Power grid maintenance resource consumption prediction method and system

Through density-based noise applications such as spatial clustering and fuzzy C-mean clustering, combined with neural network prediction model, data quality and model adaptability problems in power grid maintenance resource consumption management are solved, and more accurate and dynamic resource consumption is achieved Forecasting improves the accuracy of cost control and decision-making support.

CN119005432APending Publication Date: 2024-11-22ECONOMIC TECH RES INST OF STATE GRID HENAN ELECTRIC POWER

Patent Information

Application Number
CN202411122801.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

In the management of power grid maintenance resource consumption, the existing technology has problems such as uneven data quality, insufficient model adaptability and accuracy, lack of dynamic prediction capabilities, and low integration of decision-making systems.

Method used

The density-based noise application spatial clustering is used to identify outliers and interpolate missing values ​​for historical data, combine fuzzy C-mean clustering to divide the job type, build a neural network prediction model, and use normal model parameter estimation and chi-square matte The combined test obtains historical references for resource consumption and dynamically adjusts the weights to improve prediction accuracy and adaptability.

Benefits of technology

Enhanced data integrity and analysis reliability, improve the accuracy and adaptability of the prediction model, ensure the dynamic reflection of resource consumption and the accuracy of cost control, and provide a more reliable reference for project review and cost control. in accordance with.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119005432A_ABST
    Figure CN119005432A_ABST
Patent Text Reader

Abstract

The invention relates to a power grid maintenance resource consumption prediction method and system, belongs to the technical field of data preprocessing and identification, and solves the problems of complex data processing, insufficient model precision and adaptability, insufficient dynamic prediction capability and the like in the prior art. The method comprises the steps of obtaining data of resource consumption of multiple overhaul operations and influence factors thereof; carrying out abnormal value identification and missing value interpolation processing on the historical data; performing clustering processing on the resource consumption in combination with historical data; randomly extracting sample data from the resource consumption of the maintenance work in the same region, and performing data analysis on the sample data through model parameter estimation, fitting test and prediction interval estimation to obtain a prediction interval of the historical reference quantity of the resource consumption; constructing and training a neural network prediction model, and predicting a resource consumption prediction target in combination with real-time data; and determining the final resource consumption based on the resource consumption historical reference quantity and the resource consumption prediction target. And data processing is simplified, and model precision, fitness and the like are provided.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of data preprocessing and identification, and in particular to a method and system for predicting power grid maintenance resource consumption. Background Art

[0002] With the increasing complexity of industrial production and infrastructure maintenance activities, efficient management and optimization of maintenance operation resources have become an important challenge for enterprises and organizations. Traditional resource scheduling and cost estimation are often based on rough estimates or historical averages, lacking detailed consideration of individual differences in operations and external environmental factors, resulting in unreasonable resource allocation and inaccurate cost control. Therefore, the use of advanced data-driven methods to finely manage maintenance operation resource consumption has become a key way to improve operational efficiency and cost-effectiveness.

[0003] In recent years, the development of data science and machine learning technology has brought innovation to this field. By integrating multi-source data such as geographic information systems (GIS) and device management systems (DMS), rich information about the work site can be collected, including the amount of work resource consumption, geographic location, work environment characteristics, regional economy and power grid conditions, etc. These data provide a solid foundation for in-depth analysis of work resource consumption patterns.

[0004] The existing technologies have the following problems: (1) Data quality varies: There are often missing values ​​and outliers in actual operation data, which affect the accuracy and reliability of the analysis. Although there are clustering algorithms such as DBSCAN for outlier detection, how to efficiently and accurately handle these data problems remains a challenge. (2) Model adaptability and accuracy: The resource consumption patterns of different types of operations vary greatly, and a single model is difficult to take into account all situations. How to select or design a suitable prediction model based on different operation characteristics, and how to fully consider multi-dimensional factors such as region and environment in model construction to improve prediction accuracy, are still urgent issues to be solved. (3) Real-time and dynamic nature of resource consumption prediction: The operating environment and conditions change over time, and existing prediction models often lack sufficient dynamic adaptability and are difficult to reflect the impact of these changes on resource consumption in real time. (4) The integration of cost control and decision support is not high: Although there is historical data as a reference, how to effectively integrate the prediction results into the project review and cost budget process to achieve intelligent resource scheduling and refined cost control still needs further exploration and practice.

[0005] In summary, although the existing technologies and methods have improved the management of job resource consumption to a certain extent, they still face problems such as complex data processing, insufficient model accuracy and adaptability, lack of dynamic prediction capabilities, and insufficient integration with the decision-making system. These problems require innovative technical means and strategies to overcome. Summary of the invention

[0006] In view of the above analysis, the embodiments of the present application aim to provide a method and system for predicting power grid maintenance resource consumption, so as to solve the problems faced by the prior art, such as complex data processing, insufficient model accuracy and adaptability, lack of dynamic prediction capabilities, and insufficient integration with the decision-making system.

[0007] On the one hand, an embodiment of the present application provides a method for predicting resource consumption of power grid maintenance, including: obtaining data on resource consumption of multiple maintenance operations and their influencing factors, wherein the data includes historical data and real-time data; using density-based noise application spatial clustering to identify outliers on the historical data, wherein the historical data is interpolated for missing values; combining the historical data after outlier processing, clustering the maintenance operation resource consumption of different types of objects of the same maintenance operation; randomly sampling sample data of resource consumption of the same region and the same maintenance operation, and performing data analysis on the sample data through normal model parameter estimation, chi-square fit test and prediction interval estimation to obtain a prediction interval of a historical reference amount of resource consumption; constructing a neural network prediction model and using the historical data for training, and combining the real-time data of the maintenance operation to be predicted, predicting a resource consumption prediction target through the trained neural network prediction model; and determining the final resource consumption based on the historical reference amount of resource consumption and the resource consumption prediction target, wherein the historical reference amount of resource consumption is obtained based on the prediction interval of the historical reference amount of resource consumption.

[0008] The beneficial effects of the above technical solution are as follows: 1. Enhance outlier identification and data integrity: The DBSCAN clustering algorithm is used for outlier detection to effectively eliminate outlier data that may distort the analysis results and ensure the reliability of the analysis basis. At the same time, the improved stratified mean interpolation model is used to handle missing values ​​to ensure data integrity and improve the accuracy of model training and prediction. Refined job type division and resource consumption statistics: Combined with the fuzzy C-means clustering algorithm, the fine classification of job objects is achieved to ensure that statistical analysis and prediction are closer to actual job characteristics. Through in-depth statistical analysis of resource consumption of different types of jobs, more accurate historical reference quantities are obtained to provide high-quality basic data for subsequent predictions. 2. Improve the accuracy and adaptability of the prediction model: Use statistical methods such as normal model parameter estimation and chi-square fit test to ensure that the model is based on reasonable assumptions, and provide quantitative evaluation of the volatility of resource consumption through prediction interval estimation. Combined with the improved neural network method, the model can not only capture complex data relationships, but also dynamically adjust according to the specific information of the job to be tested, thereby improving the flexibility and accuracy of the prediction. 3. Dynamic weight adjustment and cost control optimization: In the forecasting process, by dynamically adjusting the weights of historical data and forecast results, cost volatility is taken into account to achieve a more accurate measurement of the rationality of resource consumption. This mechanism ensures that the forecasting model can reflect the latest market dynamics and changes in operating conditions, providing a more practical and reliable reference for project review and cost control.

[0009] Based on further improvements of the above method, obtaining data on resource consumption of multiple maintenance operations and its influencing factors further includes: using a database to obtain historical data on resource consumption of multiple maintenance operations and its influencing factors, wherein the influencing factors in the historical data include the geographical location, environment, region and power grid-related data of the equipment to be maintained; and using multiple sensors to obtain real-time data on resource consumption of multiple maintenance operations and its influencing factors, wherein the influencing factors in the real-time data include the geographical location, environment, region and power grid-related data of the equipment to be maintained.

[0010] Based on the further improvement of the above method, the use of density-based noise application spatial clustering to detect outliers on the historical data to interpolate missing values ​​on the historical data further includes: in a stratified identification step, the historical data is stratified according to equipment type, operation type, geographical location or environment type to identify outliers, wherein the outliers include missing values; in a layered mean calculation step, within each layer, the mean of the complete observation records is calculated; in a stratified interpolation step, when there are missing value records in the historical data, the missing value is filled with the mean within the layer according to the layer to which the missing value belongs; in a weighted or complex interpolation step, considering the variability within the layer and the sample size, different weights are assigned to different layers or weighted average or multiple interpolation is used to reflect the internal structure of the historical data; and in an iteration and verification step, the stratification and statistics are re-evaluated based on the interpolated historical data, and iteration is performed again until a predetermined convergence standard or an upper limit on the number of iterations is met.

[0011] Based on further improvements of the above method, the historical data is stratified according to equipment type, operation type, geographic location or environment type to identify outliers, including: setting a domain radius ε and a minimum number of points MinPts, wherein the domain radius ε represents the proximity relationship between points, and the minimum number of points MinPts represents the minimum number of points required to be reached in the neighborhood of a point; for each point in the historical data, the distance between each point and all other points is calculated; based on the calculated distance between each point and all other points, each data point in the historical data is divided into a core point, a boundary point and a noise point, wherein when there are at least the minimum number of points MinPts in the ε neighborhood of the current point, the current point is marked as a core point; when the current point is directly or indirectly connected to the core point, but there are less than MinPts points in the ε neighborhood of the current point, the current point is marked as a boundary point; and when the current point is a data point that is neither a core point nor a boundary point, the current point is marked as a noise point as an outlier.

