Proceedings of the Third International Conference on Smart Electronics and Communication (ICOSEC 2022)

IEEE Xplore Part Number: CFP22V90-ART; ISBN: 978-1-6654-9764-0

Non-Intrusive Load Monitoring for Energy

Consumption Disaggregation
2022 3rd International Conference on Smart Electronics and Communication (ICOSEC) | 978-1-6654-9764-0/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICOSEC54921.2022.9951891

Aravind P R Sarath T V
Department of Electrical and Electronics Engineering Department of Electrical and Electronics Engineering
Amrita School of Engineering, Coimbatore Amrita School of Engineering, Coimbatore
Amrita Vishwa Vidyapeetham, India Amrita Vishwa Vidyapeetham, India
[email protected] [email protected]

Abstract – The smart grid offers a venue for reducing the reduced tariff rates. This will motivate the consumers to plan
disparity in demand and generation by demand response their usage towards the subsidized tariff. The power
initiatives. The efficacy of demand response algorithms relies consumption of the consumer is fetched by the utility and the
on identifying the active non-essential loads at consumer options to participate in the energy market are put forward by
premises during peak hours. Hence, separating the electricity the electric utility. The installation of smart energy meters
usage of a household into its individual appliance consumption helps us in this matter. Traditional meters just measure the
is essential for facilitating demand response. Non-intrusive power consumption which had to be recorded physically.
load monitoring (NILM) is the widely adopted methodology for While smart meters can measure consumption at regular
the disaggregation of power consumption. This would
intervals and can also transmit this data back to the grid. The
consequently help the consumers to manage their energy usage.
This paper has implemented and compared two deep learning
energy meters will only provide the aggregate or total power
architectures, CNN and Bi-GRU network for energy consumption of a particular household. This will only give a
consumption disaggregation. Standard UKDALE dataset is broad picture of the scenario. If consumer wishes to reduce
used for the training and testing of these architectures. The their consumption, they should know exactly where and how
complex nature of the Bi-GRU network identified appliances their energy is getting consumed excessively. So, the next
with sporadic activity nature whereas CNN performed better step is to disaggregate the consumption at the appliance
in appliances that exhibit periodicity. level, broadly known as load monitoring.
Load monitoring can be classified into two categories
Keywords— Non-Intrusive Load Monitoring, Appliance
intrusive and non-intrusive approaches. The former approach
Monitoring, Energy Disaggregation.
deploys sensors to every appliance in every household for
I. INTRODUCTION measurement. Setting up such a process can be tedious and
time-consuming. The latter is the non-intrusive approach
The power grid enables the power transaction among which takes the help of existing smart grid infrastructure and
utilities and consumers. Various fields such as medicine, machine learning algorithms to estimate energy
manufacturing, banking, water treatment, etc. are dependent consumption. So Non-intrusive Load Monitoring (NILM) is
on power systems. The advancement in technologies has a much more feasible approach [6]. The disaggregation of
caused the increase in demand for power and hence led to the power consumption will allow the consumer to know where
phenomenon of energy crisis. The traditional power systems and how the power goes precisely instead of just having a
are not able to meet the demands of their consumers. The top-level idea. This can also help provide the utilities with
integration of renewable sources like wind and solar has the knowledge of the capability of reducing power
helped up to some extent. With the addition of newer consumption in each household.
technologies, the usage of information has also become a
crucial commodity. Appropriate communication The rise of deep learning has led to the development of a
technologies are required to handle the flow of information wide range of applications. Neural networks are designed
and tackle other challenges such as reliability, security, and specifically for a particular application to extract information
cost [1]. So, as a solution to the above-mentioned scenario, a from raw data (text, image, audio, or video). The benefit of
framework was developed called the smart grid. this approach is that no other prior knowledge regarding the
application is required other than raw data. For energy
Smart grids are power grids that exchange not only disaggregation, the neural networks can be designed to
power but also information between utilities and consumers. perform NILM by extracting the individual power
This information availability will help keep track of the consumption of appliances from the aggregated load. The
energy consumption across the whole grid. So the networks can detect features that can be associated with a
availability of feedback from the consumer end helps to particular appliance. Hence the individual consumption of
prevent conditions such as network overload. The the appliance can be formulated.
information can also be helpful in identifying power system
faults and would be needed for restoration [2]. Such a grid In this paper, two neural networks are compared in their
can also complement renewable energy resources where ability to perform NILM on the UKDALE dataset [7]. This
production is not consistent by using optimal load dispatch paper is organized as follows: Section II is the literature
strategies [3][4][5]. Even with such measures energy crisis is review of the dataset and deep learning architectures
still a concern. So, the other alternative is to minimize the implemented for NILM, Section III gives insight into the
demand through scheduled usage. The usage can be proposed methodology with the architectures being used,
maintained within a limit with the help of demand-response Section IV describes the analysis of the experiments
programs. performed, and Section V discusses the results obtained and
Section VI presents conclusions of the work.
Demand response programs provide incentives to
consumers who reduce their power consumption during
certain peak hours. The incentives will be in the form of

