Measurements for non-intrusive load monitoring through machine learning approaches


ACTA IMEKO 
ISSN: 2221-870X 
December 2021, Volume 10, Number 4, 90 - 96 

 

ACTA IMEKO | www.imeko.org December 2021 | Volume 10 | Number 4 | 90 

Measurements for non-intrusive load monitoring through 
machine learning approaches 

Giovanni Bucci1, Fabrizio Ciancetta1, Edoardo Fiorucci1, Simone Mari1, Andrea Fioravanti1 

1 University of L’Aquila, Piazzale E. Pontieri 1, 67100 L’Aquila, Italy 

 

 

Section: RESEARCH PAPER  

Keywords: Non-Intrusive Load Monitoring (NILM); energy management; Deep Learning (DL); Sweep Frequency Response Analysis (SFRA) 

Citation: Giovanni Bucci, Fabrizio Ciancetta, Edoardo Fiorucci, Simone Mari, Andrea Fioravanti, Measurements for non-intrusive load monitoring through 
machine learning approaches, Acta IMEKO, vol. 10, no. 4, article 16, December 2021, identifier: IMEKO-ACTA-10 (2021)-04-16 

Section Editors: Umberto Cesaro and Pasquale Arpaia, University of Naples Federico II, Italy 

Received October 12, 2020; In final form December 6, 2021; Published December 2021 

Copyright: This is an open-access article distributed under the terms of the Creative Commons Attribution 3.0 License, which permits unrestricted use, 
distribution, and reproduction in any medium, provided the original author and source are credited. 

Corresponding author: Simone Mari, e-mail: simone.mari@graduate.univaq.it   

 

1. INTRODUCTION 

Nowadays, economic development has led to a steady 
increase in the demand for electricity and related advanced 
services. Companies are therefore moving towards the 
conversion of their traditional electrical systems into smart 
systems. One of the advantages that can be achieved is an 
efficient use of energy through the programming of loads and 
the awareness of consumption by users. 

 To this end it is necessary to know the information on the 
status and consumption of the various loads powered by the 
system. This can be achieved through intrusive monitoring, i.e. 
by installing individual sensors for each load. But also through 
non-intrusive monitoring, i.e. measuring the total power 
absorbed by the system and deducing the contributions of the 
individual loads from this, through the use of specific algorithms. 
In this second case, an extremely simple and compact 
measurement system is obtained, at the expense of greater 
complexity from the point of view of processing [1]. 

These non-intrusive monitoring systems, however, have 
proven effective on a wide range of applications, which go 
beyond energy management alone [2]. 

Non-intrusive load monitoring systems (NILMs) are used 
successfully in many applications, including demand response 
programs, where consumers can generate profits based on their 
flexibility [3], [4]. Other applications are those of Anomaly 
Detection, to detect malfunctions based on the profiles of power 
absorbed by individual loads [5] and of Condition-based 
maintenance, which has allowed the creation of monitoring 
systems capable of helping operators in maintenance planning 
[6], [7]. Finally, the Ambient Assisted Living is also very 
important, where the NILM system monitors the switching on 
and off of household appliances to infer the position and 
activities of people, detecting the space-time context and, 
therefore, the activities of daily life. of the subject [8]-[10]. 

The first NILM system was proposed by G. Hart in 1985 [11]. 
This algorithm was based on a detection of the edges in the 
aggregate power profile, followed by a clustering operation and 
subsequent matching based on the value of the absorbed power 
and on the on and off time. Clearly this approach, while being 

ABSTRACT 
The topic of non-intrusive load monitoring (NILM) has seen a significant increase in research interest over the past decade, which has 
led to a significant increase in the performance of these systems. Nowadays, NILM systems are used in numerous applications, in 
particular by energy companies that provide users with an advanced management service of different consumption. These systems are 
mainly based on artificial intelligence algorithms that allow the disaggregation of energy by processing the absorbed power signal over 
more or less long time intervals (generally from fractions of an hour up to 24 h). Less attention was paid to the search for solutions that 
allow non-intrusive monitoring of the load in (almost) real time, that is, systems that make it possible to determine the variations in 
loads in extremely short times (seconds or fractions of a second). This paper proposes possible approaches for non-intrusive load 
monitoring systems operating in real time, analysing them from the point of view of measurement. The measurement and post-
processing techniques used are illustrated and the results discussed. In addition, the work discusses the use of the results obtained to 
train machine learning algorithms that allow you to convert the measurement results into useful information for the user. 

mailto:simone.mari@graduate.univaq.it


 

ACTA IMEKO | www.imeko.org December 2021 | Volume 10 | Number 4 | 91 

functional in certain situations, showed significant limitations as 
a multi-state appliance had to be managed as a set of distinct ON 
/ OFF appliances. Conversely, continuously variable appliances 
and appliances with permanent consumption could not be 
detected correctly. It also included strong manual feature 
extraction requirements. 