[0012] Based on the further improvement of the above method, combined with the historical data after outlier processing, clustering the maintenance operation resource consumption of different types of objects of the same maintenance operation further includes: based on the historical data after outlier processing, characteristic induction of different operation types is performed to obtain multiple feature vector sample sets to be clustered, wherein the feature vector sample sets include the number of equipment failures, equipment operating years, equipment type, labor consumption, construction machinery consumption, material consumption, equipment consumption, temperature, wind speed, terrain, regional GDP, and regional electricity consumption; using the multiple feature vector sample sets to construct operation type portraits through fuzzy C-means clustering, so that the multiple feature vector sets are re-divided into multiple cluster sample clusters, wherein the membership degree and cluster center are calculated by the following formula:

[0013]

[0014] Among them, u ij is the i-th sample x i The membership degree of the jth cluster; c j is the cluster center of the jth cluster, with a dimension of n; ||*|| represents the distance between the sample and the cluster center; m is the feature vector sample set X = {x1, x2, ..., x m}, each sample x i ={x i1 ,x i2 ,...,x in} has n eigenvectors; when X is divided into C categories, each subset is represented by C = {c1, c2, ..., c k}; The membership coefficient and cluster center are iterated through the membership and cluster center formulas until the following convergence conditions of fuzzy C-means clustering are reached:

[0015]

[0016] Where T is the number of iterations and ε is the error threshold.

[0017] Based on the further improvement of the above method, sample data of resource consumption of the same maintenance operation in the same region is randomly selected, and data analysis is performed on the sample data through normal model parameter estimation, fitting test and prediction interval estimation to obtain the historical reference of resource consumption of maintenance operations of the same type of objects. The method further includes: based on the re-divided multiple clustering sample clusters, a part of the sample data is randomly selected from the resource consumption of the same maintenance operation in the same region to obtain the resource consumption data of the maintenance operation; and the estimated values ​​of the overall expectation μ and variance σ of the resource consumption are calculated based on the resource consumption data of the maintenance operation by using maximum likelihood estimation:

[0018]

[0019] where x1,x2,...,x n is the resource consumption after processing, is the mean of the processed resource consumption; the distribution type of the resource consumption data of the maintenance operation is fitted and tested using the overall mean and variance estimate of the resource consumption to determine whether it obeys the normal distribution, wherein when the distribution type obeys the normal distribution, the test statistic satisfies the following formula:

[0020]

[0021] In this case, all Ω values ​​of x are divided into k mutually disjoint subsets A1, A2, ..., A n , the sample observations x1,x2,...,x n Appears in subset A i The number of i (i=1,2,...,k), then the event

[0022] =A i {x value in A i The frequency of the inner i / n, calculate the probability p of the event i =P(A i ),i=1,2,...,k; the observation frequency f in each category i and the expected frequency np i , the number of parameters r represents the number of parameters estimated in the model, n is the total number of samples, and kr-1 is the number of degrees of freedom; the prediction interval of the historical reference amount of resource consumption is estimated using the normal distribution, where the prediction interval of the historical reference amount of resource consumption is:

[0023]

[0024] Among them, at point On the right, the probability of the area enclosed by the student distribution probability density curve with n-1 degrees of freedom and the X-axis is The probability that the resource consumption of a maintenance operation exceeds the prediction interval is 1-α, and the sample mean is the average value of a set of sample data, and n is the number of samples.

[0025] On the other hand, an embodiment of the present application provides a power grid maintenance resource consumption prediction system, including: a data acquisition module, used to obtain data on resource consumption of multiple maintenance operations and their influencing factors, wherein the data includes historical data and real-time data; a preprocessing module, used to use density-based noise application spatial clustering to identify outliers on the historical data, wherein missing values ​​are interpolated on the historical data; a clustering module, used to cluster the maintenance operation resource consumption of different types of objects of the same maintenance operation in combination with the historical data after outlier processing; a historical reference quantity acquisition module, used to cluster the resource consumption of the same maintenance operation in the same region and the same maintenance operation A sample data is randomly extracted, and the sample data is analyzed through normal model parameter estimation, chi-square fit test and prediction interval estimation to obtain a prediction interval of a historical reference amount of resource consumption; a prediction module is used to construct a neural network prediction model and use the historical data for training, and in combination with the real-time data of the maintenance operation to be predicted, predict the resource consumption prediction target through the trained neural network prediction model; and a weighting module is used to determine the final resource consumption based on the historical reference amount of resource consumption of the maintenance operation and the resource consumption prediction target, wherein the historical reference amount of resource consumption is obtained based on the prediction interval of the historical reference amount of resource consumption.

[0026] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0027] 1. Accurately diagnose data anomalies and improve analysis reliability: This application uses the DBSCAN clustering algorithm to efficiently identify and remove outliers in the data set, reduce noise interference, and ensure that subsequent analysis is based on accurate data. Compared with traditional manual screening, this progress has greatly improved the efficiency and accuracy of anomaly detection and laid a solid foundation for the construction of resource consumption models.

[0028] 2. Innovative data filling strategy to ensure data integrity: To address the problem of missing data, this application innovatively adopts an improved stratified mean interpolation model, which takes into account the hierarchical structure and correlation of the data and more accurately restores the lost information. Compared with simple average or median filling, it can better maintain the authenticity and integrity of the data and improve the robustness of the prediction model.

[0029] (1) Considering the hierarchical structure of data: Traditional methods often ignore the structural characteristics of the data and simply apply a unified filling strategy to all missing values. The improved stratified mean interpolation model first divides the data into different levels or categories based on the intrinsic characteristics of the data (such as time series characteristics, regional differences, equipment types, etc.). This approach is based on the understanding that the similarity between similar data is higher, and the filling of missing values ​​should refer more to the statistical characteristics of similar data, thereby improving the pertinence and accuracy of the filling.

[0030] (2) Utilizing correlation information: On the basis of stratification, the model further considers the correlation between data. For data within a certain layer, in addition to calculating the mean, it is also possible to make weighted adjustments based on the correlation between other variables. For example, if two variables have historically been highly positively correlated, the missing value of one variable can be interpolated based on the current value of the other variable and the adjusted mean through the correlation coefficient. This filling strategy that considers correlation can better reflect the dynamic connection between data and reduce the information loss caused by viewing each missing value in isolation.

[0031] (3) Dynamically adjust the filling strategy: The improved model is not static, but can dynamically adjust the stratification criteria and weight distribution according to the specific situation of the data and the filling effect. For example, as the data set expands or new features are discovered, the model can re-divide the levels or adjust the correlation weights to continuously optimize the filling effect.

[0032] (4) Improve the robustness of the prediction model: Since the model can more accurately restore the true state of the data and avoid data deviation caused by simple filling, it can significantly improve the robustness and prediction accuracy of subsequent prediction models. Especially in machine learning and statistical analysis, high-quality input data is the key to model performance. The improved data filling strategy provides a more reliable foundation for model training.

[0033] 3. Intelligent classification of job types to enhance the pertinence of analysis: Through fuzzy C-means clustering, this application can intelligently classify jobs into different categories based on their multi-dimensional characteristics. This strategy goes beyond the traditional one-size-fits-all approach and achieves accurate determination of job resource consumption. The classification results tailor a resource consumption model for each type of job, improving the pertinence of analysis and the accuracy of prediction.

[0034] 4. Integrate statistics and machine learning to improve prediction accuracy: Combining statistical methods such as normal model parameter estimation and chi-square fit test with an improved neural network prediction model, this application not only ensures the scientific nature of the model establishment, but also captures the nonlinear relationship between complex data through deep learning technology, significantly improving the accuracy and robustness of resource consumption prediction, and providing more reliable data support for decision-making.

[0035] 5. Dynamic weight strategy to adapt to cost fluctuations: This application innovatively introduces a dynamic weight adjustment mechanism to automatically adjust the weights of historical data and forecast results according to market fluctuations and individual operating characteristics, ensuring that the forecast model can flexibly adapt to unpredictable environments and effectively respond to cost fluctuations, providing dynamic and accurate guidance for project cost control, which is much better than the static prediction of fixed models.

[0036] In the present application, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present application will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings are only for the purpose of illustrating specific embodiments and are not to be considered as limiting the present application. The same reference symbols denote the same components throughout the accompanying drawings.

[0038] Figure 1 A flowchart of a method for predicting power grid maintenance resource consumption according to an embodiment of the present application;

[0039] Figure 2 A simplified flow chart of a method for predicting power grid maintenance resource consumption according to an embodiment of the present application;

[0040] Figure 3 is a schematic diagram of a volatility evaluation standard according to an embodiment of the present application; and

[0041] Figure 4 4 is a block diagram of a power grid maintenance resource consumption prediction system according to an embodiment of the present application. DETAILED DESCRIPTION

[0042] The preferred embodiments of the present application are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of the present application and are used together with the embodiments of the present application to illustrate the principles of the present application, and are not used to limit the scope of the present application.

