Customer credit evaluation method and system based on multi-view features

Abstract

The invention discloses a customer credit evaluation method and system based on multi-view features, particularly relates to the field of credit evaluation, is used for solving the problem of customer default risk dynamic identification and accurate management, and is characterized in that a time sequence feature matrix is generated through preprocessing on the basis of historical electricity fee payment records and electricity consumption cycle data; extracting a multi-period overdue increasing factor and a time lag correlation factor, and depicting an accumulative effect and a time sequence characteristic of an overdue behavior; in combination with a time sequence analysis model, modeling an interaction relationship between an overdue progressive increase effect and power consumption fluctuation to realize customer default risk dynamic prediction; recursively deducing superposition influence of overdue behaviors on subsequent risks through a dynamic risk accumulation mechanism, and updating credit scores and marking high-risk customers based on a real-time monitoring window and a dynamic threshold triggering mechanism; in combination with actual payment performance and prediction result deviation quantification, the weight of the feature matrix is dynamically corrected, model rolling optimization and adaptive adjustment are completed, and the dynamic accuracy and management efficiency of credit risk identification are improved.

Description

Technical Field

[0001] The present invention relates to the field of credit evaluation, and more specifically, to a customer credit evaluation method and system based on multi-view features. Background Art

[0002] In the customer credit evaluation of the electricity fee scenario, electricity payment behavior, electricity usage habits and the resulting credit risks often have strong temporal continuity and interrelationships. For example, if a customer has paid on time for a long time in the past, it usually means a low probability of default; if there are multiple minor overdue payments and a gradual accumulation trend, a higher risk of default may occur in the subsequent account period. However, traditional credit evaluation methods often regard the customer's credit performance as a static indicator, often ignoring the fact that the customer's electricity usage and payment behavior will continue to evolve over time. Especially in the settlement and collection of electricity fees, the electricity usage behavior of different customers has obvious periodicity and volatility. If this time correlation cannot be accurately portrayed, it is easy to have evaluation lags or misjudgments, making it difficult for risk monitoring and control to respond in a timely manner at the critical point when the customer's default risk is about to rise.

[0003] The current technical problem is that it is difficult to effectively extract potential overdue increases or abnormal electricity consumption accumulation signals from multiple periods of data, and to quantify and predict them in the model in a timely manner. Specifically, traditional models only score customers' single-period or short-term overdue situations, and fail to capture the risk superposition effect caused by the spread of multiple periods of overdue payments to the future. It is also difficult to accurately identify the potential impact of customers' current electricity consumption behavior on subsequent billing periods. When customers show slight late payments or load anomalies in previous bills, the model cannot amplify them in a timely manner as key early warning factors, resulting in the lack of sensitivity and foresight in credit evaluation results to future overdue outbreaks. This evaluation lag not only weakens the credit score's ability to distinguish risks, but also causes delays in default prevention and customer management in practical applications. In order to solve the above problems, 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 customer credit evaluation method and system based on multi-perspective features. The customer credit evaluation method is constructed through multi-perspective features, based on historical electricity bill payment records and electricity consumption cycle data, and a time series feature matrix is ​​generated after unified preprocessing. Multi-period overdue incremental factors and time lag correlation factors are extracted to accurately characterize the cumulative effect and time series characteristics of overdue behavior; combined with the time series analysis model, the complex interactive relationship between the overdue incremental effect and electricity consumption fluctuation is deeply modeled to achieve dynamic prediction of customer default risk; the superposition effect of overdue behavior on subsequent risks is recursively deduced through a dynamic risk accumulation mechanism, and based on a real-time monitoring window and a dynamic threshold trigger mechanism, the credit score is updated in time and high-risk customers are marked; further, by quantifying the deviation between actual payment performance and the predicted result, the feature matrix weight is rolled over to achieve model adaptive optimization 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 customer credit evaluation method based on multi-view features includes the following steps:

[0007] S1: Obtain historical electricity bill payment records and electricity consumption cycle data from relevant platforms, and generate integrated time series data with timestamps through unified preprocessing;

[0008] S2: Conduct cumulative analysis on the continuous overdue behavior of each customer in the integrated time series data, extract the overdue incremental characteristic factors and time lag correlation factors based on the pre-set overdue delay length threshold, and construct a characteristic matrix that can characterize the customer's multi-period overdue cumulative risk;

[0009] S3: Taking the feature matrix as input, the time series analysis model is used to jointly model the multi-period overdue incremental effect and customer electricity consumption fluctuations to predict the customer's default probability in the subsequent account period;

[0010] S4: A dynamic risk accumulation mechanism is set up during the forecasting process to recursively calculate the impact of previous overdue periods on subsequent default risks, and assign higher risk weights to customers with increasing overdue phenomena;

[0011] S5: Monitor the prediction results of the time series analysis model in real time. Once the customer's overdue increase index and electricity consumption fluctuation index reach the risk threshold, the credit score update and risk warning will be triggered immediately.

[0012] S6: Compare the updated credit score with the customer’s subsequent actual payment performance, and calibrate the weight of the overdue incremental factor in the feature matrix based on the prediction error and time lag indicators to achieve rolling optimization of the risk identification model.

[0013] In a preferred embodiment, the specific processing logic of step S2 is as follows:

[0014] First, filter out the overdue amount O of each customer in the consecutive account period from the time-stamped integrated data output in step S1. k and the number of days of delay L k , define the account period sequence as {k=1, 2, ..., n}; then, determine whether the account period is overdue based on the pre-set overdue delay threshold. If the number of delayed days in the account period k is greater than the overdue delay threshold, the corresponding account period is considered overdue and the overdue amount is calculated; then, in order to characterize the cumulative upward trend of overdue, a multi-layer incremental feature factor and a nonlinear trigger mechanism are used: the first layer defines the difference feature Δ k =max(0,O k -O k-1 ), which is used to measure the increment of the overdue amount of the current account period compared with the previous account period; the second layer constructs the cumulative potential energy factor P through multi-step recursion k =P k-1 +f(Δ k , Δ k-1 , ..., Δ k-m ), where P0 = 0, m is the number of backtracking steps, and f(·) is a nonlinear activation function, which increases the cumulative potential factor when a small amount of overdue debt is detected. To reflect the hysteresis of the strong temporal correlation between accounts, a time lag correlation factor is used to distinguish between weak hysteresis areas and strong hysteresis areas: First, define the gene R k =g(L k )×(O k / O ref ); where O ref is the calibration value of the highest single-period overdue amount in the customer's history, g(·) is the piecewise logistic function; in order to accumulate the corresponding lag effect in multi-period time series, the secondary propagation mechanism H is introduced k =H k-1 +φ(R k , R k-1 , ..., R k-m );where φ(·) is a self-defined nonlinear transformation function, which maps the lag factors of the most recent account periods into a comprehensive lag risk quantity, which is used to capture the rapid growth of the customer's default risk when there is a significant delay in multiple consecutive periods; Finally, (P k , H k ) This pair of increasing and lagging factors is regarded as the key risk characterization indicators under the account period k, and is arranged together with other auxiliary features to form a feature matrix.

