Method and system for predicting user willingness to participate in demand-side response based on semi-supervised clustering

Abstract

The present invention relates to the field of electric power customer service technology, and is a method and system for predicting user willingness to participate in demand-side response based on semi-supervised clustering. The method comprises: obtaining a data set of electricity consumption characteristics of demand-side response users and performing data preprocessing; dividing labeled data into a must-connect constraint set and a do-not-connect constraint set; inputting the data into a training semi-supervised clustering model based on pairwise constraints and KL divergence, iteratively calculating the center points of the two clusters and the membership matrix; judging the category to which the electric power user belongs based on the membership matrix to obtain a clustering result; statistically analyzing the distribution of the two types of labeled samples in the clustering result; inputting the user-related characteristic data for which the willingness to participate in demand-side response needs to be predicted into the trained model, and calculating the membership of the user sample to "willingness to participate in demand-side response". The present invention uses information entropy and KL divergence to characterize the pairwise constraint relationship, fully utilizing the existing small amount of labeled data information, and improving the accuracy of the clustering algorithm.

Description

Technical Field

[0001] The present invention relates to the field of electric power customer service technology, and in particular to a method and system for predicting user willingness to participate in demand-side response based on semi-supervised clustering. Background Art

[0002] In actual production, it's difficult to obtain labeled data on electricity users' willingness to participate, so unsupervised learning algorithms are typically used. Manually collecting a small amount of labeled sample data can be used to develop a semi-supervised learning algorithm. In the demand-side response phase, using big data technology to predict users' willingness to participate can improve the accuracy of invitations and reduce the workload of grassroots staff in prefectures and cities who conduct door-to-door visits, telephone communications, and outreach.

[0003] That is to say, in the existing technology, in order to increase the number of users participating in demand-side response, the business is often promoted through door-to-door visits by grassroots employees, telephone communications, publicity visits, etc. However, blind promotion will result in a waste of manpower, so there is an urgent need to use big data to improve the accuracy of promotion, to identify as many users as possible who are willing to participate in demand-side response, and to reduce the burden on grassroots staff without causing "harassment" to users. The method currently used is mainly to combine feature extraction with the AHP hierarchical analysis method to calculate the user's willingness to participate in demand-side response. This method is highly subjective in the weight calculation process, and accordingly, has low reliability and robustness. Summary of the Invention

[0004] In order to overcome the defects of the existing technology, the present invention proposes a method and system for predicting user participation in demand-side response based on semi-supervised clustering. The semi-supervised clustering algorithm based on pairwise constraints and KL divergence expresses the relationship information between samples of different categories through pairwise constraints, and characterizes the pairwise constraint relationship through information entropy and KL divergence, making full use of the existing small amount of label data information and greatly improving the accuracy of the clustering algorithm.

[0005] The method of the present invention is implemented by the following technical solution: a method for predicting user participation in demand-side response based on semi-supervised clustering, comprising the following steps:

[0006] Obtaining an electricity consumption feature dataset of an original demand-side response user, wherein the electricity consumption feature dataset includes labeled samples, unlabeled samples, and category labels, wherein the number of unlabeled samples is much larger than the number of labeled samples;

[0007] Performing data preprocessing on the electricity consumption characteristic data set;

[0008] For the pre-processed data, the labeled data is divided into a must-link constraint set and a cannot-link constraint set;

[0009] The data is input into the training of the semi-supervised clustering model based on pairwise constraints and KL divergence. According to the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence, the center points and membership matrix of the two clusters are iteratively calculated;

[0010] According to the membership matrix, the categories to which the power users belong are determined and the clustering results are obtained;

[0011] Count the distribution of the two types of labeled samples in the clustering results and determine the categories of the corresponding clusters;

[0012] The user-related feature data whose willingness to participate in demand-side response needs to be predicted is input into the trained semi-supervised clustering model based on pairwise constraints and KL divergence, and the membership degree of the user sample whose willingness to participate in demand-side response needs to be predicted to "willing to participate in demand-side response" is calculated.

[0013] Preferably, the electricity consumption feature data set is a data sample set X={x j |j=1,2,...,n}, including n electricity users. The data of each electricity user include monthly electricity consumption, target industry, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, daily electricity consumption fluctuation coefficient, enterprise operating conditions, enterprise operating conditions relative to the industry, and business expansion application.

