Data mining based framework for exploring household electricity consumption patterns: A case study in China context

Accepted Manuscript Data mining based framework for exploring household electricity consumption patterns: A case study in China context Zhifeng Guo, Kaile Zhou, Xiaoling Zhang, Shanlin Yang, Zhen Shao PII:

S0959-6526(18)31607-X

DOI:

10.1016/j.jclepro.2018.05.254

Reference:

JCLP 13111

To appear in:

Journal of Cleaner Production

Received Date: 26 September 2017 Revised Date:

6 May 2018

Accepted Date: 29 May 2018

Please cite this article as: Guo Z, Zhou K, Zhang X, Yang S, Shao Z, Data mining based framework for exploring household electricity consumption patterns: A case study in China context, Journal of Cleaner Production (2018), doi: 10.1016/j.jclepro.2018.05.254. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Data mining based framework for exploring household electricity consumption patterns: A case study in China context

a. School of Management, Hefei University of Technology, Hefei 230009, China

RI PT

Zhifeng Guo a,b, Kaile Zhou a,b*, Xiaoling Zhang c, Shanlin Yang a,b, Zhen Shao a,b

b. Key Lab of Process Optimization and Intelligent Decision-making, Ministry of Education, Hefei 230009, China

SC

c. City University of Hong Kong, Kowloon, Hong Kong

Abstract: This study proposes a data mining based framework for exploring the electricity

M AN U

consumption patterns, which includes three consecutive stages. Firstly, electricity consumption patterns and behaviors are explored in festivals such as the Spring Festival, the Labor Day and the National Day. Secondly, seasonal electricity consumption patterns and behaviors are compared,

TE D

and the relationship between temperature and electricity demand is analyzed through data visualization. Thirdly, we focus on the phenomenon of electricity consumption patterns shifting. Finally, a case study of Nanjing and Yancheng City, Jiangsu Province, China is presented. The

EP

results indicate that: (1) Volatility of electricity consumption is higher in winter and summer than

AC C

in spring and autumn. (2) There are three typical load profiles during the Spring Festival, two typical load profiles during the Labor Day the National Day. (3) High temperature in summer and low temperature in winter have obvious influence on electricity consumption. However, the electricity consumption peak lags one or two days behind the temperature peak in summer, and consumers’ response time gets shorter as the frequency of temperature peaks increase. (4) The phenomenon of instability of household electricity consumption patterns is identified. 7.22% of *

Corresponding author. E-mail addresses: [email protected], [email protected] (K. Zhou)

1

ACCEPTED MANUSCRIPT the high volatility households transferred to low volatility households from winter to spring. 6.08% low volatility households transferred to high volatility households from summer to autumn. Finally, we proposed some suggestions for promoting energy conservation and improving energy

RI PT

efficiency.

Keywords: Household electricity consumption patterns; framework; clustering; seasonal

M AN U

SC

characteristics; temperature

1. Introduction

With the rapid development of the global economy, energy demand has increased significantly (Zhang et al., 2017; Liu et al., 2016; Liu et al., 2015; Narayan and Sharma, 2015). In recent years,

TE D

the proportion of household electricity consumption has increased obviously (Cabeza et al., 2014; Huebner et al., 2016; Pothitou et al., 2017; Tomás et al., 2015). However, traditional energy

EP

consumption structure mainly consists of thermal power that leads to serious environmental pollution (Shi et al., 2015; Tang et al., 2016; Zhang et al., 2016). Peak load shifting according to

AC C

household electricity consumption patterns is an effective solution that maintains the balance between supply and demand from the perspective of demand response (Salo et al., 2016). Currently, there have been considerable literatures that focused on the residential electricity consumption patterns and behaviors (Allen et al., 2015; Botetzagias et al., 2014; Kaplowitz et al., 2012; Nilsson et al., 2015). Some related methods that have an impact on household electricity consumption behaviors are proposed (Lin et al., 2016; Ntona et al., 2015; Tejani et al., 2011; Zhou and Yang, 2016). Intervention is an important strategy for reducing electricity consumption during 2

ACCEPTED MANUSCRIPT peak time periods compared with other strategies (Abrahamse et al., 2005; Kok et al., 2011; Nilsson et al., 2015). However, it cannot exert lasting effects on different kinds of households. A major reason is that different families have different electricity consumption patterns (Abreu et al.,

RI PT

2012), and even the same household may have different load profiles during different seasons. Traditional intervention strategies are time-invariant and lack of diversity, such that they are not applicable for different electricity consumption patterns (Guo et al., 2018). Therefore,

SC

personalized intervention strategies need to be developed according to the household

M AN U

characteristics (Zhou and Yang, 2016). From this point of view, mining electricity consumption behavior patterns is conductive to develop personalized intervention strategies and services, such that improving customer experience, utility operations, and advanced power management. There are some existing research efforts that can be mainly summarized in four aspects. The

TE D

first research stream is that different granularity of electricity consumption data are applied to obtain typical load profiles, for example, daily data, hourly data and minute data. Besides, different clustering algorithms are compared, such as k-means (Hartigan and Wong, 1979), SOM

EP

(Kohonen, 1998) and ensemble clustering (Yu et al., 2012). Panapakidis et al. (2014) implemented

AC C

clustering on load curves of nine buildings of the AUTH university campus by means of different clustering algorithms, including the FCM (Bezdek et al., 1984), SOM (Kohonen, 1998) and K-means++ (Jain, 2010). McLoughlin et al. (2015) analyzed hourly electricity consumption data using k-means, k-medoid and SOM, and the best performing technique was then applied to segment individual households into groups. The result revealed a series of load profiles that present common electricity consumption patterns. The second research direction is feature selection before data clustering. Feature selection gets

3

ACCEPTED MANUSCRIPT more attention during the big data era (Räsänen and Kolehmainen, 2009). Dimension of electricity consumption data can be very high with the popularity of smart meters, and some traditional clustering algorithms become invalid in high dimensional space, Feature selection is a feasible

RI PT

way of tackling high dimensional data. Räsänen and Kolehmainen (2009) proposed an efficient computational method for time series data based on the extraction of statistical features. The results showed that the approach could obtain more accurate load profiles, and they pointed that

SC

there are several advantages of this method. Other methods could also be applied to tackle the

M AN U

problem. For instance, Granell et al. (2015) used a Bayesian non-parametric model to group load profiles from households and business premises.

The third area of research is to investigate family characteristics through data analysis. There are various data that can be utilized to infer family characteristics such as electricity consumption

TE D

data, social economic data and temperature data (Bessec and Fouquau, 2008; Jovanović et al., 2015). The family characteristics are vital for making personalized intervention strategies from the point of demand side management. Bessec and Fouquau (2008) studied the relationship between

EP

electricity demand and temperature in the European Union, and the results indicated that there

AC C

existed a nonlinearity link between electricity demand and temperature in more limited geographical areas.

