Professional Documents
Culture Documents
This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCE.2023.3324921
JOURNAL 1
Abstract—In this work, a method for unsupervised energy (NILM), there exists a variety of methods and we give an
disaggregation in private households equipped with smart meters overview over existing literature in the following.
is proposed. The method aims to classify power consumption as
active or passive, granting the ability to report on the residents’
activity and presence without direct interaction. This lays the
foundation for applications like non-intrusive health monitoring Related works
of private homes.
The proposed method is based on minimizing a suitable energy
functional, for which the iPALM (inertial proximal alternating
Reviews of existing methods for NILM can be found in [11],
linearized minimization) algorithm is employed, demonstrating [13]. NILM methods are frequently split into the categories
that various conditions guaranteeing convergence are satisfied. supervised if they use labeled training data and unsupervised
In order to confirm feasibility of the proposed method, in the converse case. While the latter case is substantially more
experiments on semi-synthetic test data sets and a comparison difficult, it is also relevant for practical applications, due to the
to existing methods are provided.
cost and difficulty of collecting sufficient amounts of labeled
Index Terms—Energy disaggregation, non-intrusive load mon- training data.
itoring, health monitoring, convolutional sparse coding
a) Supervised methods: One of the first approaches
introduced in [14] is based on matching the changes in
I. I NTRODUCTION power consumption (event detection) to certain devices in the
MART meters – energy meters capable of digital com- household. In [16], the authors use graph signal processing
S munication – are becoming increasingly available in Eu-
ropean domestic households. The regulation Intelligente Mess-
to predict changes in the state of individual devices. There
are many works on different versions of hidden Markov
geräte-Einführungsverordnung – IME-VO [2] derived from the models (HMM), and in particular factorial versions thereof
EU directive [1] dictates that 95% of all Austrian meter points [8], [28], [45], [51]. In [26], [40], the authors propose sparse
have to be equipped with smart meters by the end of 2024. coding for NILM, and in [47], this approach is extended to
As a result, regular readings of household energy consumption a deep sparse coding model. Deep learning approaches have
will be available on a large scale, potentially enabling a variety also been applied to NILM. In [21], the authors propose to
of insights into power consumption, usage patterns and human use recurrent convolutional neural networks in a sequence-
activity. For an overview of potential health care applications to-sequence fashion. In [22], the authors employ a sequence-
of smart meters see [12], [43]. As a lack of human activity to-sequence generative adversarial network (GAN) for energy
in a household of, e.g., an elderly or care-dependent person disaggregation, and in [20], this model is improved by using
over an extended period of time could be an indication of a a deep learning recurrent classifier to model the discriminator
potential emergency, in this work we focus on the possibility component of the GAN. For more work on deep learning for
of detecting human activity, respectively a lack thereof, based NILM, see also [19], [29], [37], [38], [50].
on smart meter readings. Such an automatic detection would b) Unsupervised methods: By employing a priori as-
alleviate the need for the consumer to report incidents manu- sumptions on the device distributions HMMs can be used in
ally, and bypass the usage of intrusive surveillance devices. an unsupervised manner [3], [17], [18], [24], [27], [35], [42].
Smart meters measure the aggregate energy consumption In [41], the authors propose a method of tuning pre-trained
of a household rather than the consumptions of individual general appliance models of an HMM to devices of a specific
devices. Obtaining detailed information about the composition household by only using aggregate data from this household.
of the aggregate data is difficult, as it requires either directly In [30], the authors propose a method comprised of event
measuring the energy consumption at many different points in detection, feature extraction, and classification. The method is
the electrical circuit, or computational separation of the smart proposed in an unsupervised and a supervised manner, using
meter signal into distinct contributions, which is inherently a decision tree and dynamic time wrapping for classification,
ill-posed. For the latter problem, under the term computa- respectively. Similar strategies are considered in [32] and [33],
tional energy disaggregation or non-intrusive load monitoring and in [49], such an approach is combined with graph signal
processing methods. In [36], the authors propose to solve the
The SOLGENIUM Project: Supported by the Austrian Research Promotion
Agency (FFG) (Grant 881561). NILM problem using non-negative matrix factorization, and
Manuscript received XXX YYY, 2022; revised AAA BBB, 2022. also a shift-invariant version thereof.