[0043] refer to Figure 1A specific embodiment of the present application discloses a method for predicting power grid maintenance resource consumption, including: in step S101, obtaining data on resource consumption of multiple maintenance operations (for example, consumption of manpower, materials and machines) and its influencing factors (including the geographical location, environment, region and power grid-related data of the maintenance equipment), the data including historical data and real-time data. For example, the data includes operation type, geographical location, environmental information, equipment failure rate, equipment operating years, equipment type, regional electricity consumption, regional GDP, etc., wherein the consumption of manpower, materials and machines includes labor consumption, construction machinery consumption and material consumption; in step S102, density-based noise application spatial clustering is used to identify outliers in historical data, wherein missing value interpolation is performed on historical data; in step S103, clustering is performed on maintenance operation resource consumption of different types of objects of the same maintenance operation in combination with the historical data after outlier processing; in step S104, resource consumption of the same region and the same maintenance operation is clustered. Randomly extract sample data of source consumption, and perform data analysis on the sample data through normal model parameter estimation, chi-square fit test and prediction interval estimation to obtain a prediction interval of the historical reference amount of resource consumption; in step S105, construct a neural network prediction model and use historical data for training, and combine the real-time data of the maintenance operation to be predicted, and predict the resource consumption prediction target through the trained neural network prediction model; and in step S106, determine the final resource consumption based on the historical reference amount of resource consumption of the maintenance operation and the resource consumption prediction target, wherein the historical reference amount of resource consumption is obtained based on the prediction interval of the historical reference amount of resource consumption.

[0044] Compared with the prior art, the power grid maintenance resource consumption prediction method provided in this embodiment uses the advanced DBSCAN clustering algorithm to identify outliers in the data processing link, and can more flexibly handle clusters and noise points of various shapes. The normal model is adopted as the basic framework, and through the Bayesian parameter estimation method, not only the observed data is considered, but also the prior knowledge is incorporated to more comprehensively and probabilistically infer the distribution parameters of the historical consumption of the same type of operation in the same region. Utilizing its powerful nonlinear fitting ability, combined with historical data for training, accurate prediction of resource consumption of specific types of operations can be achieved.

[0045] In the following, reference Figure 1 Each step in the power grid maintenance resource consumption prediction method according to an embodiment of the present application is described in detail.

[0046] In step S101, data on resource consumption (e.g., consumption of manpower, materials, and machines) and influencing factors of multiple maintenance operations are obtained, and the influencing factors include the geographical location, environment, region, and power grid-related data of the maintenance equipment. The data includes historical data and real-time data, for example, the data includes operation type, consumption of manpower, materials, and machines, geographical location, environmental information, equipment failure rate, equipment operating years, equipment type, regional electricity consumption, regional GDP, etc. For example, consumption of manpower, materials, and machines includes labor consumption, construction machinery consumption, and material consumption.

[0047] Specifically, obtaining data on resource consumption of multiple maintenance operations and its influencing factors further includes: using a database to obtain historical data on resource consumption of multiple maintenance operations and its influencing factors, wherein the influencing factors in the historical data include the geographical location, environment, region and power grid-related data of the equipment to be maintained; and using multiple sensors to obtain real-time data on resource consumption of multiple maintenance operations and its influencing factors, wherein the influencing factors in the real-time data include the geographical location, environment, region and power grid-related data of the equipment to be maintained.

[0048] In step S102, density-based noise application spatial clustering is used to perform outlier detection on historical data to perform missing value interpolation processing on the historical data.

[0049] The method of using density-based noise spatial clustering to detect outliers in historical data to interpolate missing values ​​in historical data further includes: a stratified identification step, a mean calculation step within the stratum, a stratified interpolation step, a weighted or complex interpolation step, and an iteration and verification step. Specifically, in the stratified identification step, the historical data is stratified according to the equipment type, the operation type, the geographical location or the environment type to identify outliers, wherein the outliers include missing values. In the mean calculation step within the stratum, the mean of the complete observation records is calculated within each stratum. In the stratified interpolation step, when there are missing value records in the historical data, the mean within the stratum is used to fill the missing values ​​according to the stratum to which the missing values ​​belong. In the weighted or complex interpolation step, the variability within the stratum and the sample size are considered, and different weights are assigned to different strata or weighted average or multiple interpolation is used to reflect the internal structure of the historical data. In the iteration and verification step, the stratification and statistics are re-evaluated based on the interpolated historical data, and iteration is performed again until a predetermined convergence standard or an upper limit of the number of iterations is met.

[0050] Specifically, stratifying historical data according to equipment type, operation type, geographic location or environment type to identify outliers includes: setting a domain radius ε and a minimum number of points MinPts, wherein the domain radius ε represents the proximity relationship between points, and the minimum number of points MinPts represents the minimum number of points required to be reached in the neighborhood of a point; for each point in the historical data, calculating the distance between each point and all other points; dividing each data point in the historical data into core points, boundary points and noise points based on the calculated distances between each point and all other points, wherein when there are at least the minimum number of points MinPts in the ε neighborhood of the current point, marking the current point as a core point; when the current point is directly or indirectly connected to the core point, but there are less than MinPts points in the ε neighborhood of the current point, marking the current point as a boundary point; and when the current point is a data point that is neither a core point nor a boundary point, marking the current point as a noise point as an outlier.

[0051] In step S103, clustering processing is performed on maintenance operation resource consumption of different types of objects of the same maintenance operation in combination with the historical data after outlier processing.

[0052] In combination with the historical data after outlier processing, clustering the maintenance operation resource consumption of different types of objects of the same maintenance operation further includes: summarizing the characteristics of different operation types based on the historical data after outlier processing to obtain multiple feature vector sample sets to be clustered, wherein the feature vector sample sets include the number of equipment failures, equipment operating years, equipment type, labor consumption, construction machinery consumption, material consumption, equipment consumption, temperature, wind speed, terrain, regional GDP, and regional electricity consumption; using multiple feature vector sample sets to construct operation type portraits through fuzzy C-means clustering, so that the multiple feature vector sets are re-divided into multiple cluster sample clusters, wherein the membership degree and cluster center are calculated by the following formula:

[0053]

[0054] Among them, u ij is the i-th sample x i The membership degree of the jth cluster; c j is the cluster center of the jth cluster, with a dimension of n; ||*|| represents the distance between the sample and the cluster center; m is the feature vector sample set X = {x1, x2, ..., x m}, each sample x i ={x i1 ,x i2 ,...,x in} has n eigenvectors; when X is divided into C categories, each subset is represented by C = {c1, c2, ..., ck}; The membership coefficient and cluster center are iterated through the membership and cluster center formulas until the following convergence conditions of fuzzy C-means clustering are reached:

[0055]

[0056] Where T is the number of iterations and ε is the error threshold.

[0057] In step S104, sample data of resource consumption of the same region and the same maintenance operation are randomly selected, and data analysis is performed on the sample data through normal model parameter estimation, chi-square fit test and prediction interval estimation to obtain the prediction interval of the historical reference amount of resource consumption of maintenance operations of the same type of objects.

[0058] Specifically, randomly extracting sample data of resource consumption of the same maintenance operation in the same region, and performing data analysis on the sample data through normal model parameter estimation, fitting test and prediction interval estimation to obtain historical reference of resource consumption of maintenance operations of the same type of objects further includes: based on the re-divided multiple clustering sample clusters, randomly extracting part of the sample data from the resource consumption of the same maintenance operation in the same region to obtain resource consumption data of the maintenance operation; using maximum likelihood estimation to calculate the estimated values ​​of the overall expectation μ and variance σ of resource consumption based on the resource consumption data of the maintenance operation:

[0059]

[0060] where x1,x2,...,x n is the resource consumption after processing, is the mean of the resource consumption after processing; the distribution type of the resource consumption data of the maintenance operation is fitted and tested using the overall mean and variance estimate of the resource consumption to determine whether it obeys the normal distribution. When the distribution type obeys the normal distribution, the test statistic satisfies the following formula:

[0061]

[0062] The above two formulas are the basic definition and distribution description of the chi-square statistic. The basic definition of the chi-square statistic calculates the square difference between the observed frequency and the expected frequency and normalizes it. The left side of the formula is the actual observed frequency f i and the expected frequency np i The right side of the formula is simplified by some algebraic transformations to show the frequency of observations in each category f i and the expected frequency np i The sum of the normalized squared differences between the two subtracted from the total number of samples n.

[0063] The distribution description of the chi-square statistic shows that the chi-square statistic follows a chi-square distribution with kr-1 degrees of freedom. The distribution description of the chi-square statistic is a special case of the basic definition of the chi-square statistic. When the model parameters are estimated, the distribution form of the chi-square statistic. r represents the number of parameters, which represents the number of estimated parameters in the model. In statistical tests, r usually represents the number of estimated parameters in the model. For example, in the normal distribution fit test, two parameters, mean and variance, need to be estimated, r = 2.