In addition to the above mentioned aspects, household electricity consumption patterns is explored by many scholars from the perspective of regional difference. This is mainly due to that different countries have different background that have fundamental influence on people’s lifestyle, which can be reflected by electricity consumption load profiles. Kantor et al. (2017) proposed that conditioning

the indoor temperature can have an important impact on electricity use. Especially, the monthly

4

ACCEPTED MANUSCRIPT electrical usage is correlated with temperature and then predicted for electrical usage based on the weather patterns within the province of Ontario, Canada. Kobayakawa and Kandpal (2015) showed that with a reduction in the ambient temperature, though the requirement of an electric fan

RI PT

is reduced, the overall electricity consumption of the household is found to increase due to more intensive use of electricity during winter season in India. Garg et al. (2014) provided an assessment of household electricity load curves in India. The results showed that during summer,

SC

space cooling loads are the highest contributors to total electricity consumption, followed by

M AN U

refrigerators, ceiling fan and lighting load. However, in winters, refrigerators contribute the most to the total electricity consumption, followed by ceiling fans and lighting equipment. In fact, ambient temperature have fundamental effect on the usage of refrigerators and air conditioners. Obinna et al. (2017) studied the issue from the point of energy monitor by means of questionnaire

TE D

in USA. They showed that 50% of the surveyed participants were satisfied with the adapted energy monitor. Wilson (2013) proposed that electricity consumption was modeled as a function of climate, demographic, structural, technological, behavioral, and urban form factors by means of

EP

linear regression model. Ozawa et al. (2016) proposed two methods that identify a household’s

AC C

lifestyle from electricity use data in Japan, and they showed that the average daily consumption of morning-oriented lifestyles is 5.3% less than that of night-oriented lifestyles. Mizutani et al., (2018) investigated the effect of demand response on households’ electricity consumption under the existence of the reference price effect. The results suggested that the reduction of electricity usage is greater in the case where the current prices are higher than the previous prices. Faisal et al. (2016) found the bi-directional causality was observed from electricity consumption to GDP. The results suggested that both the electricity consumption and economic growth empirically had a

5

ACCEPTED MANUSCRIPT mutual and complementary relationship. However, there are still some limitations in previous studies. To explore the residential electricity patterns more comprehensively, it could be improved from three aspects. The first is

RI PT

that local electricity consumption patterns should be derived from local data instead of all data. For example, this study utilizes seasonal data instead of yearly data to get seasonal electricity consumption patterns. The second is that electricity consumption patterns in festivals and holidays

SC

should be considered. People’s activities change in festivals may lead to sharp decreases or

M AN U

increases in electricity demand. In this situation, mining electricity consumption patterns during festivals contributes to the understanding of household electricity consumption behaviors. The third point is that electricity consumption patterns shifting should be explored in the turn of seasons. Electricity consumption patterns are varying in different seasons (De Felice et al., 2015;

TE D

Taylor, 2010), so it is necessary to track the change in electricity consumption patterns. To further investigate the household electricity consumption patterns, this paper proposes a clustering-based research framework to understand the household electricity consumption patterns

EP

from the multi-perspective views. The contributions of this study lie in the following three aspects.

AC C

(1) It suggests that the local electricity consumption patterns should be derived from local data instead of all data. For example, electricity consumption patterns in the Labor Day (from May 1st to May 3rd) are derived from electricity consumption data during the Festival period is more reasonable. (2) It proposes a data mining based framework in which the electricity consumption patterns are explored at three levels. At the first level, typical load profiles in festival such as the Spring Festival, the Labor Day and the National Day are explored. At the second level, seasonal electricity consumption patterns are explored. At the third level, the relationship between

6

ACCEPTED MANUSCRIPT temperature and electricity demand is analyzed using data visualization. Besides, we analyze the electricity consumption behaviors that can be reflected by load profiles and give related suggestions on energy saving. (3) This paper proposes the phenomenon of electricity consumption

RI PT

patterns shifting and identified the phenomenon through transfer matrix, which could help to make personal intervention strategies.

The remainder of the paper is structured as follows: Section 2 presents the clustering-based

SC

framework, and elaborates the related stages in detail. Section 3 provides a case study in Nanjing

2. Framework and methodology

M AN U

and Yancheng City and discusses the results. Section 4 gives concluding remarks.

A data mining based framework is proposed to explore the electricity consumption patterns

TE D

from several aspects. The proposed framework (see Fig. 1) comprises consecutive four phases, and the first phase is smart power data collection. Residents consume electricity provided by

EP

power station, and the related electricity consumption data is recorded by smart meter and transmitted to the data center. The second phase is data cleaning. Not all electricity consumption

AC C

data is recorded and transmitted to the data center entirely and consecutively due to a variety of reasons, such as malfunction of equipment, broken circuit and short circuit. It always leads to outliers and missing values. These incomplete data has a negative impact on data analysis. In this phase, the non-residential users, missing value users, zero-value users and outlier users will be removed for the purpose of analysis. In the third phase, the typical load profiles are found in different seasons and festivals. On this basis, the relationship between temperature and electricity consumption is explored using data visualization. In the last phase, the phenomenon of electricity 7

ACCEPTED MANUSCRIPT consumption patterns shifting is studied. The following elaborates four stages mentioned above in detail.

Power transmission

Power consumption

1.Smart power data collection

Non-residential users removal

Missing-value residential user removal

Data center

Outlier users removal

M AN U

Spring Festival (2014.01.30)

Zero-use residential user removal

SC

2.Data clean & preprocessing

Data acquisition

RI PT

Power generation

Festivals

Labor Day (2014.05.01)

National Day (2014.10.01)

The patterns of power consumption in festivals

Winter (2014.01.01-2014.02.28) Spring (2014.03.01-2014.05.31)

3.Data clustering

Seasons

TE D

Summer (2014.06.01-2014.08.31) Autumn (2014.09.01-2014.11.30)

Relationship between electricity consumption and temperature

EP

Temperature

Grouped by volatility

AC C

4.Electricity consumption pattern shift

The pattern of power consumption in seasons

High volatility load profiles

Low volatility load profiles

Low volatility load profiles

High volatility load profiles

Temperature and power consumption profiles

Fig. 1 The proposed research framework.

2.1 Smart power use data collection Data transmission and storage are of great importance during big data era (Zhou and Yang, 2016). With the development of Internet, massive data is generated every day. Many new tools and platforms have been produced for handling big data in recent years, such as Amazon Cloud, Window Azure and Google Cloud. With the coming of big data era, the development of energy 8

ACCEPTED MANUSCRIPT sector has sped up by means of big data technology, which is vital to promote energy efficiency on the condition of global scarcity of energy resources. In the field of electricity, electricity consumption data is massive with the popularity of smart meters. Electricity consumption data is

RI PT

recorded by smart meter, and then transmitted to data center. Any mistake in the process of transmission and storage may causes missing values. In this sense, data collection is very

SC

important, and subsequent studies are based on this.

M AN U

2.2 Data cleaning and preprocessing

In data center, the electricity consumption data are not all perfect. Data corruption, missing values and outliers are mainly three typical problems that we need to handle. In this paper, non-residential and residential electricity users are blending. Data cleaning is necessary for the

TE D

purpose of data analysis. Four steps are proceed in turn. These are Non-residential Users removal, Missing-value Residential Users removal, Zero-use Residential Users removal and Outlier Users removals. As Zhou et al. (2017) defined, if there is at least one missing value in the daily

EP

electricity consumption data, the residential user is defined as a Missing-value Residential User. If

AC C

a residential user is not a Missing-value Residential User, and there are more than 20 days a month when its daily electricity is less than 1 kWh, the user is defined as a Zero-use Residential User. If the residential daily electricity consumption data is extremely high or low in a day, this residential will also be defined as Outlier User, and it will be removed from the data set. If there exists at least one value exceeds 30 kWh in the daily electricity consumption data, this type of residential user is defined as a Non-residential User. The Non-residential Users will also be removed from the data.

9

ACCEPTED MANUSCRIPT

2.3 Data clustering Massive electricity consumption data is collected through smart data acquisition system. Data

RI PT

cleaning is a vital step before cluster clustering. In this stage, advanced data mining methods, such as clustering analysis, can be applied on these data to discover valuable knowledge for policy makers. Residential users with same electricity consumption patterns can be clustered into one

SC

group. K-means is an unsupervised learning process and it is one of the most popular machine learning algorithm, (see Table 1). In this paper, electricity consumption data relatively smooth that

M AN U

is the assumption of k-means algorithm. Namely the joint distribution of features within each cluster is spherical, in which, k-means algorithm works. Besides, considering that we are facing big data environment, k-means algorithms works faster than other more complex cluster algorithm due to the fact that its time complexity is linear with the number of data. Furthermore, this

TE D

algorithm is easy to interpret since its clusters have similar density in each clusters compared with neural networks related models. Since k is the hyper-parameter, in this case, we select small range

EP

of k value, such as 2-10. For each k value, we take the clustering result according to inertia or within-cluster sum-of-squares, and final determination is supported by means of data visualization.

