Credit risk assessment method and system based on artificial intelligence

Abstract

The invention relates to the technical field of financial science and technology and big data analysis, and discloses a credit risk assessment method and system based on artificial intelligence, and the method comprises the steps: collecting a total power load sequence and a fund invoice sales sequence of a target enterprise, and carrying out the preprocessing; decomposing the total power load by using a signal decomposition algorithm, removing non-productive components, and extracting a productive effective load sequence; calculating a cross-correlation function of the productive effective load and the fund invoice sales sequence, and determining an optimal production and sales phase lag parameter; performing time sequence alignment on the sequence by using the parameter, calculating an energy value conversion rate, constructing a reference interval of the same industry, and calculating a deviation degree of a target enterprise; and generating a consistency risk penalty factor based on the deviation degree, correcting the basic credit score, and outputting a dynamic credit score. According to the invention, by establishing the time sequence mapping relation between the physical energy consumption and the economic output, the abnormal departure of the enterprise operation data is effectively identified, and the accuracy of credit risk assessment is improved.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and big data analytics, specifically to a credit risk assessment method and system based on artificial intelligence. Background Technology

[0002] Commercial banks and supply chain finance institutions traditionally rely primarily on financial statements, bank statements, and business and tax data provided by manufacturing companies when conducting credit approvals and risk management. However, relying solely on financial data has limitations such as outdated information and susceptibility to manipulation. To improve the objectivity of assessments, the industry has begun to introduce electricity big data as an auxiliary verification method, attempting to reconstruct the true operating conditions of companies by comparing electricity output or electricity tax.

[0003] Despite the introduction of multidimensional data, existing credit risk assessment techniques still have certain limitations. Firstly, when using electricity data to extrapolate capacity, current methods typically use a company's total electricity consumption directly, failing to effectively distinguish between productive loads directly involved in product processing and manufacturing and maintenance loads such as lighting, air conditioning, and office equipment. Since maintenance loads are significantly affected by temperature, season, and personnel activity, this mixed data processing method results in energy consumption indicators containing a large amount of noise unrelated to actual capacity, failing to accurately reflect the company's true production activity.

[0004] Secondly, existing input-output analysis models typically align data based on absolute calendar time windows, assuming that a company's current energy consumption directly corresponds to its current sales revenue. This static analytical logic ignores the inventory turnover cycle, logistics and distribution cycle, and payment delays that are common in industrial production. Because there is an objective phase difference between the time of physical input and the time of economic output, direct horizontal comparisons can lead to data mismatch, thereby reducing the confidence of correlation analysis.

[0005] Furthermore, existing technologies lack a dynamic benchmark system based on specific industry segments when conducting data cross-validation. Due to differences in energy efficiency levels among enterprises of different sizes and with different processes, simple ratio calculations alone are insufficient to quantify the degree of abnormal deviation between a single enterprise's energy consumption and output value. This makes it difficult for assessment models to effectively identify highly concealed credit fraud behaviors such as fictitious revenue (high sales without actual production) or hidden actual capacity (high energy consumption and low sales), resulting in credit scores that fail to accurately reflect the enterprise's default risk. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a credit risk assessment method and system based on artificial intelligence, which solves at least one technical problem existing in the prior art.

[0007] To achieve the above objectives, the first aspect of the present invention provides a credit risk assessment method based on artificial intelligence, comprising the following steps:

[0008] S100, obtain the processed total power load sequence and the processed cash invoice sales sequence of the target enterprise;

[0009] S200, decompose the processed total power load sequence into productive load components and non-productive load components, and output the productive effective load sequence after removing the non-productive load components;

[0010] S300, calculate the cross-correlation function between the productive effective load sequence and the capital invoice sales sequence, search for the time delay corresponding to the maximum value of the cross-correlation function, and determine the optimal production and sales phase lag parameter;

[0011] S400, the productive load sequence is time-shifted and calibrated using the optimal production and sales phase lag parameter to obtain the calibrated productive load sequence. The ratio of the processed cash invoice sales sequence to the calibrated productive load sequence is calculated as the energy value conversion rate. A conversion rate benchmark interval for companies in the same industry is constructed, and the deviation of the target company's energy value conversion rate from the conversion rate benchmark interval is calculated.

[0012] S500: Calculate the consistency risk penalty factor based on the deviation, obtain the basic credit score of the target enterprise, apply the consistency risk penalty factor to adjust and correct the basic credit score, and generate and output the final dynamic credit score result.

[0013] Further, step S100 specifically includes: collecting the total power load sequence and the cash invoice sales sequence of the target enterprise, cleaning outliers in the total power load sequence and the cash invoice sales sequence of the target enterprise, and resampling and standardizing them according to a uniform time sampling frequency to obtain the processed total power load sequence and the processed cash invoice sales sequence.

[0014] The resampling and standardization process according to a uniform time sampling frequency specifically includes:

[0015] For the total power load sequence, a uniform aggregation window is set, and the average value and volatility of the load are calculated within each aggregation window to obtain a resampled total power load sequence with the same frequency as the fund invoice sales sequence;

[0016] Adaptive standardization is performed on the resampled total power load sequence and the fund invoice sales sequence, respectively. This standardization process converts the resampled total power load sequence and the fund invoice sales sequence into dimensionless standard scores by calculating the relative position of each data point with respect to the mean and standard deviation within the historical sliding window.

[0017] Furthermore, in step S200, before decomposing the processed total power load sequence into productive load components and non-productive load components using a signal decomposition algorithm, the method further includes:

[0018] Construct an electrical fingerprint database for production equipment, which includes the steady-state active power amplitude range and power factor range, start-up transient current waveform envelope, and high-order harmonic content of equipment directly involved in product processing and manufacturing.

[0019] Construct a maintenance equipment electrical fingerprint database, which includes the power factor and waveform stability characteristics, and intermittent start-stop characteristics of equipment maintaining the factory operating environment;

[0020] Based on the principle of circuit superposition, a composite load superposition model containing the states of production equipment and maintenance equipment is constructed using the electrical fingerprint database of the production equipment and the electrical fingerprint database of the maintenance equipment.

[0021] Further, in step S200, the processed total power load sequence is decomposed into productive load components and non-productive load components using a signal decomposition algorithm, specifically including:

[0022] Construct a sparse representation objective function for the load state space, and find a set of the sparsest combination of state vectors under the L1 norm constraint so as to minimize the error between the reconstructed load curve and the measured total power load sequence.

[0023] The preliminary device state activation matrix is ​​obtained by solving the objective function using a sparse coding algorithm.

[0024] A hidden Markov model is introduced to correct the temporal constraints of the device state activation matrix. The state transition probability matrix and Viterbi algorithm are used to remove short-term state jumps that do not conform to the physical operation logic, and the corrected state sequence is obtained.

