Non-Intrusive Load Monitoring (NILM) using Deep

Neural Networks: A Review

Mohammad Irani Azad Roozbeh Rajabi Abouzar Estebsari
Faculty of ECE Faculty of ECE School of the Built Environment and Architecture
Qom University of Technology Qom University of Technology London South Bank University
Qom, Iran Qom, Iran London, United Kingdom
[email protected] [email protected] [email protected]

Abstract—Demand-side management now encompasses more change, and a negative impact on the economy [4]. It is
residential loads. To efficiently apply demand response strategies, estimated that the rise in carbon emissions will increase
it’s essential to periodically observe the contribution of various global temperatures by 2.5 to 10 ◦ C this century, causing
domestic appliances to total energy consumption. Non-intrusive
load monitoring (NILM), also known as load disaggregation, is a more frequent floods, droughts, a rise in sea level, and the
arXiv:2306.05017v1 [eess.SP] 8 Jun 2023

method for decomposing the total energy consumption profile spread of infectious illnesses [5]. Therefore, it is essential
into individual appliance load profiles within the household. to reduce carbon emissions across all sectors, including
It has multiple applications in demand-side management, construction, industry, and transportation, to mitigate climate
energy consumption monitoring, and analysis. Various methods, change. Researchers are working on developing technology
including machine learning and deep learning, have been used to
implement and improve NILM algorithms. This paper reviews solutions for energy conservation [3]. Buildings are one
some recent NILM methods based on deep learning and of the major contributors to energy consumption [6], with
introduces the most accurate methods for residential loads. It energy consumption in this sector steadily increasing over
summarizes public databases for NILM evaluation and compares time. In order to mitigate carbon emissions, optimizing
methods using standard performance metrics. energy consumption in residential and commercial buildings
Index Terms—Smart Grids, NILM, Deep Learning, Energy
Management. is crucial. This can be achieved through the construction or
design of energy-efficient structures, as well as improving
energy usage in existing buildings.
I. I NTRODUCTION
The paper is organized as follows. Section II introduces
The non-intrusive load monitoring (NILM) method has the mathematical definition of the NILM problem. Section
gained popularity in recent years as a way to monitor appliance III discusses deep learning-based NILM methods. Section IV
and electrical utility energy usage in buildings and events provides a summary of the public NILM datasets. Section V
(on/off) using a single energy meter. If consumers had data presents a comparison study of NILM methods, and finally,
on appliance-level energy usage, they could better understand Section VI concludes the paper.
their energy consumption behavior and take action to reduce
it. The aim of this study is to present an overview of the II. NILM P ROBLEM D EFINITION
latest algorithms currently being investigated by researchers A. Mathematical Problem Definition
to create a precise non-intrusive load monitoring (NILM)
method for effective energy management. The article discusses The issue at hand can be described as follows: at a given
the potential applications of NILM across different fields, time t, the total active power consumed by a system is
along with future research objectives. The development of represented by y(t), while yi (t) represents the active power
sustainable and smart cities has been made possible by consumed by the ith appliance at the same time. The overall
advancements in artificial intelligence (AI), smart meters, the load is the sum of the energy consumed by individual
internet of things (IoT), and smart grids, as cited in [1] and appliances and an unmeasured residual load, expressed as:
[2]. Effective energy management is a crucial component of
N
sustainable city development, which aims to utilize resources X
y(t) = yi (t) + e(t), (1)
responsibly, protect the environment, and enhance society’s
i=1
well-being. The objective of energy management is to promote
energy system self-reliance and sustainability [1]. where N denotes the number of appliances considered, and
Energy management involves monitoring and controlling e(t) represents the undetermined residual load. The aim is to
electrical utilities to optimize energy use and reduce estimate F (y(t)) by determining the values of yi (t), given
consumption. However, with the increase in energy needs, only the value of y(t), as:
energy conservation has become a challenge in recent years
[3]. Greater energy use can lead to an energy crisis, climate y1 (t), y2 (t), ..., yi (t), ..., yN (t) = F (y(t)), (2)
 where F is an operator that produces N distinct values when
applied to the total active power. These numbers represent
the most accurate estimate of the power consumed by each
appliance. It should be noted that yi (t) typically does not
reflect the entire set of home appliances but rather a subset of
them. As a result, the unknown term e(t) takes into account
the loads caused by unmonitored appliances. If simultaneous
measurements of the aggregate consumption and load of each
appliance are available, approximating the F operator can
be considered a supervised learning problem. When mainly Fig. 1: A pictorial representation illustrates the different types
concerned with activation times and cumulative consumption, of appliances categorized based on their operating states.
as is the case in real situations, the estimated individual
appliance consumption (ŷi (t)) can be obtained using functions
that are constant over the device’s activation period: types of appliance events, which may happen concurrently or
independently and at varying time intervals.
