A method and system for analyzing and predicting customer payment habits

Abstract

The present invention discloses a method and system for analyzing and predicting customer payment habits, which belongs to the field of electric power operation technology. The method solves the problem that the SOM neural network in the existing analysis method is sensitive to outliers and noise, resulting in large deviation and instability in the clustering results of unbalanced customer electricity consumption data. The method comprises: acquiring customer electricity consumption data, enhancing the customer electricity consumption data based on the SMOTE algorithm, using a habit analysis model to perform overall oversampling processing on an evaluation index vector, calculating an output matrix of the evaluation index, analyzing and processing the output matrix of the evaluation index based on a differential feature extraction method, and outputting the corresponding payment habit prediction result of the customer. The habit analysis model in the present invention has high accuracy and good robustness in analyzing the portrait of customer payment behavior, and enhancing the customer electricity consumption data based on the SMOTE algorithm can effectively alleviate the problem that the habit analysis model is sensitive to outliers and noise.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric power operation, and in particular relates to a method and system for analyzing and predicting customer payment habits. Background Art

[0002] In recent years, with the rapid development of science and technology and economy, the concepts of digitalization and intelligence have penetrated into various fields. With the rapid development of the power industry, the number of power customers has also shown a geometric growth, which has brought huge challenges to the traditional electricity bill collection work. The traditional collection method is inefficient. Due to the limited business outlets, the time and place of customer payment are relatively concentrated. A large amount of cash payment has greatly increased the pressure on business outlets, and the payment efficiency has been further reduced. There are differences in the payment frequency of different customer groups. The phenomenon of arrears is common among all types of customers, but the reasons and degrees of arrears are different. In order to better meet customer needs and improve the recovery rate of electricity bills, in-depth analysis of customer payment habits and taking corresponding strategic measures are of great significance to improving service quality, optimizing resource allocation, and responding to market changes.

[0003] Chinese patent CN107909288A discloses a payment behavior analysis method based on SOM neural network clustering algorithm, including the following steps: obtaining data on basic attribute information and payment habit attribute information of all paying users in the entire area to form a data set; determining the constraints of behavior indicator parameters, customer classification quantity and connection weight in the data set, and constructing a SOM neural network; selecting a part of the samples in the data set, training each learning mode of the SOM neural network in turn, and continuously optimizing and correcting each connection weight connected to the winning neuron until the correction amount meets the set value; using the optimized SOM neural network to classify the data set, obtain a classification result for the behavior indicator parameters that meets the customer classification quantity, calculate the average value of each indicator of the data set, and obtain the payment behavior clustering result. However, in the existing analysis method, the SOM neural network is sensitive to outliers and noise, resulting in large deviations and instability in the clustering results of uneven customer electricity consumption data. In view of the above problems, we propose a customer payment habit analysis and prediction method and system. Summary of the invention

[0004] The purpose of the present invention is to address the shortcomings of the prior art and provide a method and system for analyzing and predicting customer payment habits, thereby solving the problem that the SOM neural network in the existing analysis method is sensitive to outliers and noise, resulting in large deviations and instability in the clustering results of unbalanced customer electricity consumption data.

[0005] The present invention is implemented as follows: a method for analyzing and predicting customer payment habits, the method comprising:

[0006] Obtain customer electricity consumption data, where the customer electricity consumption data includes electricity load parameters, electricity payment information, and electricity arrears information, and enhance the customer electricity consumption data based on the SMOTE algorithm to obtain enhanced electricity consumption data;

[0007] Load the electricity consumption enhancement data, analyze the customer payment behavior profile based on NILM combined with Apriori algorithm, and convert the customer payment behavior profile into an evaluation index vector;

[0008] Obtain the evaluation index vector, use the habit analysis model to oversample the evaluation index vector as a whole, calculate the output matrix of the evaluation index, analyze and process the output matrix of the evaluation index based on the differential feature extraction method, and obtain the change feature quantity caused by the change of customer payment habits. Use the change feature quantity as input, execute the habit analysis model, and output the corresponding payment habit prediction result of the customer.

[0009] In response to the prediction results of the customer's corresponding payment habits, the customer's electricity fee recovery risk analysis is performed based on the payment habit prediction results by combining the long short-term memory network and the cascading failure algorithm, and the customer's electricity fee recovery confidence is calculated;

[0010] The electricity bill collection strategy library is triggered based on the customer's electricity bill recovery confidence level, triggering the customer's corresponding personalized collection strategy instructions.

[0011] Preferably, the method for enhancing the customer electricity consumption data based on the SMOTE algorithm specifically includes:

[0012] Load customer electricity consumption data, calculate the coefficient of variation of customer electricity consumption data based on the boundary mixed sampling algorithm, and define the coefficient threshold according to the coefficient of variation of customer electricity consumption data;

[0013] Obtain a coefficient threshold of the customer's electricity usage data, and find the boundary area data and non-boundary area data of the customer's electricity usage data based on the coefficient threshold of the customer's electricity usage data to suppress the imbalance of the customer's electricity usage data;

[0014] Load the boundary area data, reconstruct the boundary area data based on the SMOTE algorithm, and obtain the boundary reconstruction set;

[0015] Load the non-boundary area data, calculate the mean of the non-boundary area data, calculate the Euclidean distance between the non-boundary area data and the mean point of the non-boundary area data based on the local anomaly factor algorithm, sort the non-boundary area data by distance, and use a preset distance threshold to delete the non-boundary area data to obtain the deleted non-boundary area data;

[0016] Integrate non-boundary area data and boundary reconstruction sets to obtain electricity consumption enhancement data;

[0017] The coefficient of variation of customer electricity consumption data is calculated using the following formula:

[0018] (1)

[0019] (2)

[0020] (3)

[0021] in, represents the coefficient of variation of customer electricity consumption data, Represents a customer electricity usage dataset The data point density is Electricity usage data set for customers Electricity consumption data of Chinese customers The number of associated data points, Electricity usage data for customers The distance to the associated data point, Represents a customer electricity usage dataset The mean of .