AC C

The k-means algorithm will stop if center of each clustering do not change much or reach the maximum iterations. In this way, it and can divide similar individuals into the same group. Cluster centers are determined at random, then the iterative algorithm is applied to adjust the clustering center step by step. Thus determining an appropriate cluster number using cluster validity index is a prerequisite step in residential user segmentation. The typical electricity consumption profiles of each user group represented by the corresponding cluster centers are also obtained. Based on this, characteristic indicators are extracted from the electricity consumption profiles. 10

ACCEPTED MANUSCRIPT

Table 1 K-means algorithm. Algorithm K-Means Clustering Step 1: Select randomly k points as initial centroids.

RI PT

Step 2: Form k clusters by assigning each point to its closest centroid. Step 3: Recompute the centroid of each cluster. Step 4: Repeat Step 2 and Step 3.

SC

Step 5: Obtain the clustering results once convergence criterion is met.

M AN U

In the phase, it explores the electricity consumption patterns at three levels. At the first level, typical load profiles in festivals such as the Spring Festival, the Labor Day and the National Day are explored from the view of regions. Household electricity consumption patterns in fact are the reflection of life styles. Different types of families will participate in various activities with their

TE D

members of family or others in festivals and holidays, that leads to different electricity load profiles. Therefore, it is necessary to explore typical load profiles in festivals. At the second level, seasonal electricity consumption patterns are explored. Season interval is

EP

determined by meteorological division method. Temperature is an important indicator for different

AC C

seasons, and it has significant effect on human activity. Electricity consumption patterns among four seasons should be distinguished according to people’s lifestyle. Finally, we give a time division (see Table 2) in order to facilitate the research.

Table 2 Festivals and seasons. Festivals and seasons

Date period

Spring Festival

2014.01.25-2014.2.6

Labor Day

2014.4.30-2014.05.04

11

ACCEPTED MANUSCRIPT 2014.09.30-2014.10.08

Winter

2014.01.01-2014.02.28

Spring

2014.03.01-2014.05.31

Summer

2014.06.01-2014.08.31

Autumn

2014.09.01-2014.11.30

RI PT

National Day

We notice that the impact of festival and season may overlap from the perspective of date. Here

SC

we simply regard the seasonal and festival electricity consumption patterns as a phenomenon according to the time interval. In fact, it is different to get purely the festival electricity

M AN U

consumption patterns on the basis of eliminating the influence of seasons. And vice versa. This is largely due to that electricity consumption behaviors is complex. On one hand, it contains the people’s daily activities; and on the other hand, there exist some additional activities during the

TE D

festivals and holidays. In this case, festival electricity consumption patterns are mainly determined by people’s festival activities. Besides, we utilizes festival data instead of yearly data to get festival electricity consumption patterns.

EP

At the third level, the relationship between temperature and electricity demand is explored using

AC C

data visualization. It is well known that the temperature have influence on electricity consumption. People have different activities, diets and lifestyles under different temperatures, all of these are reflected on electricity consumption patterns. This study explores the temperature sensitivity. Special attention is paid to the relationship between the temperature peaks and the energy consumption peaks.

2.4 Electricity consumption patterns shifting Residential electricity consumption behaviors are varying in different seasons. This paper

12

ACCEPTED MANUSCRIPT focuses on the electricity consumption patterns shifting phenomenon. For example, some residents will be assigned into one group in winter by means of clustering, and will these residents be assigned into one group again in spring? In other words, there may exist some residents who could

RI PT

be assigned into one group in winter, however, could not be assigned into the same group again in spring. The reason is that these residents share the same electricity consumption patterns in winter so they have similar load profiles that are easily clustered into one group. However, for some

SC

reasons, some of them changed their electricity consumption patterns in the following season. In

M AN U

this case, these residents do not share with same electricity consumption patterns in spring, and this is why they cannot be clustered into one group again. It always occurs in the turn of seasons, such as from winters to spring, from spring to summer and from summer to autumn. The phenomenon has some negative effects for policy makers to make intervention strategies. This is

TE D

because researchers often partition electricity consumption users into groups by cluster algorithms. However, the group is unstable, and some of them transfer to other groups in the next season. It means that intervention strategies should not be made according to the results of clustering at a

EP

whole time period, due to the fact that some of them will not belong to the group next time. This

AC C

paper pays attention to these residents who are easy to transfer to other groups with the turn of seasons.

This paper defines a transfer matrix (see Table 3) and utilizes it to analyze electricity consumption patterns shifting phenomenon. In the first column, it displays the groups in winter. The first row is groups in spring. In the first column, the group Gi indicates it is the

ith group.

Each group is recognized by its volatility and proportion (Prop.). Here we measure the volatility by its standard deviation (Std.). And the first row shows the same information in spring. Pij

13

ACCEPTED MANUSCRIPT represents the transition proportion. We define Pij = Ai ∩ B j set of the

Ai , where Ai represents the

ith group of users in winter. B j represents the set of j th group of users in spring.

th Ai is the number of i group of users. Ai ∩ B j is the number of intersection set between

RI PT

group Ai and B j . In this situation, Pij represents the proportion of transfers from group Ai to group B j . For example, A1 = {id1 , id 2 , id 3 , id 5 } represents the first group A in winter, and it is composed of residential user

{id1 , id 2 , id 3 , id 5 } , B2 = {id 2 , id 3 , id 4 }

{id1 , id2 , id3 , id4 , id5} = 0.6 . It indicates

In this case,

60% of group A1 transfers

M AN U

P12 = {id2 , id3 , id4 }

{id 2 , id 3 , id 4 } .

SC

group B in spring and is composed of residential users

represents the second

to group B2 from winter to spring. In this case, we can explore transfer process between high volatility users and low volatility users.

Table 3 Transfer matrix from winter to spring. Spring

G1

G2

M

Gm

L

P1m

P12

P21

P22

M

M

O

M

Pn2

Pnj

Pnm

Pn1

P2m

AC C

Gn

L

P11

EP

G1

TE D

Winter

G2

3. Case study

3.1 Overview of the study area Regional electricity consumption in China and the sample of cities in Jiangsu Province are shown in Fig. 2 (a). Electricity consumption is displayed with different color. The more electricity consumption is, the darker the color is. Fig. 2 (b) is the sample of cities in Jiangsu, two navy blue areas are targeted cities. Namely, Nanjing and Yancheng.

14

ACCEPTED MANUSCRIPT With the process of reform and opening up, China has experienced a rapid development in economy and society. Electricity demand, which embodies the economic development level, has been increasing accordingly. The provinces in the east of China (see Fig. 2 (a)) have higher

RI PT

electricity demand, such as Guangdong, Zhejiang, Shandong, Jiangsu and Hebei. This is because that more industries are located in these regions. Besides, these regions have a very high population density. In the western regions, such as Qinghai, Gansu, Xinjiang and Tibet, electricity

SC

consumptions are obvious lower compared to the coastal cities. From an economic point of view,

M AN U

electricity consumption is related to the economic development level,

Jiangsu Province is one of several coastal provinces in China, and it is a big manufacturing province. Jiangsu has a quite developed economy. Nanjing and Yancheng are two typical cities in Jiangsu. Nanjing is provincial capital of Jiangsu, and it has 821.61 million people. Its urbanization

TE D

rate is 80.92% and ranks 7th in China. It is a very competitive city and an important national center of science and education. Nanjing is also famous for economic and tourism. Yancheng is the largest prefecture level city in Jiangsu Province, it covers an area of 1.7 million square kilometers

EP

and has 721.06 million permanent people. Its urbanization rate is 60.1%. The urbanization rate of

AC C

Nanjing is higher than that of Yancheng. It is concluded that Nanjing is more developed than Yancheng from an economic point of view.

