Electricity customer credit rating management method and system based on differentiated service strategy

Abstract

The invention discloses an electricity customer credit rating management method and system based on a differentiated service strategy, particularly relates to the field of energy service, and solves the problem of cold start user credit evaluation deviation and power service resource dynamic adaptation optimization. Seamless migration of credit scoring parameters and smooth transition of user portraits are realized by introducing multi-dimensional feature similarity analysis of virtual data and real data and a time-driven weight attenuation strategy, and continuity and adaptive optimization capability of an evaluation system are ensured; meanwhile, an intelligent service matching network is constructed based on the credit rating updated in real time, and a closed-loop service management and control system driven by data feedback is formed by dynamically adjusting a prepayment mode threshold value and an electric charge overdraft elastic interval strategy, so that the credit value of a user is deeply associated with the power resource configuration efficiency, and the user experience is improved. The operation risk and the service cost of a power grid enterprise are reduced, and finally collaborative optimization ecology considering energy safety management and user behavior guidance is established.

Description

Technical Field

[0001] The present invention relates to the field of energy services, and more specifically, to a method and system for managing credit ratings of electricity users based on a differentiated service strategy. Background Art

[0002] In the credit management scenario of electricity users, the credit assessment system cannot accurately judge the risk of users who have newly installed electricity meters (such as newly connected enterprises and relocated households) through traditional models due to the lack of historical payment, electricity consumption, default behavior and other data. In the early stage, the average data of the region or industry is usually used to generalize the rating, but the actual electricity consumption behavior of new users may deviate from the group average (for example, users whose electricity bills surge after a short-term promotion but whose income does not increase), causing the system to misjudge their credit rating, resulting in no early warning when the risk of arrears surges.

[0003] In the electricity customer credit management system based on differentiated service strategy, how to accurately construct the initial credit level of cold start users under zero sample conditions (when there is no user historical data) to avoid the failure of differentiated service strategy due to data vacuum has become a core technical bottleneck that needs to be solved urgently. In order to solve the above problem, a technical solution is now provided. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method and system for managing the credit rating of electricity users based on a differentiated service strategy. By introducing multi-dimensional feature similarity analysis of virtual data and real data and a time-driven weight decay strategy, seamless migration of credit scoring parameters and smooth transition of user portraits are achieved, thereby ensuring the continuity and adaptive optimization capability of the evaluation system. At the same time, an intelligent service matching network is constructed based on real-time updated credit ratings, and a closed-loop service management and control system driven by data feedback is formed by dynamically adjusting the prepaid mode threshold and the electricity overdraft elastic range strategy, so that the user's credit value is deeply associated with the efficiency of power resource allocation, which not only reduces the operating risks and service costs of power grid companies, but also improves the consumption experience of high-credit users through a differentiated service response mechanism, and finally establishes a collaborative optimization ecosystem that takes into account both energy security management and user behavior guidance to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The method for managing credit rating of electricity users based on differentiated service strategies includes the following steps:

[0007] Based on the user installation address coordinates and industry classification codes, the power facility topology data and historical user credit data are associated to establish a heterogeneous graph network containing geographical proximity edges and industry similarity edges;

[0008] Embed a meta-learning module in the graph neural network, dynamically select the attention weight aggregation method according to the sparsity of the edges connected to each node in the heterogeneous graph network, extract the characteristics of the payment stability of high-credit users and the default characteristics of adjacent low-credit users, and generate the initial credit vector of cold-start users;

[0009] Take the initial credit vector as the conditional input, combine the peak-valley period distribution and volatility of the historical load in the target area, and use the conditional generative adversarial network to generate virtual power consumption and payment interval time series data to fill the data vacuum period in the initial stage of cold start;

[0010] When the real electricity consumption data of users accumulates to the preset time window threshold, calculate the feature similarity between the virtual data and the real data, dynamically attenuate the weight coefficient of the virtual data in the credit scoring model, and gradually replace it with the credit evaluation parameters driven by the real data;

[0011] Based on the credit scoring level optimized by migration, automatically match the prepayment mode or dynamic electricity fee amount to complete the closed-loop management of cold-start users from virtual data inference to real behavior learning.

[0012] In a preferred embodiment, a heterogeneous graph network is constructed, and the specific implementation process is as follows:

[0013] A1, Collect the installation address coordinates, industry classification codes, power facility topology data, and historical credit data of users;

[0014] A2, Construct a heterogeneous graph network containing two types of edges through the geographical proximity between the user installation address and the power facility and the similarity between the user industry classification codes; if the distance between the user and the power facility is within the preset range, a strong geographical proximity edge is established between the two; if two users belong to industries that meet the preset similarity degree, the strength of the edge is determined according to the similarity level;

[0015] A3, Combine the geographical proximity and industry similarity edges to construct a complete heterogeneous graph structure.

[0016] In a preferred embodiment, the initial credit vector of cold-start users is generated, and the specific implementation process is as follows:

[0017] B1, Introduce a meta-learning module in the framework of the graph neural network to generate the hyperparameters of the graph attention mechanism according to the edge sparsity of each user node; the sparsity refers to the ratio of the number of edges of a node to the average number of edges of all historical user nodes in the area where the corresponding node is located; if the sparsity of the node is lower than or equal to the sparsity critical value, the parameters are generated through the meta-learning module; if the sparsity of the node is greater than the sparsity critical value, the parameters are adjusted to 4 multi-head attention heads, and the scaling factor is determined by the fourth root of the node feature dimension;

[0018] B2. Select different aggregation methods according to the sparsity of node connections. If the sparsity of user nodes is less than or equal to the sparsity threshold, use the multi-head attention mechanism to aggregate the information of neighbor nodes. If the sparsity of nodes is greater than the sparsity threshold, use the weighted average aggregation method.

[0019] In a preferred embodiment, B3. Classify the credit ratings of neighbor nodes, and extract the stability characteristics of high-credit users and the default characteristics of low-credit users respectively. The stability characteristics are obtained by calculating the coefficient of variation of the user load curve, and the default characteristics are measured by calculating the number of defaults of neighbor users, and logarithmic smoothing processing is adopted;

[0020] B4. Concatenate the stability characteristics extracted from high-credit users with the default characteristics extracted from low-credit users to form a vector containing two parts of information;

[0021] Map the concatenated feature vector into a latent space through a fully connected layer;

[0022] Perform positive and negative sample contrast learning on users in the same industry; adjust the feature space through a contrast loss function so that the credit vectors of similar users are closer, while the credit vectors of dissimilar users are farther apart;

[0023] Map the features of the target user to the final initial credit vector through the mapping of two layers of fully connected layers.

