Internet of Things 20 (2022) 100636

Contents lists available at ScienceDirect

Internet of Things
journal homepage: www.sciencedirect.com/journal/internet-of-things

Research article

A machine-learning ensemble model for predicting energy

consumption in smart homes
Ishaani Priyadarshini a, Sandipan Sahu b, Raghvendra Kumar c, David Taniar d, *
a
School of Information, University of California, Berkeley, USA
b
Department of Computer Science and Engineering, Bengal Institute of Technology, India
c
Department of Computer Science and Engineering, GIET University, India
d
Faculty of Information Technology, Monash University, Australia

A R T I C L E I N F O A B S T R A C T

Keywords: Smart homes incorporate several devices that automate tasks and make our lives easy. These
Smart Home devices can be useful for many things, like security access, lighting, temperature, etc. Using the
Internet of Things Internet of Things (IoT) platform, smart homes essentially let homeowners control appliances and
Decision tree
devices remotely. Due to their self-learning skills, smart homes can learn homeowners’ schedules
Random Forest
eXtreme gradient boosting
and adapt accordingly to make adjustments. Since convenience and cost savings is necessary in
Ensemble model such an environment, and there are multiple devices involved, there is a need to analyze power
Machine learning consumption in smart homes. Moreover, increased energy consumption leads to an increase in
carbon footprint, elevates the risk of climate, and leads to increased demand in supply. Hence,
monitoring energy consumption is crucial. In this paper, we perform an overall analysis of energy
consumption in smart homes by deploying machine learning models. We rely on machine
learning techniques, like Decision Trees (DT), Random Forest (RF), eXtreme Gradient Boosting
(XGBoost), and k-Nearest Neighbor (KNN) for predicting the power consumption of multiple
datasets. We also propose a DT-RF-XGBoost-based Ensemble Model for analyzing the consump­
tion and comparing it with the baseline algorithms. The evaluation parameters used in the study
are Mean Square Error (MSE), R-squared (R2,), Root Mean Square Error (RMSE), and Mean Ab­
solute Error (MAE), respectively. The study has been performed on multiple datasets and our
study shows that the proposed DT-RF-XG-based Ensemble Model outperforms all the other
baseline algorithms for multiple datasets with R2 around 0.99.

1. Introduction

One of the best applications of Artificial Intelligence (AI) in modern times is in the form of smart homes. Smart homes are residences
that rely on internet-connected devices for remote monitoring and device and appliance management [1]. It is also referred to as home
automation or domotics and is concerned with providing comfort, security, energy efficiency, and convenience. Smart devices are
usually controlled by a smart home application or a networked device. Smart homes are built on the Internet of Things (IoT) platform
and incorporate sensors, speakers, smart bulbs, cameras, locks, door openers, etc. [2]. Due to their self-learning skills, they are capable
of learning the homeowners’ schedules and can adjust accordingly. Additionally, these devices may operate together and share

* Corresponding author.
E-mail address: [email protected] (D. Taniar).

https://doi.org/10.1016/j.iot.2022.100636

Available online 7 November 2022

2542-6605/© 2022 Elsevier B.V. All rights reserved.
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

consumer usage data due to automation based on homeowners’ preferences [3]. Since these devices are power-driven, they can reduce
power consumption and lead to energy-related cost savings [4].
While smart home devices can save a considerable amount of energy, the efficiency can be improvised even further. Many smart
speakers and connected cameras consume more power since they add more energy load. Since power consumption relies on many
predictable factors like what devices were used before, what devices are being used presently, which product is bought, and how it is
used, a hundred percent power saving cannot be guaranteed. The methods to address the issue are being actively researched.
Another major reason that leads to inefficient power consumption is based on poorly constructed buildings. For constructions with
a single-pane window with no insulation, deploying a smart thermometer may not be beneficial. In other words, if the building is not
designed to save energy, integrating applications and components may be a tedious task. This may lead to greater power consumption.
Moreover, as more and more smart homes house intelligent lighting controls, although it may use very little electricity, the fact that it
is smart and always connected may lead to more consumption of electricity in general [5]. Hence, there is a need to inspect power
consumption for smart homes.
Monitoring power consumption is also necessary for load-balancing power plants. Performing load study is an important aspect of
energy monitoring. As there is an increase in energy consumption, it adds to the risk of climate change and increases carbon footprints
[6]. The increasing energy costs, in turn, lead to an increase in the demand for energy consumption. Owing to all these factors, there is
a need to monitor energy consumption. One of the ways of inspecting power consumption is by taking a look at the prediction data. To
reduce power consumption in smart homes, it is necessary to observe prediction trends. In the past several methods have been pro­
posed to monitor power consumption. Some of these methods are in-chip configurations in microprocessing systems, digital power
meters, energy auditing tools, delay and power monitoring schemes [7], etc. Moreover, machine learning methods have been applied
in abundance to monitor power consumption. Linear Regression [8], Support Vector Machines [9], and Long-short Term Memory [10,
11], etc., are some of the popular machine-learning models that have been relied on in the past for analyzing power consumption.
While traditional machine learning models yield satisfactory results for addressing the problem, the models often have limitations.
Moreover, overfitting and cost are prevalent in most traditional machine learning algorithms. To address limitations like these,
ensemble methods are deployed. Ensemble methods combine many machine learning algorithms to produce one optimal predictive
model, thereby enhancing the model’s performance and robustness.
The novelty and main contributions of the paper are as follows:

1 We propose an ensemble-based technique for predicting power consumption in smart homes. The ensemble proposed is a com­
bination of Decision Trees (DT), Random Forests (RF), and eXtreme Gradient Boosting (XGBoost).
2 We compare the performance of our proposed ensemble method with several other baseline models such as K Nearest Neighbors
(KNN), Decision Trees (DT), Random Forests (RF), and Gradient Boosting (GB). Deploying the ensemble model has two advantages.
First, it improves the prediction performance over the other contributing components of the ensemble. Second, it reduces the
variance of prediction errors induced by components of the ensemble, thereby addressing any kind of overfitting.
3 The performance of the models is evaluated using multiple statistical parameters such as Mean Square Error (MSE), R-squared Error
(R2), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE).
4 Extensive analysis has been performed on two different datasets that incorporate readings with a time span of 1 minute of house
applications in kiloWatts from smart meters. To the best of our knowledge, this is the first paper highlighting power consumption in
smart homes using the DT-RF-GB-based ensemble approach.

