NILM-Based Solutions for the Automation of Energy Services in

Residential Buildings

Pierre Ferrez

, Dominique Gabioud

and Pierre Roduit

Institute of Sustainable Energy, University of Applied Sciences and Arts Western Switzerland,

Rue de l’industrie 23, 1950 Sion, Switzerland

{ﬁrstname.name}@hevs.ch

Keywords:

Non Intrusive Load Monitoring (NILM), Heat Pump (HP), Space Heating (SH), Domestic Hot Water (DHW),

Energy Services, Energy Efﬁciency (EE), Fault Detection and Diagnostics (FDD).

Abstract:

In Switzerland, about 40 % (90 TWh) of the energy needs are due to buildings and about 70 % of these needs

come from heating. Therefore, improving the efﬁciency of buildings has a high potential for energy savings.

This paper presents innovative solutions for the automation of energy services in residential buildings, based

solely on the disaggregation of the centralized electricity consumption measurement of the household. The

developed method was successfully used on electricity consumption data recorded in 14 households over a

year to (1) extract the heat pump consumption from the aggregated consumption, (2) split the heat pump

consumption into space heating and domestic hot water consumption and (3) split the remaining consumption

into four categories: base load, low power, medium power, and high power. With a daily average error on the

prediction of the heat pump consumption below 3.5 % and daily average errors on the prediction of the number

of cycles and the operating time of the heat pump both around 1 %, the described method can be used for the

development of energy service prototypes allowing to better understand and optimize the energy functioning

of residential buildings with potential savings for the residents.

1 INTRODUCTION

In Switzerland, greenhouse gas emissions from build-

ings, mainly in the form of carbon dioxide (CO

), are

primarily produced by the use of fossil fuels to heat

buildings and provide hot water in both residential

and commercial buildings (Federal Ofﬁce for the En-

vironment, 2023). Buildings consume yearly around

90 TWh or about 40 % of the ﬁnal energy demand,

and heating (Space Heating (SH) and Domestic Hot

Water (DHW)) accounts for around 70 % of the en-

ergy consumption of buildings (Swiss Federal Ofﬁce

of Energy, 2023). Furthermore, due to poorly con-

ﬁgured or malfunctioning heating systems, and/or the

behaviour of residents, up to approximately 30 % of

this consumption is wasted (Rashid and Singh, 2018;

Bang et al., 2019).

Therefore, improving the efﬁciency of buildings

has a high potential for energy savings, and since

heating appliances contribute to a signiﬁcant part of

the energy consumption of buildings, both Energy

https://orcid.org/0009-0003-0926-4934

https://orcid.org/0000-0002-4279-6428

https://orcid.org/0000-0001-9747-2036

Efﬁciency (EE) and Fault Detection and Diagnos-

tics (FDD) techniques focused on heating appliances

are essential in the development of feedback systems

in order to reduce energy consumption in buildings.

Paradoxically, residents have in most cases a very lim-

ited understanding of their building’s energy system,

as digitalization and smart technologies have hardly

penetrated the building domain. Nevertheless, imple-

menting EE and FDD feedback systems can result in

20 to 30 % of energy savings in buildings (Kim and

Katipamula, 2018) and the potential savings for res-

idents who are provided with detailed feedback on

the individual consumption of their appliances can ex-

ceed 12 % (Carrie Armel et al., 2013).

In order to optimize the efﬁciency of a household,

and before launching into renovations or changes of

often expensive appliances, it is essential to under-

stand the consumption of large consumers such as

SH appliances, DHW appliances, white appliances or

continuously consuming appliances. These data can

be obtained through the installation of dedicated sub-

meters. However, if such installations are feasible in

the framework of a pilot project, they are not realis-

tic on a large scale due to their cost (intervention of

Ferrez, P., Gabioud, D. and Roduit, P.

NILM-Based Solutions for the Automation of Energy Services in Residential Buildings.

DOI: 10.5220/0012733400003714

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2024), pages 85-95

ISBN: 978-989-758-702-3; ISSN: 2184-4968

an electrician) and their complexity (electrical panels

not adapted, data communication problems).

The automatic extraction of these data from the

centralized electricity consumption measurement of

the household (electric meter) would allow their use

on a large scale in order to propose to the ﬁnal cus-

tomer automatic EE services providing, for example,

detailed information on the consumption of families

of appliances or FDD notiﬁcations in case of unusual

or faulty behaviours of appliances such as the heat

pump. This is known as Non Intrusive Load Moni-

toring (NILM) (Hart, 1992), in which individual ap-

pliance consumptions are disaggregated from a cen-

tral recording. NILM is a data-driven technique that

could be used in the implementation of both EE and

FDD feedback systems.

