Optimal maintenance position determination method and device for power grid fault prediction

Abstract

The invention relates to an optimal maintenance position determination method and device for power grid fault prediction, belongs to the technical field of data preprocessing and recognition, and solves the problem of how to determine an optimal maintenance position after a power grid fault under natural disasters. The method comprises the following steps: acquiring historical fault data and related influence factor data, and performing interpolation processing through an adversarial network; selecting a plurality of key influence factors; judging the correlation degree between the power grid fault and each key influence factor by adopting an improved anti-entropy-grey correlation analysis method so as to select part of key influence factors as influence indexes; constructing and training a neural network to obtain a prediction model, taking the geological disaster probability and the influence index as input parameters of the prediction model, and constructing a geological disaster early warning model to obtain the geological disaster probability; and an optimal maintenance position optimization model is constructed based on the predicted power grid fault, and an optimal maintenance position is obtained by combining an objective function and a constraint condition. And accurately predicting the maintenance demand by using the geological disaster probability and the key influence index.

Description

Technical Field

[0001] The present application relates to the technical field of data preprocessing and identification, and in particular to a method and device for determining an optimal maintenance location for power grid fault prediction. Background Art

[0002] With global climate change and the frequent occurrence of extreme weather events, power grid systems are facing more and more challenges caused by natural disasters, among which geological disasters caused by rainfall are particularly prominent. Traditional power grid fault prediction and maintenance strategies often rely on empirical judgment and single model analysis, which makes it difficult to accurately capture the nonlinear relationship between complex environmental factors and power grid faults, resulting in uneven distribution of maintenance resources, long response time, and high costs.

[0003] Bayesian Model Averaging (BMA) is a statistical method that improves the overall prediction accuracy by integrating the prediction results of multiple models. It is particularly suitable for dealing with data analysis problems with high uncertainty. Hierarchical hybrid models can further refine the data structure, identify influencing factors at different levels, and improve the explanatory power and predictive ability of the model.

[0004] Although data analysis technology has begun to be used for fault prediction in current power grid operation and maintenance management, most of them are concentrated on the application of a single model, lacking efficient integration of multi-source data and in-depth exploration of complex interaction mechanisms. Especially in the face of extreme weather events, how to accurately predict the impact of geological disasters on power grid facilities and then quickly locate maintenance needs is still an urgent problem to be solved.

[0005] Difficulty in data integration: Historical fault data, meteorological data and geological disaster data come from various sources and in different formats, making integration and cleaning a cumbersome task.

[0006] Identification of key factors: Power grid failures are affected by many factors, including but not limited to rainfall, geological structure, equipment aging, etc. It is difficult to accurately screen out key influencing factors.

[0007] Model prediction accuracy: It is difficult for a single model to fully capture the complex mechanism of failure occurrence, and the generalization ability and robustness of the model need to be improved.

[0008] Maintenance resource optimization: How to achieve rapid and accurate scheduling of maintenance teams under limited resource conditions and reduce power outage time and economic losses is a major challenge in actual operations.

[0009] In summary, there is an urgent need for a comprehensive technical solution that can integrate multiple technologies such as multi-source data processing, advanced statistical modeling, and intelligent algorithm optimization to improve the accuracy of power grid fault prediction and optimize maintenance resource allocation to ensure safe and stable operation of the power grid. Summary of the invention

[0010] In view of the above analysis, the embodiments of the present application aim to provide a method and device for determining the optimal maintenance location for power grid fault prediction, so as to solve the problem of how to determine the optimal maintenance location after a power grid fault under natural disasters.

[0011] On the one hand, an embodiment of the present application provides a method for determining the optimal maintenance location for power grid fault prediction, including: obtaining historical fault data and relevant influencing factor data, and performing outlier interpolation processing on the historical fault data and the relevant influencing factor data through an adversarial network, wherein the relevant influencing factor data include power grid operation data, meteorological data and geographic information data; selecting multiple key influencing factors from the relevant influencing factor data by fusing a Bayesian model with a hierarchical hybrid model and using statistical information; using an improved anti-entropy-grey correlation analysis method to determine the correlation between the power grid fault and each key influencing factor, so as to select some key influencing factors from the multiple key influencing factors according to the correlation as influencing indicators; constructing a BP neural network and using the historical fault data and the influencing indicators for training to obtain a power grid fault prediction model, wherein the probability of geological disasters and the influencing indicators are used as input parameters of the power grid fault prediction model, and a rainfall-induced geological disaster warning model is constructed to obtain the probability of geological disasters; and constructing an optimal maintenance location optimization model based on the predicted power grid fault, combining the objective function with the constraint conditions to obtain the optimal maintenance location.

[0012] The beneficial effects of the above technical solution are as follows: This application can not only make full use of statistical information and screen out key influencing factors from multiple dimensions, but also enhance the model's ability to handle uncertainty and complexity, and improve the accuracy of fault prediction and the robustness of the model. The improved anti-entropy-grey correlation analysis method further strengthens the quantitative evaluation of the correlation between key indicators and faults, so that the prediction model can better reflect the actual situation, reduce false alarms and omissions, and provide a more reliable basis for maintenance decisions. Based on the improved neural network power grid fault prediction model, combined with the geological disaster warning caused by rainfall, this application is incorporated into a unified framework to achieve forward-looking and accurate prediction of maintenance needs.

[0013] Based on the further improvement of the above method, the outlier interpolation processing of the historical fault data and the related influencing factor data through the adversarial network further includes: generating a reconstructed sample for adversarial through a generator; distinguishing the distance between the adversarial sample and the original sample through a discriminator, wherein the original sample is any sample data of the historical fault data and the related influencing factor data; and reducing the distance between the reconstructed sample G(z) and the original sample x by using a reconstruction loss function, wherein the reconstruction loss function is expressed by the following formula:

[0014] L r(z) = ||x·mG(z)·m||2;

[0015] The output probability of the discriminator is maximized by using a discriminant loss function to ensure that the reconstructed sample G(z) conforms to the distribution of the original sample, wherein the discriminant loss function is expressed by the following formula:

[0016] L d (z) = -D(G(z));

[0017] A filling loss function is generated according to the reconstruction loss function and the discrimination loss function, and the filling loss function is minimized based on the reconstructed sample and the original sample to generate a missing value for interpolation. The filling loss function and the missing value are respectively expressed by the following formulas:

[0018] L imputation (z) = L r (z)+λ·L d (z);

[0019] x imputed =x·m+(1-m)·G(z * );

[0020] Wherein, x represents the original sample; m is a mask matrix used to indicate the missing parts of the historical fault data and the related influencing factor data; G(z) represents the reconstructed sample, L r (z) represents the distance between the generated sample and the original sample; L d (z) represents the discriminant loss function, which is used to evaluate the output probability of the discriminator; D(G(z)) is the output of the discriminator, which represents the probability that the reconstructed sample G(z) is discriminated as the original sample; L imputation (z) is a comprehensive loss function, which consists of a reconstruction loss function and a discrimination loss function; λ is a hyperparameter used to control the ratio between the reconstruction loss and the discrimination loss; x imputed represents the padded data sample; x·m represents the non-missing part of the original sample; (1-m)·G(z) represents the corresponding missing part of the reconstructed sample; G(z * ) represents filling the missing values ​​of the original sample.

[0021] Based on the further improvement of the above method, the improved anti-entropy-grey correlation analysis method is used to judge the correlation between power grid faults and each key influencing factor, which further includes:

[0022] The matrix X of key influencing factors is expressed by the following formula m for:

[0023]

[0024] Among them, x ij For the value of the jth key influencing factor of the ith object, n key influencing factors, k regions, and m years of data are selected from the plurality of key influencing factors;

[0025] Based on the entropy values ​​of multiple key influencing factors, the anti-entropy of the following key influencing factors is calculated:

[0026]

[0027] Among them, c ij is the entropy value of the jth key influencing factor under the i-th object;

[0028] The anti-entropy of each key influencing factor is normalized to obtain the objective weight of the corresponding anti-entropy:

[0029]

[0030] Among them, d j is the objective weight of the jth key influencing factor;

[0031] Based on the objective weights and subjective weights, the following combined weights are calculated:

[0032]

[0033] Among them, w j is the combined weight of the key influencing factor j; c j d j are the objective weight and subjective weight of the jth key influencing factor respectively;

[0034] The correlation degree γ between the power grid fault and the key influencing factor j is calculated by the following formula j :

[0035]

[0036]

[0037] T r =1-e r ;

[0038] f(E r ,e r )=E r (1-e r );

[0039]

[0040] Among them, γ j is the correlation between the power grid fault and the key influencing factor j;rj is the correlation coefficient between the power grid fault and the key influencing factor j determined based on the combined weight; T r is the residual degree of the key influencing factor r; e r is the relative entropy, which represents the ratio of the anti-entropy of the key influencing factor r to the maximum entropy; E r 、E max are the anti-entropy and maximum entropy of the key influencing factors respectively; b(r) is the proportional weight of other key influencing factors r, f(E r ,e r ) is a function of anti-entropy and relative entropy.

[0041] Based on the further improvement of the above method, constructing a geological disaster warning model caused by rainfall to obtain the probability of the geological disaster further includes: when the time length is a discrete time interval of one year, the average annual probability of geological disasters is related to the time period T of the occurrence of geological disasters, wherein the annual probability of geological disasters is expressed by the following formula:

[0042]

[0043] Among them, T is the occurrence period of geological disaster events, λ is the frequency of geological disasters; the probability of geological disasters occurring in many years is expressed by the following formula:

[0044]

[0045] Where e is a natural constant, λ is the frequency of geological disasters under limited rainfall conditions; the probability of one or more geological disasters in the next t years is expressed by the following formula:

[0046] P=1-e -λt .

[0047] Based on further improvements of the above method, constructing a neural network power grid fault prediction model further includes: normalizing the historical fault data, the power grid operation data, the meteorological data and the geographic information data; dividing the historical fault data and the corresponding influencing indicators into multiple data subsets, wherein one subset of the multiple data subsets is used as a validation set and the remaining subsets are used as training sets; before training the neural network, using a genetic algorithm to optimize the network structure and initial weights; constructing the BP neural network according to the network structure and initial weights optimized by the genetic algorithm, the BP neural network includes an input layer, a hidden layer and an output layer, wherein the number of the input layers is determined according to the influencing indicators and the probability of geological disasters, the number and structure of the hidden layers are determined according to the genetic algorithm, and the output of the output layer is a fault rate prediction value; using the training set, applying the back propagation algorithm to adjust the network weights and thresholds of the BP neural network to obtain the power grid fault prediction model; using the validation set to calculate the prediction error of the trained BP neural network; and inputting real-time influencing indicators into the power grid fault prediction model to predict the power grid failure rate in real time.

