448 IEEE TRANSACTIONS ON SMART GRID, VOL. 10, NO.

1, JANUARY 2019

Incorporating Appliance Usage Patterns for

Non-Intrusive Load Monitoring
and Load Forecasting
Shirantha Welikala , Student Member, IEEE, Chinthaka Dinesh , Student Member, IEEE,
Mervyn Parakrama B. Ekanayake, Member, IEEE, Roshan Indika Godaliyadda, Member, IEEE,
and Janaka Ekanayake, Fellow, IEEE

Abstract—This paper proposes a novel non-intrusive load solar photovoltaic (PV) and wind. Due to the unpredictable
monitoring (NILM) method which incorporates appliance usage nature of the generation, maintaining the second-by-second
patterns (AUPs) to improve performance of active load identifi- balance between demand and generation has become a chal-
cation and forecasting. In the first stage, the AUPs of a given
residence were learned using a spectral decomposition based lenging task, if an expensive reserve service is not maintained.
standard NILM algorithm. Then, learnt AUPs were utilized to As reserve services are mainly provided by operating certain
bias the priori probabilities of the appliances through a specif- power plants below their rating, this not only underutilizes
ically constructed fuzzy system. The AUPs contain likelihood its own capacity but also results in them being operated
measures for each appliance to be active at the present instant inefficiently. As a viable solution to this problem, DSM is
based on the recent activity/inactivity of appliances and the
time of day. Hence, the priori probabilities determined through considered. DSM tries to reduce/increase the demand either
the AUPs increase the active load identification accuracy of the by shifting or reducing the consumption so that the avail-
NILM algorithm. The proposed method was successfully tested able generation can be utilized efficiently while maintaining a
for two standard databases containing real household measure- minimum reserve.
ments in USA and Germany. The proposed method demonstrates Direct Load Control (DLC) is one attractive option for
an improvement in active load estimation when applied to the
aforementioned databases as the proposed method augments the DSM which helps the utility to shape the customer energy
smart meter readings with the behavioral trends obtained from consumption profile by remotely controlling customers pre-
AUPs. Furthermore, a residential power consumption forecast- agreed set of controllable appliances such as, heat, ventilation,
ing mechanism, which can predict the total active power demand air-conditioning and smart (HVACS) systems. Even though a
of an aggregated set of houses, 5 min ahead of real time, was smart meter connected at the consumer premises could make
successfully formulated and implemented utilizing the proposed
AUP based technique. these HVACS loads flexible, unless the grid operator knows
the amount of flexible load that is available at a given time,
Index Terms—Non-intrusive load monitoring (NILM), fuzzy the utilities continue to maintain a large reserve by deloading
systems, usage patterns, smart grid, demand side management
(DSM), direct load control (DLC), demand response (DR). generators.
This paper proposes a solution to this problem in the
form of a Load Monitoring (LM) method that can predict
I. I NTRODUCTION
the amount of flexible load available at consumer premises.
N THE recent years, Demand Side Management (DSM)
I has become an essential element of the rapidly develop-
ing smart grid; mainly as a result of increasing penetration
Using LM, the set of appliances that are currently turned
ON and their individual energy contributions at a customer
premise is predicted. Even though LM could be achieved
of intermittent and variable renewable energy sources such as by attaching sensors for each appliance, due to the imple-
Manuscript received January 13, 2017; revised June 20, 2017; accepted mentation cost and the complexity, it is not a feasible
August 17, 2017. Date of publication September 18, 2017; date of cur- solution. In contrast, Non-Intrusive Load Monitoring (NILM)
rent version December 19, 2018. This work was supported by the National approaches, in which only the total power at the entry
Science Foundation Sri Lanka under Grant RG/2016/EA and Grant ICT/01.
Paper no. TSG-00064-2017. (Corresponding author: Janaka Ekanayake.) point to the consumer premise is monitored to find the load
S. Welikala, M. P. B. Ekanayake, and R. I. Godaliyadda are with activities [1], involve lower implementation cost and com-
the Department of Electrical and Electronic Engineering, University of plexity. Due to these advantages, NILM methods are gaining
Peradeniya, Kanky 20400, Sri Lanka (e-mail: [email protected];
[email protected]; [email protected]). popularity.
C. Dinesh is with the University of Peradeniya, Kanky 20400, Sri Lanka,
and also with Simon Fraser University, Burnaby, BC V5A 1S6, Canada
(e-mail: [email protected]). A. Related Work
J. Ekanayake is with the University of Peradeniya, Kanky 20400,
Sri Lanka, and also with Cardiff University, Cardiff CF24 3AA, U.K. (e-mail: Throughout literature, a number of different NILM meth-
[email protected]). ods have been proposed. In general, those NILM methods
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. can be categorized based on the type of measurements uti-
Digital Object Identifier 10.1109/TSG.2017.2743760 lized, as steady state methods and transient state methods.
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/
WELIKALA et al.: INCORPORATING AUPs FOR NILM AND LOAD FORECASTING 449

Among steady state measurement based NILM methods, learning period. This proposed NILM method uses the NILM
information of active power [2]–[5], reactive power [6], har- technique in [2] and improves upon it to gain this added adapt-
monic content [7], [8], and voltage-current trajectory [9], [10] ability. Hence, in this proposed novel approach, individual
are commonly utilized. In transient state measurement based appliance usage patterns (AUPs) are used to augment the direct
NILM methods, voltage and current [11]–[13], power [8] as smart meter measurements to identify the currently turned ON
well as harmonic information [7], [14] have been utilized. appliance combination.
NILM methods based on steady state measurements have In this approach, the priori probabilities of individual appli-
the common difficulty of identifying non resistive appli- ances at the current time instant are calculated using a devel-
ances [1], [15], [16] and appliances which cause non-discrete oped fuzzy system. This fuzzy system utilize the pre-observed
changes in power [17]. Transient measurement based NILM historical individual AUPs containing historical activity of
methods have the common drawback of requiring measure- each appliance as pertaining to the time of day and turned
ments with higher sampling rates in kilohertz range [18]–[20]. ON/OFF duration likelihoods.
Such methods require high communication bandwidths and Further, to calculate the individual appliance priori probabil-
processing power. Moreover, some of the aforementioned ities, the developed fuzzy system also uses the information of
techniques require more than one electrical measurement, the turned ON appliance combinations identified in the recent
resulting in the need for costly, multi-functional smart meters. past.
Considering all these factors, for a scalable NILM solu- The individual appliance priori probability is the probability
tion, such commercial costs and implementation complexities of that appliance being in the turned ON state at the current
should be mitigated. time instant. This key information is appropriately biased for
In more recent works, this NILM problem has been further each appliance depending on the current behavior trends and
studied along several different major avenues. In [21] and [22], then considered in the proposed NILM method to find the most
several multi-label classification techniques based on wavelet- probable currently turned ON appliance combination.
domain and time-domain feature sets have been evaluated Hence the proposed NILM method does not solely depend
to address the NILM problem. In parallel, Graph Signal on collected measurements. To decide the NILM solution at
Processing based techniques have been utilized in [23]–[25] a certain time instant, apart from collected active power mea-
for the same purpose. In [26]–[28], event detection and cluster- surements, it also considers the priori probability values (given
ing avenue have also been explored to address the same NILM by pre-observed AUPs) of each appliance combination. As a
problem, based on Subtractive Clustering, Bayesian-Viterbi result of this novel approach, it delivers more accurate NILM
Clustering and Dynamic Time Warping techniques respec- results when compared to very recent NILM methods.
tively. Further, Sparse Coding (SC) and Hidden Markov Model Further, this paper proposes a novel load forecasting tech-
(HMM) based techniques are also two emerging avenues in nique which uses the learned usage patterns of appliances
the same research. NILM methods proposed in [29]–[31] are together with the present NILM result to predict the load pro-
based on learning a basis for each individual appliance through file of an aggregated set of houses a few minutes ahead of
sparse coding and dictionary learning. These methods use current time. Since this proposed total load forecasting tech-
Deep-SC, ‘Powerlet’ Learning and Descriptive-SC techniques nique incorporates both the AUPs as well as the current NILM
respectively. In [5] and [32]–[38], number of different HMM- solution, it clearly demonstrates the viability of applying the
based NILM techniques have been discussed. Furthermore, a proposed NILM method and its incorporated AUPs in a DSM
more detailed overview on various such recent NILM methods application such as in DLC.
have been presented in [21] and [39]–[41]. Furthermore, in this paper, implementation and scalability
Even though there are many such diverse NILM approaches aspects of the proposed complete NILM solution (with fore-
suggested in [1]–[25] and [29]–[42] almost all the NILM meth- casting) have been explored. This is a clear contribution of
ods estimate the present turned ON appliance combination this paper as none of the NILM methods existing in the litera-
based on the recently collected set of measurements. For exam- ture have not extensively demonstrated such applicability and
ple, the NILM strategy proposed in [3] decides the currently practicality of NILM in DSM.
turned ON appliance combination based on ten most recent In the proceeding sections, first, the underlying NILM tech-
set of total active power measurements. Most of the proposed nique is summarized under Section II. Then, the improvement
NILM methods completely rely on smart meter measurements of the NILM algorithm by studying appliance usage pattern
without incorporating any of the activity that happened in the is examined in Section III. Finally, the proposed total power
recent past in terms of load activity and inactivity. Due to this profile forecasting mechanism is introduced in Section IV.
reason, a single erroneous or unlearned measurement has the
potential to mislead the NILM algorithm.
II. U NDERLYING NILM M ETHOD U SED
A. Overview of the NILM Algorithm
B. Contributions The overall flow of the proposed NILM algorithm is
As a remedy to the above mentioned common shortcom- described in Fig. 1. The main steps are,
ing of NILM methods, this paper proposes a NILM method 1) Feature extraction from individual appliance power pro-
which adapts itself to the user behavioral patterns rather than files (in Reading Set 1 - RS1) by the Karhunen Loève
being rigidly dependent on collected measurements or on the Expansion (KLE) based technique described in [3].
450 IEEE TRANSACTIONS ON SMART GRID, VOL. 10, NO. 1, JANUARY 2019