[0015] In a preferred embodiment, the specific processing logic of step S3 is as follows:

[0016] First, the overdue increment factor and the lag correlation factor are read from the feature matrix, and the electricity consumption fluctuation sequence extracted in step S1 is called at the same time, and the two parts of data are indexed and matched at the account period level to generate a multi-dimensional time series input sequence; then, a nonlinear time series analysis framework including wavelet coefficient decomposition is used to perform multi-scale decomposition on the electricity consumption fluctuation sequence and extract several groups of wavelet coefficients to capture the abnormal oscillation characteristics of customers' electricity consumption in the high-frequency and low-frequency ranges; then, the wavelet coefficients are input into the deep backtracking neural network together with the overdue increment factor and the time lag correlation factor; when the model training converges, the deep backtracking neural network is used to infer and predict the data of the new account period, and the corresponding default probability value is output.

[0017] In a preferred embodiment, the specific processing logic of step S4 is as follows:

[0018] First, the predicted default probability of the customer in each account period is read, denoted as p(k); at the same time, the overdue increment factor corresponding to each customer in the same account period is read from the feature matrix, denoted as A(k). The overdue increment factor is used to quantify the degree of superposition of the customer's overdue amount or overdue days in the previous account renewal period; then, the dynamic risk accumulation amount R(k) is defined and calculated using the following recursive formula: R(k+1)=R(k)+[p(k) (1+γ×A(k)) -p(k)]; where γ is a control factor used to determine the amplification strength of the overdue increase phenomenon; when iterating, starting from the starting account period, R(1)=0 and calculate the dynamic risk accumulation of subsequent account periods in sequence until the predetermined forecast period is covered; when the dynamic risk accumulation exceeds the pre-set risk threshold, the corresponding customer is marked as a high-risk customer.

[0019] In a preferred embodiment, the specific processing logic of step S5 is as follows:

[0020] First, the latest risk value sequence is extracted from the dynamic risk accumulation, and the power consumption fluctuation feature sequence and the overdue increasing factor are called to construct a dynamic monitoring window. The monitoring window covers the data of the last n billing periods, and records the sequence information of the dynamic risk accumulation, power consumption fluctuation amplitude and overdue increasing factor, which is recorded as W = {R(i), U(i), A(i)|i = k-n+1, ​​..., k}, where R(i) represents the dynamic risk accumulation of bill period i, U(i) represents the power consumption fluctuation amplitude, and A(i) represents the overdue increasing factor. In order to realize dynamic risk judgment, the real-time trigger factor Z(k) is defined as:

[0021] A dynamic threshold T(k) is set for the real-time trigger factor, and its calculation formula is: T(k) = μ+η·log(1+var(W)); where μ is the basic threshold, η is the fluctuation adjustment factor, and var(W) is the amplitude change of the characteristic sequence in the monitoring window; when the trigger factor exceeds the dynamic threshold, the credit score update module is called according to the real-time monitoring results; the credit score update module readjusts the score based on the customer's current risk level, and the specific calculation is: S(k) = S(k-1)+δ·(Z(k)-T(k)); where S(k) is the credit score of the current account period, S(k-1) is the score of the previous account period, and δ is the score adjustment coefficient, which is used to control the adjustment range.

[0022] In a preferred embodiment, the specific processing logic of step S6 is as follows:

[0023] First, based on the latest credit score and the actual payment performance of the customer in the new account period after the warning is triggered, if the customer is still overdue in the corresponding account period, the corresponding overdue duration and amount are recorded; if the customer pays on time, it is recorded as a normal account period; then, in order to quantify the degree of deviation between the prediction and reality, the matching function G(k) is used to measure the difference between the model's predicted default probability in account period k and the actual result, which is defined as follows: in Indicates the actual overdue amount or number of overdue days for the corresponding account period. is the predicted value of the default probability of the corresponding account period; then, in order to correct the impact of the overdue increasing factor in the characteristic matrix on subsequent predictions, the correction coefficient C(k+1) is defined as the rolling correction amount of the overdue increasing factor: where sign(G(k) is used to determine the direction of the deviation, and Γ is used to convert the observed overdue amount to the time lag level of the previous period. Combined, a correction amount for the overdue incremental factor of the next account period is generated, ζ is the control coefficient, which determines the sensitivity of the correction; the correction coefficient is superimposed on the field where the overdue incremental factor is located in the feature matrix, and the overdue incremental characteristic value of the corresponding customer's subsequent account period is updated.

[0024] The customer credit evaluation system based on multi-view features includes: data integration module, feature extraction module, time series modeling module, risk accumulation module, real-time monitoring module, and model optimization module;

[0025] Data integration module: obtains historical electricity bill payment records and electricity consumption cycle data from relevant platforms, and generates integrated time series data with timestamps through unified preprocessing. The output integrated time series data with timestamps is passed to the feature extraction module;

[0026] Feature extraction module: conducts cumulative analysis on the continuous overdue behavior of each customer in the integrated time series data, extracts the overdue incremental characteristic factors and time lag correlation factors based on the overdue delay length threshold, constructs a feature matrix that describes the cumulative risk of multi-period overdue, and transmits the generated feature matrix to the time series modeling module;

[0027] Time series modeling module: Using the feature matrix as input, combined with the customer's electricity consumption fluctuation characteristics, joint modeling is performed through the time series analysis model to predict the customer's default probability in the subsequent account period, generate a dynamic default probability prediction value, and output the default probability prediction value and key features to the risk accumulation module;

[0028] Risk accumulation module: Based on the default probability and overdue increment factor output by time series modeling, the dynamic risk accumulation mechanism is used to recursively infer the impact of previous overdue behaviors on subsequent risks, assign higher risk weights to high-risk customers, generate dynamic risk accumulation, and pass the dynamic risk accumulation and corresponding default probability prediction value to the real-time monitoring module;

[0029] Real-time monitoring module: real-time monitoring of dynamic risk accumulation and default probability prediction values, combined with the customer's overdue incremental factor and electricity consumption fluctuation characteristics, based on dynamic thresholds to identify high-risk customers, trigger credit score updates and risk warnings, and pass updated credit scores and high-risk customer tags to the model optimization module;

[0030] Model optimization module: compares the updated credit score with the customer's actual payment performance, calculates the prediction error and time lag index, and dynamically corrects the overdue incremental factor weights in the feature matrix through deviation quantification to achieve rolling optimization of the risk identification model.