[0048] Based on the further improvement of the above method, according to the network structure and initial weights optimized by the genetic algorithm, constructing the BP neural network further includes: expressing the initialized BP neural network by the following formula:

[0049]

[0050] Where X is the neuron input vector; n is the number of input layer nodes; Y is the neuron output vector, and m is the number of output layer nodes; the output value H of the hidden layer is expressed by the following formula: aj :

[0051]

[0052] Wherein, p is the number of nodes in the hidden layer; ω ij is the weight between the input layer and the hidden layer; a is the threshold; f is the activation function; the output value H of the output layer is expressed by the following formula bk :

[0053]

[0054] Among them, ω jk is the weight between the hidden layer and the output layer; b is the threshold; the predicted output H is expressed by the following formula bk With the expected output Y k The error between k :

[0055] ek =|Y k -H bk |;

[0056] Among them, when max(e k )>e0, the BP neural network parameters are adjusted and learning and training are performed again until all training data meet the specified error standard, where e0 is the maximum allowable error value; the weight ω between the input layer and the hidden layer is updated according to the error ij And the weight ω between the hidden layer and the output layer jk :

[0057]

[0058] ω jk =ω jk +ηH aj e k ;

[0059] Among them, η is the learning rate, which is between 0 and 1; the threshold is corrected by the following formula:

[0060]

[0061] λ b =λ b +e k .

[0062] Based on the further improvement of the above method, an optimal maintenance location optimization model is constructed based on the predicted power grid fault, and combining the objective function and the constraint conditions to obtain the optimal maintenance location further includes: determining the objective function and constraint conditions of the maintenance station optimization model, and setting the initial value set of the distribution network fault maintenance station location coordinates; calculating the optimal location coordinates of the distribution network fault maintenance station according to the initial value set, the objective function and the constraint conditions; judging whether the termination condition is met according to the predicted number of iterations, wherein when the termination condition is met, the optimal solution of the maintenance station is obtained, and the optimization of the distribution network fault maintenance station location is completed; when the termination condition is not met, the initial value set of the maintenance station location coordinates is updated; and the distance d from the maintenance station to the feeder is calculated by the following formula ij :

[0063]

[0064] Among them, x i -a j For x i and a j The Euclidean distance of j is the feeder equivalent point a j Failure coefficient; a j(j=1,2,...,n) is the feeder equivalent point, x i (i=1,2,...,m) is the maintenance station, where m and n are the number of maintenance stations and the number of feeder equivalent points respectively; the time T from the maintenance station to the maintenance location is calculated by the following formula ij The calculation formula is:

[0065] T ij =(d ij ·a·λ·r) / v;

[0066] Among them, a, λ, r are the road congestion coefficient, road tortuosity coefficient, and feeder tortuosity coefficient, respectively, and v is the average speed of the maintenance vehicle. The maximum time T required to rush from the maintenance station to the feeder is calculated by the following formula: M :

[0067] T M =T max -t1-t2-t3-t4;

[0068] t1, t2, t3, and t4 are the time taken by the service center to process the work order, the time taken by the dispatch layer to locate and handle the fault, the time taken to prepare the maintenance vehicle, and the time taken to find the maintenance location. The shortest distance required from the specific feeder to the maintenance location is calculated using the following formula:

[0069]

[0070] In this model, it is assumed that a feeder belongs to only one maintenance station, b ij is the condition that the standard range of fault maintenance time is satisfied. If the time from the i-th maintenance station to the j-th feeder is within the standard range of fault maintenance time, then b ij =1, otherwise b ij =0,a ij represents the distance from feeder i to maintenance station j; the constraint condition of the maintenance station location optimization model: the time required to rush from the maintenance station to the fault location is:

[0071] T ij <T M ;

[0072] The constraint condition for arranging all feeder maintenance work at m maintenance stations is:

[0073]

[0074] Based on further improvements of the above method, obtaining historical fault data and related influencing factor data further includes: the historical fault data includes fault occurrence time, fault type, fault location, fault duration, fault impact range, fault repair time, maintenance cost, fault cause and equipment information; the power grid operation data includes current monitoring data, voltage monitoring data, power factor, load change data, transformer oil temperature, transformer oil level, insulation resistance test record facility inspection record and maintenance and repair log; the meteorological data includes rainfall, temperature, humidity, wind speed and direction, air pressure, snowfall, ground subsidence or soil moisture content and lightning strike data; and the geographic information data includes the geographical distribution of power grid facilities, topographic data and land cover type.

[0075] On the other hand, an embodiment of the present application provides an optimal maintenance location determination device for power grid fault prediction, including: a data acquisition module, used to acquire historical fault data and related influencing factor data, wherein the related influencing factor data include power grid operation data, meteorological data and geographic information data; an interpolation module, used to perform outlier interpolation processing on the historical fault data and the related influencing factor data through an adversarial network; a primary selection module, used to select multiple key influencing factors from the related influencing factor data by fusing a Bayesian model with a hierarchical hybrid model and utilizing statistical information; a secondary selection module, used to use an improved anti-entropy-grey correlation analysis method to judge the power grid The correlation between the fault and each key influencing factor is used to select some key influencing factors from the multiple key influencing factors as influencing indicators according to the correlation; a fault prediction model is used to construct a BP neural network and use the historical fault data and the influencing indicators for training to obtain a power grid fault prediction model, wherein the probability of geological disasters and the influencing indicators are used as input parameters of the power grid fault prediction model, and a rainfall-induced geological disaster early warning model is constructed to obtain the probability of geological disasters; and an optimal position acquisition module is used to construct an optimal maintenance position optimization model based on the predicted power grid fault, combining the objective function with the constraint conditions to obtain the optimal maintenance position.

[0076] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0077] 1. Improve fault prediction accuracy and robustness: This application integrates the Bayesian model averaging strategy and the hierarchical hybrid model to achieve in-depth mining and comprehensive analysis of historical fault data and environmental factors. This fusion method can not only make full use of statistical information and screen out key influencing factors from multiple dimensions, but also enhance the model's ability to handle uncertainty and complexity, and improve the accuracy of fault prediction and the robustness of the model. The improved anti-entropy-grey correlation analysis method further strengthens the quantitative evaluation of the correlation between key indicators and faults, so that the prediction model can better reflect the actual situation, reduce false alarms and omissions, and provide a more reliable basis for maintenance decisions.

[0078] 2. Optimize maintenance resource allocation and response speed: Based on the improved neural network power grid fault prediction model and combined with the geological disaster warning caused by rainfall, this application incorporates the probability of geological disasters and key influencing indicators into a unified framework to achieve forward-looking and accurate prediction of maintenance needs. On this basis, by constructing an optimization model for the optimal maintenance location and using a small habitat genetic algorithm to solve it, it can effectively consider multi-objective optimization (such as cost, time, resource utilization) and constraints (such as maintenance team capabilities, traffic conditions), ensuring efficient allocation and rapid response of maintenance resources. This directly reduces maintenance costs, shortens fault recovery time, and improves overall maintenance efficiency.

[0079] 3. Enhance system adaptability and disaster resilience: This application significantly enhances the adaptability and resilience of the power grid system to natural disasters by comprehensively considering the impact of various natural environmental factors on the power grid, especially the in-depth analysis of rainfall and geological disasters. In the context of frequent extreme weather, this capability is particularly important. It can not only reduce the damage of natural disasters to power grid facilities, but also ensure the stability and continuity of power supply to the greatest extent through early warning and rapid response mechanisms, and reduce the negative impact on social and economic activities.

[0080] 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

[0081] 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.

[0082] Figure 1 A flowchart of a method for determining an optimal maintenance location for power grid fault prediction according to an embodiment of the present application;

[0083] Figure 2 A BP neural network of a power grid fault prediction model according to an embodiment of the present application;

[0084] Figure 3 A block diagram for determining an optimal maintenance location for power grid fault prediction according to an embodiment of the present application. DETAILED DESCRIPTION

[0085] 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.

[0086] refer to Figure 1 A specific embodiment of the present application discloses a method for determining the optimal maintenance location for power grid fault prediction, including: in step S101, obtaining historical fault data and relevant influencing factor data, and performing outlier interpolation processing on the historical fault data and the relevant influencing factor data through an adversarial network, wherein the relevant influencing factor data include power grid operation data, meteorological data and geographic information data; in step S102, by fusing the Bayesian model with the hierarchical hybrid model and using statistical information, multiple key influencing factors are selected from the relevant influencing factor data; in step S103, an improved anti-entropy-grey correlation analysis method is used to determine the optimal maintenance location for power grid fault prediction. The correlation between the grid failure and each key influencing factor is calculated, and some key influencing factors are selected from multiple key influencing factors as influencing indicators according to the correlation; in step S104, a BP neural network is constructed and trained with historical fault data and influencing indicators to obtain a grid failure prediction model, wherein the geological disaster probability and influencing indicators are used as input parameters of the grid failure prediction model, and a rainfall-induced geological disaster early warning model is constructed to obtain the geological disaster probability; and in step S105, an optimal maintenance location optimization model is constructed based on the predicted grid failure, and the objective function and constraint conditions are combined to obtain the optimal maintenance location.

[0087] Improvements in the steps of collecting and organizing historical fault data and geological disaster data: Systematically collect historical power grid fault data and geological disaster data such as rainfall, organize and integrate these data to ensure the comprehensiveness and accuracy of the data. Technical effect: Through comprehensive data collection and organization, more historical fault information and geological disaster data can be obtained, providing a solid foundation for subsequent analysis and model building.

[0088] Improvements in the step of screening key influencing factors by integrating the Bayesian model averaging strategy and the hierarchical hybrid model method: The Bayesian model averaging strategy and the hierarchical hybrid model method are used to maximize the use of statistical information to screen out effective and robust key influencing factors for power grid failures from multiple influencing factors. Technical effect: This method can more accurately identify the key factors that affect power grid failures and improve the prediction accuracy and robustness of the model.

[0089] Improved anti-entropy-grey correlation analysis method for quantitative analysis of the correlation between faults and indicators: Improved anti-entropy-grey correlation analysis method is used to quantitatively analyze the correlation between faults and their indicators and determine the correlation between the two. Technical effect: The improved analysis method can more accurately quantify the correlation between faults and key indicators, provide more reliable analysis results, and improve the accuracy of fault prediction.

[0090] Improvements in the steps of building a warning model for geological disasters caused by rainfall: The probability of geological disasters caused by rainfall and key influencing indicators are used as input parameters to build a power grid failure prediction model with an improved neural network. Technical effect: By introducing the probability of geological disasters into the model, the impact of natural disasters on power grid failures can be more comprehensively considered, improving the accuracy and practicality of the fault prediction model.