[0064] In this case, all Ω values ​​of x are divided into k mutually disjoint subsets A1, A2, ..., A n , the sample observations x1,x2,...,x n Appears in subset A i The number of i (i=1,2,...,k), then event=A i {x value in A i The frequency of the inner Calculate the probability of an event p i =P(A i ),i=1,2,...,k; the observation frequency f in each category i and the expected frequency np i , the number of parameters r, represents the number of parameters estimated in the model, n is the total number of samples, kr-1 is the number of degrees of freedom; the prediction interval of the historical reference amount of resource consumption is estimated using the normal distribution, where the prediction interval of the historical reference amount of resource consumption is:

[0065]

[0066] Among them, at point On the right, the probability of the area enclosed by the student distribution probability density curve with n-1 degrees of freedom and the X-axis is The probability that the resource consumption of a maintenance operation exceeds the prediction interval is 1-α, and the sample mean is the average value of a set of sample data, and n is the number of samples.

[0067] In step S105, a neural network prediction model is constructed and trained using historical data, and combined with real-time data of the maintenance work to be predicted, the resource consumption prediction target is predicted through the trained neural network prediction model.

[0068] Specifically, a neural network prediction model is constructed and trained with historical data, and combined with real-time data of the maintenance work to be predicted, predicting the resource consumption prediction target through the trained neural network prediction model further includes: constructing a neural network prediction model, the neural network prediction model includes an input layer, a hidden layer and an output layer of a BP neural network, wherein a genetic algorithm is used to initialize the population to obtain initial weights and thresholds of the BP neural network; training the neural network prediction model with historical data, and updating the weights and thresholds based on the deviation between the predicted output value and the expected output value during the training of the neural network prediction model; and inputting the real-time data of the maintenance work to be predicted into the trained neural network prediction model to predict the resource consumption prediction target.

[0069] Specifically, initializing the original population using a genetic algorithm to obtain the initial weights and thresholds of the BP neural network further includes: determining the individual fitness value F based on the absolute value of the error between its expected output value and the predicted output value:

[0070]

[0071] Where m is the number of output nodes; p i is the expected output value of the ith node; o i is the expected output value of the ith node; the population is selected based on individual fitness to generate a selected updated population:

[0072]

[0073] Among them, F i is the fitness value of the i-th group, k is the group size, f i is the probability of the i-th population being selected; randomly select 2 parents from the population, generation P k and P l , under a certain probability, the real number crossing mode of the jth point is:

[0074]

[0075] Among them, α is a random number between 0 and 1; randomly select an individual P in the update population i Mutation is performed, and the mutation mode of the jth point is:

[0076]

[0077] Among them, P max , P min are the individual upper and lower bounds respectively; υ is the number of iterations at this time; V maxis the maximum number of evolutions; μ is a random number ranging from 0 to 1; check whether the obtained individual meets the optimal standard, where the fitness check is performed again when the obtained individual cannot meet the standard individual; when the obtained individual meets the standard, the optimal individual is used as the weight and threshold of the BP neural network, and the learning and training times are repeated until the neural network parameters are adjusted to the set accuracy range or the learning times are met.

[0078] In step S106, the final resource consumption is determined based on the historical reference amount of maintenance operation resource consumption and the resource consumption prediction target, and the historical reference amount of resource consumption is obtained based on the prediction interval of the historical reference amount of resource consumption.

[0079] Specifically, determining the final resource consumption through the prediction interval of the historical reference amount of resource consumption of the maintenance operation and the predicted target of resource consumption further includes: calculating the historical reference amount of resource consumption based on the prediction interval of the historical reference amount of resource consumption; using the first weight and the second weight to calculate the weighted average of the historical reference amount of resource consumption and the predicted target of resource consumption as the optimal resource consumption, wherein the first weight and the second weight are respectively assigned to the historical reference amount of resource consumption and the predicted target of resource consumption, and a volatility judgment matrix is ​​constructed according to the change amplitude and change frequency of the same type of costs, so as to determine the first weight and the second weight according to the volatility judgment matrix. The comprehensive evaluation result is that the cost filled in the blank belongs to the type of weak volatility. The weight of the weak volatility type based on the historical results is set to 0.7, and the weight set based on the predicted results is 0.3. The weight of the strong volatility type based on the historical results is set to 0.3, and the weight set based on the predicted results is 0.7.

[0080] refer to Figure 3A specific embodiment of the present application discloses a power grid maintenance resource consumption prediction system, including: a data acquisition module 301, which is used to obtain data on resource consumption (for example, consumption of manpower, materials and machines) and its influencing factors for multiple maintenance operations, and the data includes historical data and real-time data. Factors affecting resource consumption include the geographical location, environment, region, and power grid related data of the maintenance equipment. For example, the data includes operation type, consumption of manpower, materials and machines, geographical location, environmental information, equipment failure rate, equipment operating life, equipment type, regional electricity consumption, regional GDP, etc. For example, the consumption of manpower, materials and machines includes labor consumption, construction machinery consumption and material consumption; the preprocessing module 302 is used to detect outliers on historical data using density-based noise application spatial clustering to interpolate missing values ​​of historical data; the clustering module 303 is used to cluster the maintenance operation resource consumption of different types of objects of the same maintenance operation in combination with the historical data after outlier processing; the historical reference quantity acquisition module 304 randomly extracts sample data of resource consumption of the same region and the same maintenance operation, and performs data analysis on the sample data through normal model parameter estimation, chi-square fit test and prediction interval estimation to obtain the prediction interval of the historical reference quantity of maintenance operation resource consumption of the same type of object; the prediction module 305 is used to construct a neural network prediction model and train it with historical data, and predict the resource consumption prediction target through the trained neural network prediction model in combination with the real-time data of the maintenance operation to be predicted; and the weighting module 306 determines the final resource consumption based on the historical reference quantity of resource consumption and the resource consumption prediction target, wherein the historical reference quantity of resource consumption is obtained based on the prediction interval of the historical reference quantity of resource consumption.

[0081] Specifically, the preprocessing module 302 includes: a stratified identification submodule, an intra-stratified mean calculation submodule, a stratified interpolation submodule, a weighted or complex interpolation submodule, and an iteration and verification submodule, wherein the stratified identification submodule is used to stratify historical data according to equipment type, operation type, geographical location or environment type to identify abnormal values, wherein abnormal values ​​include missing values; the intra-stratified mean calculation submodule is used to calculate the mean of complete observation records within each layer; the stratified interpolation submodule is used to fill the missing values ​​with the mean within the layer according to the layer to which the missing values ​​belong when there are missing value records in the historical data; the weighted or complex interpolation submodule is used to consider the variability within the layer and the sample size, assign different weights to different layers, or use weighted average or multiple interpolation to reflect the internal structure of the historical data; and the iteration and verification submodule is used to re-evaluate the stratification and statistics based on the interpolated historical data, and iterate again until the predetermined convergence standard or the upper limit of the number of iterations is met.

[0082] In the following, reference Figure 2 and Figure 3 , the power grid maintenance resource consumption prediction method according to the embodiment of the present application is described in detail in the form of a specific example.

[0083] refer to Figure 2 In step S201, the resource consumption of multiple maintenance operations and the geographical location, environment, region, power grid and other related data of the corresponding equipment are collected; outlier detection and processing are performed based on the data, and the DBSCAN clustering algorithm is used to carry out the outlier detection method, and the missing values ​​are processed according to the improved stratified mean interpolation model.

[0084] 1. The data includes operation type, human, material and machine consumption, geographic location, environmental information, equipment failure rate, equipment operating years, equipment type, regional electricity consumption, regional GDP, etc.

[0085] 2. In the outlier detection method, the density-based clustering algorithm improves the defect of the distance-based outlier algorithm that gives the same weight to all objects. It is suitable for the situation where the sample data is unevenly distributed. It can make good outlier judgments for the situation where there are different densities in the data, reducing the limitations of use and performing probability evaluation on the judgment of whether it is an outlier. In addition, clusters in reality often have various shapes. Traditional clustering methods such as K-means clustering can only process spherical clusters, which obviously cannot meet the needs of clusters of other irregular shapes. Therefore, a DBSCAN clustering algorithm that meets the needs of arbitrary cluster shapes is proposed.

[0086] DBSCAN (Density-Based Spatial Clustering of Applications with Noise, DBSCAN) is a density-based outlier detection method. It belongs to unsupervised learning, is resistant to noise, can process clusters of any shape and size, and does not require the prior division of the number of clusters. Therefore, for the collected multi-source data, the internal structure of the data is unknown, and the data type that cannot be predicted before cleaning, this method has strong operability for such data.

[0087] Abnormal data identification principle:

[0088] (1) Define density and neighborhood: First, set two key parameters, namely the neighborhood radius (ε) and the minimum number of points (MinPts). ε defines the proximity relationship between points, while MinPts defines the minimum number of points required in the neighborhood of a point to determine the density status of the point.

[0089] (2) Classified points: DBSCAN classifies each data point into one of three types by examining its neighborhood: Core points, which have at least MinPts points (including the point itself) in their ε neighborhood. Boundary points, which are not in the ε neighborhood of any core point but can be connected to a core point through a series of core points. Noise points (outliers): are neither core points nor part of the neighborhood of any core point.

[0090] (3) Clustering and anomaly detection: The algorithm starts from any unvisited point. If the point is a core point, a cluster is expanded with it as the center, including all points in its neighborhood, and the neighborhood points of these points are recursively added until no new points can be added. In the process, points that are not included in any cluster are considered outliers.

[0091] Determine the parameters: Neighborhood radius ε (epsilon), which determines the maximum distance between points that are considered "neighboring". Minimum number of points MinPts, the minimum number of points (including the point itself) that must be included in the ε neighborhood of a point to become a core point.

