Tech Science Press

DOI: 10.32604/ee.2024.041441

ARTICLE

Analysis of Electricity Consumption Pattern Clustering and Electricity

Consumption Behavior

Liang Zhu1 , Junyang Liu1 , Chen Hu1 , Yanli Zhi2 and Yupeng Liu3, *
1
Power Supply Service Management Center, State Grid Jiangxi Electric Power Corporation Limited, Nanchang, 330077, China
2
State Grid Jiangxi Electric Power Corporation Limited, Nanchang, 330077, China
3
College of Electrical and Information Engineering, Hunan University, Changsha, 410082, China
* Corresponding Author: Yupeng Liu. Email: [email protected]
Received: 23 April 2023 Accepted: 08 September 2023 Published: 19 August 2024

ABSTRACT
Studying user electricity consumption behavior is crucial for understanding their power usage patterns. However,
the traditional clustering methods fail to identify emerging types of electricity consumption behavior. To address
this issue, this paper introduces a statistical analysis of clusters and evaluates the set of indicators for power usage
patterns. The fuzzy C-means clustering algorithm is then used to analyze 6 months of electricity consumption data
in 2017 from energy storage equipment, agricultural drainage irrigation, port shore power, and electric vehicles.
Finally, the proposed method is validated through experiments, where the Davies-Bouldin index and profile
coefficient are calculated and compared. Experiments showed that the optimal number of clusters is 4. This study
demonstrates the potential of using a fuzzy C-means clustering algorithm in identifying emerging types of electricity
consumption behavior, which can help power system operators and policymakers to make informed decisions and
improve energy efficiency.

KEYWORDS
Electricity consumption; clustering; consumption behavior; fuzzy C-means

1 Introduction
The smart grid represents an innovative power technology that integrates information technology
with energy systems, offering the potential for substantial transformation in traditional electricity
services. By leveraging advanced business systems, comprehensive customer data, and intelligent
service platforms, smart grids enable the deployment of intelligent power services, ultimately enhancing
energy efficiency on a societal scale. This advancement holds the promise of creating a more sustainable
and efficient energy infrastructure that can effectively meet the evolving demands of modern society
[1–3].
During the intelligent construction and development of traditional power grids, the wide adoption
of smart meters has enabled researchers to access large-scale and diverse original electricity consump-
tion data for analyzing user electricity consumption behavior [4,5]. These data present characteristics
of high dimensionality and complex data processing, and extracting valuable information from them is
a challenge faced by the power industry. Thus, analyzing the electricity consumption behavior patterns
and demand characteristics of power users has significant research value as it can enhance the service

Copyright © 2024 The Authors. Published by Tech Science Press.

This work is licensed under a Creative Commons Attribution 4.0 International License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
2640 EE, 2024, vol.121, no.9

level of electricity sales enterprises and distribution network dispatchers, and drive the transformation
of power consumption mode towards intelligence.
Over the past decades, research on the energy consumption behavior of consumers become an
important way to improve energy efficiency and promote energy conservation. The traditional analysis
of electricity consumption behavior mainly relied on questionnaires and statistics to analyze residents’
electricity consumption preferences and utilized highly subjective fuzzy comprehensive evaluation
methods to assess residents’ attitudes towards electricity consumption, which failed to take into
account the characteristics of residents’ electricity load. Alternatively, reference [6] calculated the
difference between daily electricity consumption curves using three similarity measures: Euclidean
distance, Pearson distance, and Chebyshev distance. Reference [7] analyzed the impact of the time
granularity of original electricity consumption data on the clustering of power users’ electricity
consumption behavior. Meanwhile, reference [8] studied a Spark-based K-means clustering algorithm
parallel method for large-scale power load curves, and reference [9] proposed a multi-resolution
clustering algorithm for large amounts of electricity consumption data collected by smart meters. Ref-
erence [10] introduced a load classification method for power users combining the K-means algorithm
and Fuzzy support vector machine (FSVM) classification algorithm. Reference [11] utilized the K-
means++ clustering algorithm in conjunction with Self-Organizing Map (SOM) neural networks to
determine user power usage patterns, while Reference [12] analyzed and evaluates several clustering
algorithms, such as the traditional K-means clustering algorithm, self-organizing neural network,
weighted average fuzzy K-means clustering algorithm, and hierarchical algorithm. Additionally, refer-
ence [13] optimized the number of selected clusters through multi-criteria decision analysis, reference
[14] adopted the parametric bootstrapping algorithm to determine the number of clusters, reference
[15] fited temperature control load curves through load-temperature correlations, and reference [16]
extended non-invasive load monitoring technology to the regional level. Load pattern identification
is often studied for specific loads in different regions, different industries, and different seasons [17].
Reference [18] proposed a load identification research method based on typical load curves of different
industries, which effectively identifies load curves with common morphological characteristics. Cluster
analysis was performed on 936 residential electricity load datasets obtained from the U.S. Department
of Energy [19]. Reference [20] presented a case study to examine how the lockdowns during the
COVID-19 pandemic affected the amount of electricity and natural gas consumption in four organized
industrial zones in Turkey. Reference [21] provided the time-series forecasting of energy demand
in electric vehicles and the impact of the COVID-19 pandemic on energy demand. Reference [22]
employed the CH-K-means algorithm and Long Short-Term Memory (LSTM) neural network to
conduct industrial energy consumption behavior and forecasting analysis. Reference [23] assessed the
effects of social media in molding the consumption behaviors while considering eco-labeling, eco-
branding, social norms, and purchase intensions. Reference [24] proposed an unsupervised progressive
incremental data mining mechanism applied to smart meters energy consumption data through
frequent pattern mining to overcome these challenges. Reference [25] analyzed energy usage behavior in
households by using a mix of primary and secondary techniques, which revealed seasonal consumption
trends, with peculiar ownership and usage patterns of appliances. Reference [26] conducted an analysis
to understand differences between consumer categories and what this could mean e.g., to develop
effective demand response measures. Reference [27] took the Energy Saving and Emission Reduction
Plan (ESER) since 2006 as a policy shock to explore the impact of the ESER on industrial firms’
energy consumption behavior and explores the impacts of the ESER on firms’ energy consumption
behavior based on industrial firm data. Reference [28] described a methodology of integrating the
energy consumption estimation associated with equipment operation into a digital twin model based
EE, 2024, vol.121, no.9 2641