[0025] Based on the corrected state sequence, the components belonging to the set of productive equipment are selected and their power is combined to obtain the productive effective load sequence.

[0026] Further, in step S300, calculating the cross-correlation function of the two sequences and searching for the time delay corresponding to the maximum value of the cross-correlation function specifically includes:

[0027] Set the length and step size of the sliding time window, and extract the production load subsequence and the fund invoice sales subsequence within the current window;

[0028] Within each sliding time window, calculate the normalized cross-correlation function between the productive load subsequence and the fund invoice sales subsequence;

[0029] Within a preset range of physical lag days, the search is performed to find the lag days that maximize the normalized cross-correlation function, generating a time-delay correlation map composed of a series of locally optimal lag parameters.

[0030] Furthermore, the method for determining the optimal production and sales phase lag parameter specifically includes:

[0031] The local optimal lag parameters in the time-delay correlation map are screened, and invalid sample points with correlation coefficient values ​​lower than the confidence threshold are removed.

[0032] A time decay weighting factor is introduced to perform weighted statistical analysis on the selected effective sample points. The sample points that are closer to the current evaluation time and have a larger correlation coefficient value have a higher weight.

[0033] The weighted average value is calculated as the optimal production and sales phase lag parameter for the entire cycle.

[0034] Furthermore, in step S400, constructing a benchmark range for conversion rates among companies in the same industry specifically includes:

[0035] Extract a set of sample companies in the same industry as the target company that have the same national economic industry classification code and similar electricity connection capacity;

[0036] Calculate the energy value conversion rate of each sample enterprise in the aforementioned industry sample enterprise set;

[0037] A density-based clustering algorithm was used to perform cluster analysis on the energy value conversion rate of all sample enterprises, and the core cluster with the highest sample density was identified as the mainstream interval.

[0038] Calculate the mean and standard deviation of the sample points within the core cluster, and generate the conversion rate benchmark interval containing upper and lower bounds based on the normal distribution assumption.

[0039] Further, in step S400, calculating the deviation of the target enterprise's energy value conversion rate from the conversion rate benchmark range specifically includes:

[0040] When the energy value conversion rate is lower than the lower limit of the conversion rate benchmark range, the ratio of the difference between the lower limit and the energy value conversion rate to the lower limit is calculated as the deviation in the first direction, which is used to characterize the risk of hidden income.

[0041] When the energy value conversion rate is higher than the upper limit of the conversion rate benchmark range, the ratio of the difference between the energy value conversion rate and the upper limit to the upper limit is calculated as the deviation in the second direction, which is used to characterize the risk of fictitious income.

[0042] When the energy value conversion rate is between the upper and lower boundaries of the conversion rate benchmark range, the deviation is determined to be zero.

[0043] Furthermore, step S500 specifically includes:

[0044] The deviation is converted into a dimensionless anomaly index using an exponential function;

[0045] A nonlinear risk mapping model is constructed based on the Sigmoid function, which maps the anomaly index to the initial risk probability.

[0046] Calculate the coupling confidence factor based on the maximum correlation coefficient value of the cross-correlation function;

[0047] The initial risk probability is weighted and corrected using the coupling credibility factor to generate a comprehensive consistency risk coefficient, wherein the lower the maximum correlation coefficient value, the lower the coupling credibility factor, and the higher the comprehensive consistency risk coefficient.

[0048] The consistency risk penalty factor in multiplicative decay form is calculated using the comprehensive consistency risk coefficient, and the basic credit score is multiplied by the consistency risk penalty factor to obtain the dynamic credit score result.

[0049] A second aspect of the present invention provides an artificial intelligence-based credit risk assessment system for performing any of the foregoing methods, the system comprising:

[0050] The multi-source heterogeneous data acquisition and preprocessing module is used to acquire the total power load sequence and the cash invoice sales sequence of the target enterprise, clean the outliers of the total power load sequence and the cash invoice sales sequence of the target enterprise, and resample and standardize them according to a uniform time sampling frequency to obtain the processed total power load sequence and the processed cash invoice sales sequence.

[0051] The productive effective load extraction module is used to receive the processed total power load sequence, decompose the processed total power load sequence into productive load components and non-productive load components using a signal decomposition algorithm, and output the productive effective load sequence after removing the non-productive load components.

[0052] The timing phase lag calculation module is used to receive the productive effective load sequence and the processed fund invoice sales sequence, calculate the cross-correlation function of the two sequences, search for the time delay corresponding to the maximum value of the cross-correlation function, and determine the optimal production and sales phase lag parameter.

[0053] The energy value conversion consistency verification module is used to perform time-series shift calibration on the productive effective load sequence using the optimal production and sales phase lag parameter to obtain the calibrated productive effective load sequence, calculate the ratio of the processed fund invoice sales sequence to the calibrated productive effective load sequence as the energy value conversion rate, construct a conversion rate benchmark interval for companies in the same industry, and calculate the deviation of the target company's energy value conversion rate from the conversion rate benchmark interval.

[0054] The dynamic credit score generation module is used to calculate the consistency risk penalty factor based on the deviation, obtain the basic credit score of the target enterprise, apply the consistency risk penalty factor to adjust the basic credit score downward, and generate and output the final dynamic credit score result.

[0055] This invention provides a credit risk assessment method and system based on artificial intelligence, which has the following beneficial effects:

[0056] This invention employs a signal decomposition algorithm based on sparse coding and hidden Markov models, combined with an electrical fingerprint database of equipment, to accurately extract non-productive load components that are significantly affected by environmental factors from the total power load. The productive effective load sequence extracted by this method directly reflects the actual processing and manufacturing activities of the enterprise, eliminating the interference of maintenance equipment such as lighting and air conditioning on energy consumption data, thereby improving the purity and accuracy of physical data sources in subsequent input-output analysis.

[0057] This invention introduces a time-series phase lag calculation mechanism. By calculating the cross-correlation function between productive load and sales sequence and applying time decay weighting, it adaptively determines the optimal production and sales phase lag parameters and performs sequence alignment. This corrects the logical bias in traditional evaluation methods that leads to a mismatch between current energy consumption and current sales data due to neglecting inventory turnover cycles, and establishes an objective mapping relationship between physical inputs and economic outputs over time.

[0058] This invention constructs a dynamic energy-value conversion rate benchmark interval based on industry clustering and calculates a consistency risk penalty factor using a two-way deviation index and a Sigmoid nonlinear risk mapping model. It can quantitatively identify operational anomalies of target companies relative to the industry mainstream, effectively distinguish between fictitious income and concealed income risks, and achieve effective verification of the authenticity of financial data using physical data through dynamic correction of the basic credit score. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the artificial intelligence-based credit risk assessment method according to an embodiment of the present invention.

[0060] Figure 2 This is a detailed flowchart illustrating the data acquisition and preprocessing process according to an embodiment of the present invention;

[0061] Figure 3 This is a detailed flowchart illustrating the extraction process of productive effective load according to an embodiment of the present invention;

[0062] Figure 4 This is a detailed flowchart illustrating the timing phase lag calculation process according to an embodiment of the present invention.