The rest of the paper has been organized as follows. Section 2 presents some relevant related works in the area. Section 3 denotes
the machine learning algorithms deployed for the ensemble method proposed. In Section 4, we detail the experimental analysis,
including the datasets and evaluation parameters. This section also includes the results obtained from the extensive study, along with a
comparative analysis. Section 5 discusses conclusions and future works.

2. Related works

In this section, we present a detailed survey of past related works. Since energy management is a global issue, the area has witnessed
extensive research over the last few decades. The progress in technology has led to several methods being proposed for energy
monitoring as well as management. We highlight some proposed methods in this section and propose an ensemble-based approach in
the next section.
Ref. [12] proposed a Multi-output Adaptive neuro-fuzzy inference system (MANFIS) based smart home energy management system
for efficient energy management in smart grids. The proposed design aims to reduce electricity costs and reverse power flow. The
system is tested with daily data concerning temperature, wind speed, isolation, and controllable and uncontrollable appliances power
as input. The output is used for handling energy production and consumption. Results show that electricity cost and power con­
sumption are reduced significantly. Ref. [13] suggested another energy-saving method by integrating big data and machine learning
with smart homes. The study deploys the J48 machine learning algorithm, along with WEKA API, for learning user behavior, in terms
of energy consumption patterns and thus classifies houses based on energy consumption. RuleML and Apache Mahout have generated
recommendations based on user preferences. A case study has been incorporated to manifest reduced energy consumption. Ref. [14]
recommended a hybrid robust-stochastic optimization technique for energy management in smart homes. The study conducted for
“day ahead” and real-time energy markets incorporate a robust optimization approach for managing the day-ahead market prices. The
proposed optimization framework takes into account the real-time energy market, as well as the associated uncertainties, by relying on

2
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

stochastic programming. The study estimates a profit of day-ahead and real-time markets to be $2.5/day. Ref. [15] proposed a
Bio-Inspired Dragonfly Algorithm and Genetic Algorithm for optimizing energy consumption. The study considers two classes of
appliances, i.e., Shiftable appliances and Non-shiftable appliances. Simulation results show that the proposed algorithm can minimize
electricity costs and a tolerable waiting time. The only drawback of the proposed system is the waiting time. This is because as
electricity cost decreases, there is an increase in the waiting time and vice versa. The proposed method also ensures the stability of the
grid since the stability of the grid is dependent on the peak-to-average ratio.
Ref. [16] suggested a distributed artificial bee colony for connected appliances for efficient energy consumption in smart homes. In
this study, swarm intelligence has been applied to connected devices. Overall decentralized management leads to sharing of infor­
mation so that individual decisions can be taken, thereby optimizing electricity prices. The proposed approach has been evaluated in a
smart home-connected environment, and the simulation manifests load-balancing optimization [17]. recommended a combined Deep
learning-IoT-based platform for effective energy management in smart buildings. The YOLO v3 algorithm is used to detect people and
count the number of people. Likewise, it is possible to manage the operation of air conditioners in a building. Hence the number of
people and status of air conditioners are published on the IoT platform, and decision-making is performed based on energy con­
sumption. Intensive test scenarios support the validation of the study. The proposed model shows appreciable accuracy. [18] presented
an artificial bee colony based on non-intrusive appliance monitoring for smart homes. The study, which considers a group of connected
consumer electronics loads, has been carried out in two parts. First, data for individual appliances are collected and stored with varying
loads. Second, the stored data is used to estimate the individual load current. Simulations in a practical household system have
validated search-based optimization.
Ref. [19] proposed multi-objective energy management due to uncertainty in wind power forecasting. The electricity cost is
formulated for achieving the best schedule of devices in smart homes. The study employs a multi-objective dragonfly algorithm for
optimizing the technical and economic objective functions. Once the optimal Pareto front is deduced, the study relies on an analytical
hierarchy process for selecting the best operational schedule for smart homes. The suggested approach is evaluated in a sample smart
grid, and numerical results validate that the proposed management method efficiently improves the performance of the smart grid.
[20] suggested a fuzzy logic-based approach for optimal household appliance scheduling. The proposed method considers electricity
price and load consumption. The daily electricity usage is predicted by a predictive model and Demand Response (DR) scheme. After
deploying a Long Short-Term Memory-based (LSTM) optimized predictive model, data is transmitted to a DR fuzzy logic-based
controller. The LSTM model outperforms other baseline models and reduces electricity costs significantly. [21] recommended a

Table 1
Summary of the existing research works.
Research Methodology Strength Weakness

Smart Home Energy MANFIS Significant reduction in cost of Curse of dimensionality, computational
Management System electricity expense.
[12]
Smart Home Energy saving Big data and machine learning Reduces energy consumption Platform specific, compatibility issues
system [13]
Energy Management [14] Hybrid robust-stochastic optimization model Profitable energy management Method is brittle, sensitive to change in
parameters,Time consuming
Home Energy Management Bio-Inspired Dragonfly Algorithm and the Significant decrease in Increase in wait time, user discomfort, does
Optimization [15] Genetic Algorithm electricity cost not work well for all scenarios
Smart Home Energy Distributed artificial bee colony algorithm Optimized performance of Requires new fitness tests on new
Management System energy management system parameters, slow
[16]
Effective energy Deep Learning and IoT based approach Enhanced decision making Segregating small objects in a group setting
management for smart (YOLOv3) about energy consumption is challenging
buildings [17]
Enhanced Load Monitoring Artificial bee colony algorithm Efficiant Load Monitoring Requires new fitness tests on new
in Smart Homes [18] parameters, slow
Multi object Energy Uncertainty model, dragonfly algorithm Improved performance of smart Increase in wait time
Mangament in Smart grid
Homes [19]
Otimal household appliance Fuzzy Logic and Machine Learning Reduction in electricity cost, Inaccuracy in results due to assumptions
scheduling [20] optimal scheduling
Smart Energy Mangament in Fuzzy Control System Reduced energy consumption Inaccuracy in results due to assumptions
Residential Areas [21]
Predicting Energy Convolutional Neural Networks (CNN), Improved performance, Trial and Error experimemnts for selecting
Consumption [22] Bidirectional Long Short Term memory computation time and load optimal hyperparameter values,
(LSTM), Auto Encoders distribution insufficient data
Predicting Energy Fruit fly optimizer, simulated annealing High Prediction accuracy and Algorithm can easily fall into local
Consumption of Oil algorithm reduced complexity optimum leading to low convergence
Pipelines [23]
Forecasting building energy Wavelet Transformation, LSTM Efficient forecasting for real Specific forecasting framework, training
consumption [24] case electricity consumption latkes longer and more memory
Prediction of heating energy Pattern Analysis, LSTM Improvised prediction Training latkes longer and more memory,
consumption [25] performance and energy prone to overfitting
concumption