XXXX–XXXX/XX$XX.XX © 2022 IEEE
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on January 22,2024 at 06:00:39 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Consumer Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCE.2023.3324921
JOURNAL 2
D] especially since a sufficiently extensive dictionary is not For the sake of simplifying notation we also denote PNthe sum-
yet available [39, Section 3.5.7]. Obtaining a dictionary by of-convolutions map for multiple devices c∗p := i=1 ci ∗pi .
measuring one activity cycle of each electrical device in the In the application of activity detection, we are particularly
household, on the other hand, poses an excessive and intrusive interested in the case where we have several active channels,
installation effort, and is consequently undesirable. Altogether, representing the characteristics of all devices that require
we therefore believe that methods not relying on training data human interaction – thus indicating human activity – and
are crucially needed for NILM. Moreover, our experiments one passive channel, consisting of the characteristics of all
suggest that, even when training data is available, our un- remaining devices. Examples of devices contributing to the
supervised approach is more robust compared to supervised active channels include, but are not limited to, ovens, washing
ones with respect to shifts in distribution between training and machines and hair dryers, while devices such as fridges and
testing data. (see Table II, Figure 1 and also [31], [46]). routers might contribute to the passive channel. Devices in the
Motivated by the aforementioned monitoring applications, active channels are assumed to be needed only a few times in
in this work we will focus on experimental results for activity the span of a single day, in the sense that the number of times
prediction, but provide also results for full disaggregation into they are activated each day is relatively low. This leads to a
individual devices. For activity prediction, we disaggregate natural sparsity of incidents of each device characteristic on
into an active and a passive channel, where the presence of the temporal axis (see Figure 1, top right). In summary we
the former will be used as an indicator of human activity. desire the following properties for our atoms and coefficients:
• The atoms should be bounded, so that scaling is left to the
coefficients. Atoms and coefficients should be positive.
II. M ODELING AND P ROBLEM FORMULATION 1 1
• With (c , p ) representing the passive channel, we let all
The power consumption u ∈ Rm of a household, with m ∈ remaining pairs, that is, (ci , pi ) for i ≥ 2, serve to explain
N the number of evenly spaced time points, is given as the sum the active channels; P thus, active energy consumption
N
of the power consumptions of all individual devices ui ∈ Rm is approximated by i i
i=2 c ∗ p . The coefficients c ,
i
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on January 22,2024 at 06:00:39 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Consumer Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCE.2023.3324921
JOURNAL 3
minimization problem. In that respect, we face two main is clear that ∇c R(c, p) and ∇p R(c, p) are differentiable with
difficulties with our problem formulation: It is neither jointly respect to (c, p) jointly, with bounded derivatives on bounded
convex nor differentiable in (c, p), meaning standard convex subsets of (Rm+2q )N × (R2q+1 )N . The mean value inequality
or gradient descent-based algorithms cannot be used as is. To immediately yields Lipschitz continuity on bounded sets.
overcome this, we employ the inertial Proximal Alternating
Linearized Minimization (iPALM) algorithm as proposed in
Remark 1. The preceding theorem ensures that Assumption
[44], which is suitable for nonconvex, nonsmooth optimization
A of [44] is satisfied (note that the lower bound in condition
problems of a particular structure. Therefore, we first rewrite
(iv) of this assumption can be easily satisfied by choosing
(1) into a form where the iPALM algorithm is applicable, i.e.,
sub-optimal Lipschitz constants that are bounded from below
we consider
by a positive number). Further, using results from [7], it is
min F (c, p), where F (c, p) := R(c, p) + f1 (c) + f2 (p). (2) clear that the objective function in (2) is semi-algebraic; it
c,p
thus satisfies the Kurdyka-Łojasiewicz property. Hence, under
Here, R(c, p) := ku − c ∗ pk22 , suitable upper bounds on the stepsizes (see Assumption B of
( PN [44]), the iPALM algorithm can be applied to (2), and it is
λp kc1 k2 + λa i=2 kci k1 , c ≥ 0, guaranteed that any bounded sequence of iterates converges
f1 (c) :=
∞, otherwise, to a critical point of (2) ([44, Theorem 4.1]).