[0091] Improvements in the steps of constructing the maintenance optimal location optimization model: Based on the fault prediction model, the maintenance optimal location optimization model is constructed, and the objective function and constraint conditions are set. The niche genetic algorithm is used to solve the problem and obtain the optimal maintenance location. Technical effect: This optimization model can effectively reduce maintenance costs, improve maintenance efficiency, and ensure the rapid recovery and stable operation of the power grid.

[0092] Compared with the prior art, the method for determining the optimal maintenance location for power grid fault prediction provided in this embodiment can not only make full use of statistical information and screen out key influencing factors from multiple dimensions, but also enhance the model's ability to handle uncertainty and complexity, thereby improving the accuracy of fault prediction and the robustness of the model. The improved anti-entropy-grey correlation analysis method further strengthens the quantitative evaluation of the correlation between key indicators and faults, so that the prediction model can better reflect the actual situation, reduce false alarms and missed alarms, and provide a more reliable basis for maintenance decisions. Based on the improved neural network power grid fault prediction model, combined with the geological disaster warning caused by rainfall, this application incorporates the probability of geological disasters and key influencing indicators into a unified framework to achieve forward-looking and accurate prediction of maintenance needs.

[0093] In the following, reference Figure 1 and Figure 2 , each step in the method for determining the optimal maintenance location for power grid fault prediction according to an embodiment of the present invention is described in detail.

[0094] In step S101, historical fault data and relevant influencing factor data are obtained, and outlier interpolation processing is performed on the historical fault data and relevant influencing factor data through an adversarial network, wherein the relevant influencing factor data includes power grid operation data, meteorological data and geographic information data.

[0095] Specifically, obtaining historical fault data and related influencing factor data further includes: historical fault data including fault occurrence time, fault type, fault location, fault duration, fault impact range, fault repair time, maintenance cost, fault cause and equipment information; power grid operation data including current monitoring data, voltage monitoring data, power factor, load change data, transformer oil temperature, transformer oil level, insulation resistance test records, facility inspection records and maintenance and repair logs; meteorological data including rainfall, temperature, humidity, wind speed and direction, air pressure, snowfall, ground subsidence or soil moisture content and lightning strike data; and geographic information data including the geographical distribution of power grid facilities, topographic data and land cover type.

[0096] Specifically, performing outlier interpolation processing on historical fault data and related influencing factor data through an adversarial network further includes: generating a reconstructed sample for adversarial use through a generator; distinguishing the distance between the adversarial sample and the original sample through a discriminator, wherein the original sample is any sample data of the historical fault data and the related influencing factor data; and reducing the distance between the reconstructed sample G(z) and the original sample x using a reconstruction loss function, wherein the reconstruction loss function is represented by the following formula:

[0097] L r (z) = ||x·mG(z)·m||2;

[0098] The discriminant loss function is used to maximize the output probability of the discriminator to ensure that the reconstructed sample G(z) conforms to the distribution of the original sample. The discriminant loss function is expressed by the following formula:

[0099] L d (z) = -D(G(z));

[0100] A filling loss function is generated according to the reconstruction loss function and the discriminant loss function, and the filling loss function is minimized based on the reconstructed samples and the original samples to generate missing values ​​for interpolation. The filling loss function and the missing value are respectively expressed by the following formulas:

[0101] L imputation (z) = L r (z)+λ·L d (z);

[0102] x imputed =x·m+(1-m)·G(z * );

[0103] Among them, x represents the original sample; m is the mask matrix, which is used to indicate the missing parts in the historical fault data and related influencing factor data; G(z) represents the reconstructed sample, L r (z) represents the distance between the generated sample and the original sample; L d (z) represents the discriminant loss function, which is used to evaluate the output probability of the discriminator; D(G(z)) is the output of the discriminator, which represents the probability that the reconstructed sample G(z) is discriminated as the original sample; L imputation (z) is a comprehensive loss function, which consists of a reconstruction loss function and a discrimination loss function; λ is a hyperparameter used to control the ratio between the reconstruction loss and the discrimination loss; x imputed represents the padded data sample; x·m represents the non-missing part of the original sample; (1-m)·G(z) represents the corresponding missing part of the reconstructed sample; G(z * ) represents filling the missing values ​​of the original sample.

[0104] In step S102, a plurality of key influencing factors are selected from the relevant influencing factor data as a key influencing factor set by fusing the Bayesian model with the hierarchical hybrid model and using statistical information.

[0105] Specifically, by fusing the Bayesian model with the hierarchical hybrid model and utilizing statistical information, multiple key influencing factors are selected from the relevant influencing factor data as the key influencing factor set, which further includes: defining the model space step: building multiple regression models based on the historical fault data and the relevant influencing factor data; specifying the prior distribution step: assigning the appropriate prior distribution to the regression model parameters based on the historical fault data and the domain expert knowledge; calculating the posterior probability step: using the Bayesian theorem in combination with the historical fault data to calculate the posterior probability of each regression model; model averaging step: weighted averaging the prediction results of the regression model according to the posterior probability of each regression model; inference and prediction step: using the weighted average results of the regression model to perform parameter estimation, hypothesis testing or prediction of future observations.

[0106] Define the model space step: Based on the historical failure data and related influencing factors (such as meteorological conditions, geographical location, equipment aging, etc.), build multiple possible regression models. These models can consider different fixed effects (such as linear effects or nonlinear effects of specific influencing factors), random effects (such as random effects of equipment models or random effects of regions), and different covariance structures (such as the correlation of data within the same region). Assume that there are the following three influencing factors: temperature, humidity, and equipment age. The model set can include the following models:

[0107] Model 1: Fixed effects are temperature and humidity, and the random effect is device age. Model 2: Fixed effects are temperature and device age, and the random effect is humidity. Model 3: Fixed effects are all three factors, and the random effect is null. Model 4: Includes a quadratic term for temperature and has a complex covariance structure.

[0108] Specify prior distribution step: In Bayesian inference, the prior distribution represents the knowledge or assumptions about the model parameters before observing the data. For each model, specify a reasonable prior distribution for the model parameters based on historical failure data and the knowledge of domain experts. You can specify a normal distribution for the regression coefficient and an inverse gamma distribution for the variance parameter. If historical data indicates that temperature may have a large effect on the failure rate, you can specify a normal distribution prior with a large mean for the temperature coefficient. For uncertain random effects, you can choose a prior distribution with a large variance.

[0109] Steps to calculate the posterior probability: Calculate the posterior probability of each model by using Bayes' theorem and combining historical fault data. The posterior probability reflects the relative applicability of each model given the data. When calculating the posterior probability, if a model better explains the data (such as a better fit with the historical fault data), the posterior probability of the model will be higher.

[0110] Model averaging step: Use the posterior probabilities of all models as weights and perform a weighted average of the predictions of these models. In this way, the uncertainty of a single model can be avoided and the robustness of the prediction can be improved. If the posterior probability of model 2 is the highest, followed by model 1 and model 3, then the prediction of model 2 will account for a larger proportion of the final weighted average, but model 1 and model 3 will also contribute to the final result.

[0111] Inference and prediction steps: Use the model averaged results to perform parameter estimation, hypothesis testing, or prediction of future observations.

[0112] In step S103, an improved anti-entropy-grey correlation analysis method is used to determine the correlation between the power grid fault and each key influencing factor, so as to select some key influencing factors from the key influencing factor set as influencing indicators according to the correlation.

[0113] Specifically, the improved anti-entropy-grey correlation analysis method is used to determine the correlation between the power grid fault and each key influencing factor, further including: calculating the entropy value for each of the multiple key influencing factors; converting the entropy value of each key influencing factor into an anti-entropy value through an anti-entropy formula, wherein the higher the anti-entropy value, the more certain the information provided by the key influencing factor and the greater the difference, so as to give a greater weight; generating an objective weight based on the anti-entropy value, and calculating a combined weight based on the objective weight and the subjective weight; calculating a correlation coefficient between each comparison sequence and a reference sequence based on the combined weight, wherein the historical fault data is selected as the reference sequence, and the multiple key influencing factors are selected as multiple comparison sequences; and calculating the correlation between the reference sequence and each comparison sequence based on the correlation coefficient as the correlation between the power grid fault and each key influencing factor.

[0114] Specifically, the improved anti-entropy-grey correlation analysis method is used to determine the correlation between the power grid fault and each key influencing factor, further comprising: the matrix X of the key influencing factors is expressed by the following formula: m for:

[0115]

[0116] Among them, x ij The value of the jth key influencing factor of the i-th object is selected from multiple key influencing factors, n key influencing factors, k regions, and m years of data; based on the entropy values ​​of multiple key influencing factors, the anti-entropy of the following key influencing factors is calculated:

[0117]

[0118] Among them, c ij is the entropy value of the jth key influencing factor under the ith object; the anti-entropy of each key influencing factor is normalized to obtain the objective weight of the corresponding anti-entropy:

[0119]

[0120] Among them, d j is the objective weight of the jth key influencing factor; the following combined weight is calculated based on the objective weight and the subjective weight:

[0121]

[0122] Among them, w j is the combined weight of the key influencing factor j; c j d j are the objective weight and subjective weight of the jth key influencing factor respectively; the correlation degree γ between the power grid fault and the key influencing factor j is calculated by the following formulaj :

[0123] The correlation degree γ between the power grid fault and the key influencing factor j is calculated by the following formula j :

[0124]

[0125] T r =1-e r ; f(E r ,e r )=E r (1-e r );

[0126]

[0127] Among them, γ j is the correlation between the power grid fault and the key influencing factor j; rj is the correlation coefficient between the power grid fault and the key influencing factor j determined based on the combined weight; T r is the residual degree of the key influencing factor r; e r is the relative entropy, which represents the ratio of the anti-entropy of the key influencing factor r to the maximum entropy; E r 、E max are the anti-entropy and maximum entropy of the key influencing factors respectively; b(r) is the proportional weight of other key influencing factors r, f(E r ,e r ) is a function of anti-entropy and relative entropy.

[0128] In step S104, a BP neural network is constructed and trained using historical fault data and impact indicators to obtain a power grid fault prediction model, wherein the geological disaster probability and impact indicators are used as input parameters of the power grid fault prediction model, and a rainfall-induced geological disaster warning model is constructed to obtain the geological disaster probability.

[0129] Specifically, constructing a geological disaster warning model caused by rainfall to obtain the probability of geological disasters further includes: when the time length is a discrete time interval of one year, the average annual probability of geological disasters is related to the time period T of the occurrence of geological disasters, wherein the annual probability of geological disasters is expressed by the following formula:

[0130]

[0131] Among them, T is the occurrence period of geological disaster events, λ is the frequency of geological disasters; the probability of geological disasters occurring in many years is expressed by the following formula:

[0132]

[0133] Where e is a natural constant, λ is the frequency of geological disasters under limited rainfall conditions; the probability of one or more geological disasters in the next t years is expressed by the following formula:

[0134] P=1-e - λt .