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on
978-1-6654-9764-0/22/$31.00 May 30,2023
©2022 IEEE at 20:18:38 UTC from IEEE Xplore. Restrictions14apply.
 Proceedings of the Third International Conference on Smart Electronics and Communication (ICOSEC 2022)
IEEE Xplore Part Number: CFP22V90-ART; ISBN: 978-1-6654-9764-0

II. LITERATURE REVIEW appliance current profiles [15]. The load is identified by
The concept of NILM was first introduced by G. W. Hart finding out which combination of current profiles matches
[6]. The goal was to figure out an appliance's state of the given power input. The disadvantage of such techniques
operation by looking at the total aggregated load is that these methods need specific features which are needed
consumption and knowing how much each appliance to be extracted manually, and a database has to be formulated
consumes in each state. Combinatorial optimization was the based on these criteria. So, such methods can be challenging
initial method for NILM [8]; however, the disadvantage of to implement in practice. Another approach uses
this methodology is that it disaggregated power at each spectrograms of current signals as input to convolutional
moment and did not account for load evolution over time. neural networks (CNN) for classification [16]. The
This method is also computationally complex and noise- spectrograms are obtained by applying STFT. The load is
sensitive, thus did not work well with many appliances. As a identified by detecting the extracted transient features. This
result, it was unsuitable for real-world scenarios. With the approach fails to detect if the current signature of any two
standard NILM datasets, the focus shifted from devices is superimposed.
combinatorial optimization to supervised machine learning In [17], LSTM and denoising autoencoders are used to
approaches. Such algorithms analyze the power consumption facilitate NILM. The LSTM networks have memory cells
over an interval of time and the power consumption pattern contained within layers. These cells have gates within them
or signature of the appliances. that act according to their corresponding activation functions.
This will help remove the problem of vanishing gradients
A. Datasets
which is a critical problem in deep neural networks.
Over the years, various public datasets were released for Denoising autoencoders treat energy disaggregation as a
performing energy disaggregation in NILM. The features denoising problem. The signal belonging to a particular
considered in NILM datasets can be summarized as steady- appliance is filtered out, and other signals are treated as
state, transient, and non-traditional features [9]. Some noise. These autoencoders try to reconstruct the input signal
steady-state features include active power, reactive power, i.e. the power consumption of the target appliance. No noise
power factor and shape factor, etc while transient features is added externally instead the demand of other appliances is
include transient voltage noise and high order harmonics, treated as noise. The sub-components present in the
Short-Term Fourier Transform (STFT), and Fast Fourier aggregate power signal are extracted which is then associated
Transform (FFT) features, etc. Non-traditional features with the corresponding appliances for load identification.
include data like start and end-time of usage, usage The size of the layer first decreases up to a particular layer
frequency, the peak time of day, etc. The REDD dataset was and it further increases after that point to the output layer.
one of the prominent public datasets, and it included power, The size of the input and the output layers of the network is
current, and voltage measurements from six different the same. So the model applies dimensionality reduction to
households [10]. UKDALE is another residential dataset that reduce features and further reconstructs the input signal. This
contains power measurements from five different homes, and ensures that the noises i.e. power demand of other appliances
the appliance monitors are turned off whenever the are removed.
appliances are not functioning [7]. This is done to conserve
electricity, but there are blank sections in measurements LSTM – Recurrent Neural Network (RNN) is a neural
indicating appliances are being turned off. The COOLL network with a memory element that is suitable for time-
dataset contains high sampled current and voltage series data. Hence, it is applied for processing NILM datasets
measurements indicating individual consumption of [7][10][11][12][13]. The LSTM-RNN networks incorporated
appliances [11]. The measurements were taken in a the better memory capability of LSTM in RNN. LSTM
laboratory environment with precise control over the turning networks are capable of retaining selective information and
ON and OFF of appliances. WHITED is a dataset that hence can reconstruct the appliance signal from the
includes current and voltage measurements from both aggregate power signal. Gated Recurrent Unit (GRU) is a
residential and industrial buildings [12]. SynD is a synthetic superior version of LSTM as it can process the data faster
dataset that can be automatically generated by performing [18]. GRU has a lesser number of gates than LSTM leading
simulations using power traces of real appliances [13]. The to faster computations. An improved version of the GRU
bright side of synthetic datasets is that we can obtain a wide network, i.e., Bi-Directional Gated Recurrent Unit (Bi-GRU)
range of data by changing various parameters without was also suggested for NILM [19]. Bi-GRU networks extract
allocating any resources for measurements. features in both forward and backward directions if applied
to the input time-series data, which would help to provide
B. Machine Learning Architectures better predictions. Bi-GRU performs much better than LSTM
The initial model that was proposed for NILM is based and GRU due to its complex architecture [20]. However, the
on Factorial Hidden Markov Models (FHMM) [14]. In the overfitting problem observed in Bi-GRU poses a serious
Hidden Markov Model (HMM) the load is identified by challenge to the classification. To rectify the same, drop-out
analyzing the occurrence and transition probabilities layers are implemented.
observed in power input. Several such models are combined
It is a challenging task to select the model from the
to form the factorial model. The drawback with such models
plethora of deep learning models suitable for a particular
is that the complexity of the problems increases with the
application. To ensure a standardized comparison, all the
number of appliances and hence affects the efficiency.
competing models are to be evaluated against the same
However, it finds difficulty in classifying multistate
performance measures. The metrics that were commonly
appliances. Appliances whose power consumption depends
used in networks involved in forecasting applications on time
on human behavior can also be challenging to analyze.
series data were Mean Absolute Error (MAE) and Root
Another method involves comparing the aggregate current
Mean Squared Error (RMSE).
signal with a database made from superimposing individual

