Tech Science Press

DOI: 10.32604/cmc.2024.051289

REVIEW

A Review of NILM Applications with Machine Learning Approaches

Maheesha Dhashantha Silva* and Qi Liu

School of Electronic and Information Engineering, Nanjing University of Information Science and Technology,
Nanjing, 210044, China
*Corresponding Author: Maheesha Dhashantha Silva. Email: [email protected]
Received: 01 March 2024 Accepted: 11 April 2024 Published: 15 May 2024

ABSTRACT
In recent years, Non-Intrusive Load Monitoring (NILM) has become an emerging approach that provides affordable
energy management solutions using aggregated load obtained from a single smart meter in the power grid.
Furthermore, by integrating Machine Learning (ML), NILM can efficiently use electrical energy and offer less of
a burden for the energy monitoring process. However, conducted research works have limitations for real-time
implementation due to the practical issues. This paper aims to identify the contribution of ML approaches to
developing a reliable Energy Management (EM) solution with NILM. Firstly, phases of the NILM are discussed,
along with the research works that have been conducted in the domain. Secondly, the contribution of machine
learning approaches in three aspects is discussed: Supervised learning, unsupervised learning, and hybrid modeling.
It highlights the limitations in the applicability of ML approaches in the field. Then, the challenges in the real-
time implementation are concerned with six use cases: Difficulty in recognizing multiple loads at a given time,
cost of running the NILM system, lack of universal framework for appliance detection, anomaly detection and
new appliance identification, and complexity of the electricity loads and real-time demand side management.
Furthermore, options for selecting an approach for an efficient NILM framework are suggested. Finally, suggestions
are provided for future research directions.

KEYWORDS
Non-intrusive load monitoring; transfer learning; machine learning; feature extraction

1 Introduction
Universal access to electricity is a crucial task which ensures sustainable economic growth in a
country. According to [1], the number of people with access to electricity is very low in sub-Saharan
Africa, which was 589.46 million in 2019. Furthermore, 77% of the population uses solid fuel for
household work, leading to non-communicable diseases due to air pollution. The use of non-renewable
sources increases global warming and climate change. On the other hand, the availability of renewable
energy sources directly impacts the country’s income. The study [2] shows that there is a relation-
ship between energy production and Gross Domestic Product (GDP) and having steady electricity
production directly affects economic growth. Because of the global population and economic and
technological development, electricity usage has increased. To solve the energy crisis, the paper [3]

This work is licensed under a Creative Commons Attribution 4.0 International License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.
2972 CMC, 2024, vol.79, no.2

suggested using control-optimized methods, passive design in transportation and buildings and green
alternatives to save energy.
Microgrids (MGs) were developed by combining conventional and non-conventional energy
sources with the integration of the Internet of Things (IoT) to address the electricity demand
effectively. Conventional energy sources include non-renewable energy sources such as fossil fuels,
while renewable energy options are composed of non-conventional energy sources, including wind,
biogas, and hydropower [4]. Further, microgrids have been introduced to achieve the higher impact
of renewable energy sources in the power grid. The energy storage system saves the excess energy
generated in the microgrid to improve load consumption. However, the paper [5] highlights Microgrid
(MG) configuration is limited in rural areas due to the unavailability of internet resources. On the other
hand, load transition identification in the MG is difficult due to the non-linear behavior of loads [6].
Several models [7] and algorithms are proposed in the literature to identify the load variations [8].
Furthermore, studies [9] found that the accuracy of the forecasting results depends on the demand.
However, the microgrid’s demand side response can be improved by implementing policies, power flow
constraints and energy monitoring.
Understanding energy consumption benefits consumers and suppliers in the power grid. From the
consumers’ point of view, they can identify the appliances that contribute to the electricity bill and take
the initiative to reduce that. On the other hand, from the suppliers’ point of view, they can identify the
electricity demand of each consumer and expand the service accordingly. Related statistics published in
the paper [10] shows that direct feedback from consumers can save energy wastage by more than 15%.
However, energy monitoring is flexible by installing smart meters attached to the appliance. Overall,
load monitoring can be classified into three approaches: Intrusive (ILM), Semi-Intrusive (Semi-ILM)
and Non-Intrusive (NILM) based on the number of smart meters installed in the power grid. The
complexity of the energy disaggregation in the NILM approaches is difficult because a single smart
meter is used to collect the load. However, the cost of installation and configuration is lower in the
NILM-based energy management solutions.
ML models have been presented in the literature to address the complexity of the NILM load
disaggregation [11]. Such as Hidden Markov Models (HMMs) [12,13], Neural Networks [14,15] and
Convolutional Neural Networks (CNNs) [16]. Approaches applied in the load disaggregation, such
as supervised, semi-supervised and unsupervised, are labeled based on the learning process of the
data. Supervised learning requires a large amount of data to learn the appliance. On the contrary,
unsupervised learning does not require labeling before the training and is idle for identifying new
appliances. However, studies [17] found that unsupervised learning is effective with small application
areas since anomaly detection does not affect appliance recognition.
The smart grid compromises with different appliances that consume different loads at different
frequencies. Therefore, the availability of public datasets such as REDD [18], PLAID [19], WHITED
[20], UK-DALE [21], and COOLL [22] help to train the supervised model to learn the ground truth
about appliances in the NILM research area. A summary of the available datasets in the NILM
research area is available in Table 1.
CMC, 2024, vol.79, no.2 2973

Table 1: Summary of the publicly available datasets

Data set Period Features Ref.
Number of Sampling Load Location Challenges
appliances rates type
REDD 3 weeks Six different 1 Hz All MIT Diverse appliance [18]
homes with Sustainable behavior.
appliances lab Varying sampling rates
of appliances.
PLAID One year 55 homes 30 kHz All US No ground-truth labels [19]
for individual
appliances.
WHITED – 110 appliances 44 kHz All Germany Only contain start-up [20]
transients.
UK-DALE One year Five different 6s All UK Variability in appliance [21]
homes usage patterns.
COOLL One month 42 appliances 100 kHz All France Variability in appliance [22]
signals is difficult to
record.
LIT – Three subsets, 15.3 kHz I, II, III Brazil [23]
including
different
numbers of
appliances
AMPds Two years One home with 1 Hz All British Real-time applicability [24]
21 sub meters Columbia
IAWE 73 days One building 1 Hz I, II, III India Real-time applicability [25]
iAWE 16 bimonthly 9923 homes 1 Hz I, II, III Ireland Multiple meters per [26]
periods house are available.
Houses with less than
six observations are
dropped from the data
collection.
ECO Eight Six homes 1 Hz All Switzerland [27]
months
BLUED 2–4 weeks One home with 1 Hz I, II, III US It is unsuitable to [28]
30 appliances evaluate classification
and energy estimation
(EE) algorithms.
HFED – 24 appliances 10 kHz All India Only contain start-up [29]
transients [24].
BLOND 213 days 53 appliances 50 kHz All Germany Only high-frequency [30]
data is available.
RBSA – 101 houses 15 min All USA Only low-frequency [31]
data is available.
IDEAL 22 months 255 houses – All UK Only low-frequency [32]
data is available.
HUE Three years 22 houses – All Canada Only low-frequency [33]
data is available.
(Continued)
2974 CMC, 2024, vol.79, no.2

Table 1 (continued)
Data set Period Features Ref.
Number of Sampling Load Location Challenges
appliances rates type
ENERTALK 29–122 days 22 houses 15 Hz All Korea Only low-frequency [34]
data is available.
SynD 180 days 21 appliances 5 Hz All Austria Only low-frequency [35]
data is available.

In addition to the mentioned challenges in Table 1, preprocessing the dataset before applying the
model is required because of the contradiction of data; missing and noisy data may affect the load
identification in the NILM approach [36].
This paper provides an inclusive review of the NILM applications with ML approaches, and
challenges are discussed for future research directions. The main contributions of this paper are as
follows:
1. To identify different Supervised and Unsupervised Learning approaches used in Electricity
Management with NILM
2. To highlight the current research challenges to find the feasibility of developing a reliable ML
Approach in the NILM research field.
The remaining sections of this article are organized as follows: In Section 2, the NILM framework
is explained with load identification using different ML approaches, and it discusses how demand-
side management helps improve awareness. In Section 3, the different research works conducted
with various ML approaches are discussed, and it further explains implemented hybrid frameworks.
Section 4 outlines open challenges and issues in the conducted research and explains why the hybrid
framework is feasible for real-time implementation. Finally, Section 5 concludes the article.