[0014] Preferably, the data preprocessing process includes:

[0015] Change the data type of the numerical features in the electricity consumption characteristic data to floating point type. The numerical features include monthly electricity consumption, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, and daily electricity consumption fluctuation coefficient.

[0016] The data types of text features in the electricity consumption characteristic data are converted into digital types through discretization processing. The text features include the target industry, business operation status of the enterprise, business operation status of the enterprise relative to the industry, and business expansion registration.

[0017] Preferably, for the pre-processed data, the labeled data is divided into a must-link constraint set and a cannot-link constraint set, specifically including:

[0018] Put the labeled samples of the same cluster into the must-link constraint set;

[0019] Put the labeled samples of different clusters into the cannot-link constraint set;

[0020] In light of the demand response business context, parent companies and subsidiaries belonging to the same group are placed in the must-link constraint set.

[0021] Preferably, information entropy and KL divergence are added to the maximum entropy clustering algorithm MEC as penalty terms of the objective function, and the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence is obtained as follows:

[0022]

[0023]

[0024] Among them, γ jk is the penalty coefficient of the sample points with pairwise constraint relationship, γ jj is the penalty coefficient of the sample point belonging to the jth class; x j 、x k is the sample point, μ ik is the sample point x k The membership degree to the i-th class, μ ij is the sample point x j The membership degree of the i-th class; ML is the must-link constraint set, CL is the cannot-link constraint set; v i is the center point of the i-th cluster, c is the number of clusters, and n is the number of samples in the electricity consumption feature dataset.

[0025] Preferably, iteratively calculating the center points and membership matrices of the two clusters specifically includes:

[0026] Randomly generate the initial membership matrix U (0) ;

[0027] The initial membership matrix U (0) Substitute into the following formula to calculate the center points of the two types of samples and

[0028]

[0029] The sample center point and Substitute into the following formula and update the membership matrix to U (1) :

[0030]

[0031] Among them, v h is the center point of the h-th cluster, μ hj is the sample point x j The membership degree to the h-th class, μ hk is the sample point x k Membership degree to the hth class; γ jj is the penalty coefficient of the sample point belonging to the jth class, γ hh is the penalty coefficient of the sample point belonging to the hth category;

[0032] Iterate and loop to calculate the center points of the two types of samples and update the membership matrix until the number of iterations reaches the maximum number of iterations T, or the error between adjacent membership matrices is less than or equal to the minimum error ε;

[0033] Output the cluster center points and membership matrix calculated by the final iteration result.

[0034] The system of the present invention adopts the following technical solutions to achieve: a user participation demand-side response willingness prediction system based on semi-supervised clustering, including the following modules:

[0035] A data set acquisition module is used to obtain an electricity consumption feature data set of the original demand-side response user, wherein the electricity consumption feature data set includes labeled samples, unlabeled samples, and category labels, wherein the number of unlabeled samples is much larger than the number of labeled samples;

[0036] A preprocessing module, configured to perform data preprocessing on the electricity consumption feature data set;

[0037] The data partitioning module divides the pre-processed data into a must-link constraint set and a cannot-link constraint set;

[0038] The model training module is used to input data into the training of the semi-supervised clustering model based on pairwise constraints and KL divergence, and iteratively calculate the center points and membership matrix of the two clusters according to the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence;

[0039] The clustering module is used to determine the category to which the power user belongs based on the membership matrix and obtain the clustering results;

[0040] The category determination module is used to count the distribution of the two types of labeled samples in the clustering results and determine the category of the corresponding cluster;

[0041] The prediction module is used to input the user-related feature data whose willingness to participate in demand-side response needs to be predicted into the trained semi-supervised clustering model based on pairwise constraints and KL divergence, and calculate the membership of the user sample whose willingness to participate in demand-side response needs to be predicted to "willing to participate in demand-side response".

