T. Edler, On global electricity usage of communication technology:
trends to 2030, Challenges 6 (2015), https://doi.org/10.3390/challe6010117.
[5] J. Koomey, Growth in Data Center Electricity Use 2005 to 2010, 2011, https://doi.
Declaration of competing interest org/10.1088/1748-9326/3/3/034008. Oakland, CA.
[6] Y. Fulpagare, A. Bhargav, Advances in data center thermal management, Renew.
Sustain. Energy Rev. 43 (2015) 981–996, https://doi.org/10.1016/j.
The authors declare that they have no known competing financial rser.2014.11.056.
interests or personal relationships that could have appeared to influence [7] Y. Fulpagare, G. Mahamuni, A. Bhargav, Effect of plenum chamber obstructions on
data center performance, Appl. Therm. Eng. 80 (2015) 187–195, https://doi.org/
the work reported in this paper. 10.1016/j.applthermaleng.2015.01.065.
[8] Y. Fulpagare, Y. Joshi, A. Bhargav, Rack Level Transient CFD Modeling of Data
Acknowledgements Center, 2016.
[9] J. Rambo, Y. Joshi, Modeling of data center airflow and heat transfer: state of the
art and future trends, Distributed Parallel Databases 21 (2–3) (2007) 193–225,
All persons who have made substantial contributions to the work https://doi.org/10.1007/s10619-006-7007-3.
reported in the manuscript (e.g., technical help, writing and editing [10] A. Saiyad, Y. Fulpagare, A. Bhargav, Data center rack analysis using detached Eddy
simulations, in: Fifth International Conference on Computational Methods for
assistance, general support), but who do not meet the criteria for Thermal Problems THERMACOMP2018, 2018, pp. 172–176.
authorship, are named in the Acknowledgements and have given us their [11] Y. Fulpagare, Y. Joshi, A. Bhargav, Rack level transient CFD modeling of data
written permission to be named. If we have not included an Acknowl center, Int. J. Numer. Methods Heat Fluid Flow 28 (2) (Feb. 2018) 381–394,
https://doi.org/10.1108/HFF-10-2016-0426.
edgements, then that indicates that we have not received substantial
[12] P. Shirbhate, Y. Fulpagare, A. Bhargav, Effect of rack layout on data center thermal
contributions from non-authors. performance using cfd and experimental studies, in: Accepted Proc. 23rd Natl. Heat
Mass Transf. Conf. 1st Int. ISHMT-ASTFE Heat Mass Transf. Conf. IHMTC2015 17-
20 December, Thiruvananthapuram, India, 2015, pp. 1–8.
Nomenclature [13] Y. Fulpagare, Y. Joshi, A. Bhargav, Transient Characterization of Data Center
Racks, Nov. 2016, https://doi.org/10.1115/IMECE2016-66870.
[14] H.S. Erden, H.E. Khalifa, R.R. Schmidt, Room-Level Transient CFD Modeling of
Symbol Rack Shutdown, Jul. 2013, https://doi.org/10.1115/IPACK2013-73282.
Y Output of the neuron [15] M. Iyengar, R.R. Schmidt, Analytical modeling of energy consumption and thermal
F Transfer function performance of data center cooling systems - from the chip to the environment, in:
2007 Proc. ASME InterPack Conf. IPACK, 1, 2007, pp. 877–886, https://doi.org/
W Weight 10.1115/IPACK2007-33924, 2007.
P Input at the node [16] R. Lloyd, J. Hayes, M. Rebow, B. Norton, A data centre air flow model for
B Bias value predicting computer server inlet temperatures, in: Proc. 16th Intersoc. Conf.
Therm. Thermomechanical Phenom. Electron. Syst. ITherm, 2017, pp. 830–839,
N Number of inputs https://doi.org/10.1109/ITHERM.2017.7992572, 2017.
N Number of occurrences [17] Y. Fulpagare, Y. Joshi, A. Bhargav, Rack level forecasting model of Data center, in:
X Weight and bias matrices 2017 16th IEEE Intersociety Conference on Thermal and Thermomechanical
Phenomena in Electronic Systems, ITherm, 2017, pp. 824–829, https://doi.org/
J Jacobian matrix
10.1109/ITHERM.2017.7992571.
I Identity matrix [18] Y. Fulpagare, A. Bhargav, Y. Joshi, Predictive model development and validation
М Scalar for raised floor plenum data center, J. Electron. Packag. 142 (2) (Jun. 2020),
https://doi.org/10.1115/1.4046554.
E Network errors vector
[19] E. Samadiani, Y. Joshi, Proper orthogonal decomposition for reduced order
thermal modeling of air cooled data centers, J. Heat Tran. 132 (7) (Jul. 2010),
https://doi.org/10.1115/1.4000978.
Subscripts [20] N. Kansara, R. Katti, K. Nemati, A.P. Bowling, B. Sammakia, Neural Network
I Input index Modeling in Model-Based Control of a Data Center, 2015, https://doi.org/
J Neuron index 10.1115/IPACK2015-48684.