[0024] In a preferred embodiment, fill the data vacuum period at the initial stage of cold start, and the specific implementation process is as follows:

[0025] C1. Perform a Flatten operation on the peak-valley period distribution matrix, convert it into a 168-dimensional vector, and concatenate it with the initial credit vector and load volatility to form a conditional feature vector with a total length of 233. The conditional feature vector is processed through a fully connected layer to reduce the dimension to 128 to obtain the final conditional embedding as the conditional input of the generator;

[0026] C2. The generator uses the input conditional feature vector and noise vector to generate virtual electricity consumption sequences and payment interval sequences through a stacked temporal convolutional network;

[0027] C3. The discriminator judges the authenticity of the electricity consumption sequence and the payment interval sequence by dividing into two channels respectively, and uses the Wasserstein loss function to stabilize the training process;

[0028] C4. Adopt a dual-training strategy of a pre-training stage and a joint-training stage, and optimize the generator and discriminator in stages to ensure that the generator can effectively learn the regional electricity consumption pattern and generate realistic electricity consumption and payment interval sequences;

[0029] C5. After the generated data is trained, post - processing and verification are carried out to ensure that the generated virtual power consumption and payment interval data are physically reasonable and consistent with the regional load characteristics.

[0030] In a preferred embodiment, it is gradually replaced by credit evaluation parameters driven by real data, and the specific implementation process is as follows:

[0031] D1. When the log time span of the user reaches a predetermined number of days and the total amount of collected data exceeds the rated number of items, the calculation of feature similarity is triggered; before that, the weight of the virtual data will always remain 1.0, indicating that the virtual data dominates in this stage;

[0032] D2. The similarity is evaluated by calculating the multi - dimensional feature similarity between the virtual data and the real data; through the similarity metrics of multiple dimensions, each similarity metric obtains the matching degree between the virtual data and the real data through weighted average;

[0033] D3. According to the cumulative number of days t of the data and the similarity score S, the specific formula for calculating the weight w(t, S) of the virtual data is:

[0034]

[0035] , where T win is the time window threshold, and T trans is the maximum number of days in the weight transition period, is an exponent adaptively adjusted according to the similarity score S, controlling the attenuation rate.

[0036] In a preferred embodiment, D4. Gradually switch the input of the credit scoring model from virtual data to real data; for each feature, calculate the weighted fusion value so that the features of the virtual data and the real data are weighted and fused according to the weight. When the weight of the virtual data drops below the rated value, start the incremental learning mechanism and use the latest real data to online - train the model, so that the model gradually transitions to a state completely dependent on real data; when the weight of the virtual data completely decays to 0, it means that the model has completely migrated from virtual data to real data, and the generated data will no longer affect the input features of the credit scoring model;

[0037] D5. During the migration process of the virtual data, if the attenuation rate of the virtual data weight changes abnormally, call the conditional generative adversarial network to regenerate the virtual data sequence according to the latest real data to ensure that the generated data conforms to the actual situation and meets the quality requirements.

[0038] In a preferred embodiment, the user learns from virtual data inference to real behavior, and the specific implementation process is as follows:

[0039] E1. By setting dynamic credit scoring binning rules, divide credit levels according to the user's credit score and map them to corresponding business strategies;

[0040] E2. Automatically select a suitable prepayment mode based on the user's credit level and calculate the corresponding prepayment amount. The selection of the prepayment mode and the calculation of the prepayment amount are dynamically adjusted according to the user's credit level and the predicted value of electricity consumption;

[0041] E3. Calculate the user's overdraft limit according to the user's credit level and regional load fluctuation conditions; users with different credit levels correspond to different overdraft limits, and a mechanism for offsetting liquidated damages is introduced for users with low credit levels.

[0042] In a preferred embodiment, E4. Every quarter, according to the actual user data, recalibrate the percentile value of the credit scoring binning and the overdraft limit coefficient, and optimize through the gradient descent method with the goal of minimizing the bad debt rate.

[0043] The electricity customer credit level management system based on the differentiated service strategy includes: a topology modeling module, an intelligent aggregation module, a data generation module, a feature matching module, and a closed-loop management module;

[0044] Topology modeling module: Based on the user's installation address coordinates and industry classification codes, associate the power facility topology data with the historical user credit data, construct a heterogeneous graph network including geographical proximity edges and industry similarity edges, including node and edge data, and transfer it to the intelligent aggregation module;

[0045] Intelligent aggregation module: Embed a meta-learning module in the graph neural network, dynamically select the attention weight aggregation method according to the sparsity of the node edges, extract the payment stability features of high-credit users and the default features of adjacent low-credit users, generate the initial credit vector of cold-start users, and transfer it to the data generation module;

[0046] Data generation module: Take the initial credit vector as the conditional input, combine the peak-valley period distribution and volatility of the historical load in the target area, and use the conditional generative adversarial network to generate virtual electricity consumption and payment interval time series data to fill the data vacuum period in the initial stage of cold start. The generated virtual electricity consumption and payment interval time series data are used as the input of the feature matching module;

[0047] Feature matching module: When the user's actual electricity consumption data accumulates to the preset time window threshold, calculate the feature similarity between the virtual data and the actual data, dynamically attenuate the weight coefficient of the virtual data in the credit scoring model, and gradually replace it with the credit evaluation parameters driven by the actual data, and transfer the optimized credit evaluation parameters to the closed-loop management module;

[0048] Closed-loop management module: Based on the migrated and optimized credit scoring level, automatically match the prepayment mode or dynamic electricity fee quota, and complete the closed-loop management of cold-start users from virtual data inference to real behavior learning.

[0049] Technical effects and advantages of the electricity customer credit rating management method and system based on the differential service strategy of the present invention:

[0050] The present invention combines the virtual data generation mechanism of the generative adversarial network with the transferable dynamic credit scoring model, overcomes the credit evaluation deviation caused by the lack of user historical behavior data in the cold start stage, and realizes the seamless migration of credit scoring parameters and the smooth transition of user portraits by introducing the multi-dimensional feature similarity analysis of virtual data and actual data and the time-driven weight attenuation strategy, ensuring the continuity and adaptive optimization ability of the evaluation system; At the same time, based on the real-time updated credit rating, an intelligent service matching network is constructed, and by dynamically adjusting the prepayment mode threshold and the electricity fee overdraft elasticity interval strategy, a data feedback-driven closed-loop service control system is formed, so that the user credit value is deeply associated with the power resource allocation efficiency, which not only reduces the operation risk and service cost of the power grid enterprise, but also improves the consumption experience of high-credit users through the differential service response mechanism, and finally establishes a collaborative optimization ecosystem that takes into account energy security management and user behavior guidance. Brief Description of the Drawings

[0051] Figure 1 It is a schematic flowchart of the electricity customer credit rating management method based on the differential service strategy of the present invention;

[0052] Figure 2 It is a schematic structural diagram of the electricity customer credit rating management system based on the differential service strategy of the present invention. Detailed Embodiments

[0053] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0054] Embodiment 1: Figure 1The present invention provides a method for managing the credit rating of electricity customers based on a differentiated service strategy, including:

[0055] Based on the installation address coordinates and industry classification codes of users, associate the power facility topology data with the historical user credit data to establish a heterogeneous graph network containing geographical proximity edges and industry similarity edges, where the nodes represent users and regional facilities.