[0022] Preferably, the method for obtaining a customer payment behavior profile based on NILM combined with Apriori algorithm analysis specifically includes:

[0023] Load the power consumption enhancement data, identify the load characteristics in the power consumption enhancement data based on the NILM model, calculate the total load characteristics of the customers within the sampling period, and determine the customer load type information through the total load characteristics of the customers within the sampling period;

[0024] The calculation formula for the total customer load characteristics during the sampling period is as follows:

[0025] (4)

[0026] (5)

[0027] in, Indicates the total customer load characteristics during the sampling period, is the total number of electrical appliances, Respectively indicate the current electrical appliances In the sampling period Load characteristics and load correction values ​​within is the sampling interval, is the state change coefficient within the sampling interval, is the electrical reactive power within the sampling interval;

[0028] Obtain customer load type information within the sampling period, use the deformable convolutional neural network (DCN) model to extract features from power consumption enhancement data and customer load type information, and obtain a feature set of evaluation indicators;

[0029] Based on the Apriori algorithm, the inter-cluster entropy of behavioral portraits is introduced into the evaluation index feature set, and the clustering points of behavioral portraits are determined by calculating the inter-cluster entropy values ​​of behavioral portraits.

[0030] The entropy value between behavioral profile clusters is calculated by the following formula:

[0031] (6)

[0032] (7)

[0033] in, represents the entropy value between behavioral profile clusters, Characterize the inter-cluster entropy of adjacent behaviors The correlation value between is the correlation matrix between different evaluation index features, is the image cluster entropy The Euclidean distance between is the number of features of the behavior profile cluster, Represents the determinant of the correlation matrix between different evaluation indicator features, Represents the entropy between adjacent behavior profile clusters Similarity Gaussian function between them;

[0034] The hierarchical analysis method is used to calculate the feature similarity between the behavior portrait clustering points, and the portrait clustering points with convergent similarity are selected as the similarity matrix of the customer payment behavior portrait. At least one set of similarity matrices is integrated to obtain the customer payment behavior portrait;

[0035] Load the customer payment behavior profile and convert it into an evaluation indicator vector based on the Hierarchical-Softmax function.

[0036] Preferably, the habit analysis model uses an extreme learning machine as the initial model, and the initial model consists of an input layer, a hidden layer, and an output layer. When constructing the habit analysis model, a random forest model is introduced between the hidden layer and the output layer, and a maximum likelihood estimation algorithm is combined with the random forest model to form a maximum likelihood-random forest model architecture. The output layer of the initial model is frozen, and a classifier of cost-sensitive elements is used to replace the output layer of the initial model to complete the construction of the habit analysis model.

[0037] Preferably, the training method of the habit analysis model specifically includes:

[0038] Load the extreme learning machine model as the initial model of the habit analysis model, select the number of nodes, weights, and node thresholds of the input layer and hidden layer of the initial model, and preset the activation functions of the input layer and hidden layer. The activation functions of the input layer and hidden layer are sinusoidal activation function and hard threshold activation function respectively;

[0039] Set the hyperparameters and loss function of the maximum likelihood-random forest model architecture in the initial model, and define the training rounds and sample size of a single round of training for the initial model;

[0040] Based on the crawler crawling technology, 1,000 sets of training samples are randomly crawled. Based on NILM combined with the Apriori algorithm, the sample behavior portraits of the training samples are obtained, and the sample behavior portraits are converted into sample indicator vectors. The sample indicator vectors are divided into training sets and test sets.

[0041] The initial model is iteratively trained using the training set. During training, the initial model is improved based on the minimum risk Bayesian decision until the initial model converges.

[0042] Obtain a test set, use the test set as input, execute the initial model, the initial model predicts and classifies the payment habits of the test set customers, and outputs the payment habit prediction classification results. If the payment habit prediction classification results meet the preset test result threshold, output a converged habit analysis model;

[0043] If the classification result of payment habit prediction does not meet the preset test result threshold, the Adam optimizer is used to adjust the hyperparameters of the maximum likelihood-random forest model architecture in the initial model, and the number of nodes, weights, and node thresholds of the input layer and hidden layer of the initial model are changed to continue iterative training of the initial model.

[0044] Preferably, the method of executing the habit analysis model with the change feature quantity as input specifically includes:

[0045] Loading the evaluation indicator vector, inputting the evaluation indicator vector into the input layer of the habit analysis model, and the input layer performs overall oversampling processing on the evaluation indicator vector to obtain an oversampling set;

[0046] Obtain an oversampled set, input the oversampled set into the hidden layer of the habit analysis model, and obtain an output matrix of the hidden layer;

[0047] The output matrix of the hidden layer is expressed as:

[0048] (8)

[0049] in, represents the output matrix of the hidden layer, is the activation function of the hidden layer, is the input representation of the oversampled set, is the node threshold of the hidden layer, is the connection weight of the hidden layer;

[0050] The truncated singular value generalized inverse of the output matrix is ​​calculated based on the difference feature extraction method, and the truncated singular value generalized inverse is used as the change feature quantity to obtain the change feature quantity caused by the change of customer payment habits;

[0051] Load the change feature quantity and input the change feature quantity into the maximum likelihood-random forest model architecture. The maximum likelihood-random forest model calculates the dependency cost matrix based on the change feature quantity.

[0052] The classifier of cost-sensitive elements is used to calculate the average cost of classification prediction that depends on the cost matrix, and the corresponding payment habit prediction result of the customer is output.

[0053] On the other hand, the present invention also provides a customer payment habit analysis and prediction system, the customer payment habit analysis and prediction system specifically comprising:

[0054] The data collection module obtains customer electricity consumption data, including electricity load parameters, electricity payment information, and electricity arrears information. The customer electricity consumption data is enhanced based on the SMOTE algorithm to obtain enhanced electricity consumption data.

[0055] The behavior profile module is used to load the electricity consumption enhancement data, obtain the customer payment behavior profile based on NILM combined with the Apriori algorithm, and convert the customer payment behavior profile into an evaluation index vector;

[0056] The habit prediction module is used to obtain the evaluation index vector, use the habit analysis model to perform overall oversampling on the evaluation index vector, calculate the output matrix of the evaluation index, analyze and process the output matrix of the evaluation index based on the difference feature extraction method, obtain the change feature quantity caused by the change of the customer's payment habit, use the change feature quantity as input, execute the habit analysis model, and output the corresponding payment habit prediction result of the customer;

[0057] The confidence calculation module, in response to the prediction results of the payment habits of the customers, combines the long short-term memory network and the cascading failure algorithm to analyze the risk of customer electricity fee recovery based on the payment habit prediction results, and calculates the confidence of customer electricity fee recovery;

[0058] The instruction trigger module triggers the electricity bill collection strategy library based on the customer's electricity bill recovery confidence level, and triggers the customer's corresponding personalized collection strategy instruction.

