0% found this document useful (0 votes)

6 views

Towards_Feasible_Solutions_for_Load_Monitoring_in_

Uploaded by

habibi

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

6 views

Towards_Feasible_Solutions_for_Load_Monitoring_in_

Uploaded by

habibi

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 24

sensors

Article
Towards Feasible Solutions for Load Monitoring in
Quebec Residences †
Sayed Saeed Hosseini 1, * , Benoit Delcroix 2 , Nilson Henao 1 , Kodjo Agbossou 1 and Sousso Kelouwani 3

1 Department of Electrical and Computer Engineering, Laboratoire d’innovation et de Recherche en Énergie

Intelligent, Hydrogen Research Institute, Université du Québec à Trois-Rivières,
Trois-Rivières, QC G8Z 4M3, Canada; [email protected] (N.H.); [email protected] (K.A.)
2 Hydro-Quebec Research Institute, Laboratoire des Technologies de l’énergie d’Hydro-Québec,
Shawinigan, QC G9N 0C5, Canada; [email protected]
3 Department of Mechanical Engineering, Laboratoire d’innovation et de Recherche en Énergie Intelligent,
Hydrogen Research Institute, Université du Québec à Trois-Rivières, Trois-Rivières, QC G8Z 4M3, Canada;
[email protected]
* Correspondence: [email protected]
† This paper is an extended version of our conference proceedings contribution, “A case study on obstacles to
feasible NILM solutions for energy disaggregation in Quebec residences”, presented at the 9th ACM
International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation, Boston, MA,
USA, 9–10 November 2022, https://doi.org/10.1145/3563357.3566151.

Abstract: For many years, energy monitoring at the most disaggregate level has been mainly sought
through the idea of Non-Intrusive Load Monitoring (NILM). Developing a practical application of
this concept in the residential sector can be impeded by the technical characteristics of case studies.
Accordingly, several databases, mainly from Europe and the US, have been publicly released to
enable basic research to address NILM issues raised by their challenging features. Nevertheless,
the resultant enhancements are limited to the properties of these datasets. Such a restriction has
caused NILM studies to overlook residential scenarios related to geographically-specific regions and
existent practices to face unexplored situations. This paper presents applied research on NILM in
Quebec residences to reveal its barriers to feasible implementations. It commences with a concise
discussion about a successful NILM idea to highlight its essential requirements. Afterward, it
Citation: Hosseini, S.S.; Delcroix, B.; provides a comparative statistical analysis to represent the specificity of the case study by exploiting
Henao, N.; Agbossou, K.; Kelouwani, real data. Subsequently, this study proposes a combinatory approach to load identification that
S. Towards Feasible Solutions for utilizes the promise of sub-meter smart technologies and integrates the intrusive aspect of load
Load Monitoring in Quebec monitoring with the non-intrusive one to alleviate NILM difficulties in Quebec residences. A load
Residences. Sensors 2023, 23, 7288. disaggregation technique is suggested to manifest these complications based on supervised and
https://doi.org/10.3390/s23167288 unsupervised machine learning designs. The former is aimed at extracting overall heating demand
Academic Editor: Anastasios from the aggregate one while the latter is designed for disaggregating the residual load. The results
Doulamis demonstrate that geographically-dependent cases create electricity consumption scenarios that can
deteriorate the performance of existing NILM methods. From a realistic standpoint, this research
Received: 18 July 2023
elaborates on critical remarks to realize viable NILM systems, particularly in Quebec houses.
Revised: 8 August 2023
Accepted: 13 August 2023
Keywords: Non-Intrusive Load Monitoring (NILM); low-sampling load disaggregation; statistical
Published: 21 August 2023
analysis; machine learning algorithms; electric baseboards; electric water heaters

Copyright: © 2023 by the authors.

Licensee MDPI, Basel, Switzerland. 1. Introduction
This article is an open access article Home Energy Management Systems (HEMSs) are the key enabler of the residential
distributed under the terms and
demand-side role in power system flexibility towards the energy transition [1]. By mon-
conditions of the Creative Commons
itoring household energy consumption, HEMSs can provide consumers and the utility
Attribution (CC BY) license (https://
with energy-saving opportunities through control, scheduling, feedback, and recommenda-
creativecommons.org/licenses/by/
tions [2]. Accordingly, this service is an essential prerequisite of HEMSs to observe electrical
4.0/).

Sensors 2023, 23, 7288. https://doi.org/10.3390/s23167288 https://www.mdpi.com/journal/sensors

Sensors 2023, 1, 0 2 of 24
Sensors 2023, 23, 7288 2 of 24

load and actuate energy management strategies. Load monitoring can track energy use at
load and actuate energy management strategies. Load monitoring can track energy use at
different electrical layers comprising household level (main entry), circuit level (electrical
different electrical layers comprising household level (main entry), circuit level (electrical
panel), plug level, and/or appliance level [3]. From the viewpoint of potential applications,
panel), plug level, and/or appliance level [3]. From the viewpoint of potential applications,
information about energy usage behavior at the most disaggregate level, i.e., the bottom
information about energy usage behavior at the most disaggregate level, i.e., the bottom
layer, is preferred. From a technical point of view, exploiting only the aggregate data mea-
layer, is preferred. From a technical point of view, exploiting only the aggregate data mea-
sured at the entry point, i.e., the top layer, is suggested [4]. These two standpoints stimulate
sured at the entry point, i.e., the top layer, is suggested [4]. These two standpoints stimulate
the non-intrusive approach to load monitoring that identifies operating appliances and
the non-intrusive approach to load monitoring that identifies operating appliances and
estimates their individual energy consumption in the aggregate signal. Non-Intrusive Load
estimates their individual energy consumption in the aggregate signal. Non-Intrusive Load
Monitoring (NILM) emerges as an alternative to troublesome intrusive practices (ILM) de-
Monitoring (NILM) emerges as an alternative to troublesome intrusive practices (ILM) de-
pending on sensing systems and sub-meter measurements located after the main entry [5].
pending on sensing systems and sub-meter measurements located after the main entry [5].
It can provide viable assistance to HEMSs. NILM can characterize customers based on
It can provide viable assistance to HEMSs. NILM can characterize customers based on
their appliance usage patterns, describe them according to their tendency towards utilizing
their appliance usage patterns, describe them according to their tendency towards utilizing
their devices, and identify those with a higher propensity for participating in power system
their devices, and identify those with a higher propensity for participating in power system
operation [6]. Consequently, it can facilitate recognizing the topmost end-users for energy
operation [6]. Consequently, it can facilitate recognizing the topmost end-users for energy
efficiency services, Demand Response (DR) programs, and electrification. Research studies
efficiency services, Demand Response (DR) programs, and electrification. Research studies
show that NILM can result in a significant energy saving of up to 15% by giving detailed
show that NILM can result in a significant energy saving of up to 15% by giving detailed
energy feedback to consumers [7]. Figure 1 explains intrusive and non-intrusive means to
energy feedback to consumers [7]. Figure 1 explains intrusive and non-intrusive means to
execute load monitoring at different layers of household electric circuits [8].
execute load monitoring at different layers of household electric circuits [8].

Smart meter

Cloud based
Edge-device based

Household level Data flow

ILM NILM
Circuit level Sensor
Plug level Smart plug

A simple
Figure 1.1. A
Figure simple representation
representation of
of intrusive
intrusive and
and non-intrusive
non-intrusive approaches
approaches to
to household
household load
load
monitoringand
monitoring andtheir
theirtechnical
technicalmeans
means[8].
[8].

Motivation
Motivation
Fromaafeasible
From feasibleviewpoint,
viewpoint,smartsmartmeters
metersare
arethethemodern
modernmetering
meteringinfrastructures
infrastructuresto to
enable cost-effective energy-saving solutions through NILM in the
enable cost-effective energy-saving solutions through NILM in the residential sector [1]. residential sector [1].
These emergent
These emergent technologies
technologies stimulate
stimulate NILM
NILM operations
operations adopted
adopted for for data
data acquisition
acquisition
systems with
systems with aa low
low sampling
sampling raterate due
due to
to their
their technical
technical limitations
limitations [9].[9]. In
In this
this context,
context,
onlysteady-state
only steady-statecharacteristics
characteristicsof ofpower
powerdemand
demandcan canbebeexploited
exploitedfor forload
loaddisaggregation.
disaggregation.
These features
These features mostly
mostly belong
belong to to active
active power
power since
since supplying
supplying otherother readings
readings increases
increases
theprice
the priceofofsmart
smartmeters.
meters. For
Forexample,
example,monitoring
monitoringreactive
reactivepower,
power,in inaddition,
addition,can canraise
raise
expenses by 50% [10]. The literature has a similar trend in providing
expenses by 50% [10]. The literature has a similar trend in providing the information space,the information space,
thussupporting
thus supportingaarealistic
realisticconcept.
concept. However,
However,ititdraws
drawson ondata
datawith
withfine
finegranularity
granularityinduced
induced
bythe
by theelectrical
electricalproperties
propertiesof ofpublic
publicdatabases.
databases. The
Thesampling
samplingfrequency
frequencyof ofregular
regulardatasets
datasets
on which
on which relevant
relevant research
research relies
relies can
can be
be acknowledged
acknowledged as as high
high for
for actual
actual measurement
measurement
apparatus [11].
apparatus [11]. Regardless
Regardless of of the
the mathematical
mathematical model,
model, the
the most
most common
common databases
databases forfor
NILM practices
NILM practices contain
contain measurements
measurements with with aa sampling
sampling rate rate of
of ≥ ≥11 Hz
Hz [12].
[12]. This
This fact
fact
isis confirmed
confirmed by by the
the four
four most
most cited
cited datasets,
datasets, i.e.,
i.e., REDD,
REDD, UMass
UMass Smart,
Smart, UK-DALE,
UK-DALE, and and
BLUED [13]. Furthermore, among the ten most cited ones, only
BLUED [13]. Furthermore, among the ten most cited ones, only two have a sampling time two have a sampling time
of≥
of ≥11min,
min,i.e.,
i.e.,AMPDs
AMPDsand andHES,
HES,as asshown
shownin inTable
Table11[1,11].
[1,11]. The
Thepublic
publicbodybodyof ofinformation
information
can also
can also influence
influence other
other NILM
NILM exercises.
exercises. For
For example,
example, aa technical
technical reason
reason forfor suggesting
suggesting
event-less methods can be that most databases do not provide a means for the analysis of the
Sensors 2023, 23, 7288 3 of 24

event detection phase [14]. On the other side, the NILM evaluation framework, specifically
Deep Learning (DL)-based analyses as the state-of-the-art, has mainly focused on the
energy estimation and load reconstruction of fridges, kettles, microwaves, dishwashers,
and washing machines as the targeted loads [15–19]. Among these appliances, only the
last two have sensible flexibility potentials considering their controllability and usage,
regardless of their identification precision, as another challenge [20]. The first three devices
are essentially uncontrollable since their manipulation can jeopardize customers’ activities.
In addition, the set of intended loads neglect equipment like Air Conditioning (AC) systems
and Electrical Vehicles (EVs) with a high degree of control that can bring about practicable
grid services, particularly DR [21]. This inadequacy can also be attributed to the nature of
public data. Such circumstances narrow NILM’s potential for useful applications.

Table 1. Ten most cited public datasets for NILM studies in residential houses in descending order [1].

Number of Measuring Duration Sampling Frequency

Dataset Site
Houses per House Appliance Aggregate
REDD 6 3–19 days 3s 1 s & 15 kHz USA
UMass Smart 3 3 months 1s 1s USA
UK-DALE 5 3–17 months 6s 1–6 s & 16 kHz UK
BLUED 1 8 days event label 12 kHz USA
AMPDs 1 1 year 1 min 1 min CDN
ECO 6 8 months 1s 1s CH
Tracebase 15 N/A 1–10 s N/A DE
HES 251 1–12 months 2–10 min 2–10 min UK
iAWE 1 73 days 1–6 s 1s IND
GreenD 9 1 year 1s 1s AT/IT

The issues above can cause insufficiency in contemporary NILM techniques for prac-
tical applications. They can question the generalization, effectiveness, and dependability
of case studies to make NILM technically feasible. The concern is raised when relevant
research looks for solutions in geographically-specific regions with unexplored energy
consumption cases [22]. This situation matters to the Quebec area, where the type of house-
hold electrical appliances composes an energy demand profile that cannot be typified by
existing public databases. Accordingly, applicable data become imperative to investigating
NILM scenarios created by residential load behavior in this region, particularly based on
low-sampling disaggregation methods. Such information can bring about an opportunity
for analyzing the high flexibility potentials of Quebec household energy usage manifested
by loads that rarely exist in public datasets. These aspects signify geographically-related
NILM analyses based on data from smart meters. Such examinations can assist in observing
actual barriers to achieving a fruitful NILM system. Accordingly, this paper tackles the
problem of low-sampling-rate NILM in the context of Quebec. Its realistic case study
utilizes on-site smart meter readings of the active power of a set of houses with a sampling
interval of 15 min. In order to provide a clear picture of the specificity of the case study,
it commences with a comparative statistical analysis by exploiting data from European
regions. Afterward, the study proposes a load identification mechanism that takes advan-
tage of smart sub-metering technologies and offers a combination of ILM and NILM. The
proposed approach aims to relieve the complications related to a system solely based on
NILM in Quebec residences. In accordance with the recommended solution, a load disag-
gregation method is also suggested on the basis of supervised and unsupervised machine
learning techniques to demonstrate these difficulties. In addition, this work elaborates
on other practical solutions suggested worldwide to enable a feasibility study through its
algorithmic design. As a result, it launches applied research toward practical NILM in
Quebec and other districts with similar circumstances. The rest of the paper is organized
as follows. Section 2 briefly discusses basic research on NILM. Section 3 represents the
statistical analysis aimed at comparing household energy usage properties of Quebec data
Sensors2023,
Sensors 1, 07288
2023,23, 4 of 24 4 of 24

as follows. Section 2 briefly discusses basic research on NILM. Section 3 represents the
with public databases. Section 4 explains the proposed method fortified by the suggested
statistical analysis aimed at comparing household energy usage properties of Quebec data
NILM practice in Section 5. Section 6 presents the results and discussion, followed by
with public databases. Section 4 explains the proposed method fortified by the suggested
conclusions
NILM practicein
inSection
Section 7.
5. Section 6 presents the results and discussion, followed by
conclusions in Section 7.
2. Fundamentals of NILM Concept
2. Fundamentals of NILMby
NILM, proposed Concept
Hart in 1992, has been practiced for many years based on load
NILM, proposed
disaggregation by Hart to
algorithms in 1992, has the
separate beenhousehold
practiced for many years
electrical loadbased
into aon
setload
of individual
disaggregation algorithms to separate the household electrical load into a set of individual
appliances [23]. The common facet of these methods comprises feature selection, classi-
appliances [23]. recognition
fication, and The commonthrough
facet of these methods
exploiting comprises
data feature
collected by anselection, classi-system [1].
acquisition
fication, and recognition through exploiting data collected by an acquisition system [1].
Figure 2 presents these steps along with their general analytical processes, which outline
Figure 2 presents these steps along with their general analytical processes, which outline
the disaggregation engine. Different aspects considered to configure these procedures are
the disaggregation engine. Different aspects considered to configure these procedures are
briefly
briefly discussed
discussed with
with relevant
relevant references
references for further
for further exploration.
exploration.

Common training methods

Sub-metered General
Aggregated
appliance appliance-level
signal
measurement information

Supervised Semi-supervised Feature

training training extraction

1 2 Load
identification
Household
appliance-level
information Energy
services

Figure
Figure 2. 2. NILM
NILM procedure
procedure alongalong
with with its common
its common choice
choice of of learning
learning methods methods practiced
practiced by the by the
fundamental
fundamental research
research [6]. [6].

