Beruflich Dokumente
Kultur Dokumente
Zeinab Moshavasha
a
Department of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Iran
b
Department of Management & Innovation Systems, University of Salerno, Fisciano, Italy
c
Dipartimento Energia “Galileo Ferraris”, Politecnico di Torino, Italy
Keywords: This paper proposes a novel model to cluster similar load consumption patterns and identify time periods with
Electricity customer clustering similar consumption levels. The model represents the customer’s load pattern as an image and takes into account
Intuitionistic fuzzy divergence the load variation and uncertainty by using exponential intuitionistic fuzzy entropy. The advantage is that the
Load pattern proposed method can handle the uncertain nature of customer’s load, by adding a hesitation index to the
Smart meters
membership and non-membership functions. A multi-level representation of the load patterns is then provided
Time period clustering
by creating specific bands for the load pattern amplitudes using intuitionistic fuzzy divergence-based thresh-
Typical load pattern
olding. The typical load pattern is then determined for each customer. In order to reduce the number of features
to represent each load pattern with respect to the time-domain data, the discrete wavelet transform is used to
extract some spectral features. To cope with the data representation with fuzzy rules, the fuzzy c-means is
implemented as the clustering algorithm. The proposed approach also identifies the time periods associated to
different load pattern levels, providing useful hints for demand side management policies. The proposed method
has been tested on ninety low voltage distribution grid customers, and its superior effectiveness with respect to
the classical k-means algorithm has been represented by showing the better values obtained for a set of clus-
tering validity indicators. The combination of load pattern clusters and time periods associated with the seg-
mented load pattern amplitudes provides exploitable information for the efficient design and implementation of
innovative energy services such as demand response for different customer categories.
⁎
Corresponding author at: Department of Electrical and Electronic Engineering, Shiraz University of Technology, Modares Blvd., Shiraz P.O. 71555-313, Iran.
E-mail address: gitizadeh@sutech.ac.ir (M. Gitizadeh).
https://doi.org/10.1016/j.ijepes.2019.105624
Received 11 April 2018; Received in revised form 10 August 2019; Accepted 10 October 2019
0142-0615/ © 2019 Published by Elsevier Ltd.
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
N number of occurrences clustering the daily load patterns of a set of customers supplied by the
R number of thresholds
low voltage distribution grid, to group similar consumption patterns.
W number of DWT levels
Y number of hours The main motivation of the study is to model together the uncertainty
Div divergence of load data and the non-determinacy (hesitation) in the definition of
probability of occurrence the data. This is the first application to electrical load pattern clustering
α conditional probability coefficient of a full approach that handles together uncertainty and non-determi-
γ time resolution data in one hour
δ integer number to identify the time lag used for the days
nacy.
∊ integer number to identify the time lag used for the time samples In this framework, the contributions of this paper are various and
l daily load pattern are aimed at customizing the treatment of uncertain data and im-
μ degree of membership precise/unknown information to the specific needs of the electrical load
degree of non-membership
ν
pattern data analysis. In particular:
π degree of uncertainty
standard deviation
τ threshold - The leading idea of this paper is that the load pattern of each cus-
l dummy index tomer is considered as an image, in which each load value is assigned
d detail coefficient vector in DWT as a pixel. A discrete set of colors for the pixels is identified to re-
l daily load pattern vector
present a given number of load levels.
A, B load matrices
F, G load matrices - A thresholding method based on an image processing technique –
M load matrix the intuitionistic fuzzy divergence (IFD) – is then applied to segment
Q, T matrices for time period clustering the images. The IFD technique is exploited for evaluating the load
L set of daily load patterns
uncertainty and variation within membership and non-membership
X generic set
functions, and incorporates a further hesitation term to model the
lack of knowledge on the membership or non-membership (the de-
1. Introduction tails are indicated in Section 2.5).
- In order to guarantee high separation accuracy for uncertain cus-
In smart grid applications, implementation of advanced metering tomer’s load, a new divergence formula based on intuitionistic fuzzy
infrastructure as one of the key technologies leads to the availability of entropy is derived. The intuitionistic fuzzy entropy is then used in a
a large amount of data that represent the consumer electricity load minimization procedure to extract for each customer a typical load
patterns, which is useful for both energy system planning and operation pattern, by using neighbor information (2-dimensional daily load
[1]. The extracted information can be used in demand response pro- values).
grams [2], load profiling [3], prediction of the customers’ consumption - For feature extraction, the discrete wavelet transform (DWT) is then
[4], and theft detection [5]. In these applications, the customer clus- implemented to pass from time domain data to a lower number of
tering in which the similar patterns are grouped together is considered features, making the execution of the clustering algorithm faster.
to be an essential step. - In order to cope with the data representation with fuzzy rules, the
A key point that has been addressed in a few previous works is the fuzzy c-means clustering algorithm is used to handle the data con-
ability of the clustering algorithms to consider the uncertainties asso- structed in the proposed framework. Eventually, appropriate clus-
ciated with the data. These uncertainties may condition the clustering tering validity indicators are calculated to assess the clustering re-
results. It has been widely recognized that the quality of the clustering sults.
results can be improved in a significant way by taking into account - Finally, a method based again on IFD thresholding is proposed to
information on data uncertainty [6]. However, there are different ways identify the time periods with similar consumption levels and is
to include uncertainty. Applications of the probability theory, and the introduced in the framework formulated for load pattern analysis.
data representation through fuzzy variables, are suitable ways to in- The hours of the day that are similar with respect to the demand
corporate the effects of the uncertainty. Moreover, a number of soft profile are grouped together.
clustering approaches have been proposed in the literature to take into
account also imprecise and/or incomplete information. The similar day issue is often encountered in the electricity utility
There are basically three ways to include the effects of uncertainty industry. However, only a few references focus on time periods clus-
and imprecise/incomplete information: tering [9]. This also depends on the fact that in most of the proposed
works the partitioning of the time periods is given as an input. Table 1
(a) Use of deterministic data in non-deterministic clustering proce- shows an example of time block division used in time-of-use pricing
dures. [10].
(b) Exploitation of data uncertainty without considering imprecise/in- The time periods may be grouped to determine peak, shoulder,
complete information. valley, or other types of aggregate periods defined for different pur-
(c) Exploitation of data uncertainty together with imprecise/in- poses in DSM policies. The identification of the time periods with dif-
complete information. ferent levels of consumption is becoming more and more relevant in the
context of studying the formulation of demand response programs. In
Considering the applications to load pattern grouping and profiling, this paper, starting from the clustering results, the IFD-based thresh-
in case a) many references use deterministic data as inputs to solution olding concepts are used to formulate an adaptive procedure to identify
methods based on probability or fuzzy logic. The most common ex- the time periods associated to different load pattern levels. In this way,
amples are various applications of fuzzy c-means [1], or load profiling
methods based on the combination of fuzzy logic and probability neural
networks [7] or on the weighted fuzzy average k-means [8]. In case b), Table 1
Example of time block division throughout a day.