[0042] Compared with the prior art, the present invention has the following advantages and effects:

[0043] 1. The semi-supervised clustering algorithm based on pairwise constraints and KL divergence is introduced in this paper. KL divergence is added as a penalty term in traditional fuzzy clustering algorithms (such as maximum entropy clustering). The relationship information between samples of different categories is expressed through pairwise constraints, and the pairwise constraint relationship is characterized by information entropy and KL divergence. It makes full use of the existing small amount of label data information and greatly improves the accuracy of the clustering algorithm.

[0044] 2. The present invention improves MEC and can make full use of the existing small amount of label information, solving the defect that the MEC algorithm cannot effectively use pairwise constraint information to guide the clustering process.

[0045] 3. Compared to other semi-supervised clustering algorithms based on pairwise constraints, this paper notes that when samples with both positive and negative correlation constraints are converted into entropy, redundant terms may interfere with the clustering process (typically, the cross-entropy semi-supervised clustering algorithm based on pairwise constraints). Therefore, this paper addresses this problem by separately characterizing the sample's self-information (information entropy) and the information between samples (relative entropy, also known as KL divergence).

[0046] 4. The present invention uses a semi-supervised clustering algorithm, which can fully utilize label information while overcoming the high subjectivity of algorithms such as the Analytic Hierarchy Process (AHP). In addition, traditional unsupervised algorithms such as clustering can only output clustering results. The membership output of the present invention can be regarded as the probability of belonging to a certain category. According to the output probability sorting, users with a high response probability are prioritized for promotion, which can further improve the promotion success rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 4 is a flowchart of a method for predicting user willingness to participate in demand-side response in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0049] Example 1

[0050] Some technical means involved in this embodiment are introduced as follows:

[0051] (1) Information entropy and KL divergence

[0052] In fuzzy theory, entropy can be used to describe the membership function. The larger the entropy value of the membership function, the more disordered the category is, the greater the difference between samples in the same category, and the worse the clustering effect. Suppose the sample point x j The membership vector is μ j T =(μ 1j ,…,μ cj), c is the number of clusters, then the sample point x j The information entropy is defined as follows:

[0053]

[0054] KL divergence can measure the distance between two probability vectors. The higher the similarity between the probability vectors, the smaller the corresponding KL divergence. Sample point x j With the sample point x k The KL divergence of is defined as follows:

[0055]

[0056] Among them, μ ij is the sample point x j The membership degree to the i-th class, μ ik is the sample point x k The degree of membership to the i-th class. It is worth noting that the KL divergence is not commutative, that is, D||x j ,x k ||≠D||x k ,x j ||.

[0057] (2) Pairwise constraints

[0058] Pairwise constraints represent the class relationship between different samples, expressed as a constraint relationship between two samples. Commonly used pairwise constraints are must-link and cannot-link. A penalty term is added to the membership corresponding to the pairwise constraint to represent the pairwise constraint information. The specific definitions of must-link and cannot-link constraints are as follows:

[0059] Definition 1: Must-link set ML = {(x j ,x k )|j<k}, if (x j ,x k )∈ML, then x j and x k Must belong to the same category, also called x j and x k There is a positive association constraint.

[0060] Definition 2: Cannot-link set CL = {(x j ,x k )|j<k}, if (x j ,x k )∈CL, then x j and x k must belong to different clusters, also called xj and x k There is a negative association constraint.

[0061] When there is a positive correlation constraint relationship between any two samples, it can be regarded as a kind of internal information, and the information entropy H(x j ,x j ) to express; when there is a negative correlation constraint between any two samples, it can be regarded as a kind of external information, and the KL divergence D(x j ||x k ). Therefore, information entropy and KL divergence can well describe the pairwise constraint information between samples.

[0062] (3) Semi-supervised clustering algorithm based on pairwise constraints and KL divergence

[0063] Semi-supervised clustering algorithms based on pairwise constraints are generally based on the maximum entropy clustering algorithm. The maximum entropy clustering algorithm (MEC) is a widely used fuzzy clustering algorithm. Its principle is to use the maximum entropy criterion and add entropy as a penalty term in the objective function of the fuzzy clustering (FCM) algorithm. Its objective function is:

[0064]

[0065] Among them, γ is the constant coefficient of the non-negative penalty term, v i is the center point of cluster i, μ ij is the jth sample point x j The degree of membership to the i-th class.