[0031] The technical effects and advantages of the customer credit evaluation method and system based on multi-view features of the present invention are as follows:

[0032] The present invention constructs a customer credit evaluation method based on multi-perspective features, relies on the integrated preprocessing of historical electricity bill payment records and electricity consumption cycle data, constructs a feature matrix that can accurately characterize the cumulative effect and time lag characteristics of customer multi-period overdue behavior, and combines with the time series analysis model to achieve accurate modeling of the complex interactive relationship between multi-period overdue incremental effect and electricity consumption fluctuation, and effectively predicts the customer's default probability in the subsequent account period; by introducing a dynamic risk accumulation mechanism, the recursive impact of previous overdue behaviors on subsequent period risks is quantified and amplified, providing a strong dynamic basis for the identification of high-risk customers; further, through real-time monitoring and dynamic threshold triggering, timely updating of credit scores and early warning marking of high-risk customers are achieved, significantly improving the sensitivity and response speed of risk management; finally, by matching actual payment performance with prediction results, the weight of overdue incremental factors is rolled optimized, and a feedback mechanism for model adaptive adjustment is constructed, which ensures the accuracy and robustness of the system in long-term dynamic scenarios, thereby comprehensively improving the efficiency of customer credit risk identification and management, and providing power companies with an accurate, dynamic and efficient credit evaluation tool. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the customer credit evaluation method based on multi-view features of the present invention;

[0034] Figure 2 It is a schematic diagram of the structure of the customer credit evaluation system based on multi-view features of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] Embodiment 1: Figure 1 The present invention provides a customer credit evaluation method based on multi-view features, including:

[0037] S1: Obtain historical electricity bill payment records and electricity consumption cycle data from relevant platforms, and generate integrated time series data with timestamps through unified preprocessing;

[0038] S2: Conduct cumulative analysis on the continuous overdue behavior of each customer in the integrated time series data, extract the overdue incremental characteristic factors and time lag correlation factors based on the pre-set overdue delay length threshold, and construct a characteristic matrix that can characterize the customer's multi-period overdue cumulative risk;

[0039] S3: Taking the feature matrix as input, the time series analysis model is used to jointly model the multi-period overdue incremental effect and customer electricity consumption fluctuations to predict the customer's default probability in the subsequent account period;

[0040] S4: A dynamic risk accumulation mechanism is set up during the forecasting process to recursively calculate the impact of previous overdue periods on subsequent default risks, and assign higher risk weights to customers with increasing overdue phenomena;

[0041] S5: Monitor the prediction results of the time series analysis model in real time. Once the customer's overdue increase index and electricity consumption fluctuation index reach the risk threshold, the credit score update and risk warning will be triggered immediately.

[0042] S6: Compare the updated credit score with the customer’s subsequent actual payment performance, and calibrate the weight of the overdue incremental factor in the feature matrix based on the prediction error and time lag indicators to achieve rolling optimization of the risk identification model.

[0043] In the electricity fee scenario, the accuracy of customer credit evaluation directly affects the risk management and customer service quality of power companies. Traditional credit evaluation methods often ignore the temporal dynamics of customer electricity consumption behavior and payment habits, resulting in a lag in the identification of potential default risks. The present invention uses a customer credit evaluation method based on multi-perspective features, especially strengthening the identification of credit evaluation hysteresis under time correlation, to achieve forward-looking identification and timely warning of customer default risks through fine-grained time series data analysis and dynamic risk modeling, thereby improving the accuracy and response speed of credit scoring and enhancing the risk control capabilities of power companies.

[0044] S1: Obtain historical electricity bill payment records and electricity usage cycle data from the customer bill management platform and electricity collection terminal, and pre-process multi-source data uniformly through data deduplication, time alignment and format conversion to generate integrated time series data with timestamps. The specific processing logic is as follows:

[0045] First, the customer bill management platform and electricity consumption collection terminal are connected through the API interface respectively, and the electricity bill payment records containing fields such as customer ID, bill date, payment amount, payment status, etc., as well as electricity consumption cycle data including fields such as customer ID, electricity consumption date, electricity consumption, and electricity consumption period are automatically extracted.

[0046] Next, a deduplication mechanism based on the SHA-256 hash algorithm is applied to ensure the uniqueness of each record in the customer ID and timestamp dimensions and eliminate all duplicate or redundant data.

[0047] Subsequently, a timestamp standardization method was used to convert all time fields into a unified ISO8601 format, and through the time window sliding technology, electricity consumption data of different frequencies and bill payment records were aligned to the preset monthly time window to ensure the consistency and comparability of the data in time.

[0048] Afterwards, use the custom script in the ETL (Extract, Transform, Load) process to convert data from different sources into a unified Parquet format, standardize field names and data types, and ensure the consistency of the data structure.

[0049] Next, missing value detection is performed, and the polynomial interpolation method is used to fill the missing values ​​in the electricity consumption data. A denoising algorithm based on wavelet transform is applied to filter out the high-frequency noise in the electricity consumption data and retain the main electricity consumption trends and periodic fluctuation characteristics.

[0050] Finally, the preprocessed data is arranged in order of customer ID and time, stored in a time series database (such as TimescaleDB), and a composite index based on customer ID and timestamp is established to optimize subsequent data query and access speed, and generate an integrated time series data set with precise timestamps as the basic input for subsequent feature extraction and risk modeling.

[0051] Step S1 generates integrated time series data with accurate timestamps through a series of refined processing such as systematic data acquisition, deduplication, consistency check, time alignment, format conversion and data fusion. This process ensures the high quality and consistency of multi-source data, and lays a solid data foundation for subsequent multi-period overdue feature accumulation analysis, time series modeling and dynamic risk assessment. Through rigorous preprocessing logic, step S1 effectively eliminates data redundancy and noise interference, improves the input data quality of the credit evaluation model, and thus improves the accuracy and responsiveness of the overall credit scoring system.