on discrete-event simulation through a data-driven pipeline and discussed the implementation of the
methodology in a laboratory environment including industrial equipment. Reference [29] evaluated
how prosumers adapt their consumption behavior in response to such political uncertainty and
increasing electricity prices. Reference [30] explored behavioral determinants of the demand for water
and electricity, contributing a set of categories and items that may be useful for understanding
knowledge, attitudes, and practices related to consumption and constructing valid scales.
It can be seen from the above research that load pattern recognition is often classified and
trained by specific loads such as regions, industries, seasons, etc., and its classifier model has low
generalization ability and weak model migration ability, which is only applicable to specific user load
models. Whenever a new load object is studied, it cannot be directly transferred to the new user load
without re-doing tedious long-term training sample collection. At present, the problem that is difficult
to solve in load clustering analysis is that the initial center of clustering is randomly selected, resulting
in discrete points or boundary points in the data sample that easily affect the determination of the
final cluster center, making it difficult to classify the sample points near the decision boundary.
To address the aforementioned issues, this study proposes a solution using statistical analysis of
clusters to analyze power usage patterns based on a set of indicators. Specifically, the fuzzy C-means
clustering algorithm is applied to investigate the electricity consumption modes of energy storage
equipment, agricultural drainage irrigation, port shore power, and electric vehicles in the low-voltage
power supply station area. The proposed method can provide support for line loss control and energy
optimization in low-voltage supply areas. Experimental results demonstrate the effectiveness of our
approach.
The rest of this paper is organized as follows. Section 2 introduces the basics of statistical analysis
of clusters. Section 3 describes the FCM algorithm and indicators for power usage patterns. Section 4
presents the load characteristics of low-voltage power supply station areas, including energy storage
equipment, agricultural drainage irrigation, electric vehicles, and port ship shore power. Section 5
provides the experimental results. Section 6 concludes this paper.

2 Basics of Statistical Analysis of Clusters

Cluster analysis is a modern multivariate statistical technique that combines contemporary
taxonomy and multivariate statistical analysis, and is considered a crucial branch of unsupervised
pattern recognition. Its purpose is to categorize a disordered set of samples into several subclasses
with distinct characteristic differences by using specific rules. This process aims to group samples with
similar feature quantities and classify them as accurately as possible, while also separating samples with
dissimilar feature quantities into different categories. There are several types of clustering methods,
including division-based clustering, hierarchical-based clustering, density-based clustering, grid-based
clustering, and model-based clustering.
Steps for clustering can be listed as follows:
Step 1: Select the clustering variable. These variables should have the following characteristics:
1) They are related to the goal of cluster analysis; 2) Reflect the characteristics of the object to be
classified; 3) There are obvious differences in the values of different objects; 4) Variables should not
be highly correlated with each other. For highly correlated variables (two): 1) Before the case cluster
analysis, the variables are clustered and selected as clustering variables in each category; 2) Do factor
analysis on variables to produce a set of uncorrelated variables as clustering variables.
2642 EE, 2024, vol.121, no.9

Step 2: Calculate the similarity. The similarity is a fundamental concept in cluster analysis and
reflects the degree of affinity between study subjects. Cluster analysis is based on the similarity between
the study objects.
Step 3: Clustering. Select a clustering method to determine the number of classes formed.
Step 4: Interpretation of clustering results. After obtaining the clustering results, verify and
interpret the results to ensure that the clustering solution is trustworthy.