[0063] Figure 5 This is a detailed flowchart illustrating the energy value conversion consistency verification process according to an embodiment of the present invention.

[0064] Figure 6 This is a detailed flowchart illustrating the dynamic credit score generation process according to an embodiment of the present invention;

[0065] Figure 7 This is a schematic diagram of the architecture of an artificial intelligence-based credit risk assessment system according to an embodiment of the present invention.

[0066] Explanation of icon numbers:

[0067] 101. Multi-source heterogeneous data acquisition and preprocessing module; 102. Productive effective load extraction module; 103. Time series phase lag calculation module; 104. Energy value conversion consistency verification module; 105. Dynamic credit score generation module. Detailed Implementation

[0068] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] like Figure 1As shown, this embodiment of the invention provides a credit risk assessment method based on artificial intelligence, including the following steps:

[0070] S100, Collect the total power load sequence of the target enterprise. and fund invoice sales sequence The collected raw sequences are cleaned of outliers, and then sampled at a uniform time frequency. Perform resampling;

[0071] S200 will process the total power load sequence Input the productive load extraction module to construct a dictionary or probabilistic model containing the electrical characteristics of industrial equipment. Use a signal decomposition algorithm to decompose the total power load sequence into productive load components and non-productive load components, and output the productive load sequence after removing the non-productive components. ;

[0072] S300 will sequence productive effective loads. and fund invoice sales sequence Input the timing phase lag calculation module, set the sliding time window length, calculate the cross-correlation function of the two sequences within the window, search for the time delay corresponding to the maximum value of the cross-correlation function, and determine the optimal production and sales phase lag parameters. ;

[0073] S400, utilizing phase lag parameter For productive effective load series Perform time-shift calibration and calculate the sales output value corresponding to the unit effective energy consumption after calibration, i.e., the energy value conversion rate. Construct a benchmark range for conversion rates among companies in the same industry, and calculate the conversion rate of the target company. Deviation from the reference interval ;

[0074] S500, based on deviation Calculate the consistency risk penalty factor, obtain the target company's basic credit score, apply the risk penalty factor to adjust the basic credit score downward, and generate and output the final dynamic credit score result.

[0075] To further illustrate the specific implementation principles and technical details of each step in the embodiments of the present invention, the steps S100 to S500 will be described in detail below with reference to specific formulas and implementation methods.

[0076] like Figure 2 As shown, the following is a detailed description of the implementation method for the multi-source heterogeneous data access and preprocessing process in step S100.

[0077] When implementing the credit risk assessment method of this invention, the multi-source heterogeneous data acquisition and preprocessing module performs data access and basic cleaning operations, which specifically include the following steps:

[0078] Step S101: Collect the total power load sequence. The system acquires real-time power data through intelligent power metering terminals deployed on the incoming side of the target enterprise's distribution cabinet. The intelligent power metering terminals support Modbus-RTU or DL / T645 communication protocols, used to record power parameters at a high-frequency sampling rate on a minute-by-minute basis. The collected data dimensions specifically include three-phase active power, three-phase reactive power, RMS voltage and current values, and total power factor. For the total power load sequence... The numerical units are uniformly converted to kilowatts (kW). This represents discrete sampling time points. To ensure data integrity, the system is configured with a breakpoint resume mechanism, automatically retransmitting missing historical data segments when communication is restored.

[0079] Step S102: Collect the sales sequence of invoices. The system establishes an encrypted connection with the enterprise's tax system or ERP financial system through an Application Programming Interface (API). It periodically (e.g., daily at midnight) retrieves the target enterprise's VAT invoice issuance records or bank transaction records. The retrieved data fields include invoice date, invoice amount, buyer information, and product name. The system then sums up all valid invoice amounts within the same collection period (e.g., daily) to generate the sales sequence of invoices. All numerical values ​​are uniformly converted to the local currency. For cases involving red-ink invoices (negative invoices) or voided invoices, the system will perform corresponding deductions according to tax rules during the summary calculation to ensure... It reflects the true daily sales output.

[0080] Step S103: Data anomaly cleaning and missing value imputation. For the collected raw sequences, the system uses statistical methods to identify and process outliers. For the total power load sequence... The Raida criterion (3σ criterion) is used to detect and remove instantaneous spike noise caused by sensor malfunctions. For detected missing data points, if the missing duration is less than a preset threshold (e.g., 1 hour), linear interpolation is used for completion; if the missing duration exceeds the preset threshold, the year-on-year / month-on-month average substitution method is used, that is, load data from the same period of the previous day or the previous week is used for filling to maintain the temporal integrity of the sequence. The cleaned high-frequency power load sequence is output for subsequent step S200 for fine decomposition.

[0081] Step S104: Multi-source data resampling and time alignment. To meet the correlation analysis requirements in step S300, the differences in time granularity between power data and financial data need to be addressed. The system performs total power load sequence... Perform downsampling to calculate the daily average load or cumulative electricity consumption, thus improving its time resolution. Sales sequence with fund invoices Maintain consistency (e.g., unify to daily granularity). Using a unified timestamp index, logically align the two heterogeneous sequences on the same timeline to form a standardized dataset that can be used by subsequent modules. For data of different scales, perform Z-Score standardization separately, as shown in the following formula:

[0082] ;

[0083] ;

[0084] in, and These are the mean and standard deviation of the total electricity load sequence, respectively. and These are the mean and standard deviation of the sales series of financial invoices, respectively. and This is the standardized output sequence. This step eliminates the order-of-magnitude difference between physical and economic dimensions, preventing features with larger values ​​from dominating subsequent calculations.

[0085] After completing basic data cleaning and imputation, in order to eliminate the dimensional differences between the physical dimension (kilowatts, amperes) and the economic dimension (yuan, ten thousand yuan) and to solve the time alignment problem of data with different sampling frequencies, the multi-source heterogeneous data acquisition and preprocessing module further performs standardization and resampling operations. This process specifically includes the following steps:

[0086] Step S105: Time window aggregation and resampling of the high-frequency sequence. This involves resampling the total power load sequence sampled at the minute level. The system is based on the set uniform time granularity. (For example, 24 hours) Construct an aggregation window. Within each aggregation window, instead of simply calculating the mean, extract a statistical vector that reflects the load characteristics of the day. Specifically, for the... Calculate the daily average load for each time window. As a representative value for that day, the daily load fluctuation rate is also calculated. This is to preserve the stable electricity consumption characteristics of that day. The calculation formula is as follows:

[0087] ;

[0088] ;

[0089] in, This represents the total number of high-frequency sampling points within a single time window (e.g., 1440 minute points per day). Indicates the first day The load values ​​at each sampling time point. Through the above aggregation operation, the original high-frequency sequence is converted into a daily load sequence with the same frequency as the fund invoice sequence. For the fund invoice sales sequence If there are multiple transaction records on the same day, they are also aggregated and summarized at this time to ensure that the indices of the two sequences on the time axis correspond one-to-one.

