A method for mining individual behavior-group energy consumption impact pattern

Abstract

The present invention discloses a method for mining individual behavior-group energy consumption impact patterns, comprising the following steps: S1, converting energy consumption time series data into point cloud data; S2, performing continuous homology feature extraction and cluster analysis on energy consumption point cloud data; S3, performing association analysis for individual behavior-group energy consumption impact pattern research, the steps are as follows: S31, multi-attribute energy consumption behavior feature design; S32, group multi-attribute energy consumption behavior data collection; S33, group multi-attribute energy consumption behavior data extraction; S34, performing association analysis; S35, interpreting group energy consumption impact patterns based on association analysis results. The present invention helps to comprehensively utilize monitoring data and empirical data to quantify the impact of user behavior on building energy consumption, and also helps to improve user energy consumption behavior in a targeted manner, thereby achieving building energy conservation. The present invention can effectively and accurately mine and explain the impact of individual energy-related behaviors on group energy consumption.

Description

Technical Field

[0001] The present invention belongs to the technical field of data mining, and particularly designs an individual behavior-group energy consumption influence pattern mining method based on continuous homology feature clustering and association analysis. Background Art

[0002] Building energy consumption accounts for 32% of the world's total energy, and the resulting greenhouse gases account for 19% of total emissions. During the entire life cycle of a building, more than 80% of the energy consumption is generated during the operation phase. The energy consumption behavior of users plays a key role in the energy consumption of the building during the operation phase. According to statistics, the energy saving potential of user behavior can reach 30% in commercial buildings and 25% in residential buildings. In buildings, the internal space structure is mostly a shared model, that is, multiple individuals use the same space together to meet their daily life, study, work and other social needs through energy sharing, such as residences, dormitories, shared offices, conference rooms, classrooms, multi-purpose halls and rest areas. Energy consumption is not only affected by the decisions made by group members, but also the result of the joint actions of group members. Therefore, energy consumption is largely a group phenomenon. Fully tapping the energy-saving potential of user behavior is crucial to promoting energy conservation and emission reduction in buildings. Fully understanding the characteristics of users' energy-using behavior and its role in energy consumption is a prerequisite for realizing behavioral energy-saving potential.

[0003] As more and more building energy consumption-related data are collected and stored, the use of data mining technology to analyze building energy consumption data and provide a basis for decision-making for energy conservation has attracted more and more attention from domestic and foreign scholars. Yu et al. (2013) proposed a theoretical framework for building energy consumption analysis based on data mining, explored how to use different data mining algorithms, such as clustering, classification and correlation analysis, to predict energy demand, identify energy consumption influencing factors and optimize energy consumption operation, and used the collected data to verify the application of the theoretical framework, and found that data mining technology has good applicability in energy consumption analysis. In order to improve the energy efficiency of building operation, Xiao et al. (2014) used association rules to analyze the data in the building automation management system of the highest building in Hong Kong, China, and found that data mining methods have important application value in the study of building energy efficiency improvement. Qing Xiaoxia, Xiao Dan, et al. (2012) used cluster analysis methods to analyze energy consumption patterns. In terms of user behavior, the application of data mining can be roughly divided into two categories. The first category is to directly analyze user behavior data and obtain user behavior patterns. The second is to indirectly study user behavior patterns by mining energy use or environmental data. In 2021, Khaniet al. developed a data mining-based framework for the first time to study the relationship between occupants’ energy behavior and energy intensity. The difference between the target energy intensity and the actual energy intensity was used to identify residents’ wasteful / energy-saving behaviors in different areas of the building.

[0004] However, on the one hand, the energy consumption pattern analysis based on data mining in the past often ignores the characteristics of long duration and single characteristics of energy consumption data. It is difficult to actually mine specific patterns of energy consumption by directly using conventional clustering methods on energy consumption time series data. On the other hand, existing studies mainly carry out data mining analysis based on building energy consumption data and energy use behavior data respectively, and there are few methods that use data mining technology to establish a direct quantitative relationship between user energy use behavior and energy consumption. Especially in a group environment, although energy consumption is the result of the behavior of all residents in the building, existing studies only measure the behavior of one or some residents. Resident behavior is simplified to "average" or several "typical" patterns to represent group-level behavior, without considering the individual differences in user energy use behavior, resulting in energy use behavior that cannot well explain energy consumption. Therefore, it is necessary to develop clustering algorithms for the characteristics of energy consumption time series data, and further develop information mining methods and processes for individual behavior-group energy consumption impact patterns. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for mining individual behavior-group energy consumption impact patterns based on continuous homology feature clustering and association analysis, which is suitable for explaining group energy consumption based on individual behavior. It helps to quantify the impact of user behavior on building energy consumption with empirical data, and also helps to improve user energy consumption behavior in a targeted manner, thereby achieving building energy saving. The present invention can effectively and accurately mine and explain the impact of individual energy-related behaviors on group energy consumption.