978-1-6654-9764-0/22/$31.00 ©2022 IEEE 15

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on May 30,2023 at 20:18:38 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Smart Electronics and Communication (ICOSEC 2022)
IEEE Xplore Part Number: CFP22V90-ART; ISBN: 978-1-6654-9764-0

III. PROPOSED METHODOLOGY The UKDALE dataset contains the power demand of five
The framework as shown in Figure 1 employs the NILM houses. The active power drawn is measured every six
algorithm, which accepts the total power consumption data seconds from each appliance and the whole house. The
from households and disaggregates the power. The meter voltage and current were also recorded at 44.1 kHz, which
reading of the house is applied as input, and the output is the was further down-sampled to 16 kHz for storage. The active
power consumed by each appliance. power, apparent power, and Root Mean Square (RMS)
voltage were also calculated from this at 1 kHz. This
procedure was carried out for 5 houses. The measurement
was done longest for house 1, for 786 days with 54
appliances. While for the rest of the houses measurement
was done over a few months with appliances ranging from 5
to 26.

Fig. 1. Overview of the proposed framework for NILM.

Two models were considered for the implementation of

NILM architecture. The former is a convolutional neural
network with temporal pooling layers and the latter is a Bi-
GRU network. To extract features from time-series data, the
convolution kernel is applied along the time dimension from
beginning to end.
A. Convolutional Neural Network
Fig. 2. Temporal pooling layer extracting features at different resolutions.
The convolutional model used is inspired by the work
[19]. The model is fed an input vector of length 510 which is
the total aggregated power of a house during an interval of
8.5 hrs. The input is passed through an encoder module of
alternating convolution and pooling layers. Convolution
layers have a kernel size of 3 without padding. This is
followed by applying batch normalization and activation
function, which is Rectified Linear Unit (ReLU). Pooling
layers have a kernel size of 2 with a stride of 2 units.
We have added two temporal pooling layers in a cascade
that allow the extraction of features at different resolutions
(as depicted in Figure 2). This is done by feeding the same
data to four different branches having different filter sizes.
Four different average pooling layers are applied with kernel Fig. 3. A Bi-GRU layer.
size equal to strides. The extracted features are then The active power data of appliances and the whole house
concatenated to get the resultant set of features. This is
is measured using plug-in appliance monitors and current
followed by a convolution layer with kernel size one and no transformers. A sound-card power meter is used to measure
padding. Further, batch normalization and activation via
active power, reactive power, and RMS voltage. All the data
ReLU are carried out. In the end, one more convolution layer
is recorded as Comma Separated Values (CSV) in .dat
is there with kernel size and stride both 8. The output layer of
format for each appliance and the aggregate load. The data is
length 480 follows this.
stored in a data-logging computer. The active power data
B. Bi-GRU Network sampled at a lower frequency is only considered for our
The second model combines the concept of convolution work.
and recurrent layers. Figure 3 shows a section of the Bi-GRU In the pre-processing stage data recorded every six
layer. Bi-GRU is a processing layer that contains two GRU seconds is down-sampled to a frequency of 1 minute. The
layers, and the former extracts feature in the forward original time series input is converted to input sequences
direction while the latter extracts in the backward direction. having an overlap of 30 input sequences so that the output is
The input and output layer is similar to the earlier continuous. The data is finally converted from .dat into .csv
mentioned CNN model. The input is passed through two format. Table I shows the composition of the dataset. The
convolution layers of kernel size 20 with 2 units padding and target appliances chosen for detection are the fridge,
1 unit stride. This is followed by three bidirectional GRU microwave, and washing machine. 80% of the dataset is used
layers of sizes 8, 128, and 256. This is followed by a dense for training, 10% of the dataset is used for validation, and
layer and then the output layer. 10% of the dataset is used for testing.

IV. EXPERIMENTS
A. Dataset

978-1-6654-9764-0/22/$31.00 ©2022 IEEE 16

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on May 30,2023 at 20:18:38 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Smart Electronics and Communication (ICOSEC 2022)
IEEE Xplore Part Number: CFP22V90-ART; ISBN: 978-1-6654-9764-0

TABLE I. THE COMPOSITION OF THE UKDALE DATASET which can affect the predictions. The training time for Bi-
Building 1 2 3 4 5 GRU is more than that of CNN due to the complex gated
Number of days data architecture.
786 234 39 205 137
collected
Number of Appliances 53 18 4 11 24

B. Performance Metrics
The performance metrics chosen were MAE and RMSE.
MAE measures the average magnitude of prediction errors. It
is used in measuring the accuracy of continuous quantities. It
does not consider the concept of direction, but it is of no
concern since power is a scalar quantity. RMSE takes the
square root of the average of the square of the deviations
from the correct value. This gives a substantial value for
significant errors and vice versa.
C. Model Setup
The model was trained and validated on data from houses
1 and 5. The data from house two is used for testing. The Fig. 5. Power consumption: (a) actual and (b) predicted output for
learning rate was varied between 1𝑒 and 1𝑒 and the
microwave by CNN.
optimum value of 1𝑒 was chosen as the loss was minimal.
For the optimizer function, Adam or Adaptive Moment
Estimation was chosen. Adam was chosen as it is efficient in
computation and is suitable for handling a large volume of
data. Mean Square Error (MSE) was chosen to display after
each epoch as an accuracy metric. Both the CNN and Bi-
GRU was run for 100 epoch with chosen parameters.
V. RESULTS
The following results show the actual versus predicted
power consumption of appliances: microwave, fridge, and
washing machine using both CNN and Bi-GRU network.
Figure 4, 5, and 6 gives the output of CNN, while Figure 7,
8, and 9 gives the output of Bi-GRU.

Fig. 6. Power consumption: (a) actual and (b) predicted output for washing
machine by CNN.
The total aggregate load of the house was disaggregated
to obtain the power consumption of individual appliances.
Tables IV and V depict the power consumption
disaggregation output obtained using CNN and Bi-GRU
networks.