[0090] Step S106: Adaptive Z-Score Standardization of Heterogeneous Data. Considering the significant differences in the operating scale of different enterprises, directly using absolute values ​​for calculation would lead to model bias for large-scale enterprises. Therefore, the system standardizes the resampled sequences, converting them into dimensionless standard scores. The system uses a sliding window statistical method to calculate the relative position of each data point with respect to its historical time window (e.g., the past 30 days). For any given time... resampled load value Its standardized value The calculation is as follows:

[0091] ;

[0092] in, and They respectively represent in the context of The deadline and the length are The mean and standard deviation within the sliding window; To prevent extremely small positive numbers with a denominator of zero (e.g., 10) ﹣6 Similarly, for the sales sequence of fund invoices... The same formula was used to process and obtain the result. This local standardization method, compared to global standardization, is better able to adapt to seasonal changes and long-term growth trends in business operations, preserving the local fluctuation characteristics of the data and providing more accurate waveform input for subsequent related analyses.

[0093] Step S107, Time Sequence Alignment and Synchronization Verification. After resampling and standardization, the system performs alignment verification on the timestamps of the two sequences. It checks for missing fund invoice sequences due to holidays or non-trading days. If such cases exist, the system marks the corresponding power load data as non-trading day load and specifically marks or removes it when calculating phase lag in subsequent calculations to prevent logical misjudgments caused by zero values ​​due to bank system shutdowns and normal factory production power consumption. The final output is a standardized load sequence with strictly aligned time indices. and standardized sales sequence Move to the next level module.

[0094] like Figure 3 As shown, before performing non-intrusive load decomposition, the productive effective load extraction module first establishes a feature model to distinguish between productive and non-productive loads. This process specifically includes the following steps:

[0095] Step S201: Construct an electrical fingerprint database for production equipment. The system pre-sets or establishes electrical feature templates for general industrial equipment through on-site data collection. Production equipment mainly refers to equipment directly involved in product processing and manufacturing, such as CNC machine tools, injection molding machines, stamping machines, and conveyor belt motors on assembly lines. Its electrical fingerprint features include steady-state and transient characteristics. Steady-state characteristics are defined as the active power amplitude range and power factor range of the production equipment during stable operation; transient characteristics are defined as the current waveform envelope and high-order harmonic content at the moment of startup or state transition of the production equipment. For example, for typical inductive production loads (such as three-phase asynchronous motors), there is an inrush current lasting from hundreds of milliseconds to several seconds at startup, accompanied by specific 3rd, 5th, and 7th harmonic components. The system parameterizes these features to construct the feature vector space of the production equipment. .

[0096] Step S202: Construct an electrical fingerprint database for sustaining equipment. Sustaining equipment refers to equipment that does not directly produce products but maintains the factory's operating environment, such as lighting systems, central air conditioning, office computers, and servers. The characteristics of this type of equipment are as follows: lighting systems exhibit purely resistive or capacitive characteristics, with a power factor close to 1.0 and stable waveforms; central air conditioning systems are controlled by a closed-loop system based on ambient temperature, exhibiting long-cycle intermittent start-stop cycles and lacking the short-term pulse characteristics strongly correlated with process cycle time found in production equipment. Based on this, the system constructs a feature vector space for sustaining equipment. .

[0097] Step S203: Establish a composite load superposition model. To extract the components from the total load, the system constructs a mathematical model based on the principle of circuit superposition. Assume that at any given time... Total power load It is by Production equipment and It is a linear superposition of a sustaining device and random noise. Define the first... The status of each production piece of equipment is as follows: (0 indicates shutdown, 1 indicates operation), rated power is ;No. The state of each maintenance device is Rated power is The mathematical expression for the composite load superposition model is:

[0098] ;

[0099] in, To measure noise and small, unmodeled loads.

[0100] This step solves for the state matrix using a subsequent algorithm. This allows for the identification of which power consumption belongs to This item. To improve recognition accuracy, the system constructed instantaneous power change rate features. As a supplementary criterion. Given the instantaneous power change rate of production equipment. The system exhibits a steep step response, while some sustaining devices (such as inverter air conditioners) show a smooth ramp response to power changes. The system utilizes this difference to set a threshold for the rate of power change. This threshold is not directly involved in the power superposition calculation, but is used as a subsequent state variable to distinguish between the two types of loads. and () constraints.

[0101] After establishing the model, the productive load extraction module uses an algorithm combining sparse coding and a Hidden Markov Model (HMM) to parse the productive load sequence from the total load sequence. This process specifically includes the following steps:

[0102] Step S204: Construct a sparse representation objective function for the load state space based on the composite load superposition model. Considering that within any given short time window, the number of devices undergoing state switching (on or off) within the factory is much smaller than the total number of devices, i.e., load changes exhibit sparsity, the system constructs an objective function to find the sparsest combination of state vectors that minimizes the error between the reconstructed load curve and the measured total power load curve. The mathematical expression of the optimization problem is as follows:

[0103] ;

[0104] in, The observed total load vector; It is a dictionary matrix, whose column vectors are composed of the normalized equipment electrical feature vectors constructed in steps S201 and S202; The device state activation matrix (corresponding to the device's activation level or operating status) is to be solved. This is the sparsity penalty parameter, whose value is usually determined by the cross-validation method to balance the reconstruction error with the sparsity of the solution; Represents the L2 norm; Let L1 norm be denoted. The objective function is solved using a sparse coding algorithm to obtain the preliminary device state activation matrix S. In this embodiment, the specific sparse coding algorithm can be the Orthogonal Matching Pursuit (OMP) algorithm or the Basis Pursuit algorithm.

[0105] Step S205: Temporal constraint correction based on Hidden Markov Model. Since sparse encoding only considers instantaneous features, noise can lead to misjudgments (e.g., misjudging random fluctuations in air conditioning units as the start-up of a low-power motor). The system introduces a Hidden Markov Model to perform temporal smoothing and correction of the state sequence. For each identified potential production equipment... Construct the corresponding HMM model .in, It represents the first The parameter set of a Hidden Markov Model (HMM) for a potential productive facility. The state transition probability matrix describes the physical probability of a production device going from shutdown to operation or from operation to shutdown (reflecting the typical working cycle of the production device; for example, a complete injection molding cycle of an injection molding machine is 45 seconds, so its state does not frequently jump within 1 second). This is the observation probability matrix, which describes the probability of observing a specific power value under a specific operating condition. Given the initial state distribution, the Viterbi algorithm is used to calculate the state sequence path. Eliminate short-term state transitions that do not conform to the physical operating logic.