[0056] Embed a meta-learning module in the graph neural network, and dynamically select the attention weight aggregation method according to the sparsity of the edges of each node in the heterogeneous graph network, extract the payment stability characteristics of high-credit users and the default characteristics of adjacent low-credit users, and generate the initial credit vector of cold-start users.

[0057] Taking the initial credit vector as the conditional input, combined with the peak-valley period distribution and volatility of the historical load in the target area, use the conditional generative adversarial network to generate virtual electricity consumption and payment interval time series data to fill the data vacuum period in the initial stage of cold start.

[0058] When the real electricity consumption data of users accumulates to the preset time window threshold, calculate the feature similarity between the virtual data and the real data, dynamically decay the weight coefficient of the virtual data in the credit scoring model, and gradually replace it with the credit evaluation parameters driven by the real data.

[0059] Based on the credit scoring level after migration optimization, automatically match the prepayment mode or dynamic electricity fee amount to complete the closed-loop management of cold-start users from virtual data inference to real behavior learning.

[0060] To establish the association relationship between cold-start users and regional power characteristics and historical credit data under zero-shot conditions, it is necessary to construct a heterogeneous graph network that quantitatively represents the topological dependence between users, facilities, and regions, providing a structured input for the subsequent graph neural network to extract implicit credit features.

[0061] The construction process of the heterogeneous graph network is as follows:

[0062] A1. Collect the installation address coordinates, industry classification codes, power facility topology data, and historical credit data of users;

[0063] Collect the installation address coordinates of users: Extract the GPS coordinate information of users from the system to locate the geographical location of each user (such as longitude and latitude) to ensure that the geographical location analysis of users can be carried out.

[0064] Collect the industry classification codes: Obtain the classification codes of the industries where each user is located (such as the standard industry classification code), and distinguish the types of industries through these codes to facilitate the industry similarity analysis in the subsequent steps.

[0065] Obtain topological data of power facilities: including information such as the geographical location, facility type, service scope, etc. of power facilities. These data help to understand the distribution of power facilities and their connection relationships with users.

[0066] Collect historical user credit data: including information such as users' historical electricity consumption, payment records, default situations, etc. These data will provide a basis for historical comparison for subsequent credit assessments.

[0067] A2. Construct a heterogeneous graph network containing two types of edges through the geographical proximity between the user installation address and power facilities and the similarity between user industry classification codes.

[0068] In the heterogeneous graph, each user and each power facility are defined as a node. The user node contains information such as its ID, address coordinates, and industry code, etc., while the facility node contains information such as the facility ID, geographical location, etc.

[0069] Based on the geographical coordinates of users and power facilities, by calculating the geographical distance between users and facilities, determine whether to establish an edge. If the distance between the user and the power facility is within the preset range, a strong geographical proximity edge is established between the two.

[0070] Establish edges between users by calculating the similarity between the industries where users are located. The similarity of industries can be calculated through the matching degree of industry classification codes. If two users belong to industries that meet the preset similarity degree, determine the strength of the edge according to the similarity level.

[0071] A3. Combine the geographical proximity and industry similarity edges to construct a complete heterogeneous graph structure, reflecting the relationships between users and facilities, and between users and users.

[0072] According to the nodes and edges defined in step A2, construct a heterogeneous graph containing geographical proximity and industry similarity edges. In the graph, the nodes represent users and facilities, and the edges represent the relationship strength between these nodes.

[0073] The weight of each edge is jointly determined by the geographical distance and industry similarity. The weight of the geographical proximity edge is calculated based on the distance between the user and the facility, and the weight of the industry similarity edge is calculated based on the similarity of the industry codes between users.

[0074] By constructing the heterogeneous graph network, the relationship between users and power facilities is effectively quantified, providing an accurate preliminary credit assessment framework for cold-start users. The combination of geographical and industry information makes up for the deficiencies of traditional credit assessment models in the absence of historical data, ensuring that cold-start users can obtain a reasonable initial credit assessment value without historical data.

[0075] Based on the constructed heterogeneous graph network, it is necessary to use the meta-learning module to dynamically adapt to the differences in node connection sparsity, fuse the payment patterns of high-credit users with the default risks of adjacent low-credit users, and generate an initial credit vector for cold-start users to fill the zero-sample data gap.

[0076] To generate the initial credit vector for cold-start users, the specific implementation process is as follows:

[0077] B1. Introduce the meta-learning module in the framework of the graph neural network and dynamically initialize the parameters of the attention mechanism according to the connection sparsity of user nodes. The processing process:

[0078] The task of the meta-learning module is to generate the hyperparameters of the graph attention mechanism according to the connection sparsity of each user node. First, calculate the sparsity of the user node. Sparsity refers to the ratio of the number of connections of a node to the average number of connections of all historical user nodes in the area where the node is located. Specifically, the higher the sparsity, the fewer neighbor nodes the node is connected to, while the lower the sparsity, the more connections the node has with neighbor nodes.

[0079] If the sparsity of the node is lower than or equal to the sparsity threshold (indicating fewer connections of the node and a sparser graph), parameters are generated through the meta-learning module, including 8 multi-head attention heads and a scaling factor, which is determined by dividing the node feature dimension by the square root of the number of attention heads.

[0080] If the sparsity of the node is greater than the sparsity threshold (indicating more connections of the node and a denser graph), the parameters are adjusted to 4 multi-head attention heads, and the scaling factor is determined by the fourth root of the node feature dimension.

[0081] The weights of the query matrix, key matrix, and value matrix in the graph neural network, as well as the bias vector, will be dynamically initialized according to the parameters generated by the meta-learning module. This initialization process provides appropriate initial weights for the graph attention layer according to different node sparsities, enabling the network to adaptively process different types of graph structures.

[0082] B2. Select different aggregation methods according to the sparsity of node connections. When the sparsity is high, use the multi-head attention mechanism to strengthen feature fusion; when the sparsity is low, use the weighted average method for feature aggregation. The processing process:

[0083] When the sparsity is high: If the sparsity of the user node is less than or equal to the sparsity critical value (high sparsity), the multi-head attention mechanism is adopted to aggregate the information of neighbor nodes. In this process, each attention head assigns weights according to the similarity between the user node and neighbor nodes, so as to weighted-aggregate the features of neighbor nodes. The results output by all attention heads are concatenated and then linearly transformed to generate the updated features of the user node.

[0084] When the sparsity is low: If the sparsity of the node is greater than the sparsity critical value (low sparsity), the weighted average aggregation method is adopted. Specifically, the weights of the geographical proximity edges and the weights of the industry similarity edges are normalized, indicating the contributions of geographical factors and industry factors in node feature aggregation. Then, the features of geographically proximate nodes and industry-similar nodes are averaged respectively and weighted according to the corresponding normalized weights to obtain the final aggregated features.