The general objective of this paper is the devel-

opment of innovative NILM-based solutions for the

automation of new EE and FDD services via the im-

plementation of robust and yet simple data analysis

and machine learning methods allowing the extraction

of a maximum of information from the centralized

electricity consumption measurement of the house-

hold (electric meter). More precisely, the objective

of this paper is the development of NILM techniques,

later used in the implementation of new EE and FDD

services, based solely on the global load curve to pro-

vide information to residents on:

• The consumption of the heat pump providing en-

ergy for SH and DHW,

• The SH and DHW consumption,

• The consumption of other families of appliances

for a global feedback on the consumption of the

household,

• Unusual or faulty behaviour of the heat pump,

The focus was put on robust detection of the heat

pump and other families of appliances using simple

solutions enabling its future use in real-time remote

EE and FDD feedback systems. The described dis-

aggregation methods, although simple, are valid and

lead to good detection performance, validated by ref-

erence values obtained using dedicated meters. In-

deed, the reported results on 14 households show a

daily average error on the prediction of the heat pump

consumption below 3.5 %. Furthermore, the daily

average error on the prediction of the number of cy-

cles and the operating time of the heat pump are both

around 1%. The development of EE and FDD ser-

vice prototypes allowing to better understand and op-

timize the energy functioning of residential buildings

with potential savings for the residents can be based

on these results.

2 NILM OVERVIEW

First introduced in 1992 (Hart, 1992), Non Intru-

sive Load Monitoring (NILM) refers to the analysis

of variations in the voltage and/or current going in

a household and deducing what appliances are used

as well as their individual energy consumption. In

other words, NILM consists in the load disaggrega-

tion in order to extract itemized information about a

household’s appliances only using the global electric-

ity consumption measurement.

The most common application of NILM is en-

ergy savings in both residential (Kelly and Knotten-

belt, 2016) and industrial sectors (Garc

ıa-P

erez et al.,

2021). Other applications include smart-grid man-

agement (C¸ imen et al., 2021), load forecasting (We-

likala et al., 2019), fault detection and diagnostics

again in both residential (Aboulian et al., 2019) and

industrial sectors (Rafati et al., 2022). It also in-

cludes human monitoring and assisted living (Her-

nandez et al., 2019), and extraction of consumer be-

haviour (Schirmer and Mporas, 2021).

The NILM approach is a low-cost alternative to

attaching individual meters to each appliance (Intru-

sive Load Monitoring) (Egarter et al., 2015), but be-

ing a data-driven method, it requires high quality con-

sumption measurements in order to provide satisfac-

tory performance.

The complexity of the disaggregation problem is

directly linked to the type of appliances which are

present in the monitored household. Based on their

unique energy consumption signature, electrical ap-

pliances can be grouped into the following four cate-

gories (Zeifman and Roth, 2011):

1. Two State Appliances. These are appliances with

only two operating states: ON and OFF (e.g., re-

sistive loads like lamps).

2. Multiple State Appliances. These are appliances

with a multiple, strictly greater than two but ﬁ-

nite, number of operating states (e.g., washing

machines or stoves).

3. Variable Consumption Appliances. These are

appliances without a ﬁxed number of operating

states, their consumption varies constantly (e.g.,

dimmers).

4. Continuously Consuming Appliances. These

are appliances that remain active at all times (e.g.,

wireless routers).

The sampling frequency of the collected aggregated

consumption data is one of the most important aspects

of NILM since different frequency ranges contain dif-

ferent potentially exploitable signal features. Low

frequencies (0-100Hz) are typically used for features

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

like active power, reactive power or phase angle. Mid

frequencies (0.1-2kHz) are used for features like low

order harmonics or DC component. High frequencies

(2-20kHz) are used for features like harmonic spec-

trum, rise/fall times or transient energy. (Carrie Armel

et al., 2013).

NILM methods can be split into three main cat-

egories for appliance identiﬁcation, namely Machine

Learning, Pattern Matching, and Source Separation.

NILM methods based on Machine Learning use fea-

ture extraction methods to train a model using Sup-

port Vector Machines (SVMs) (Li et al., 2021), De-

cision Trees (DTs) (Al–Khadher et al., 2022), Hid-

den Markov Models (HMMs) (Wu et al., 2021), or

K-Nearest Neighbours (KNNs) (Himeur et al., 2021)

algorithms amongst others. NILM methods based on

Pattern Matching rely on the generation of a database

consisting of a set of reference signatures (Schirmer

et al., 2020). The aim of these methods is to ﬁnd the

best match between an unknown observed signature

and the signatures from the database. With NILM

methods based on Source Separation, the disaggrega-

tion problem is formulated as an optimization prob-

lem (Rahimpour et al., 2017). Constraints like sparse-

ness are used on the optimization algorithm in order to

extract the individual power consumption signatures

of the appliances from the aggregated signal.

The use of public datasets is essential to com-

pare different NILM methods and their performance.

Several public datasets from different countries with

various types and number of appliances, sampling

frequencies, and data collection durations are avail-

able (Kriechbaumer and Jacobsen, 2018; Beckel

et al., 2014; Kolter and Johnson, 2011).

The deﬁnition and use of standardized perfor-

mance metrics is also essential in the comparison of

different NILM methods. For classiﬁcation NILM

algorithms, performance metrics include Accuracy,

False positive and False Negative while for regression

NILM methods, performance metrics include Mean

Absolute Error, Root-Mean-Square Error and Estima-

tion Accuracy.

With accuracy ranging from 60 % to over 96

% (Zhou et al., 2022), the reported performance of

NILM systems strongly vary according to the used

method and mainly to the considered appliances. Un-

surprisingly, performance is also inversely propor-

tional to the number of appliances to be disaggre-

gated. Given the ever growing number of electrical

appliances in modern households, the complexity of

the NILM problem is not likely to decrease in the fu-

ture.