[0052] Customer overdue behavior often presents the characteristics of cumulative increase and lag triggering. It is difficult to capture the potential default trend by simply using a single-period overdue statistical method. In step S2, the present invention performs calculations around the overdue increase characteristic factor and the time lag correlation factor. Through a series of recursion, transformation and nonlinear mapping methods, the connection relationship between customer overdue behavior in different account periods is fully explored to form a characteristic matrix that can quantify the cumulative risk of multi-period overdue, providing accurate input for the time series analysis model in step S3.

[0053] Step S2: Perform cumulative analysis on the continuous overdue behavior of each customer in the integrated time series data, extract the overdue incremental characteristic factors and time lag correlation factors based on the pre-set overdue delay length threshold, and construct a characteristic matrix that can characterize the customer's multi-period overdue cumulative risk. The specific processing logic is as follows:

[0054] First, filter out the overdue amount O of each customer in the consecutive account period from the time-stamped integrated data output in step S1. k and the number of days of delay L k , define the account period sequence as {k=1, 2, ..., n} and ensure that the time sequence is not missing; then, based on the pre-set overdue delay threshold T, determine whether the account period is overdue. If the number of delayed days in the account period k is greater than the overdue delay threshold, the corresponding account period is considered to be overdue and the overdue amount is further calculated; then, in order to characterize the cumulative upward trend of overdue, use multi-layer incremental feature factors and nonlinear trigger mechanism: the first layer defines the difference feature Δ k =max(0,O k -O k-1 ), which is used to measure the increment of the overdue amount of the current account period compared with the previous account period; the second layer constructs the cumulative potential energy factor P through multi-step recursion k =P k-1 +f(Δ k , Δ k-1 , ..., Δ k-m ); where P0 = 0, m is the number of self-defined backtracking steps, and f(·) is a nonlinear activation function (such as using the piecewise Sigmoid form to merge the increments of several consecutive periods into a cumulative amount to amplify the superposition effect of multiple periods of small overdue payments), so as to significantly improve the cumulative potential factor when detecting continuously increasing small overdue payments; in order to further reflect the hysteresis of the strong temporal correlation between accounts, the time lag correlation factor is used to distinguish between the weak hysteresis area and the strong hysteresis area: first define the gene R k =g(L k )×(O k / O ref ); where O ref is the calibration value of the highest single-period overdue amount in the customer's history, g(·) is the piecewise logistic function, when the number of delay days is in the weak hysteresis zone, that is, the number of delay days does not exceed the overdue delay threshold, g(L k ) tends to grow slowly, and when the number of delayed days exceeds the overdue delay threshold and enters the strong hysteresis zone, g(L k ) will rise rapidly, forming an enhanced amplification effect on the long-delayed overdue accounts; in order to accumulate the lag effect over multiple periods, a secondary propagation mechanism H is introduced k =H k-1 +φ(R k , R k-1 , ..., R k-m);where φ(·) is a self-defined nonlinear transformation function (such as a cubic polynomial form), which can map the lag factors of the most recent account periods into a comprehensive lag risk amount to capture the rapid growth of the customer's default risk when there is a significant delay in multiple consecutive periods; finally, (P k , H k ) This pair of increasing and lagging factors is regarded as the key risk characterization indicator under the account period k, and is combined with other auxiliary characteristics (such as the current overdue amount O k , account period number k, customer ID, etc.) are arranged together to form a feature matrix

[0055] After the matrix is ​​generated for all customers, a completeness check and possible missing value interpolation (such as correction through multi-period trend extrapolation) are performed, and finally the feature matrix is ​​written into the database with customer ID and account period number as the primary key, providing the time series analysis model in step S3 with directly callable depth increment and time series lag feature inputs.

[0056] Step S2 accurately captures the accumulation and increasing trend of customers' overdue behavior in continuous account periods, as well as the time lag effect of overdue behavior, by constructing multiple layers of increasing characteristic factors and complex time-lag correlation factors. This process significantly enhances the ability to identify the accumulation of multiple periods of small overdues and sudden large delayed overdues, so that the characteristic matrix can truly reflect the dynamic changes of customers' default risk. Specifically, the increasing characteristic factor can quantify the incremental increase of customers' overdue amounts and timely identify customers with gradually increasing default risks; the lagged correlation factor ensures that the model is highly sensitive to sudden high-risk behaviors by strengthening the impact of large delayed overdue accounts. The construction of this characteristic matrix not only improves the accuracy of risk prediction, but also provides high-resolution and fine-grained input data for subsequent time series analysis models, ensuring that the default probability prediction can more accurately reflect the actual risk status of customers. In addition, through integrity verification and missing value interpolation, the continuity and integrity of the characteristic matrix are ensured, further improving the stability and reliability of the overall credit evaluation system. Finally, the characteristic matrix generated in step S2 provides power companies with a scientific and detailed customer default risk quantification tool, which significantly improves the foresight and response speed of risk identification, and helps power companies achieve more efficient risk control and accurate customer service in credit management.

[0057] It is difficult to fully capture the complex interactive relationship between customers' multi-period overdue accumulation behavior and electricity consumption fluctuations by relying solely on simple linear analysis. The previous step has generated a feature matrix with overdue increment factors and time lag correlation factors. Combined with the customer's electricity load fluctuation information, if a deep modeling method for time series characteristics is not adopted, it is difficult to achieve an accurate prediction of the default probability of subsequent account periods. In step S3, the present invention further amplifies and captures the time dependence between multi-period overdue increments and electricity consumption fluctuations through deep coupling of multi-dimensional feature fusion and nonlinear time series analysis model, thereby providing a high-precision default probability prediction result for the subsequent dynamic risk accumulation (step S4).

[0058] Step S3: Take the feature matrix as input, use the time series analysis model to jointly model the multi-period overdue incremental effect and customer electricity consumption fluctuations, and predict the customer's default probability in the subsequent account period. The specific processing logic is as follows:

[0059] First, the overdue increment factor and the lag correlation factor are read from the feature matrix generated in step S2, and the electricity consumption fluctuation sequence extracted in step S1 is called at the same time, and these two parts of data are indexed and matched at the account period level to generate a multi-dimensional time series input sequence.

[0060] Then, a nonlinear time series analysis framework including wavelet coefficient decomposition is used to perform multi-scale decomposition of the electricity consumption fluctuation sequence and extract several groups of wavelet coefficients to capture the abnormal oscillation characteristics of customers' electricity consumption in the high-frequency and low-frequency ranges.