Subsequently other algorithms based on combinatorial 
optimization were proposed [12], whose main assumption is that 
each load can be in a given state (one of a reduced number K), in 
which each state is associated with a different energy 
consumption. The goal of the algorithm is to assign states to 
household appliances in such a way as to minimize the difference 
between the aggregate power reading and the sum of the energy 
consumption of the different loads. 

In the last decade, thanks to the increase in available 
computing powers, attention has shifted to artificial intelligence 
algorithms, such as hidden-state Markov chains [13]-[15] and 
Deep Learning models [16]-[18]. In particular, the use of Deep 
Learning algorithms has overcome many of the limits that have 
characterized previous methods, thus allowing measurement 
systems to adapt to homes never analysed during the training 
phase. Furthermore, in terms of accuracy, systems based on 
convolutional neural networks have overcome other state-of-
the-art methods, such as those based on Factorial Hidden 
Markov Models [19]. 

However, the state of the art of these systems is represented 
almost exclusively by monitoring systems that process signals 
over time intervals of the order of hours, consequently the 
resulting feedback will not be in real time. Instead, it is often 
necessary to know in real time the status changes of the 
monitored loads. Non-intrusive load monitoring systems are 
therefore required, capable of recognizing the different powered 
devices based on signal processing, during intervals of seconds 
or fractions of seconds.  

In this paper, two approaches for the recognition of electrical 
loads in real time will be presented. The first is a passive 
measurement system, based on the acquisition and processing of 
the current absorbed by the system. The second is an active 
measurement system, based on the measurement of the response 
to a variable frequency signal injected into the system. 

2. NILM SYSTEM BASED ON PASSIVE MEASUREMENTS 

A first attempt was made by creating a recognition system for 
electrical loads based on the analysis of the total absorbed 
current. A system of this type allows to obtain a low cost and 
galvanically isolated measuring system. In steady state conditions, 
the absorbed current does not provide sufficient information to 
characterize a wide range of different loads. This is because the 
waveform of the current absorbed by domestic loads hardly has 
a significant harmonic content. Therefore, the only 
considerations that can be made are based on the difference in 
amplitude. It was therefore decided to characterize the loads on 
the basis of their transient characteristics. Previous studies have 
tried to create NILM systems based on transient characteristics 
[20]-[23], but all have limited the analysis to a reduced number of 
loads and particular cases. In this study, on the other hand, tests 
were conducted on signals acquired from a test plant in which 
five commonly used household appliances were activated and 
deactivated, but above all the performance was also evaluated on 
the basis of the Building-Level fUlly-labeled public dataset for 
Electricity Disaggregation (BLUED) [24]. 

First, the rms value of the current is calculated by processing 
the acquired raw current with a sliding window technique, as 
follows: 

𝐼rms (𝑘) = √
1

𝑁
 ∑ 𝑖(𝑛)

2

𝑘+(𝑁−1)

𝑛=𝑘

 , (1)  

where 𝑘 is the 𝑘-th measured current sample, 𝑁 is the number 
of samples per cycle, 𝑖(𝑛) is the sampled signal, and 𝑛 is the 
summation index. 

The rms value is then derived, and the resulting signal 𝐼rms(𝑘)
′  

is an impulsive signal in which each pulse corresponds to a 
change in state of one of the powered loads. An example is 
shown in Figure 1. 

𝐼rms(𝑘)
′ = 𝐼rms (𝑘) − 𝐼rms (𝑘−1) (2)  

 The position of the pulse in the derived signal identifies the 
moment in which a certain event occurred. 

In this way, the information relating to the steady state is 
filtered and only the information relating to the transients is kept. 
This impulsive signal is successively processed by the Short-time 
Fourier transform (STFT), through the following known 
transformation [25], [26]: 

𝑆𝑇𝐹𝑇(𝑚,𝜔) = ∑ 𝐼rms(𝑛)
′  𝑤(𝑛−𝑚) 𝑒

−𝑗𝜔𝑛

∞

𝑛=−∞

 (3)  

 
Each change in the states of the powered loads is 

characterized and discriminated on the basis of the spectral 
content of the derived signal. The current is processed cyclically 

 

Figure 1. Variation with time of the rms current 𝐼rms (k) (top) and its derivative 

𝐼rms(k)
′  (bottom).  