(
0, p ≥ 0 and max1≤i≤N kpi k2 ≤ 1, Remark 2. The Lipschitz constants from 1) in Proposition
f2 (p) :=
∞, otherwise. 1 are needed for the stepsize choice in the iPALM algorithm.
Those correspond to the operator norms kP (p)k resp. kC(c)k,
To ensure convergence of iPALM according to [44], several which can be computed explicitly via spectral radii, but this
properties of the functionals involved in (2) need to be may prove impractical for numerics. Denoting the matrix
satisfied. It is clear that f1 , f2 and R are proper, lower semi- representation of ci 7→ ci ∗pi as P i (pi ), norm estimates on the
continuous, bounded from below, and that R is differentiable. P i (pi ) can be obtained as follows: With J : Rm+2q → `2 (R),
Moreover, f1 and f2 are both convex, with f1 being a (Jx)i := xi for 1 ≤ i ≤ m + 2q and (Jx)i := 0 otherwise,
linear combination of componentwise norms under a convex with K : R2q+1 → `1 (R), (Kx)i := xi+2q+1 for −2q ≤ i ≤ 0
constraint, and f2 simply being a convex constraint. We further and (Kx)i := 0 otherwise, and with ˜∗ the discrete convolution
need Lipschitz continuity of the derivative of R as stated in between `2 (R) and `1 (R), Young’s inequality implies
the following proposition.
kP i (pi )ci k2 = kci ∗ pi k2 ≤ kJci ˜∗Kpi k`2
Proposition 1. For R as above, the following holds:
≤ kJci k`2 kKpi k`1 = kci k2 kkpi k1 .
1) For any fixed p ∈ (R2q+1 )N , c 7→ ∇c R(c, p) is Lipschitz
continuous. Similarly, for any fixed c ∈ (Rm+2q )N , p 7→ Taking the supremum yields kP i (pi )k ≤ kpi k1 . A similar
∇p R(c, p) is Lipschitz continuous. argument shows kC i (ci )k ≤ kci k1 for C i (ci ) the matrix
2) ∇R is Lipschitz continuous on bounded subsets of representation of pi 7→ ci ∗ pi . Applying standard matrix norm
(Rm+2q )N × (R2q+1 )N . estimates, one altogether has
( m
Proof. Write R(c, p) := k · k22 ◦ [u − ·] ◦ (c ∗ p). As the map p X
c 7→ c ∗ p is linear for fixed p, one has c ∗ p = P (p)c with kC (c )k ≤ min kci k1 , 2q + 1 max
i i
cij+∆j ,
0≤∆j≤2q
P : (R2q+1 )N → Rm×N (m+2q) . Simultaneously, for fixed c, j=1
k∇p R(c, p) − ∇p R(c, p̃)k2 ≤2kC(c)k2 kp − p̃k2 Applied to our setting, the iPALM algorithm takes
(4) the following form: Fix a suitable non-zero initialisation
=:L2 (c)kp − p̃k2 , (c−1 , p−1 ) = (c0 , p0 ) ∈ (Rm+2q )N × (R2q+1 )N , all non-
showing 1). negative componentwise. Fix ε ∈ (0, 1) small, and for i ∈
To show 2), we note that as the maps p 7→ P (p), p 7→ {1, 2}, choose ᾱi ∈ (0, 1 − ε), β̄i > 0 arbitrary. Define
ᾱ1 +2β̄1 + + +
P (p)T , c 7→ C(c) and c 7→ C(c)T are linear and bounded, it γi := 2(1−ε− ᾱ1 ) λi , where λ1 resp. λ2 corresponds to the
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on January 22,2024 at 06:00:39 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Consumer Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCE.2023.3324921
JOURNAL 4
upper bound over all k on L1 (pk ) as in (3), resp. L2 (ck ) as in (proxft 1 (c))i for i ≥ 2: Writing f1 (c)ij = |cij | + δ+ (cij ) for
(4); when such a bound is not analytically apparent, a viable 1 ≤ j ≤ N , a combination of [6, Example 6.8] with [9,
substitute in the k-th iteration step is max1≤j≤k L1 (pj ) resp. Lemma 6.14(b)] immediately yields the claimed form.