[0135] Specifically, constructing a neural network power grid fault prediction model further includes: normalizing historical fault data, power grid operation data, meteorological data and geographic information data; dividing historical fault data and corresponding influencing indicators into multiple data subsets, wherein one subset of the multiple data subsets is used as a validation set and the remaining subsets are used as training sets; before training the neural network, using a genetic algorithm to optimize the network structure and initial weights; constructing a BP neural network based on the network structure and initial weights optimized by the genetic algorithm, the BP neural network includes an input layer, a hidden layer and an output layer, wherein the number of input layers is determined based on the influencing indicators and the probability of geological disasters, the number and structure of the hidden layers are determined based on the genetic algorithm, and the output of the output layer is the failure rate prediction value; using the training set, applying the back propagation algorithm to adjust the network weights and thresholds of the BP neural network to obtain a power grid fault prediction model; using the validation set to calculate the prediction error of the trained BP neural network; and inputting the real-time influencing indicators into the power grid fault prediction model to predict the power grid failure rate in real time.

[0136] Specifically, according to the network structure and initial weights optimized by the genetic algorithm, constructing the BP neural network further includes: expressing the initialized BP neural network by the following formula:

[0137]

[0138] Among them, X is the neuron input vector; n is the number of input layer nodes; Y is the neuron output vector, and m is the number of output layer nodes; the output value H of the hidden layer is expressed by the following formula aj :

[0139]

[0140] Where p is the number of nodes in the hidden layer; ω ij is the weight between the input layer and the hidden layer; a is the threshold; f is the activation function; the output value H of the output layer is expressed by the following formula bk :

[0141]

[0142] Among them, ω jk is the weight between the hidden layer and the output layer; b is the threshold; the predicted output H is expressed by the following formulabk With the expected output Y k The error between k :

[0143] e k =Y k -H bk ;

[0144] Among them, when max(e k )>e0, the BP neural network parameters are adjusted and learning and training are performed again until all training data meet the specified error standard, where e0 is the maximum allowable error value; the weight ω between the input layer and the hidden layer is updated according to the error ij And the weight ω between the hidden layer and the output layer jk :

[0145]

[0146] ω jk =ω jk +ηH aj e k ;

[0147] Among them, η is the learning rate, which is between 0 and 1; the threshold is corrected by the following formula:

[0148]

[0149] λ b =λ b +e k .

[0150] In step S105, an optimal maintenance location optimization model is constructed based on the predicted power grid fault, and the objective function and constraint conditions are combined to obtain the optimal maintenance location.

[0151] Specifically, an optimal maintenance location optimization model is constructed based on predicted power grid faults, and combining the objective function with the constraint conditions to obtain the optimal maintenance location further includes: determining the objective function and constraint conditions of the maintenance station optimization model, and setting the initial value set of the distribution network fault maintenance station location coordinates; calculating the optimal location coordinates of the distribution network fault maintenance station according to the initial value set, the objective function and the constraint conditions; judging whether the termination condition is met according to the predicted number of iterations, wherein when the termination condition is met, the optimal solution of the maintenance station is obtained, and the optimization of the distribution network fault maintenance station location is completed; when the termination condition is not met, the initial value set of the maintenance station location coordinates is updated; and the distance d from the maintenance station to the feeder is calculated by the following formula ij :

[0152]

[0153] Among them, x i -a j For x i and a j The Euclidean distance of j is the feeder equivalent point a j Failure coefficient; a j (j=1,2,...,n) is the feeder equivalent point, x i (i=1,2,...,m) is the maintenance station, where m and n are the number of maintenance stations and the number of feeder equivalent points respectively; the time T from the maintenance station to the maintenance location is calculated by the following formula ij The calculation formula is:

[0154] T ij =(d ij ·a·λ·r) / v;

[0155] Among them, a, λ, r are the road congestion coefficient, road tortuosity coefficient, and feeder tortuosity coefficient, respectively, and v is the average speed of the maintenance vehicle. The maximum time T required to rush from the maintenance station to the feeder is calculated by the following formula: M :

[0156] T M =T max -t1-t2-t3-t4;

[0157] Among them, t1, t2, t3, and t4 are the time taken by the service center to process the work order, the time taken by the dispatch layer to locate and handle the fault, the time taken to prepare the maintenance vehicle, and the time taken to find the maintenance location. The shortest distance required from the specific feeder to the maintenance station is calculated using the following formula:

[0158]

[0159] In this model, it is assumed that a feeder belongs to only one maintenance station, b ij is the condition that the standard range of fault maintenance time is satisfied. If the time from the i-th maintenance station to the j-th feeder is within the standard range of fault maintenance time, then b ij =1, otherwise b ij =0,a ij represents the distance from feeder i to maintenance station j; the time required to rush from the maintenance station to the fault location is constrained as: T ij <T M ; The constraint condition for arranging all feeder maintenance work at m maintenance stations is:

[0160] In the following, reference Figure 1 and Figure 2The optimal maintenance location determination method for power grid fault prediction according to an embodiment of the present invention is described in detail by way of a specific example.

[0161] In step S101, historical fault data is collected and sorted, geological disaster data such as rainfall is collected, and the data is processed to achieve outlier interpolation through generative adversarial networks.

[0162] The collected data include: (1) Historical fault data: fault occurrence time (date and time); fault type (such as short circuit, overload, insulation fault, etc.); fault location (latitude and longitude, substation name, line number); fault duration; impact range (number of affected users, power outage area); fault repair time; maintenance cost; fault cause analysis (preliminary diagnosis and final confirmation); equipment information (equipment model, service life, maintenance record). (2) Meteorological data: rainfall (accumulated amount per unit time, such as hour, day, week); temperature (highest temperature, lowest temperature, average temperature); humidity; wind speed and direction; air pressure; snowfall (applicable to cold areas); ground subsidence or soil moisture content (for areas sensitive to geological disasters); lightning strike data (frequency and intensity of thunderstorm activity). (3) Power grid operation data: current and voltage monitoring data; power factor; load change data (load during peak and valley periods); transformer oil temperature and oil level; insulation resistance test records; facility inspection records; maintenance and overhaul logs. (4) Geographic information data: geographical distribution of power grid facilities (location of substations, transmission lines, and distribution stations); topographic data (altitude, terrain slope); land cover type (urban, farmland, forest land, and water area); after data collection, fill in the blank values ​​in the data.

[0163] For example, historical disaster data: collect data on landslides and mudslides that occurred in a certain area in the past 10 years. The data includes information such as the time, location, intensity, and rainfall of the disaster. Meteorological data: collect meteorological data within the same time period, including rainfall, rainfall intensity, duration, wind speed, etc. Geological data: collect geological data of the area where the transmission line tower base is located, including soil type, geological structure, terrain slope, etc.

[0164] Generative Adversarial Networks (GANs) consist of two parts: a generator and a discriminator. The generator tries to generate realistic data, while the discriminator tries to distinguish between generated data and real data. The two improve each other's ability through adversarial training. The generator (G) observes certain components of the real data vector, estimates the missing components based on what is actually observed, and outputs a complete vector. The discriminator (D) then uses a complete vector and tries to determine which components are actually observed and which are interpolated.

[0165] In order to reduce the difference between the original time series data x and the newly generated sample G(z), increase the similarity between the two, and enable G(z) to be used for missing value filling, this algorithm adopts the strategy of optimizing the input vector z. For any missing time series data x, this algorithm attempts to find a corresponding optimal z* from the standard normal distribution so that the generated complete data G(z*) is most similar to the original data x. To achieve this goal, this algorithm defines a filling loss function. The filling loss function consists of two parts: the first part is the reconstruction loss function, and the second part is the discriminant loss function. The reconstruction loss function is committed to reducing the distance between the new sample G(z) and the original sample x, and the discriminant loss function is committed to ensuring that the new sample G(z) still conforms to the distribution of the original data set and can still be mistaken for the real thing.

[0166] The definition of the reconstruction loss function is shown in the formula:

[0167] L r (z) = ∥x·mG(z)·m∥2;

[0168] Where G(z) represents the newly generated sample, L r (z) represents the distance between the time series data and the sample.

[0169] The specific definition of the discriminant loss function is shown in the formula:

[0170] L d (z) = -D(G(z));

[0171] The discriminant loss function attempts to maximize the output probability of the discriminator to ensure that the new sample G(z) still conforms to the distribution of the original data set.

[0172] The filling loss function defined by the algorithm is shown in the formula:

[0173] L imputation (z) = L r (z)+λ·L d (z);

[0174] The filling loss function is composed of the reconstruction loss function and the discriminant loss function. λ is a hyperparameter that controls the ratio between the two. For any missing time series data x, this algorithm fixes all weight parameters of the entire generative adversarial network, takes the input vector z of the generator as the optimization object, and obtains the corresponding optimal input vector z* by minimizing the filling loss function, so that the corresponding new sample G(z*) is most similar to the original sample x, and the missing values ​​in the original sample x are filled by G(z*).

[0175] Missing value filling is as shown in the formula:

[0176] ximputed =x·m+(1-m)·G(z * );

[0177] Reconstruction loss function L r (z), represents the distance between the newly generated sample G(z) and the original data sample. x: original data sample. m mask matrix, used to indicate which values ​​in the data are missing. G(z) is the output of the generative model, representing the generated new sample. L d (z) Discriminant loss function, used to evaluate the output probability of the discriminator. D(G(z)) The output of the discriminator represents the probability that the generated sample G(z) is judged as a real sample. L imputation (z) Comprehensive loss function, which consists of reconstruction loss function and discrimination loss function. The λ hyperparameter controls the ratio between reconstruction loss and discrimination loss. imputed The data sample after filling. x·m represents the non-missing part of the original data sample. The corresponding missing part in the new sample generated by (1-m)·G(z). Through the above filling process, this algorithm can ensure that for any time series data, the most similar new sample can be found to fill the missing values.

[0178] In step S102, by integrating the Bayesian model averaging strategy and the hierarchical hybrid model method, statistical information is used to the maximum extent, and effective and robust key influencing factors for power grid failures are screened out from multiple influencing factors.

[0179] Bayesian Model Averaging (BMA) is a statistical method that deals with model uncertainty by ensembling multiple models on a given dataset. Instead of selecting a single best model, BMA assigns a posterior probability to each model and then weights the model's predictions based on these probabilities to obtain the final prediction or inference result. This approach recognizes that in the real world, due to data limitations or model simplifications, no single model can perfectly capture the process of generating all data.

[0180] Mixed regression models (such as mixed linear models, mixed effects models, etc.) are a collection of various types of regression models that can be used to capture complex structures in the data, such as the combination of fixed effects and random effects, or heterogeneity in the data. Mixed models are particularly suitable for data analysis with hierarchical structures or clustering effects.