Gaussian models and other approaches have been typically used. Con-
cerning case c), dealing with imprecise/unknown information has not Period Hours
been addressed yet in the contributions currently available for electrical
load pattern clustering. The details referring to case b) and case c) are Peak 2:00 p.m. to 7:00 p.m.
Shoulder (semi-peak) 5:00 a.m. to 2:00 p.m., and 7:00 p.m. to 12:00 p.m.
discussed in the next section.
Valley (off-peak) 12:00 p.m. to 5:00 a.m.
In this context, this paper introduces an innovative framework for
2
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
a complete and consistent application of the IFD-based approach is 2.2. Review of recent works on load pattern clustering without uncertainty
provided.
The rest of this paper is organized as follows. Section 2 reviews the In the following, a brief review of recent literature contributions on
literature and presents a background on load pattern clustering load pattern clustering is presented, to highlight the variety of con-
methods and intuitionistic fuzzy sets. In this section, it is clarified that tributions that do not handle data uncertainty (the approaches that
uncertainty modeling has been incorporated in the clustering algo- consider uncertainty are addressed in Section 2.4). A data-driven ap-
rithms in different ways. A basic distinction is made between general proach is applied in [2] to determine the shape of seasonally-resolved
aspects of uncertainty modeling for clustering (with a discussion on the residential demand profiles, as well as the optimal number of normal-
main approaches proposed to address imprecise/incomplete informa- ized representative residential electricity use profiles. A support vector
tion), and specific clustering applications for electrical load pattern machine is proposed in [5] considering a consumption pattern-based
analysis and profiling (where no existing approach handles imprecise/ energy theft detector in an advanced metering infrastructure. A clus-
incomplete information). The details of the proposed clustering ap- tering model based on self-organizing maps is presented in [16] for
proach are elaborated in Section 3. In Section 4, the performance of the creating a series of representative electricity load profile classes for the
proposed methodology is illustrated through a realistic case study and domestic sector in Ireland.
thoroughly discussed. Section 5 contains the relevant conclusions. A comprehensive study of the electricity consumption profiles is
presented in [17] for a Southwest European city, through the combi-
nation of data from smart meters with door-to-door question surveys for
2. Background a sample of 265 households. A method based on frequency analysis is
proposed in [18] to recognize the relation between the energy con-
2.1. Load pattern clustering process sumption and the lifestyle of the residents in a household. An entropy-
based model is proposed in [19] to cluster the electricity consumers
Generally, the load pattern clustering process can be divided into according to their daily load patterns and to identify the outliers, while
five stages: load data preparation, typical load pattern (TLP) determi- the effect of the temporal resolution of load profiles on the quality and
nation per customer, feature extraction, clustering algorithm, and efficiency of the clustering process is analyzed in [20]. Ant colony in
clustering evaluation [1]. In the first stage, the electrical consumption [21] and g-means in [22] are proposed to promote the quality of the
data of individual customers, read by the smart meters at given time clustering process. Clusters that are applicable to real conditions based
steps (e.g., 15 min, 30 min or one hour), are combined on a load pattern on domain expert knowledge are formulated in [23]. A dynamic clus-
defined for a certain time period (e.g., one day to form daily load tering analysis is performed in [24] as an efficient tool for customers’
patterns). Possible bad data are then eliminated. In the second stage, classification and trend behavior. A classification of residential elec-
the TLP of each customer is determined from the load patterns available tricity customers is proposed in [25] by using model-based feature se-
in similar conditions (e.g., weekdays) by combining the load values that lection.
correspond to the same time step according to a statistical criterion A segmentation model, which uses features derived from the en-
(e.g., mean or median). In this regard, the TLP becomes a useful re- coded data, is proposed in [26]. A classification model is applied to
presentative of the customer’s behavior in normal operating conditions, enable the classification of new consumers using a set of normalized
in which the amount of data to be clustered is significantly reduced. shape indices as features in [27]. Patterns and trends in energy usage
Load profiling is traditionally used to group together similar TLPs and profiles of commercial and industrial customers are identified in [28]
create a load profile for each group. for energy efficiency programs. A computational method capable of
In the third stage, the most commonly used features are the TLP handling large amounts of data is proposed in [29] by dividing custo-
time series values themselves. In alternative to the raw time domain mers into user groups based on the similarities of their electricity use
data, some approaches use frequency domain features. For example, in behavior. Hopfield k-means clustering algorithm is proposed in [30] in
[11] triplets of features have been introduced for each positive har- order to overcome the limitations of other algorithms such as ran-
monic, together with one feature for the continuous component, then domness of the solution (k-means), the lack of pre-allocation of the
the relevant features selected correspond to the most significant har- number of clusters (follow the leader), and the improvement of the
monics. In [12] the major components in the frequency domain are solution obtained (hierarchical algorithm).
selected by using a classification and regression tree. In [13] the fea- Subspace clustering is used in [31], where projected clustering
tures are selected from the truncated Fourier series. While determining methods are proposed to capture such subspaces of load patterns that
the features, an interesting possibility refers to reducing the number of allow reaching the desired subspace clustering result in fewer steps.
data that are stored for each customer and are sent to the clustering Feature construction and calibration methods for clustering daily
step. Data size reduction can be performed by using methods such as electricity load curves are proposed and compared in [32].