 

ACTA IMEKO | www.imeko.org December 2021 | Volume 10 | Number 4 | 92 

at 1 second acquisition intervals, following the described 
procedure. Each acquisition slot is processed (to calculate rms 
and the derivative) by adopting an overlap of 500 ms to ensure 
correct analysis. It is also processed for transient events, which 
can be fragmented into two successive slots. The STFT is 
implemented by processing 10-cycles (200 ms) windows with an 
overlap of 4/5 of the processing window. This results in a 
spectrogram with 101 points in frequency at 26 different instants 
of time. To take into account the sign of the change (switching 
on, switching off or passing to a different consumption state), 
the spectrogram is multiplied by the sign of the cumulative sum, 
evaluated on the rms signal as follows: 

𝑆𝑁 = ∑(𝐼rms (𝑛) − 𝐼rms (𝑛−1))

𝑁

𝑛=1

 (4)  

where 𝐼rms (𝑛)  is the rms value of the current described in (1), 𝑁 

is the number of samples, and 𝑆𝑁  is the value of the cumulative 
sum. The final signal 𝑆(𝑖, 𝑗) can be obtained in the form of a 
101 × 26 matrix, as follows: 

𝑆(𝑖, 𝑗) = 𝑆𝑇𝐹𝑇(𝑚,𝜔) ∙ sgn(𝑆𝑁 )

= ∑ 𝐼rms(𝑛)
′  𝑤(𝑛−𝑚) e

−𝑗𝜔𝑛

∞

𝑛=−∞

∙ sgn (∑(𝐼rms (𝑛) − 𝐼rms (𝑛−1))

𝑁

𝑛=1

) 

(5)  

An example of the spectrogram obtained from this procedure 
is shown in Figure 2. 

This spectrogram is used as inputs to a neural network which 
provides a response every 500 ms, indicating the presence or 
absence of events in the signal, and the type of device involved. 

2.1. The adopted Artificial Neural Network 

The deduction of the loads, starting from the spectrogram 
described above, is traced back to a multiclass classification 
problem, that is, a single unique label must be associated with 
each spectrogram. Analysing the current using such a small 
sliding window (1 second with 0.5 second overlap) makes it 
possible to state that within a single window there is no change 
of state of more than one load. Artificial Neural Networks 
(ANNs) are an example of algorithms that natively support 
multiclass classification problems.  

In this work, a particular ANN type, namely, the 
convolutional neural network (CNN), is adopted [27] because of 

its capability of processing complex inputs such as 
multidimensional arrays. More specifically, CNNs are designed 
to exploit the intrinsic properties of some two-dimensional data 
structures, in which there is a correlation between spatially close 
elements (local connectivity). 

The proposed system [28] includes different layers: an input 
level (for signal loading), three groups of layers, each of which 
consisting of convolution, Relu, and max pooling layers (for 
feature extraction from the input), and a group of flatten, fully 
connected, and softmax layers, which uses data from 
convolution layers to generate the output. 

2.2. The proposed system setup 

The proposed measurement system uses an Agilent U2542A 
data acquisition module with a 16-bit resolution. The current 
signal was acquired using a TA SCT-013 current transducer and 
the sampling frequency was set to 10 kHz. The CNN network 
was implemented on a desktop computer (based on the 
Windows 10 x64 operating system) using the open-source 
Python 3.7 from Anaconda.  

Tests were conducted on signals acquired directly from a real 
system, in order to have flexibility both as regards the sampling 
frequency and for the generation of multiple events. Other tests 
were conducted on signals belonging to the public BLUED 
dataset, which features 34 different types of devices. 

The proposed measurement system was installed on a test 
system, designed to generate electrical loads made by domestic 
users, as part of the research project "Non-intrusive 
infrastructures for monitoring loads in residential users". The 
system, which is located in the Electrical Engineering Laboratory 
of the University of L'Aquila (I), allows the generation of 
electrical loads in a single or simultaneous way. These loads 
correspond to the loads generated by the most common 
household appliances and are integrated in a structure similar to 
that of a residential building to reproduce the real problems of 
conditioning and measurement of the signals. 

2.3. The obtained results 

The performance of the NILM system was assessed by 
conducting acquisitions, during which various loads were turned 
ON and OFF for a total of over 519 events. 

Next, BLUED, a public dataset on residential electricity 
usage, was used. This dataset includes voltage and current 
measurements for a single-family house in the United States, 
sampled at 12 kHz for an entire week. 

Regarding NILM systems, no standard and consolidated 
techniques can be found in the literature to evaluate the 
performance of event detectors. Since the purpose of a NILM 
system is to disaggregate consumption for each of the devices in 
question, their performances were analysed to verify the 
achievement of these objectives, which in summary are correct 
identification and classification of the events. These parameters 
were obtained using the number of true positive (TP), false 
positive (FP), true negative (TN), and false negative (FN). In 
addition, the accuracy was assessed as follows: 

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 % =
𝐶𝑜𝑟𝑟𝑒𝑐𝑡 𝑚𝑎𝑡𝑐ℎ𝑒𝑠

𝑇𝑜𝑡𝑎𝑙 𝑝𝑜𝑠𝑠𝑖𝑏𝑙𝑒 𝑚𝑎𝑡𝑐ℎ𝑒𝑠
100 % (6)  

The obtained results are summarized in Table 1. 
 
 
 
 

 

Figure 2. Spectrogram obtained during the switch ON of a microwave oven.  



