Innovative Computing Review (ICR)

Volume 2 Issue 1, Spring 2022

ISSN(P): 2791-0024 ISSN(E): 2791-0032
Homepage: https://journals.umt.edu.pk/index.php/UMT-AIR

Article QR

Title: Energy Prediction of Appliances using Supervised ML Algorithms

Author (s): Robbia Gulnar Robbia1

Affiliation (s): Department of Computer Science, Government College of Faisalabad (GCUF),

Punjab, Pakistan.
DOI: https://doi.org/10.32350.icr.21.05

History: Received: April 8, 2022, Revised: May 22, 2022, Accepted: June 14, 2022

R. Gulnar and H. Tariq, “Energy prediction of appliances using supervised ML

Citation: algorithms,” UMT Artif. Intell. Rev., vol. 2, no. 1, pp. 00-00, 2022, doi:
https://doi.org/10.32350.icr.21.05

Copyright: © The Authors

Licensing: This article is open access and is distributed under the terms of
Creative Commons Attribution 4.0 International License
Conflict of
Interest: Author(s) declared no conflict of interest

A publication of
School of Systems and Technology
University of Management and Technology, Lahore, Pakistan
 “Energy Prediction of Home Appliances using Supervised
Machine Learning (ML) Algorithms”
Robbia Gulnar Robbia * 0F

Department of Computer Science, Government College of Faisalabad (GCUF),

Punjab, Pakistan.

Abstract-The amount of energy I.Introduction

consumed by domestic appliances is
It is important to predict the
an important area of research. amount of energy consumed by
Hence, the main goal of this study is household electrical appliances,
to produce very precise forecasts since improper appliance use wastes
about energy consumption by home energy in the residential sector.
appliances using the least amount of
Hence, an accurate assessment
processing power. The algorithms
of energy demand in the housing
used in this study for predicting
sector is crucial in order to
energy usage/consumption included
determine the amount of energy that
regression, K-nearest neighbor, may be saved. The amount of
decision trees, and random forest. energy saved mostly depends on the
These algorithms were applied on the type of device being used; for
appliances’ energy prediction example, some devices may cause
dataset made available for public use an imbalance state, others may
at the UCI Machine Learning operate more slowly, and some have
Repository. To compare the data sets fixed running times. The exact
and choose the optimal machine energy forecasting model, from the
learning (ML) algorithm for them, perspective of energy providers,
root mean square error (RMSE) was may assist in determining the ideal
computed. time to employ various devices to
lower total carbon emissions and
Index Terms- energy consumption, also to save money. By making
prediction energy utility, root mean
good use of consumer business
square error (RMSE), supervised
machine learning.
models, agents may arrange the
functioning of various gadgets. The
energy forecasting solutions also
provide consumers with an in-depth
analysis of home energy
*
Corresponding author: robiagulnar@gmail.com
Innovative Computing Review
74
Volume 2 Issue 1, Spring 2022
 Gulnar

consumption patterns, allowing select the optimum regression

them to better manage and control model for a particular dataset [2].
their energy use and energy
II.Literature Review
expenses. Reactive power
management has been the subject of Load estimation and forecasting
extensive research from the are key factors when it comes to
viewpoint of an industry consumer, efficiently distribute power and
although little research has been keeping reserves for the future. In
conducted about it from the order to meet the electricity demand
standpoint of a household and not to resort to load shedding,
consumer. The analysis of elements of load forecasting
residential energy use and customer mechanisms should be integrated
behavior provides valuable insights with the techniques utilized by the
that may help to develop more dominant power utility companies.
efficient energy consumption There are many ML techniques that
tactics. It is a difficult optimization can be utilized to achieve the desired
problem to plan the operations of results. So, an in-depth analysis was
home appliances in various smart carried out using multiple ML
homes, since it is fundamentally a algorithms in this paper to identify
complicated nonlinear the best technique. The techniques
combinatorial issue [1]. used to differentiate between all
these algorithms and to list their
Models that predict the energy respective advantages and
consumption of home appliances disadvantages included mean
have been the subject of several absolute error (MAE), root mean
researches. The decision tree (DT) square error (RMSE), and mean
method, the decision table classifier absolute percentage error (MAPE).
(DTC), and the Bayesian network The most efficient way determined
(BN) are a few examples of was to not only use one of these
machine learning (ML) approaches algorithms independently but to
used to provide a model for combine them in various
predicting the next-hour and next combinations and use these
24-hour energy consumption of combinations instead. This
home appliances. These methods approach tentatively yields the most
codify expert information on energy accurate results. Indeed, hybrid
usage and provide a suitable data algorithms are the ones that work
structure for the regressor. They the best with the understanding that
also show how challenging it is to all algorithms have their specific