[0085] B3. By classifying the credit ratings of neighbor nodes, the stability features of high-credit users and the default features of low-credit users are extracted respectively, enhancing the association between the target user and the behaviors of surrounding users. Processing procedure:

[0086] Classify user credit ratings: The credit ratings are classified according to the user's payment punctuality rate. The payment punctuality rate of high-credit users is higher than the mean of all historical users plus one standard deviation, while that of low-credit users is lower than the mean minus one standard deviation. In this way, users are classified into two categories: high-credit and low-credit.

[0087] Extract high-credit features: For the neighbor nodes of high-credit users, calculate their stability features. The stability features are obtained by calculating the coefficient of variation (the ratio of the standard deviation to the mean) of the user load curve. Users with less load fluctuation usually show higher payment stability. By calculating the stability indicators of high-credit neighbor users, stable payment behavior features can be extracted for the target user.

[0088] Extract low-credit features: For the neighbor nodes of low-credit users, extract their default features. The default features are measured by calculating the number of defaults of neighbor users, and logarithmic smoothing is adopted to avoid the excessive influence of users with fewer default times on the aggregation result. By calculating the default features of low-credit users, potential risk signals can be identified for the target user.

[0089] B4. Fuse the stability and default features, generate the initial credit vector of the target user through non-linear mapping, and introduce contrastive learning to optimize the feature distribution. Processing procedure:

[0090] Stability and Default Feature Fusion: Concatenate the stability features extracted from high-credit users with the default features extracted from low-credit users to form a vector containing two parts of information.

[0091] This vector represents the comprehensive characteristics of the target user in terms of stability and default.

[0092]

[0093] Among them, The concatenated vector containing the stability feature s stability and the default feature s risk By concatenating these two types of features, a two-dimensional vector is formed.

[0094] Nonlinear Mapping and Latent Space Mapping: Map the concatenated feature vector into a latent space through a fully connected layer, and use a nonlinear activation function (such as GeLU) to ensure the nonlinear transformation of the features, so that the model can handle complex credit assessment tasks.

[0095] Mapped to the latent space through the fully connected layer:

[0096]

[0097] Among them:

[0098] h i is the intermediate feature vector processed by the GeLU activation function.

[0099] W p is the weight matrix used to project the concatenated feature vector into a higher-dimensional space. Through this matrix, the concatenated feature vector is transformed into a representation that the model can use. Among them, 64 represents the target dimension after projection, and 2 is the dimension from the concatenated features of stability and risk.

[0100] b p is the bias term related to the projection layer. The bias term ensures that there is an offset in the output after the linear transformation. Matches the dimension of the output after projection.

[0101] Contrastive Loss Optimization: After generating the initial credit vector, use the contrastive learning method for optimization. Through positive and negative sample contrastive learning of users in the same industry, ensure that the initial credit vectors of users from the same industry have a small gap, while the gap between users from different major industries is large. Adjust the feature space through the contrastive loss function so that the credit vectors of similar users are closer, while the credit vectors of dissimilar users are farther apart.

[0102] For users in the same industry pair (u i ,u j ), if the industry codes of the two users match to the subclass, they are regarded as positive sample pairs, and the loss function is:

[0103]

[0104] where u k is a randomly sampled negative sample (the industry major categories do not match), and the margin parameter α = 0.5 is used to balance the differences between positive and negative sample pairs.

[0105] is the contrastive loss function used to train the model. The contrastive loss function calculates the difference in feature vectors between user α and user i. If they belong to the same subclass industry, they are regarded as positive sample pairs for contrastive optimization.

[0106] u i ,u j are user indices representing the two users to be compared (i and j). The feature vectors of these users will be compared in the contrastive loss function.

[0107] Final initial credit vector: Through the mapping of two fully connected layers, the features of the target user are mapped to the final initial credit vector, outputting an initial credit vector suitable for the credit scoring model. This vector can serve as the basis for subsequent credit evaluation and prediction and be provided to the model for further learning.

[0108] Mapped to the final credit vector dimension d credit = 32 through two fully connected layers:

[0109] v i = Sigmoid(W out2 ·GeLU(W out1 h i +b out1 )+b out2 )

[0110] where

[0111] v i is the final vector output by the model, representing the initial credit vector or credit score of the target user.

[0112] W out1 ,W out2 are the weight matrices of the final fully connected layer, used to map the intermediate feature vector h i to the final output v i .

[0113] b out1 ,b out2It is the bias term related to the final fully connected layer, which helps adjust the offset of the output.

[0114] Through the Sigmoid activation function, the values of the credit vectors are compressed into the interval [0, 1], which is usually used to limit the model output within a probability range, representing the final credit score.

[0115] Finally, it constitutes the feature extraction and credit scoring process of the model. By extracting and combining the stability and default features of users, an initial credit vector is generated. This vector not only reflects the user's behavior pattern, but also ensures that the feature vectors of users from similar industries are similar and those of users from different industries are quite different through contrastive learning.

[0116] Taking the initial credit vector as the conditional input, combining the peak-valley period distribution and volatility of the historical load in the target area, and using the conditional generative adversarial network to generate virtual power consumption and payment interval time series data for filling the data vacuum period in the initial stage of cold start. The specific implementation process is as follows:

[0117] C1, by fusing the initial credit vector of cold start users with the historical load characteristics of the target area, generates the conditional feature vector of the generative adversarial network. By analyzing the period distribution and volatility of the historical load, it is ensured that the generated virtual power consumption and payment interval sequences can truly reflect the regional power consumption pattern and the credit characteristics of users. The processing process:

[0118] Initial credit vector: The initial credit vector of cold start users is obtained from step S2, representing the credit status of users. The dimension of this vector is 64, containing various features describing the user's credit level.

[0119] Peak-valley period distribution matrix: It contains the load data for 24 hours each day over a past period of time, organized by 7 days a week. Therefore, the dimension of this matrix is 24×7. Each element in the matrix represents the average load during a certain period.

[0120] Load volatility: Calculate the volatility of the historical load data to measure the degree of load change. The specific method is to measure the load fluctuation by calculating the standard deviation of the load. The larger the standard deviation value, the stronger the fluctuation.

[0121] Perform a Flatten operation on the peak-valley period distribution matrix, convert it into a 168-dimensional vector, and splice it with the initial credit vector and the load volatility to form a conditional feature vector with a total length of 233. The conditional feature vector is processed through a fully connected layer to reduce the dimension to 128, obtaining the final conditional embedding C embed , and this embedding is used as the conditional input of the generator.