ŷi (t) = pi âi (t), (3) III. D EEP L EARNING BASED NILM METHODS
where pi represents the average consumption of appliance NILM techniques can be broadly categorized into two
i, and âi (t) represents an estimate of the activation state of the groups: supervised and unsupervised methods [9]. In
particular appliance at time t. Its value is one if the device is supervised NILM, individual appliance power usage is used
in use and uses energy, and zero otherwise. Therefore, starting to train the models. On the other hand, unsupervised
with the aggregate load, a technique is provided to derive the methods can only utilize aggregate power usage data.
most accurate and feasible assessment of the activation state Examples of unsupervised NILM techniques include Hidden
of the appliances: Markov models (HMM) [10], [11], factorial HMM (FHMM)
[12], [13], and techniques based on event detection and
â1 (t), â2 (t), ..., âi (t), ..., âN (t) = Fa (y(t)), (4) clustering [14], [15]. These techniques have been thoroughly
examined in previous studies [9], [16]. With the advent of
After learning the average nominal consumption of the deep neural networks (DNNs), many neural network-based
considered equipment, one can use Equation 3 to estimate supervised NILM techniques have been developed [17], [18].
consumption. Convolutional neural networks (CNN) have also recently
made significant advances [19], [20]. Graph signal processing
B. Appliance types [21], HMM [12], [13], [15], [22], [23], and DNNs [24],
Based on their operational characteristics, appliances can be [25] are commonly used in suggested NILM approaches.
classified into four types as discussed in [7]. Type I appliances As the cost of employing appliance data for training has
have two modes of operation - on and off. These include grown dramatically, researchers have focused on developing
appliances such as kettles, toasters, and light bulbs, which unsupervised approaches and incorporating appliance models.
consume energy only when turned on. Type I appliances are Despite the significant progress made in NILM research
predominantly resistive with few linear reactive components. in recent years, challenges remain in terms of application,
Type II appliances are characterized as multi-state or finite identification accuracy, training time, and online deployment
state machines with a limited number of operational states techniques in smart metering frameworks.
that may be run repeatedly. Changes in these appliances’
states can be observed by monitoring the power consumption’s A. Event-Based Non-Intrusive Detection
falling/rising edges over time. Stove burners, refrigerators, and The event-based NILM method is based on the concept of
washing machines are some examples of Type II appliances detecting and categorizing events within a combined electrical
[7], [8]. Figure 1 demonstrates the distinct appliance operation signal. Figure 2 shows the block diagram of the this approach.
conditions. A robust event detector should be developed to cope with
Category III appliances, also known as Continuously noisy fluctuations and identify events with decay and growth
Variable Devices (CVDs), exhibit a non-repetitive power usage patterns, which is a bottleneck and inherent difficulty in
pattern, which poses a challenge for energy consumption existing event detectors [26]. One approach includes the steps
disaggregation. Examples of Type III appliances include power of event detection, extraction, clustering, and matching in the
drills and dimmer lights [8]. Type IV appliances are those event-based block [27]. It should be noted that the accuracy of
that run continuously for extended periods of time, typically previous event-based frameworks is dependent on the power
lasting several days or weeks. Examples of Type IV equipment features that are introduced. Since some appliances may have
include wireless telephone devices and cable TV receivers [8]. identical active power curves but radically distinct reactive
Therefore, the Non-Intrusive Load Monitoring (NILM) system power trends, increasing the number of features can enhance
is required to have the ability to differentiate between various the accuracy of the appliance model, particularly for non-linear
 Fig. 2: Block diagram scheme of event-based NILM.
Fig. 3: Outline of network structure for NILM by CNN.