3 Electricity Consumption Pattern Recognition Method Based on Cluster Analysis Theory

3.1 Set of Indicators for Power Usage Patterns
The user load characteristics and power consumption mode can be broadly classified into three
types: user attribute class, load curve class, and electrical index class. The user attribute class comprises
the electricity consumption attributes that are determined by the user at the time of using electricity.
These attributes are typically stable or undergo minimal changes and are assigned during clustering.
The load curve class comprises the power and electricity consumption change curve of the user over
a specific period of time. This class provides information on the user’s peak and valley characteristics,
load change trends, and fluctuations, making it a rich source of power consumption data. The
electrical index class describes the user’s electricity consumption behavior in numeric form, with load
fluctuation, growth, peak and valley, and other aspects being quantified over different time frames
such as day, week, month, season, and year. Table 1 shows the different types of indicators and their
respective meanings.

Table 1: Description of user electricity behavior labels

No. Type Indicator Description
1 Electricity consumption Reflects the industry in which the
type user is located and the price of
electricity
2 Measurement method Reflects the measurement method
of the user’s electric energy meter
3 User parameters Contract capacity Reflects the user’s production scale
or electricity consumption capacity
4 Power factor Reflects the user’s electricity
efficiency
5 Load Load curve Reflects the user’s load change trend
6 Electricity growth rate Reflects the growth degree and
development trend of users’
electricity consumption
7 Current unbalance rate Reflects the degree of asymmetry of
three-phase load
8 Load volatility Reflects the degree of stability of the
load
9 Electrical indicators Peak-to-valley difference Reflects the user’s sensitivity to
rate peak-valley electricity prices
(Continued)
EE, 2024, vol.121, no.9 2643

Table 1 (continued)
No. Type Indicator Description
10 Peak-to-valley load ratio Reflects the degree of uneven user
load
11 Load ratio Reflects user capacity utilization
efficiency
12 Quarter imbalance Reflects the degree of imbalance in
coefficient the user’s seasonal electricity
consumption

1) User attribute class. Type of electricity consumption (agricultural production, agricultural

drainage and irrigation, large industry, general industry, and commerce, residential lighting, etc.),
measurement method (single-phase, three-phase three-wire, three-phase four-wire), contract capacity,
power factor.
2) Load curve class. Daily load curve, monthly load curve, annual load curve.
3) Electrical indicators. Electricity growth rate, current imbalance rate, load fluctuation rate, peak-
to-valley difference rate, seasonal imbalance coefficient, load ratio, peak-to-valley load ratio.
Some electrical indicators have time scale malleability, so that different index types are formed
to comprehensively describe the user’s power consumption mode, and the content of the extended
indicators is shown in Table 1.

3.2 FCM Clustering Algorithm

The Fuzzy C-means (FCM) clustering algorithm is designed to assign a degree of membership to
each data point in a cluster, based on their similarity, thereby grouping the most similar objects into
the same cluster and separating the least similar objects into different clusters. Unlike the traditional
C-means algorithm that is rigid in data partitioning, FCM provides a flexible approach for fuzzy
data partitioning. Fuzzy clustering methods are particularly useful in cases where clustering involves
ambiguous boundaries between objects. The FCM algorithm is commonly employed in cluster analysis
to identify various electricity consumption patterns within an industry, where load characteristics are
classified for each industry. The flowchart of the FCM algorithm is shown in Fig. 1.
The FCM clustering process involves several steps:
1) Initialization. Let the number of clustering categories be c (2 ≤ c ≤ n, n is the number
of samples), set the iteration stop threshold ε and the iteration counter b = 0, and initialize the
membership matrix U (0) ;
2) Update the cluster center matrix P(b) by the following equation:
n  m
Uik(b) xk
= 
k=1
Pi(b) n  m (1)
Uik(b)
k=1
2644 EE, 2024, vol.121, no.9

where m is the hyper-parameters that controls how fuzzy the cluster will be. xk is the samples. b denotes
the iteration counter. U (b) is the membership matrix.