[0122] C2. The generator uses the input conditional feature vector and noise vector to generate virtual electricity consumption sequences and payment interval sequences through a stacked temporal convolutional network (TCN). The design of the generator emphasizes the generation ability of time series data to ensure that the generated data conforms to the real regional load distribution and user behavior. Processing procedure:

[0123] Input layer: The generator receives two inputs, one is a random noise vector, and the other is an embedded vector from conditional features. The noise vector is sampled from a standard normal distribution and has a dimension of 100, which is used to introduce randomness into the generated data. The conditional embedded vector is a 128-dimensional vector obtained in step C1, which contains user credit features and regional load features.

[0124] Temporal feature extraction: The generator performs temporal feature extraction through a stacked causal convolutional layer. The convolutional layer captures the dependencies in the time series data through filters, and the GroupNorm and LeakyReLU activation functions are used during the processing to enhance the representation ability of features. Each convolutional operation deepens the ability to model the temporal dependencies of electricity consumption and payment intervals.

[0125] Output branch 1 (electricity consumption generation): After convolutional processing, the generator generates virtual electricity consumption sequences through a transposed convolutional layer. The transposed convolution is used to enlarge the feature map and restore the dimension of the time series data. The generated electricity consumption is processed through the Sigmoid activation function to keep the generated electricity consumption between 0 and 1, and then scaled to conform to the range of the regional historical load.

[0126] Output branch 2 (payment interval generation): The generator also generates payment interval sequences through a lightweight LSTM layer. The LSTM can capture long-term dependencies in the sequence. The generated payment intervals are ensured to be non-negative through the Softplus function, and then the output is adjusted using the GeLU activation function and scaling factor to ensure that the generated payment intervals conform to the mean and fluctuation range of the region.

[0127] C3. The design of the discriminator aims to guide the generator to improve the authenticity of the generated data by discriminating the generated data. The discriminator judges the authenticity of the electricity consumption sequence and the payment interval sequence through two channels respectively, and uses the Wasserstein loss function to stabilize the training process. Processing procedure:

[0128] Electricity consumption discrimination channel: The real electricity consumption sequence or the generated electricity consumption sequence is concatenated with the conditional embedding and processed through multiple 1D convolutional layers. These convolutional layers gradually extract important features in the electricity consumption sequence through different numbers of channels (from 64 to 1024), and finally calculate the authenticity probability of the electricity consumption sequence through a global pooling layer.

[0129] Payment interval discrimination channel: The real payment interval sequence or the generated payment interval sequence is concatenated with the conditional embedding and processed through an LSTM network. The LSTM network can capture the long-term dependencies in the payment interval sequence, and finally outputs the authenticity probability of the payment interval sequence through a fully connected layer.

[0130] Adversarial loss function: To optimize the discriminator, the Wasserstein loss function is used for training. The loss function includes the authenticity judgment of the generated data and the gradient penalty. The gradient penalty term is used to punish the situation where the gradient is too large to maintain the stability of the generation process. At the same time, the generator loss function further improves the generation ability of the generator by maximizing the misjudgment probability of the discriminator for the generated data.

[0131] C4, adopts a dual-training strategy of a pre-training stage and a joint-training stage. By optimizing the generator and discriminator in stages, it ensures that the generator can effectively learn the regional electricity consumption patterns and generate realistic electricity consumption and payment interval sequences. Processing process:

[0132] Pre-training stage: In the initial stage, only the electricity consumption generation module of the generator is trained, the payment interval generation module is frozen, the learning rate is set to 3×10 -4 , the batch size is 64, and the Adam optimizer is used for training.

[0133] Joint-training stage: Unfreeze all modules of the generator and train the generator and discriminator simultaneously. Add the sequence coherence loss and the correlation loss between electricity consumption and payment intervals to ensure a reasonable temporal dependence relationship between the generated electricity consumption and payment interval sequences. Through the joint optimization of the adversarial loss and the correlation loss, the generated data is highly matched with the real data both statistically and temporally.

[0134] C5, after the generated data is trained, post-processing and verification steps are carried out to ensure that the generated virtual electricity consumption and payment interval data are physically reasonable and consistent with the regional load characteristics. Processing process:

[0135] Electricity consumption range correction: If the generated electricity consumption exceeds the reasonable range (less than 10% of the minimum value or higher than 120% of the maximum value), the electricity consumption is corrected to ensure that it conforms to the actual regional load range.

[0136] Payment interval smoothing: The generated payment interval data is smoothed using a moving average filtering method to remove extreme fluctuations and ensure that the sum of the generated payment interval sequence matches the monthly cycle (30 days).

[0137] Distribution verification: By calculating the variance and peak-valley ratio of the generated sequence, verify whether the distribution of the generated data is consistent with the regional historical data. Ensure that the generated virtual data is consistent with the actual load data in statistical attributes such as volatility and peak-valley ratio.

[0138] When the user's real electricity consumption data accumulates to the preset time window threshold, calculate the feature similarity between the virtual data and the real data, dynamically decay the weight coefficient of the virtual data in the credit scoring model, and gradually replace it with the credit evaluation parameters driven by the real data. The specific implementation process is as follows:

[0139] D1. Set conditions for the accumulation of real data to ensure that after meeting the requirements of the preset data time span and data volume, start the transition from virtual data to real data. Set a threshold to determine when to start the process of decaying the weight of virtual data to ensure the appropriate timing of weight decay. Processing process:

[0140] It is set that the user must continuously collect real electricity consumption data for at least a predetermined number of days. This means that if the user's electricity consumption data records are continuous within 30 days, the requirements for triggering subsequent processing conditions are met.

[0141] It is required that each user collects at least one complete electricity consumption and payment record per day, and the total sample size must reach the rated number of records. This requirement ensures that the generated data has sufficient representativeness and integrity to reflect relatively real electricity consumption behavior.

[0142] Monitor the real data of each user in real time. When the log time span of the user reaches the predetermined number of days and the total amount of collected data exceeds the rated number of records, the module for calculating feature similarity is triggered. Before this, the weight of the virtual data will always remain 1.0, indicating that the virtual data dominates in this stage.

[0143] D2. Evaluate the similarity by calculating the multi-dimensional feature similarity between the virtual data and the real data. Through similarity measures in multiple dimensions, such as mean difference, volatility, peak-valley period coincidence degree, and distribution difference of payment behavior, each similarity measure obtains the matching degree between the virtual data and the real data through weighted average. The calculation of the similarity score reflects the similarity between the virtual data and the real data in multiple aspects and determines the weight decay speed of the virtual data. Processing process:

[0144] Mean difference degree: Calculate the difference between the average daily electricity consumption means of the virtual data and the real data. This difference degree is used to measure the closeness between the two in the overall level. The closer the daily means of the virtual data and the real data are, the higher the similarity between the two.

[0145] Fluctuation matching degree: Calculate the variance of the daily electricity consumption of virtual data and real data to measure the degree of fluctuation. Evaluate the consistency of fluctuations by comparing the variances of the two. If the fluctuations of virtual data and real data are similar, it indicates consistent performance in terms of load fluctuations.

