IEEE Power & Energy Magazine - May/June 2018 - 57

pli

tio

consumption behaviors so that they can provide diversified similarity-oriented and classification-oriented indices that
and personalized services and encourage customers to par- are used to determine the optimal clustering method and the
corresponding optimized parameters. the similarity-oriented
ticipate in demand response and energy efficiency programs.
Prior to the current decade, utilities collected most demand indices include the Davies bouldin index (Dbi), mean index
information from supervisory control and data acquisition adequacy (Mia), clustering dispersion indicator (cDi), and
systems, with the demand data being used mostly for system- others. the classification-oriented indices include purity, recall,
or bus-level load forecasting. During the past few years, nearly and F-precision, among others.
finally, cluster analysis of electricity consumption can be
one billion smart meters have been installed. the popularization of smart meters has resulted in increasingly detailed elec- applied for various purposes, such as load pattern recognition,
tricity consumption data that enable extensive applications of evaluation of demand response potential, energy efficiency
management, rate-making, load forecasting, and consumer
big data analytics on the demand side.
because it is impossible to analyze all these meters one by sociodemographic information identification. for example,
one, it is necessary to group similar ones before developing fur- consumers with no fixed-time occupation, no access to a netther models. the resulting groups are called clusters; such work, and no computers (e.g., farmers) might exhibit peak
tasks are also known as clustering or cluster analysis in data consumption at noon; consumers with computers (e.g., whitemining. figure 3 summarizes and explains load profiling from collar workers) might exhibit peak electricity consumption at
three patches: 1) the input load profiles of clustering, 2) spe- night; and consumers with access to a network but no fixedcific clustering techniques and evaluation indices, and 3) po- time occupation (e.g., online youth) might exhibit peak electricity consumption before dawn. thus, the correspondence
tential applications.
first, the input load profiles vary in terms of time and loca- between load patterns obtained by clustering and consumer
tion. the time scale of the clustered data is driven by the target sociodemographic information can be detected.
state grid Jiangsu electric Power company is pioneering
of the analysis, which can vary from daily, weekly, and monthly
to yearly time frames. for example, we can design a daily tariff demand response applications in china, especially for large
based on the clusters of daily load profiles. if we need a better consumers. there are two clustering-based software modunderstanding of the dynamics of consumers' electricity con- ules integrated into a demand response scheduling system in
sumption behavior, load profiles
over a longer period are required.
in terms of spatial scales, the load
profiles range from a single consumer, substation, feeder, or bus to
the entire system, according to the
Time Scale
specific objectives.
In
pu
* Daily
second, we can categorize the
t
clustering methods into 1) direct
* Weekly
clustering and 2) indirect clustering, according to whether features
* Monthly
Spatial Scale
have been extracted from original
Energy Efficiency
load profiles. in particular, direct
Demand Response
Areas
* Total Load
* Yearly
clustering methods include k-means,
* Industries
* Areas
Tariff Design
fuzzy clustering, hierarchical clus* Single
Load Profiling
tering, and density-based clustering
Time-Series Analysis
technologies such as density-based
* Indirect Feature Exaction
spatial clustering of applications
Dimension Reduction
with noise (Dbscan). indirect
* Similarity Oriented
Fuzzy Clustering
clustering includes clustering methDBI
* Direct K-Means
MIA
ods with data preprocessing proceCDI
DBSCAN
dures achieved by using time-series
* Classification Oriented
analysis techniques (e.g., the hidden
Purity
Recall
Markov model), feature extracIndices
Methods
F-Precision
tion methods (e.g., sparse coding),
dimensionality reduction methods
(e.g., principal component analyClustering
sis), and so forth. current evaluation indices for clustering performance can be categorized into figure 3. A summary of load profiling.
*

may/june 2018

ieee power & energy magazine

Table of Contents for the Digital Edition of IEEE Power & Energy Magazine - May/June 2018