2.1.
2.1.Mathematical
MathematicalMethods
Methods
Numerical
Numerical schemes
schemesusedused
to formulate an NILM
to formulate task cantask
an NILM be mainly
can beclassified into
mainly classified into
event-based and event-less algorithms. Unlike the latter, which inspects all data instances,
event-based and event-less algorithms. Unlike the latter, which inspects all data instances,
the former is more computationally efficient, since it draws inferences only about detected
the former is more computationally efficient, since it draws inferences only about detected
events. On the other hand, the second can rectify incorrect estimations since it evaluates
events. On the other hand, the second can rectify incorrect estimations since it evaluates
every sample for classification dissimilar to the first, which can contain inaccuracies due to
every sample
misdetection for classification
or false-detection [24].dissimilar to the
Additionally, thefirst, which can
mathematical contain inaccuracies
approaches to NILM due to
misdetection or false-detection [24]. Additionally, the mathematical
can be categorized into optimization, traditional machine learning, and DL models. Vari- approaches to NILM
can be categorized into optimization, traditional machine learning,
ants of Hidden Markov Models (HMMs) are the most famous former NILM method that and DL models. Vari-
ants
can of Hidden
provide Markov
an efficient Models (HMMs)
representation are the[6].
of appliances most famous
Linear and former NILM
non-linear method that
mixed
integer optimization
can provide are otherrepresentation
an efficient popular means ofofload identification
appliances [6]. inLinear
the literature [25,26].
and non-linear mixed
Recently,
integer DL, mostly Recurrent
optimization Neural
are other Networks
popular means (RNNs)
of load and Convolutional
identification in Neural Net- [25,26].
the literature
works (CNN),
Recently, DL,has drawnRecurrent
mostly significantNeural
attention for performing
Networks (RNNs) NILM
and[27]. Mathematical
Convolutional Neural Net-
techniques for load disaggregation can be fundamentally exercised through Open Source
works (CNN), has drawn significant attention for performing NILM [27]. Mathematical
NILM Toolkit (NILMTK). NILMTK provides a public framework for executing NILM based
techniques for load disaggregation can be fundamentally exercised through Open Source
on different methods over several public databases [28].
NILM Toolkit (NILMTK). NILMTK provides a public framework for executing NILM based
on different methods over several public databases [28].

2.2. Learning Procedures

An NILM system established by machine learning methods can be divided according
to its learning phase. In this regard, it is developed based on supervised, semi-supervised,
Sensors 2023, 23, 7288 5 of 24

and unsupervised means. Supervised techniques, as the most common strategy, use labeled
data to train the disaggregation model. Semi-supervised methods use partially labeled
data or a set of prior knowledge for the same purpose. Unsupervised practices otherwise
consider no labeled data for training [29]. The operating systems based on the first learning
concept form a supervised NILM, while those based on the two latter training mechanisms
create an unsupervised NILM exercise [30]. Supervised NILM requires relatively more
human intervention and computational resources. It is more promising at the cost of
sufficient data, where it attempts to identify a wide range of appliances. Such a system
uses static information to create load models, thus is indifferent to variations in appliances’
temporal behavior and the presence of unknown devices in the total signal (see Figure 2).
Unsupervised NILM demands manual labeling. However, it is dynamic regarding temporal
changes in the electrical signal since its identification task is independent of targeting a
specific set of appliances through their models. This operation can assist with the low-cost
generalization of accurate algorithms appealing to business [6].

2.3. Operation Modes

The operating mode is another important aspect of an NILM system. Generally, a
real-time operation is favored to enable practical applications for HEMS services. A low-
complex highly-accurate algorithm is essential to operate a real-time disaggregation. All
research that intends real-world implementations should adopt such a manner. In this
regard, the learning phase of an NILM exercise can be carried out through either online or
offline means according to its algorithmic process, as shown in Figure 3. Off-line training
accounts for a manual configuration that uses sub-measurement readings while online
learning presents an automatic configuration that utilizes only general information (see
Figure 2). Supervised NILM methods are basically offline processes with two separate
steps of model learning and load identification. On the other hand, unsupervised NILM
schemes promote online procedures with adaptive structures since they utilize no/less
training [5]. From the perspective of specific applications, both frameworks can offer a
real-time service if they guarantee pre-defined deadlines [31]. Enabling the real-time aspect
is a great deal for any load monitoring system since delayed/late usage information is
restricted to empower customers for near real-time load modification. For example, major
loads, such as HVAC systems, are influenced by immediate end-users’ demand and, thus,
modifying their current patterns requires on-time actions [6]. It is important to know that a
level of prior knowledge is the least requirement of any effective NILM. This, a priori, can
simply be the number of states, the power levels, or the operation time of appliances so
that their electrical characteristics are not subject to large variations across households [5].
Primary information can be acquired from manufacturers or experts. Regardless of the
Sensors 2023, 1, 0 6 of 24
operation mode, a totally blind NILM process can lose crucial features and fail in identifying
loads accurately [6].

NILM
operating modes

Off-line On-line
model construction model construction

Off/On-line On-line
disaggregation disaggregation

Supervised system
Off-line
labeling

Unsupervised system

Figure Operation
3.Operation
Figure 3. modes
modes of aofcommon
a common NILM
NILM system
system regarding
regarding its learning
its learning phase [6].
phase [6].

2.4. Data Source

The development of a fruitful NILM vastly depends on the quality of data, principally
its sampling intervals. Data frequency plays a key role in the algorithmic design and
technical configuration of an NILM system, discussed above [22]. High-frequency data are
collected at either thousands or millions of times per second using specific hardware. Lower
sampling rates of 1–10 per second and 15–30 per minute are extracted from smart meters
over a Home Area Network (HAN) and a Wide Area Network (WAN), respectively [32]. For
high sampling rate data, methods that can deal with harmonics and transient (microscopic)
features have been explored. These techniques require excessive learning and face diffi-
culties in measuring energy consumption and realizing robust procedures [33]. Contrarily,
algorithms that can handle power features, such as being active and reactive, have been in-
vestigated for lower data granularity. Such means are generally applied to the steady-state
(macroscopic) electrical characteristics of appliances’ operating states in the context of a
power-based load disaggregation [34]. Such cost-effective NILM systems process electric
power signals that are easier to measure and transmit [35]. However, steady-state-based
mechanisms can undergo challenges related to spatial and time overlapping of appliances
operation, specifically those with lower power amplitude and higher occurrence [36].
In addition, NILM can be assisted by high-level macroscopic properties of appliances’
operation habits as secondary information. For instance, Time of Use (ToU) is an addi-
tional feature that can be utilized for identifying loads with certain operation times more
accurately [37]. Nevertheless, different concerns should be noticed while using secondary
information. The enhancement of appliances’ load recognition should be worth the cost of
processing their additional characteristics. The improvement in energy estimation should
not hinder other applications. For example, utilizing ToU information should not avoid
uncovering temporal abnormality intended by anomaly detection services. Data required
for integrating secondary knowledge into the training phase of the disaggregation method
should be accessible. For instance, capturing the ToU characteristics of devices with a
lower occurrence, like a washing machine, needs a considerable amount of data; this is a
crucial issue since NILM already suffers from reliable data limitations. Notwithstanding,
secondary information, such as time of day, duration of usage, and correlation effects, can
be used to deal with challenging disaggregation practices [3].

3. Quebec Residential Energy Usage Context

The difficulty of NILM in Quebec can be ascribed to load specifications that lead to
a disaggregation scenario with multiple complexities [38]. NILM complication is mainly
rooted in the number of appliances, their power level differences, and the frequency of their
state of operation changes, which are all exposed by the case study. Except for common
loads, Quebec dwellings are mostly equipped with electric space and water heating systems
due to particular geographical conditions. Each house is equipped with several Electric
Baseboard Heaters (EBHs) (8–12 numbers) with high switching frequencies. These loads
can have power levels that are similar, not only to each other (almost identical for the same
products) but also to other energy-extensive appliances like Electric Water Heaters (EWHs),
washing machines, and dryers. Additionally, EBHs can distort low-power trajectories
associated with a wide range of devices, such as fridges, freezers, lighting, entertainment
equipment, and even kettles and microwaves, due to their high power usage and lengthy
operation time. This set includes more than half of the appliances (three out of five) targeted
by load disaggregation in the basic research, as mentioned previously. EWHs deteriorate
this situation with their large demand, short duration, and regular presence. These loads
share similar issues with EBHs, like challenging other loads’ operation detection [22].
EWHs are also a dominant load in other geographical locations. For instance, in New
Zealand households, water heating presents the highest demand with 27% and, thus, is
considered a crucial element of demand-side management strategies [39]. The aggregate
load profiled under such conditions yields a monitoring scenario rarely experienced by
current disaggregation algorithms. As described, the main reason for this negligence is the
Sensors 2023, 23, 7288 7 of 24

focus of the methods on public databases, hardly exemplifying cases like Quebec. In [40],
the authors investigate the impact of EBHs on aggregate power consumption by adding
the overall demand of only four baseboards to a daily load profile from the ECO dataset.
For further analysis of the Quebec context, smart meter data from ten residences with
15 min sampling intervals are utilized. For these houses, aggregate and circuit-level power
consumption at a 1 min sampling rate is also available for an elaborated investigation,
which is not a real-world condition.

3.1. Quebec Residential Data Features

The primary challenge with developing an NILM system for Quebec houses is the
diversity in energy statistics across seasons. This situation is observed by exploiting 1 min
measurement data from different houses, numbered for ease of understanding. Figure 4
exemplifies the behavior within the warm seasons. The first aspect of the aggregate power
profile measured through the ’main’ input of the panel is high fluctuations, especially in
lower demands. In all instances, this can be mainly attributed to the tedious time-varying
load of the Heat Pump (HP). In combination with lower demands from appliances like
the fridge, this behavior creates a challenging pattern to disaggregate (Houses 1 and 2).
The second feature of the overall use in summer is the similarity of the power usage
between energy-expensive devices comprising EWH, dryers, spas (House 3), and even
stoves (House 2). This characteristic can increase the overlapping rate, mostly with EWH
due to its regular operation, sometimes with two different levels of demand (House 2).
Furthermore, the consistent occurrence of the pool (House 4) can expand the concern with
this region of demand. Indeed, overlapping among major appliances is hardly experienced
in public databases, knowing that washing machines and dishwashers are not measured
in the case study, which can boost the problem. Notwithstanding, the case of Quebec
demonstrates its specificity by presenting the load of several EBHs. Figure 5 exemplifies
this condition for two houses with different numbers of baseboards. Each circuit-based
reading, the Heat Channel (HC), represents two heating devices that share the Total Heat
(TH). There is no doubt that such usage can have a massive impact on the domestic load.
This influence can be noticed through two visible features comprising dramatic variations
and significant demand compared to the overall domestic load (see Figure 4), which can
dim their patterns. It forms a ’main’ power consumption as the aggregate usage in cold
seasons that is thoroughly different than in warm periods. For example, the effective
fluctuations of HP, sometimes bigger in winter, are not observable under the new situation.

3.2. Quebec Comparative Data Statistics

The careful observation of power consumption patterns in the Quebec house samples
can be elaborated by a comparative analysis based on public databases. For this purpose,
UK-DALE and ECO are exploited, which belong to the ten most cited datasets. The former
has been broadly utilized in recent studies based on DL and the latter provides challenging
cases for an NILM task [5,15]. In addition, these popular datasets hold data measurements
for an extended period of time that can reveal power consumption patterns impacted by
weather conditions and customers’ activities. This feature, as the reason for excluding
REDD, is essential for a sensible comparison with the Quebec data, since its behavior
widely varies in accordance with seasonal changes. Insufficient measurement is also the
logic behind omitting House 3 in UK-DALE from the analysis. The statistical exploration is
extended by using other data from a load monitoring study funded by the CleverGuard
(CG) project (“CleverGuard—ICT solution based on energy data for protecting the elderly
at home, staying safe and independently” an EU funded AALJP project, 2021–2023, contact:
Pascal Kienast, [email protected]) in Switzerland [41]. This confidential information
was obtained from a set of customers and was recently shared with the authors under
collaborative research work. Although CG data are private, they are referred to as public
for simplification. For the Quebec case, the statistics are explored for both aggregate and
Sensors 2023, 23, 7288 8 of 24

Sensors 2023, 1, 0 8 of 24
Sensors 2023, 1, 0 group-level data comprising domestic and TH loads. Subsequently, the following statistics
8 of 24
are applied to four selected databases to enable an effective comparison.

Main EWH Dryer HP Fridge Main EWH Spa HP Fridge

Main EWH Dryer HP Fridge
7 Main EWH Spa HP Fridge
10 Stove Stove Dryer
7
10 Stove Stove Dryer

(kW)

(kW)
6

(kW)

(kW)
6
8
8 5
consumption

consumption
consumption 5

consumption
6 4
6 4
3
4 3
4
Power

Power
2
Power

Power
2
2
2 1
1
0 0
0 0
00

8
0

1
060

066

162

168

070

076

172

178

000

006

102

108

010

016

112

118
6.0

6.0

7.0

0.1

1.1

1.1
.
.007

.007

.07

.007

.017

.017
0

0
07

07
(a) House 1. (b) House 2.
(a) House 1. (b) House 2.
12 12 Main EWH HP Fridge Pool
12 Main EWH Dryer Spa HP 12
Main EWH Dryer Spa HP Main
Stove EWH
Dryer HP Fridge Pool
Stove
Stove Stove Dryer
10 10

(kW)
(kW)

10 10

(kW)
(kW)

consumption
8
consumption

6 6
6 6

4 4
4 4
Power
Power

Power
Power

2 2
2 2

0 0
0 0
0

18
0

1
070

076

178

080

086

182

188

030

036

132

138

046

142

148
07

14
7.0

7.0

8.0

3.1

4.1

41
7.

.
.007

.007

.017

.17

.17
0

0
07

07
(c) House 3. (d) House 4.
(c) House 3. (d) House 4.
Figure 4. An exemplification of the household power consumption profile in Quebec residences
Figure 4. An
Figure 4. An exemplification
exemplificationofofthe
thehousehold
householdpower
power consumption
consumptionprofile in in
profile Quebec residences
Quebec residences
within two days in warm seasons at 1 min sampling intervals.
within two days in warm seasons at 1 min sampling intervals.
within two days in warm seasons at 1 min sampling intervals.
16
12 Main HC1 HC3 HC5
16 Main HC1 HC2 HC3
12 Main
TH HC1
HC2 HC3
HC4 HC5 Main
TH HC1 HC2 HC3
TH HC2 HC4 14 TH
14
(kW)

10
(kW)
(kW)

10
(kW)

12
12
consumption

consumption

8 10
consumption

consumption

8 10
6 8
6 8
6
4 6
Power

4
Power

Power

4
2
2 2
2
0 0
0 0
00

8
0

1
050

056

158

060

066

162

168

050

056

152

158

066

162

168
05

06
5.0

5.0

6.0

5.0

6.0

6.0
5.

6.
.002

.002

.002
02

(a) House 1. (b) House 5.