[0061] Then, the above wavelet coefficients are input into the deep backtracking neural network together with the overdue increment factor and the time lag correlation factor. The time layer unit of the deep backtracking neural network receives the multi-dimensional feature information of the previous several account periods at each step to explore the nonlinear impact of multi-period increments and large delayed overdue on the default risk of subsequent account periods. During the network training process, an objective function based on cross entropy is defined, and the predicted default probability is compared with the actual bill settlement situation account by account period. The back propagation algorithm is used to continuously update the parameters of each layer in the timing network, so that the model gradually learns to output more accurate default probabilities under different overdue modes and electricity consumption fluctuation modes. When the model training converges, the deep backtracking neural network is used to infer and predict the data of the new account period, and the corresponding default probability value is output, providing the necessary input basis for the dynamic risk accumulation mechanism of step S4.

[0062] Finally, the output default probability is indexed and stored according to the customer and account period, and the time series state of the model is retained in the database to ensure that dynamic iteration and risk assessment can be continued in subsequent steps in combination with the intermediate prediction results that have been generated.

[0063] By deeply coupling the multi-period overdue increment factor and time lag correlation factor generated in step S2 with the electricity consumption fluctuation characteristics obtained by wavelet decomposition in the same time series model, step S3 successfully constructs a prediction system that can comprehensively characterize the complex interactive relationship between overdue increment and abnormal electricity consumption oscillation. The system realizes refined default probability estimation at the account period level, significantly improves the forward-looking identification ability of risks such as multiple accumulation of small overdues and sudden large-amount delays in a single period, and provides a stable and high-precision default probability input for the dynamic risk accumulation and subsequent risk warning in step S4. By combining hierarchical wavelet decomposition with deep backtracking neural networks, this step effectively avoids the problem of traditional linear models ignoring complex time dependencies, and realizes the precise quantification of multi-scale, nonlinear and time-series interactions of default risks, bringing more forward-looking and adaptive technical support to the credit management system of power companies.

[0064] It is difficult to achieve dynamic monitoring of multi-period overdue behavior only through a single default probability prediction, especially when the customer has shown signs of increasing overdue in multiple account periods, the subsequent default risk will often accumulate exponentially. Previously, step S2 generated a feature matrix that depicts the cumulative trend of multi-period overdue, and step S3 output the default probability of each account period of the customer based on the feature matrix and electricity consumption fluctuation information. If the risk of overdue history in previous periods is not superimposed in the further evaluation process, it will be difficult to identify high-risk customers in time and intervene in advance. On this basis, step S4 introduces a dynamic risk accumulation mechanism. By combining the key results of steps S2 and S3, the cumulative effect of multi-period risks is recursively calculated, and customers with increasing overdue phenomena are given higher risk weights, thereby achieving dynamic correction and amplification of the customer's default probability, providing a highly sensitive risk quantification indicator for real-time monitoring and risk warning in subsequent steps.

[0065] Step S4: Set up a dynamic risk accumulation mechanism during the prediction process, recursively calculate the impact of previous overdue periods on subsequent default risks, and assign higher risk weights to customers with increasing overdue phenomena. The specific processing logic is as follows:

[0066] First, the predicted default probability of the customer in each account period is read from step S3, recorded as p(k); at the same time, the overdue increment factor corresponding to each customer in the same account period is read from the feature matrix generated in step S2, recorded as A(k). The overdue increment factor is used to quantify the degree of superposition of the customer's overdue amount or overdue days in the previous account period. Then, in order to reflect the amplifying effect of incremental overdue on subsequent risks, a dynamic risk accumulation amount R(k) is defined and calculated using the following recursive formula: R(k+1)=R(k)+[p(k) (1+γ×A(k))-p(k)]; where R(k) represents the dynamic risk accumulation at the kth account period, p(k) represents the predicted value of the default probability of the corresponding account period, A(k) represents the overdue increasing factor of the corresponding account period, and γ is a regulatory factor used to determine the amplification strength of the overdue increasing phenomenon; if the overdue increasing factor is 0, indicating that there is no sign of increasing overdue in the account period, then the impact of the predicted value of the default probability remains at the original level, and when the overdue increasing factor is large, the predicted value of the default probability will be raised to a high-order power, greatly amplifying the risk contribution to the subsequent account periods, thereby achieving a doubling response to the possibility of default when multiple small overdue periods occur consecutively.

[0067] When iterating, R(1)=0 is set from the starting account period and the dynamic risk accumulation of subsequent account periods is calculated in sequence until the predetermined forecast period is covered; when the dynamic risk accumulation exceeds the pre-set risk threshold, the corresponding customer is marked as a high-risk customer and a note is made in the internal registration form so that the real-time monitoring and early warning triggering of the subsequent step (S5) can give priority to such customers.

[0068] The final calculated dynamic risk accumulation value and the corresponding default probability prediction value are stored in the database, and the result after incremental overdue amplification is provided to the next step for more targeted risk intervention in the monitoring or early warning stage.

[0069] Through the above-mentioned dynamic risk accumulation mechanism, step S4 nonlinearly couples the default probability output by step S3 with the overdue incremental factor extracted by step S2, significantly enhancing the risk amplification effect on multiple periods of small overdues and single period of large overdues; this mechanism accumulates overdue risks in real time during multiple account period iterations. When there is an obvious incremental overdue trajectory in the customer's history, the predicted value of the default probability will increase significantly with the account period recursion, thereby achieving more sensitive and accurate risk identification in the time dimension. This cumulative risk characterization method effectively avoids the defect that single-period predictions are prone to diluting the impact of continuous overdues, and provides a reliable digital basis for rapid processing in the real-time monitoring and early warning process in subsequent steps, helping power companies to minimize financial losses caused by defaults.

[0070] Customer credit risk is dynamic and sudden, especially when there are multiple periods of incremental overdue payments and abnormal fluctuations in electricity consumption. It is impossible to achieve real-time reflection and precise control of risks by relying solely on static scoring. The previous step generates the recursive results of the customer's multi-period risk through a dynamic risk accumulation mechanism. Step S5 needs to be combined with a real-time monitoring window and a dynamic threshold mechanism to achieve sensitive capture of risks, instant updates of credit scores, and early warning marks for high-risk customers, providing data support for subsequent enterprise intervention and model optimization.