[0092] Calculate distance: For each point in the data set, calculate the distance between it and all other points. Usually use Euclidean distance or other distance metric.

[0093] Identify core points and border points: If a point has at least MinPts points in its ε neighborhood, the point is marked as a core point. Border points are those points that are directly or indirectly connected to core points, but have less than MinPts points in their own ε neighborhood.

[0094] Marking noise points: Any data point that is neither a core point nor a boundary point is considered a noise point, i.e., an outlier.

[0095] Output results: Normal data points will be classified into one or more clusters (as core points or boundary points), while noise points will be identified separately and regarded as outliers.

[0096] 3. The stratified mean interpolation method is a further refinement of the single interpolation method. It selects appropriate attributes according to specific data to divide the population. If missing data appears in a layer, the missing data in that layer is used to fill it. The improved stratified mean interpolation model is a more advanced missing value processing technology developed on the basis of the traditional mean interpolation method. This model mainly targets the hierarchical structure in the data set (such as stratification by time, geographic location, category and other characteristics), and uses these hierarchical information to guide the missing value filling process to improve the accuracy and effectiveness of interpolation. The main steps are as follows:

[0097] Layer identification: First, identify the natural hierarchical structure in the data set, such as dividing the data into different layers or subsets based on characteristics such as equipment type, job type, region, etc. This is done because data points within the same layer may have stronger similarities and correlations.

[0098] Calculate the mean within the layer: Within each layer, calculate the mean (or median, mode, etc., depending on the specific situation) of the records containing complete observations. This statistic reflects the central tendency of the data in this layer and is used as a benchmark for interpolation.

[0099] Stratified interpolation: For each record with missing values, first determine the level to which it belongs, and then use the mean (or other statistics) calculated within the level to fill in the missing values. Compared with global mean interpolation, this method takes into account the structural differences in the data and can provide more reasonable estimates. The stratified interpolation strategy is suitable for missing value processing scenarios of various complex data sets because it can take into account the structural differences within the data, especially for those situations where there are obvious grouping characteristics or hierarchical structures between the data. Combined with the information provided previously, we can explain in detail how to apply the stratified interpolation method in specific scenarios:

[0100] Scope of application: Multi-dimensional data analysis. When a data set contains multiple features and these features can define different groups (such as equipment maintenance data in different regions, maintenance records of different types of mechanical equipment), stratified interpolation can more accurately reflect the characteristics within each group. Time series data. For data that changes over time, stratification by time window can retain the continuity and seasonal characteristics of the time series. Heterogeneous data sets. The data set contains samples from different sources and under different conditions. Stratified interpolation can reduce the deviation caused by heterogeneity. Analysis of the correlation between resource consumption and geographical environment. For example, equipment maintenance resource consumption data is closely related to geographical location and environmental conditions, and can be stratified according to geographical location, environmental type, etc.

[0101] Assume that you are processing resource consumption records for a series of equipment maintenance operations. The data includes but is not limited to operation time, quantity of materials consumed, labor cost, equipment type, operation location (geographic location), environmental factors (such as climate conditions), etc. Some records may have missing values ​​for certain fields, such as the labor cost of a specific operation is not recorded.

[0102] The stratification steps include: a. Determine the stratification standard: First, determine a reasonable stratification standard based on the characteristics of the data. For example, stratification can be performed according to equipment type (such as power equipment, mechanical devices), geographical location (province, city level), and environmental type (such as coastal areas, inland areas). b. Calculate intra-level statistics: In each layer, calculate the mean (or median, mode, etc., depending on the data distribution characteristics) for non-missing data. For example, calculate the average labor cost of all records of the same equipment type in coastal areas. c. Perform interpolation: For each record with missing values, fill in the missing values ​​with the statistics calculated in the layer according to the layer to which it belongs. For example, if the labor cost in the maintenance record of power equipment in a coastal area is missing, it is filled with the average labor cost of power equipment maintenance in the area.

[0103] Compared with global mean interpolation, stratified interpolation takes into account the huge differences that may exist between different levels, avoiding the systematic errors introduced by treating all data as having the same distribution. Improved accuracy: When the data has obvious stratified characteristics, stratified interpolation can better reflect the actual situation, improve the accuracy of the fill value and the robustness of the model prediction. Enhanced practicality: For resource consumption analysis that is sensitive to geographical location and environmental factors, stratified interpolation can better preserve the differences in the impact of these factors on resource consumption, providing a more practical reference for decision-making.

[0104] Consider weights or complex imputation strategies: In some cases, the improved model will further consider the variability within the level, sample size or other relevant factors, assign different weights to different levels or records, or adopt more complex imputation strategies, such as weighted averaging, multiple imputation, etc., to better reflect the inherent structure of the data.

[0105] Iteration and validation: To improve the imputation results, the process is sometimes performed iteratively, i.e., a preliminary imputation is performed first, then the stratification and statistics are reassessed based on the imputed data, and if necessary, the imputation is performed again until the predetermined convergence criteria or the upper limit of the number of iterations is met.

[0106] Let n' be the number of non-missing samples in the population, the total number of samples is n, and use Represents the sample population mean estimate, and the calculation formula is as follows

[0107]

[0108] Among them, the resource consumption data y of the i-th sample i Indicates the resource consumption value in a single sample. In actual calculation, this is the resource consumption of each sample point in the data set. i is an indicator variable, when a i When it is 0, it means data is missing. i1 means that the data is not missing. is the mean of the non-missing samples, the estimated value of the overall mean is the interpolation value of the missing data, and the variance is assumed to be The variance after data interpolation is s 2 , the variance changes before and after interpolation are as follows:

[0109]

[0110] The above display This shows that the mean interpolation method can reduce sample variance and improve overall data quality, but there will be a situation where all missing data are interpolated by the same data, resulting in a distorted sample distribution. The stratified mean interpolation method is a further refinement of the single interpolation method. The population is divided according to the specific data, and if there is missing data in the layer, the missing data in the layer is used to fill in the missing data. This method is better than the above method, but it still cannot avoid the problem of distorted sample distribution.

[0111] Assume the total number of samples is n, H is the total number of layers, n h is the total number of samples in the hth layer, and the mean of the non-missing data in the hth layer is Estimation of the population mean The calculation formula is as follows:

[0112]

[0113] After imputation, the change in the overall variance is as follows, where n 1h is the number of non-missing data in the h-th layer:

[0114]

[0115] In step S202, different information of the same operation is classified and processed in combination with data collection and processing. Statistical analysis is performed on the operation resource consumption of different types of objects respectively, and the historical reference amount of operation resource consumption of the same type of objects is obtained through data analysis.

[0116] (1) Design of the portrait label system: For each job instance, a "job portrait" is formed based on its feature value, that is, a collection of feature vectors that reflect the typical characteristics and requirements of the job. A job instance is a maintenance job, such as transformer maintenance. The feature value is the data information indicator in Table 1 below. The job portrait is a collection of feature vectors, each of which represents an indicator of the job.

[0117] Table 1

[0118]

[0119] (2) Portrait construction based on fuzzy C-means clustering: Fuzzy C-means (FCM) is a clustering algorithm based on objective function partitioning. It can assign a weight to each object and each cluster to indicate the degree to which the object belongs to the cluster. It can avoid hard clustering phenomenon and provide more flexible clustering results.

[0120] The main principle of FCM is to maximize the similarity between objects classified into the same cluster and minimize the similarity between objects in different clusters. By calculating the membership degree of each data sample to be clustered with each cluster center, it is divided into clusters with obvious separation of characteristics to achieve data sample clustering. A weight is assigned to each sample and each cluster, and the degree to which the sample belongs to the cluster is judged by the cluster distance. Through multiple iterations, the distance within each cluster sample is minimized and the distance between clusters is maximized.

[0121] Suppose the set of m samples to be clustered is X = {x1, x2, ..., x m}, where each sample x i There are n characteristic indicators, let it be x i ={x i1 ,x i2 ,...,x in If we want to divide X into C categories, each subset can be represented as C = {c1, c2, ..., c k}Then the clustering loss function is as follows:

[0122]

[0123] Where: J is the value function of FCM; b is the weighting coefficient, which is used to adjust the influence of the membership index in the optimization target, and its value is generally 2; u ij is the membership degree of the i-th sample to the j-th cluster; x i is the i-th sample, with dimension n; c j is the cluster center of the jth cluster, with a dimension of n; ||*|| represents the distance between the sample and the cluster center. To solve this optimization problem, the Lagrange multiplier method is used to introduce the equality constraint into the objective function, and the conditional extreme value is converted into the unconditional extreme value, resulting in the following formula:

[0124]

[0125] Where: i The Lagrange multiplier of the ith sample. j and membership Find partial derivatives:

[0126]

[0127] The fuzzy C-means clustering algorithm calculates the membership degree and cluster center through continuous iteration until it reaches the optimal value. The expressions of membership degree and cluster center can be obtained from the following formulas, as shown in the following formulas:

[0128]

[0129] According to the above two formulas, the membership coefficient and cluster center are iterated until the convergence condition of FCM is reached. The convergence condition is:

[0130]

[0131] Where: T is the number of iterations; ε is the error threshold.

[0132] In step S203, the historical reference amount of operation resource consumption of objects of the same type is obtained through data analysis using methods such as normal model parameter estimation, chi-square fit test, and prediction interval estimation.