2 NILM Framework
Load Monitoring enables energy consumption monitoring by a smart meter attached to the
electricity appliance. The concept of NILM was invented by Hart [37] in 1984 to extract the load
consumption of devices using aggregated power load collected at the end of the power source. The
initial approach is suitable to capture the ON/OFF state of the electrical appliances in the smart grid.
After that, NILM implementation has become an emerging area for load monitoring in the home and
commercial setting.
NILM contributes to identifying the energy consumption of appliances, leading consumers to
reduce usage and lower electricity bills. Further, it helps to identify the appliance’s health and then
recognize the need for maintenance or replacement. As a result of demand-side management features
provided in the NILM architecture, the operational times of each appliance can be recognized and
controlled [38]. Overall, NILM works as a powerful and cost-effective tool to increase energy efficiency,
minimize energy crises, and environmental impact.
The Typical NILM framework comprises five stages: Data acquisition and preprocessing, feature
extraction, load identification through a learning process, and demand-side management. Since
the NILM architecture depends on the aggregated data, computational work is high throughout
CMC, 2024, vol.79, no.2 2975

the disaggregation process. This section discusses the framework of NILM with the methods and
techniques used in each stage.

2.1 Data Acquisition

In general, different types of electrical appliances are available in the power grid. Based on the load
behavior, devices can be classified into four categories: Type I (with ON/OFF states), Type II (with
multi-states), Type III (continuously varies) and Type IV (constant power). And different sampling
rates are used to collect aggregated data: Low sampling (less than 1 Hz) and high sampling rate
(10 kHz). Conducted research works with low sampling rates are limited due to the feature loss [39,40].
Furthermore, the study [41] highlights that the results of the load disaggregation can be improved
using the data collected on low sampling frequency. On the other hand, high-frequency applications
are limited due to the high computational cost [42]. However, it supports feature analysis with ML to
improve load identification [43]. However, the load identification performance depends on the input
data quality in the disaggregation process.
In NILM, data preprocessing is crucial since the entire framework depends on the aggregated
data collected at the end. It involves data cleaning, feature engineering, temporal alignment of data,
resampling for a consistent time interval, and normalization and scaling for a similar interval. The
paper [44] presented a hybrid statistical filtering method for outlier detection on the aggregated data
collected with a high-frequency rate. However, Low-frequency outlier detection remains questionable
for practical settings. The article [45] suggested the Transform NILM method to handle the data
balancing. It replaces aggregated data with fake samples and creates masked aggregate data. Then,
the Transform generator takes the masked aggregated signal as input and trains the data to predict the
original values at the masked aggregate data positions.

2.2 Feature Extraction

In NILM, the feature extraction stage refers to deriving relevant information from the prepro-
cessed aggregated load data. Features are characteristics of the aggregated load data, which can be
identified from the time-series data of the electric signal. In overall, approaches used in the feature
extraction can be divided into two use cases based on the changes in the electrical signal: Steady state
(appliances with stable operation) and transient state (appliances with sudden changes). Steady-state
approaches use changes in the load transitions, and the system only works with the finite forms of
the load [46–48]. On the contrary, transient state approaches monitor the load transitions in real-time
[49]. Therefore, transient state approaches are idle for real-time feature extraction.

2.3 Load Identification

The load identification stage uses aggregated load features to identify individual appliances by
applying ML algorithms or signal processing techniques. ML models in this stage can be categorized
into supervised and unsupervised learning. The model can be selected based on the requirements
of the NILM solution, such as cost, accuracy and time. Studies show that supervised approaches
such as Deep Neural Networks (DNN) [50–54], k-Nearest Neighbor (k-NN) [55–58], Support Vector
Machine (SVM) [59–61], Decision Tree are commonly used for classification of aggregated load data.
However, a supervised NILM approach for all appliances remains questionable due to the variability
of the appliances in PowerGrid, and it requires prior training to detect appliances. For practical
settings, unsupervised learning approaches have been recommended by various studies [62–66]. Neural
Networks have significantly impacted the development of unsupervised NILM approaches [67]. CNNs
2976 CMC, 2024, vol.79, no.2

[68], Long Short-Term Memory Networks (LSTM) [69–71], Recurrent Neural Networks (RNN) [72],
Encoders [73], Graph Signal Processing [74–75] are commonly used to develop a unsupervised NILM
approaches. However, conducted research works are limited to the lab, and a solution for a large-scale
environment is unexplored.
Several research works have been done as hybrid works to improve generalizability and efficiency.

2.4 Demand Side Management to Improve Awareness

Appliance level energy consumption monitoring motivates consumers to reduce electricity usage
[76]. The article [77] reviewed the social impact of the NILM on financial savings, quality of life,
security, time management and energy saving. The study [78] developed a real-time NILM solution
that provides a web-based appliance-level consumption monitoring system. Further, the study [79]
developed a mobile application to detect anomalies.
This section discussed the stages of the NILM framework with the models and technologies used
in the domain. Furthermore, the work that has been conducted is concerned with drawbacks. The
following section will highlight the ML approaches in the context of energy management.

3 Machine Learning in the Energy Management Context

ML plays a significant role in the NILM domain to optimize energy consumption, reduce energy
bills, and improve the efficiency of the disaggregation process. This section will explore the different
machine-learning approaches used in the NILM domain, mainly for appliance detection in residential
and commercial buildings. Furthermore, it will describe the hybrid modeling approaches used in
NILM to improve performance.

3.1 Supervised Learning Approaches in NILM

The model should be trained with aggregate and disaggregate loads in supervised learning.
Different supervised approaches such as Naïve Bayes, Linear Regression, Logistic Regression, k-NN,
SVM, Random Forest and DNN are used to identify electrical appliances in the NILM context.
However, the supervised learning results lead to incomplete decisions because of the unlabeled data
in the aggregated load. To address the incomplete supervision issue, the paper [80] presented a
Multiple Instance Regression (MIR) approach, which estimates the appliance state of the unlabeled
data. The UK-DALE dataset evaluated the model’s performance with different training conditions.
Furthermore, the results show that the suggested model significantly improves compared to the
Convolutional Recurrent Neural Network and Sequence-to-point (Seq2Point) using annotated data
with unlabeled data. However, research is conducted with Type I and Type II appliances, and the
model should be trained with transfer learning techniques to make it work in real-time conditions.
Furthermore, it is required to collect data with high frequencies to obtain accurate results in supervised
learning. Therefore, solutions may not be applicable in real time because having a one-to-one training
profile for each appliance is recommended. The study [81] presented a supervised learning approach
based on SVM with a linear kernel for load identification. However, the research is limited to Type
II appliance recognition. The study [82] compared different supervised NILM algorithms for load
classification. Results show that Decision Tree (DT) performed better than k-NN, NN, Ensemble and
SVM with high-frequency data for steady-state appliance recognition. The paper [83] compared the
status prediction of home residential appliances with six supervised learning approaches: DT, Linear
Discriminant, k-NN, Quadratic Discriminant, SVM, and Ensemble. Results show that k-NN and
SVM predict the status of Type II appliances than other algorithms.
CMC, 2024, vol.79, no.2 2977

Supervised learning approaches are common in the NILM domain for load classification and
appliance detection. The selection of a supervised learning approach depends on the objective of the
research work. Furthermore, a rich aggregated load profile with a high computational environment
is required in the training phase of supervised learning. However, the real-time implementation of
supervised NILM is limited due to its cost.

3.2 Unsupervised Learning Approaches in NILM

Unsupervised learning does not require labelled data for the model’s training. In NILM,
K-means clustering, Principal Component Analysis (PCA) and Density-Based Spatial Clustering
of Applications with Noise (DBSCAN) are typical to discover the hidden patterns, structures and
clusters within the data. An unsupervised NILM approach using odd harmonic current amplitudes is
presented in the paper [84], and it uses k-means clustering to group appliances with similar electrical
patterns. The proposed approach explores the appliance based on the 3rd and 5th odd harmonics
current. Therefore, the research is limited to Type I and II appliances. To address the generalizability
and data shortage problem in the NILM domain, the study [85] introduced a self-supervised learning
approach based on Seq2Point, Bidirectional Recurrent Gated Unit (Bi-GRU) and LSTM. However,
data preprocessing should be addressed during the implementation.
Overall, unsupervised learning minimizes the training phase’s computational time and cost.
However, unexpected behavior of the appliance limits the general adoption of the model in the
study [17].