the principal component analysis, the canonical variate analysis [14],
and others [15]. 2.3. General aspects of uncertainty modeling for clustering
At the fourth stage, on the basis of the features defined, clustering
techniques are used to perform load pattern grouping. Many re- In general terms, the main techniques used are based on the prob-
searchers have focused on the clustering techniques applied on elec- ability theory and on soft clustering approaches (in particular, fuzzy
trical load data [1], for example k-means [2], k-medoids [16], hier- logic and rough sets [33]). In the applications of the probability theory,
archical [17], fuzzy c-means [18], entropy-based [19], mixture models the data uncertainty can be determined from the technique used to
[20], ant colony clustering [21], and g-means [22]. A summary of the measure or construct the data. This uncertainty can be expressed by
techniques used in various literature papers is reviewed in [1]. using error variances, or by assessing the probability density function
In the last stage, different clustering validity indicators have been with which the data is available. Fuzzy logic applications have been
defined to evaluate the effectiveness of the clustering method, as illu- introduced to represent the ambiguity due to the similarity between
strated in [1] and [22]. Each indicator can be calculated on a data set input data and clusters, by enabling each input data to belong to dif-
formed by using either time-domain or frequency-domain data. Com- ferent clusters with a degree of membership to each cluster. In parti-
parison among the clustering results is meaningful only when the same cular, in the fuzzy c-means clustering [34] each input can be assigned to
type of feature is considered to calculate the clustering validity in- different clusters, in such a way that the sum of the degrees of mem-
dicators. berships of each input to all the clusters is equal to unity [35]. An
evolution of the fuzzy set theory has been presented in [36] with the
3
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
introduction of the intuitionistic fuzzy sets. The intuitionistic fuzzy set certain customer characteristics. The spectral clustering and the Ex-
theory includes the definition of three terms summing up to unity, pectation maximization are applied in [56] together with fuzzy c-means
namely, a membership function, a non-membership function, and a to assess the most informative load pattern features with a supervised
further contribution called hesitancy function, which represents im- approach. Probabilistic baseline estimations are proposed in [57]
precise and/or incomplete information. Clustering methods based on through a probabilistic Gaussian process, and in [58] through a quantile
intuitionistic fuzzy sets have been developed only at a later time, regression forests model. Moreover, daily load profiles are used in [59]
starting from the approach based on the intuitionistic fuzzy equivalence to construct the most probable weekly or yearly load profiles.
matrix developed in [37] (other contributions are reviewed in [38;39]). This literature review clearly confirms that there is no application to
Further approaches have discussed the hypotheses used in the fuzzy electrical load pattern grouping or load profiling of methods like in-
c-means approach, considering that in the presence of noisy data or tuitionistic fuzzy sets, possibilistic clustering, rough sets, shadowed
outliers it is important to introduce a distinction between the concepts sets, credal partition, or evidential c-means methods. Some applications
“equally likely” and “unknown” membership to the clusters. This dis- of these methods refer to load forecasting, whose review and discussion
tinction has been introduced in the possibilistic clustering [40], where is outside the scope of this paper.
the sum of the membership functions over the clusters is no longer
constrained to be equal to unity. In this way, the importance of the 2.5. Intuitionistic fuzzy sets
outliers in the creation of the centroids is reduced. On another point of
view, a noise cluster that has a constant distance from all the inputs has The fuzzy set theory was proposed by Zadeh [60] to deal with un-
been introduced in [41] in order to collect the inputs having reduced certainty of information. To better represent lack of significant in-
representation in the clusters. The hypothesis of constant distance has formation, Atanassov [36] extended the theory by introducing the in-
been later removed in [42] to obtain a generalization of the possibilistic tuitionistic fuzzy set (IFS), which is a very useful tool in handling non-
clustering technique. determinacy (hesitation) in the system.
Rough sets applications have been introduced to add a further di- To date, IFS has been applied to a range of problems, for example,
mension to characterize the uncertainty due to missing or wrong in- medical images segmentation [61–64], multilevel programming [65],
formation. In particular, the rough k-means has been introduced in pattern recognition [66], etc. Along with the membership function, IFS
[43]. The applications of rough k-means have been deemed to be su- also contains the non-membership function and the hesitancy function.
perior to the sole use of the basic k-means in reducing the number of The latter is capable of representing lack of information more accu-
incorrectly clustered inputs [44]. Further evolutions have been based rately. As a result, IFS is able to model situations where the classical
on hybridizations of soft computing approaches, to construct algorithms fuzzy set theory fails to use all the available information. This is pre-
more robust with respect to outliers and initial parameter settings with cisely the idea behind the usage of IFS for the purpose of load pattern
respect to the individual approaches [33]. Various hybridizations of clustering and time period segmentation (see Table 2).
fuzzy and rough concepts have been presented [45]. The technique In every IFS, the degrees of membership and non-membership, as
developed in [46] integrates c-means, probabilistic and possibilistic well as the hesitation, are defined for every element x ∈ X in the finite
memberships of fuzzy sets, and rough sets. The rough-fuzzy collabora- universe X. Hence, the IFS set S is defined as follows [36]:
tive clustering described in [47] applies collaboration principles to the
analysis of clustered subsets, in order to exchange information among S = { x , µS (x ), S (x ) |x X} (1)
these subsets with the aim to modify the cluster centroids and reach a
where the functions µS : X → [0, 1] and S : X → [0, 1] denote a degree
stable structure. The concept of shadowed sets introduced in [48] to
of membership and non-membership of the element x ∈ X to set S, re-
support a three-value logic able to represent three interpretations of an
spectively; such that 0 ≤ µS (x) + S (x) ≤ 1. The function
outcome (yes, no, and unknown) has been used in the approaches
presented in [49] and in [50]. S (x ) =1 µS (x ) S (x ); x X (2)
Furthermore, the concept of credal partition, based on the belief
functions theory (also called evidence theory) has been introduced in is called the intuitionistic fuzzy index or the hesitation index (πS(x)),
[51] to extend the concepts of hard, fuzzy (or probabilistic), possibi- denoting lack of knowledge on whether x belongs to S or not.
listic and rough partition. The credal partition is obtained by quanti- The distances between IFSs should be calculated by taking into
fying the uncertainty of the cluster membership by using mass functions account three parameters that describe each IFS. Between two IFSs, the
for each input that cannot be assigned with certainty to clusters. The intuitionistic fuzzy divergence (IFD) measures the extent to which the
concepts of credal partition and noise cluster have been used in the two sets differ from each other. There are different formulas for cal-
evidential c-means clustering technique proposed in [52]. culating the divergence [66] and [67]. The Hamming distance is the
most popular divergence formula. The Hamming distance DivSP be-
2.4. Literature review on uncertainty in load pattern clustering methods tween two sets S and P, whose elements belong to the universe X, is
defined as follows:
In spite of the wide evolution of the concepts to address uncertain
DivSP = (|µS (x ) µP (x )| + | S (x ) P (x )| + | S (x ) P (x )|)
information, only a few applications have been proposed for managing x X
uncertainties for electrical load patterns grouping or load profiling
(3)
purposes. Some applications refer to Gaussian models. In particular,
linear Gaussian models are presented in [3] in order to capture multiple In particular, if S = P, then DivSP = 0.