[0106] Step S206, Productive Effective Load Sequence Synthesis. Based on the corrected state sequence. The system filters out all sets of production equipment. The components are summed and their power values ​​are combined to obtain the final productive effective load sequence. :

[0107] ;

[0108] in, This is a set of production equipment indexes, which is determined based on the production equipment electrical fingerprint database constructed in step S201, and includes all equipment indexes identified as directly involved in product processing and manufacturing. For the first Each device at time The corrected state value (takes a value of 0 or 1) is taken from the optimal state path corrected by the Viterbi algorithm in step S205. For the first The rated power of each production equipment is defined in the composite load superposition model in step S203.

[0109] At the same time, the system calculates the non-productive load sequence. As a verification:

[0110] ;

[0111] like If the data still contains periodic pulse components, a feedback mechanism is triggered, indicating that the feature dictionary needs to be updated. Through the above steps, the system successfully separated the maintenance load, which is greatly affected by environmental factors (temperature and sunlight), and retained only the effective energy consumption data that reflects the company's actual processing and manufacturing capabilities, providing a high signal-to-noise ratio input source for subsequent input-output analysis.

[0112] like Figure 4 As shown, high-purity productive payload sequences were extracted. Subsequently, the time-series phase lag calculation module performs a sliding window cross-correlation analysis to quantify the time correlation characteristics between production inputs and sales outputs. This process specifically includes the following steps:

[0113] Step S301: Define the sliding time window parameters. To capture the dynamic characteristics of the business cycle (e.g., fluctuations in inventory turnover due to peak and off-peak seasons), the system does not perform a one-time calculation using all data, but instead introduces a sliding window mechanism. The basic window length is set to... (e.g., 90 days), step size is (For example, 7 days). At any given time Extract the productive load subsequence within the current window. and the sales subsequence of fund invoices Window length The selection of the criterion must meet the statistical significance requirement, and can be set to cover at least 1.5 estimated production and sales cycles.

[0114] Step S302: Calculate the locally normalized cross-correlation function. Within each time window, the system calculates the cross-correlation function (CCF) between the two subsequences. To eliminate the influence of mean and variance differences on the correlation magnitude, a normalized cross-correlation algorithm is used. Define the time lag variable. Its value range is set to ,in This is the preset maximum physical lag days (e.g., 180 days, covering the inventory cycle of most industrial products). Cross-correlation function. The calculation formula is as follows:

[0115] ;

[0116] in, This indicates the length of the sliding time window (corresponding to the number of data points). This length is determined by multiplying a preset time span (e.g., 30 days) by the resampling frequency; This represents the index of the time sampling point within the window, with a value ranging from 0 to... ; Indicates the current number The cutoff time of each sliding window; Indicates at time The productive effective load value, which is derived from the data synthesized in step S206. The sequence reflects the actual production energy consumption of an enterprise at a specific point in history; Indicates the current number The arithmetic mean (mean) of the productive load sequence within a sliding window. Indicates at time The sales value of invoices. This data comes from the sales sequence of invoices processed in step S100. Note that lagged variables are introduced here. This indicates that the system is attempting to shift the sales data forward. Days, to match the timing of production; Indicates the introduction of lag Then, the arithmetic mean of the corresponding sales data window.

[0117] This formula calculates the time lag at a given time. The Pearson correlation coefficient of the two sequences after removing their respective DC components (mean) is used to quantify the similarity of their waveforms.

[0118] Step S303: Construct a time-delay correlation map. The system performs the above calculations for all sliding windows to generate a two-dimensional time-delay correlation matrix. The row index corresponds to the time progress of the sliding window, and the column index corresponds to the number of lag days. Elements in the matrix That is The value of . This matrix visually displays the trajectory of a company's production and sales cycle over time. For example, if the highly correlated areas (bright areas) in the matrix correspond to A value that gradually increases over time indicates that the company's inventory turnover is slowing down and it faces the risk of product backlog; conversely, if... Stable values ​​indicate that the company has good control over its production and sales rhythm. This graph is not only used for subsequent parameter optimization, but also stored in the database as a visualized intermediate analysis result.

[0119] After calculating the time delay correlation distribution, the time-series phase lag calculation module further performs parameter optimization to determine a globally optimal phase lag parameter that maximizes the logical matching degree between production and sales sequences. The process specifically includes the following steps:

[0120] Step S304: Extract the trajectory with the highest correlation. Based on the time-delay correlation matrix generated above. The system performs an extremum search for each row of the matrix (i.e., each time window). For the Within a time window, find the numbers that are mutually related. Number of days after reaching the maximum value :

[0121] ;

[0122] in, Indicates the first Locally optimal production and sales phase lag parameters (unit: days) calculated within a time window; The operator for maximizing the value of variables is called (Arguments of the Maxima). As the objective function for optimization, namely the normalized cross-correlation function calculated in step S302, the value of this function reflects the performance under the assumed lag. Under the condition of 10 days, the similarity in shape between the production curve and the sales curve; Represents the time-delay variable to be optimized; This indicates the preset physical lag search interval; It is usually set to 0 to indicate that production and sales are synchronized; The maximum physical lag days (e.g., 180 days) constrain the search boundary of the algorithm, preventing mathematical solutions that do not conform to common industrial sense (e.g., sales occur before production, or inventory cycles last for years).

[0123] Simultaneously record the corresponding maximum correlation coefficient value. This operation generates a discrete trajectory sequence consisting of a series of locally optimal hysteresis parameters. To eliminate artificially high correlations caused by chance factors (such as false coincidences due to specific holidays), the system sets a correlation confidence threshold. (e.g., 0.6). Only when At that time, the window calculates Only those points that are not included in the data are considered valid sample points; otherwise, the window is marked as a weakly correlated interval and will not be included in subsequent parameter statistics.

[0124] Step S305: Weighted estimation of the globally optimal phase lag parameter. To obtain a stable lag parameter representative of the assessment period, the system performs weighted statistical analysis on the effective sample point sequence. Considering that recent data has a more significant impact on the current credit assessment, the system introduces a time decay weighting factor. Global optimal phase lag parameter The calculation is as follows:

[0125] ;

[0126] in, The set of windows that satisfy the confidence threshold condition. Weights , The attenuation coefficient is... At the current assessment point, For window The point in time. Furthermore. As one of the weighting factors, this ensures that a highly correlated window contributes more to the results. This parameter... Physically, it represents the average inventory turnover days of a target company in its current operating phase, which is the average time interval from consuming electricity for production and processing to the sale and invoicing of finished products.

