A method and system for intelligent verification and deviation correction of electricity sales daily clearing results

Abstract

The application discloses a kind of intelligent verification and deviation correction method and system of power selling day clearing result.The method collects power selling day clearing data, which is decomposed into metering, price, loss and settlement rules and other business components and evaluated uncertainty, according to tail thickness, divided into light tail type or near Gaussian type component and heavy tail type component;Through first-order Taylor expansion to obtain sensitivity coefficient, saddle point approximation second-order semi-analytic expansion is used for heavy tail component to obtain second-order cumulant curvature;Spearman dependence matrix is constructed, and positive definite correlation matrix is generated by ridge parameter disturbance and normalization, and combined with standard deviation to form covariance matrix, calculate the combined standard uncertainty and the corrected component degree of freedom, generate asymmetric verification threshold;When day clearing result is over threshold, based on contribution degree vector, deviation is counter-distributed, and kalman filter with residual auto-regression is used to update component, and output correction result.

Description

Technical Field

[0001] This application belongs to the field of power transaction settlement data processing technology, and in particular relates to a method and system for intelligent verification and deviation correction of daily power sales results. Background Technology

[0002] Electricity sales daily reconciliation data comes from a wide range of sources and is highly interrelated to various business operations, typically encompassing multiple dimensions such as metering, pricing, losses, and settlement rules. The daily reconciliation accounting process is affected by a confluence of factors, including meter measurement errors, market spot price fluctuations, time-varying grid losses, and frequent adjustments to settlement rules, resulting in significant randomness and uncertainty. Existing daily reconciliation verification mechanisms often employ deterministic calculation models or fixed tolerance ranges, making it difficult to accurately characterize the compound deviations formed by the accumulation of errors from multiple data sources. This can easily lead to false alarms, exceeding limits, or missed anomalies in the daily reconciliation results.

[0003] Existing methods typically combine uncertainty propagation theory with filtering techniques. This involves first conducting a preliminary uncertainty assessment of various business data sources, then synthesizing a total error interval based on classical first-order Taylor expansion and the standard uncertainty propagation law, using this interval as a verification threshold. When data exceeds the threshold, a basic Kalman filter state estimation algorithm is used to smooth and correct the deviation data. However, these propagation models usually assume that each uncertainty component follows an ideal symmetric Gaussian distribution, easily neglecting the heavy-tailed distribution characteristics caused by sudden price changes and losses in actual electricity sales scenarios. Furthermore, they fail to fully extract higher-order cumulative information from heavy-tailed fluctuations, resulting in a fixed and symmetrical confidence interval that is difficult to adapt to real business fluctuations. In addition, when dealing with the dependence of multiple business components, existing methods often use simple linear correlation coefficients, which are insufficient to characterize nonlinear or monotonic dependence characteristics and may lead to non-positive definite covariance matrices due to numerical errors, thus affecting the stability of the synthesized standard uncertainty calculation. When the deviation correction is triggered, the process noise and observation noise parameters of the conventional filtering algorithm are mostly fixed based on experience. They fail to be effectively correlated with the true contribution of each business component to the total deviation and lack characterization of the autoregressive characteristics of the correction residual sequence. This can easily lead to unreasonable reverse allocation correction coefficients and slow convergence of state updates, making it difficult to achieve stable and accurate correction of the daily electricity sales results. Summary of the Invention

[0004] To improve the quality of daily electricity sales data correction and the reliability of settlement results, this invention provides an intelligent verification and deviation correction method and system for daily electricity sales results.

[0005] According to one aspect of the present invention, a method for intelligent verification and deviation correction of daily electricity sales results is provided, the method comprising:

[0006] Collect daily electricity sales data, decompose the daily electricity sales data into metering business components, price business components, loss business components and settlement rule business components, and evaluate the corresponding uncertainty components. According to the tail thickness index, the uncertainty components are divided into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components. The sensitivity coefficient of each business component is obtained by using a first-order Taylor expansion. The second-order cumulant curvature is obtained by using a saddle point approximation second-order semi-analytical expansion for heavy-tailed uncertainty components.

[0007] Construct a dependency matrix with Spearman coefficients as elements. Apply ridge parameter perturbation to the diagonal elements of the dependency matrix and reconstruct a positive definite correlation matrix. Scale the matrix with the standard deviation of each uncertainty component to a positive definite covariance matrix. Substitute the covariance matrix into the transfer formula to calculate the combined standard uncertainty. Use the Welch-Satterthwaite formula to correct the expansion factor. Feed back the ratio of the second-order cumulant curvature to the square of the corresponding standard uncertainty to the corrected component degrees of freedom to generate a verification threshold with asymmetric confidence intervals.

[0008] When the daily electricity sales result exceeds the verification threshold, a correction is triggered. The contribution magnitude vector is constructed by the square of the product of the sensitivity coefficient of each business component and the corresponding standard uncertainty. The correction direction is determined by combining the sign of the deviation of the daily electricity sales result from the verification threshold and the sign of the sensitivity coefficient of each business component. The deviation is then distributed in reverse to each business component. A Kalman filter with residual autoregression is used to update the business components. The process noise is scaled by the root mean square of the contribution vector, and the observation noise is adjusted by the out-of-bounds amplitude. The correction result is then output.

[0009] According to another aspect of the present invention, an intelligent verification and deviation correction system for daily electricity sales results is provided, the system comprising the following modules:

[0010] The solution module is used to collect daily electricity sales data, decompose the daily electricity sales data into metering business components, price business components, loss business components and settlement rule business components, and evaluate the corresponding uncertainty components. According to the tail thickness index, the uncertainty components are divided into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components. The sensitivity coefficient of each business component is obtained by using a first-order Taylor expansion. For the heavy-tailed uncertainty components, a saddle point approximation second-order semi-analytical expansion is used to obtain the second-order cumulative curvature.

[0011] The generation module is used to construct a dependency matrix with Spearman coefficients as elements. After applying ridge parameter perturbation to the diagonal elements of the dependency matrix, it reconstructs a positive definite correlation matrix and scales it to a positive definite covariance matrix by combining the standard deviation of each uncertainty component. The combined standard uncertainty is calculated by substituting it into the transfer formula. The expansion factor is corrected using the Welch-Satterthwaite formula. The ratio of the curvature of the second-order cumulant to the square of the corresponding standard uncertainty is fed back to the corrected component degrees of freedom to generate a verification threshold with asymmetric confidence intervals.

[0012] The correction module is used to trigger correction when the daily electricity sales result exceeds the verification threshold. It constructs a contribution magnitude vector by multiplying the sensitivity coefficient of each business component by the square of the product of the corresponding standard uncertainty. It determines the correction direction by combining the sign of the deviation of the daily electricity sales result from the verification threshold and the sign of the sensitivity coefficient of each business component. The deviation is then distributed in reverse to each business component. The business components are updated using a Kalman filter that includes residual autoregression. The process noise is scaled by the root mean square of the contribution vector, and the observation noise is adjusted by the out-of-bounds magnitude. The correction result is then output.