[0133] (1) Data preprocessing: Random sampling of resource consumption for maintenance operations of the same type and in the same region. First, sample data of maintenance operations of the same type and in the same region are taken, and the original data of resource consumption of the random sampling is converted to the same time point.

[0134]

[0135] Among them, F is the consumption, and the resource consumption after processing is recorded as x1,x2,...,x n For example: personnel consumption days and machine consumption days. Since the model uses samples to study the population, and the population may have any situation, continuity correction can be used. The coverage interval of the horizontal axis of the frequency distribution histogram is R; secondly, sort the resource consumption by size and determine the appropriate group interval, and count the sample frequency f in each interval. i , calculate the frequency f i / n; Finally, a sample distribution frequency histogram is plotted with the group interval as the horizontal axis and the frequency as the vertical axis. The observed sample data approximately comes from a certain distribution population.

[0136] (2) Maximum likelihood estimation: Random sampling is used to collect observations of resource consumption samples. Since resource consumption can be regarded as a statistic similar to the sample mean, according to the central limit theorem, the sample distribution frequency histogram drawn is most likely similar to the normal distribution. The model assumes that the random samples come from a normal population with parameters μ and σ, that is, the resource consumption population obeys the normal distribution, and the distribution function is denoted as F(x, μ, σ). According to the principle of maximum likelihood estimation, the probability density function of X is as follows:

[0137]

[0138] The likelihood function is shown as follows:

[0139]

[0140] Taking the logarithm of it gives the following formula:

[0141]

[0142] Taking partial derivatives, we get the following formula:

[0143]

[0144] The following is an estimate of the overall expectation and variance of resource consumption:

[0145]

[0146] (3) Chi-square fit test: The assumption that the population belongs to a distribution type is based on the observed random sample frequency distribution diagram, which needs further testing. The estimated values ​​calculated above are used as parameter values ​​to perform a fit test on the distribution type to test whether it obeys the normal distribution. Assume that x obeys the normal distribution, and the distribution function is denoted by F(x, μ, σ). Divide all Ω values ​​of x into k mutually exclusive subsets A1, A2, ..., A n , the sample observations x1, x2, ..., x n Appears in A i The number of i (i=1,2,...,k), then event=A i {x value in A i The frequency of the inner i / n. Calculate the probability of the event and get p i =P(A i ),i=1,2,...,k. Usually there is a difference between frequency and probability. When the number of trials is large enough, this difference will decrease and become negligible, that is, Should not be too large, and the following statistic is used to measure the degree of fit between the sample and the hypothesized distribution, where C i (Multiply each term by the appropriate C i , so that the statistic has an ideal limiting distribution) is a constant.

[0147]

[0148] According to Pearson's theorem, The following formula is used as the test statistic:

[0149]

[0150] When X follows a normal distribution, the above equation holds.

[0151] At this time x2 Should not be too large, the rejection region is x 2 ≥G=x 2 (kr-1). Finally, check whether the following equation is true.

[0152] x 2 ≥x a 2 (kr-1);

[0153] a is the significance level. When the inequality holds, it is rejected that the original distribution follows the normal distribution. Otherwise, it is “not rejected” that the original distribution follows the normal distribution, that is, the random sample comes from a normal population.

[0154] (4) Prediction interval estimation: After testing the normality of random samples, prediction interval estimation can be performed. In practical applications, the overall expectation and variance of resource consumption are unknown. The estimated values ​​of the overall expectation and variance are both statistics, not constants. If the estimated values ​​are directly imported into the statistics as parameter values ​​for prediction interval estimation, it will inevitably affect the accuracy of the calculation results and cause systematic errors. When n is large, the student distribution is similar to the normal distribution. The statistical measurement model uses the student distribution instead of the normal distribution, and uses s instead of σ for calculations. A new statistic is constructed based on the prediction interval estimation theory, as shown in the following formula:

[0155]

[0156] T follows the student distribution with n-1 degrees of freedom. The probability of using the T statistic is as follows:

[0157] p(-t a / 2 <T<t a / 2 ) = 1 - α;

[0158] Placing x0 at the center of the probability statement, the probability of the event described by the following equation occurring is 1-α.

[0159]

[0160] The 100(1-α)% prediction interval for future observations x0 is:

[0161]

[0162] Among them, To the right of the point, the area (probability) enclosed by the student distribution probability density curve with degrees of freedom v = n-1 and the X-axis is The probability that the resource consumption of a job exceeds this interval is 1-α.

[0163] In step S204, a resource consumption prediction model is constructed based on historical data, and the resource consumption prediction target is determined in combination with the job-related data to be measured. The rationality of resource consumption is accurately measured by predicting the reference target and the historical reference results, thereby providing a reference for project review and cost determination.

[0164] 1. Determination based on weighted average historical value: Based on the average value calculation, the historical consumption results, for example, the interval is a closed interval [a, b], the average value m can be simply calculated as:

[0165]

[0166] This value is exactly in the middle of the interval and reflects the "average level" of the values ​​at both ends of the interval.

[0167] 2. Determination of predicted value based on improved neural network: The structure of BP neural network includes input layer, hidden layer and output layer. Input parameters enter the input layer, then pass through the hidden layer, and finally output by the output layer. The steps of using K-fold-GA-BP neural network method to calculate resource consumption are as follows.

[0168] (1) Design of input and output layers. According to the design characteristics of BP neural network, the mutual influence relationship between various factors is not considered. The consumption influencing factors are taken as the input layer, and the output layer is the resource consumption.

[0169] (2) Selection of target value: Based on the resource consumption of implemented projects, the target resource consumption is estimated by processing data in horizontal and vertical directions as the output target value.

[0170] (3) Model establishment: including genetic algorithm part and neural network part.

[0171] Genetic algorithm part: First, use GA to initialize the population to obtain the initial weights and thresholds of the BP neural network, and then substitute the individuals in the initialized population into the BP neural network to iterate and find the optimal initial weights and thresholds.

[0172] ① Encode the population of real number strings consisting of the weights between the input layer, hidden layer and output layer, and the thresholds of the hidden layer and the output layer.

[0173] ②The fitness value F of an individual is composed of the absolute value of the error between its expected output value and the predicted output value:

[0174]

[0175] Where m is the number of output nodes; p i is the expected output value of the ith node; o i is the expected output value of the ith node.

[0176] ③ According to the individual fitness results calculated by the above formula, select better individuals to form a new population. The calculation method is:

[0177]

[0178] Among them, F i is the fitness value of the i-th group, k is the group size, f i is the probability of group i being selected.

[0179] ④ Randomly select 2 fathers from the population k and P l , under a certain probability, the real number crossing mode of the jth point is:

[0180]

[0181] Among them, α is a random number between 0 and 1.

[0182] ⑤ Through probability calculation, randomly select an individual P from the new population i Mutation is performed, and the mutation mode of the jth point is:

[0183]

[0184] Where P max , P min are the upper and lower bounds of the individual respectively; υ is the number of iterations at this time; V max is the maximum number of evolutions; μ is a random number ranging from 0 to 1.

[0185] ⑥ Check whether the obtained individuals meet the optimal standards, and re-check the fitness of individuals that cannot meet the standards; if they meet the standards, the optimal individuals are used as the weights and thresholds of the BP neural network, repeat the learning and training times, and adjust the neural network parameters to the set accuracy range or until the learning times are met.

[0186] Neural network part: After obtaining the initial weights and thresholds of the BP neural network, the steps for analyzing its internal structure model are as follows.

[0187] ①Initialize the network.

[0188]

[0189] Among them, X is the neuron input vector, is the factor affecting resource consumption; n is the number of input layer nodes; Y is the neuron output vector, is the resource consumption, and m is the number of output layer nodes.

[0190] ②Calculate the hidden layer output. Hidden layer output value H a The calculation method is:

[0191]

[0192] Where p is the number of hidden layer nodes; ω ij is the weight between the input layer and the hidden layer; a is the threshold; f is the activation function.

[0193] ③Calculate the output of the output layer. Output layer output value H b The calculation method is:

[0194]

[0195] Among them, ω jk is the weight between the hidden layer and the output layer; b is the threshold value.

[0196] ④ Calculate the error and verify. The error e is expressed as the predicted output H b The difference from the expected output Y is calculated as:

[0197] e k =|Y k -H bk |;

[0198] Let e0 be the maximum allowable error value. If max(e m )>e0, the BP neural network parameters are adjusted and the learning and training are repeated until all the training data meet the specified error standard.

[0199] ⑤Update weights and thresholds. Update ω by the required error ij and ω jk :

[0200] ω jk =ω jk +ηH aj e k ;

[0201] Where η is the learning rate, which is between 0 and 1. The threshold correction is:

[0202] λ b =λ b +e k ;

[0203] (4) Model calculation. The model calculation was performed using MATLAB software. During the calculation process, the existing data on the maintenance and operation costs of substation equipment were used as samples and trained using the GA-BP neural network. To avoid overfitting and improve the accuracy of the GA-BP neural network model, the K-fold cross-validation method was used on the training data to determine the accuracy of the model.

[0204] (5) Prediction effect evaluation indicators. When testing the results of the model, it is necessary to consider the reliability and validity of the model. Reliability refers to the reliability of the model, which is manifested in the reproducibility and stability of the model results; validity refers to the effectiveness of the model, which is manifested in the degree of consistency between the model results and the content to be examined.