(a) House 1. (b) House 5.
Figure 5. An example of the household heating consumption profile in Quebec residences within
Figure 5. An example of the household heating consumption profile in Quebec residences within
Figure 5. An
two days example
in cold seasonsof at
the household
1 min heating
sampling consumption profile in Quebec residences within
intervals.
two days in cold seasons at 1 min sampling intervals.
two days in cold seasons at 1 min sampling intervals.
3.2. Quebec Comparative Data Statistics
3.2. Quebec Comparative Data Statistics
Figure 6 depicts
The careful the public
observation and Quebec
of power data distribution
consumption patterns in the in corresponding houses at
Quebec house samples
The careful observation of power consumption patterns in the Quebec house samples
acan
15 be
min samplingbyrate
elaborated to manage its
a comparative size. The
analysis based terms in this
on public figure represent
databases. For this data name
purpose,
can be elaborated by a comparative analysis based on public databases. For this purpose,
and house and
UK-DALE number;
ECO areforexploited,
example, whichQH2 stands
belong for Quebec
to the ten mostHouse
cited2.datasets.
Generally, the level
The former
UK-DALE and ECO are exploited, which belong to the ten most cited datasets. The former
of
haspower demands
been broadly is notincomparable
utilized between
recent studies based onbothDLcases in any
and the latterregion of scattered
provides challengingdata
has been broadly utilized in recent studies based on DL and the latter provides challenging
from
casesall
forinstances, even
an NILM task with
[5,15]. the focus only
In addition, on
these the Quebec
popular domestic
datasets holdload, which can
data measurements
cases for an NILM task [5,15]. In addition, these popular datasets hold data measurements
share
for an extended
similar period
appliances withof of
the time that can revealWith
public power consumption patterns impacted by
for an extended period time that databases.
can reveal powerregard to the interquartile
consumption and whisker
patterns impacted by
weather conditions
range, representing and customers’ activities. This feature, as the reason for excluding
weather conditions half
and and all data,activities.
customers’ respectively,
Thisit feature,
can be deduced that EBHs,
as the reason EWHs, and
for excluding
REDD, is essential
case-specific devicesfor a sensiblechange
remarkably comparison
the with the
spread of Quebec in
demand data, sincehouses.
Quebec its behavior
REDD, is essential for a sensible comparison with the Quebec data, since its behavior
widely varies in accordance with seasonal changes. Insufficient measurement is also the
widely varies in accordance with seasonal changes. Insufficient measurement is also the
logic behind omitting House 3 in UK-DALE from the analysis. The statistical exploration is
logic behind omitting House 3 in UK-DALE from the analysis. The statistical exploration is
a 15 min sampling rate to manage its size. The terms in this figure represent data name
and house number; for example, QH2 stands for Quebec House 2. Generally, the level
of power demands is not comparable between both cases in any region of scattered data
from all instances, even with the focus only on the Quebec domestic load, which can share
similar appliances with the public databases. With regard to the interquartile and whisker
Sensors 2023, 23, 7288 9 of 24
range, representing half and all data, respectively, it can be deduced that EBHs, EWHs, and
case-specific devices remarkably change the spread of demand in Quebec houses.

QH1 QH3 QH5 QH7 QH9

QH2 QH4 QH6 QH8 QH10
8

Power conumptoin (kW)

0
1

5
KH

H
O

CG
U

U
Main Domestic TH
EC

EC
(a) Public data instances. (b) Quebec data instances.
Figure 6. The distribution of public and Quebec data in targeted houses along with the range of the
Figure 6. The distribution of public and Quebec data in targeted houses along with the range of the
domestic and TH loads from the former.
domestic and TH loads from the former.
With reference to the outlier ranges in public data, it is more likely that this region
With reference to the outlier ranges in public data, it is more likely that this region
Sensors 2023, 1, 0 includes samples related to energy-extensive loads due to their power levels and operation 10 of 24
includes samples related to energy-extensive loads due to their power levels and operation
schedules, except for ECO Houses 4 and 5 (ECOH4 and ECOH5) and CG House 4 (CGH4).
schedules,
Therefore, itexcept
can be forstated
ECO that,
Houses 4 andcases,
in most 5 (ECOH4
major and ECOH5)
appliances and CG
operate in House 4 (CGH4).
distinguishable
Therefore, it can be stated that, in most cases, major appliances operate
regions of public load profiles since outliers are data points with significant differences in distinguishable
Knowing
regions the fact load
that these loads advertise an operation schedule, such specificity can
from theofrestpublic profiles
of the samples. since
Such aoutliers are
circumstance data points
facilitates with significant
identifying thesedifferences
types of
assist
from with their
thesuch
rest as load
ofwashing identification.
the samples. Indeed, in
Suchdishwashers, all
a circumstance cases, a cluster
facilitates with power
identifying quantities over
devices, machines, and kettles, as targeted loadsthese
in thetypes
NILMof
1300 W can
devices, suchbeasapproximated,
washing whichdishwashers,
machines, stands out ofand 90%kettles,
of all the
as data. Onloads
targeted the other
in thehand,
NILM
literature. Indeed, the outlier extent in Quebec data distinguishes no appliances, either
none of the
literature. above
Indeed, distinctive
the patterns
outlier extent is
inmanifested
Quebec by
data the frequency
distinguishes histogram
no of
appliances, Quebec
either
targeted or non-targeted.
data, specifically
targeted or regarding the appliance target space. In this case, only 50% of samples
non-targeted.
Figure 7 indicates the frequency histograms of public and Quebec data.
coverFigure
a wide range of up tofrequency
3 kW splithistograms
into severalof groups with significant frequencies.
For the 7former,
indicates the
a 1 min sampling period has beenpublic
used toand Quebec
better data.
approximate the active
power of the operation state of existing appliances and to provide insights into probable
0.00200
UK-DALE ECO CG Main Domestic TH
groups of targeted ones. It can be observed that a substantial portion of samples has a
0.00175
0.40

power demand of less than 500 W in the public data. Such similarities can be challenging
0.00150
0.35
Normalized values
Normalized values

for NILM if it represents the demands of several targeted devices. However, the only major 0.30
0.00125
operation in this power band relates to the fridge. This can be noted by investigating these 0.25

0.00100
databases at the appliance level and reducing power intervals in the analysis. For example, 0.20

close to 50% of samples carry a load of less than 200 W in around 70% of cases. Additionally,
0.00075
0.15

a minor fraction of instances lies over 1 kW which, interestingly, contains power values
0.00050
0.10

of other targeted appliances. Particularly, the washing machine, dishwasher, and kettle
0.00025 0.05

operate in this boundary according to appliance-level information of associated datasets.

0.00000
0.0 0.5 0.9 1.3 1.7 2.1 2.5 2.9 3.4 3.8 4.2 4.6 5.0 5.4 5.9 6.3 6.7 7.1 7.5 7.9 8.3 8.8 9.2 9.6 10.0
0.00
0 1 2 3 4 5 6 7 8 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 22
to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to to
0.5 0.9 1.3 1.7 2.1 2.5 2.9 3.4 3.8 4.2 4.6 5.0 5.4 5.9 6.3 6.7 7.1 7.5 7.9 8.3 8.8 9.2 9.6 10.010.4 1 2 3 4 5 6 7 8 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 22 23

Power ranges (kW) Power ranges (kW)

(a) Public databases. (b) Quebec data.

Figure 7. The frequency histogram of public and Quebec data along with relevant domestic and TH
Figure 7. The frequency histogram of public and Quebec data along with relevant domestic and TH
shares of the latter.
shares of the latter.
The power consumption pattern is another characteristic that can provide sensible
For the former, a 1 min sampling period has been used to better approximate the active
insights into the data. Exploring this property can help improve demand-side management
power of the operation state of existing appliances and to provide insights into probable
strategies by understanding customers’ behavior toward utilizing their electrical appliances,
groups ofbased
especially targeted ones. It
on activity can be
cycles andobserved that a substantial
climate conditions portion
[42]. Although of samples
pattern has a
recognition
power demand of less than 500 W in the public data. Such similarities
should be an essential service of any load monitoring system regarding energy-saving can be challenging
for NILM if popular
awareness, it represents the demands
databases of severaltotargeted
are inadequate enable devices. However,
such a practice thetoonly
due major
limited
operation
data length in and
this power
quality.bandThisrelates
can beto observed
the fridge.inThis can be
Figure noted by
8, where investigating
it is challengingthese
to
databases at the appliance level and reducing power intervals in
determine a common period to draw inferences about behavioral differences among the analysis. For example,
end-
close toNotwithstanding
users. 50% of samples carry threeayears
load of
of less
datathan 200 W inthe
acquisition, around 70% ofdatabase
UK-DALE cases. Additionally,
seriously
a minor fraction of instances lies over 1 kW which, interestingly, contains
suffers from missing data, and the available data are scattered across dissimilar time power values
periods,
of other targeted appliances. Particularly, the washing machine,
except for House 1. However, a continuous pattern has been extracted by combining dishwasher, and kettle
operate in this boundary according to appliance-level information of
relevant houses for six months. House 3 does not offer sufficient readings, even for an associated datasets.
Knowing the
individual fact that
analysis. Thethese loads advertise
ECO dataset anthe
is subject to operation schedule,
same problem withsuch specificity
less severity can
since,
assist with
except their load
for Houses identification.
3 and 6 with notableIndeed, in all
missing data,cases,
powera cluster
valueswith power quantities
are available for roughlyover
the entire measurement period. Nonetheless, it can be noticed that diurnal behavior per
month is similar for each case study in the public data. Slightly higher variations can be
detected within time progress in the ECO data.
Since the selected period covers partially warm and cold seasons, such a similar occur-
rence demonstrates less correlation with environmental factors and a stronger relationship
Sensors 2023, 23, 7288 10 of 24

1300 W can be approximated, which stands out of 90% of all the data. On the other hand,
none of the above distinctive patterns is manifested by the frequency histogram of Quebec
data, specifically regarding the appliance target space. In this case, only 50% of samples
cover a wide range of up to 3 kW split into several groups with significant frequencies.
The power consumption pattern is another characteristic that can provide sensible
insights into the data. Exploring this property can help improve demand-side management
strategies by understanding customers’ behavior toward utilizing their electrical appliances,
especially based on activity cycles and climate conditions [42]. Although pattern recognition
should be an essential service of any load monitoring system regarding energy-saving
awareness, popular databases are inadequate to enable such a practice due to limited
data length and quality. This can be observed in Figure 8, where it is challenging to
determine a common period to draw inferences about behavioral differences among end-
users. Notwithstanding three years of data acquisition, the UK-DALE database seriously
suffers from missing data, and the available data are scattered across dissimilar time periods,
except for House 1. However, a continuous pattern has been extracted by combining
relevant houses for six months. House 3 does not offer sufficient readings, even for an
individual analysis. The ECO dataset is subject to the same problem with less severity since,
except for Houses 3 and 6 with notable missing data, power values are available for roughly
the entire measurement period. Nonetheless, it can be noticed that diurnal behavior per
Sensors 2023, 1, 0 11 of 24
month is similar for each case study in the public data. Slightly higher variations can be
detected within time progress in the ECO data.

μ μ ± 2σ 2.5 μ μ ± 2σ
2.0
Power consumption (kW)
Power consumption (kW)

2.0
1.5

1.5

1.0
1.0

0.5
0.5

0.0 0.0
0

0
0

:0
:0

12
00

.
ov
pt

ct
ay

Au
Ju

O
Se

N
Ju
M

(a) UK-DALE Houses 2 and 4. (b) UK-DALE Houses 1 and 5.

1.75
3.0 μ μ ± 2σ μ μ ± 2σ
Power consumption (kW)

1.50
Power consumption (kW)

2.5
1.25

2.0
1.00

1.5
0.75

1.0
0.50

0.5
0.25

0.0 0.00
0

0
:0

:0
00

12
l.

n.
ov
pt

ay
.

l.
Ju

ar
b
Au

Ju
Ap
O
Se

Ju
Fe
N

M
M

(c) ECO Houses 1, 2, 4 and 5. (d) CG houses.

Figure 8. The diurnal behavior of energy consumption in public databases according to data avail-
Figure 8. The diurnal behavior of energy consumption in public databases according to data avail-
ability and time interval similarity.
ability and time interval similarity.
100 Domestic TH
μ μ ± 2σ
Since the selected period covers partially warm and cold seasons, such a similar occur-
12
Power consumption (kW)

rence demonstrates less correlation with environmental factors and a stronger relationship
10 80 45 45 44
Demand percentage

with calendar ones. The pattern tendency can be attributed to both the type of in-use
53
59 59 62 60
8 65 66

appliances, especially heating/cooling systems, and the climate condition. These are the
6
60

same reasons for which the Quebec data illustrates a significantly different usage pattern,
4
40

2
55 55 56
47
0 41 41 38 40
20 35 34
12 0
00 0
12 0
00 0
12 0
00 0
12 0
00 0
12 0
00 0
12 0
00 0
12 0
00 0
12 0
00 0
12 0
00 0
12 0
00 0
12 0
00 0
12 0
0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
:0
00

0
1

10
n.

.
b.

n.
ar

ov
pt

H
Ju
Ap
Ja

Ja
Fe

H
Q

Q
M
M

O
Se

Q
2.5

Power consumption (k

Power consumption (k
1.25

2.0
1.00

1.5
0.75

1.0
0.50

Sensors 2023, 23, 7288 0.5

0.25 11 of 24
0.0 0.00

0
:0

:0
00

12
as shown in Figure 9a. Weather and calendar components strongly influence the power

n.
ov
pt

l.
b.

n.
Ju

ar
Au

Ju
Ap
O
Se

Ju
Fe
N

M
M
consumption behavior in Quebec houses. A major share of this impact can be assigned to
the notable (c)load
ECO of
Houses
EBHs,1,according
2, 4 and 5. to Figure 9b [43]. It should
(d)be
CG houses.
noted that a total number
of eight houses have been exploited for this statistic due to the lack ofaccording
Figure 8. The diurnal behavior of energy consumption in public databases to datadata
measurement avail-at
ability and time interval similarity.
the main circuit for two houses.
100 Domestic TH
μ μ ± 2σ
12
Power consumption (kW)

10 80 45 44
45

Demand percentage
53
59 59 62 60
8 65 66

60
6

4
40

10
n.

n.
b.

.
ar

ov
pt

H
Ju
Ap
Ja

Ja
Fe

H
Q

Q
M
M

O
Se

Q
(a) Main readings. (b) Quebec data instances.
Figure 9. The diurnal behavior of energy consumption in eight Quebec houses according to data
Figure 9. The diurnal behavior of energy consumption in eight Quebec houses according to data
available from the main reading.
available from the main reading.
Exploring power demand time series according to their systematic and unsystematic
Exploring power demand time series according to their systematic and unsystematic
components is another useful analysis regarding data characterization. A seasonality
components
study can be is another for
employed useful analysis regarding
this purpose. data characterization.
Figure 10 exemplifies A seasonality
this exercise for two cases.
study can be employed for this purpose. Figure 10 exemplifies
From public databases, ECO is considered regarding its demand pattern, and this exercise for two
Housecases.
4
From
is selected considering its power distribution, which signifies the presence of energy-4
public databases, ECO is considered regarding its demand pattern, and House
is selected considering
demanding its power
loads, especially seasonaldistribution, which signifies
ones. The examination the presence
is carried of energy-
out by the use of
demanding loads,model
the multiplicative especially
sinceseasonal ones.
it is a better The examination
choice for time-varying is carried
behavioroutand
by the use of
removes
the multiplicative model since it is a better choice for time-varying behavior
difficulty in interpreting negative values. Furthermore, this proves to perform better in and removes
difficulty
capturingin interpreting
peaks throughnegative
its seasonal values. Furthermore,
component. this proves
With regard to thetoresidual
performelement,
better in
capturing peaks through its seasonal component. With regard to
it can be noticed that the public case contains a huge amount of unpredictable/noisy the residual element,
itinformation,
can be noticed that the public case contains a huge amount of unpredictable/noisy
which exposes fluctuations almost similar to the main signal. This shows that
information, which
a notable amount ofexposes fluctuations
data is not consistentalmost similar
with the rest oftoit.the main
The trendsignal. This shows
and seasonal that
factors
a notable amount of data is not consistent with the rest of it. The trend and seasonal factors
show a general upward slope and a clear recurring/periodic pattern, respectively, that are
relatively poor considering the residual. This can be estimated by multiplying these two
components. It is observed that the systematic information contributes inadequately to
explaining the usage. On the other hand, the Quebec data are characterized by valuable
systematic information, particularly in winter, where the model is able to strongly describe
the demand. In addition, it can be realized that TH has a great influence on the seasonality
of the aggregate load. This impact, along with the level of overall systematic information,
promotes a seasonality-based NILM approach to disaggregating EBHs load. In addition,
the results demonstrate that a classic decomposition is not an efficient choice for the Quebec
case, since the seasonality of the data strongly changes within the year. It should be noted
that the seasonality of other public instances is inferior to the selected one, specifically for
UK-DALE and CG data.
In order to discover the real-world relationship among data instances and suggest
generalizable hypotheses, a correlation analysis can be used. Figure 11 presents the results
of this investigation into existing houses from all datasets, except for UK-DALE due to
dissimilarity over available data periods. It can be seen that the correlation between
household usage in the public data is not even moderate, evidencing a notable difference
in their tendency toward operating electrical appliances. Similar behavior can be detected
in the Quebec domestic load. However, a moderate correlation can be noticed across the
overall load that is rooted in the medium to high correspondence between TH demand,
similar to seasonality.
the demand. In addition, it can be realized that TH has a great influence on the seasonality
of the aggregate load. This impact, along with the level of overall systematic information,
promotes a seasonality-based NILM approach to disaggregating EBHs load. In addition,
the results demonstrate that a classic decomposition is not an efficient choice for the Quebec
case, since the seasonality of the data strongly changes within the year. It should be noted
Sensors 2023, 23, 7288 12 of 24
that the seasonality of other public instances is inferior to the selected one, specifically for
UK-DALE and CG data.