 

ACTA IMEKO | www.imeko.org December 2021 | Volume 10 | Number 4 | 93 

3. NILM SYSTEM BASED ON ACTIVE MEASUREMENTS 

Subsequently, we tried to recognize the loads powered by an 
electrical system through the Sweep Frequency Response 
Analysis (SFRA). The SFRA is a non-destructive diagnostic 
technique that detects the displacement and deformation of 
windings, among other mechanical and electrical failures, in 
power and distribution transformers. SFRA proceeds by 
applying a sinusoidal voltage signal of constant amplitude and 
variable frequency between one terminal of the bipole under test 
and ground. The response is measured between the other bipole 
terminal and ground. Both input and output signals are acquired 
and processed. The obtained result is the Transfer Function (TF) 
of the bipole over a wide frequency range. A failure is detected 
when a change in the TF is observed.  

The possibility of using these traces to identify which devices 
are powered at the time of measurement was evaluated [29]. The 
basic idea is to detect a change in the load, starting from the 
change in the measured TF. For this purpose, a variable 
frequency sinusoidal signal is applied between the terminal of the 
power phase conductor and ground, by means of the 
instrumentation shown in Figure 3, then both the applied input 
signal and the output signal between the neutral conductor 
terminal and ground are measured and processed. The test 
instrument generates a sinusoidal input signal of constant 
amplitude (a few volt) and frequency variable in the range 
between 10 kHz and 1.5 MHz. The results obtained on the 
electrical system are analysed on a temporal basis, comparing 
them with those previously obtained on the same system. The 
measurement techniques follow the IEC 60076-18 standard [30], 
which regulates the test execution methods, the characteristics of 
the instruments used, the connection methods and the analysis 
of the results. Figure 4 shows the signatures obtained for the 
different types of loads. 

Tests were also conducted in order to evaluate the ability to 
discriminate through these traces different loads when they are 

powered simultaneously. From Figure 5 it is possible to see the 
possibility to discriminate, through these traces, if the heater is 
powered individually or in combination with other loads. 

3.1. The Machine Learning approaches 

In order to translate these traces into useful information for 
the users, the problem was formulated as a multi-label 
classification problem. This is a variant of the classification 
problem, where multiple labels (or multiple classes) may be 
assigned to each instance. Multi-label classification is a 
generalization of multiclass classification, which is the single-
label problem of categorizing instances into precisely one of 
more than two classes. In the multi-label problem, there is no 
constraint on how many of the classes the instance can be 
assigned to. The problem was initially addressed with an ANN 
[31] similar to the one described in the previous section, with 
good results. However, a limitation of the ANNs is the large 
amount of training data required, which makes it difficult to 
apply them in real cases. An attempt was therefore made to use 
another machine learning algorithm, the Support Vector 
Machine (SVM). 

SVM is one of the most popular artificial intelligence 
algorithms and is a supervised learning algorithm used primarily 
for solving classification problems. Unlike generic classification 
algorithms that discriminate on the basis of characteristics 
common to each class, SVM focuses on the samples that are 
most similar to each other but belonging to different classes, 
which are therefore the most difficult samples to discriminate. 

On the basis of these samples, the algorithm constructs an 
optimal hyperplane capable of separating them, and which can 
then be used to discriminate the new samples. These samples are 

Table 1. Scores achieved with the acquired signal and the BLUED dataset. 

 Acquired signal BLUED dataset 

Precision 0.981 0.998 

Recall 0.998 0.998 

F1-Score 0.989 0.998 

Accuracy % 98.0% 87.9% 

 

Figure 3. Instrumentation used for the SFRA.  

 

Figure 4. SFRA tests of different household appliances.  

 

Figure 5. SFRA tests with simultaneous loads powered.  



 

ACTA IMEKO | www.imeko.org December 2021 | Volume 10 | Number 4 | 94 

called support vectors, because they are the only samples that 
support model creation, while all other samples are useless. 

In a two-dimensional case, where the examples to be classified 
are defined by only two characteristics, the optimal hyperplane is 
reduced to a straight line as shown in Figure 6. The algorithm 
searches for the line that maximizes the margin between the two 
examples indicated as support vectors. If it is not possible to 
separate the classes with a straight line, as in the case of nonlinear 
classification problems, the algorithm uses the kernel trick [32]. 
In particular, a polynomial nucleus was chosen for this work, 
thus examining not only the given characteristics of the input 
samples to determine their similarity, but also their 
combinations. 

In the case of the proposed NILM system, the problem is 
obviously not two-dimensional. In fact, the SFRA measuring 
system returns an array of 320 points, which represent the 
transfer functions corresponding to the different frequency 
values. Therefore, the inputs of the SVM is 320, and 
consequently also the size of the problem is 320. To solve the 
problem, to identify which devices are powered starting from the 
result of the SFRA measurement, four SVM classifiers are used, 
each of which performs a binary classification, identifying the 
presence or absence of the device associated with it. 