[0059] The data acquisition module comprises:

[0060] Data capture unit, used for distributed collection of customer electricity consumption data;

[0061] A coefficient of variation determination unit is used to load customer electricity consumption data, calculate the coefficient of variation of the customer electricity consumption data based on a boundary mixed sampling algorithm, and define a coefficient threshold value according to the coefficient of variation of the customer electricity consumption data;

[0062] A boundary determination unit, used to obtain a coefficient threshold of the customer's electricity usage data, find the boundary area data and non-boundary area data of the customer's electricity usage data based on the coefficient threshold of the customer's electricity usage data, and suppress the imbalance degree of the customer's electricity usage data;

[0063] The data enhancement unit is used to load the boundary area data, reconstruct the boundary area data based on the SMOTE algorithm to obtain the boundary reconstruction set, load the non-boundary area data, calculate the mean of the non-boundary area data, calculate the Euclidean distance between the non-boundary area data and the mean point of the non-boundary area data based on the local anomaly factor algorithm, sort the non-boundary area data by distance, delete the non-boundary area data using a preset distance threshold to obtain the deleted non-boundary area data, integrate the non-boundary area data and the boundary reconstruction set to obtain the electricity consumption enhancement data.

[0064] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0065] In the embodiment of the present invention, the customer payment behavior portrait is obtained by combining NILM with the Apriori algorithm, and the overall oversampling of the evaluation index vector is processed by the habit analysis model to obtain the corresponding payment habit prediction result of the customer. The habit analysis model has high accuracy and good robustness in analyzing the customer payment behavior portrait, so that the distribution of the electricity consumption enhancement data can be more obvious without overfitting. And the enhancement processing of customer electricity consumption data based on the SMOTE algorithm can effectively alleviate the problem of the habit analysis model being sensitive to outliers and noise, and overcome the problem that the SOM neural network in the existing analysis method is sensitive to outliers and noise, resulting in large deviations and instability in the clustering results of unbalanced customer electricity consumption data.

[0066] In an embodiment of the present invention, customer electricity consumption data is enhanced based on the SMOTE algorithm. By dividing the minority class samples into boundary area data and non-boundary area data, and combining the SMOTE algorithm and the local anomaly factor algorithm to generate and reconstruct new synthetic samples, the number of minority class samples is increased, and the number of samples in each category is balanced, thereby avoiding the customer payment behavior profiling and habit analysis model from being overly biased towards the majority class due to data skew, and overcoming the problem of increasing the imbalance of customer electricity consumption data due to the large difference in the number of positive and negative samples in the imbalanced data.

[0067] In the embodiment of the present invention, the customer payment behavior portrait is obtained based on NILM combined with Apriori algorithm analysis, so that the characteristics of multiple dimensions can be considered at the same time, rather than being limited to the traditional single or a few characteristics. By analyzing and modeling a large amount of high-dimensional data, it is possible to more accurately identify the subtle differences between different customer groups and achieve a more refined customer payment behavior portrait construction. Based on the accurate customer payment behavior portrait, the enterprise can provide customers with personalized payment method recommendations, promotional activity recommendations, etc. The customer payment behavior portrait can also provide richer information and basis for the habit analysis model evaluation, which is conducive to the establishment of a more comprehensive habit analysis evaluation system.

[0068] In an embodiment of the present invention, a habit analysis model and a training method are provided. The habit analysis model uses an extreme learning machine as the initial model, introduces a maximum likelihood-random forest model architecture, and uses a classifier of cost-sensitive elements to replace the output layer of the initial model, thereby realizing a more refined analysis of customer payment habits, which can effectively reduce the complexity of calculations, reduce operation time, and improve the fluency of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a schematic diagram of the implementation flow of the customer payment habit analysis and prediction method provided by the present invention.

[0070] Figure 2 The figure shows a schematic diagram of the implementation process of the method for enhancing the customer electricity consumption data based on the SMOTE algorithm.

[0071] Figure 3 The figure shows a schematic diagram of the implementation process of the method for obtaining a customer payment behavior profile based on NILM combined with Apriori algorithm analysis.

[0072] Figure 4 A schematic diagram of the implementation process of the habit analysis model training method is shown.

[0073] Figure 5 The figure shows a schematic diagram of the implementation flow of the habit analysis model method using the change feature quantity as input.

[0074] Figure 6 The schematic diagram of the structure of the customer payment habit analysis and prediction system is shown. DETAILED DESCRIPTION

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0076] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0077] In the existing analysis methods, the SOM neural network is sensitive to outliers and noise, resulting in large deviations and instability in the clustering results of unbalanced customer electricity consumption data. To address the above problems, we propose a customer payment habit analysis and prediction method and system. In short, when implementing the method, the customer electricity consumption data is first obtained, and the customer electricity consumption data is enhanced based on the SMOTE algorithm to obtain enhanced electricity consumption data. Then, the customer payment behavior portrait is obtained based on the NILM combined with the Apriori algorithm analysis. The habit analysis model is used to perform overall oversampling on the evaluation index vector. The output matrix of the evaluation index is analyzed and processed based on the differential feature extraction method to obtain the change feature quantity caused by the change in the customer's payment habit, and the corresponding payment habit prediction result of the customer is output. Finally, the long short-term memory network and the cascading failure algorithm are combined to analyze the customer's electricity bill recovery risk based on the payment habit prediction result, and the customer's electricity bill recovery confidence is calculated. In the embodiment of the present invention, the customer payment behavior portrait is obtained by combining NILM with the Apriori algorithm, and the overall oversampling of the evaluation index vector is processed by the habit analysis model to obtain the corresponding payment habit prediction result of the customer. The habit analysis model has high accuracy and good robustness in analyzing the customer payment behavior portrait, so that the distribution of the electricity consumption enhancement data can be more obvious without overfitting. And the enhancement processing of customer electricity consumption data based on the SMOTE algorithm can effectively alleviate the problem of the habit analysis model being sensitive to outliers and noise, and overcome the problem that the SOM neural network in the existing analysis method is sensitive to outliers and noise, resulting in large deviations and instability in the clustering results of unbalanced customer electricity consumption data.

[0078] The embodiment of the present invention provides a method for analyzing and predicting customer payment habits. Figure 1The following is a schematic diagram of the implementation process of the method for analyzing and predicting customer payment habits. The method for analyzing and predicting customer payment habits specifically includes:

[0079] Step S10, obtaining customer electricity consumption data, and enhancing the customer electricity consumption data based on the SMOTE algorithm to obtain enhanced electricity consumption data;

[0080] In an embodiment of the present invention, customer electricity consumption data includes but is not limited to electricity load parameters, electricity payment information, and electricity arrears information, wherein the electricity load parameters may be active power, reactive power, apparent power, instantaneous load, average load, peak load, valley load, power factor, harmonic content, and three-phase imbalance of the customer's electrical equipment; the customer's electrical equipment may be an air conditioner, a kettle, a television, a computer, a terminal, an electric fan, industrial equipment, computing equipment, audio equipment, an electric rice cooker, and a microwave oven.