loads. One of the advantages of incorporating reactive power

convolutioned with a unit filter size. This reduces the number
is that it enables discrimination between different types of
of characteristics to one-quarter of those used in the input.
devices.
The convolutional results are used as the input for a rectified
B. NILM Disaggregation by CNN linear unit activation function, which is then batch normalized.
Lastly, linear up-sampling yields a temporal resolution at the
The proposed method [28] employs a convolutional neural TP block’s output that is equal to the encoder’s output. A
network (CNN) that takes the time interval of a home’s energy dropout is also added to the output, allowing the network to
consumption as input and predicts the activation status of each be controlled. The context features generated by this block are
device at every time step. The network architecture, referred to connected to the encoder’s detail features, doubling the total
as Temporal Pooling NILM (TP-NILM), is an updated version number of features in the decoder’s input. To raise the signal’s
of Zhao et al.’s PSPNet (Pyramid Scene Parsing Network) temporal resolution and lower the number of features, the
used for semantic image segmentation [29]. The TP-NILM decoder contains a transposed convolutional layer with a stride
follows the conventional approach to image segmentation, and kernel size of 8. The ReLU is still used as the activation
with an encoder comprising pooling and convolutional layers function, followed by an extra convolutional layer with an
that enhance the feature space of the signal but decrease its unified kernel size that keeps the temporal resolution and
temporal resolution, and a decoder module that uses these increases the number of output channels to match the number
features to approximate the activation state of the devices at of devices being analyzed. In the output, a sigmoid function is
the original resolution. To establish a temporal context that used. This is because in the semantic segmentation of photos,
covers extended periods without compromising the signal’s each pixel is connected to a single class, however in the current
resolution, the TP-NILM incorporates a Temporal Pooling application, many appliances might be in use at the same time.
module that accumulates features at various resolutions, While working in this manner, the network decomposes all
enabling accurate reconstruction of the activation state. appliances at the same time. This should enable an encoder
Figure 3 illustrates the architecture of the network used in to use more broad convolutive filters that aren’t specialized
this study. The encoder uses a rectified linear unit activation for a single kind of appliance, boosting the neural network’s
function, batch-independent normalization downstream of the capacity to generalize. Gradient descent optimization may be
activations, and a regularization dropout layer, with three used to find the net’s weights. The loss function is a binary
convolutional filters interleaved by max-pooling layers. The crossentropy applied to each output channel that assesses the
encoder reduces the signal’s temporal resolution by a factor disparity between the activations predicted by the net âi (t) and
of eight and raises the number of output characteristics the actual ones ai (t) for each appliance under consideration
from a single aggregate power consumption value to 256. and for each instant of the period under consideration.
The TP block provides context information to the decoding
block, allowing it to create extra features for decoding by IV. NILM P UBLIC DATASETS
aggregating encoder output with various resolutions. The To develop NILM (non-intrusive load monitoring)
encoder output is passed through four average pooling modules algorithms and assess their performance, the research
with various filter sizes, which degrade the temporal resolution community provides various NILM datasets in the public
while maintaining the number of features, before being domain [30]. Since each dataset monitors different appliances
in diverse environments and buildings over a varied time electrical signals in order to extract fundamental and harmonic
period, each dataset has its own specific criteria [31], [32]. characteristics.
However, it has been observed that many public datasets have
structural variations that require pre-processing before usage. V. E XPERIMENTAL R ESULTS
To address this issue, the dsCleaner Python module was A. Evaluation Metrics
developed to standardize, clean, and convert time series data To evaluate the performance of algorithms in recognizing
into a consistent file format, and also includes a resampling appliance switching ON or OFF, the classification metrics
method for datasets. Typically, NILM datasets consist of presented in Eqs. 5-8 were used. The metrics are calculated
aggregated energy data from a single meter and the actual using True Positive (TP), True Negative (TN), False Positive
energy consumption of each appliance, which is measured by (FP), and False Negative (FN). TP represented the number
plug-level meters and serves as the ground truth for evaluating of times a device was correctly recognized as ON, whereas
NILM algorithms. Table I lists the most popular publicly TN represented the number of correctly identified OFF
accessible NILM datasets for research purposes [30]. occurrences. FP highlighted instances when ON states were
In 2011, the Reference Energy Disaggregation Dataset recorded despite the appliance not consuming power. On the
(REDD) [32] became available as the first openly accessible other hand, FN displayed the number of OFF occurrences that
dataset designed specifically to aid NILM research. Following were incorrectly recognized.