Overall demand Trend×Seasonality

Overall demand Trend×Seasonality
6 6

Observed
Observed
4 4

2 2
Sensors 2023, 1, 0 13 of 24
0

8 TH Overall demand
5

6 4

Residual
Residual
3

can also be acknowledged from the statistical study that disaggregating the choice of
4
2
2
appliances in the literature, especially fridges, kettles, and microwaves, from the Quebec
1

0 0

data is a burdensome task. This exercise becomes more difficult knowing the fact that
1.75 5 TH Overall demand

actual readings have a sampling rate of 15 min. Figure 12 shows daily household load
1.50

1.25
4

Trend
3
Trend

profiles recorded by smart meters within warm and cold seasons. Complex operational
1.00 2

curves, seasonal variations, and continuous changes at lower demands are the challenging
0.75

0.50
1

0
features of these profiles. In summer, the EWH represents the most notable demand, TH Overall demand

contributing to almost all load peaks solely or partially. As a result, it gives the total usage a
2.0
1.5

Seasonal
Seasonal

similar pattern to its own in both shape and magnitude. In winter, the EBHs illustrate their
1.5
1.0

remarkable influence by transforming the domestic load as shown in this figure. Another
1.0

0.5
important property that can be realized is the significant level of unknown demand as the
0.5
Jul
2012
Aug Sep Oct Nov Dec Jun Jul Aug Sep Oct Nov Dec Jan
2018
Feb Mar Apr

difference between the main circuit and the total loads and domestic and aggregate usages
(a) respectively.
in Figure 10a,b, ECO House 4.This underlying issue can seriously
(b) Quebec House
impact 1. performance
the
Figure 10. Thepractice.
of an NILM seasonal decomposition
Accordingly, itofcan
public and Quebec
be stated data for
that NILM two a
faces fine examples based
completely on
different
Figure 10. The seasonal decomposition of public and Quebec data for two fine examples based on
the multiplicative
case model.
in Quebec dwellings.
the multiplicative model.
In order to discover the real-world relationship among data instances and suggest
ECOH1

CGH1

0.18
generalizable hypotheses, a correlation analysis can be used. Figure 11 presents the results
0.06

0.16
of this investigation into existing houses from all datasets, except for UK-DALE due to
ECOH2

0.04
CGOH2

0.07
0.19
0.14

dissimilarity over available data periods. It can be seen that the correlation between 0.02
ECOH3

0.07 0.05
CGH3

0.12

household usage in the public data is not even moderate, evidencing a notable difference
0.08 0.07
0.00
ECOH4

-0.02 -0.03 -0.00

0.10

in their tendency toward operating electrical appliances. Similar behavior can be detected −0.02
CGH4

0.05 0.11 0.08

ECOH5

0.08

in the Quebec domestic load. However, a moderate correlation can be noticed across the
0.01 0.03 -0.04 -0.03
−0.04
CGH5

0.06

overall load that is rooted in the medium to high correspondence between TH demand,
ECOH6

0.09 0.10 0.04 0.11

0.02 0.06 0.03 -0.07 -0.03
−0.06

similar to seasonality. ECOH1 ECOH2 ECOH3 ECOH4 ECOH5 ECOH6 CGH1 CGOH2 CGH3 CGH4 CGH5

The (a)
above
ECOanalysis
data. has been aimed at(b)
revealing
CG data.statistics that have not been fairly taken
into consideration in related research and can be applied to any other case studies. Indeed,
QH1

it is evident that the Quebec data encounter a vastly higher amount of events compared to
QH1

QH1

0.8
0.60
QH2

0.20 0.73
QH2

0.58

public data due to a bigger number of appliances with high switching frequency in relevant
QH2

0.07
QH3
QH3

0.47 0.50 0.61 0.57

QH3

0.09 0.00 0.15

0.55

operations, i.e., EBHs. From every targeted aspect, the statistical analysis demonstrates 0.7
QH4
QH4

0.48 0.48 0.46 0.69 0.61 0.68

QH4

0.15 0.11 0.08

massive differences between the public and Quebec data. From a practical standpoint,
QH5

0.10
QH5

0.74 0.62 0.59 0.81

QH5

0.52 0.54 0.43 0.55 0.50 0.07 -0.02 0.20 0.09

QH6
QH6

0.81 0.71 0.71 0.83 0.83 0.6

a load monitoring practice should approach this case with a different set of targeted
QH6

0.45 0.43 0.43 0.53 0.44 0.19 0.12 0.09 0.24 0.06 0.05
0.45
QH7
QH7

0.77 0.67 0.61 0.72 0.76 0.80

QH7

0.42 0.46 0.37 0.40 0.42 0.34 0.06 0.02 0.11 0.09 0.13 0.05

appliances. The group of interest must certainly contain EBHs due to their share of 0.00
QH8
QH8

QH8

0.60 0.59 0.52 0.48 0.53 0.42 0.46 0.08 0.04 0.04 0.08 0.02 0.07 0.06 0.82 0.69 0.69 0.78 0.80 0.85 0.79
0.5
0.40

demand, their impact on load characteristics, and their potential applications for HEMSs.
QH10 QH9
QH10 QH9

QH10 QH9

0.60 0.63 0.46 0.50 0.61 0.45 0.46 0.62 0.10 0.10 0.02 0.08 -0.10 0.16 0.03 0.05 0.56 0.47 0.39 0.48 0.52 0.60 0.42 0.58
−0.05

It should also include EWHs due to their major usage and regular presence that makes
0.54 0.62 0.49 0.44 0.49 0.40 0.43 0.54 0.57 0.03 0.06 0.08 0.05 0.09 0.05 0.06 -0.00 -0.01 0.75 0.73 0.66 0.70 0.71 0.80 0.73 0.72 0.51
0.35 0.4
QH1 QH2 QH3 QH4 QH5 QH6 QH7 QH8 QH9 QH10 QH1 QH2 QH3 QH4 QH5 QH6 QH7 QH8 QH9 QH10 QH1 QH2 QH3 QH4 QH5 QH6 QH7 QH8 QH9 QH10

them responsible for the most rapid rises in household demand all across the year. It
(c) Quebec aggregate data. (d) Quebec domestic data. (e) Quebec TH data.
Figure 11. Correlation between available instances in public and Quebec databases along with similar
Figure 11. Correlation between available instances in public and Quebec databases along with similar
information for domestic and TH loads of the latter.
information for domestic and TH loads of the latter.
HC1 HC4 Overall heating Aggregate demand
Stove Fridge Heat pump Main circuit
The above analysis has been aimed at revealing
6
10
EWH statistics that have not been fairly taken
Dryer Domestic demand
HC2
HC3
HC5 Domestic demand Main circuit
Power consumption (kW)

into consideration in related research and can be 8applied to any other case studies. Indeed,
Power consumption (kW)

5
it is evident that the Quebec data encounter a vastly higher amount of events compared to
4
public data due to a bigger number of appliances 6with high switching frequency in relevant
operations,
3
i.e., EBHs. From every targeted aspect,
4 the statistical analysis demonstrates
massive
2 differences between the public and Quebec data. From a practical standpoint,
a load
1 monitoring practice should approach this 2
case with a different set of targeted
appliances.
0
The group of interest must certainly
0
contain EBHs due to their share of
demand, their impact on load characteristics, and their potential applications for HEMSs.
00

21
00

It should also include EWHs due to their major usage and regular presence that makes
3

3
0

-1

-1
-1

-1

02
07

(a) Summer case. (b) Winter case.

Figure 12. Exemplification of seasonal behavior of power consumption profiles in a Quebec house at
a 15 min sampling rate.
case in Quebec dwellings.

ECOH1

CGH1
0.06 0.18

0.16

ECOH2
0.04

CGOH2
0.07
0.19
0.14

Sensors 2023, 23, 7288 13 of 24

0.02

ECOH3
0.07 0.05

CGH3
0.12
0.08 0.07
0.00

ECOH4
-0.02 -0.03 -0.00
0.10
−0.02

CGH4
0.05 0.11 0.08

ECOH5
0.08
0.01 0.03 -0.04 -0.03

them responsible for the most rapid rises in household demand all across the year. It
−0.04

CGH5
ECOH6 0.06
0.09 0.10 0.04 0.11
0.02 0.06 0.03 -0.07 -0.03

can also be acknowledged from the statistical study that disaggregating the choice of
−0.06

ECOH1 ECOH2 ECOH3 ECOH4 ECOH5 ECOH6 CGH1 CGOH2 CGH3 CGH4 CGH5

appliances in the literature, especially fridges, kettles, and microwaves, from the Quebec
(a) ECOdata
data.is a burdensome task. (b) CG data.
This exercise becomes more difficult knowing the fact that
actual readings have a sampling rate of 15 min. Figure 12 shows daily household load

QH1
QH1

QH1
profiles recorded by smart meters within warm and cold seasons. Complex operational
0.60
0.8

QH2
0.20 0.73
QH2

0.58

QH2
0.07

curves, seasonal variations, and continuous changes at lower demands are the challenging

QH3
QH3

0.47 0.50 0.61 0.57

QH3
0.09 0.00 0.15
0.55
0.7

QH4
features of these profiles. In summer, the EWH represents the most notable demand,
QH4

0.48 0.48 0.46 0.69 0.61 0.68

QH4
0.15 0.11 0.08

QH5
0.10
QH5

0.74 0.62 0.59 0.81

QH5
0.52 0.54 0.43 0.55 0.50

contributing to almost all load peaks solely or partially. As a result, it gives the total usage a
0.07 -0.02 0.20 0.09

QH6
QH6

0.81 0.71 0.71 0.83 0.83 0.6

QH6
0.45 0.43 0.43 0.53 0.44 0.19 0.12 0.09 0.24 0.06 0.05

similar pattern to its own in both shape and magnitude. In winter, the EBHs illustrate their
0.45

QH7
QH7

0.77 0.67 0.61 0.72 0.76 0.80

QH7
0.42 0.46 0.37 0.40 0.42 0.34 0.06 0.02 0.11 0.09 0.13 0.05

0.00

remarkable influence by transforming the domestic load as shown in this figure. Another

QH8
QH8

QH8
0.60 0.59 0.52 0.48 0.53 0.42 0.46 0.08 0.04 0.04 0.08 0.02 0.07 0.06 0.82 0.69 0.69 0.78 0.80 0.85 0.79
0.5
0.40

QH10 QH9
QH10 QH9

QH10 QH9
0.60 0.63 0.46 0.50 0.61 0.45 0.46 0.62 0.56 0.47 0.39 0.48 0.52 0.60 0.42 0.58

important property that can be realized is the significant level of unknown demand as the
0.10 0.10 0.02 0.08 -0.10 0.16 0.03 0.05 −0.05

0.54 0.62 0.49 0.44 0.49 0.40 0.43 0.54 0.57 0.03 0.06 0.08 0.05 0.09 0.05 0.06 -0.00 -0.01 0.75 0.73 0.66 0.70 0.71 0.80 0.73 0.72 0.51
0.35

difference between the main circuit and the total loads and domestic and aggregate usages
0.4
QH1 QH2 QH3 QH4 QH5 QH6 QH7 QH8 QH9 QH10 QH1 QH2 QH3 QH4 QH5 QH6 QH7 QH8 QH9 QH10 QH1 QH2 QH3 QH4 QH5 QH6 QH7 QH8 QH9 QH10

in Figure
(c) Quebec aggregate 10a,b, respectively.
data. This underlying
(d) Quebec domestic data. issue (e)
canQuebec
seriously impact the performance
TH data.
Figure 11. Correlation between available instances in public and Quebec databases along withasimilar
of an NILM practice. Accordingly, it can be stated that NILM faces completely different
case in Quebec dwellings.
information for domestic and TH loads of the latter.

HC1 HC4 Overall heating Aggregate demand

Stove Fridge Heat pump Main circuit
10 HC2 HC5 Domestic demand Main circuit
6 EWH Dryer Domestic demand
HC3
Power consumption (kW)
Power consumption (kW)

5 8

4
6

3
4
2

2
1

0 0
00

21
00

3
0

-1

-1
-1

-1

02
07

(a) Summer case. (b) Winter case.

Figure 12. Exemplification of seasonal behavior of power consumption profiles in a Quebec house at
Figure 12. Exemplification of seasonal behavior of power consumption profiles in a Quebec house at
a 15 min sampling rate.
a 15 min sampling rate.

4. A Disaggregation Approach to Quebec Household Power Consumption

As discussed, the designation of an NILM model highly relies upon data sampling
intervals. The least requirements reported for a disaggregation task based on regression and
classification methods are 3 Hz and 1 Hz, respectively, which is not the case in this study [11].
From the viewpoint of feature extraction, power-based steady-state electrical features at
a macroscopic level are recommended for load disaggregation purposes. Nevertheless,
relevant research indicates that even a sampling time of 1 min is not sufficient to capture
EBHs’ behavior due to their rapid state transitions. From the load identification point of
view, an unsupervised method is favored for the Quebec case due to data shortage in both
length and quality, e.g., appliance-level information. Such an approach can be combined
with a supervised technique to relieve the impracticality of a completely blind practice at
the cost of sufficient data, particularly for EBHs with seasonal presence. These aspects help
outline an NILM system for the Quebec case. Notwithstanding, general inadequacies of
existing methods in the literature on one side, and potential challenges of deploying NILM
in the Quebec region on the other side, stimulate alternative solutions.
As illustrated in Figure 1, ILM is another technology that can take part in HEMS
strategies. This method of load monitoring can accurately determine the operation state of
a device with the use of one or several sensors. Although the barriers related to complex in-
stallation and multiple sensor configurations can limit this approach, affordable technology
and reliable implementation can support it. The ILM can have different levels according to
the number of meters, which can be one meter per zone, plug, or appliance. The former has
the lowest complexity in terms of both hardware and software as well as data transition re-
Sensors 2023, 23, 7288 14 of 24

quirements. Therefore, it can be simply combined with the entry point meter (non-intrusive
concept) to boost monitoring efficiency, particularly in small-scale installations, for example,
only for major appliances [10,38]. Similarly, in Quebec households, there is one thermostat
per zone for controlling corresponding EBHs. The information on the thermostats can
be combined with that on total energy demand to improve NILM results. HEMS can
communicate with each EBH (if all thermostats are connected) or zone (otherwise) and use
the acquired data to disaggregate the overall heating use from the aggregate load profile.
Subsequently, the remaining usage can be analyzed to uncover the power demand of
EWHs as the second targeted load in Quebec households. Extracting these two demanding
loads allows for the exploration of the residual to disaggregate common home appliances,
Sensors 2023, 1, 0
i.e., those from the literature, based on popular NILM methods. This potential solution 15 of 24
draws interest to the approach of combining ILM and NILM systems in Quebec houses
for alleviating its complicated scenario. This method is also supported by advancements
in IoT technologies that link smart appliances throughout the house. Figure 13 illustrates
as
thepool pumps,approach,
proposed air conditioning, furnaces,
where EBHs andalso
data can EBHsbein our case,
obtained while
from theanother
electric scheme
panel to
can use a training phase based on specific signatures/parameters to disaggregate
extract their total demand. In this illustration, a piece of prior knowledge acquired power
from
consumption patterns of other categories of appliances [44].
either the same dwelling (1) or other houses (2) can assist the whole procedure.