[0006] The object of the present invention is achieved through the following technical solution: a method for mining individual behavior-group energy consumption influence pattern, comprising the following steps:

[0007] S1. Converting energy consumption time series data into point cloud data includes the following sub-steps:

[0008] S11. Obtain group energy consumption data, recorded as N is the number of energy consumption data; among them, Y i =(Y i1 ,Y i2 ,...,Y im ) indicates that the energy consumption data of the i-th group has m energy consumption amplitude sample points;

[0009] S12, for Y i Perform Hilbert transform to obtain the amplitude envelope time series X i =(X i1 ,X i2 ,...,X im ), let X i =X i(t) ;

[0010] S13, amplitude envelope time series X i =X i (t) Delay embedding is performed to obtain the point cloud space X i (t) = {X i (t),X i (t+τ),X i (t+2τ),...,X i [t+(d-1)τ]}; where τ is the time delay and d is the embedding dimension of the point cloud space; finally, we obtain the N (t) Point cloud dataset C composed of;

[0011] S2, continuous coherent feature extraction and cluster analysis of energy consumption point cloud data;

[0012] The continuous homology feature extraction method is: construct Rips complex for N groups of energy consumption point cloud data C, and obtain the corresponding N continuous homology graphs PD i And continuous bar code PB i ; and extract the following vectorized features:

[0013] 1) Wasserstein distance;

[0014] 2) The number of 0D barcodes;

[0015] 3) The length of the second longest 0-dimensional barcode;

[0016] 4) Number of 1D barcodes;

[0017] 5) The longest length of a 1D barcode;

[0018] Finally, we get N groups of corresponding energy consumption time series data. The continuous homogeneity feature in,

[0019] A i =(A i1 ,A i2 ,...,A i5 ) represents the five features extracted from the i-th group of data;

[0020] The clustering analysis method is: Apply the K-means algorithm to transform the vectors of the new sample space Divide into k clusters and obtain each energy consumption time series data Y i Category C i ∈[1,2,...,k];

[0021] S3. Conduct a correlation analysis on the individual behavior-group energy consumption impact model. The steps are as follows:

[0022] S31. Multi-attribute energy consumption behavior feature design: Assuming there is energy consumption time series data The energy consumption behavior characteristics of each individual are: Among them, the number of individuals corresponding to each energy consumption time series data is M, L is the number of behavioral characteristics, i = 1, 2, ..., N, j = 1, 2, ..., M;

[0023] S32. Collecting data on multi-attribute energy consumption behavior of groups: Collect individual energy consumption behavior through questionnaire surveys, and quantify the characteristic attributes of energy consumption behavior in combination with Likert scale design rules; each characteristic indicator consists of a group of statements, and each group of statements has five answers: "always", "often", "occasionally", "rarely", and "never", which are recorded as 5, 4, 3, 2, and 1 respectively.

[0024] S33. Extraction of multi-attribute energy consumption behavior data for groups: Based on the characteristics of all individual behaviors in the group, five statistical indicators, namely the maximum value, minimum value, mode, median and average value of each characteristic attribute in the group, are extracted, as shown below:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030] Where, l=1,2,...,L,i=1,2,...,N,j=1,2,...,M;

[0031] S34, perform correlation analysis: Use the grey correlation analysis method to cluster the energy consumption patterns into i ∈[1,2,...,k] and multi-attribute energy consumption behavior characteristic data Q il (max),Q il (min),Q il (mode),Q il (median),Q il (mean) for correlation analysis; finally, the group behavior characteristic index with the largest correlation coefficient is selected, namely W l =maxf(Q il ,C), f(.,.) represents correlation analysis;

[0032] S35. Explain the group energy consumption impact pattern based on the association analysis results: The correlation vector W of the group-level behavior of each behavioral feature after screening is l ,l=1,2,...,L, and perform correlation analysis with the energy consumption clustering result C again; according to the degree of correlation, rank the importance of each behavior and find out the typical energy-related behavior pattern.