[0081] Step S20, loading the electricity consumption enhancement data, analyzing the customer payment behavior profile based on NILM combined with the Apriori algorithm, and converting the customer payment behavior profile into an evaluation index vector;

[0082] It should be noted that the customer payment behavior portrait is a model / knowledge graph that comprehensively describes and analyzes the various behavioral characteristics displayed by customers during the payment process. It builds a comprehensive and three-dimensional customer payment behavior portrait by collecting, organizing and mining customer payment data, such as payment amount, payment time, payment frequency, payment channel, etc., as well as other factors that may affect payment behavior, such as age, gender, region, occupation, consumption habits, etc.

[0083] Among them, the basic information dimensions of customer payment behavior portraits include the customer's age, gender, region, occupation, income level, etc. This information can help companies understand the basic attributes of different types of customers and provide a basis for subsequent behavior analysis. For example, young customers may prefer to use online payment channels, while older customers may be more accustomed to offline payment methods.

[0084] Step S30, obtain the evaluation index vector, use the habit analysis model to perform overall oversampling processing on the evaluation index vector, calculate the output matrix of the evaluation index, analyze and process the output matrix of the evaluation index based on the differential feature extraction method, obtain the change feature quantity caused by the change of the customer's payment habit, use the change feature quantity as input, execute the habit analysis model, and output the corresponding payment habit prediction result of the customer.

[0085] Step S40, in response to the prediction result of the payment habit corresponding to the customer, combining the long short-term memory network and the cascading failure algorithm to analyze the risk of customer electricity fee recovery based on the payment habit prediction result, and calculating the confidence of customer electricity fee recovery;

[0086] It should be noted that when analyzing the risk of customer electricity bill recovery based on the payment habit prediction results by combining the long short-term memory network and the cascading failure algorithm, the historical payment data and the payment habit prediction result features are input into the LSTM model for training, and the model parameters are adjusted through the back propagation algorithm so that the model can learn the time series characteristics and potential laws of the customer's payment behavior. Then the cascading failure algorithm is incorporated into the long short-term memory network to establish a customer payment association network: the customer is regarded as a node in the network, and the payment relationship or common characteristics between customers are used as edges to build a complex network structure. For example, if two customers are in the same region, the same industry, or have similar payment patterns, a connection is established between them. Assuming that a customer has arrears or defaults, the degree and scope of the impact of the event on other customers and the entire network are calculated based on the weights of the nodes and edges and the topological structure of the network. For example, if an important customer's arrears may lead to limited electricity consumption of its upstream and downstream customers, which in turn affects their payment ability and willingness, forming a cascade reaction, and then the payment amount and payment time of the customer in the future period are predicted to obtain the predicted value of the customer's payment possibility and amount at different time nodes.

[0087] Step S50, triggering the electricity fee collection strategy library based on the customer's electricity fee collection confidence level, triggering the customer's corresponding personalized collection strategy instruction.

[0088] In this embodiment, the electricity bill collection strategy library is a set of systematic strategies established by the power company to improve the efficiency of electricity bill collection and reduce the risk of arrears. The electricity bill collection strategy library defines the classification of collection methods: according to different collection methods and intensities, the collection methods are divided into multiple categories, such as SMS collection, telephone collection, door-to-door collection, lawyer letter collection, power outage collection, etc. Collection strategy formulation: for each collection method category, formulate a specific collection strategy, including the time interval of collection, collection content template, collection tone intensity, etc. For example, for SMS collection, it can be set to send the first collection SMS on the first day after the customer is overdue, with a friendly reminder; if the customer still has not paid, send the second collection SMS on the third day, with a slightly stronger tone; send the third collection SMS on the seventh day, clearly informing the possible consequences, etc.

[0089] In the embodiment of the present invention, the customer payment behavior portrait is obtained by combining NILM with the Apriori algorithm, and the overall oversampling of the evaluation index vector is processed by the habit analysis model to obtain the corresponding payment habit prediction result of the customer. The habit analysis model has high accuracy and good robustness in analyzing the customer payment behavior portrait, so that the distribution of the electricity consumption enhancement data can be more obvious without overfitting. And the enhancement processing of customer electricity consumption data based on the SMOTE algorithm can effectively alleviate the problem of the habit analysis model being sensitive to outliers and noise, and overcome the problem that the SOM neural network in the existing analysis method is sensitive to outliers and noise, resulting in large deviations and instability in the clustering results of unbalanced customer electricity consumption data.

[0090] SMOTE (Synthetic Minority Over-sampling Technique) is a synthetic minority over-sampling technique, which is a commonly used algorithm for processing data category imbalance problems. Considering that there are fewer payment behavior data samples in customer electricity consumption data and more unbalanced data, the difference in the number of positive and negative samples in unbalanced data is large, which increases the unbalanced degree of customer electricity consumption data. In order to solve the above problems, the embodiment of the present invention provides a method for enhancing customer electricity consumption data based on the SMOTE algorithm. Figure 2 The schematic diagram of the implementation process of the method for enhancing the customer power consumption data based on the SMOTE algorithm is shown. The method for enhancing the customer power consumption data based on the SMOTE algorithm specifically includes:

[0091] Step S101, loading customer electricity consumption data, calculating the coefficient of variation of the customer electricity consumption data based on a boundary hybrid sampling algorithm, and defining a coefficient threshold by the coefficient of variation of the customer electricity consumption data;

[0092] The coefficient of variation of the customer's electricity consumption data is a statistic that measures the degree of dispersion of a set of electricity consumption data. In this embodiment, the coefficient of variation of the customer's electricity consumption data is calculated by the following formula:

[0093] (1)

[0094] (2)

[0095] (3)

[0096] in, represents the coefficient of variation of customer electricity consumption data, Represents a customer electricity usage dataset The data point density is Electricity usage data set for customers Electricity consumption data of Chinese customers The number of associated data points, Electricity usage data for customers The distance to the associated data point, Represents a customer electricity usage dataset The mean of .