[0013] This invention refines and decomposes daily electricity sales data, classifies uncertainty components based on tail thickness indices, and employs a saddle-point approximation second-order semi-analytical expansion for heavy-tailed data, enhancing the model's ability to characterize heavy-tailed fluctuations such as sudden price changes and sudden losses. By applying ridge parameter perturbations to the Spearman dependency matrix and generating a positive-definite covariance matrix, the stability and reliability of uncertainty synthesis calculations are improved. An asymmetric confidence interval verification threshold is constructed using the Welch-Satterthwaite formula and curvature feedback mechanism, making the threshold setting more consistent with the actual fluctuation characteristics of electricity sales data. A contribution vector is constructed using sensitivity coefficients and uncertainties to achieve quantitative location and reverse allocation of deviation sources. A Kalman filter including residual autoregressive terms is used, combined with process noise scaling and observation noise adjustment mechanisms to suppress noise interference, thereby improving the correction accuracy of daily electricity sales data and the reliability of settlement results. Attached Figure Description

[0014] Figure 1 A flowchart of the first embodiment;

[0015] Figure 2 A comparison of the probability density distributions of the uncertainty components in light-tailed or near-Gaussian and heavy-tailed uncertainties;

[0016] Figure 3 This is a real-time comparison chart of the daily electricity sales settlement value and the asymmetric verification threshold. Detailed Implementation

[0017] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0018] In the first embodiment, the present invention proposes an intelligent verification and deviation correction method for daily electricity sales results, such as... Figure 1 As shown, it includes:

[0019] S1. Collect daily electricity sales data, decompose the daily electricity sales data into metering business components, price business components, loss business components, and settlement rule business components, and evaluate the corresponding uncertainty components. According to the tail thickness index, the uncertainty components are divided into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components. The sensitivity coefficient of each business component is obtained by using a first-order Taylor expansion. The second-order cumulant curvature is obtained by using a saddle-point approximation second-order semi-analytical expansion for the heavy-tailed uncertainty components.

[0020] By calling the power system's monitoring and data acquisition interface, or using structured query language, the daily electricity sales data, including daily user electricity consumption, real-time active and reactive power, time-of-use pricing, catalog pricing, line loss rate, estimated technical line loss, estimated management line loss, tiered pricing base, deviation assessment rate, and peak-valley fluctuation ratio, can be extracted from the marketing business application system database, metering and acquisition system, line loss management system, and electricity price rule configuration table.

[0021] The extracted data is loaded and broken down according to business dimensions to obtain the metering business component, which includes real-time active and reactive power collected by smart meters; the pricing business component, which includes time-of-use pricing and catalog pricing; the loss business component, which includes estimated values ​​of technical and management line losses; and the settlement rule business component, which includes deviation assessment rates and peak-valley fluctuation ratios.

[0022] For measurement and loss business components with sufficient historical observation data, the sample standard deviation is calculated, and Type A uncertainty assessment is performed. For price and settlement rule business components lacking historical observation data, the square root of variance is calculated using the uniform distribution variance calculation formula based on the preset maximum error limit and uniform distribution assumption, and Type B uncertainty assessment is performed to obtain the uncertainty components corresponding to each business component. For uncertainty components with historical observation samples, kurtosis or excess kurtosis is calculated based on the historical samples; for uncertainty components lacking historical observation samples and assessed using Type B, the theoretical kurtosis of the preset distribution is used, or a substitute sample is constructed based on rule change records and price boundary disturbance ranges, and then the tail thickness is assessed; when no substitute sample exists, the component is treated as a light-tailed or regular uncertainty component and is not included in the heavy-tailed determination.

[0023] Calculate the kurtosis value of the historical error sample, historical residual sample, or substitute disturbance sample corresponding to each business component, or use the theoretical kurtosis value of a preset probability distribution as a tail thickness indicator. Set the standard kurtosis threshold to 3. If the calculated kurtosis value is less than or equal to 3, the corresponding uncertainty component is determined to be a light-tailed or near-Gaussian uncertainty component; if the kurtosis value is greater than 3, it is determined to be a heavy-tailed uncertainty component.

[0024] Calculate the first-order partial derivatives of the comprehensive electricity sales fee function that includes the above four business components, and substitute the actual values ​​of each component in the current verification period into the partial derivative expression to construct the Jacobian matrix. Extract each element in the matrix as the sensitivity coefficient of the corresponding business component.

[0025] For data sequences divided into heavy-tailed uncertainty components, standardization and finite-interval truncation are first performed. An empirical cumulative generating function is constructed based on the truncated empirical samples. When this cumulative generating function exists and is differentiable within the solution interval, the saddle-point equation of the cumulative generating function is solved to obtain the saddle point. Using the first and second derivatives, the second derivative value of the cumulative generating function at the saddle point is calculated. This second derivative value is then used for a semi-analytic expansion to obtain the curvature of the second cumulant of the probability density distribution.

[0026] In some embodiments, dividing the uncertainty components into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components based on the tail thickness index includes:

[0027] Calculate the kurtosis value of the historical error sample, historical residual sample or substitute disturbance sample corresponding to each business component, or use the theoretical kurtosis value of the preset probability distribution as the tail thickness index.

[0028] Set a standard kurtosis threshold and determine whether the kurtosis value or theoretical kurtosis value is greater than the standard kurtosis threshold;

[0029] If it is not greater than the standard kurtosis threshold, the corresponding uncertainty component is determined to be a light-tailed or near-Gaussian uncertainty component.

[0030] If the uncertainty exceeds the standard kurtosis threshold, the corresponding uncertainty component is classified as a heavy-tailed uncertainty component.

[0031] For each collected business component, corresponding historical error samples, historical residual samples, or substitute disturbance samples, the sample kurtosis is calculated by the ratio of the fourth central moment to the square of the second central moment. The excess kurtosis is obtained by subtracting 3 from the sample kurtosis, serving as a tail thickness indicator. An excess kurtosis threshold is set between 0.5 and 1.5, preferably 1.0. If the excess kurtosis is not greater than this threshold, the corresponding uncertainty component is determined to be a light-tailed or near-Gaussian uncertainty component; if the excess kurtosis is greater than this threshold, it is determined to be a heavy-tailed uncertainty component. A comparison of the probability density distributions of light-tailed or near-Gaussian and heavy-tailed uncertainty components is shown below. Figure 2 As shown.

[0032] In some embodiments, the step of using a first-order Taylor expansion to obtain the sensitivity coefficients of each service component, and using a saddle-point approximation second-order semi-analytical expansion to solve for the heavy-tailed uncertainty component to obtain the second-order cumulant curvature includes:

[0033] For each business component, establish the first-order partial derivative function of the electricity sales daily clearing model, and substitute the expected value of each business component into the partial derivative function to obtain the corresponding first-order sensitivity coefficient.

[0034] For the heavy-tailed uncertainty components, construct the cumulative generating function of the heavy-tailed uncertainty components, solve the saddle point equation to obtain the saddle point, and take the second derivative of the cumulative generating function at the saddle point to obtain the corresponding second-order cumulant curvature.