3.3 Hybrid Modeling in NILM

In NILM, hybrid modeling combines multiple approaches, such as classification and regression
models, to perform better in the training stage. Article [86] examined the feasibility of a multi-
output model for appliance identification. The model is tested with high-consumption appliances,
and generalizability is unexplored. The study [87] presented a scalable hybrid classification-regression
method, which identifies the appliance load based on the state transition of the data.
Moreover, model selection can be different based on the objectives of the NILM research
and the applicable domain. Unsupervised hybrid models are suitable for large domains such as
commercial buildings. This section highlighted the standard ML approaches used in the NILM and
their limitations. The following section will discover the open challenges of real-time implementation.

4 Open Challenges in Real-Time Implementation

This section will explain the open challenges in the real-time implementation of the NILM
solution.

4.1 Difficulty in Recognizing Multiple Loads at a Given Time

The traditional smart grid consists of different types of appliances which perform different
operating patterns. The study [88] presented that there is a shortcoming for a universal NILM
framework. To improve the efficiency of multiple loads recognition, the study [16] proposed a CNN-
based architecture based on signal analysis of real-time aggregated data. However, a large sample is
required to train the model. Further, the presented framework cannot differentiate appliances with
similar electrical loads. On the other hand, the suggested model is validated and tested with the
LIT database and UK-DALE, PLAID and REDD datasets cannot be applicable because aggregated
2978 CMC, 2024, vol.79, no.2

data is collected as the summation of watts. To overcome that issue, the paper [89] presented an
optimization-based approach with mixed-integer formulation to detect the state transition of the
appliance automatically. However, the research is limited to a low sampling load, but the work has
not been investigated in a noisy environment. In addition, the paper [90] introduced a correlation-
based feature extraction from the current waveforms to identify multiple electrical appliances in the
grid. However, the research is conducted with fewer devices in the lab. In conclusion, Difficulties in
multiple load recognition should be addressed to find a scalable solution in the NILM domain.

4.2 Cost of Running the NILM System

Studies show that consumers’ awareness of the usage can reduce energy consumption. On the
other hand, the cost of smart meters can be minimized with the NILM approach. However, a robust
computational environment is required to implement NILM systems, which consumes energy to run
the system. To address that, the paper [91] presented a cloud based NILM algorithm that provides
real-time load disaggregation. The experiment is validated and tested with two servers, a dedicated
server, and a virtual server, to evaluate the efficiency of cost modeling in two use cases. The suggested
model combines HMM-based algorithms and sliding-window functions, providing accurate results
for the large-scale aggregated data. Therefore, the processing time and noises in the network should
be addressed in the implementation. However, a reliable network is expected to transmit vast amounts
of data to the cloud, which would be expensive [92]. Therefore, a research gap exists for a cost-effective,
low-sampling NILM solution.

4.3 Lack of a Universal Framework for Appliance Detection

Different types of appliances can be used in residential and commercial buildings and suggested
NILM frameworks in the literature use different devices to validate results. They may not be applicable
in another domain. To address that, relative research works have used transfer learning methods
to learn the appliance with the existing dataset and manipulate the knowledge to another domain.
The research [93] presented two transfer learning approaches based on seq2point learning. Appliance
transfer learning: CNN layers are trained on the washing machine (Type III appliance) for other
devices. However, the results have not been presented for Type IV appliances. Cross-Domain Transfer
Learning (CTL): CNN layers re-trained on REFIT [94] and tested with UK-DALE and REDD. If the
domains are not similar, fine-tuning should be done when applying the CTL. The study [95] proposed
a transfer learning approach based on the Long Short-Term Probabilistic Neural Network (LSTM-
PNN) algorithm but has not been tested with multistate appliances. However, research works [96]
conducted with transfer learning predicted accurate results when the source and target domain have
similar or less different types of devices. Furthermore, the paper [97] stated that the lack of a dataset
with all kinds of appliances is the major issue when evaluating the performance of the transfer learning
approaches in the NILM area.
On the contrary, the study [98] suggested a pre-trained model that can train with less data to
identify the appliance. The proposed scheme quickly learns appliances and avoids negative learning.
Further, the model was investigated using the same types of appliances.
A NILM framework that can identify completely unknown appliances using aggregated data
remains challenging.
CMC, 2024, vol.79, no.2 2979

4.4 Anomaly Detection and New Appliance Identification

Overall, more research has been done assuming the collected aggregated data has fixed appliances.
But in real-time, appliances may be replaced, and the appliances may work with faults. This abnormal
behavior of the aggregated data may affect the efficiency of the appliance recognition. Submetering
is widely used to overcome that. However, installing additional meters to the power grid costs extra
money. As a solution, the study [99] presented work to detect anomalies based on the appliances’ power
distribution and participation index. Further, the consumer can obtain feedback about the devices
in the grid using the mobile application. However, the detailed description of the anomalies is not
addressed.
In practice, NILM approaches can identify unknown load features when training the model. These
novel features can be faulty in an existing or new appliance. To address that, the paper [100] proposed
adaptive NILM based on the Siamese network. It uses similarity information between training and
testing V-I trajectories. It recognizes new appliances accurately with a smaller number of training
samples. However, the Siamese network should be run simultaneously to match the features, and
computational efficiency is low.
There should be appliance signatures to train the model for the abnormal operation. Therefore,
such a dataset should be available for anomaly detection and new appliance recognition.

4.5 Complexity of the Electricity Loads

NILM research has generally been conducted using public datasets, including different residential
building appliances. Then, it is extended to commercial buildings. However, the unavailability of a
shared dataset in commercial buildings leads to addressing a NILM approach in commercial buildings.
The study [101] highlighted the practical limitations of a NILM approach in Commercial buildings:
The existence of small and similar electrical load appliances, difficulty in detecting small changes in
load consumption, and availability of Type IV appliance readings. As a solution for those limitations,
they suggested applying NILM through audits. However, adding a new appliance to the PowerGrid
may be difficult since each appliance requires submetering to learn the ground truth about the device.

4.6 Real-Time Demand Side Energy Management

The integration of energy meters in modern electrical appliance has revolutionized how consumers
interact with and manage their energy usage. These smart appliances allow consumers to track their
energy usage patterns, identify higher load consumption devices, and make informed decisions to opti-
mize energy efficiency. The convenience of accessing energy consumption data through smartphone
apps has contributed to increased consumer awareness and engagement in energy conservation efforts.
Therefore, the study [102] introduced cloud-based demand side EM platform that uses disaggregated
loads in cloud and access via smart phone or web. Further, it supports to control the energy
consumption via platform by changing the operating status of appliances. Furthermore, the paper
[103] introduced WeChat applet to control the devices remotely. The study [104] accessed database
through the Bluetooth to monitor the energy consumption of appliances using smart app. Further,
the article [105] introduced a mobile app to view cost of the energy usage. In addition to that, the
paper [106] added a feature to visualize the load behavior by changing the appliance status via app.
But adding new appliances and anomaly detection are not supported in suggested platforms. On the
other hand, load forecasting is developed in the paper [107] with appliance controlling options. but
anomaly detection is remaining a challenge.
2980 CMC, 2024, vol.79, no.2

Furthermore, Challenges in the development lead to identifying criteria for selecting an approach
for an efficient NILM framework. Table 2 highlights the available options for the selection.

Table 2: Selection measurements for the NILM implementation

Selection criteria Options
Scope Residential or commercial buildings.
Predict energy usage based on historical data or
real-time data
Understand the daily usage of energy
Purpose of the energy disaggregation
Develop an intelligent health energy platform
Minimize the cost of the implementation
Increase the efficiency of the energy
disaggregation
Early detection of faults
Fault diagnosis and anomaly detection Detection of the anomalies through the
preprocessing
Identification of faulty appliances [107]
Use of ML model Chose a model or combine several models based
on the sampling rate and appliance types
Representation of load data characterization Graphical or numerical representation [108]

This section highlights open challenges in the real-time implementation of the NILM framework.
Potential future research directions in NILM will be discussed in the next section.