[0127] Step S306, Multi-peak distribution detection and anomaly early warning. While calculating the average lag, the system analyzes... The distribution pattern of the sequence. If The distribution exhibits a clear bimodal or multimodal characteristic (e.g., mainly concentrated in the 30-day and 90-day intervals), indicating that the company has multiple product lines with significantly different turnover cycles, or that its business model has undergone a structural change. In this case, the system will trigger a segmented evaluation mechanism, dividing the evaluation period into multiple sub-intervals and calculating the local optimal lag parameters for each, to avoid a single mean masking the actual operational volatility risk. For the case of a unimodal distribution, the output is directly... This serves as a unified phase calibration parameter for the entire lifecycle, which is then passed to the next level module.

[0128] like Figure 5 As shown, after determining the optimal phase lag parameter Subsequently, the energy value conversion consistency verification module performs time-series reconstruction of the data and calculation of core performance indicators. This process specifically includes the following steps:

[0129] Step S401: Perform phase shift calibration of the productive load sequence. To establish a direct mapping relationship between physical inputs and economic output over time, the system performs phase shift calibration on the productive load sequence. Perform a rightward time shift operation. Define the calibrated load sequence as follows: Its mathematical expression is:

[0130] ;

[0131] This operation will The production energy consumption that occurs at any given moment is transferred to its corresponding sales revenue collection moment. For the head gap data resulting from translation (i.e., the front...) (days), the system uses zero-padding or truncation to retain only the valid intervals where the two sequences overlap on the time axis. Further analysis will be conducted. Among them, This indicates the starting time of the original data sequence collection. This is the common starting time of the total power load sequence and the fund invoice sales sequence collected in step S100 (e.g., January 1, 2025). This indicates the cutoff time for the collection of the original data sequence, which is the end time for data collection in step S100 (e.g., December 31, 2025). This represents the best production-sales phase lag parameter. This parameter is calculated in step S300 (specifically S306) and represents the average physical turnover days (e.g., 30 days) from the start of production to the realization of sales revenue for the company.

[0132] This alignment step corrects the logical error in traditional assessment methods that leads to a mismatch between current energy consumption and current sales due to neglecting inventory turnover periods. For example, it prevents the misinterpretation of high energy consumption (stocking period) and low sales (off-season) in the current month as low energy efficiency.

[0133] Step S402: Calculate the cumulative energy and value conversion rate over the interval. Based on the aligned data sequence, the system calculates the total energy and value conversion rate within the evaluation period, denoted as... This indicator represents the economic value converted from a unit of effective production energy consumption, reflecting a company's technological added value level and market premium capability. The calculation formula is as follows:

[0134] ;

[0135] The numerator represents the total cumulative sales during the evaluation period, and the denominator represents the cumulative effective energy consumption corresponding to the production of these sold products. The system also calculates the discrete distribution characteristics of this indicator, generating a time series of the conversion rate. To monitor fluctuations in a company's operational efficiency at different points in time:

[0136] ;

[0137] This uses a sliding summation window. (For example, 30 days) are smoothed to avoid the randomness of daily sales data causing drastic fluctuations in the indicators.

[0138] Step S403: Eliminate logical discrepancies between production and sales logic and idling energy consumption during logic verification. In the calculation... During this process, the system synchronously performs logical verification. If within a certain time period... The value is consistently positive and relatively large, while the corresponding A value that remains consistently zero or extremely low can lead to localized... If the output approaches zero, the system will identify this phenomenon as high energy consumption with no output, corresponding to product backlog or production line idling for debugging; conversely, if... extremely low Very high, leading to local If the value is abnormally high, it is identified as high output without energy consumption, indicating the risk of mixed trade income (i.e., buying and selling non-self-produced products) or false sales invoices. The system records the start and end times and deviation range of these abnormal intervals, which are used as auxiliary features input into the subsequent risk scoring model.

[0139] The energy value conversion rate of the target enterprise was calculated. Subsequently, the energy value conversion consistency verification module constructs a density-based dynamic benchmark interval to quantitatively assess the deviation of the target company's operating data. This process specifically includes the following steps:

[0140] Step S404: Construct the feature space of peer samples. The system accesses the historical database and extracts a set of sample enterprises that have the same National Economic Industry Classification Code (SIC) and similar contracted electricity capacity as the target enterprise. To eliminate interference from abnormal operating samples, the system pre-filters out enterprises with records of power outages due to arrears or tax violations in the most recent assessment period. For the selected... Each of the sample companies calculated its energy value conversion rate after phase calibration using steps S300 to S402, and this result is denoted as the set. .

[0141] Step S405: Generate a dynamic baseline interval based on density clustering. Given the differences in operational levels among companies within the industry and the presence of outliers, a simple arithmetic mean cannot accurately reflect the mainstream industry level. This embodiment employs a density-based noise applied spatial clustering algorithm (DBSCAN) to cluster the set. Perform cluster analysis. Set the neighborhood radius. and the minimum number of samples contained in the core point Algorithm for traversing the set The system identifies core objects and expands them into several connectivity density clusters based on density reachability relationships. The cluster containing the largest number of samples is selected as the core cluster, representing the mainstream operating range within that specific industry segment. The mean conversion rate of all sample points within this core cluster is calculated. and standard deviation Based on the normal distribution assumption, a dynamic benchmark interval is constructed. :

[0142] ;

[0143] ;

[0144] in, The confidence coefficient is set to 1.96 to cover the 95% confidence interval. This baseline interval fluctuates dynamically as the sample data is updated, reflecting the impact of the macroeconomic environment on the overall input-output ratio of the industry.

[0145] Step S406: Calculate the two-way deviation index. The system will calculate the conversion rate of the target company. Mapping to the baseline interval A comparison is performed. To simultaneously capture both types of risk—fictitious revenue (high deviation) and inflated costs / hidden revenue (low deviation)—the system defines a two-way deviation degree. .when When the deviation is found to be below the industry average, it indicates that the enterprise's output per unit of energy consumption is lower than the industry average, suggesting extremely low production efficiency or that some products are sold without invoices (concealing income). In this case, the deviation is calculated as follows:

[0146] ;

[0147] when When this occurs, it indicates that the value of the enterprise's output per unit of energy consumption is higher than the industry average, suggesting an excessively high proportion of trade revenue or the issuance of false sales invoices (without actual production). In this case, the deviation is calculated as follows:

[0148] ;

[0149] when When this occurs, it indicates that the company is operating within its normal scope, and the deviation is within acceptable limits. .

[0150] Step S407, Deviation Normalization and Rating. To unify the deviation from different industry backgrounds into a standardized risk metric, the system normalizes and rates the calculated deviation. Perform nonlinear mapping. Set an industry tolerance threshold. The physical deviation is transformed into a dimensionless anomalous index using an exponential function. :

[0151] ;

[0152] in, This is the sensitivity adjustment coefficient. The range of values ​​is A higher value indicates a greater degree of discrepancy between the data and the actual situation, and a greater risk of doubt regarding the authenticity of the operational data. This indicator is directly used as input for the penalty item in the subsequent scoring model; This is the industry tolerance threshold. It is a preset non-negative empirical parameter (e.g., 0.1, which allows for a 10% baseline deviation). This indicates the deviation of the target enterprise's energy value conversion rate calculated in the preceding step S406.