[0035] Based on the accounting logic of daily electricity sales, a functional relationship expression of the output variable with respect to the input variable is constructed, and the first-order partial derivatives are calculated to obtain the analytical sensitivity matrix. The historical average of each business component over the past 30 days is extracted as the expected value and substituted into the partial derivative function. For example, when the average price component is 0.52 yuan / kWh and the average metering component is 10000 kWh, the first-order sensitivity coefficient of the metering component is calculated to be 0.52. This process completes the sensitivity detection of the overall system with respect to the linear changes of each variable. For uncertainty components determined to be heavy-tailed through kurtosis testing, the moment generating function of the probability density function is constructed using Laplace transform, and the natural logarithm is used to generate the cumulative generating function. The first-order derivative of this cumulative generating function is set equal to the target state value to establish a saddle-point equation. The Newton-Raphson iteration method is used with a maximum iteration count of 50 and an error tolerance of [missing information]. Solving the equation under the given conditions yields real roots, which are the saddle points. Substituting the coordinates of the saddle points into the expression for the second derivative of the cumulative generating function, the result is the curvature of the second cumulant, such as an output of 14.62. This curvature represents the effect of tail fluctuations in a heavy-tailed distribution on the amplification of the combined uncertainty.

[0036] S2. Construct a dependency matrix with Spearman coefficients as elements. Apply ridge parameter perturbation to the diagonal elements of the dependency matrix and reconstruct a positive definite correlation matrix. Combine the standard deviation of each uncertainty component and scale it to a positive definite covariance matrix. Substitute it into the transfer formula to calculate the combined standard uncertainty. Use the Welch-Satterthwaite formula to correct the expansion factor. Feed back the ratio of the curvature of the second-order cumulant to the square of the corresponding standard uncertainty to the corrected component degrees of freedom to generate a verification threshold with asymmetric confidence intervals.

[0037] Extract the historical time series of each business component, calculate the Spearman coefficient between each pair of business component time series, combine all correlation coefficients, and construct an initial dependency matrix.

[0038] Calculate the eigenvalues ​​of the initial dependency matrix and perform eigenvalue decomposition. As a basic implementation, if negative eigenvalues ​​exist, a ridge parameter is uniformly added to the main diagonal elements of the initial dependency matrix. This ridge parameter is taken as the absolute value of the smallest negative eigenvalue plus a minimal constant. The reconstructed positive definite correlation matrix is ​​generated, ensuring that all eigenvalues ​​are greater than 0.

[0039] Extract the standard deviations of each uncertainty component calculated previously and construct a standard deviation diagonal matrix. Perform matrix multiplication operations on the standard deviation diagonal matrix, the positive definite correlation matrix, and the standard deviation diagonal matrix consecutively to obtain the positive definite covariance matrix.

[0040] The obtained sensitivity coefficients are used to construct a sensitivity vector. The uncertainty propagation formula is used to calculate the product of the transpose matrix of the sensitivity vector, the positive definite covariance matrix, and the sensitivity vector. The square root of the product is then taken to obtain the combined standard uncertainty.

[0041] Before calculating the overall effective degrees of freedom, the strength of dependency is determined based on the off-diagonal correlation coefficient of the positive definite correlation matrix, and the equivalent components participating in the Welch-Satterthwaite formula calculation are identified: when the dependency is weak, the original business components are used as equivalent components; when the dependency is strong, the principal components obtained from the eigenvalue decomposition of the covariance matrix are used as equivalent components. Then, the degrees of freedom or equivalent degrees of freedom of each equivalent component are obtained, and the Welch-Satterthwaite formula is called to calculate the initial effective degrees of freedom, which is calculated by dividing the fourth power of the combined standard uncertainty by the sum of the ratios of the products of the standard uncertainties of each sensitivity coefficient to their respective independent degrees of freedom. For terms containing heavy-tailed uncertainty components, extract the previously calculated second-order cumulant curvature κi, divide it by the sum of the squared standard uncertainty of the heavy-tailed uncertainty component and the stability constant, and obtain the curvature correction coefficient ri, i.e. ;according to The corrected component degrees of freedom are obtained, where νi is the initial degree of freedom, νmin is the lower limit of the minimum degree of freedom, and α is the curvature adjustment coefficient; then the corrected component degrees of freedom νi′ are substituted into the Welch-Satterthwaite formula to calculate the overall effective degrees of freedom.

[0042] Based on the corrected effective degrees of freedom and the set 95% confidence level, the expansion factor is obtained by calling the 0.975 quantile function of the t-distribution in the two-sided verification scenario; the expansion factor is also obtained by calling the 0.95 quantile function of the t-distribution in the one-sided verification scenario. The product of the expansion factor and the combined standard uncertainty is used as the basic confidence half-width, and the upper and lower half-width coefficients are determined based on the weighted skewness γs of the heavy-tailed uncertainty components. Let the basic confidence half-width be U, and the skewness adjustment coefficient be ρ, then the upper and lower half-widths are respectively: , When γs is positive, the upper threshold is wider than the lower threshold; when γs is negative, the lower threshold is wider than the upper threshold, thus generating upper and lower check thresholds with asymmetric confidence intervals.

[0043] In some embodiments, the construction of a dependency matrix with Spearman coefficients as elements, the application of ridge parameter perturbations to the diagonal elements of the dependency matrix to reconstruct a positive definite correlation matrix, and the scaling to a positive definite covariance matrix by combining the standard deviations of each uncertainty component, includes:

[0044] Calculate the Spearman coefficients between each pair of uncertainty components to construct the initial dependency matrix;

[0045] The initial dependency matrix is ​​decomposed into eigenvalues ​​to obtain the minimum eigenvalue and condition number. The sum of the absolute value of the minimum negative eigenvalue and the stability constant is used as the positive definite guarantee term. The logarithm of the condition number is used as the ill-conditioned basis value. The weighting is increased exponentially by combining the energy proportion of the heavy-tailed uncertainty component in the total uncertainty to obtain the ridge parameter that increases with the increase of the energy proportion.

[0046] The diagonal elements of the initial dependency matrix are perturbed by adding ridge parameters, and the perturbed matrix is ​​then diagonally normalized to obtain a candidate positive definite correlation matrix. Cholesky decomposition is performed on the candidate positive definite correlation matrix to verify its positive definiteness. If the decomposition fails, the ridge parameters are increased and the diagonal perturbation and diagonal normalization are repeated until the positive definite condition is met. Then, the positive definite covariance matrix is ​​generated by left-multiplying and right-multiplying the standard deviation diagonal matrix of each uncertainty component.