3.2. The obtained results 

Unlike the NILM system based on passive measurements, in 
this case there is no public dataset that has the necessary 
characteristics, i. e. there is no public dataset of measurements 
obtained through the SFRA technique. Therefore, the 
performance evaluation was made solely on the basis of our 
acquired measurements. As for the previously described system, 
also in this case the same test parameters were used. The 
proposed algorithm was subjected to different scenarios, each for 

a certain number of tests, in which the different appliances were 
powered individually or simultaneously.  

Since, as already explained above, each appliance has an 
associated SVM algorithm that reveals its presence or not, the 
performance of the four algorithms were assessed individually. 
To allow a comparison with the algorithms developed by other 
researchers, precision, recall, F1-Score and accuracy during 
classification were evaluated [33]. The results are shown in Table 
2. 

As far as the ANN is concerned, the evaluation measures for 
a multiclass, hence single-label classification problem, are 
generally different from those for the multiple label. In single-
label classification we can use simple metrics such as precision, 
recall, and accuracy [34]. However, in the multi-label 
classification, an incorrect classification is no longer a real error, 
as a forecast containing a subset of the actual classes is certainly 
better than a forecast that does not, i.e. correctly predicting two 
of the four labels is better than foresee the absence of labels. To 
evaluate the performance of a multi-label classifier we have to 
calculate the average of the classes. There are two different 
methods of doing this called micro-averaging and macro-
averaging [35]. The current is processed cyclically at 1 second 
acquisition the metric independently for each class and then take 
the average hence treating all classes equally, whereas the micro-
average will aggregate the contributions of all classes to compute 
the average metric. In a multi-label classification setup, micro-
average is preferable if there is a suspicion that there may be a 
class imbalance (i.e. the possibility of having many more 
examples of a class than other classes). In the cases under 
examination, this problem does not exist as the examples used 
for training and testing are sufficiently uniform, so micro-average 
and macro-average can both be considered reliable. 

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛micro−averaging =
∑ 𝑇𝑃𝑛

𝑁
𝑛=1

∑ 𝑇𝑃𝑛 + 𝐹𝑃𝑛
𝑁
𝑛=1

 (7)  

𝑅𝑒𝑐𝑎𝑙𝑙micro−averaging =
∑ 𝑇𝑃𝑛

𝑁
𝑛=1

∑ 𝑇𝑃𝑛 + 𝐹𝑁𝑛
𝑁
𝑛=1

 (8)  

𝐹1 − 𝑠𝑐𝑜𝑟𝑒micro−averaging

=
2 ×  𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛micro−averaging  ×  𝑅𝑒𝑐𝑎𝑙𝑙micro−averaging

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛micro−averaging +  𝑅𝑒𝑐𝑎𝑙𝑙micro−averaging
 (9)  

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛macro−averaging =
∑ 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑛

𝑁
𝑛=1

𝑁
 (10)  

𝑅𝑒𝑐𝑎𝑙𝑙macro−averaging =
∑ 𝑅𝑒𝑐𝑎𝑙𝑙𝑛

𝑁
𝑛=1

𝑁
 (11)  

𝐹1 − 𝑠𝑐𝑜𝑟𝑒macro−averaging

=
2 ×  𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛macro−averaging  ×  𝑅𝑒𝑐𝑎𝑙𝑙macro−averaging

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛macro−averaging +  𝑅𝑒𝑐𝑎𝑙𝑙macro−averaging
 (12)  

 

Figure 6. Representation of a linear classification problem (top) and a non-
linear classification problem (bottom) in which the samples are defined by 
only two features.  

Table 2. Achieved scores with the SVM. 

 SVM 
Lamp 

SVM 
Hairdryer 

SVM 
Induction Hob 

SVM 
Heater 

TP 42 200 200 200 

FP 0 0 0 2 

TN 400 250 250 248 

FN 8 0 0 0 

Precision 1 1 1 0.99 

Recall 0.84 1 1 1 

F1-Score 0.91 1 1 0.99 



 

ACTA IMEKO | www.imeko.org December 2021 | Volume 10 | Number 4 | 95 

Table 3 shows the scores achieved according to the two 
criteria both with the ANN and with the SVM. 

4. CONCLUSIONS AND FINAL REMARKS 

In this paper a brief introduction of the state of the art of 
NILM systems has been presented. Two different types of 
systems for real-time identification of electrical loads, based on 
different measurement techniques, were then presented. Both 
systems were proven excellent identification performance. 

More in detail, the first system, based on the spectrogram 
analysis of the effective current through the CNN, has been 
proven excellent performance in correspondence with both the 
acquired measurements and those available in the BLUED 
dataset, reaching F1-Score respectively equal to at 0.989 and 
0.998 and Accuracy% respectively equal to 98.0% and 87.9%.  

The greatest difficulty encountered in the classification phase 
with the BLUED dataset is attributable to the significantly 
greater number of devices that the network is required to 
recognize, compared to those used for the acquired 
measurements. Furthermore, the value obtained for the F1-score 
is higher than that obtained with other systems using the same 
dataset, as those proposed in [36] (0.915) and [37] (0.932).  