[0033] The beneficial effects of the present invention are as follows: the present invention is a method for mining individual behavior-group energy consumption influence patterns based on continuous homophonic feature clustering and association analysis, which is suitable for explaining group energy consumption based on individual behavior. It helps to quantify the impact of user behavior on building energy consumption with empirical data, and also helps to improve user energy consumption behavior in a targeted manner, thereby achieving building energy saving. At the same time, based on point cloud processing, continuous homophonic feature extraction and vectorization conversion of time series energy consumption data, energy consumption patterns can be extracted and mined. The process of this method can effectively and accurately mine and explain the impact of individual energy-related behaviors on group energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The principle of continuous homology characteristics, where (a) is an example of a topological structure graph; (b) is a continuous barcode graph; (c) is a topological continuous graph;

[0035] Figure 2 This is a schematic diagram of the principle of continuous synchronization;

[0036] Figure 3 is a flow chart of the present invention;

[0037] Figure 4 This is an example diagram of energy consumption time series data in this embodiment;

[0038] Figure 5 Schematic diagram of behavior-energy consumption impact model in this embodiment. DETAILED DESCRIPTION

[0039] Continuous homology is the core concept of topological data analysis, which is used to study the characteristics of the topological structure in the data as the filtering parameters change. It involves several basic concepts such as simplex, simplicial complex, and Betti number. The k-simplex is the convex hull of k+1 affine independent points. When k∈[0,3], the k-simplex represents vertices, edges, triangles, and tetrahedrons respectively. A simplicial complex refers to a combination of simplexes under specific rules. The k-simplex generates high-dimensional complexes by gradually connecting data points, thereby showing the topological structure hidden in the original data.

[0040] Betty number β k It is an important parameter to characterize the topological structure of the complex, and is used to count the number of k-dimensional homology features of the complex B(ε). Among them, the 0-dimensional, 1-dimensional, and 2-dimensional Betti numbers β0, β1, and β2 are the numbers of connected branches, closed loops, and three-dimensional holes, respectively. Figure 1 As shown in (a), a ring unit represents a 0-dimensional connected branch and a 1-dimensional closed loop, and a curve unit represents a 0-dimensional connected branch. Therefore, the 0-dimensional Betti number β0 of this graph is 3, and the 1-dimensional Betti number β1 is 2.

[0041] Betti numbers are not enough to fully represent the topological characteristics of data. It is necessary to introduce the concept of continuous homology to study the dynamic topological structure characteristics of data. Continuous homology extracts topological features by constructing simplicial complexes of data. Based on the continuous filter value ε, a series of monotonically increasing simplicial complexes B(ε) are constructed to approximate the topological space of the original data:

[0042]

[0043] Among them, ε1≥ε2≥...ε n ,ε∈[0,1]. As the filter value ε increases, the topological characteristics of B(ε) will change, that is, a new topological structure will be generated, accompanied by the disappearance of the existing topological structure, such as connected branches, closed loops, etc. Figure 2 As shown in the figure, when the filter value is ε1, the image has 11 connected branches; when the filter value increases to ε2, the connected branches are reduced to 1, and a new closed-loop structure is generated; when the filter value increases to ε3, the closed loop disappears, and only 1 connected branch remains.

[0044] The persistent homology feature obtains the topological structure information under all ε distance parameters, and further uses the persistent barcode graph and persistence graph to visualize the topological information. The ε filter value corresponding to the generation and disappearance of the l-th k-dimensional topological structure of the complex is represented, and the barcode is a finite interval [b k,l ,d k,l ) or [b k,l ,+∞). Taking a simple graph as an example ( Figure 1 (a)), its continuous barcode diagram is as follows Figure 1 (b) shows that the horizontal axis represents the filter value ε. The three red barcodes in the vertical direction represent three 0-dimensional topological structures, corresponding to one curve and two closed loops, a total of three connected branches. 0,1 ,d 0,1 ,b 0,2 ,d 0,2 ,b 0,3 ,d 0,3 are the filter values ​​of the starting and ending points of the three connected branches. The blue barcode represents a 1D closed loop structure in the figure, corresponding to two ring units. 1,1 ,d 1,1 ,b 1,2 ,d 1,2are the ε filter values ​​of the blue barcode at the start and end of the horizontal axis. Taking the duration of the closed-loop structure as an example, it can be seen that as the ε filter value increases, the smaller circle in the upper left corner is filled first and the closed-loop structure disappears. The corresponding barcode b 1,1 ,d 1,1 The circle in the lower right corner is larger and takes longer to fill, so the corresponding barcode is b 1,2 ,d 1,2 Relatively long. Figure 1 (c) is the topological persistence diagram of the image. The red and blue dots in the diagram represent 0-dimensional and 1-dimensional topological structural units, respectively. The horizontal and vertical axes represent the corresponding filter values ​​when the topological units appear and disappear, respectively. It can be seen that the more the dots deviate from the y=x line in the diagram, the longer the period of the topological unit.