Figure 1: The flowchart of the FCM algorithm

3) Update the membership matrix U (b + 1) by following equation:

⎧ ⎡ ⎤⎫−1
⎨ c (b+1)
dik
2
m−1 ⎬
Uik =
(b+1) ⎣ ⎦ (2)
⎩ h=1 dhk(b+1) ⎭

where d ik denotes the distance between the sample k and the cluster center i, dik = ||xi − xk ||2 .
4) Determine whether ||U (b) − U (b+1) || < ε is satisfied, and if so, the algorithm stops and outputs
matrices U and P. Otherwise, let b = b + 1 and go to step 2 to continue, where ‘|| ||’ is one type of the
suitable matrix norms.
After the algorithm iteration stops, the final membership degree and cluster center form are:
1
Uik = 2
(3)

c dik m−1

h=1 dhk

n
(Uik )m xk
pi = 
k=1
n
(4)
m
(Uik )
k=1

The characteristics of user electricity mainly cover the following six types and grades:
1) Seasonal characteristics: insensitive, spring sensitive, summer sensitive, autumn sensitive, winter
sensitive.
EE, 2024, vol.121, no.9 2645

2) Temperature characteristics: insensitive, high-temperature peak, low-temperature peak.

3) Weekly rest characteristics: irregular, five-day working system, six-day working system, seven-
day working system.
4) Peak-valley characteristics: peak preference, valley preference, peak-valley equilibrium.
5) Fluctuation characteristics: low, medium, high, high.
6) Benefit characteristics: low, medium, high, high.

4 Load Characteristics of Low-Voltage Power Supply Station Area

In this section, we analyze the electricity consumption of four scenarios, including energy storage
equipment, agricultural drainage irrigation, electric vehicles, and port ship shore power. Data comes
from State Grid Jiangxi Electric Power Co., Ltd. (Nanchang, China), and 6 months of data in 2017
are selected.

4.1 Energy Storage Equipment

To analyze how the peak-to-valley electricity price ratio affects the response characteristics of
energy storage equipment users, various peak-to-valley electricity price ratios are separately set, and
the demand response electricity consumption characteristics of users are calculated. Subsequently,
the impact of the peak-to-valley electricity price ratio on the demand response characteristics of
energy storage users is analyzed. As shown in Table 2, the initial peak-to-valley electricity price ratio
is 4.45, and a higher peak-to-valley electricity price ratio of 9.90 is introduced. For each energy
storage equipment user, an electricity consumption scheme is developed based on this electricity
price ratio, which is then compared to the original peak-to-valley electricity price. Users A and B,
who have different economic sensitivities, were chosen for a comparative analysis of electrical load
characteristics. User A utilizes a medium-sized energy storage device, while User B utilizes a large
energy storage device.

Table 2: Electricity consumption schemes for various types of users at different peak-to-valley
electricity price ratios
Type Original scheme New scheme
(Price/, °C) (Price/, °C)
User A with small-sized energy storage device and sensitive (15.5, 1.83) (11.13, 1.96)
to price
User B with small-sized energy storage device and very (11.46, 2.6) (7.85, 2.5)
sensitive to price
User C with medium-sized energy storage device and (19.96, 1.9) (14.02, 1.86)
sensitive to price
User D with large-sized energy storage device and not (24.21, 1.4) (17.1, 1.56)
sensitive to price
User E with non energy storage device and sensitive to price (12.05, 2.5) (7.93, 2.51)

Based on Fig. 2, it can be observed that for users C and D, their electricity consumption slightly
increases from 20:00 to 08:00 the next day with an increase in electricity price, while during the peak
2646 EE, 2024, vol.121, no.9

period of electricity prices from 08:00 to 20:00, the electricity consumption of user C remains stable,
whereas the electricity consumption of user D decreases significantly. Thus, for user C, who is more
sensitive to the economy, under the influence of the original peak-to-valley electricity price ratio, the
response to peak electricity prices is more pronounced, and the load characteristics do not change with
a further increase in the electricity price ratio. However, for user D with weaker economic sensitivity,
the increase in peak-to-valley electricity price ratio results in significant adjustments in electricity load
during peak periods, leading to an increase in the degree of demand response, until the electricity
price reaches a certain level, after which the user’s response potential becomes maximum, and the load
characteristics remain constant.

Figure 2: Daily load characteristics of users under different peak-valley electricity prices

4.2 Agricultural Drainage Irrigation

The seasonal imbalance coefficient and annual load rate are used as the electricity consumption
labels of the irrigation load of agricultural drainage. The quarterly imbalance coefficient is:

12
Pi
Kc =
i=1
(5)
12Pyear,max
where Pyear,max is the maximum load of the year, and Pi is the maximum load of the ith month.
The annual load factor is:
12
Wi
λyear =
i=1
(6)
Pmax × 8760
where W i is the electricity consumption in the ith month (kWh) and Pmax is the maximum annual load
(kW).
The typical daily load characteristics of agricultural drainage irrigation are shown in Fig. 3.
The figure illustrates that the daily fluctuation of electricity consumption for agricultural irriga-
tion is minimal. This is because the electricity consumption cycle for agricultural irrigation is long and
continuous and not restricted to day or night. The irrigation system is typically kept running until the
end of the irrigation season, resulting in a low daily load fluctuation rate.
EE, 2024, vol.121, no.9 2647