15

ACCEPTED MANUSCRIPT (a) Regional electricity consumption in China

Heilongjiang

Jiling

Inner Mongolia Xingjiang

Liaoning Beijing Tianjing ShanxiHebei

Gansu Ningxia

Henan Tibet

Jiangsu Anhui Shanghai

Hubei Chongqing

Sichuan

Zhejiang Guizhou

Jiangxi

Hunan

Fujian Yunnan Guangdong Macao

Taiwan

Hongkong

Hainan

M AN U

(b) The sample of cities in Jiangsu province

SC

Guangxi

RI PT

Shandong Shannxi

Qinhai

Lianyungang

Xuzhou

Suqian

Yancheng

Huaian

Yangzhou

TE D

Haozhou

Nanjing

Nantong Zhengjiang Wuxi Suzhou

EP

Changzhou

Fig. 2 Regional electricity consumption in China and the sample of cities in Jiangsu Province.

AC C

Note: The regional electricity consumption data of Fig. 2 (a) come from China Electric Power Yearbook in 2015.

3.2 Data selection and preprocessing We select daily electricity consumption data of Nanjing and Yancheng City from January 1, 2014 to December 31, 2014 from large data center in State Grid which is responsible for deploying smart grid, smart meter, power supply and collecting electricity consumption data as we mentioned in section 2.1. The number of initial sample of Nanjing and Yancheng numbers are

16

ACCEPTED MANUSCRIPT 73674 and 2458, in the data cleaning and preprocessing stage, Outlier Users, Non-residential Users, Missing-value Residential users and Zero-use Residential users were totally removed. After data cleaning and preprocessing, the valid number of residential users for analysis of Nanjing and

RI PT

Yancheng City are 3000 and 1399, respectively.

3.3 Experimental setup

M AN U

an Intel Core3 CPU M380 at 2.53 GHz with 4 G of memory.

SC

The k-means algorithm is implemented in python 3.5, and experiments were performed using

3.4 Electricity consumption patterns during festivals

3.4.1 The electricity consumption patterns during the Spring Festival The Spring Festival is one of most grand Chinese traditional festivals, and many people will

TE D

return to their hometown and reunite with their family members after a year’s hard work. In China, the largest population flow occurs during the Spring Festival. In this situation, some families who

EP

live in the cities will head for hometowns, and the population flow out of cities will cause a significant decrease of electricity consumption in cities. In another case, Family members are

AC C

reunited at home during the Spring Festival, which causes a significant increase of electricity consumption. In this Section, we focus on finding typical load profiles during the Spring Festival. We determine nine typical load profiles both in Nanjing (see Fig. 3) and the descriptive statistics of each group are displayed in Table 4.

17

ACCEPTED MANUSCRIPT

2

3

4

5

6

7

8

9

The Spring Festival(from 2014-01-15 to 2014-02-06 )

RI PT

Daily electricity consumption(kwh)

Nanjing City 1

SC

Fig. 3 Electricity consumption patterns during the Spring Festival.

In Fig. 3, the red vertical line indicates the day of the Spring Festival. The maximum average

M AN U

daily electricity consumption is 14.48 kWh in Nanjing, and its proportion is 3.8%. The largest proportion is 22% residential users corresponding to the 2nd subplot in the panel in Nanjing. Focusing on the changing trend, we can find that there are three types of load profiles during the Spring Festival according to the shape of cures, as summarized in Table 4. The first column

TE D

indicates the initial cluster labels in Fig. 3, and the fifth column indicates the types that the difference in quantity is ignored.

EP

Type I: The electricity consumption increased significantly before vertical line and decreased significantly after vertical line. It looks like an upper convex curve, such as the 1st, 4th, 5th and 8th

AC C

subplot in the panel.

Type II: The electricity consumption decreased significantly before vertical line and increased significantly after vertical line. It looks like a lower convex curve, such as the 3rd, 7th and 9th subplot in the panel. Type III: The electricity consumption neither increased significantly before vertical line nor decreased significantly after vertical line. It looks like a horizontal line, such as the 2nd and 6th subplot in the panel. 18

ACCEPTED MANUSCRIPT

Table 4 Descriptive statistics of each group in Nanjing City during the Spring Festival. Mean

Std.

Prop.

Type

1

10.94

1.26

0.07

I

4

7.76

0.71

0.13

I

5

5.43

0.53

0.18

I

8

14.48

2.08

0.04

I

3

5.06

3.43

0.04

II

7

4.93

2.67

0.07

II

9

9.10

3.31

0.06

2

3.31

0.33

0.22

6

1.37

0.34

Prop. (%)

RI PT

Cluster No.

SC

42

II

III

M AN U 0.20

16

42

III

In general, electricity consumption patterns are the reflection of electricity consumption

TE D

behaviors. For the first type load profile, most families prepare for family reunion dinner and the kitchen that equipped with many electrical equipments, such as refrigerator, microwave oven and electromagnetic furnace operate more than at ordinary times. This is largely due to that the Spring

EP

Festival is a big event to Chinese. Besides, decorations are hanging up everywhere for celebration

AC C

of the festival, the same as Christmas in the West. People sit in front of the TV and watch the Spring Festival Gala and TV programs. The number of family population is more than usual because of the Spring Festival. Thousands of Chinese people are sitting at home to gather for the special day. This will inevitably lead to an increase in electricity consumption for not only a family but a whole country. That’s why electricity consumption rises on the eve of the Lunar New Year holiday as type I load profile. After the Spring Festival, the situation is contrary to that before the Spring Festival. People felt exhausted after a busy year at work and relax themselves. In this

19

ACCEPTED MANUSCRIPT case, electricity consumption will drop a little. For the type II, it is still suitable for family life in China, specifically, many people will leave for their hometown that will lead a decrease even a drop-off in electricity consumption. People

RI PT

will return their dwelling houses after getting together with relatives and friends. Electricity consumption naturally return to its original level. One possible situation is that people travel or take part in other relaxation activities with the development of China’s economy during the Spring

SC

Festival. Type III are likely refer to some specific groups, such as an elderly person of no family

M AN U

or poor population. Their family members and diet will not change much compared to the usual. In this situation, the trend line of electricity consumption is almost flat-feeling. There are some suggestions for reducing energy consumption and improving energy efficiency. Firstly, providing some energy saving recommendations, particularly, for the type I users.

TE D

Household appliances are used frequently during the Spring Festival, reasonable use can reduce unnecessary energy consumption, for example, the temperature of the air conditioner neither is too high nor too low, and turning off the air conditioning each morning when we leave for work,

EP

Secondly, it is one of most effective means to adopt an efficient bulb, the energy consumption of

AC C

night lighting is significantly reduced through the installation of energy saving bulbs. Thirdly, for the type II, Turn off your home’s electricity when there is no one at home is most useful practice. For the type III load profiles, energy saving tips still work. 3.4.2 Electricity consumption patterns in the Labor Day The Labor Day is another important festival in China. It is a legal three-day holiday. Many families travel or visit friends in the holiday. It has influence on electricity consumption. This section focuses on the electricity consumption patterns during the Labor Day. The clustering

20

ACCEPTED MANUSCRIPT results are displayed in Fig. 4 and Table 5.

2

3

4

5

6

7

8

9

RI PT

Daily electricity consumption(kwh)

Nanjing City 1

SC

The Labor day(from 2014-04-30 to 2014-05-04)

Fig. 4 Electricity consumption patterns during the Labor Day.

M AN U

In Fig. 4, the red vertical line indicates the day of the Labor Day. We identify nine typical load profiles in Nanjing during the Labor Day. Most residents’ load profiles look like a horizontal line although the average daily electricity consumption is different. The largest proportion is 21% users corresponding to the 4th subplot that daily electricity consumption is 3.18 kWh in Nanjing.