[0205] Mainly adopt e MSE Evaluate the impact of K-fold cross validation on model accuracy. The calculation formula is:

[0206]

[0207] The mean absolute percentage error and root mean square error (RMSE) are used to measure the validity of the network, that is, the prediction effect. The calculation formula is:

[0208]

[0209] Among them, D t is the actual output of the model, Y t is the expected output. In order to test the reliability of the model, this paper takes the e after multiple runs. MAPE 、e RMSE The standard deviation is used as the final test indicator.

[0210] 3. Final consumption measurement based on historical results and forecast results.

[0211] By assigning different weights to historical results and forecast results, and then calculating the weighted average. The choice of weights can be determined based on factors such as the credibility of historical data and the accuracy of the forecast model. For example, if the forecast model is proven to be very accurate, a higher weight can be given to the forecast results; conversely, if the historical data is more representative, the weight of the historical results can be increased. The formula is as follows:

[0212] Optimal result = α × historical result + (1-α) × predicted result;

[0213] Here, α is a weight coefficient between 0 and 1.

[0214] Weights are assigned based on the volatility characteristics of costs of the same type.

[0215] (1) Calculation of change range: The change range is determined by the average value of the change rate of each data and the data average value. The formula is: where w n is the variation range, and n is the number of variables. x e for x1,x2,...x n The average value of .

[0216] (2) Change frequency calculation: Data has a virtual falsehood (data wave-particle duality). When the data layout is relatively sparse (for example, the data near the average fluctuation has a relatively large fluctuation amplitude), the difference between the average variance of the data and the large variance is larger. When the data layout is relatively concentrated, the square of the difference between the private data and the smaller average data. Therefore, the larger the variance, the larger the data amplitude; the smaller the variance, the smaller the data amplitude. Therefore, by using the variance method, combined with settlement data, the variance of the actual cost of different types of operations can be statistically analyzed to obtain the change frequency of each operation cost.

[0217] The average of the sum of the squares of the differences between each data in the sample and the sample mean is called the sample variance; the arithmetic square root of the sample variance is called the sample standard deviation. Both the sample variance and the sample standard deviation are quantities that measure the size of a sample's fluctuation. The larger the sample variance or the sample standard deviation, the greater the fluctuation of the sample data. In statistical descriptions, variance is used to calculate the difference between each variable (observation) and the overall mean. In order to avoid the sum of the deviations from the mean being zero and the sum of the squares of the deviations from the mean being affected by the sample content, statistics uses the average sum of the squares of the deviations from the mean to describe the degree of variation of the variable. The formula for calculating the overall variance is:

[0218]

[0219] σ 2 is the population variance, X is the variable, μ is the population mean, and N is the number of columns in the population.

[0220] When the data distribution is relatively dispersed (that is, the data fluctuates greatly around the mean), the sum of the squares of the differences between each data and the mean is large, and the variance is large; when the data distribution is relatively concentrated, the sum of the squares of the differences between each data and the mean is small. Therefore, the larger the variance, the greater the fluctuation of the data; the smaller the variance, the smaller the fluctuation of the data.

[0221] (3) Construction of volatility judgment matrix: The judgment matrix refers to a two-dimensional judgment matrix consisting of the frequency and amplitude of standard cost changes. Specific judgment criteria are as follows: Figure 3 As shown, the part filled with slashes is disordered fluctuation with strong volatility, and the part filled with blanks is orderly fluctuation with weak volatility. The evaluation criteria for the amplitude of change are: 0-10% is general, 10%-15% is medium, 15%-29% is high, and above 20% is high. The evaluation criteria for the frequency of change are: 0-1 is general, 1-3 is medium, 3-6 is high, and above 6 is high.

[0222] The cost type with a comprehensive evaluation result of a slash filling belongs to the strong volatility type, and the cost volatility of this type of business is relatively large. The cost type with a comprehensive evaluation result of a blank filling belongs to the weak volatility type. The weight of the weak volatility type based on historical results is set to 0.7, and the weight set based on forecast results is set to 0.3. The weight of the strong volatility type based on historical results is set to 0.3, and the weight set based on forecast results is set to 0.7.

[0223] Compared with the prior art, the improvements of the embodiments of the present application include: 1. Advanced outlier detection and intelligent interpolation processing. In the data processing stage, the advanced DBSCAN clustering algorithm is used to identify outliers, which can handle clusters and noise points of various shapes more flexibly than traditional methods. For missing values, an improved stratified mean interpolation model is innovatively introduced. The model is stratified according to the natural classification of the data (such as equipment type, geographical location, etc.) to ensure that the interpolation value is closer to the actual situation and enhance the integrity and accuracy of the data. 2. Construction of job type portraits and historical reference quantities. Specific details: By analyzing the processed data and combining the portrait concept, different job types are characterized and divided into types to form job type portraits. This process not only deepens the understanding of the job mode, but also provides a clear framework for the historical consumption statistics of similar jobs, which facilitates the extraction of more valuable historical reference quantities. 3. Determination of historical reference indicators based on Bayesian theory. First, we use the detailed data collected, adopt the normal model as the basic framework, and use the Bayesian parameter estimation method to not only consider the observed data, but also incorporate prior knowledge to more comprehensively and probabilistically infer the distribution parameters of historical consumption in the same region and similar operations. In order to ensure the applicability and accuracy of the model, a chi-square fit test is implemented to rigorously verify whether the data conforms to the normal distribution assumption to ensure the scientific nature of the analysis. 4. Determination of prediction reference indicators based on neural networks. For the clustered job types, an independent neural network prediction model is constructed, using its powerful nonlinear fitting capabilities and combined with historical data for training to achieve accurate prediction of resource consumption for specific job types.

[0224] Improve forecasting accuracy and decision support quality: Through the above improvements, resource consumption forecasts are more accurate, providing a scientific basis for project review and cost budgeting, and reducing the risk of budget overruns. Enhance the robustness and practicality of data processing: Advanced data cleaning technology (outlier detection and intelligent interpolation) improves the reliability and effectiveness of data analysis, and can provide more accurate analysis results even when the data is incomplete or noisy. Promote optimal resource allocation: Job type profiling and refined prediction models help identify the patterns of resource consumption and potential optimization space, promote a more reasonable allocation of resources among different jobs, and improve overall operational efficiency. Improve the ability to cope with uncertainty: The application of Bayesian methods enhances the model's adaptability to future changes, especially when external conditions such as the environment and equipment conditions change, it can still provide reliable prediction intervals and decision support.

[0225] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0226] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed in the present application should be covered within the protection scope of the present application.

Claims

1. A method for predicting power grid maintenance resource consumption, characterized in that: include: Obtaining data on resource consumption and influencing factors of multiple maintenance operations, wherein the data includes historical data and real-time data; Using density-based noise application spatial clustering to identify outliers on the historical data, wherein the historical data is interpolated for missing values; Combined with the historical data after outlier processing, cluster processing is performed on the maintenance operation resource consumption of different types of objects in the same maintenance operation; Randomly extract sample data of resource consumption in the same region and the same maintenance operation, and perform data analysis on the sample data through normal model parameter estimation, chi-square fit test and prediction interval estimation to obtain the prediction interval of the historical reference amount of resource consumption; Constructing a neural network prediction model and using the historical data for training, and combining the real-time data of the maintenance operation to be predicted, predicting the resource consumption prediction target through the trained neural network prediction model; and Based on the resource consumption history reference amount and the resource consumption prediction target, a final resource consumption amount is determined, wherein the resource consumption history reference amount is obtained based on a prediction interval of the resource consumption history reference amount.

2. The method for predicting power grid maintenance resource consumption according to claim 1, characterized in that: The data of resource consumption and its influencing factors for multiple maintenance operations further include: Using a database to obtain historical data of resource consumption and influencing factors of multiple maintenance operations, wherein the influencing factors in the historical data include the geographical location, environment, region and power grid-related data of the equipment being maintained; and A variety of sensors are used to obtain real-time data on resource consumption and influencing factors of multiple maintenance operations, wherein the influencing factors in the real-time data include the geographical location, environment, region and power grid-related data of the equipment to be maintained.

3. The method for predicting power grid maintenance resource consumption according to claim 2, characterized in that: Using density-based noise application spatial clustering to perform outlier detection on the historical data to perform missing value interpolation processing on the historical data further includes: In the stratified identification step, the historical data is stratified according to equipment type, operation type, geographic location or environment type to identify abnormal values, wherein the abnormal values ​​include missing values; In the intra-stratum mean calculation step, within each stratum, the mean of the records containing complete observations is calculated; In the stratified interpolation step, when there are missing value records in the historical data, the missing values ​​are filled using the mean value within the stratum according to the stratum to which the missing values ​​belong; In the weighted or complex imputation steps, consider the variability within the stratum, the sample size, assign different weights to different strata or use weighted average or multiple imputation to reflect the inherent structure of the historical data; and In the iteration and validation step, the stratification and statistics are re-evaluated based on the imputed historical data, and iterations are performed again until the predetermined convergence criteria or the upper limit of the number of iterations is met.

