A multi-modal fusion and dynamic game driven power grid special cost intelligent review method

Abstract

The present application relates to the technical field of power system, specifically relates to a kind of multi-modal fusion and dynamic game driven power grid special cost intelligent review method, comprising the following steps: S1, the multi-source data of power grid special cost project is uniformly collected and processed, including field structured, check, deduplication, fill and exception identification, generate standardized review dataset;S2, based on attention PCA and density peak clustering Weighted feature space is constructed, and same kind cost category is divided;S3, establish log-normal mixed effect model and fuse multi-level factors, estimate reasonable unit price interval in combination with experience bayes method;S4, construct Markov decision process, optimize review strategy and output conclusion and adjustment suggestion.The present application, by constructing multi-source data processing, attention clustering modeling, mixed effect statistical analysis and dynamic game decision fusion Intelligent review system, realizes the high precision, automation and strategy explainability of power grid special cost review.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method for intelligent review of power grid special costs driven by multimodal fusion and dynamic game theory. Background Technology

[0002] In the process of power grid construction and operation and maintenance, special cost review is an important link in controlling investment risks and ensuring the rationality of projects. Its accuracy is directly related to the economic benefits and resource allocation efficiency of power grid enterprises. In order to improve review efficiency, in recent years, various units have gradually established information-based review platforms. Through cost field extraction and structured analysis, combined with historical data reference, the rationality of the declared unit price can be assessed. Especially in the context of multiple regions and multiple projects being promoted in parallel, the formation of a data-driven intelligent review system has become an industry development trend.

[0003] However, existing methods still have several technical shortcomings. First, the data sources are highly heterogeneous, with inconsistent field definitions and insufficient handling of missing and outlier values, making it difficult to guarantee the quality of the basic data for review. Second, the classification of cost categories relies heavily on manual rules, which cannot fully reflect the differences in characteristics such as region and project type, affecting the representativeness of clustering results. Third, most current unit price range estimation methods use static statistics or fixed models, failing to effectively integrate the project hierarchy and uncertainty, which can easily lead to misjudgments. Fourth, in the review decision-making process, there is a lack of systematic modeling of review actions and strategies based on dynamic optimization mechanisms such as Markov processes, making it difficult to support the generation of automated and personalized review suggestions. Summary of the Invention

[0004] This invention provides a multimodal fusion and dynamic game-driven intelligent review method for power grid special costs. Through multi-source data standardization processing, feature weighted clustering, multi-level Log-normal hybrid modeling, and Markov strategy optimization, it constructs an integrated review system covering the entire process of data preprocessing, category division, unit price evaluation, and intelligent decision-making, thereby achieving a synergistic improvement in the rationality, accuracy, and efficiency of the review and enhancing the level of intelligent cost control.

[0005] A multimodal fusion and dynamic game-driven intelligent review method for power grid special costs includes the following steps:

[0006] S1. Collect and preprocess multi-source data of power grid special cost projects in a unified manner. The multi-source data includes project information and cost details. Complete the structuring of cost fields, arithmetic consistency verification, merging of duplicate items, filling of missing values ​​and identification of outliers to form a standardized review dataset.

[0007] S2, based on attention-weighted PCA and density peak clustering, constructs a weighted feature space by combining project type, regional characteristics and cost details, and divides similar cost categories by comprehensive distance and local density;

[0008] S3, based on similar cost categories, constructs a multi-level log-normal mixed effect model, integrates the hierarchical factors of region, department, and project, and combines empirical Bayesian methods to output a reasonable unit price range and confidence level;

[0009] S4 constructs a Markov decision process with declared unit price, reasonable range, abnormal marker and confidence level as states, defines a variety of review actions and payoff functions, optimizes the strategy and outputs the final review conclusion and adjustment suggestions.

[0010] Optionally, S1 includes:

[0011] S11, uniformly collect project information and expense details of power grid special cost projects. The project information includes project number, project name, year, unit, region and project type. The expense details include expense line number, expense name or expense item, specifications, unit of measurement, quantity, declared unit price and total amount. The field format and expression of multi-source data are standardized.

[0012] S12 performs an arithmetic consistency check on the total amount, quantity, and declared unit price for each expense detail, calculates the error value and compares it with the set allowable error threshold. If the error value is within the error threshold range, it is marked as passed; otherwise, it is marked as failed. It also identifies duplicate records in the same project that are highly similar in expense name, specifications, and unit of measurement, performs quantity merging and total price merging operations, and backtracks the declared unit price after merging.

[0013] S13, For missing quantities or declared unit prices in the cost details, a median imputation strategy based on project type, region and year is used for statistical imputation, and the imputation confidence weight is calculated based on the imputation sample size.

[0014] S14. In the same group of expense details, identify abnormal declared unit prices based on robust Z-scores and box plot rules.

[0015] Optionally, S2 includes:

[0016] S21. The project type, region, and cost details are encoded into feature vectors. A comprehensive feature vector is constructed, and regional and time weights are introduced. Combined with prior weights, a weighted feature matrix is ​​generated. PCA is then performed on the weighted feature matrix to reduce its dimensionality and extract the principal components that explain the highest proportion of variance, thus forming the dimensionality-reduced feature representation.

[0017] S22. In the principal component space, define the comprehensive distance that combines region, time and cost, and calculate the local density of each cost detail and the minimum distance to the high-density point. Select cluster centers based on local density and minimum distance, and dynamically estimate the number of clusters based on project type, region and cost fluctuation. All cost details are assigned to the corresponding cluster centers according to the density increase path to complete the clustering of the same cost category.

[0018] Optionally, S21 includes:

[0019] S211 maps project type, region, and cost details into numerical feature vectors, and constructs a comprehensive feature vector. ;

[0020] S212 determines the regional weight of each expense detail in clustering dimensionality reduction by constructing a regional feature vector that includes basic attributes, load density, and terrain type, and calculating its similarity to the average regional feature vector. ;

[0021] S213, set the time decay coefficient according to the project cycle type of the cost details, and calculate the time weight in combination with the recorded timestamp. ;

[0022] S214 introduces prior weights for the feature dimensions, combining them with the time and region weights of the cost details to construct a weighted feature matrix that integrates multiple weight information. ;

[0023] S215, for the weighted characteristic matrix Principal component analysis is performed to compress high-dimensional features into low-dimensional irrelevant principal components.

[0024] Optionally, S22 includes:

[0025] S221, combining region, time, and cost details, calculates the three-dimensional normalized distance and introduces a scale parameter to construct a comprehensive distance. ;

[0026] S222, calculates the local density of each cost detail based on the comprehensive distance. ;

[0027] S223, calculate the minimum distance from each cost detail to cost details with a higher density than itself. ;

[0028] S224, constructing a comprehensive index based on local density and minimum distance. In conjunction with the number of project types, regions, responsible departments, and cost fluctuation levels, the number of cluster categories is adaptively estimated. ;

[0029] S225. Select cluster centers based on the comprehensive index ranking, and allocate the non-center cost details to the clusters of the nearest high-density samples according to the density-guided principle, thus completing the classification of cost categories of the same type.