HEMS

Data flow Information

HEMS Electric panel
Zone 1 Zone 2 1

OR
EBHs
2
EWH
& others

First level Second level

Smart meter Zone 4 Zone 3 Priors
Priors
ILM NILM

Aggregate load

Figure13.
Figure Anexample
13.An exampleof
ofthe
theproposed
proposedapproach
approachin
inaahouse
housewith
withfour
fourthermal
thermalzones.
zones.

5. AnAvoiding
Introductorythe disaggregation
NILM Practiceof indevices
QuebecatResidences
one single point is a feasible approach that
can increase accuracy, since with more than
Although the proposed approach implies a mechanism one accumulating point,
that takesmore classifications
advantage of ILM
can be executed. In the Quebec case, one point can disaggregate
to extract EBHs demand from the aggregate usage, it is practiced by a procedure total heating demand into
entirely
individual EBHs and another can disassociate EWH and other loads.
based on NILM to demonstrate load disaggregation complications in Quebec houses by Indeed, a practical
load monitoring
numerical outcomes. structure should focus
The methodology on both
follows thedevelopment
steps indicated cost
in and
Figureproper
13. Asaccuracy.
shown
A combinatory method can realize these two goals with an effective
in Figure 14, the initial task focuses on extracting TH demand from the aggregate number of meters,
load
simplerthrough
profile signatures, straightforward
a supervised method, algorithms,
while the and less complicated
successive accumulating
one concentrates profiles.
on capturing
The feasibility
other of this approach
possible devices can be fortified
in the remaining by applicable
signal within NILM products
an unsupervised process. like Bidgely,
In addition
which use programmable thermostat data to improve the performance of
to facilities brought about by thermostats, the statistical analysis also proves that heating disaggregation
techniques
load should be andtheoffer
firstoptimal
target insetpoint schedules. It is also supported by the Bidgely
Quebec houses.
low-resolution disaggregation technique.
The energy estimation of electric heating Thesystems
relevanthas
product uses the
also drawn low-granularity
attention to other
whole-house profile, non-electrical information, and training
implementations. In this regard, analyzing heating systems’ relationships with data to identify individual
external
appliances’
factors load profiles.
and time-series As stated
patterns have beenby Bidgely,
favored.aIn reliable
France,NILM on low-resolution
the Hello Watt Company data has
shouldout
carried be the
carried out demand
heating by reasonably specific underlying
disaggregation according totechniques.
its relationshipOnewith
method
outdoor can
extract a portion of energy from overall usage related to a specific
temperature by means of the piece-wise linear regression method [45]. However, such class of loads such
as pool pumps,
techniques cannotairefficiently
conditioning, furnaces,
explain and EBHs
the heating loadininour case, while
Quebec another
residences duescheme
to its
can use a training phase based on specific signatures/parameters
complex behavior. Our studies on Quebec data show that individual household demands to disaggregate power
consumption
and patterns ofare
outside temperature other
notcategories
sufficiently of correlated
appliancesto[44].
be treated by linear designs [43].
To be specific, heating demand is significantly influenced by calendar components related
5. An Introductory NILM Practice in Quebec Residences
to occupants’ activities as well. Therefore, meteorological variables are inadequate for
Although
accurately the proposed
estimating approach implies
this consumption for load aidentification
mechanism that takes advantage
purposes. of ILM
In Switzerland, a
to extract EBHs demand from the aggregate usage, it is practiced by a procedure
similar examination has been performed by means of time-series clustering methods [46]. entirely
Nevertheless, the utilized scheme is not applicable to the Quebec case since it is aimed at
recognizing a water-based heating system with a large distinct demand of 18 kW occurring
at periods with the lowest power consumption, i.e., between 23:00 and 6:00. As shown in
Figure 2 of the related study, the pattern of an electrical load with these properties is not
Sensors 2023, 23, 7288 15 of 24

based on NILM to demonstrate load disaggregation complications in Quebec houses by

numerical outcomes. The methodology follows the steps indicated in Figure 13. As shown
in Figure 14, the initial task focuses on extracting TH demand from the aggregate load
profile through a supervised method, while the successive one concentrates on capturing
other possible devices in the remaining signal within an unsupervised process. In addition
to facilities brought about by thermostats, the statistical analysis also proves that heating
load should be the first target in Quebec houses.
The energy estimation of electric heating systems has also drawn attention to other
implementations. In this regard, analyzing heating systems’ relationships with external
factors and time-series patterns have been favored. In France, the Hello Watt Company has
carried out the heating demand disaggregation according to its relationship with outdoor
temperature by means of the piece-wise linear regression method [45]. However, such
techniques cannot efficiently explain the heating load in Quebec residences due to its
complex behavior. Our studies on Quebec data show that individual household demands
and outside temperature are not sufficiently correlated to be treated by linear designs [43].
To be specific, heating demand is significantly influenced by calendar components related
to occupants’ activities as well. Therefore, meteorological variables are inadequate for
accurately estimating this consumption for load identification purposes. In Switzerland, a
similar examination has been performed by means of time-series clustering methods [46].
Nevertheless, the utilized scheme is not applicable to the Quebec case since it is aimed at
recognizing a water-based heating system with a large distinct demand of 18 kW occurring
at periods with the lowest power consumption, i.e., between 23:00 and 6:00. As shown in
Figure 2 of the related study, the pattern of an electrical load with these properties is not
ambiguous and is thus difficult to extract.
As a state-of-the-art approach to NILM, DL is another promising method for iden-
tifying the total demand of EBHs and disaggregating it from the aggregate load profile.
Particularly, Long Short-Term Memory (LSTM) networks can be employed for this task due
to their capability to learn long-term temporal dependencies in time-series. Accordingly,
an LSTM network is considered to execute the supervised phase of the suggested method-
Sensors 2023, 1, 0 16 of 24
ology. The structure is developed in Python by using Keras as a DL API. In this regard,
a stacked LSTM architecture through a sequential model is constructed. The selected
arrangement has five hidden layers and one output layer. The hidden architecture contains
three fully
three fully connected
connected LSTMs
LSTMs withwith30,
30,30,
30,and
and2020neurons,
neurons, and
and twotwoDense
Dense layers with
layers 40
with
andand
40 20 20
neurons, progressively.
neurons, progressively.Adam
Adam andand
mean square
mean error
square in terms
error of optimization
in terms of optimizationand
loss loss
and functions, respectively,
functions, are used
respectively, to compile
are used the network.
to compile In addition,
the network. tanhtanh
In addition, andand
relurelu
are
utilized as non-linear and rectified-linear activation functions in the first and last
are utilized as non-linear and rectified-linear activation functions in the first and last layers, layers,
respectively. Training
respectively. Training isis carried
carried out
out within
within 100
100 epochs
epochs running
running aa sliding
sliding window
window that
that maps
maps
seven samples of aggregate power demand to the next (eighth) one of the
seven samples of aggregate power demand to the next (eighth) one of the overall heating overall heating
load. In
load. In addition,
addition, its
its convergence
convergence is is enhanced
enhanced by by utilizing
utilizing aa dynamic
dynamic learning
learning rate.
rate.

Supervised phase

Submetered
Aggregate signal
ESHs readings

LSTM
(learning process)
Unsupervised phase

OPTICS+DBSCAN
Residual signal
(clustering process)

Other loads
ESHs extraction
identification

Figure 14. The

Figure 14. The block
block diagram
diagram of
of the
the NILM
NILM practice
practice proposed
proposed to
to tackle
tackle the
the Quebec
Quebec case
case [47].
[47].

After subtracting the EBHs energy use, the remaining signal is analyzed for identifying
other possible appliances in an unsupervised context. This framework takes advantage
of a clustering scheme that focuses on grouping similar demands according to a proba-
bility threshold, determining their occurring time in the power profile, and creating their
related operation (ON-OFF) sequence, which can possibly represent a load. The on-state of
each created sequence is defined by calculating the average values of the corresponding
cluster. This process can be carried out by using either power level or power difference
information [5]. Particularly in an unsupervised framework, the granularity of data has
a significant impact on the soundness of sequences, captured across the time-series data.
Figure 15 illustrates the power level and difference information of EWH in House 1 at one
and 15 min sampling intervals. The most evident statistic that can be acknowledged in
this figure is the critical change in the power demand pattern at the lower frequency. An
unsupervised system can observe this trajectory as a combination of several loads. In a
supervised practice, such behavior can cause different loads to be represented with similar
models. Therefore, in both cases, recognizing a load with no prior knowledge becomes a
burdensome task, and exploiting expert/secondary information is necessitated. Moreover,
it can be noticed that the sampling rate can notably alter the ON-OFF power transitions of
EWH. They are quite different even in the higher sampling frequency, knowing the fact that
this resistive appliance has one of the most uniform steady-state behaviors among house-
hold devices [22]. In this situation, utilizing power difference for sequence construction
can result in incorrect state matching. It can also lose high operation states and, in turn,
neglect an important share of demand. As a result, employing power level is preferred
since it creates sensible operation sequences of potential devices despite its variations
2023, 1, 0 17 of 24 under a low
and overlapping events. All appliances can be subject to this transformation
sampling rate measurement, especially those with short operation duration.

Power level Power difference Power level Power difference

4
4
Power consumption (kW)

Power consumption (kW)

2 2

0 0

−2 −2

−4
−4
0

0
:0

:0
00

(a) One minute sampling rate. (b) Fifteen minute sampling rate.
Figure 15. The EWH power profile from Quebec House 1 in (a) 1 and (b) 15 min sampling time.
Figure 15. The EWH power profile from Quebec House 1 in (a) 1 and (b) 15 min sampling time.
Two different techniques have been evaluated for executing the clustering phase.
Two different
The first method employs techniques
a Kernel-based have been
Subtractive evaluated(K-SC)
Clustering for executing
algorithm the[29].
clustering phase.
The second scheme takes advantage of DBSCAN (Density-Based Spatial Clustering ofalgorithm [29].
The first method employs a Kernel-based Subtractive Clustering (K-SC)
TheNoise),
Applications with secondwhich
scheme takestime-series
enables advantageclustering.
of DBSCAN (Density-Based
Accordingly, Spatial Clustering of
its advantage
over the K-SC method is that it executes the first two steps of the sequence creation process its advantage
Applications with Noise), which enables time-series clustering. Accordingly,
over and
(load classification the K-SC methoddetection)
occurrence is that it executes the first
within one two
stage. steps of the sequence
Nevertheless, DBSCANcreation process
(load classification
requires the definition and occurrence
of two parameters detection)
that account within one
for Epsilon andstage.
MinPts.Nevertheless,
These DBSCAN
requires the definition of two parameters that account
parameters explain that the Epsilon-neighborhood of each point must contain a minimumfor Epsilon and MinPts. These
parameters explain that the Epsilon-neighborhood of each
number of points (MinPts). A specific way to estimate MinPts has not been suggested inpoint must contain a minimum
number
the literature. For example,of points (MinPts).
the authors A specific
of [48] way to estimate
have recommended usingMinPts hastwice
a value not been
the suggested in
dimension of the data. This suggestion is adopted for this analysis and, thus, a MinPts aofvalue twice the
the literature. For example, the authors of [48] have recommended using
dimension ofparticular
4 is selected. Additionally, the data. methods
This suggestion is adopted
have been for this
developed analysisEpsilon.
to compute and, thus, a MinPts of
4 is selected. Additionally, particular methods have been developed
In [49], the authors have provided a procedure within which, first, the average distance to compute Epsilon.
In [49],and
between each point theits
authors have
k (equal provided
to MinPts a procedure
defined withinnearest
previously) which,neighbors
first, the average
is distance
between each point and its k (equal to MinPts defined previously)
calculated, second, k-distances are plotted in ascending order, and third, the biggest slope nearest neighbors is
(greatest curvature) is considered as an Epsilon value. This process has been evaluated
for the case study in order to provide an effective application of DBSCAN. Figure 16
exemplifies the outcomes for the power level and difference data from House 1 on a typical
day. It can be noticed that the method is not effective for the power difference data since
Sensors 2023, 23, 7288 17 of 24

calculated, second, k-distances are plotted in ascending order, and third, the biggest slope
(greatest curvature) is considered as an Epsilon value. This process has been evaluated
for the case study in order to provide an effective application of DBSCAN. Figure 16
exemplifies the outcomes for the power level and difference data from House 1 on a typical
day. It can be noticed that the method is not effective for the power difference data since
the distance plot fails to realize an elbow at which a specific point can be chosen as the
Epsilon value. On the other side, the elbow plot of the power level data yields a sharp
slope. Nonetheless, this curvature covers a wide range of points corresponding to samples
from 70 to higher than 250, all of which can be used for the Epsilon parameter. The analysis
evidences that moving across this range causes DBSCAN to create different clusters. This,
in turn, makes the choice of Epsilon a difficult task, especially in an unsupervised system.
It should be noted that this challenge is not necessarily the case for other types of data and
can be attributed to Quebec data characteristics (see Section 3.2). However, it necessitates
exploring other schemes. In this regard, OPTICS (Ordering Points To Identify the Clustering
Structure) is considered as another density-based clustering method. OPTICS is able to
provide information about cluster structure in the time-series data that can be observed
through the reachability plot. This plot is illustrated in Figure 17 for the same MinPts and
instance as in Figure 16. It presents the structure of the clusters in the data according to
the peaks and valleys. For instance, in Figure 17a, three major groups can be collected
considering the significant peaks. In fact, the interpretation of the significance can define
the number of clusters as noticed in this figure. It can be observed that OPTICS not only
18 of 24 can
facilitates the choice of Epsilon but also defines the prospective clusters. This technique
manifest the change in the number of clusters based on the range of Epsilon values. This
can be simply practiced by drawing horizontal lines on the reachability plots in Figure 17.
As a result, a combination of the OPTICS and DBSCAN algorithms is suggested for the
the least amount aresequence
considered the step.
creation best Generally,
candidatesthefor Epsilon.
closest valuesThis
to thevalue
top of is
theset to 0.4 inpeak
significant all with
cases. Different analyses demonstrate
the least that the combined
amount are considered clustering
the best candidates approach
for Epsilon. This proposed and
value is set to 0.4 in all
the K-SC method generally capture comparable operations. However, the former uncovers and
cases. Different analyses demonstrate that the combined clustering approach proposed
a higher number of the K-SC method
sequences for generally
a similarcapture comparable
cluster, which canoperations.
reflectHowever,
a higherthe former uncovers
estimation
a higher number of sequences for a similar cluster, which can reflect a higher estimation
of energy demand for a possible load. Moreover, DBSCAN provides a class of outliers
of energy demand for a possible load. Moreover, DBSCAN provides a class of outliers
belonging to no cluster that can
belonging to nobe further
cluster thatanalyzed.
can be further analyzed.

0.8 0.20

0.6 0.15
K distance

K distance

0.4 0.10

0.2 0.05

0.0 0.00
0

Number of data samples Number of data samples

(a) Power level samples. (b) Power difference samples.