3
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

fuzzy control system for energy management in residential buildings. The study is based on environmental data, which is processed by
a fuzz control system to recommend minimum energy consumption values. The system relies on a forward chaining Mamdani
approach, along with decision tree linearization. As the proposed system generates fuzzy rules, the energy consumption behavior is
highlighted. The proposed method manifests improved accuracy as well as faster computation.
Ref. [22] presented a study on using deep learning methods for predicting energy consumption. The study considers commercial
and domestic buildings, and the proposed architecture incorporates a hybrid framework constructed using a convolutional neural
network (CNN), autoencoder, and bidirectional long short-term memory networks. Experimental analysis suggests improving
computation time, and the method achieves satisfactory performance. Ref. [23] proposed a data-driven model for predicting pipeline
energy consumption. The study deploys a hybrid support vector that relies on the fruit fly optimizer. The model shows high prediction
accuracy and outperforms the other baseline models considered. [24] presented a study on forecasting energy consumption. The study
is based on two decomposition algorithms such as empirical mode decomposition and wavelet transformation, long short-term
memory networks (LSTM). The analysis has been conducted for twenty buildings in various locations with different functionalities.
Results show that LSTM with empirical mode decomposition shows the best performance. [25] suggested LSTM networks for pre­
dicting heating energy consumption on operation patterns of buildings. The study highlights three neural networks applied to different
operation pattern data. The inputs taken for the three LSTMs were different, and adding additional variables to inputs yielded better
results.
Table 1 summarizes the overall strengths and weaknesses of the existing works discussed.
Energy Management is a widely researched problem, and in the past several machine-learning methods have been proposed to
encourage efficient energy usage. Since it is a global concern, a much more challenging task is to find methods of energy management
that also consider performance efficiency and robustness. While the methods proposed previously shed light on the existence of various
methods that can lead to energy management, there is always scope for increased accuracy and endurance, which we present in this
study. The study emphasizes ensemble-based techniques (DT-RF-XGBoost) for analyzing energy consumption by combining multiple
machine-learning techniques into one predictive model. Finding a good balance between bias and variance is necessary to minimize
the total error. An optimal balance between bias and variance ensures no overfitting or underfitting. To understand the behavior of
prediction models, there is a need to find an optimal balance between bias and variance, and ensemble models establish the same.

3. Machine learning methods and our proposed ensemble-based method

In this paper, we use machine learning models extensively, including DT, RF, and XGBoost. In addition, we proposed a hybrid
method based on a Decision Tree, Random Forest, and eXtreme Gradient Boost, called the DT-RF-XGBoost Ensemble method, which
gives much better performance compared to the individual machine learning method.

3.1. Decision Tree (DT)

Decision trees may be defined as supervised machine learning models capable of predicting targets by learning decision rules from
features. A decision tree model learns a set of questions based on deducing class labels and is very useful for interpretation. The root
node or the first parent of a decision tree undergoes recursive partitioning [26]. Every node in this stage may be split into left and right
child nodes, respectively. These nodes can become parents and be split into other nodes [27]. While this can result in a very deep tree
due to overfitting, there is a need to prune the nodes. The optimal splitting is decided by the information gain, defined as an objective
function that requires optimization using the tree learning algorithm. Decision trees are easy to read, learn and prepare. By creating a
comprehensive analysis, decision trees consider all possible outcomes of a decision to conclude.

3.2. Random Forest (RF)

Random Forest regression may be defined as an ensemble learning method based on supervised learning algorithms. An ensemble
learning technique incorporates predictions from several machine learning algorithms for making a more accurate prediction con­
cerning a single model. In a random forest model, the trees run parallel, and there is no interaction among them [28]. Several decision
trees are constructed while training. The output is the mean of the classes. The algorithm works by picking k data points from a training
set, followed by building a decision tree from the points [29]. After choosing the number of trees that must be built, the previous steps
are repeated for all the trees. These new trees will predict the output and assign a new data point. These new data points are averaged
to get the mean output. A random forest regressor is robust and works with features having non-linear relationships too. However,
overfitting is a common problem, hence the number of trees must be chosen correctly.

3.3. eXtreme gradient boosting (XGBoost)

XGBoost is an ensemble learning method where the trees are built sequentially. As the sequencing continues, the errors get reduced.
Each tree learns from the previous tree and reconditions the residual error. The most recent tree will have the least residual error in this
manner. While the base learners in XGBoost are weak, they contribute vital information for prediction [30]. Hence, the overall
boosting technique produces robust learning by combining weak learners. A strong learner also brings down bias and variance.
Boosting uses fewer splits for prediction; thus, even small trees are highly interpretable.It is also possible to optimally select parameters
through validation techniques, such as k-fold cross-validation. Since many trees may lead to overfitting, it is necessary to choose the

4
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

stopping criteria. Thus XGBoost may lead to regularization and can effectively handle sparse data.