For further and more precise information, several

publications provide thorough reviews of the NILM

technique including recent developments and chal-

lenges (Angelis et al., 2022; Schirmer and Mporas,

2023; Mari et al., 2022; Dash and Sahoo, 2022). In

particular, Rafati recently reviewed the use of NILM

for fault detection and efﬁciency assessment of heat,

ventilation and air conditioning systems (Rafati et al.,

2022). Their investigation shows that NILM tech-

niques can successfully identify electro-mechanical

and electrical faults in such systems that are difﬁcult

to detect by thermal measurements.

3 EXPERIMENTAL METHOD

3.1 Framework

The EU-funded domOS (Operating System for Smart

Services in Buildings) project (domOS Consortium,

2024) takes a closer look at the smart building sec-

tor by researching two axes. The ﬁrst is the technol-

ogy and secure connection of smart devices and smart

appliances so that building owners can enforce pri-

vacy rules. The second addresses the development

of smart services that increase efﬁciency of space

heating. For instance, the project studies how build-

ings can become active nodes of an electricity grid

or a district heating grid. The project’s proposals are

tested on ﬁve demonstration sites, one of them is lo-

cated in Sion, Switzerland. This demonstration site

was initially deployed in the framework of the EU-

funded GOFLEX (Generalized Operational FLEXi-

bility for Integrating Renewables in the Distribution

Grid) project (GOFLEX Consortium, 2020) which

aimed to develop a vertical ﬂexibility chain, from

buildings to grid control centres and energy markets.

This demonstration site is currently being upgraded

and extended in the framework of the domOS project.

3.2 Data Collection

Of the 52 households currently equipped with and

running the domOS monitoring and control system,

14 were selected for this study. The selected house-

holds passed both the following selection criteria:

• Relevance. The household features a heating sys-

tem relevant for the study, i.e., a heat pump pro-

viding energy for both SH and DHW. Further-

more, both the centralized electricity consump-

tion and the speciﬁc electricity consumption of the

heat pump were collected throughout 2022.

• Data Quality. The household features at least

80 % days with good quality data throughout 2022

(at least 292 days). Is considered a good data day

NILM-Based Solutions for the Automation of Energy Services in Residential Buildings

Figure 1: Typical collected data for a winter day. The heat

pump features several SH cycles (around 2 kW) and ﬁve

DHW cycles (around 2.3 kW).

any day (midnight to midnight) with at least 80 %

of the expected data points and no windows of

missing data points longer than an hour.

The centralized electricity consumption was collected

every 5 seconds and features the total active power as

well as the total reactive power. This was collected

using a Landis+Gyr E450 residential advanced me-

ter with an accuracy over 99.7 %. The heat pump

electricity consumption was collected every 15 sec-

onds and only features the active power. This was

collected using a Aeotec Home Energy Meter Gen5

with an accuracy over 99 %. This means that for

any day in 2022 to be kept, the centralized electricity

consumption data must contain at least 13’824 data

points for both the active and reactive power and the

heat pump electricity consumption data must contain

at least 4’608 data points. Figure 1 shows the col-

lected data for a typical winter day.

3.3 Data Processing

For each valid day, the collected data were ﬁrst

cleaned (removal of duplicates, removal of outliers,

etc.) and then downsampled to one data point per

minute. The objective of this downsampling is

twofold, not only it allows synchronisation between

the collected data, but it also reduces the amount of

data being processed without loosing essential infor-

mation. This could become critical for future real-

time remote applications. The resampled data were

ﬁnally ﬁltered using a median ﬁlter with a kernel size

of 5 in order to remove the high frequency compo-

nents of the signal.

3.4 Heat Pump Extraction

Due to the use of induction motors, most heat pumps

consume large amounts of reactive power in compar-

Figure 2: Estimation of the heat pump cycle power

heat pump

. The power consumed by all the other appliances

other

during the heat pump cycle is estimated as the aver-

age of P

be f ore

and P

a f ter

. P

other

is then subtracted from the

average power of the given cycle P

cycle

to provide the esti-

mation of P

heat pump

ison with other typical household appliances. This

feature is particularly interesting when trying to ex-

tract the heat pump consumption from the centralized

consumption. Indeed, it can be seen in Figure 1 that

large amounts of reactive power, typically well over

1.5 kVAR in this case, correspond almost perfectly to

the operation of the heat pump.

The time-wise extraction of the heat pump con-

sumption was therefore simply based on a household-

speciﬁc reactive power threshold. For each times-

tamp, the heat pump is considered in operation if the

value of the global reactive power is above the thresh-

old. For now, the estimation of the household-speciﬁc

reactive power threshold is done by manually inspect-

ing the centralized electricity consumption data of the

ﬁrst available day. This will however be automated in

the future.

The daily number of cycles and operating time of

the heat pump are directly deducted from the time-

wise extraction of the heat pump consumption. To

achieve the full extraction of the heat pump from the

centralized electricity consumption, the system still

needs to estimate the power level of each cycle. This

estimation procedure, detailed below, is illustrated in

Figure 2.

1. The power before the beginning of a heat pump

cycle P

be f ore

is simply estimated as the average

power between the end of the previous cycle of

the heat pump and the beginning of the considered

one. The power after the end of a heat pump cycle

a fter

is estimated the same way as the average

power between the end of the considered cycle of

the heat pump and the beginning of the next one.