Fig. 1. Overall Flow of the Proposed NILM Algorithm.

2) Creation of Appliance/Combination/Power consumption According to (1), signal X was decomposed into N̄ number
level signature databases using extracted features. of mutually uncorrelated spectral components which are also
3) Turned ON appliance combination identification using known as Subspace Components (SCs) of X, named hereafter
initial priori unbiased NILM step (for Reading as, x1 , x2 , . . . , xN̄ where xi = qTi Xqi . Here, qi can be thought
Set 2 - RS2). of as a narrow band eigen-filter whose output is sinusoidal
4) Appliance Usage Pattern (AUP) extraction using the with a center freq of fci and phase angle of θi . Thereafter, the
obtained initial result (of RS2). average amplitude of SC, which is incidentally the eigenvalue
5) Evaluation of priori biased NILM method using con- λi , and the phase angle θi formed the complex features for
structed AUP based fuzzily priori biasing technique (for each SC. This is converted to rectangular form via relations,
Reading Set 3 - RS3). Rei = λi cos(θi ) and Imi = λi sin(θi ). With that, for each SW
6) Aggregated residential power profile forecasting method of length N there are N̄ number of SCs denoted by xi and each
(using the priori biased NILM result and usage patterns). of these SCs have three features, namely fci , Rei and Imi . In
order to establish the stationarity within the SW, parameters
For the ease of explanation, steps 1-3 are referred to N and N̄ has been chosen as ten and five respectively.
as “Stage A” of the proposed NILM method. This section
describes the processes involved in Stage A. An in-depth anal-
ysis about this stage can be found in [2]. The final three C. Signature Database Construction
steps are referred to as “Stage B” and it is introduced in Using the feature data obtained from each SW of train-
Sections III and IV. ing data (i.e., the RS1) for an appliance, 2D histograms
Individual appliance power profiles taken from two publicly were formulated for each center frequency fc . Once these
available datasets containing real measurements collected from histograms are normalized, they yield the probability of hav-
U.S. and German households were considered for this study. ing a feature; for instance, (Re1 , Im1 ) at fc = 0.2 Hz
They are Tracebase Database [43] and Reference Energy for the learned appliance (say A1 ). This is denoted sim-
Disaggregation Dataset (REDD) [34]. A publicly available ply as an appliance level Probability Mass Function (PMF)
toolkit named NILM Toolkit (NILMTK) [44] was utilized for P(A1 ,fc =0.2Hz) (Re1 , Im1 ). Utilizing each of these constructed
dataset conversion and data pre-processing. appliance specific PMFs, a set of appliance combination spe-
cific PMFs for each fc for each possible appliance combination
B. KLE Based Appliance Feature Extraction were constructed through mathematical convolution opera-
If X = [X(n) X(n−1) . . . X(n−i+1) . . . X(n−N+1) ]T is a sliding tion between corresponding appliance specific PMFs [3]. This
window (SW) of an individual appliance active power trace gives the probability of obtaining a feature when that appliance
(taken from RS1, at time instant n), its KLE is given by, combination is currently turned ON.
These generated sets of appliance specific PMFs and appli-

N̄
ance combination specific PMFs form the appliance level
X = Qx̄ = qTi Xqi (1) signature database (ALSD) and the combination level signa-
i=1 ture databases (CLSD) respectively [45]. Furthermore, in order
where, q1 , q2 , . . . , qN̄ are the eigenvectors and Q is the eigen- to perform the power level disaggregation according to [46]
vector matrix of the Autocorrelation Matrix (ACM) of X. after the active appliance combination is identified, the power
Further, x̄ is the Karhunen Loève Transform of X. consumption levels of each and every appliance were studied
WELIKALA et al.: INCORPORATING AUPs FOR NILM AND LOAD FORECASTING 451

Algorithm 1 Work Flow of Appliance Combination After that, at the SES, for every remaining appliance combi-
Identification in Initial Priori Unbiased NILM Step (for RS2) nation, probability of generating the current SC of the OSW
1: for each SW in RS2 do was calculated using the PMFs stored in the CLSD. These
2: Extract features from the SW; #5 SCs: Zi ; i = 1, 2, ..., 5;
3: Set i = 1; #Iteration or SC Number values were denoted by P(Zi /Cj ) where Zi is the features of
4: Set execution = 1; #Matching Incomplete Yet ith SC of an OSW and Cj is the jth appliance combination. If
5: Set S0 = {Cj : j = 2Napps } #Set of All Combinations
6: Apply the PES to S0 and obtain S1 ; this likelihood value is not larger than 0.1, those combinations
7: while execution do #Iterative SC Matching were also eliminated.
8: Consider the ith dominant SC : Zi ; Then at the MAP criteria, probabilities obtained after the
9: Apply the FES to S1 or S3 and obtain S2 ;
10: Apply the SES to S2 and obtain S3 ; SES were utilized to calculate the most probable appliance
11: Apply the MAP criteria to S3 and obtain S4 ; #Get γCj,i combination which matched all OSW SCs up to the current
12: if γCj,i > 0.99 then #Matching of Cj upto ith SC iteration. This MAP criteria value is given by,
13: Output: Turned ON Appliance Combination = Cj ;  
14: Set execution = 0; #Matching Complete P Cj /Z1,...,i
15: else if (i == 5) ∪ (S4 ∈ ø) then γCj,i =   , (2)
16: Output: Most Probable Solution: argmax(γCj,i , Cj ); P Cj /Z1,...,i
17: Set execution = 0; #Matching Complete ∀j
18: else
19: i=i+1 #Go to Next SC Matching where,
20: end if
       
21: end while
Cj
i
Cj
i P ZCkj P Cj
end for = =  Zk   .
22: P P (3)
Z1,...,i Zk P Cj P Cj
k=1 k=1
∀j

and a power consumption level signature database (PCLSD) Here, P(Zk /Cj ) values are taken from the pre-constructed
was constructed [46]. CLSD and P(Cj ) denotes the priori probability of the appliance
combination Cj . In this step, since every appliance combina-
tion was assumed to have an equal priori probability value, (3)
D. Initial Priori Unbiased NILM Step was simplified further by assuming,
 