this, the building-Level fully-labeled dataset for Electricity
Disaggregation (BLUED) [33] was released in 2012, which TP + TN
Accuracy = (5)
contained data from a single household. TP + TN + FP + FN
The Almanac of Minutely Power dataset (AMPds) [34],
on the other hand, was made public in 2013 and comprised TP
Presicion = (6)
both aggregate and sub-metered power data from a single TP + FP
household. The Almanac of Minutely Power dataset Version
2 (AMPds2) [35] is another dataset that captures all three TP
Recall = (7)
primary types of consumption, including electricity, water, and TP + FN
natural gas, over an extended period of 2 years. Furthermore,
Presicion ∗ Recall
it provides 11 measurement characteristics for electricity. The F1 = 2 ∗ (8)
data in AMPds2 has been pre-cleaned to ensure consistent and Presicion + Recall
comparable accuracy results among researchers and machine The recall metric in Eq. 7 measures the ratio of correctly
learning algorithms. identified positive instances (TP) to the total number of
The REFIT Electrical Load Measurements [36] dataset positive instances in the dataset. On the other hand, the
is another one that includes cleaned electrical consumption precision metric in Eq. 6 represents the ratio of correctly
data in Watts for 20 households at both aggregate and identified positive instances (TP) to the total number of
appliance levels, timestamped and sampled at 8-second instances identified as positive by the algorithm. The F1 score
intervals. It is designed to support research into energy is a weighted mean of precision and recall, which is used
conservation and advanced energy services, ranging from to determine the accuracy of the algorithm in identifying
non-intrusive appliance load monitoring, demand response appliance states. Higher F1 scores indicate better algorithm
measures, tailored energy and retrofit advice, appliance performance in recognizing appliance state transitions.
usage analysis, consumption and time-use statistics, and The mean absolute error (MAE) and proportion of energy
smart home/building automation. Finally, the UK Domestic correctly allocated (PECA) metrics in Eqs. 9 and 10,
Appliance-Level Electricity data set (UK-DALE) [37] was respectively, are non-event-based metrics used to evaluate
released, containing data from four households. the accuracy of load disaggregation systems in calculating
The available NILM datasets have varying sample rates and assigning electricity usage. MAE measures the average
ranging from 1 Hz to 100 kHz and cover individual appliances absolute difference between the estimated and actual energy
as well as residential complexes. While data collected usage, while PECA evaluates the percentage of energy
from individual appliances can be valuable for modeling correctly allocated to individual appliances.
and training the NILM system, its performance may not
T
be optimal when tested on the entire residential building. X
MAE = 1/T ∗ |ŷti − yti | (9)
Conversely, relying solely on whole-household datasets may
t=1
not be appropriate for training the algorithms, especially when
PT PN
individual appliance data is unavailable. t=1 i=1 |ŷti − yti |
Furthermore, certain databases provide primary current and FECA = 1 − PT (10)
2 t=1 ȳt
voltage signals, whereas others provide calculated electrical
parameters such as active power, reactive power, apparent In the preceding equations, ŷti and yti are the estimated
power, and power factor. However, in order to create an and ground-truth power of the ith device at time step t,
effective NILM system, it is crucial to obtain unprocessed respectively. Furthermore, ȳt is the total power at time t [24].