[0181] To combine Bayesian model averaging with a mixed regression model, the following steps can be taken:

[0182] Define the model space: First, identify your model set, which may include different mixed regression models, each with different fixed effects terms, random effects structures, or different covariance structures.

[0183] Specify prior distributions: Specify appropriate prior distributions for each model and parameter in the model. This step is critical to Bayesian analysis because it encodes our beliefs about the model parameters before we see the data.

[0184] Calculate the posterior probability: Using Bayes' theorem, calculate the posterior probability of each model using the data. This involves calculating the likelihood function and prior probability of the model, as well as a normalization constant to ensure that the sum of the posterior probabilities of all models is 1.

[0185] Model Averaging: Each model is weighted according to its posterior probability, and then the predictions of the models are weighted averaged. This means that the more likely models will contribute more to their final predictions.

[0186] Inference and prediction: Use the averaged model results to estimate parameters, test hypotheses, or predict future observations. This step takes into account all information about model uncertainty and is usually more robust than predictions based on a single model.

[0187] Sensitivity analysis: Analyze the impact of different prior choices on the results to ensure the robustness of the results.

[0188] For example, we collected historical failure data of power grid equipment in a certain area, as well as environmental data of the same period, including temperature, humidity, wind speed, etc. We applied a hierarchical mixed model to analyze the contribution of environmental factors at each level to the failure rate, and identified that temperature and wind speed had a significant impact on the failure rate.

[0189] For example, historical operating data of transformers are collected, including parameters such as load current, oil temperature, vibration frequency, etc. Finally, load current and oil temperature are selected as key operating parameters because they have strong explanatory power and robustness for equipment failures.

[0190] For example, historical failure data of power grid equipment in a certain area and geological disaster data of the same period are collected, including earthquake frequency, rainfall, geological structure, etc. Finally, earthquake frequency and rainfall are selected as key geological disaster factors because they have strong explanatory power and robustness for power grid failures.

[0191] In step S103, the correlation between the power grid fault and its key influencing factors is quantitatively analyzed, and the improved anti-entropy-grey correlation analysis method is used to determine the correlation between the two. The improved anti-entropy-grey correlation analysis method is a method that aims to improve the accuracy and objectivity of traditional grey correlation analysis, especially in the scenario of dealing with the correlation analysis between power grid faults and their multiple key influencing factors.

[0192] The reasons for introducing anti-entropy are as follows: Enhance objectivity: Traditional grey correlation analysis adopts equal weighting, that is, it considers all key influencing factors to be equally important, which may lead to the neglect of key influencing factors' information. Anti-entropy method quantifies the information of key influencing factors and automatically adjusts the weights of key influencing factors, so that the analysis results can better reflect the actual situation; Reveal internal associations: Anti-entropy can highlight those key influencing factors with large variations and rich information, so as to give them more attention in correlation analysis, which is helpful to discover more refined correlation patterns; Reduce bias: When the correlation between key influencing factors is complex and scattered, anti-entropy-grey correlation analysis can effectively avoid the analysis bias caused by local correlation and improve the overall accuracy of the analysis.

[0193] The basic steps of the improved anti-entropy-grey correlation analysis method are as follows: Data preprocessing: Collect data on power grid faults and their related key influencing factors, and perform necessary data cleaning, standardization or normalization to ensure the comparability between key influencing factors. Calculate the anti-entropy of key influencing factors: Calculate the entropy value of each key influencing factor to reflect the uncertainty of its information. Through the anti-entropy formula conversion, the higher the anti-entropy value, the more certain the information provided by the key influencing factor, the greater the difference, and the higher the weight should be given. Weight determination: Weight the key influencing factors based on the anti-entropy value. The weight reflects the relative importance of each key influencing factor in the analysis, overcoming the subjectivity of equal treatment. Construct reference sequence and comparison sequence: Usually, faults are selected as reference sequences, and other key influencing factors are selected as comparison sequences to construct a grey system model. Calculate the correlation coefficient and correlation degree: Apply the weighted difference sequence to calculate the correlation coefficient between each comparison sequence and the reference sequence, taking into account the weight determined by anti-entropy. Calculate the grey correlation degree of each comparison sequence to obtain the correlation matrix.

[0194] In order to avoid the problem of weight imbalance caused by the excessive sensitivity of key influencing factors in the entropy weight method, the anti-entropy value method is used to objectively weight the key influencing factors. The specific steps are as follows:

[0195] (1) Construct the evaluation matrix of key influencing factors. Select n key influencing factors, k regions, and m years of data to establish the corresponding initial matrix, x ij is the value of the jth key influencing factor of the ith object, and the key influencing factor matrix X is evaluated. m for

[0196]

[0197] (2) Calculate the anti-entropy of each key influencing factor. Define the anti-entropy h′ as

[0198]

[0199] Among them, pj is the entropy value of the jth key influencing factor, 0≤p j ≤1, and

[0200] The anti-entropy h′ of each key influencing factor is

[0201]

[0202] Where: c ij is the entropy value of the jth key influencing factor under the i-th object,

[0203] (3) Normalization. The anti-entropy of each key influencing factor is further normalized to obtain:

[0204]

[0205] Where: d j is the objective weight of the jth key influencing factor.

[0206] In order to scientifically and reasonably evaluate the impact of each key influencing factor on distribution network failure, the key influencing factors are combined and weighted as follows:

[0207]

[0208] Where: w j is the combined weight of the key influencing factor j; c j d j are the objective weight (objective weight calculated according to entropy value) and subjective weight of the jth key influencing factor, respectively. In power grid fault analysis and prediction, the combined weights of key influencing factors are crucial for accurately assessing the correlation between these factors and faults. The combined weight is the result of a reasonable integration of the subjective weight and objective weight of each key influencing factor, which can more comprehensively reflect the importance of each factor. The subsequent calculation of the correlation depends on these combined weights, reflecting the relationship and interaction between the key influencing factors.

[0209] The fuzzy mathematics comprehensive evaluation method uses the ascending semi-trapezoidal distribution function as the membership function to fit and score each key influencing factor. In order to further quantitatively analyze the correlation between the fault and its key influencing factors, the improved anti-entropy-grey correlation analysis method is used to determine the correlation between the two. Traditional grey correlation analysis establishes a correlation matrix by obtaining a differential sequence, and generally adopts equal weighting. This equal weighting method easily conceals the characteristic relationship between each key influencing factor and ignores the rich information contained in the key influencing factors themselves. Since the key influencing factors come from multiple aspects and levels, when the correlation relationship is relatively discrete, a trend of local correlation is likely to appear, resulting in analysis deviations. Therefore, anti-entropy is introduced on the basis of the entropy method to overcome the existing subjective problems. The specific steps are as follows. The calculation formula for the correlation between sequences in the traditional grey correlation analysis method is:

[0210]

[0211] Where: γ j is the correlation between power grid fault and key influencing factor j; rj is the correlation coefficient between the power grid fault and the key influencing factor j determined based on the combined weight, and the fixed weight value is 1 / n.

[0212] Define the residual degree of the jth key influencing factor as

[0213] T j =1-e j ;

[0214] Where: T j is the residual degree of the jth key influencing factor; e j is the ratio of the anti-entropy of the jth key influencing factor to the maximum entropy, that is, the relative entropy. The specific expression is

[0215]

[0216] Where: E j 、E max are the anti-entropy and maximum entropy of the jth key influencing factor respectively. j The larger the value is, the more obvious the characteristics of the key influencing factor are and the higher its importance is.

[0217] The proportional weight of other key influencing factors r is set to b(r), and the specific calculation formula is:

[0218]

[0219] f(E r ,e r )=E r (1-e r );

[0220] Among them, T r is the residual degree of a key influencing factor r; f(E r ,e r ) is a function of anti-entropy and relative entropy; replacing 1 / n in the formula with b(r), the correlation γ between the power grid fault and the key influencing factor j is obtained j The final expression is

[0221]

[0222] In step S104, a rainfall-induced geological disaster warning model is constructed, and the probability of geological disasters and key influencing indicators are used as input parameters to construct a power grid fault prediction model of an improved neural network.

[0223] Geological disaster prediction model: By establishing a geological model and conducting numerical simulation to analyze the deformation and movement characteristics of geological disasters, the displacement of geological disasters under different rainfall conditions is obtained as an intensity indicator; historical disasters and previous rainfall conditions are statistically analyzed, and combined with the results of regional geological disaster susceptibility zoning, short-term meteorological warning criteria for geological disasters at transmission line tower bases are established, and a geological disaster risk warning method based on complex weather elements is proposed.

[0224] 1. Establish geological model and numerical simulation. Geological model: Construct a geological model of the tower base area to simulate the geological deformation and movement characteristics under different rainfall conditions. Through numerical simulation, analyze the displacement of the tower base under different rainfall amounts and durations, and obtain the geological disaster displacement as an intensity indicator.

[0225] 2. Analyze the susceptibility zoning of geological disasters. Regional susceptibility zoning: Based on geological data and historical disaster data, the tower base area is divided into high, medium and low susceptibility zones. Combined with the numerical simulation results of the geological model, the specific scope and characteristics of each zone are determined.

[0226] 3. Establish meteorological warning criteria. Rainfall intensity criteria: Analyze historical data to determine the rainfall intensity threshold that triggers geological disasters. For example, if a landslide disaster in the past occurred when the rainfall in 24 hours exceeded 100 mm, set 100 mm as the rainfall intensity threshold. Rainfall duration criteria: Count the duration of rainfall when the disaster occurs. For example, if multiple landslide disasters occur when continuous rainfall exceeds 48 hours, set 48 hours as the rainfall duration threshold. Combined criteria: Establish comprehensive criteria, such as a combination of rainfall intensity and duration. For example, when the rainfall in 24 hours exceeds 80 mm and the rainfall duration exceeds 24 hours, warn of the risk of geological disasters.

[0227] 4. Meteorological warning method. The early warning system establishes a real-time monitoring system to monitor meteorological data (rainfall, rainfall intensity, duration, etc.). The system compares the current meteorological data with the early warning criteria in real time, and issues an early warning signal when the early warning conditions are met. Early warning level: Different early warning levels (such as yellow warning, orange warning, and red warning) are set according to the degree of deviation between the meteorological data and the criteria. For example, a yellow warning is issued when the rainfall approaches or reaches the threshold, an orange warning is issued when it exceeds the threshold by 50%, and a red warning is issued when it exceeds the threshold by 100%.