8
7
6

Power/MW
June
5
July
4 Aug
3
2
1

00

00

00

00

00

0
:0

:0

:0

:0

:0

:0

:0
0:

2:

4:

6:

8:

10

12

14

16

18

20

22
Time

Figure 3: Daily load characteristics of typical agricultural irrigation substations

4.3 Electric Vehicle

We examine real-world historical data pertaining to electric vehicles, which includes records of
each EV charging session. The crucial information captured in these records consists of the starting
and ending times of the charging, as well as the state of charge (SOC) at the beginning and end of the
charging session.
For the typical load of electric vehicles on weekdays and weekends, we consider two scenarios.
(1) Electric vehicle charging load scenario
Three charging scenarios were considered to generate electric vehicle charging load scenarios, and
three of them are presented in Fig. 4. The first scenario (in blue) and the second scenario (in green)
both depict a single charging event that occurs at night, with the only difference being that the charging
time in the first scenario spans 24 h. The third scenario (in red) depicts two charging events, one in the
morning and one in the evening.

Figure 4: Electric vehicle charging load scenario

(2) Electric vehicle load under the modeling of different charging cycles
500 electric vehicle grid load power scenarios are generated by considering the key variables of
tertiary, secondary, and primary charging. The weighted average electric vehicle load curve is obtained
and shown in Fig. 5. To provide a benchmark for comparison, the load curve of electric vehicles is
obtained based on historical data, considering all charge and discharge times.
The load model shows the closest approximation to the load curve based on historical data when
considering the key variable of tertiary charging. However, when considering only the key variables of
2648 EE, 2024, vol.121, no.9

two charging sessions and one charging session, the load curve tends to be underestimated, particularly
during the evening peak hours.

Figure 5: EV load curve under modeling of different charging cycles (working days)

4.4 Port Ship Shore Power

When the ship is connected, the load experiences a sudden and significant increase, rather than a
gradual one, which has a notable impact on the power system. This is a rapidly changing process that
occurs over approximately 7 h of docking time, during which the load is maintained at a high operating
state and exhibits minimal fluctuations.
Active and reactive load curve of a single ship connected to shore power is shown in Fig. 6.

Figure 6: Daily load curve for a single vessel

The ships usually access shore power for 7 to 9 h between 07:00 and 21:00, in a 24-h cycle for a total
of 7 days. When a single berth is connected, the shore power load starts from zero and fluctuates slightly
during the access period due to the use of various power loads on the ship. The change characteristics
of active and reactive power are similar.
When two berths are connected to shore power, there are differences in port call and residence
times of the vessels, as well as differences in load operation characteristics, leading to fluctuations in
power load when both ships are connected to shore power simultaneously. The change characteristics
of active and reactive power are similar. Although the daily electricity consumption period is extended,
the overall electricity consumption period remains between 07:00 and 21:00.
EE, 2024, vol.121, no.9 2649

After conducting the above analysis, the following characteristics of port shore power load have
been identified:
(1) There is a significant disparity in load between day and night. As port ship operations primarily
occur during the daytime, the load during this period is substantially higher than that during
nighttime. The day-night load ratio can be used as an indicator to describe the shore power
load characteristics of the port.
(2) The load fluctuation during peak periods is minimal. Once the ship is docked, the overall
load variation is not abrupt. During peak periods of port shore power, a diverse range of
constant power equipment occupies a substantial portion of the electrical load. However, the
load variation is very sharp in the case of certain loads like pole-changing motors, which are
utilized as drive motors for cargo machinery.
(3) The power factor is low. Port ships involve a large number of inductive loads, which result in a
low power factor, typically ranging between 0.75 to 0.9, with the total power factor being close
to 0.8.
By analyzing the relationship between peak-to-valley electricity price ratio and the response
characteristics of energy storage equipment users, we can conclude that users with different economic
sensitivities exhibit varying responses to increasing peak-to-valley electricity price ratios, leading to
adjustments in electricity load during peak periods. The analysis shows that users more sensitive
to the economy maintain stable load characteristics during peak price periods, while those with
weaker economic sensitivity demonstrate a higher degree of demand response. In terms of agricultural
irrigation, electricity consumption exhibits minimal daily fluctuation. This is attributed to the long and
continuous electricity consumption cycle associated with irrigation, which is not limited to specific day
or night periods. The irrigation system typically operates continuously until the end of the irrigation
season, leading to a low rate of daily load fluctuation. Our findings related to EV charging load
scenario suggest that including tertiary charging as a key variable provides the closest approximation
to the historical load curve. Finally, the load patterns at the port exhibit a significant disparity
between day and night, minimal load fluctuation during peak periods, and a low power factor due
to the presence of inductive loads. Understanding these characteristics is crucial for effective load
management and optimizing the use of shore power resources at the port.