[0030] Optionally, S3 includes:

[0031] S31, perform logarithmic transformation on the declared unit price and construct a fixed-effects feature vector. This includes project type, responsible department, province, city / county, geographical features, specifications, and time offset;

[0032] S32, construct a multi-level log-normal linear model with both fixed and random effects to explain the offset of the unit price of the cost details under different geographical, project, department and category dimensions;

[0033] S33, the restricted maximum likelihood (REML) method is used to estimate the variance components at each level. The variance components include the variances of the province, city, county, project type, responsible department, category, and error. The posterior mean of each random effect is calculated using the empirical Bayes method to obtain the bias term at each level. Finally, the conditional mean of each cost detail is obtained by superimposing the fixed effect and the random bias term at all levels.

[0034] S34. Using the conditional mean and predicted variance output from the multilevel log-normal linear model, a log-normal distribution is constructed, and a central estimate is generated. confidence interval ;

[0035] S35, Construct the historical average unit price of similar expense details Compared with the overall average unit price And based on the historical average unit price Compared with the overall average unit price The mean deviation is adjusted by the empirical Bayesian weighting factor to narrow the upper edge of the unit price range;

[0036] S36, Introduction of Dimensional Factor And combined with empirical Bayesian factors Construct a comprehensive confidence level .

[0037] Optionally, S4 includes:

[0038] S41, construct a state vector based on the current cost details, define operation actions and set relevant parameters, and construct a state transition function based on changes in declared unit price and anomaly marker update rules;

[0039] S42, for cost details with a comprehensive confidence level lower than the lower limit of comprehensive confidence level, a conservative mechanism is triggered to adjust the upper limit of the reasonable unit price range to the weighted lower limit of the current upper limit, the upper limit threshold and the correction target;

[0040] S43, construct an immediate benefit function that includes rewards and penalties as well as adjustment costs, and use maximizing the cumulative expected benefit throughout the review process as the optimization objective to guide the dynamic selection of strategies;

[0041] S44, set the minimum compression ratio Check the feasibility of the pressure reduction; if it is met, generate the compression ratio for each sub-item according to the full budget proportion. To achieve a reasonable allocation of resources for reduction and rectification;

[0042] S45. Based on Markov processes, construct Q and V functions, and use Bellman equations and policy iteration methods to solve for the optimal execution strategy of cost compression actions.

[0043] S46 terminates the strategy when the unit price falls back to a reasonable range, the confidence level meets the standard, or the number of rounds of review exceeds the limit, and outputs the final reduction suggestion, corrected unit price, and range adjustment results.

[0044] Optionally, S41 includes:

[0045] S411, the state vector is composed of the declared unit price of the current cost details, the upper and lower limits of the reasonable unit price range, the price deviation, the anomaly marker and the overall confidence level;

[0046] S412 sets the execution action for the current status of the expense details, including veto, uniform weight reduction and overall item compression, in order to control the declared unit price;

[0047] S413, recalculate the state vector for the next moment based on the declared unit price and normalized deviation after the action is executed, and construct the state transition relationship by combining the anomaly label and confidence adjustment rules.

[0048] Optionally, S43 includes:

[0049] S431, the immediate benefit function takes the comprehensive score brought by the action in the current state as its core. It quantifies the immediate benefit of the current review adjustment by rewarding price adjustments within a reasonable price range, penalizing exceeding the upper limit and insufficient confidence, and taking into account the action cost caused by weight reduction or compression. ;

[0050] S432, multi-step expected return targets use discount factors to weigh current and future returns, throughout the entire review process. arrive Calculate the cumulative expected return And by maximizing This guides the strategy to select the optimal sequence of review actions.

[0051] The beneficial effects of this invention are:

[0052] This invention introduces structured cleaning, arithmetic consistency testing, duplicate item merging, intelligent missing value filling, and anomaly identification mechanisms into the multi-source data processing of power grid special cost projects. This enables the efficient conversion of cost detail data from its original heterogeneous representation to a unified and standardized structure, significantly improving the quality and reliability of pre-review data.

[0053] This invention constructs a weighted feature matrix that integrates project type, regional characteristics, and cost attributes, and combines attention PCA and density peak clustering methods to achieve fine-grained classification of similar costs based on local density and comprehensive distance. This effectively solves the problems of large heterogeneity in cost structure and uncontrollable intra-class differences in traditional classification methods.

[0054] This invention constructs a reasonable unit price range based on a multi-level log-normal mixed-effects model, and combines empirical Bayesian estimation and confidence level construction to build a dynamic range based on capturing deviations in geographical, project, and management dimensions. Then, it integrates Markov game strategy to optimize the payoff of review actions, and finally realizes a closed-loop decision-making process that automatically generates suggestions for reduction, rejection, and rectification, which significantly improves the intelligence, refinement, and interpretability of the review process. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a schematic diagram of the review method flow according to an embodiment of the present invention. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0058] like Figure 1 As shown, a multimodal fusion and dynamic game-driven intelligent review method for power grid special costs includes the following steps:

[0059] S1, Conduct data collection and organization for the data cost project. Collect project feasibility study reports, cost lists, contract documents, and relevant policy standards, standardize data formats and expressions, and check and process arithmetic errors, duplicates, missing values, and outliers in the data.

[0060] (1) Data collection and field specifications:

[0061] First, all fields related to cost items are standardized and organized to form a minimal structured record. The project information section mainly includes fields such as project number, project name, year, unit, region, and project type, used to locate the project subject and its scope. The expense details section mainly includes fields such as expense line number, expense name or expense account, specifications, unit of measurement, quantity, declared unit price, and total amount, used to reflect the pricing basis and amount of a single expense.

[0062] Based on available relevant materials, including project proposals, feasibility studies, cost lists, draft contracts, meeting minutes, records of similar projects from previous years, and policy reference databases, data collection will be conducted. The main content collected includes the following:

[0063]

[0064] (2) Basic consistency and arithmetic verification:

[0065] After completing the field specifications, a basic arithmetic check is performed on each expense list to verify the consistency between the total amount and the quantity and unit price. The system calculates the theoretical total amount based on the declared quantity and the declared value, and compares it with the original declared total amount, setting a tolerance threshold that combines percentage and absolute value. When the difference falls within the tolerance range, the expense line is marked as arithmetic checked and passed; when the difference exceeds the threshold, it is only marked as arithmetically inconsistent at the data level, and the original value is retained. No pass or fail judgment is made directly in this step, but it becomes a key focus for subsequent model calculations and manual review.