The next step of the proposed NILM method is to evaluate ∀j, ∀OSWs : P Cj = constant. (4)
the initial priori unbiased NILM step for the aggregated power
profile of RS2. In this step, every appliance combination was If the calculated γCj,i value in (2) is larger than 0.99 for a cer-
considered to have an equal priori probability value to be the tain appliance combination Cj , in a certain iteration i, then that
currently active appliance combination. Thus, it is called the combination was taken as the identified initial NILM solution
priori unbiased NILM step in this paper. (i.e., the active set of appliances) for that OSW [2].
First, as in [3], sliding windows of 10 samples taken from Finally, after evaluating Algorithm 1 for the RS2, through
the aggregated power profile in RS2 were considered sequen- the obtained results at the end of Stage A, each Appliance
tially. Such a sliding window is referred to as an observation Usage Pattern (AUP) was observed.
sliding window (OSW). Then for each OSW, its corresponding Next section describes how the proposed NILM method in
set of features were extracted in the same way as described this paper utilized these observed AUPs to use the full MAP
in Section II-B. After that, these extracted features (i.e., the criteria definition given in (3) without assuming (4).
five SCs) were used to find the matching turned ON appliance
combination corresponding to that particular OSW. The con- III. S TUDY OF A PPLIANCE U SAGE PATTERNS (AUP S )
structed CLSD was utilized for this matching purpose, where In the NILM method proposed in [2], which was also used
the possible features corresponding to every viable appliance in stage A, the turned ON appliance combination at a partic-
combination had been stored. This identification process is ular time instant was found by considering the most recent
summarized in Algorithm 1. few samples of measurements (i.e., by using the OSW). This
Initially, all the possible appliance combinations were con- direct dependence between NILM result and the sensor mea-
sidered as viable solutions and then appliance combinations surements hinders the accuracy levels. Therefore, as a remedy,
were rapidly reduced based on a “pre-elimination stage” historical AUPs were used to enhance the accuracy of the
(PES), a “first elimination stage” (FES), and a “second elimi- NILM method.
nation stage” (SES). Finally, a Maximum a Posteriori (MAP) For example, due to an anomaly in measurements such as,
criteria was applied to evaluate the most likely solution. sensor measurement noises, interferences or unlearned behav-
At PES, the static level of the OSW (i.e., the first SC) is ior of appliances and residential voltage level fluctuations,
compared with the minimum static levels obtained for every correctly identified appliance combination given by a NILM
possible appliance combination. Then every appliance combi- method may get altered by yielding an incorrect appliance
nation with a minimum static level larger than the measured combination for a small duration of time. However, it is under-
active power signals static level (i.e., within the OSW in con- standable that a sudden changes in the turned ON appliance
cern) was eliminated. At FES, average power level of each combination within a residential building is unlikely.
SC was used to eliminate the appliance combinations which In the proposed NILM method, an avenue was created to
showed lower maximum possible average SC power levels [2]. utilize the observed AUPs from the results of Stage A, in order
452 IEEE TRANSACTIONS ON SMART GRID, VOL. 10, NO. 1, JANUARY 2019

to obtain the priori probability values for each appliance com-

bination for each OSW in Stage B. This section describes the
extracted AUPs, the technique used to calculate the priori prob-
abilities and the method which utilized the calculated priori
probabilities referred to as the “priori biased NILM method”.

A. Extracted Different Appliance Usage Pattern Profiles Fig. 2. (a) PONDu (t) and (b) POFFDu (t) for FR.
Vital information was revealed when analyzing the AUPs
obtained from Stage A (See Fig. 1) of the proposed NILM
method. First, by exploring the individual appliance usage pro-
files given by the results of Stage A (for the RS2), it was
found that most of the appliances show certain ON durations
followed by certain OFF durations more often. For example,
appliances like Refrigerators, Freezers and Water Fountains
show specific ON and OFF duration occurrences more fre-
quently. Further, it was noticed that, some appliances are in
Fig. 3. PTOD (t) for WM and TV.
the ON state more commonly during a particular time period
of the day. For example, certain Lamps are more likely to be
in the ON state during the night and in the early morning.
switched ON in the current time instant after being in the OFF
Therefore, in order to interpret AUPs, these three parameters
state for a duration of time t seconds up to now is,
ON Duration (OnDu), OFF Duration (OFFDu) and Time of
Day (TOD) were used. POFFDu,Ak (t)
POFF,Ak (t) = 1 − . (8)
In Stage B, results from Stage A were utilized to extract POFFDu,Ak (t − 1)
the ON and OFF durations characteristics of each appli-
Finally, through tracking the time of day on which the
ance. Then the appliance specific histograms for ON and
appliance Ak has been mostly used, a likelihood function was
OFF durations were constructed and converted to correspond-
constructed as PTOD,Ak (t) which gave the likelihood of an
ing Complementary Cumulative Mass Functions (CCMFs) as
appliance being used at a given time in a day. Fig. 3 shows
respectively given by,
the constructed PTOD (t) for the Television (TV) and wash-
PONDu,Ak (t) = P(ONDu ≥ t), (5) ing machine (WM). Following this technique, AUPs were
characterized in terms of three appliance specific likelihood
and, functions: PON,Ak (t), POFF,Ak (t) and PTOD,Ak (t).

POFFDu,Ak (t) = P(OFFDu ≥ t). (6) B. Fuzzy Based Priori Probability (PP) Calculating Strategy
Here, (5) gives the probability of the kth appliance Ak being As the step after the AUP extraction described in
used for a duration of time more than t seconds. For exam- Section III-A, through the constructed likelihood functions,
ple, ON duration histogram and respective CCMF curve a fuzzy logic based priori probability (PP) calculating tech-
obtained (PONDu (t)) for the Freezer (FR) is given in Fig. 2(a). nique was used to obtain the appliance combination specific
According to that histogram, ON duration of the FR takes a PP values P(Cj ). Those PP values were used to generalize
bimodal behavior where it is most likely to show an ON dura- the evaluation of MAP criteria in (3) without assuming the
tion of 5 or 10 minutes. Also FR is less likely to show an ON constraint in (4).
duration more than 12 minutes or less than 4 minutes. For an appliance combination Cj = {A1 , A2 , . . . , An } where
Similarly, (6) describes the probability of appliance Ak being n ∈ {1, 2, . . . , NA }; NA = number of appliances; and j ∈
in switched OFF state for a duration of time more than t {1, 2, . . . , 2NA }; its priori probability at the time instant t = t0
seconds. As an example, OFF Duration histogram and the was obtained from under the assumption that all appliances
obtained POFFDu (t) curve for the FR is shown in Fig. 2(b). are independent as,
Note that CCMF curves in Fig. 2 should be read using the   
n
right hand side y-axis. P Cj |t=t0 = PPP,Ak (t0 ). (9)
Then, the first CCMF in (5) was converted into a conditional k=1
Probability Mass Function (PMF) using, Here, PPP,Ak (t0 ) denotes the PP value of the appliance Ak at
  time instant t = t0 . In order to get this value, first, the history
OnDu > t PONDu,Ak (t)
PON,Ak (t) = P = , (7) of the given appliance state (ON or OFF) was used to obtain
OnDu > (t − 1) PONDu,Ak (t − 1)
the corresponding time duration that the appliance remained
which gave the probability of the appliance Ak being in the in that state. While executing the proposed NILM method for
turned ON state in the current time instant, given that it has each appliance, the state history and the corresponding time
been in the turned ON state for a duration of time t seconds durations were updated and stored. Now, through the con-
up till now. Similarly, the probability of appliance Ak being structed likelihood functions in (7) or (8), either PON,Ak (t0 )
WELIKALA et al.: INCORPORATING AUPs FOR NILM AND LOAD FORECASTING 453

Fig. 4. FIS used to get the PP when Ak is ON : (FISON,Ak ).

Fig. 6. Output Surfaces for (a) FISON,Ak and (b) FISOFF,Ak .

only for priori probability biasing. Thus it is not the sole cri-
teria for determining the turned ON appliance combination.
Therefore, even under random behaviors, this priori biased
NILM algorithm works accurately.