[0047] The Spearman coefficients of the pairwise pairs of the historical time series of the four business components are calculated using the rank correlation algorithm, generating a 4×4 symmetric initial dependency matrix. Eigenvalue decomposition is performed on this initial matrix using the Jacobi iteration method. The condition number is obtained by the ratio of the absolute values ​​of the largest eigenvalue (e.g., 3.2) and the smallest eigenvalue (e.g., 0.01), resulting in a condition number of 320. The natural logarithm of this condition number, approximately 5.77, is taken as the base value. The proportion of the sum of the variances of the heavy-tailed components to the total variance of all components is extracted as the energy percentage, set to 0.3. Ridge parameters are calculated using an exponentially increasing weighting formula. The ridge parameters satisfy... Where λmin is the minimum eigenvalue of the initial dependency matrix, ε is the stability constant, κ is the condition number, κ0 is the upper limit of the condition number, Eh is the energy proportion of the heavy-tailed component, η is a scaling factor less than 1, and β is an adjustment factor. When the minimum eigenvalue is positive and the condition number does not exceed the upper limit of the condition number, the ridge parameter is either the stability constant or no additional perturbation is applied; when perturbation is required, the ridge parameter is preferably limited to between 0.001 and 0.1. After obtaining the ridge parameter, it is added to each element of the main diagonal of the initial dependency matrix for perturbation. After completing the diagonal element correction, the perturbed matrix is ​​first diagonally normalized to obtain the candidate positive definite dependency matrix; then, the Cholesky decomposition is called to verify the positive definiteness of the candidate positive definite dependency matrix. If the Cholesky decomposition is unsuccessful, the stability constant is increased or the diagonal perturbation and diagonal normalization are repeated until the candidate positive definite dependency matrix satisfies the positive definiteness condition. Extract the standard deviation of the uncertainty samples of each business component to construct a 4×4 diagonal matrix. Multiply the standard deviation diagonal matrix on the left by the correlation matrix, and then multiply it on the right by the standard deviation diagonal matrix to complete the dimensional recovery and numerical scaling, and output a non-singular and completely positive definite 4×4 covariance matrix.

[0048] In some embodiments, the calculation of the combined standard uncertainty by substituting the transfer formula, correcting the expansion factor using the Welch-Satterthwaite formula, and feeding back the ratio of the second-order cumulant curvature to the square of the corresponding standard uncertainty to the corrected component degrees of freedom, generating a verification threshold with an asymmetric confidence interval, includes:

[0049] The combined standard uncertainty is calculated based on the sensitivity coefficients and positive definite covariance matrices of each business component.

[0050] The ratio of the curvature of the second-order cumulant of each heavy-tailed uncertainty component to the square of the corresponding standard uncertainty is used as the degree of freedom reduction factor. The initial degree of freedom of the corresponding component is reduced and a minimum degree of freedom lower limit is set to obtain the corrected component degree of freedom.

[0051] Substitute the corrected component degrees of freedom and the combined standard uncertainty into the Welch-Satterthwaite formula to calculate the overall degrees of freedom, and obtain the expansion factor from the table based on the overall degrees of freedom;

[0052] The combined standard uncertainty is multiplied by the expansion factor, and an asymmetric confidence interval is constructed based on the skewness characteristics of the heavy-tailed uncertainty components. The boundary of the asymmetric confidence interval is used as the verification threshold.

[0053] The row vector composed of first-order sensitivity coefficients, the positive definite covariance matrix, and the column vector of sensitivity coefficients are multiplied sequentially, and the square root of the result is taken to obtain the combined standard uncertainty of the entire daily electricity bill result. For example, the combined standard uncertainty is 150.5 yuan. For each previously marked heavy-tailed uncertainty component, the curvature correction factor is calculated by dividing the pre-calculated second-order cumulant curvature (e.g., 14.62) by the square of the standard uncertainty corresponding to that component (e.g., 16.0), which is approximately 0.91. Assuming the initial degrees of freedom of this heavy-tailed component is 30 and the curvature adjustment factor α is 0.1, then according to... The calculated corrected degrees of freedom are approximately 30 / (1+0.1×0.91)=27.50. This operation reduces the effectiveness of statistical information due to tail-weighting, effectively decreasing the amount of information in the sample.

[0054] Substituting the corrected component degrees of freedom νi′ and the combined standard uncertainty into the Welch-Satterthwaite formula, the overall effective degrees of freedom of the entire system are calculated, which is 118 if aggregated by denominator weights. Based on this overall effective degrees of freedom and the preset 95% confidence level, the corresponding expansion factor is obtained from the t-distribution critical value table. Since the degrees of freedom of 118 are close to infinity, the expansion factor can be determined to be 1.98.

[0055] Multiplying the combined standard uncertainty of 150.5 by the expansion factor of 1.98 yields the basic confidence half-width of 297.99. Combining this with the skewness calculated from the heavy-tailed uncertainty sample, if the skewness is positive (e.g., 1.2) and the skewness adjustment coefficient ρ is 0.1, then the upper half-width is U+ = 1.12U, or 333.75, and the lower half-width is U- = U, or 297.99. An asymmetric confidence interval is constructed, such as the expected value of 1000 plus the upper tolerance threshold of 333.75, and the expected value of 1000 minus the lower tolerance threshold of 297.99, outputting [702.01, 1333.75] as the daily clearing verification threshold with an asymmetric confidence interval for the judgment. The real-time comparison of the daily electricity sales settlement value with the above asymmetric verification threshold and the triggering of out-of-bounds conditions are as follows: Figure 3 As shown.

[0056] S3, when the daily electricity sales result exceeds the verification threshold, a correction is triggered. The contribution magnitude vector is constructed by the square of the product of the sensitivity coefficient of each business component and the corresponding standard uncertainty. The correction direction is determined by combining the sign of the deviation of the daily electricity sales result from the verification threshold and the sign of the sensitivity coefficient of each business component. The deviation is then distributed in reverse to each business component. The business components are updated using a Kalman filter that includes residual autoregression. The process noise is scaled by the root mean square of the contribution vector, and the observation noise is adjusted by the out-of-bounds amplitude. The correction result is then output.

[0057] Compare the actual daily electricity sales results of the current verification period with the calculated upper and lower verification thresholds. If the actual result is greater than the upper verification threshold or less than the lower verification threshold, the correction process is activated, and the difference between the actual result and the verification threshold exceeding the threshold is calculated as the total deviation value.

[0058] Extract the sensitivity coefficient and corresponding standard uncertainty of each business component, multiply the two and square them to calculate the variance contribution value of each business component. Normalize all variance contribution values ​​by dividing each variance contribution value by the sum of all variance contribution values ​​to construct a contribution vector whose sum of elements is 1.

[0059] The calculated total deviation value is multiplied by the corresponding element value in the contribution vector to calculate the output-side deviation share that each service component should bear. Then, each output-side deviation share is divided by the sum of the absolute value of the corresponding sensitivity coefficient and the preset stability constant to convert it into a correction amount in the dimensions of that service component itself. Specifically, for the i-th service component, its correction amount is: Where Δy is the algebraic difference between the daily electricity sales result and the boundary of the side verification threshold, ci is the sensitivity coefficient of the i-th business component, wi is the corresponding contribution magnitude vector element, and ε is the preset stability constant; according to After obtaining the reverse-allocated business component value, the reverse correction of the original business component value is completed.