Figure 16. The k-nearest neighbor analysis for one-day data from House 1 with MinPts equal to 4.
Figure 16. The k-nearest neighbor analysis for one-day data from House 1 with MinPts equal to 4.

0.45
0.8
0.40
achability (Epsilon distance)

achability (Epsilon distance)

0.35
0.6
0.30
0.25
0.4
0.20
0.15
0.2
0.10
0.0 0.00

35
Number of data samples Number of data samples

Sensors 2023, 23, 7288 (a) Power level samples. (b) Power difference samples. 18 of 24
Figure 16. The k-nearest neighbor analysis for one-day data from House 1 with MinPts equal to 4.

0.45
0.8
0.40
Reachability (Epsilon distance)

Reachability (Epsilon distance)

0.35
0.6
0.30
0.25
0.4
0.20
0.15
0.2
0.10
0.05
0.0
20

35
Number of data samples Number of data samples

(a) Power level samples. (b) Power difference samples.

Figure 17. The reachability plot for one-day data from House 1 based on the OPTICS algorithm.
Figure 17. The reachability plot for one-day data from House 1 based on the OPTICS algorithm.

6. Results and Discussion

6. Results and Discussion
The developed NILM system isNILM
The developed applied to the
system nQuebec
is applied to thedata.
nQuebec Fordata.
the supervised
For the supervised
phase, all ten houses are evaluated. The developed LSTM model is trained for eachfor
phase, all ten houses are evaluated. The developed LSTM model is trained each
case bycase by
exploiting EBHs data. The main concern over this network, like the other DL structures, is
exploiting EBHs data. The main concern over this network, like the other DL structures, is
the need for sufficient data. This matters in all case studies since the data span only one
the need for sufficient
year. data. This matters
The amount of usefulindata
all reduces
case studies
knowing since the that
the fact datathespan only patterns
abundant one of
year. The amount ofEBHs useful
for learning and testing are mostly available during cold seasons. In order to of
data reduces knowing the fact that the abundant patterns alleviate
EBHs for learning and thistesting
problem,are90%
mostly
of theavailable during cold
data are considered for seasons. In order
training, which can to alleviate
cover more winter
this problem, 90% of samples.
the data For are
the quantification
considered of forthetraining,
outcomes,which
the mean absolute
can covererror
more(MAE), the mean
winter
square error (MSE), the root mean square error (RMSE), and the symmetric mean absolute
samples. For the quantification of the outcomes, the mean absolute error (MAE), the mean
percentage error (sMAPE) are exploited to provide a thorough examination [50]. Table 2
square error (MSE), presents
the rootthe mean square
test results error
of the (RMSE),
Quebec and
houses. ThetheMAE symmetric mean absolute
should be analyzed comparatively.
percentage error (sMAPE)
For instance, the heating demand reaches up to 18kW in House 2, which causes2a high
are exploited to provide a thorough examination [50]. Table
presents the test results of Additionally,
error. the Quebec houses. The MAE
the percentage score should
improves betheanalyzed comparatively.
interpretation of the outcomes.
For instance, the heating demand reaches up to 18kW in House 2, which causes a highin TH
The values of sMAPE along with MAE demonstrate that the model competence
error. Additionally,demand estimation is acceptable in Houses 1, 5, and 9. The difficulty in capturing load
the percentage score improves the interpretation of the outcomes.
peaks entirely is the main source of inaccuracy in these three houses. Nevertheless, their
The values of sMAPE LSTM along
design with MAE demonstrate
outperforms that the
the others regarding MSEmodel competence
and RMSE rates. In in TH the
addition,
demand estimationLSTM is acceptable in Houses
performs fairly 1, 5,3, and
in Houses 6, and9.8 The difficulty
considering in capturing
the large load
power levels of EBHs
peaks entirely is theusage.
main On source of inaccuracy
the other in these
hand, the results three4,houses.
in Houses 7, and 10Nevertheless, theirthe DL
are poor. Generally,
LSTM design outperforms the others regarding MSE and RMSE rates. In addition, the
model’s inadequacy in estimating power spikes can be improved by analyzing thesource
of such behavior. The study shows that these patterns can be related to specific activities
LSTM performs fairly in Houses 3, 6, and 8 considering the large power levels of EBHs
due to their periodicity. Moreover, the network scheme can be enhanced by adding other
sources of information, e.g., descriptive components of power demand since the ongoing
practice relies upon only one input. Such variables can be simply supplied to the HEMS
and used in different operations. It should be added that several LSTM networks with
different structures, data scalers, demand types (energy), and input shapes have been
exercised among which the current model provides better results. Notwithstanding the
limited amount of data, the share of EBHs’ demand in overall usage, and the heating trend
pattern, its sudden fluctuations with large values are critical issues that should be dealt
with in any intended structure.
Sensors 2023, 23, 7288 19 of 24

Table 2. The LSTM results of estimating the overall load of EBHs in Quebec houses.

Data MAE (kW) MSE RMSE (kW) sMAPE (%)

House 1 0.39 0.41 0.64 46
House 2 1.03 2.68 1.65 88
House 3 0.72 1.71 1.31 56
House 4 0.97 2.26 1.50 76
House 5 0.48 0.69 0.83 46
House 6 0.79 1.37 1.17 58
House 7 0.67 0.99 0.99 69
House 8 0.66 0.69 0.83 54
House 9 0.56 0.66 0.81 45
House 10 0.72 1.62 1.27 72

Furthermore, Table 3 lists the results for House 1 over one week in winter based on
MAE, MSE, and the Proportion of Total Energy Correctly Assigned (TECA) metrics [50].
With regard to TECA values, it can be deduced that the regression model performs EBHs
load profiling adequately knowing that a satisfying load disaggregation task should yield a
minimum accuracy of 80% [5]. From the results, it can be stated that the method carries
out the supervised step efficiently since the TH used in House 1 has a high value of 10 kW.
Nevertheless, the assessment error results in two inevitable issues that are passed along
to the next step accounting for inefficiency in catching peaks and variations in estimating
lower demand. Considering the fluctuations in the overall demand, the second matter
has a higher impact on the remaining load. The situation worsens in the presence of
unknown demand. In fact, the share of this type of usage varies from 16% to 49% in
targeted houses (26% in House 1), which is remarkable regarding the power consumption
ranges. The adverse impact of these circumstances on the unsupervised phase is creating
impractical clusters with wide variances in the corresponding region of the load profile.
This is acknowledged by applying the proposed clustering technique to residual demand
across the same week of Table 3 to uncover valid load operations, as shown in Table 4. It is
evident that the first loads represent no actual devices. This is the effect of the error in EBHs
load estimation that makes the recognition of devices with low power use doubtful. This
condition can be the case for any system with a similar strategy since the power of small
loads, such as refrigerators, is normally close to the inaccuracy of examining the demand of
energy-expensive devices like EBHs.

Table 3. The LSTM results of estimating the overall load of EBHs in Quebec House 1.

Day TECA (%) MAE (kW) MSE

1st 86 0.48 0.57
2nd 86 0.41 0.52
3rd 86 0.47 0.37
4th 84 0.53 0.58
5th 84 0.47 0.73
6th 85 0.37 0.36
7th 85 0.38 0.33

Table 4. The proposed clustering method results of identifying other loads in residual signal.

Detected Load (kW) and Related Device

Day
1st 2st 3rd 4th 5th Energy (%)
3.8 2.3 1.9 4.6
1st 0.92 EWH EWH EWH EWH 30
Stove Stove
Sensors 2023, 23, 7288 20 of 24

Table 4. Cont.

Detected Load (kW) and Related Device

Day
1st 2st 3rd 4th 5th Energy (%)
2.9 2.5 1.8
2nd 0.75 - 37
EWH EWH EWH
3rd 0.89 - - - - -
2.9 4.4 3.7
EWH EWH EWH
4th 0.93 - 38
Stove Stove Stove
Dryer
3.8 3.2
EWH EWH
5th 0.82 - - 23
Stove Stove
Dryer
3.1
6th 0.81 - - - 34
EWH
2.9 4.9
7th 0.87 EWH EWH - - 35
Stove

Accordingly, the four last loads are considered for appliance recognition. The worst
outcome occurs on the third day when the sequence creation is not able to identify any
possible load. Nonetheless, the outcomes for such days can be improved since a fixed
Epsilon value is intended for all the days. Besides, the weighted clustering demonstrates
the existence of a pattern in the occurrence of specific loads, for example, ≈2.9 kW and
≈1.9 kw in the second and fourth categories, respectively, which increases their feasibility
as actual appliances. In a 15 min interval, the clustering outcomes reveal the concerns,
discussed above, about load identification in the residual signal. First, the notable loads
become very similar in such a way that their separation is almost impossible within an
unsupervised system. This causes a created sequence to describe more than one device.
Second, the behavior of an appliance changes within consecutive intervals and, thus, the
sequence construction either explains its operation by different loads or fails to capture it
completely; third, the rate of overlapping increases. These situations cause a complicated
interpretation of valid loads in Table 4. It can be noticed that, on the 4th day, one possible
load stands for three actual appliances, while on the 1st day, one actual device is captured
through four possible loads. As expected from the statistical analysis, all the loads in
Table 4 are related to EWH since it holds the highest portion of overall usage. In fact, EWH
either represents a load or shares it with other devices. It should be noted that, in some
cases, this share is very delicate. For instance, on the 4th day, only two sub-sequences of
the relevant created operation are associated with the stove and dryer. On the 7th day,
only one sub-sequence corresponds to the stove. However, the overlapping of the biggest
usages, like the high demands of EWH and dryers, is rarely faced since the developed
clustering procedure considers these kinds of events as outliers due to their low-spatial
density. Furthermore, the last column describes the total amount of energy that is estimated
by all possible loads within a day regarding the residual demand.
According to the results, EWH is the only appliance that can be represented by specific
loads. Such instances that have been made bold in Table 4 bring about an opportunity
to evaluate the identification accuracy of this appliance, as the second major device in
Quebec residences, by NILM metrics. Table 5 presents the results of such an examination
based on the F1-score and TECA measures [50]. It should be noted that the 7th day is
added to this evaluation because of the minor effect of the stove, as explained. In this Table,
F1-scores of total outcomes are also presented to estimate the overall constructed sequence,
i.e., ON-OFF operation states. The last step is subject to all issues, gathered across the whole
Sensors 2023, 23, 7288 21 of 24

process. Therefore, the total results can be expressed as fair for half of the days in Table 5.
Nonetheless, the process should be enhanced with regard to all the opportunities and
challenges, discussed through the analysis of the case studies. Particularly, the challenges
related to blind load disaggregation (the unsupervised step) must be addressed. Load
construction from scratch is a difficult task even in high sampling intervals of less than 1 min.
Indeed, the thorough investigation, conducted by this study, not only reveals the technical
obstacles to developing NILM systems in Quebec but also stimulates the implementation
of the proposed combinatory approach of ILM and NILM. The Quebec case is restricted
by in-use appliances’ characteristics, data availability, and smart meter properties. Under
the suggested load monitoring scheme, suitable data with higher sampling intervals using
smart plugs technologies and additional information through HEMS can be supplied to
accomplish practical applications.

Table 5. The results of the load identification practice applied to only EWH potential loads.

Day Load (kW) F1-Score (%) TECA (%)

3.8 25 54
1st
4.6 25 55
Total 44 -
2.9 30 54
2nd 2.5 21 55
1.8 39 59
Total 62 -
6th 3.1 69 66
2.9 47 57
7th
4.9 39 59
Total 70 -

7. Conclusions
This paper provides a thorough investigation into NILM difficulties in Quebec houses.
It exploits real-world low-sampling rate data from a set of houses, located in Quebec, to
demonstrate the challenging scenario of the case study. It discusses the most important
elements of an NILM for actual applications and deepens its exploration by a comparative
statistical analysis based on popular public databases. In addition, this work proposes a
realistic approach for the case of Quebec by combining the intrusive and non-intrusive
aspects of load monitoring. In order to stimulate its suggestion, the study carries out a
low-sampling load disaggregation practice to manifest the complications of targeting an
NILM-based oriented system for houses in this region. The exercise utilizes a supervised
method to extract the power consumption of EBHs in the first step based on deep learning
models and, subsequently, employs an unsupervised scheme to identify the power demand
of other devices in the remaining load based on clustering techniques. This exhaustive
manuscript explores different techniques for every facet of the problem to provide a fruitful
discussion with elaborated remarks beneficial to practical implementations. Indeed, the
continuation of this study can lead to enhanced load monitoring mechanisms pertinent to
the Quebec case.

Author Contributions: The author’s contribution to the body of this research can be summarized in
terms of conceptualization, S.S.H. and B.D.; methodology, S.S.H., B.D. and N.H.; software, S.S.H.;
validation, S.S.H., B.D., N.H., K.A. and S.K.; formal analysis, S.S.H. and B.D.; investigation, S.S.H.;
resources, B.D., K.A. and S.K.; data curation, S.S.H.; writing—original draft preparation, S.S.H.;
writing—review and editing, S.S.H., B.D. and K.A.; visualization, S.S.H., B.D. and N.H.; supervision,
B.D., N.H., K.A. and S.K.; project administration, B.D. and K.A. All authors have read and agreed to
the published version of the manuscript.
Sensors 2023, 23, 7288 22 of 24

Funding: This research received no external funding.