[0097] Step S102, obtaining a coefficient threshold of the customer's electricity usage data, finding the boundary area data and the non-boundary area data of the customer's electricity usage data based on the coefficient threshold of the customer's electricity usage data, and suppressing the imbalance degree of the customer's electricity usage data;

[0098] Step S103, loading the boundary area data, reconstructing the boundary area data based on the SMOTE algorithm, and obtaining a boundary reconstruction set;

[0099] Step S104, loading the non-border area data, calculating the mean of the non-border area data, calculating the Euclidean distance between the non-border area data and the mean point of the non-border area data based on the local anomaly factor algorithm, sorting the non-border area data by distance, and using a preset distance threshold to delete the non-border area data to obtain deleted non-border area data;

[0100] Step S105, integrating the non-boundary area data and the boundary reconstruction set to obtain the power consumption enhancement data.

[0101] In an embodiment of the present invention, customer electricity consumption data is enhanced based on the SMOTE algorithm. By dividing the minority class samples into boundary area data and non-boundary area data, and combining the SMOTE algorithm and the local anomaly factor algorithm to generate and reconstruct new synthetic samples, the number of minority class samples is increased, and the number of samples in each category is balanced, thereby avoiding the customer payment behavior profiling and habit analysis model from being overly biased towards the majority class due to data skew, and overcoming the problem of increasing the imbalance of customer electricity consumption data due to the large difference in the number of positive and negative samples in the imbalanced data.

[0102] The embodiment of the present invention provides a method for analyzing and obtaining a customer payment behavior profile based on NILM combined with Apriori algorithm. Figure 3 The following is a schematic diagram of the implementation process of a method for analyzing a customer payment behavior profile based on NILM combined with an Apriori algorithm. The method for analyzing a customer payment behavior profile based on NILM combined with an Apriori algorithm specifically includes:

[0103] Step S201, loading the power consumption enhancement data, identifying the load characteristics in the power consumption enhancement data based on the NILM model, calculating the total load characteristics of the customers within the sampling period, and determining the customer load type information through the total load characteristics of the customers within the sampling period;

[0104] In this embodiment, when calculating the total customer load characteristics within the sampling period, the NILM model can be used to directly extract features from the time series data, such as mean, variance, kurtosis, skewness, etc. These features can reflect the usage status and usage habits of electrical equipment. Statistical information such as power consumption and power consumption frequency in different time periods can also be calculated as time domain features. The time domain signal is converted to the frequency domain, and features such as frequency, amplitude, phase, etc. are extracted from the frequency domain signal. Fourier transform is a commonly used method for time domain to frequency domain conversion. Different types of electrical equipment can be identified by analyzing the spectrum characteristics.

[0105] The calculation formula for the total customer load characteristics during the sampling period is as follows:

[0106] (4)

[0107] (5)

[0108] in, Indicates the total customer load characteristics during the sampling period, is the total number of electrical appliances, Respectively indicate the current electrical appliances In the sampling period Load characteristics and load correction values ​​within is the sampling interval, is the state change coefficient within the sampling interval, which can be 0.01-0.05. is the electrical reactive power within the sampling interval;

[0109] Step S202, obtaining customer load type information within a sampling period, using a deformable convolutional neural network DCN model to extract features from power consumption enhancement data and customer load type information, and obtaining an evaluation index feature set;

[0110] Step S203, introducing the behavior portrait inter-cluster entropy into the evaluation index feature set based on the Apriori algorithm, and determining the behavior portrait clustering points by calculating the behavior portrait inter-cluster entropy values;

[0111] The entropy value between behavioral profile clusters is calculated by the following formula:

[0112] (6)

[0113] (7)

[0114] in, represents the entropy value between behavioral profile clusters, Characterize the inter-cluster entropy of adjacent behaviors The correlation value between is the correlation matrix between different evaluation index features, is the image cluster entropy The Euclidean distance between is the number of features of the behavior profile cluster, Represents the determinant of the correlation matrix between different evaluation indicator features, Represents the entropy between adjacent behavior profile clusters The similarity Gaussian function between them is used to measure the entropy between adjacent behavior profile clusters. This Gaussian function takes into account the correlation matrix between different eigenvalues , which reflects the correlation between the characteristics of different evaluation indicators;

[0115] Step S204, using the hierarchical analysis method to calculate the feature similarity between the behavior portrait cluster points, selecting the portrait cluster points with convergent similarity as the similarity matrix of the customer payment behavior portrait, integrating at least one set of similarity matrices, and obtaining the customer payment behavior portrait;

[0116] In this embodiment, when selecting the portrait cluster points with convergent similarity as the similarity matrix of the customer payment behavior portrait, the selected distance measurement method is used to calculate the distance between each behavior portrait cluster point to obtain a distance matrix. The distance measurement method includes but is not limited to Euclidean distance, Manhattan distance, cosine similarity, etc. The cluster merging strategies when using the hierarchical analysis method to calculate the feature similarity between the behavior portrait cluster points include single link, full link, and average link.

[0117] Step S205, loading the customer payment behavior portrait, and converting the customer payment behavior portrait into an evaluation index vector based on the Hierarchical-Softmax function.

[0118] In the embodiment of the present invention, the customer payment behavior portrait is obtained based on NILM combined with Apriori algorithm analysis, so that the characteristics of multiple dimensions can be considered at the same time, rather than being limited to the traditional single or a few characteristics. By analyzing and modeling a large amount of high-dimensional data, it is possible to more accurately identify the subtle differences between different customer groups and achieve a more refined customer payment behavior portrait construction. Based on the accurate customer payment behavior portrait, the enterprise can provide customers with personalized payment method recommendations, promotional activity recommendations, etc. The customer payment behavior portrait can also provide richer information and basis for the habit analysis model evaluation, which is conducive to the establishment of a more comprehensive habit analysis evaluation system.

[0119] The embodiment of the present invention provides a method for training a habit analysis model. Figure 4 The following is a schematic diagram of the implementation process of the habit analysis model training method, which specifically includes:

[0120] Step S301, loading an extreme learning machine model as the initial model of the habit analysis model, selecting the number of nodes, weights, and node thresholds of the input layer and hidden layer of the initial model, and presetting the activation functions of the input layer and hidden layer. The activation functions of the input layer and hidden layer are sinusoidal activation function and hard threshold activation function, respectively;

[0121] In the embodiment of the present invention, the number of nodes in the input layer is 30, the weight is 0.02, and the number of nodes in the hidden layer is 60, the weight is 0.01.