PN
max1≤j≤k+1 L2 (cj ), possibly dropping early indices to avoid proxft2 (p): Observe that f2 (p) = i
i=1 δCi (p ), where
excessive size. For each k ∈ N0 , repeat the following process: Ci := x ∈ R 2q+1
| x ≥ 0, kxk2 ≤ 1 . Employing [6, The-
orems 6.6, 6.24 and 6.30], as well as noting that Ci =
Require: Pick α1k ∈ [0, ᾱ1 ], β1k ∈ [0, β̄1 ].
x ∈ R2q+1 | gi (x) ≤ 1 , where gi (x) := δ+ (x) + kxk22 with
y1k := ck + α1k (ck − ck−1 )
z1k := ck + β1k (ck − ck−1 ) effective domain R2q+1 + , one has
(1+ε)γ1 +(1+β1k )L1 (pk )
τ1k := (
k1
2−α [pi ]+ , gi ([pi ]+ ) ≤ 1,
(proxft 2 (p))i i
f1
ck+1 ∈ proxτ k y1k − τ1k ∇c R(z1k , pk ) = PCi (p ) = i
1 1
proxλgi (p ), else,
Require: Pick α2k ∈ [0, ᾱ2 ], β2k ∈ [0, β̄2 ].
y2k := pk + α2k (pk − pk−1 ) where λ > 0 solves
z2k := pk + β2k (pk − pk−1 )
(1+ε)γ2 +(1+β2k )L2 (ck+1 ) pi
τ2k := 1 = gi (proxλgi (pi )) = gi prox
1 ( )
2−αk2 2λ+1 δ+ 2λ +1
f2
pk+1 ∈ proxτ k y2k − τ1k ∇p R(ck+1 , z2k ) [pi ] 1
+
2 2
= gi = k[pi ]+ k22 ,
2λ + 1 (2λ + 1)2
The mappings proxfτ1k and proxfτ2k above are proximal
1 2
mappings (see [6, Chapter 6]), which are given explicitly as with the last three equalities following from [6, Theorems 6.13,
follows. 6.24 and Lemma 6.26]. Simplifying leads to the desired form.
Proposition 2. For t > 0, the functions f1 , f2 as above satisfy
( λp /t
1
1 − max{k[c ]+ k2 ,λp /t} [c ]+ , i = 1,
(proxft 1 (c))i =
1
IV. E XPERIMENTS
ci − λta + ,
i > 1,
[pi ]+ In this section, we provide experimental results of the
(proxft 2 (p))i = , proposed method (CSC) for two different scenarios, namely
max {k[pi ]+ k2 , 1}
activity prediction, where we try to detect whenever energy
where [·]+ , | · |, − and sign are understood componentwise.
is consumed actively by a person, and multichannel disag-
Proof. Denote for any set E by δE the indicator function gregation, where we aim to reconstruct all individual device
satisfying δE (x) = 0 for x ∈ E and δE (x) = ∞ for x ∈ / E, signals that amount to an aggregate energy signal. In particular,
and by δ+ the indicators for the positive cones. PE denotes Table II in Section IV-D1, and Section IV-D2 contain results
projection onto E. for the setting that no ground truth training data is available
We begin by noting that [6, Theorem 6.6], and the fact that for the household of interest, constituting a realistic use case
f1 and f2 can both be written as componentwise sums enables as explained in Section I.
us to compute each proximal mapping componentwise, with
the i-th component depending only on ci resp. pi .