School of System and Technology

75
Volume 2 Issue 1, Spring 2022
Energy Prediction of Home…

pros and cons. Load estimation learning along with distributed

requires very frequent checks on demand response programs. In the
different load types. So, different current paper, the techniques
technologies have been discussed discussed for this purpose include
with respect to different load logistic regression (LR), support
horizons. As far as performance and vector machine (SVM), naïve Bayes
accuracy are concerned, predictive (NB), decision tree classifier
models that are a combination of (DTC), K-nearest neighbor (KNN),
more than two existing models have and neural networks (NNs). The
been proven to be the most fruitful. point is to propose a model well
Support vector machine (SVM), suited for the estimation and
artificial neural network (ANN), prediction of short term load
and other relevant models have forecasting (STLF). After extensive
achieved a well-organized power research and analysis, DTC was
system utility along with the identified as the best technique.
minimum percentage of error. The Enhanced DTC (EDTC) was
authors of the current study proposed that utilizes integrated
conducted numerous tests to filter function, loss function, and
determine which techniques go well gradient boosting to fine tune the
together in a hybrid model and yield already existing DTC mathematical
the most accurate results [3]. model. The resulting EDTC
algorithm yields a better forecasting
World population is increasing
result.
day by day and with it the overall
electricity consumption is also The forecasting and stabilizing
increasing. In the current situation, of smart grid (SG) remains a
where supply struggles to meet challenge in today’s landscape. ML
demand, the best course of action is comes into play when we discuss the
to implement those techniques resilience of the SGs. Indeed, there
which may help to predict the are different ML algorithms that
overall electricity consumption at have the capability to predict the
any given time in order to take future energy needs. Amongst all
necessary measures beforehand to the suitable algorithms, selecting
meet the demand. The energy and adopting the best one poses a
consumption trends throughout are challenge. Numerous tests were
nonlinear and remain dynamic. To conducted to select the best model
predict short-term and long-term and DTC outperformed all the other
consumption with high accuracy, it algorithms. EDTC was established
is imperative to use machine to further amplify the predicting
Innovative Computing Review
76
Volume 2 Issue 1, Spring 2022
 Gulnar

capabilities of DTC. EDTC was LabVIEW making a new VI. In the

found to be superior to SVM, KNN, model, blocks containing any SB
NN, LR, and DT, when it came to appliance could be selected any time
accuracy, precision loss, and ROC of the day. The model functioned on
curve metrics [2]. the basis of ANN. The algorithm
was not very accurate as the dataset
It is crucial to predict and
fed to the model was not big enough
schedule the energy needs in smart
for it to make proper computations.
buildings (SBs) and to meet them
Hence, training and validation
accordingly in order to deploy
remained a challenge. It was
energy-efficient management
determined that ANN is not the best
systems In the current paper, several
model when it comes to prediction
approaches were explored and
[4].
amongst them artificial neural
network (ANN) and genetic In this day and age, STLF is
algorithms remained in focus. To needed to fulfil the demand of
get the most accurate result, ANN power consumption. This helps to
was implemented in a real SB predict the pattern on which the
testbed. The tests were conducted power system operates. The
using poly-voltaic panel installation methods of load forecasting
and SB electronic appliances and employed in non-residential areas
the data was collected from them. are based on customer demand or on
To implement ANN, the authors experience and historical data. The
used CompactRIO and their model best way to make the best prediction
exhibited subpar prediction is to use machine learning. In the
accuracy [4]. current paper, the authors sought the
In the proposed paper, authors best ML model to generate an
accurate short term algorithm for
built a model for accurate prediction
non-residential areas. For this
of the energy consumption and also
purpose, the authors conducted
worked on the scheduling process.
experiments and rooted out the best
The proposed model utilized
model for industrial users.
machine learning. The research was
Recurrent neural network based on
intended as a roadmap towards a
sliding window approach turned out
better model which can achieve
to be the best model for both short-
accuracy greater to that of this
term and long-term prediction. After
model.
three months of testing, the model
The proposed model was that gave the best results was gated
implemented using python in recurrent unit (GRU), with long-
School of System and Technology
77
Volume 2 Issue 1, Spring 2022
Energy Prediction of Home…