[0122] Step S302, setting the hyperparameters and loss function of the maximum likelihood-random forest model architecture in the initial model, and defining the training rounds and single-round training sample size of the initial model;

[0123] It should be noted that the hyperparameter of the maximum likelihood-random forest model architecture can be 0.001, and the training rounds and single-round training sample size of the initial model are 100-150 and 10-15 respectively.

[0124] Step S303: randomly crawl 1000 groups of training samples based on crawler crawling technology, analyze the sample behavior portraits of the training samples based on NILM combined with Apriori algorithm, convert the sample behavior portraits into sample index vectors, and divide the sample index vectors into training sets and test sets;

[0125] In this embodiment, the crawler capture technology can be a Selenium automated crawler, and the training samples are sourced from the electricity consumption data of 50 customers of a power grid in China. The training is completed on a personal computer with a single CPU of 2.5 GHz and 4 GB of memory, and all training samples are processed by MATLAB.

[0126] Step S304, iteratively training the initial model using the training set, and improving the initial model based on the minimum risk Bayesian decision during training until the initial model converges;

[0127] Step S305, obtaining a test set, taking the test set as input, executing the initial model, the initial model predicts and classifies the payment habits of the test set customers, and outputs the payment habit prediction classification results;

[0128] Determine whether the payment habit prediction classification result meets the preset test result threshold;

[0129] Step S306, if the payment habit prediction classification result meets the preset test result threshold, output a converged habit analysis model;

[0130] Step S307, if the payment habit prediction classification result does not meet the preset test result threshold, the Adam optimizer is used to adjust the hyperparameters of the maximum likelihood-random forest model architecture in the initial model, and the number of nodes, weights, and node thresholds of the input layer and hidden layer of the initial model are changed, and the process returns to step S304 to continue iterative training of the initial model.

[0131] In this embodiment, the habit analysis model uses an extreme learning machine as the initial model. The initial model consists of an input layer, a hidden layer, and an output layer. When constructing the habit analysis model, a random forest model is introduced between the hidden layer and the output layer. The maximum likelihood estimation algorithm is combined with the random forest model to form a maximum likelihood-random forest model architecture. The output layer of the initial model is frozen, and a classifier of cost-sensitive elements is used to replace the output layer of the initial model to complete the construction of the habit analysis model.

[0132] In an embodiment of the present invention, a habit analysis model and a training method are provided. The habit analysis model uses an extreme learning machine as the initial model, introduces a maximum likelihood-random forest model architecture, and uses a classifier of cost-sensitive elements to replace the output layer of the initial model, thereby realizing a more refined analysis of customer payment habits, which can effectively reduce the complexity of calculations, reduce operation time, and improve the fluency of the system.

[0133] The embodiment of the present invention provides a method for executing a habit analysis model using a change feature quantity as input. Figure 5 The schematic diagram of the implementation flow of the method for executing the habit analysis model with the change feature quantity as input is shown. The method for executing the habit analysis model with the change feature quantity as input specifically includes:

[0134] Step S401, loading the evaluation index vector, inputting the evaluation index vector into the input layer of the habit analysis model, and the input layer performs overall oversampling processing on the evaluation index vector to obtain an oversampling set;

[0135] It should be noted that oversampling is a data enhancement technique, which is mainly used to solve the class imbalance problem in the data set. That is, when the number of samples of one class (usually called the minority class) is much smaller than that of other classes (majority classes), the data set is balanced by increasing the number of minority class samples.

[0136] Step S402, obtaining an oversampled set, inputting the oversampled set into a hidden layer of a habit analysis model, and obtaining an output matrix of the hidden layer;

[0137] The output matrix of the hidden layer is expressed as:

[0138] (8)

[0139] in, represents the output matrix of the hidden layer, is the activation function of the hidden layer, is the input representation of the oversampled set, is the node threshold of the hidden layer, is the connection weight of the hidden layer;

[0140] Step S403, calculating the truncated singular value generalized inverse of the output matrix based on the difference feature extraction method, taking the truncated singular value generalized inverse as the change feature quantity, and obtaining the change feature quantity caused by the change of the customer's payment habit;

[0141] Step S404, loading the change feature quantity, inputting the change feature quantity into the maximum likelihood-random forest model framework, and the maximum likelihood-random forest model calculates the dependency cost matrix based on the change feature quantity;

[0142] Step S405 , using a classifier of cost-sensitive elements to calculate the average cost of classification predictions that depend on the cost matrix, and outputting the corresponding payment habit prediction result of the customer.

[0143] In this embodiment, the payment habit prediction result corresponding to the customer refers to the classification of payment habits into different levels according to certain standards using a cost-sensitive element classifier, such as "excellent", "good", "average", "poor", etc. For example, customers with high payment timeliness and amount stability can be marked as "excellent".

[0144] The embodiment of the present invention provides a customer payment habit analysis and prediction system. Figure 6 The schematic diagram of the structure of the customer payment habit analysis and prediction system is shown, and the customer payment habit analysis and prediction system specifically includes:

[0145] The data collection module 100 acquires customer electricity consumption data, wherein the customer electricity consumption data includes electricity load parameters, electricity payment information, and electricity arrears information, and enhances the customer electricity consumption data based on the SMOTE algorithm to obtain enhanced electricity consumption data;

[0146] The behavior profile module 200 is used to load the electricity consumption enhancement data, obtain the customer payment behavior profile based on NILM combined with Apriori algorithm analysis, and convert the customer payment behavior profile into an evaluation index vector;

[0147] The habit prediction module 300 is used to obtain the evaluation index vector, use the habit analysis model to perform overall oversampling processing on the evaluation index vector, calculate the output matrix of the evaluation index, analyze and process the output matrix of the evaluation index based on the difference feature extraction method, obtain the change feature quantity generated by the change of the customer's payment habit, use the change feature quantity as input, execute the habit analysis model, and output the payment habit prediction result corresponding to the customer;

[0148] The confidence calculation module 400, in response to the prediction result of the payment habit corresponding to the customer, combines the long short-term memory network and the cascading failure algorithm to analyze the risk of customer electricity fee recovery based on the payment habit prediction result, and calculates the confidence of customer electricity fee recovery;

[0149] The instruction triggering module 500 triggers the electricity fee collection strategy library based on the customer's electricity fee collection confidence level, and triggers the customer's corresponding personalized collection strategy instruction.