behaviors exhibited by homogenous residential customers. Further- In the IFS-based segmentation problem, the segments can be shown
more, a Gaussian mixture model (GMM) is used in [53] to capture as multi-intuitionistic fuzzy subsets. Some measures can be optimized
random effects, as well as in [54] to provide a probabilistic cluster as objective functions of segmentation, among them there are the fuz-
membership. A finite mixture model of Gaussian multivariate dis- ziness index, the non-fuzziness index, and the entropy index. The ex-
tributions is proposed in [55] for clustering and analyzing the peak ponential entropy for a matrix G of size I × J, with components gi,j and
demand and identifying the major sources of variability in the elec- having L different values (the matrix entries are discrete, hence the
tricity usage behavior for residential customers. Moreover, a hidden same value may be found more times in the matrix) is defined as
Markov model is presented in [4] to estimate the randomness in the L 1
customers’ consumption and to demonstrate that temporal patterns in H (G) = (Probl e (1 Probl ) )
the customer’s consumption data can predict with good accuracy l =0 (4)
4
Table 2
Comparison of the reviewed load profile clustering papers.
k-means Hierarchical FCM Follow the k-medoids Mixture Others 0–500 500–1000 1000–2500 > 2500 MIA CDI DBI Others
leader models
[1] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
[2] ✓ ✓
[3] ✓ ✓ ✓
[4] ✓ ✓
[5] ✓ ✓ ✓
[11] ✓ ✓ ✓ ✓ ✓ ✓
[12] ✓ ✓ ✓
[14] ✓ ✓ ✓
[15] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
[16] ✓ ✓ ✓ ✓ ✓
[17] ✓ ✓
[18] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[19] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[20] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
[21] ✓ ✓ ✓ ✓ ✓
[22] ✓ ✓ ✓ ✓ ✓ ✓ ✓
[23] ✓ ✓ ✓ ✓ ✓
[24] ✓ ✓
[25] ✓ ✓ ✓ ✓
[26] ✓ ✓ ✓ ✓
[27] ✓ ✓ ✓ ✓ ✓ ✓
5
[28] ✓ ✓ ✓
[29] ✓ ✓ ✓ ✓
[30] ✓ ✓ ✓ ✓
[31] ✓ ✓ ✓ ✓
[32] ✓ ✓ ✓ ✓
[53] ✓ ✓ ✓
[54] ✓ ✓ ✓ ✓
[55] ✓ ✓ ✓
This paper ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Ref. Data region Time sample resolution Type of customer Period of measurement
Europe North Asia Other 1 min 10 min 15 min 30 min 60 min Industrial Residential Mix 1 day – 3–6 months 7–12 months years > 2 years
America 3 months
[1] ✓ ✓ ✓ ✓
[2] ✓ ✓ ✓ ✓
[3] ✓ ✓ ✓ ✓
[4] ✓ ✓ ✓ ✓
[5] ✓ ✓ ✓ ✓
[11] ✓ ✓ ✓ ✓
[12] ✓ ✓ ✓ ✓
[14] ✓ ✓ ✓ ✓
[15] ✓ ✓ ✓ ✓
[16] ✓ ✓ ✓ ✓
[17] ✓ ✓ ✓ ✓
[18] ✓ ✓ ✓ ✓
[19] ✓ ✓ ✓ ✓
[20] ✓ ✓ ✓ ✓
(continued on next page)
Electrical Power and Energy Systems 117 (2020) 105624
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
(5)
✓ where µG ( gi, j ) is the membership degree of the (i,j)th element. Based on
✓
✓
✓
Eq. (5) and extending to an IFD, the fuzzy divergence formula between
7–12 months
DivGF
I J
✓
✓
✓
✓
✓
1
= (4 {1 µG (gi, j ) + µ F (fi, j )} e µG (gi, j) µF (fi, j )
2
3–6 months
i=1 j=1
Period of measurement
✓
e G (gi, j ) F (fi, j ) {1 F (fi, j ) G (gi, j ) )
F (fi, j ) + G (gi, j )} e (6)
where µ( g i,j) and µ(fi,j) are the membership values of the (i,j)th element
3 months
1 day –
The input data includes the actual measured data of all types of
✓
✓
✓
✓
✓
sibility of exploiting the variability that exists in the load pattern data.
For this purpose, starting from the load patterns of an individual cus-
Type of customer
✓
✓
✓
averaging the load pattern data at each time step, as it happens in the
pre-clustering phase in most literature contributions. Rather, the
Industrial
✓
✓
(1) The first stage finds the optimal thresholds using IFD to segment per-
✓
✓
✓
✓
✓
customer load values and then determines the TLP based on the
30 min
✓
✓
✓
(2) The second stage clusters TLPs using the fuzzy c-means (FCM) al-
gorithm, in which DWT-based features are used instead of time
15 min
(3) The third stage segments the time periods for each cluster adap-
Time sample resolution
This stage deals with time series data gathered from smart meters,
and aims at deriving the TLP for each customer. As mentioned in the
Introduction, the main idea of this paper is that the load consumption
Other
in which the columns of Mc correspond to the days (i ∈ {1, 2, .., I}) and
✓
✓
✓
✓
✓
✓
✓
✓
valley ranges of the load pattern). Hence, S-1 thresholds (i.e., τ1 and τ2
This paper
find the optimum set of thresholds. This method converts the load
6
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
(i) The algorithm searches in S-1 loops to find the optimum thresholds
(τc,1, …, τc,S-1) for each customer c, varying from the minimum load
value l to the maximum load value l¯c throughout the planning
_c
horizon.
(ii) For a certain threshold set (τc,1, …, τc,S-1), the load matrix is divided
into S bands (regions). If lc(s) is the load value of customer c located
in the band s, then:
(iii) The mean value of each block s (mc(s) ) can be expressed as follows:
(s )
l¯c
l (s ) × Nc (lc(s) )
l (s ) c
_c
mc(s) = (s )
l¯c
Nc (lc(s) )
l (s )
_c (8)
(s )
For each customer c, the values l (s) and l¯c denote the lower and
_c
upper bound of band s, respectively, while the entry Nc (lc(s) ) defines the
number of occurrences of the load values of customer c in band s.
(iv) The membership value μc,i,j of each element of the load matrix
denotes the closeness of its pixel (load value) to the mean of its
band (mc(s) ). In this paper, a Gaussian like membership function is
adopted:
(lc(,si), j mc(s) ) 2
µc, i, j = exp 2
2 c,s (9)
The variance σ2c,s is considered as in [68]:
(s )
2
c, s = l¯c l (s )
_c (10)
(v) The non-membership value υc,i,j of each element of the load matrix
is evaluated using the Sugeno’s intuitionistic fuzzy generator [69],
given by Valley
Peak Shoulder
1 µc, i, j
c, i, j = ;z>0
1 + zµc, i, j (11)
1 2 3
In this paper, the parameter z is set to 2 (as in [62]).
(vi) Hence, the input load matrix is defined as an IFS. The IFD between
two load matrices A and B of size I × J is defined as in Eq. (6). Let 4 5 6
A be the load matrix thresholded by (τc,1, …, τc,S-1) and B be the
ITLM, then μB(bi,j) = 1 and νB(bi,j) = 0. Hence Eq. (6) is simplified
(the components referring to the matrix B are replaced with either
0 or 1) as:
7 8 9
DivAB Fig. 3. Scheme to determine the 3 × 3 neighborhood-based membership
I J
1 function of element #5, with ε = 1, δ = 1, and S = 3.