Traditional NILM systems perform the loads classification 
based on the analysis of quantities also related to voltage (e.g. 
analysis in the P-Q or V-I plan [38]). The proposed system has 
the advantage of measuring only the overall current in a house. 
As a result, the complexity of the processing system is reduced. 
Another advantage is that the measuring system can be 
implemented with a galvanically isolated at low-cost system, 
using a clamp current transducer. 

The second proposed system, based on the trace analysis 
provided by the SFRA, also proved excellent performance. The 
traces were initially processed through an artificial neural 
network similar to that used for the previous system, reaching 
F1-Score of 0.96. In order to reduce the number of training 
examples needed, it was decided to use a Support Vector 
Machine. Despite a significant reduction in the examples needed 
for training (from over 2000 to 90), the F1-Score achieved with 
this second machine learning structure was even higher than that 
obtained with the artificial neural network.  

A system of this type is particularly interesting as it allows the 
creation of a plug-in solution that can be installed in any 
domestic, industrial or commercial environment. Furthermore, 
the detection technique takes into account the physical 
characteristics of household appliances and the resulting transfer 
function. Consequently, the identification of multi-state or 
continuously variable appliances is simplified, compared to 
processing time-varying signals such as real power, current, etc. 

REFERENCES 

[1] G. Bucci, F. Ciancetta, E. Fiorucci and S. Mari, Load identification 
system for residential applications based on the NILM technique, 

2020 IEEE International Instrumentation and Measurement 
Technology Conference (I2MTC), 25-28 May 2020, pp. 1-6. 
DOI: 10.1109/I2MTC43012.2020.9128599 

[2] G. Bucci, F. Ciancetta, E. Fiorucci, S. Mari, A. Fioravanti, State of 
art overview of Non-Intrusive Load Monitoring applications in 
smart grids, Measurement: Sensors 18(2021), art. no. 100145. 
DOI: 10.1016/j.measen.2021.100145 

[3] A. Lucas, L. Jansen, N. Andreadou, E. Kotsakis, M. Masera, Load 
flexibility forecast for DR using non-intrusive load monitoring in 
the residential sector, Energies 12(14) (2019), art. no. 2725. 
DOI: 10.3390/en12142725 

[4] W. Schneider, F. Campello de Souza, Non-intrusive load 
monitoring for smart grids, 2018. Technical report. DELL EMC. 

[5] H. Rashid, P. Singh, V. Stankovic, L. Stankovic, Can non-intrusive 
load monitoring be used for identifying an appliance's anomalous 
behaviour?, Applied Energy 238 (2019), pp. 796-805. 
DOI: 10.1016/j.apenergy.2019.01.061 

[6] D. Green, T. Kane, S. Kidwell, P. Lindahl, J. Donnal, S. Leeb, 
NILM dashboard: Actionable feedback for condition-based 
maintenance, IEEE Instrumentation & Measurement Magazine 
23(5) (2020), pp. 3-10. 
DOI: 10.1109/MIM.2020.9153467 

[7] A. Aboulian et al., NILM dashboard: A power system monitor for 
electromechanical equipment diagnostics, IEEE Transactions on 
Industrial Informatics 15(3) (2019), pp. 1405-1414. 
DOI: 10.1109/TII.2018.2843770 

[8] C. Belley, S. Gaboury, B. Bouchard, A. Bouzouane, An efficient 
and inexpensive method for activity recognition within a smart 
home based on load signatures of appliances, Pervasive and 
Mobile Computing 12 (2014), pp. 58-78. 
DOI: 10.1016/j.pmcj.2013.02.002 

[9] N. Noury, M. Berenguer, H. Teyssier, M. Bouzid,M. Giordani, 
Building an index of activity of inhabitants from their activity on 
the residential electrical power line, IEEE Transactions on 
Information Technology in Biomedicine 15(5) (2011), pp. 758-
766. 
DOI: 10.1109/TITB.2011.2138149 

[10] X. Zhang, T. Kato, T. Matsuyama, Learning a context-aware 
personal model of appliance usage patterns in smart home,2014 
IEEE Innovative Smart Grid Technologies - Asia (ISGT ASIA), 
Kuala Lumpur, Malaysia, 20-23 May 2014, pp. 73-78. 
DOI: 10.1109/ISGT-Asia.2014.6873767 

[11] G. Hart. 1985. Prototype nonintrusive appliance load monitor. In 
MIT Energy Laboratory Technical Report, and Electric Power 
Research Institute Technical Report. 

[12] G. W. Hart. 1992. Nonintrusive appliance load monitoring. Proc. 
IEEE 80, 12 (Dec. 1992), 1870–1891. 
DOI: 10.1109/5.192069 

[13] J Zico Kolter and Matthew J Johnson. 2011. REDD: A public data 
set for energy disaggregation research. 

[14] Oliver Parson, Siddhartha Ghosh, Mark Weal, Alex Rogers, 
Nonintrusive load monitoring using prior models of general 
appliance types, 2012 Twenty-Sixth AAAI Conference on 
Artificial Intelligence, Toronto, Canada, 22-26 July 2012. 