[0066] 1) Arithmetic consistency error:

[0067] To quantify the difference between the total cost of the i-th row and the quantity × unit price, the following definition is used:

[0068] ;

[0069] in, For the first Line price For the first Number of rows For the first The unit price is declared in the line. This represents the absolute error of arithmetic consistency.

[0070] 2) Arithmetic pass / fail judgment:

[0071] To mark at the data level whether the arithmetic relationship of each row of costs conforms to the allowable error range, arithmetic consistency is defined by a marker variable:

[0072] ;

[0073] in, To indicate whether arithmetic consistency is passed, 1 indicates pass; 0 indicates fail. This is the allowable error threshold for this row.

[0074] 3) Error tolerance:

[0075] To balance the reasonable tolerance for both large and small fee rows in arithmetic verification, the allowable error threshold for the i-th line of fees is determined using a combination of proportional and absolute value methods:

[0076] ;

[0077] in, A proportional coefficient of 0.5% of the amount. The lower limit is fixed (in yuan).

[0078] 4) Merging duplicate rows:

[0079] To identify expense rows in the same project that have identical expense names, specifications, and units of measurement, potentially indicating duplicate entries, calculate the similarity score based on the key-value combination of the two expense records. The similarity criterion is key-value pairing. :

[0080] The quantities are combined as follows: ;

[0081] The total price is combined as follows: ;

[0082] The unit price is calculated as follows: ;

[0083] in, This represents the total number after the merger. The number of each duplicate candidate row. This is the total price after the merger; The total price for each candidate row, This refers to the combined unit price.

[0084] (3) Identification and handling of missing values:

[0085] 1) Numerical missing imputation:

[0086] In order to objectively and reproducibly impute missing values ​​in key numerical fields such as quantity and declared unit price when there is no reliable documentary evidence, the grouped median method is used to calculate the imputed value for the i-th row.

[0087] ;

[0088] in, For the first Fill in missing values ​​for rows. For variables to be filled, the three conditions within the criteria are of the same type, region, and adjacent years. The backtracking rule for grouping is as follows: if the sample size of the group is... If insufficient, it degenerates to "same type × same region"; if still insufficient, it degenerates to the global median of "same type". All imputation flags ImputeFlag=1.

[0089] 2) Filling in credibility:

[0090] Define a confidence weight for each record to be filled, and use it in subsequent models to distinguish the reliability of the filling results under different sample sizes.

[0091] ;

[0092] in, For the first Line filling credibility, The grouped sample size used to calculate the median. This is the smoothing constant.

[0093] If text is missing, the description of the same ItemID will be searched first on the same page / adjacent pages; if it cannot be found, a suggestion to fill in the missing text will be generated instead of guessing.

[0094] (4) Outlier identification and handling:

[0095] Within each group, a robust Z-score based on the median and absolute median difference is constructed to identify outliers in the reported unit price that deviate significantly from the typical level of the same group.

[0096] 1) Robust Z-score:

[0097] ;

[0098] in, For the first Stable Z-score The median unit price within the group. The absolute median difference in unit price within the group. For the first The unit price is declared in the line;

[0099] The judgment is as follows: ;

[0100] in, For statistical anomaly marking, For robust Z-threshold, a value of 3.5 is recommended.

[0101] 2) Box line rules:

[0102] The box plot rules are then used to identify outliers in the unit price distribution within the group, under the following conditions:

[0103] ;

[0104] in, IQR anomaly marker, Lower / upper quartiles, It is the interquartile range.

[0105] S2, based on spatiotemporal fusion clustering, constructs similar cost categories. Its function is to organize the originally scattered individual cost lists into a set of similar cost clusters that conform to the power grid business logic and are convenient for subsequent modeling, providing a sample space for subsequent cost calculation and game decision-making. In terms of specific methods, firstly, attention-weighted PCA is used to reduce the dimensionality of multi-dimensional indicators such as project type, regional characteristics, and cost level, introducing time weights and regional weights. Industry characteristics such as differences in project cycle, regional load density, and terrain type in power grid special costs are explicitly integrated into the feature space, highlighting the influence of representative samples and key features. Secondly, density peak clustering is used in the dimensionality-reduced feature space. Based on the comprehensive distance of time-space-cost and the local density-distance joint index, the cluster centers are automatically identified, and the number of project types, regions, and responsible departments are incorporated into the cluster number calculation to form a set of cost categories.

[0106] (1) Feature construction and attention-weighted PCA dimensionality reduction:

[0107] 1) Construction of comprehensive feature vectors:

[0108] Information such as project type, regional characteristics, and cost characteristics are uniformly mapped into numerical feature vectors for subsequent weighted dimensionality reduction and clustering.

[0109] For the The comprehensive feature vector constructed from the records is as follows: ;

[0110] in, For the first Comprehensive feature vector of each expense record For the encoding or embedding vector of the item type, one-hot encoding or low-dimensional embedding can be used. The region's encoding or embedding vector is composed of basic attributes such as the city and province to which the region belongs. This is a cost-based feature vector, which can include cost-related indicators such as normalized unit price, logarithmic unit price, quantity, and percentage of total price.

[0111] All The feature matrix is ​​formed as follows:

[0112] ;

[0113] in, The sample feature matrix, For the sample size, For feature dimensions.

[0114] 2) Regional characteristics and regional weights:

[0115] The indicators of "load density" and terrain type are explicitly introduced into the regional characteristics to construct regional attention weights, reflecting the differences in cost structure among different regions.

[0116] First, construct the region feature vector for each region as follows: ;

[0117] in, In order to be with the first The region feature vector corresponding to each record These are the basic attributes and characteristics of a region, such as the scale of its power grid, its economic level, and its price index. For the regional power grid load density characteristics, The terrain type is represented by coding methods such as plains, hills, and mountains.

[0118] The average value of the regional characteristics calculated in the current sample set is: ;

[0119] To reflect the representativeness of regions that are closer to the overall characteristics in dimensionality reduction, the 1st... The region weight is defined for each record as follows:

[0120] ;

[0121] in, For the first The regional weight of each record For regional feature vectors With average characteristics The similarity between regions is preferably calculated using cosine similarity, with the denominator being the sum of the similarity indices of all sample regions. This sum is used for normalization; the closer a region's characteristics are to the overall average level, the better. The larger the sample size, the more complete the sample size. The sum is 1.

[0122] 3) Project cycle type and time weighting:

[0123] Based on the cyclical characteristics of power grid special cost projects, a project cycle type is introduced for time weighting, so that annual, quarterly and irregular projects have different time windows in sample selection.