5 Potential Future Research Directions in NILM

The initial version of the NILM framework focused only on identifying the status of the appliance
through disaggregation. Later, it was developed with a vast scope and various applications. This section
will discuss the future directions for a NILM framework based on the literature.
1. Update the load profile: Commercial and residential buildings have a vast diversity of appliance
and aggregate loads. Some appliances in the grid may not perform well because their life span
may be near the end. Therefore, filtering the health status of the appliance during the data
collection would be effective for a reliable load profile. Furthermore, the load dataset can
have faulty and unknown data because of the addition of new appliances and faults of the
existing appliances. Therefore, effective implementation to update the load profile would offer
productive solutions in the NILM domain.
2. Minimize the complexity of the aggregated dataset: NILM models require high-frequency load
profiles to implement disaggregation techniques. However, training the model requires a more
robust computational environment and consumes cost. As a solution, a cloud-based NILM
system can be developed. However, having a higher data transmission rate is compulsory
to maintain the communication between the server and the grid. Furthermore, a supervised
NILM system takes a few days or weeks to disaggregate the load to provide results. On
CMC, 2024, vol.79, no.2 2981

the contrary, unsupervised NILM systems are highly computationally expensive. However, a
preprocessed aggregated dataset would be helpful to minimize the issues mentioned above.
3. Develop a load profile for commercial buildings: There is a considerable diversity of appliances
used in the retail sector. Moreover, it uses Type IV appliances. The available public dataset
includes appliances used in residential buildings. In addition, commercial sector consumers
expect a NILM solution with a software-based monitoring platform. Therefore, a commercial
load profile is required to develop such a solution.
4. Recognition of similar load features: Developed NILM solutions have selected appliances with
different load signatures in the training phase. Further, the NILM is a promising research area
in energy management; however, developed solutions lack recognition of similar appliances.
5. Develop a Transfer Learning approach for a real-time solution: There is a research shortcom-
ing in developing a universal solution for energy disaggregation. In supervised learning, similar
load signatures are expected in the training and testing dataset. On the contrary, unsupervised
approaches address only a few appliance types in the literature. In addition, transfer learning
has been developed to expand the disaggregation for a real-time solution. However, the one-to-
one profile is required in the testing phase. Therefore, the one-to-one appliance profile remains
questionable in the energy management sector.
6. Develop a software-based monitoring platform: Integration of real-time energy metering in
smart appliances offers unique opportunities for advancing NILM research. The availability of
fine-tuned disaggregated loads using ML allows users to monitor and control the energy usage.
In addition to that, adding new appliances, updating the load profile and detecting anomalies
using existing profile can be included when implementing real-time smart app.
This section provided suggestions for future research in the NILM area, and the next section will
conclude the paper.

6 Conclusion
NILM is an effective energy management solution that motivates energy usage control by moni-
toring appliance behavior. Compared to Intrusive Load Monitoring, the installation and configuration
are less expensive, with a single meter attached to the power grid. Different public datasets are
available and presented as an overview in this paper to study disaggregation. However, integrating
ML models makes the load identification stage less effortless since the diversity and complexity of
the available public datasets are enormous. This paper used the current literature to introduce the
framework of NILM with comprehensive literature about the ML approaches used in the NILM
sector. Accordingly, NILM based models provides benefits in the energy disaggregation domain
due to its non-intrusive nature with cost-effectiveness deployment. Further, NILM Based models
are inherently scalable and can be deployed across a wide range of residential and commercial
environments without significant infrastructure modification. Further, it exposed various reasons
for applying ML for energy disaggregation: Automated appliance disaggregation, scalability and
efficiency, adaptability to new appliances, learning complex patterns, and handling noisy and missing
data in the load profile. However, model selection can depend on the purpose of the NILM research.
Then, this paper highlighted the challenges in the energy management sector in developing a real-time
ML approach. Therefore, achieving robust performance across different households and appliance
types remains a significant challenge for NILM-based models. Variations in appliance behavior, energy
usage patterns, and household demographics pose obstacles to generalization, requiring tailored
approaches and extensive training data to ensure reliable performance. Further, NILM algorithms
2982 CMC, 2024, vol.79, no.2

face challenges in accurately disaggregating energy consumption data, especially in environments with
multiple appliances operating simultaneously. The complex interplay between appliance signatures,
noise, and variability in energy usage patterns can hinder the accuracy of NILM-based models,
leading to misclassification errors and false positives. Further, different appliances exhibit unique
power consumption patterns, which can be further complicated by variations in usage patterns and
the addition of new devices over time. Consequently, accurately disaggregating these loads and
identifying individual appliances from aggregate power readings becomes a challenging task, requiring
sophisticated ML algorithms and data preprocessing techniques. Additionally, this paper discussed
the challenges in real-time demand-side management. Finally, it provided options for selecting a ML
criterion for the NILM research.
In the residential and commercial sectors, the appliance diversity is high in real-time. In addition,
the appliance’s operating status may not be as expected. Furthermore, new devices can be added to
the power grid anytime. Therefore, a research gap exists to develop an approach to update the load
profile accordingly. On the other hand, ML approaches require high computational conditions to train
models. In addition, the complexity of the load profile contributes to the computational requirement
for running the model. A few suggestions have been provided as future research directions to address
issues in the current literature. Monitoring the appliance’s health status helps identify faulty and
missing data in the load profile. Therefore, the load profile should be updated to maintain error-
free aggregated data. Furthermore, Implementing NILM techniques is expensive since it requires high
computational power. However, a preprocessed aggregated dataset would minimize the complexity
issue in the implementation. In addition, NILM research in the commercial sector has been limited due
to the great diversity of the appliance load profile and the unavailability of a commercial load profile.
Therefore, future research work can be expanded to design a detailed load profile for the commercial
sector. Suggestions have been provided at the end as future research directions to address the issues in
the NILM sector, such as the inability of similar load recognition and failure to learn new features.

Acknowledgement: I would like to express my heartfelt gratitude to Prof. Qi Liu for his invaluable
guidance, unwavering support, and mentorship throughout the completion of this research paper. His
expertise, encouragement, and insightful feedback played a pivotal role in shaping the direction of this
study and navigating the complexities of the submission process.

Funding Statement: The authors received no specific funding for this study.

Author Contributions: The authors confirm contribution to the paper as follows: Conceptualization,
methodology, data collection, data analysis and draft manuscript preparation: Maheesha Dhashantha;
Review and editing, supervision: Qi Liu. All authors reviewed the results and approved the final version
of the manuscript.

Availability of Data and Materials: Not applicable.

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the
present study.
CMC, 2024, vol.79, no.2 2983