[0071] Step S5: Monitor the prediction results of the model in real time. Once it is found that the customer's overdue increase index and electricity consumption fluctuation index reach the risk threshold, the credit score update and risk warning will be triggered immediately. The specific processing logic is as follows:

[0072] When the dynamic risk accumulation exceeds the pre-set risk threshold, first, extract the latest risk value sequence from the dynamic risk accumulation in step S4, and call the power consumption fluctuation feature sequence in step S3 and the overdue increasing factor in step S2 to build a dynamic monitoring window. The monitoring window covers the data of the last n billing periods, records the sequence information of the dynamic risk accumulation, power consumption fluctuation amplitude and overdue increasing factor, and is recorded as W = {R(i), U(i), A(i)|i = k-n+1, ​​..., k}, where R(i) represents the dynamic risk accumulation of billing period i, U(i) represents the power consumption fluctuation amplitude, and A(i) represents the overdue increasing factor. Where n represents the number of billing periods in the monitoring window, and the monitoring window of n billing periods means that when the system makes a risk judgment in the current billing period k, it will trace back the dynamic risk accumulation, power consumption fluctuation amplitude and overdue increasing factor of the last n billing periods. For example, if n = 3, the system will use the characteristic data of billing periods k, k-1, and k-2 for calculation.

[0073] In order to realize dynamic risk determination, the real-time trigger factor Z(k) is defined as:

[0074] In this formula, the dynamic risk accumulation and the overdue incremental factor in the numerator strengthen the superposition effect of incremental risk through exponential mapping, while the power consumption fluctuation amplitude constructs a dynamic adjustment factor of risk sensitivity through the integral form of the denominator. The goal of the formula design is to comprehensively measure the multi-period overdue incremental and abnormal power consumption risks to avoid the dominant influence of a single factor on the judgment.

[0075] The dynamic threshold T(k) is set for the real-time trigger factor, and its calculation formula is: T(k) = μ+η·log(1+var(W)); where μ is the basic threshold, η is the fluctuation adjustment factor, and var(W) is the amplitude change of the characteristic sequence in the monitoring window. By dynamically adjusting the dynamic threshold, the system can adaptively adjust the threshold to adapt to the changing characteristics of different customer risk levels and account periods. When the real-time trigger factor exceeds the dynamic threshold, the system marks the current customer as high risk and immediately enters the next processing flow.

[0076] When the trigger factor exceeds the dynamic threshold, the system calls the credit score update module based on the real-time monitoring results. The credit score update module readjusts the score based on the customer's current risk level. The specific calculation is: S(k) = S(k-1) + δ·(Z(k)-T(k)); where S(k) is the credit score of the current account period, S(k-1) is the score of the previous account period, and δ is the score adjustment coefficient, which is used to control the adjustment range. The updated score and risk tag are written to the database at the same time, and an early warning notification is generated and pushed to the operation and maintenance and customer service teams. The warning content includes customer number, risk level, main risk factors and recommended handling measures.

[0077] The timestamp of the triggering event, the triggering factor value and the score change are recorded in the monitoring log, and the monitoring window data before the triggering is archived for subsequent model optimization and improvement of the risk identification mechanism.

[0078] Step S5 integrates core features such as multi-period overdue increments and power consumption fluctuations through real-time monitoring windows and dynamic threshold mechanisms, dynamically calculates trigger factors and adaptively adjusts judgment thresholds to ensure sensitive capture and real-time response to rising risks. By triggering credit score updates and high-risk markings, this step achieves accurate positioning and timely warning of risky customers, providing stable data support for subsequent risk management. At the same time, the recording and feedback mechanism of triggering events provides detailed basic data for the iterative optimization of the model, making the overall credit evaluation process more efficient and accurate, and injecting real-time and dynamic technical guarantees into the risk control and customer management of power companies.

[0079] If only a one-way prediction of the probability of default is made without comparing it with the actual payment performance of the subsequent account period, it will be difficult to timely correct the key factors within the risk identification model. The previous steps have achieved recursive estimation and real-time warning of the customer's probability of default through multiple periods of overdue incremental factors, electricity consumption fluctuation characteristics and dynamic risk accumulation mechanism. On this basis, step S6 compares the latest actual payment results with the model prediction value, calculates the prediction deviation and specifically corrects the influence coefficient of the overdue incremental factor in the feature matrix, gradually realizing the rolling optimization of the risk identification model and improving the accuracy and sensitivity of the overall credit evaluation.

[0080] Step S6: Compare the updated credit score with the customer's subsequent actual payment performance, and calibrate the overdue incremental factor weight in the feature matrix based on the prediction error and time lag index to achieve rolling optimization of the risk identification model. The specific processing logic is as follows:

[0081] First, the latest credit score and the actual payment performance of the customer in the new account period after the warning is triggered are obtained from step S5. If the customer is still overdue in the corresponding account period, the corresponding overdue duration and amount are recorded; if the customer pays on time, it is recorded as a normal account period. Then, in order to quantify the degree of deviation between the prediction and reality, the matching function G(k) is used to measure the difference between the model's predicted default probability in account period k and the actual result, which is defined as follows: in Indicates the actual overdue amount or number of overdue days for the corresponding account period. is the predicted value of the default probability for the corresponding account period. If the value of the matching function is positive and large, it means that the actual default level significantly exceeds the model prediction. If the value of the difference between the predicted default probability and the actual result is negative and large, it means that the model overestimates the default risk.

[0082] Then, in order to correct the impact of the overdue increment factor in the feature matrix on subsequent predictions, the correction coefficient C(k+1) is defined as the rolling correction amount for the overdue increment factor: where sign(G(k) is used to determine the direction of the deviation, and Γ is used to convert the observed overdue amount to the time lag level of the previous period. Combined, the correction amount for the overdue increment factor of the next account period is generated. ζ is the control coefficient, which determines the sensitivity of the correction. The correction coefficient is superimposed on the field where the overdue increment factor is located in the feature matrix, and the overdue increment feature value of the corresponding customer's subsequent account period is updated. This update process is implemented in real time in the model training database and triggers a small-scale incremental training or iterative update, so that the model can better adapt to the feedback of previous deviations in the prediction of subsequent account periods, thereby improving the flexible identification ability of continuous or temporary overdue.

[0083] Finally, the detailed information of this correction process (including the value of the difference between the predicted default probability and the actual result, the final result of the correction coefficient, and the updated overdue incremental factor value) is stored in the correction log table, and the model correction record of the current customer is displayed on the operation and maintenance end to provide a basis for subsequent operation and maintenance strategies or customer service communication. Through this rolling optimization process, the system can continuously track and correct the prediction deviation of the overdue factor, improve the adaptive ability to multi-period incremental overdue and time lag problems, and gradually enhance the accuracy and response speed of credit scoring in dynamic scenarios.