[0124] First, based on the project type or business rules, mark each record with the project cycle type as follows: ;

[0125] Then, the corresponding time decay coefficients are configured for different period types as follows:

[0126] ;

[0127] in, For annual, quarterly, and irregular projects, the time decay parameter should satisfy the following condition: Smaller, used to emphasize data from the last 1-2 years. Larger values ​​are used to emphasize data from the most recent 1-2 quarters. Take the median value for regular, irregular projects.

[0128] For the The time weight is defined for each record as follows: ;

[0129] in, For the first Time weight of each record This is the time label corresponding to the current review point in time. For the first The record shows the time stamp of the expense. In order to be with the first The shorter the time decay coefficient interval corresponding to the project cycle type, the better. The closer to 1; the longer the interval, Approaching 0. By distinguishing... It can achieve differentiated time sampling for different scenarios such as annual projects and seasonal projects.

[0130] 4) Feature prior weights and weighted sample matrix:

[0131] Prior weights are set for different feature dimensions, and the feature matrix is ​​weighted by combining the sample layer weights and the feature layer weights.

[0132] First, set the prior weight vector for the feature dimension as follows: This vector is used to represent the importance of different features in classification, where, For the first The prior weights for each feature can be set, with larger values ​​for cost-related features and smaller values ​​for auxiliary descriptive features. Normalization can be performed according to set rules, for example, making .

[0133] Secondly, the time weight and the region weight are combined to form the comprehensive weight of the sample layer: ;

[0134] This formula is used to generate the overall impact weight for each record, which controls the contribution of the sample to the overall dimensionality reduction.

[0135] The sample weight diagonal matrix and the feature weight diagonal matrix are constructed as follows:

[0136] .

[0137] Finally, the feature matrix is ​​weighted as follows: ;

[0138] in, The weighted eigenma matrix is ​​multiplied on the left. Equivalent to sorting each row of samples by Scaling, right multiplication Equivalent to applying each column of features by Scaling is performed.

[0139] 5) PCA dimensionality reduction:

[0140] For the weighted characteristic matrix Principal component analysis is performed to compress high-dimensional features into low-dimensional irrelevant principal components. The PCA is performed as follows: ;

[0141] in, The sample feature matrix after PCA dimensionality reduction, the first... OK For the first The representation of each record in the principal component space. The number of principal components to be retained is generally determined by the cumulative explained variance ratio, preferably such that the cumulative explained variance of the first k principal components is not less than 98%.

[0142] (2) Cluster analysis based on DPC:

[0143] 1) Definition of distance:

[0144] Based on the principal component space, and combined with regional, temporal, and cost characteristics, a comprehensive distance for density estimation is defined.

[0145] For any two records Define the distance between the three components as:

[0146] ;

[0147] in, For regional distances, calculations can be based on the Euclidean distance of the city's latitude and longitude or a preset regional similarity matrix. For time distance, calculate directly based on the absolute value of the difference between time stamps. To eliminate dimensional differences, three scale parameters are introduced to determine the cost distance, for example, based on the absolute difference between logarithmic or normalized unit prices. And define the comprehensive distance as:

[0148] ;

[0149] in, For record and The overall distance The scale parameters are for the three dimensions of region, time, and cost, and can be taken as the median or a specified quantile of the distance between the corresponding components.

[0150] 2) Local density calculation:

[0151] The local density of each record is calculated based on the comprehensive distance to identify high-density regions. For the first... The local density of a record is defined as follows:

[0152] ;

[0153] in, For the first Local density of records For record and The overall distance, the smaller the distance, The closer to 1, the better. The greater the contribution, the greater the local density, indicating that there are more samples of the same type, from the same region, and with similar cost levels around the record.

[0154] 3) Minimum distance to high-density points:

[0155] By calculating the minimum distance from each sample to samples with higher density than itself, candidate cluster centers that simultaneously possess high density and relative isolation are identified. If the density is not the largest among all samples, then it is defined as: ;

[0156] like If it is the global maximum density (or one of them), then it is defined as: ;

[0157] in, For the first The minimum distance from a record to a higher density sample, when At maximum density, Set it to the maximum distance from other samples, so that it naturally stands out as the center in subsequent judgments. Typically, Larger and Larger points are more likely to be cluster centers.

[0158] 4) Comprehensive index and cluster number calculation:

[0159] Cluster centers are selected by combining local density and distance indicators. Furthermore, the number of project types, regions, responsible departments, and cost fluctuation levels are incorporated into the cluster count calculation. Firstly, for the... The comprehensive index is defined as follows: ; The larger the value, the more suitable the point is as a cluster center, as it is located in a high-density region and is also far away from other high-density points.

[0160] Then, the number of clusters is estimated based on the number of project types, regions, responsible departments, and cost variation coefficients. for:

[0161] ;

[0162] in, This represents the number of item types in the current sample. This represents the number of categories in the current sample region. This represents the number of department categories in the current sample, such as the Equipment Department, Marketing Department, etc. Cost characteristics The coefficient of variation is defined as: ;

[0163] And can be cut off at Intervals are used for stability:

[0164] This is an adjustment factor used to control the overall cluster size.

[0165] These are the lower and upper limits for the number of clusters, used to prevent the number of clusters from being too few or too many;

[0166] This is an estimate of the initial cluster number without truncation;

[0167] To determine the final number of clusters, first... Take the top whole number, then crop it using the top and bottom bounds;

[0168] By introducing This can reflect the differences in management standards among different departments at the cluster level, preventing expense records that should belong to different departments from being classified into the same category.

[0169] 5) Cluster center selection:

[0170] According to comprehensive indicators Sort the samples and select the cluster centers.

[0171] All samples Sort by largest to smallest, and take the first few. Each sample is taken as the cluster center, and the set is denoted as:

[0172] ;

[0173] For each cluster center Assign cluster labels as: ;

[0174] in, This is the cluster center index set.

[0175] 6) Non-central sample cluster allocation:

[0176] Non-central samples are assigned to the corresponding cluster centers according to the principle of moving towards high-density points.

[0177] For each non-center sample Its high-density neighbor set is defined as: ;

[0178] exist Find the sample The sample with the smallest overall distance is: ;

[0179] Sample The cluster label is set to the cluster label of the high-density neighbor: ;

[0180] Through the above process, all samples are assigned to a cluster center along the direction of increasing density, forming categories corresponding to multiple density peaks.

[0181] S3, cost range calculation of multi-level log-normal mixed effects and empirical Bayesian benchmark.

[0182] Based on the similar expense categories obtained in S2, a multi-level log-normal mixed-effects model is constructed. This model incorporates multiple hierarchical factors, including a three-tiered geographical structure (province-city-county), project type, responsible department, expense category, and time period. It provides point and interval estimates of the reasonable unit price for each expense list row. Furthermore, it uses an empirical Bayesian method to benchmark and correct historical data from similar projects, forming a cost range that simultaneously reflects statistical regularities and enterprise experience. The output prediction center, correction interval, and confidence level will directly serve as inputs for subsequent dynamic decision-making, achieving a natural connection from category clustering to interval reference and then to dynamic game theory.