C. The Priori Biased NILM Method

Fig. 5. FIS used to get the PP when Ak is OFF : (FISOFF,Ak ).
Using the AUP based PP calculation technique discussed in
Sections III-A & B, PP values required to evaluate the MAP
or POFF,Ak (t0 ) was found for each appliance Ak depending criteria given in (3) were found. To incorporate this priori bias-
on its ON/OFF state. Further, a time of day based likelihood ing approach into the priori unbiased NILM method described
value for the appliance Ak was also evaluated based on the con- in Section II, two additional stages were incorporated into the
structed likelihood function PTOD,Ak . Based on the most recent Algorithm 1.
state of the appliance Ak , one of constructed Fuzzy Inference First, each appliance state and its present ON or OFF dura-
Systems (FISs) out of FISON,Ak in Fig. 4 or FISOFF,Ak in Fig. 5 tion was updated before the PES (in Algorithm 1: line 6)
was used to get the PPP,Ak (t0 ) value appropriately. using the NILM solution in the previous time instant. Then,
For the FISs, as shown in Fig. 4 and Fig. 5, piecewise linear in the first iteration (i = 1), after the SES (in Algorithm 1:
membership functions (MFs) were used. Since the appliance line 10), PP values of all remaining possible appliance combi-
Ak has the PON,Ak (t) values biased towards 1, these MFs were nations were calculated using (9), by utilizing the constructed
also automatically shifted towards that region in order to get appliance specific FISs: FISON,Ak and FISOFF,Ak .
a better input sensitivity [47], [48]. Similarly, other input MFs With these modifications, the NILM algorithm with the pri-
(3 per each appliance) were also decided based on that auto- ori biasing technique as shown in Algorithm 2 was deployed
mated logic. All output MFs of FISs (2 FISs per appliance) as step 5 of the proposed NILM algorithm to carry out the
have the same MF arrangement where MFs have been shifted load combination identification for the aggregated power pro-
outward from 0.5 towards both 0 and 1. This was done in order file: RS3. Here, from the obtained results, an improvement in
to have a more information rich output rather than having a the NILM accuracy levels was observed with the introduction
result similar to PPP,Ak = 0.5 where the information content of the AUP based priori biasing technique for NILM.
is less [49].
Furthermore, for each appliance, same two different Fuzzy IV. T OTAL P OWER D EMAND F ORECASTING M ETHOD
Rule Bases (FRBs) have been used for respective two appli- Total power demand forecasting is considered as one key
ance specific FISs. These two FRBs were designed in such a application of NILM for DSM [1]. However, to the best of
way that they can eliminate NILM solutions which indicates authors knowledge, there are no known cases of NILM or even
sudden appliance state changes that occur at unusual time of AUP based approaches been utilized to tackle this problem of
day. This fact is illustrated by the FIS Output Surfaces shown demand forecasting reported in literature. As a viable solution,
in Fig. 6. There, since in such incidents PTOD,Ak value is near this paper proposes a NILM and AUP based approach to fore-
0, while the PON,Ak or the POFF,Ak value is near 1, FRB has cast the total demand of a number of houses, 5 minutes ahead
been designed to have a low priori probability value output for of current time instant. This information enables a DSM aggre-
such inputs. Thus, it increases the stability and the robustness gator to inform the Transmission System Operator the amount
of the NILM solution. of DSM available in case of a system emergency.
The proposed priori biasing technique in this Section In the proposed NILM technique, identified ON and OFF
depends on extracted common appliance usage behaviors. sets of appliances as well as their respective ON or OFF dura-
However, as a result of randomness in the human behav- tions are available for a given time instant. Since pre-calculated
ior, actual appliance usages might occasionally deviate from ON or OFF duration based CCMFs given in (5) and (6) are
these pre-constructed appliance usage patterns. Therefore, the also available, for each appliance, most probable ON or OFF
usage patterns may not solely be able to decide the turned durations into the future were calculated for a constant con-
ON appliance combination. As a remedy to this scenario, the fidence level value, α, for each time instant. For illustrative
overall priori biased NILM technique uses the usage patterns purposes, confidence levels have been chosen as 0.9 and 0.5.
454 IEEE TRANSACTIONS ON SMART GRID, VOL. 10, NO. 1, JANUARY 2019

Algorithm 2 Work Flow of Appliance Combination power level disaggregation solution [46], to predict the indi-
Identification in Priori Biased NILM Step (for RS3) vidual appliance power profiles ahead from the current time.
1: for each SW in RS2 do Through aggregation of these results for appliances in a num-
2: Extract features from the SW; #5 SCs: Zi ; i = 1, 2, ..., 5;
3: Set i = 1; #Iteration or SC Number ber of houses, the proposed total power demand forecasting
4: Set execution = 1; #Matching Incomplete Yet technique was successfully validated.
5: Set S0 = {Cj : j = 2Napps } #Set of All Combinations
6: Update ON and OFF durations of each appliance;
7: Apply the PES to S0 and obtain S1 ;
8: while execution do #Iterative SC Matching V. C ASE S TUDY
9: Consider the ith dominant SC : Zi ;
10: Apply the FES to S1 or S3 and obtain S2 ; Two case studies were carried out to evaluate the proposed
11: Apply the SES to S2 and obtain S3 ; NILM method and the demand forecasting technique.
12: if i == 1 then #only in 1st iteration
13: Calculate priori probabilities for every appliance combi-
nation in S3 using (9) and constructed FISs;
14: end if A. Performance Metrics
15: Apply the MAP criteria to S3 and obtain S4 ; #Get γCj,i
The case study utilized the following performance metrics.
16: if γCj,i > 0.99 then #Matching of Cj upto ith SC 1) Appliance Combination Identification Accuracy (Aci ):
17: Output: Turned ON Appliance Combination = Cj ;
18: Set execution = 0; #Matching Complete To asses the overall performance of the NILM method, Aci
19: else if (i == 5) ∪ (S4 ∈ ø) then value was calculated as the percentage of OSWs where the
20: Output: Most Probable Solution: argmax(γCj,i , Cj );
turned ON appliance combination was found correctly [2].
21: Set execution = 0; #Matching Complete
22: else 2) F-Measure (Fm ): The F-measure (Fm ) [50] was used to
23: i=i+1 #Go to Next SC Matching evaluate the accuracy of identifying the states of combination
24: end if
25: end while (Cj ), and is given by,
26: end for
Fm,Cj = 2TP/(2TP + FN + FP), (11)

where TP, FN and FP, for each identified turned on appliance

combination, are the True Positives, False Negatives and False
Positives. The average Fm,Cj over ∀Cj in a given aggregated
active power signal is denoted as Afm .
3) Total Power Correctly Assigned (Apa ): To evaluate the
performance of the power disaggregation, the “Total Power
Correctly Assigned” (Apa ) metric described in [1] and [34]
was used. Metric Apa is formally defined as follows [34]:

Fig. 7. tf vs td Profiles of RF for (a) ON and (b) OFF States. 
T 
n
(i)

T
Apa,Cj = 1 − |yˆt − yit |/ 2 y¯t × 100%,
t=1 i=1 t=1
These confidence levels are adjustable and the two values were (12)
specifically selected to highlight the impact of this parameter.
For example, if the NILM method detected that, appliance where yˆt (i) denotes the calculated mean power level of the
Ak has been in the ON state for a duration of td up till the proposed method for ith appliance at the tth OSW while yit
current instant, then, using the CCMF given in (5) the turned denotes the measured mean power level for ith appliance  at the
ON duration tf into the future for that appliance with a confi- tth OSW in a given aggregated signal. Moreover, ȳ = ni=1 yit .
dence level of α was found using the conditional probability The average Apa,Cj over ∀Cj in a given aggregated active power
given by, signal is denoted as Apd . Further, the metric Apa for estimation
    of power demand forecasting is denoted as Apf .
ONDu > td + tf PONDu,Ak td + tf 4) Average Execution Time (Aet ): All algorithms in this
α=P = . (10)
ONDu > td PONDu,Ak (td ) paper were executed on a workstation with Intel Core i5 pro-
cessor and 16 GB RAM running at 2.3 GHz, with Windows
Further, to reduce the execution time, for each appliance 10 OS. In order to demonstrate the speed of the solution, the
Ak and for each possible ON duration td , corresponding tf metric Average Execution Time (AET) taken to process an
values were pre-stored for few confidence levels such as OSW to generate the NILM result was used.
α = 0.3, 0.5, 0.7, 0.8, 0.9, & 0.95, during AUP extrac-
tion step. Due to this pre-storing techniques, forecasts with
different confidence levels can be achieved. Same procedure B. Case Study 1
was repeated to predict the likelihood of the OFF state tf into This case study was carried out to evaluate the NILM
the future for a given confidence level. Fig. 7 shows such accuracy improvement achieved by AUP based fuzzily pri-
pre-stored curves for the Refrigerator (RF). ori biasing. Here, seven tests were carried out by deploying
Now, these pre-constructed sets of profiles were used along the proposed NILM method exactly as illustrated in Fig. 1 for
with the proposed NILM method’s present solution and the the real data taken from Tracebase and REDD datasets.
WELIKALA et al.: INCORPORATING AUPs FOR NILM AND LOAD FORECASTING 455

TABLE I TABLE II
P ERFORMANCE M ETRICS C OMPARISON B ETWEEN BASIC NILM (PUN) C OMPARISON OF O BTAINED F-M EASURE VALUES
VS AUP BASED E NRICHED NILM (PBN)