5 User Electricity Clustering Experiments

We use the 936 residential electricity load datasets from the U.S. Department of Energy for
cluster analysis [19]. The dataset spans one year, with data collected every hour, resulting in 8760 data
points per user. Note that the data do not include the time of COVID. The COVID-19 lockdowns
have adversely affected energy demand and consumption. The changes in energy requirements are
compared and analyzed from multiple perspectives according to available data and information.
Interested readers are referred to [20,21] for better the better understanding of the impacts of COVID-
19 pandemic.
The data exhibit clear peak and off-peak periods, with high consumption from January to mid-
March and June to September, and low consumption from mid-March to June and September to mid-
November. Based on these periods, four initial cluster medoids were determined. The experiments
involved comparing the FCM clustering method and the distance cost function.
2650 EE, 2024, vol.121, no.9

The Davies-Bouldin index (DBI) is used as the cluster effectiveness evaluation index, and the
calculation formula is as follows:
1 
K

IDBI = Rk (7)
K k
 
d (xk ) + d xj
Rk = max   (8)
k=j d ck , cj
where d(xk ) and d(xj ) represent the average distance from the same category of data to their clustering
centers, respectively; d(ck , ck ) represents the vector distance between different categories, so it can be
seen that the smaller the DBI, the higher the clustering accuracy.
Different K values are set on the same scale dataset, and the results are shown in Fig. 7. The
experimental results show that the FCM clustering algorithm shows good clustering performance
under different K values.
Distance cost function FCM
5

3
DBI

0
2 4 6 8 10
K value

Figure 7: Cluster accuracy at different K values

In order to determine if the 4 initial cluster clusters obtained by FCM are the optimal number of
clusters, K values ranging from 2 to 9 are tested. Contour coefficients are calculated for each K value,
and Fig. 8 shows the corresponding profile coefficient map. The profile coefficient ranges from −1
to 1, with values closer to 1 indicating better clustering results. From the results of Fig. 8, the largest
profile coefficient value of 0.4978 is obtained when the number of clusters is 4, followed by a value of
4.852 when the number of clusters is 5. Combining the contour coefficient method and DBI value, it
is determined that the optimal number of clusters is 4.
In conclusion, this study successfully applies the FCM clustering algorithm to analyze user
electricity consumption behavior. The statistical analysis of clusters and the set of indicators for
power usage patterns provide a comprehensive understanding of power consumption modes in various
scenarios. The findings reveal specific load characteristics in energy storage equipment, agricultural
drainage irrigation, port shore power, and electric vehicles in the low-voltage power supply station
area. The FCM algorithm proves effective in clustering consumption patterns and analyzing power
consumption behavior, offering valuable insights for load management, energy optimization, and the
development of intelligent power services. The study’s findings have implications for enhancing energy
efficiency and supporting informed decision-making by power system operators and policymakers.
EE, 2024, vol.121, no.9 2651

FCM Distance cost function

0.6

Profile coefficient
0.4

0.2

0.0
1 2 3 4 5 6 7 8 9 10
K value

Figure 8: Clustering K value and profile coefficient plot

6 Conclusion
This paper explores the power consumption behavior of users by using a clustering algorithm to
analyze their electricity consumption data. The analysis includes statistical analysis of clusters and the
introduction of a set of indicators for power usage patterns. The fuzzy C-means clustering algorithm
is applied to analyze the electricity consumption patterns of various scenarios, such as energy storage
equipment, agricultural drainage irrigation, port shore power, and electric vehicles in the low-voltage
power supply station area. We selected 6 months of electricity consumption data in 2017 from Jiangxi
Province, China, and 936 residential electricity load datasets obtained from the U.S. Department of
Energy for analysis. For energy storage equipment, the impact of the peak-to-valley electricity price
ratio from 4.45 to 9.90 is analyzed. We used the Davies-Bouldin index and profile coefficient to select
the K value. The largest profile coefficient value of 0.4978 is obtained when the number of clusters is
4, followed by a value of 4.852 when the number of clusters is 5. Combining these two results, we get
the number of clusters, K = 4. The experimental results show that the proposed FCM is suitable for
clustering consumption patterns and analyzing consumption behavior.

Acknowledgement: None.