[0183] (1) Definition of variables and logarithmic transformation:

[0184] To accommodate the common positive skewed distribution and long-tail characteristics of unit prices for special costs, this step involves modeling in logarithmic space and taking the logarithm of the unit price.

[0185] For the One expense record, recorded as To observe the unit price, The unit price is logarithmic. The fixed-effects feature vector contains:

[0186] Dummy variables such as project type, responsible department, and S2 category target signature;

[0187] Regional characteristics, including provincial-level, prefectural-level, and county-level companies;

[0188] Specifications (capacity, voltage level, line length unit, etc.);

[0189] Scale characteristics (quantity, total price, unit size, etc.);

[0190] Time characteristics (year, time difference relative to the base period, etc.);

[0191] The regional stratum is clearly divided into three levels: "province-city-county," and each level is assigned a random effects index:

[0192] For the first The record belongs to the provincial-level company;

[0193] For the local city-level company;

[0194] For the county-level company or district / county power supply unit;

[0195] At the same time, other hierarchical indexes are retained;

[0196] Group by project type;

[0197] Grouping by department;

[0198] Group the S2 cluster categories;

[0199] (2) Multilevel log-normal mixed-effects model:

[0200] The fixed effects component, consisting of the global intercept and coefficients corresponding to the eigenvectors, characterizes the systematic impact of factors such as project type, responsible department, province / city / county, scale, and time on the average cost level. The random effects component sequentially sets up provincial, municipal, county, project type, department, and category random effects to characterize the degree of deviation relative to the overall network average at each level. By introducing the assumption of independent normal distribution for each level of random effects and introducing a variance parameter for each level, the model can quantify the contribution of each level to overall cost fluctuations.

[0201] A multi-level linear mixed-effects model is constructed for unit price in logarithmic space as follows:

[0202] ;

[0203] in, For the global intercept term, This is a fixed-effects coefficient vector, reflecting the systematic impact of project type, department, region, specifications, scale, time, etc., on the logarithmic unit price. To represent provincial random effects, this characterizes the deviation of the overall cost level of different provincial companies from the network-wide average. This represents the random effect at the prefecture-level city level, characterizing the cost differences between different prefecture-level cities within the same province. This represents a county-level random effect, reflecting the differences between remote counties, mountainous companies, and the average level of their respective cities. For random effects at the project type level, For departmental level random effects, This is a random effect in the S2 category layer. The observation error term characterizes the residual fluctuation of the unit price after considering the factors mentioned above:

[0204] The random effects and error term are assumed to be:

[0205] ;

[0206] ;

[0207] The random effects and errors at each level are independent of each other. Variance components It depicts the contribution of different levels to cost fluctuations.

[0208] The matrix form can be written as:

[0209] ;

[0210] in, Design a matrix for fixed effects. Design matrices for random effects at each level. This corresponds to the random effects vector.

[0211] (3) Parameter estimation and empirical Bayesian estimation of random effects:

[0212] In the parameter estimation phase, the variance components of the random effects at each level are first estimated using the restricted maximum likelihood method, including the variances at the provincial, municipal, county, project type, responsible department, category, and error levels. Then, the posterior mean of the random effects at each level, i.e., the best linear unbiased prediction, is calculated using empirical Bayesian methods. For each level, the random offset is derived based on the difference between the actual observed cost data for that province and the overall network model.

[0213] 1) Variance component estimation:

[0214] Within the REML framework, Perform maximum likelihood or restricted maximum likelihood estimation to obtain ;

[0215] 2) Fixed effects estimation:

[0216] Given the variance component estimates, the generalized least squares estimate of the fixed effects coefficient is:

[0217] ;

[0218] in, It consists of the variance components of each layer and the design matrix.

[0219] 3) Empirical Bayesian estimation of random effects:

[0220] In the estimated With respect to the variance component, the posterior mean (best linear unbiased prediction, BLUP) of the random effects at each level is:

[0221] ;

[0222] in, This is the diagonal matrix of the variance blocks of random effects in each layer.

[0223] For the The total random effects of the records are estimated as follows:

[0224] ;

[0225] The conditional mean estimate of the logarithmic unit price is:

[0226] ;

[0227] (4) Unit price point estimation and basic interval under log-normal distribution

[0228] The conditional mean and prediction variance obtained in step (3) are transformed into the central unit price and basic interval of the original unit price space. Under the modeling assumption, the logarithmic unit price follows a normal distribution given the features and hierarchical structure; therefore, the original unit price follows a log normal distribution. After the model is fitted, it will provide a conditional mean of the logarithmic unit price and the corresponding prediction variance for each record.

[0229] Within the framework of the mixed-effects model, logarithmic unit price Given the characteristics, it approximately follows a normal distribution as follows: ;

[0230] in, The predicted variance is given by the model covariance structure. Theoretically, the predicted variance can be obtained from the random effects variance and the error variance through matrix operations, and can be written in the form of:

[0231] ;

[0232] In this method, reviewers are not required to manually derive the above matrix formulas. Instead, statistical software is used to automatically export the prediction variance of each record after the model converges, and the reasonableness is verified by combining the residual distribution test results.

[0233] Under the original unit price space It follows a log-normal distribution:

[0234] ;

[0235] Therefore, we can conclude that:

[0236] 1) Median (as an estimate of the cost center): ;

[0237] 2) Given the confidence level, select the lower quantile. and upper percentile (e.g., 0.1 and 0.9), let the quantiles of the standard normal distribution be... Then it is:

[0238] ;

[0239] in, This corresponds to approximately 80% of the central interval.

[0240] This yields the initial reasonable unit price range based on LMM. and central estimation .

[0241] (5) Empirical Bayesian peer-to-peer correction and upper edge tightening:

[0242] By further integrating similar experiences, Bayesian benchmarking, and institutional upper limit constraints, the basic range is tightened.

[0243] 1) Benchmarking grouping and within-group statistics:

[0244] According to management standards, the benchmarking group is defined as follows: ;

[0245] To ensure sufficient sample size, the regional level uses the province as the basic unit for benchmarking, while the category level is still divided according to the S2 category, so that the benchmarking is carried out under the scope of "same type – same department – ​​same province – same category".

[0246] For each benchmarking group Extract its historical reference sample set:

[0247] It has a historically reasonable unit price: ;

[0248] The mean within the logarithmic space is: ;

[0249] And calculate the logarithmic mean of the entire network or the upper layer aggregation: ;

[0250] in, This represents the total number of reference samples.