(proxft 1 (c))1 : Writing it out explicitly, one has A. Data
(proxft 1 (c))1 = For our experiments, we employ the SynD data set generator
v
um+2q m+2q [25], an algorithm to generate synthetic energy signals. This
uX X t enables us to work with clean, yet realistic, data. The algorithm
argmin t d2i + δ+ (di ) + |di − c1i |2 .
d∈Rm+2q i=1 i=1
2 uses real device energy signatures, which it randomly alters
and places throughout the day. It became apparent, however,
From this, it is apparent that any solution dˆ ∈ Rm+2q will that the fridge’s energy signal generated with SynD is unreal-
satisfy dˆi = 0 if c1i ≤ 0. As such, one can replace c1 by istically regular, being identical each day; for this reason, we
[c1 ]+ in the above minimization problem without changing replaced the fridge by an energy signal taken from the UK-
the solution. But with [c1 ]+ instead of c1 , it is clear that DALE 6 seconds data set [23], and resampled to the desired
any solution of the above problem will be non-negative even sampling rate by weighted averaging.
without the non-negativity constraint, making it equivalent to Explicitly, we generate two different synthetic households,
λ k·k2
proxt p ([c1 ]+ ) = referred to as synd1 and synd2. Each of the two data sets
v consists of four devices, synd1 of the fridge from UK-
um+2q m+2q
uX t X DALE’s house 1 as well as SynD’s electric space heater,
argmin t d2i + |di − ([c1 ]+ )i |2 . washing machine, and hair dryer, and synd2 of the fridge
d∈Rm+2q i=1
2 i=1 from UK-DALE’s house 2 and SynD’s dishwasher, iron, and
By [6, Example 6.19], the solution of the latter is of the watercooker. We use a sampling rate of 1 minute, and the data
claimed form. sets are normalized to have values between 0 and 1.
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on January 22,2024 at 06:00:39 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Consumer Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCE.2023.3324921
JOURNAL 5
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on January 22,2024 at 06:00:39 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Consumer Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCE.2023.3324921
JOURNAL 6
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on January 22,2024 at 06:00:39 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Consumer Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCE.2023.3324921
JOURNAL 7
V. C ONCLUSION
[2] Intelligente Messgeräte-Einführungsverordnung. Online Source:
In this work, we developed a new unsupervised method for www.ris.bka.gv.at, retrieved on May 23, 2022.
energy disaggregation based on convolutional sparse coding [3] M. Aiad and P. H. Lee. Unsupervised approach for load disaggregation
with a focus on classifying energy consumption as active or with devices interactions. Energy and Buildings, 116:96–103, 2016.
[4] N. Batra, J. Kelly, O. Parson, H. Dutta, W. Knottenbelt, A. Rogers,
passive. Prior training is not needed, as the method relies A. Singh, and M. Srivastava. Nilmtk: An open source toolkit for
only on the choice of hyperparameters, which can be fixed non-intrusive load monitoring. In Proceedings of the 5th international
for a broad class of input data. We employ the iPALM conference on Future energy systems, pages 265–276, 2014.
[5] N. Batra, R. Kukunuri, A. Pandey, R. Malakar, R. Kumar, O. Krys-
algorithm to minimize the used objective functional, ensuring talakos, M. Zhong, P. Meira, and O. Parson. Towards reproducible
strong convergence guarantees compared to many competing state-of-the-art energy disaggregation. In Proceedings of the 6th ACM
methods. As the results of our numerical experiments show, international conference on systems for energy-efficient buildings, cities,
and transportation, pages 193–202, 2019.
our unsupervised approach performs comparable to the state [6] A. Beck. First-order methods in optimization. 2017.
of the art in activity prediction in terms of classification [7] J. Bolte, S. Sabach, and M. Teboulle. Proximal alternating linearized
scores, and yields superior performance in terms of NMSE minimization for nonconvex and nonsmooth problems. Mathematical
and NMAE scores (see Tables I and II). Further, it recognizes Programming, 146:459–494, 2014.