= (4 {2 µA (ai,j )} e µA (ai, j) 1 µA (ai,j ) e1 µA (ai, j) {1 A (ai, j )}
2 i= 1 j =1
e A (ai, j ) {1 + A (ai, j )} e A (ai, j ) ) (12) 3.1.2. Derivation of the TLP for each customer
In the previous step, the daily load patterns of each customer have
been clustered using IFD-based thresholding. The S bands have been
(vii) The algorithm searches for every possible set of thresholds (τc,1, used as input to find the representative patterns. A representative pat-
…, τc,S-1), and the IFD is calculated. The best thresholds are those tern should keep as much information as possible. The most commonly
that minimize the IFD. The detailed derivation of the IFD formula used representative patterns are the center of the largest cluster of pre-
and thresholding can be found in [62]. clustered customers and the mean of daily load of customers [1]. The
centroid of the largest cluster cannot represent the complete real be-
As shown in Fig. 2, the algorithm uses two loops to find the op- havior of the customer. Also, in the mean of the daily load approach,
timum set (τc,1, …, τc,S-1) for each customer c, varying the load from the outlier’s data affect the results. Efficient detection of outliers is an im-
minimum value l to the maximum value l¯c . In the inner loop, the portant feature of well-performing clustering schemes. To overcome
_c
membership (µc,i,j) and non-membership (νc,i,j) degrees (including the this problem, in this paper, a classification model based on the posterior
hesitation factor) are calculated to determine the IFD. At the first stage, probability is proposed for construction of representative patterns. The
the load matrices are constructed for each customer, then the optimal representative value should be proportional to the number of load va-
thresholds from IFD for segmenting customer load values are de- lues (i.e., the frequency of events) that occur in each band.
termined. Note that the IFD algorithm uses daily values and neighbor The representative load value for customer c at time sample j is:
information to extract the optimal thresholds. Neighbor information
means relation between the hours of the days (2-dimensional relation: 1
l (s )
s c , j, s i c , i, j
for example, for hour 4 of day 3, it considers hours 2, 3, 5 and 6 of day 4 Nc(,si), j
Lc, j =
(horizontal relation) and also hours 3, 4, and 5 of days 3 and 5 (vertical s c, j , s (14)
relation)). For each customer, the TLP is determined from the in-
formation of all days. As such, it is not useful to extract the thresholds where is the number of elements (pixels) that i and j belonging to
Nc(,si), j
for each day separately. Therefore, the load value at day i and time the band s, and αc,j,s is a coefficient for each customer c at time sample j
sample j is correlated with the values at time samples j − ε to j + ε, and and band s, which is expressed as:
also for the days i − δ to i + δ, where ε and δ are the integer numbers
c , j, s = (band(s )|sample(j)) × (sample(j)|band(s ))
used to determine the time lags for the time samples and for the days,
respectively. To segment the correlated load matrix, the membership (sample(j) band(s )) (sample(j) band(s ))
= ×
function described in Eq. (9) has been modified by a neighborhood- (sample(j )) (band(s ))
based membership function as: I
i=1
Nc(,si), j
I
N (s)
i = 1 c, i, j
= × I J
I N (s)
i=1 j = 1 c, i, j (15)
( mc , i , j mc(s) ) 2
µc, i, j = exp
2 2
c,s (13) In Eq. (15), the term (band(s )|sample(j)) is the conditional prob-
ability of band(s ) given sample(j ) , and (sample(j)|band(s )) is the
where mc,i,j is the mean of the load matrix for customer c at time sample conditional probability of sample(j ) given band(s ) . Actually, the first
j of day i and its effective neighbors. The effective neighbors are those term represents the effect of events’ repetition that occurs in each band,
elements in the neighborhood that belong to the same band as the and the latter term represents the density of band s in each time sample
center element. The proposed strategy incorporates explicit correlation j. Based on Eq. (15), the effect of events’ repetition can be seen in the
between different uncertain loads, such as spatial and temporal corre- representative load values.
lation between the consumption levels, rendering the algorithm more
robust to different kinds of variation, as well as to outliers. 3.2. Stage 2: feature extraction and clustering
In order to provide a simple example, let us refer to Fig. 3 with
S = 3. For the calculations related to the membership function for The second stage comprises two steps:
element #5 in its neighborhood defined for ε = 1 and δ = 1, the values
of the elements 1, 6 and 7 are used, because they are in the same band (i) feature extraction from TLPs (via DWT), and
(shoulder). As such, because of using neighborhood-based membership (ii) clustering of daily pattern of customers.
function, the proposed strategy incorporates explicit correlation be-
tween consumption levels, rendering the algorithm more robust to For preprocessing before clustering, commonly, raw time series
different kinds of variation. domain data are used. To reduce the dimensionality of the feature space
8
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
9
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
1 1
real
mean
TLP
0.5 0.5
0 0
10 20 30 40 10 20 30 40
1 1
0 0
10 20 30 40 10 20 30 40
1 1
0 0
10 20 30 40 10 20 30 40
Time sample
Fig. 8. TLPs obtained from the proposed algorithm (example for six customers).
4.2.2. Stage 2
After feature extraction of TLPs using 6-level DWT, FCM is used to
cluster TLPs. The 48-dimensional original time series vector is trans-
70
formed into an upsampled vector (26 = 64 samples, 6-level DWT), then
each TLP is represented through 6 features, determined as indicated in
Section 3.2. Fig. 9 gives an example of using DWT to decompose an
90 individual customer's TLP. The first row reports the resampled version
10 20 30 40 of the original TLP. The other rows are related to one level of the DWT
Time sample decomposition; the left column shows the approximation coefficients
A1 to A6 (not used to calculate the features), and the right column
Fig. 7. Optimal segmented load matrix of the single customer. shows the detail coefficients D1 to D6.
The analysis carried out on electrical load patterns typically creates
into three ranges of values, to separate the load matrix values into three an initial partitioning in two types of days, namely, weekdays and
bands (valley, shoulder and peak). These three ranges are very weekends, also taking into account that in some DSM policies the
common, and are implemented in several works. selling prices are different in these two types of day [9]. Thereby, a pre-
After the execution of the proposed algorithm, the optimal partitioning into weekdays and weekends has been considered here.