[0066] This embodiment improves the semi-supervised clustering algorithm by adding information entropy and KL divergence as penalty terms of the objective function to the maximum entropy clustering algorithm MEC. The objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence is:

[0067]

[0068]

[0069] Among them, γ jk is the penalty coefficient of the sample points with pairwise constraint relationship, γ jj is the penalty coefficient of the sample point itself; μ ik is the sample point x k The membership degree of the i-th class; ML is the must-link constraint set, CL is the cannot-link constraint set; c is the number of clusters, and n is the number of samples in the electricity consumption feature dataset.

[0070] like Figure 1As shown, this embodiment provides a method for predicting user participation in demand-side response based on semi-supervised clustering, with the input being a data sample set X = {x j |j=1,2,...,n}, including n user electricity consumption characteristics; the number of clusters is 2, one of which is the users who indicated that they would definitely participate in the response during the survey phase, and the other is the users who indicated that they would definitely not participate in the response during the survey phase. The maximum number of iterations is T, and the minimum error is ε; the output is the user's willingness to participate in the demand-side response membership μ j (is the probability). Specifically, the following steps are included:

[0071] S1. Obtain the electricity consumption feature dataset of the original demand-side response user. The original electricity consumption feature dataset includes labeled samples, unlabeled samples, and category labels, where the number of unlabeled samples is much larger than the number of labeled samples.

[0072] In this embodiment, the power consumption feature data set is a data sample set X={x j |j=1,2,...,n}, including n electricity users. The data of each electricity user include monthly electricity consumption, target industry, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, daily electricity consumption fluctuation coefficient, enterprise operating conditions, enterprise operating conditions relative to the industry, business expansion registration and other user electricity consumption characteristics.

[0073] S2. Perform data preprocessing on the acquired electricity consumption characteristic data set.

[0074] In this embodiment, the data preprocessing process is specifically as follows:

[0075] De-duplicate electricity usage feature data based on user ID;

[0076] Change the data type of the numerical features in the electricity consumption characteristic data (including monthly electricity consumption, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, and daily electricity consumption fluctuation coefficient) to floating point type;

[0077] The data types of text features in the electricity consumption feature data (including target industry, enterprise operating conditions, enterprise relative industry operating conditions, and business expansion registration) are converted into digital types through discretization processing;

[0078] Users with a large number of missing feature fields (missing fields ≥ 5) will be deleted;

[0079] For users with fewer missing feature fields (less than 5 missing fields), we perform completion processing. Floating-point data (such as monthly electricity consumption, peak electricity consumption ratio, peak-valley-to-flat fluctuation coefficient, and daily electricity consumption fluctuation coefficient) are completed using the average, while discrete numeric data (such as target industry, enterprise operating conditions, enterprise relative to industry operating conditions, and business expansion registration) are completed using the mode.

[0080] Combined with the box plot, abnormal data is deleted.

[0081] S3. For the preprocessed data, divide the labeled data into a must-link constraint set and a cannot-link constraint set.

[0082] Put the label samples of the same cluster into the must-link constraint set. For example, if user A and user B both have the label of "must participate in the response" (or "must not participate in the response"), then record user A and user B into the must-link constraint set. At this time, the must-link constraint set ML = {(x A ,x B )}.

[0083] Put the label samples of different clusters into the cannot-link constraint set. For example, if user A is labeled "must participate in the response" and user C is labeled "must not participate in the response", then record user A and user C into the cannot-link constraint set. At this time, the cannot-link constraint set CL = {(x A ,x C )}.

[0084] In particular, combined with the demand response business background, the parent company and subsidiaries belonging to the same group are placed in the must-link constraint set. For example, if user A is the parent company of user D and user A’s label is “must participate in the response”, then user D is considered to be “must participate in the response” as well. User A and user D are recorded in the must-link constraint set. At this time, the must-link constraint set ML = {(x A ,x B ),(x A ,x D )}.