[0045] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0046] like Figure 3 As shown, a method for mining individual behavior-group energy consumption influence patterns of the present invention comprises the following steps:

[0047] S1. Convert the energy consumption time series data into point cloud data. In view of the characteristics of dense sampling points and long duration of energy consumption data, in order to better mine the pattern of this type of time series data, the energy consumption time series data needs to be converted into point cloud data; specifically, the following sub-steps are included:

[0048] S11. Obtain group energy consumption data, recorded as N is the number of energy consumption data; among them, Y i =(Y i1 ,Y i2 ,...,Y im ) indicates that the energy consumption data of the i-th group has m energy consumption amplitude sample points;

[0049] S12, for Y i Perform Hilbert transform to obtain the amplitude envelope time series X i =(X i1 ,X i2 ,...,X im ), let X i =X i (t);

[0050] S13, amplitude envelope time series X i =X i (t) Delay embedding is performed to obtain the point cloud space X i (t) = {X i (t),X i (t+τ),X i (t+2τ),...,X i[t+(d-1)τ]}; where τ is the time delay and d is the embedding dimension of the point cloud space; finally, we obtain the N (t) Point cloud dataset C composed of;

[0051] S2, continuous coherent feature extraction and cluster analysis of energy consumption point cloud data;

[0052] The continuous homology feature extraction method is: construct Rips complex for N groups of energy consumption point cloud data C, and obtain the corresponding N continuous homology graphs PD i And continuous bar code PB i ; and extract the following vectorized features:

[0053] 1) Wasserstein distance;

[0054] 2) The number of 0D barcodes;

[0055] 3) The length of the second longest 0-dimensional barcode;

[0056] 4) Number of 1D barcodes;

[0057] 5) The longest length of a 1D barcode;

[0058] Finally, we get N groups of corresponding energy consumption time series data. The continuous homogeneity feature in,

[0059] A i =(A i1 ,A i2 ,...,A i5 ) represents the five features extracted from the i-th group of data;

[0060] The clustering analysis method is: the extracted topological features can fully express the data structure characteristics of the original data space, so the conventional K-means clustering algorithm can achieve an ideal clustering effect. The K-means algorithm is used to transform the vectors of the new sample space into Divide into k clusters and obtain each energy consumption time series data Y i Category C i ∈[1,2,...,k];

[0061] S3. Conducting correlation analysis for the study of individual behavior-group energy consumption impact patterns. The purpose of this invention is to extract group-level behavior characteristics based on individual multi-attribute energy consumption behavior characteristics, and compare their impact on group energy consumption, find typical energy-related behavior patterns according to the degree of impact, and thus reveal the individual behavior-group energy consumption impact pattern. The specific steps are as follows:

[0062] S31. Design of multi-attribute energy consumption behavior characteristics: The present invention summarizes the factors that affect the user's energy consumption behavior into environmental triggers, event triggers and habit triggers, thereby designing multi-attribute energy consumption behavior characteristics, as shown in Table 1. Assume that there is energy consumption time series data N is the number of energy consumption data. The energy consumption behavior feature set of each individual is Among them, the number of individuals corresponding to each energy consumption time series data is M, L is the number of behavioral characteristics, i = 1, 2, ..., N, j = 1, 2, ..., M;

[0063] Table 1 Characteristics of multi-attribute energy consumption behavior

[0064]

[0065] S32. Collecting data on multi-attribute energy consumption behavior of groups: Collect individual energy consumption behavior through questionnaire surveys, and quantify the characteristic attributes of energy consumption behavior in combination with Likert scale design rules; each characteristic indicator consists of a group of statements, and each group of statements has five answers: "always", "often", "occasionally", "rarely", and "never", which are recorded as 5, 4, 3, 2, and 1 respectively.