[21] J. Athavale, M. Yoda, Y. Joshi, Comparison of data driven modeling approaches for
Predicted Predicted value of the variable
temperature prediction in data centers, Int. J. Heat Mass Tran. 135 (Jun. 2019)
Actual Actual value of the variable 1039–1052, https://doi.org/10.1016/j.ijheatmasstransfer.2019.02.041.
K Step index [22] C. Deb, L.S. Eang, J. Yang, M. Santamouris, Forecasting diurnal cooling energy load
for institutional buildings using Artificial Neural Networks, Energy Build. 121
(2016) 284–297, https://doi.org/10.1016/j.enbuild.2015.12.050.
Superscript [23] Y.T. Chae, R. Horesh, Y. Hwang, Y.M. Lee, Artificial neural network model for
forecasting sub-hourly electricity usage in commercial buildings, Energy Build. 111
T Transpose of the matrix (2016) 184–194, https://doi.org/10.1016/j.enbuild.2015.11.045.
[24] K. Li, C. Hu, G. Liu, W. Xue, “Building’s electricity consumption prediction using
optimized artificial neural networks and principal component analysis, Energy
Abbreviations Build. 108 (2015) 106–113, https://doi.org/10.1016/j.enbuild.2015.09.002.
ANN Artificial Neural Network [25] F. de Lorenzi, C. Vömel, Neural network-based prediction and control of air flow in
CFD Computational Fluid Dynamics a data center, J. Therm. Sci. Eng. Appl. 4 (2) (2012), 021005, https://doi.org/
10.1115/1.4005605.
CNN Convolutional Neural Network [26] Z. Song, B.T. Murray, B. Sammakia, Airflow and temperature distribution
GPR Gaussian Process Regression optimization in data centers using artificial neural networks, Int. J. Heat Mass
LM Levenberg-Marquardt Tran. 64 (2013) 80–90, https://doi.org/10.1016/j.
ijheatmasstransfer.2013.04.017.
MIMO Multi Input Multi Output [27] S.K. Shrivastava, J.W. VanGilder, B.G. Sammakia, Data center cooling prediction
MLP Multi-layer Perceptron using artificial neural network, in: Proceedings of the 2007 InterPACK Conference,
MSE Mean Square Error 2007, pp. 1–7.
[28] Z. Song, B.T. Murray, B. Sammakia, Multivariate prediction OF airflow and
POD Proper Orthogonal Decomposition
temperature distributions, in: Proceedings of the ASME 2011 Pacific Rim Technical
R Regression Conference & Exposition on Proceedings of the ASME InterPACK Conference
SVR Support Vector Regression
10
A. Saiyad et al. Journal of Building Engineering 42 (2021) 102397
Packaging and Integration of Electronic and Photonic Systems InterPACK 2011, [32] R. Evans, J. Gao, DeepMind AI Reduces Google Data Centre Cooling Bill by 40%,
2017, pp. 1–10, https://doi.org/10.1115/IPACK2011-52167. 2016.
[29] N. Liu, X. Lin, Y. Wang, Data center power management for regulation service [33] J. Gao, Machine Learning Applications for Data Center Optimization, 2014.
using neural network-based power prediction, in: 2017 18th International [34] R. A. Evans, J. Gao, M. C. Ryan, G. D.- Arnold, J. K. Scholz, and T. A. Hester,
Symposium on Quality Electronic Design, ISQED, 2017, pp. 367–372, https://doi. “Optimizing Data Center Controls using Neural Networks,” US 2018/0204116 A1,
org/10.1109/ISQED.2017.7918343. 2018.
[30] Y.W. Foo, C. Goh, H.C. Lim, Z.-H. Zhan, Y. Li, Evolutionary neural network based [35] Y. Fulpagare, A. Bhargav, Y. Joshi, Dynamic thermal characterization of raised
energy consumption forecast for cloud computing, in: 2015 International floor plenum data centers: experiments and CFD, J. Build. Eng. 25 (Sep. 2019)
Conference on Cloud Computing Research and Innovation, ICCCRI, 2015, 100783, https://doi.org/10.1016/j.jobe.2019.100783.
pp. 53–64, https://doi.org/10.1109/ICCCRI.2015.17. [36] S.N. Sivanandam, S. Sumathi, S.N. Deepa, Introduction to Neural Networks Using
[31] Z. Song, B.T. Murray, B. Sammakia, A dynamic compact thermal model for data MATLAB 6.0, Tata McGraw-Hill Education Private Limited, 2005.
center analysis and control using the zonal method and artificial neural networks, [37] S. Butterworth, On the theory of filter amplifiers, Exp. Wirel. Wirel. Eng. 7 (1930)
Appl. Therm. Eng. 62 (1) (2014) 48–57, https://doi.org/10.1016/j. 536–541.
applthermaleng.2013.09.006. [38] G. Nelson, Development of an Experimentally-Validated Compact Model of a
Server Rack, Georgia Institute of Technology, 2007.
11