[8] R. Bonfigli, E. Principi, M. Fagiani, M. Severini, S. Squartini, and
specific energy consumption signatures better (see Figure 1), F. Piazza. Non-intrusive load monitoring by using active and reactive
and offers more flexibility due to the untrained approach. Ap- power in additive factorial hidden markov models. Applied Energy,
plied to unsupervised multichannel disaggregation, our method 208:1590–1607, 2017.
[9] K. Bredies and M. Holler. Higher-order total variation approaches and
performs slightly superior to the comparison method GSP [49]. generalisations. Inverse Problems, 36:123001, 2020.
[10] D. Chicco and G. Jurman. The advantages of the matthews correlation
R EFERENCES coefficient (mcc) over f1 score and accuracy in binary classification
[1] Directive (EU) 2019/944 of the European Parliament and of the Coun- evaluation. BMC Genomics, 21(6), 2020.
cil of 5 June 2019 on common rules for the internal market for [11] S. Dash and N. Sahoo. Electric energy disaggregation via non-intrusive
electricity and amending Directive 2012/27/EU. Online Source: eur- load monitoring: A state-of-the-art systematic review. Electric Power
lex.europa.eu/legal-content/EN/ALL/?uri=CELEX%3A32019L0944, re- Systems Research, 213:108673, 2022.
trieved on June 23, 2022. [12] M. Fell, H. Kennard, G. Huebner, M. Nicolson, S. Elam, and D. Ship-
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on January 22,2024 at 06:00:39 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Consumer Electronics. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCE.2023.3324921
JOURNAL 8
worth. Energising health: A review of the health and care applications graph signal processing”. https://github.com/loneharoon/GSP energy
of smart meter data. London, UK: SMART Energy GB, 2017. disaggregator, 2018.
[13] R. Gopinath, M. Kumar, C. P. C. Joshua, and K. Srinivas. Energy [35] S. Makonin, F. Popowich, I. V. Bajić, B. Gill, and L. Bartram. Exploiting
management using non-intrusive load monitoring techniques–state-of- hmm sparsity to perform online real-time nonintrusive load monitoring.
the-art and future research directions. Sustainable Cities and Society, IEEE Transactions on smart grid, 7(6):2575–2585, 2015.
62:102411, 2020. [36] A. Miyasawa, Y. Fujimoto, and Y. Hayashi. Energy disaggregation based
[14] G. W. Hart. Nonintrusive appliance load monitoring. Proceedings of on smart metering data via semi-binary nonnegative matrix factorization.
the IEEE, 80(12):1870–1891, 1992. Energy and Buildings, 183:547–558, 2019.
[15] T. Hastie, R. Tibshirani, and M. Wainwright. Statistical Learning with [37] A. Moradzadeh, B. Mohammadi-Ivatloo, M. Abapour, A. Anvari-
Sparsity–The Lasso and Generalizations. Chapman and Hall/CRC, 2015. Moghaddam, S. Gholami Farkoush, and S.-B. Rhee. A practical
[16] K. He, L. Stankovic, J. Liao, and V. Stankovic. Non-intrusive load solution based on convolutional neural network for non-intrusive load
disaggregation using graph signal processing. IEEE Transactions on monitoring. Journal of Ambient Intelligence and Humanized Computing,
Smart Grid, 9(3):1739–1747, 2016. 12(10):9775–9789, 2021.
[17] R. Jia, Y. Gao, and C. J. Spanos. A fully unsupervised non-intrusive [38] D. Murray, L. Stankovic, V. Stankovic, S. Lulic, and S. Sladojevic.
load monitoring framework. In 2015 IEEE international conference on Transferability of neural network approaches for low-rate energy dis-
smart grid communications (SmartGridComm), pages 872–878. IEEE, aggregation. In ICASSP 2019-2019 IEEE International Conference on
2015. Acoustics, Speech and Signal Processing (ICASSP), pages 8330–8334.
[18] M. J. Johnson and A. S. Willsky. Bayesian nonparametric hidden semi- IEEE, 2019.