[0150] In this embodiment, the data acquisition module 100 includes:

[0151] A data capture unit 110, used for distributed collection of customer power consumption data;

[0152] A coefficient of variation determination unit 120 is configured to load customer power consumption data, calculate the coefficient of variation of the customer power consumption data based on a boundary hybrid sampling algorithm, and define a coefficient threshold value according to the coefficient of variation of the customer power consumption data;

[0153] The boundary determination unit 130 is used to obtain a coefficient threshold of the customer power usage data, find the boundary area data and the non-boundary area data of the customer power usage data based on the coefficient threshold of the customer power usage data, and suppress the imbalance degree of the customer power usage data;

[0154] The data enhancement unit 140 is used to load the boundary area data, reconstruct the boundary area data based on the SMOTE algorithm to obtain a boundary reconstruction set, load the non-boundary area data, calculate the mean of the non-boundary area data, calculate the Euclidean distance between the non-boundary area data and the mean point of the non-boundary area data based on the local anomaly factor algorithm, sort the non-boundary area data by distance, use a preset distance threshold to delete the non-boundary area data to obtain deleted non-boundary area data, integrate the non-boundary area data and the boundary reconstruction set to obtain electricity consumption enhancement data.

[0155] It should be noted that the customer payment habit analysis and prediction system provided in the embodiment of the present invention corresponds to the above-mentioned customer payment habit analysis and prediction method. The explanations, examples, beneficial effects and other parts of its relevant contents can refer to the corresponding contents in the customer payment habit analysis and prediction method, and will not be repeated here.

[0156] In summary, the present invention provides a method and system for analyzing and predicting customer payment habits. In the embodiment of the present invention, the customer payment behavior portrait is obtained by combining NILM with the Apriori algorithm, and the overall oversampling of the evaluation index vector is processed by the habit analysis model to analyze and obtain the corresponding payment habit prediction result of the customer. The habit analysis model has high accuracy and good robustness in analyzing the customer payment behavior portrait, so that the distribution of electricity consumption enhancement data can be made more obvious without overfitting. And the enhancement processing of customer electricity consumption data based on the SMOTE algorithm can effectively alleviate the problem that the habit analysis model is sensitive to outliers and noise, and overcomes the problem that the SOM neural network in the existing analysis method is sensitive to outliers and noise, resulting in large deviations and instability in the clustering results of unbalanced customer electricity consumption data.

[0157] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0158] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A method for analyzing and predicting customer payment habits, characterized in that: The customer payment habit analysis and prediction method comprises: Obtain customer electricity consumption data, where the customer electricity consumption data includes electricity load parameters, electricity payment information, and electricity arrears information, and enhance the customer electricity consumption data based on the SMOTE algorithm to obtain enhanced electricity consumption data; Load the electricity consumption enhancement data, analyze the customer payment behavior profile based on NILM combined with Apriori algorithm, and convert the customer payment behavior profile into an evaluation index vector; Obtain the evaluation index vector, use the habit analysis model to perform overall oversampling on the evaluation index vector, calculate the output matrix of the evaluation index, analyze and process the output matrix of the evaluation index based on the difference feature extraction method, obtain the change feature quantity caused by the change of the customer's payment habit, use the change feature quantity as input, execute the habit analysis model, and output the corresponding payment habit prediction result of the customer; The method for obtaining a customer payment behavior profile based on NILM combined with Apriori algorithm analysis specifically includes: Load the power consumption enhancement data, identify the load characteristics in the power consumption enhancement data based on the NILM model, calculate the total load characteristics of the customers within the sampling period, and determine the customer load type information through the total load characteristics of the customers within the sampling period; The calculation formula for the total customer load characteristics during the sampling period is as follows: （4） （5） in, Indicates the total customer load characteristics during the sampling period, is the total number of electrical appliances, Respectively indicate the current electrical appliances In the sampling period Load characteristics and load correction values ​​within is the sampling interval, is the state change coefficient within the sampling interval, is the electrical reactive power within the sampling interval; Obtain customer load type information within the sampling period, use the deformable convolutional neural network (DCN) model to extract features from power consumption enhancement data and customer load type information, and obtain a feature set of evaluation indicators; Based on the Apriori algorithm, the inter-cluster entropy of behavioral portraits is introduced into the evaluation index feature set, and the clustering points of behavioral portraits are determined by calculating the inter-cluster entropy values ​​of behavioral portraits. The entropy value between behavioral profile clusters is calculated by the following formula: （6） （7） in, represents the entropy value between behavioral profile clusters, Characterize the inter-cluster entropy of adjacent behaviors The correlation value between is the correlation matrix between different evaluation index features, is the image cluster entropy The Euclidean distance between is the number of features of the behavior profile cluster, Represents the determinant of the correlation matrix between different evaluation indicator features, Represents the entropy between adjacent behavior profile clusters Similarity Gaussian function between them; The hierarchical analysis method is used to calculate the feature similarity between the behavior portrait clustering points, and the portrait clustering points with convergent similarity are selected as the similarity matrix of the customer payment behavior portrait. At least one set of similarity matrices is integrated to obtain the customer payment behavior portrait; Load the customer payment behavior profile and convert it into an evaluation indicator vector based on the Hierarchical-Softmax function.

2. The method for analyzing and predicting customer payment habits according to claim 1, characterized in that: The method further comprises: In response to the prediction results of the customer's corresponding payment habits, the customer's electricity fee recovery risk analysis is performed based on the payment habit prediction results by combining the long short-term memory network and the cascading failure algorithm, and the customer's electricity fee recovery confidence is calculated; The electricity bill collection strategy library is triggered based on the customer's electricity bill recovery confidence level, triggering the customer's corresponding personalized collection strategy instructions.

3. The method for analyzing and predicting customer payment habits according to claim 1, characterized in that: The method for enhancing the processing of customer electricity consumption data based on the SMOTE algorithm specifically includes: Load customer electricity consumption data, calculate the coefficient of variation of customer electricity consumption data based on the boundary mixed sampling algorithm, and define the coefficient threshold according to the coefficient of variation of customer electricity consumption data; Obtain a coefficient threshold of the customer's electricity usage data, and find the boundary area data and non-boundary area data of the customer's electricity usage data based on the coefficient threshold of the customer's electricity usage data to suppress the imbalance of the customer's electricity usage data; Load the boundary area data, reconstruct the boundary area data based on the SMOTE algorithm, and obtain the boundary reconstruction set; Load the non-boundary area data, calculate the mean of the non-boundary area data, calculate the Euclidean distance between the non-boundary area data and the mean point of the non-boundary area data based on the local anomaly factor algorithm, sort the non-boundary area data by distance, and use a preset distance threshold to delete the non-boundary area data to obtain the deleted non-boundary area data; The non-boundary area data and boundary reconstruction set are integrated to obtain the electricity consumption enhancement data.