[0251] 2) Empirical Bayesian reduction of dynamic smoothing constant and benchmark coefficient:

[0252] Using the empirical Bayesian approach, the bias of group means is reduced to: ;

[0253] in, The group credibility weight is defined as:

[0254] ;

[0255] Smoothing constant Dynamic values ​​are adopted to take into account the difference in sample size between popular and unpopular categories in the power grid special cost.

[0256] For popular categories with a large historical sample size, when When the value is significantly greater than the preset threshold, take This allows large sample categories to retain more of their own characteristics; for less popular categories with smaller sample sizes, when When the value is small or below the threshold, take Increase the reduction efforts towards the average level of the entire network to avoid excessive benchmarking bias in small sample groups.

[0257] Will Mapping back to the original unit price space, the correction coefficient for similar benchmarks is defined as follows: ;

[0258] For the The record has a benchmarking correction factor of: .

[0259] 3) Tightening of benchmarks at the upper limit of the range and embedding of policy ceilings:

[0260] Given an initial upper bound Based on this, and combined with the benchmarking coefficient and the upper limit of the clauses, the upper limit is conservatively tightened. The lower limit is determined based on the limit of each cost item.

[0261] Based on the comprehensive model range, comparable benchmarks, and policy ceilings, the upper limit is constrained as follows:

[0262] ;

[0263] in, To align with the corrected upper edge, This represents the upper quantile of the mixed-effects model under the log-normal assumption. This represents the upper limit of the empirical range of the central value multiplied by the benchmark coefficient, based on the experience of similar projects. The maximum unit price of the clauses is automatically extracted from the policy clause library and serves as a rigid constraint.

[0264] The lower edge will not be tightened in this step, and will be directly recorded as: ;

[0265] Therefore, the final reasonable unit price range for each record is updated as follows: ;

[0266] This range simultaneously reflects information from three aspects: statistical models, high-low stratification benchmarking, and institutional constraints.

[0267] (6) Model confidence construction:

[0268] To differentiate between reliable model results and those requiring stricter handling, a quantitative confidence index needs to be constructed. On one hand, prediction variance reflects the model's uncertainty regarding the unit price of a given record; a larger variance indicates less stable price predictions. This variance can be mapped to a confidence factor using a simple exponential decay function. On the other hand, a larger benchmark sample size and a more reasonable empirical Bayesian reduction coefficient indicate more sufficient historical information. A benchmark confidence factor can be obtained through sample size normalization. Multiplying these two results in a comprehensive confidence level between 0 and 1.

[0269] 1) Measurement of local uncertainty: ;

[0270] The larger the value, the higher the prediction uncertainty of the record in the logarithmic space.

[0271] 2) Measurement of the credibility of the benchmark group: ;

[0272] The more samples, The smaller, The closer it is to 1.

[0273] 3) Overall confidence level: ;

[0274] in, Follow An increase followed by a monotonically decreasing variance indicates a higher confidence level. This demonstrates the impact of sample size and temperature attributes of the benchmark group on the confidence level. This can serve as a basis for determining the need for a conservative strategy when identifying low-confidence samples in the next step.

[0275] S4, Dynamic cost decision-making based on Markov review path and strategy optimization:

[0276] This step, based on the data output from S1–S3, introduces a Markov decision process to model the state evolution, action selection, and benefit outcomes of a single expense item during the review process. First, a review state vector is constructed using information such as the reasonable unit price range, the current declared unit price, arithmetic check results, anomaly markers, and confidence levels to characterize the current level and risk of this expense item. Then, several practical handling methods are abstracted into a finite number of actions, including keeping it unchanged, proportionally reducing it uniformly, splitting and rectifying it by sub-item, and direct rejection. The changes in unit price and markers under different actions are specified, thus forming a Markov process where state plus action yields the next state. A simple scoring function is then designed to uniformly convert factors such as whether the reduced cost has entered a reasonable range, whether it still exceeds limits, data reliability, and workload into scores. Given a state, the action with the lowest score is selected as the priority strategy. Finally, termination conditions are set, such as the cost item has entered a reasonable range, still exceeds limits after multiple rounds of adjustments, and the maximum number of rounds has been reached, completing the dynamic review loop for this expense item.

[0277] (1) State-Action-Transition Structure:

[0278] 1) Definition of review status:

[0279] For the For a single cost record during a round of review, the state vector is defined as follows:

[0280] ;

[0281] The price deviation is defined as follows:

[0282] ;

[0283] in, For the first The unit price declared during the shift, The lower and upper limits of the reasonable unit price range are to be output. The midpoint of the reasonable range reflects the reasonable level of the benchmark. The reasonable range width reflects the allowable fluctuation range. To normalize price deviations, the compliance and anomaly markers used to measure the relative reasonableness of the current declared unit price are as follows:

[0284] ;

[0285] in, To determine whether the unit price exceeds the maximum limit under the terms (0 / 1). Whether it is near the upper limit of the terms (e.g., exceeding 94% of the upper limit) (0 / 1). To determine whether arithmetic consistency is passed (total price and quantity × unit price error are within the allowable range, 0 / 1). Anomalies (0 / 1) detected by robust Z-scores are marked as abnormalities. The anomaly markers (0 / 1) detected by the interquartile range.

[0286] The model and data credibility section is as follows:

[0287] Confidence level of the cost model generated for S3;

[0288] The confidence level of the missing fill obtained in S1;

[0289] The benchmarking and policy constraints section is as follows: ;

[0290] in, If graph attention alignment is used, then this represents the alignment convergence coefficient of the S4 output; if not, it can be... Or replace with the initial value, The maximum unit price is specified in the terms and conditions. This is the ratio of the upper limit to the current declared unit price, used to reflect the distance from the red line.

[0291] 2) Action set:

[0292] For a single fee record, the set of optional actions is defined as follows: ;

[0293] in, If rejected, the corresponding item will not be counted or will be returned. To ensure uniform weight reduction, this fee will be reduced proportionally across the board, with a reduction factor of [value missing]. To facilitate structured rectification, rectification and reduction plans are provided for several sub-items, with the overall reduction equivalent to: ;

[0294] The parameter constraints are: ;

[0295] in, To ensure a consistent retention ratio after compression, This represents the overall equivalent reduction ratio after rectification.

[0296] 3) Unit price update and state transition:

[0297] For the Unit price declared during the shift The new unit price after the action is executed is:

[0298] ;

[0299] The normalization bias is then updated as follows:

[0300] ;

[0301] in, The unit price declared after the action is executed. For the first The action selected in the step.

[0302] state arrive The transition relationship is jointly determined by the unit price update, the anomaly marker update, and the confidence adjustment, and can be expressed as the transition probability as: .