2. The power consumed by all the other appliances

other

during the considered heat pump cycle is

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

estimated as the average of P

be f ore

and P

a fter

3. Finally, P

other

is subtracted from the average

power of the given cycle P

cycle

to provide the esti-

mation of the heat pump power P

heat pump

3.5 SH and DHW Separation

Since every heat pump cycle has been assigned a

power level P

heat pump

, it is now possible to classify

them as SH or DHW. This was performed using a two-

class k-means classiﬁer (k=2). Each heat pump cycle

characterised by its P

heat pump

belongs to the cluster

(SH or DHW) with the nearest mean and the clus-

ters means were updated in a pseudo-online system

through every successive day of 2022.

Given a set of cycle powers (P

, P

, ..., P

) and

an initial set of two means m

and m

for two sets

and S

(SH or DHW), the algorithm proceeds by

alternating between two steps:

• Assignation. Each cycle is assigned to the cluster

with the nearest mean.

: |P

− m

| ≤ |P

− m

| i, j = 1, 2

• Update. The means are recalculated with the

powers assigned to each cluster.

∑

∈S

The initial means of the clusters were roughly manu-

ally estimated based on the centralized electricity con-

sumption of the very ﬁrst days of 2022. For example,

from Figure 1, the SH cluster’s mean would have been

initialized to 2000 (2400-400) and the DHW cluster’s

mean would have been initialized to 2300 (2700-400).

As for the estimation of the household-speciﬁc reac-

tive power threshold, this will be automated in the fu-

ture.

3.6 Power Bands Disaggregation

The extraction of the consumption of the heat pump

from the centralized electricity consumption allows

the reconstruction of the heat pump load curve. This

speciﬁc load can then be subtracted from the global

load in order to obtain the aggregated load curve of

the other household’s appliances. Based on an opti-

mized and adapted NILM algorithm previously pub-

lished (Ferrez and Roduit, 2014), the consumption of

the remaining non-heating appliances were disaggre-

gated in to the following four power bands:

• Base Load. This is the daily minimum power

level (> 0), it represents the continuously con-

suming appliances.

• Low Power (≤ 250 W). This includes lighting

and some home electronics.

• Medium Power (> 250 W and < 1 kW). This

includes micro-waves, vacuum cleaners and some

home electronics.

• High Power (≥ 1 kW). This include white appli-

ances and cooking appliances.

3.7 Performance Evaluation

While the extraction of the heat pump from the cen-

tralized electricity consumption and the disaggrega-

tion of the remaining load was performed exclusively

using the collected global active and reactive power,

a speciﬁc measure of the heat pump consumption is

also available. This speciﬁc measure was exclusively

used for performance evaluation purposes. Indeed,

the a priori daily heat pump consumption, daily num-

ber of cycles and daily total operating time can easily

be extracted from this speciﬁc measure and compared

to the results obtained using the centralized electric-

ity consumption. Three different performance metrics

were used, namely the Relative Error (RE), the Mean

Absolute Error (MAE) and the Root Mean Square Er-

ror (RMSE).

4 RESULTS

The centralized electricity consumption of 14 house-

holds was collected throughout 2022 and daily dis-

aggregated into six categories. This section ﬁrst pro-

vides detailed results about one of these households

before showing global results about the 14 house-

holds.

4.1 Single Household

Daily disaggregation results for one of the 14 house-

holds are presented in this ﬁrst subsection. Fig-

ure 3 shows the daily disaggregation throughout 2022

into six categories, namely, base load, low power,

medium power, high power, DHW and SH. Both the

base load and the low power are relatively stable at

roughly 5 kWh (equivalent to a constant power of

200 W). Both the medium and high power are small

but roughly constant over the year. The DHW is

roughly at 5 kWh in the winter, but plummets to 1

to 2 kWh in the summer. Finally, the SH is nil be-

tween May and September, but reaches 20 to 25 kWh

in January and December. These results are conform

to expectations.

NILM-Based Solutions for the Automation of Energy Services in Residential Buildings

Figure 3: Daily disaggregation results for a single house-

hold throughout 2022. The centralized consumption is split

into six categories: base load, low power, medium power,

high power, HDW and SH.

Figure 4: Daily absolute energy error (top), daily absolute

cycle error (center), daily absolute operating time error (bot-

tom) for a single household throughout 2022. A positive

error means overestimation whereas a negative error means

underestimation.

Figure 4 shows the daily absolute error of the pre-

diction of the energy (top), the number of cycles (cen-

ter), and the operating time (bottom) of the heat pump.

A positive error means overestimation whereas a neg-

ative error means underestimation. Overall, the en-

ergy daily absolute error is under 2 kWh and there’s a

tendency to underestimate the energy in winter. These

absolute errors correspond to daily relative errors un-

der 10 %. With regard to cycles, there were only ﬁve

false negatives and three false positive, whereas with

regard to operating time, the absolute error is under

three minutes for most days, corresponding to relative

errors under 5 % for most days.