[0153] like Figure 6 As shown, after obtaining the bidirectional deviation and anomaly index through the energy value conversion consistency test module, the dynamic credit scoring generation module further performs risk quantification and penalty factor calculation. This process transforms the degree of divergence between physical and economic data into a mathematical correction for the enterprise credit score, specifically including the following steps:

[0154] Step S501: Construct a nonlinear risk mapping function based on the Sigmoid function. Considering the nonlinear nature of credit risk's response to data deviations (i.e., small deviations are measurement errors and should not incur penalties; however, once the deviation exceeds a critical point, the risk probability rises sharply in an S-shape), the system is configured with a risk probability mapping model based on the Sigmoid function. The input variable is the anomaly index calculated in the previous steps. The output is the initial risk probability. The calculation formula is as follows:

[0155] ;

[0156] in, The growth rate parameter controls the steepness of the function curve near the critical point, reflecting the system's sensitivity to abnormal data. This represents the risk threshold, signifying the boundary of normal fluctuations the system can tolerate. The function maps the input to the (0, 1) interval, ensuring the normalization of the risk measurement.

[0157] Step S502 introduces a correction to the temporal coupling weight. Relying solely on the total conversion rate deviation is insufficient to fully capture the risk; therefore, the system further incorporates the maximum cross-correlation coefficient calculated in step S300. . This characterizes the degree of coupling between the production curve and the sales curve in terms of waveform shape. If A low level, even if the total amount matches, suggests a risk of companies engaging in ticket-buying schemes or disorderly operations. The system defines a coupled credibility factor. :

[0158] ;

[0159] in, The minimum acceptable correlation coefficient (e.g., 0.3) is used. The initial risk probability is weighted and corrected using a coupling credibility factor to obtain the comprehensive consistency risk coefficient. ,

[0160] To reflect the cumulative risks caused by morphological mismatch, the following formula is used:

[0161] ;

[0162] in, This is the penalty weight for shape mismatch. The formula indicates that when the waveform shapes of production and sales are highly uncorrelated (…), Approaching 0 will amplify the risk factor caused by the deviation in quantity.

[0163] Step S503: Calculate the final dynamic penalty factor. This is to incorporate the dimensionless risk coefficient. When applied to standard credit scoring systems (e.g., a percentage system), the system converts it into a multiplicative correction factor or an additive deduction item. In this embodiment, a multiplicative decay factor is used. Calculations are performed to dynamically adjust the baseline credit score:

[0164] ;

[0165] in, This is the maximum penalty coefficient (e.g., 0.4, meaning a maximum deduction of 40% of the credit score). This is the curvature adjustment parameter. If the calculated... If the value is below the preset circuit breaker threshold, the system will generate a mandatory warning signal. The final output... This will be used as a coefficient to directly affect the static financial credit score of enterprises, realizing a closed-loop assessment based on electronic verification of invoices and invoices to prove credit.

[0166] The calculated risk penalty factor reflects the physical and economic consistency. Subsequently, the dynamic credit scoring generation module combines the company's traditional static financial assessment results to perform the final credit score synthesis and grading process. This process specifically includes the following steps:

[0167] Step S504: Obtain the multi-dimensional basic credit score. The system first retrieves the basic credit score of the target company calculated based on traditional financial statements and business registration data. This score is constructed based on four dimensions: a company's solvency, profitability, operational efficiency, and growth potential. For the specific calculation model of the basic score, those skilled in the art can use well-known algorithms such as Logistic Regression or Random Forest, which will not be elaborated upon here. The value range is [0, 100]. To ensure the timeliness of the assessment, the system will verify the timestamp of the basic data to ensure that the financial data is the audit result of the most recent accounting period.

[0168] Step S505: Perform physical consistency correction calculation. The system will apply the risk penalty factor calculated in the previous steps. Effect on basic credit score This generates a dynamic credit score that integrates energy-side authenticity verification. The calculation formula is as follows:

[0169] ;

[0170] in, It is a rounding function; This is the historical default risk coefficient. If the target company has overdue electricity payment records, this coefficient is set to a non-zero value (e.g., 0.05); otherwise, it is 0. This formula achieves numerical correction for data distortion caused by relying solely on financial statements. When a company's production energy consumption and sales output are significantly out of sync (i.e., ... When the score is less than 1, even if the profit margin shown in its financial statements is high, the final score will be lower. This will also be significantly lowered, thus objectively reflecting its potential operational risks.

[0171] Step S506, Dynamic Credit Rating Mapping. To facilitate financial institutions' direct use of assessment results, the system constructs a non-linear mapping table from credit scores to credit ratings. Define the set of rating intervals. and the corresponding set of level tags The system is based on The final credit rating is determined by the range in which the credit rating falls. :

[0172] ;

[0173] During this process, the system introduces a forced degradation mechanism: if the duration of the false boom or idling detected in step S403 exceeds a preset severity threshold (e.g., 30 consecutive days), regardless of the calculated severity threshold... The higher or lower the numerical value, the more likely the credit rating will be forcibly downgraded to a high-risk warning level (such as C or D), with an additional risk label explanation.

[0174] Step S507: Track rating trends and generate reports. The system does not simply output a single rating, but rather... Store the data in a time-series database to construct the enterprise's credit evolution trajectory. Calculate the rate of change in credit scores. .like If the decline exceeds a preset slope threshold (e.g., a 10% week-on-week decrease), the system generates a dynamic risk warning report. The report includes the final score and the correction percentage. The system transmits key deduction items (such as excessively large phase lag in energy consumption sales or abnormally low conversion rates) and visualized time-lag correlation graphs to the downstream credit decision-making system through an encrypted interface, providing a quantitative basis for the dynamic adjustment of bank credit limits.

[0175] like Figure 7 As shown, this embodiment of the invention provides an artificial intelligence-based credit risk assessment system, which may include: a multi-source heterogeneous data acquisition and preprocessing module 101, a productive effective load extraction module 102, a time-series phase lag calculation module 103, an energy value conversion consistency verification module 104, and a dynamic credit score generation module 105.

[0176] The multi-source heterogeneous data acquisition and preprocessing module 101 connects to external data sources, including intelligent power metering terminals deployed at the enterprise site and digital interfaces of the tax management system. The multi-source heterogeneous data acquisition and preprocessing module 101 acquires high-frequency time-series power load data and low-frequency time-series financial invoice data of the target enterprise, and performs noise reduction, missing value imputation, and time granularity unification processing on the data to generate standardized total load sequences and sales data sequences.

[0177] The productive load extraction module 102 is connected to the multi-source heterogeneous data acquisition and preprocessing module 101. The productive load extraction module 102 receives the total load sequence and uses a non-intrusive load monitoring algorithm to decompose the total load sequence. By identifying preset electrical characteristics of industrial equipment, the productive load extraction module 102 separates and removes the maintenance load component from the total load sequence, extracting the productive load sequence that characterizes the actual production activities of the enterprise.