short term memory (LSTM) as the loads, respectively. Research was

second best. GRU minimized conducted on multiple ML
5326.17 euros compared to LSTM algorithms and again the algorithm
in these three months and resulted in that yielded the best result was RF.
5.28% MAPE. The proposed model The error rate determined from the
was made to justify and evaluate the testing turned out to be 0.404 and
gap between evaluation matrices 1.128 for heating and cooling loads,
and the impact of forecast errors in respectively. Contributing factors
power market. The implementation were determined to make the best
involved three-month data of prediction using univariate
different ML algorithms. GRU algorithm. Height was determined
turned out to be the best and the as the most notable feature that
authors considered the data as contributed towards the prediction
sufficient [5]. of the overall power consumption
[6].
Considering the overall increase
in population and the depletion of Home energy management
energy resources around the world, systems (HEMS) can be further
we need to utilize our energy enhanced by the use of load
resources efficiently and develop a forecasting. This can be achieved by
model to accurately predict the utilizing machine learning
overall energy consumption considering the increase in the
according to the given factors. To relevant data in the recent years. The
figure out the best factors utilized in current authors propose two
the prediction process, univariate methods for load forecasting,
regression algorithm was employed enhancing the traditional long short-
by the authors. The algorithm term memory (S2S-LSTM) model.
predicted that the factors with the In the first method, three algorithms
most impact were overall height, are applied: density based spatial
roof area, surface, and relative clustering of applications with noise
compaction. The models tested for (DBSCAN), K-means, and Pearson
the given factors were DT, RF, and correlation coefficient (PCC).
K-NN. The testing of these Amongst all these techniques, PCC
algorithms was conducted on proved itself to be the best one. PCC
Orange software. After extensive was better at accommodating a large
testing, the algorithm that gave the number of consumers. The second
most accurate prediction was RF. method constitutes an extension to
The forecasting error was 1.128 and method one and increases its
0.404 for cooling load and heating performance. It utilizes NN
Innovative Computing Review
78
Volume 2 Issue 1, Spring 2022
 Gulnar

architecture with softmax layers sources of the energy which in most

which are fully-connected, dropout, cases are hydro-electric power
and stable. In LSTM, it further plants, petroleum products, nuclear
optimizes supervised learning and wind energy plants. Under and
which results in a more stable and overproduction are also the causes
accurate model for prediction. The of the fluctuation in price, but the
findings were reached by ones that contribute the most
conducting an 8-week long research towards that fluctuation are
with 2337 consumers. metrological parameters, economics
and industrial activities. That’s why
In the current paper, two
load estimation needs to performed
methods were proposed to make
on a regional level which will help
accurate predictions when it comes
in efficiently manage, scheduling
to energy consumption. These
and planning. All of this would
methods enhance S2S-LTSM model
result in overall low cost. In
to make the predictions. Method one
machine learning there are multiple
uses an amalgamation of human
algorithms that can be used to
pattern recognitions which is
accurately measure and estimate the
extracted by three cluster analysis
overall energy requirement. There
algorithm. Amongst all the
are different algorithms that were
algorithms that were employed PCC
used in the proposed paper. The
proved itself to be the best. The idea
different supervised learning
behind method one is to make a
algorithms were linear regression
weight matrix by energy utilization
(LR), support vector repressor’s
habits and calculate the distance of
the cluster. Three layer optimizing (SVR), K-nearest neighbor (KNN),
architecture was used to further random forest (RF), and AdaBoost.
enhance method one [7]. The performance that was
associated with all these algorithms
Electricity is a facility that is varied with different times and data.
utilized daily by almost everyone To minimize the price per unit we
and the lack of this energy will lead utilize machine learning with
major disasters. We should generate correlated metrological parameters
only the required amount of kept in consideration. The model
electricity for utilization and cannot that was the focus of the proposed
make more that required because paper was least cost electric load
utilizing large sums of energy is not forecasting model (lcELFM). The
possible. The price that is associated model was implemented by
with the energy depends on the minimizing root mean square error
School of System and Technology
79
Volume 2 Issue 1, Spring 2022
Energy Prediction of Home…