In a slightly different register, Figure 5 shows the

evolution of the daily predicted heat pump power for

both SH and DHW. Far from being constant, the SH

power ranges from about 2.2 to 2.6 kW with an in-

Figure 5: Daily heat pump power for both SH and DHW for

a single household. Far from being constant, the SH power

ranges from about 2.2 to 2.6 kW whereas the DHW power

ranges from about 2.5 to 3.9 kW. There was no detected SH

between May and September.

terruption between May and September, whereas the

DHW power ranges from about 2.5 to 3.9 kW. Fig-

ure 6 shows the collected data and the results of the

disaggregation for 20 January 2022 (vertical red line

on Figure 5). This ﬁgure shows that all the rather

narrow strictly SH heat pump cycles with an average

power of about 2.4 kW were successfully detected. It

also shows two hybrid cycles with an average power

of about 2.7 kW and mainly contributing to DHW be-

ing detected as such. The nature of the used classiﬁer

forces every heat pump cycle to be either classiﬁed

as SH or DHW. Therefore, hybrid cycles are a source

of error in the estimation of both the SH and DHW

consumption. This will be addresses in the future. Fi-

nally, Figure 6 also shows a base load of 250 W and

the band disaggregation of the remaining load.

Similarly, Figure 7 shows the collected data and

the results of the disaggregation for 21 July 2022 (ver-

tical green line on Figure 5). As expected, there are no

SH cycles, but simply two DHW cycles with a signiﬁ-

cantly higher average operating power, about 3.6 kW.

This ﬁgure also shows a base load of about 180 W and

the band disaggregation of the remaining load.

4.2 All Households

This second subsection provides detailed information

about the performance of the detection of the heat

pump as well as the overall consumption for the 14

households. Table 1 shows RE, MAE and RMSE for

the energy, number of cycles and operating time of the

heat pump for the 14 households in 2022. Absolute

values of the daily errors were used in order to avoid

compensation between positive (overestimation) and

negative (underestimation) errors. The average daily

energy RE is below 5 % for all but one household

whereas both the average daily cycles RE and the the

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

Figure 6: Collected data and disaggregation results for 20

January 2022. All strictly SH cycles (about 2.4 kW) were

successfully detected and two hybrid cycles (about 2.7 kW)

mainly contributing to DHW were detected as such.

Figure 7: Collected data and disaggregation results for 21

July 2022. There was no SH cycles and two successfully

detected DHW cycles (about 3.6 kW).

average daily operating time RE are below 2 % for all

but two households.

The box plots in Figure 8 show the overall disag-

gregation results of the 14 households in 2022 through

their quartiles. The horizontal whiskers indicate vari-

ability outside the lower and upper quartiles and cir-

cles show outliers. The low power appliances and the

SH are the two main contributors to the electricity bill

and both show a large variance. The consumption of

the other four categories is relatively low with a low

to moderate variance.

The pie chart in Figure 9 shows the distribution

of the overall detected consumed energy by the 14

households in 2022. The average total consumption is

8’050 kWh and the heat pump (SH and DHW) is re-

sponsible for 44 % of this consumption (3’585 kWh).

The base load and the low power appliances, cate-

gories with a high potential of consumption reduc-

tion, represent together over a third of the consump-

tion (1’120 and 1’740 kWh, respectively).

Finally, the lowest total consumption was

Figure 8: Overall disaggregation results for the 14 house-

holds for 2022. The horizontal whiskers indicate variability

outside the lower and upper quartiles and circles show out-

liers.

Figure 9: Distribution of the overall detected consumed en-

ergy by the 14 households in 2022. The average total con-

sumption is 8’050 kWh.

4’700 kWh, which is less than 60 % of the av-

erage whereas the highest total consumption was

15’940 kWh, almost twice the average consumption.

These variations reﬂect large differences of consump-

tion behaviours.

5 DISCUSSION

5.1 Algorithm Performance

The general error rates presented in Table 1 as well

as the daily error rates presented in Figure 4 are very

encouraging. Despite the fact that the described de-

tection algorithms are relatively simple, they lead to

good detection performance, not only for the con-

sumed energy, but also for the number of cycles and

the operating time of the heat pump for all 14 house-

holds.

Indeed, with an overall average energy RE below

3.5 %, an overall average cycles RE below 1 % and an

overall average operating time RE just above 1%, the

described heat pump detection method match the per-

NILM-Based Solutions for the Automation of Energy Services in Residential Buildings

Table 1: RE, MAE and RMSE for the energy, number of cycles and operating time of the heat pump for the 14 households

for 2022. RE is expressed in % and both MAE and RMSE are expressed in kWh, cycles and minutes, respectively.

House

Heat Pump Energy Heat Pump Cycles Heat Pump Op. Time

C MAE RMSE RE MAE RMSE RE MAE RMSE

H01 3.18 0.21 0.31 0.07 0.02 0.17 0.68 1.69 4.54

H02 5.56 0.44 0.73 0.03 0.01 0.08 1.81 0.96 1.88

H03 4.80 0.42 0.52 0.00 0.00 0.00 0.23 0.6 1.17

H04 3.38 0.39 0.55 0.84 0.01 0.16 0.23 0.45 1.53

H05 2.88 0.19 0.34 0.72 0.10 0.37 1.5 5.49 11.45

H06 2.69 0.23 0.34 0.02 0.00 0.06 0.79 1.04 2.06

H07 2.25 0.22 0.34 0.03 0.00 0.05 0.33 0.54 0.91

H08 3.56 0.28 0.39 2.15 0.10 0.36 1.03 1.56 3.38

H09 2.02 0.28 0.60 1.25 0.27 0.69 1.99 7.26 12.51

H10 3.83 0.22 0.34 0.78 0.11 0.36 0.41 0.39 0.86

H11 4.81 0.52 1.05 3.48 0.34 0.79 2.21 4.42 9.76

H12 2.51 0.24 0.35 0.00 0.00 0.00 0.43 0.8 1.41

H13 2.23 0.22 0.37 0.00 0.00 0.00 0.44 0.66 1.11

H14 4.36 0.21 0.30 1.47 0.06 0.25 2.59 1.26 1.85

Average 3.43 0.29 0.47 0.77 0.07 0.24 1.05 1.94 3.89

formance of state-of-the-art methods and could there-

fore be used not only for energy efﬁciency services,

but also for fault detection and efﬁciency assessment

of heating system. Indeed, by tracking deviations in

the daily energy consumption and the daily number

of cycles of the heat pump, keeping in mind the natu-

ral seasonal variations, it should be possible to detect

malfunctions.