[0228] Improved neural network: The K-fold-GA-BP neural network method has become a powerful tool in the field of power grid fault prediction due to its high adaptability, optimization ability and prediction accuracy. Integrated optimization: This method combines three powerful technologies—K-fold cross validation, genetic algorithm and BP neural network—to form a highly optimized prediction model. K-fold cross validation improves the reliability and generalization ability of model evaluation; the genetic algorithm automatically optimizes the network structure and parameters, avoiding the tediousness and subjectivity of manual parameter adjustment; the BP neural network is good at learning nonlinear problems and can better capture complex data relationships. Reduce overfitting: K-fold cross validation splits the data for multiple training and verification to ensure that the model is not only overfitted to the training data, but can also perform well on unseen data, thereby enhancing the versatility of the model. Parameter optimization: As a global optimization technology, genetic algorithm (GA) can efficiently explore the parameter space of BP neural network (such as the number of hidden layer nodes, learning rate, etc.), find the optimal or near-optimal network configuration, thereby improving prediction accuracy and training efficiency, and avoiding the blindness and inefficiency of traditional trial and error methods. Strong adaptability: BP neural network has good nonlinear mapping ability and can handle complex and nonlinear data relationships that may exist in fault prediction. It can provide relatively accurate predictions even in the face of a large number of input variables and complex interactions. Automation and flexibility: The entire process from feature selection to model optimization is highly automated, reducing manual intervention. At the same time, the introduction of genetic algorithms increases the flexibility of model adjustment, allowing the model to be quickly adjusted according to different data sets and prediction requirements.

[0229] (1) Improvements in K-fold cross validation Using K-fold cross validation instead of a single training set and validation set division makes full use of data and provides robust model evaluation. Technical effect: Improves the generalization ability of the model: avoids overfitting and underfitting, and provides more robust performance evaluation. Makes full use of data: Each data point has the opportunity to be used as a training set and a validation set, which improves data utilization.

[0230] (2) Improvements in Genetic Algorithm Optimization GA is used to optimize the initial weights and hyperparameters of the neural network and enhance the optimization capability of the network. Technical Effect: Strong global optimization capability: GA can jump out of the local optimum and find a better initial weight and hyperparameter configuration. Automated parameter selection: Reduces the workload of manual parameter adjustment and improves the efficiency and effect of parameter selection.

[0231] (3) Improvements in BP neural network Combining K-fold cross validation and GA optimized BP neural network enhances the network's adaptability and prediction accuracy. Technical effect: Improved prediction accuracy: Combining GA optimization and cross validation improves the prediction accuracy of BP neural network. Strong adaptability: The network structure and parameters can be adaptively optimized to adapt to different data sets and problems.

[0232] The duration of the rainfall event that induces geological disasters is recorded as D, the cumulative rainfall during this period is recorded as E, and the ratio of E to D is the average rainfall intensity. The principle is to depict the previous rainfall E and the duration condition D of the geological disaster in double logarithmic coordinates, and carry out fitting analysis on the data points of the rainfall event that induces geological disasters.

[0233] E=c+α×D β ;

[0234] Where E is the cumulative rainfall (mm); D is the duration of rainfall, and α, β, and c are statistical parameters. However, the process of rainfall events inducing geological disasters is not an average effect. It is generally believed that only the rainfall on the day of the geological disaster plays a full role in the occurrence of geological disasters, and the rest is considered to be only partially effective for geological disasters. Therefore, effective rainfall is usually used to characterize the threshold index of rainfall-induced geological disasters. This application mainly uses the power exponential formula to calculate the corresponding effective rainfall.

[0235] R c =R0+αR1+α 2 R2+...+α n R n ;

[0236] Among them, R c is the effective rainfall, R0 is the rainfall on the day of the geological disaster, and R n is the rainfall n days before the disaster. α is the attenuation coefficient, and n is the number of days.

[0237] Among them, the risk calculation method of geological disasters at the transmission line tower base is as follows.

[0238] R = H × V;

[0239] Among them, R is the geological disaster risk, H is the hazard probability analysis, and V is the vulnerability analysis.

[0240] Based on Bayesian theory, H is obtained by combining the ED rainfall threshold model and the rainfall threshold exceedance probability calculation. The time probability of rainfall-induced geological disasters is divided into two parts, namely the probability of a certain rainfall condition occurring and the probability of geological disasters occurring under the rainfall condition. First, the rainfall corresponding to different extreme value recurrence conditions in a certain area is determined by statistical methods; then, the rainfall threshold for inducing geological disasters in each local area is obtained by statistical analysis, and the probability of geological disasters occurring under rainfall conditions exceeding the threshold is analyzed by statistical analysis of historical geological disaster events; the time probability of geological disasters is obtained by combining the two.

[0241] Assume that the rainfall event is A and the geological disaster event is B, P[A] and P[B|A] are the probability of the rainfall event and the geological disaster under the rainfall conditions respectively. The time probability P(AB) of rainfall-induced geological disasters in Bayesian conditional probability calculation is expressed as:

[0242] P(AB)=P[A]·P[BA];

[0243] In the exceedance probability calculation determined based on the ED effective rainfall threshold model, after obtaining the geological disaster threshold of each research unit, combined with its historical rainfall records, the number of rainfall events that exceeded the threshold rainfall conditions in history, as well as the number of geological disasters induced by such events, are analyzed. The two can be combined to obtain the probability of geological disasters P[B|A]. The probability distribution of geological disasters over the same long period of time can be evaluated using binomial distribution and Poisson distribution. When the time length is a discrete time interval of one year, the average annual probability of geological disasters is related to the time period T of the occurrence of geological disasters under this condition. The annual probability of geological disasters can be expressed as

[0244]

[0245] Where T is the occurrence period of geological disaster events, and λ is the frequency of geological disasters. This method is also applicable to calculating the probability of geological disasters in a given time period. Let P0 be the probability of no geological disasters occurring in a given time period (within 1 year). Then P 0t is the probability that no geological disaster occurs within t years, so the probability of geological disasters occurring in many years can be expressed as:

[0246]

[0247] Among them, e is a natural constant, and λ is the frequency of geological disasters under limited rainfall conditions. The probability of one or more geological disasters occurring in the next t years is:

[0248] P=1-e- λt ;

[0249] The risk probability V of qualified disasters is mainly calculated through the Monte Carlo simulation method. Based on the statistical sampling theory, the Monte Carlo method studies the numerical calculation method of generating random variables through computers, which is also called random simulation method or statistical experimental method.

[0250] First, establish a function to represent the stable state of geological hazards:

[0251]

[0252] All parameters are random variables that control the stability of geological disasters and are determined by factors such as the rock and soil properties and deformation mechanism of geological disasters. They are defined as rainfall, single maximum rainfall, rainfall duration, and rock and soil type. It is generally believed that these parameters have a certain range of values, and they are random within the range of values. The reasons for this randomness include the heterogeneity of geological bodies, changes in external environmental factors, and measurement errors.

[0253] Fs represents the stability coefficient of geological disasters. When Fs≤1, geological disasters are considered to be in an unstable state. In N groups of random numbers, there are M times when Fs≤1 occurs. When N is large enough, it is considered that the frequency of Fs≤1 is close to the probability. The dangerous probability of geological disasters is expressed as follows:

[0254]

[0255] When N is large enough, its mean μ F and standard deviation σ F They are:

[0256]

[0257]

[0258] The risk probability of geological disasters is:

[0259] P f =1-Φ(β).

[0260] The structure of BP neural network includes input layer, hidden layer and output layer. The input parameters enter the input layer, then pass through the hidden layer, and finally output by the output layer. The steps of fault prediction using K-fold-GA-BP neural network method are as follows.

[0261] 1. Design of input and output layers. According to the design characteristics of BP neural network, without considering the mutual influence relationship between various factors, the constituent factors can be selected as the input layer, and the output layer is the power grid failure rate.

[0262] 2. Selection of target value: Based on the historical fault data, the method of processing data in horizontal and vertical directions is used to estimate the probability of power grid failure as the output target value.

[0263] 3. Model establishment.

[0264] Genetic algorithm part: Use GA to initialize the population, obtain the initial weights and thresholds of the BP neural network, and substitute the individuals after population initialization into the BP neural network to iterate and find the optimal initial weights and thresholds.

[0265] (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.

[0266] (2) The 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:

[0267]

[0268] 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.

[0269] (3) According to the individual fitness results calculated by the above formula, select the better individuals to form a new population. The calculation method is:

[0270]

[0271] 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.

[0272] (4) Randomly select two parents from the population k and P l , under a certain probability, the real number crossing mode of the jth point is:

[0273]

[0274] Among them, α is a random number between 0 and 1.

[0275] (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:

[0276]

[0277] Among them, P max , P minare 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.

[0278] (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, and the learning and training times are repeated to adjust the neural network parameters to within the set accuracy range or until the learning times are met.

[0279] 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.

[0280] (1) Initialize the network.

[0281]

[0282] Among them, X is the neuron input vector, which is the factor constituting the power grid fault; n is the number of input layer nodes; Y is the neuron output vector, which is the maintenance and operation cost; m is the number of output layer nodes.

[0283] (2) Calculate the hidden layer output. The hidden layer output value H a The calculation method is:

[0284]

[0285] 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.

[0286] (3) Calculate the output of the output layer. Output layer output value H b The calculation method is:

[0287]

[0288] Among them, ω jk is the weight between the hidden layer and the output layer; b is the threshold value.

[0289] (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:

[0290] e k =Y k -H bk ;

[0291] Let e0 be the maximum allowable error value. If max(e k)>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.

[0292] (5) Update weights and thresholds. Update ω based on the error ij and ω jk :

[0293]

[0294] ω jk =ω jk +ηH aj e k ;

[0295] Where η is the learning rate, which is between 0 and 1. The threshold correction is:

[0296]

[0297] λ b =λ b +e k ;

[0298] 4. Model calculation. The model calculation uses MATLAB software. During the calculation process, the existing data of power grid fault information is used as samples and the GA-BP neural network is used for training. In order to avoid overfitting and improve the accuracy of the GA-BP neural network model, the K-fold cross-validation method is used for the training data to determine the accuracy of the model.

[0299] 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.

[0300] Mainly adopt e MSE Evaluate the impact of K-fold cross validation on model accuracy. The calculation formula is:

[0301]

[0302] 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:

[0303]

[0304] 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, we take the e after multiple runs. MAPE 、eRMSE The standard deviation is used as the final test indicator.

[0305] In step S105, based on the fault prediction model, an optimization model for the optimal maintenance position is constructed, and the optimal maintenance position is obtained by using a niche genetic algorithm in combination with the objective function and constraint setting.

[0306] The artificial immune optimization algorithm is an intelligent algorithm inspired by the human immune system, which has the ability to learn, remember and self-adapt. In the process of selecting the location of the maintenance station, the artificial immune optimization algorithm selects, crosses and mutates antibodies according to the affinity of individuals by analogy with the diversity in the immune system, ensuring that the algorithm converges quickly and finally obtains the optimal solution for the maintenance station. The process of using the artificial immune optimization algorithm to solve the problem of maintenance station distribution and maintenance task allocation is as follows: Figure 1 As shown, the specific steps are as follows:

[0307] 1. Determine the objective function and constraints of the maintenance station optimization model.