3.4. Proposed ensemble model: DT-RF-XGBoost ensemble

Our proposed model is an Ensemble Model based on a Decision Tree, Random Forest, and Extreme Gradient Boost (DT-RF-XGBoost
Ensemble Model). Traditional machine learning models often run into issues related to performance, efficiency, and overfitting, which
can be easily addressed using ensemble methods along with producing extremely accurate predictive models. These learning methods
can combine multiple learners. They have supervised learning methods combining weak learners to produce strong ones. Ensemble
Learning relies on many machine learning algorithms for building more efficient models for improving the overall prediction accuracy.
While weak learners work individually to predict the target outcome, they are not the most optimal models as they are not generalized.
They can predict few cases accurately and may not predict target classes and expected cases efficiently. Hence, combining these weak
learners leads to formation of a generalized strong model which is optimized well enough to predict the target classes efficiently. Weak
learners can be used as building blocks to design complex models as they do not perform well by themselves. When weak learners are
combined, the bias variance trade-off is maintained and the ensemble achieves a better performance. The proposed Ensemble Model
combines the Decision tree, Random Forest, and XGBoost, which have been discussed in detail previously. Fig. 1 depicts the basic
ensemble architecture where multiple models run independently to produce a combined output.
Ensemble-based models ensure the best combination of machine learning algorithms [31]. More than that, ensemble-based models
assert that a combination of algorithms will lead to fewer chances of error than a single algorithm. Hence, machine learning models
working together will have a better potential for gaining accuracy. This is because diverse classifiers combined together will have a
greater potential of gaining higher accuracy compared to non-diverse classifiers. Deploying an ensemble model will lead to improved
average prediction performance over the other contributing components of the ensemble. The improved performance of the ensemble
is associated with a reduction in the variance of prediction errors induced by ensemble components. Hence it adds to the robustness of
the model. In our proposed model, each of the machine learning models (DT-RF-XGBoost) is trained using the same training dataset. All
the models are trained and ready to predict. We have passed the same x_test date to each of the trained models (model-1, model-2,
model-3), and each model predicts an output value (p1, p2, p3). Next, we combined all the predicted values by generating a mean.
Finally, prediction using an ensemble model has been generated.
In the proposed ensemble model, we use a combination of decision trees, random forests and XGBoost algorithms. The decision tree
is capable of handling both numeric and categorical data and work well with multi output problems. It does not require extensive
preparation of data and the easily explained whitebox model is cost effective. Moreover, it works well eve if assumptions are violated.
The limitations of this model lie in its overfitting and unstable nature which is taken care by the random forest component. Random
forests work well with large datasets and provide high accuracy. The major limitation associated with random forests is the training
time which is conveniently handled by the XGBoost algorithm, which provides a direct route to minimum error. Since it is based on
gradient descent, it converges quickly in fewer steps and also leads to improved speed and lower computation cost [39–41]. Thus the

Fig. 1. Proposed ensemble model.

5
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

three components of the ensemble model complement provide immensely to the overall ensemble model, in their own ways and handle
the limitations of each other, thereby making it an overall robust strong learner [42–44].

4. Experimentation results

In this section, we highlight the datasets followed by the evaluation metrics. Based on the experimental analysis, we also present the
experimental results in scatter plots, pair plots, tables, and bar graphs.

4.1. Datasets

In our experimentations, we used two publicly available datasets from ‘smart-home-dataset-with-weather-information1’ Kaggle,
namely HomeC.csv and HomeC1.csv. Both files incorporate readings with a time span of 1 minute of house appliances in kiloWatts
(kW) from smart meters and weather conditions of the specific region. The datasets have 32 features each and a total of 503910 data
points per dataset. Some of the characteristics are based on time, use [kW], gen [kW], House overall [kW], Home office [kW], fridge
[kW]. For the experiment analysis, we are considering two groups, i.e., House overall [kW] and Home office [kW], for both datasets.
The dataset has been split into 80% training data and 20% testing data. The datasets have been split into 80% training data and 20%
test data taking into account the Pareto principle or the 80-20 rule. It is the most common split when the dataset is large. This split
yields statistically meaningful results and good prediction accuracy. Moreover, it contributes to some optimization within the learning.
We apply the data to baseline machine learning models DT, RF, XGB, and KNN and evaluate it against the proposed DT-RF-XGB
ensemble model, using statistical parameters like MSE, R2, RMSE, and MAE, respectively.

4.2. Evaluation metrics

In the experimentations, we used four evaluation metrics: Mean Square Error (MSE), R-squared Error (R2), Root Mean Square Error
(RMSE), and Mean Absolute Error (MEA).

a Mean Square Error (MSE): Mean Square Error or MSE is the mean or average of the square of differences between the real and
predicted values:
/
∑
n
MSE = 1 n (Yi − Ŷi )2
i=1

where MSE is the mean squared error, n is a number of data points, Yi -observed values, and Ŷi -predicted values.

a R-squared Error: R-squared error or R2 is used to determine how close data is to the regression fitted line. It gives the goodness of fit
for regression models.
RSS
R2 = 1 −
TSS

where R2 is the coefficient of determination, RSS is the sum of squares of residuals, and TSS is the total sum of squares.

a Root Mean Square Error (RMSE): Root Mean Square Error or RMSE is used to determine the error in a model for estimating predictive
data.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑n 2
i=1 (Predicatedi − Actuali )
RMSE =
N

b Mean Absolute Error (MAE): Mean Absolute Error or MAE may be defined as the measure of errors for paired observations that
express the same phenomenon
∑n
|yi − xi |
MAE = i=1
n

where MAE is the mean absolute error, yi is the prediction, xi is the true value, and n is the total number of data points.

6
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

4.3. Experimental results

In this subsection, we will depict the results for both datasets: Dataset1 (HomeC) and Dataset2 (HomeC1). For Dataset1 (D1) and
Dataset2 (D2), we will be specifically analyzing power consumption for fields Home Office [kW] and House overall [kW].
While the datasets look similar and have the same features, the major difference between the two datasets lies in their values.
Moreover, the domain values for Home Office and House Overall are different for both datasets. Thus the datasets and the field values
have been used independently to perform the experimental analysis. The underlying idea behind using multiple datasets with different
field values is to ensure non-independence and observe the patterns and results across multiple datasets for better overall analysis.
The results have been depicted using two data visualization tools, i.e., Pair Plots and Bar graphs. Using pair plots, we can obtain
scatter plots for each of the machine learning models for both fields in both datasets. The scatter plot depicts the relationship between
two variables, i.e., power consumption concerning time. The comparison between the values has been depicted using bar graphs.