markov models. 2013. [39] B. Najafi, S. Moaveninejad, and F. Rinaldi. Data analytics for energy
[19] M. Kaselimi, N. Doulamis, A. Doulamis, A. Voulodimos, and E. Pro- disaggregation: Methods and applications. In Big data application in
topapadakis. Bayesian-optimized bidirectional lstm regression model for power systems, pages 377–408. Elsevier, 2018.
non-intrusive load monitoring. In ICASSP 2019-2019 IEEE International [40] S. Pandey and G. Karypis. Structured dictionary learning for energy dis-
Conference on Acoustics, Speech and Signal Processing (ICASSP), pages aggregation. In Proceedings of the Tenth ACM International Conference
2747–2751. IEEE, 2019. on Future Energy Systems, pages 24–34, 2019.
[20] M. Kaselimi, N. Doulamis, A. Voulodimos, A. Doulamis, and E. Pro- [41] O. Parson, S. Ghosh, M. Weal, and A. Rogers. An unsupervised
topapadakis. Energan++: A generative adversarial gated recurrent training method for non-intrusive appliance load monitoring. Artificial
network for robust energy disaggregation. IEEE Open Journal of Signal Intelligence, 217:1–19, 2014.
Processing, 2:1–16, 2020. [42] S. Pattem. Unsupervised disaggregation for non-intrusive load monitor-
[21] M. Kaselimi, E. Protopapadakis, A. Voulodimos, N. Doulamis, and ing. In 2012 11th International Conference on Machine Learning and
A. Doulamis. Multi-channel recurrent convolutional neural networks Applications, volume 2, pages 515–520. IEEE, 2012.
for energy disaggregation. IEEE Access, 7:81047–81056, 2019. [43] J. Paxman, M. James, E. Costanza, and J. Manning. Smart
future of healthcare. 20/20health, 2020. Online Source:
[22] M. Kaselimi, A. Voulodimos, E. Protopapadakis, N. Doulamis, and
2020health.org/publication/smart-future-of-healthcare/, retrieved on June
A. Doulamis. Energan: A generative adversarial network for energy
28, 2022.
disaggregation. In ICASSP 2020-2020 IEEE International Conference
[44] T. Pock and S. Sabach. Inertial proximal alternating linearized mini-
on Acoustics, Speech and Signal Processing (ICASSP), pages 1578–
mization (ipalm) for nonconvex and nonsmooth problems. Sustainable
1582. IEEE, 2020.
Cities and Society, 9:1756–1787, 2016.
[23] J. Kelly and W. Knottenbelt. The UK-DALE dataset, domestic
[45] G. A. Raiker, S. B. Reddy, L. Umanand, A. Yadav, and M. M.
appliance-level electricity demand and whole-house demand from five
Shaikh. Approach to non-intrusive load monitoring using factorial
UK homes. Scientific Data, 2(150007), 2015.
hidden markov model. In 2018 IEEE 13th International Conference
[24] H. Kim, M. Marwah, M. Arlitt, G. Lyon, and J. Han. Unsupervised
on Industrial and Information Systems (ICIIS), pages 381–386. IEEE,
disaggregation of low frequency power measurements. In Proceedings
2018.
of the 2011 SIAM international conference on data mining, pages 747–
[46] Y. Shi, I. Daunhawer, J. E. Vogt, P. Torr, and A. Sanyal. How robust
758. SIAM, 2011.
is unsupervised representation learning to distribution shift? In The
[25] C. Klemenjak, C. Kovatsch, M. Herold, and W. Elmenreich. A synthetic Eleventh International Conference on Learning Representations, 2022.
energy dataset for non-intrusive load monitoring in households. Scientific [47] S. Singh and A. Majumdar. Deep sparse coding for non–intrusive load
Data, 7(1):1–17, 2020. monitoring. IEEE Transactions on Smart Grid, 9(5):4669–4678, 2017.