[0060] A joint Kalman filter model is constructed for all traffic components after reverse allocation. The state vector includes the correction values ​​for each traffic component and the residual autoregressive term, i.e., Xk=[x1,k, x2,k, ..., xn,k, ek] T The observations are either the current daily electricity sales result or the daily results recalculated from each business component after reverse allocation. The observation matrix consists of the first-order sensitivity coefficients and residual coefficients of each business component, i.e., Hk=[c1, c2, ..., cn, 1]. The residual autoregressive term is used to characterize the impact of unresolved deviations from the previous verification period on the current period. First-order autoregressive modeling is performed on the historical deviation residuals, and the autoregressive coefficients are extracted as parameters of the state transition matrix for Kalman filtering. The root mean square of the normalized contribution vector generated in the current verification period is calculated. Based on this root mean square, a process noise scaling factor of not less than 1 is constructed. For example, 1 plus the product of the root mean square of the contribution and the process noise adjustment coefficient is used as the scaling factor. This scaling factor is then multiplied by the initial process noise covariance matrix to complete the adjustment of the process noise matrix. The absolute distance by which the actual daily electricity sales result exceeds the verification threshold is calculated as the out-of-bounds amplitude. This out-of-bounds amplitude is used to scale the preset initial observation noise covariance matrix through an exponential mapping function to obtain the updated observation noise covariance matrix.

[0061] The time update step of the Kalman filter is performed to predict the current state based on the prior state estimate and the state transition matrix. The measurement update step is then performed, using the Kalman gain formula and combining the updated process noise covariance matrix and observation noise covariance matrix to correct the prior state estimate. The output is the filtered and smoothed value of each traffic component, serving as the correction result.

[0062] In some embodiments, when the daily electricity sales result exceeds the verification threshold, a correction is triggered by constructing a contribution magnitude vector using the square of the product of the sensitivity coefficient of each business component and the corresponding standard uncertainty. The correction direction is determined by combining the sign of the deviation of the daily electricity sales result from the exceeding verification threshold with the sign of the sensitivity coefficient of each business component, and the deviation is then distributed in reverse to each business component. This includes:

[0063] When the daily electricity sales result exceeds the verification threshold, the algebraic difference between the daily electricity sales result and the boundary of the verification threshold on the exceeding side is calculated as the total deviation.

[0064] Calculate the square of the product of the sensitivity coefficient and the corresponding standard uncertainty of each business component, and obtain the contribution magnitude vector after normalization.

[0065] Based on the sign of the total deviation and the sign of the sensitivity coefficient of each service component, the correction direction of each service component is determined.

[0066] The output-side deviation share borne by each service component is determined by the contribution magnitude vector, and the deviation is distributed in reverse to each corresponding service component by scaling according to the sum of the absolute value of the corresponding sensitivity coefficient and the preset stability constant.

[0067] The real-time monitoring system outputs daily electricity sales data. When the daily electricity sales result for a certain day, such as 1500, exceeds the preset upper limit of verification, such as 1333.7, or falls below the lower limit of verification, the correction procedure is activated, and the algebraic difference between the actual daily sales result and the threshold boundary that exceeded the limit, i.e., 166.3, is extracted as the total system-level deviation value that needs to be eliminated.

[0068] During the error allocation stage, the sensitivity coefficients and standard uncertainties of the four business components are extracted respectively. The sensitivity coefficient and standard uncertainty of the same component are multiplied, and the square of the product is used as the variance contribution value of that business component. Assume the calculated variance contribution values ​​of the four components are 50, 20, 10, and 120, respectively. After squaring, the variance contribution values ​​of each business component are added together to obtain a weighted sum, i.e., 200. The variance contribution values ​​of each component are then divided by this weighted sum for normalization, generating a contribution vector with a sum of 1, i.e., [0.25, 0.10, 0.05, 0.60]. This contribution vector will be mapped and used as the correction coefficient that each underlying business component must bear, reflecting that parameters with higher uncertainty and greater sensitivity to the overall output need to bear greater responsibility for deviations. Multiply the total deviation value of 166.3 by the respective proportions in the correction coefficient vector to calculate the output-side deviation share that the four business components of measurement, price, loss, and settlement rules need to bear, which are 41.575, 16.63, 8.315, and 99.78, respectively. Then divide each of these values ​​by the sum of the absolute value of the corresponding sensitivity coefficient and the preset stability constant to convert them into the compensation amount under the dimensions of each business component, and determine the direction of addition or subtraction according to the sign of the sensitivity.

[0069] In some embodiments, the Kalman filter with residual autoregression is used to update the traffic components, wherein process noise is scaled by the root mean square of the contribution vector, observation noise is adjusted by the out-of-bounds magnitude, and the output correction result includes:

[0070] Establish the Kalman filter state equation and observation equation that include historical residual autoregressive terms;

[0071] Calculate the root mean square of the contribution vector of the current verification period, construct a process noise scaling factor of not less than 1 based on the root mean square, adjust the initial process noise covariance using the process noise scaling factor, and obtain the updated process noise covariance.

[0072] Calculate the out-of-bounds range of the current verification threshold, and use the out-of-bounds range as an adjustment factor to update the observation noise covariance.

[0073] By using Kalman filtering time update and measurement update steps, recursive filtering is performed on each service component after inverse amortization to obtain the smoothed and denoised service component correction value, and the correction result is merged and output.

[0074] For all business components after reverse allocation of bias, a discrete-time joint Kalman filter model is constructed. Based on the state equation, a first-order autoregressive historical residual term is added, meaning the current state value depends not only on the state and control input of the previous time step but also on the residual value of the previous time step; the autoregressive coefficient is preferably set between 0.6 and 0.8, for example, 0.7. Simultaneously, a corresponding observation equation is constructed to map the internal state to the actual daily clearing results. The normalized contribution vector generated in the current verification period is extracted and its root mean square is calculated; for example, the square root of the squared mean of four elements [0.25, 0.10, 0.05, 0.60] is approximately 0.327. Construct a process noise scaling factor of not less than 1, where aq is the process noise adjustment coefficient and RMS(w) is the root mean square of the contribution vector. Multiply the process noise scaling factor with the initial process noise covariance matrix to complete the adjustment of the process noise matrix, which represents the decrease in confidence level caused by parameter anomalies when the model changes.

[0075] At the level of updating observation noise, the absolute value of the amplitude of the daily electricity sales settlement exceeding the verification threshold in the current verification period is calculated. The ratio of this amplitude of the exceedance to the original verification threshold span is used as an adjustment factor to amplify the initial observation noise covariance matrix in an exponential manner, generating an updated observation noise covariance matrix. This process indicates that when the exceedance result is judged as an abnormal observation or suspected data quality deviation, the more severe the exceedance, the greater the unreliability of the observation system is considered. When multiple consecutive verification periods show exceedances in the same direction, the process noise covariance is increased simultaneously to characterize the possibility of a real drift in the business status.

[0076] After the parameters are ready, the execution time prediction step derives the prior state estimate and prior error covariance, and uses the calculated Kalman gain for measurement correction. The correction amount previously back-allocated to each business component is incorporated as innovative information into the model recursion to suppress high-frequency random fluctuations and transient interference. The joint filtering outputs the smoothed correction values ​​of each business component, and the updated components are recombined and forward-mapped, merging to output the corrected electricity charges or settled electricity results of the daily electricity sales system.