[15] Mingjun Zhong, Nigel Goddard, and Charles Sutton. 2014. Signal 
Aggregate Constraints in Additive Factorial HMMs, with 
Application to Energy Disaggregation. In Advances in Neural 
Information Processing Systems 27, Z. Ghahramani, M. Welling, 
C. Cortes, N. D. Lawrence, and K. Q. Weinberger (Eds.). Curran 
Associates, Inc., 3590–3598. 

[16] J. Kelly, W. Knottenbelt, Neural NILM: Deep neural networks 
applied to energy disaggregation, 2nd ACM International 
Conference on Embedded Systems for Energy-Efficient Built 
Environments (BuildSys '15). Association for Computing 
Machinery, New York, NY, USA, pp. 55–64.  
DOI: 10.1145/2821650.2821672 

[17] Z. Jia, L. Yang, Z. Zhang, H. Liu, F. Kong, Sequence to point 
learning based on bidirectional dilated residual network for non-
intrusive load monitoring, International Journal of Electrical 
Power & Energy Systems 129(2021), art. no. 106837. 

Table 3. Comparison between SVM and ANN. 

 SVM ANN 
 Micro-

averaging 
Macro-

averaging 
Micro-

averaging 
Macro-

averaging 

Precision 0.99 0.99 0.94 0.91 

Recall 0.98 0.96 0.99 0.99 

F1-Score 0.98 0.97 0.96 0.95 

https://doi.org/10.1109/I2MTC43012.2020.9128599
https://doi.org/10.1016/j.measen.2021.100145
https://doi.org/10.3390/en12142725
https://doi.org/10.1016/j.apenergy.2019.01.061
http://dx.doi.org/10.1109/MIM.2020.9153467
https://doi.org/10.1109/TII.2018.2843770
https://doi.org/10.1016/j.pmcj.2013.02.002
https://doi.org/10.1109/TITB.2011.2138149
https://doi.org/10.1109/ISGT-Asia.2014.6873767
https://doi.org/10.1109/5.192069
https://doi.org/10.1145/2821650.2821672


 

ACTA IMEKO | www.imeko.org December 2021 | Volume 10 | Number 4 | 96 

DOI: 10.1016/j.ijepes.2021.106837 
[18] G. Bucci, F. Ciancetta, E. Fiorucci, S. Mari, A. Fioravanti, Multi-

state appliances identification through a NILM system based on 
convolutional neural network, 2021 IEEE Instrumentation and 
Measurement Technology Conference I2MTC 2021, 17-21 May 
2021. 
DOI: 10.1109/I2MTC50364.2021.9460038 

[19] C. Zhang, M. Zhong, Z. Wang, N. Goddard, C. Sutton, Sequence-
to-point learning with neural networks for non-intrusive load 
monitoring, AAAI Conf. Artif. Intell., New Orleans, LA, USA, 
Feb. 2018, pp. 2604–2611. 

[20] H. Chang, K. Chen, Y. Tsai and W. Lee, A new measurement 
method for power signatures of nonintrusive demand monitoring 
and load identification, IEEE Transactions on Industry 
Applications 48(2) (2012), pp. 764-771. 
DOI: 10.1109/TIA.2011.2180497 

[21] Y. Lin, M. Tsai, Development of an improved time–frequency 
analysis-based nonintrusive load monitor for load demand 
Identification, IEEE Transactions on Instrumentation and 
Measurement 63(6) (2014), pp. 1470-1483. 
DOI: 10.1109/TIM.2013.2289700 

[22] H. Chang, K. Lian, Y. Su, W. Lee, Power-spectrum-based wavelet 
transform for nonintrusive demand monitoring and load 
identification, IEEE Transactions on Industry Applications 50(3) 
(2014), pp. 2081-2089. 
DOI: 10.1109/TIA.2013.2283318 

[23] S. B. Leeb, S. R. Shaw, J. L. Kirtley, Transient event detection in 
spectral envelope estimates for nonintrusive load monitoring, 
IEEE Transactions on Power Delivery 10(3) (1995), pp. 1200-
1210. 
DOI: 10.1109/61.400897 

[24] K. Anderson, A. Ocneanu, D. Benitez, D. Carlson, A. Rowe, M. 
Bergés, BLUED: a fully labeled public dataset for event-based 
non-intrusive load monitoring research, 2nd KDD Workshop on 
Data Mining Applications in Sustainability (SustKDD), Beijing, 
China, Aug. 2012, pp. 1–5. 