Thanks to high quality data collection and down-

sampling to one data point per minute, the time-wise

extraction of the heat pump is performing extremely

well as shown by the cycle and operating time er-

ror rates. The rare errors of the time-wise extraction

based simply on a reactive power threshold can be

due to other appliances consuming large mounts of

reactive power. This time-wise detection could be en-

hanced by adding rules regarding what is and is not

a heat pump cycle (e.g., duration, power, number ac-

cording to the season). These rules could be learnt and

updated in a similar way to the SH and DHW power

levels.

As for the error on the predicted energy consumed

by the heat pump, it is the most challenging part of the

detection. Indeed, “guessing” the power level of the

heat pump from the aggregated power is a complex

task. Both data collection systems show accuracies

above 99 % and can therefore only explain part of the

errors. The power estimation of the heat pump cy-

cles could be enhanced ba a better estimation of the

subtracted power of the other appliances or by a more

complex model of the evolution of the power during a

cycle. Indeed, the relatively simple model presented

in this study assumes a constant power during a cycle

whereas power tends to increase during a cycle.

Since only the speciﬁc consumption of the heat

pump was collected using dedicated meters, it is im-

possible to precisely assess the performance of the

disaggregation of the other household’s appliances.

However, this power band disaggregation was per-

formed using an optimized and adapted NILM algo-

rithm previously published [reference blinded] and it

was applied and veriﬁed on the whole set of 14 house-

holds. The daily disaggregation results shown in Fig-

ure 3 and the reconstruction of the speciﬁc load curves

corresponding to the four power bands shown in Fig-

ure 6 and Figure 7 conﬁrm it’s validity for a particular

house.

The objective of this study was not the develop-

ment of new complex NILM algorithms outperform-

ing the best previously published techniques. The

general objective of this study was the robust detec-

tion of the heat pump and other families of appliances

using simple solutions enabling its future use in real-

time remote EE and FDD feedback systems. The er-

ror rates presented for the 14 households in Table 1

show that although the detection method is not per-

fect, it is robust. Furthermore, the main advantage of

this technique is its simplicity and its low computa-

tional requirements. Indeed, it only requires a limited

number of household speciﬁc parameters easily esti-

mated with a couple of days of data, namely the level

of reactive power corresponding to the heat pump op-

eration and the initial means of the clusters of the two-

class k-means classiﬁer. Furthermore, the whole daily

detection can be performed on a standard ofﬁce com-

puter within a few seconds. This is a key issue for a

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

future application in real-time remote feedback sys-

tems.

Finally, to enhance the evaluation of the heat

pump detection technique as well as that of the power

band disaggregation, it would be useful to extend the

analysis to more households, with a broader variety of

heat pumps and other electrical appliances. It would

also be useful to assess the performance of the pro-

posed technique on publicly available datasets featur-

ing households with a heat pump providing energy for

both SH and DHW for a more accurate comparison

with other state-of-the-art methods.

5.2 Overall Consumption

Figure 8 show the overall disaggregation results of the

14 households for 2022 through their quartiles and

Figure 9 presents the distribution of the groups of

appliances (power bands) according to their average

electricity consumption.

Unsurprisingly, the main contributor is SH with

28 %. Maybe more so, the second largest contribu-

tor with 22 % is the low power appliances (below 250

W) mainly including lighting and home electronics.

Adding the 14% of the base load, we reach almost

two third of the overall consumption and these three

categories show a rather large variance, as shown in

Figure 8). These three categories include a large vari-

ety of appliances, but more importantly, the behaviour

of the residents has a strong impact on their impor-

tance. Therefore, these categories have a high poten-

tial for energy savings. Indeed, with a feedback with

the residents in the framework of energy efﬁciency

services, it should be possible to point out which ap-

pliances contribute to the base load in order to reduce

their impact on the energy bill. Similarly, although

the electricity consumption due to home electronics

is ever growing, its impact can be reduced or at least

controlled by behavioural adaptation (e.g., turn off the

TV when nobody is watching it).

With 16 % of the overall consumption, the DHW

is the third largest, just above both the base load and

the high power appliances. More interestingly, it has a

very small variance suggesting that the consumption

of hot water was very similar for the 14 considered

households.