[0308] 2. Generate the initial value set of the location coordinates of the existing distribution network fault maintenance station. Suppose the k-th iteration maintenance station coordinate set is in Represents the position coordinates of the mth maintenance station.

[0309] 3. According to the initial value set, objective function and constraints, calculate and save the optimal location coordinates of the existing distribution network fault maintenance station.

[0310] 4. According to the predicted number of iterations, determine whether the termination condition is met. If the termination condition is met, the optimal solution of the maintenance station point is obtained, and the optimization of the location of the maintenance station point for the existing distribution network fault is completed; otherwise, the initial value set of the maintenance station point location coordinates is updated and return to step 3.

[0311] In order to reduce the total time of fault maintenance and improve maintenance efficiency, it is necessary to set the feeder fault coefficient according to the predicted feeder fault level to ensure that the maintenance station that meets the constraint conditions is close to the feeder with a higher fault level. The feeder is regarded as a point in the plane, and the distance between the maintenance station and the feeder is regarded as the distance between the two points. Assume a j (j=1,2,...,n) is the feeder equivalent point, x i (i=1,2,...,m) is the maintenance station, where m and n are the number of maintenance stations and the number of feeder equivalent points respectively. The distance d from the maintenance station to the feeder ij The calculation formula is:

[0312]

[0313] Where: x i-a j For x i and a j The Euclidean distance of j is the feeder equivalent a j Failure coefficient.

[0314] Considering the fault repair process comprehensively, the road congestion coefficient, road tortuosity coefficient, and feeder tortuosity coefficient are set to a, λ, and r respectively, and the average speed of the repair vehicle is set to v. The time T from the repair station to the repair location is ij The calculation formula is:

[0315] T ij =(d ij ·a·λ·r) / v;

[0316] Assume that the total time it takes for the maintenance team to arrive at the maintenance site is T max , and the time to reach the urban area is stipulated to be no more than 40 minutes (i.e. T max =40min), the longest time to reach the suburbs is no more than 60min (i.e. T max = 60min). Assume that the maximum time required to reach the feeder from the maintenance station is T M , let the time taken by the service center to process the work order, the time taken by the dispatch layer to locate and handle the fault, the time taken to prepare the repair vehicle, and the time taken to find the repair location be t1, t2, t3, and t4 respectively, then T M The calculation formula is:

[0317] T M =T max -t1-t2-t3-t4;

[0318] After obtaining the time required for the maintenance journey, it is necessary to divide the feeder and the fault maintenance station. ij is the condition that the standard range of fault maintenance time is satisfied. If the time from the i-th maintenance station to the j-th feeder is within the standard range of fault maintenance time, then b ij =1, otherwise b ij =0,a ij represents the distance from feeder i to maintenance station j. Assuming that a feeder in this model belongs to only one maintenance station, the shortest distance required from the specific feeder to the maintenance station can be expressed as:

[0319]

[0320] The constraints of the maintenance station location optimization model are as follows:

[0321] (1)T ij <T M , which represents the time constraint required to rush from the maintenance station to the fault location.

[0322] (2) Represents the constraints for arranging all feeder maintenance work at m maintenance stations.

[0323] Niche Genetic Algorithm (NGA) is an improved algorithm developed on the basis of traditional genetic algorithm, which is particularly suitable for solving multi-objective optimization problems. It introduces the concept of "niche" to imitate the survival and competition mechanism of biological populations in nature under different ecological environments to maintain the diversity of populations and avoid premature convergence, so as to more effectively search for the Pareto frontier solution of multi-objective optimization problems. The specific solution process is as follows:

[0324] 1. Input system data, parameter values ​​and variable values. The input microhabitat genetic algorithm parameters include the size of the individual population, the number of genes contained in the individual, the crossover rate, the mutation rate and the maximum genetic generation.

[0325] 2. Use binary to encode chromosomes.

[0326] 3. Randomly generate the initial population household seal.

[0327] 4. Verify constraints.

[0328] 5. Calculate the fitness function value of each chromosome in the population. In order to minimize the total risk of the objective function described and adapt to the characteristics of the genetic algorithm to seek the maximum value, the fitness function is:

[0329]

[0330] F=R+μW;

[0331] Among them, GF is the fitness function; F is the correction value of the objective function; F max is a given larger constant, which can be selected from the maximum value of the objective function F obtained so far or obtained based on test data. W is the network overload capacity (MW).

[0332] 6. Arrange each chromosome in descending order according to its fitness function value, and remember the first LK chromosomes (M<LK).

[0333] 7. Perform proportional selection operation on the population P(tv) to obtain P′(tv). The fitness function value proportional method is used as the selection strategy to implement the selection operation.

[0334] 8. Perform a single-point crossover operation on P′(tv) to obtain P″(tv). The crossover operation can change the topological structure of the two networks, thereby obtaining a new structure. The basic steps to achieve crossover are:

[0335] (1) Select the parent chromosome: The process is: repeat the following process from ai=1 to LK: generate a random number r from [0,1] ai , if r ai ≤P C , then select chromosome ai as the parent for crossover. C is the crossover rate, usually between 0.6 and 0.9, reflecting the idea of ​​information exchange in nature.

[0336] (2) Selecting the crossover point: Using one-point crossover, the specific operation is to randomly select a crossover point in a chromosome, and the two chromosomes are partially exchanged before or after this point to produce a new chromosome.

[0337] 9. Perform two-point mutation operation on P″(tv) to obtain P″′(tv).

[0338] 10. Calculate the fitness of P″′(tv).

[0339] 11. Perform the niche elimination algorithm. Combine the LK chromosomes obtained in step 9 and the M chromosomes memorized in step 6 to obtain a new population containing (LK+M) chromosomes. For these (LK+M) chromosomes, two chromosomes are obtained according to the following calculation formula: X ai and X aj The Hamming distance between:

[0340]

[0341] When X ai -X aj <DL (L is the specified distance), compare chromosome X ai and X aj The fitness of the chromosomes is calculated by the penalty function Fmin(X ai ,X aj )=Penalty.

[0342] 12. Arrange the new fitness of (N+M) chromosomes in descending order and remember the first M chromosomes.

[0343] 13. Implement the optimal preservation strategy.

[0344] 14. Determine whether the termination condition is met. If the termination condition is met, go to step 15, otherwise update the evolutionary generation counter Gen=Gen+1, take the first N chromosomes in the arrangement of step (11) as the new next generation population and go to step 7.

[0345] 15. Decode and output the optimization calculation results. Output several chromosomes with the highest or second highest fitness function values, and decode them to restore them to the optimal position points.

[0346] refer to Figure 3 Another specific embodiment of the present application discloses an optimal maintenance location determination device for power grid fault prediction, including: a data acquisition module 301, used to obtain historical fault data and related influencing factor data, wherein the related influencing factor data includes power grid operation data, meteorological data and geographic information data; an interpolation module 302, used to perform outlier interpolation processing on the historical fault data and the related influencing factor data through an adversarial network; a primary selection module 303, used to select multiple key influencing factors from the related influencing factor data by fusing the Bayesian model with the hierarchical hybrid model and using statistical information; a secondary selection module 304, used to select multiple key influencing factors by using an improved anti-entropy-grey correlation analysis method; An analysis method is used to determine the correlation between power grid faults and each key influencing factor, so as to select some key influencing factors from multiple key influencing factors as influencing indicators according to the correlation; a fault prediction model 305 is used to construct a BP neural network and use historical fault data and influencing indicators for training to obtain a power grid fault prediction model, wherein the geological disaster probability and the influencing indicator are used as input parameters of the power grid fault prediction model, and a rainfall-induced geological disaster early warning model is constructed to obtain the geological disaster probability; and an optimal position acquisition module 306 is used to construct an optimal maintenance position optimization model based on the predicted power grid fault, and combine the objective function with the constraint conditions to obtain the optimal maintenance position.

[0347] 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.

[0348] 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 determining the optimal maintenance location for power grid fault prediction, It is characterized in that include: Acquire historical fault data and relevant influencing factor data, and perform outlier interpolation processing on the historical fault data and the relevant influencing factor data through an adversarial network, wherein the relevant influencing factor data includes power grid operation data, meteorological data, and geographic information data; Selecting a plurality of key influencing factors from the relevant influencing factor data by fusing the Bayesian model with the hierarchical hybrid model and utilizing statistical information; An improved anti-entropy-grey correlation analysis method is used to determine the correlation between the power grid fault and each key influencing factor, so as to select some key influencing factors from the multiple key influencing factors as influencing indicators according to the correlation; Constructing a BP neural network and using the historical fault data and the influencing index to train to obtain a power grid fault prediction model, wherein the probability of geological disasters and the influencing index are used as input parameters of the power grid fault prediction model, and constructing an early warning model of geological disasters caused by rainfall to obtain the probability of geological disasters; and An optimal maintenance location optimization model is constructed based on the predicted power grid faults, and the optimal maintenance location is obtained by combining the objective function with the constraints.

2. The method for determining the optimal maintenance location for power grid fault prediction according to claim 1, It is characterized in that Performing outlier interpolation processing on the historical fault data and the related influencing factor data through the adversarial network further includes: Generate adversarial reconstruction samples through the generator; Distinguishing the distance between the adversarial sample and the original sample by a discriminator, wherein the original sample is any sample data of the historical fault data and the relevant influencing factor data; The distance between the reconstructed sample G(z) and the original sample x is reduced by using a reconstruction loss function, wherein the reconstruction loss function is expressed by the following formula: L r (z)=||x·m-G(z)·m|| 2 ; The output probability of the discriminator is maximized by using a discriminant loss function to ensure that the reconstructed sample G(z) conforms to the distribution of the original sample, wherein the discriminant loss function is expressed by the following formula: L d (z)=-D(G(z)); A filling loss function is generated according to the reconstruction loss function and the discrimination loss function, and the filling loss function is minimized based on the reconstructed sample and the original sample to generate a missing value for interpolation. The filling loss function and the missing value are respectively expressed by the following formulas: L imputation (from)=L r (z)+λ L d (With); x imputed =x·m+(1-m)·G(z * ); Wherein, x represents the original sample; m is a mask matrix used to indicate the missing parts of the historical fault data and the related influencing factor data; G(z) represents the reconstructed sample, L r (z) represents the distance between the generated sample and the original sample; L d (z) represents the discriminant loss function, which is used to evaluate the output probability of the discriminator; D(G(z)) is the output of the discriminator, which represents the probability that the reconstructed sample G(z) is discriminated as the original sample; L imputation (z) is a comprehensive loss function, which consists of a reconstruction loss function and a discrimination loss function; λ is a hyperparameter used to control the ratio between the reconstruction loss and the discrimination loss; x imputed represents the padded data sample; x·m represents the non-missing part of the original sample; (1-m)·G(z) represents the corresponding missing part of the reconstructed sample; G(z * ) represents filling the missing values ​​of the original sample.