[0084] Step S6 compares the actual payment performance with the previous risk forecast, uses a matching function to measure the deviation, and uses the correction coefficient to roll over the overdue increment factor, so that the risk identification model can learn and respond to the latest overdue information more promptly. This self-feedback mechanism effectively prevents the continuous accumulation of misjudgments of overdue trends, ensuring that the model has updated overdue feature characterization capabilities when facing new account period data, so as to more accurately capture risk fluctuations in subsequent forecasts. Through this rolling optimization model, power companies can implement continuous iteration and adaptive correction of customer credit scores, so that the deviation between multi-period overdue predictions and real-time payment behaviors continues to narrow, providing more reliable technical support for credit management, and also providing more accurate decision-making basis for subsequent collection strategies and customer service plans.

[0085] Embodiment 2: Figure 2 The present invention provides a customer credit evaluation system based on multi-view features, including: a data integration module, a feature extraction module, a time series modeling module, a risk accumulation module, a real-time monitoring module, and a model optimization module;

[0086] Data integration module: obtains historical electricity bill payment records and electricity consumption cycle data from relevant platforms, and generates integrated time series data with timestamps through unified preprocessing. The output integrated time series data with timestamps is passed to the feature extraction module;

[0087] Feature extraction module: conducts cumulative analysis on the continuous overdue behavior of each customer in the integrated time series data, extracts the overdue incremental characteristic factors and time lag correlation factors based on the overdue delay length threshold, constructs a feature matrix that describes the cumulative risk of multi-period overdue, and transmits the generated feature matrix to the time series modeling module;

[0088] Time series modeling module: Using the feature matrix as input, combined with the customer's electricity consumption fluctuation characteristics, joint modeling is performed through the time series analysis model to predict the customer's default probability in the subsequent account period, generate a dynamic default probability prediction value, and output the default probability prediction value and key features to the risk accumulation module;

[0089] Risk accumulation module: Based on the default probability and overdue increment factor output by time series modeling, the dynamic risk accumulation mechanism is used to recursively infer the impact of previous overdue behaviors on subsequent risks, assign higher risk weights to high-risk customers, generate dynamic risk accumulation, and pass the dynamic risk accumulation and corresponding default probability prediction value to the real-time monitoring module;

[0090] Real-time monitoring module: real-time monitoring of dynamic risk accumulation and default probability prediction values, combined with the customer's overdue incremental factor and electricity consumption fluctuation characteristics, based on dynamic thresholds to identify high-risk customers, trigger credit score updates and risk warnings, and pass updated credit scores and high-risk customer tags to the model optimization module;

[0091] Model optimization module: compares the updated credit score with the customer's actual payment performance, calculates the prediction error and time lag index, and dynamically corrects the overdue incremental factor weights in the feature matrix through deviation quantification to achieve rolling optimization of the risk identification model.

[0092] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0093] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0094] It should be noted that, in this article, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including a..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0095] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A customer credit evaluation method based on multi-view features, characterized in that: Includes steps: S1: Obtain historical electricity bill payment records and electricity consumption cycle data from relevant platforms, and generate integrated time series data with timestamps through unified preprocessing; S2: Conduct cumulative analysis on the continuous overdue behavior of each customer in the integrated time series data, extract the overdue incremental characteristic factors and time lag correlation factors based on the pre-set overdue delay length threshold, and construct a characteristic matrix that can characterize the customer's multi-period overdue cumulative risk; S3: Taking the feature matrix as input, the time series analysis model is used to jointly model the multi-period overdue incremental effect and customer electricity consumption fluctuations to predict the customer's default probability in the subsequent account period; S4: A dynamic risk accumulation mechanism is set up during the forecasting process to recursively calculate the impact of previous overdue periods on subsequent default risks, and assign higher risk weights to customers with increasing overdue phenomena; S5: Monitor the prediction results of the time series analysis model in real time. Once the customer's overdue increase index and electricity consumption fluctuation index reach the risk threshold, the credit score update and risk warning will be triggered immediately. S6: Compare the updated credit score with the customer’s subsequent actual payment performance, and calibrate the weight of the overdue incremental factor in the feature matrix based on the prediction error and time lag indicators to achieve rolling optimization of the risk identification model.

2. The customer credit evaluation method based on multi-view features according to claim 1 is characterized in that: The specific processing logic of step S2 is as follows: First, filter out the overdue amount O of each customer in the consecutive account period from the time-stamped integrated data output in step S1. k and the number of days of delay L k , define the account period sequence as {k=1,2,...,n}; then, determine whether the account period is overdue based on the pre-set overdue delay threshold. If the number of delayed days in the account period k is greater than the overdue delay threshold, the corresponding account period is considered overdue and the overdue amount is calculated; then, in order to characterize the cumulative upward trend of overdue, use multi-layer incremental feature factors and nonlinear trigger mechanism: the first layer defines the difference feature Δ k =max(0,O k -O k-1 ), used to measure the increment of the overdue amount of the current account period compared with the previous account period; The second layer constructs the cumulative potential energy factor P through multi-step recursion k =P k-1 +f(Δ k , Δ k-1 , ..., Δ k-m ), where P0 = 0, m is the number of backtracking steps, and f(·) is a nonlinear activation function, which increases the cumulative potential factor when a small amount of overdue debt is detected. To reflect the hysteresis of the strong temporal correlation between accounts, a time lag correlation factor is used to distinguish between weak hysteresis areas and strong hysteresis areas: First, define the gene R k =g(L k )×(O k / O ref );O ref is the calibration value of the highest single-period overdue amount in the customer's history, g(·) is the piecewise logistic function; in order to accumulate the corresponding lag effect in multi-period time series, the secondary propagation mechanism H is introduced k =H k-1 +φ(R k , R k-1 , ..., R k-m );where φ(·) is a self-defined nonlinear transformation function, which maps the lag factors of the most recent account periods into a comprehensive lag risk quantity, which is used to capture the rapid growth of the customer's default risk when there is a significant delay in multiple consecutive periods; Finally, (P k , H k ) This pair of increasing and lagging factors is regarded as the key risk characterization indicators under the account period k, and is arranged together with other auxiliary features to form a feature matrix.