5.3 Energy Feedback Prototype

From the centralized electricity consumption mea-

surement of the household, the described disaggre-

gation methods allowed the reconstruction of the

speciﬁc load curve of six categories of appliances,

namely base load, low power, medium power, high

power, SH and DHW. With these speciﬁc load curves,

it is possible to estimate the energy consumed by any

of these categories for any time window. This is par-

ticularly useful to provide residents with direct feed-

back on their energy consumption, or provide them

with a comparison between days or relatively to ref-

erence values (EE services). Beyond the estimation of

the different consumed energies, the described disag-

gregation methods also allowed the accurate estima-

tion of the number of cycles and operating time of the

heat pump for any time window. This could be the

base of FDD services by tracking the number of cy-

cles of the heat pump, their duration and power over

time and therefore detect anomalies. A screenshot of

a very ﬁrst prototype of a webpage providing energy

consumption feedback is presented in Figure 10. The

top chart shows the energy consumption of the six cat-

egories for a speciﬁc household over a year. The bot-

tom chart shows the six disaggregated load curves for

25 and 26 October 2022 for the same household.

6 CONCLUSION

The general objective of this study was the robust de-

tection of the heat pump and other families of appli-

ances using simple solutions enabling its future use in

real-time remote EE and FDD feedback systems. The

presented NILM technique is simple and yet robust

enough to be used in the implementation of such ser-

vices in order to optimize the energy efﬁciency of res-

idential buildings with potential savings for the resi-

dents. Indeed, with an overall average energy RE be-

low 3.5 %, an overall average cycles RE below 1 %

and an overall average operating time RE just above

1%, the described heat pump detection method is per-

forming well.

It is nevertheless not perfect and can probably be

enhanced without jeopardizing its usability in real-

time remote systems. As mentioned before, the time-

wise detection of the heat pump based on a reactive

power threshold could be enhanced by adding rules

regarding what is and is not a heat pump cycle (e.g.,

duration, power, number according to the season).

The prediction of the energy consumed by the heat

pump could also be enhanced by a more accurate es-

timation of the power consumed by the other appli-

ances during a heat pump cycle and by a more real-

istic power evolution model during a heat pump cy-

cle. Furthermore, the reactive power threshold and

both the initial means of the clusters of the two-class

k-means classiﬁer are so far estimated manually but

this should be automated relatively easily. Finally,

the challenge of hybrid heat pump cycles (providing

NILM-Based Solutions for the Automation of Energy Services in Residential Buildings

Figure 10: First prototype of the energy feedback webpage. The top chart shows the energy consumption of the six categories

for a speciﬁc household over a year. The bottom chart shows the six disaggregated load curves for 25 and 26 October 2022

for the same household.

energy to both SH and DHW) has to be addressed

and the described technique should be tested on more

households on a longer time frame, as well as on

publicly available datasets featuring relevant house-

holds, to better assess its validity and robustness in the

long run and in comparison with other state-of-the-art

methods.

ACKNOWLEDGEMENTS

This work was supported by the Engineering and Ar-

chitecture Faculty and the Institue of Sustainable En-

ergy of HES-SO University of Applied Sciences and

Arts Western Switzerland as well as by the domOS

project funded under the European Union’s Horizon

2020 Programme (GA 894240).

REFERENCES

Aboulian, A., Green, D. H., Switzer, J. F., Kane, T. J.,

Bredariol, G. V., Lindahl, P., Donnal, J. S., and Leeb,

S. B. (2019). NILM Dashboard: A Power System

Monitor for Electromechanical Equipment Diagnos-

tics. IEEE Transactions on Industrial Informatics,

15(3):1405–1414.

Al–Khadher, O., Mukhtaruddin, A., Hashim, F. R., Az-

izan, M. M., Mamat, H., and Saleh, N. (2022). Non-

intrusive load monitoring based on bagging decision

trees and the selective features for commercial build-

ing loads. In 2022 IEEE International Conference on

Power and Energy (PECon), pages 293–298.

Angelis, G.-F., Timplalexis, C., Krinidis, S., Ioannidis, D.,

and Tzovaras, D. (2022). NILM applications: Liter-

ature review of learning approaches, recent develop-

ments and challenges. Energy and Buildings, 261.

Bang, M., Engelsgaard, S. S., Alexandersen, E. K., Skydt,

M. R., Shaker, H. R., and Jradi, M. (2019). Novel real-

time model-based fault detection method for auto-

matic identiﬁcation of abnormal energy performance

in building ventilation units. Energy and Buildings,

183:238–251.

Beckel, C., Kleiminger, W., Cicchetti, R., Staake, T., and

Santini, S. (2014). The ECO Data Set and the Per-

formance of Non-Intrusive Load Monitoring Algo-

rithms. In BuildSys 2014 - Proceedings of the 1st

ACM Conference on Embedded Systems for Energy-

Efﬁcient Buildings.

Carrie Armel, K., Gupta, A., Shrimali, G., and Albert, A.

(2013). Is disaggregation the holy grail of energy

efﬁciency? The case of electricity. Energy Policy,

52:213–234.

Dash, S. and Sahoo, N. (2022). Electric energy disaggre-

gation via non-intrusive load monitoring: A state-of-

the-art systematic review. Electric Power Systems Re-

search, 213:108673.

domOS Consortium (2024). Operating System for Smart

Services in Buildings.

Egarter, D., Bhuvana, V. P., and Elmenreich, W. (2015).

PALDi: Online Load Disaggregation via Particle Fil-

tering. IEEE Transactions on Instrumentation and

Measurement, 64(2):467–477.

Federal Ofﬁce for the Environment (2023). Greenhouse gas

emissions from buildings.

Ferrez, P. and Roduit, P. (2014). Non-Intrusive Appliance

Load Curve Disaggregation for Service Development.