3. The method for determining the optimal maintenance location for power grid fault prediction according to claim 1, It is characterized in that Selecting a plurality of key influencing factors from the relevant influencing factor data by fusing the Bayesian model with the hierarchical hybrid model and utilizing statistical information further includes: Define the model space step: construct multiple regression models according to the historical fault data and the relevant influencing factor data; Specifying prior distribution step: assigning appropriate prior distribution to regression model parameters based on the historical failure data and domain expert knowledge; Calculating the posterior probability step: using the Bayesian theorem combined with the historical fault data to calculate the posterior probability of each regression model; Model averaging step: performing weighted averaging on the prediction results of the regression models according to the posterior probability of each regression model; Inference and prediction steps: Use the weighted average results of the regression model to estimate parameters, test hypotheses, or predict future observations; The improved anti-entropy-grey correlation analysis method is used to determine the correlation between power grid failure and each key influencing factor, further including: Calculating an entropy value for each of the multiple key influencing factors; The entropy value of each key influencing factor is converted into an anti-entropy value by using an anti-entropy formula, wherein the higher the anti-entropy value, the more certain the information provided by the key influencing factor and the greater the difference, so as to be given a greater weight; Generate an objective weight based on the anti-entropy value, and calculate a combined weight based on the objective weight and the subjective weight; Calculating the correlation coefficient between each comparison sequence and the reference sequence based on the combined weight, wherein the historical fault data is selected as the reference sequence and the multiple key influencing factors are used as multiple comparison sequences; and The correlation degree between the reference sequence and each comparison sequence is calculated based on the correlation coefficient as the correlation degree between the power grid fault and each key influencing factor.

4. The method for determining the optimal maintenance location for power grid fault prediction according to claim 3, It is characterized in that The improved anti-entropy-grey correlation analysis method is used to determine the correlation between power grid failure and each key influencing factor, further including: The matrix X of key influencing factors is expressed by the following formula m for: Among them, x ij For the value of the jth key influencing factor of the ith object, n key influencing factors, k regions, and m years of data are selected from the plurality of key influencing factors; Based on the entropy values ​​of multiple key influencing factors, the anti-entropy of the following key influencing factors is calculated: Among them, c ij is the entropy value of the jth key influencing factor under the i-th object; The anti-entropy of each key influencing factor is normalized to obtain the objective weight of the corresponding anti-entropy: Among them, d j is the objective weight of the jth key influencing factor; Based on the objective weights and subjective weights, the following combined weights are calculated: Among them, w j is the combined weight of the key influencing factor j; c j ,d j are the objective weight and subjective weight of the jth key influencing factor respectively; The correlation degree γ between the power grid fault and the key influencing factor j is calculated by the following formula j : T r =1-e r ; f(E r ,And r )=And r (1-e r ); Among them, γ j is the correlation between the power grid fault and the key influencing factor j; rj is the correlation coefficient between the power grid fault and the key influencing factor j determined based on the combined weight; T r is the residual degree of the key influencing factor r; e r is the relative entropy, which represents the ratio of the anti-entropy of the key influencing factor r to the maximum entropy; E r 、E max are the anti-entropy and maximum entropy of the key influencing factors respectively; b(r) is the proportional weight of other key influencing factors r, f(E r ,e r ) is a function of anti-entropy and relative entropy.

5. The method for determining the optimal maintenance location for power grid fault prediction according to claim 1, It is characterized in that Constructing an early warning model for geological disasters caused by rainfall to obtain the probability of the geological disasters further includes: When the time length is a discrete time interval of one year, the average annual probability of geological disasters is related to the time period T of the occurrence of geological disasters, where the annual probability of geological disasters is expressed by the following formula: Among them, T is the occurrence period of geological disaster events, and λ is the frequency of geological disasters; The probability of geological disasters occurring over many years is expressed by the following formula: Where e is a natural constant, and λ is the frequency of geological disasters under specified rainfall conditions; The probability of one or more geological disasters in the next t years is expressed by the following formula: P=1-e- λt 。 6. The method for determining the optimal maintenance location for power grid fault prediction according to claim 5, It is characterized in that Building a neural network power grid fault prediction model further includes: Normalizing the historical fault data, the power grid operation data, the meteorological data, and the geographic information data; Dividing the historical fault data and the corresponding influencing indicators into multiple data subsets, wherein one subset of the multiple data subsets is used as a validation set, and the remaining subsets are used as training sets; Before training the neural network, a genetic algorithm is used to optimize the network structure and initial weights; According to the network structure and initial weights optimized by the genetic algorithm, the BP neural network is constructed, and the BP neural network includes an input layer, a hidden layer and an output layer, wherein the number of the input layer is determined according to the influencing index and the probability of geological disasters, the number and structure of the hidden layer are determined according to the genetic algorithm, and the output of the output layer is the failure rate prediction value; Using the training set, applying a back propagation algorithm to adjust the network weights and thresholds of the BP neural network to obtain the power grid fault prediction model; Calculating the prediction error of the trained BP neural network using the verification set; and The real-time impact index is input into the power grid fault prediction model to predict the power grid failure rate in real time.

7. The method for determining the optimal maintenance position for power grid fault prediction according to claim 6, It is characterized in that According to the network structure and initial weights optimized by the genetic algorithm, constructing the BP neural network further includes: The initialized BP neural network is expressed by the following formula: Among them, X is the neuron input vector; n is the number of input layer nodes; Y is the neuron output vector, and m is the number of output layer nodes; The output value H of the hidden layer is expressed by the following formula: aj : Wherein, p is the number of nodes in the hidden layer; ω ij is the weight between the input layer and the hidden layer; a is the threshold; f is the activation function; The output value H of the output layer is expressed by the following formula bk : Among them, ω jk is the weight between the hidden layer and the output layer; b is the threshold value; The predicted output H is expressed by the following formula bk With the expected output Y k The error between k : e k =|Y k -H bk |; Among them, when max(e k )>e 0 When , the BP neural network parameters are adjusted and learning and training are performed again until all training data meet the specified error standard, where e 0 is the maximum allowable error value; Update the weight ω between the input layer and the hidden layer according to the error i j and the weight ω between the hidden layer and the output layer jk : oh jk =ω jk +ηH aj e k ; Among them, η is the learning rate, which is between 0 and 1; The threshold is corrected by the following formula: l b =λ b +e k 。 8. The method for determining the optimal maintenance location for power grid fault prediction according to claim 1, It is characterized in that Building an optimal maintenance location optimization model based on predicted power grid faults, combining the objective function with constraints to obtain the optimal maintenance location further includes: Determine the objective function and constraint conditions of the maintenance station point optimization model, and set the initial value set of the distribution network fault maintenance station point location coordinates; Calculating the optimal position coordinates of the distribution network fault maintenance station point according to the initial value set, the objective function and the constraint conditions; According to the predicted number of iterations, it is judged whether the termination condition is met. When the termination condition is met, the optimal solution of the maintenance station point is obtained, and the optimization of the location of the maintenance station point of the distribution network fault is completed; when the termination condition is not met, the initial value set of the maintenance station point location coordinates is updated; The distance d from the maintenance station to the feeder is calculated by the following formula ij : Among them, ||x i -a j || is x i and a j The Euclidean distance of j is the feeder equivalent point a j Failure coefficient; a j (j=1,2,...,n) is the feeder equivalent point, x i (i=1,2,...,m) is the maintenance station point, where m and n are the number of maintenance stations and the number of feeder equivalent points respectively; The time T from the maintenance station to the maintenance location is calculated by the following formula ij The calculation formula is: T ij =(d ij ·a·λ·r) / v; Among them, a, λ, r are the road congestion coefficient, road tortuosity coefficient, and feeder tortuosity coefficient, respectively, and v is the average speed of the maintenance vehicle; The maximum time T required to reach the feeder from the maintenance station is calculated by the following formula M : T M =T max -t 1 -t 2 -t 3 -t 4 ; t 1 ,t 2 ,t 3 ,t 4 They are the time taken by the service center to process the work order, the time taken by the dispatch layer to locate and handle the fault, the time taken to prepare the repair vehicle, and the time taken to find the repair location; The shortest distance required from a specific feeder to the maintenance station is calculated using the following formula: In this model, it is assumed that a feeder belongs to only one maintenance station, b ij is the condition that the standard range of fault maintenance time is satisfied. If the time from the i-th maintenance station to the j-th feeder is within the standard range of fault maintenance time, then b ij =1, otherwise b ij =0,a ij represents the distance from feeder i to maintenance station j; Constraints of the maintenance station location optimization model: The time required to travel from the maintenance station to the fault location is constrained as: T ij <T M ; The constraint condition for arranging all feeder maintenance work at m maintenance stations is:

9. The method for determining the optimal maintenance position for power grid fault prediction according to any one of claims 1 to 8, It is characterized in that Acquiring historical fault data and related influencing factor data further includes: The historical fault data includes fault occurrence time, fault type, fault location, fault duration, fault impact range, fault repair time, repair cost, fault cause and equipment information; The grid operation data includes current monitoring data, voltage monitoring data, power factor, load change data, transformer oil temperature, transformer oil level, insulation resistance test record facility inspection records and maintenance and repair logs; The meteorological data include rainfall, temperature, humidity, wind speed and direction, air pressure, snowfall, ground subsidence or soil moisture content and lightning strike data; and The geographic information data includes the geographical distribution of power grid facilities, topographic data and land cover types.

10. An optimal maintenance location determination device for power grid fault prediction, It is characterized in that include: A data acquisition module, used to acquire historical fault data and relevant influencing factor data, wherein the relevant influencing factor data includes power grid operation data, meteorological data and geographic information data; An interpolation module, used for performing outlier interpolation processing on the historical fault data and the relevant influencing factor data through an adversarial network; A primary selection module, for selecting a plurality of key influencing factors from the relevant influencing factor data by fusing a Bayesian model with a hierarchical hybrid model and utilizing statistical information; A secondary selection module is used to use an improved anti-entropy-grey correlation analysis method to determine the correlation between the power grid fault and each key influencing factor, so as to select some key influencing factors from the multiple key influencing factors as influencing indicators according to the correlation; A fault prediction model is used to construct a BP neural network and use the historical fault data and the impact index to train to obtain a power grid fault prediction model, wherein the probability of geological disasters and the impact index are used as input parameters of the power grid fault prediction model, and a rainfall-induced geological disaster warning model is constructed to obtain the geological disaster probability; and The optimal location acquisition module is used to construct an optimal maintenance location optimization model based on the predicted power grid faults, and combine the objective function with the constraint conditions to obtain the optimal maintenance location.

Images