 TABLE I: Summary of NILM public datasets

Dataset Sampling Duration Type Number of Number of Country URL

Rate Time Households Appliances
Resolution
REDD 16.5 KHz 19 days Residential 6 6 US https://tokhub.github.io/dbecd/links/redd.html
BLUED 12 KHz 1 week Residential 1 50 US https://tokhub.github.io/dbecd/links/Blued.html
UK-DALE 16 KHz 2 years Residential 5 5 UK https://jack-kelly.com/data/
REFIT 8 seconds 2 years Residential 20 9 UK https://pureportal.strath.ac.uk/en/datasets/refit-
electrical-load-measurements
AMPds2 1 min 2 years Residential 1 21 Canada http://dx.doi.org/10.7910/DVN/FIE0S4

B. Performance Evaluation and Comparison Study appliances to energy consumption is beneficial. NILM is
In this part, the disaggregation findings for the NILM demonstrated to be a good approach to this end. The accuracy
approaches using the REDD, UK-DALE, and REFIT datasets of NILM depends on the method applied. This paper reviewed
are shown. The performance indicators obtained by executing some deep learning-based methods which outperform other
the tests on these datasets shows that F1 produced the best existing NILM algorithms. The paper compared the results of
results for refrigerators, air conditioners, freezers, televisions, applying these advanced methods to provide a basis for future
and washing machines across the three datasets, with values implementation. These datasets have public access and are
greater than 0.70. Toasters and electronics, on the other hand, widely used in NILM literature. Several performance criteria
have lower F1 scores of roughly 0.25 owing to misclassifica- are formulated to analyze the performance of the methods.
tion caused by the non-uniform pattern of these items. R EFERENCES
The accuracy metrics of the findings were compared for
[1] A. Garulli, S. Paoletti, and A. Vicino, “Models and techniques for
published methods: on-line NILM [27], NILM-TK [14], an electric load forecasting in the presence of demand response,” IEEE
FHMM implementation; Neural-NILM [7], a DNN adaption Transactions on Control Systems Technology, vol. 23, no. 3, pp. 1087–
for energy estimation. The Neural-NILM used three DNN 1097, 2014.
[2] A. Estebsari and R. Rajabi, “Single residential load forecasting using
architectures: i) long short-term memory, ii) de-noising deep learning and image encoding techniques,” Electronics, vol. 9, no. 1,
auto-encoders, and iii) rectangles. Rectangle networks, in p. 68, 2020.
particular, regress the start-time, end-time, and average power [3] R. Rajabi and A. Estebsari, “Deep learning based forecasting of
individual residential loads using recurrence plots,” in 2019 IEEE Milan
of appliance activation. PowerTech. IEEE, 2019, pp. 1–5.
In experiments using UK-DALE dataset, a comparison of [4] N. Rezaei, R. Rajabi, and A. Estebsari, “Electricity price forecasting
on-line NILM, NILM-TK and Neural-NILM methods for five model based on gated recurrent units,” in 2022 IEEE International
Conference on Environment and Electrical Engineering and 2022 IEEE
appliances (fridge, washing machine, dishwasher, microwave, Industrial and Commercial Power Systems Europe (EEEIC/I&CPS
and kettle) is done. The microwave gives the lowest marks Europe), 2022, pp. 1–5.
for all three methods. MAE and an F1 score of roughly 195 [5] T. Wang, B. Shen, C. H. Springer, and J. Hou, “What prevents us
from taking low-carbon actions? a comprehensive review of influencing
Watts and 0.01, respectively, are reported by NILM-TK. The factors affecting low-carbon behaviors,” Energy Research & Social
best MAE of 6 Watts and an F1 score of 0.21 is shown by Science, vol. 71, p. 101844, 2021.
the neural-NILM. With an F1 score of about 0.35, the on-line [6] I. Ishak, N. S. Othman, and N. H. Harun, “Forecasting electricity
consumption of malaysia’s residential sector: Evidence from an
NILM method outperforms the other two. The MAE and F1 exponential smoothing model,” F1000Research, vol. 11, no. 54, p. 54,
scores reported by NILM-TK are roughly 67 watts and 0.55, 2022.
respectively. The Neural-NILM has an MAE of 18 Watts and [7] D. Shi, R. Li, R. Shi, and F. Li, “Analysis of the relationship between
load profile and weather condition,” in 2014 IEEE PES General
an F1 score of 0.82. In terms of energy estimate, the Neural- Meeting— Conference & Exposition. IEEE, 2014, pp. 1–5.
NILM outperformed the suggested technique, particularly [8] M. U. Fahad and N. Arbab, “Factor affecting short term load
for complicated equipment like dishwashers and washing forecasting,” Journal of Clean Energy Technologies, vol. 2, no. 4, pp.