[0066] S33, group-oriented multi-attribute energy consumption behavior data extraction: In order to explore the relationship between individual energy consumption behavior and energy consumption, the key point is how to use individual behavior data to determine the relationship between it and group energy consumption. The present invention proposes a method for group-oriented multi-attribute energy consumption behavior data extraction: based on the characteristics of all individual behaviors in the group, extract the maximum value, minimum value, mode, median and average of each characteristic attribute in the group, which are as follows:

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] Among them, l=1,2,...,L,i=1,2,...,N,j=1,2,...,M.

[0073] S34, perform correlation analysis: Use the grey correlation analysis method to cluster the energy consumption patterns into i ∈[1,2,...,k] and multi-attribute energy consumption behavior characteristic data Q il (max),Q il(min),Q il (mode),Q il (median),Q il (mean) for correlation analysis; finally, the group behavior characteristic index with the largest correlation coefficient is selected, that is, f(.,.) represents correlation analysis; the larger the correlation coefficient, the more representative the group behavior. For example, if there are five people in a group, each person has a different energy consumption behavior. First, the characteristic index of the group behavior index is calculated based on the individual behavior performance, and then the correlation coefficient between the group characteristic index and the energy consumption pattern is calculated by correlation analysis. Finally, according to the size of the correlation coefficient, the final group characteristics are selected from the five types of characteristic indicators. If the average value of the group characteristic (0.68) shows the greatest correlation with energy consumption, its average value can be selected as the final group behavior characteristic of the behavior characteristic.

[0074] S35. Explain the group energy consumption impact pattern based on the association analysis results: The correlation vector W of the group-level behavior of each behavioral feature after screening is l ,l=1,2,...,L, and perform correlation analysis again with the energy consumption clustering result C. According to the degree of correlation, the importance of each behavior is ranked to find out the typical energy-related behavior pattern. Finally, through two correlations, the individual-group action pattern and energy consumption behavior characteristics that affect the final energy consumption pattern are obtained, thereby revealing the individual behavior-group energy consumption influence pattern.

[0075] Taking the energy consumption of a university student dormitory as an example, the energy consumption time series is as follows: Figure 4 As shown. According to the energy consumption of each group, the energy consumption pattern mining method based on continuous homophonic feature extraction and vectorization is used to cluster and encode the energy consumption of each group. Then, a correlation analysis is conducted for the study of individual behavior-group energy consumption impact pattern. First, the multi-attribute energy consumption behavior characteristics are designed (Table 1), and the energy consumption behavior performance of each dormitory group member is collected through a questionnaire survey, and then the energy consumption of each group is collected through the building energy management system. Based on the behavioral performance and energy consumption level of each individual in the group, the method of extracting multi-attribute energy consumption behavior data for the group is used to mine the group-level behavioral characteristics. Based on the characteristics of all individual behaviors in the group, five statistical indicators such as the maximum value, minimum value, mean value, mode and median of each group-level behavioral characteristic attribute are extracted, and gray correlation analysis is performed with the group energy consumption level. According to the results of the correlation analysis (correlation degree), find the statistical indicator that best reflects the group-level behavioral characteristic attributes.

[0076] Take room NO.7604 as an example. There are five people in this group who share the dormitory. Assuming that the behaviors that affect energy consumption are {P1, P2, …, P10}, the behavior performance level of each individual in the group is shown in Table 2. The annual average cooling energy consumption of the dormitory is 1316.76Kw.h, which is generated by the behaviors of all individuals at the group level.

[0077] Table 2 User behavior in a dormitory

[0078]

[0079] First, the characteristic index of each behavior indicator is calculated based on the individual behavior performance. For example, the characteristic indexes of the behavior P10 of the group are max = 3.00, min = 2.33, mean = 2.67, mode = 2.67, and median = 2.33. Then, the correlation coefficient (i.e., gray relationship level) between the group characteristic index and the cooling energy consumption is calculated by the gray correlation analysis method, as shown in Table 3. Finally, according to the largest correlation coefficient in Table 3, the final group characteristics are selected from the five categories of characteristic indicators. For variable P1, its maximum value (0.7) shows the greatest correlation between group behavior and cooling energy consumption (the maximum value is coded with *), so its maximum value is selected as the final group-level behavior characteristic (for example, among the five students, one student is the most wasteful, then the overall behavior is most affected by the most wasteful behavior, and the entire dormitory also exhibits wasteful behavior).