4.3.1. Experimental results from dataset1 (HomeC)

This dataset has two fields: Home Office [kW] and House Overall [kW]. Firstly, we conducted experiments for the Home Office
[kW] using individual Machine Learning techniques (e.g., Decision Tree, Random Forest, XGBoost, and KNN) and our proposed
Ensemble Model (DT-RF-XGBoost). The results are presented in Fig. 2a-e.
Fig. 2a-d depict the scatterplot relationship between electricity consumption (Home Office) (x-axis) concerning time (y-axis) for
Dataset1 (D1) for Decision Tree, Random Forest, XGBoost, and KNN, respectively. All the scatterplots show the values lying in the grid,
and the horizontal line at the bottom implies how close the predicted values are to the original values. The line plot is an extension of
the same scatter plot graph and shows that most values lie toward the bottom.
Comparing with the competitor methods, namely Decision Tree, Random Forest, XGBoost, and KNN (shown previously in Figs. 2-
5), Fig. 6 depicts the Ensemble-based scatterplot relationship between electricity consumption (Home Office) (X-axis) concerning time
(Y-axis) for Dataset 1 (D1). Here, the data is less scattered and based on the results, we observe that the predicted values here seem
the closest to the original values concerning the other models. Again, the line plot is an extension of the same scatter plot graph and
shows that maximum values lie towards the bottom. Hence, the proposed Ensemble Model exhibits the best performance.
In the second experiment, we conducted experiments for the House Overall [kW] using individual Machine Learning techniques (e.
g., Decision Tree, Random Forest, XGBoost, and KNN), and our proposed Ensemble Model (DT-RF-XGBoost). The results are presented
in Figs. 7-11.
Figs. 7-10 shows a scatterplot relationship between electricity consumption (House Overall) for (x-axis) concerning time (y-axis)
for Dataset1 (D1) using Decision Tree, Random Forest, XGBoost, and KNN, respectively. The scatterplot shows the values lying in the
grid, and it appears that the predicted values are close to the original values as they are clustered toward the bottom of the graph. The
line plot is an extension of the same scatter plot graph and shows that most values lie toward the bottom. Since the changes occur on a
microscopic scale, representing minute changes is almost impossible. Hence, while the graphs seem similar, there are significant
differences in the values of performance parameters.
Comparing the four individual methods: Decision Tree, Random Forest, XGBoost, and KNN, as shown in Figs. 7-10, Fig. 11 depicts
the Ensemble-based scatterplot. The predicted values, in this case, are the closest to the original values, and the line plot is an
extension of the same scatter plot graph. Due to changes in the values on a microscopic scale, depicting the subtle differences is
challenging; however, the values depict significant differences in the performance metrics of all the models.

4.3.2. Experimental results from Dataset2 (HomeC1)

We conducted the same experiments for Dataset2 (HomeC1); particularly, we ran experiments to evaluate Home Office [kW] and
House Overall [kW] using individual Machine Learning techniques, as well as using our proposed Ensemble Model. The results

Fig. 2. Decision tree scatter plot and pair plot for home office (D1).

7
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

Fig. 3. Random forest scatter plot and pair plot for home office (D1).

presented in Fig. 4a-e are for Home Office [kW].

Figs. 12-15 depicts the scatterplot relationship between electricity consumption (Home Office) for (x-axis) concerning time (y-axis)
for Dataset2 (D2), using Decision Tree, Random Forest, XGBoost, and KNN, respectively. The scatterplot shows the values in the grid,
and the horizontal line at the bottom implies how close the predicted values are to the original values. For the XGBoost scatterplot (see
Fig. 14), while the data points look much more scattered here, the predicted values are seemingly close enough to the original values at
the bottom. For the KNN scatter plot (see Fig. 15), fewer are scattered here, and the predicted values are close to the original values at
the bottom.
The line plot is an extension of the same scatter plot graph and shows that most values lie toward the bottom.
Fig. 16 depicts the Ensemble-based scatterplot relationship between electricity consumption (Home Office) for (X-axis) concerning
time (Y-axis) for Dataset 2 (D2). In this case, the predicted values are the closest to the original values at the bottom. As the changes are
made on a microscopic scale, depicting the same on a grid is challenging. However, the performance evaluation values differ
significantly from previous models. The line plot is an extension of the same scatter plot graph and shows that most values lie toward
the bottom.
For the remaining part of the experiment, we conducted an analysis of the House Overall [kW] using individual Machine Learning
techniques (e.g., Decision Tree, Random Forest, XGBoost, and KNN), and our proposed Ensemble Model (DT-RF-XGBoost). The results
are presented in Figs. 17-21.
Figs. 17-20 depicts the scatterplot relationship between electricity consumption (House Overall) for (X-axis) concerning time (Y-
axis) for Dataset 2 (D2) using Decision Tree, Random Forest, XGBoost, and KNN, respectively. The scatterplot shows the values lying in
the grid, and the predicted values are close to the original values. The line plot is an extension of the same scatter plot graph and shows
that most values lie toward the bottom. Since the changes occur on a microscopic scale, representing minute changes is almost
impossible. Hence, while the graphs seem similar, there are significant differences in the values of performance parameters.
Finally, Fig. 21 depicts the Ensemble-based scatterplot relationship between electricity consumption (House Overall) for (x-axis)
concerning time (y-axis) for Dataset2 (D2). In this case, the predicted values are the closest to the original values at the bottom. As the
changes are made on a microscopic scale, depicting the same on a grid is challenging. However, the performance evaluation values

Fig. 4. XGB scatter plot and pair plot for home office (D1).

8
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

Fig. 5. KNN scatter plot and pair plot for home office (D1).

Fig. 6. Proposed ensemble model scatter plot and pair plot for home office (D1).

Fig. 7. Decision tree scatter plot and pair plot for house overall (D1).

9
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

Fig. 8. Random forest scatter plot and pair plot for house overall (D1).

Fig. 9. XG-Boost scatter plot and pair plot for house overall (D1).

Fig. 10. KNN scatter plot and pair plot for house overall (D1).

10
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

Fig. 11. Proposed ensemble model scatter plot and pair plot for house overall (D1).

Fig. 12. Decision scatter plot and pair plot for home office (D2).

Fig. 13. Random forest scatter plot and pair plot for home office (D2).

differ significantly concerning previous models. The line plot is an extension of the same scatter plot graph and shows that most values
lie toward the bottom.

4.3.3. Experimental results from the performance metric

As mentioned in the previous section, the performance evaluation metrics considered for the study are MSE, R2, RMSE, and MAE. In
this section, we will present the performance evaluation values based on the experimental analysis to better compare the models.

11
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

Fig. 14. XGBoost Scatter Plot and Pair Plot for Home Office (D2).

Fig. 15. KNN Scatter Plot and Pair Plot for Home Office (D2).

Fig. 16. Proposed ensemble model scatter plot and pair plot for home office (D2).