Data Availability Statement: The data for this research is unavailable due to privacy restrictions.
Acknowledgments: The author would like to thank the Laboratoire des technologies de l’énergie
d’Hydro-Québec, the Natural Science and Engineering Research Council of Canada, and the Foun-
dation of Université du Québec à Trois-Rivières, the iHomeLab—Lucerne University of Applied
Sciences and Arts, and the CLEMAP AG.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Hosseini, S.S.; Agbossou, K.; Kelouwani, S.; Cardenas, A. Non-intrusive load monitoring through home energy management
systems: A comprehensive review. Renew. Sustain. Energy Rev. 2017, 79, 1266–1274. [CrossRef]
2. Zandi, H.; Kuruganti, P.T.V.; Vineyard, E.A.; Fugate, D.L. Home Energy Management Systems: An Overview. In Proceedings of
the Proceedings of the 9th international conference on Energy Efficiency in Domestic Appliances and Lighting, Irvine, CA, USA,
13–15 September 2018; pp. 606–614.
3. Hosseini, S.S.; Agbossou, K.; Kelouwani, S.; Cardenas, A.; Henao, N. A Practical Approach to Residential Appliances on-Line
Anomaly Detection: A Case Study of Standard and Smart Refrigerators. IEEE Access 2020, 8, 57905–57922. [CrossRef]
4. Sakib, N.; Hossain, E.; Ahamed, S.I. A Qualitative Study on the United States Internet of Energy: A Step Towards Computational
Sustainability. IEEE Access 2020, 8, 69003–69037. [CrossRef]
5. Hosseini, S.S.; Kelouwani, S.; Agbossou, K.; Cardenas, A.; Henao, N. Adaptive on-line unsupervised appliance modeling for
autonomous household database construction. Int. J. Electr. Power Energy Syst. 2019, 112, 156–168. [CrossRef]
6. Hosseini, S.S. Load Monitoring and Diagnosis of Residential Electrical Appliances. Ph.D. Thesis, Electrical and Computer
Engineering Department, Université du Québec à Trois-Rivières, Trois-Rivières, QC, Canada, 2020.
7. Oliver Parson. Disaggregated Homes: Overview of the NILM Field. Available online: https://blog.oliverparson.co.uk/2015/03/
overview-of-nilm-field.html (accessed on 17 August 2023).
8. Liaqat, R.; Sajjad, I.A.; Waseem, M.; Alhelou, H.H. Appliance Level Energy Characterization of Residential Electricity Demand:
Prospects, Challenges and Recommendations. IEEE Access 2021, 9, 148676–148697. [CrossRef]
9. Huchtkoetter, J.; Reinhardt, A. A study on the impact of data sampling rates on load signature event detection. Energy Inform.
2019, 2, 24. [CrossRef]
10. Xu, F.; Huang, B.; Cun, X.; Wang, F.; Yuan, H.; Lai, L.L.; Vaccaro, A. Classifier economics of Semi-Intrusive Load Monitoring. Int. J.
Electr. Power Energy Syst. 2018, 103, 224–232. [CrossRef]
11. Angelis, G.F.; Timplalexis, C.; Krinidis, S.; Ioannidis, D.; Tzovaras, D. NILM applications: Literature review of learning approaches,
recent developments and challenges. Energy Build. 2022, 261, 111951. [CrossRef]
12. Faustine, A.; Pereira, L.; Bousbiat, H.; Kulkarni, S. UNet-NILM: A Deep Neural Network for Multi-tasks Appliances State
Detection and Power Estimation in NILM. In Proceedings of the 5th International Workshop on Non-Intrusive Load Monitoring,
Virtual Event, 18 November 2020; pp. 84–88. [CrossRef]
13. Iqbal, H.K.; Malik, F.H.; Muhammad, A.; Qureshi, M.A.; Abbasi, M.N.; Chishti, A.R. A critical review of state-of-the-art
non-intrusive load monitoring datasets. Electr. Power Syst. Res. 2021, 192, 106921. [CrossRef]
14. Völker, B.; Pfeifer, M.; Scholl, P.M.; Becker, B. Annoticity: A Smart Annotation Tool and Data Browser for Electricity Datasets. In
Proceedings of the 5th International Workshop on Non-Intrusive Load Monitoring, Virtual Event, 18 November 2020; pp. 1–5.
[CrossRef]
15. Lin, J.; Ma, J.; Zhu, J.; Liang, H. Deep Domain Adaptation for Non-Intrusive Load Monitoring Based on a Knowledge Transfer
Learning Network. IEEE Trans. Smart Grid 2022, 13, 280–292. [CrossRef]
16. Zhou, G.; Li, Z.; Fu, M.; Feng, Y.; Wang, X.; Huang, C. Sequence-to-Sequence Load Disaggregation Using Multiscale Residual
Neural Network. IEEE Trans. Instrum. Meas. 2021, 70, 2503310. [CrossRef]
17. D’Incecco, M.; Squartini, S.; Zhong, M. Transfer Learning for Non-Intrusive Load Monitoring. IEEE Trans. Smart Grid 2020, 11,
1419–1429. [CrossRef]
18. Rafiq, H.; Shi, X.; Zhang, H.; Li, H.; Ochani, M.K.; Shah, A.A. Generalizability improvement of deep learning-based non-intrusive
load monitoring system using data augmentation. IEEE Trans. Smart Grid 2021, 12, 3265–3277. [CrossRef]
19. Cui, G.; Liu, B.; Luan, W.; Yu, Y. Estimation of Target Appliance Electricity Consumption Using Background Filtering. IEEE Trans.
Smart Grid 2019, 10, 5920–5929. [CrossRef]
20. Afzalan, M.; Jazizadeh, F. Residential loads flexibility potential for demand response using energy consumption patterns and
user segments. Appl. Energy 2019, 254, 113693. [CrossRef]
21. Soares, A.; Gomes, Á.; Antunes, C.H. Categorization of residential electricity consumption as a basis for the assessment of the
impacts of demand response actions. Renew. Sustain. Energy Rev. 2014, 30, 490–503. [CrossRef]
22. Hosseini, S.; Kelouwani, S.; Agbossou, K.; Cardenas, A.; Henao, N. A semi-synthetic dataset development tool for household
energy consumption analysis. In Proceedings of the 2017 IEEE International Conference on Industrial Technology (ICIT), Toronto,
ON, Canada, 22–25 March 2017; pp. 564–569. [CrossRef]
Sensors 2023, 23, 7288 23 of 24

23. Hart, G. Nonintrusive appliance load monitoring. Proc. IEEE 1992, 80, 1870–1891. [CrossRef]
24. Wong, Y.F.; Ahmet Sekercioglu, Y.; Drummond, T.; Wong, V.S. Recent approaches to non-intrusive load monitoring techniques in
residential settings. In Proceedings of the 2013 IEEE Computational Intelligence Applications in Smart Grid (CIASG), Singapore,
16–19 April 2013; pp. 73–79. [CrossRef]
25. Bhotto, M.Z.A.; Makonin, S.; Bajić, I.V. Load Disaggregation Based on Aided Linear Integer Programming. IEEE Trans. Circuits
Syst. II Express Briefs 2017, 64, 792–796. [CrossRef]
26. Balletti, M.; Piccialli, V.; Sudoso, A.M. Mixed-Integer Nonlinear Programming for State-Based Non-Intrusive Load Monitoring.
IEEE Trans. Smart Grid 2022, 13, 3301–3314. [CrossRef]
27. Kaselimi, M.; Protopapadakis, E.; Voulodimos, A.; Doulamis, N.; Doulamis, A. Towards Trustworthy Energy Disaggregation: A
Review of Challenges, Methods, and Perspectives for Non-Intrusive Load Monitoring. Sensors 2022, 22, 5872. [CrossRef]
28. GitHub—Nilmtk/Nilmtk: Non-Intrusive Load Monitoring Toolkit (Nilmtk). Available online: https://nilmtk.github.io/ (ac-
cessed on 17 August 2023).
29. Henao, N.; Agbossou, K.; Kelouwani, S.; Dube, Y.; Fournier, M. Approach in Nonintrusive Type I Load Monitoring Using
Subtractive Clustering. IEEE Trans. Smart Grid 2015, 8, 812–821. [CrossRef]
30. Zhao, B.; Stankovic, L.; Stankovic, V. On a Training-Less Solution for Non-Intrusive Appliance Load Monitoring Using Graph
Signal Processing. IEEE Access 2016, 4, 1784–1799. [CrossRef]
31. Rammig, F.; Ditze, M.; Janacik, P.; Heimfarth, T.; Kerstan, T.; Oberthuer, S.; Stahl, K. Basic Concepts of Real Time Operating
Systems. In Hardware-Dependent Software; Springer: Dordrecht, The Netherlands, 2009; pp. 15–45. [CrossRef]
32. Shin, C.; Rho, S.; Lee, H.; Rhee, W. Data Requirements for Applying Machine Learning to Energy Disaggregation. Energies 2019,
12, 1696. [CrossRef]
33. Zeifman, M.; Roth, K. Nonintrusive appliance load monitoring: Review and outlook. IEEE Trans. Consum. Electron. 2011, 57,
76–84. [CrossRef]
34. Guo, Z.; Wang, Z.J.; Kashani, A. Home appliance load modeling from aggregated smart meter data. IEEE Trans. Power Syst. 2015,
30, 254–262. [CrossRef]
35. Barker, S.; Kalra, S.; Irwin, D.; Shenoy, P. Empirical Characterization, Modeling, and Analysis of Smart Meter Data. IEEE J. Sel.
Areas Commun. 2014, 32, 1312–1327. [CrossRef]
36. Pöchacker, M.; Egarter, D.; Elmenreich, W. Proficiency of power values for load disaggregation. IEEE Trans. Instrum. Meas. 2016,
65, 46–55. [CrossRef]
37. Dinesh, C.; Makonin, S.; Bajic, I.V. Incorporating time-of-day usage patterns into non-intrusive load monitoring. In Proceedings
of the 2017 IEEE Global Conference on Signal and Information Processing (GlobalSIP), Montreal, QC, Canada, 14–16 November
2017; pp. 1110–1114. [CrossRef]
38. Restrepo, J.; Naranjo, F.; Barzola, J.; Otero, C.; Garcia, P. Load Identification for Smart Grid Management. In Proceedings of the
2019 IEEE PES Innovative Smart Grid Technologies Conference—Latin America (ISGT Latin America), Gramado, Brazil, 15–18
September 2019; pp. 1–5. [CrossRef]
39. Rehman, A.U.; Rahman Tito, S.; Nieuwoudt, P.; Imran, G.; Lie, T.T.; Valles, B.; Ahmad, W. Applications of Non-Intrusive Load
Monitoring Towards Smart and Sustainable Power Grids: A System Perspective. In Proceedings of the IEEE 29th Australasian
Universities Power Engineering Conference (AUPEC), Nadi, Fiji, 26–29 November 2019; pp. 1–6. [CrossRef]
40. Henao, N.; Agbossou, K.; Kelouwani, S.; Hosseini, S.; Fournier, M. Power Estimation of Multiple Two-State Loads Using A
Probabilistic Non-Intrusive Approach. Energies 2018, 11, 88. [CrossRef]
41. CleverGuard—Invisible Protection for Pensioners Based on Energy Data—CLEMAP Blog. Available online: https://www.
cleverguard.care (accessed on 17 August 2023).
42. Rafiq, H.; Manandhar, P.; Rodriguez-Ubinas, E.; Barbosa, J.D.; Qureshi, O.A. Analysis of residential electricity consumption
patterns utilizing smart-meter data: Dubai as a case study. Energy Build. 2023, 291, 113103. [CrossRef]
43. Dab, K.; Agbossou, K.; Henao, N.; Dubé, Y.; Kelouwani, S.; Hosseini, S.S. A compositional kernel based gaussian process approach
to day-ahead residential load forecasting. Energy Build. 2022, 254, 111459. [CrossRef]
44. Energy Disaggregation—Bidgely Utility AI—Energy Analytics. Available online: https://www.bidgely.com (accessed on 17
August 2023).
45. Culière, F.; Leduc, L.; Belikov, A. Bayesian model of electrical heating disaggregation. In Proceedings of the 5th International
Workshop on Non-Intrusive Load Monitoring, Virtual Event, 18 November 2020; pp. 25–29. [CrossRef]
46. Deb, C.; Frei, M.; Hofer, J.; Schlueter, A. Automated load disaggregation for residences with electrical resistance heating. Energy
Build. 2019, 182, 61–74. [CrossRef]
47. Hosseini, S.S.; Delcroix, B.; Henao, N.; Agbossou, K.; Kelouwani, S. A case study on obstacles to feasible NILM solutions for energy
disaggregation in quebec residences. In Proceedings of the 9th ACM International Conference on Systems for Energy-Efficient
Buildings, Cities, and Transportation, Boston, MA, USA, 2–3 November 2022; pp. 363–367.
48. Sander, J.; Ester, M.; Kriegel, H.P.; Xu, X. Density-based clustering in spatial databases: The algorithm GDBSCAN and its
applications. Data Min. Knowl. Discov. 1998, 2, 169–194. [CrossRef]
Sensors 2023, 23, 7288 24 of 24

49. Rahmah, N.; Sitanggang, I.S. Determination of Optimal Epsilon (Eps) Value on DBSCAN Algorithm to Clustering Data on
Peatland Hotspots in Sumatra. Iop Conf. Ser. Earth Environ. Sci. 2016, 31, 012012. [CrossRef]
50. Laouali, I.H.; Qassemi, H.; Marzouq, M.; Ruano, A.; Bennani, S.D.; El Fadili, H. A Survey On Computational Intelligence
Techniques For Non Intrusive Load Monitoring. In Proceedings of the 2020 IEEE 2nd International Conference on Electronics,
Control, Optimization and Computer Science (ICECOCS), Kenitra, Morocco, 2–3 December 2020; pp. 1–6. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.