[0085] S4, the data processed by steps S2 and S3 are input into the semi-supervised clustering model based on pairwise constraints and KL divergence (KL-sSC) training, and according to the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence, the center points v1 and v2 of the two clusters and the membership matrix U = (μ ij ) 2*n . Specifically including:

[0086] (1) Randomly generate the initial membership matrix U (0) ;

[0087] (2) The initial membership matrix U (0) Substitute into the following formula to calculate the center points of the two types of samples and

[0088]

[0089] (3) The sample center point and Substitute into the following formula and update the membership matrix to U (1) :

[0090]

[0091] Among them, v i is the center point of cluster i, v j is the center point of the hth cluster, x j 、x k is the sample point, μ ij is the sample point x j The membership degree to the i-th class, μ ik is the sample point x k The membership degree to the i-th class, μ hj is the sample point x j The membership degree to the h-th class, μ hk is the sample point x k The membership degree of the h-th class; γ is the penalty coefficient, which is a constant parameter that can be adjusted during model training, where γ jj is the penalty coefficient of the sample point belonging to the jth class, γ jk is the penalty coefficient of the sample points with pairwise constraint relationship, γ hh is the penalty coefficient for the sample point belonging to the hth class. Usually, the penalty coefficient values ​​for samples with positive and negative pairwise constraints and samples without labels are not equal.

[0092] (4) Iterate steps (2) and (3); if the number of iterations reaches the maximum number of iterations T, or the error between adjacent membership matrices is less than or equal to the minimum error ε, then terminate the iteration loop. In this embodiment, set T = 100 and ε = 0.0001.

[0093] (5) Output the cluster center points v1, v2 and membership matrix U = (μ ij ) 2*n .

[0094] S5. According to the membership matrix output in step S4, the categories to which the n power users belong are determined to obtain a clustering result.

[0095] If the final calculated membership of user A is μ1=0.8, μ2=0.2, it is determined that user A belongs to the first category, that is, a user who has participated in a certain response.

[0096] S6. Count the distribution of the two types of labeled samples in the clustering results and determine the categories of the corresponding clusters.

[0097] If the proportion of users with certain participation response tags in Category 1 is high, then Category 1 is set as "willing to participate in demand-side response."

[0098] S7. Input the relevant feature data of users whose willingness to participate in demand-side response needs to be predicted into the trained semi-supervised clustering model (KL-sSC) based on pairwise constraints and KL divergence, and calculate the membership of the sample that needs to be predicted to "willingness to participate in demand-side response". The closer the membership value is to 1, the higher the willingness of the sample user to participate in demand-side response.

[0099] Example 2

[0100] Based on the same inventive concept as Example 1, this embodiment provides a user participation demand-side response willingness prediction system based on semi-supervised clustering, including the following modules:

[0101] A data set acquisition module is used to obtain an electricity consumption feature data set of the original demand-side response user, wherein the electricity consumption feature data set includes labeled samples, unlabeled samples, and category labels, wherein the number of unlabeled samples is much larger than the number of labeled samples;

[0102] A preprocessing module, configured to perform data preprocessing on the electricity consumption feature data set;

[0103] The data partitioning module divides the pre-processed data into a must-link constraint set and a cannot-link constraint set;

[0104] The model training module is used to input data into the training of the semi-supervised clustering model based on pairwise constraints and KL divergence, and iteratively calculate the center points and membership matrix of the two clusters according to the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence;

[0105] The clustering module is used to determine the category to which the power user belongs based on the membership matrix and obtain the clustering results;

[0106] The category determination module is used to count the distribution of the two types of labeled samples in the clustering results and determine the category of the corresponding cluster;

[0107] The prediction module is used to input the user-related feature data whose willingness to participate in demand-side response needs to be predicted into the trained semi-supervised clustering model based on pairwise constraints and KL divergence, and calculate the membership of the user sample whose willingness to participate in demand-side response needs to be predicted to "willing to participate in demand-side response".

[0108] Among them, the electricity consumption feature data set is the data sample set X = {x j|j=1,2,...,n}, including n electricity users. The data of each electricity user include monthly electricity consumption, target industry, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, daily electricity consumption fluctuation coefficient, enterprise operating conditions, enterprise operating conditions relative to the industry, and business expansion application.

[0109] The data preprocessing process of the preprocessing module includes:

[0110] Change the data type of the numerical features in the electricity consumption characteristic data to floating point type. The numerical features include monthly electricity consumption, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, and daily electricity consumption fluctuation coefficient.

