See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/342956749

A Smart Building Energy Management using Internet of Things (IoT) and

Machine Learning

Article in Test Engineering and Management · July 2020

CITATIONS READS

4 553

6 authors, including:

Nor Azuana Ramli Mohd Badrulhisham Ismail

Universiti Malaysia Pahang University of Kuala Lumpur
40 PUBLICATIONS 99 CITATIONS 6 PUBLICATIONS 8 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The Optimal Location of STATCOM and BESS in Smart Grid System for Voltage Stability Improvement using PSO Adaptive Neural Network Hybrid Algorithm View project

DEVELOPMENT OF A PREDICTIVE ALGORITHM FOR HEART DISEASE USING AN ENSEMBLE OF LOGIT MODEL AND NEAREST NEIGHBOUR TREE METHOD BASED ON
SMARTWATCH DATA View project

All content following this page was uploaded by Nor Azuana Ramli on 15 July 2020.

The user has requested enhancement of the downloaded file.

 A Smart Building Energy Management using
Internet of Things (IoT) and Machine Learning
Nursyura Laili Mazlan, Nor Azuana Ramli, Lilik Jamilatul Awalin, Mohd Badrulhisham Ismail*, Azizah Kassim~,
Ashwin Menon°
*Electrical Engineering Section, Universiti Kuala Lumpur British Malaysian Institute, Gombak, Selangor, Malaysia.
([email protected], [email protected], [email protected], [email protected])
~
Green 3 Sdn Bhd, Petaling Jaya, Selangor, Malaysia. ([email protected])
°Reneon Technologies Sdn. Bhd. ([email protected])

Abstract—Energy consumed by smart buildings in efficiency by simplifying the system connection in using
Malaysia accelerates continuously due to the growth of computerized process [4]. All the electrical appliances can be
country’s population. This occurrence creates awareness controlled digitally and offer handiness towards consumers
towards energy management team to organize and manage in managing their energy management. A smart building
their energy consumption systematically. The main purpose of
this research is to analyse and predict the energy consumptions
with the help of Internet of Things (IoT) is the most effective
in order to achieve a better energy management of a ways in supporting energy efficiency management and
commercial smart building towards efficiency. In this paper, monitoring energy consumption [5]. IoT provides a digital
historical data of hourly consumption of maximum demand platform to observe and supervise the energy consumption in
collected at the selected tenants of the smart commercial a huge scale of data in the presence of internet. This
building implemented with Internet of Things (IoT) has been technology also helps in reducing the energy costs and
analysed using statistical method computed with the formula of charges by enhancing the way in managing the energy
mean, variance, skewness and kurtosis. k-nearest neighbour consumption of building [6]. It is crucial for consumers
(k-NN) method has been applied on the data of consumptions especially energy management team to have a deep
for prediction process by using three different values of nearest
neighbour. The predicted data has been separated into
comprehension on the pattern of their smart building energy
different training and testing ratios which are 70% and 30%. consumption. Statistical method is one of the analysis
Root mean square error (RMSE) is proposed in this paper to techniques that are used to obtain data regarding the energy
evaluate the performance of predicted data. The results consumption of building [7]. Analysing of energy
showed that the nearest neighbour with k = 5 is the most consumption is really important in improving energy