Fig. 4. Power consumption: (a) actual and (b) predicted output for fridge by
CNN.
Tables II and III below are the observations that were
recorded after implementing both the neural networks. The
MAE and RMSE of the Bi-GRU network are lower than that
of CNN, except for the case of the microwave. CNN
performed better predictions for microwaves. The appliances
fridge and washing machine have an irregular pattern in their
power consumption graph because of the nature of their Fig. 7. Power consumption: (a) actual and (b) predicted output for fridge by
usage. Bi-GRU is better at predicting such patterns as the
Bi-GRU.
network has memory cells that can correlate the future
predictions with previous values. In the case of the
microwave, there are these spikes in the consumption pattern

978-1-6654-9764-0/22/$31.00 ©2022 IEEE 17

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on May 30,2023 at 20:18:38 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Smart Electronics and Communication (ICOSEC 2022)
IEEE Xplore Part Number: CFP22V90-ART; ISBN: 978-1-6654-9764-0

The error metrics for the Bi-GRU network were less than
that of CNN. Bi-GRU performed better when the power
consumption pattern of the appliance was sporadic, while
CNN performed better when it was periodic. The training
time for Bi-GRU is more, as architecture is more complex
than CNN. Bi-GRU takes approximately six times longer to
train than CNN. Power disaggregation was better for Bi-
GRU as the predicted power consumption of the selected
appliances was very much closer to the ground truth than
CNN.
VI. CONCLUSION
In this paper, two deep learning architectures, CNN and
Bi-GRU were implemented for NILM. The power
consumption data of selected appliances is also obtained via
Fig. 8. Power consumption: (a) actual and (b) predicted output for disaggregation. The consumer having access to such
information can plan and adapt to reduce power
microwave by Bi-GRU.
consumption. Thus, giving a detailed insight into the power
usage.
As a future scope, the power disaggregation data can be
integrated into smart grids to design and develop detailed
demand-response programs for each consumer. With a
proper communication framework, the smart meter data can
be made available to the utility server where the NILM
algorithms are implemented. Further, the utility can share
disaggregated information with the consumers to ensure their
participation in the energy market.
REFERENCES
[1] S. T.V., S. P. and S. K. Kottayil, "Communication Framework for
Real-Time Monitoring of a Smart Grid Emulator," Inventive
Communication and Computational Technologies, vol 311, pp. 287–
Fig. 9. Power consumption: (a) actual and (b) predicted output for washing 304, Springer, 2022.