305–309, 2014.
machines. Nonetheless, the time and computational resources [9] R. Bonfigli, S. Squartini, M. Fagiani, and F. Piazza, “Unsupervised
required to train the neural network and generate the models algorithms for non-intrusive load monitoring: An up-to-date overview,”
need a large amount of appliance-level data. The on-line NILM in 2015 IEEE 15th international conference on environment and
electrical engineering (EEEIC). IEEE, 2015, pp. 1175–1180.
technique, on the other hand, may generate appliance models [10] O. Parson, S. Ghosh, M. Weal, and A. Rogers, “Non-intrusive load
using aggregate data without the necessity for appliance-level monitoring using prior models of general appliance types,” in Twenty-
sub-metered data. Sixth AAAI Conference on Artificial Intelligence, 2012.
[11] ——, “An unsupervised training method for non-intrusive appliance load
VI. C ONCLUSION monitoring,” Artificial Intelligence, vol. 217, pp. 1–19, 2014.
[12] J. Z. Kolter and T. Jaakkola, “Approximate inference in additive
The three-phase distribution network often feeds the factorial hmms with application to energy disaggregation,” in Artificial
final residential customers through single phase cables. It intelligence and statistics. PMLR, 2012, pp. 1472–1482.
is important to make the loads balanced. Demand side [13] X. Sun, P. B. Luh, K. W. Cheung, W. Guan, L. D. Michel, S. Venkata,
and M. T. Miller, “An efficient approach to short-term load forecasting
management and different optimisation techniques would help at the distribution level,” IEEE Transactions on Power Systems, vol. 31,
on this. However, understanding the contribution of different no. 4, pp. 2526–2537, 2015.
 TABLE II: Comparison of Non-Intrusive Load Monitoring (NILM) Methods
Method Pros Cons Data Real-time Interpretability Evaluation
Requirements Performance Performance
Template Simple, easy to Requires pre-defined None Fast High Limited by
Matching implement templates for each pre-defined
appliance, limited templates
accuracy
Steady-State Accurate for steady- Limited accuracy Detailed Fast Medium Limited
Analysis state loads for transient loads, power by power
requires detailed system data system
knowledge of the model
power system
HMM and Can handle transient Requires training data Labeled Slow Low Sensitive to
fractional loads, can identify for each appliance, training data initialization
HMM multiple appliances can be computation- and noise
simultaneously ally expensive
Deep Neural High accuracy, can Requires a large Labeled Fast Low Dependent
Networks handle complex amount of training training data on quality
appliance behavior, data, can be and quantity
can identify computationally of training
multiple appliances expensive, limited data
simultaneously interpretability
Autoencoders Can handle variable Requires a large Unlabeled Fast Low Dependent
appliance behavior, amount of training training data on quality
can identify data, can be and quantity
multiple appliances computationally of training
simultaneously, can expensive data
learn without labeled
data

[14] H. Gonçalves, A. Ocneanu, M. Bergés, and R. Fan, “Unsupervised [25] P. P. M. do Nascimento, “Applications of deep learning techniques on
disaggregation of appliances using aggregated consumption data,” in nilm,” Diss. Universidade Federal do Rio de Janeiro, 2016.
The 1st KDD workshop on data mining applications in sustainability [26] S. Henriet, U. Şimşekli, B. Fuentes, and G. Richard, “A generative model
(SustKDD), 2011. for non-intrusive load monitoring in commercial buildings,” Energy and
[15] J. Zheng, C. Xu, Z. Zhang, and X. Li, “Electric load forecasting in smart Buildings, vol. 177, pp. 268–278, 2018.
grids using long-short-term-memory based recurrent neural network,” [27] M. A. Mengistu, A. A. Girmay, C. Camarda, A. Acquaviva, and E. Patti,
in 2017 51st Annual Conference on Information Sciences and Systems “A cloud-based on-line disaggregation algorithm for home appliance
(CISS). IEEE, 2017, pp. 1–6. loads,” IEEE Transactions on Smart Grid, vol. 10, no. 3, pp. 3430–
[16] M. Zhuang, M. Shahidehpour, and Z. Li, “An overview of non- 3439, 2019.
intrusive load monitoring: Approaches, business applications, and [28] L. Massidda, M. Marrocu, and S. Manca, “Non-intrusive load
challenges,” in 2018 international conference on power system disaggregation by convolutional neural network and multilabel classifica-
technology (POWERCON). IEEE, 2018, pp. 4291–4299. tion,” Applied Sciences, vol. 10, no. 4, 2020.