1) Procedure: The Tracebase dataset contains 22 differ-

ent individual appliance power profiles collected at 1 samples It should be noted that accuracy of the power level disag-
per second from German households and office spaces. For gregation has also been increased by around two percentage
the first test, 12 residentially used appliances were purpo- points from using this strategy. Since both priori unbiased
sively selected from this dataset such that, selected set of and biased NILM algorithms had used the same power break-
appliances contained 2 or more number of appliances from down technique in [46], this slight improvement should be
each appliance category [2]: “single-state” (SS), “multi-state” due to the increased appliance combination identification accu-
(MS), and “continuously varying” (CV). From the selected set racy achieved by the proposed NILM method. Although AET
of appliances by manually aggregating the appliance power has been increased by around 30% due to the introduced
profiles a house was created (HouseT) which had individual priori biasing step in the proposed NILM method, still the
and aggregated power profiles for 52 days. As higher number AET is well inside 1s or 3s sampling periods. Thus, real-time
of different types of appliances have been used in the case implementation of this strategy is clearly viable.
study, real-world complex conditions have been emulated in Furthermore, REDD houses 3-6 contains several unknown
the constructed HouseT. appliances and plug sockets. So, activation of such appli-
First ten days of individual appliance power data (RS1) were ances should decrease the NILM accuracy levels significantly.
used in feature extraction and creation of signature databases. The proposed NILM method was able to produce an accu-
Then, next 21 days of the aggregated power signal (RS2) were racy improvement even under such a challenging scenario.
evaluated by Stage A of Figure 1 to study the AUPs. Then Also, it should be noted that, in the absence of such unknown
from the remaining aggregated power profile data (RS3), priori appliances, the proposed method generates NILM results with
biased NILM method was evaluated. Finally, for performance accuracy levels higher than 93%. In addition, most NILM solu-
comparison purposes, priori unbiased NILM method was also tions in the literature have actually not used the REDD houses
evaluated for RS3. in concern (4, 5 and 6) due to the presence of the unknown
The REDD dataset contains real measured active power sig- appliances. Even compared to the few that have used these
nals taken from 6 real households in USA, with whole home, houses [5], [51], the accuracy level of the method proposed in
appliance/circuit level data at 1/3 samples per second. These this paper is significantly higher [41], [42].
data from the 6 REDD houses (House1-6) were used for the 3) Comparison With State of the Art: In order to demon-
evaluation of the proposed NILM method as the next six tests. strate the strength of the proposed NILM method, obtained
For each house, reading sets RS1, RS2 and RS3 were selected F-Measure values (Afm ) and power disaggregation accuracy
in a similar manner to the previous case (HouseT). Since these (Apa ) values were compared among other state of the art
houses consists of different types of loads and even some NILM methods. In this comparison, all considered algorithms
unknown loads, viability of the proposed NILM algorithm have utilized the data taken from publicly available REDD
under real measurements was validated through this real world dataset [34]. Moreover, use of common accuracy metrics such
scenario. as Afm and Apa suggested in [34] and [50] enabled this direct
2) Results and Discussion: Appliance combination iden- comparison.
tification accuracy (Aci , Afm ), power disaggregation accuracy Table II summaries the overall average F-measure values
(Apa ) and the Average Execution Time (AET-Aet ) obtained for achieved by different state of the art NILM methods including
described test cases are given in Table I. From these results, the proposed method in this paper. Here, for the completeness
it was observed that, with the proposed priori biased NILM of the comparison, fundamentally diverse set of recent and
method (denoted by PBN), for each house tested, turned on benchmark NILM methods were used. These approaches have
appliance combinations were identified with more than 86% been summarized in the following paragraph.
accuracy in-terms of Aci and more than 84% accuracy in- In [24] and [25], supervised and unsupervised Graph Signal
terms of Afm . However, with the priori unbiased NILM method Processing (GSP) based two NILM methods have been dis-
(denoted by PUN), these two values were 76% and 75% cussed. Two benchmark NILM algorithms based on Hidden
respectively. So, the improvement due to the introduction of Markov Models (HMMs) have been introduced in [5] and [35].
AUP based priori biasing technique is substantial. There, [5] uses an unsupervised HMM algorithm while [35]
456 IEEE TRANSACTIONS ON SMART GRID, VOL. 10, NO. 1, JANUARY 2019

TABLE III
C OMPARISON OF O BTAINED P OWER D ISAGGREGATION which uses ‘Powerlets’ Learning (PED), Temporal Multi-Label
ACCURACY VALUES Classification (ML) and Factorial HMM (FHMM) respectively.
Moreover, three HMM based techniques, named Factorial-
Hierarchical Dirichlet Process HMM (F-HDP-HMM), F-
HDP Hidden Semi-Markov Model (F-HDP-HSMM) and
Expectation Maximization FHMM (EM-FHMM) have been
proposed in [33]. Furthermore, in [26], Subtractive Clustering
technique have been evaluated.
From the comparison of the results it is clear that, F-HDP-
HSMM method proposed in [33] and Subtractive Clustering
method proposed in [26] have slightly outperformed the pro-
posed NILM method. However, it should be noted that, these
two methods only have been evaluated for a limited number
of appliances considering only a single household, as reported
in [25], [26], [29], and [33]. In contrast, the proposed NILM
method was evaluated on all six houses. Also, in each house,
power profiles were disaggregated for the seven highest power
consuming appliances in each household.
uses an additive factorial-HMM to identify the turned ON From both comparisons, it can be concluded that, both appli-
appliance combination. Further, NILM via event (i.e., ON/OFF ance identification accuracy and the power level disaggregation
transition) classification is also an emerging technique in the accuracy of the proposed NILM method are comparable or
research. In [28] two such methods have been introduced based superior compared to existing state of the art NILM methods.
on supervised Decision Tree (DT) classifier and a Dynamic
Time Warping (DTW) based classier. Similarly, [27] define
Bayesian classifier and a Viterbi algorithm to address same C. Case Study 2
even classification task. Furthermore, [21] explores multi label This case study was carried out to validate the proposed total
classification (ML-KNN) based NILM methods based on both power demand forecasting technique described in Section IV.
time domain and wavelet domain feature sets. 1) Procedure: Real household data taken from REDD
As reported in [5], [21], [25], [27], [28], and [39], it should Houses 1, 2 & 3 were used for this study. For each of these
be noted that, the F-measure value of each aforementioned chosen houses, signature databases, and extracted AUPs had
NILM method has been evaluated considering only certain already been constructed (using respective RS1, RS2 data in
set of appliances. In most cases, this specific set is selected the Stage A of the proposed NILM method) for the previous
based on the identification accuracy level of each appliance case study. So, for each house, only the priori biased NILM
under the considered NILM method. Further, as mentioned method was evaluated (i.e., the step 5) while generating load
before, some of these NILM methods do not consider REDD forecasts for five minutes ahead of the current time instant.
houses such as House 4, 5 and 6 due to the presence of So, for each house, using its RS2 data, the 5th and 6th steps
unknown appliances. In contrast, the proposed NILM method of the proposed NILM method were re-evaluated.
have been evaluated for all houses of the REDD dataset consid- Next, for the chosen three houses, data corresponding to
ering all appliances in each household. Therefore the proposed 21 different days were selected from their whole house power
NILM algorithm have considered in essence a more challeng- profiles (from RS2 of each house). Then these 21 residential
ing dataset. Despite this challenging nature of the used data, power data profiles (each of length of 24 hours) were con-
according to the results shown in Table II, proposed NILM sidered as data which belongs to 21 different houses for one
method have also outperformed all the other NILM methods complete day. This was done to demonstrate the viability of
in terms of appliance identification accuracy. the prediction technique for a large area.
Another comparison was carried out to compare the Finally, using appliance level power demand predictions
achieved individual appliance power level disaggregation accu- of each of the 21 constructed houses (during a day), the
racies (Apa ) by different state of the art NILM methods total aggregated power demand forecast was estimated with
including the proposed NILM method in this paper. Results confidence levels (i.e., α) of 50%, 70%, 90% and 95%.
of this comparison are summarized in Table III. Here also, 2) Results and Discussion: Actual and predicted total
for the comparison, a fundamentally diverse set of recent and power profiles of the 21 houses are presented in Fig. 8. For
benchmark NILM methods have been used. These approaches demonstration purposes, Fig. 8 displays the prediction result of
are summarized as follows. a 15-minute window out of the power profile which was actu-
In [29], two load disaggregation schemes named as the ally predicted for one complete day with a confidence level
Greedy Solution and the Exact Solution have been proposed of 90%. The accuracy of the total aggregated power predic-
based on Deep Sparse Coding (SC). In the work [31] another tion of all 21 houses for each 3 hour period from 06:00 to
two NILM methods, named General SC and Discriminative SC 24:00 is presented in Table IV. Further, sensitivity of the fore-
have been proposed. For the load disaggregation task, another casting accuracy was studied by varying the confidence level
three avenues have been explored in [22], [30], and [34] parameter α and those results are also shown in Table IV.
WELIKALA et al.: INCORPORATING AUPs FOR NILM AND LOAD FORECASTING 457

Fig. 8. Actual Vs Predicted Total Power Profiles.