4. The method for predicting power grid maintenance resource consumption according to claim 3, characterized in that: Stratifying the historical data by equipment type, job type, geographic location, or environment type to identify outliers includes: Set a domain radius ε and a minimum number of points MinPts, wherein the domain radius ε represents the proximity relationship between points, and the minimum number of points MinPts represents the minimum number of points that need to be reached in the neighborhood of a certain point; For each point in the historical data, calculate the distance between each point and all other points; According to the calculated distance between each point and all other points, each data point in the historical data is divided into core points, boundary points and noise points, wherein when there are at least a minimum number of points MinPts in the ε neighborhood of the current point, the current point is marked as a core point; when the current point is directly or indirectly connected to the core point, but there are less than MinPts points in the ε neighborhood of the current point, the current point is marked as a boundary point; and when the current point is neither a core point nor a boundary point, the current point is marked as a noise point as an outlier.

5. The method for predicting power grid maintenance resource consumption according to claim 3, characterized in that: Combining the historical data after outlier processing, clustering the maintenance operation resource consumption of different types of objects of the same maintenance operation further includes: Based on the historical data after the outlier processing, characteristics of different operation types are summarized to obtain multiple feature vector sample sets to be clustered, wherein the feature vector sample sets include the number of equipment failures, equipment operating years, equipment type, labor consumption, construction machinery consumption, material consumption, equipment consumption, temperature, wind speed, terrain, regional GDP, and regional electricity consumption; The job type portrait is constructed by using the multiple feature vector sample sets through fuzzy C-means clustering, so that the multiple feature vector sets are re-divided into multiple cluster sample clusters, wherein the membership degree and the cluster center are calculated by the following formula: Among them, u ij is the i-th sample x i The membership degree of the jth cluster; c j is the cluster center of the jth cluster, with a dimension of n; ||*|| represents the distance between the sample and the cluster center; m is the feature vector sample set X = {x1, x2, ..., x m }, each sample x i ={x i1 ,x i2 ,...,x in } has n eigenvectors; when X is divided into C categories, each subset is represented by C = {c1, c2, ..., c k }; The membership coefficient and cluster center are iterated through the membership and cluster center formulas until the following convergence conditions of fuzzy C-means clustering are reached: Where T is the number of iterations and ε is the error threshold.

6. The method for predicting power grid maintenance resource consumption according to claim 5, characterized in that: Randomly extracting sample data of resource consumption of the same maintenance operation in the same region, and performing data analysis on the sample data through normal model parameter estimation, fitting test and prediction interval estimation to obtain the prediction interval of the historical reference amount of maintenance operation resource consumption of the same type of object further includes: Based on the re-divided multiple cluster sample clusters, some sample data are randomly selected from the resource consumption of the same region and the same maintenance operation to obtain the resource consumption data of the maintenance operation; The maximum likelihood estimation is used to calculate the estimated values ​​of the overall expectation μ and variance σ of resource consumption based on the resource consumption data of maintenance operations: where x1,x2,...,x n is the resource consumption after processing, is the mean of resource consumption after processing; The distribution type of the resource consumption data of the maintenance operation is fitted by using the overall mean and variance estimates of the resource consumption to determine whether it obeys the normal distribution. When the distribution type obeys the normal distribution, the test statistic satisfies the following formula: In this case, all Ω values ​​of x are divided into k mutually disjoint subsets A1, A2, ..., A n , the sample observations x1,x2,...,x n Appears in subset A i The number of i (i=1,2,...,k), then event=A i {x value in A i The frequency of the inner Calculate the probability of an event p i =P(A i ),i=1,2,...,k; the observation frequency f in each category i and the expected frequency np i , the number of parameters r, represents the number of estimated parameters in the model, n is the total number of samples, and kr-1 is the number of degrees of freedom; The prediction interval of the historical reference amount of resource consumption is estimated using normal distribution, wherein the prediction interval of the historical reference amount of resource consumption is: Among them, at point On the right, the probability of the area enclosed by the student distribution probability density curve with n-1 degrees of freedom and the X-axis is The probability that the resource consumption of a maintenance operation exceeds the prediction interval is 1-α, and the sample mean is the average value of a set of sample data, and n is the number of samples.

7. The method for predicting power grid maintenance resource consumption according to claim 6, characterized in that: Constructing a neural network prediction model and using the historical data for training, and combining the real-time data of the maintenance operation to be predicted, the resource consumption prediction target predicted by the trained neural network prediction model further includes: Constructing the neural network prediction model, the neural network prediction model includes an input layer, a hidden layer and an output layer of a BP neural network, wherein a population is initialized using a genetic algorithm to obtain an initial weight and threshold of the BP neural network; Training the neural network prediction model using the historical data, and updating weights and thresholds based on deviations between predicted output values ​​and expected output values ​​during the training of the neural network prediction model; and Input the real-time data of the maintenance work to be predicted into the trained neural network prediction model to predict the resource consumption prediction target; The method of initializing the original population using a genetic algorithm to obtain the initial weights and thresholds of the BP neural network further includes: determining the individual fitness value F based on the absolute value of the error between the expected output value and the predicted output value: Where m is the number of output nodes; p i is the expected output value of the ith node; o i is the expected output value of the i-th node; The population is selected based on the individual fitness to generate a selected updated population: Among them, F i is the fitness value of the i-th group, k is the group size, f i is the probability of the i-th group being selected; Randomly select 2 parents from the population k and P l , under a certain probability, the real number crossing mode of the jth point is: Among them, α is a random number between 0 and 1; Randomly select an individual P from the update population i Mutation is performed, and the mutation mode of the jth point is: Among them, P max , P min are the individual upper and lower bounds respectively; υ is the number of iterations at this time; V max is the maximum number of evolutions; μ is a random number ranging from 0 to 1; Check whether the obtained individuals meet the optimal standards. When the obtained individuals cannot meet the standard individuals, the fitness check is performed again. When the obtained individuals meet the standards, the optimal individuals are used as the weights and thresholds of the BP neural network, and the learning and training times are repeated until the neural network parameters are adjusted to the set accuracy range or the learning times are met.

8. The method for predicting power grid maintenance resource consumption according to claim 7, characterized in that: Determining the final resource consumption by using the maintenance operation resource consumption history reference amount and the resource consumption prediction target further includes: Calculating a historical reference amount of resource consumption based on a prediction interval of the historical reference amount of resource consumption; A weighted average of the historical reference amount of resource consumption and the predicted target of resource consumption is calculated using a first weight and a second weight as the optimal resource consumption, wherein the first weight and the second weight are respectively assigned to the historical reference amount of resource consumption and the predicted target of resource consumption, and a volatility judgment matrix is ​​constructed according to the change amplitude and change frequency of the same type of costs to determine the first weight and the second weight according to the volatility judgment matrix.

9. A power grid maintenance resource consumption prediction system, characterized in that: A data acquisition module, used to acquire data on resource consumption and influencing factors of multiple maintenance operations, wherein the data includes historical data and real-time data; A preprocessing module, used for identifying outliers on the historical data by using density-based noise application spatial clustering, wherein the historical data is interpolated for missing values; A clustering module is used to perform clustering processing on maintenance operation resource consumption of different types of objects of the same maintenance operation in combination with historical data after outlier processing; The historical reference quantity acquisition module randomly extracts sample data of resource consumption of the same region and the same maintenance operation, and performs data analysis on the sample data through normal model parameter estimation, chi-square fit test and prediction interval estimation to obtain the prediction interval of the historical reference quantity of resource consumption; A prediction module, which is used to construct a neural network prediction model and use the historical data for training, and combine the real-time data of the maintenance operation to be predicted to predict the resource consumption prediction target through the trained neural network prediction model; and A weighting module determines a final resource consumption based on the historical reference amount of resource consumption of the maintenance operation and the resource consumption prediction target, wherein the historical reference amount of resource consumption is obtained based on a prediction interval of the historical reference amount of resource consumption.

10. The power grid maintenance resource consumption prediction system according to claim 9, characterized in that: The preprocessing module comprises: a stratified identification submodule, for stratifying the historical data according to equipment type, operation type, geographic location or environment type to identify abnormal values, wherein the abnormal values ​​include missing values; The intra-layer mean calculation submodule is used to calculate the mean of the complete observation records within each layer; A hierarchical interpolation submodule is used to fill in missing values ​​using the mean value within the hierarchy according to the hierarchy to which the missing values ​​belong when there are missing value records in the historical data; Weighted or complex imputation submodules to account for intra-stratum variability, sample size, assigning different weights to different strata or using weighted averages or multiple imputations to reflect the inherent structure of the historical data; and The iteration and verification submodule is used to re-evaluate the stratification and statistics based on the imputed historical data and iterate again until the predetermined convergence criteria or the upper limit of the number of iterations is met.

Citation Information

Patent Citations

  • Park power system net load combined prediction method based on adaptive error feedback

    CN110909912A

  • Power grid primary equipment operation and maintenance cost determination method based on support vector machine

    CN112633528A

  • Method for intelligently determining operation and maintenance debugging cost of primary equipment

    CN115879361A

Cited By

  • Lyocell fiber production and manufacturing cost prediction method and application

    CN119941341A

  • Building power consumption data layered cleaning method, medium and electronic equipment

    CN120086213A

  • Experimental method based on foundation pit dewatering simulation device and foundation pit dewatering simulation device

    CN120873929A

  • Resource command control optimization method and system

    CN121279755A

  • A resource command control optimization method and system

    CN121279755B