[0111] The data types of text features in the electricity consumption characteristic data are converted into digital types through discretization processing. The text features include the target industry, business operation status of the enterprise, business operation status of the enterprise relative to the industry, and business expansion registration.

[0112] The partitioning process of the data partitioning module is as follows:

[0113] Put the labeled samples of the same cluster into the must-link constraint set;

[0114] Put the labeled samples of different clusters into the cannot-link constraint set;

[0115] In light of the demand response business context, parent companies and subsidiaries belonging to the same group are placed in the must-link constraint set.

[0116] Preferably, in the model training module, information entropy and KL divergence are added to the maximum entropy clustering algorithm MEC as penalty terms of the objective function, and the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence is obtained as follows:

[0117]

[0118]

[0119] Among them, γ jk is the penalty coefficient of the sample points with pairwise constraint relationship, γ jj is the penalty coefficient of the sample point belonging to the jth class; x j 、x k is the sample point, μ ik is the sample point x k The membership degree to the i-th class, μ ij is the sample point x j The membership degree of the i-th class; ML is the must-link constraint set, CL is the cannot-link constraint set; v i is the center point of the i-th cluster, c is the number of clusters, and n is the number of samples in the electricity consumption feature dataset.

[0120] The above modules of this embodiment are respectively used to implement the steps of Example 1. The detailed implementation process can be found in Example 1 and will not be described in detail here.

[0121] The above embodiments are preferred implementations of the present invention, but the implementations of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for predicting user willingness to participate in demand-side response based on semi-supervised clustering, characterized in that: The following steps are involved: Obtaining an electricity consumption feature dataset of an original demand-side response user, wherein the electricity consumption feature dataset includes labeled samples, unlabeled samples, and category labels, wherein the number of unlabeled samples is much larger than the number of labeled samples; Performing data preprocessing on the electricity consumption characteristic data set; For the pre-processed data, the labeled data is divided into a must-link constraint set and a cannot-link constraint set; The data is input into the training of the semi-supervised clustering model based on pairwise constraints and KL divergence. According to the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence, the center points and membership matrix of the two clusters are iteratively calculated; According to the membership matrix, the categories to which the power users belong are determined and the clustering results are obtained; Count the distribution of the two types of labeled samples in the clustering results and determine the categories of the corresponding clusters; The user-related feature data for which the willingness to participate in demand-side response needs to be predicted is input into a trained semi-supervised clustering model based on pairwise constraints and KL divergence. The membership degree of the user sample for which the willingness to participate in demand-side response needs to be predicted to "willing to participate in demand-side response" is calculated. Information entropy and KL divergence are added to the maximum entropy clustering algorithm MEC as penalty terms of the objective function, and the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence is obtained as follows: Among them, γ jk is the penalty coefficient of the sample points with pairwise constraint relationship, γ jj is the penalty coefficient of the sample point belonging to the jth class; x j 、x k is the sample point, μ ik is the sample point x k The membership degree to the i-th class, μ ij is the sample point x j The membership degree of the i-th class; ML is the must-link constraint set, CL is the cannot-link constraint set; v i is the center point of the i-th cluster, c is the number of clusters, and n is the number of samples in the electricity consumption feature dataset.

2. The prediction method according to claim 1, characterized in that The electricity consumption feature data set is a data sample set X={x j |j=1,2,...,n}, including n electricity users. The data of each electricity user include monthly electricity consumption, target industry, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, daily electricity consumption fluctuation coefficient, enterprise operating conditions, enterprise operating conditions relative to the industry, and business expansion application.

3. The prediction method according to claim 1 or 2, characterized in that: The data preprocessing process includes: Change the data type of the numerical features in the electricity consumption characteristic data to floating point type. The numerical features include monthly electricity consumption, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, and daily electricity consumption fluctuation coefficient. The data types of text features in the electricity consumption characteristic data are converted into digital types through discretization processing. The text features include the target industry, business operation status of the enterprise, business operation status of the enterprise relative to the industry, and business expansion registration.