[0303] (2) Rigid boundary and conservative tightening:

[0304] For samples with low model confidence and imputation confidence, a conservative tightening mechanism at the upper edge of the interval is adopted. A lower bound threshold for the overall confidence is defined. for: ;

[0305] Under these conditions, tightening the upper edge is as follows: ;

[0306] in, This is a confidence threshold; a higher threshold indicates stricter requirements for the model and data. This is a conservative coefficient used to limit the degree to which the upper edge of the interval deviates from the midpoint. This is the upper limit after the conservative tightening, and it will be used as a reference in the subsequent profit function.

[0307] (3) Immediate revenue function and overall optimization objective:

[0308] 1) Immediate return function:

[0309] Defined in state Next action The immediate benefits are:

[0310] ;

[0311] The action cost item is defined as follows:

[0312] ;

[0313] in, The reward weight is within a reasonable range. For penalties exceeding the maximum allowed under the terms, The penalty weight is higher than the upper limit of the conservative range. Penalty weights for insufficient model confidence. To compensate for insufficient credibility, As the weight of the penalty for the action cost, To standardize the cost coefficient of weight reduction actions, A larger value indicates a greater pressure drop, which in turn leads to higher communication and adjustment costs. The cost coefficient for rectification actions. The larger the value, the greater the scope of rectification and the higher the corresponding work cost. This is an indicator function that takes the value 1 if the condition is true and 0 otherwise. The immediate benefits are determined by comprehensively considering rationality, compliance, confidence level, and workload costs.

[0314] 2) Multi-step expected return maximization objective:

[0315] Given a strategy The review process begins from... arrive The expected cumulative return is:

[0316] ;

[0317] The goal of strategy optimization is: ;

[0318] in, As a discount factor, it controls the weight of future returns. This represents the maximum number of review rounds or the termination step. Decision rules (strategies) for moving from state to action. In strategy The expected cumulative return.

[0319] (4) Minimum feasible compression coefficient and rectification allocation:

[0320] 1) Minimum feasible compression factor:

[0321] Given a conservative upper edge and the upper limit of terms Under the premise of simplifying the unified weight reduction decision, a "minimum feasible compression coefficient" can be given first:

[0322] ;

[0323] in, The current declared unit price (initially...) (If there have been multiple adjustments, then the price is the most recent one). The safety margin coefficient is used to reserve a safety margin between the upper limit and the lower limit, such as... This indicates that a 2% safety margin is reserved. To determine the minimum feasible retention ratio while simultaneously satisfying both the conservative upper limit and the safety boundary conditions, if Not lower than a certain preset threshold (e.g.) If so, a unified deweighting action can be directly adopted. As a preferred option.

[0324] 2) Structured rectification and allocation:

[0325] When simply reducing the weight of expenses uniformly is insufficient to meet the requirements of reasonableness and upper limit constraints, it is necessary to break down the reduction target into several sub-items or constituent items. Assume the expense includes several adjustable sub-items. Define the monetary weight of each sub-item as follows:

[0326] Total Amount: ;

[0327] Let the pressure drop ratio of each sub-item be... Then the overall pressure drop constraint is:

[0328] ;

[0329] Among them, the amount For sub-items The corresponding claim amount, the total amount is the total claim amount for this expense line. The percentage of the amount for each sub-item For pairs of items Required pressure drop ratio The overall target reduction ratio;

[0330] Based on this, upper limit constraints can be set according to the rigidity of sub-items and business necessity. A set of feasible methods can be obtained through simple linear programming. This serves as a structured rectification list.

[0331] 5) Strategy solution method:

[0332] In offline scenarios, key continuous state dimensions can be discretized into a finite set of grids to construct a finite state-action space, and the optimal policy can be solved using value iteration or policy iteration methods.

[0333] Define the state-action value function as follows:

[0334] ;

[0335] Through iterative updates and The process continues until convergence, yielding the optimal strategy:

[0336] ;

[0337] in, In the state Next action And the expected return when acting according to the optimal strategy, For state Maximum expected return In the state ,action Transition to state The probability (which can be estimated from historical data or set according to simplified rules).

[0338] The solution results can be directly stored as a "state-action decision table" or "deviation-compression curve", which can be easily called directly in the system.

[0339] 6) Termination conditions and result feedback:

[0340] Decision termination conditions:

[0341] During the review process, dynamic decision-making for this expense item will terminate and a rigid veto will be triggered if any of the following conditions are met: the declared unit price still cannot be reduced after multiple rounds of reduction or rectification. The following may still clearly exceed the limits of the terms. .

[0342] Successfully returning to a reasonable range means that, after unified weight reduction or rectification, the declared unit price meets the requirements. Furthermore, arithmetic consistency and anomaly marking are already in an acceptable state.

[0343] The maximum number of rounds is the number of review rounds. Reaching the preset limit To avoid excessive round trips, the result corresponding to the current optimal action is forcibly output.

[0344] Upon termination, this step will output, including but not limited to, the following information:

[0345] The final review conclusion was Reject.

[0346] Downweighting (uniform weight reduction) and providing compression coefficients. ;

[0347] Rectify (accept after rectification) and provide a set of sub-item reduction ratios. .

[0348] The corresponding reasonable unit price range is Or conservative upper edge .

[0349] The decision-making criteria include the initial declared unit price, the midpoint and width of the reasonable range, the benchmarking correction factor, the upper limit of policy provisions, anomaly markers, and confidence levels.

[0350] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0351] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multimodal fusion and dynamic game-driven intelligent review method for power grid special costs, characterized in that, Includes the following steps: S1. Collect and preprocess multi-source data of power grid special cost projects in a unified manner. The multi-source data includes project information and cost details. Complete the structuring of cost fields, arithmetic consistency verification, merging of duplicate items, filling of missing values ​​and identification of outliers to form a standardized review dataset. S2, based on attention-weighted PCA and density peak clustering, constructs a weighted feature space by combining project type, regional characteristics and cost details, and divides similar cost categories by comprehensive distance and local density; S3, based on similar cost categories, constructs a multi-level log-normal mixed effect model, integrates the hierarchical factors of region, department, and project, and combines empirical Bayesian methods to output a reasonable unit price range and confidence level; S4 constructs a Markov decision process with declared unit price, reasonable range, abnormal marker and confidence level as states, defines a variety of review actions and payoff functions, optimizes the strategy and outputs the final review conclusion and adjustment suggestions.

2. The intelligent review method for power grid special costs driven by multimodal fusion and dynamic game theory as described in claim 1, characterized in that, S1 includes: S11, uniformly collect project information and expense details of power grid special cost projects. The project information includes project number, project name, year, unit, region and project type. The expense details include expense line number, expense name or expense item, specifications, unit of measurement, quantity, declared unit price and total amount. The field format and expression of multi-source data are standardized. S12 performs an arithmetic consistency check on the total amount, quantity, and declared unit price for each expense detail, calculates the error value and compares it with the set allowable error threshold. If the error value is within the error threshold range, it is marked as passed; otherwise, it is marked as failed. It also identifies duplicate records in the same project that are highly similar in expense name, specifications, and unit of measurement, performs quantity merging and total price merging operations, and backtracks the declared unit price after merging. S13, For missing quantities or declared unit prices in the cost details, a median imputation strategy based on project type, region and year is used for statistical imputation, and the imputation confidence weight is calculated based on the imputation sample size. S14. In the same group of expense details, identify abnormal declared unit prices based on robust Z-scores and box plot rules.