[0077] The test sample library uses nearly six months of real daily electricity sales settlement data from a provincial power trading center, including 180 days of data on metering, pricing, loss, and settlement rules. The experiment compares a baseline scheme, ablation scheme one, and a complete scheme. The baseline scheme uses a traditional normal distribution with symmetric verification intervals and proportional error allocation; ablation scheme one retains the asymmetric confidence interval threshold generation module of this invention but discards the Kalman filter state update process; the complete scheme uses a complete mechanism including heavy-tail judgment, second-order cumulative penalty, and Kalman filter reverse allocation with residual autoregressive terms. Hardware conditions are uniformly deployed on an eight-core server with a 3.0 GHz clock speed, and the convergence error tolerance of the iterative algorithm is fixed at [value missing]. .

[0078] After 180 days of playback testing, the baseline scheme accumulated 56 invalid over-limit alarms, with a total average absolute deviation of 185.6 yuan in daily electricity sales accounting and an overall normalized root mean square error of 26.3. The invalid over-limit alarms in ablation scheme one decreased to 31, the total average deviation of daily electricity sales accounting decreased to 166.3 yuan, and the overall root mean square error of prediction decreased to 19.2. The complete scheme, after triggering verification within the asymmetric confidence interval, performed real-time updates, reducing invalid over-limit alarms to 25, decreasing the total average absolute deviation of daily electricity sales accounting to 41.5 yuan, and the maximum underlying business component compensation value smoothly converged to 99.78 yuan, with the overall root mean square error of prediction controlled at 6.8.

[0079] The ablation scheme 1 reduced the root mean square error by 26.9% compared to the baseline scheme, confirming that isolating heavy-tailed components based on excess kurtosis and implementing degree-of-freedom penalties in conjunction with second-order cumulant curvature can align with the asymmetric distribution characteristics of electricity billing and reduce boundary misjudgment rates. The complete scheme further reduced the root mean square error by 64.5% compared to ablation scheme 1, indicating that incorporating Kalman filtering with noisy covariance and residual autoregressive terms in the reverse allocation stage not only rationally allocated the correction weights of underlying sensitive variables but also filtered out transient interference from high-frequency observations, improving the smoothness and convergence speed of daily electricity sales error correction.

[0080] In the second embodiment, the present invention also proposes an intelligent verification and deviation correction system for daily electricity sales results, comprising the following modules:

[0081] The solution module is used to collect daily electricity sales data, decompose the daily electricity sales data into metering business components, price business components, loss business components and settlement rule business components, and evaluate the corresponding uncertainty components. According to the tail thickness index, the uncertainty components are divided into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components. The sensitivity coefficient of each business component is obtained by using a first-order Taylor expansion. For the heavy-tailed uncertainty components, a saddle point approximation second-order semi-analytical expansion is used to obtain the second-order cumulative curvature.

[0082] The generation module is used to construct a dependency matrix with Spearman coefficients as elements. After applying ridge parameter perturbation to the diagonal elements of the dependency matrix, it reconstructs a positive definite correlation matrix and scales it to a positive definite covariance matrix by combining the standard deviation of each uncertainty component. The combined standard uncertainty is calculated by substituting it into the transfer formula. The expansion factor is corrected using the Welch-Satterthwaite formula. The ratio of the curvature of the second-order cumulant to the square of the corresponding standard uncertainty is fed back to the corrected component degrees of freedom to generate a verification threshold with asymmetric confidence intervals.

[0083] The correction module is used to trigger correction when the daily electricity sales result exceeds the verification threshold. It constructs a contribution magnitude vector by multiplying the sensitivity coefficient of each business component by the square of the product of the corresponding standard uncertainty. It determines the correction direction by combining the sign of the deviation of the daily electricity sales result from the verification threshold and the sign of the sensitivity coefficient of each business component. The deviation is then distributed in reverse to each business component. The business components are updated using a Kalman filter that includes residual autoregression. The process noise is scaled by the root mean square of the contribution vector, and the observation noise is adjusted by the out-of-bounds magnitude. The correction result is then output.

[0084] In some embodiments, dividing the uncertainty components into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components based on the tail thickness index includes:

[0085] Calculate the kurtosis value of the historical error sample, historical residual sample or substitute disturbance sample corresponding to each business component, or use the theoretical kurtosis value of the preset probability distribution as the tail thickness index.

[0086] Set a standard kurtosis threshold and determine whether the kurtosis value or theoretical kurtosis value is greater than the standard kurtosis threshold;

[0087] If it is not greater than the standard kurtosis threshold, the corresponding uncertainty component is determined to be a light-tailed or near-Gaussian uncertainty component.

[0088] If the uncertainty exceeds the standard kurtosis threshold, the corresponding uncertainty component is classified as a heavy-tailed uncertainty component.

[0089] In some embodiments, the step of using a first-order Taylor expansion to obtain the sensitivity coefficients of each service component, and using a saddle-point approximation second-order semi-analytical expansion to solve for the heavy-tailed uncertainty component to obtain the second-order cumulant curvature includes:

[0090] For each business component, establish the first-order partial derivative function of the electricity sales daily clearing model, and substitute the expected value of each business component into the partial derivative function to obtain the corresponding first-order sensitivity coefficient.

[0091] For the heavy-tailed uncertainty components, construct the cumulative generating function of the heavy-tailed uncertainty components, solve the saddle point equation to obtain the saddle point, and take the second derivative of the cumulative generating function at the saddle point to obtain the corresponding second-order cumulant curvature.

[0092] In some embodiments, the construction of a dependency matrix with Spearman coefficients as elements, the application of ridge parameter perturbations to the diagonal elements of the dependency matrix to reconstruct a positive definite correlation matrix, and the scaling to a positive definite covariance matrix by combining the standard deviations of each uncertainty component, includes:

[0093] Calculate the Spearman coefficients between each pair of uncertainty components to construct the initial dependency matrix;

[0094] The initial dependency matrix is ​​decomposed into eigenvalues ​​to obtain the minimum eigenvalue and condition number. The sum of the absolute value of the minimum negative eigenvalue and the stability constant is used as the positive definite guarantee term. The logarithm of the condition number is used as the ill-conditioned basis value. The weighting is increased exponentially by combining the energy proportion of the heavy-tailed uncertainty component in the total uncertainty to obtain the ridge parameter that increases with the increase of the energy proportion.

[0095] The diagonal elements of the initial dependency matrix are perturbed by adding ridge parameters, and the perturbed matrix is ​​then diagonally normalized to obtain a candidate positive definite correlation matrix. Cholesky decomposition is performed on the candidate positive definite correlation matrix to verify its positive definiteness. If the decomposition fails, the ridge parameters are increased and the diagonal perturbation and diagonal normalization are repeated until the positive definite condition is met. Then, the positive definite covariance matrix is ​​generated by left-multiplying and right-multiplying the standard deviation diagonal matrix of each uncertainty component.