[0178] The time-series phase lag calculation module 103 is connected to the productive load extraction module 102 and the multi-source heterogeneous data acquisition and preprocessing module 101. The time-series phase lag calculation module 103 receives the productive load sequence and the sales data sequence. Based on the sliding window cross-correlation analysis method, the time-series phase lag calculation module 103 calculates the correlation distribution of the productive load sequence and the sales data sequence on the time axis, and determines the optimal production and sales phase lag parameters based on the maximum correlation coefficient. The phase lag parameter characterizes the inventory turnover cycle of the target enterprise from production input to sales output.

[0179] The energy value conversion consistency verification module 104 is connected to the time-series phase lag calculation module 103. The energy value conversion consistency verification module 104 is used to perform a time-axis shift operation on the productive effective load sequence based on the phase lag parameter, achieving logical time-series alignment between production data and sales data. The energy value conversion consistency verification module 104 calculates the aligned interval energy value conversion rate and compares this rate with the industry's dynamic benchmark interval to calculate the deviation index.

[0180] The dynamic credit scoring generation module 105 is connected to the energy value conversion consistency verification module 104. The dynamic credit scoring generation module 105 generates a risk penalty coefficient based on the deviation index. Combining the basic credit score based on static indicators, the dynamic credit scoring module 105 uses the risk penalty coefficient to correct the basic credit score and outputs the final dynamic credit score result.

[0181] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A credit risk assessment method based on artificial intelligence, characterized in that, include: S100, obtain the processed total power load sequence and the processed cash invoice sales sequence of the target enterprise; S200 decomposes the processed total power load sequence into productive load components and non-productive load components, and outputs the productive effective load sequence. S300, calculate the cross-correlation function between the productive effective load sequence and the capital invoice sales sequence, search for the time delay corresponding to the maximum value of the cross-correlation function, and determine the optimal production and sales phase lag parameter; S400, the productive load sequence is time-shifted and calibrated using the optimal production and sales phase lag parameter to obtain the calibrated productive load sequence. The ratio of the processed cash invoice sales sequence to the calibrated productive load sequence is calculated as the energy value conversion rate. A conversion rate benchmark interval for companies in the same industry is constructed, and the deviation of the target company's energy value conversion rate from the conversion rate benchmark interval is calculated. S500: Calculate the consistency risk penalty factor based on the deviation, obtain the basic credit score of the target enterprise, apply the consistency risk penalty factor to adjust and correct the basic credit score, and output the dynamic credit score result.

2. The method according to claim 1, characterized in that, Step S100 specifically includes: collecting the total power load sequence and the cash invoice sales sequence of the target enterprise; cleaning outliers from the total power load sequence and the cash invoice sales sequence of the target enterprise; and resampling and standardizing the sequence according to a uniform time sampling frequency to obtain the processed total power load sequence and the processed cash invoice sales sequence.

3. The method according to claim 1, characterized in that, Before step S200, the following are also included: Construct an electrical fingerprint database for production equipment; Construct a maintenance equipment electrical fingerprint database; Based on the principle of circuit superposition, a composite load superposition model is constructed using the electrical fingerprint database of the production equipment and the electrical fingerprint database of the maintenance equipment, which includes the state of the production equipment and the state of the maintenance equipment.

4. The method according to claim 3, characterized in that, Step S200 specifically includes: Based on the aforementioned composite load superposition model, a sparse representation objective function for the load state space is constructed. The objective function is solved using a sparse coding algorithm to obtain the device state activation matrix; A hidden Markov model is introduced to correct the temporal constraints of the device state activation matrix, resulting in a corrected state sequence. Based on the corrected state sequence, the components belonging to the set of productive equipment are selected and their power is combined to obtain the productive effective load sequence.

5. The method according to claim 1, characterized in that, The specific methods for determining the time delay in step S300 include: Set the length and step size of the sliding time window, and extract the production load subsequence and the fund invoice sales subsequence within the current sliding time window; Within each sliding time window, calculate the normalized cross-correlation function between the productive load subsequence and the fund invoice sales subsequence; Within a preset range of physical lag days, the search is performed to find the lag days that maximize the normalized cross-correlation function, generating a time-delay correlation map composed of a series of locally optimal lag parameters.

6. The method according to claim 5, characterized in that, The method for determining the optimal production and sales phase lag parameter specifically includes: The local optimal lag parameters in the time-delay correlation map are screened; A time decay weighting factor is introduced to perform weighted statistical analysis on the selected effective sample points; The weighted average value is calculated as the optimal production and sales phase lag parameter for the entire cycle.

7. The method according to claim 1, characterized in that, In step S400, a benchmark range for conversion rates among companies in the same industry is constructed, specifically including: Extract a set of sample companies in the same industry as the target company that have the same national economic industry classification code and similar electricity connection capacity; Calculate the energy value conversion rate of each sample enterprise in the aforementioned industry sample enterprise set; Cluster analysis was performed on the energy value conversion rate of all sample enterprises to identify the core cluster with the highest sample density as the mainstream interval; Calculate the mean and standard deviation of the sample points within the core cluster, and generate the conversion rate benchmark interval containing upper and lower bounds based on the normal distribution assumption.

8. The method according to claim 1, characterized in that, In step S400, the deviation of the target enterprise's energy value conversion rate from the conversion rate benchmark range is calculated, specifically including: When the energy value conversion rate is lower than the lower limit of the conversion rate benchmark range, the ratio of the difference between the lower limit and the energy value conversion rate to the lower limit is calculated as the deviation in the first direction. When the energy value conversion rate is higher than the upper limit boundary of the conversion rate benchmark range, the ratio of the difference between the energy value conversion rate and the upper limit boundary to the upper limit boundary is calculated as the deviation in the second direction. When the energy value conversion rate is between the upper and lower boundaries of the conversion rate benchmark range, the deviation is determined to be zero.

9. The method according to claim 1, characterized in that, In step S500, the consistency risk penalty factor is applied to lower the basic credit score, specifically including: The deviation is converted into a dimensionless anomaly index; A nonlinear risk mapping model is constructed to map the anomaly index to the initial risk probability; Calculate the coupling confidence factor based on the maximum correlation coefficient value of the cross-correlation function; The initial risk probability is weighted and corrected using the coupling credibility factor to generate a comprehensive consistency risk coefficient. The consistency risk penalty factor in multiplicative decay form is calculated using the comprehensive consistency risk coefficient, and the basic credit score is multiplied by the consistency risk penalty factor to obtain the dynamic credit score result.

10. A credit risk assessment system based on artificial intelligence, characterized in that, The system includes a functional module that performs the method according to any one of claims 1-9.

Images