[17] L. Mauch and B. Yang, “A new approach for supervised power [29] B. Zhao, L. Stankovic, and V. Stankovic, “On a training-less solution for
disaggregation by using a deep recurrent lstm network,” in 2015 IEEE non-intrusive appliance load monitoring using graph signal processing,”
Global Conference on Signal and Information Processing (GlobalSIP). IEEE Access, vol. 4, pp. 1784–1799, 2016.
IEEE, 2015, pp. 63–67. [30] T. Kriechbaumer, D. Jorde, and H.-A. Jacobsen, “Waveform signal
[18] J. Kelly and W. Knottenbelt, “Neural nilm: Deep neural networks applied entropy and compression study of whole-building energy datasets,” in
to energy disaggregation,” in Proceedings of the 2nd ACM international Proceedings of the Tenth ACM International Conference on Future
conference on embedded systems for energy-efficient built environments, Energy Systems, 2019, pp. 58–67.
2015, pp. 55–64. [31] L. Wen, K. Zhou, and S. Yang, “Load demand forecasting of residential
[19] C. Shin, S. Joo, J. Yim, H. Lee, T. Moon, and W. Rhee, “Subtask gated buildings using a deep learning model,” Electric Power Systems
networks for non-intrusive load monitoring,” in Proceedings of the AAAI Research, vol. 179, p. 106073, 2020.
Conference on Artificial Intelligence, vol. 33, no. 01, 2019, pp. 1150– [32] J. Z. Kolter and M. J. Johnson, “Redd: A public data set for energy
1157. disaggregation research,” in Workshop on data mining applications in
[20] C. Zhang, M. Zhong, Z. Wang, N. Goddard, and C. Sutton, sustainability (SIGKDD), San Diego, CA, vol. 25, no. Citeseer, 2011,
“Sequence-to-point learning with neural networks for non-intrusive pp. 59–62.
load monitoring,” in Proceedings of the AAAI Conference on Artificial [33] K. Anderson, A. Ocneanu, D. Benitez, D. Carlson, A. Rowe, and
Intelligence, vol. 32, no. 1, 2018. M. Berges, “Blued: A fully labeled public dataset for event-based non-
[21] E. Busseti, I. Osband, and S. Wong, “Deep learning for time series intrusive load monitoring research,” in Proceedings of the 2nd KDD
modeling,” Technical report, Stanford University, pp. 1–5, 2012. workshop on data mining applications in sustainability (SustKDD),
[22] S. Makonin, F. Popowich, I. V. Bajić, B. Gill, and L. Bartram, vol. 7. ACM New York, 2012, pp. 1–5.
“Exploiting hmm sparsity to perform online real-time nonintrusive load [34] S. Makonin, F. Popowich, L. Bartram, B. Gill, and I. V. Bajić, “Ampds:
monitoring,” IEEE Transactions on smart grid, vol. 7, no. 6, pp. 2575– A public dataset for load disaggregation and eco-feedback research,” in
2585, 2015. 2013 IEEE Electrical Power & Energy Conference, 2013, pp. 1–6.
[23] H. Kim, M. Marwah, M. Arlitt, G. Lyon, and J. Han, “Unsupervised [35] S. Makonin, “Ampds2: The almanac of minutely power dataset (version
disaggregation of low frequency power measurements,” in Proceedings 2),” 2021. [Online]. Available: https://dx.doi.org/10.21227/c2rq-bx53
of the 2011 SIAM international conference on data mining. SIAM, [36] D. Murray, L. Stankovic, and V. Stankovic, “An electrical load
2011, pp. 747–758. measurements dataset of united kingdom households from a two-year
[24] J. Kelly and W. Knottenbelt, “Neural nilm: Deep neural networks applied longitudinal study,” Scientific data, vol. 4, no. 1, pp. 1–12, 2017.
to energy disaggregation,” in Proceedings of the 2nd ACM international [37] J. Kelly and W. Knottenbelt, “The uk-dale dataset, domestic appliance-
conference on embedded systems for energy-efficient built environments, level electricity demand and whole-house demand from five uk homes,”
2015, pp. 55–64. Scientific data, vol. 2, no. 1, pp. 1–14, 2015.