TABLE IV
T OTAL P OWER D EMAND P REDICTION ACCURACIES VS C ONFIDENCE Fig. 9. Prediction Accuracy Vs Confidence Level.
L EVEL (α) W HEN T IME S TEP A HEAD = 5 M INUTES

15 minutes ahead. For such time steps, optimum predic-

tion accuracy levels were achieved when the confidence level
is selected as 80% and 50% consecutively. Results of this
investigation is illustrated in Fig. 9.
According to the obtained results, predictions with increased
time step durations ahead are achieved through lowering the
used confidence level. This will slightly decrease the pre-
diction accuracy as shown in Fig. 9. On the other hand,
According to the actual and predicted total power profiles
more accurate predictions for the near future can be obtained
(See Fig. 8), the proposed NILM algorithm has the ability
via increasing the used confidence level. This is a logical
to identify downward steps in the total power demand more
observation as in any case, with higher levels of confidence,
efficiently when compared to identifying the upward trends.
it is not possible to predict the behavior far beyond the
Here, for many appliances as shown in Fig. 2 and in Fig. 7,
near future and vice versa. So, the introduced parameter
observed ON durations would be shorter compared to their
α enabled changing the prediction duration further into the
OFF durations. As a result, it improves the predictability of
future.
instants where appliances are turning OFF compared to the
instants where appliances are turned ON. Thus, prediction
of downward trends in the total power demand by the pro- D. Some Aspects of Installation and Scalability
posed NILM method is more efficient compared to forecasting Several key aspects have been discussed in this chapter on
upward steps. the installation and scalability of the proposed NILM method.
In general, according to the results in Fig. 8 and Table IV, it 1) Installation: The architecture and operation of its hard-
is clear that the proposed NILM method identifies the turned ware installation is illustrated in Fig. 10.
on appliance combinations and predicts the total power con- Here, for each household, only a dedicated processor unit
sumption of number of houses five minutes into the future with a communication channel is required apart from the smart
with reasonable accuracy. meter. Once this residential NILM processor unit is installed
Furthermore, since the predicted breakdown is available at and connected to the smart meter, individual appliance sig-
the appliance level in each house, it enables a power sys- nature learning and signature database construction phases
tem aggregator to predict both the non-critical appliances that (i.e., steps 1 & 2 in Fig. 1) are initiated as described in
could be turned off and the amount of Demand Response Sections II-B and II-C. Since most of the appliance models are
achievable, ahead of an event that could create a possible sys- commonly used in many households, for a large scale imple-
tem emergency. A proper architecture to implement such a mentation, signatures for such appliance models can be taken
mechanism in large scale is shown in Fig. 10 and discussed from global databases. Apart from that, unique appliances for
in Section V-D. the considered house should be learned individually by turn-
3) Inferences From Sensitivity Analysis: From the sensi- ing each of them ON and observing their power profile for a
tivity analysis, a trade off between the used confidence level duration of few hours. At the end of this signature learning
α and the prediction accuracy was observed. Observations in stage, House specific set of signature databases will be stored
Table IV revealed that in order to forecast the total power in the processor unit.
demand of 21 houses 5 minutes ahead, using a 90% confidence After that, steps 3 & 4 (See Fig. 1) of the proposed NILM
level is more accurate than using a 95% or 70% confidence method will be automatically completed in the residentially
levels. installed processor unit. There, first, priori unbiased NILM
In order to further investigate this dependence between method as described in Section II-D will evaluate the appliance
accuracy level, confidence level and the forecasting time usages. This method have been verified by the authors, both
step ahead, same experiment was carried out to forecast in simulations [2] as well as in real-time implementation [45].
the total power demand of 21 houses, both 10 minutes and Next, using the obtained results, in step 4, appliance usage
458 IEEE TRANSACTIONS ON SMART GRID, VOL. 10, NO. 1, JANUARY 2019

Fig. 10. Proposing Overall Hardware Architecture.

TABLE V
patterns of the installed household are extracted and, priori M EMORY R EQUIREMENT FOR DATABASES
biasing technique is formulated in the corresponding NILM
processor as described in Sections III-A and III-B.
There onwards, step 5 & 6 (See Fig. 1) are continu-
ously evaluated in the residentially installed processing unit.
Here, the novel priori biased NILM technique described in sizes will not go beyond the sizes of conventional data storage
Section III-C will evaluate the currently turned ON appliance device.
combination as well as the appliance power level disaggrega- Furthermore, when identifying the currently turned ON
tion. Further, as described in Section IV, power level forecast appliance combination for an observed sliding window
for five minutes ahead of current time value can also be (OSW), the proposed NILM method (i.e., the Algorithm 2)
evaluated for each appliance in the considering household. first starts by considering all possible appliance combinations
2) Increasing Number of Appliances Per House: In the pro- as viable solutions for that OSW. Then onwards, this solution
posed NILM method, all the house specific signature databases space is continuously reduced using one pre elimination stage
are kept pre-stored in the residentially installed processor unit. (PES) and two iterative first and second elimination stages
There, in constructing the appliance combination level signa- (FES & SES) as described in the Section II-D & Algorithm 2.
ture database (CLSD) as described in Section II-C, all possible In order to illustrate the strength of these elimination stages,
appliance combinations for the given set of residential appli- first, a house was synthetically created with 15 different appli-
ances are considered. So, it was observed that the size of this ances taken from REDD dataset. Then, corresponding CLSD
CLSD tends to grow exponentially with the increase of number of that house was constructed as described in II-C. Then, for
of appliances in the household. that CLSD, for all possible static levels of an OSW, the number
Table V was obtained by constructing CLSDs for different of appliance combinations left in the solution space after each
number of residential appliances. For this task, real measure- elimination stage in the 1st iteration was observed. This result
ments obtained from a laboratory setup was utilized. From the is illustrated using a logarithmic plot in Fig. 11 (a). From this
obtained CLSD sizes and data reading times shown in Table V, result, an exponential drop in the solution space was observed
it is clear that even for higher number of appliances, database for every possible static level of an OSW.
WELIKALA et al.: INCORPORATING AUPs FOR NILM AND LOAD FORECASTING 459

Fig. 11. (a) Remaining no. of appliance combinations after each elimination Fig. 12. No. of appliance combinations remaining after each elimination
stage of the first iteration of Algorithm 2 and (b) Likelihood Vs different stage of the first iteration of Algorithm 2 Vs Total no. of appliances in the
possible OSW static levels. household when the Static Level of OSW is (a) 400W and (b) 1000W.

Further, considering the actual total power profile of the (i.e., 32Bps). Further, monthly consumed data amount is cal-
created household, a likelihood function was created for the culated to be 80 MB per house. From the data concentrators
OSW-Static Level. This is shown in Fig. 11 (b). This behavior point of view, in order to monitor 400 houses by one aggrega-
conveys the fact that for an actual household, static levels of tor [54], a communication bandwidth of 100 kbps is required
observed sliding windows, taken from the total power profile between the aggregator and the concentrator. Furthermore,
are more likely to have lower values. For this particular case, inside the data aggregator, in order to carryout the data acqui-
most likely OSW static levels are in the range 0 W - 1000 sition, manipulation and storing for 400 households for 1 s
W (Fig. 11 (b)). From this range, to further study the strength sampling interval, the average execution time was evaluated
of elimination stages, 400 W and 1000 W static levels were to be 0.1225 s for the processor mentioned in Section V-A4.
arbitrarily chosen. All estimated parameter values confirms the ability to
Then, same experiment was carried out while changing the deploy the proposed NILM technique using normal processor
number of appliances in the household. After that, for the units and conventional communication methods. This proves
selected OSW static levels of 400 W and 1000 W, the num- the scalability as well as the feasibility of the proposed NILM
ber of remaining appliance combinations in the solution space technique in a large scale setup.
after each elimination stage was observed. This is shown in
Fig. 12.
These observations conveys that, even though the number of VI. C ONCLUSION
possible appliance combinations grow exponentially with the This paper proposes a novel NILM method with enriched
number of appliances present in the house, the used elimina- capabilities to not only identify turned-on appliances and their
tion stages are capable of eliminating appliance combinations power consumption levels, but also to adapt itself according
in the solution space in an exponential manner so that final to AUPs. Since this NILM solution does not depend solely on
solution is achieved within few iterations. collected measurements, it produces more accurate and robust
3) Increasing Number of Houses Per Aggregator: Once the results compared to existing NILM techniques. The ability to
residential appliances have been classified into critical and non use AUPs in NILM allowed this method to be used to predict
critical categories [52], [53], using the appliance level break- the total power consumption of a number of houses a few
down of the current and forecast power consumptions, the minutes ahead of the present time instant (i.e., real-time). This
values of total / critical / non critical power demands can be has an important practical interest as utilities are reluctant to
calculated for the household as shown in Fig. 10. Then, only utilize DLC for DR due to difficulty of estimating amount of
these six values are to be transmitted from each household to load available for DR ahead of the real time.
the concentrator per every 1 s interval. The method utilizes the KL expansion to separate uncor-
For this purpose, required communication bandwidth related spectral information in active power profiles and
between a house and the concentrator is estimated as 256 bps construct signature databases. Further, it incorporates AUPs
460 IEEE TRANSACTIONS ON SMART GRID, VOL. 10, NO. 1, JANUARY 2019