accurate since it provides the lowest average RMSE value with efficiency and cost saving for a building [8]. Hence,
5.73, 8.54 and 0.35 for each tenant of the building respectively. statistical method is proposed as the main objective in this
This model will be used as a reference to predict the energy research is to do statistical analysis for energy consumption
usage for the upcoming period. In the future, data that has of a smart building.
been predicted can be integrated into the available system for
user monitoring and controlling purposes. Prediction also brings convenience towards consumers
and energy management team in estimating their usage and
Keywords—Internet of Things, smart building, maximum planning for upcoming period to improve the energy
demand, machine learning, k-nearest neighbour
efficiency of a smart building. To achieve the next objective
of the research, k-nearest neighbour (k-NN) method is
I. INTRODUCTION proposed in this study in order to predict energy consumption
Energy consumptions in Malaysia increase gradually due for the upcoming period. This method known as one of the
to the growth of population which is in year of 2018, it hits prediction technique that helps in energy management
32 million people and will be estimated to increase more in efficiency by minimize the energy consumption towards a
the future [1]. The growth of population lead to the building [9]. Moreover, this method is considered to be one
increasing of energy demand in this country and have been of the simplest classification techniques used for prediction
estimated to reach 116 million tons of oil equivalents (mtoe) [10]. This method also is a non-parametric and simple
by the year 2020. Energy provided in Malaysia is influenced learning algorithm that uses a database concerning the
by the main fossil fuel sources which included coal, natural energy consumption to predict and make a classification of
gas and fuel oil. Buildings which including commercial, new sample data [11]. Prediction method will encourage
residential and industrial in our country utilizes a total of energy management team to supervise their energy
48% of the electricity that have been created [2]. The consumption that will lead the management system to
increasing of energy consumptions towards buildings from become more manageable and systematic towards achieving
day to day create enforcement to this country in managing energy efficiency. Predicting their energy consumption also
and reducing the energy consumption as much as possible in becomes substantial in order to control the energy usage and
order to improve energy efficiency. reduce the expenses on energy cost for billing purposes.
Energy efficiency can be improved by applying a smart All the data used for analysis and prediction method were
system management towards a building. Smart building is a collected based on real time data from a smart commercial
cutting edge technology structures that can improve energy building located in Klang Valley, Malaysia. This paper is
separated into four sections. Comparison between present User can observe and monitor the collected data internally
studies with literatures that has been reviewed will be through an online platform by giving an ID number and
discussed in the first section. Next, all methods for analysis password.
and prediction that have been applied in this study will be
deliberated in details. The results and findings obtained from
both methods will be discussed in the last section.