3. The intelligent review method for power grid special costs driven by multimodal fusion and dynamic game theory as described in claim 2, characterized in that, S2 includes: S21. The project type, region, and cost details are encoded into feature vectors. A comprehensive feature vector is constructed, and regional and time weights are introduced. Combined with prior weights, a weighted feature matrix is ​​generated. PCA is then performed on the weighted feature matrix to reduce its dimensionality and extract the principal components that explain the highest proportion of variance, thus forming the dimensionality-reduced feature representation. S22. In the principal component space, define the comprehensive distance that combines region, time and cost, and calculate the local density of each cost detail and the minimum distance to the high-density point. Select cluster centers based on local density and minimum distance, and dynamically estimate the number of clusters based on project type, region and cost fluctuation. All cost details are assigned to the corresponding cluster centers according to the density increase path to complete the clustering of the same cost category.

4. The intelligent review method for power grid special costs driven by multimodal fusion and dynamic game theory as described in claim 3, characterized in that, S21 includes: S211 maps project type, region, and cost details into numerical feature vectors, and constructs a comprehensive feature vector. ; S212 determines the regional weight of each expense detail in clustering dimensionality reduction by constructing a regional feature vector that includes basic attributes, load density, and terrain type, and calculating its similarity to the average regional feature vector. ; S213, set the time decay coefficient according to the project cycle type of the cost details, and calculate the time weight in combination with the recorded timestamp. ; S214 introduces prior weights for the feature dimensions, combining them with the time and region weights of the cost details to construct a weighted feature matrix that integrates multiple weight information. ; S215, for the weighted characteristic matrix Principal component analysis is performed to compress high-dimensional features into low-dimensional irrelevant principal components.

5. The intelligent review method for power grid special costs driven by multimodal fusion and dynamic game theory as described in claim 4, characterized in that, S22 includes: S221, combining region, time, and cost details, calculates the three-dimensional normalized distance and introduces a scale parameter to construct a comprehensive distance. ; S222, calculates the local density of each cost detail based on the comprehensive distance. ; S223, calculate the minimum distance from each cost detail to cost details with a higher density than itself. ; S224, constructing a comprehensive index based on local density and minimum distance. In conjunction with the number of project types, regions, responsible departments, and cost fluctuation levels, the number of cluster categories is adaptively estimated. ; S225. Select cluster centers based on the comprehensive index ranking, and allocate the non-center cost details to the clusters of the nearest high-density samples according to the density-guided principle, thus completing the classification of cost categories of the same type.

6. The intelligent review method for power grid special costs driven by multimodal fusion and dynamic game theory as described in claim 5, characterized in that, S3 includes: S31, perform logarithmic transformation on the declared unit price and construct a fixed-effects feature vector. This includes project type, responsible department, province, city / county, geographical features, specifications, and time offset; S32, construct a multi-level log-normal linear model with both fixed and random effects to explain the offset of the unit price of the cost details under different geographical, project, department and category dimensions; S33, the restricted maximum likelihood (REML) method is used to estimate the variance components at each level. The variance components include the variances of the province, city, county, project type, responsible department, category, and error. The posterior mean of each random effect is calculated using the empirical Bayes method to obtain the bias term at each level. Finally, the conditional mean of each cost detail is obtained by superimposing the fixed effect and the random bias term at all levels. S34. Using the conditional mean and predicted variance output from the multilevel log-normal linear model, a log-normal distribution is constructed, and a central estimate is generated. confidence interval ; S35, Construct the historical average unit price of similar expense details Compared with the overall average unit price And based on the historical average unit price Compared with the overall average unit price The mean deviation is adjusted by the empirical Bayesian weighting factor to narrow the upper edge of the unit price range; S36, Introduction of Dimensional Factor And combined with empirical Bayesian factors Construct a comprehensive confidence level .

7. The intelligent review method for power grid special costs driven by multimodal fusion and dynamic game theory as described in claim 6, characterized in that, S4 includes: S41, construct a state vector based on the current cost details, define operation actions and set relevant parameters, and construct a state transition function based on changes in declared unit price and anomaly marker update rules; S42, for cost details with a comprehensive confidence level lower than the lower limit of comprehensive confidence level, a conservative mechanism is triggered to adjust the upper limit of the reasonable unit price range to the weighted lower limit of the current upper limit, the upper limit threshold and the correction target; S43, construct an immediate benefit function that includes rewards and penalties as well as adjustment costs, and use maximizing the cumulative expected benefit throughout the review process as the optimization objective to guide the dynamic selection of strategies; S44, set the minimum compression ratio Check the feasibility of the pressure reduction; if it is met, generate the compression ratio for each sub-item according to the full budget proportion. To achieve a reasonable allocation of resources for reduction and rectification; S45. Based on Markov processes, construct Q and V functions, and use Bellman equations and policy iteration methods to solve for the optimal execution strategy of cost compression actions. S46 terminates the strategy when the unit price falls back to a reasonable range, the confidence level meets the standard, or the number of rounds of review exceeds the limit, and outputs the final reduction suggestion, corrected unit price, and range adjustment results.

8. The intelligent review method for power grid special costs driven by multimodal fusion and dynamic game theory as described in claim 7, characterized in that, S41 includes: S411, the state vector is composed of the declared unit price of the current cost details, the upper and lower limits of the reasonable unit price range, the price deviation, the anomaly marker and the overall confidence level; S412 sets the execution action for the current status of the expense details, including veto, uniform weight reduction and overall item compression, in order to control the declared unit price; S413, recalculate the state vector for the next moment based on the declared unit price and normalized deviation after the action is executed, and construct the state transition relationship by combining the anomaly label and confidence adjustment rules.

9. The intelligent review method for power grid special costs driven by multimodal fusion and dynamic game theory as described in claim 8, characterized in that, S43 includes: S431, the immediate benefit function takes the comprehensive score brought by the action in the current state as its core. It quantifies the immediate benefit of the current review adjustment by rewarding price adjustments within a reasonable price range, penalizing exceeding the upper limit and insufficient confidence, and taking into account the action cost caused by weight reduction or compression. ; S432, multi-step expected return targets use discount factors to weigh current and future returns, throughout the entire review process. arrive Calculate the cumulative expected return And by maximizing This guides the strategy to select the optimal sequence of review actions.

Images