machine by Bi-GRU. [2] S. V. and S. T.V., "Comparative Analysis of Machine Learning
Algorithms for the Classification of Power System Faults," 2019 3rd
International conference on Electronics, Communication and
TABLE II. OBSERVATIONS OF CNN NETWORK Aerospace Technology (ICECA), 2019, pp. 1085-1090.
Appliances Training [3] A. R., S. P. and S. K. Kottayil, "Load Management and Smart
MAE RMSE
Time Monitoring For Rooftop PV in Academic Building," 2020
Fridge 34.75 169.71 International Conference on Power Electronics and Renewable
Microwave 22.17 29.75 2013.22s Energy Applications (PEREA), 2020, pp. 1-6.
Washing Machine 18.21 116.82 [4] N. Somasundaran, V. Kayani, R. Chandrasekhar and S. K. Kottayil,
"Automated Dispatch of Wind Power On Microgrid For Voltage
Regulation," International Journal of Automation and Smart
Technology, vol 9, 3, pp. 115-120, 2019.
TABLE III. OBSERVATIONS OF BI-GRU NETWORK
[5] A. Bashar and S. Smys, "Integrated Renewable Energy System for
Appliances Training Stand-Alone Operations with Optimal Load Dispatch Strategy,"
MAE RMSE
Time Journal of Electronics and Informatics, vol 3, 2, pp. 89-98, 2021.
Fridge 23.1 131.93
11681.69s [6] G. W. Hart, "Non-intrusive appliance load monitoring," Proceedings
Microwave 28.71 33.95
of the IEEE, vol. 80, no. 12, pp. 1870-1891, Dec. 1992.
Washing Machine 15.83 104.59
[7] J. Kelly and W. Knottenbelt, "The UK-DALE dataset, domestic
appliance-level electricity demand and whole-house demand from
five UK homes," Scientific Data, vol.2, pp.1-14, 2015.
TABLE IV. PERCENTAGE POWER CONSUMPTION DISAGGREGATION BY [8] G. Berrettoni, C. Bourelly, D. Capriglione, L. Ferrigno and G. Miele,
CNN "Preliminary Sensitivity Analysis of Combinatorial Optimization
Appliances Actual Predicted (CO) for NILM Applications: Effect of the Meter Accuracy, " 2021
Fridge 45.68 40.79 IEEE 6th International Forum on Research and Technology for
Society and Industry (RTSI), 2021, pp. 486-490.
Microwave 44.44 43.5
Washing Machine 9.87 15.69 [9] R. Gopinath, M. Kumar, C. P. C. Joshua, and K. Srinivas, "Energy
management using non-intrusive load monitoring techniques—State-
ofthe-art and future research directions," Sustain. Cities Soc., vol. 62,
Nov. 2020, Art. no. 102411.
TABLE V. PERCENTAGE POWER CONSUMPTION DISAGGREGATION BY
BI-GRU [10] J. Kolter and M. Johnson, "REDD: A Public Data Set for Energy
Disaggregation Research," Artif. Intell. vol. 25, 2011.
Appliances Actual Predicted [11] T. Picon, M. Nait Meziane, P. Ravier, G. Lamarque, C. Novello, J.-C.
Fridge 45.68 43.92 Le Bunetel, and Y. Raingeaud, "COOLL: Controlled On/Off loads
Microwave 44.44 43.63 library, a public dataset of high-sampled electrical signals for
Washing Machine 9.87 12.43 appliance identification," 2016, arXiv:1611.05803. [Online].
Available: http://arxiv.org/abs/1611.05803.