4. The prediction method according to claim 1, wherein: For the preprocessed data, the labeled data is divided into a must-link constraint set and a cannot-link constraint set, specifically including: Put the labeled samples of the same cluster into the must-link constraint set; Put the labeled samples of different clusters into the cannot-link constraint set; In light of the demand response business context, parent companies and subsidiaries belonging to the same group are placed in the must-link constraint set.

5. The prediction method according to claim 1, wherein: Iteratively calculate the center points and membership matrices of the two clusters, including: Randomly generate the initial membership matrix U (0) ; The initial membership matrix U (0) Substitute into the following formula to calculate the center points of the two types of samples and The sample center point and Substitute into the following formula and update the membership matrix to U (1) : Among them, v h is the center point of the h-th cluster, μ hj is the sample point x j The membership degree to the h-th class, μ hk is the sample point x k Membership degree to class j; γ jj is the penalty coefficient of the sample point belonging to the jth class, γ hh is the penalty coefficient of the sample point belonging to the hth category; Iterate and loop to calculate the center points of the two types of samples and update the membership matrix until the number of iterations reaches the maximum number of iterations T, or the error between adjacent membership matrices is less than or equal to the minimum error ε; Output the cluster center points and membership matrix calculated by the final iteration result.

6. A user participation demand-side response willingness prediction system based on semi-supervised clustering, characterized by: Includes the following modules: A data set acquisition module is used to obtain an electricity consumption feature data set of the original demand-side response user, wherein the electricity consumption feature data set includes labeled samples, unlabeled samples, and category labels, wherein the number of unlabeled samples is much larger than the number of labeled samples; A preprocessing module, configured to perform data preprocessing on the electricity consumption feature data set; The data partitioning module divides the pre-processed data into a must-link constraint set and a cannot-link constraint set; The model training module is used to input data into the training of the semi-supervised clustering model based on pairwise constraints and KL divergence, and iteratively calculate the center points and membership matrix of the two clusters according to the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence; The clustering module is used to determine the category to which the power user belongs based on the membership matrix and obtain the clustering results; The category determination module is used to count the distribution of the two types of labeled samples in the clustering results and determine the category of the corresponding cluster; The prediction module is used to input the relevant characteristic data of users whose willingness to participate in demand-side response needs to be predicted into the trained semi-supervised clustering model based on pairwise constraints and KL divergence, and calculate the membership degree of the user sample whose willingness to participate in demand-side response needs to be predicted to "willing to participate in demand-side response"; In the model training module, information entropy and KL divergence are added to the maximum entropy clustering algorithm MEC as penalty terms of the objective function, and the objective function of the semi-supervised clustering model based on pairwise constraints and KL divergence is obtained as follows: Among them, γ jk is the penalty coefficient of the sample points with pairwise constraint relationship, γ jj is the penalty coefficient of the sample point belonging to the jth class; x j 、x k is the sample point, μ ik is the sample point x k The membership degree to the i-th class, μ ij is the sample point x j The membership degree of the i-th class; ML is the must-link constraint set, CL is the cannot-link constraint set; v i is the center point of the i-th cluster, c is the number of clusters, and n is the number of samples in the electricity consumption feature dataset.

7. The prediction system according to claim 6, characterized in that The electricity consumption feature data set is a data sample set X={x j |j=1,2,...,n}, including n electricity users. The data of each electricity user includes monthly electricity consumption, target industry, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, daily electricity consumption fluctuation coefficient, enterprise operating conditions, enterprise operating conditions relative to the industry, and business expansion application; The data preprocessing process of the preprocessing module includes: Change the data type of the numerical features in the electricity consumption characteristic data to floating point type. The numerical features include monthly electricity consumption, peak electricity consumption ratio, peak-valley-flat fluctuation coefficient, and daily electricity consumption fluctuation coefficient. The data types of text features in the electricity consumption characteristic data are converted into digital types through discretization processing. The text features include the target industry, business operation status of the enterprise, business operation status of the enterprise relative to the industry, and business expansion registration.

8. The prediction system according to claim 6, characterized in that The partitioning process of the data partitioning module is as follows: Put the labeled samples of the same cluster into the must-link constraint set; Put the labeled samples of different clusters into the cannot-link constraint set; In light of the demand response business context, parent companies and subsidiaries belonging to the same group are placed in the must-link constraint set.

Images