References
[1] H. Ritchie, P. Rosado, and M. Roser, “Access to energy,” Our World Data, 2019.
[2] G. Szustak, P. Dąbrowski, W. Gradoń, and Ł. Szewczyk, “The relationship between energy production
and GDP: Evidence from selected European economies,” Energies, vol. 15, no. 1, pp. 50, 2021. doi:
10.3390/en15010050.
[3] M. Farghali et al., “Strategies to save energy in the context of the energy crisis: A review,” in Environmental
Chemistry Letters. Springer Science and Business Media Deutschland GmbH, Aug. 2023, vol. 21, pp.
2003–2039. doi: 10.1007/s10311-023-01591-5.
[4] H. Shayeghi, E. Shahryari, M. Moradzadeh, and P. Siano, “A survey on microgrid energy man-
agement considering flexible energy sources,” Energies, vol. 12, no. 11, pp. 2156, Jun. 2019. doi:
10.3390/en12112156.
[5] Q. Liu, K. Michael Kamoto, and X. Liu, “Microgrids-as-a-service for rural electrification in Sub-Saharan
Africa,” Comput. Mater. Contin., vol. 63, no. 3, pp. 1249–1261, Jan. 2019. doi: 10.32604/cmc.2020.05598.
[6] P. Zeng, H. Li, H. He, and S. Li, “Dynamic energy management of a microgrid using approximate dynamic
programming and deep recurrent neural network learning,” IEEE Trans. Smart Grid, vol. 10, no. 4, pp.
4435–4445, Jul. 2019. doi: 10.1109/TSG.2018.2859821.
[7] S. A. G. K. Abadi, S. I. Habibi, T. Khalili, and A. Bidram, “A model predictive control strategy for
performance improvement of hybrid energy storage systems in DC microgrids,” IEEE Access, vol. 10,
pp. 25400–25421, Mar. 2022. doi: 10.1109/ACCESS.2022.3155668.
[8] X. Liu, T. Zhao, H. Deng, P. Wang, J. Liu and F. Blaabjerg, “Microgrid energy management with energy
storage systems: A review,” CSEE J. Power Energy Syst., vol. 9, no. 2, pp. 483–504, Mar. 2023. doi:
10.17775/CSEEJPES.2022.04290.
[9] F. Wang et al., “Smart households’ aggregated capacity forecasting for load aggregators under incentive-
based demand response programs,” IEEE Trans. Ind. Appl., vol. 56, no. 2, pp. 1086–1097, Jan. 2020. doi:
10.1109/TIA.2020.2966426.
[10] C. Nalmpantis and D. Vrakas, “Machine learning approaches for non-intrusive load monitoring: From
qualitative to quantitative comparation,” Artif. Intell. Rev., vol. 52, no. 1, pp. 217–243, Jun. 2019. doi:
10.1007/s10462-018-9613-7.
[11] J. Li, M. S. Herdem, J. Nathwani, and J. Z. Wen, “Methods and applications for artificial intelligence, big
data, internet of things, and blockchain in smart energy management,” Energy AI, vol. 11, pp. 100208,
Jan. 2023. doi: 10.1016/j.egyai.2022.100208.
[12] P. Kumar and A. R. Abhyankar, “A time efficient factorial hidden Markov model-based approach for
non-intrusive load monitoring,” IEEE Trans. Smart Grid, vol. 14, no. 5, pp. 3627–3639, Feb. 2023. doi:
10.1109/TSG.2023.3245019.
[13] L. Yan, W. Tian, J. Han, and Z. Li, “eFHMM: Event-based factorial hidden Markov model for real-
time load disaggregation,” IEEE Trans. Smart Grid, vol. 13, no. 5, pp. 3844–3847, Apr. 2022. doi:
10.1109/TSG.2022.3169398.
[14] J. Chen, X. Wang, X. Zhang, and W. Zhang, “Temporal and spectral feature learning with two-stream
convolutional neural networks for appliance recognition in NILM,” IEEE Trans. Smart Grid, vol. 13, no.
1, pp. 762–772, Sep. 2021. doi: 10.1109/TSG.2021.3112341.
[15] D. H. Green, A. W. Langham, R. A. Agustin, D. W. Quinn, and S. B. Leeb, “Adaptation for automated
drift detection in electromechanical machine monitoring,” IEEE Trans. Neural Netw. Learn. Syst., vol.
34, no. 10, pp. 6768–6782, Oct. 2023. doi: 10.1109/TNNLS.2022.3184011.
[16] L. S. Nolasco, A. E. Lazzaretti, and B. M. Mulinari, “DeepDFML-NILM: A new CNN-based architec-
ture for detection, feature extraction and multi-label classification in NILM signals,” IEEE Sens. J., vol.
22, no. 1, pp. 501–509, Nov. 2021. doi: 10.1109/JSEN.2021.3127322.
[17] N. Jawdat and J. Donnal, “Unsupervised identification of electrical loads from aggregate power measure-
ments,” in 2022 IEEE Power Energy Soc. Innov. Smart Grid Technol. Conf. (ISGT), New Orleans, LA,
USA, 2022, pp. 1–6. doi: 10.1109/ISGT50606.2022.9817487.
2984 CMC, 2024, vol.79, no.2

[18] J. Z. Kolter and M. J. Johnson, “REDD: A public data set for energy disaggregation research,” Artificial
Intelligence, vol. 25, 2011.
[19] J. Gao, S. Giri, E. C. Kara, and M. Bergés, “PLAID: A public dataset of high-resoultion electrical
appliance measurements for load identification research: Demo abstract,” in Proc. 1st ACM Conf. Embed.
Syst. Energy-Effic. Build., Memphis, Tennesse, 2014, pp. 198–199. doi: 10.1145/2674061.2675032.
[20] M. Kahl, A. U. Haq, T. Kriechbaumer, and H. A. Jacobsen, “WHITED-A worldwide household and
industry transient energy data set,” in 3rd Int. Workshop on Non-Intrusive Load Monitori., 2016.
[21] J. Kelly and W. Knottenbelt, “The UK-DALE dataset, domestic appliance-level electricity demand
and whole-house demand from five UK homes,” Sci. Data, vol. 2, no. 1, pp. 150007, Mar. 2015. doi:
10.1038/sdata.2015.7.
[22] T. Picon et al., “COOLL: Controlled on/off loads library, a public dataset of high-sampled electrical signals
for appliance identification,” arXiv:1611.05803, 2016.
[23] D. P. B. Renaux et al., “A dataset for non-intrusive load monitoring: Design and implementation,”
Energies, vol. 13, no. 20, pp. 5371, Oct. 2020. doi: 10.3390/en13205371.
[24] S. Makonin, F. Popowich, L. Bartram, B. Gill, and I. V. Bajić, “AMPds: A public dataset for load
disaggregation and eco-feedback research,” in 2013 IEEE Electr. Power Energy Conf., Halifax, NS,
Canada, 2013, pp. 1–6. doi: 10.1109/EPEC.2013.6802949.
[25] S. Makonin et al., “Inspiring energy conservation through open source metering hardware and embedded
real-time load disaggregation,” in 2013 IEEE PES Asia-Pac. Power Energy Eng. Conf. (APPEEC), Hong
Kong, China, 2013, pp. 1–6. doi: 10.1109/APPEEC.2013.6837120.
[26] B. Coyne and E. Denny, “Mind the energy performance gap: Testing the accuracy of building
energy performance certificates in Ireland,” Energy Effic., vol. 14, no. 6, pp. 57, Jul. 2021. doi:
10.1007/s12053-021-09960-1.
[27] C. Beckel, W. Kleiminger, R. Cicchetti, T. Staake, and S. Santini, “The ECO data set and the performance
of non-intrusive load monitoring algorithms,” in BuildSys 2014—Proc. 1st ACM Conf. Embed. Syst.
Energy-Effic. Build., New York, NY, USA, Association for Computing Machinery, 2014, pp. 80–89. doi:
10.1145/2674061.2674064.
[28] K. D. Anderson, A. Ocneanu, D. R. Carlson, A. G. Rowe, and M. E. Berges, “BLUED: A fully labeled
public dataset for event-based non-intrusive load monitoring research,” in ACM SustKDD’12, Beijing,
China, Aug. 2012.
[29] M. Gulati, S. Ram, and A. Singh, “An in depth study into using EMI signatures for appliance identifica-
tion,” in BuildSys 2014—Proc. 1st ACM Conf. Embed. Syst. Energy-Effic. Build., New York, NY, USA,
Association for Computing Machinery, 2014. doi: 10.1145/2674061.2674070.
[30] T. Kriechbaumer and H. A. Jacobsen, “BLOND, a building-level office environment dataset of typical
electrical appliances,” Sci. Data, vol. 5, no. 1, pp. 180048, Mar. 2018. doi: 10.1038/sdata.2018.48.
[31] J. Chen et al., “Stochastic simulation of occupant-driven energy use in a bottom-up residential building
stock model,” Appl. Energy, vol. 325, no. 3, pp. 119890, Nov. 2022. doi: 10.1016/j.apenergy.2022.119890.
[32] M. Pullinger et al., “The IDEAL household energy dataset, electricity, gas, contextual sensor data and sur-
vey data for 255 UK homes,” Sci. Data, vol. 8, no. 1, pp. 146, May 2021. doi: 10.1038/s41597-021-00921-y.
[33] S. Makonin, “HUE: The hourly usage of energy dataset for buildings in British Columbia,” Data Brief,
vol. 23, no. 1, pp. 103744, Apr. 2019. doi: 10.1016/j.dib.2019.103744.
[34] C. Shin, E. Lee, J. Han, J. Yim, W. Rhee and H. Lee, “The ENERTALK dataset, 15 Hz electric-
ity consumption data from 22 houses in Korea,” Sci. Data, vol. 6, no. 1, pp. 193, Oct. 2019. doi:
10.1038/s41597-019-0212-5.
[35] C. Klemenjak, C. Kovatsch, M. Herold, and W. Elmenreich, “A synthetic energy dataset for
non-intrusive load monitoring in households,” Sci. Data, vol. 7, no. 1, pp. 108, Apr. 2020. doi:
10.1038/s41597-020-0434-6.
[36] H. K. Iqbal, F. H. Malik, A. Muhammad, M. A. Qureshi, M. N. Abbasi and A. R. Chishti, “A critical
review of state-of-the-art non-intrusive load monitoring datasets,” Elect. Power Syst. Res., vol. 192, no.
12, pp. 106921, Mar. 2021. doi: 10.1016/j.epsr.2020.106921.
CMC, 2024, vol.79, no.2 2985