(RMSE), mean absolute error The proposed model generated

(MAE), and mean absolute forecasting reduced prediction error
percentage error (MAPE). The data for ELnMPFModels. There was a
for testing was taken from significant reduction in MAPE and
Muzaffarabad from start of January with the implementation of the
2014 till end of December 2015. proposed model Muzaffarabad will
Pakistan meteorological department save 0.303 million rupees daily.
provided author with Although this study was conducted
meteorological time series data for keeping Muzaffarabad in mind but
the same period and time. The the same model can be implemented
proposed model turned out to be the in any city of Pakistan to get
best when compared to other accurate estimation on load [4].
models. Estimation of the power
Human energy consumption consumption is a very important
patterns change consistently with tasks when we are supposed to
changes in habits that develop due to generated energy in advance or plan
the change in weather. This energy for the generation of power
is generated using sources like beforehand. With the
water, petroleum, wind, and natural implementation of smart grid, the
energy. Consumers want the price to need of energy consumption
be less whereas the providers want estimation is dire. Estimation of a
the profits to be maximum [8]. future event is always a difficult task
and to do it with high precession is
AELF model was developed
an even bigger challenge. There
taking in account the weather
have been many attempts at
conditions and human behaviors.
This model is also used to reduce the estimation of power consumption
price of electricity. The proposed accurately but none of them very
accurate. Machine learning has
paper suggests a least cost
multiple algorithm that are well
estimation model and utilized
suited for this task. Machine
meteorological parameters driven
learning has been recognized to
electrical load demand of
predict failure before it even occurs.
Muzaffarabad. The study suggested
Machine learning is artificial
that meteorological parameters like
intelligence (AI) which develop a
temperature influence the overall
model based studying a given data.
consumption. This proved that
The algorithms that were explored
factors like time and season
were artificial neural network
drastically impact the consumption.
(ANN), multiple linear regression
Innovative Computing Review
80
Volume 2 Issue 1, Spring 2022
 Gulnar

(MLR), adaptive neuro fuzzy executed, included decision trees

interface system (ANFIS), and (DT), support vector machine
support vector machine (SVM). The (SVM), logistics regression, linear
criteria selected in proposed paper regression, K-nearest neighbor
for power generation is Cyprus. (KNN), and random forest (RF).
Testing was done on real data SVM is a type of generalized linear
accumulated over 2016 and 2017 for classifier that uses supervised
the motivation of using in long term learning to classify data into binary
and short term analysis. It was categories. SVM was first
determined that the factors that presented/introduced in 1964 and it
affect the consumption of electricity grew in popularity during the 1990s,
the most are temperature, humidity, resulting in a number of enhanced
solar irradiation, population, gross and expanded algorithms.
national income (GNI) per capita, Regression issues may also be
and the electricity price per solved with SVM. KNN may be
kilowatt-hour. By doing multiple used to study regression by getting a
computations it was later discovered sample's nearest neighbors and
that SVM and ANN were superior to assigning the average of these
other machine learning algorithms neighbors’ properties to the sample
which had fewer estimation error in order to obtain the sample’s
[9]. properties. Another (improved) way
is to assign different weights to the
With smart grid implementation
effects of neighbors at various
load estimation is more important
distances from the sample, with the
than ever. The prediction of load in
weight being inversely proportional
any given time is difficult
to the distance. Among ML
considering the dynamic nature of
techniques, it was found that RF is
the consumer. For prior planning of
faster in the training process and
the energy consumption it is crucial
powerful for/more effective in
to estimate it beforehand. In this
solving high dimensional data and
study multiple methods were tested
complex problems in the industry.
where ANN and SVM were more
Its performance remains stable and
accurate and provided better
accurate, due to which it creates
estimation of the energy required
multiple decision trees and
[10].
combines them to produce output.
III.Research Methods Figure 1 summarizes the proposed
ML algorithms, on which the methodology.
data set was implemented and
School of System and Technology
81
Volume 2 Issue 1, Spring 2022
Energy Prediction of Home…

IV.Experimental Setups A. Data Set

and Results Data set consisted of
The experimental setup included temperature (measured in Celsius)
the tools needed to identify the best and humidity (measured in
ML algorithm for making prediction percentage). The data was collected
and doing anticipatory tasks.