via a fuzzy logic based priori biasing technique. Since the [22] K. Basu, V. Debusschere, S. Bacha, U. Maulik, and S. Bondyopadhyay,
algorithm performs with high accuracy even on power pro- “Nonintrusive load monitoring: A temporal multilabel classification
approach,” IEEE Trans. Ind. Informat., vol. 11, no. 1, pp. 262–270,
files sampled at low rates, expensive hardware is unnecessary. Feb. 2015.
Furthermore, from the execution speeds achieved, this is a [23] K. He, L. Stankovic, J. Liao, and V. Stankovic, “Non-intrusive load
viable algorithm for a real-time implementation. disaggregation using graph signal processing,” IEEE Trans. Smart Grid,
to be published, doi: 10.1109/TSG.2016.2598872.
[24] V. Stankovic, J. Liao, and L. Stankovic, “A graph-based signal process-
R EFERENCES ing approach for low-rate energy disaggregation,” in Proc. IEEE Symp.
Comput. Intell. Eng. Solutions (CIES), Orlando, FL, USA, Dec. 2014,
[1] G. W. Hart, “Nonintrusive appliance load monitoring,” Proc. IEEE, pp. 81–87.
vol. 80, no. 12, pp. 1870–1891, Dec. 1992. [25] B. Zhao, L. Stankovic, and V. Stankovic, “On a training-less solution for
[2] C. Dinesh et al., “Residential appliance identification based on spectral non-intrusive appliance load monitoring using graph signal processing,”
information of low frequency smart meter measurements,” IEEE Trans. IEEE Access, vol. 4, pp. 1784–1799, 2016.
Smart Grid, vol. 7, no. 6, pp. 2781–2792, Nov. 2016. [26] N. Henao, K. Agbossou, S. Kelouwani, Y. Dubé, and M. Fournier,
[3] H. G. C. P. Dinesh et al., “A subspace signature based approach for “Approach in nonintrusive type I load monitoring using subtractive
residential appliances identification using less informative and low res- clustering,” IEEE Trans. Smart Grid, vol. 8, no. 2, pp. 812–821,
olution smart meter data,” in Proc. 9th Int. Conf. Ind. Inf. Syst. (ICIIS), Mar. 2017.
Gwalior, India, Dec. 2014, pp. 1–6. [27] M. Zeifman, “Disaggregation of home energy display data using prob-
[4] H. G. C. P. Dinesh, P. H. Perera, G. M. R. I. Godaliyadda, abilistic approach,” IEEE Trans. Consum. Electron., vol. 58, no. 1,
M. P. B. Ekanayake, and J. B. Ekanayake, “Residential appliance mon- pp. 23–31, Feb. 2012.
itoring based on low frequency smart meter measurements,” in Proc. [28] J. Liao, G. Elafoudi, L. Stankovic, and V. Stankovic, “Power disag-
IEEE Int. Conf. Smart Grid Commun. (SmartGridComm), Miami, FL, gregation for low-sampling rate data,” in Proc. 2nd Int. Non Intrusive
USA, Nov. 2015, pp. 878–884. Appliance Load Monitor. Workshop, Austin, TX, USA, 2014.
[5] O. Parson, S. Ghosh, M. Weal, and A. Rogers, “Non-intrusive load mon- [29] S. Singh and A. Majumdar, “Deep sparse coding for non-intrusive
itoring using prior models of general appliance types,” in Proc. 26th load monitoring,” IEEE Trans. Smart Grid, to be published,
Conf. Artif. Intell. (AAAI), 2012, pp. 356–362. doi: 10.1109/TSG.2017.2666220.
[6] G. W. Hart, Nonintrusive Appliance Load Data Acquisition Method: [30] E. Elhamifar and S. Sastry, “Energy disaggregation via learning
Progress Report. Concord, MA, USA: MIT Energy Lab., 1984. [Online]. ‘powerlets’ and sparse coding,” in Proc. 29th AAAI Conf. Artif.
Available: https://books.google.lk/books?id=gYlYtwAACAAJ Intell. (AAAI), Austin, TX, USA, 2015, pp. 629–635. [Online].
[7] W. Wichakool, A.-T. Avestruz, R. W. Cox, and S. B. Leeb, “Modeling Available: http://dl.acm.org/citation.cfm?id=2887007.2887095
and estimating current harmonics of variable electronic loads,” IEEE [31] J. Z. Kolter, S. Batra, and A. Y. Ng, “Energy disaggregation via dis-
Trans. Power Electron., vol. 24, no. 12, pp. 2803–2811, Dec. 2009. criminative sparse coding,” in Proc. Adv. Neural Inf. Process. Syst.,
[8] S. R. Shaw, S. B. Leeb, L. K. Norford, and R. W. Cox, “Nonintrusive Vancouver, BC, Canada, 2010, pp. 1153–1161.
load monitoring and diagnostics in power systems,” IEEE Trans. [32] S. Makonin, F. Popowich, I. V. Bajić, B. Gill, and L. Bartram,
Instrum. Meas., vol. 57, no. 7, pp. 1445–1454, Jul. 2008. “Exploiting HMM sparsity to perform online real-time nonintrusive load
[9] T. Hassan, F. Javed, and N. Arshad, “An empirical investigation of V-I monitoring,” IEEE Trans. Smart Grid, vol. 7, no. 6, pp. 2575–2585,
trajectory based load signatures for non-intrusive load monitoring,” IEEE Nov. 2016.
Trans. Smart Grid, vol. 5, no. 2, pp. 870–878, Mar. 2014. [33] M. J. Johnson and A. S. Willsky, “Bayesian nonparametric hid-
[10] W. Wichakool, Z. Remscrim, U. A. Orji, and S. B. Leeb, “Smart meter- den semi-Markov models,” J. Mach. Learn. Res., vol. 14, no. 1,
ing of variable power loads,” IEEE Trans. Smart Grid, vol. 6, no. 1, pp. 673–701, Jan. 2013. [Online]. Available: http://dl.acm.org/ cita-
pp. 189–198, Jan. 2015. tion.cfm?id=2502581.2502602
[11] Y.-H. Lin and M.-S. Tsai, “Non-intrusive load monitoring by novel [34] J. Kolter and M. J. Johnson, “Redd: A public data set for
neuro-fuzzy classification considering uncertainties,” IEEE Trans. Smart energy disaggregation research,” in Proc. Workshop Data Min. Appl.
Grid, vol. 5, no. 5, pp. 2376–2384, Sep. 2014. Sustain. (SIGKDD), San Diego, CA, USA, 2011.
[12] M. Hazas, A. Friday, and J. Scott, “Look back before leaping forward: [35] J. Z. Kolter and T. Jaakkola, “Approximate inference in additive factorial
Four decades of domestic energy inquiry,” IEEE Pervasive Comput., HMMs with application to energy disaggregation,” in Proc. Int. Conf.
vol. 10, no. 1, pp. 13–19, Jan./Mar. 2011. Artif. Intell. Stat., 2012, pp. 1472–1482.
[13] H. Y. Lam, G. S. K. Fung, and W. K. Lee, “A novel method to construct [36] M. Aiad and P. H. Lee, “Unsupervised approach for load dis-
taxonomy electrical appliances based on load signaturesof,” IEEE Trans. aggregation with devices interactions,” Energy Build., vol. 116,
Consum. Electron., vol. 53, no. 2, pp. 653–660, May 2007. pp. 96–103, 2016. [Online]. Available: http://www.sciencedirect.com/
[14] Y.-H. Lin and M.-S. Tsai, “Development of an improved time–frequency science/article/pii/S0378778815304795
analysis-based nonintrusive load monitor for load demand identifica- [37] O. Parson, S. Ghosh, M. Weal, and A. Rogers, “Using hidden Markov
tion,” IEEE Trans. Instrum. Meas., vol. 63, no. 6, pp. 1470–1483, models for iterative non-intrusive appliance monitoring,” in presented
Jun. 2014. at the Neural Inf. Process. Syst. Workshop Mach. Learn. Sustain.,
[15] F. Sultanem, “Using appliance signatures for monitoring residential 2011.
loads at meter panel level,” IEEE Trans. Power Del., vol. 6, no. 4, [38] S. Pattem, “Unsupervised disaggregation for non-intrusive load monitor-
pp. 1380–1385, Oct. 1991. ing,” in Proc. 11th Int. Conf. Mach. Learn. Appl., vol. 2. Boca Raton,
[16] C. Laughman et al., “Power signature analysis,” IEEE Power Energy FL, USA, Dec. 2012, pp. 515–520.
Mag., vol. 1, no. 2, pp. 56–63, Mar./Apr. 2003. [39] R. Bonfigli, S. Squartini, M. Fagiani, and F. Piazza, “Unsupervised algo-
[17] H.-H. Chang, K.-L. Chen, Y.-P. Tsai, and W.-J. Lee, “A new measure- rithms for non-intrusive load monitoring: An up-to-date overview,” in
ment method for power signatures of nonintrusive demand monitoring Proc. IEEE 15th Int. Conf. Environ. Elect. Eng. (EEEIC), Rome, Italy,
and load identification,” IEEE Trans. Ind. Appl., vol. 48, no. 2, Jun. 2015, pp. 1175–1180.
pp. 764–771, Mar./Apr. 2012. [40] M. Zeifman and K. Roth, “Nonintrusive appliance load monitoring:
[18] S. R. Shaw and C. R. Laughman, “A Kalman-filter spectral enve- Review and outlook,” IEEE Trans. Consum. Electron., vol. 57, no. 1,
lope preprocessor,” IEEE Trans. Instrum. Meas., vol. 56, no. 5, pp. 76–84, Feb. 2011.
pp. 2010–2017, Oct. 2007. [41] A. Zoha, A. Gluhak, M. A. Imran, and S. Rajasegarar, “Non-intrusive
[19] Y.-H. Lin and M.-S. Tsai, “Development of an improved–time frequency load monitoring approaches for disaggregated energy sensing: A survey,”
analysis-based nonintrusive load monitor for load demand identifica- Sensors, vol. 12, no. 12, pp. 16838–16866, Dec. 2012.
tion,” IEEE Trans. Instrum. Meas., vol. 63, no. 6, pp. 1470–1483, [42] C. Dinesh, P. Perera, R. I. Godaliyadda, M. P. B. Ekanayake, and
Jun. 2014. J. B. Ekanayake, “Non-intrusive load monitoring based on low fre-
[20] M. Dong, P. C. M. Meira, W. Xu, and W. Freitas, “An event window quency active power measurements,” AIMS Energy, vol. 4, no. 3,
based load monitoring technique for smart meters,” IEEE Trans. Smart pp. 414–443, 2016. [Online]. Available: http://www.aimspress.com/
Grid, vol. 3, no. 2, pp. 782–796, Jun. 2012. energy/article/708.html
[21] S. M. Tabatabaei, S. Dick, and W. Xu, “Toward non-intrusive load mon- [43] A. Reinhardt et al., “On the accuracy of appliance identification based
itoring via multi-label classification,” IEEE Trans. Smart Grid, vol. 8, on distributed load metering data,” in Proc. 2nd IFIP Conf. Sustain.
no. 1, pp. 26–40, Jan. 2017. Internet ICT Sustain. (SustainIT), Pisa, Italy, 2012, pp. 1–9.
WELIKALA et al.: INCORPORATING AUPs FOR NILM AND LOAD FORECASTING 461