[37] G. W. Hart, “Nonintrusive appliance load monitoring,” Proc. IEEE, vol. 80, no. 12, pp. 1870–1891, Dec.
1992. doi: 10.1109/5.192069.
[38] Q. Liu, K. M. Kamoto, X. Liu, M. Sun, and N. Linge, “Low-complexity non-intrusive load monitoring
using unsupervised learning and generalized appliance models,” IEEE Trans. Consum. Electron., vol. 65,
no. 1, pp. 28–37, Feb. 2019. doi: 10.1109/TCE.2019.2891160.
[39] K. K. Kee, Y. S. Lim, J. Wong, and K. H. Chua, “Non-Intrusive Load Monitoring (NILM)–A recent
review with cloud computing,” in 2019 IEEE Int. Conf. Smart Instrum., Meas. Appl. (ICSIMA), Kuala
Lumpur, Malaysia, 2019, pp. 1–6. doi: 10.1109/ICSIMA47653.2019.9057316.
[40] A. E. Ruano, Á. Hernández, J. Ureña, M. D. G. Ruano, and J. J. García, “NILM techniques for intelligent
home energy management and ambient assisted living: A review,” Energies, vol. 12, no. 11, pp. 2203, 2019.
doi: 10.3390/en12112203.
[41] M. A. Ahajjam, C. Essayeh, M. Ghogho, and A. Kobbane, “On multi-label classification for non-intrusive
load identification using low sampling frequency datasets,” in 2021 IEEE Int. Instrum. Meas. Technol.
Conf. (I2MTC), Glasgow, UK, 2021, pp. 1–6. doi: 10.1109/I2MTC50364.2021.9460059.
[42] L. Zhang and L. Zhu, “Basic summary of non-intrusive load monitoring,” in 2019 IEEE 10th Int. Conf.
Softw. Eng. Serv. Sci. (ICSESS), Beijing, China, 2019, pp. 1–5. doi: 10.1109/ICSESS47205.2019.9040726.
[43] S. Steephen, R. Sheeba, and N. Naufal, “A review on non-intrusive load monitoring using deep learning,”
in 2022 Int. Conf. Innov. Sci. Technol. Sustain. Dev. (ICISTSD), Kollam, India, 2022, pp. 381–385. doi:
10.1109/ICISTSD55159.2022.10010467.
[44] S. Ronaghi, A. Ferrero, S. Salicone, and H. V. Jetti, “Novel algorithms for filtering and event detection
in non-intrusive load monitoring,” in 2023 IEEE 13th Int. Workshop on Appli. Measurem. Power Syst.
(AMPS), Bern, Switzerland, IEEE, 2023. doi: 10.1109/AMPS59207.2023.10297257.
[45] S. Sykiotis, M. Kaselimi, A. Doulamis, and N. Doulamis, “ELECTRIcity: An efficient transformer for
non-intrusive load monitoring,” Sensors, vol. 22, no. 8, pp. 2926, Apr. 2022. doi: 10.3390/s22082926.
[46] C. Wang et al., “A switching event detection method for household appliances based on improved sliding
window,” in 2023 6th Int. Conf. Energy, Electr. Power Eng.(CEEPE), Guangzhou, China, 2023, pp. 683–
689. doi: 10.1109/CEEPE58418.2023.10167223.
[47] A. U. Rehman, T. T. Lie, B. Vallès, and S. R. Tito, “Event-detection algorithms for low sampling
nonintrusive load monitoring systems based on low complexity statistical features,” IEEE Trans. Instrum.
Meas., vol. 69, no. 3, pp. 751–759, Mar. 2020. doi: 10.1109/TIM.2019.2904351.
[48] B. Liu, F. Chang, W. Luan, and B. Zhao, “Improved appliance transient feature extraction via template
matching,” in 2023 IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), Jun. 2023, pp. 4–10. doi:
10.1109/ICASSP49357.2023.10094638.
[49] H. H. Chang, H. J. Chang, and C. L. Lin, “Novel high-frequency transient techniques in non-intrusive
load monitoring techniques,” in 2023 8th Asia Conf. Power Electr. Eng. (ACPEE), Tianjin, China, 2023,
pp. 1533–1538. doi: 10.1109/ACPEE56931.2023.10135761.
[50] H. Çimen, N. Çetinkaya, J. C. Vasquez, and J. M. Guerrero, “A microgrid energy management system
based on non-intrusive load monitoring via multitask learning,” IEEE Trans. Smart Grid, vol. 12, no. 2,
pp. 977–987, Mar. 2021. doi: 10.1109/TSG.2020.3027491.
[51] Y. Zhang, G. Tang, Q. Huang, Y. Wang, H. Xu and X. Wang, “More behind your electricity bill: A dual-
DNN approach to non-intrusive load monitoring,” in 2022 IEEE Int. Conf. Internet Things (iThings),
IEEE Green Sustain. Dev. (GreenCom), IEEE Cyber, Phys. Soc. Comput. (CPSCom), IEEE Smart
Data (SmartData), IEEE Congr. Cybermatics (Cybermatics), Espoo, Finland, 2022, pp. 292–299. doi:
10.1109/iThings-GreenCom-CPSCom-SmartData-Cybermatics55523.2022.00023.
[52] H. Xiong, J. Chen, S. Lan, and Y. Zhang, “Research on industrial load decomposition based on correlation
analysis and DNN,” in IEEE 5th Int.Conf. Power, Intelligent Comput. Syst. (ICPICS), Shenyang, China,
2023, pp. 730–735. doi: 10.1109/ICPICS58376.2023.10235433.
[53] C. Mariño, E. Masquil, F. Marchesoni, A. Fernandez, and P. Massaferro, “NILM: Multivariate DNN
performance analysis with high frequency features,” in 2021 IEEE PES Innov. Smart Grid Technol.
2986 CMC, 2024, vol.79, no.2