Learning AI Mastery Guide
No ratings yet
Learning AI Mastery Guide
4 pages
8 - 23 - Image Captioning Based On Scene Graphs - A Survey
No ratings yet
8 - 23 - Image Captioning Based On Scene Graphs - A Survey
24 pages
Matlab/Simulink Modeling and Simulation of Electric Appliances Based On Their Actual Current Waveforms
No ratings yet
Matlab/Simulink Modeling and Simulation of Electric Appliances Based On Their Actual Current Waveforms
7 pages
Non-Intrusive Load Monitoring (NILM) Using Deep
No ratings yet
Non-Intrusive Load Monitoring (NILM) Using Deep
6 pages
Transfer Learning For Multiobjective Non Intrusive Load Monitoring in Smart Building
No ratings yet
Transfer Learning For Multiobjective Non Intrusive Load Monitoring in Smart Building
11 pages
Load Monitoring and Activity Recognition in Smart
No ratings yet
Load Monitoring and Activity Recognition in Smart
14 pages
A Framework To Generate and Label Datasets For Non
No ratings yet
A Framework To Generate and Label Datasets For Non
26 pages
Non Intrusive Energy Disaggregation by Detecting Similarities in Consumption Patterns
No ratings yet
Non Intrusive Energy Disaggregation by Detecting Similarities in Consumption Patterns
20 pages
electronics-13-01420
No ratings yet
electronics-13-01420
21 pages
Towards Identification of Appliances in Conventional Homes Using ML and Descriptive Statistics
No ratings yet
Towards Identification of Appliances in Conventional Homes Using ML and Descriptive Statistics
6 pages
Electronics 07 00235 PDF
No ratings yet
Electronics 07 00235 PDF
15 pages
Nonintrusive Load Monitoring Based On Self-Supervised Learning
No ratings yet
Nonintrusive Load Monitoring Based On Self-Supervised Learning
13 pages
Unsupervised Energy Disaggregation Via Convolutional Sparse Coding
No ratings yet
Unsupervised Energy Disaggregation Via Convolutional Sparse Coding
8 pages
Incorporating Appliance Usage Patterns For Non-Intrusive Load Monitoring and Load Forecasting
No ratings yet
Incorporating Appliance Usage Patterns For Non-Intrusive Load Monitoring and Load Forecasting
14 pages
A Review of NILM Applications Withmachine Learning Approaches
No ratings yet
A Review of NILM Applications Withmachine Learning Approaches
19 pages
Sciencedirect Sciencedirect Sciencedirect
No ratings yet
Sciencedirect Sciencedirect Sciencedirect
10 pages
Multi-State Energy Classifier To Evaluate The NILM
No ratings yet
Multi-State Energy Classifier To Evaluate The NILM
18 pages
synopsis IoT
No ratings yet
synopsis IoT
10 pages
Machine Learning Approaches To Non-Intrusive Load Monitoring (Roberto Bonfigli, Stefano Squartini) (Z-Library)
No ratings yet
Machine Learning Approaches To Non-Intrusive Load Monitoring (Roberto Bonfigli, Stefano Squartini) (Z-Library)
142 pages
MATNilm
No ratings yet
MATNilm
10 pages
Residential Electrical Consumption Disaggregation On A Single Low-Cost Meter
No ratings yet
Residential Electrical Consumption Disaggregation On A Single Low-Cost Meter
7 pages
NILM 2017 Paper
No ratings yet
NILM 2017 Paper
23 pages
2022 - 1 s2.0 S2352467722001370 Main
No ratings yet
2022 - 1 s2.0 S2352467722001370 Main
12 pages
Paldi: Online Load Disaggregation Via Particle Filtering
No ratings yet
Paldi: Online Load Disaggregation Via Particle Filtering
11 pages
Murray 2019
No ratings yet
Murray 2019
5 pages
Non-Intrusive Load Monitoring For Energy Consumption Disaggregation
No ratings yet
Non-Intrusive Load Monitoring For Energy Consumption Disaggregation
6 pages
A Practical Solution For Non-Intrusive Type II Load Monitoring Based On Deep Learning and Post-Processing
No ratings yet
A Practical Solution For Non-Intrusive Type II Load Monitoring Based On Deep Learning and Post-Processing
13 pages
Machine Learning Approaches For Non-Intrusive Load
No ratings yet
Machine Learning Approaches For Non-Intrusive Load
27 pages
Unsupervised Algorithm For Disaggregating Low-Sampling-Rate Electricity Consumption of Households
No ratings yet
Unsupervised Algorithm For Disaggregating Low-Sampling-Rate Electricity Consumption of Households
38 pages
Self-Supervised Feature Learning For Appliance Recognition in Non-Intrusive Load Monitoring
No ratings yet
Self-Supervised Feature Learning For Appliance Recognition in Non-Intrusive Load Monitoring
13 pages
kmk.26Nov2018.TCE-2018-09-0422.rev.9
No ratings yet
kmk.26Nov2018.TCE-2018-09-0422.rev.9
10 pages
sensors-23-05226
No ratings yet
sensors-23-05226
25 pages
PHD Thesis
No ratings yet
PHD Thesis
205 pages
Energy Sustainability in Smart Cities Artificial Intelligence, Smart Monitoring, and Optimization of Energy Consumption
No ratings yet
Energy Sustainability in Smart Cities Artificial Intelligence, Smart Monitoring, and Optimization of Energy Consumption
20 pages
Regularized_NCA_based_Prominent_Feature_Selection_
No ratings yet
Regularized_NCA_based_Prominent_Feature_Selection_
12 pages
energies-11-02869-v2
No ratings yet
energies-11-02869-v2
20 pages
10 3390@en12071217
No ratings yet
10 3390@en12071217
18 pages
Applied Sciences: Non-Intrusive Load Disaggregation by Convolutional Neural Network and Multilabel Classification
No ratings yet
Applied Sciences: Non-Intrusive Load Disaggregation by Convolutional Neural Network and Multilabel Classification
17 pages
A Recent Review of NILM Framework Development and Challenges
No ratings yet
A Recent Review of NILM Framework Development and Challenges
7 pages
MTP_p2_aashish
No ratings yet
MTP_p2_aashish
44 pages
Energy Disaggregation From Non-Intrusive Load Monitoring
No ratings yet
Energy Disaggregation From Non-Intrusive Load Monitoring
11 pages
87290 HENRIET 2020 Archivage
No ratings yet
87290 HENRIET 2020 Archivage
187 pages
Dissertation NILM
No ratings yet
Dissertation NILM
112 pages
Buildings 12 01034 With Cover
No ratings yet
Buildings 12 01034 With Cover
36 pages
A Load Identification Method Based On Active Deep Learning and Discrete Wavelet Transform
No ratings yet
A Load Identification Method Based On Active Deep Learning and Discrete Wavelet Transform
11 pages
Journal Pre-Proof: Energy & Buildings
No ratings yet
Journal Pre-Proof: Energy & Buildings
42 pages
A New High-Resolution Dataset For Nilm
No ratings yet
A New High-Resolution Dataset For Nilm
1 page
Energy Efficiency and Predictive Maintenance Appli
No ratings yet
Energy Efficiency and Predictive Maintenance Appli
8 pages
Document 7867
No ratings yet
Document 7867
11 pages
Performance_evaluation_in_non-intrusive_load_monit
No ratings yet
Performance_evaluation_in_non-intrusive_load_monit
17 pages
ApplianceNet a neural network based framework to recognize daily activities_2022
No ratings yet
ApplianceNet a neural network based framework to recognize daily activities_2022
15 pages
Nonintrusive Load Monitoring
No ratings yet
Nonintrusive Load Monitoring
7 pages
Buy ebook Machine Learning Approaches to Non Intrusive Load Monitoring Roberto Bonfigli cheap price
No ratings yet
Buy ebook Machine Learning Approaches to Non Intrusive Load Monitoring Roberto Bonfigli cheap price
55 pages
IET Smart Cities - 2024 - Kommey - An Artificial Intelligence Based Non Intrusive Load Monitoring of Energy Consumption in
No ratings yet
IET Smart Cities - 2024 - Kommey - An Artificial Intelligence Based Non Intrusive Load Monitoring of Energy Consumption in
24 pages
Distributed Facts Device for Flow Controls
From Everand
Distributed Facts Device for Flow Controls
Dr.V.V.L.N. Sastry
No ratings yet
Deep Learning-Based Non-Intrusive Commercial
No ratings yet
Deep Learning-Based Non-Intrusive Commercial
18 pages
Machine Learning Approaches to Non Intrusive Load Monitoring Roberto Bonfigli - Download the ebook now to start reading without waiting
100% (1)
Machine Learning Approaches to Non Intrusive Load Monitoring Roberto Bonfigli - Download the ebook now to start reading without waiting
70 pages
[Ebooks PDF] download (Ebook) Machine Learning Approaches to Non-Intrusive Load Monitoring by Roberto Bonfigli, Stefano Squartini ISBN 9783030307813, 9783030307820, 3030307816, 3030307824 full chapters
100% (8)
[Ebooks PDF] download (Ebook) Machine Learning Approaches to Non-Intrusive Load Monitoring by Roberto Bonfigli, Stefano Squartini ISBN 9783030307813, 9783030307820, 3030307816, 3030307824 full chapters
67 pages
2402.17809v1
No ratings yet
2402.17809v1
9 pages
Analysis and Demand Forecasting of Residential Energy Consumption at Multiple Time Scales
No ratings yet
Analysis and Demand Forecasting of Residential Energy Consumption at Multiple Time Scales
6 pages
Approaches To Non-Intrusive Load Monitoring (Nilm) in The Home
No ratings yet
Approaches To Non-Intrusive Load Monitoring (Nilm) in The Home
53 pages
Non-Intrusive Appliance Load Monitoring Using Low-Resolution Smart Meter Data
No ratings yet
Non-Intrusive Appliance Load Monitoring Using Low-Resolution Smart Meter Data
6 pages
1-s2.0-S2214785322035957-main
No ratings yet
1-s2.0-S2214785322035957-main
7 pages
SZanchetti-LKissIFACMMM07_final2-draft
No ratings yet
SZanchetti-LKissIFACMMM07_final2-draft
7 pages
TheInfluenceofIndustrialEnvironmentalFactorsonSoftRobotMaterials
No ratings yet
TheInfluenceofIndustrialEnvironmentalFactorsonSoftRobotMaterials
16 pages
SoftRoboticsASystematicReviewandBibliometricAnalysis
No ratings yet
SoftRoboticsASystematicReviewandBibliometricAnalysis
44 pages
Characterization_of_Slaughterhouse_Wastewater_and_
No ratings yet
Characterization_of_Slaughterhouse_Wastewater_and_
28 pages
Influence_of_3D-Printing_Strategies_on_the_Compres
No ratings yet
Influence_of_3D-Printing_Strategies_on_the_Compres
6 pages
Recent_progress_in_caron-based_stimulus-responsive
No ratings yet
Recent_progress_in_caron-based_stimulus-responsive
18 pages
JAECh-2006-01
No ratings yet
JAECh-2006-01
10 pages
Evaluating_the_Flexural_Properties_of_Polyoxymethy
No ratings yet
Evaluating_the_Flexural_Properties_of_Polyoxymethy
6 pages
3D_and_4D_Printing_of_PETG-ABS-Fe3O4_Nanocomposite
No ratings yet
3D_and_4D_Printing_of_PETG-ABS-Fe3O4_Nanocomposite
16 pages
Durand-2023_CompositesPartC
No ratings yet
Durand-2023_CompositesPartC
10 pages
Topological_engineered_3D_printing_of_Architectura
No ratings yet
Topological_engineered_3D_printing_of_Architectura
22 pages
2024CompAGDRMecaFib
No ratings yet
2024CompAGDRMecaFib
15 pages
1-s2.0-S2666682023000245-main
No ratings yet
1-s2.0-S2666682023000245-main
11 pages
In_Situ_Tensile_Testing_under_High_Speed_Optical_R
No ratings yet
In_Situ_Tensile_Testing_under_High_Speed_Optical_R
20 pages
Energetic Materials in 3D An In-Depth Exploration
No ratings yet
Energetic Materials in 3D An In-Depth Exploration
23 pages
A_review_on_cost-effective_polyhydroxyalkanoate_PH
No ratings yet
A_review_on_cost-effective_polyhydroxyalkanoate_PH
15 pages
1stPaper
No ratings yet
1stPaper
16 pages
Geometry-Induced Process and Part Characteristics
No ratings yet
Geometry-Induced Process and Part Characteristics
11 pages
Antenna Systems For IoT Applications A Review
No ratings yet
Antenna Systems For IoT Applications A Review
28 pages
Impact of Process Parameters On Improving The Perf
No ratings yet
Impact of Process Parameters On Improving The Perf
30 pages
Assessment of The Development Performance of Addit
No ratings yet
Assessment of The Development Performance of Addit
17 pages
TriboMat2024VorkapiTM 2024 01 04
No ratings yet
TriboMat2024VorkapiTM 2024 01 04
10 pages
In Situ Growth of Carbon Nanotubes On Fly Ash Subs
No ratings yet
In Situ Growth of Carbon Nanotubes On Fly Ash Subs
18 pages
Jin 2024 Machine
No ratings yet
Jin 2024 Machine
18 pages
In-Situ Study of Understanding The Resistive Switc
No ratings yet
In-Situ Study of Understanding The Resistive Switc
30 pages
Investigation of The Mechanical Properties Surface
No ratings yet
Investigation of The Mechanical Properties Surface
23 pages
LBachusACustodi 2003 UsefulWorkFromAPump KnowAndUnderstandCent
No ratings yet
LBachusACustodi 2003 UsefulWorkFromAPump KnowAndUnderstandCent
2 pages
Printing The Future Layer by Layer A Comprehensive
No ratings yet
Printing The Future Layer by Layer A Comprehensive
39 pages
1 s2.0 S0888327024006605 Main
No ratings yet
1 s2.0 S0888327024006605 Main
16 pages
Introducing Machine Learning To Parameter Estimation
No ratings yet
Introducing Machine Learning To Parameter Estimation
6 pages
THESis of NLP
No ratings yet
THESis of NLP
68 pages
CVDL Cae1
No ratings yet
CVDL Cae1
28 pages
Synopsis Phase 1
No ratings yet
Synopsis Phase 1
14 pages
mlt ese
No ratings yet
mlt ese
21 pages
Efficient MIMO Detection With Imperfect Channel Knowledge - A Deep Learning Approach
No ratings yet
Efficient MIMO Detection With Imperfect Channel Knowledge - A Deep Learning Approach
6 pages
Face Recognition Based Attendance System: Problem Statement
No ratings yet
Face Recognition Based Attendance System: Problem Statement
4 pages
Sri Ram - Week 3 Assignment
No ratings yet
Sri Ram - Week 3 Assignment
14 pages
Eswa D 23 08549
No ratings yet
Eswa D 23 08549
29 pages
Artificial Intelligence Midterm Exam
No ratings yet
Artificial Intelligence Midterm Exam
5 pages
PhD_thesis_Fouad_Sakr
100% (1)
PhD_thesis_Fouad_Sakr
170 pages
TS2Vec - Towards Universal Representation of Time Series
No ratings yet
TS2Vec - Towards Universal Representation of Time Series
20 pages
Cellular Neural Network
No ratings yet
Cellular Neural Network
7 pages
Efficient Lipophilicity Prediction of Molecules Employing Deep-Learning Models
No ratings yet
Efficient Lipophilicity Prediction of Molecules Employing Deep-Learning Models
13 pages
Python Title List-2023-2024
No ratings yet
Python Title List-2023-2024
8 pages
SaranPreeth Resume
No ratings yet
SaranPreeth Resume
2 pages
Python Odoo 14 NRW Version Book PDF
No ratings yet
Python Odoo 14 NRW Version Book PDF
22 pages
80066cajournal May2024 17
No ratings yet
80066cajournal May2024 17
8 pages
7. Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes
No ratings yet
7. Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes
12 pages
COVID-19 Image Classification Using VGG-16 & CNN Based On CT Scans
No ratings yet
COVID-19 Image Classification Using VGG-16 & CNN Based On CT Scans
9 pages
A Lip Reading Method Based On 3D Convolutional Vision Transformer
No ratings yet
A Lip Reading Method Based On 3D Convolutional Vision Transformer
8 pages
Fall Detection
No ratings yet
Fall Detection
8 pages
Gesture Recognition for Interactive Presentation Control a Deep Learning and Edge Computing Approach on Raspberry Pi
No ratings yet
Gesture Recognition for Interactive Presentation Control a Deep Learning and Edge Computing Approach on Raspberry Pi
5 pages
Arfaso Birhanu
No ratings yet
Arfaso Birhanu
99 pages
Deep Learning by Author of LSTM
No ratings yet
Deep Learning by Author of LSTM
26 pages
Paper Name: Oid:27992:59096314 Similarity Report ID
No ratings yet
Paper Name: Oid:27992:59096314 Similarity Report ID
36 pages
Value Added Program (VAP) on AI & ML
No ratings yet
Value Added Program (VAP) on AI & ML
8 pages
Project Report Format DBUU
No ratings yet
Project Report Format DBUU
27 pages

Towards_Feasible_Solutions_for_Load_Monitoring_in_

Uploaded by

Towards_Feasible_Solutions_for_Load_Monitoring_in_

Uploaded by

sensors

1 Department of Electrical and Computer Engineering, Laboratoire d’innovation et de Recherche en Énergie

Copyright: © 2023 by the authors.

Sensors 2023, 23, 7288. https://doi.org/10.3390/s23167288 https://www.mdpi.com/journal/sensors

Household level Data flow

Number of Measuring Duration Sampling Frequency

Common training methods

Supervised Semi-supervised Feature

2.2. Learning Procedures

2.3. Operation Modes

2.4. Data Source

2.4. Data Source

3. Quebec Residential Energy Usage Context

3.1. Quebec Residential Data Features

3.2. Quebec Comparative Data Statistics

Main EWH Dryer HP Fridge Main EWH Spa HP Fridge

(a) House 1. (b) House 5.

QH1 QH3 QH5 QH7 QH9

Power conumptoin (kW)

Power conumptoin (kW)

operate in this boundary according to appliance-level information of associated datasets.

Power ranges (kW) Power ranges (kW)

(a) Public databases. (b) Quebec data.

(a) UK-DALE Houses 2 and 4. (b) UK-DALE Houses 1 and 5.

(c) ECO Houses 1, 2, 4 and 5. (d) CG houses.

Sensors 2023, 23, 7288 0.5

Overall demand Trend×Seasonality

-0.02 -0.03 -0.00

0.05 0.11 0.08

0.09 0.10 0.04 0.11

0.47 0.50 0.61 0.57

0.09 0.00 0.15

0.48 0.48 0.46 0.69 0.61 0.68

0.15 0.11 0.08

0.74 0.62 0.59 0.81

0.52 0.54 0.43 0.55 0.50 0.07 -0.02 0.20 0.09

0.81 0.71 0.71 0.83 0.83 0.6

0.77 0.67 0.61 0.72 0.76 0.80

(a) Summer case. (b) Winter case.

Sensors 2023, 23, 7288 13 of 24

0.47 0.50 0.61 0.57

0.48 0.48 0.46 0.69 0.61 0.68

0.74 0.62 0.59 0.81

0.81 0.71 0.71 0.83 0.83 0.6

0.77 0.67 0.61 0.72 0.76 0.80

HC1 HC4 Overall heating Aggregate demand

(a) Summer case. (b) Winter case.

4. A Disaggregation Approach to Quebec Household Power Consumption

Data flow Information

First level Second level

based on NILM to demonstrate load disaggregation complications in Quebec houses by

Figure 14. The

Power level Power difference Power level Power difference

Power consumption (kW)

Number of data samples Number of data samples

(a) Power level samples. (b) Power difference samples.

achability (Epsilon distance)

Reachability (Epsilon distance)

(a) Power level samples. (b) Power difference samples.

6. Results and Discussion

Data MAE (kW) MSE RMSE (kW) sMAPE (%)

Day TECA (%) MAE (kW) MSE

Detected Load (kW) and Related Device

Detected Load (kW) and Related Device

Day Load (kW) F1-Score (%) TECA (%)

Funding: This research received no external funding.

You might also like