[0096] In some embodiments, the calculation of the combined standard uncertainty by substituting the transfer formula, correcting the expansion factor using the Welch-Satterthwaite formula, and feeding back the ratio of the second-order cumulant curvature to the square of the corresponding standard uncertainty to the corrected component degrees of freedom, generating a verification threshold with an asymmetric confidence interval, includes:

[0097] The combined standard uncertainty is calculated based on the sensitivity coefficients and positive definite covariance matrices of each business component.

[0098] The ratio of the curvature of the second-order cumulant of each heavy-tailed uncertainty component to the square of the corresponding standard uncertainty is used as the degree of freedom reduction factor. The initial degree of freedom of the corresponding component is reduced and a minimum degree of freedom lower limit is set to obtain the corrected component degree of freedom.

[0099] Substitute the corrected component degrees of freedom and the combined standard uncertainty into the Welch-Satterthwaite formula to calculate the overall degrees of freedom, and obtain the expansion factor from the table based on the overall degrees of freedom;

[0100] The combined standard uncertainty is multiplied by the expansion factor, and an asymmetric confidence interval is constructed based on the skewness characteristics of the heavy-tailed uncertainty components. The boundary of the asymmetric confidence interval is used as the verification threshold.

[0101] In some embodiments, when the daily electricity sales result exceeds the verification threshold, a correction is triggered by constructing a contribution magnitude vector using the square of the product of the sensitivity coefficient of each business component and the corresponding standard uncertainty. The correction direction is determined by combining the sign of the deviation of the daily electricity sales result from the exceeding verification threshold with the sign of the sensitivity coefficient of each business component, and the deviation is then distributed in reverse to each business component. This includes:

[0102] When the daily electricity sales result exceeds the verification threshold, the algebraic difference between the daily electricity sales result and the boundary of the verification threshold on the exceeding side is calculated as the total deviation.

[0103] Calculate the square of the product of the sensitivity coefficient and the corresponding standard uncertainty of each business component, and obtain the contribution magnitude vector after normalization.

[0104] Based on the sign of the total deviation and the sign of the sensitivity coefficient of each service component, the correction direction of each service component is determined.

[0105] The output-side deviation share borne by each service component is determined by the contribution magnitude vector, and the deviation is distributed in reverse to each corresponding service component by scaling according to the sum of the absolute value of the corresponding sensitivity coefficient and the preset stability constant.

[0106] In some embodiments, the Kalman filter with residual autoregression is used to update the traffic components, wherein process noise is scaled by the root mean square of the contribution vector, observation noise is adjusted by the out-of-bounds magnitude, and the output correction result includes:

[0107] Establish the Kalman filter state equation and observation equation that include historical residual autoregressive terms;

[0108] Calculate the root mean square of the contribution vector of the current verification period, construct a process noise scaling factor of not less than 1 based on the root mean square, adjust the initial process noise covariance using the process noise scaling factor, and obtain the updated process noise covariance.

[0109] Calculate the out-of-bounds range of the current verification threshold, and use the out-of-bounds range as an adjustment factor to update the observation noise covariance.

[0110] By using Kalman filtering time update and measurement update steps, recursive filtering is performed on each service component after inverse amortization to obtain the smoothed and denoised service component correction value, and the correction result is merged and output.

[0111] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0112] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for intelligent verification and deviation correction of daily electricity sales results, characterized in that, include: Collect daily electricity sales data, decompose the daily electricity sales data into metering business components, price business components, loss business components and settlement rule business components, and evaluate the corresponding uncertainty components. According to the tail thickness index, the uncertainty components are divided into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components. The sensitivity coefficient of each business component is obtained by using a first-order Taylor expansion. The second-order cumulant curvature is obtained by using a saddle point approximation second-order semi-analytical expansion for heavy-tailed uncertainty components. Construct a dependency matrix with Spearman coefficients as elements. Apply ridge parameter perturbation to the diagonal elements of the dependency matrix and reconstruct a positive definite correlation matrix. Scale the matrix with the standard deviation of each uncertainty component to a positive definite covariance matrix. Substitute the covariance matrix into the transfer formula to calculate the combined standard uncertainty. Use the Welch-Satterthwaite formula to correct the expansion factor. Feed back the ratio of the second-order cumulant curvature to the square of the corresponding standard uncertainty to the corrected component degrees of freedom to generate a verification threshold with asymmetric confidence intervals. When the daily electricity sales result exceeds the verification threshold, a correction is triggered. The contribution magnitude vector is constructed by the square of the product of the sensitivity coefficient of each business component and the corresponding standard uncertainty. The correction direction is determined by combining the sign of the deviation of the daily electricity sales result from the verification threshold and the sign of the sensitivity coefficient of each business component. The deviation is then distributed in reverse to each business component. A Kalman filter with residual autoregression is used to update the business components. The process noise is scaled by the root mean square of the contribution vector, and the observation noise is adjusted by the out-of-bounds amplitude. The correction result is then output.

2. The method according to claim 1, characterized in that, The method of classifying uncertainty components into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components based on the tail thickness index includes: Calculate the kurtosis value of the historical error sample, historical residual sample or substitute disturbance sample corresponding to each business component, or use the theoretical kurtosis value of the preset probability distribution as the tail thickness index. Set a standard kurtosis threshold and determine whether the kurtosis value or the theoretical kurtosis value is greater than the standard kurtosis threshold; If the uncertainty is not greater than the standard kurtosis threshold, the corresponding uncertainty component is determined to be a light-tailed or near-Gaussian uncertainty component. If the uncertainty exceeds the standard kurtosis threshold, the corresponding uncertainty component is classified as a heavy-tailed uncertainty component.

3. The method according to claim 1, characterized in that, The process of using a first-order Taylor expansion to obtain the sensitivity coefficients of each business component, and using a saddle-point approximation second-order semi-analytical expansion to solve for the heavy-tailed uncertainty components to obtain the second-order cumulant curvature, includes: For each business component, establish the first-order partial derivative function of the electricity sales daily clearing model, and substitute the expected value of each business component into the partial derivative function to obtain the corresponding first-order sensitivity coefficient. For the heavy-tailed uncertainty components, construct the cumulative generating function of the heavy-tailed uncertainty components, solve the saddle point equation to obtain the saddle point, and take the second derivative of the cumulative generating function at the saddle point to obtain the corresponding second-order cumulant curvature.

4. The method according to claim 2, characterized in that, The construction of the dependency matrix with Spearman coefficients as elements, the application of ridge parameter perturbations to the diagonal elements of the dependency matrix, and the reconstruction to generate a positive definite correlation matrix, combined with the standard deviations of each uncertainty component, to form a positive definite covariance matrix, includes: Calculate the Spearman coefficients between each pair of uncertainty components to construct the initial dependency matrix; The initial dependency matrix is ​​decomposed into eigenvalues ​​to obtain the minimum eigenvalue and condition number. The sum of the absolute value of the minimum negative eigenvalue and the stability constant is used as the positive definite guarantee term. The logarithm of the condition number is used as the ill-conditioned basis value. The weighting is increased exponentially by combining the energy proportion of the heavy-tailed uncertainty component in the total uncertainty to obtain the ridge parameter that increases with the increase of the energy proportion. The diagonal elements of the initial dependency matrix are perturbed by adding ridge parameters, and the perturbed matrix is ​​then diagonally normalized to obtain a candidate positive definite correlation matrix. Cholesky decomposition is performed on the candidate positive definite correlation matrix to verify its positive definiteness. If the decomposition fails, the ridge parameters are increased and the diagonal perturbation and diagonal normalization are repeated until the positive definite condition is met. Then, the positive definite covariance matrix is ​​generated by left-multiplying and right-multiplying the standard deviation diagonal matrix of each uncertainty component.