II. LITERATURE REVIEW

There are many different approaches for energy
consumption analysis and prediction that has been done
from the previous studies. Energy forecasting is really
important in improving the energy utilization rate and helps
the energy management system of a building to make a
better evaluation [12]. For this present study, analysis and Fig. 1. System architecture of the smart commercial building
prediction method using statistical and k-NN method give
attention on maximum demand measured in kilowatt for The online website really helps the users and team
selected tenants of a two smart commercial building, management in monitoring their building energy
building A and B. Study done by [13] analysed energy consumption data using systematic and manageable online
consumption of electricity, oil and gas using time series system. It provides graph visualisation that shows the data
combined with regression method. Data that has been collected for every hour in order to help the users in
analysed will be used for prediction method affected by visualizing the behaviour or trend of the maximum demand
seasonal factors. Result for this study demonstrates the of each tenant. For example, Figure 9 and 10 in the
accuracy of the method used. Building A shows that the appendix shows the images that represented hourly data of
error for predicted annual electricity, oil and gas are 6.326% maximum demand (MD) for tenant A1 and A2 collected by
and 1.753% meanwhile for building B the predicted the smart meter reading that has been provided via the
consumption error are 1.0221% and 5.885%. website in June 2018.
Analysis using statistical method has been proposed by Collected data can be extracted manually from the online
[10] on energy consumption of apartment’s building in order platform in the form of CSV file. Data of maximum demand
to split the power consumption used into two categories consumption measured in kilowatt starting from June 2018
which are lower power consumption and high power until December 2018 of each tenant were chosen as the main
consumption. Data of hourly power consumption for all data used for analysis and prediction purposes that will be
home equipment based on their daily used has been explained comprehensively in next subsection.
collected in order to undergo data processing similar with
A. Data Analysis
this research which are mean, variance, skewness and
kurtosis. k-NN method employed with Euclidian distance Statistical method is applied to complete the analysis
has been applied in this study for prediction process. process computed with formula of mean, variance, skewness
Furthermore, predicted data undergo training and testing and kurtosis as equations below where N is the total number
ratio of 60%-40% with the help of 10-Fold and 5-Fold cross of hours. Total power consumption will determine the value
validation to evaluate the findings. Results for [10] of xi over ith hour of the day where i = 0, 1, 2,…., 23.
concluded that training and testing ratio used giving Highest values of collected maximum demand data for each
95.9615% accurate results. month are retrieved in the form of CSV file. Afterward, all
There are some other methods that have been applied in the data were imported into SPSS as this analytic software
order to do the analysis and energy prediction for different already provides all the mathematical formulas. The purpose
types of building done by previous studies but they were not of doing the analysis is to know the normality of the
included in this study since this paper only focuses on maximum demand data itself before proceed to the next
analysis and prediction method that undergoes analysis method.
using statistical method and k-NN.
Mean, µ = (1)
III. RESEARCH METHODOLOGY
Variance, σ = 2 (2)
This research starts with the study on the smart building
system using Internet of Things (IoT). The smart building in Skewness, S = 3 (3)
this case study is a shopping complex building located in Kurtosis, K = 4 (4)
Klang Valley, Malaysia. Figure 1 shows the system
architecture of the smart commercial building that
demonstrated the flow of data that have been measured at B. Prediction by using k-Nearest Neighbour (k-NN) Method
the smart building. The smart building equipped with IoT In this study, k-nearest neighbour (k-NN) method has
meters which are connected to the power inlet socket at two been chosen in predicting maximum demand. k-NN method
major tenants of the building. Each tenant is divided into is known as one of the machine learning algorithm as it is
two area consists of two IoT meters named tenant A1, A2, identified as a simple algorithm that able to do a
B1 and B2. Collected data per minute that mapped into classification and regression by storing all available cases
TNB requirements were saved in an open source web server. and categorizes new cases based on a similarity
measurement [10]. It is also known as lazy learning because December 2018 by using statistical method. Descriptive
of its training is held up to run time [14]. For classification, statistics analysis table is the output provided by SPSS
the output is determined by the majority of the value of its software that comprehends value of mean, variance,
neighbours while regression’s output is determined by the skewness and kurtosis. Figure 3 shows outcomes for
average of the k nearest neighbours. This method will be analysis process on highest value of maximum demand data
applied to predict the maximum demand based on the that has been completed for tenant A1, A2, B1 and B2.
electrical usage measured in kilowatt (kW) at both tenants. Mean values provided in the table indicates the central
Figure 2 shows the illustration in applying k-NN method in values of maximum demand data. Value of 156.427, 58.766,
predicting maximum demand of the building. 108.731 and 2.755 measured in kilowatt also indicates the
Features of maximum demand used for k-NN method peak position for graph distribution. As for variance, a large
are voltage and current. This is because the multiplication of value of variance indicates that the numbers in the data set is
these two features will result the value of power which is far from the mean and each other while small variance
resulting a maximum demand value. Voltage and current indicates the opposite. For tenant A1, A2, and B2 the
value are being selected from the CSV file that has been variance value calculated with 7.388, 7.805 and 1.366 shows
downloaded through the web browser. The numbers of that the data is closed to each other. Data for tenant B2 the
neighbour is setting to 3, 4 and 5 which means, the data is far from mean and each other due to the large value
classification of prediction will be included 3, 4 and 5 of variance with 953.578.
nearest neighbour of features to maximum demand. All the
analysis will be done in SPSS software.