Table 2 shows the results from Dataset1 (HomeC) Home Office, whereas Table 3 shows the results from Dataset1 (HomeC) House
Overall.
From Table 2, we can deduce that the MSE value for the proposed Ensemble Model is up to one order of magnitude better than the
Decision Trees and XGBoost. The R2 value for the proposed Ensemble Model is also the best among all other methods. For the RMSE
value, the Ensemble Model performs better than all other methods except KNN. Finally, for the MAE value, the Ensemble Model is
better than others, except XGBoost. While RMSE for KNN seems to be the lowest, it seems reasonable for the proposed model. Hence the
proposed model outperforms the other baseline models.

12
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

Fig. 17. DT Scatter plot and pair plot for house overall (D2).

Fig. 18. RF Scatter plot and pair plot for house overall (D2).

Fig. 19. XGB scatter plot and pair plot for house overall (D2).

13
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

Fig. 20. KNN scatter plot and pair plot for house overall (D2).

Fig. 21. Proposed ensemble model scatter plot and pair plot for house overall (D2).

Table 2
Performance evaluation for home office (D1).
Machine Learning Algorithm MSE R2 RMSE MAE

Decision Trees 0.0014 0.8674 0.0383 0.0087

Random Forest 0.0006 0.9393 0.0259 0.0086
XGBoost 0.0012 0.8839 0.0358 0.0202
k-Nearest Neighbor 0.0003 0.9477 0.0116 0.0018
Proposed Ensemble Model 0.0001 0.9846 0.0173 0.0108

Table 3
Performance evaluation for house overall (D1).
Machine Learning Algorithm MSE R2 RMSE MAE

Decision Trees 0.049800 0.98010 0.00128 0.00026

Random Forest 0.000001 0.98820 0.00082 0.00010
XGBoost 0.000006 0.98390 0.00272 0.00164
k-Nearest Neighbor 0.069850 0.99010 0.26430 0.08870
Proposed Ensemble Model 0.000008 0.99999 0.00072 0.00033

14
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

Table 4
Performance Evaluation for Home Office (D2).
Machine Learning Algorithm MSE R2 RMSE MAE

Decision Trees 0.00147 0.86666 0.03841 0.00878

Random Forest 0.00069 0.93437 0.02630 0.00069
XGBoost 0.00128 0.88398 0.03583 0.02022
k-Nearest Neighbor 0.00113 0.94142 0.01163 0.00187
Proposed Ensemble Model 0.00061 0.98776 0.02486 0.0109

Table 5
Performance evaluation for house overall (D2).
Machine Learning Algorithm MSE R2 RMSE MAE

Decision Trees 0.01418 0.98998 0.00142 0.00026

Random Forest 0.00004 0.98099 0.00088 0.0001
XGBoost 0.000006 0.96969 0.00272 0.00164
k-Nearest Neighbor 0.06985 0.93556 0.2643 0.0887
Proposed Ensemble Model 0.000002 0.99999 0.00078 0.00033

From Table 3, we can easily see that The R2 and RMSE values for the proposed Ensemble Model seem to be the best. While Random
Forest has the best MSE value and the MAE value, the proposed ensemble model has reasonable MAE and MSE values. Hence the
proposed model outperforms the other models in this case.
Tables 4 and 5 focus on Dataset2 (HomeC1). Again, we experimented with the Home Office and House Overall elements of the
dataset.
From Table 4, it shows that the MSE and R2 values for the proposed model seem to be the best in this case. While the RMSE value is
the best for KNN, and the MAE value is best for RF, the overall performance of the proposed Ensemble model seems better.
From Table 5, it shows that the MSE, R2 and RMSE values for the proposed Ensemble model seem to be the best in this case.
Although the MAE value is the best for Random Forest, the overall performance of the proposed model seems better.
Based on the experiments above, we can confirm that the proposed Ensemble model outperforms the other baseline models.
The same may be explained using bar graphs (see Figs. 22-23). From Fig. 22, it is evident that the proposed ensemble model
achieves the best RMSE value for Dataset1 for both Home Office and House Overall.
Similarly, from Fig. 23, it is evident that the proposed ensemble model achieves the best RMSE value for Dataset2 in terms of Home
Office as well as House Overall.

4.4. Comparative analysis

While the experimental analysis on multiple datasets indicates that the overall performance of the proposed Ensemble model
outperforms the baseline models for both the fields in both the datasets, we present a comparative analysis of our work with some of
the relevant research works done in the past (Table 6).
Based on the overall analysis, we can make certain observations. The significance of this work is multi-fold. We have introduced yet
another machine-learning methodology for monitoring. power consumption in smart homes. The proposed method is an ensemble
approach incorporating three machine learning models, i.e., DT, RF and XGBoost The approach was adopted to enhance performance
issues and robustness, which are often a limitation in traditional machine learning approaches. The performance of the proposed
ensemble method was compared to traditional machine learning methods, including the individual components of the ensemble

Fig. 22. Performance evaluation for home office and house overall (D1).

15
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

Fig. 23. Performance evaluation for home office (D2).

Table 6
Comparative analysis of our proposed work with related works.
Authors Proposed Work Methodologies/ Parameters Result

Culaba et al. Energy Consumption for mix Clustering and Forecasting (k-means, support vector R2~0.0201, MAE~ 0.0163
(2020) [32] use buildings machines)
Khan et al. (2020) Predicting Energy Consumption Multilayer perceptron (MLP), support vector regression MAE~15.72, MSE~472.96, RMSE~
[33] for renewable/non-renewable (SVR), and CatBoost, 21.74
power sources
Chou and Truong, Energy Consumption Time Series ML models R2~ 0.86, RMSE~ 8.421, MAE~70.919
(2021) [34] Forecasting (MWh)
Shapi et al. (2021) Energy Consumption prediction Support Vector Machine, Artificial Neural Network, and k- Lowest RMSE for KNN~0.5439, SVM~
[35] for smart building Nearest Neighbour 0.5558, ANN~0.5471

Lee et al. (2021) Prediction of Heating Energy Deep Neural Network R2~0.961
[36] Consumption
Amasyali and El- Energy consumption prediction ML algorithms like classification and regression trees DNN performs the best with 2.97%
Gohary in office building (CART), ensemble bagging trees (EBT), artificial neural coefficient of variation
(2021) [37] networks (ANN), and deep neural networks (DNN)
Khan et al (2021) Energy Consumption Ensemble, LSTM and Gated Recurrent Units (GRU) Mean Absolute Percentage Error (MAPE)
[38] Forecasting Model values are 4.182 and 4.54
Proposed Model Predicting energy consumption DT-RF-XGBoost based Ensemble Model R2~0.9999 across both datasets, likewise,
in smart homes MSE, RMSE and MAE values low across all
fields in both datasets

model.
Moreover, the experimental analysis was conducted on multiple datasets to avoid any kind of bias. The analysis yields consistent
results across multiple datasets depicting that the proposed method performs better than the individual machine-learning models.
Moreover, the significance has been justified in the comparative analysis as our proposed work shows satisfactory results compared
to similar research works in the past.