5. The method according to claim 1, characterized in that, The combined standard uncertainty is calculated by substituting the transfer formula, and the expansion factor is corrected using the Welch-Satterthwaite formula. The ratio of the second-order cumulant curvature to the square of the corresponding standard uncertainty is fed back to the corrected component degrees of freedom, generating a verification threshold with asymmetric confidence intervals, including: The combined standard uncertainty is calculated based on the sensitivity coefficients and positive definite covariance matrices of each business component. The ratio of the curvature of the second-order cumulant of each heavy-tailed uncertainty component to the square of the corresponding standard uncertainty is used as the degree of freedom reduction factor. The initial degree of freedom of the corresponding component is reduced and a minimum degree of freedom lower limit is set to obtain the corrected component degree of freedom. Substitute the corrected component degrees of freedom and the combined standard uncertainty into the Welch-Satterthwaite formula to calculate the overall degrees of freedom, and obtain the expansion factor from the table based on the overall degrees of freedom; The combined standard uncertainty is multiplied by the expansion factor, and an asymmetric confidence interval is constructed based on the skewness characteristics of the heavy-tailed uncertainty components. The boundary of the asymmetric confidence interval is used as the verification threshold.

6. The method according to claim 1 or 2, characterized in that, When the daily electricity sales result exceeds the verification threshold, a correction is triggered. A contribution magnitude vector is constructed using the square of the product of the sensitivity coefficient and the corresponding standard uncertainty of each business component. The correction direction is determined by combining the sign of the deviation of the daily electricity sales result from the exceeding verification threshold with the sign of the sensitivity coefficient of each business component. The deviation is then distributed in reverse to each business component, including: When the daily electricity sales result exceeds the verification threshold, the algebraic difference between the daily electricity sales result and the boundary of the verification threshold on the exceeding side is calculated as the total deviation. Calculate the square of the product of the sensitivity coefficient and the corresponding standard uncertainty of each business component, and obtain the contribution magnitude vector after normalization. Based on the sign of the total deviation and the sign of the sensitivity coefficient of each service component, the correction direction of each service component is determined. The output-side deviation share borne by each service component is determined by the contribution magnitude vector, and the deviation is distributed in reverse to each corresponding service component by scaling according to the sum of the absolute value of the corresponding sensitivity coefficient and the preset stability constant.

7. The method according to claim 1, characterized in that, The process component is updated using a Kalman filter incorporating residual autoregression, where process noise is scaled by the root mean square of the contribution vector and observation noise is adjusted by the out-of-bounds magnitude. The output correction result includes: Establish the Kalman filter state equation and observation equation that include historical residual autoregressive terms; Calculate the root mean square of the contribution vector of the current verification period, construct a process noise scaling factor of not less than 1 based on the root mean square, adjust the initial process noise covariance using the process noise scaling factor, and obtain the updated process noise covariance. Calculate the out-of-bounds range of the current verification threshold, and use the out-of-bounds range as an adjustment factor to update the observation noise covariance. By using Kalman filtering time update and measurement update steps, recursive filtering is performed on each service component after inverse amortization to obtain the smoothed and denoised service component correction value, and the correction result is merged and output.

8. A smart verification and deviation correction system for daily electricity sales results, characterized in that, Includes the following modules: The solution module is used to collect daily electricity sales data, decompose the daily electricity sales data into metering business components, price business components, loss business components and settlement rule business components, and evaluate the corresponding uncertainty components. According to the tail thickness index, the uncertainty components are divided into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components. The sensitivity coefficient of each business component is obtained by using a first-order Taylor expansion. For the heavy-tailed uncertainty components, a saddle point approximation second-order semi-analytical expansion is used to obtain the second-order cumulative curvature. The generation module is used to construct a dependency matrix with Spearman coefficients as elements. After applying ridge parameter perturbation to the diagonal elements of the dependency matrix, it reconstructs a positive definite correlation matrix and scales it to a positive definite covariance matrix by combining the standard deviation of each uncertainty component. The combined standard uncertainty is calculated by substituting it into the transfer formula. The expansion factor is corrected using the Welch-Satterthwaite formula. The ratio of the curvature of the second-order cumulant to the square of the corresponding standard uncertainty is fed back to the corrected component degrees of freedom to generate a verification threshold with asymmetric confidence intervals. The correction module is used to trigger correction when the daily electricity sales result exceeds the verification threshold. It constructs a contribution magnitude vector by multiplying the sensitivity coefficient of each business component by the square of the product of the corresponding standard uncertainty. It determines the correction direction by combining the sign of the deviation of the daily electricity sales result from the verification threshold and the sign of the sensitivity coefficient of each business component. The deviation is then distributed in reverse to each business component. The business components are updated using a Kalman filter that includes residual autoregression. The process noise is scaled by the root mean square of the contribution vector, and the observation noise is adjusted by the out-of-bounds magnitude. The correction result is then output.

9. The system according to claim 8, characterized in that, The method of classifying uncertainty components into light-tailed or near-Gaussian uncertainty components and heavy-tailed uncertainty components based on the tail thickness index includes: Calculate the kurtosis value of the historical error sample, historical residual sample or substitute disturbance sample corresponding to each business component, or use the theoretical kurtosis value of the preset probability distribution as the tail thickness index. Set a standard kurtosis threshold and determine whether the kurtosis value or the theoretical kurtosis value is greater than the standard kurtosis threshold; If the uncertainty is not greater than the standard kurtosis threshold, the corresponding uncertainty component is determined to be a light-tailed or near-Gaussian uncertainty component. If the uncertainty exceeds the standard kurtosis threshold, the corresponding uncertainty component is classified as a heavy-tailed uncertainty component.

10. The system according to claim 8, characterized in that, The process of using a first-order Taylor expansion to obtain the sensitivity coefficients of each business component, and using a saddle-point approximation second-order semi-analytical expansion to solve for the heavy-tailed uncertainty components to obtain the second-order cumulant curvature, includes: For each business component, establish the first-order partial derivative function of the electricity sales daily clearing model, and substitute the expected value of each business component into the partial derivative function to obtain the corresponding first-order sensitivity coefficient. For the heavy-tailed uncertainty components, construct the cumulative generating function of the heavy-tailed uncertainty components, solve the saddle point equation to obtain the saddle point, and take the second derivative of the cumulative generating function at the saddle point to obtain the corresponding second-order cumulant curvature.

Images