[25] W. Yuegang, J. Shao, X. Hongtao, Non-stationary signals 
processing based on STFT, 8th International Conference on 
Electronic Measurement and Instruments, Xi'an, 2007, pp. 3-301–
3-304.  
DOI: 10.1109/ICEMI.2007.4350914 

[26] S. Zhang, D. Yu, S. Sheng, A discrete STFT processor for real-
time spectrum analysis, APCCAS 2006 - 2006 IEEE Asia Pacific 
Conference on Circuits and Systems, Singapore, 4-7 Dec. 2006, 
pp. 1943–1946. 
DOI: 10.1109/APCCAS.2006.342241 

[27] S. Albawi, T. A. Mohammed, S. Al-Zawi, Understanding of a 
convolutional neural network, International Conference on 
Engineering and Technology (ICET), Antalya, Turkey, 21-23 Aug. 
2017, pp. 1–6.  

DOI: 10.1109/ICEngTechnol.2017.8308186 
[28] F. Ciancetta, G. Bucci, E. Fiorucci, S. Mari, A. Fioravanti, A new 

convolutional neural network-based system for NILM 
applications, IEEE Transactions on Instrumentation and 
Measurement 70 (2021), art no. 1501112.  
DOI: 10.1109/TIM.2020.3035193 

[29] A. Fioravanti, A. Prudenzi, G. Bucci, E. Fiorucci, F. Ciancetta, S. 
Mari, Non-intrusive electrical load identification through an online 
SFRA based approach, 2020 International Symposium on Power 
Electronics, Electrical Drives, Automation and Motion 
(SPEEDAM), Sorrento (Italy), 24–26 June 2020, pp. 694–698. 
DOI: 10.1109/SPEEDAM48782.2020.9161856 

[30] IEC 60076-18:2012 Power transformers - Part 18: Measurement 
of frequency response 

[31] G. Bucci, F. Ciancetta, E. Fiorucci, S. Mari, A. Fioravanti, Deep 
Learning applied to SFRA Results: a Preliminary Study, 7th 
International Conference on Computing and Artificial Intelligence 
ICCAI 2021, Tianjin, China, 23-26 April 2021, pp. 302-307 
DOI: 10.1145/3467707.3467753 

[32] T. Hofmann et al., Kernel methods in machine learning, Ann. 
Statist. 36(3) (2008), pp. 1171–1220. 
DOI: 10.1214/009053607000000677 

[33] G. Bucci, F. Ciancetta, E. Fiorucci, S. Mari, A. Fioravanti, A non-
intrusive load identification system based on frequency response 
analysis, 2021 IEEE International Workshop on Metrology for 
Industry 4.0 & IoT (MetroInd4.0&IoT), 7-9 June 2021, pp. 254-
258. 
DOI: 10.1109/MetroInd4.0IoT51437.2021.9488472 

[34] S Makonin, F Popowich, Nonintrusive load monitoring (NILM) 
performance evaluation, Energy Efficiency 8(4) (2014), pp. 809–
814. 
DOI: 10.1007/s12053-014-9306-2 

[35] O. Koyejo, N. Natarajan, P. K. Ravikumar, I. S. Dhillon, 
Consistent multilabel classification, in Proc. NIPS, 2015, pp. 
3321–3329. 

[36] M. A.Peng, H. Lee, Energy disaggregation of overlapping home 
appliance consumptions using a cluster splitting approach, 
Sustainable Cities and Society 43 (2018), pp. 487–494. 
DOI: 10.1016/j.scs.2018.08.020 

[37] K. Jain, S. S. Ahmed, P. Sundaramoorthy, R. Thiruvengadam, V. 
Vijayaraghavan, Current peak based device classification in NILM 
on a low-cost embedded platform using extra-trees, IEEE MIT 
Undergraduate Research Technology Conference (URTC), 
Cambridge, MA, -5 Nov. 2017, pp. 1–4.  
DOI: 10.1109/URTC.2017.8284200 

[38] T. Hassan, F. Javed, N. Arshad, An empirical investigation of V-I 
trajectory-based load signatures for non-intrusive load monitoring, 
IEEE Transactions on Smart Grid 5(2) (2014), pp. 870–878. 
DOI: 10.1109/PESGM.2014.6938824 

 

https://doi.org/10.1016/j.ijepes.2021.106837
https://doi.org/10.1109/I2MTC50364.2021.9460038
https://doi.org/10.1109/TIA.2011.2180497
https://doi.org/10.1109/TIM.2013.2289700
https://doi.org/10.1109/TIA.2013.2283318
https://doi.org/10.1109/61.400897
https://doi.org/10.1109/ICEMI.2007.4350914
https://doi.org/10.1109/APCCAS.2006.342241
https://doi.org/10.1109/ICEngTechnol.2017.8308186
https://doi.org/10.1109/TIM.2020.3035193
https://doi.org/10.1109/SPEEDAM48782.2020.9161856
https://doi.org/10.1145/3467707.3467753
https://doi.org/10.1214/009053607000000677
https://doi.org/10.1109/MetroInd4.0IoT51437.2021.9488472
https://doi.org/10.1007/s12053-014-9306-2
https://doi.org/10.1016/j.schres.2018.08.020
https://doi.org/10.1109/URTC.2017.8284200
https://doi.org/10.1109/PESGM.2014.6938824