Funding Statement: The work is supported by the Science and Technology Project of State Grid
Jiangxi Electric Power Corporation Limited ‘Research on Key Technologies for Non-Intrusive Load
Identification for Typical Power Industry Users in Jiangxi Province’ (521852220004), where the
principle investigator is Liang Zhu.
Author Contributions: The authors confirm contribution to the paper as follows: study conception
and design: Liang Zhu, Junyang Liu, Yupeng Liu; data collection: Chen Hu, Yanli Zhi; analysis
and interpretation of results: Liang Zhu, Junyang Liu, Chen Hu, Yupeng Liu; draft manuscript
preparation: Liang Zhu, Yanli Zhi. All authors reviewed the results and approved the final version
of the manuscript.
Availability of Data and Materials: The 936 residential electricity load datasets used for cluster analysis
in this paper is from the U.S. Department of Energy from reference [19].
Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the
present study.
2652 EE, 2024, vol.121, no.9

References
1. M. Liserre, T. Sauter, and J. Y. Hung, “Future energy systems: Integrating renewable energy sources into the
smart power grid through industrial electronics,” Ind. Electron. Mag. IEEE, vol. 4, no. 1, pp. 18–37, Mar.
2010. doi: 10.1109/MIE.2010.935861.
2. X. Lin, R. Yu, J. Yu, and H. Wen, “Constant coupling effect-based PLL for synchronization stability
enhancement of grid-connected converter under weak grids,” IEEE Trans. Ind. Electron., vol. 70, no. 11,
pp. 11310–11323, Nov. 2023. doi: 10.1109/TIE.2022.3227268.
3. J. Song, J. Zhang, H. Kuang, and H. Wen, “Dynamic synchrophasor estimation based on weighted
real-valued sinc interpolation method,” IEEE Sens. J., vol. 23, no. 1, pp. 588–598, Jan. 2023. doi:
10.1109/JSEN.2022.3224365.
4. A. Mohsenianrad and A. Leongarcia, “Coordination of cloud computing and smart power grids,” in
Proc. 2010 First IEEE Int. Conf. Smart Grid Commun., Gaithersburg, MD, USA, 2010, pp. 368–372. doi:
10.1109/SMARTGRID.2010.5622069.
5. Y. W. Nie, W. Yang, L. S. Huang, and X. R. Lu, “A privacy preserving scheme for change increment of
electricity big data base on symmetric cryptography,” in Proc. 2014 IEEE 3rd Int. Conf. Cloud Comput.
Intell. Syst., Shenzhen, China, Nov. 2014, pp. 487–490. doi: 10.1109/CCIS.2014.7175784.
6. M. Zatti et al., “k-MILP: A novel clustering approach to select typical and extreme days for multi-energy
systems design optimization,” Energy, vol. 181, pp. 1051–1063, Aug. 2019. doi: 10.1016/j.energy.2019.05.044.
7. R. Granell, C. J. Axon, and D. Wallom, “Impacts of raw data temporal resolution using selected clustering
methods on residential electricity load profiles,” IEEE Trans. Power Syst., vol. 30, no. 6, pp. 3217–3224, Jun.
2015. doi: 10.1109/TPWRS.2014.2377213.
8. S. Pan, J. Qiao, and L. Zhu, “Application of parallel clustering algorithm based on R in power customer
classification,” in IEEE 4th Int. Conf. Cloud Comput. Big Data Anal. (ICCCBDA), Chengdu, China, 2019,
pp. 165–169. doi: 10.1109/ICCCBDA.2019.8725760.
9. R. Li, N. Smith, and F. Li, “Multi-resolution load profile clustering for smart metering data,” IEEE Trans.
Power Syst., vol. 31, no. 6, pp. 4473–4482, 2016. doi: 10.1109/TPWRS.2016.2536781.
10. R. Q. Sun, X. Q. Xiao, F. Zhou, and Y. S. Zhou, “Research of power user load classification method based
on K-means and FSVM,” in Proc. 2019 IEEE 3rd Inf. Technol., Netw., Electron. Autom. Control Conf.
(ITNEC), Chengdu, China, 2019, pp. 2138–2142. doi: 10.1109/ITNEC.2019.8729085.
11. I. Panapakidis, T. Papadopoulos, G. Christoforidis, and G. Papagiannis, “Pattern recognition algorithms
for electricity load curve analysis of buildings,” Energy Build., vol. 73, pp. 137–145, Apr. 2014. doi:
10.1016/j.enbuild.2014.01.002.
12. S. M. Bidoki, N. Mahmoudi-Kohan, M. H. Sadreddini, Z. P. Jahromi, and M. P. Moghaddam, “Evaluating
different clustering techniques for electricity customer classification,” in Proc. IEEE PES T&D 2010, New
Orleans, LA, USA, Apr. 2010, pp. 1–5. doi: 10.1109/TDC.2010.5484234.
13. P. Laurinec and M. Lucká, “Clustering-based forecasting method for disaggregated end-consumer electric-
ity load using smart grid data,” in Proc. 2017 IEEE 14th Int. Sci. Conf. Inf., Poprad, Slovakia, Nov. 2017,
pp. 210–215. doi: 10.1109/INFORMATICS.2017.8327248.
14. X. Luo, X. Zhu, and E. G. Lim, “A parametric bootstrap algorithm for cluster number determination of
load pattern categorization,” Energy, vol. 180, pp. 50–60, May 2019. doi: 10.1016/j.energy.2019.04.089.
15. J. Cui, R. Yu, D. Zhao, J. Yang, W. Ge and X. Zhou, “Intelligent load pattern modeling and denoising
using improved variational mode decomposition for various calendar periods,” Appl. Energy, vol. 247, pp.
480–491, Aug. 2019. doi: 10.1016/j.apenergy.2019.03.163.
16. S. Wang, R. Li, A. Evans, and F. Li, “Regional nonintrusive load monitoring for low voltage
substations and distributed energy resources,” Appl. Energy, vol. 260, 2020, Art. no. 114225. doi:
10.1016/j.apenergy.2019.114225.
EE, 2024, vol.121, no.9 2653