Fig. 3. Descriptive statisctis for energy demand consumption for each

tenant

Fig. 2. The illustration of k-nearest neighbour method Skewness determines the symmetrical of the data
distribution while kurtosis defines the shape of the
The prediction process required historic data of the distribution curves for graph distribution. Standard error of
electrical consumption that has been collected at both skewness and kurtosis for each tenant in the table has been
tenants in a particular time to undergo prediction using multiplied with 3 and all the values are compared to the
nearest neighbour values which are 3, 4 and 5. The predicted absolute value of skewness and kurtosis in order to know the
data will be compared with the historical data for each normality of dataset of maximum demand consumption used
values of nearest neighbour used. This prediction method to undergo analysis process. If the values of absolute
used Euclidian distance function that is set as default in skewness and kurtosis are lesser than the calculated standard
SPSS software. error that has been multiplied earlier, which are known as
score values, it shows that the data of hourly consumption of
C. Performance Evaluation maximum demand used to undergo analysis using SPSS
software is normal.
Root mean square error (RMSE) is used to evaluate the
Figure 4 and 5 shows the skewness and kurtosis for all
value of maximum demand that have been predicted using
tenants. Bar graph of absolute values for skewness and
k-NN method. RMSE will calculates the difference between
kurtosis are lesser than bar graph of calculated score values.
the predicted data and historical data value of maximum
It can be concluded that all the data used for analysis is
demand consumption for each month of every tenant using
normal and can proceed to undergo prediction process.
equation provided below where n is the total number of
data, At is the historical values of maximum demand data
and Ft is the predicted values. The accuracy of prediction
value calculated by difference value of k was determined by
RMSE calculation.

Root mean square (RMSE) = (5)

IV. RESULTS AND DISCUSSION

This section consists of results analysis of maximum
demand consumption for each tenant from June 2018 until
 Fig. 4. Skewness graph for all tenants

Fig. 6. Average RMSE value for different k tested on all tenants

Next, value of k = 5 is applied on the maximum demand

value for each month of all tenants. Predicted values are
affected by the average of power factor, single phase current
and single L-N voltage. All these features are used to
complete the prediction process with the help of SPSS
software. Graphs in Figure 7 (a), 7 (b), 7 (c) and 7 (d)
illustrate and compare the outline of line graph between
historical and predicted maximum demand for tenant A1,
A2, B1 and B2. From the graphs shown below, it can be
Fig. 5. Kurtosis graph for all tenants concluded that predicted maximum demand are not much
different compare to actual value.
For prediction process that used k-nearest neighbour (k-
NN) method, historical hourly consumption of maximum
demand data for all tenants were computed into SPSS
software with values of k = 3, 4 and 5. Training and test
with the ratio of 70 to 30 was applied to avoid any
overestimation or underestimation of the predicted data.
Next, the most accurate value of k was selected by using
RMSE calculation. Value of k that gave prediction value
similar to the historical value of maximum demand data
provides least value of RMSE. The comparison of average
RMSE between values of k = 3, 4 and 5 for each tenant are
shown in Figure 6. RMSE value for tenant A1, B1 and B2 Fig. 7 (a). Comparison between actual and predicted maximum demand for
tenant A1
have the lowest average value of RMSE for k = 5 with 5.73,
8.54 and 0.35. It can be concluded that k = 5 is the most
accurate value used in this research compare to k = 3 and 4.
The k value obtained in this process is used as a reference to
predict maximum demand for upcoming period.

Fig. 7 (b). Comparison between actual and predicted maximum demand for
tenant A2
 formulas consists of mean, variance, skewness and kurtosis
in order to know the normality of collected dataset.
Prediction in this research is completed by using k-nearest
neighbour (k-NN) method with three different values of k.
Predicted data undergo training and testing data in the ratio
of 70 to 30 to avoid overestimating. Root mean square error
(RMSE) is applied on the predicted data to determine the
accuracy value of k. As a conclusion, value of k = 5 is the
most accurate among other values due to the least value of
RMSE calculated. Prediction on maximum demand for
Fig. 7 (c). Comparison between actual and predicted maximum demand for
tenant B1 upcoming period also had been done by using k = 5 as a
reference. To conclude, energy consumption analysis brings
convenience towards user and energy management team to
have a deep understanding on their energy behaviour.
Prediction creates awareness and helps them to enhance
their energy management system in order to reduce the
energy bills. For future study, predicted data of maximum
demand can be integrated into the available system for user
monitoring and controlling purposes in achieving energy
efficiency and sustainability.