3. The customer credit evaluation method based on multi-view features according to claim 2 is characterized in that: The specific processing logic of step S3 is as follows: First, the overdue increment factor and the lag correlation factor are read from the feature matrix, and the electricity consumption fluctuation sequence extracted in step S1 is called at the same time. The two parts of data are indexed and matched at the account period level to generate a multi-dimensional time series input sequence. Then, a nonlinear time series analysis framework including wavelet coefficient decomposition is used to perform multi-scale decomposition on the electricity consumption fluctuation sequence and extract several groups of wavelet coefficients to capture the abnormal oscillation characteristics of customers' electricity consumption in the high-frequency and low-frequency ranges. Then, the wavelet coefficients, overdue increment factors, and time lag correlation factors are input into the deep backtracking neural network together. When the model training converges, the deep backtracking neural network is used to infer and predict the data of the new account period and output the corresponding default probability value.

4. The customer credit evaluation method based on multi-view features according to claim 3 is characterized in that: The specific processing logic of step S4 is as follows: First, the predicted default probability of the customer in each account period is read, recorded as p(k); at the same time, the corresponding overdue increment factor of each customer in the same account period is read from the feature matrix, recorded as A(k). The overdue increment factor is used to quantify the degree of superposition of the customer's overdue amount or overdue days in the previous account renewal period; Then, define the dynamic risk accumulation amount R(k) and use the following recursive formula to calculate: R(k+1)=R(k)+[p(k) (1+γ×A(k)) -p(k)]; where γ is a control factor used to determine the amplification strength of the overdue increase phenomenon; when iterating, starting from the starting account period, R(1)=0 and calculate the dynamic risk accumulation of subsequent account periods in sequence until the predetermined forecast period is covered; when the dynamic risk accumulation exceeds the pre-set risk threshold, the corresponding customer is marked as a high-risk customer.

5. The customer credit evaluation method based on multi-view features according to claim 4 is characterized in that: The specific processing logic of step S5 is as follows: First, the latest risk value sequence is extracted from the dynamic risk accumulation, and the power consumption fluctuation feature sequence and the overdue increasing factor are called to construct a dynamic monitoring window. The monitoring window covers the data of the last n billing periods, and records the sequence information of the dynamic risk accumulation, power consumption fluctuation amplitude and overdue increasing factor, which is recorded as W = {R(i), U(i), A(i)|i = k-n+1, ​​..., k}, where R(i) represents the dynamic risk accumulation of bill period i, U(i) represents the power consumption fluctuation amplitude, and A(i) represents the overdue increasing factor. In order to realize dynamic risk judgment, the real-time trigger factor Z(k) is defined as: A dynamic threshold T(k) is set for the real-time trigger factor, and its calculation formula is: T(k) = μ+η·log(1+var(W)); where μ is the basic threshold, η is the fluctuation adjustment factor, and var(W) is the amplitude change of the characteristic sequence in the monitoring window; when the trigger factor exceeds the dynamic threshold, the credit score update module is called according to the real-time monitoring results; the credit score update module readjusts the score based on the customer's current risk level, and the specific calculation is: S(k) = S(k-1)+δ·(Z(k)-T(k)); where S(k) is the credit score of the current account period, S(k-1) is the score of the previous account period, and δ is the score adjustment coefficient, which is used to control the adjustment range.

6. The customer credit evaluation method based on multi-view features according to claim 5 is characterized in that: The specific processing logic of step S6 is as follows: First, based on the latest credit score and the customer’s actual payment performance in the new billing period after the warning is triggered, if the customer is still overdue in the corresponding billing period, the corresponding overdue duration and amount will be recorded; If the customer pays the bill on time, it is recorded as a normal account period. Then, in order to quantify the degree of deviation between the prediction and reality, the matching function G(k) is used to measure the difference between the model's predicted default probability in account period k and the actual result, which is defined as follows: in Indicates the actual overdue amount or number of overdue days for the corresponding account period. is the predicted value of the default probability of the corresponding account period; then, in order to correct the impact of the overdue increasing factor in the characteristic matrix on subsequent predictions, the correction coefficient C(k+1) is defined as the rolling correction amount of the overdue increasing factor: where sign(G(k) is used to determine the direction of the deviation, and Γ is used to convert the observed overdue amount to the time lag level of the previous period. Combined, a correction amount for the overdue incremental factor of the next account period is generated, ζ is the control coefficient, which determines the sensitivity of the correction; the correction coefficient is superimposed on the field where the overdue incremental factor is located in the feature matrix, and the overdue incremental characteristic value of the corresponding customer's subsequent account period is updated.

7. A customer credit evaluation system based on multi-view features, used to implement the customer credit evaluation method based on multi-view features as described in any one of claims 1 to 6, characterized in that: include: Data integration module, feature extraction module, time series modeling module, risk accumulation module, real-time monitoring module, model optimization module; Data integration module: obtains historical electricity bill payment records and electricity consumption cycle data from relevant platforms, and generates integrated time series data with timestamps through unified preprocessing. The output integrated time series data with timestamps is passed to the feature extraction module; Feature extraction module: conducts cumulative analysis on the continuous overdue behavior of each customer in the integrated time series data, extracts the overdue incremental characteristic factors and time lag correlation factors based on the overdue delay length threshold, constructs a feature matrix that describes the cumulative risk of multi-period overdue, and transmits the generated feature matrix to the time series modeling module; Timing Modeling Module: Using the feature matrix as input, combined with the customer's electricity consumption fluctuation characteristics, a joint model is built through a time series analysis model to predict the customer's default probability in the subsequent account period, generate a dynamic default probability prediction value, and pass the output default probability prediction value and key features to the risk accumulation module; Risk accumulation module: Based on the default probability and overdue increment factor output by time series modeling, the dynamic risk accumulation mechanism is used to recursively infer the impact of previous overdue behaviors on subsequent risks, assign higher risk weights to high-risk customers, generate dynamic risk accumulation, and pass the dynamic risk accumulation and corresponding default probability prediction value to the real-time monitoring module; Real-time monitoring module: real-time monitoring of dynamic risk accumulation and default probability prediction values, combined with the customer's overdue incremental factor and electricity consumption fluctuation characteristics, based on dynamic thresholds to identify high-risk customers, trigger credit score updates and risk warnings, and pass updated credit scores and high-risk customer tags to the model optimization module; Model optimization module: compares the updated credit score with the customer's actual payment performance, calculates the prediction error and time lag index, and dynamically corrects the overdue incremental factor weights in the feature matrix through deviation quantification to achieve rolling optimization of the risk identification model.

Images