[0146] Peak-valley period coincidence rate: Define the peak period set, which represents the periods when the electricity consumption is greater than 70% of the maximum regional load. Evaluate the consistency of the two during high-load periods by calculating the coincidence rate of the peak period sets of virtual data and real data. The higher the coincidence rate, the better the matching of virtual data and real data during high-load periods.

[0147] KL divergence of payment behavior: Discretize the distribution of payment intervals and calculate the KL divergence between virtual data and real data. The KL divergence measures the degree of difference between the two distributions, and understand the similarity of payment behavior by comparing the discretized distributions.

[0148] Comprehensive similarity score: Synthesize the above similarity measures through weighted average to obtain the final similarity score. Each measure (such as mean difference, volatility, peak-valley period coincidence, and KL divergence of payment behavior) will be assigned different weights according to its importance. This comprehensive score ultimately reflects the matching degree of virtual data and real data at various levels.

[0149] D3, Dynamically adjust the weight of virtual data through a non-linear decay function according to the calculated similarity score and the cumulative time of real data. This process ensures that the weight of virtual data gradually decays as the similarity and time change, thus gradually switching the model from relying on virtual data to relying on real data. Processing process:

[0150] Decay function design: Design a non-linear decay function according to the cumulative number of days t of the data and the similarity score S. This decay function controls how the weight of virtual data gradually decreases as real data accumulates. The specific formula for calculating the weight w(t, S) of virtual data is:

[0151]

[0152] where T win is the time window threshold, T trans is the maximum number of days in the weight transition period, is an exponent adaptively adjusted according to the similarity score S, which controls the decay rate.

[0153] Specifically, when the similarity score is high, the decay rate of virtual data is slow; when the similarity score is low, the decay rate accelerates.

[0154] Decay Behavior: When the similarity score is higher than the set value, the weight of the virtual data will gradually decrease, maintaining a smooth transition; if the similarity score is lower than the set value, it indicates that there is a large difference between the virtual data and the real data. At this time, an alarm will be triggered and the decay of the virtual data weight will be accelerated, and finally the weight will quickly drop to 0.

[0155] D4. Gradually switch the input of the credit scoring model from virtual data to real data. Adopt a weighted fusion method to adjust the weight of each feature according to the similarity score between the virtual data and the real data, so as to achieve a smooth transition. Processing process:

[0156] Feature-level Fusion: For each feature, calculate the weighted fusion value so that the features of the virtual data and the real data can be weighted and fused according to the weight. Specifically, as the weight decays, the features of the real data gradually dominate.

[0157] Model Retraining: When the weight of the virtual data drops below the rated value, start the incremental learning mechanism and use the latest real data to perform online training on the model, so that the model gradually transitions to a state that completely depends on the real data.

[0158] Full Switch: When the weight of the virtual data completely decays to 0, it means that the model has completely migrated from virtual data to real data, and the generated data will no longer affect the input features of the credit scoring model.

[0159] D5. During the migration process of the virtual data, if the decay rate of the virtual data weight changes abnormally, by calling the conditional generative adversarial network, regenerate the virtual data sequence according to the latest real data to ensure that the generated data meets the actual situation and quality requirements. Processing process:

[0160] Threshold Alarm: If the similarity score is lower than the set threshold, it is considered that there is a serious mismatch between the virtual data and the real data, and start the regeneration or manual review process to ensure the validity of the data.

[0161] Trend Anomaly Detection: If the decay rate of the virtual data weight changes abnormally (such as the decay rate is greater than a certain set threshold), then trigger the trend anomaly detection and enter the anomaly handling process.

[0162] Anomaly Handling Process: By calling the conditional generative adversarial network (GAN), regenerate the virtual data sequence according to the latest real data. At the same time, the manual review team will intervene to check the matching degree between the newly generated data and the real data to ensure the accuracy of the generated data. Finally, adjust the parameters of the virtual data generation to ensure the quality and validity of the virtual data.

[0163] Based on the credit score level optimized by migration, automatically match the prepayment mode, dynamic electricity charge quota or emergency repair response priority instruction, and complete the closed-loop management of cold-start users from virtual data inference to real behavior learning. The specific implementation process is as follows:

[0164] E1. By setting dynamic credit score binning rules, divide the credit level according to the user's credit score and map it to the corresponding business strategy. In the range of credit scores from 0 to 100, dynamically adjust the binning threshold according to the distribution of the user group to ensure that credit management is flexible and conforms to the latest user behavior data. The example is as follows:

[0165] First, according to the credit score distribution of users, calculate the 30th percentile and 70th percentile of the user distribution, which are used to set the demarcation points of the credit score. Specifically, the credit score of the 30th percentile is set to 65, and the credit score of the 70th percentile is set to 85. Based on this, users are divided into three categories:

[0166] Low credit (C level): The score is less than 65, and it is applicable to high-risk control strategies.

[0167] Medium credit (B level): The score is between 65 and 85, and neutral strategies are applicable.

[0168] High credit (A level): The score is greater than or equal to 85, and preferential strategies are enabled.

[0169] Demarcation point p 30 and p 70 Are updated quarterly, and the percentile values of the current user group are recalculated to adapt to the changes in the user group and ensure that the score binning can accurately reflect the current user credit distribution.

[0170] E2. Automatically select a suitable prepayment mode based on the user's credit level and calculate the corresponding prepayment amount. The selection of the prepayment mode and the calculation of the prepayment amount are dynamically adjusted according to the user's credit level and the predicted value of electricity consumption. The example is as follows:

[0171] A level: For A-level users, a post-payment mode without prepayment is adopted, provided that the user's credit score is greater than or equal to 85 for three consecutive months.

[0172] B level: For B-level users, a dynamically floating prepayment mode is adopted. The prepayment amount D is calculated according to the predicted value of electricity consumption last month min and the electricity price p, and a minimum prepayment amount D

[0173]

[0174] C level: For C-level users, a fixed-ratio prepayment mode is adopted, and the prepayment amount is That is, it is calculated by multiplying the predicted electricity consumption by 1.5 times the electricity price.

[0175] Dynamic adjustment: If the actual electricity consumption L real exceeds 20% of the predicted value, then update the predicted value to the weighted average of the original predicted value and the actual electricity consumption:

[0176]

[0177] Subsequently, recalculate the pre - stored amount.

[0178] E3. According to the user's credit rating and the regional load fluctuation situation, calculate the user's overdraft limit. Users with different credit ratings correspond to different overdraft limits, and a mechanism for offsetting liquidated damages is introduced for users with low credit ratings. The example is as follows:

[0179] For Class - A users, the overdraft limit Q credit is calculated based on the average electricity consumption in the past 30 days electricity price p, and the regional load fluctuation coefficient Δ region as follows:

[0180]

[0181] For Class - B users, the calculation of the overdraft limit does not consider the load fluctuation coefficient and is only calculated based on the user's average electricity consumption and electricity price:

[0182]

[0183] For Class - C users, when calculating the overdraft limit, if the calculation result is less than zero, then use 0 as the overdraft limit. At the same time, if the user's overdraft limit exceeds the historical liquidated damages threshold θ penalty (for example, 20 yuan), then deduct the liquidated damages from the limit:

[0184]

[0185] where k A , k B , k C are overdraft limit coefficients for different credit ratings, used to adjust the overdraft limits of users at different levels.