[26] J. Kolter, S. Batra, and A. Ng. Energy disaggregation via discriminative [48] F. Wiese, I. Schlecht, W.-D. Bunke, C. Gerbaulet, L. Hirth, M. Jahn,
sparse coding. Advances in neural information processing systems, 23, F. Kunz, C. Lorenz, J. Mühlenpfordt, J. Reimann, et al. Open power
2010. system data–frictionless data for electricity system modelling. Applied
[27] J. Z. Kolter and T. Jaakkola. Approximate inference in additive factorial Energy, 236:401–409, 2019.
hmms with application to energy disaggregation. In Artificial intelligence [49] B. Zhao, L. Stankovic, and V. Stankovic. On a training-less solution for
and statistics, pages 1472–1482. PMLR, 2012. non-intrusive appliance load monitoring using graph signal processing.
[28] W. Kong, Z. Y. Dong, D. J. Hill, J. Ma, J. Zhao, and F. Luo. IEEE Access, 4:1784–1799, 2016.
A hierarchical hidden markov model framework for home appliance [50] G. Zhou, Z. Li, M. Fu, Y. Feng, X. Wang, and C. Huang. Sequence-to-
modeling. IEEE Transactions on Smart Grid, 9(4):3079–3090, 2016. sequence load disaggregation using multiscale residual neural network.
[29] D. Li, J. Li, X. Zeng, V. Stankovic, L. Stankovic, and Q. Shi. Non- IEEE Transactions on Instrumentation and Measurement, 70:1–10,
intrusive load monitoring for multi-objects in smart building. In 2021 2020.
International Balkan Conference on Communications and Networking [51] A. Zoha, A. Gluhak, M. Nati, and M. A. Imran. Low-power appliance
(BalkanCom), pages 117–121. IEEE, 2021. monitoring using factorial hidden markov models. In 2013 IEEE Eighth
[30] J. Liao, G. Elafoudi, L. Stankovic, and V. Stankovic. Non-intrusive International Conference on Intelligent Sensors, Sensor Networks and
appliance load monitoring using low-resolution smart meter data. In Information Processing, pages 527–532. IEEE, 2013.
2014 IEEE International Conference on Smart Grid Communications Christian Aarset concluded his MSc in Mathematics at the University of
(SmartGridComm), pages 535–540. IEEE, 2014. Oslo, and his PhD in Mathematics at the University of Klagenfurt.
[31] H. Liu, J. Z. HaoChen, A. Gaidon, and T. Ma. Self-supervised learning Andreas Habring received his MSc in Mathematics from Graz University
is more robust to dataset imbalance. arXiv preprint arXiv:2110.05025, of Technology. He currently does his PhD in applied Mathematics at the
2021. University of Graz supervised by Martin Holler.
[32] Q. Liu, K. M. Kamoto, X. Liu, M. Sun, and N. Linge. Low- Martin Holler received his MSc and his PhD with a ”promotio sub auspiciis”
complexity non-intrusive load monitoring using unsupervised learning in Mathematics from the University of Graz. After research stays at the
and generalized appliance models. IEEE Transactions on Consumer University of Cambridge, UK, and the Ecole Polytechnique, Paris, he currently
Electronics, 65(1):28–37, 2019. holds a Assistant Professor position at the Institute of Mathematics and
[33] Q. Liu, F. M. Nakoty, X. Wu, R. Anaadumba, X. Liu, Y. Zhang, and Scientific Computing of the University of Graz.
L. Qi. A secure edge monitoring approach to unsupervised energy Mario Mitter received his MSc and PhD in Physics and his MSc in
disaggregation using mean shift algorithm in residential buildings. Mathematics, all ”with honors”, from the University of Graz. After research
Computer Communications, 162:187–195, 2020. stays at the University of Heidelberg, DE, and the Brookhaven National
[34] H. R. Lone, A. Barbosa, and A. Renaud. Source code to the paper ”On a Laboratory, Upton, NY, he switched to the field of Data Science in the private
training-less solution for non-intrusive appliance load monitoring using sector in 2019.
Authorized licensed use limited to: UNIV OF ENGINEERING AND TECHNOLOGY TAXILA. Downloaded on January 22,2024 at 06:00:39 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.