Conf.—Latin America (ISGT Latin America), Lima, Peru, 2021, pp. 1–5. doi: 10.1109/ISGTLatinAmer-
ica52371.2021.9543016.
[54] Y. Jiang and M. Liu, “Design of a reliable algorithmic with deep learning and transfer learning for load
combination recognition,” in 2021 8th Int. Conf. Dependable Syst. Their Appl. (DSA), Yinchuan, China,
2021, pp. 541–550. doi: 10.1109/DSA52907.2021.00080.
[55] N. C. Yang and K. L. Sung, “Non-intrusive load classification and recognition using soft-voting ensemble
learning algorithm with decision tree, k-nearest neighbor algorithm and multilayer perceptron,” IEEE
Access, vol. 11, pp. 94506–94520, Jan. 2023. doi: 10.1109/ACCESS.2023.3311641.
[56] F. B. Gurbuz, R. Bayindir, and S. Vadi, “Comprehensive non-intrusive load monitoring process: Device
event detection, device feature extraction and device identification using KNN, random forest and decision
tree,” in 2021 10th Int. Conf. Renewable Energy Res. Appl. (ICRERA), Istanbul, Turkey, 2021, pp. 447–
452. doi: 10.1109/ICRERA52334.2021.9598679.
[57] D. A. M. Lemes, T. W. Cabral, G. Fraidenraich, L. G. P. Meloni, E. R. de Lima and F. B. Neto, “Load
disaggregation based on time window for HEMS application,” IEEE Access, vol. 9, pp. 70746–70757, May
2021. doi: 10.1109/ACCESS.2021.3078340.
[58] A. F. M. Jaramillo et al., “Distributed energy resources electric profile identification in low voltage
networks using supervised machine learning techniques,” IEEE Access, vol. 11, no. 2, pp. 19469–19486,
Feb. 2023. doi: 10.1109/ACCESS.2023.3247977.
[59] P. Franco, J. M. Martínez, Y. C. Kim, and M. A. Ahmed, “IoT based approach for load monitoring
and activity recognition in smart homes,” IEEE Access, vol. 9, pp. 45325–45339, Mar. 2021. doi:
10.1109/ACCESS.2021.3067029.
[60] Musrinah, M. A. Murti, and F. Budiman, “Identification of household electric load based on harmonic
parameters using SVM,” in 2023 IEEE Int. Conf. Ind. 4.0, Artif. Intell., Commun. Technol, (IAICT), Bali,
Indonesia, 2023, pp. 196–200. doi: 10.1109/IAICT59002.2023.10205529.
[61] A. U. Rehman, T. T. Lie, B. Vallès, and S. R. Tito, “Comparative evaluation of machine learning models
and input feature space for non-intrusive load monitoring,” J. Mod. Power Syst. Clean Energy, vol. 9, no.
5, pp. 1161–1171, Sep. 2021. doi: 10.35833/MPCE.2020.000741.
[62] Z. Zhou, Y. Xiang, H. Xu, Y. Wang, and D. Shi, “Unsupervised learning for non-intrusive load monitoring
in smart grid based on spiking deep neural network,” J. Mod. Power Syst. Clean Energy, vol. 10, no. 3, pp.
606–616, May 2022. doi: 10.35833/MPCE.2020.000569.
[63] Y. Zhang et al., “FedNILM: Applying federated learning to NILM applications at the edge,” IEEE Trans.
Green Commun. Netw., vol. 7, no. 2, pp. 857–868, Jun. 2023. doi: 10.1109/TGCN.2022.3167392.
[64] B. Liu, W. Luan, J. Yang, and Y. Yu, “The balanced window-based load event optimal match-
ing for NILM,” IEEE Trans. Smart Grid, vol. 13, no. 6, pp. 4690–4703, Nov. 2022. doi:
10.1109/TSG.2022.3179440.
[65] B. Liu, J. Zheng, W. Luan, and Z. Liu, “Appliance power pattern mining via motif discovery in
unsupervised NILM,” in 2021 IEEE 5th Conf. Energy Internet Energy Syst. Integr. (EI2), Taiyuan, China,
2021, pp. 3172–3177. doi: 10.1109/EI252483.2021.9713097.
[66] A. Majumdar, “Trainingless energy disaggregation without plug-Level sensing,” IEEE Trans. Instrum.
Meas., vol. 71, pp. 1–8, Feb. 2022. doi: 10.1109/TIM.2022.3152311.
[67] M. Irani Azad, R. Rajabi, and A. Estebsari, “Nonintrusive load monitoring (NILM) using a deep learning
model with a transformer-based attention mechanism and temporal pooling,” Electron., vol. 13, no. 2, pp.
407, Jan. 2024. doi: 10.3390/electronics13020407.
[68] D. Yang, X. Gao, L. Kong, Y. Pang, and B. Zhou, “An event-driven convolutional neural architecture for
non-intrusive load monitoring of residential appliance,” IEEE Trans. Consum. Electron., vol. 66, no. 2,
pp. 173–182, May 2020. doi: 10.1109/TCE.2020.2977964.
[69] M. Kaselimi, N. Doulamis, A. Voulodimos, E. Protopapadakis, and A. Doulamis, “Context aware energy
disaggregation using adaptive bidirectional LSTM models,” IEEE Trans. Smart Grid, vol. 11, no. 4, pp.
3054–3067, Jul. 2020. doi: 10.1109/TSG.2020.2974347.
CMC, 2024, vol.79, no.2 2987

[70] Y. Belguermi, P. Wira, and G. Hermann, “A multi-output LSTM-CNN learning scheme for power
disaggregation within a NILM framework,” in 2023 IEEE 21st Int. Conf. Ind. Inform. (INDIN), Lemgo,
Germany, 2023, pp. 1–6. doi: 10.1109/INDIN51400.2023.10217936.
[71] H. Hwang and S. Kang, “Nonintrusive load monitoring using an LSTM with feedback structure,” IEEE
Trans. Instrum. Meas., vol. 71, no. 1, pp. 1–11, Apr. 2022. doi: 10.1109/TIM.2022.3169536.
[72] J. He et al., “InFocus: Amplifying critical feature influence on non-intrusive load monitoring through
self-attention mechanisms,” IEEE Trans. Smart Grid, vol. 14, no. 5, pp. 3828–3840, Sep. 2023. doi:
10.1109/TSG.2023.3236792.
[73] H. Çimen, Y. Wu, Y. P. Wu, Y. Terriche, J. C. Vasquez and J. M. Guerrero, “Deep learning-based
probabilistic autoencoder for residential energy disaggregation: An adversarial approach,” IEEE Trans.
Ind. Inform., vol. 18, no. 12, pp. 8399–8408, Dec. 2022. doi: 10.1109/TII.2022.3150334.
[74] R. Feng, W. Yuan, L. Ge, and S. Ji, “Nonintrusive load disaggregation for residential users based on
alternating optimization and downsampling,” IEEE Trans. Instrum. Meas., vol. 70, pp. 1–12, Oct. 2021.
doi: 10.1109/TIM.2021.3117068.
[75] J. Yuan, H. Wang, P. Wu, C. Wang, J. Chen and H. Jiao, “A novel current signal feature and its
application in noninvasive load monitoring,” IEEE Trans. Instrum. Meas., vol. 70, pp. 1–10, Nov. 2020.
doi: 10.1109/TIM.2020.3036651.
[76] R. Ramadan, Q. Huang, and O. Bamisile, “Non-intrusive load monitoring for household energy disaggre-
gation: A state-of-the-art,” in 2022 IEEE 6th Conf. Energy Internet Energy Syst. Integr. (EI2), Chengdu,
China, 2022, pp. 1397–1403. doi: 10.1109/EI256261.2022.10116809.
[77] F. B. Gurbuz, R. Bayindir, and H. I. Bulbul, “A brief review of non-intrusive load monitoring and its
impact on social life,” in 2021 9th Int. Conf. Smart Grid (icSmartGrid), Setubal, Portugal, 2021, pp. 289–
294. doi: 10.1109/icSmartGrid52357.2021.9551258.
[78] M. T. Ahammed et al., “Real-time non-intrusive electrical load classification over iot using machine
learning,” IEEE Access, vol. 9, pp. 115053–115067, Aug. 2021. doi: 10.1109/ACCESS.2021.3104263.
[79] E. Azizi, M. T. H. Beheshti, and S. Bolouki, “Appliance-level anomaly detection in nonintrusive load
monitoring via power consumption-based feature analysis,” IEEE Trans. Consum. Electron., vol. 67, no.
4, pp. 363–371, Nov. 2021. doi: 10.1109/TCE.2021.3129356.
[80] L. Serafini, G. Tanoni, E. Principi, S. Spinsante, and S. Squartini, “A multiple instance regression
approach to electrical load disaggregation,” in 2022 30th Eur. Signal Process. Conf. (EUSIPCO),
Belgrade, Serbia, 2022, pp. 1666–1670. doi: 10.23919/EUSIPCO55093.2022.9909747.
[81] N. Madhushan, N. Dharmaweera, and U. Wijewardhana, “Supervised non-intrusive load monitor-
ing algorithm for identifying different operating states of type-II residential appliances,” in 2022
Moratuwa Eng. Res. Conf. (MERCon), Moratuwa, Sri Lanka, 2022, pp. 1–6. doi: 10.1109/MER-
Con55799.2022.9906271.
[82] O. Al-Khadher, A. Mukhtaruddin, F. R. Hashim, M. M. Azizan, H. Mamat and M. Mani, “Comparison
of non-intrusive load monitoring supervised methods using harmonics as feature,” in 2022 Int. Conf.
Electr., Comput. Energy Tech. (ICECET), Prague, Czech Republic, 2022, pp. 1–6. doi: 10.1109/ICE-
CET55527.2022.9872689.
[83] M. Z. Abbas, I. A. Sajjad, R. Liaqat, T. Khursheed, M. Awais and M. S. Javed, “Non-intrusive status
prediction of residential appliances using supervised learning techniques,” in 2023 6th Int. Conf. Energy
Conserv. Eff. (ICECE), Lahore, Pakistan, 2023, pp. 1–6. doi: 10.1109/ICECE58062.2023.10092511.
[84] P. G. Papageorgiou, P. A. Gkaidatzis, G. C. Christoforidis, and A. S. Bouhouras, “Unsupervised NILM
implementation using odd harmonic currents,” in 2021 56th Int. Universities Power Eng. Conf. (UPEC),
Middlesbrough, UK, 2021, pp. 1–6. doi: 10.1109/UPEC50034.2021.9548250.
[85] S. Chen, B. Zhao, M. Zhong, W. Luan, and Y. Yu, “Nonintrusive load monitoring based on self-supervised
learning,” IEEE Trans. Instrum. Meas., vol. 72, no. 4, pp. 1–13, Oct. 2023. doi: 10.1109/TIM.2023.3246504.
[86] D. Precioso and D. Gómez-Ullate, “Non-intrusive load monitoring using multi-output CNNs,” in 2021
IEEE Madrid PowerTech, Madrid, Spain, 2021, pp. 1–6. doi: 10.1109/PowerTech46648.2021.9494943.
2988 CMC, 2024, vol.79, no.2