10
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
1
Upsampled TLP
0.5
0
10 20 30 40 50 60
Samples
A1 1 D1
1
0 -1
10 20 30 40 50 60 10 20 30 40 50 60
A2 1
D2
1
0 -1
10 20 30 40 50 60 10 20 30 40 50 60
A3 D3
2 1
0 -1
10 20 30 40 50 60 10 20 30 40 50 60
A4 1
D4
2
0 -1
10 20 30 40 50 60 10 20 30 40 50 60
A5 D5
1
2
0 -1
10 20 30 40 50 60 10 20 30 40 50 60
A6 D6
1
2
0 -1
10 20 30 40 50 60 10 20 30 40 50 60
Samples Samples
Fig. 9. TLP decomposition by 6-level DWT (Aj and Dj correspond to the approximation and detail coefficients, respectively).
A further issue to be addressed before running the clustering algo- (with 3 TLPs) exhibits the shape with peak consumption occurring in
rithm is to determine the number of clusters. Specific indicators have the afternoon hours. Also, profiles of cluster 3 (with 35 TLPs) and
been defined in the literature to deal with the particular nature of the cluster 5 (with 15 TLPs) have the similar shape as cluster 1, but with
data use for fuzzy set-based clustering. These indicators are defined to slightly lower and higher consumption levels, respectively. Clusters 1, 3
work in situations with overlapping clusters. This paper uses the cluster and 5 contain the residents with relatively high consumption, especially
validity index named DWSC, presented in [72] (the proposers did not in the evening hours and at midnight hours, whereas in clusters 1 and 3
provide an explicit interpretation of the acronym). The index DWSC is the consumptions have fewer fluctuations than in other clusters. On the
based on a dynamic weighted sum of two separation and compactness contrary, large variability is shown in cluster 4, with abruptly low TLP
measures. Its compactness and separation properties are considered as values between 6 am and 1 pm. Similar considerations apply to the
major characteristics in the fuzzy c-means clustering technique. The results obtained for the weekends.
compactness measure refers to the variation or scattering of the data
within a cluster. The separation measure quantifies how much the 4.2.3. Stage 3
clusters are isolated among them. By minimizing the DWSC index Time period clustering is an important aspect of DSM, a general
value, the optimal number of clusters can be obtained. The results in- strategy for influencing demand patterns in use. The time steps of a day
dicate that the minimum value of the DWSC index corresponds to 5 that are most similar with respect to the demand profile are grouped in
clusters for weekdays and 4 clusters for weekends, respectively. a cluster. The time periods may be grouped to determine peak,
Fig. 10 illustrates the optimal clusters obtained from stage 2, vi- shoulder, valley, or other types of aggregate periods for different pur-
sualized in the time domain for weekdays and weekends, respectively. poses of DSM policies (e.g., pricing schemes). Two cases of time period
There are clusters that have a very large number of customers, while clustering are applied here. Case 1 considers that the periods are equal
some clusters include outlier consumption behavior. In weekdays, (i.e., 8-h periods) and can be used in time-of-use pricing schemes. Case
cluster 1 is the largest cluster, with 36 TLPs. Cluster 2 has only one 2 considers that the periods are determined based on the highest
member, which corresponds to an industrial load and its consumption is probability of each band based on Eq. (14) (where the periods are not
completely shifted in time with respect to the other load patterns. This necessarily equal). Fig. 11 shows the results of the two cases. In this
cluster exhibits low electrical energy consumption during the daytime specific example, Case 2 leads to a more regular partitioning of the time
and has a sudden increase in nighttime hours. The profile of cluster 4 periods.
11
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
0 0 0 0 0
12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00
hour hour hour hour hour
(a)
0 0 0 0
12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00
hour hou r h ou r hour
(b)
Fig. 10. The clusters as found in Stage 2 for TLPs, (a) weekdays and (b) weekends.
1 1
2
2
Case 1
3
3
4
5 4
00:00 12:00 24:00 00:00 12:00 24:00
1 1
2 2
Case 2
3
3
4
5 4
00:00 12:00 24:00 00:00 12:00 24:00
12
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
0 0 0 0 0 0 0
12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00
hour hour hour hour hour hour hour
(a) weekdays
0 0 0 0 0 0 0 0
12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00 12:00 24:00
hour hour hour hour hour hour hour hour
(b) weekends
Fig. 12. The clusters as found in Stage 2 for TLPs without DWT, (a) weekdays and (b) weekends.
Table 4 1
Comparison of performance indicators (with and without DWT). real
MIA CDI DBI SMI WCBCR
mean
Power consumption (p.u.) TLP (S = 2)
Weekdays With DWT 0.654 0.317 0.859 0.293 0.235
Without DWT 0.829 0.877 2.699 0.376 1.699 TLP (S = 3)
Weekends With DWT 0.856 0.710 1.937 0.241 0.822 TLP (S = 4)
Without DWT 0.916 0.721 2.120 0.775 0.561
0.5
13
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
Table 6 [4] Albert A, Rajagopal R. Smart meter driven segmentation: what your consumption
Performance indicators in the case with S = 4. says about you. IEEE Trans Power Syst 2013;28(4):4019–30.
[5] Jokar P, Arianpoo N, Leung VCM. Electricity theft detection in AMI using customers’
consumption patterns. IEEE Trans Smart Grid 2016;7(1):216–26.
MIA CDI DBI SMI WCBCR
[6] Aggarwal CC, Yu PS. A framework for clustering uncertain data streams. In: IEEE
24th international conference on data engineering; 2008, p. 150–9.
Weekdays FCM 0.770 0.391 1.136 0.146 0.773 [7] Gerbec D, Gasperic S, Smon I, Gubina F. Determining the load profiles of consumers
k-means 0.931 0.851 2.608 0.220 2.450 based on fuzzy logic and probability neural networks. IEE Proc - Gen, Trans Distrib
k-medoids 0.967 0.830 2.290 0.169 1.696 2004;151:395–400.
[8] Mahmoudi-Kohan N, Parsa Moghaddam M, Sheikh-El-Eslami MK. An annual fra-
Weekends FCM 0.523 0.282 0.772 0.144 0.301 mework for clustering-based pricing for an electricity retailer. Electr Power Syst Res
k-means 0.794 0.567 1.201 0.293 0.343 2010;80:1042–8.
k-medoids 0.535 0.317 0.938 0.278 0.178 [9] Charwand M, Gitizadeh M. Optimal TOU tariff design using robust intuitionistic
fuzzy divergence based thresholding. Energy 2018;147:655–62.
[10] Yang P, Tang G, Nehorai A. A game-theoretic approach for optimal time-of-use
case S = 2, the optimal number of clusters are 5 and 4 for weekdays and electricity pricing. IEEE Trans Power Syst 2013;28(2):884–92.