TE D

It can identify two typical load profiles in Nanjing if we ignore the difference in quantity and only consider the shape of curves.

EP

Type I: The electricity consumption neither increased significantly before vertical line nor decreased significantly after vertical line. It looks like a horizontal line, such as the 1st, 2nd, 3rd, 6th,

AC C

7th, 8th and 9th subplot in the panel.

Type II: The electricity consumption was invariable before May 3rd and increased significantly in the May 3rd, such as the 5th subplot in the panel.

Table 5 Descriptive statistics of each group in Nanjing City during the Labor Day. Cluster No.

Means

Std.

Prop.

Type

Prop. (%)

1

2.17

0.07

0.20

I

98

21

5.39

0.24

0.11

I

3

6.92

0.37

0.07

I

4

3.18

0.17

0.21

I

9

5.72

1.16

0.05

I

6

4.20

0.14

0.19

I

7

0.95

0.04

0.10

I

8

8.48

0.41

0.04

I

5

10.71

0.18

0.02

II

2

SC

2

RI PT

ACCEPTED MANUSCRIPT

Here, we identified two types of load profiles, and the electricity consumption patterns are

M AN U

relatively homogenous compared with the case in the Spring Festival. We observed that 98% users have the trend line as usual (see Table 5), only 2% users have slight fluctuation. Take type I for example, this is manly due that there are three days in the labor day according to the Chinese law. It is relatively short compared with other holidays such as the Spring Festival and the National

TE D

Day. Most people will take a rest at home or simply go to the park and museum with their relatives and friends. In this setting, electricity consumption neither increases nor declines significantly. For

EP

the type II, possible reasons is that a part of people have a party or other activities that lead to an increase in electricity consumption with relatives and friends, we cannot deduce much because we

AC C

do not access to more detailed data. We cannot provide special suggestions on energy conservation since the lack of diversity in electricity consumption patterns. However, energy saving tips still suitable in the situation.

3.4.3 The Electricity consumption patterns during the National Day The National Day is another legal holiday that people can enjoy 7 days’ vacation. Many families travel at home or abroad. In this case, electricity consumption pattern is different according to people’s holiday life style. We identify nine typical load profiles in Nanjing (see Fig. 22

ACCEPTED MANUSCRIPT 5). Maximum average daily electricity consumption was 9.58 kWh in Nanjing, and the proportion of both cities was about 1%. The largest proportion was 26% corresponding to the 3rd subplot that

RI PT

daily electricity consumption was 3.70 kWh in the panel. There are two typical load profiles from the shape of the curves in Nanjing during the National Day.

SC

Type I: The electricity consumption neither increased significantly nor decreased significantly

M AN U

during the National Day. It looks like a horizontal line, such as the 1st, 3rd, 4th, 5th, 6th, 7th and 8th subplot in the panel.

Type II: The electricity consumption firstly decreased and then increased significantly during

TE D

the National Day. It looks like a lower convex line, such as 2nd, 9th subplot in the panel.

Nanjing City

2

3

4

5

6

7

8

9

AC C

EP

Daily electricity consumption(kwh)

1

The National Day (from 2014.09.30 to 2014.10.07)

Fig. 5 Electricity consumption patterns during the National Day in Nanjing and Yancheng.

Table 6 Descriptive statistics of each group in Nanjing City during the National Day. Cluster No.

Mean

Std.

Prop.

Type

1

1.33

0.26

0.09

I

3

3.70

0.12

0.26

I

4

2.16

0.72

0.24

I

23

Prop. (%)

93

ACCEPTED MANUSCRIPT 5

6.62

1.22

0.09

I

6

9.58

1.54

0.02

I

7

0.51

0.50

0.05

I

8

4.72

1.24

0.18

I

2

4.38

1.84

0.02

II

9

2.07

1.13

0.05

II

RI PT

7

The electricity consumption patterns during the National Day is consistent with the fact that

SC

people take advantage of this relatively long holiday to travel according to the information from National Tourism Administration of the People's Republic of China. The holiday is one of the best

M AN U

times in China and we can evidently observe that there are two types of load profiles. For the first type, people just walk around or take part in some local activities, such as watch a movie with their friends or go to restaurants. In this setting, electricity consumption do not change much. For the second type, people spend several days on visit tourist areas at home and abroad. In general,

TE D

whether leave a home is one of most important factors that have influence on electricity consumption. In other words, the reduction of electricity usage is greater in the case where the go

EP

out to travel than in the case where people participate in activities near home. The most effective mean for reducing the electricity consumption is that remind people who go out to travel turn off

AC C

all electrical appliances that is not necessary to run.

3.5 The seasonal electricity consumption patterns 3.5.1. Electricity consumption patterns in winter and spring For further study of electricity consumption patterns from the perspective of seasons, we focus on electricity consumption patterns in a season. We distinguish six typical load profiles in Nanjing during winter (see Fig. 6). Volatility of each typical load profile is high. It is mainly affected by 24

ACCEPTED MANUSCRIPT cold weather and the Spring Festival. On the one hand, fluctuation of temperature brings changes in electricity consumption, which we will discuss in the next section. On the other hand, the electricity consumption changes according to the occurrence of population flow during the Spring

RI PT

Festival. However, we notice that there is a low volatility load profile in Nanjing, such as the 3rd subplot in the left panel and its proportion is 18% (see Table 7). Average daily electricity consumption is about 2 kWh. Descriptive statistics of each group are summarized in Table 7. The

SC

largest proportion is 30% and corresponding average daily electricity consumption is 7.01 kWh in

2

4

5

3

4

Daily electricity consumption(kwh)

1

Nanjing City

TE D

Daily electricity consumption(kwh)

Nanjing City

M AN U

Nanjing.

3

1

6

Winter (High volatility)

2

3

5

6

EP

Spring (Low volatility)

Fig. 6 Electricity consumption patterns during the winter and spring in Nanjing.

AC C

Table 7 Descriptive statistics of each group in Nanjing City during the winter. 1

2

3

4

5

6

Prop.

0.30

0.11

0.18

0.13

0.04

0.24

Mean

7.01

7.59

2.06

4.50

10.16

13.96

Std.

0.59

2.80

0.15

0.64

1.39

2.50

The electricity consumption patterns in spring are different compared with it in winter (see Fig. 6). The load profiles get smooth, besides there is a slight downward trend in electricity

25

ACCEPTED MANUSCRIPT consumption. The reason is that the weather gets warmer gradually and people’s outdoor activities increase, so the heating equipment turns off. It contributes to the decrement of the electricity consumption.

RI PT

We notice that 33% families’ average daily consumed electricity 5.03 kWh (see Table 8) in Nanjing. It is the largest proportion group.

Table 8 Descriptive statistics of each group in Nanjing City during the spring. 3

4

5

6

Prop.

0.21

0.23

0.12

0.06

0.33

0.05

Mean

5.07

1.69

7.29

10.25

5.03

3.33

Std.

0.49

0.23

0.86

1.30

2.20

0.41

SC

2

M AN U

1

There are some measures can be taken to reduce electricity consumption and promote energy

TE D

conservation. Specifically, In the case of winter, High volatility of consumption curve is greater that shows more elasticity than its in spring. On one hand, time-of-use electricity price could be an

EP

effective means of intervention. Facing a flexible tariff, a large number of electricity customers shift their energy use to off-peak times because the peak prices are more expensive than a flat rate

AC C

besides, there's less need for expensive upgrades as the current infrastructure can be used more efficiently, flexible pricing will not suit everyone. However, it will be most beneficial for households that use a lot of energy in the cheaper off-peak periods. On the other hand, strengthen energy conservation propaganda is necessary by means of the Internet, As the number of net citizens swell and the time they spent online increases, Internet has become a main platform for people to pay attention to various policies and news. Over the last few years, mobile Internet has gained rapid development and has been studied and applied in all kinds of fields. Therefore, many 26

ACCEPTED MANUSCRIPT energy saving behaviors, for instance use energy-efficient lighting, shut off idle equipment and lighting, decrease heat usage, and turn down your water heater and so on, could be advocated through the network.