17. M. A. Asadi, H. Samet, and T. Ghanbari, “Ghanbari, k-NN based fault detection and classification
methods for power transmission systems,” Prot. Control Mod. Power Syst., vol. 2, no. 32, Aug. 2017. doi:
10.1186/s41601-017-0063-z.
18. X. Liu et al., “A data mining research on office building energy pattern based on time-series energy
consumption data,” Energy Build., vol. 259, Mar. 2022, Art. no. 111888. doi: 10.1016/j.enbuild.2022.111888.
19. The US Department of Energy. Data from the Open Energy Information Website. Washington, DC,
USA: National Renewable Energy Laboratory, 2023. Accessed: Nov. 29, 2023. [Online]. Available:
http://en.openei.org/datasets/dataset
20. P. Cihan, “Impact of the COVID-19 lockdowns on electricity and natural gas consumption in the different
industrial zones and forecasting consumption amounts: Turkey case study,” Int. J. Electr. Power Energy
Syst., vol. 134, Jan. 2022, Art. no. 107369. doi: 10.1016/j.ijepes.2021.107369.
21. P. Cihan, “Time-series forecasting of energy demand in electric vehicles and impact of the COVID-19
pandemic on energy demand,” Sakarya Univ. J. Comput. Inf. Sci., vol. 6, no. 1, pp. 10–21, Apr. 2023. doi:
10.35377/saucis.1209519.
22. Q. Wu, H. Ren, S. Shi, C. Fang, S. Wan and Q. Li, “Analysis and prediction of industrial energy consumption
behavior based on big data and artificial intelligence,” Energy Rep., vol. 9, pp. 395–402, 2023. doi:
10.1016/j.egyr.2023.01.007.
23. M. Ali, S. Ullah, M. S. Ahmad, M. Y. Cheok, and H. Alenezi, “Assessing the impact of green consumption
behavior and green purchase intention among millennials toward sustainable environment,” Environ. Sci.
Pollut. Res., vol. 30, no. 9, pp. 23335–23347, Feb. 2023. doi: 10.1007/s11356-022-23811-1.
24. S. Singh and A. Yassine, “Mining energy consumption behavior patterns for households in smart grid,”
IEEE Trans. Emerg. Top., vol. 7, no. 3, pp. 404–419, Apr. 2017. doi: 10.1109/TETC.2017.2692098.
25. S. Thapar, “Energy consumption behavior: A data-based analysis of urban Indian households,” Energy
Policy, vol. 143, Aug. 2020, Art. no. 111571. doi: 10.1016/j.enpol.2020.111571.
26. I. Vassileva, F. Wallin, and E. Dahlquist, “Understanding energy consumption behavior for future
demand response strategy development,” Energy, vol. 46, no. 1, pp. 94–100, Oct. 2012. doi:
10.1016/j.energy.2012.02.069.
27. Z. Dong, S. Wang, W. Zhang, and H. Shen, “The dynamic effect of environmental regulation on firms’
energy consumption behavior-evidence from China’s industrial firms,” Renew. Sustain. Energ. Rev., vol.
156, Mar. 2022, Art. no. 111966. doi: 10.1016/j.rser.2021.111966.
28. F. Abdoune, L. Ragazzini, M. Nouiri, E. Negri, and O. Cardin, “Toward Digital twin for sustainable
manufacturing: A data-driven approach for energy consumption behavior model generation,” Comput. Ind.,
vol. 150, Sep. 2023, Art. no. 103949. doi: 10.1016/j.compind.2023.103949.
29. S. Pelka et al., “Self-consumption rises due to energy crises? An evaluation of prosumers’ consumption
behavior in 2022,” in Proc. 19th Int. Conf. Eur. Energy Market (EEM), Lappeenranta, Finland, Jul. 2023,
pp. 1–6. doi: 10.1109/EEM58374.2023.10161968.
30. W. Salas-Zapata, L. Hoyos-Medina, and D. Mejía-Durango, “Urban residential water and electricity
consumption behavior: A systematic literature review,” Utilities Policy, vol. 83, Aug. 2023, Art. no. 101590.
doi: 10.2139/ssrn.4314049.