[11] Carpaneto E, Chicco G, Napoli R, Scutariu M. Electricity customer classification
weekends, respectively (same as S = 3). Also, in the case S = 4, the using frequency-domain load pattern data. Int J Electr Power & Energy Syst
optimal number of clusters are 4 and 7, respectively. 2006;28(1):13–20.
[12] Zhong S, Tam KS. Hierarchical classification of load profiles based on their char-
acteristic attributes in frequency domain. IEEE Trans Power Syst
5. Conclusions 2015;30(5):2434–41.
[13] Li Y, Wolfs PJ. A hybrid model for residential loads in a distribution system with
high PV penetration. IEEE Trans Power Syst 2013;28(3):3372–9.
This paper has introduced an innovative framework for clustering [14] Li X, Bowers C, Schnier T. Classification of energy consumption in buildings with
the daily load patterns of a set of low voltage distribution grid custo- outlier detection. IEEE Trans Ind Electron 2010;57(11):3639–44.
[15] Chicco G, Napoli R, Piglione F. Comparison among clustering techniques for elec-
mers in which the load pattern data are considered with their possible
tricity customer classification. IEEE Trans Power Syst 2006;21(2):933–40.
uncertainty and non-determinacy. The framework presented covers the [16] McLoughlin F, Duffy A, Conlon M. A clustering approach to domestic electricity
research gap of adding the modeling of the incomplete/imprecise in- load profile characterisation using smart metering data. Appl Energy
2015;141:190–9.
formation in electrical load pattern clustering applications. [17] Gouveia JP, Seixas J. Unraveling electricity consumption profiles in households
In the first stage, the daily customer's behavior is synthesized with through clusters: combining smart meters and door-to-door surveys. Energy Build
only one TLP for each customer. This TLP is constructed by using in- 2016;116:666–76.
[18] Ozawa A, Furusato R, Yoshida Y. Determining the relationship between a house-
tuitionistic fuzzy divergence-based thresholding. With the proposed hold’s lifestyle and its electricity consumption in Japan by analyzing measured
model, the extracted TLPs represent a detailed view of a customer, in- electric load profiles. Energy Build 2016;119:200–10.
[19] Chicco G, Sumaili Akilimali J. Renyi entropy-based classification of daily electrical
cluding the effect of the uncertainty and non-determinacy implicitly load patterns. IET Gener Transm Distrib 2010;4:736–45.
embedded in the load pattern data. [20] Granell R, Axon CJ, Wallom DCH. Impacts of raw data temporal resolution using
In the second stage, the results show that the proposed clustering selected clustering methods on residential electricity load profiles. IEEE Trans
Power Syst 2015;30(6):3217–24.
method demonstrates to be more accurate over the classical k-means [21] Chicco G, Ionel O-M, Porumb R. Electrical load pattern grouping based on centroid
and k-medoids algorithm. In addition, with the preparation of the model with ant colony clustering. IEEE Trans Power Syst 2013;28(2):1706–15.
features by using DWT, a lower number of features for each TLP are [22] Mets K, Depuydt F, Develder C. Two-stage load pattern clustering using fast wavelet
transformation. IEEE Trans Smart Grid 2016;7(5):2250–9.
sent to the clustering procedure, thus reducing the computation time [23] Kang J, Lee J. Electricity customer clustering following experts’ principle for de-
and improving most of the clustering performance indicators. mand response applications. Energies 2015;8:12242–65.
[24] Benítez I, Quijano A, Díez J, Delgado I. Dynamic clustering segmentation applied to
In the third stage, an adaptive mechanism is developed to group the load profiles of energy consumption from Spanish customers. Int J Electr Power
time periods of the clustered load patterns. These time periods provide Energy Syst 2014;55:437–48.
further information on the periods of the day in which the customers [25] Viegas JL, Vieira SM, Melício R, Mendes VMF, Sousa JMC. Classification of new
electricity customers based on surveys and smart metering data commission for
are expected to belong to different load levels. energy regulation. Energy 2016;107:804–17.
As shown, the proposed algorithm is flexible concerning the selec- [26] Kwac J, Flora J, Rajagopal R. Household energy consumption segmentation using
hourly data. IEEE Trans Smart Grid 2014;5(1):420–30.
tion of the number of ranges determined by the decision maker. With
[27] Ramos S, Duarte JM, Duarte FJ, Vale Z. A data-mining-based methodology to
less thresholds (ranges), the simulation time is reduced, but on the support MV electricity customers’ characterization. Energy Build 2015;91:16–25.
other hand, TLP accuracy is reduced. Conversely, with the increase in [28] Lavin A, Klabjan D. Clustering time-series energy data from smart meters. Energy
Effic 2014;8:681–9.
the number of ranges, the TLP accuracy improvement could become [29] Räsänen T, Voukantsis D, Niska H, Karatzas K, Kolehmainen M. Data-based method
progressively lower. As such, the increase in the number of ranges may for creating electricity use load profiles using large amount of customer-specific
be limited to avoid an excessive computational burden. The choice of a hourly measured electricity use data. Appl Energy 2010;87:3538–45.
[30] López JJ, Aguado JA, Martín F, Mu F, Rodríguez A, Ruiz JE. Hopfield-K-Means
number of ranges not higher than three or four is advisable, also to keep clustering algorithm: a proposal for the segmentation of electricity customers. Electr
consistency with the number of the load levels typically used in time-of- Power Syst Res 2011;81:716–24.
[31] Piao M, Shon H, Lee J, Ryu K. Subspace projection method based clustering analysis
use tariffs and intuitively understandable by the operators. in load profiling. IEEE Trans Power Syst 2014;29:2628–35.
Future work will focus on the application of the proposed model to [32] Al-otaibi R, Jin N, Wilcox T, Flach P. Feature construction and calibration for
study the formulation and implementation of demand response pro- clustering daily load curves from smart-meter data. IEEE Trans Ind Inform
2016;12:645–54.
grams. [33] Peters G, Crespo F, Lingras P, Weber R. Soft clustering – Fuzzy and rough ap-
proaches and their extensions and derivatives. Int J Approximate Reasoning
Declaration of Competing Interest 2013;54:307–22.
[34] Bezdek J. Pattern recognition with fuzzy objective algorithms. New York: Plenum
Press; 1981.
The authors declare that they have no known competing financial [35] Ruspini E. A new approach to clustering. Inf Control 1969;15:22–32.
[36] Atanassov KT. Intuitionistic fuzzy sets. Fuzzy Sets Syst 1986;20:87–96.
interests or personal relationships that could have appeared to influ- [37] Zhang HM, Xu ZS, Chen Q. On clustering approach to intuitionistic fuzzy sets.
ence the work reported in this paper. Control Decision 2007;22:882–8.
