The document presents a methodology for predicting energy consumption using a non-invasive hardware prototype based on embedded systems to optimize energy use in data centers. It details a system model utilizing ESP32 hardware for data collection and various regression algorithms for prediction accuracy, highlighting results that demonstrate varying levels of fit depending on the time window analyzed. The conclusion indicates that while hourly predictions show high accuracy, daily predictions indicate an underfitting issue, suggesting further refinement may be needed.

2. Learning-based Energy
Consumption Prediction
Presenter by: Víctor Asanza, Ph.D
Escuela Superior Politécnica del Litoral, ESPOL, Guayaquil, Ecuador
Centro de Tecnologías de Información, CTI
Facultad de Ingeniería en Electricidad y Computación, FIEC
Rebeca Estrada Pico , Víctor Asanza , Adrián Bazurto , Irving Valeriano , Danny Torres

3. Published in:
https://doi.org/10.1016/j.procs.2022.07.035

4. Learning-based Energy Consumption Prediction
OUTLINE
• Objetive
• Related Work
• System Model
• Methodology
• Numerical results
• Conclusion

5. OBJECTIVE
Emerging technology of embedded systems
enables the implementation of prediction
models.
A non-invasive hardware prototype to
record the energy consumption of a
workstation.
With this prototype we can optimize and
reduce energy consumption and
associated maintenance costs IT
Equipment in the Data Center.

6. RELATED WORK
Energy
Prediction
Methods
INPUT
VARIABLES
OUTPUT
VARIABLES
A.I
ALGOROITHM
• CPU Workload, Memory disk workload, I/O Unit Workload [9].
• Quantity Data to process [10].
• Number of file downloads on Servers [12].
• Speed of energy consumption from a server [13].
• Energy consumption [8].
• Relative and absolute average of workloads [9].
• CPU frequency [10].
• Number of abstract views on the servers [12].
• Energy consumption from data center servers [13].
• Linear prediction weighted [8]. Tests were done on 5, 10 and
20minute interval.
• Neuronal Network BP, Elman and LSTM TIME lapses in seconds [9].
• Support vector regression data was collected for 1 year [13].

7. SYSTEM MODEL: ARCHITECTURE
CTI (https://www.cti.espol.edu.ec/)

8. SYSTEM MODEL: ARCHITECTURE
Data Sources Data Collector Output Data

9. SYSTEM MODEL: ARCHITECTURE
Data Source
Data Collector
Output Data
Voltage Source (110V-220V) + Module
PZM004T V3.0 + Microcontroller ESP32
MQTT BROKER + NODE RED + MySQL Database
Telegram + Dashboard
WIFI / MQTT
BOT TELEGRAM/MQTT/HTTP

10. SYSTEM MODEL: HARDWARE DESIGN
Data acquisition equipment is based on the ESP32 hardware.
To measure the voltage, current, power, frequency, energy and power
factor, we use PZMT-004. (Fig.1)
The energy measurement system consists of:
◦ Sensor Network with the ESP32 module.
◦ The communication protocol MQTT is used.
A script running as a daemon on the workstation, which is enabled to
record the measurements from CPU and Memory RAM data.
Fig. 1. 3D design and PCB of the energy consumption meter
developed

11. METHODOLY
DATA ACQUISITION
PREPROCESSING
DATA
VARIABLE
SELECTION
REGRESSION
LEANER

12. DATA ACQUISITION
2. Data Acquisition
Data is registered on the database ”Data
Server Energy Consumption”. (Fig. 3)
Data was collected from a workstation for
120 days.
Fig. 3 Hourly Recorded Variables
Fig. 2 Procesos in Workstation
Data Collector
Data Source Output Data
BOT TELEGRAM
MQTT
1. Servers/Workstation
Two processes:
◦ Telegram BOT
◦ MQTT Protocol. (Fig.2)

13. The dataset used for data processing are available in:
https://ieee-dataport.org/open-access/data-server-energy-consumption-dataset
*120 days with a sampling
frequency of a value per
hour.

14. METHODOLY
DATA ACQUISITION
PREPROCESSING
DATA
VARIABLE
SELECTION
REGRESSION
LEANER

15. PREPROCESSING DATA
A normalization of the dataset values was carried out as pre-processing.
The normalization is needed because the value range for the selected
features are different. (Fig.4)
The normalization was carried out considering the maximum and
minimum values of each of the variables.(Eq. 1)
Fig. 4. Normalization of the dataset values
Eq.1. Min–max formula used for normalization

16. METHODOLY
DATA ACQUISITION
PREPROCESSING
DATA
VARIABLE
SELECTION
REGRESSION
LEANER

17. VARIABLE SELECTION
Correlation matrix of the proposed
variables is estimated using a
MATLAB script. (Fig. 5)
Selected Features:
1. Voltage
2. Frequency
3. Active Power RMS
4. Active Energy RMS
5. ESP32_temp.
Fig. 5: Correlation Matrix Fig. 6. Normalized Data (hour)

18. METHODOLY
DATA ACQUISITION
PREPROCESSING
DATA
VARIABLE
SELECTION
REGRESSION
LEANER

19. REGRESSION LEARNER
The dataset is distributed as follows:
◦ 70% to train the algorithms
◦ 15% to test the accuracy degree (Table1.)
◦ 15% to validate the prediction model
Table 1. RMSE Testing Comparison of Linear
Regression algorithms

20. The matlab code used for data processing are available in:
https://github.com/vasanza/Matlab_Code/tree/EnergyConsumptionPredictionDatacenter

21. RESULTS
During the validation process:
◦ Hourly predicted values are similar to real
energy consumption with a MAE value of
0.0885 [kWh] and a RMSE of 0.025712
[kWh] .
◦ Daily prediction values showed that the
model is an underfitted model because of its
MAE of 3.2255 [kWh] and a RMSE of
3.25029 [kWh],

22. CONCLUSION
1. Energy consumption in days, presents an RMSE of 3.25029 [kWh], which is an
indicator that the model is underfitting in this time window.
2. Energy consumption in One-Hours, presents an RMSE of 0.025712 [kWh], which is
an indicator that the model is not over-fitted in this time window.
3. Robust Linear Regression Model, was selected based on the RMSE of the energy
consumption predicted value.

23. For more information
Mail: {restrada, vasanza,abazurto, ivaleria, daaltorr}@espol.edu.ec
Centro de Tecnologías de Información, CTI
Escuela Superior Politécnica del Litoral, ESPOL
Campus Gustavo Galindo Km 30.5 Vía Perimetral, P.O. Box 09-01-5863
090150 Guayaquil, Ecuador
Rebeca Estrada Pico , Víctor Asanza , Adrián Bazurto , Irving Valeriano , Danny Torres

24. Thank you!