RI PT

In the case of spring, the trend line is smooth that have less flexibility compared with that in winter due to mild climate and regular life. However, energy conservation tips still work well in this situation.

SC

3.5.2 Electricity consumption patterns in summer and autumn

M AN U

The electricity consumption patterns of summer and winter are similar from the perspective of volatility (see Fig. 8). We notice that there is a group which have low volatility users such as the 6th subplot in Nanjing. Their average daily electricity consumption were 1.96 kWh. The Proportion of the group was 0.28 in Nanjing. The most high volatility group consumed 8 kWh in

TE D

Nanjing. Most families consumed 1.96 and 6.65 kWh in Nanjing.

3

5

6

Nanjing City

Daily electricity consumption(kwh)

4

2

EP

1

AC C

Daily electricity consumption(kwh)

Nanjing City

1

2

4

3

5

Summer (High volatility)

6

Autumn (Low volatility)

Fig. 7 Electricity consumption patterns during the summer and autumn in Nanjing City.

Table 9 Descriptive statistics of each group in Nanjing City during the summer. 1

2

3

4

5

6

Prop.

0.16

0.08

0.15

0.13

0.20

0.28

Mean

5.42

11.03

4.16

8.00

6.65

1.96

Std.

3.13

3.38

1.33

3.74

1.60

0.34

27

ACCEPTED MANUSCRIPT

The general trend of electricity consumption in autumn is flat (see Fig. 7) in Nanjing, and it has a little upward tendency. Maximal average electricity consumption is 10.10 kWh. Most families

RI PT

consume 2.7 (about 26%) and 5.45 (about 26%) kWh in Nanjing.

Table 10 Descriptive statistics of each group in Nanjing City during the autumn. 2

3

4

Prop.

0.26

0.20

0.12

0.04

Mean

5.45

1.21

10.10

7.33

Std.

0.40

0.14

1.40

0.68

5

6

0.20

0.26

4.00

2.70

0.29

0.19

M AN U

SC

1

Electricity consumption patterns of spring and autumn are similar from the perspective of volatility, and the load profiles are relatively smooth. Electricity consumption patterns vary according to seasons. Seasonal consumption patterns differences are obvious, and the result is

TE D

meaningful for policy makers. Different power supply schemes should be made in different

EP

seasons, and also the elasticity of load profile is a key factor for making price strategies.

3.6 Temperature sensitivity of electricity consumption

AC C

Household electricity consumption patterns heavily depend on household behavior patterns. Temperature is a key factor which affect people’s behaviors. For example, there is a decrease in outdoor activities significantly at high temperature or in cold day, and people tend to stay at home. In this situation, air conditioner will open for cooling or heating, and it causes sharp changes in electricity consumption. Besides, volatility of electricity consumption is related to the volatility of temperature. People have regular electricity consumption patterns when the temperature is moderate. 28

ACCEPTED MANUSCRIPT Season is a natural division for temperature, the difference of temperature is obvious among seasons. This section focuses on the relationship between temperature and electricity consumption

Temperature

Summer (from 2014.06.01 to 2014.08.31)

M AN U

Summer (from 2014.06.01 to 2014.08.31)

SC

Daily electricity consumption(kwh)

Yancheng City Temperature and power load profiles

Temperature

Daily electricity consumption(kwh)

Nanjing City Temperature and power load profiles

RI PT

based on clustering and data visualization.

Fig. 8 Temperature curves and typical load profiles in Nanjing (left) and Yancheng (right).

We identify six typical load profiles in both Nanjing and Yancheng by means of clustering on the daily electricity consumption data in summer. We plot typical load profiles and temperature

TE D

curves in the same figure (see Fig. 8). The vertical line indicates the day of local highest temperature. In this way, there are three local highest temperature in Nanjing and two local highest

EP

temperature in Yancheng. And the electricity demand peaks lag behind the temperature peaks as shown in Fig. 8, and the lag period gets shorter with the number of highest temperatures increases.

AC C

In fact, the lag period is people’s reaction time to high temperature. The result is very meaningful for Power Company. It indicates that the temperature peaks are always accompanied with electricity consumption peaks. However, they do not arrive peaks at the same time, and the electricity consumption arrives peak later. Lag period is varying with the number of highest temperature weather increase. In the sense, hot weather is a good predictor for electricity consumption for Power Company. The company should prepare for electricity consumption peaks according to the regular patterns of extreme weather. 29

ACCEPTED MANUSCRIPT

Winter (from 2014.09.01 to 2014.11.30)

Temperature

RI PT

Daily electricity consumption(kwh)

Yancheng City Temperature and power load profiles

Temperature

Daily electricity consumption(kwh)

Nanjing City Temperature and power load profiles

Winter (from 2014.09.01 to 2014.11.30)

SC

Fig. 9 Temperature curves and typical load profiles in Nanjing (left) and Yancheng (right).

M AN U

The situation is similar in winter, i.e., cold weather also causes an increase in electricity consumption. There are some differences between summer and winter, since electricity consumption peaks either arrive before the cold wave or after cold wave. Besides, the influence degree of low temperature in winter on electricity consumption is less than that of high

TE D

temperature in summer.

The temperature related results make sense for Power Company, highly accurate load forecasting, especially the prediction of extreme values, is important for the power system

AC C

EP

planning and operational decisions.

3.7 Electricity consumption patterns shifting The volatility of the typical load profiles are distinguished in different seasons, and it is meaningful to make intervention strategies. The volatility of load profiles is obviously higher in winter and summer than in spring and autumn as we have pointed above. Here we analyze the electricity consumption patterns shifting that defined in section 2, and we want to know that if there are some residents who have low volatility in winter but high volatility in spring. In this way,

30

ACCEPTED MANUSCRIPT it can track change in consumption patterns. We regard Yancheng as the object of analysis for convenience of study.

RI PT

(1) Electricity consumption patterns shifting from winter to spring Table 11 is transfer matrix from winter to spring. The Std. indicates the volatility of cluster center in each group and the Prop. indicates the proportion of each group. Most high volatility

SC

groups in winter will transfer to high volatility groups in spring (see Table 11). For example, in

M AN U

winter group G6’s volatility was 2.14. It is the maximum value compared with other groups. Obviously, it belongs to high volatility groups. 43% users transferred to group G6 and 31% users transferred to group G1 in spring. Group G1 and group G6 in spring still belong to high volatility groups. The situation is similar with group G3 in winter. However, there are some low volatility

TE D

groups in winter tend to be high volatility groups in spring. We notice that the group G2’s standard deviation is 0.12 in winter, it belongs to low volatility groups. However 38% of the group G2 in winter transfer to group G3 in spring. The volatility of group G3 was 0.33 and ranks 5th in the

EP

spring. Most residents have high volatility load profiles because of the temperature and the Spring

AC C

Festival in winter, and volatility of load profiles gets small in spring and autumn due to moderate temperature. However, some residents have low volatility profiles in winter when most residents have high volatility profiles owing to the influence of temperature and holidays. A part of these residents tend to be high volatility groups in spring when most residents have low volatility profiles, such as the group G1, G2 and G4.