4. The method for analyzing and predicting customer payment habits according to claim 3, characterized in that: The coefficient of variation of customer electricity consumption data is calculated using the following formula: （1） （2） （3） in, represents the coefficient of variation of customer electricity consumption data, Represents a customer electricity usage dataset The data point density is Electricity usage data set for customers Electricity consumption data of Chinese customers The number of associated data points, Electricity usage data for customers The distance to the associated data point, Represents a customer electricity usage dataset The mean of .

5. The method for analyzing and predicting customer payment habits according to claim 1, characterized in that: The habit analysis model uses an extreme learning machine as the initial model. The initial model consists of an input layer, a hidden layer, and an output layer. When constructing the habit analysis model, a random forest model is introduced between the hidden layer and the output layer. The maximum likelihood estimation algorithm is combined with the random forest model to form a maximum likelihood-random forest model architecture. The output layer of the initial model is frozen, and a classifier of cost-sensitive elements is used to replace the output layer of the initial model to complete the construction of the habit analysis model.

6. The method for analyzing and predicting customer payment habits according to claim 5, characterized in that: The training method of the habit analysis model specifically includes: Load the extreme learning machine model as the initial model of the habit analysis model, select the number of nodes, weights, and node thresholds of the input layer and hidden layer of the initial model, and preset the activation functions of the input layer and hidden layer. The activation functions of the input layer and hidden layer are sinusoidal activation function and hard threshold activation function respectively; Set the hyperparameters and loss function of the maximum likelihood-random forest model architecture in the initial model, and define the training rounds and sample size of a single round of training for the initial model; Based on the crawler crawling technology, 1,000 sets of training samples are randomly crawled. Based on NILM combined with the Apriori algorithm, the sample behavior portraits of the training samples are obtained, and the sample behavior portraits are converted into sample indicator vectors. The sample indicator vectors are divided into training sets and test sets. The initial model is iteratively trained using the training set. During training, the initial model is improved based on the minimum risk Bayesian decision until the initial model converges. Obtain a test set, use the test set as input, execute the initial model, the initial model predicts and classifies the payment habits of the test set customers, and outputs the payment habit prediction classification results. If the payment habit prediction classification results meet the preset test result threshold, output a converged habit analysis model; If the classification result of payment habit prediction does not meet the preset test result threshold, the Adam optimizer is used to adjust the hyperparameters of the maximum likelihood-random forest model architecture in the initial model, and the number of nodes, weights, and node thresholds of the input layer and hidden layer of the initial model are changed to continue iterative training of the initial model.

7. The method for analyzing and predicting customer payment habits according to claim 6, characterized in that: The method of executing the habit analysis model with the change feature quantity as input specifically includes: Loading the evaluation indicator vector, inputting the evaluation indicator vector into the input layer of the habit analysis model, and the input layer performs overall oversampling processing on the evaluation indicator vector to obtain an oversampling set; Obtain an oversampled set, input the oversampled set into the hidden layer of the habit analysis model, and obtain an output matrix of the hidden layer; The output matrix of the hidden layer is expressed as: （8） in, represents the output matrix of the hidden layer, is the activation function of the hidden layer, is the input representation of the oversampled set, is the node threshold of the hidden layer, is the connection weight of the hidden layer; The truncated singular value generalized inverse of the output matrix is ​​calculated based on the difference feature extraction method, and the truncated singular value generalized inverse is used as the change feature quantity to obtain the change feature quantity caused by the change of customer payment habits; Load the change feature quantity and input the change feature quantity into the maximum likelihood-random forest model architecture. The maximum likelihood-random forest model calculates the dependency cost matrix based on the change feature quantity. The classifier of cost-sensitive elements is used to calculate the average cost of classification prediction that depends on the cost matrix, and the corresponding payment habit prediction result of the customer is output.

8. A customer payment habit analysis and prediction system, used to implement the customer payment habit analysis and prediction method according to any one of claims 1 to 7, characterized in that: The customer payment habit analysis and prediction system specifically includes: The data collection module obtains customer electricity consumption data, including electricity load parameters, electricity payment information, and electricity arrears information. The customer electricity consumption data is enhanced based on the SMOTE algorithm to obtain enhanced electricity consumption data. The behavior profile module is used to load the electricity consumption enhancement data, obtain the customer payment behavior profile based on NILM combined with the Apriori algorithm, and convert the customer payment behavior profile into an evaluation index vector; The habit prediction module is used to obtain the evaluation index vector, use the habit analysis model to perform overall oversampling on the evaluation index vector, calculate the output matrix of the evaluation index, analyze and process the output matrix of the evaluation index based on the difference feature extraction method, obtain the change feature quantity caused by the change of the customer's payment habit, use the change feature quantity as input, execute the habit analysis model, and output the corresponding payment habit prediction result of the customer; The confidence calculation module, in response to the prediction results of the payment habits of the customers, combines the long short-term memory network and the cascading failure algorithm to analyze the risk of customer electricity fee recovery based on the payment habit prediction results, and calculates the confidence of customer electricity fee recovery; The instruction trigger module triggers the electricity bill collection strategy library based on the customer's electricity bill recovery confidence level, and triggers the customer's corresponding personalized collection strategy instruction.

9. The customer payment habit analysis and prediction system according to claim 8, characterized in that: The data acquisition module comprises: Data capture unit, used for distributed collection of customer electricity consumption data; A coefficient of variation determination unit is used to load customer electricity consumption data, calculate the coefficient of variation of the customer electricity consumption data based on a boundary mixed sampling algorithm, and define a coefficient threshold value according to the coefficient of variation of the customer electricity consumption data; A boundary determination unit, used to obtain a coefficient threshold of the customer's electricity usage data, find the boundary area data and non-boundary area data of the customer's electricity usage data based on the coefficient threshold of the customer's electricity usage data, and suppress the imbalance degree of the customer's electricity usage data; The data enhancement unit is used to load the boundary area data, reconstruct the boundary area data based on the SMOTE algorithm to obtain the boundary reconstruction set, load the non-boundary area data, calculate the mean of the non-boundary area data, calculate the Euclidean distance between the non-boundary area data and the mean point of the non-boundary area data based on the local anomaly factor algorithm, sort the non-boundary area data by distance, delete the non-boundary area data using a preset distance threshold to obtain the deleted non-boundary area data, integrate the non-boundary area data and the boundary reconstruction set to obtain the electricity consumption enhancement data.

Images