Fig. 1. Proposed method

by installing temperature sensor and about the outside weather, pressure,
humidity sensor in all the rooms of humidity, wind speed, and visibility
a low energy house. The values on were taken from the weather station.
the temperature sensor and humidity This data was linked with the
sensor were monitored with the help experimental data of temperature
of ZigBee wireless sensor network. and humidity of each room with the
Each sensor transmitted the data help of date and time columns. The
(condition of temperature and source of energy prediction for
humidity) for 3.3 mins, which was appliances data set is UCI Machine
averaged for a 10 mins period. The Learning Repository. Table 1
data was monitored for every 10 summarizes the characteristics of
mins and logged in. The details the dataset.
Innovative Computing Review
82
Volume 2 Issue 1, Spring 2022
 Gulnar

From a total of 19,735 data Table 3 shows the performance

samples, 15,788 samples (80%) evaluation of the prediction models
were randomly assigned to the used to measure the energy
training set and 3,947 samples consumption of household
(20%) to the testing set. Table 2 appliances. RMSE of linear
gives the data set description. regression is 0.046 (which is very
low). It shows the better
Table I
performance of this model as
Characteristics of Dataset compared to other models. RF has
Characteristic Values RMSE of 0.047 and it also shows
better performance than SVM and
No. of data logistic regression.
19,735
sample Table III
No. of features 29 Model Performance
Models Accuracy RMSE
Sampling
10 min
time/rate
Linear
B. Results Regression 0.953
0.046
All prediction models are
regression models. Root mean Random
square error (RMSE), mean 0.939
Forest 0.047
absolute error (MAE), mean square
error (MSE), and decision Logistic
coefficient (R2) are all regularly 0.732
Regression 11.71
used metrics for assessing
regression models. In this SVM
experiment, the RMSE assessment 0.740 10.37
indicator was employed. It was used
to calculate the difference between
the observed value and true value.
RMS=�1/𝑚𝑚 ∑𝑚𝑚 ′
𝑖𝑖=1(𝑌𝑌𝑌𝑌 − 𝑌𝑌𝑖𝑖 )

Here, Yi is the real value of the

data at time i and Yi’ is the predicted
value of the data at time i.

School of System and Technology

83
Volume 2 Issue 1, Spring 2022
Energy Prediction of Home…

Table II
Dataset Variables and Description

Feature # Variable Description Unit

1 T1 Temperature in kitchen °C
2 T2 Temperature in living room °C
3 T3 Temperature in laundry room °C
4 T4 Temperature in office room °C
5 T5 Temperature in bathroom °C
6 T6 Temperature outside the building °C
7 T7 Temperature in ironing room °C
8 T8 Temperature in teenager room °C
9 T9 Temperature in parents’ room °C
10 R1 Humidity in kitchen %
11 R2 Humidity in living room %
12 R3 Humidity in laundry room %
13 R4 Humidity in office room %
14 R5 Humidity in bathroom %
15 R6 Humidity outside the building %
16 R7 Humidity in ironing room %
17 R8 Humidity in teenager room %
18 R9 Humidity in parents’ room %
19 L Light energy consumption Wh
20 RO Humidity outside the airport %
21 Td Dew point temperature °C
22 V Visibility km
23 W Wind speed m/s
24 TO Temperature outside the airport °C
25 rv1 Random Variable1 \
26 rv2 Random Variable2 \
27 P Pressure mmHG

Innovative Computing Review

84
Volume 2 Issue 1, Spring 2022
 Gulnar

Fig 2. Accuracy of Models

Fig 3. Root Mean Square Error of Models

School of System and Technology

85
Volume 2 Issue 1, Spring 2022
Energy Prediction of Home…

Fig 4. Energy consumption by home appliances in the given period

Fig 5. Energy Consumption of Lights

Innovative Computing Review

86
Volume 2 Issue 1, Spring 2022
 Gulnar

Fig 6. Pressure and Humidity during Energy Consumption

Fig 7. Graph Depicts Visibility (Scatter Plot)