Table 11 Transfer matrix from summer to autumn in Yancheng City. Spring

G1

G2

G3

G4

31

G5

G6

ACCEPTED MANUSCRIPT

Winter

Std. (0.68)

Std. (0.18)

Std. (0.33)

Std. (0.77)

Std. (0.67)

Std. (0.41)

Prop. (0.20)

Prop. (0.14)

Prop. (0.25)

Prop. (0.09)

Prop. (0.03)

Prop. (0.29)

0.0464

0.1093

0.4481

0.0055

0.0000

0.3907

0.0038

0.5496

0.3817

0.0038

0.2824

0.0118

0.0118

0.4471

0.3312

0.0162

0.4198

0.0047

G1 Std. (0.57)

G2 Std. (0.13) Prop. (0.19)

Std. (1.66)

G4 Std. (0.55) Prop. (0.22)

Std. (1.05) Prop. (0.15) G6

0.0542

0.0353

0.0552

0.0097

0.4708

0.0755

0.2689

0.0660

0.1651

0.0482

0.0000

0.4277

0.1627

AC C

Prop. (0.12)

0.3072

0.2118

EP

Std. (2.14)

0.0573

0.1169

TE D

G5

M AN U

Prop. (0.06)

0.0038

SC

G3

RI PT

Prop. (0.26)

(2) Electricity consumption patterns shifting from summer to autumn Table 12 is transfer matrix from summer to autumn, we notice that the group G4’s volatility is 0.23 in summer, it obviously belongs to low volatility load profiles in summer when most residents have a high volatility load profiles. However, 32% of the group transforms to group G6 in autumn. The volatility of group G6 is 0.70 in autumn. Obviously, and group G6 belongs to high volatility load profiles users. Group G4 in summer is the group with minimum volatility. However

32

ACCEPTED MANUSCRIPT they tend to be active in autumn when most residents’ load profile tend to be smooth.

Table 12 Transfer matrix from summer to autumn in Yancheng City.

Summer

G1

G2

G3

G4

G5

G6

Std. (0.23)

Std. (0.46)

Std. (0.30)

Std. (0.10)

Std. (0.70)

Std. (0.70)

Prop. (0.31)

Prop. (0.20)

Prop. (0.19)

Prop. (0.16)

Prop. (0.02)

Prop. (0.26)

0.4706

0.0224

0.2465

0.0308

0.0028

0.2269

0.3905

0.0949

0.3658

0.0206

Std. (2.12)

SC

G1

Prop. (0.26)

0.3613

0.0292

0.0073

0.1168

0.0590

0.1003

0.0029

0.4513

0.0184

0.6066

0.0000

0.3199

0.4015

0.3650

0.0146

0.0438

0.0219

0.1500

0.1000

0.0000

0.7000

0.0500

Prop. (0.20) G3 Std. (0.94)

Prop. (0.19) G5 Std. (3.73)

0.1533

AC C

Prop. (0.10)

0.0478

0.0074

EP

Std. (0.23)

TE D

Prop. (0.24) G4

M AN U

G2 Std. (3.16)

RI PT

Autumn

G6

Std. (2.96)

0.0000

Prop. (0.01)

These two phenomena mentioned above are interesting. Specifically, about 7.22% of the high volatility users transferred to low volatility user from winter to spring. And 6.08% low volatility users transferred to high volatility groups from summer to autumn which indicates indicate there

33

ACCEPTED MANUSCRIPT exists the phenomenon of user instability, this paper define it as ‘electricity consumption patterns shifting’. There are some reasons for electricity consumption patterns shifting. For example, some houses are renting houses, and different people live in the house in different seasons that will show

RI PT

different electricity consumption patterns. This phenomenon is meaningful for policy makers. On one way, volatility is an important factor for making intervention strategies. High volatility implies high elasticity, which are advantageous to good intervention results, the researchers prefer

SC

to intervene these families that has load profiles with high volatility. However, if we select the

M AN U

target residents according to cluster analysis in winter, then the group G4 in winter is likely to be mistaken for low volatility group. In this situation, it is not regarded as an intervention family. In fact, they tend be active in spring.

TE D

4. Conclusions

This paper proposes a data mining based framework for exploring electricity consumption

EP

patterns, and a case study of Nanjing and Yancheng City, Jiangsu Province, China is presented. In the proposed framework, it analyzes electricity consumption patterns from several levels.

AC C

At the first level, typical load profiles in festivals such as the Spring Festival, the Labor Day and the National Day are explored, we observed that there are three typical load profiles during the Spring Festival, two typical load profiles during the Labor Day the National Day. At the second level, seasonal electricity consumption pattern is explored. According to the results of clustering, the volatility of load profiles in spring and autumn is lower than that in summer and winter. Besides there is a slight downward trend in spring and upward trend in in autumn of electricity consumption. 34

ACCEPTED MANUSCRIPT At the third level, the relationship between temperature and electricity demand is explored using data visualization. Temperature is a key factor that has influence on electricity consumption, and there exists a time lag between temperature peaks and electricity consumption peaks. The

RI PT

electricity consumption peak arrive one or two days later after the temperature peak. This indicates that there is a response time. The response time get short when local temperature peaks arrive once more, it means that people response to high temperature more quickly. The cold temperature also

SC

brings an increase in electricity consumption. There is a close relationship between seasons and

M AN U

electricity consumption patterns.

At the fourth level, electricity consumption pattern shifting is explored. The phenomenon is Identify by means of transfer matrix. For the convenience of purpose, two turns of seasons are selected, namely from winter to spring and from summer to autumn. During the winter, most of

TE D

families have high volatility of load profiles due to factors such as temperature and festivals. 7.22% of the high volatility families from winter to spring. Besides, from summer to autumn, for most of families, the volatility of load profiles changes from high to low, however, 6.08% families that

EP

have low volatility load profiles, tend to become active in autumn when most families have low

AC C

volatility load profiles.

The results in this paper are meaningful for Power Company. Firstly, people have different electricity consumption patterns in different festivals and seasons. For example, electricity consumption will soar due to a lot of cooking and lighting for the most families. In this case, a lot of energy-saving suggestions are useful and can be presented by means of the Internet. Secondly, the number of typical load profiles in different seasons is different. In this situation, The Power Company should make personalized intervention strategies for residents during different periods

35

ACCEPTED MANUSCRIPT from the point of view of DSM. Thirdly, the temperature is a key factor that has great influence on electricity consumption. Power Company should predict local electricity consumption peak according to the local temperature peaks. Fourthly, there exists the phenomenon of electricity

RI PT

consumption pattern shifting. Generally, there are two types of users, namely stable and unstable users. Different strategies should be applied for these two types of users.

SC

Acknowledgments

M AN U

This work is supported by the National Natural Science Foundation of China under grant nos. 71501056 and 71601063, Anhui Science and Technology Major Project under grant no. 17030901024, Hong Kong Scholars Program under grant no. 2017-167, China Postdoctoral Science Foundation under grant no. 2017M612072, and Anhui Provincial Natural Science

EP

References

TE D

Foundation Program under grant no. 1608085QG165.

AC C

Abrahamse, W., Steg, L., Vlek, C., Rothengatter, T., 2005. A review of intervention studies aimed at household energy conservation. Journal of Environmental Psychology 25(3), 273-291.

Abreu, J.M., Câmara Pereira, F., Ferrão, P., 2012. Using pattern recognition to identify habitual behavior in residential electricity consumption. Energy and Buildings 49, 479-487.

Allen, S., Dietz, T., McCright, A.M., 2015. Measuring household energy efficiency behaviors with attention to behavioral plasticity in the United States. Energy Research & Social Science 10, 133-140. Bessec, M., Fouquau, J., 2008. The non-linear link between electricity consumption and temperature in Europe: a threshold panel approach. Energy Economics 30(5), 2705-2721. Bezdek, J.C., Ehrlich, R., Full,W., 1984. FCM: the fuzzy c-means clustering algorithm.Comput. Geosci. 36

ACCEPTED MANUSCRIPT 10, 191-203. Botetzagias, I., Malesios, C., Poulou, D., 2014. Electricity curtailment behaviors in Greek households: different behaviors, different predictors. Energy Policy 69, 415-424. Cabeza, L.F., Urge-Vorsatz, D., McNeil, M.A., Barreneche, C., Serrano, S., 2014. Investigating

buildings. Renewable and Sustainable Energy Reviews 36, 188-193.