[87] G. Saraswat, B. Lundstrom, and M. V. Salapaka, “Scalable hybrid classification-regression solution

for high-frequency nonintrusive load monitoring,” in 2023 IEEE Power Energy Soc. Innov. Smart Grid
Technol. Conf. (ISGT), Jan. 2023, pp. 1–5. doi: 10.1109/ISGT51731.2023.10066447.
[88] M. Dhashantha Silva, Q. Liu, and O. Famous Darteh, “A recent review of NILM framework:
Development and challenges,” in 2022 IEEE Int. Conf. Depend., Autonom. Secur. Comput., Int. Conf.
Pervasive Intell. Comput., Int. Conf. Cloud Big Data Comput., Int. Conf. Cyber Sci. Technol. Congr.
(DASC/PiCom/CBDCom/CyberSciTech), Falerna, Italy, 2022, pp. 1–7. doi: 10.1109/DASC/PiCom/CB-
DCom/Cy55231.2022.9927902.
[89] M. Balletti, V. Piccialli, and A. M. Sudoso, “Mixed-integer nonlinear programming for state-based
non-intrusive load monitoring,” IEEE Trans. Smart Grid, vol. 13, no. 4, pp. 3301–3314, Jul. 2022. doi:
10.1109/TSG.2022.3152147.
[90] S. Ghosh, D. K. Panda, S. Das, and D. Chatterjee, “Cross-correlation based classification of electrical
appliances for non-intrusive load monitoring,” in 2021 Int. Conf. Sustain. Energy Future Electric Transp.
(SEFET), Hyderabad, India, 2021, pp. 1–6. doi: 10.1109/SeFet48154.2021.9375687.
[91] M. W. Asres, L. Ardito, and E. Patti, “Computational cost analysis and data-driven predictive modeling
of cloud-based online-NILM algorithm,” IEEE Trans. Cloud Comput., vol. 10, no. 4, pp. 2409–2423, Jan.
2021. doi: 10.1109/TCC.2021.3051766.
[92] G. Rajendiran, “Non-intrusive load monitoring: A promising path to the society for responsible energy
utilization and sustainability,” in Proc. 13th ACM Int. Conf. Future Energy Syst., New York, NY, USA,
2022, pp. 594–596. doi: 10.1145/3538637.3539635.
[93] M. D’Incecco, S. Squartini, and M. Zhong, “Transfer learning for non-intrusive load monitoring,” IEEE
Trans. Smart Grid, vol. 11, no. 2, pp. 1419–1429, Aug. 2019. doi: 10.1109/TSG.2019.2938068.
[94] D. Murray, L. Stankovic, and V. Stankovic, “An electrical load measurements dataset of United Kingdom
households from a two-year longitudinal study,” Sci. Data, vol. 4, Jan. 2017. doi: 10.1038/sdata.2016.122.
[95] Z. Zhou, Y. Xiang, H. Xu, Z. Yi, D. Shi and Z. Wang, “A novel transfer learning-based intelligent
nonintrusive load-monitoring with limited measurements,” IEEE Trans. Instrum. Meas., vol. 70, pp. 1–8,
Jul. 2020. doi: 10.1109/TIM.2020.3011335.
[96] J. Lin, J. Ma, J. Zhu, and H. Liang, “Deep domain adaptation for non-intrusive load monitoring based on
a knowledge transfer learning network,” IEEE Trans. Smart Grid, vol. 13, no. 1, pp. 280–292, Sep. 2021.
doi: 10.1109/TSG.2021.3115910.
[97] X. Chang et al., “Transferable tree-based ensemble model for non-intrusive load monitoring,” IEEE Trans.
Sustain. Comput., vol. 7, no. 4, pp. 970–981, May 2022. doi: 10.1109/TSUSC.2022.3175941.
[98] L. Wang, S. Mao, B. M. Wilamowski, and R. M. Nelms, “Pre-trained models for non-intrusive appli-
ance load monitoring,” IEEE Trans. Green Commun. Netw., vol. 6, no. 1, pp. 56–68, Mar. 2022. doi:
10.1109/TGCN.2021.3087702.
[99] A. Aboulian et al., “NILM dashboard: A power system monitor for electromechanical equip-
ment diagnostics,” IEEE Trans. Ind. Inform., vol. 15, no. 3, pp. 1405–1414, Mar. 2019. doi:
10.1109/TII.2018.2843770.
[100] M. Yu, B. Wang, L. Lu, Z. Bao, and D. Qi, “Non-intrusive adaptive load identification based on siamese
network,” IEEE Access, vol. 10, pp. 11564–11573, Jan. 2022. doi: 10.1109/ACCESS.2022.3145982.
[101] A. Meier and D. Cautley, “Practical limits to the use of non-intrusive load monitoring in commercial
buildings,” Energy Build., vol. 251, no. 12, pp. 111308, Nov. 2021. doi: 10.1016/j.enbuild.2021.111308.
[102] G. A. Raiker et al., “Internet of things based demand side energy management system using non-intrusive
load monitoring,” in 2020 IEEE Int. Conf. Power Electron., Smart Grid Renew. Energy (PESGRE2020),
Cochin, India, 2020, pp. 1–5. doi: 10.1109/PESGRE45664.2020.9070739.
[103] H. Xiao, W. Hu, D. Zhou, Q. Deng, and H. Zhou, “Remote monitoring and control system for domestic
electricity consumption based on WeChat applet,” in 2019 Chinese Control Conf. (CCC), Guangzhou,
China, 2019, pp. 5321–5326. doi: 10.23919/ChiCC.2019.8866512.
CMC, 2024, vol.79, no.2 2989

[104] G. A. Raiker, S. B. Reddy, L. Umanand, A. Yadav, and M. M. Shaikh, “Approach to non-intrusive load
monitoring using factorial hidden markov model,” in 2018 IEEE 13th Int. Conf. Ind. Inf. Syst. (ICIIS),
Rupnagar, India, 2018, pp. 381–386. doi: 10.1109/ICIINFS.2018.8721436.
[105] M. Oinonen, O. Gaus, T. Pereira, A. Walia, and W. G. Morsi, “Non-intrusive load monitoring using
machine learning accelerator hardware for smart meters,” in 2022 IEEE Can. Conf. Electrical Comput.
Eng. (CCECE), Halifax, NS, Canada, 2022, pp. 483–488. doi: 10.1109/CCECE49351.2022.9918403.
[106] K. Dowalla, P. Bilski, R. Łukaszewski, A. Wójcik, and R. Kowalik, “NILM application for real
time monitoring of appliances energy consumption used,” in 2021 11th IEEE Int. Conf. Intell. Data
Acquisit. Adv. Comput. Syst.: Technol. Appl. (IDAACS), Cracow, Poland, 2021, pp. 1022–1025. doi:
10.1109/IDAACS53288.2021.9660389.
[107] H. Rashid, P. Singh, V. Stankovic, and L. Stankovic, “Can non-intrusive load monitoring be used for
identifying an appliance’s anomalous behaviour?,” Appl. Energy, vol. 238, no. 1, pp. 796–805, Mar. 2019.
doi: 10.1016/j.apenergy.2019.01.061.
[108] R. Liaqat, M. I. A. Sajjad, M. Waseem, and H. Haes Alhelou, “Appliance level energy characterization
of residential electricity demand: Prospects, challenges and recommendations,” IEEE Access, vol. 9, pp.
148676–148697, Oct. 2021. doi: 10.1109/ACCESS.2021.3123196.