[44] N. Batra et al., “NILMTK: An open source toolkit for non- Mervyn Parakrama B. Ekanayake (M’13)
intrusive load monitoring,” in Proc. 5th Int. Conf. Future Energy received the B.Sc.Eng. degree in electrical and
Syst. (e-Energy), Cambridge, U.K., 2014, pp. 265–276. [Online]. electronic engineering from the University of
Available: http://doi.acm.org/10.1145/2602044.2602051 Peradeniya, Sri Lanka, in 2006, and the Ph.D. degree
[45] S. Welikala, C. Dinesh, M. P. B. Ekanayake, R. I. Godaliyadda, and from Texas Tech University, in 2011. He is cur-
J. Ekanayake, “A real-time non intrusive load monitoring system,” in rently a Faculty Member with the Department of
Proc. IEEE 11th Int. Conf. Ind. Inf. Syst. (ICIIS), Roorkee, India, Electrical and Electronic Engineering, University of
Dec. 2016. Peradeniya. His current research activities include
[46] H. G. C. P. Dinesh, P. H. Perera, G. M. R. I. Godaliyadda, blind source separation and blind event separa-
M. P. B. Ekanayake, and J. B. Ekanayake, “Individual power profile tion, visual systems and sensor fusion techniques
estimation of residential appliances using low frequency smart meter for mobile robots, bio realistic modeling of visual
data,” in Proc. IEEE 10th Int. Conf. Ind. Inf. Syst. (ICIIS), Peradeniya, systems, and coding of sensory information in biological systems.
Sri Lanka, Dec. 2015, pp. 140–145.
[47] B. Bouchon-Meunier, M. Dotoli, and B. Maione, “On the choice of
membership functions in a mamdani-type fuzzy controller,” in Proc. 1st
Online Workshop Soft Comput. (WSC1), Nagoya, Japan, 1996.
[48] S. S. Jamsandekar and R. R. Mudholkar, “Fuzzy classification system by
self generated membership function using clustering technique,” BIJIT
BVICAM’s Int. J. Inf. Technol., vol. 6, no. 1, pp. 697–704, 2014.
[49] T. M. Cover and J. A. Thomas, Elements of Information Theory (Wiley
Series in Telecommunications and Signal Processing). Hoboken, NJ,
USA: Wiley, 2006.
[50] D. L. Olson and D. Delen, Advanced Data Mining Techniques,
1st ed. Heidelberg, Germany: Springer, 2008. [Online]. Available:
https://books.google.com/books?id=2vb-LZEn8uUC&printsec=
copyright&source=gbs_pub_info_r#v=onepage&q&f=false
[51] O. Parson, S. Ghosh, M. Weal, and A. Rogers, “An unsupervised train- Roshan Indika Godaliyadda (M’14) received
ing method for non-intrusive appliance load monitoring,” Artif. Intell., the B.Sc.Eng. degree in electrical and electronic
vol. 217, pp. 1–19, Dec. 2014. engineering from the University of Peradeniya,
[52] M. Shad et al., “Identification and estimation for electric water heaters Sri Lanka, in 2005, and the Ph.D. degree from the
in direct load control programs,” IEEE Trans. Smart Grid, vol. 8, no. 2, National University of Singapore, in 2011. He was
pp. 947–955, Mar. 2017. with Dialog Telekom. He is currently a Faculty
[53] A. Gholian, H. Mohsenian-Rad, and Y. Hua, “Optimal industrial Member with the Department of Electrical and
load control in smart grid,” IEEE Trans. Smart Grid, vol. 7, no. 5, Electronic Engineering, University of Peradeniya.
pp. 2305–2316, Sep. 2016. His current research interests include image and sig-
[54] R. R. Mohassel, A. Fung, F. Mohammadi, and K. Raahemifar, nal processing, subspace techniques, pattern recogni-
“A survey on advanced metering infrastructure,” Int. J. Elect. Power tion, blind source separation techniques, positioning
Energy Syst., vol. 63, pp. 473–484, Dec. 2014. [Online]. Available: systems, computer vision, adaptive techniques, nonlinear systems and control,
http://www.sciencedirect.com/science/article/pii/S0142061514003743 and wireless communications.

Shirantha Welikala (S’13) received the B.Sc.Eng.

degree in electrical and electronic engineering from
the University of Peradeniya, Sri Lanka, in 2015,
where he joined the Department of Electrical and
Electronic Engineering as a Temporary Instructor
and is currently a Research Assistant. His main
research interests include control systems, robotics,
intelligent systems, electronics, embedded systems,
and signal processing.

Janaka Ekanayake (F’17) received the B.Sc. degree

in electrical and electronic engineering from the
University of Peradeniya, Sri Lanka, and the Ph.D.
Chinthaka Dinesh (S’15) received the B.Sc.Eng. degree from UMIST, U.K. He joined the University
degree in electrical and electronic engineering from of Peradeniya as a Lecturer and was promoted to a
the University of Peradeniya, Sri Lanka, in 2012. Professor in electrical engineering in 2003. In 2008,
He is currently pursuing the M.A.Sc. degree with he joined the Cardiff School of Engineering, U.K.
the School of Engineering Science, Simon Fraser He is currently with the University of Peradeniya and
University, Canada. His current research interests Cardiff University. He has published over 50 papers
include signal processing for smart grid applica- in refereed journals and has also co-authored four
tions, 3-D image processing, and graph-based signal books. His main research interests include power
processing for various applications. electronic applications for power systems, renewable energy generation and
its integration, and smart grid applications. He is a fellow of IET.