In 2014 IEEE International Energy Conference (EN-

ERGYCON), pages 813–820.

Garc

ıa-P

erez, D., P

erez-L

opez, D., D

ıaz-Blanco, I.,

Gonz

alez-Mu

niz, A., Dom

ınguez-Gonz

alez, M., and

Cuadrado Vega, A. A. (2021). Fully-Convolutional

Denoising Auto-Encoders for NILM in Large Non-

Residential Buildings. IEEE Transactions on Smart

Grid, 12(3):2722–2731.

GOFLEX Consortium (2020). Generalized Operational

SMARTGREENS 2024 - 13th International Conference on Smart Cities and Green ICT Systems

FLEXibility for Integrating Renewables in the Distri-

bution Grid.

Hart, G. (1992). Nonintrusive appliance load monitoring.

Proceedings of the IEEE, 80(12):1870–1891.

Hernandez, A., Ruano, A., Urena, J., Ruano, M., and Gar-

cia, J. (2019). Applications of NILM Techniques

to Energy Management and Assisted Living. IFAC-

PapersOnLine, 52(11):164–171.

Himeur, Y., Alsalemi, A., Bensaali, F., and Amira, A.

(2021). Smart non-intrusive appliance identiﬁcation

using a novel local power histogramming descriptor

with an improved k-nearest neighbors classiﬁer. Sus-

tainable Cities and Society, 67:102764.

Kelly, J. and Knottenbelt, W. (2016). Does disaggregated

electricity feedback reduce domestic electricity con-

sumption? A systematic review of the literature. In

3rd International NILM Workshop.

Kim, W. and Katipamula, S. (2018). A review of fault

detection and diagnostics methods for building sys-

tems. Science and Technology for the Built Environ-

ment, 24(1):3–21.

Kolter, J. and Johnson, M. (2011). REDD: A Public Data

Set for Energy Disaggregation Research. Artiﬁcial In-

telligence, 25.

Kriechbaumer, T. and Jacobsen, H.-A. (2018). BLOND,

a building-level ofﬁce environment dataset of typical

electrical appliances. Scientiﬁc Data, 5.

Li, Y., Yang, Y., Sima, K., Li, B., Sun, T., and Li, X.

(2021). Non-intrusive load monitoring based on har-

monic characteristics. Procedia Computer Science,

183:776–782.

Mari, S., Bucci, G., Ciancetta, F., Fiorucci, E., and Fiora-

vanti, A. (2022). A Review of Non-Intrusive Load

Monitoring Applications in Industrial and Residential

Contexts. Energies, 15(23).

Rafati, A., Shaker, H. R., and Ghahghahzadeh, S. (2022).

Fault Detection and Efﬁciency Assessment for HVAC

Systems Using Non-Intrusive Load Monitoring: A

Review. Energies, 15(1).

Rahimpour, A., Qi, H., Fugate, D., and Kuruganti, T.

(2017). Non-Intrusive Energy Disaggregation Us-

ing Non-Negative Matrix Factorization With Sum-to-

k Constraint. IEEE Transactions on Power Systems,

32(6):4430–4441.

Rashid, H. and Singh, P. (2018). Monitor: An Abnormality

Detection Approach in Buildings Energy Consump-

tion. In 2018 IEEE 4th International Conference on

Collaboration and Internet Computing (CIC), pages

16–25.

Schirmer, P. and Mporas, I. (2021). On the non-intrusive

extraction of residents’ privacy- and security-sensitive

information from energy smart meters. Neural Com-

puting and Applications, 35:1–14.

Schirmer, P., Mporas, I., and Paraskevas, M. (2020). Energy

Disaggregation Using Elastic Matching Algorithms.

Entropy, 22.

Schirmer, P. A. and Mporas, I. (2023). Non-Intrusive Load

Monitoring: A Review. IEEE Transactions on Smart

Grid, 14(1):769–784.

Swiss Federal Ofﬁce of Energy (2023). Buildings.

Welikala, S., Dinesh, C., Ekanayake, M. P. B., Godaliyadda,

R. I., and Ekanayake, J. (2019). Incorporating Ap-

pliance Usage Patterns for Non-Intrusive Load Mon-

itoring and Load Forecasting. IEEE Transactions on

Smart Grid, 10(1):448–461.

Wu, Z., Wang, C., Peng, W., Liu, W., and Zhang, H. (2021).

Non-intrusive load monitoring using factorial hidden

markov model based on adaptive density peak cluster-

ing. Energy and Buildings, 244:111025.

Zeifman, M. and Roth, K. (2011). Nonintrusive appliance

load monitoring: Review and outlook. IEEE Transac-

tions on Consumer Electronics, 57(1):76–84.

Zhou, Y., Li, F., Liu, L., Wang, T., Cheng, Z., Li, R., and

Gao, J. (2022). Non-intrusive load monitoring method

based on the time-segmented state probability. Energy

Reports, 8:1418–1423.

C¸ imen, H., C¸ etinkaya, N., Vasquez, J. C., and Guerrero,

J. M. (2021). A Microgrid Energy Management Sys-

tem Based on Non-Intrusive Load Monitoring via

Multitask Learning. IEEE Transactions on Smart

Grid, 12(2):977–987.

NILM-Based Solutions for the Automation of Energy Services in Residential Buildings