Fig. 7 (d). Comparison between actual and predicted maximum demand for
tenant B2

Differences between prediction and actual values are

calculated to see the difference between actual and predicted
values when k-NN method with k = 5 is applied. Based on
Figure 8 below, tenant A1, A2, B1 and B2 show slightest
differences which are 0.43, 0.11, 5.45 and 0.15 respectively.
Hence, based on the results above it can be concluded that
the prediction method used for this research are able to
make improvement in maximum demand values for each
tenants.

Fig. 8. Differences value for all tenants

V. CONCLUSION
In this paper, a research is conducted based on real data
collection at the smart commercial building located in Klang
Valley, Malaysia. Analysis on maximum demand
consumption measured in kilowatt had been done
concerning at major tenants of the building. Statistical
method is proposed computed with several mathematical
 APPENDIX

Fig. 9. Hourly consumption of maximum demand graph for Tenant A1 of June 2018

Fig. 10. Hourly consumption of maximum demand graph for Tenant A2 of June 2018

REFERENCES
[1] “Malaysia Population (LIVE).” Malaysia Population (2018) – [8] F. Farahnakian, T. Pahikkala, P. Liljeberg, and J. Plosila, “Energy
Worldometers, 2018 [Online]. aware consolidation algorithm based on K-nearest neighbor regression
[2] J. S. Hassan, R. M. Zin, M. Z. A. Majid, S. Balubaid, and M. R. for cloud data centers,” Proc. - 2013 IEEE/ACM 6th Int. Conf. Util.
Hainin, “Building energy consumption in Malaysia: An overview,” J. Cloud Comput. UCC 2013.
Teknol., 2014. [9] F. Wahid and D. Kim, “A Prediction Approach for Demand Analysis
[3] P. Tracy and P. Tracy, "What is a smart building and how can it benefit of Energy Consumption Using K-Nearest Neighbor in Residential
you?", RCR Wireless News, 2019 [Online] Buildings,” Int. J. Smart Home, 2016.
[4] K. Akkaya, I. Guvenc, R. Aygun, N. Pala, and A. Kadri, “IoT-based [10] "A Quick Introduction to K-Nearest Neighbors Algorithm",
occupancy monitoring techniques for energy-efficient smart buildings,” Noteworthy - The Journal Blog, 2019 [Online].
2015 IEEE Wirel. Commun. Netw. Conf. Work. WCNCW, 2015 [11] C. Li, Z. Ding, D. Zhao, J. Yi, and G. Zhang, “Building energy
[5] A. R. Al-Ali, I. A. Zualkernan, M. Rashid, R. Gupta, and M. Alikarar, consumption prediction: An extreme deep learning approach,”
“A smart home energy management system using IoT and big data Energies, 2017.
analytics approach,” IEEE Trans. Consum. Electron., 2017. [12] R. Zhou, Y. Pan, Z. Huang, and Q. Wang, “Building Energy Use
[6] V. V. Afanas’ev, V. G. Kovalev, V. V. Tarasova, and V. A. Prediction Using Time Series Analysis,” IEEE 6th Int. Conf.
Tarasov,“The use of statistical methods to analyze electric-energy Serv. Comput. Appl, 2013.
consumption by budget institutions,” Russ. Electr. Eng., 2017. [13] N. A. Ramli, M.F.A. Hamid, N. H. Azhan, and M. A. A. Ishak,
[7] M. N. . Salleh, A. Hariri, M. Z. . Yusof, A. . Leman, M. Faris Mubin, “Solar Power Generation Prediction by using k- Nearest Neighbor
and K. . Rahman, “Energy Consumption Analysis Based on Energy Method,” unpublished.
Efficiency Approach: A Case of Suburban Area,” 2016.

View publication stats