[0186] E4. Every quarter, according to the actual user data, recalibrate the percentile values p 30 and p 70 , as well as the overdraft limit coefficients k A , k B , k C , and optimize through the gradient - descent method, with the goal of minimizing the bad debt rate.

[0187]

[0188] Among them, β1 is the risk tolerance coefficient, which is used to adjust the weight between the bad debt rate and the overdraft limit.

[0189] It means optimizing the parameter k, and the goal is to minimize the subsequent loss function.

[0190] Badget i It represents the bad debt rate of the i-th user, usually referring to the default or arrears situation of the user. Q credit,i It represents the overdraft limit of the i-th user, which is calculated based on factors such as the user's credit rating.

[0191] The loss function (Badget i -β1·Q credit,i ) 2 is the objective function to be minimized, which is used to measure the difference between the bad debt rate and the overdraft limit. Specifically, the loss is the square of the difference between the bad debt rate and the overdraft limit (the adjusted overdraft limit). Minimizing this difference means that the optimized overdraft limit can better balance the bad debt rate.

[0192] When a user has a default behavior, a credit score penalty mechanism is initiated to reduce their credit score. For example, the penalty rules are as follows:

[0193] If the number of times the user's monthly prepaid payment is in arrears reaches 2 or more, the penalty value is 10 points.

[0194] If the user's overdraft limit overrun rate exceeds 30%, the penalty value is 5 points.

[0195] Positive feedback reinforcement: If the user has not had any arrears for 3 consecutive months and the utilization rate of their credit limit is less than 50%, then a scoring reward is triggered, and the user's credit score will increase by 10% up to 100 points.

[0196] Embodiment 2: Figure 2 It provides a power consumption customer credit rating management system based on the differential service strategy of the present invention, including: a topology modeling module, an intelligent aggregation module, a data generation module, a feature matching module, and a closed-loop management module;

[0197] Topology modeling module: Based on the user installation address coordinates and industry classification codes, associate the power facility topology data with the historical user credit data, construct a heterogeneous graph network including geographical proximity edges and industry similarity edges, including node and edge data, and transfer it to the intelligent aggregation module;

[0198] Intelligent Aggregation Module: Embed a meta-learning module in the graph neural network, dynamically select the attention weight aggregation method according to the sparsity of node connections, extract the payment stability characteristics of high-credit users and the default characteristics of adjacent low-credit users, generate the initial credit vector of cold-start users, and transmit it to the data generation module;

[0199] Data Generation Module: Using the initial credit vector as the conditional input, combined with the peak-valley period distribution and volatility of the historical load in the target area, use the conditional generative adversarial network to generate virtual power consumption and payment interval time series data to fill the data vacuum period in the initial stage of cold start. The generated virtual power consumption and payment interval time series data are used as the input of the feature matching module;

[0200] Feature Matching Module: When the user's real power consumption data accumulates to the preset time window threshold, calculate the feature similarity between the virtual data and the real data, dynamically decay the weight coefficient of the virtual data in the credit scoring model, and gradually replace it with the credit evaluation parameters driven by the real data. Transmit the optimized credit evaluation parameters to the closed-loop management module;

[0201] Closed-loop Management Module: Based on the migrated and optimized credit scoring level, automatically match the prepayment mode or dynamic electricity fee amount to complete the closed-loop management of cold-start users from virtual data inference to real behavior learning.

[0202] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data for software simulation to get a formula closest to the real situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0203] Only some exemplary embodiments of the present invention are described by way of illustration above. Undoubtedly, for those of ordinary skill in the art, without departing from the spirit and scope of the present invention, the described embodiments can be modified in various different ways. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the protection scope of the claims of the present invention.

[0204] It should be noted that in this text, if there are relational terms such as first and second, they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0205] As described above, the above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claimed rights.

Claims

1. A method for managing the credit rating of electricity users based on a differentiated service strategy, characterized in that: Includes steps: Based on the user installation address coordinates and industry classification codes, the power facility topology data and historical user credit data are associated to establish a heterogeneous graph network containing geographical proximity edges and industry similarity edges; A meta-learning module is embedded in the graph neural network. According to the sparsity of the edges between nodes in the heterogeneous graph network, the attention weight aggregation method is dynamically selected to extract the payment stability characteristics of high-credit users and the default characteristics of adjacent low-credit users, and generate the initial credit vector of cold start users. Taking the initial credit vector as the conditional input, combined with the peak and valley time distribution and volatility of the historical load in the target area, the conditional generative adversarial network is used to generate virtual electricity consumption and payment interval time series data to fill the data vacuum period at the beginning of the cold start; When the user's real electricity consumption data accumulates to the preset time window threshold, the feature similarity between the virtual data and the real data is calculated, and the weight coefficient of the virtual data in the credit scoring model is dynamically attenuated and gradually replaced with the credit assessment parameters driven by the real data; Based on the credit score level after migration optimization, the prepaid mode or dynamic electricity fee quota is automatically matched to complete the closed-loop management of cold start users from virtual data inference to real behavior learning.

2. The electricity customer credit rating management method based on differentiated service strategy according to claim 1 is characterized in that: Construct a heterogeneous graph network. The specific implementation process is as follows: A1, collect the user's installation address coordinates, industry classification code, power facility topology data and historical credit data; A2, through the geographical proximity between the user's installation address and the power facility, and the similarity between the user's industry classification codes, a heterogeneous graph network containing two types of edges is constructed; if the distance between the user and the power facility is within the preset range, a strong geographical proximity edge is established between the two; if the two users belong to industries that meet the preset similarity, the strength of the edge is determined according to the similarity; A3, combines the geographical proximity and industry similarity edges to build a complete heterogeneous graph structure.

3. The electricity customer credit rating management method based on differentiated service strategy according to claim 2 is characterized in that: Generate the initial credit vector of the cold start user. The specific implementation process is as follows: B1, introduce a meta-learning module into the framework of graph neural network, and generate the hyperparameters of graph attention mechanism according to the sparsity of the edges of each user node; sparsity refers to the ratio of the number of edges of a node to the average number of edges of all historical user nodes in the region where the corresponding node is located; if the sparsity of the node is lower than or equal to the sparsity critical value, the parameters are generated by the meta-learning module; if the sparsity of the node is greater than the sparsity critical value, the parameters are adjusted to 4 multi-head attention heads, and the scaling factor is determined by the quarter power of the node feature dimension; B2, according to the sparsity of the node edges, different aggregation methods are selected. If the sparsity of the user node is lower than or equal to the sparsity critical value, the multi-head attention mechanism is used to aggregate the information of the neighboring nodes. If the sparsity of the node is greater than the sparsity critical value, the weighted average aggregation method is used.