5. Conclusions and future work

Increased energy consumption has led to an increased carbon footprint and elevated climate change risk. Due to the higher demand
for energy across the globe, not only higher costs of energy are incurred, but also there is a constant demand for supply. Hence
monitoring energy consumption is necessary to manage energy costs and realize saving opportunities. One of the common ways of
monitoring energy consumption is by predicting its usage.
In this study, we have deployed four machine learning algorithms to study the energy consumption in smart homes, i.e., DT, RF,
XGBoost and KNN
Moreover, we have proposed a novel Decision Tree- Random Forest- XGBoost-based Ensemble model (DT-RF-XGBoost) for
comparing it to the four baseline machine learning algorithms. The study considered two datasets that incorporate readings with a time
span of one minute of house appliances in kiloWatts (kW) from smart meters. Each of these datasets has multiple fields. For this study,
we have considered two fields from each dataset, which results in an overall four different experimental analyses, hence strengthening
our claim. The performance evaluation metrics considered for the study are MSE, R2, RMSE, and MAE. Our study depicts that the
proposed ensemble model outperforms all the baseline algorithms across multiple fields in multiple datasets.
In the future, we would like to monitor energy consumption by relying on other machine learning techniques like Neural Networks
and Optimization methods. As machine learning methods continue gaining popularity, it would be interesting to solve such problems
using deep learning, time series analysis, and other advanced machine learning algorithms. Energy consumption can be studied on a

16
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

global scale to analyze the largest sources of energy consumption. Consequently, techniques may be introduced to address the same.
Moreover, once the consumption is analyzed, we can look forward to building trustworthy systems for the reliable consumption of
energy.

Declaration of Competing Interest

None.

Data availability

Data will be made available on request.

References

[1] H. Kim, H. Choi, H. Kang, J. An, S. Yeom, T. Hong, A systematic review of the smart energy conservation system: From smart homes to sustainable smart cities,
Renewable Sustainable Energy Rev. 140 (2021), 110755.
[2] V. Puri, S. Jha, R. Kumar, I. Priyadarshini, M. Abdel-Basset, M. Elhoseny, H.V. Long, A hybrid artificial intelligence and internet of things model for generation of
renewable resource of energy, IEEE Access 7 (2019) 111181–111191.
[3] S. Alani, S.N. Mahmood, S.Z. Attaallah, H.S. Mahmood, Z.A. Khudhur, A.A. Dhannoon, IoT based implemented comparison analysis of two well-known network
platforms for smart home automation, Int. J. Electr. Comput. Eng. 11 (1) (2021), 2088-8708.
[4] X. Wang, X. Mao, H. Khodaei, A multi-objective home energy management system based on internet of things and optimization algorithms, J. Build. Eng. 33
(2021), 101603.
[5] Y.B. Hamdan, Smart home environment future challenges and issues-a survey, J. Electron. 3 (01) (2021) 239–246.
[6] D. Chakraborty, A. Alam, S. Chaudhuri, H. Başağaoğlu, T. Sulbaran, S. Langar, Scenario-based prediction of climate change impacts on building cooling energy
consumption with explainable artificial intelligence, Appl. Energy 291 (2021), 116807.
[7] A. Babuta, B. Gupta, A. Kumar, S. Ganguli, Power and energy measurement devices: a review, comparison, discussion, and the future of research, Measurement
172 (2021), 108961.
[8] E. García-Martín, C.F. Rodrigues, G. Riley, H. Grahn, Estimation of energy consumption in machine learning, J. Parallel Distrib. Comput. 134 (2019) 75–88.
[9] X. Dong, S. Deng, D. Wang, A short-term power load forecasting method based on k-means and SVM, J Ambient Intell Humaniz Comput (2021) 1–15.
[10] S. Verma, S. Singh, A. Majumdar, Multi-label LSTM autoencoder for non-intrusive appliance load monitoring, Electr. Power Syst. Res. 199 (2021), 107414.
[11] S. Kaushik, K. Srinivasan, B. Sharmila, D. Devasena, M. Suresh, H. Panchal, N. Srimali, Continuous monitoring of power consumption in urban buildings based
on Internet of Things, Int. J. Ambient Energy (2021) 1–7.
[12] J.S. GK, J. Jasper, MANFIS based SMART home energy management system to support SMART grid, Peer Peer Netw Appl (2020) 1–12.
[13] I. Machorro-Cano, G. Alor-Hernández, M.A. Paredes-Valverde, L. Rodríguez-Mazahua, J.L. Sánchez-Cervantes, J.O. Olmedo-Aguirre, HEMS-IoT: A big data and
machine learning-based smart home system for energy saving, Energies 13 (5) (2020) 1097.
[14] A. Akbari-Dibavar, S. Nojavan, B. Mohammadi-Ivatloo, K. Zare, Smart home energy management using hybrid robust-stochastic optimization, Comput. Ind. Eng.
143 (2020), 106425.
[15] I. Hussain, M. Ullah, I. Ullah, A. Bibi, M. Naeem, M. Singh, D. Singh, Optimizing energy consumption in the home energy management system via a bio-inspired
dragonfly algorithm and the genetic algorithm, Electronics 9 (3) (2020) 406.
[16] K.H.N. Bui, I.E. Agbehadji, R. Millham, D. Camacho, J.J. Jung, Distributed artificial bee colony approach for connected appliances in smart home energy
management system, Process. Expert Syst., Technol. Value Sugar Beet, Prog. Sugar Technol.: Proc. Gen. Assem. C.I.T.S. , 20th 37 (6) (2020) e12521.
[17] M. Elsisi, M.Q. Tran, K. Mahmoud, M. Lehtonen, M.M. Darwish, Deep learning-based industry 4.0 and Internet of Things towards effective energy management
for smart buildings, Sensors 21 (4) (2021) 1038.
[18] S. Ghosh, D. Chatterjee, Artificial bee colony optimization based non-intrusive appliances load monitoring technique in a smart home, IEEE Trans. Consum.
Electron. 67 (1) (2021) 77–86.
[19] M. Alilou, B. Tousi, H. Shayeghi, Multi-objective energy management of smart homes considering uncertainty in wind power forecasting, Electr. Eng. (2021)
1–17.
[20] S. Atef, N. Ismail, A.B. Eltawil, A new fuzzy logic based approach for optimal household appliance scheduling based on electricity price and load consumption
prediction, Adv. Build. Energy Res. (2021) 1–19.
[21] D. Kontogiannis, D. Bargiotas, A. Daskalopulu, Fuzzy control system for smart energy management in residential buildings based on environmental data,
Energies 14 (3) (2021) 752.
[22] O. Jogunola, B. Adebisi, K.V. Hoang, Y. Tsado, S.I. Popoola, M. Hammoudeh, R. Nawaz, CBLSTM-AE: A Hybrid deep learning framework for predicting energy
consumption, Energies 15 (3) (2022) 810.
[23] H. Lu, Z.D. Xu, M. Azimi, L. Fu, Y. Wang, An effective data-driven model for predicting energy consumption of long-distance oil pipelines, J. Pipeline Syst. Eng.
Pract. 13 (2) (2022), 04022005.
[24] S.Y. Chou, A. Dewabharata, F.E. Zulvia, M. Fadil, Forecasting building energy consumption using ensemble empirical mode decomposition, wavelet
transformation, and long short-term memory algorithms, Energies 15 (3) (2022) 1035.
[25] J. Jang, J. Han, S.B. Leigh, Prediction of heating energy consumption with operation pattern variables for non-residential buildings using LSTM networks,
Energy Build. 255 (2022), 111647.
[26] S. Jha, R. Kumar, M. Abdel-Basset, I. Priyadarshini, R. Sharma, H.V. Long, Deep learning approach for software maintainability metrics prediction, IEEE Access 7
(2019) 61840–61855.
[27] T.A. Tuan, H.V. Long, R. Kumar, I. Priyadarshini, N.T.K. Son, Performance evaluation of Botnet DDoS attack detection using machine learning, Evolut. Intell.
(2019) 1–12.
[28] I. Priyadarshini, V. Puri, A convolutional neural network (CNN) based ensemble model for exoplanet detection, Earth Sci. Inf. (2021) 1–13.
[29] N. Pritam, M. Khari, R. Kumar, S. Jha, I. Priyadarshini, M. Abdel-Basset, H.V. Long, Assessment of code smell for predicting class change proneness using
machine learning, IEEE Access 7 (2019) 37414–37425.
[30] T. Chen, C. Guestrin, XGBoost: A scalable tree boosting system, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, 2016, pp. 785–794.
[31] K.W. Hsu, A theoretical analysis of why hybrid ensembles work, Comput. Intell. Neurosci. 2017 (2017).
[32] A.B. Culaba, A.J.R. Del Rosario, A.T. Ubando, J.S Chang, Machine learning-based energy consumption clustering and forecasting for mixed-use buildings, Int. J.
Energy Res. 44 (12) (2020) 9659–9673.
[33] P.W. Khan, Y.C. Byun, S.J. Lee, D.H. Kang, J.Y. Kang, H.S. Park, Machine learning-based approach to predict energy consumption of renewable and non-
renewable power sources, Energies 13 (18) (2020) 4870.