[0080] Table 3 Example association analysis table

[0081]

[0082]

[0083] After extracting the group-level behavioral features of each behavioral feature, the correlation vector W l The correlation analysis was conducted again with the group energy consumption clustering results. The analysis results are as follows Figure 5 As shown, typical energy-related behaviors such as P10, P2, P4, etc. are found, thereby revealing the individual behavior-group energy consumption influence pattern.

[0084] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for mining individual behavior-group energy consumption impact patterns, characterized in that: The steps include: S1. Converting energy consumption time series data into point cloud data includes the following sub-steps: S11. Obtain group energy consumption data, recorded as N is the number of energy consumption data; among them, Y i =(Y i1 ,Y i2 ,...,Y im ) indicates that the energy consumption data of the i-th group has m energy consumption amplitude sample points; S12, for Y i Perform Hilbert transform to obtain the amplitude envelope time series Xi = (Xi1, Xi2, ..., Xim), let Xi = Xi (t); S13, amplitude envelope time series X i =X i (t) Delay embedding is performed to obtain the point cloud space X i (t) = {X i (t),X i (t+τ),X i (t+2τ),...,X i [t+(d-1)τ]}; where τ is the time delay and d is the embedding dimension of the point cloud space; finally, we obtain the N (t) Point cloud dataset C composed of; S2, continuous coherent feature extraction and cluster analysis of energy consumption point cloud data; The continuous homology feature extraction method is: construct Rips complex for N groups of energy consumption point cloud data C, and obtain the corresponding N continuous homology graphs PD i And continuous bar code PB i ; and extract the following vectorized features: 1) Wasserstein distance; 2) The number of 0D barcodes; 3) The length of the second longest 0-dimensional barcode; 4) Number of 1D barcodes; 5) The longest length of a 1D barcode; Finally, we get N groups of corresponding energy consumption time series data. The continuous homogeneity feature in, A i =(A i1 ,A i2 ,...,A i5 ) represents the five features extracted from the i-th group of data; The clustering analysis method is: Apply the K-means algorithm to transform the vectors of the new sample space Divide into k clusters and obtain each energy consumption time series data Y i Category C i ∈[1,2,...,k]; S3. Conduct a correlation analysis on the individual behavior-group energy consumption impact model. The steps are as follows: S31. Multi-attribute energy consumption behavior feature design: Assuming there is energy consumption time series data The energy consumption behavior characteristics of each individual are: Among them, the number of individuals corresponding to each energy consumption time series data is M, L is the number of behavioral characteristics, i = 1, 2, ..., N, j = 1, 2, ..., M; S32. Collecting data on multi-attribute energy consumption behavior of groups: Collect individual energy consumption behavior through questionnaire surveys, and quantify the characteristic attributes of energy consumption behavior in combination with Likert scale design rules; each characteristic indicator consists of a group of statements, and each group of statements has five answers: "always", "often", "occasionally", "rarely", and "never", which are recorded as 5, 4, 3, 2, and 1 respectively. S33. Extraction of multi-attribute energy consumption behavior data for groups: Based on the characteristics of all individual behaviors in the group, five statistical indicators, namely the maximum value, minimum value, mode, median and average value of each characteristic attribute in the group, are extracted, as shown below: Where, l=1,2,...,L,i=1,2,...,N,j=1,2,...,M; S34, perform correlation analysis: Use the grey correlation analysis method to cluster the energy consumption patterns into i ∈[1,2,...,k] and multi-attribute energy consumption behavior characteristic data Q il (max),Q il (min),Q il (mode),Q il (median),Q il (mean) for correlation analysis; finally, the group behavior characteristic index with the largest correlation coefficient is selected, namely W l =maxf(Q il ,C), f(.,.) represents correlation analysis; S35. Explain the group energy consumption impact pattern based on the association analysis results: The correlation vector W of the group-level behavior of each behavioral feature after screening is l ,l=1,2,...,L, and perform correlation analysis with the energy consumption clustering result C again; according to the degree of correlation, rank the importance of each behavior and find out the typical energy-related behavior pattern.

Images