School of System and Technology

87
Volume 2 Issue 1, Spring 2022
Energy Prediction of Home…

V. Conclusion https://doi.org/10.1016/j.enbuil
Prediction models for energy d.2017.03.064
usage by home appliances based on
SVM, KNN, RF, linear regression, [2] T. Alquthami, M. Zulfiqar, M.
and logistic regression were Kamran, A. H. Milyani, and M.
investigated. Firstly, the authors B. Rasheed, “A performance
reviewed the data pretreatment to comparison of machine learning
remove certain features from the algorithms for load forecasting
filtered data and to normalize it. in smart grid,” IEEE Access,
Secondly, the grid search technique vol. 10, pp. 48419-48433, 2022,
was utilized to find the best doi:
parameters for the model and https://doi.org/10.1109/ACCES
models based on several ML
S.2022.3171270
algorithms were created. Finally,
each model's prediction
performance was tested and [3] A. A. Mamun, M. Sohel, N.
compared. The results revealed that Mohammad, M. S. Haque
linear regression and RF obtained Sunny, D. R. Dipta, and E.
good results in both the training and Hossain, “A comprehensive
testing data sets, with the best review of the load forecasting
prediction performance among the techniques using single and
four prediction models developed hybrid predictive models,”
using the classic ML approach. In IEEE Access, vol. 8, pp.
the testing set, KNN, RF, and SVM 134911-134939, 2020, doi:
all performed equally well, https://doi.org/10.1109/ACCES
however, SVM performed the S.2020.3010702
poorest in the training data set.
References [4] J. Zhua, Y. Lin, W. Lei, Y. Liu,
and M. Tao, "Optimal
[1] S. Hu, D. Yan, S. Guo, Y. Cui, household appliances
and B. Dong, “A survey on scheduling of multiple smart
energy consumption and energy homes using an improved
usage behavior of households cooperative algorithm," Energy,
and residential building in urban vol. 171, pp. 944-955, Mar.
China,” Energy Build., vol. 148, 2019, doi:
pp. 366–378, Aug. 2017, doi:

Innovative Computing Review

88
Volume 2 Issue 1, Spring 2022
 Gulnar

https://doi.org/10.1016/j.energy [8] R. Olu-Ajayi, H. Alaka, I.

.2019.01.025 Sulaimon, F. Sunmola, and S.
Ajayi, "Machine learning for
[5] S. Ungureanu, V. Topa, and A. energy performance prediction
C. Cziker, "Analysis for non- at the design stage of buildings.
residential short-term load Energy Sustain. Dev., vol. 66,
forecasting using machine pp. 12-25, Feb. 2022, doi:
learning and statistical methods https://doi.org/10.1016/j.esd.20
with financial impact on the 21.11.002
power market. Energies, vol. 14,
no. 21, Art. no. 6966, doi: [9] D. Solyali, "A comparative
https://doi.org/10.3390/en14216 analysis of machine learning
966 approaches for short-/long-term
electricity load forecasting in
[6] S. Hosseini and R. H. Fard, Cyprus. Sustainability, vol. 12,
"Machine learning algorithms no. 9, Art. no. 3612, 2020, doi:
for predicting electricity
https://doi.org/10.3390/su12093
consumption of buildings,"
612
Wirel. Pers. Commun., vol. 121,
no. 4, pp. 3329-3341, Aug.
[10] K. Thiyagarajan and R. S.
2021, doi:
Kumar, "Real time energy
https://doi.org/10.1007/s11277-
021-08879-1 management and load
forecasting in smart grid using
[7] L. Fan, J. Li, and X.-P. Zhang, compact RIO," Procedia
"Machine learning load Comput. Sci., vol. 85, pp. 656–
prediction methods for home 66, 2016, doi:
energy management systems https://doi.org/10.1016/j.procs.2
based on human behavior 016.05.250
patterns recognition," CSEE J.
Power Energy Syst., vol. 6, no.
3, Sep. 2020, doi:
https://doi.org/10.17775/CSEEJ
PES.2018.01130

School of System and Technology

89
Volume 2 Issue 1, Spring 2022