17
I. Priyadarshini et al. Internet of Things 20 (2022) 100636

[34] J.S. Chou, D.N. Truong, Multistep energy consumption forecasting by metaheuristic optimization of time-series analysis and machine learning, Int. J. Energy
Res. 45 (3) (2021) 4581–4612.
[35] M.K.M. Shapi, N.A. Ramli, L.J. Awalin, Energy consumption prediction by using machine learning for smart building: Case study in Malaysia, Dev. Built Environ.
5 (2021), 100037.
[36] S. Lee, S. Cho, S.H. Kim, J. Kim, S. Chae, H. Jeong, T. Kim, Deep neural network approach for prediction of heating energy consumption in old houses, Energies
14 (1) (2021) 122.
[37] K. Amasyali, N. El-Gohary, Machine learning for occupant-behavior-sensitive cooling energy consumption prediction in office buildings, Renew. Sustain. Energy
Rev. 142 (2021), 110714.
[38] A.N. Khan, N. Iqbal, A. Rizwan, R. Ahmad, D.H. Kim, An ensemble energy consumption forecasting model based on spatial-temporal clustering analysis in
residential buildings, Energies 14 (11) (2021) 3020.
[39] A. Lakhan, MA. Mohammed, AN. Rashid, Seifedine Kadry and Karrar Hameed Abdulkareem, Deadline aware and energy-efficient scheduling algorithm for fine-
grained tasks in mobile edge computing, Int. J. Web Grid Serv. 18 (2) (2022).
[40] B. Bhola, R. Kumar, BK. Mishra, Internet of things-based low cost water meter with multi functionality, Int. J. Web Grid Serv. 18 (3) (2022) 250–265.
[41] H. Sun, M. Liu, Z. Qing, X. Li, L. Li, Energy consumption optimisation based on mobile edge computing in power grid internet of things nodes, Int. J. Web Grid
Serv. 16 (3) (2020) 238–253, https://doi.org/10.1504/IJWGS.2020.109468.
[42] A. Balamane, Scalable Biclustering algorithm considers the presence or absence of properties, Int. J. Data Warehous. Min. (IJDWM) 17 (1) (2021) 39–56, https://
doi.org/10.4018/IJDWM.2021010103.
[43] H. Li, Z. Liu, P. Zhu, An engineering domain knowledge-based framework for modelling highly incomplete industrial data, Int. J. Data Warehous. Min. (IJDWM)
17 (4) (2021) 48–66, https://doi.org/10.4018/IJDWM.2021100103.
[44] T.T. Nguyen, N.L. Giang, D.T. Tran, T.T. Nguyen, H.Q. Nguyen, A.V. Pham, T.D. Vu, A novel filter-wrapper algorithm on intuitionistic fuzzy set for attribute
reduction from decision tables, Int. J. Data Warehous. Min. (IJDWM) 17 (4) (2021) 67–100, https://doi.org/10.4018/IJDWM.2021100104.

18