4. The electricity customer credit rating management method based on differentiated service strategy according to claim 3 is characterized by: B3, divide the credit level of neighbor nodes, extract the stability characteristics of high-credit users and the default characteristics of low-credit users respectively. The stability characteristics are obtained by calculating the coefficient of variation of the user load curve, and the default characteristics are measured by calculating the number of defaults of neighbor users, using logarithmic smoothing; B4, concatenate the stability features extracted from high-credit users with the default features extracted from low-credit users to form a vector containing two parts of information; The concatenated feature vector is mapped into a latent space through a fully connected layer; By comparing positive and negative samples of users in the same industry, the feature space is adjusted by comparing the loss function, so that the credit vectors of similar users are closer and the credit vectors of dissimilar users are farther apart. Through the mapping of two fully connected layers, the target user’s features are mapped to the final initial credit vector.

5. The electricity customer credit rating management method based on differentiated service strategy according to claim 4 is characterized in that: Fill the data vacuum period at the beginning of cold start. The specific implementation process is as follows: C1, Flatten the peak-valley period distribution matrix to convert it into a 168-dimensional vector, and concatenate it with the initial credit vector and load volatility to form a conditional feature vector with a total length of 233. The conditional feature vector is processed by the fully connected layer to reduce the dimension to 128, and the final conditional embedding is obtained as the conditional input of the generator; C2, the generator uses the input conditional feature vector and noise vector to generate virtual electricity consumption sequence and payment interval sequence through a stacked temporal convolutional network; C3, the discriminator judges the authenticity of the electricity consumption sequence and the payment interval sequence by dividing into two channels, and uses the Wasserstein loss function to stabilize the training process; C4, adopts a dual training strategy of pre-training stage and joint training stage, and optimizes the generator and discriminator in stages to ensure that the generator can effectively learn the regional electricity consumption pattern and generate realistic electricity consumption and payment interval sequences; C5, after the generated data is trained, it is post-processed and verified to ensure that the generated virtual electricity consumption and payment interval data are physically reasonable and consistent with the regional load characteristics.

6. The electricity customer credit rating management method based on differentiated service strategy according to claim 5 is characterized in that: Gradually replace it with credit assessment parameters driven by real data. The specific implementation process is as follows: D1: When the user's log time span reaches the predetermined number of days and the total amount of collected data exceeds the rated number of records, the feature similarity calculation is triggered; before this, the weight of the virtual data will always remain at 1.0, indicating that the virtual data occupies a dominant position in this stage; D2, evaluates the similarity between virtual data and real data by calculating the multi-dimensional feature similarity; Through the similarity measurement of multiple dimensions, each similarity measurement is weighted averaged to obtain the matching degree between virtual data and real data; D3, based on the cumulative number of days t of the data and the similarity score S; the specific formula for calculating the weight w(t,S) of the virtual data is: Among them, T win is the time window threshold, T trans is the maximum number of days for the weight transition period, It is an exponent that is adaptively adjusted according to the similarity score S and controls the decay rate.

7. The electricity customer credit rating management method based on differentiated service strategy according to claim 6 is characterized by: D4, gradually switch the input of the credit scoring model from virtual data to real data; for each feature, calculate the weighted fusion value so that the features of virtual data and real data are weighted and fused according to the weight. When the weight of the virtual data drops below the rated value, start the incremental learning mechanism and use the latest real data to train the model online, so that the model gradually transitions to a state that is completely dependent on real data; when the weight of the virtual data completely decays to 0, it means that the model has completely migrated from virtual data to real data, and the generated data will no longer affect the input features of the credit scoring model; D5, during the migration of virtual data, if the rate of virtual data weight decay changes abnormally, the conditional generative adversarial network is called to regenerate the virtual data sequence based on the latest real data to ensure that the generated data conforms to the actual situation and meets the quality requirements.

8. The electricity customer credit rating management method based on differentiated service strategy according to claim 7 is characterized in that: Users learn from virtual data to real behavior. The specific implementation process is as follows: E1, by setting dynamic credit score binning rules, divide the credit level according to the user's credit score and map it to the corresponding business strategy; E2, automatically selects the appropriate prepayment mode based on the user's credit rating and calculates the corresponding prepaid amount. The selection of the prepayment mode and the calculation of the prepaid amount are dynamically adjusted according to the user's credit rating and the predicted value of electricity consumption; E3, calculate the user's overdraft limit based on the user's credit rating and regional load fluctuations; users with different credit ratings have different overdraft limits, and a penalty deduction mechanism is introduced for users with low credit ratings.

9. The electricity customer credit rating management method based on differentiated service strategy according to claim 8 is characterized by: E4, every quarter, based on actual user data, recalibrates the percentile values ​​of the credit score bins and the overdraft limit coefficient, and optimizes them through the gradient descent method, with the goal of minimizing the bad debt rate.

10. A credit rating management system for electricity users based on a differentiated service strategy, used to implement a credit rating management method for electricity users based on a differentiated service strategy as described in any one of claims 1 to 9, characterized in that: include: Topological modeling module, intelligent aggregation module, data generation module, feature matching module and closed-loop management module; Topology modeling module: Based on the user installation address coordinates and industry classification codes, the power facility topology data is associated with the historical user credit data, and a heterogeneous graph network containing geographic proximity edges and industry similarity edges is constructed, including node and edge data, which are then passed to the intelligent aggregation module; Intelligent aggregation module: Embed a meta-learning module in the graph neural network, dynamically select the attention weight aggregation method according to the sparsity of the node edges, extract the payment stability characteristics of high-credit users and the default characteristics of adjacent low-credit users, generate the initial credit vector of the cold start user, and pass it to the data generation module; Data generation module: Taking the initial credit vector as the conditional input, combined with the peak and valley time distribution and volatility of the historical load in the target area, the conditional generative adversarial network is used to generate virtual electricity consumption and payment interval time series data to fill the data vacuum period in the initial cold start. The generated virtual electricity consumption and payment interval time series data are used as the input of the feature matching module; Feature matching module: When the user's real electricity consumption data accumulates to the preset time window threshold, the feature similarity between the virtual data and the real data is calculated, the weight coefficient of the virtual data in the credit scoring model is dynamically attenuated, and gradually replaced with the credit assessment parameters driven by the real data, and the optimized credit assessment parameters are passed to the closed-loop management module; Closed-loop management module: Based on the credit score level after migration optimization, it automatically matches the prepaid mode or dynamic electricity fee quota to complete the closed-loop management of cold start users from virtual data inference to real behavior learning.

Images