RI PT

greenhouse challenge from growing trends of electricity consumption through home appliances in

De Felice, M., Alessandri, A., Catalano, F., 2015. Seasonal climate forecasts for medium-term electricity demand forecasting. Applied Energy 137, 435-444.

SC

Faisal, Tursoy, T., Resatoglu, N.G., 2016. Energy Consumption, Electricity, and GDP Causality; The Case of Russia, 1990-2011. Procedia Economics and Finance 39, 653-659.

M AN U

Garg, A., Shukla, P.R., Maheshwari, J., Upadhyay, J., 2014. An assessment of household electricity load curves and corresponding CO2 marginal abatement cost curves for Gujarat state, India. Energy Policy 66, 568-584.

Granell, R., Axon, C.J., Wallom, D.C.H., 2015. Clustering disaggregated load profiles using a Dirichlet process mixture model. Energy Conversion and Management 92, 507-516.

TE D

Guo, Z., Zhou, K., Zhang, C., Lu, X., Chen, W., Yang, S., 2018. Residential electricity consumption behavior: Influencing factors, related theories and intervention strategies. Renewable and Sustainable Energy Reviews 81, 399-412.

EP

Hartigan, J., Wong, M. 1979. A K-means clustering algorithm. Applied Statistics 28, 100-108. Huebner, G., Shipworth, D., Hamilton, I., Chalabi, Z., Oreszczyn, T., 2016. Understanding electricity

AC C

consumption: A comparative contribution of building factors, socio-demographics, appliances, behaviours and attitudes. Applied Energy 177, 692-702.

Jain, A.K., 2010. Data clustering: 50 years beyond k-means. Pattern Recognition Letters 31(8), 651-666.

Jovanović, S., Savić, S., Bojić, M., Djordjević, Z., Nikolić, D., 2015. The impact of the mean daily air temperature change on electricity consumption. Energy 88, 604-609. Kantor, I., Rowlands, I.H., Parker, P., 2017. Aggregated and disaggregated correlations of household electricity consumption with time-of-use shifting and conservation. Energy and Buildings 139, 326-339. Kaplowitz, M.D., Thorp, L., Coleman, K., Kwame Yeboah, F., 2012. Energy conservation attitudes, 37

ACCEPTED MANUSCRIPT knowledge, and behaviors in science laboratories. Energy Policy 50, 581-591. Kobayakawa, T., Kandpal, T.C., 2015. Analysis of electricity consumption under a photovoltaic micro-grid system in India. Solar Energy 116, 177-183. Kohonen, T. (1990). The self-organizing map. Proceedings of the IEEE, 78, 1464–1480

intervention mapping approach. Energy Policy 39(9), 5280-5286.

RI PT

Kok, G., Lo, S.H., Peters, G.-J.Y., Ruiter, R.A.C., 2011. Changing energy-related behavior: an

Lin, B., Omoju, O.E., Okonkwo, J.U., 2016. Factors influencing renewable electricity consumption in China. Renewable and Sustainable Energy Reviews 55, 687-696.

SC

Liu, G., Hao, Y., Zhou, Y., Yang, Z., Zhang, Y., Su, M., 2016. China's low-carbon industrial transformation assessment based on logarithmic mean divisia index model. Resources,

M AN U

Conservation and Recycling 108, 156-170.

Liu, L., Chen, C., Zhao, Y., Zhao, E., 2015. China‫׳‬s carbon-emissions trading: overview, challenges and future. Renewable and Sustainable Energy Reviews 49, 254-266.

McLoughlin, F., Duffy, A., Conlon, M., 2015. A clustering approach to domestic electricity load profile characterisation using smart metering data. Applied Energy 141, 190-199.

TE D

Mizutani, F., Tanaka, T., Nakamura, E., 2018. The effect of demand response on electricity consumption under the existence of the reference price effect: Evidence from a dynamic pricing experiment in Japan. The Electricity Journal 31(1), 16-22.

84-92.

EP

Narayan, P.K., Sharma, S.S., 2015. Is carbon emissions trading profitable? Economic Modelling 47,

AC C

Nilsson, A., Andersson, K., Bergstad, C.J., 2015. Energy behaviors at the office: An intervention study on the use of equipment. Applied Energy 146, 434-441.

Ntona, E., Arabatzis, G., Kyriakopoulos, G.L., 2015. Energy saving: Views and attitudes of students in secondary education. Renewable and Sustainable Energy Reviews 46, 1-15.

Obinna, U., Joore, P., Wauben, L., Reinders, A., 2017. Comparison of two residential Smart Grid pilots in the Netherlands and in the USA, focusing on energy performance and user experiences. Applied Energy 191, 264-275. Ozawa, A., Furusato, R., Yoshida, Y., 2016. Determining the relationship between a household’s lifestyle and its electricity consumption in Japan by analyzing measured electric load profiles. Energy and Buildings 119, 200-210. 38

ACCEPTED MANUSCRIPT Panapakidis, I.P., Papadopoulos, T.A., Christoforidis, G.C., Papagiannis, G.K., 2014. Pattern recognition algorithms for electricity load curve analysis of buildings. Energy and Buildings 73, 137-145. Pothitou, M., Hanna, R.F., Chalvatzis, K.J., 2017. ICT entertainment appliances’ impact on domestic

RI PT

electricity consumption. Renewable and Sustainable Energy Reviews 69, 843-853. Räsänen, T., Kolehmainen, M., 2009. Feature-Based clustering for electricity use time series data. Lecture Notes in Computer Science 5495, 401-412.

Salo, M., Nissinen, A., Lilja, R., Olkanen, E., O'Neill, M., Uotinen, M., 2016. Tailored advice and

SC

services to enhance sustainable household consumption in Finland. Journal of Cleaner Production 121, 200-207.

M AN U

Shi, J., Tang, L., Yu, L., 2015. Economic and environmental effects of coal resource tax reform in China: based on a dynamic CGE approach. Procedia Computer Science 55, 1313-1317. Tang, X., Jin, Y., McLellan, B.C., Wang, J., Li, S., 2016. China’s coal consumption declining-Impermanent

or

permanent?

Resources,

Conservation

and

Recycling,

https://doi.org/10.1016/j.resconrec.2016.07.018.

TE D

Taylor, J.W., 2010. Triple seasonal methods for short-term electricity demand forecasting. European Journal of Operational Research 204(1), 139-152. Tejani, D., Al-Kuwari, A.M.A.H., Potdar, V., 2011. Energy conservation in a smart home. 241-246.

EP

Tomás, A.S., Sumner, M., Lamparter, S., Robinson, D., 2015. On the energy demands of small appliances in homes. Energy Procedia 78, 3384-3390.

AC C

Wilson, B., 2013. Urban form and residential electricity consumption: Evidence from Illinois, USA. Landscape and Urban Planning 115, 62-71.

Yu, Z., You, J., Wong, H.-S., Han, G., 2012. From cluster ensemble to structure ensemble. Information Sciences 198, 81-99.

Zhang, C., Zhou, K., Yang, S., Shao, Z., 2017. On electricity consumption and economic growth in China. Renewable and Sustainable Energy Reviews 76, 353-368. Zhang, Y., Nie, R., Shi, R., Zhang, M., 2016. Measuring the capacity utilization of the coal sector and its decoupling with economic growth in China’s supply-side reform. Resources, Conservation and Recycling, https://doi.org/10.1016/j.resconrec.2016.09.022. Zhou, K., Yang, C., Shen, J., 2017. Discovering residential electricity consumption patterns through 39

ACCEPTED MANUSCRIPT smart-meter data mining: A case study from China. Utilities Policy 44, 73-84. Zhou, K., Yang, S., 2016. Understanding household energy consumption behavior: The contribution of

AC C

EP

TE D

M AN U

SC

RI PT

energy big data analytics. Renewable and Sustainable Energy Reviews 56, 810-819.

40