978-1-6654-9764-0/22/$31.00 ©2022 IEEE 18

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on May 30,2023 at 20:18:38 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Third International Conference on Smart Electronics and Communication (ICOSEC 2022)
IEEE Xplore Part Number: CFP22V90-ART; ISBN: 978-1-6654-9764-0

[12] M. Kahl, A. U. Haq, T. Kriechbaumer and H. A. Jacobsen, "Whited-a Applications," in IEEE Transactions on Instrumentation and
worldwide household and industry transient energy dataset," In Measurement, vol. 70, pp. 1-12, 2021.
Workshop on Non-Intrusive Load Monitoring (NILM), [17] J. Kelly and W. Knottenbelt, "Neural NILM: Deep Neural Networks
Proceedings of the 3rd International, 2016. Applied to Energy Disaggregation," Proceedings of the 2nd ACM
[13] C. Klemenjak, C. Kovatsch, M. Herold and W. Elmenreich, "A International Conference on Embedded Systems for Energy-Efficient
synthetic energy dataset for non-intrusive load monitoring in Built Environments (BuildSys '15). Association for Computing
households," Scientific Data,vol.7,pp.1-17,2020. Machinery, 2015, pp.55–64.
[14] G. A. Raiker, S. B. Reddy, L. Umanand, A. Yadav and M. M. Shaikh, [18] S. Yang, X. Yu and Y. Zhou, "LSTM and GRU Neural Network
“Approach to Non-Intrusive Load Monitoring using Factorial Hidden Performance Comparison Study: Taking Yelp Review Dataset as an
Markov Model,” 2018 IEEE 13th International Conference on Example," 2020 International Workshop on Electronic
Industrial and Information Systems (ICIIS), 2018, pp. 381-386. Communication and Artificial Intelligence (IWECAI), 2020, pp. 98-
[15] A. M. H. S. Abeykoon, A. P. S. Perera, R. K. S. Prabodhanie, M. D. 101.
N. V. Matharage and A. P. Abeysinghe, "A Machine [19] D. Precioso and D. Gómez-Ullate, "NILM as a regression versus
Learning Approach for NILM based on Superimposed Current classification problem: the importance of thresholding," 2020,
Profiles," 2020 Moratuwa Engineering Research Conference arXiv:2010.16050. [Online]. Available:
(MERCon), 2020, pp. 584-589. https://arxiv.org/abs/2010.16050.
[16] F. Ciancetta, G. Bucci, E. Fiorucci, S. Mari and A. Fioravanti, "A [20] N. Bhalaji, "Cloud Load Estimation with Deep Logarithmic Network
New Convolutional Neural Network-Based System for NILM for Workload and Time Series Optimization," Journal of Soft
Computing Paradigm, vol 3, 3, pp.234-248, 2021.

978-1-6654-9764-0/22/$31.00 ©2022 IEEE 19

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on May 30,2023 at 20:18:38 UTC from IEEE Xplore. Restrictions apply.