[38] Xu Z. Intuitionistic fuzzy aggregation and clustering. Studies in fuzziness and soft
computing 2012;279.
References [39] Danish Lohani QM, Solanki R, Muhuri PK. Novel adaptive clustering algorithms
based on a probabilistic similarity measure over Atanassov intuitionistic fuzzy set.
[1] Chicco G. Overview and performance assessment of the clustering methods for IEEE Trans Fuzzy Syst 2018;26:3715–29.
electrical load pattern grouping. Energy 2012;42(1):68–80. [40] Krishnapuram R, Keller JM. A possibilistic approach to clustering. IEEE Trans Fuzzy
[2] Rhodes JD, Cole WJ, Upshaw CR, Edgar TF, Webber ME. Clustering analysis of Syst 1993;1:98–110.
residential electricity demand profiles. Appl Energy 2014;135:461–71. [41] Davé RN. Clustering relational data containing noise and outliers. Pattern Recogn
[3] Stephen B, Mutanen A, Galloway S, Burt G, Jarventausta P. Enhanced load profiling Lett 1991;12:657–64.
for residential network customers. IEEE Trans Power Delivery 2014;29(1):88–96. [42] Dave RN, Sen S. Noise clustering algorithm revisited. In: 1997 annual meeting of
14
M. Charwand, et al. Electrical Power and Energy Systems 117 (2020) 105624
the North American fuzzy information processing society – NAFIPS; 1997, p. [58] Sun M, Wang Y, Teng F, Ye Y, Strbac G, Kang C. Clustering-based residential
199–204. baseline estimation: a probabilistic perspective. IEEE Trans Smart Grid 2019. in
[43] Lingras P, West C. Interval set clustering of web users with rough k-means. press.
Department of Mathematics and Computer Science, St. Mary’s University, Halifax, [59] Le Ray G, Pinson P. Online adaptive clustering algorithm for load profiling.
Canada, Tech. Rep. 2002-002; 2002. Sustainable Energy Grids Networks 2019;17:100181.
[44] Peters G. Is there any need for rough clustering? Pattern Recogn Lett 2015;53:31–7. [60] Zadeh LA. Fuzzy sets. Inf Control 1965;8:338–53.
[45] Lai JZC, Juan EYT, Lai FJC. Rough clustering using generalized fuzzy clustering [61] Ghosh M, Das D, Chakraborty C, Ray AK. Development of Renyi’s entropy based
algorithm. Pattern Recogn 2013;46:2538–47. fuzzy divergence measure for leukocyte segmentation. J Med Imaging Health Inf
[46] Maji P, Pal SK. Rough set based generalized fuzzy c-means algorithm and quanti- 2011;1:334–40.
tative indices. IEEE Trans Syst, Man, Cybernet – Part B: Cybernet 2007;37:1529–40. [62] Jati A, Singh G, Mukherjee R, Ghosh M, Konar A, Chakraborty C, et al. Automatic
[47] Mitra S, Banka H, Pedrycz W. Rough-fuzzy collaborative clustering. IEEE Trans Syst, leukocyte nucleus segmentation by intuitionistic fuzzy divergence based thresh-
Man, Cybernet – Part B: Cybernet 2006;36:795–805. olding. Micron 2014;58:55–65.
[48] Pedrycz W. Shadowed sets: Representing and processing fuzzy sets. IEEE Trans Syst, [63] Moshavash Z, Danyali H, Helfroush MS. An automatic and robust decision support
Man, Cybernet – Part B: Cybernet 1998;28:103–9. system for accurate acute Leukemia diagnosis from blood microscopic images. J
[49] Mitra S, Pedrycz W, Barman B. Shadowed c-means: Integrating fuzzy and rough Digital Imaging 2018;31:702–17.
clustering. Pattern Recogn 2010;43:1282–91. [64] Danyali H, Helfroush MS, Moshavash Z. Robust leukocyte segmentation in blood
[50] Pedrycz W. Shadowed sets in the characterization of rough-fuzzy clustering. Pattern microscopic images based on intuitionistic fuzzy divergence. In: 22nd biomedical
Recogn 2011;44:1738–49. engineering conference (ICBME); 2015, p. 275–280.
[51] Denoeux T, Masson M-H. “EVCLUS: evidential clustering of proximity data. IEEE [65] Zhao X, Zheng Y, Wan Z. Interactive intuitionistic fuzzy methods for multilevel
Trans Syst, Man, Cybernet – Part B: Cybernet 2004;34:95–109. programming problems. Expert Syst Appl 2017;72:258–68.
[52] Masson MH, Denoeux T. ECM: an evidential version of the fuzzy c-means algorithm. [66] Nguyen H. A novel similarity/dissimilarity measure for intuitionistic fuzzy sets and
Pattern Recogn 2008;41:1384–97. its application in pattern recognition. Expert Syst Appl 2016;45:97–107.
[53] Coke G, Tsao M. Random effects mixture models for clustering electrical load series. [67] Szmidt E, Kacprzyk J. Distances between intuitionistic fuzzy sets. Fuzzy Sets and
J Time Ser Anal 2010;31:451–64. Syst 2000;114(3):505–18.
[54] Li R, Li F, Smith ND. Multi-resolution load profile clustering for smart metering [68] Verma R, Sharma BD. On generalized intuitionistic fuzzy divergence (relative in-
Data. IEEE Trans Power Syst 2016;31:4473–82. formation) and their properties. J Uncertain Syst 2012;6(4):308–20.
[55] Haben S, Singleton C, Grindrod P. Analysis and clustering of residential customers [69] Grabisch M, Murofushi T, Sugeno M, Kacprzyk J. Fuzzy measures and integrals,
energy behavioral demand using smart meter data. IEEE Trans Smart Grid theory and applications. Berlin: Physica Verlag; 2000.
2015;7:136–44. [70] Mallat SG. A theory for multi-resolution signal decomposition: the wavelet re-
[56] Ferraro P, Crisostomi E, Tucci M, Raugi M. Comparison and clustering analysis of presentation. IEEE Trans Patt Anal Mach Intell 1989;11:674–93.
the daily electrical load in eight European countries. Electr Power Syst Res [71] Deng Z, Jiang Y, Chung F, Ishibuchi H, Choi K, Wang S. Transfer prototype-based
2016;141:114–23. fuzzy clustering. IEEE Trans Fuzzy Syst 2016;24:1210–32.
[57] Weng Y, Yu J, Rajagopal R. Probabilistic baseline estimation based on load patterns [72] Zhang F, Qian X. A new validity index for fuzzy clustering. J Comput Inf Syst
for better residential customer rewards. Int J Electr Power Energy Syst 2012;8:5875–83.
2018;100:508–16.
15